Course Hero has millions of student submitted documents similar to the one

below including study guides, practice problems, reference materials, practice exams, textbook help and tutor support.

Course Hero has millions of student submitted documents similar to the one below including study guides, practice problems, reference materials, practice exams, textbook help and tutor support.

page This intentionally left blank
Acquisitions Editor: Matt Goldstein
Project Editor: Maite Suarez-Rivas
Production Supervisor: Marilyn Lloyd
Marketing Manager: Michelle Brown
Marketing Coordinator: Jake Zavracky
Project Management: Windfall Software
Composition: Windfall Software, using ZzTEX
Copyeditor: Carol Leyba
Technical Illustration: Dartmouth Publishing
Proofreader: Jennifer McClain
Indexer: Ted Laux
Cover Design: Joyce 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 Westford
Access the latest information about Addison-Wesley titles 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, Eva Tardos.1st 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, Eva.
II. Title.
QA76.9.A43K54 2005
005.1dc22
2005000401
Copyright 2006 by Pearson Education, Inc.
For information on obtaining permission for use of material in this work, please
submit a written request to Pearson Education, Inc., Rights and Contract Department,
75 Arlington Street, Suite 300, Boston, MA 02116 or fax your request to (617) 848-7047.
All rights reserved. No part of this publication may be reproduced, stored in a
retrieval system, or transmitted, in any form or by any means, electronic, mechanical,
photocopying, recording, or any toher media embodiments now known or hereafter to
become known, without the prior written permission of the publisher. Printed in the
United States of America.
ISBN 0-321-29535-8
1 2 3 4 5 6 7 8 9 10-CRW-08 07 06 05
About the Authors
Jon Kleinberg is a professor of Computer Science at
Cornell University. He received his Ph.D. from M.I.T.
in 1996. He is the recipient of an NSF Career Award,
an ONR Young Investigator Award, an IBM Outstanding Innovation Award, the National Academy of Sciences Award for Initiatives in Research, research fellowships from the Packard and Sloan Foundations,
and teaching awards from the Cornell Engineering
College and Computer Science Department.
Kleinbergs research is centered around algorithms, particularly those concerned with the structure of networks and information, and with applications
to information science, optimization, data mining, and computational biology. His work on network analysis using hubs and authorities helped form the
foundation for the current generation of Internet search engines.
Eva Tardos is a professor of Computer Science at Cornell University. She received her Ph.D. from E tv s
oo
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 Guggenheim, 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-ow 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 selsh users.
This page intentionally left blank
Contents
About the Authors
Preface
v
xiii
1
Introduction: Some Representative Problems
1.1
A First Problem: Stable Matching 1
1.2
Five Representative Problems 12
Solved Exercises 19
Exercises 22
Notes and Further Reading 28
2
Basics of Algorithm Analysis
29
2.1
Computational Tractability 29
2.2
Asymptotic Order of Growth 35
2.3
Implementing the Stable Matching Algorithm Using Lists and
Arrays 42
2.4
A Survey of Common Running Times 47
2.5
A More Complex Data Structure: Priority Queues 57
Solved Exercises 65
Exercises 67
Notes and Further Reading 70
3
Graphs
73
3.1
Basic Denitions and Applications 73
3.2
Graph Connectivity and Graph Traversal 78
3.3
Implementing Graph Traversal Using Queues and Stacks 87
3.4
Testing Bipartiteness: An Application of Breadth-First Search 94
3.5
Connectivity in Directed Graphs 97
1
viii
Contents
3.6
4
5
6
Directed Acyclic Graphs and Topological Ordering
Solved Exercises 104
Exercises 107
Notes and Further Reading 112
99
Greedy Algorithms
115
4.1
Interval Scheduling: The Greedy Algorithm Stays Ahead 116
4.2
Scheduling to Minimize Lateness: An Exchange Argument 125
4.3
Optimal Caching: A More Complex Exchange Argument 131
4.4
Shortest Paths in a Graph 137
4.5
The Minimum Spanning Tree Problem 142
4.6
Implementing Kruskals Algorithm: The Union-Find Data
Structure 151
4.7
Clustering 157
4.8
Huffman Codes and Data Compression 161
4.9
Minimum-Cost Arborescences: A Multi-Phase Greedy
Algorithm 177
Solved Exercises 183
Exercises 188
Notes and Further Reading 205
Divide and Conquer
5.1
A First Recurrence: The Mergesort Algorithm
5.2
Further Recurrence Relations 214
5.3
Counting Inversions 221
5.4
Finding the Closest Pair of Points 225
5.5
Integer Multiplication 231
5.6
Convolutions and the Fast Fourier Transform
Solved Exercises 242
Exercises 246
Notes and Further Reading 249
209
210
234
Dynamic Programming
251
6.1
Weighted Interval Scheduling: A Recursive Procedure 252
6.2
Principles of Dynamic Programming: Memoization or Iteration
over Subproblems 258
6.3
Segmented Least Squares: Multi-way Choices 261
The star indicates an optional section. (See the Preface for more information about the relationships
among the chapters and sections.)
Contents
6.4
6.5
6.6
6.7
6.8
6.9
6.10
7
8
Subset Sums and Knapsacks: Adding a Variable 266
RNA Secondary Structure: Dynamic Programming over
Intervals 272
Sequence Alignment 278
Sequence Alignment in Linear Space via Divide and
Conquer 284
Shortest Paths in a Graph 290
Shortest Paths and Distance Vector Protocols 297
Negative Cycles in a Graph 301
Solved Exercises 307
Exercises 312
Notes and Further Reading 335
Network Flow
7.1
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 Preow-Push Maximum-Flow Algorithm 357
7.5
A First Application: The Bipartite Matching Problem 367
7.6
Disjoint Paths in Directed and Undirected Graphs 373
7.7
Extensions to the Maximum-Flow Problem 378
7.8
Survey Design 384
7.9
Airline Scheduling 387
7.10 Image Segmentation 391
7.11 Project Selection 396
7.12 Baseball Elimination 400
7.13
A Further Direction: Adding Costs to the Matching Problem
Solved Exercises 411
Exercises 415
Notes and Further Reading 448
NP and Computational Intractability
8.1
Polynomial-Time Reductions 452
8.2
Reductions via Gadgets: The Satisability Problem
8.3
Efcient Certication and the Denition of NP 463
8.4
NP-Complete Problems 466
8.5
Sequencing Problems 473
8.6
Partitioning Problems 481
8.7
Graph Coloring 485
337
404
451
459
ix
x
Contents
8.8
8.9
8.10
9
Numerical Problems 490
Co-NP and the Asymmetry of NP 495
A Partial Taxonomy of Hard Problems 497
Solved Exercises 500
Exercises 505
Notes and Further Reading 529
PSPACE: A Class of Problems beyond NP
9.1
PSPACE 531
9.2
Some Hard Problems in PSPACE 533
9.3
Solving Quantied Problems and Games in Polynomial
Space 536
9.4
Solving the Planning Problem in Polynomial Space 538
9.5
Proving Problems PSPACE-Complete 543
Solved Exercises 547
Exercises 550
Notes and Further Reading 551
10 Extending the Limits of Tractability
10.1
10.2
10.3
10.4
10.5
11.2
11.3
11.4
11.5
11.6
11.7
553
Finding Small Vertex Covers 554
Solving NP-Hard Problems on Trees 558
Coloring a Set of Circular Arcs 563
Tree Decompositions of Graphs 572
Constructing a Tree Decomposition 584
Solved Exercises 591
Exercises 594
Notes and Further Reading 598
11 Approximation Algorithms
11.1
531
599
Greedy Algorithms and Bounds on the Optimum: A Load
Balancing Problem 600
The Center Selection Problem 606
Set Cover: A General Greedy Heuristic 612
The Pricing Method: Vertex Cover 618
Maximization via the Pricing Method: The Disjoint Paths
Problem 624
Linear Programming and Rounding: An Application to Vertex
Cover 630
Load Balancing Revisited: A More Advanced LP Application 637
Contents
11.8
Arbitrarily Good Approximations: The Knapsack Problem
Solved Exercises 649
Exercises 651
Notes and Further Reading 659
12 Local Search
12.1
12.2
12.3
12.4
12.5
12.6
12.7
The Landscape of an Optimization Problem 662
The Metropolis Algorithm and Simulated Annealing 666
An Application of Local Search to Hopeld Neural Networks
Maximum-Cut Approximation via Local Search 676
Choosing a Neighbor Relation 679
Classication via Local Search 681
Best-Response Dynamics and Nash Equilibria 690
Solved Exercises 700
Exercises 702
Notes and Further Reading 705
13 Randomized Algorithms
644
661
671
707
13.1
13.2
13.3
13.4
13.5
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
13.11 Packet Routing 762
13.12 Background: Some Basic Probability Denitions 769
Solved Exercises 776
Exercises 782
Notes and Further Reading 793
Epilogue: Algorithms That Run Forever
795
References
805
Index
815
xi
This page intentionally left blank
Preface
Algorithmic ideas 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 deciencies 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 denitions. 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 reections 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 ow
problem. So was the former student, now a management consultant working
on stafng 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 eld 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-specic 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
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 efcient 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 efcient algorithms for the resulting
problems. Sophisticated algorithms are often best understood by reconstructing the sequence of ideasincluding false starts and dead endsthat 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 reects 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 rst algorithms course is not entirely
standard, we begin the book with self-contained coverage of topics that at
some institutions are familiar to students from CS1/CS2, but which at other
institutions are included in the syllabi of the rst 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
exibility in the prerequisite knowledge 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 elds. 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), articial
Preface
intelligence (planning, game playing, Hopeld networks), computer vision
(image segmentation), data mining (change-point detection, clustering), operations research (airline scheduling), and computational biology (sequence
alignment, RNA secondary structure).
The notion of computational intractability, and NP-completeness in particular, plays a large role in the book. This is consistent with how we think
about the overall process of algorithm design. Some of the time, an interesting problem arising in an application area will be amenable to an efcient
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 options 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 reects this theme. The discovery that a problem is NPcomplete should not be taken as the end of the story, but as an invitation to
begin looking for approximation algorithms, heuristic local search techniques,
or tractable special cases. We include extensive coverage of each of these three
approaches.
Problems and Solved Exercises
An important feature of the book is the collection of problems. Across all
chapters, the book includes over 200 problems, almost all 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 fully explained algorithm, an analysis of
the running 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 signicantly longer than what
would be needed simply to write a complete, correct solution (in other words,
xv
xvi
Preface
signicantly 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 specication 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 difculty, 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 difculty.
Second, aside from the lowest-numbered ones, the problems are designed to
involve some investment of time, both to relate the problem description to the
algorithmic techniques in the chapter, and then to actually design the necessary
algorithm. In our undergraduate class, we have tended to assign roughly three
of these problems per week.
Pedagogical Features and Supplements
In addition to the problems and solved exercises, the book has a number of
further pedagogical features, as well as additional supplements to facilitate its
use for teaching.
As noted earlier, a large number of the sections in the book are devoted
to the formulation of an algorithmic problemincluding its background and
underlying motivationand the design and analysis of an algorithm for this
problem. To reect 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 efciency. 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 all the problems, providing full 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
les are available at www.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 nal 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 1 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 ve representative problems
that foreshadow topics from the remainder of the course. These ve problems
are interrelated in the sense that they are all variations and/or special cases
of the Independent Set Problem; but one is solvable by a greedy algorithm,
one by dynamic programming, one by network ow, 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 ve 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 denitions and
notations used for analyzing algorithms, as well 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 efciency. It then discusses growth rates of functions and asymptotic analysis more formally, and offers a guide to commonly
occurring functions in algorithm analysis, together with standard applications
in which they arise. Chapter 3 covers the basic denitions 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 denitions, graph traversal techniques such
as breadth-rst search and depth-rst search, and directed graph concepts
including strong connectivity and topological ordering.
xvii
xviii
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 here will be familiar to students from the CS1/CS2 sequence,
our focus is on these data structures in the broader context of algorithm design
and analysis.
Chapters 4 through 7 cover four major algorithm design techniques: greedy
algorithms, divide and conquer, dynamic programming, and network ow.
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 the design of algorithms that
improve over straightforward approaches to a number of basic problems, including the comparison of rankings, the computation of closest 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 ow problems, devoting much of our
focus in this chapter to discussing a large array of different ow applications.
To the extent that network ow 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 justice 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 difcult
reduction. We also consider types of computational hardness beyond NPcompleteness, particularly through the topic of PSPACE-completeness. We
Preface
nd this is a valuable way to emphasize that intractability doesnt end at
NP-completeness, and PSPACE-completeness also forms the underpinning for
some central notions from articial intelligenceplanning and game playing
that would otherwise not nd a place in the algorithmic landscape we are
surveying.
Chapters 10 through 12 cover three major techniques for dealing with computationally intractable problems: identication 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 efcient algorithm. We illustrate how NP-complete problems are often efciently 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 nd 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, and a third method we refer to as pricing, which
incorporates ideas from each of the rst 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 rst undergraduate course on
algorithms, but it can also be used as the basis for an introductory graduate
course.
When we use the book at the undergraduate level, we spend roughly
one lecture per numbered section; in cases where there is more than one
xix
xx
Preface
lectures worth of material in a section (for example, when a section provides
further applications as additional examples), we treat this extra material as a
supplement that students can 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 rst half of
the book (for example, we tend to skip Sections 4.3, 4.74.8, 5.55.6, 6.5, 7.6,
and 7.11). We cover roughly half of each of Chapters 1113.
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 rst 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 specic course to its prerequisites.
The resulting syllabus looks roughly as follows: Chapter 1; Chapters 48
(excluding 4.3, 4.74.9, 5.55.6, 6.5, 6.10, 7.4, 7.6, 7.11, and 7.13); Chapter 9
(briey); Chapter 10, Sections.10.1 and 10.2; Chapter 11, Sections 11.1, 11.2,
11.6, and 11.8; Chapter 12, Sections 12.112.3; 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 nd the emphasis on formulating problems to be
useful as well, since students will soon be trying to dene their own research
problems in many different subelds. For this type of course, we cover the
later topics in Chapters 4 and 6 (Sections 4.54.9 and 6.56.10), 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 1013. Although our focus in an introductory graduate course is
on the more advanced sections, we nd it useful for the students to have the
full book to consult for reviewing or lling in background knowledge, given
the range of different undergraduate backgrounds among the students in such
a course.
Finally, the book can be used to support self-study by graduate students,
researchers, or computer professionals who want to get a sense for how they
Preface
might be able to use particular algorithm design techniques in the context of
their own work. A number of graduate students and colleagues have used
portions of the book in this way.
Acknowledgments
This book grew out of the sequence of algorithms courses that we have taught
at Cornell. These courses have grown, as the eld has grown, over a number of
years, and they reect the inuence of the Cornell 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 all our colleagues at Cornell for countless discussions both
on the material here and on broader issues about the nature of the eld.
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 Evmievski, Ariful Gani, 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 l, Jon
a
Peress, Matt Piotrowski, Joe Polastre, Mike Priscott, Xin Qi, Venu Ramasubramanian, Aditya Rao, David Richardson, Brian Sabino, Rachit Siamwalla, Sebastian Silgardo, 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 beneted
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 still 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
reects the infuence of specic colleagues. Many of these contributions have
undoubtedly escaped our notice, but we especially thank Yuri Boykov, Ron
Elber, Dan Huttenlocher, Bobby Kleinberg, Evie Kleinberg, Lillian Lee, David
McAllester, Mark Newman, Prabhakar Raghavan, Bart Selman, David Shmoys,
Steve 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 project. We further 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 gures, Ted Laux for
the indexing, and Carol Leyba and Jennifer McClain for the copyediting 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 nal version of the text.
Finally, we thank our familiesLillian 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, briey to pass
through a place traditionally occupied by celebrities and other inhabitants of
the pop-cultural rmament. (It was probably just 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 eld 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 nd this book an enjoyable and useful guide wherever your computational
pursuits may take you.
Jon Kleinberg
Eva Tardos
Ithaca, 2005
xxiii
This page intentionally left blank
Chapter 1
Introduction: Some
Representative Problems
1.1 A First Problem: Stable Matching
As an opening topic, we look at an algorithmic problem that nicely illustrates
many of the themes we will be emphasizing. It is motivated by some very
natural and practical concerns, and from these we formulate a clean and
simple statement of a problem. The algorithm to solve the problem is very
clean as well, and most of our work will be spent in proving that it is correct
and giving an acceptable bound on the amount of time it takes to terminate
with an answer. The problem itselfthe Stable Matching Problemhas 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 rst think informally about the kind of situation
that might arise as a group of friends, all juniors 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 companyonce
the applications come informs 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.
2
Chapter 1 Introduction: Some Representative Problems
Gale and Shapley considered the sorts of things that could start going
wrong with this process, in the absence of any mechanism to enforce the status
quo. Suppose, for example, that your friend Raj has just 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 nal decisions, calls up Raj and offers him a summer job as well. Now, Raj
actually prefers WebExodus to CluNetwon over perhaps by the laid-back,
anything-can-happen atmosphereand 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 nd 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 nd 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 employerscan 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-enforcingif 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 stablethere
are no further outside deals that can be made.
So this is the question Gale and Shapley asked: Given a set of preferences
among employers and applicants, can we assign applicants to employers so
that for every employer E , and every applicant A who is not scheduled to work
for E , at least one of the following two things is the case?
1.1 A First Problem: Stable Matching
(i) E prefers every one of its accepted applicants to A; or
(ii) A prefers her current situation over working for employer E .
If this holds, the outcome is stable: individual self-interest will prevent any
applicant/employer deal from being made behind the scenes.
Gale and Shapley proceeded to develop a striking algorithmic solution to
this problem, which we will discuss presently. Before doing this, lets note that
this is not the only origin of the Stable Matching Problem. It turns out that for
a decade before the work of Gale and Shapley, unbeknownst to them, the
National Resident Matching Program had been using a very similar procedure,
with the same underlying motivation, to match residents to hospitals. Indeed,
this system, with relatively little change, is still in use today.
This is one testament to the problems fundamental appeal. And from the
point of view of this book, it provides us with a nice rst domain in which
to reason about some basic combinatorial denitions and the algorithms that
build on them.
Formulating the Problem To get at the essence of this concept, it helps to
make the problem as clean as possible. The world of companies and applicants
contains some distracting asymmetries. Each applicant is looking for a single
company, but each company is looking for many applicants; moreover, there
may be more (or, as is sometimes the case, fewer) applicants than there are
available slots for summer jobs. Finally, each applicant does not typically 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 simplied 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 gendersthe applicants and the companiesand 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 dierent at a
technical level. Given the applicant-employer application were considering here, well be focusing
on the version with two genders.
3
4
Chapter 1 Introduction: Some Representative Problems
So consider a set M = {m1, . . . , mn } of n men, and a set W = {w1, . . . , wn }
of n women. Let M W denote the set of all possible ordered pairs of the form
(m , w), where m M and w W . A matching S is a set of ordered pairs, each
from M 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 .
An instability: m and w
each prefer the other to
their current partners.
m
w
m
w
Figure 1.1 Perfect matching
S with instability (m , w ).
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 personthere 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 w to w if m ranks w higher
than w . We will refer to the ordered ranking of m as his preference list . We will
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 , w) and (m , w ) in S (as
depicted in Figure 1.1) with the property that m prefers w to w, and w prefers
m to m . In this case, theres nothing to stop m and w from abandoning their
current partners and heading off together; the set of marriages is not selfenforcing. Well say that such a pair (m , w ) is an instability with respect to S:
(m , w ) does not belong to S, but each of m and w 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 efciently construct a stable
matching if there is one?
Some Examples To illustrate these denitions, consider the following two
very simple instances of the Stable Matching Problem.
First, suppose we have a set of two men, {m , m }, and a set of two women,
{w , w }. The preference lists are as follows:
m prefers w to w .
m prefers w to w .
1.1 A First Problem: Stable Matching
w prefers m to m .
w 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 , w) and (m , w ). The other perfect matching, consisting of the
pairs (m , w) and (m , w ), would not be a stable matching, because the pair
(m , w) would form an instability with respect to this matching. (Both m and
w 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 w to w .
m prefers w to w.
w prefers m to m.
w prefers m to m .
Whats going on in this case? The two mens preferences mesh perfectly with
each other (they rank different women rst), 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 forwardits 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 efcient
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 w who ranks highest on his preference list and proposes to
her. Can we declare immediately that (m , w) will be one of the pairs in our
nal stable matching? Not necessarily: at some point in the future, a man
m whom w prefers may propose to her. On the other hand, it would be
5
6
Chapter 1 Introduction: Some Representative Problems
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 stateengagement .
Woman w will become
engaged to m if she
prefers him to m .
.
m
w
m
Figure 1.2 An intermediate
state of the G-S algorithm
when a free man m is proposing to a woman w.
Suppose we are now at a state in which some men and women are free
not engagedand some are engaged. The next step could look like this.
An arbitrary free 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 free.
.
Finally, the algorithm will terminate when no one is free; at this moment,
all engagements are declared nal, and the resulting perfect matching is
returned.
Here is a concrete description of the Gale-Shapley algorithm, with Figure 1.2 depicting a state of the algorithm.
Initially all m M and w 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 w is free then
(m , w) become engaged
Else w is currently engaged to m
If w prefers m to m then
m remains free
Else w prefers m to m
(m , w) become engaged
m becomes free
Endif
Endif
Endwhile
Return the set S of engaged pairs
An intriguing thing is that, although the G-S algorithm is quite simple
to state, it is not immediately obvious that it returns a stable matching, or
even a perfect matching. We proceed to prove this now, through a sequence
of intermediate facts.
1.1 A First Problem: Stable Matching
Analyzing the Algorithm
First consider the view of a woman w during the execution of the algorithm.
For a while, no one has proposed to her, and she is free. Then a man m may
propose to her, and she becomes engaged. As time goes on, she may receive
additional proposals, accepting those that increase the rank of her partner. So
we discover the following.
(1.1) w remains engaged from the point at which she receives her rst
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)
loop.
The G-S algorithm terminates after at most n2 iterations of the While
Proof. A useful strategy for upper-bounding the running time of an algorithm,
as we are trying to do here, is to nd 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
7
8
Chapter 1 Introduction: Some Representative Problems
iteration. For example, the number of free individuals could remain constant
from one iteration to the next, as could the number of engaged pairs. Thus,
these quantities could not be used directly in giving an upper bound on the
maximum possible number of iterations, in the style of the previous paragraph.
Let us now establish that the set S returned at the termination of the
algorithm is in fact a perfect matching. Why is this not immediately obvious?
Essentially, we have to show that no man can fall off the end of his preference
list; the only way for the While loop to exit is for there to be no free 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 free 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.
(1.5)
The set S returned at termination is a perfect matching.
Proof. The set of engaged pairs always forms a matching. Let us suppose that
the algorithm terminates with a free man m. At termination, it must be the
case that m had already proposed to every woman, for otherwise the While
loop would not have exited. But this contradicts (1.4), which says that there
cannot be a free man who has proposed to every woman.
Finally, we prove the main property of the algorithmnamely, that it
results in a stable matching.
(1.6) Consider an execution of 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 dened earlier, such an instability
would involve two pairs, (m , w) and (m , w ), in S with the properties that
.
m prefers w to w, and
.
w prefers m to m .
In the execution of the algorithm that produced S, ms last proposal was, by
denition, to w. Now we ask: Did m propose to w at some earlier point in
1.1 A First Problem: Stable Matching
this execution? If he didnt, then w must occur higher on ms preference list
than w , contradicting 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 nal partner of w , so either m = m or, by (1.1), w prefers her nal
partner m to m ; either way this contradicts our assumption that w prefers
m to m .
It follows that S is a stable matching.
Extensions
We began by dening 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 with, 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:
m prefers w to w .
m prefers w to w.
w prefers m to m.
w prefers m to m .
Now, in any execution of the Gale-Shapley algorithm, m will become engaged
to w, m will become engaged to w (perhaps in the other order), and things
will stop there. Thus, the other stable matching, consisting of the pairs (m , w)
and (m , w ), is not attainable from an execution of the G-S algorithm in which
the men propose. On the other hand, it would be reached if we ran a version of
the algorithm in which the women propose. And in larger examples, with more
than two people on each side, we can have an even larger collection of possible
stable matchings, many of them not achievable by any natural algorithm.
This example shows a certain unfairness in the G-S algorithm, favoring
men. If the mens preferences mesh perfectly (they all list different women as
their rst choice), then in all runs of the G-S algorithm all men end up matched
with their rst 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.
9
10
Chapter 1 Introduction: Some Representative Problems
To begin with, our example reinforces the point that the G-S algorithm
is actually underspecied: as long as there is a free man, we are allowed to
choose any free man to make the next proposal. Different choices specify
different executions of the algorithm; 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 nal 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 all 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 w is a valid partner
of a man m if there is a stable matching that contains the pair (m , w). We will
say that w is the best valid partner of m if w is a valid partner of m, and no
woman whom m ranks higher than w 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 G-S algorithm results in the set S.
This statement is surprising at a number of levels. First of all, as dened,
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 nally, it answers
1.1 A First Problem: Stable Matching
our question above by showing that the order of proposals in the G-S algorithm
has absolutely no effect on the nal outcome.
Despite all this, the proof is not so difcult.
Proof. Let us suppose, by way of contradiction, that some execution E 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 E of the algorithm. So consider the rst moment during
the execution E in which some man, say m, is rejected by a valid partner w.
Again, since men propose in decreasing order of preference, and since this is
the rst time such a rejection has occurred, it must be that w is ms best valid
partner best (m).
The rejection of m by w may have happened either because m proposed
and was turned down in favor of ws existing engagement, or because w broke
her engagement to m in favor of a better proposal. But either way, at this
moment w forms or continues an engagement with a man m whom she prefers
to m.
Since w is a valid partner of m, there exists a stable matching S containing
the pair (m , w). Now we ask: Who is m paired with in this matching? Suppose
it is a woman w = w.
Since the rejection of m by w was the rst rejection of a man by a valid
partner in the execution E , it must be that m had not been rejected by any valid
partner at the point in E when he became engaged to w. Since he proposed in
decreasing order of preference, and since w is clearly a valid partner of m , it
must be that m prefers w to w . But we have already seen that w prefers m
to m, for in execution E she rejected m in favor of m . Since (m , w) S , it
follows that (m , w) 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
worst valid partner of w if m is a valid partner of w, and no man whom w
ranks lower than m is a valid partner of hers.
(1.8) In the stable matching S, each woman is paired with her worst valid
partner.
Proof. Suppose there were a pair (m , w) in S such that m is not the worst
valid partner of w. Then there is a stable matching S in which w is paired
11
12
Chapter 1 Introduction: Some Representative Problems
with a man m whom she likes less than m. In S , m is paired with a woman
w = w; since w is the best valid partner of m, and w is a valid partner of m,
we see that m prefers w to w .
But from this it follows that (m , w) is an instability in S , contradicting the
claim that S is stable and hence contradicting our initial assumption.
Thus, we nd that our simple example above, in which the mens preferences clashed with the womens, hinted at a very general phenomenon: for
any input, the side that does the proposing in the G-S algorithm ends up with
the best possible stable matching (from their perspective), while the side that
does not do the proposing correspondingly ends up with the worst possible
stable matching.
1.2 Five Representative Problems
The Stable Matching Problem provides us with a rich example of the process of
algorithm design. For many problems, this process involves a few signicant
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 efciency.
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 identiable
genres and appreciating how subtle changes in the statement of a problem can
have an enormous effect on its computational difculty.
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 difculty and in the kinds of approaches that one brings
to bear on them. The rst three will be solvable efciently 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 efcient algorithm; and the fth hints at a class of problems
believed to be harder still.
The problems are self-contained and are all motivated by computing
applications. To talk about some of them, though, it will help to use the
terminology of graphs. While graphs are a common topic in earlier computer
13
1.2 Five Representative Problems
science courses, well be introducing them in a fair amount of depth in
Chapter 3; due to their enormous expressive power, well also be using them
extensively throughout the book. For the discussion here, its enough to think
of a graph G as simply a way of encoding pairwise relationships among a set
of objects. Thus, G consists of a pair of sets (V , E )a collection V of nodes
and a collection E of edges, each of which joins two of the nodes. We thus
represent an edge e E as a two-element subset of V : e = {u , v} for some
u , v V , where we call u and v the ends of e . We typically 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.
(a)
Lets now turn to a discussion of the ve representative problems.
Interval Scheduling
Consider the following very simple scheduling problem. You have a resource
it may be a lecture room, a supercomputer, or an electron microscopeand
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 all 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 nish time fi . Naturally, we have si < fi for all
i. Two requests i and j are compatible if the requested intervals do not overlap:
that is, either request i is for an earlier time interval than request j (fi sj ),
or request i is for a later time than request j (fj 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 Interval Scheduling Problem in Figure 1.4.
Note that there is a single compatible set of size 4, and this is the largest
compatible set.
Figure 1.4 An instance of the Interval Scheduling Problem.
(b)
Figure 1.3 Each of (a) and
(b) depicts a graph on four
nodes.
14
Chapter 1 Introduction: Some Representative Problems
We will see shortly that this problem can be solved by a very natural
algorithm that orders the set of requests according to a certain heuristic and
then greedily processes them in one pass, selecting as large a compatible
subset as it can. This will be typical of a class of greedy algorithms that we
will consider for various problemsmyopic rules that process the input one
piece at a time with no apparent look-ahead. When a greedy algorithm can be
shown to nd an optimal solution for all instances of a problem, its often fairly
surprising. We typically learn something about the structure of the underlying
problem from the fact that such a simple approach can be optimal.
Weighted Interval Scheduling
In the Interval Scheduling Problem, we sought 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 > 0; 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 nd a
compatible subset of intervals of maximum total value.
The case in which vi = 1 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 conguration of the full 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 efcient algorithm.
Bipartite Matching
When we considered the Stable Matching Problem, we dened 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 dened
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 dene the notion of a bipartite graph. We say that
a graph G = (V , E ) is bipartite if its node set V can be partitioned into sets X
15
1.2 Five Representative Problems
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 nding 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 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 .
To see that this does capture the same notion we encountered in the Stable
Matching Problem, consider a bipartite graph G with a set X of n men, a set Y
of n women, and an edge from every node in X to every node in Y . Then the
matchings and perfect matchings in G are precisely the matchings and perfect
matchings among the set of men and women.
In the Stable Matching Problem, we added preferences to this picture. Here,
we do not consider preferences; but the nature of the problem in arbitrary
bipartite graphs adds a different source of complexity: there is not necessarily
an edge from every x X to every y Y , so the set of possible matchings has
quite a complicated structure. In other words, it is as though only certain pairs
of men and women are willing to be paired off, and we want to gure 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 (xi , yj ) can indicate that
machine yj 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 xi is capable of teaching
course yj . A perfect matching in this graph consists of an assignment of each
professor to a course that he or she can teach, in such a way that every course
is covered.
Thus the Bipartite Matching Problem is the following: Given an arbitrary
bipartite graph G , nd a matching of maximum size. If |X | = |Y | = n, then there
is a perfect matching if and only if the maximum matching has size n. We will
nd that the algorithmic techniques discussed earlier do not seem adequate
x1
y1
x2
y2
x3
y3
x4
y4
x5
y5
Figure 1.5 A bipartite graph.
16
Chapter 1 Introduction: Some Representative Problems
for providing an efcient algorithm for this problem. There is, however, a very
elegant and efcient algorithm to nd a maximum matching; it inductively
builds up larger and larger matchings, selectively backtracking along the way.
This process is called augmentation, and it forms the central component in a
large class of efciently solvable problems called network ow problems.
1
2
Independent Set
3
4
6
5
7
Figure 1.6 A graph whose
largest independent set has
size 4.
Now lets talk about an extremely general problem, which includes most of
these earlier problems as special cases. Given a graph G = (V , E ), we say
a set of nodes S V is independent if no two nodes in S are joined by an
edge. The Independent Set Problem is, then, the following: Given G , nd 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
conicts 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 conicting pair.
Interval Scheduling and Bipartite Matching can both be encoded as special
cases of the Independent Set Problem. For Interval Scheduling, dene 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 conicts arise between two edges
that share an end. (These, indeed, are the pairs of edges that cannot belong
to a common matching.) So we dene a graph G = (V , E ) in which the node
set V is equal to the edge set E of G . We dene 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 conict graph in an instance of
1.2 Five Representative Problems
Given the generality of the Independent Set Problem, an efcient algorithm
to solve it would be quite impressive. It would have to implicitly contain
algorithms for Interval Scheduling, Bipartite Matching, and a host of other
natural optimization problems.
The current status of Independent Set is this: no efcient 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-complete . No efcient 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 difculty between checking that something is a large independent
set and actually nding a large independent set. This may look like a very basic
observationand it isbut it turns out to be crucial in understanding this class
of problems. Furthermore, as well see next, its possible for a problem to be
so hard that there isnt even an easy way to check solutions in this sense.
Competitive Facility Location
Finally, we come to our fth problem, which is based on the following twoplayer game. Consider two large companies that operate caf franchises across
e
the countrylets call them JavaPlanet and Queequegs Coffeeand 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 conict graph in an instance
of Bipartite Matching.
17
18
Chapter 1 Introduction: Some Representative Problems
10
1
5
15
5
1
5
1
15
10
Figure 1.7 An instance of the Competitive Facility 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 , j ) 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 conicts via a graph
G = (V , E ), where V is the set of zones, and (i , j ) is an edge in E if the
zones i and j are adjacent. 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 rst. 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 gure for a while; but it requires
some amount of case-checking of the form, If P1 goes here, then P2 will go
there; but if P1 goes over there, then P2 will go here. . . . And this appears to
be intrinsic to the problem: not only is it computationally difcult 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
nding 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 articial intelligence.
Solved Exercises
Solved Exercises
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
rst k entries are the good people (of the opposite gender) in some order, and
its next n k are the bad people (of the opposite gender) in some order.
Show that in every stable matching, every good man is married to a good
woman.
Solution A natural way to get started thinking about this problem is to
assume the claim is false and try to work toward obtaining a contradiction.
What would it mean for the claim to be false? There would exist some stable
matching M in which a good man m was married to a bad woman w.
Now, lets consider what the other pairs in M look like. There are k good
men and k good women. Could it be the case that every good woman is married
to a good man in this matching M ? No: one of the good men (namely, m) is
already married to a bad woman, and that leaves only k 1 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
qualications or certications, and so they cannot be employed at certain
companies, however desirable they may seem. Using the analogy to marriage
between men and women, we have a set M of n men, a set W of n women,
19
20
Chapter 1 Introduction: Some Representative Problems
and a set F M W of pairs who are simply not allowed to get married. Each
man m ranks all the women w for which (m , w) F , and each woman w ranks
all 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) 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) 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 denitions, 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 nd 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 difculty, of course, is
that the G-S algorithm doesnt know anything about forbidden pairs, and so
the condition in the While 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 While 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 such a man m
Let w be the highest-ranked woman in ms preference list
to which m has not yet proposed
If w is free then
(m , w) become engaged
Else w is currently engaged to m
If w prefers m to m then
m remains free
Else w prefers m to m
(m , w) become engaged
m becomes free
Endif
Endif
Endwhile
Return the set S of engaged pairs
We now prove that this yields a stable matching, under our new denition
of stability.
To begin with, facts (1.1), (1.2), and (1.3) 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 part of a pair in S, then m must have proposed to every
nonforbidden woman; and if w is a woman who is not part of a pair in S, then
it must be that no man ever proposed to w.
Finally, we need only show
(1.9)
There is no instability with respect to the returned matching S.
21
22
Chapter 1 Introduction: Some Representative Problems
Proof. Our general denition 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 m,
and thus she prefers her nal partner to ma 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 nal partner to m a
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 m 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 w a
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 m to be single, he must have proposed to every
nonforbidden woman; in particular, he must have proposed to w, which means
she would no longer be singlea contradiction.
Exercises
1. 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 rst on the
preference list of w and w is ranked rst on the preference list of m.
2. 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? Consider an instance of the Stable Matching Problem in which
there exists a man m and a woman w such that m is ranked rst on the
preference list of w and w is ranked rst 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, involving 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 slotso 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 A 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.
4. 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
Chapter 1 Introduction: Some Representative Problems
The interest, naturally, was in finding a way of assigning each student
to at most one hospital, in such a way that all available positions in all
hospitals were filled. (Since we are assuming a surplus of students, there
would be some students who do not get assigned to any hospital.)
We say that an assignment of students to hospitals is stable if 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
h and h , so that
s is assigned to h, and
s is assigned to h , and
h prefers s to s, and
s prefers h 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.
5. 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 m1 is ranked in first
place; second place is a tie between m2 and m3 (she has no preference
between them); and m4 is in last place. We will say that w prefers m to m
if m is ranked higher than m on her preference list (they are not tied).
With indifferences in the rankings, 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 w, such that each of m and w prefers the other to their
partner in S. Does there always exist a perfect matching with no
Exercises
strong instability? Either give an example of a set of men and women
with preference lists for which every perfect matching has a strong
instability; or give an algorithm that is guaranteed to find a perfect
matching with no strong instability.
(b) A weak instability in a perfect matching S consists of a man m and
a woman w, such that their partners in S are w and m , respectively,
and one of the following holds:
m prefers w to w , and w either prefers m to m or is indifferent
between these two choices; or
w prefers m to m , and m either prefers w to w or is indifferent
between these two choices.
In other words, the pairing between m and w 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:
() 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 Si , 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 Si will not visit the remaining ports on
its schedule (if any) that month, but this is okay. So the truncation of
Si 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 () 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
Chapter 1 Introduction: Some Representative Problems
Example. Suppose we have two ships and two ports, and the month has
four days. Suppose the first ships schedule is
port P1; at sea; port P2; at sea
and the second ships schedule is
at sea; port P1; at sea; port P2
Then the (only) way to choose truncations would be to have the first ship
remain in port P2 starting on day 3, and have the second ship remain in
port P1 starting on day 2.
7. 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 n output wires , each
directed from a source to a terminus . Each input wire meets each output
wire in exactly one distinct point, at a special piece of hardware called
a junction box . Points on the wire 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 wires 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 1.8 gives an
example of such a collection of input and output wires.
Now, heres the switching component of this situation. Each input
wire is carrying a distinct data stream, and this data stream must be
switched onto one of the output wires. 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
which input data stream gets switched onto which output wire, but
each input data stream must be switched onto a different output wire.
Furthermoreand this is the tricky constraintno 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
foundone 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.
Exercises
Junction
Junction
Junction
Junction
Output 1
(meets Input 2
before Input 1)
Output 2
(meets Input 2
before Input 1)
Input 1
Input 2
(meets Output 2 (meets Output 1
before 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 2and this is not
allowed.
8. 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 m on her list of preferences (i.e.,
by falsely claiming that she prefers m to m) and running the algorithm
with this false preference list, w will end up with a man m that she truly
prefers to both m and m ? (We can ask the same question for men, but
will focus on the case of women for purposes of this question.)
Resolve this question by doing one of the following two things:
(a) Give a proof that, for any set of preference lists, switching the
order of a pair on the list cannot improve a womans partner in the GaleShapley algorithm; or
27
28
Chapter 1 Introduction: Some Representative Problems
(b) Give an example of a set of preference lists for which there is
a switch that would improve the partner of a woman who switched
preferences.
Notes and Further Reading
The Stable Matching Problem was rst dened 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 New 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 Guseld and Irving
(1989) and Knuth (1997c). Guseld and Irving also provide a nice survey of
the parallel 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 ve representative problems will be
central to the books discussions, respectively, of greedy algorithms, dynamic
programming, network ow, NP-completeness, and PSPACE-completeness.
We will discuss the problems in these contexts later in the book.
Chapter 2
Basics of Algorithm Analysis
Analyzing algorithms involves thinking about how their resource requirementsthe amount of time and space they usewill scale with increasing
input size. We begin this chapter by talking about how to put this notion on a
concrete footing, as making it concrete opens the door to a rich understanding
of computational tractability. Having done this, we develop the mathematical
machinery needed to talk about the way in which different functions scale
with increasing input size, making precise what it means for one function to
grow faster than another.
We then develop running-time bounds for some basic algorithms, beginning with an implementation of the Gale-Shapley algorithm from Chapter 1
and continuing to a survey of many different running times and certain characteristic types of algorithms that achieve these running times. In some cases,
obtaining a good running-time bound relies on the use of more sophisticated
data structures, and we conclude this chapter with a very useful example of
such a data structure: priority queues and their implementation using heaps.
2.1 Computational Tractability
A major focus of this book is to nd efcient algorithms for computational
problems. At this level of generality, our topic seems to encompass the whole
of computer science; so what is specic to our approach here?
First, we will try 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 efcient algorithms. At the same time, it would be
pointless to pursue these design principles in a vacuum, so the problems and
30
Chapter 2 Basics of Algorithm Analysis
approaches we consider are drawn from fundamental issues that arise throughout computer science, and a general study of algorithms turns out to serve as
a nice survey of computational 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 efciently nd a solution that satises certain clearly
delineated conditions.
As we seek to understand the general notion of computational efciency,
we will focus primarily on efciency in running time: we want algorithms that
run quickly. But it is important that algorithms be efcient in their use of other
resources as well. In particular, the amount of space (or memory) used by an
algorithm is an issue that will also arise at a number of points in the book, and
we will see techniques for reducing the amount of space needed to perform a
computation.
Some Initial Attempts at Dening Efciency
The rst major question we need to answer is the following: How should we
turn the fuzzy notion of an efcient algorithm into something more concrete?
A rst attempt at a working denition of efciency is the following.
Proposed Denition of Efciency (1): An algorithm is efcient if, when
implemented, it runs quickly on real input instances.
Lets spend a little time considering this denition. At a certain level, 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 signicant area of research
devoted to the careful implementation and proling of different algorithms for
discrete computational problems.
But there are some crucial things missing from this denition, even if our
main goal is to solve real problem instances quickly on real computers. The
rst 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 full range of
input instances that will be encountered in practice, and some input instances
can be much harder than others. Finally, this proposed denition 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 still run quickly while the other consumes a huge amount of time.
2.1 Computational Tractability
So what we could ask for is a concrete denition of efciency that is
platform-independent, instance-independent, and of predictive value with
respect to increasing input sizes. Before focusing on any specic 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 will 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 ne-grained details of how the data is represented. In considering the
problem, we will seek to describe an algorithm at a high level, and then analyze
its running time mathematically as a function of this input size N .
Worst-Case Running Times and Brute-Force Search
To begin with, we will focus on analyzing the worst-case running time: we will
look for a bound on the largest possible running time the algorithm could have
over all inputs of a given size N , and see how this scales with N . The focus on
worst-case performance initially seems quite draconian: what if an algorithm
performs well on most instances and just has a few pathological inputs on
which it is very slow? This certainly is an issue in some cases, but in general
the worst-case analysis of an algorithm has been found to do a reasonable job
of capturing its efciency in practice. Moreover, once we have decided to go
the route of mathematical analysis, it is hard to nd an effective alternative to
worst-case analysis. Average-case analysisthe obvious appealing alternative,
in which one studies the performance of an algorithm averaged over random
instancescan 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 rst simple guide
31
32
Chapter 2 Basics of Algorithm Analysis
is by comparison with brute-force search over the search space of possible
solutions.
Lets return to the example of the Stable Matching Problem. Even when
the size of a Stable Matching input instance is relatively small, the search
space it denes is enormous (there are n! possible perfect matchings between
n men and n women), and we need to nd 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 nding a stable matching from
among 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 denition of efciency.
Proposed Denition of Efciency (2): An algorithm is efcient if it achieves
qualitatively better worst-case performance, at an analytical level, than
brute-force search.
This will turn out to be a very useful working denition 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 denition, it is vagueness. What do we mean by qualitatively better performance? This suggests
that we consider the actual running time of algorithms more carefully, and try
to quantify what a reasonable running time would be.
Polynomial Time as a Denition of Efciency
When people rst began analyzing discrete algorithms mathematicallya
thread of research that began gathering momentum through the 1960s
2.1 Computational Tractability
a consensus began to emerge on how to quantify the notion of a reasonable
running time. Search spaces for natural combinatorial problems tend to grow
exponentially in the size N of the input; if the input size increases by one, the
number of possibilities increases multiplicatively. Wed like a good algorithm
for such a problem to have a better scaling property: when the input size
increases by a constant factorsay, a factor of 2the algorithm should only
slow down by some constant factor C .
Arithmetically, we can formulate this scaling behavior as follows. Suppose
an algorithm has the following property: There are absolute constants c > 0
and d > 0 so that on every input instance of size N , its running time is
bounded by cN d primitive computational steps. (In other words, its running
time is at most proportional to N d .) 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 cN d to c (2N )d = c 2d N d , which is a slow-down by a factor of 2d . Since d is
a constant, so is 2d ; 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 denition of efciency.
Proposed Denition of Efciency (3): An algorithm is efcient if it has a
polynomial running time.
Where our previous denition seemed overly vague, this one seems much
too prescriptive. Wouldnt an algorithm with running time proportional to
n100and hence polynomialbe hopelessly inefcient? Wouldnt we be relatively pleased with a nonpolynomial running time of n1+ .02(log n)? The answers are, of course, yes and yes. And indeed, however much one may
try to abstractly motivate the denition of efciency in terms of polynomial
time, a primary justication 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 difcult in practice. There are certainly
exceptions to this principle in both directions: there are cases, for example, in
33
34
Chapter 2 Basics of Algorithm Analysis
Table 2.1 The running times (rounded up) of different algorithms on inputs of
increasing size, for a processor performing a million high-level instructions per second.
In cases where the running time exceeds 1025 years, we simply record the algorithm as
taking a very long time.
n
n log2 n
n2
n3
1.5n
n = 10
< 1 sec
< 1 sec
n = 30
< 1 sec
n = 50
< 1 sec
n = 100
< 1 sec
2n
< 1 sec
< 1 sec
< 1 sec
< 1 sec
< 1 sec
< 1 sec
< 1 sec
18 min
< 1 sec
< 1 sec
< 1 sec
11 min
36 years
very long
< 1 sec
< 1 sec
1 sec
12,892 years
1017 years
very long
n!
< 1 sec
4 sec
1025
years
n = 1,000
< 1 sec
< 1 sec
1 sec
18 min
very long
very long
very long
n = 10,000
< 1 sec
< 1 sec
2 min
12 days
very long
very long
very long
n = 100,000
< 1 sec
2 sec
3 hours
32 years
very long
very long
very long
1 sec
20 sec
12 days
31,710 years
very long
very long
very long
n = 1,000,000
which an algorithm with exponential worst-case behavior generally runs well
on the kinds of instances that arise in practice; and there are also cases where
the best polynomial-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 denition of polynomial time has turned out to correspond surprisingly well in practice to what we observe about the efciency of
algorithms, and the tractability of problems, in real 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, 30, 50, 100, 1,000, 10,000, 100,000, and
1,000,000.
There is a nal, fundamental benet to making our denition of efciency
so specic: it becomes negatable. It becomes possible to express the notion
that there is no efcient 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 efcient algorithms as a well-dened question. In contrast, both of our
2.2 Asymptotic Order of Growth
previous denitions were completely subjective, and hence limited the extent
to which we could discuss certain issues in concrete terms.
In particular, the rst of our denitions, which was tied to the specic
implementation of an algorithm, turned efciency into a moving target: as
processor speeds increase, more and more algorithms fall under this notion of
efciency. Our denition 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 efcient solutions, and others
do not.
2.2 Asymptotic Order of Growth
Our discussion of computational tractability has turned out to be intrinsically
based on our ability to express the notion that an algorithms worst-case
running time on inputs of size n grows at a rate that is at most proportional to
some function f (n). The function f (n) then becomes a bound on the running
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 will 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 xed-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 nally, extremely detailed
statements about the number of steps an algorithm executes are oftenin
a strong sensemeaningless. As just discussed, we will generally be counting
steps in a pseudo-code specication of an algorithm that resembles a highlevel programming language. Each one of these steps will typically unfold
into some xed number of primitive steps when the program is compiled into
35
36
Chapter 2 Basics of Algorithm Analysis
an intermediate representation, and then into some further number of steps
depending on the particular architecture being used to do the computing. So
the most we can safely say is that as we look at different levels of computational
abstraction, the notion of a step may grow or shrink by a constant factor
for example, if it takes 25 low-level machine instructions to perform one
operation in our high-level language, then our algorithm that took at most
1.62n2 + 3.5n + 8 steps can also be viewed as taking 40.5n2 + 87.5n + 200 steps
when we analyze it at a level that is closer to the actual hardware.
O,
, 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 functionsay, the 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 O(f (n)) (read as T (n) is order f (n)) if, for
sufciently 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) = O(f (n)). More precisely,
T (n) is O(f (n)) if there exist constants c > 0 and n0 0 so that for all n n0,
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 denition requires a constant c
to exist that works for all n; in particular, c cannot depend on n.
As an example of how this denition 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) = pn2 + qn + r pn2 + qn2 + rn2 = (p + q + r )n2
for all n 1. This inequality is exactly what the denition 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 )n3 as the denition of O(n3) requires. The
fact that a function can have many upper bounds is not just a trick of the
notation; it shows up in the analysis of running times as well. There are cases
2.2 Asymptotic Order of Growth
where an algorithm has been proved to have running time O(n3); 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 rst 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 algorithmsay 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 specic function f (n). (In this example, f (n) happens to be n2.) Thus, we
say that T (n) is (f (n)) (also written T (n) = (f (n))) if there exist constants
> 0 and n0 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 xed, independent
of n.
This denition works just like O(), 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) = (n2). Whereas establishing the upper bound involved
inating 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 0, we have
T (n) = pn2 + qn + r pn2 ,
which meets what is required by the denition of
() with
= p > 0.
Just as we discussed the notion of tighter and weaker upper bounds,
the same issue arises for lower bounds. For example, it is correct to say that
our function T (n) = pn2 + qn + r is (n), since T (n) pn2 pn.
Asymptotically Tight Bounds If we can show that a running time T (n) is
both O(f (n)) and also (f (n)), then in a natural sense weve found the right
bound: T (n) grows exactly like f (n) to within a constant factor. This, for
example, is the conclusion we can draw from the fact that T (n) = pn2 + qn + r
is both O(n2) and (n2).
There is a notation to express this: if a function T (n) is both O(f (n)) and
(f (n)), we say that T (n) is (f (n)). In this case, we say that f (n) is an
asymptotically tight bound for T (n). So, for example, our analysis above shows
that T (n) = pn2 + qn + r is (n2).
Asymptotically tight bounds on worst-case running times are nice things
to nd, since they characterize the worst-case performance of an algorithm
37
38
Chapter 2 Basics of Algorithm Analysis
precisely up to constant factors. And as the denition 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 (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 innity. Essentially, if the ratio of functions
f (n) and g (n) converges to a positive constant as n goes to innity, then
f (n) = (g (n)).
(2.1)
Let f and g be two functions that
lim
n
f (n)
g (n)
exists and is equal to some number c > 0. Then f (n) =
(g (n)).
Proof. We will use the fact that the limit exists and is positive to show that
f (n) = O(g (n)) and f (n) = (g (n)), as required by the denition of ().
Since
lim
n
f (n)
= c > 0,
g (n)
it follows from the denition of a limit that there is some n0 beyond which the
1
ratio is always between 2 c and 2c . Thus, f (n) 2cg (n) for all n n0, which
1
implies that f (n) = O(g (n)); and f (n) 2 cg (n) for all n n0, which implies
that f (n) = (g (n)).
Properties of Asymptotic Growth Rates
Having seen the denitions of O,
basic properties.
, and
, it is useful to explore some of their
Transitivity A rst 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.
(2.2)
(a) If f = O(g ) and g = O(h), then f = O(h).
(b) If f =
(g ) and g =
(h), then f =
(h).
2.2 Asymptotic Order of Growth
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 n0, we
have g (n) c h(n) for all n n0. So consider any number n that is at least as
large as both n0 and n0. We have f (n) cg (n) cc h(n), and so f (n) cc h(n)
for all n max(n0 , n0). 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 f = (g ) and that
g = (h). Then since f = O(g ) and g = O(h), we know from part (a) that
f = O(h); since f = (g ) and g = (h), we know from part (b) that f = (h).
It follows that f = (h). Thus we have shown
(2.3)
If f =
(g ) and g =
(h), then f =
(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 n0, we have f (n) ch(n)
for all n n0. Also, for some (potentially different) constants c and n0,
we have g (n) c h(n) for all n n0. So consider any number n that is at
least as large as both n0 and n0. We have f (n) + g (n) ch(n) + c h(n). Thus
f (n) + g (n) (c + c )h(n) for all n max(n0 , n0), which is exactly what is
required for showing that f + g = O(h).
There is a generalization of this to sums of a xed 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 essentially 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 xed constant, and let f1, f2 , . . . , fk and h be functions such
that fi = O(h) for all i. Then f1 + f2 + . . . + fk = O(h).
There is also a consequence of (2.4) that covers the following kind of
situation. It frequently happens that were analyzing an algorithm with two
high-level parts, and it is easy to show that one of the two parts is slower
than the other. Wed like to be able to say that the running time of the whole
algorithm is asymptotically comparable to the running time of the slow part.
Since the overall running time is a sum of two functions (the running times of
39
40
Chapter 2 Basics of Algorithm Analysis
the two parts), results on asymptotic bounds for sums of functions are directly
relevant.
(2.6) Suppose that f and g are two functions (taking nonnegative values)
such that g = O(f ). Then f + g = (f ). In other words, f is an asymptotically
tight bound for the combined function f + g.
Proof. Clearly f + g = (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 (2.4) we have f + g = O(f ).
This result also extends to the sum of any xed, constant number of
functions: the most rapidly growing among the functions is an asymptotically
tight bound for the sum.
Asymptotic Bounds for Some Common Functions
There are a number of functions that come up repeatedly in the analysis of
algorithms, and it is useful to consider the asymptotic properties of some of
the most basic of these: polynomials, logarithms, and exponentials.
Polynomials Recall that a polynomial is a function that can be written in
the form f (n) = a0 + a1n + a2n2 + . . . + ad nd for some integer constant d > 0,
where the nal coefcient ad 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 termthe 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 coefcient ad > 0.
(2.7) Let f be a polynomial of degree d, in which the coefcient ad is positive.
Then f = O(nd ).
Proof. We write f = a0 + a1n + a2n2 + . . . + ad nd , where ad > 0. The upper
bound is a direct application of (2.5). First, notice that coefcients aj for j < d
may be negative, but in any case we have aj nj |aj |nd for all n 1. Thus each
term in the polynomial is O(nd ). Since f is a sum of a constant number of
functions, each of which is O(nd ), it follows from (2.5) that f is O(nd ).
One can also show that under the conditions of (2.7), we have f =
and hence it follows that in fact f = (nd ).
(nd ),
2.2 Asymptotic Order of Growth
This is a good point at which to discuss the relationship between these
types of asymptotic bounds and the notion of polynomial time , which we
arrived at in the previous section as a way to formalize the more elusive concept
of efciency. Using O() notation, its easy to formally dene polynomial time:
a polynomial-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 O( n) = O(n1/2).
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 running time
O(n log n), then it also has running time O(n2), and so it is a polynomial-time
algorithm.
Logarithms Recall that logb 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 logb n is asymptotically bounded by every function of the
form nx , even for (noninteger) values of x arbitrary close to 0.
(2.8)
For every b > 1 and every x > 0, we have logb n = O(nx ).
One can directly translate between logarithms of different bases using the
following fundamental identity:
loga n =
logb n
.
logb a
This equation explains why youll often notice people writing bounds like
O(log n) without indicating the base of the logarithm. This is not sloppy
usage: the identity above says that loga n = log1 a logb n, so the point is that
b
loga n = (logb n), and the base of the logarithm is not important when writing
bounds using asymptotic notation.
41
42
Chapter 2 Basics of Algorithm Analysis
Exponentials Exponential functions are functions of the form f (n) = r n for
some constant base r . Here we will be concerned with the case in which r > 1,
which results in a very fast-growing function.
In particular, where polynomials raise n to a xed exponent, exponentials
raise a xed 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 > 0, we have nd = O(r n ).
In particular, every exponential grows faster than every polynomial. And as
we saw in Table 2.1, when you plug in actual values of n, the differences in
growth rates are really quite impressive.
Just as people write O(log n) 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 justied 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 r n = (sn ). Indeed, this would require that
for some constant c > 0, we would have r n csn for all sufciently large n.
But rearranging this inequality would give (r /s)n c for all sufciently large
n. Since r > s, the expression (r /s)n is tending to innity with n, and so it
cannot possibly remain bounded by a xed 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 exponentialthey 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
further details of the exact running time. This is not entirely fair. Occasionally
theres more going on with an exponential algorithm than rst appears, as
well see, for example, in Chapter 10; but as we argued in the rst section of
this chapter, its a reasonable rule of thumb.
Taken together, then, logarithms, polynomials, and exponentials serve as
useful landmarks in the range of possible functions that you encounter when
analyzing running times. Logarithms grow more slowly than polynomials, and
polynomials grow more slowly than exponentials.
2.3 Implementing the Stable Matching Algorithm
Using 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 fashionas we expressed the GaleShapley Stable Matching algorithm in Chapter 1, for exampleone 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 specic 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 n2 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 rst question we need to
discuss is how such a ranking will 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 efcient 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 will 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 n 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
43
44
Chapter 2 Basics of Algorithm Analysis
elements one by one in O(n) time, assuming we dont know anything
about the order in which the elements appear in A.
.
If the array elements are sorted in some clear way (either numerically
or alphabetically), then we can determine whether an element e belongs
to the list in O(log n) time using binary search; we will not need to use
binary search for any part of our stable matching implementation, but
we will have more to say about it in the next section.
An array is less good for dynamically maintaining a list of elements that
changes over time, such as the list of free men in the Stable Matching algorithm;
since men go from being free to engaged, and potentially back again, a list of
free 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 null if i is the last element. We also have a pointer First
that points to the rst 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 xed 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 eld e .val
that contains the value of the element, and a eld 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 eld e .Prev that contains
a pointer to the previous element in the list. (e .Prev = null if e is the rst
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 rst line of Figure 2.1.
A doubly linked list can be modied as follows.
.
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 .Next, point directly to each other. The
deletion operation is illustrated in Figure 2.1.
.
Insertion. To insert element e between elements d and f in a list, we
splice it in by updating d.Next and f .Prev to point to e , and the Next
and Prev pointers of e to point to d and f , respectively. This operation is
2.3 Implementing the Stable Matching Algorithm Using Lists and Arrays
Before deleting e:
Element e
val
val
val
After deleting e:
Element e
val
val
val
Figure 2.1 A schematic representation of a doubly linked list, 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 nd the ith element of the list in
O(1) time: to nd the ith element, we have to follow the Next pointers starting
from the beginning of the list, which takes a total of O(i) time.
Given the relative advantages and disadvantages of arrays and lists, it may
happen that we receive the input to a problem in one of the two formats and
want to convert it into the other. As discussed earlier, such preprocessing is
often useful; and in this case, it is easy to convert between the array and
list representations in O(n) time. This allows us to freely choose the data
structure that suits the algorithm better and not be constrained by the way
the information is given as input.
Implementing the Stable Matching Algorithm
Next we will use arrays and linked lists to implement the Stable Matching algorithm from Chapter 1. We have already shown that the algorithm terminates in
at most n2 iterations, and this provides a type of upper bound on the running
time. However, if we actually want to implement the G-S algorithm so that it
runs in time proportional to n2, we need to be able to implement each iteration
in constant time. We discuss how to do this now.
For simplicity, assume that the set of men and women are both {1, . . . , n}.
To ensure this, we can order the men and women (say, alphabetically), and
associate number i with the ith man mi or ith women wi in this order. This
45
46
Chapter 2 Basics of Algorithm Analysis
assumption (or notation) allows us to dene an array indexed by all men
or all women. We need to have a preference list for each man and for each
woman. To do this we will have two arrays, one for womens preference lists
and one for the mens preference lists; we will use ManPref[m , i] to denote
the ith woman on man ms preference list, 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 2n individuals is O(n2), 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 efciently. 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 by maintaining the set
of free men as a linked list. When we need to select a free man, we take the
rst man m on this list. We delete m from the list if he becomes engaged, and
possibly insert a different man m , 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
will propose to on his list. We initialize Next[m] = 1 for all men m. If a man m
needs to propose to a woman, hell propose to w = ManPref[m ,Next[m]], and
once he proposes 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 identify
the man m that w is engaged to (if there is such a man). We can do this by
maintaining an array Current of length n, where Current[w] is the woman
ws current partner m . We set Current[w] to a special null symbol when we
need to indicate that woman w is not currently engaged; at the start of the
algorithm, Current[w] is initialized to this null symbol for all women w.
To sum up, the data structures we have set up thus far can implement the
operations (1)(3) in O(1) time each.
2.4 A Survey of Common Running Times
Maybe the trickiest question is how to maintain womens preferences to
keep step (4) efcient. Consider a step of the algorithm, when man m proposes
to a woman w. Assume w is already engaged, and her current partner is
m =Current[w]. We would like to decide in O(1) time if woman w prefers m
or m . Keeping the womens preferences in an array WomanPref, 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 nd m and m on the list. While O(n)
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 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
n2. Then, to decide which of m or m is preferred by w, we simply compare
the values Ranking[w , m] and Ranking[w , m ].
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 implement the G-S
algorithm in O(n2) time.
2.4 A Survey of Common Running Times
When trying to analyze a new algorithm, it helps to have a rough sense of
the landscape of different running times. Indeed, there are styles of analysis
that recur frequently, and so when one sees running-time bounds like O(n),
O(n log n), and O(n2) appearing over and over, its often for one of a very
small number of distinct reasons. Learning to recognize these common styles
of analysis is a long-term goal. To get things under way, we offer the following
survey of common running-time bounds and some of the typical approaches
that lead to them.
Earlier we discussed the notion that most problems have a natural search
spacethe set of all possible solutionsand we noted that a unifying theme
in algorithm design is the search for algorithms whose performance is more
efcient 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
many cases with an analysis of the brute-force algorithm, since it is a useful
47
48
Chapter 2 Basics of Algorithm Analysis
way to get ones bearings with respect to a problem; the task of improving on
such algorithms will be our goal in most of the book.
Linear Time
An algorithm that runs in O(n), or linear, time has a very natural property:
its running time is at most a constant factor times the size of the input. One
basic way to get an algorithm with this running time is to process the input
in a single pass, spending a constant amount of time on each item of input
encountered. Other algorithms achieve a linear time bound for more subtle
reasons. To illustrate some of the ideas here, we consider two simple lineartime algorithms as examples.
Computing the Maximum 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
a1, a2 , . . . , 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 ai .
max = a1
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 youfor example, an algorithm running on a high-speed switch
on the Internet may see a stream of packets ying 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.
Suppose we are given two lists of n numbers each, a1, a2 , . . . , an and
b1, b2 , . . . , bn , and each is already arranged in ascending order. Wed like to
2.4 A Survey of Common Running Times
merge these into a single list c1, c2 , . . . , c2n that is also arranged in ascending
order. For example, merging the lists 2, 3, 11, 19 and 4, 9, 16, 25 results in the
output 2, 3, 4, 9, 11, 16, 19, 25.
To do this, we could just throw the two lists together, ignore the fact that
theyre separately arranged in ascending order, and run a sorting algorithm.
But this clearly seems wasteful; wed like to make use of the existing order in
the input. One way to think about designing a better algorithm is to imagine
performing the merging of the two lists by hand: suppose youre given two
piles of numbered cards, each arranged in ascending order, and youd like to
produce a single ordered pile containing all the cards. If you look at the top
card on each stack, you know that the smaller of these two should go rst 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.
To merge sorted lists A = a1, . . . , an and B = b1, . . . , bn :
Maintain a Current pointer into each list, initialized to
point to the front elements
While both lists are nonempty:
Let ai and bj be the elements pointed to by the Current pointer
Append the smaller of these two to the output list
Advance the Current pointer in the list from which the
smaller element was selected
EndWhile
Once one list is empty, append the remainder of the other list
to the output
See Figure 2.2 for a picture of this process.
Append the smaller of
ai and bj to the output.
////// ai
A
/// bj
B
Merged result
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 output.
49
50
Chapter 2 Basics of Algorithm Analysis
Now, to show a linear-time bound, one is tempted to describe an argument
like what worked for the maximum-nding 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 n/2
iterations while elements from A are repeatedly being selected, and hence
it will be involved in (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 bound, 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 rst 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 ow 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.
O(n log n) Time
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 equal-sized pieces, solves each piece
recursively, and then combines the two solutions in linear time.
Sorting is perhaps the most well-known example of a problem that can be
solved this way. Specically, 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.
2.4 A Survey of Common Running Times
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 x1, x2 , . . . , xn
on which copies of a le arrived at a server, and wed like to nd the largest
interval of time between the rst and last of these time-stamps during which
no copy of the le arrived. A simple solution to this problem is to rst sort the
time-stamps x1, 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(n 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
specied by (x , y) coordinates, and youd like to nd 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
1
is n = n(n2 1) , and since this quantity is bounded by 2 n2, it is O(n2). More
2
2) because we multiply the number of
crudely, the number of pairs is O(n
ways of choosing the rst 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
(xi xj )2 + (yi yj )2 in constant time, so the overall running time is O(n2).
This example illustrates a very common way in which a running time of O(n2)
arises: performing a search over all pairs of input items and spending constant
time per pair.
Quadratic time also arises naturally from a pair of nested loops: An algorithm consists of a loop 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.
The brute-force algorithm for nding the closest pair of points can be
written in an equivalent way with two nested loops:
For each input point (xi , yi )
For each other input point (xj , yj )
Compute distance d = (xi xj )2 + (yi yj )2
51
52
Chapter 2 Basics of Algorithm Analysis
If d is less than the current minimum, update minimum to d
Endfor
Endfor
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
rst, one feels there is a certain inevitability about this 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 nds 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
S1, S2 , . . . , Sn , each of which is a subset of {1, 2, . . . , n}, and we would like
to know whether some pair of these sets is disjointin 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 constant time whether a
given number p belongs to Si . The following is a direct way to approach the
problem.
For pair of sets Si and Sj
Determine whether Si and Sj have an element in common
Endfor
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 Sj
For each element p of Si
Determine whether p also belongs to Sj
Endfor
If no element of Si belongs to Sj then
2.4 A Survey of Common Running Times
Report that Si and Sj are disjoint
Endif
Endfor
Endfor
Each of the sets has maximum size O(n), so the innermost loop takes time
O(n). Looping over the sets Sj involves O(n) iterations around this innermost
loop; and looping over the sets Si involves O(n) iterations around this. Multiplying these three factors of n together, we get the running time of O(n3).
For this problem, there are algorithms that improve on O(n3) running
time, but they are quite complicated. Furthermore, it is not clear whether
the improved algorithms for this problem are practical on inputs of reasonable
size.
O(nk ) Time
In the same way that we obtained a running time of O(n2) by performing bruteforce search over all pairs formed from a set of n items, we obtain a running
time of O(nk ) for any constant k when we search over all subsets of size k.
Consider, for example, the problem of nding independent sets in a graph,
which we discussed in Chapter 1. Recall that a set of nodes is independent
if no two are joined by an edge. Suppose, in particular, that for some xed
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 algorithm 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
Endfor
If no k-node independent set was found then
Declare failure
Endif
To understand the running time of this algorithm, we need to consider two
quantities. First, the total number of k-element subsets in an n-element set is
n
n(n 1)(n 2) . . . (n k + 1) nk
=
.
k
k(k 1)(k 2) . . . (2)(1)
k!
53
54
Chapter 2 Basics of Algorithm Analysis
Since we are treating k as a constant, this quantity is O(nk ). Thus, the outer
loop in the algorithm above will run for O(nk ) 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 denition 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 k , that is, O(k2), pairs and spending constant time
2
on each.
Thus the total running time is O(k2nk ). 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 nd 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 signicant differences in the efciency 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 come up very frequently are 2n and n!,
and we now discuss why this is so.
Suppose, for example, that we are given a graph and want to nd 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 signicantly on brute-force search, which in this case would look as
follows.
For each subset S of nodes
Check whether S constitutes an independent set
If S is a larger independent set than the largest seen so far then
Record the size of S as the current maximum
Endif
Endfor
This is very much like the brute-force algorithm for k-node independent sets,
except that now we are iterating over all subsets of the graph. The total number
2.4 A Survey of Common Running Times
of subsets of an n-element set is 2n , and so the outer loop in this algorithm
will run for 2n iterations as it tries all these subsets. Inside the loop, we are
checking all pairs from a set S that can be as large as n nodes, so each iteration
of the loop takes at most O(n2) time. Multiplying these two together, we get a
running 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 inefcient appears to be necessary; but its important
to keep in mind that 2n is the size of the search space for many problems,
and for many of them we will be able to nd highly efcient 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 nd 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 nd 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 itemsfor 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 rst 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 nd 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 rst city, so the crux of the problem is the
55
56
Chapter 2 Basics of Algorithm Analysis
implicit search over all orders of the remaining n 1 cities, leading to a search
space of size (n 1)!. In Chapter 8, we will see that Traveling Salesman
is another problem that, like Independent Set, belongs to the class of NPcomplete problems and is believed to have no efcient 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 efciently, 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 valuesay it is qand 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, if 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
1
n, so after k probes it has size at most ( 2 )k n.
Given this, how long will it take for the size of the active region to be
1
reduced to a constant? We need k to be large enough so that ( 2 )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 bound
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 sufce to shrink the input down to constant
size, at which point the problem can generally be solved directly.
2.5 A More Complex Data Structure: Priority Queues
2.5 A More Complex Data Structure:
Priority Queues
Our primary goal in this book was expressed at the outset of the chapter:
we seek algorithms that improve qualitatively on brute-force search, and in
general we use polynomial-time solvability as the concrete formulation of
this. Typically, achieving a polynomial-time solution to a nontrivial problem
is not something that depends on ne-grained implementation details; rather,
the difference between exponential and polynomial is based on overcoming
higher-level obstacles. Once one has an efcient algorithm to solve a problem,
however, it is often possible to achieve further improvements in running time
by being careful with the implementation details, and sometimes by using
more complex data structures.
Some complex data structures are essentially tailored for use in a single
kind of algorithm, while others are more generally applicable. In this section,
we describe one of the most broadly useful sophisticated data structures,
the priority queue. Priority queues will be useful when we describe how to
implement some of the graph algorithms developed later in the book. For our
purposes here, it is a useful illustration of the analysis of a data structure that,
unlike lists and arrays, must perform some nontrivial processing each time it
is invoked.
The Problem
In the implementation of the Stable Matching algorithm in Section 2.3, we
discussed the need to maintain a dynamically changing set S (such as the set
of all free men in that case). In such situations, we want to be able to add
elements to and delete elements from the set S, and we want to be able to
select an element from S when the algorithm calls for it. A priority queue is
designed for applications in which elements have a priority value, or key, and
each time we need to select an element from S, we want to take the one with
highest priority.
A priority queue is a data structure that maintains a set of elements S,
where each element v S has an associated value key(v) that denotes the
priority of element v; smaller keys represent higher priorities. Priority queues
support the addition and deletion of elements from the set, and also the
selection of the element with smallest key. Our implementation of priority
queues will also support some additional operations that we summarize at the
end of the section.
A motivating application for priority queues, and one that is useful to keep
in mind when considering their general function, is the problem of managing
57
58
Chapter 2 Basics of Algorithm Analysis
real-time events such as the scheduling of processes on a computer. Each
process has a priority, or urgency, but processes do not arrive in order of
their priorities. Rather, we have a current set of active processes, and we want
to be able to extract the one with the currently highest priority and run it.
We can maintain the set of processes in a priority queue, with the key of a
process representing its priority value. Scheduling the highest-priority process
corresponds to selecting the element with minimum key from the priority
queue; concurrent with this, we will also be inserting new processes as they
arrive, according to their priority values.
How efciently do we hope to be able to execute the operations in a priority
queue? We will show how to implement a priority queue containing at most
n elements at any time so that elements can be added and deleted, and the
element with minimum key selected, in O(log n) time per operation.
Before discussing the implementation, let us point out a very basic application of priority queues that highlights why O(log n) time per operation is
essentially the right bound to aim for.
(2.11) A sequence of O(n) priority queue operations can be used to sort a set
of n numbers.
Proof. Set up a priority queue H , and insert each number into H with its value
as a key. Then extract the smallest number one by one until all numbers have
been extracted; this way, the numbers will come out of the priority queue in
sorted order.
Thus, with a priority queue that can perform insertion and the extraction
of minima in O(log n) per operation, we can sort n numbers in O(n log n)
time. It is known that, in a comparison-based model of computation (when
each operation accesses the input only by comparing a pair of numbers),
the time needed to sort must be at least proportional to n log n, so (2.11)
highlights a sense in which O(log n) time per operation is the best we can
hope for. We should note that the situation is a bit more complicated than
this: implementations of priority queues more sophisticated than the one we
present here can improve the running time needed for certain operations, and
add extra functionality. But (2.11) shows that any sequence of priority queue
operations that results in the sorting of n numbers must take time at least
proportional to n log n in total.
A Data Structure for Implementing a Priority Queue
We will use a data structure called a heap to implement a priority queue.
Before we discuss the structure of heaps, we should consider what happens
with some simpler, more natural approaches to implementing the functions
2.5 A More Complex Data Structure: Priority Queues
of a priority queue. We could just have the elements in a list, and separately
have a pointer labeled Min to the one with minimum key. This makes adding
new elements easy, but extraction of the minimum hard. Specically, nding
the minimum is quickwe just consult the Min pointerbut after removing
this minimum element, we need to update the Min pointer to be ready for the
next operation, and this would require a scan of all elements in O(n) time to
nd 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 nd 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 nd the position where s should
be inserted.
The Denition of a Heap So in all these simple approaches, at least one of
the operations can take up to O(n) timemuch more than the O(log n) per
operation that were hoping for. This is where heaps come in. The heap data
structure combines the benets 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 tree. In other words,
Heap order: For every element v, at a node i, the element w at is parent
satises 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 pointers: 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
60
Chapter 2 Basics of Algorithm Analysis
Each nodes key is at least
as large as its parents.
1
2
5
10
15
3
17
20
7
9
15
1
2
5
10
3
7
11 15 17 20
9
15
8
16
X
11
8
16
Figure 2.3 Values in a heap shown as a binary 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.
at position i, the children are the nodes at positions leftChild(i) = 2i and
rightChild(i) = 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(i) = i/2 . If
the heap has n < N elements at some time, we will use the rst n positions
of the array to store the n heap elements, and use length(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.
Implementing the Heap Operations
The heap element with smallest key is at the root, so it takes O(1) time to
identify the minimal element. How do we add or delete heap elements? First
consider 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 nal 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 Heapify-up to x our heap. Let j = parent(i) =
i/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 will x 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
61
2.5 A More Complex Data Structure: Priority Queues
The Heapify-up process is moving
element v toward the root.
2
4
5
10
15
9
17
20
4
7
17
15
11
8
2
16
10
w
3
5
v
15
9
17
20
7
17
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 violation moves one step closer to the root of
the tree (on the right).
j = parent(i) to continue xing the heap by pushing the damaged part upward.
Figure 2.4 shows the rst two steps of the process after an insertion.
Heapify-up(H,i):
If i > 1 then
let j = parent(i) = i/2
If key[H[i]]<key[H[j]] then
swap the array entries H[i] and H[j]
Heapify-up(H,j)
Endif
Endif
To see why Heapify-up works, eventually restoring the heap order, it
helps to understand more fully the structure of our slightly damaged heap in
the middle of this process. Assume that H is an array, and v is the element in
position i. We say that H is almost a heap with the key of H [i] too small, if there
is a value key(v) such that raising the value of key(v) to would make
the resulting array satisfy the heap property. (In other words, element v in H [i]
is too small, but raising it to would x 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 would make H a heap, then the value of H [1] must
also be smaller than both its children, and hence it already has the heap-order
property.
15
3
8
16
v
11 w
62
Chapter 2 Basics of Algorithm Analysis
(2.12) The procedure Heapify-up(H , i) xes the heap property in O(log i)
time, assuming that the array H is almost a heap with the key of H [i] too small.
Using Heapify-up we can insert a new element in a heap of n elements in
O(log n) time.
Proof. We prove the statement by induction on i. If i = 1 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 = 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 Heapify-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 rst 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 x 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(H , 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 rst 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
rst 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-down, closely analogous to Heapify-up, that
63
2.5 A More Complex Data Structure: Priority Queues
The Heapify-down process
is moving element w down,
toward the leaves.
4
7
21
10
15
16
17
20
7
17
15
w
7
11
8
16
4
7
10
15
16
17
20
w
21
17
15
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 rst steps of this process.
Heapify-down(H,i):
Let n = length(H )
If 2i > n then
Terminate with H unchanged
Else if 2i < n then
Let left = 2i, and right = 2i + 1
Let j be the index that minimizes key[H [left]] and key[H [right]]
Else if 2i = n then
Let j = 2i
Endif
If key[H [j ]] < key[H [i]] then
swap the array entries H [i] and H [j ]
Heapify-down(H , j )
Endif
Assume that H is an array and w is the element in position i. We say that
H is almost a heap with the key of H [i] too big , if there is a value key(w)
such that lowering the value of key(w) to 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
8
11
16
64
Chapter 2 Basics of Algorithm Analysis
H [i] is already larger than its parent and hence it already has the heap-order
property.
(2.13) The procedure Heapify-down(H , i) xes 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
Heapify-up or Heapify-down we can delete a new element in a heap of n
elements in O(log n) time.
Proof. We prove that the process xes 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 Heapify-down xes 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
H [i] 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 x the heap property in O(log n)
time.
Implementing Priority Queues with Heaps
The heap data structure with the Heapify-down and Heapify-up operations
can efciently 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.
.
.
.
.
.
StartHeap(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.
Insert(H , v) inserts the item v into heap H . If the heap currently has n
elements, this takes O(log n) time.
FindMin(H ) identies the minimum element in the heap H but does not
remove it. This takes O(1) 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 ) identies 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.
Solved Exercises
There is a second class of operations in which we want to operate on
elements by name, rather than by their position in the heap. For example, in
a number of graph algorithms that use heaps, the heap elements are nodes of
the graph with key values that are computed during the algorithm. At various
points in these algorithms, we want to operate on a particular node, regardless
of where it happens to be in the heap.
To be able to access given elements of the priority queue efciently, we
simply maintain an additional array Position that stores the current position
of each element (each node) in the heap. We can now implement the following
further operations.
.
.
To delete the element v, we apply Delete(H ,Position[v]). Maintaining
this array does not increase the overall running time, and so we can
delete an element v from a heap with n nodes in O(log n) time.
An additional operation that is used by some algorithms is ChangeKey
(H , v , ), which changes the key value of element v to key(v) = . To
implement this operation in O(log n) time, we rst 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 identied the position of element v,
we change the key and then apply Heapify-up or Heapify-down 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 f (n) is O(g (n)).
f1(n) = 10n
f2(n) = n1/3
f3(n) = nn
f4(n) = log2 n
f5(n) = 2 log2 n
Solution We can deal with functions f1, 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(f1(n)).
65
66
Chapter 2 Basics of Algorithm Analysis
Now, the function f3 isnt so hard to deal with. It starts out smaller than
10n , but once n 10, then clearly 10n nn . This is exactly what we need for
the denition of O() notation: for all n 10, we have 10n cnn , where in this
case c = 1, and so 10n = O(nn ).
Finally, we come to function f5, 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 f5(n) = log2 n =
(log2 n)1/2. What do the logarithms of the other functions look like? log f4(n) =
1
log2 log2 n, while log f2(n) = 3 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 f2(n) = z
3
log f4(n) = log2 z
log f5(n) = z 1/2
Now its easier to see whats going on. First, for z 16, we have log2 z
z 1/2. But the condition z 16 is the same as n 216 = 65, 536; thus once
n 216 we have log f4(n) log f5(n), and so f4(n) f5(n). Thus we can write
1
f4(n) = O(f5(n)). Similarly we have z 1/2 3 z once z 9in other words,
9 = 512. For n above this bound we have log f (n) log f (n) and
once n 2
5
2
hence f5(n) f2(n), and so we can write f5(n) = O(f2(n)). Essentially, we
have discovered that 2 log2 n is a function whose growth rate lies somewhere
between that of logarithms and polynomials.
Since we have sandwiched f5 between f4 and f2, this nishes 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 = (f ).
Solution This exercise is a way to formalize the intuition that O() and ()
are in a sense opposites. It is, in fact, not difcult to prove; it is just a matter
of unwinding the denitions.
Were given that, for some constants c and n0, we have f (n) cg (n) for
1
all n n0. Dividing both sides by c , we can conclude that g (n) c f (n) for
all n n0. But this is exactly what is required to show that g = (f ): we have
established that g (n) is at least a constant multiple of f (n) (where the constant
1
is c ), for all sufciently large n (at least n0).
Exercises
Exercises
1. 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
(b) n3
(c)
100n2
(d) n log n
(e)
2n
2. Suppose you have algorithms with the six 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
1010 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)
n2
(b) n3
(c)
100n2
(d) n log n
(e)
2n
(f)
22
n
3. 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)).
f1(n) = n2.5
f2(n) = 2n
f3(n) = n + 10
f4(n) = 10n
f5(n) = 100n
f6(n) = n2 log n
4. Take the following list of functions and arrange them in ascending order
of growth rate. That is, if function g (n) immediately follows function f (n)
in your list, then it should be the case that f (n) is O(g (n)).
67
68
Chapter 2 Basics of Algorithm Analysis
g1(n) = 2 log n
g2(n) = 2n
g4(n) = n4/3
g3(n) = n(log n)3
g5(n) = nlog n
n
g6(n) = 22
g7(n) = 2n
2
5. 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(log2 g (n)).
(b) 2f (n) is O(2g (n)).
(c)
f (n)2 is O(g (n)2).
6. Consider the following basic problem. Youre given an array A consisting
of n integers A[1], A[2], . . . , 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[j ]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, 2, . . . , n
For j = i + 1, i + 2, . . . , n
Add up array entries A[i] through A[j ]
Store the result in B[i , j ]
Endfor
Endfor
(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 the running time of the algorithm
on an input of size n is also (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 problemafter all, it just iterates through
Exercises
the relevant entries of the array B, filling in a value for eachit
contains some highly unnecessary sources of inefficiency. Give a
different algorithm to solve this problem, with an asymptotically
better running time. In other words, you should design an algorithm
with running time O(g (n)), where limn g (n)/f (n) = 0.
7. 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 example, 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 the pear
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 having to write out all the previous lines each time. (So the phrase
five golden rings only has to be written once, even though it will appear
in verses five and onward.)
Theres something asymptotic that can be analyzed here. Suppose,
for concreteness, that each line has a length that is 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.
8. Youre doing some stress-testing on various models of glass jars to
determine the height from which they can be dropped and still not break.
The setup for this experiment, on a particular type of jar, is as follows.
You have a ladder with n rungs, and you want to find the highest rung
from which you can drop a copy of the jar and not have it break. We call
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 you 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 jarat the moment
69
70
Chapter 2 Basics of Algorithm Analysis
it breaks, you have the correct answerbut you may have to drop it n
times (rather than log n 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 answerthe highest safe rungand 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
according to your strategy, then the functions f1, f2 , f3 , . . . should have
the property that each grows asymptotically slower than the previous
one: limn fk (n)/fk1(n) = 0 for each k.
Notes and Further Reading
Polynomial-time solvability emerged as a formal notion of efciency by a
gradual process, motivated by the work of a number of researchers including Cobham, Rabin, 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 algorithmsas opposed to working out exact formulas with
leading coefcients and lower-order termsis 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, Tarjan, 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 nd minimum-
Notes and Further Reading
cost spanning trees, and we will discuss randomized hashing in Chapter 13.
A number of other data structures are discussed in the book by Tarjan (1983).
The LEDA library (Library of Efcient Datatypes and Algorithms) of Mehlhorn
and N her (1999) offers an extensive library of data structures useful in
a
combinatorial and geometric applications.
Notes on the Exercises Exercise 8 is based on a problem we learned from
Sam Toueg.
71
This page intentionally left blank
Chapter 3
Graphs
Our focus in this book is on problems with a discrete avor. 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 subject. 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 denitions 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.
3.1 Basic Denitions and Applications
Recall 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 as a two-element subset of V : e = {u , v} for some
u , v V , where we call 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
closely 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 (u , 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 call 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 , v}, 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 dene: 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 specic 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 useful examples against which to check the basic denitions and algorithmic problems
that well be encountering later in the chapter. Also, in going through the list,
its useful 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 v if there is a nonstop ight that departs from u and arrives at v.
Described this way, the graph is directed; but in practice when there is an
edge (u , v), there is almost always an edge (v , 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 ights each way. Looking at
such a graph (you can generally nd them depicted in the backs of inight 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.
2. Communication networks. A collection of computers connected via a
communication network can be naturally modeled as a graph in a few
3.1 Basic Denitions and Applications
different ways. First, we could have a node for each computer and
an edge joining u and v if there is a direct physical link connecting
them. Alternatively, for studying the large-scale structure of the Internet,
people often dene 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 themroughly, an agreement
to exchange data under the standard BGP 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 denes a graph where
the nodes are computing devices situated at locations in physical space,
and there is an edge from u to v if v 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 v can hear us signal but u cannot hear vs 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.
3. Information networks. The World Wide Web can be naturally viewed as a
directed graph, in which nodes correspond to Web pages and there is an
edge from u to v if u has a hyperlink to v. The directedness of the graph
is crucial here; many pages, for example, link to popular news sites,
but these sites clearly do not reciprocate all these links. The structure of
all these hyperlinks can be used by algorithms to try inferring the most
important pages on the Web, a technique employed by most current
search engines.
The hypertextual structure of the Web is anticipated by a number of
information networks that predate the Internet by many decades. These
include the network of cross-references among articles in an encyclopedia
or other reference work, and the network of bibliographic citations
among scientic papers.
4. 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 dene 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 nancial 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
75
76
Chapter 3 Graphs
notion of afliation: given a set X of people and a set Y of organizations,
we could dene an edge between u 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 inuential people in a company or organization, to model
trust relationships in a nancial or political setting, and to track the
spread of fads, rumors, jokes, diseases, and e-mail viruses.
5. Dependency networks. It is natural to dene 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 call. Or given a set of species in an ecosystem, we could dene
a grapha food webin 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 Connectivity One of the fundamental operations in a graph is
that of traversing a sequence of nodes connected by edges. In the examples
just 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 ights.
With this notion in mind, we dene a path in an undirected graph
G = (V , E ) to be a sequence P of nodes v1, v2 , . . . , vk1, vk with the property
that each consecutive pair vi , vi+1 is joined by an edge in G . P is often called
a path from v1 to vk , or a v1-vk 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 v1, v2 , . . . , vk1, vk in which k > 2, the rst k 1
nodes are all distinct, and v1 = vk in other words, the sequence of nodes
cycles back to where it began. All of these denitions carry over naturally
to directed graphs, with the following change: in a directed path or cycle,
each pair of consecutive nodes has the property that (vi , vi+1) 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 dene connectivity of a
77
3.1 Basic Denitions and Applications
3
1
4
2
8
1
6
2
5
7
7
5
9
3
4
6
8
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
v while v has no path to u. We say 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.
In addition to simply knowing about the existence of a path between some
pair of nodes u and v, we may also want to know whether there is a short path.
Thus we dene the distance between two nodes u and v to be the minimum
number of edges in a u-v path. (We can designate some symbol like 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 u to v, 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 3.1 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 from r ; 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 edgesbut the drawing
on the right represents the result of rooting T at node 1.
9
78
Chapter 3 Graphs
Rooted trees are fundamental objects in computer science, because they
encode the notion of a hierarchy. For example, we can imagine the rooted tree
in Figure 3.1 as corresponding to the organizational structure of a tiny nineperson company; employees 3 and 4 report to employee 2; employees 2, 5,
and 7 report to employee 1; and so on. Many Web sites are organized according
to a tree-like structure, to facilitate navigation. A typical 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, rooting a tree T can make certain questions about T
conceptually easy to answer. For example, given a tree T on n nodes, how many
edges does it have? Each node other than the root has a single edge leading
upward to its parent; and conversely, each edge leads upward from precisely
one non-root node. Thus we have very easily proved the following fact.
(3.1)
Every n-node tree has exactly n 1 edges.
In fact, the following stronger statement is true, although we do not prove
it here.
(3.2) Let G be an undirected graph on n nodes. Any two of the following
statements implies the third.
(i) G is connected.
(ii) G does not contain a cycle.
(iii) G has n 1 edges.
We now turn to the role of trees in the fundamental algorithmic idea of
graph traversal.
3.2 Graph Connectivity and Graph Traversal
Having built up some fundamental notions regarding graphs, we turn to a very
basic algorithmic question: node-to-node connectivity. Suppose we are given a
graph G = (V , E ) and two particular nodes s and t . Wed like to nd an efcient
algorithm that answers the question: Is there a path from s to t in G ? We will
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 Connectivity 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,
3.2 Graph Connectivity and Graph Traversal
1
2
4
7
9
11
8
10
12
3
5
6
13
Figure 3.2 In this graph, node 1 has paths to nodes 2 through 8, but not to nodes 9
through 13.
then the problem is to start in a room s and nd your way to another designated
room t . How efcient an algorithm can we design for this task?
In this section, we describe two natural algorithms for this problem at a
high level: breadth-rst search (BFS) and depth-rst search (DFS). In the next
section we discuss how to implement each of these efciently, 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-rst
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 sthis is the rst layer of the search. We then
include all additional nodes that are joined by an edge to any node in the rst
layerthis 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 rst 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 ood 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 dene the layers L1, L2 , L3, . . . constructed by the BFS algorithm
more precisely as follows.
79
80
Chapter 3 Graphs
.
Layer L1 consists of all nodes that are neighbors of s. (For notational
reasons, we will sometimes use layer L0 to denote the set consisting just
of s.)
.
Assuming that we have dened layers L1, . . . , Lj , then layer Lj+1 consists
of all nodes that do not belong to an earlier layer and that have an edge
to a node in layer Lj .
Recalling our denition 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 all nodes
at distance exactly j from s. A node fails 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
following fact.
(3.3) For each j 1, layer Lj produced by BFS consists of all nodes at distance
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-rst search is that it produces, in a very
natural way, a tree T rooted at s on the set of nodes reachable from s.
Specically, for each such node v (other than s), consider the moment when
v is rst discovered by the BFS algorithm; this happens when some node u
in layer Lj is being examined, and we nd that it has an edge to the previously
unseen node v. At this moment, we add the edge (u , v) to the tree T u
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-rst 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}.
(b) Layer L2 is then grown by considering the nodes in layer L1 in order (say,
rst 2, then 3). Thus we discover nodes 4 and 5 as soon as we look at 2,
so 2 becomes their parent. When we consider node 2, we also discover
an edge to 3, but this isnt added to the BFS tree, since we already know
about node 3.
We rst 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
81
3.2 Graph Connectivity and Graph Traversal
1
1
2
1
2
3
4
3
5
7
2
8
4
3
5
7
6
(a)
(b)
Figure 3.3 The construction of a breadth-first search tree T for the graph 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
1, 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 property of BFS trees in general.
(3.4) Let T be a breadth-rst 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.
(c)
8
82
Chapter 3 Graphs
Current component
containing s
s
It is safe to add v.
u
v
Figure 3.4 When growing the connected component containing s, we look for nodes
like v that have not yet been visited.
Exploring a Connected Component
The set of nodes discovered by the BFS algorithm is precisely those reachable
from the starting node s. We will refer to this set R as the connected component
of G containing s; and once we know the connected component containing s,
we can simply check whether t belongs to it so as to answer the question of
s-t connectivity.
Now, if one thinks about it, its clear that BFS is just 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 dene R = {s}.
Then at any point in time, if we nd an edge (u , v) where u R and v R, we
can add v to R. Indeed, if there is a path P from s to u, then there is a path
from s to v obtained by rst 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 an edge (u , v) where u R and v R
Add v to R
Endwhile
Here is the key property of this algorithm.
(3.5) The set R produced at the end of the algorithm is precisely the connected
component of G containing s.
3.2 Graph Connectivity and Graph Traversal
Proof. We have already argued that for any node v R, there is a path from s
to v.
Now, consider a node w R, and suppose by way of contradiction, that
there is an s-w path P in G . Since s R but w R, there must be a rst node v
on P that does not belong to R; and this node v is not equal to s. Thus there is
a node u immediately preceding v on P , so (u , v) is an edge. Moreover, since v
is the rst 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 v R; this contradicts the stopping rule
for the algorithm.
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 v, the edge (u , v) that was considered in the iteration in which
v 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 denes an s-t path.
To conclude, we notice that the general algorithm we have dened to
grow R is underspecied, 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 visitin successive layers, based on their distance from s. But
there are other natural ways to grow the component, several of which lead
to efcient 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-rst search, and develop some of its basic properties.
Depth-First Search
Another natural method to nd 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 rst edge leading out
of it, to a node v. Youd then follow the rst edge leading out of v, and continue
in this way until you reached a dead enda 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 depthrst 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 , v) incident to u
If v is not marked "Explored" then
Recursively invoke DFS(v)
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 efciency.
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 reection 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 call to DFS(u), we add the
edge (u , v) to T . The resulting tree is called a depth-rst 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 nd, and so it backs up to 5, nds node 6, backs up
again to 3, and nds nodes 7 and 8. At this point there are no new nodes to nd
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
gure, nontree edges can only connect ancestors of T to descendants.
3.2 Graph Connectivity and Graph Traversal
1
1
1
1
2
2
2
2
3
3
3
5
5
4
(a)
(b)
(c)
(d)
1
1
1
2
2
2
3
3
3
5
5
7
5
7
4
6
4
6
4
6
(e)
(f)
8
(g)
Figure 3.5 The construction of a depth-first search tree T for the graph in Figure 3.2,
with (a) through (g) depicting the nodes as they are discovered in sequence. The solid
edges are the edges of T ; the dotted edges are edges of G that do not belong to T .
To establish this, we rst observe the following property of the DFS
algorithm and the tree that it produces.
(3.6) For a given recursive call DFS(u), all nodes that are marked Explored
between the invocation and end of this recursive call are descendants of u
in T.
Using (3.6), we prove
(3.7) Let T be a depth-rst 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 or y 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 rst 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 rst invoked, it is a node that was discovered between the
invocation and end of the recursive call DFS(x). It follows from (3.6) that y is
a descendant of x.
The Set of All Connected Components
So far we have been talking about the connected component containing a
particular node s. But there is a connected component associated with each
node in the graph. What is the relationship between these components?
In fact, this relationship is highly structured and is expressed in the
following claim.
(3.8) For any two nodes s and t in a graph, their connected components are
either identical or disjoint.
This is a statement that is very clear intuitively, if one looks at a graph like
the example in Figure 3.2. The graph is divided into multiple pieces with no
edges between them; the largest piece is the connected component of nodes
1 through 8, the medium piece is the connected component of nodes 11, 12,
and 13, and the smallest piece is the connected component of nodes 9 and 10.
To prove the statement in general, we just need to show how to dene 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 just 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 disjoint.
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 arbitrary node s, and we use BFS (or DFS) to generate its connected
3.3 Implementing Graph Traversal Using Queues and Stacks
component. We then nd a node v (if any) that was not visited by the search
from s, and iterate, using BFS starting from v, to generate its connected
componentwhich, by (3.8), will be disjoint from the component of s. We
continue in this way until all nodes have been visited.
3.3 Implementing Graph Traversal Using Queues
and Stacks
So far we have been discussing basic algorithmic primitives for working with
graphs without mentioning any implementation details. Here we discuss how
to use lists and arrays to represent graphs, and we discuss the trade-offs
between the different representations. Then we use these data structures to
implement the graph traversal algorithms breadth-rst search (BFS) and depthrst search (DFS) efciently. 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 will 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
|V |, and the number of edges |E |. We will use n = |V | and m = |E | 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
n
2
2 n . 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 1 edges.
But these comparisons do not always tell us which of two running times (such
as m2 and n3) are better, so we will tend to keep the running times in terms
of both of these parameters. In this section we aim to implement the basic
graph search algorithms in time O(m + n). We will refer to this as linear time ,
since it takes O(m + n) time simply to read the input. Note that when we work
with connected graphs, a running time of O(m + n) is the same as O(m), since
m n 1.
Consider a graph G = (V , E ) with n nodes, and assume the set of nodes
is V = {1, . . . , n}. The simplest way to represent a graph is by an adjacency
87
88
Chapter 3 Graphs
matrix, which is an n n matrix A where A[u , v] is equal to 1 if the graph
contains the edge (u , v) and 0 otherwise. If the graph is undirected, the matrix A
is symmetric, with A[u , v] = A[v , u] for all nodes u , v V . The adjacency
matrix representation allows us to check in O(1) time if a given edge (u , v) is
present in the graph. However, the representation has two basic disadvantages.
.
.
The representation takes (n2) space. When the graph has many fewer
edges than n2, more compact representations are possible.
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 presentand 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 signicantly fewer edges incident to most nodes, and so it
would be good to be able to nd all these incident edges more efciently.
The representation of graphs used throughout the book is the adjacency
list, which works better for sparse graphsthat is, those with many fewer than
n2 edges. In the adjacency list representation there is a record for each node v,
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 v appears
on the list for node w.
Lets compare the adjacency matrix and adjacency list representations.
First consider the space required by the representation. An adjacency matrix
requires O(n2) space, since it uses an n 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 v and the one
for w. Thus the total length of all lists is 2m = O(m).
Another (essentially equivalent) way to justify this bound is as follows.
We dene the degree nv of a node v to be the number of incident edges it has.
The length of the list at Adj[v] is list is nv , so the total length over all nodes is
O
vV nv . 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)
vV
nv = 2m.
3.3 Implementing Graph Traversal Using Queues and Stacks
Proof. Each edge e = (v , w) contributes exactly twice to this sum: once in the
quantity nv and once in the quantity nw . Since the sum is the total of the
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) space.
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 v 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 exploring
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 v once it
encounters it on this list in constant time; and then be ready to read the list
associated with node v. The list representation thus corresponds to a physical
notion of exploring the graph, in which you learn the neighbors of a node
u once you arrive at u, and can read them off in constant time per neighbor.
Queues and Stacks
Many algorithms have an inner step in which they need to process a set of
elements, such the set of all edges adjacent to a node in a graph, the set of
visited nodes in BFS and DFS, or the set of all free men in the Stable Matching
algorithm. For this purpose, it is natural to maintain the set of elements to be
considered in a linked list, as we have done for maintaining the set of free men
in the Stable Matching algorithm.
One important issue that arises is the order in which to consider the
elements in such a list. In the Stable Matching algorithm, the order in which
we considered the free men did not affect the outcome, although this required
a fairly subtle proof to verify. In many other algorithms, such as DFS and BFS,
the order in which elements are considered is crucial.
89
90
Chapter 3 Graphs
Two of the simplest and most natural options are to maintain a set of
elements as either a queue or a stack. A queue is a set from which we extract
elements in rst-in, rst-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, rst-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 rst
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 rst position on the list. Recall that a
doubly linked list has explicit First and Last pointers to the beginning and
end, respectively, so each of these insertions can be done in constant time.
Next we will discuss how to implement the search algorithms of the
previous section in linear time. We will see that BFS can be thought of as
using a queue to select which node to consider next, while DFS is effectively
using a stack.
Implementing Breadth-First Search
The adjacency list data structure is ideal for implementing breadth-rst 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 , v), we need to
know whether or not node v has been previously discovered by the search.
To make this simple, we maintain an array Discovered of length n and set
Discovered[v] = true as soon as our search rst sees v. The algorithm, as
described in the previous section, constructs layers of nodes L1, 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, 1, 2, . . . .
BFS(s):
Set Discovered[s] = true and Discovered[v] = false for all other v
Initialize L[0] to consist of the single element s
Set the layer counter i = 0
Set the current BFS tree T =
While L[i] is not empty
Initialize an empty list L[i + 1]
For each node u L[i]
Consider each edge (u , v) incident to u
If Discovered[v] = false then
Set Discovered[v] = true
Add edge (u , v) to the tree T
3.3 Implementing Graph Traversal Using Queues and Stacks
Add v to the list L[i + 1]
Endif
Endfor
Increment the layer counter i by one
Endwhile
In this implementation it does not matter whether we manage each list
L[i] as a queue or a stack, since the algorithm is allowed to consider the nodes
in a layer Li in any order.
(3.11) The above implementation of the BFS algorithm runs 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 rst 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 all iterations of the
While loop. When we consider a node u, we need to look through all edges
(u , v) 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( uV nu ). Recall
from (3.9) that uV 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
Li . 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 rst 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 rst in the queue.
91
92
Chapter 3 Graphs
If we maintain the discovered nodes in this order, then all nodes in layer Li
will appear in the queue ahead of all nodes in layer Li+1, 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+1, and so forth. Hence this implementation in terms
of a single queue will produce the same result as the BFS implementation
above.
Implementing Depth-First Search
We now consider the depth-rst 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 vthe rst time it is seen, when the algorithm nds an edge leading
to vand 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 Li+1, and we deferred actually
exploring these neighbors until we got to the processing of layer Li+1. In
contrast, DFS is more impulsive: when it explores a node u, it scans the
neighbors of u until it nds the rst not-yet-explored node v (if any), and
then it immediately shifts attention to exploring v.
To implement the exploration strategy of DFS, we rst 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 rst discovered. The implementation
in full looks as follows.
3.3 Implementing Graph Traversal Using Queues and Stacks
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
Endfor
Endif
Endwhile
There is one nal wrinkle to mention. Depth-rst search is underspecied,
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 adjacency list is processed in reverse order).
If we want the algorithm to also nd 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 sufces 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 the running time, we
need to bound the number of these operations. To count the number of stack
operations, it sufces 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 nv denote the
degree of node v. Node v will be added to the stack S every time one of its
nv adjacent nodes is explored, so the total number of nodes added to S is at
93
94
Chapter 3 Graphs
most u nv = 2m. This proves the desired O(m + n) bound on the running
time of DFS.
(3.13) The above implementation of the DFS algorithm runs in time O(m + n)
(i.e., linear in the input size), if the graph is given by the adjacency list
representation.
Finding the Set of All Connected Components
In the previous section we talked about how one can use BFS (or DFS) to nd
all connected components of a graph. We start with an arbitrary node s, and
we use BFS (or DFS) to generate its connected component. We then nd 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 componentwhich, by
(3.8), will be disjoint from the component of s. We continue in this way until
all nodes have been visited.
Although we earlier expressed the running time of BFS and DFS as O(m +
n), where m and n are the total number of edges and nodes in the graph, both
BFS and DFS in fact spend work only on edges and nodes in the connected
component containing the starting node. (They never see any of the other
nodes or edges.) Thus the above algorithm, although it may run BFS or
DFS a number of times, only spends a constant amount of work on a given
edge or node in the iteration when the connected component it belongs to is
under consideration. Hence the overall running time of this algorithm is still
O(m + n).
3.4 Testing Bipartiteness: An Application of
Breadth-First Search
Recall the denition of a bipartite graph: it is one where the node set V can
be partitioned into sets X and Y in such a way that every edge has one end
in X and the other end in Y . To make the discussion a little smoother, we can
imagine that the nodes in the set X are colored red, and the nodes in the set
Y are colored blue. With this imagery, we can say a graph is bipartite if it is
possible to color its nodes red and blue so that every edge has one red end
and one blue end.
The Problem
In the earlier chapters, we saw examples of bipartite graphs. Here we start by
asking: What are some natural examples of a nonbipartite graph, one where
no such partition of V is possible?
3.4 Testing Bipartiteness: An Application of Breadth-First Search
Clearly a triangle is not bipartite, since we can color one node red, another
one blue, and then we cant do anything with the third node. More generally,
consider a cycle C of odd length, 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 oncoloring 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 a graph G is bipartite, then it cannot contain an odd cycle.
It is easy to recognize that a graph is bipartite when appropriate sets X
and Y (i.e., red and blue nodes) have actually been identied 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 bipartitethat is, whether there exists a
partition into red and blue nodes, as required. How difcult is this? We see from
(3.14) that an odd cycle is one simple obstacle to a graphs being bipartite.
Are there other, more complex obstacles to bipartitness?
Designing the Algorithm
In fact, there is a very simple procedure to test for bipartiteness, and its analysis
can be used to show that odd cycles are the only obstacle. First we assume
the graph G is connected, since otherwise we can rst 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, until 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 done: G
simply is not bipartite. We now want to argue this point precisely and also
work out an efcient way to perform the coloring.
The rst thing to notice is that the coloring procedure we have just
described is essentially identical to the description of BFS: we move outward
from s, coloring nodes as soon as we rst encounter them. Indeed, another
way to describe the coloring algorithm is as follows: we perform BFS, coloring
95
96
Chapter 3 Graphs
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[v] = red if i + 1 is an even number, and Color[v] = blue if i + 1 is an
odd number. At the end of this procedure, we simply scan all 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), just
as it is for BFS.
Analyzing the Algorithm
We now prove a claim that shows this algorithm correctly determines whether
G is bipartite, and it also shows that we can nd an odd cycle in G whenever
it is not bipartite.
(3.15) Let G be a connected graph, and let L1, L2 , . . . be the layers produced
by BFS starting at node s. Then exactly one of the following two things must
hold.
(i) 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.
The cycle through x, y,
and z has odd length.
s
Layer Li
Layer Lj
x
(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.
Proof. 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 joins nodes
either in the same layer or in adjacent layers. Our assumption for case (i) is
precisely that the rst 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.
z
y
Figure 3.6 If two nodes x and
y in the same layer are 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.
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 call 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 dened by following the z -x path in T , then the edge e ,
3.5 Connectivity in Directed Graphs
and then the y-z path in T . The length of this cycle is (j i) + 1 + (j i), adding
the length of its three parts separately; this is equal to 2(j i) + 1, which is an
odd number.
3.5 Connectivity in Directed Graphs
Thus far, we have been looking at problems on undirected graphs; we now
consider the extent to which these ideas carry over to the case of directed
graphs.
Recall that in a directed graph, the edge (u , v) has a direction: it goes from
u to v. In this way, the relationship between u and v is asymmetric, and this
has qualitative effects on the structure of the resulting graph. In Section 3.1, for
example, we discussed the World Wide Web as an instance of a large, complex
directed graph whose nodes are pages and whose edges are hyperlinks. The act
of browsing the Web is based on following a sequence of edges in this directed
graph; and the directionality is crucial, since its not generally possible to
browse backwards by following hyperlinks in the reverse direction.
At the same time, a number of basic denitions and algorithms have
natural analogues in the directed case. This includes the adjacency list representation and graph search algorithms such as BFS and DFS. We now discuss
these in turn.
Representing Directed Graphs
In order to represent a directed graph for purposes of designing algorithms,
we use a version of the adjacency 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-rst search and depth-rst search are almost the same in directed
graphs as they are in undirected graphs. We will focus here on BFS. We start
at a node s, dene a rst layer of nodes to consist of all those to which s has
an edge, dene a second layer to consist of all additional nodes to which these
rst-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-rst 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-rst 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 dene a new directed graph, G rev , that we obtain from G simply
by reversing the direction of every edge. We could then run BFS or DFS in G rev ;
a node has a path from s in G rev 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 denition; 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 reachability has a number of nice properties, many of them stemming from the following simple fact.
(3.16) If u and v are mutually reachable, and v and w are mutually reachable,
then u and w are mutually reachable.
Proof. To construct a path from u to w, we rst go from u to v (along the
path guaranteed by the mutual reachability of u and v), and then on from v
to w (along the path guaranteed by the mutual reachability of v and w). To
construct a path from w to u, we just reverse this reasoning: we rst go from
w to v (along the path guaranteed by the mutual reachability of v and w), and
then on from v to u (along the path guaranteed by the mutual reachability of
u and v).
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 G rev . Now, if one of
these two searches fails to reach every node, then clearly G is not strongly
connected. But suppose we nd that s has a path to every node, and that
3.6 Directed Acyclic Graphs and Topological Ordering
every node has a path to s. Then s and v are mutually reachable for every v,
and so it follows that every two nodes u and v are mutually reachable: s and
u are mutually reachable, and s and v are mutually reachable, so by (3.16) we
also have that u and v are mutually reachable.
By analogy with connected components in an undirected graph, we can
dene 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 G rev ; 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 well. 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 v would be mutually reachable, and v and t would be
mutually reachable, so from (3.16) it would follow that s and t were mutually
reachable.
In fact, although we will not discuss the details of this here, with more
work it is possible to compute the strong components for all nodes in a total
time of O(m + n).
3.6 Directed Acyclic Graphs and
Topological Ordering
If an undirected graph has no cycles, then it has an extremely simple structure:
each of its connected components is a tree. But it is possible for a directed graph
to have no (directed) cycles and still have a very rich structure. For example,
such graphs can have a large number of edges: if we start with the node
99
100
Chapter 3 Graphs
In a topological ordering, all
edges point from left to right.
v2
v6
v3
v5
v7
(a)
v4
v1
v2
v3
v4
v5
v6
v7
v1
(b)
(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 n edges but no cycles.
2
If a directed graph has no cycles, we call itnaturally enougha directed
acyclic 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
jobs, with assertions that the output of job i is used in determining the input
to job j , and hence job i must be done before job 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 rst.
3.6 Directed Acyclic Graphs and Topological Ordering
Lets continue a little further with this picture of DAGs as precedence
relations. Given a set of tasks with dependencies, it would be natural to seek
a valid order in which the tasks could be performed, so that all dependencies
are respected. Specically, for a directed graph G , we say that a topological
ordering of G is an ordering of its nodes as v1, v2 , . . . , vn so that for every edge
(vi , vj ), 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
v1, v2 , . . . , vn , and also has a cycle C . Let vi be the lowest-indexed node on C ,
and let vj be the node on C just before vi thus (vj , vi ) is an edge. But by our
choice of i, we have j > i, which contradicts the assumption that v1, v2 , . . . , vn
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 The main question we consider here is
the converse of (3.18): Does every DAG have a topological ordering, and if so,
how do we nd one efciently? 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 efciently computable order in which to perform
the tasks.
Designing and Analyzing the Algorithm
In fact, the converse of (3.18) does hold, and we establish this via an efcient
algorithm to compute a topological ordering. The key to this lies in nding a
way to get started: which node do we put at the beginning of the topological
ordering? Such a node v1 would need to have no incoming edges, since any
such incoming edge would violate the dening 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 nd a cycle in G ; this will prove the claim. We pick
any node v, and begin following edges backward from v: since 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 indenitely, since every node we encounter has an
incoming edge. But after n + 1 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.
In fact, the existence of such a node v is all we need to produce a topological
ordering of G by induction. Specically, 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 nd a node v with no incoming edges,
as guaranteed by (3.19). We place v rst 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)
If 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 v
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
103
3.6 Directed Acyclic Graphs and Topological Ordering
v2
v6
v3
v5
v7
v2
v4
v1
v3
v6
v5
v5
v7
v4
v7
v6
v5
v7
(a)
v6
v3
(b)
v4
v6
(c)
v5
v6
v7
(d)
v7
(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 topological 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 of O(m + n) could be a signicant improvement over (n2).
In fact, we can achieve a running time of O(m + n) using the same highlevel algorithmiteratively deleting nodes with no incoming edges. We simply
have to be more efcient in nding these nodes, and we do this as follows.
We declare a node to be active if it has not yet been deleted by the
algorithm, and we explicitly maintain two things:
(a) for each node w, the number of incoming edges that w has from active
nodes; and
(b) the set S of all active nodes in G that have no incoming edges from other
active nodes.
v4
104
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 v from the set S and deleting it. After deleting v, we go through all
nodes w to which v 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 to w to drop to zero, then we add w 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.
b
a
e
Solved Exercises
c
d
Solved Exercise 1
Figure 3.9 How many topological orderings does this
graph have?
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 v1, 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 2
1 = 120 possible orderings and check whether each is a topological ordering.
But this would take a while.
Instead, we think about this as follows. As we saw in the text (or reasoning
directly from the denition), the rst 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 rst and the node e must come last.
Now we have to gure 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 all 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
Solved Exercises
with a base station, and your friends nd 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 oor 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 species that one of the robots moves across a
single edge, from one node to a neighboring node; at the end of the schedule,
the robot from node a should be sitting on c , and the robot from b should be
sitting on d.
A schedule is interference-free if there is no point at which the two robots
occupy nodes that are at a distance r from one another in the graph, for a
given parameter r . Well assume that the two starting nodes a and b are at a
distance greater than r , and so are the two ending nodes c and d.
Give a polynomial-time algorithm that decides whether there exists an
interference-free schedule by which each robot can get to its destination.
Solution This is a problem of the following general avor. We have a set
of possible congurations for the robots, where we dene a conguration
to be a choice of location for each one. We are trying to get from a given
starting conguration (a , b) to a given ending conguration (c , d), subject to
constraints on how we can move between congurations (we can only change
one robots location to a neighboring node), and also subject to constraints on
which congurations 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 conguration of the robotsthat is, the
current location of each oneits 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-nding problem,
not in the original graph G but in the space of all possible congurations.
Let us dene the following (larger) graph H . The node set of H is the set
of all possible congurations of the robots; that is, H consists of all possible
pairs of nodes in G . We join two nodes of H by an edge if they represent
congurations that could be consecutive in a schedule; that is, (u , v) and
(u , v ) will be joined by an edge in H if one of the pairs u , u or v , v are equal,
and the other pair corresponds to an edge in G .
105
106
Chapter 3 Graphs
We can already observe that paths in H from (a , b) to (c , d) correspond
to schedules for the robots: such a path consists precisely of a sequence of
congurations 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 congurations in which there would be interference. Thus we dene 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 connectivity 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 congurations 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 n2 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 , v), 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 rst build H by
enumerating all pairs of nodes in G in time O(n2), and constructing edges
using the denition above in time O(n) per node, for a total of O(n3).
Now we need to gure 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 breadthrst search from u and identify all nodes v within distance r of u. We list
all these pairs (u , v) and delete them from H . Each breadth-rst search
in G takes time O(m + n), and were doing one from each node, so the
total time for this part is O(mn + n2).
107
Exercises
3. Now we have H , and so we just need to decide whether there is a path
from (a , b) to (c , d). This can be done using the connectivity algorithm
from the text in time that is linear in the number of nodes and edges
of H . Since H has O(n2) nodes and O(n3) edges, this nal step takes
polynomial time as well.
Exercises
1. Consider the directed acyclic graph G in Figure 3.10. How many topological orderings does it have?
2. Give an algorithm to detect whether a given undirected graph contains
a cycle. If the graph contains a cycle, then your algorithm should output
one. (It should not output all cycles in the graph, just one of them.) The
running time of your algorithm should be O(m + n) for a graph with n
nodes and m edges.
3. The algorithm described in Section 3.6 for computing a topological ordering of a DAG repeatedly 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 G is a DAG; or (b) a cycle in G , thus
establishing that G is not a DAG. The running time of your algorithm
should be O(m + n) for a directed graph with n nodes and m edges.
4. Inspired by the example of that great Cornellian, Vladimir Nabokov, some
of your friends have become amateur lepidopterists (they study butterflies). Often when they return from a trip with specimens of butterflies,
it is very difficult for them to tell how many distinct species theyve
caughtthanks to the fact that many species look very similar to one
another.
One day they return with n butterflies, and they 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 groupsthose that belong to A and those that belong to Bbut its
very hard for them to directly label any one specimen. So they decide to
adopt the following approach.
b
c
a
f
d
e
Figure 3.10 How many topological orderings does this
graph have?
108
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 label the
pair (i , j ) 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 well call the pair ambiguous .
So now they have the collection of n specimens, as well 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 B in such a way that for each pair (i , j )
labeled same, it is the case that i and j have the same label; and for each
pair (i , j ) labeled different, it is the case that i and j 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.
5. 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.
6. We have a connected graph G = (V , E ), and a specific vertex u V . Suppose
we compute a depth-first search tree rooted at u, and obtain a tree T that
includes all nodes of G . Suppose we then compute a breadth-first search
tree rooted at u, 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 u, then G cannot contain any edges that do not belong to
T .)
7. 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 if the
physical locations of i and j are no more than 500 meters apart. (If so, we
say that i and j 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
Exercises
the following property: at all times, each 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 n/2, 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.
8. 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 u and v
in a graph G = (V , E ) is the minimum number of edges in a path joining
them; well denote this by dist (u , v). We say that the diameter of G is
the maximum distance between any pair of nodes; and well denote this
quantity by diam(G ).
Lets define a related quantity, which well call the average pairwise
distance in G (denoted apd(G )). We define apd(G ) to be the average, over
all n sets of two distinct nodes u and v, of the distance between u and v.
2
That is,
n
dist (u , v) /
.
apd(G ) =
2
{u , v}V
Heres a simple example to convince yourself that there are graphs G
for which diam(G ) = apd(G ). Let G be a graph with three nodes u , v , w, and
with the two edges {u , v} and {v , w}. Then
diam(G ) = dist (u , w) = 2,
while
apd(G ) = [dist (u , v) + dist (u , w) + dist (v , w)]/3 = 4/3.
109
110
Chapter 3 Graphs
Of course, these two numbers arent all that far apart in the case of
this three-node graph, and so its natural to ask whether theres always a
close 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 )
c.
apd(G )
Decide whether you think the claim is true or false, and give a proof of
either the claim or its negation.
9. Theres a natural intuition that two nodes that are far apart in a communication networkseparated by many hopshave 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 susceptibility 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 v, 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 running time O(m + n) to find such a node v.
10. A number of art museums around the country have been featuring work
by an artist named Mark Lombardi (19512000), consisting of a set of
intricately rendered graphs. Building on a great deal of research, these
graphs encode the relationships among people involved in major political
scandals over the past several decades: the nodes correspond to participants, and each edge indicates some type of relationship between a pair
of participants. And so, if you peer closely 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.
Exercises
Of course, a single, spurious short path between nodes v and w in
such a network may be more coincidental than anything else; a large
number of short paths between v 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 v-w paths.
This turns out to be a problem that can be solved efficiently. Suppose
we are given an undirected graph G = (V , E ), and we identify two nodes v
and w in G . Give an algorithm that computes the number of shortest v-w
paths in G . (The algorithm should not list all the paths; just the number
suffices.) The running time of your algorithm should be O(m + n) for a
graph with n nodes and m edges.
11. Youre helping some security analysts monitor a collection of networked
computers, tracking the spread of an online virus. There are n computers
in the system, labeled C1, C2 , . . . , Cn , and as input youre given a collection
of trace data indicating the times at which pairs of computers communicated. Thus the data is a sequence of ordered triples (Ci , Cj , tk ); such a
triple indicates that Ci and Cj exchanged bits at time tk . There are m triples
total.
Well assume that the triples 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 tk (in other
words, if one of the triples (Ci , Cj , tk ) or (Cj , Ci , tk ) appears in the trace
data), then computer Cj becomes infected as well, starting at time tk .
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 , tk ) and (Cj , Cq , tr ), where tk tr ,
then Cq will become infected via Cj . (Note that it is okay for tk to be equal
to tr ; 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 Cq .)
For example, suppose n = 4, the trace data consists of the triples
(C1, C2 , 4) , (C2 , C4 , 8) , (C3 , C4 , 8) , (C1, 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 infected
at time 4, then C4 gets the virus from C2 at time 8, and then C3 gets the
virus from C4 at time 8. On the other hand, if the trace data were
(C2 , C3 , 8) , (C1, C4 , 12) , (C1, C2 , 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
C2 becomes infected at time 14, we see that C3 only communicates with C2
before C2 was infected. There is no sequence of communications moving
forward in time by which the virus could get from C1 to C3 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 should run in time O(m + n).
12. Youre helping a group of ethnographers analyze some oral history data
theyve collected by interviewing members of a village to learn about the
lives of people whove lived there over the past two hundred years.
From these interviews, theyve learned about a set of n people (all
of them now deceased), whom well denote P1, P2 , . . . , Pn . 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 Pi died before person Pj was born; or
.
for some i and j , the life spans of Pi 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 existthat
is, the facts collected by the ethnographers are not internally consistent.
Notes and Further Reading
The theory of graphs is a large topic, encompassing both algorithmic and nonalgorithmic issues. It is generally considered to have begun with a paper by
Notes and Further Reading
Euler (1736), grown through interest in graph representations of maps and
chemical compounds in the nineteenth century, and emerged as a systematic
area of study in the twentieth century, rst 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
This page intentionally left blank
Chapter 4
Greedy Algorithms
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 dene precisely what is meant by a greedy
algorithm. An algorithm is greedy if it builds up a solution in small 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 rule 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; nding cases in which they
work well, and proving that they work well, is the interesting challenge.
The rst 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 rst approach as establishing that the greedy algorithm stays
ahead. By this we mean that if one measures the greedy algorithms progress
116
Chapter 4 Greedy Algorithms
in a step-by-step fashion, one sees that it does better than any other algorithm
at each step; it then follows that it produces an optimal solution. The second
approach is known as an exchange argument , and it is more general: one
considers any possible solution to the problem and gradually transforms it
into the solution found by the greedy algorithm without hurting its quality.
Again, it will follow that the greedy algorithm must have found a solution that
is at least as good as any other solution.
Following our introduction of these two styles of analysis, we focus on
several of the most well-known applications of greedy algorithms: shortest
paths in a graph, the Minimum Spanning Tree Problem, and the construction of Huffman codes for performing data compression. They each provide
nice examples of our analysis techniques. We also explore an interesting relationship between minimum spanning trees and the long-studied problem of
clustering . Finally, we consider a more complex application, the MinimumCost Arborescence Problem, which further extends our notion of what a greedy
algorithm is.
4.1 Interval Scheduling: The Greedy Algorithm
Stays Ahead
Lets recall the Interval Scheduling Problem, which was the rst of the ve
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 nishing 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.
Designing a Greedy Algorithm
Using the Interval Scheduling Problem, we can make our discussion of greedy
algorithms much more concrete. The basic idea in a greedy algorithm for
interval scheduling is to use a simple rule to select a rst request i1. Once
a request i1 is accepted, we reject all requests that are not compatible with i1.
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 challenge in designing a good greedy algorithm is in deciding
which simple rule to use for the selectionand 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.
4.1 Interval Scheduling: The Greedy Algorithm Stays Ahead
.
The most obvious rule might be to always select the available request
that starts earliestthat 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 casesay, when the nish time f (i) is the maximum
among all requeststhe 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 timenamely, 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 4.1(b), accepting the short interval in
the middle would prevent us from accepting the other two, which form
an optimal solution.
(a)
(b)
(c)
Figure 4.1 Some instances of the Interval Scheduling Problem on which natural greedy
algorithms fail to find the optimal solution. In (a), it does not work to select the interval
that starts earliest; in (b), it does not work to select the shortest interval; and in (c), it
does not work to select the interval with the fewest conflicts.
117
118
Chapter 4 Greedy Algorithms
In the previous greedy rule, our problem was that the second request
competes with both the rst and the thirdthat 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 conicts.) 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 rst the request that nishes rst, that is, the request 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 not yet empty
Choose a request i R that has the smallest finishing time
Add request i to A
Delete all requests from R that are not compatible with request i
EndWhile
Return the set A as the set of accepted requests
Analyzing the Algorithm
While this greedy method is quite natural, it is certainly not obvious that it
returns an optimal set of intervals. Indeed, it would only be sensible to reserve
judgment on its optimality: the ideas that led to the previous nonoptimal
versions of the greedy method also seemed promising at rst.
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.
119
4.1 Interval Scheduling: The Greedy Algorithm Stays Ahead
6
Intervals numbered in order
1
3
2
8
5
4
9
7
8
Selecting interval 1
1
3
5
4
9
7
8
Selecting interval 3
1
3
5
9
7
8
Selecting interval 5
1
3
5
1
3
9
5
8
Selecting interval 8
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 O be an optimal set of intervals. Ideally one might want to show
that A = O, 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| = |O|, that is, that A contains the same number of intervals as O and
hence is also an optimal solution.
The idea underlying the proof, as we suggested initially, will be to nd
a sense in which our greedy algorithm stays ahead of this solution O. We
will compare the partial solutions that the greedy algorithm constructs to initial
segments of the solution O, 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 i1, . . . , ik be the set
of requests in A in the order they were added to A. Note that |A| = k. Similarly,
let the set of requests in O be denoted by j1, . . . , jm . Our goal is to prove that
k = m. Assume that the requests in O are also ordered in the natural left-toright order of the corresponding intervals, that is, in the order of the start and
nish points. Note that the requests in O are compatible, which implies that
the start points have the same order as the nish points.
120
Chapter 4 Greedy Algorithms
Can the greedy algorithms
r th interval really finish later?
ir 1
ir ?
jr 1
jr
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 free again as soon as possible after satisfying the rst request. And
indeed, our greedy rule guarantees that f (i1) f (j1). This is the sense in which
we want to show that our greedy rule stays aheadthat each of its intervals
nishes at least as soon as the corresponding interval in the set O. Thus we now
prove that for each r 1, the r th accepted request in the algorithms schedule
nishes no later than the r th request in the optimal schedule.
(4.2)
For all indices r k we have f (ir ) f (jr ).
Proof. We will prove this statement by induction. For r = 1 the statement is
clearly true: the algorithm starts by selecting the request i1 with minimum
nish 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 r th interval not to nish 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-nishing interval, the greedy algorithm
always has the option (at worst) of choosing jr and thus fullling the induction
step.
We can make this argument precise as follows. We know (since O 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 ir .
The greedy algorithm selects the available interval with smallest nish time;
since interval jr is one of these available intervals, we have f (ir ) f (jr ). This
completes the induction step.
Thus we have formalized the sense in which the greedy algorithm is
remaining ahead of O: for each r , the r th interval it selects nishes at least
as soon as the r th interval in O. We now see why this implies the optimality
of the greedy algorithms set A.
4.1 Interval Scheduling: The Greedy Algorithm Stays Ahead
(4.3)
The greedy algorithm returns an optimal set A.
Proof. We will prove the statement by contradiction. If A is not optimal, then
an optimal set O 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+1 in O. This request starts after request jk ends, and hence after
ik ends. So after deleting all requests that are not compatible with requests
i1, . . . , ik , the set of possible requests R still contains jk+1. But the greedy
algorithm stops with request ik , and it is only supposed to stop when R is
emptya 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 nishing
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 rst interval; we then iterate through the intervals in
order until reaching the rst interval j 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 rst j for which s(j ) f . In this way, we implement the greedy algorithm
analyzed above in one pass through the intervals, spending constant time per
interval. Thus this part of the algorithm takes time O(n).
Extensions
The Interval Scheduling Problem we considered here is a quite simple scheduling problem. There are many further complications that could arise in practical
settings. The following point out issues that we will see later in the book in
various forms.
.
In dening the problem, we assumed that all requests were known to
the scheduling algorithm when it was choosing the compatible subset.
It would also be natural, of course, to think about the version of the
problem in which the scheduler needs to make decisions about accepting
or rejecting certain requests before knowing about the full set of requests.
Customers (requestors) may well be impatient, and they may give up
and leave if the scheduler waits too long to gather information about all
other requests. An active area of research is concerned with such online algorithms, which must make decisions as time proceeds, without
knowledge of future input.
121
122
Chapter 4 Greedy Algorithms
.
Our goal was to maximize the number of satised 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 satised 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 have
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.1
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 all 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 jobs that need to be processed for
a specic 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 ber-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 rst row contains all the intervals
1
The problem is also referred to as the Interval Coloring Problem; the terminology arises from
thinking of the dierent resources as having distinct colorsall the intervals assigned to a particular
resource are given the corresponding color.
4.1 Interval Scheduling: The Greedy Algorithm Stays Ahead
e
c
j
d
g
b
h
f
i
f
a
j
(a)
c
d
g
b
a
e
i
h
(b)
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 rst 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 dene 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 I1, . . . , 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 efcient 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 locala set of intervals all piled over the same point. Its
not immediately clear that there couldnt exist other, long-range obstacles
that push the number of required resources even higher.
123
124
Chapter 4 Greedy Algorithms
We now design a simple greedy algorithm that schedules all intervals
using a number of resources equal to the depth. This immediately implies the
optimality of the algorithm: in view of (4.4), no solution could use a number
of resources that is smaller than the depth. The analysis of our algorithm
will therefore illustrate another general approach to proving optimality: one
nds 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. Specically, we have
the following description.
Sort the intervals by their start times, breaking ties arbitrarily
Let I1, I2 , . . . , In denote the intervals in this order
For j = 1, 2, 3, . . . , n
For each interval Ii that precedes Ij in sorted order and overlaps it
Exclude the label of Ii from consideration for Ij
Endfor
If there is any label from {1, 2, . . . , d} that has not been excluded then
Assign a nonexcluded label to Ij
Else
Leave Ij unlabeled
Endif
Endfor
Analyzing the Algorithm We claim the following.
(4.5) If we use the greedy algorithm above, every interval will be assigned a
label, and no two overlapping intervals will receive the same label.
Proof. First lets argue that no interval ends up unlabeled. Consider one of
the intervals Ij , and suppose there are t intervals earlier in the sorted order
that overlap it. These t intervals, together with Ij , form a set of t + 1 intervals
that all pass over a common point on the time-line (namely, the start time of
4.2 Scheduling to Minimize Lateness: An Exchange Argument
Ij ), and so t + 1 d. Thus t d 1. It follows that at least one of the d labels
is not excluded by this set of t intervals, and so there is a label that can be
assigned to Ij .
Next we claim that no two overlapping intervals are assigned the same
label. Indeed, consider any two intervals I and I that overlap, and suppose I
precedes I in the sorted order. Then when I is considered by the algorithm,
I is in the set of intervals whose labels are excluded from consideration;
consequently, the algorithm will not assign to I the label that it used for I .
The algorithm and its analysis are very simple. Essentially, if you have
d labels at your disposal, then as you sweep through the intervals from left
to right, assigning an available label to each interval you encounter, you can
never reach a point where all the labels are currently in use.
Since our algorithm is using d labels, we can use (4.4) to conclude that it
is, in fact, always using the minimum possible number of labels. We sum this
up as follows.
(4.6) The greedy algorithm above schedules every interval on a resource,
using a number of resources equal to the depth of the set of intervals. This
is the optimal number of resources needed.
4.2 Scheduling to Minimize Lateness: An Exchange
Argument
We now discuss a scheduling 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 exible. Instead of a start time and nish 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 ti , and different requests
must be assigned nonoverlapping intervals.
There are many objective functions we might seek to optimize when faced
with this situation, and some are computationally much more difcult than
125
126
Chapter 4 Greedy Algorithms
Length 1
Deadline 2
Job 1
Length 2
Deadline 4
Job 2
Deadline 6
Length 3
Job 3
Solution:
Job 1:
done at
time 1
Job 2:
done at
time 1+2=3
Job 3:
done at
time 1+2+3=6
Figure 4.5 A sample instance of scheduling to minimize lateness.
others. Here we consider a very natural goal that can be optimized by a greedy
algorithm. Suppose that we plan to satisfy each request, but we are allowed
to let certain requests run late. Thus, beginning at our overall start time s, we
will assign each request i an interval of time of length ti ; let us denote this
interval by [s(i) , f (i)], with f (i) = s(i) + ti . Unlike the previous problem, then,
the algorithm must actually determine a start time (and hence a nish 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 dened to be li = f (i) di . We will 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
jobs: the rst has length t1 = 1 and deadline d1 = 2; the second has t2 = 2
and d2 = 4; and the third has t3 = 3 and d3 = 6. It is not hard to check that
scheduling the jobs in the order 1, 2, 3 incurs a maximum lateness of 0.
Designing the Algorithm
What would a greedy algorithm for this problem look like? There are several
natural greedy approaches in which we look at the data (ti , di ) about the jobs
and use this to order them according to some simple rule.
.
One approach would be to schedule the jobs in order of increasing length
ti , so as to get the short jobs out of the way quickly. This immediately
4.2 Scheduling to Minimize Lateness: An Exchange Argument
looks too simplistic, since it completely ignores the deadlines of the jobs.
And indeed, consider a two-job instance where the rst job has t1 = 1 and
d1 = 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 rst is indeed the optimal solution.
.
The previous example suggests that we should be concerned about jobs
whose available slack time di ti is very smalltheyre 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 rst job has t1 = 1 and d1 = 2, while the second job has
t2 = 10 and d2 = 10. Sorting by increasing slack would place the second
job rst in the schedule, and the rst job would incur a lateness of 9. (It
nishes at time 11, nine units beyond its deadline.) On the other hand,
if we schedule the rst job rst, then it nishes 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
deadlines di , and schedule them in this order. (This rule 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
specically 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
d1 . . . 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 f (1) = s(1) + t1; Job 2
will start at time s(2) = f (1) and end at time f (2) = s(2) + t2; and so forth. We
will use f to denote the nishing 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 d1 . . . dn
Initially, f = s
127
128
Chapter 4 Greedy Algorithms
Consider the jobs i = 1, . . . , n in this order
Assign job i to the time interval from s(i) = f to f (i) = f + ti
Let f = f + ti
End
Return the set of scheduled intervals [s(i) , f (i)] for i = 1, . . . , n
Analyzing the Algorithm
To reason about the optimality of the algorithm, we rst observe that the
schedule it produces has no gapstimes when the machine is not working
yet there are jobs left. The time that passes during a gap will be called idle
time: there is work to be done, yet for some reason the machine is sitting idle.
Not only does the schedule A produced by our algorithm have no idle time;
it is also very easy to see that there is an optimal schedule with this property.
We do not write down a proof for this.
(4.7)
There is an optimal schedule with no idle time.
Now, how can we prove that our schedule A is optimal, that is, its
maximum lateness L is as small as possible? As in previous analyses, we will
start by considering an optimal schedule O. Our plan here is to gradually
modify O, 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 rst 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 dj < di . Notice that, by denition, 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.
4.2 Scheduling to Minimize Lateness: An Exchange Argument
The main step in showing the optimality of our algorithm is to establish
that there is an optimal schedule that has no inversions and no idle time. To 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 scheduling after this conversion will be optimal
as well.
(4.9)
There is an optimal schedule that has no inversions and no idle time.
Proof. By (4.7), there is an optimal schedule O with no idle time. The proof
will consist of a sequence of statements. The rst of these is simple to establish.
(a) If O has an inversion, then there is a pair of jobs i and j such that j is
scheduled immediately after i and has dj < di .
Indeed, consider an inversion in which a job a is scheduled sometime before
a job 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 rst time. This corresponds to a pair of consecutive jobs that form an
inversion.
Now suppose O has at least one inversion, and by (a), let i and j be a pair of
inverted requests that are consecutive in the scheduled order. We will decrease
the number of inversions in O by swapping the requests i and j in the schedule
O. The pair (i , j ) formed an inversion in O, this inversion is eliminated by the
swap, and no new inversions are created. Thus we have
(b) After swapping i and j 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 O
can have at most n inversions (if all pairs are inverted), and hence after at
2
most n swaps we get an optimal schedule with no inversions.
2
So we now conclude by proving (c), showing that by 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 O: assume
that each request r is scheduled for the time interval [s(r ) , f (r )] and has
lateness lr . Let L = maxr lr denote the maximum lateness of this schedule.
129
130
Chapter 4 Greedy Algorithms
Only the finishing times of i and j
are affected by the swap.
Before swapping:
Job i
dj
Job j
di
(a)
After swapping:
Job j
dj
Job i
di
(b)
Figure 4.6 The effect of swapping two consecutive, inverted jobs.
Let O denote the swapped schedule; we will use s(r ), f (r ), lr , and L to denote
the corresponding quantities in the swapped schedule.
Now recall our two adjacent, inverted jobs i and j . The situation is roughly
as pictured in Figure 4.6. The nishing time of j before the swap is exactly equal
to the nishing time of i after the swap. Thus all jobs other than jobs i and j
nish at the same time in the two schedules. Moreover, job j will get nished
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 nishes at time f (j ), when job j was
nished in the schedule O. If job i is late in this new schedule, its lateness
is li = f (i) di = f (j ) di . But the crucial point is that i cannot be more late
in the schedule O than j was in the schedule O. Specically, our assumption
di > dj implies that
li = f (j ) di < f (j ) dj = lj .
Since the lateness of the schedule O was L lj > li , this shows that the swap
does not increase the maximum lateness of the schedule.
The optimality of our greedy algorithm now follows immediately.
4.3 Optimal Caching: A More Complex Exchange Argument
(4.10) The schedule A produced by the greedy algorithm has optimal maximum lateness L.
Proof. Statement (4.9) proves that an optimal schedule with no inversions
exists. 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.
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 di and the requested time ti , 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 naturally
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 nds 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 will see later in the book, in Chapter 8, this more general
version of the problem is much more difcult to solve optimally.
4.3 Optimal Caching: A More Complex Exchange
Argument
We now consider a problem that involves processing a sequence of requests
of a different form, and we develop an algorithm whose analysis requires
a more subtle use of the exchange argument. The problem is that of cache
maintenance .
The Problem
To motivate caching, consider the following situation. Youre working on a
long research paper, and your draconian library will only allow you to have
eight books checked out at once. You know that youll probably need more
than this over the course of working on the paper, but at any point in time,
youd like to have ready access to the eight books that are most relevant at
that time. 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,
131
132
Chapter 4 Greedy Algorithms
and a large amount of data that requires more time to access; and you must
decide which pieces of data to have close at hand.
Memory hierarchies have been a ubiquitous feature of computers since
very early in their history. To begin with, data in the main memory of a
processor can be accessed much more quickly than the data on its hard disk;
but the disk has much more storage capacity. Thus, it is important to keep
the most regularly used pieces of 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: small 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 still much faster than
downloading something over the Internet.
Caching is a general term for the process of storing a small amount of data
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 looking 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 = d1, d2 , . . . , dm drawn from U is
presented to usthis 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 full, to evict some other piece of data that is
currently in the cache to make room for di . This is called a cache miss, and we
want to have as few of these as possible.
4.3 Optimal Caching: A More Complex Exchange Argument
Thus, on a particular sequence of memory references, a cache maintenance algorithm determines an eviction schedule specifying which items
should be evicted from the cache at which points in the sequenceand this
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 d1, 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 full sequence S of memory
references, what is the eviction schedule that incurs as few cache misses as
possible?
Designing and Analyzing the Algorithm
In the 1960s, Les Belady showed that the following simple rule will always
incur the minimum number of misses:
When di needs to be brought into the cache,
evict the item that is needed the farthest into the future
We will call this the Farthest-in-Future Algorithm. When it is time to evict
something, we look at the next time that each item in the cache will 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 rst appears. Why evict the
item that is needed farthest in the future, as opposed, for example, to the one
that will be used least frequently in the future? Moreover, consider a sequence
like
a , b, c , d , a , d , e , a , d , b, c
133
134
Chapter 4 Greedy Algorithms
with k = 3 and items {a , b, c } initially in the cache. The Farthest-in-Future rule
will produce a schedule S that evicts c on the fourth step and b on the seventh
step. But there are other eviction schedules that are just as good. Consider
the schedule S that evicts b on the fourth step and c on the seventh step,
incurring the same number of misses. So in fact its easy to nd cases where
schedules produced by rules other than Farthest-in-Future are also optimal;
and given this exibility, 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 rst, we can swap
the choices of b and c without changing the cost. This reasoningswapping
one decision for anotherforms the rst outline of an exchange argument that
proves the optimality of Farthest-in-Future.
Before delving into this analysis, lets clear up one important issue. All
the cache maintenance algorithms weve been considering so far produce
schedules that only bring an item d into the cache in a step i if there is a
request to d in step i, and d is not already in the cache. Let us call such a
schedule reducedit 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 dene a new schedule
Sthe reduction of Sas follows. In any step i where S brings in an item d
that has not been requested, our construction of S 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 S 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) S is a reduced schedule that brings in at most as many items as the
schedule S.
Note that for any reduced schedule, the number of items that are brought
in is exactly the number of misses.
4.3 Optimal Caching: A More Complex Exchange Argument
Proving the Optimalthy of Farthest-in-Future We now proceed with the
exchange argument showing that Farthest-in-Future is optimal. Consider an
arbitrary sequence D of memory references; let SFF 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 SFF , 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 schedule that makes the same eviction decisions
as SFF through the rst j items in the sequence, for a number j. Then there is a
reduced schedule S that makes the same eviction decisions as SFF through the
rst j + 1 items, and incurs no more misses than S does.
Proof. Consider the (j + 1)st request, to item d = dj+1. Since S and SFF 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 SFF through step j + 1, and we can set S = S.
Similarly, if d needs to be brought into the cache, but S and SFF 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 SFF evicts item e = f . Here S and SFF do
not already agree through step j + 1 since S has e in cache while SFF has f in
cache. Hence we must actually do something nontrivial to construct S .
As a rst 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 nish the construction of S by just having it
behave like S.
Specically, from request j + 2 onward, S behaves exactly like S until one
of the following things happens for the rst 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
Chapter 4 Greedy Algorithms
of S and S will be the same. If e = e , then we have S evict e as well, and
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 nish
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 rst j + 1 items and incurs no more misses than S does.
And cruciallyhere is where we use the dening property of the Farthest-inFuture Algorithmone of these two cases will arise before there is a reference
to e . This is because in step j + 1, Farthest-in-Future evicted the item (e ) that
would be needed farthest in the future; so before there could be a request to
e , there would have to be a request to f , and then case (ii) above would apply.
Using this result, it is easy to complete the proof of optimality. We begin
with an optimal schedule S, and use (4.12) to construct a schedule S1 that
agrees with SFF through the rst step. We continue applying (4.12) inductively
for j = 1, 2, 3, . . . , m, producing schedules Sj that agree with SFF through the
rst j steps. Each schedule incurs no more misses than the previous one; and
by denition Sm = SFF , since it agrees with it through the whole sequence.
Thus we have
(4.13) SFF 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 y 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 reversedlongest 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.
4.4 Shortest Paths in a Graph
Long after the adoption of LRU in practice, Sleator and Tarjan showed that
one could actually provide some theoretical analysis of the performance of
LRU, bounding the number of misses it incurs relative to Farthest-in-Future.
We will discuss this analysis, as well as the analysis of a randomized variant
on LRU, when we return to the caching problem in Chapter 13.
4.4 Shortest Paths in a Graph
Some of the basic algorithms for graphs are based on greedy design principles.
Here we apply a greedy algorithm to the problem of nding shortest paths, and
in the next section we look at the construction of minimum-cost spanning trees.
The Problem
As weve seen, graphs are often used to model networks in which one travels from one point to anothertraversing a sequence of highways through
interchanges, or traversing a sequence of communication links through intermediate routers. As a result, a basic algorithmic problem is to determine the
shortest path between nodes in a graph. We may ask this as a point-to-point
question: Given nodes u and v, what is the shortest u-v path? Or we may ask
for more information: Given a start node s, what is the shortest path from s to
each other node?
The concrete setup of the shortest paths problem is as follows. We are
given a directed graph G = (V , E ), with a designated start node s. We assume
that s has a path to every other node in G . Each edge e has a length e 0,
indicating the time (or distance, or cost) it takes to traverse e . For a path P ,
the length of P denoted (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 specied for a directed
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 (v , u),
each of length e .
Designing the Algorithm
In 1959, Edsger Dijkstra proposed a very simple greedy algorithm to solve the
single-source shortest-paths problem. We begin by describing an algorithm that
just determines the length of the shortest path from s to each other node in the
graph; it is then easy to produce the paths as well. The algorithm maintains a
set S of vertices u for which we have determined a shortest-path distance d(u)
from s; this is the explored part of the graph. Initially S = {s}, and d(s) = 0.
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 S, followed by the single edge (u , v). That is, we consider the quantity
137
138
Chapter 4 Greedy Algorithms
d (v) = mine=(u , v):uS d(u) + e . We choose the node v V S for which this
quantity is minimized, add v to S, and dene d(v) to be the value d (v).
Dijkstras Algorithm (G , )
Let S be the set of explored nodes
For each u S, we store a distance d(u)
Initially S = {s} and d(s) = 0
While S = V
Select a node v S with at least one edge from S for which
d (v) = mine=(u , v):uS d(u) + e is as small as possible
Add v to S and define d(v) = d (v)
EndWhile
It is simple to produce the s-u paths corresponding to the distances found
by Dijkstras Algorithm. As each node v is added to the set S, we simply record
the edge (u , v) on which it achieved the value mine=(u , v):uS d(u) + e . The
path Pv is implicitly represented by these edges: if (u , v) is the edge we have
stored for v, then Pv is just (recursively) the path Pu followed by the single
edge (u , v). In other words, to construct Pv , we simply start at v; follow the
edge we have stored for v in the reverse direction to u; then follow the edge we
have stored for u 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 v 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 rst added node u, and the second added
node v. In the iteration that is about to be performed, the node x will be added
because it achieves the smallest value of d (x); thanks to the edge (u , x), we
have d (x) = d(u) + lux = 2. Note that attempting to add y or z to the set S at
this point would lead to an incorrect value for their shortest-path distances;
ultimately, they will be added because of their edges from x.
Analyzing the Algorithm
We see in this example that Dijkstras Algorithm is doing the right 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 v?
We now answer this by proving the correctness of the algorithm, showing
that the paths Pu really are shortest paths. Dijkstras Algorithm is greedy in
4.4 Shortest Paths in a Graph
y
3
u
1
1
1
4
s
2
Set S:
nodes already
explored
x
2
2
v
3
z
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-v path we can make from a
path in S followed by a single edge. We prove its correctness using a variant of
our rst 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 v, 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
Algorithm, 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 |S| = 1 is easy,
since then we have S = {s} and d(s) = 0. Suppose the claim holds when |S| = k
for some value of k 1; we now grow S to size k + 1 by adding the node v.
Let (u , v) be the nal edge on our s-v path Pv .
By induction hypothesis, Pu is the shortest s-u path for each u S. Now
consider any other s-v path P ; we wish to show that it is at least as long as Pv .
In order to reach v, this path P must leave the set S somewhere ; let y be the
rst 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 Pv because it is already at least as
139
140
Chapter 4 Greedy Algorithms
x
y
P
The alternate sv path P through
x and y is already too long by
the time it has left the set S.
s
Pu
Set S
u
v
Figure 4.8 The shortest path Pv and an alternate s-v path P through the node y.
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 v. 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 Pv . Since edge lengths
are nonnegative, the full path P is at least as long as Pv 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 (P ) (Px ) = d(x). Thus the subpath
of P out to node y has length (P ) + (x , y) d(x) + (x , y) d (y), and the
full path P is at least as long as this subpath. Finally, since Dijkstras Algorithm
selected v in this iteration, we know that d (y) d (v) = (Pv ). Combining these
inequalities shows that (P ) (P ) + (x , y) (Pv ).
Here are two observations about Dijkstras Algorithm and its analysis.
First, the algorithm does not always nd 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 algorithmdue to Bellman and Fordis 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-rst 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 lled with water, joined together at
the nodes; each edge e has length e and a xed 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
4.4 Shortest Paths in a Graph
of wavefront reaches nodes in increasing order of their distance from s. It is
easy to believe (and also true) that the path taken by the wavefront to get to
any node v 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.
Implementation and Running Time To conclude our discussion of Dijkstras
Algorithm, we consider its running time. There are n 1 iterations of the
While loop for a graph with n nodes, as each iteration adds a new node v
to S. Selecting the correct node v efciently is a more subtle issue. Ones rst
impression is that each iteration would have to consider each node v S,
and go through all the edges between S and v to determine the minimum
mine=(u , v):uS d(u) + e , so that we can select the node v for which this
minimum is smallest. For a graph with m edges, computing all these minima
can take O(m) time, so this would lead to an implementation that runs in
O(mn) time.
We can do considerably better if we use the right data structures. First, we
will explicitly maintain the values of the minima d (v) = mine=(u , v):uS d(u) +
e for each node v V S, rather than recomputing them in each iteration.
We can further improve the efciency by keeping the nodes V S in a priority
queue with d (v) 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 efciently 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 ExtractMin.
How do we implement Dijkstras Algorithm using a priority queue? We put
the nodes V in a priority queue with d (v) as the key for v V . To select the node
v that should be added to the set S, we need the ExtractMin operation. To see
how to update the keys, consider an iteration in which node v is added to S, and
let w 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 mine=(u , w):uS 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(v) + e ). If
d (w) > d(v) + 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
Chapter 4 Greedy Algorithms
(4.15) Using a priority queue, Dijkstras Algorithm can be implemented on
a graph with n nodes and m edges to run in O(m) time, plus the time for n
ExtractMin and m ChangeKey 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 n).
4.5 The Minimum Spanning Tree Problem
We now apply an exchange argument in the context of a second fundamental
problem on graphs: the Minimum Spanning Tree Problem.
The Problem
Suppose we have a set of locations V = {v1, 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 nodesbut subject 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 nd a subset of the edges T E so
that the graph (V , T ) is connected, and the total cost eT ce is as small as
possible. (We will assume that the full 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
dened above. Then (V , T ) is a tree.
Proof. By denition, (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 cheapera contradiction.
If we allow some edges to have 0 cost (that is, we assume only that the
costs ce are nonnegative), then a minimum-cost solution to the network design
problem may have extra edgesedges that have 0 cost and could optionally
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
4.5 The Minimum Spanning Tree Problem
cycles until we had a tree; with nonnegative edges, the cost would not increase
during this process.
We will call a subset T 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 nding 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 efciently nd the cheapest tree from among all these options.
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 rst 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 nds 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 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 , v):uS 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. Specically, 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
Chapter 4 Greedy Algorithms
c
a
r
f
d
b
h
c
r
b
e
(a)
a
g
f
d
g
e
(b)
h
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 next edge to be added is a dashed line.
long as doing so would not actually disconnect the graph we currently
have. For want of a better name, this approach is generally called the
Reverse-Delete Algorithm (as far as we can tell, its never been named
after a specic person).
For example, Figure 4.9 shows the rst 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
Problemthere 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.
Analyzing the Algorithms
All these algorithms work by repeatedly inserting or deleting edges from a
partial solution. So, to analyze them, it would be useful to have in hand some
basic facts saying when it is safe to include an edge in the minimum spanning
tree, and, correspondingly, when it is safe to eliminate an edge on the grounds
that it couldnt possibly be in the minimum spanning tree. For purposes of
the analysis, we will make the simplifying assumption that all edge costs are
distinct from one another (i.e., no two are equal). This assumption makes it
4.5 The Minimum Spanning Tree Problem
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 insertion is the following statement, which we will
refer to as the Cut Property.
(4.17) Assume that all edge costs are distinct. Let S be any subset of nodes that
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.
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. Well 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 nd 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 v, suppose we follow the nodes
of P in sequence; there is a rst node w on P that is in V S. Let v 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 } {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 = (v , 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 {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 rst 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 {f } {e } is a spanning tree that is cheaper than T .
145
146
Chapter 4 Greedy Algorithms
S
v
v
e
e
w
w
e can be swapped for e .
f
h
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 } {e } is cheaper than T . The difculty is that T {f } {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 nd the right one.
The Optimality of Kruskals and Prims Algorithms We can now easily
prove the optimality of both Kruskals Algorithm and Prims Algorithm. The
point is that both algorithms only include an edge when it is justied by the
Cut Property (4.17).
(4.18)
Kruskals Algorithm produces a minimum spanning tree of G.
Proof. Consider any edge e = (v , w) added by Kruskals Algorithm, and let
S be the set of all nodes to which v has a path at the moment just before
e is added. Clearly v S, but w 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.
4.5 The Minimum Spanning Tree Problem
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
rst of these that it encounters.
(4.19)
Prims Algorithm produces a minimum spanning tree of G.
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 v and edge e are added that minimize the quantity
mine=(u , v):uS ce . By denition, 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 will 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 nd 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 v; and V S, containing
node w. 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 nd such an edge by following the cycle C . The edges of C other
than e form, by denition, a path P with one end at v and the other at w. If
we follow P from v to w, we begin in S and end up in V S, so there is some
147
148
Chapter 4 Greedy Algorithms
S
v
e
w
Cycle C
e
e can be swapped for e.
Figure 4.11 Swapping the edge e for the edge e in the spanning tree T , as described in
the proof of (4.20).
edge e on P that crosses from S to V S. See Figure 4.11 for an illustration of
this.
Now consider the set of edges T = T {e } {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
justied by (4.20).
(4.21)
of G.
The Reverse-Delete Algorithm produces a minimum spanning tree
Proof. Consider any edge e = (v , w) removed by Reverse-Delete. At the time
that e is removed, it lies on a cycle C ; and since it is the rst 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
4.5 The Minimum Spanning Tree Problem
contradiction that (V , T ) contains a cycle C . Consider the most expensive edge
e on C , which would be the rst one encountered by the algorithm. This edge
should have been removed, since its removal would not have disconnected
the graph, and this contradicts the behavior of Reverse-Delete.
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 justied by the Cut Property and deleting
edges when justied by the Cycle Propertyin any order at allwill 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 still have
the same relative order, since the perturbations are so small; and since all
of our algorithms are based on just comparing edge costs, the perturbations
effectively serve simply as tie-breakers to resolve comparisons among costs
that used to be equal.
Moreover, we claim that any minimum spanning tree T for the new,
perturbed instance must have also been a minimum spanning tree for the
original instance. To see this, we note that if T cost more than some tree T in
the original instance, then for small enough perturbations, the change in the
cost of T cannot be enough to make it better than T under the new costs. Thus,
if we run any of our minimum spanning tree algorithms, using the perturbed
costs for comparing edges, we will produce a minimum spanning tree T that
is also optimal for the original instance.
Implementing Prims Algorithm
We next discuss how to implement the algorithms we have been considering
so as to obtain good running-time bounds. We will see that both Prims and
Kruskals Algorithms can be implemented, with the right choice of data structures, to run in O(m log n) time. We will see how to do this for Prims Algorithm
149
150
Chapter 4 Greedy Algorithms
here, and defer discussing the implementation of Kruskals Algorithm to the
next section. Obtaining a running time close to this for the Reverse-Delete
Algorithm is difcult, so we do not focus on Reverse-Delete in this discussion.
For Prims 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):uS 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 ExtractMin
operation, and update the attachment costs using ChangeKey operations.
There are n 1 iterations in which we perform ExtractMin, 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
ExtractMin, and m ChangeKey operations.
As with Dijkstras Algorithm, if we use a heap-based priority queue we
can implement both ExtractMin 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 goalnding a good way to connect a set of
sites by installing edges between them. A minimum spanning tree optimizes
a particular goal, achieving connectedness with minimum total edge cost. But
there are a range of further 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
trafc 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 trafc. Here too, it is easy to nd cases in which the minimum spanning
tree ends up concentrating a lot of trafc on a single edge.
More generally, it is reasonable to ask whether a spanning tree is even the
right kind of solution to our network design problem. A tree has the property
that destroying any one edge disconnects it, which means that trees are not at
4.6 Implementing Kruskals Algorithm: The Union-Find Data Structure
all robust against failures. One could instead make resilience an explicit goal,
for example seeking the cheapest connected network on the set of sites that
remains connected after the deletion of any one edge.
All of these extensions lead to problems that are computationally much
harder than the basic Minimum Spanning Tree problem, though due to their
importance in practice there has been research on good heuristics for them.
4.6 Implementing Kruskals Algorithm:
The Union-Find Data Structure
One of the most basic graph problems is to nd the set of connected components. In Chapter 3 we discussed linear-time algorithms using BFS or DFS for
nding 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 xed population of nodes, but it
grows over time by having edges appear between certain pairs of nodes. Our
goal is to maintain the set of connected components of such a graph throughout
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 call 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
efciently. As each edge e = (v , w) is considered, we need to efciently nd
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 efcient
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
Chapter 4 Greedy Algorithms
name of the component containing u. If we add an edge (u , v) to the graph,
then we rst 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 operations.
.
.
.
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 = |S|.
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 O(1) time
for this operation.
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 briey discuss what we mean by the name of a setfor example,
as returned by the Find operation. There is a fair amount of exibility in
dening 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 currently
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[s] = s for all s S. This implementation
makes Find(v) 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[s] for all elements in sets A and B.
To improve this bound, we will do a few simple optimizations. 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 nd the elements that need updating. Further,
4.6 Implementing Kruskals Algorithm: The Union-Find Data Structure
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 Component 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 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(1) 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
v throughout 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 v at least doubles. The size of vs set starts out at
1, 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 log2(2k) times throughout the process.
Moreover, at most 2k elements are involved in any Union operations at all, so
153
154
Chapter 4 Greedy Algorithms
we get a bound of O(k log k) for the time spent updating Component values
in a sequence of k Union operations.
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 operation
to 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 initialize a record for each element v S
with a pointer that points to itself (or is dened as a null pointer), to indicate
that v is in its own set.
Consider a Union operation for two sets A and B, 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 v 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 result, for elements w B other than u, the name of the set they
belong to must be computed by following a sequence of pointers, rst leading
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 , v), Union(i , x),
Union(y , j ), Union(x , j ), and Union(u , v).
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 will
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
4.6 Implementing Kruskals Algorithm: The Union-Find Data Structure
The set {s, u, w} was merged into {t, v, z}.
u
w
v
s
t
j
z
x
y
i
Figure 4.12 A Union-Find data structure using pointers. The data structure has only
two sets at the moment, named after nodes v and j . The dashed arrow from u to v is the
result of the last Union operation. To answer a Find query, we follow the arrows until
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 j .
structure in Figure 4.12 followed this convention. To implement this choice
efciently, we will maintain an additional eld with the nodes: the size of the
corresponding set.
(4.24) Consider the above pointer-based implementation of the Union-Find
data structure for some set S of size n, where unions keep the name of the larger
set. A Union operation takes O(1) time, 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(v) for a node v is the number of times the set
containing node v 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 v changes, the size of this set at least doubles.
Since the set containing v starts at size 1 and is never larger than n, its size can
double at most log2 n times, and so there can be at most log2 n name changes.
Further Improvements
Next we will briey 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
Chapter 4 Greedy Algorithms
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 rst 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 n time. Now we can
issue Find(v) repeatedly, and it takes log n for each such call. Having to follow
the same sequence of log n pointers every time for nding the name of the set
containing v is quite redundant: after the rst request for Find(v), we already
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(n) time to set up a data structure for a set of size n. 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
Everything on the path from v to x
now points directly to x.
x
v
x
v
(a)
(b)
Figure 4.13 (a) An instance of a Union-Find data structure; and (b) the result of the
operation Find(v) on this structure, using path compression.
4.7 Clustering
the time, since after nding 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 time 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 n; the actual
upper bound is O(n(n)), where (n) is an extremely slow-growing function
of n called the inverse Ackermann function. (In particular, (n) 4 for any
value of n that could be encountered in practice.)
Implementing Kruskals Algorithm
Now well use the Union-Find data structure to implement Kruskals Algorithm. First we need to sort the edges by cost. This takes time O(m log m).
Since we have at most one edge between any pair of nodes, we have m n2
and hence this running time is also O(m log n).
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 n) time. (While
more efcient 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
(4.25) Kruskals Algorithm can be implemented on a graph with n nodes and
m edges to run in O(m log n) time.
4.7 Clustering
We motivated the construction of minimum spanning trees through the problem of nding a low-cost network connecting a set of sites. But minimum
157
158
Chapter 4 Greedy Algorithms
spanning trees arise in a range of different settings, several of which appear
on the surface to be quite different from one another. An appealing example
is the role that minimum spanning trees play in the area of clustering .
The Problem
Clustering arises whenever one has a collection of objectssay, a set of
photographs, documents, or microorganismsthat one is trying to classify
or organize into coherent groups. Faced with such a situation, it is natural
to look rst for measures of how similar or dissimilar each pair of objects is.
One common approach is to dene a distance function on the objects, with
the interpretation that objects 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 dene the distance
between two species to be the number of years since they diverged in the
course of evolution; we could dene the distance between two images in a
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, objects within the same
group are close, and objects in different groups are far apart. Starting from
this vague set of goals, the eld 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.
Clusterings 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 objects, labeled p1, p2 , . . . , pn . For each pair, pi and pj , we have a
numerical distance d(pi , pj ). We require only that d(pi , pi ) = 0; that d(pi , pj ) > 0
for distinct pi and pj ; and that distances are symmetric: d(pi , pj ) = d(pj , pi ).
Suppose we are seeking to divide the objects 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 , . . . , Ck . We dene 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 efciently nd the one that has
maximum spacing?
4.7 Clustering
Designing the Algorithm
To nd 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(pi , pj ).
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 nd that pi and pj already belong to the
same cluster, we will refrain from adding the edgeits 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 full 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 dening the k-clustering to be the resulting connected components C1, C2 , . . . , Ck .
Thus, iteratively merging clusters is equivalent to computing a minimum spanning tree and deleting the most expensive edges.
Analyzing the Algorithm
Have we achieved our goal of producing clusters that are as spaced apart as
possible? The following claim shows that we have.
159
160
Chapter 4 Greedy Algorithms
Cluster 1
Cluster 2
Cluster 3
Figure 4.14 An example of single-linkage clustering with k = 3 clusters. The clusters
are formed by adding edges between points in order of increasing distance.
(4.26) The components C1, C2 , . . . , Ck formed by deleting the k 1 most
expensive edges of the minimum spanning tree T constitute a k-clustering of
maximum spacing.
Proof. Let C denote the clustering C1, C2 , . . . , Ck . The spacing of C is precisely
the length d of the (k 1)st most expensive edge in the minimum spanning
tree; this is the length of the edge that Kruskals Algorithm would have added
next, at the moment we stopped it.
Now consider some other k-clustering C , which partitions U into nonempty sets C1, C2 , . . . , Ck . We must show that the spacing of C is at most
d.
Since the two clusterings C and C are not the same, it must be that one
of our clusters Cr is not a subset of any of the k sets Cs in C . Hence there
are points pi , pj Cr that belong to different clusters in C say, pi Cs and
pj Ct = Cs.
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
pi -pj path P before we stopped it. In particular, this means that each edge on
4.8 Huffman Codes and Data Compression
Cluster C r
pi
p
pj
p
Cluster C t
Cluster C s
Figure 4.15 An illustration of the proof of (4.26), showing that the spacing of any
other clustering can be no larger than that of the clustering found by the single-linkage
algorithm.
P has length at most d. Now, we know that pi Cs but pj Cs; so let p be
the rst node on P that does not belong to Cs, 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 C , and hence the spacing of C is at most d(p , p ) d. This
completes the proof.
4.8 Huffman Codes and Data Compression
In the Shortest-Path and Minimum Spanning Tree Problems, weve seen how
greedy algorithms can be used to commit to certain parts of a solution (edges
in a graph, in these cases), based entirely on relatively short-sighted considerations. We now consider a problem in which this style of committing is
carried out in an even looser sense: a greedy rule is used, essentially, to shrink
the size of the problem instance, so that an equivalent smaller problem can
then be solved by recursion. The greedy operation here is proved to be safe,
in the sense that solving the smaller instance still leads to an optimal solution for the original instance, but the global consequences of the initial greedy
decision do not become fully apparent until the full recursion is complete.
The problem itself is one of the basic questions in the area of data compression, an area that forms part of the foundations for digital communication.
161
162
Chapter 4 Greedy Algorithms
The Problem
Encoding Symbols Using Bits Since computers ultimately operate on sequences of bits (i.e., sequences consisting only of the symbols 0 and 1), one
needs encoding schemes that take text written in richer alphabets (such as the
alphabets underpinning human languages) and converts this text into long
strings of bits.
The simplest way to do this would be to use a xed 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 ve
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 25 = 32 symbolsjust 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 ve bits per symbol is sufcient. 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 symbol using just four bits, since 24 is only 16not 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 ve 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 ve 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 les 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 le should be able to correctly reconstruct it.
A huge amount of research is devoted to the design of compression algorithms
4.8 Huffman Codes and Data Compression
that can take les as input and reduce their space through efcient 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 further
progress in compression, taking advantage of features other than nonuniform
frequencies.
Variable-Length Encoding Schemes Before the Internet, before the digital
computer, before the radio and telephone, there was the telegraph. Communicating by telegraph was a lot faster than the contemporary alternatives of
hand-delivering messages by railroad or on horseback. But telegraphs were
only capable of transmitting pulses down a wire, and so if you wanted to send
a message, you needed a way to encode the text of your message as a sequence
of pulses.
To deal with this issue, the pioneer of telegraphic communication, Samuel
Morse, developed Morse code , translating each letter into a sequence of dots
(short pulses) and dashes (long pulses). For our purposes, we can think of
dots and dashes as zeros and ones, and so this is simply a mapping of symbols
into bit strings, just as in ASCII. Morse understood the point that one could
communicate more efciently 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 letters (so the encoding of
aa would actually be dot-dash-pause-dot-dash-pause). This is a reasonable
solutionusing very short bit strings and then introducing pausesbut 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 1, there would
need to be some further encoding in which the pause got mapped to bits.
163
164
Chapter 4 Greedy Algorithms
Prex Codes The ambiguity problem in Morse code arises because there exist
pairs of letters where the bit string that encodes one letter is a prex of the bit
string that encodes another. To eliminate this problem, and hence to obtain an
encoding scheme that has a well-dened interpretation for every sequence of
bits, it is enough to map letters to bit strings in such a way that no encoding
is a prex of any other.
Concretely, we say that a prex code for a set S of letters is a function
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 prex of the sequence (y).
Now suppose we have a text consisting of a sequence of letters x1x2x3 . . .
xn . 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:
(x1) (x2) . . . (xn ). If we then hand this message to a recipient who knows the
function , they will be able to reconstruct the text according to the following
rule.
.
Scan the bit sequence from left to right.
.
As soon as youve seen enough bits to match the encoding of some letter,
output this as the rst letter of the text. This must be the correct rst letter,
since no shorter or longer prex of the bit sequence could encode any
other letter.
.
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 articial devices like pauses to separate the letters.
For example, suppose we are trying to encode the set of ve letters
S = {a , b, c , d , e }. The encoding 1 specied by
1(a ) = 11
1(b) = 01
1(c ) = 001
1(d) = 10
1(e ) = 000
is a prex code, since we can check that no encoding is a prex of any other.
Now, for example, the string cecab would be encoded as 0010000011101. A
recipient of this message, knowing 1, would begin reading from left to right.
Neither 0 nor 00 encodes a letter, but 001 does, so the recipient concludes that
the rst 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
4.8 Huffman Codes and Data Compression
on the rest of the message, 0000011101; next they will conclude that the second
letter is e , encoded as 000.
Optimal Prex 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, xS fx = 1.
Now, if we use a prex 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 | (x)| to denote the length (x), we can write this as
nfx | (x)| = n
encoding length =
xS
fx | (x)|.
xS
Dropping the leading coefcient of n from the nal expression gives us
xS fx | (x)|, the average number of bits required per letter. We denote this
quantity by ABL( ).
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:
fa = .32,
fb = .25,
fc = .20,
fd = .18,
fe = .05.
Then the average number of bits per letter using the prex code 1 dened
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 xed-length encoding. (Note that a xed-length encoding is a prex code:
if all letters have encodings of the same length, then clearly no encoding can
be a prex of any other.) With a set S of ve letters, we would need three bits
per letter for a xed-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, 1 is not the best we can do in this example. Consider the
prex code 2 given by
165
166
Chapter 4 Greedy Algorithms
2(a ) = 11
2(b) = 10
2(c ) = 01
2(d) = 001
2(e ) = 000
The average number of bits per letter using 2 is
.32 2 + .25 2 + .20 2 + .18 3 + .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 prex
code that is as efcient as possiblenamely, a prex code that minimizes the
average number of bits per letter ABL( ) = xS fx | (x)|. We will call such a
prex code optimal.
Designing the Algorithm
The search space for this problem is fairly complicated; it includes all possible
ways of mapping letters to bit strings, subject to the dening property of prex
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 prex code
very efciently. As a rst step, it is useful to develop a tree-based means of
representing prex codes that exposes their structure more clearly than simply
the lists of function values we used in our previous examples.
Representing Prex 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 prex 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 prex code.
Proof. In order for the encoding of x to be a prex of the encoding of y, the
path from the root to x would have to be a prex of the path from the root
4.8 Huffman Codes and Data Compression
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 prex codes works in the other
direction as well. Given a prex code , 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
encodings 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 prex
code 0 specied by
0(a ) = 1
0(b) = 011
0(c ) = 010
0(d) = 001
0(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 1); the leaf labeled e is
obtained by taking 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 prex code 1
dened earlier, and the labeled tree in Figure 4.16(c) corresponds to the prex
code 2 dened earlier. Note also that the binary trees for the two prex codes
1 and 2 are identical in structure; only the labeling of the leaves is different.
The tree for 0, on the other hand, has a different structure.
Thus the search for an optimal prex 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 v in T simply by depthT (v). (As two 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 are 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: xS fx depthT (x). We will use ABL(T ) to
denote this quantity.
167
168
Chapter 4 Greedy Algorithms
a
Code:
a11
b01
c001
d10
e000
Code:
a1
b011
c010
d001
e000
d
b
e
d
c
b
e
(a)
a
c
(b)
Code:
a11
b10
c01
d001
e000
b
c
e
a
d
(c)
Figure 4.16 Parts (a), (b), and (c) of the figure depict three different prefix codes for
the alphabet S = {a , b, c , d , e }.
As a rst step in considering algorithms for this problem, lets note a simple
fact about the optimal tree. For this fact, we need a denition: 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 child.) Note that all three binary
trees in Figure 4.16 are full.
(4.28)
The binary tree corresponding to the optimal prex code is full.
Proof. This is easy to prove using an exchange argument. Let T denote the
binary tree corresponding to the optimal prex code, and suppose it contains
4.8 Huffman Codes and Data Compression
a node u with exactly one child v. 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 v 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 not affect other leaves. So the
prex code corresponding to T has a smaller average number of bits per letter
than the prex 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
1
in each set is exactly 2 . 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
prex 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 S2.)
It is not entirely clear how we should concretely dene this nearly
balanced split of the 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 gures in the area
of information theory, which deals with representing and encoding digital
information. These types of prex 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 prex
code. Consider again our example with the ve-letter alphabet S = {a , b, c , d , e }
and frequencies
fa = .32,
fb = .25,
fc = .20,
fd = .18,
fe = .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 1, given by the labeled tree in Figure 4.16(b); and weve
already seen that 1 is not as efcient as the prex code 2 corresponding to
the labeled tree in Figure 4.16(c).
169
170
Chapter 4 Greedy Algorithms
Shannon and Fano knew that their approach did not always yield the
optimal prex 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 Huffman
discovered for producing optimal prex codes.
What If We Knew the Tree Structure of the Optimal Prex Code? A technique that is often helpful in searching for an efcient 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 nding 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 prex code, but not the labeling
of the leaves? To complete the solution, we would need to gure 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 prex
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, ABL(T ) =
xS 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 prex code that we had before the exchange.
We can see the idea behind (4.29) in Figure 4.16(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.
4.8 Huffman Codes and Data Compression
Statement (4.29) gives us the following intuitively natural, and optimal,
way to label the tree T if someone should give it to us. We rst take all leaves
of depth 1 (if there are any) and 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 xS fx | (x)| 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 depththe ones that
receive the letters with lowest frequency. Specically, consider a leaf v in T
whose depth is as large as possible. Leaf v has a parent u, and by (4.28) T is
a full 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 justied 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 v and w get the two lowest-frequency letters of all.
We sum this up in the following claim.
(4.31) There is an optimal prex code, with corresponding tree T , in which
the two lowest-frequency letters are assigned to leaves that are siblings in T .
171
172
Chapter 4 Greedy Algorithms
New merged letter
with sum of frequencies
Two lowest-frequency letters
Figure 4.17 There is an optimal solution in which the two lowest-frequency letters
label sibling leaves; deleting them and labeling their parent with a new letter having the
combined frequency yields an instance with a smaller alphabet.
An Algorithm to Construct an Optimal Prex Code Suppose that y and z
are the two lowest-frequency letters in S. (We can break ties in the frequencies
arbitrarily.) Statement (4.31) is important because it tells us something about
where y and z go in the optimal 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 nd a
prex code for the smaller alphabet, and then open up the meta-letter back
into y and z to obtain a prex 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 1
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 fy + fz
Recursively construct a prefix code for S , with tree T
Define a prefix code for S as follows:
Start with T
4.8 Huffman Codes and Data Compression
Take the leaf labeled and add two children below it
labeled y and z
Endif
We refer to this as Huffmans Algorithm, and the prex code that it
produces for a given alphabet is accordingly referred to as a Huffman code .
In general, it is clear that this algorithm always terminates, since it simply
invokes a recursive call on an alphabet that is one letter smaller. Moreover,
using (4.31), it will not be difcult to prove that the algorithm in fact produces
an optimal prex 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
fa = .32,
fb = .25,
fc = .20,
fd = .18,
fe = .05.
The algorithm would rst merge d and e into a single letterlets 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
Algorithmthe merging of the two lowest-frequency lettersts 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 t 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 rst and foremost about the top-level split in
the binary treenamely, 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 Algorithm
The Optimality of the Algorithm We rst prove the optimality of Huffmans
Algorithm. Since the algorithm operates recursively, invoking itself on smaller
and smaller alphabets, it is natural to try establishing optimality by induction
173
174
Chapter 4 Greedy Algorithms
on the size of the alphabet. Clearly it is optimal for all two-letter alphabets
(since it uses only one bit per letter). So suppose by induction that it is optimal
for all alphabets of size k 1, and consider an input instance consisting of an
alphabet S of size k.
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
, calls itself recursively on the smaller alphabet S (in which y and z are
replaced by ), and by induction produces an optimal prex 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 ) f .
Proof. The depth of each letter 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
in T . Using this, plus the fact that f = fy + fz , we have
ABL(T ) =
fx depthT (x)
xS
= fy depthT (y) + fz depthT (z ) +
fx depthT (x)
x =y , z
= (fy + fz ) (1 + depthT ()) +
fx depthT (x)
x =y , z
= f (1 + depthT ()) +
= f + f depthT () +
= f +
fx depthT (x)
x =y , z
fx depthT (x)
x =y , z
fx depthT (x)
xS
= f + ABL(T ).
Using this, we now prove optimality as follows.
(4.33) The Huffman code for a given alphabet achieves the minimum average
number of bits per letter of any prex 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
4.8 Huffman Codes and Data Compression
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 , we get a tree
Z that denes a prex code for S . In the same way that T is obtained from
T , the tree Z is obtained from Z by adding leaves for y and z below ; thus
the identity in (4.32) applies to Z and Z as well: ABL(Z ) = ABL(Z ) f .
But we have assumed that ABL(Z ) < ABL(T ); subtracting f from both sides
of this inequality we get ABL(Z ) < ABL(T ), which contradicts the optimality
of T as a prex 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 dene 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 k 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 iterationwhich performs just three of these operationstakes
time O(log k). Summing over all k iterations, we get a total running time of
O(k log k).
Extensions
The structure of optimal prex 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
Chapter 4 Greedy Algorithms
What more could we want beyond an optimal prex code? First, consider
an application in which we are transmitting black-and-white images: each
image is a 1,000-by-1,000 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 prex 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 letterthe 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 dene an encoding scheme
where the notion of using fractions of bits is well-dened. There are results
in the area of data compression, however, that do just this; arithmetic coding
and a range of other techniques have been developed to handle settings like
this.
A second drawback of prex codes, as dened here, is that they cannot
adapt to changes in the text. Again lets consider a simple example. Suppose we
are trying to encode the output of a program that produces a long sequence
of letters from the set {a , b, c , d}. Further suppose that for the rst 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 just one bit per letter, plus a bit of extra overhead, as follows.
.
Begin with an encoding in which the bit 0 represents a and the bit 1
represents b.
.
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 1 represents d.
.
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
4.9 Minimum-Cost Arborescences: A Multi-Phase Greedy Algorithm
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 signicant improvements over the static
method weve considered here.
These issues suggest some of the directions in which work on data compression has proceeded. In many of these cases, there is a trade-off between
the power of the compression technique and its computational cost. In particular, many of the improvements to Huffman codes just described come with
a corresponding increase in the computational effort needed both to produce
the compressed version of the data and also to decompress it and restore the
original text. Finding the right balance among these trade-offs is a topic of
active research.
* 4.9 Minimum-Cost Arborescences: A Multi-Phase
Greedy Algorithm
As weve seen more and more examples of greedy algorithms, weve come to
appreciate that there can be considerable diversity in the way they operate.
Many greedy algorithms make some sort of an initial ordering decision on
the input, and then process everything in a one-pass fashion. Others make
more incremental decisionsstill local and opportunistic, but without a global
plan in advance. In this section, we consider a problem that stresses our
intuitive view of greedy algorithms still further.
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 signicant new complications. At the same time, the style
of the algorithm has a strongly greedy avor, since it still constructs a solution
according to a local, myopic rule.
We begin with the basic denitions. 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 . Specically, 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
Chapter 4 Greedy Algorithms
r
r
r
(a)
(b)
(c)
Figure 4.18 A directed graph can have many different arborescences. Parts (b) and (c)
depict two different aborescences, both rooted at node r , for the graph in part (a).
(4.34) A subgraph T = (V , F ) of G is an arborescence with respect to root r if
and only if T has no cycles, and for each node v = r, there is exactly one edge
in F that enters v.
Proof. If T is an arborescence with root r , then indeed every other node v
has exactly one edge entering it: this is simply the last edge on the unique r -v
path.
Conversely, suppose T has no cycles, and each node v = r has exactly
one entering edge. In order to establish that T is an arborescence, we need
only show that there is a directed path from r to each other node v. Here is
how to construct such a path. We start at v and repeatedly follow edges in
the backward direction. Since T has no cycles, we can never return to a node
weve previously visited, and thus this process must terminate. But r is the
only node without incoming edges, and so the process must in fact terminate
by reaching r ; the sequence of nodes thus visited yields a path (in the reverse
direction) from r to v.
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-rst search tree rooted at r
will form an arborescence.
(4.35) A directed graph G has an arborescence rooted at r if and only if there
is a directed path from r to each other node.
4.9 Minimum-Cost Arborescences: A Multi-Phase Greedy Algorithm
The basic problem we consider here is the following. We are given a
directed graph G = (V , E ), with a distinguished root node r and with a nonnegative cost ce 0 on each edge, and we wish to compute an arborescence
rooted at r of minimum total cost. (We will refer to this as an optimal arborescence.) We will assume throughout that G at least has an arborescence rooted
at r ; by (4.35), this can be easily checked at the outset.
Designing the Algorithm
Given the relationship between arborescences and trees, the minimum-cost
arborescence problem certainly has 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, condent 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 4.19 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
r
r
2
2
10
10
1
2
4
4
2
8
2
4
4
(a)
2
(b)
Figure 4.19 (a) A directed graph with costs on its edges, and (b) an optimal arborescence
rooted at r for this graph.
179
180
Chapter 4 Greedy Algorithms
the other edges out of r . This kind of argument never clouded our thinking in
the Minimum Spanning Tree Problem, where it was always safe to plunge
ahead and include the cheapest edge; it suggests that nding the optimal
arborescence may be a signicantly 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 fact.
(4.36)
If (V , F ) is an arborescence, then it is a minimum-cost arborescence.
So the difculty 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 cost of any
edge entering v. For each edge e = (u , v), with cost ce 0, we dene its modied cost ce to be ce yv . Note that since ce yv , all the modied costs are still
nonnegative. 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 modied costs {ce }.
Proof. Consider an arbitrary arborescence T . The difference between its cost
with costs {ce } and {ce } is exactly v =r yv that is,
ce
e T
ce =
e T
yv .
v =r
4.9 Minimum-Cost Arborescences: A Multi-Phase Greedy Algorithm
This is because an arborescence has exactly one edge entering each node v
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 subject to {ce }
if and only if it has minimum cost subject to {ce }.
We now consider the problem in terms of the costs {ce }. All the edges in
our set F have cost 0 under these modied 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
supernode , 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 to
an edge 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 ne; however, we delete self-loops from E edges that have both
ends equal to c . We recursively nd 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 v = r
Let yv be the minimum cost of an edge entering node v
Modify the costs of all edges e entering v to ce = ce yv
Choose one 0-cost edge entering each v = r , obtaining a set F
If F forms an arborescence, then return it
Else there is a directed cycle C F
Contract C to a single supernode, yielding a graph G = (V , E )
Recursively find an optimal arborescence (V , F ) in G
with costs {ce }
Extend (V , F ) to an arborescence (V , F ) in G
by adding all but one edge of C
Analyzing the Algorithm
It is easy to implement this algorithm so that it runs in polynomial time. But
does it lead to an optimal arborescence? Before concluding that it does, we need
to worry about the following point: not every arborescence in G corresponds to
an arborescence in the contracted graph G . Could we perhaps miss the true
optimal arborescence in G by focusing on G ? What is true is the following.
181
182
Chapter 4 Greedy Algorithms
The arborescences of G are in one-to-one correspondence with arborescences
of G that have exactly one edge entering the cycle C ; and these corresponding
arborescences have the same cost with respect to {ce }, since C consists of 0cost edges. (We say that an edge e = (u , v) enters C if v belongs to C but u does
not.) So to prove that our algorithm nds 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 0, such that r C.
Then there is an optimal arborescence rooted at r that has exactly one edge
entering C.
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 an 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 r -v
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 denition 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 rst 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 still exists in T . Otherwise,
let w be the last node in P C , and let P be the subpath of P from w to v.
Observe that all the edges in P still exist in T . We have already argued that
w is reachable from r in T , since it belongs to C . Concatenating this path
to w 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 Exercises
(4.39)
The algorithm nds an optimal arborescence rooted at r in G.
Proof. The proof is by induction on the number of nodes in G . If the edges
of F form an arborescence, then the algorithm returns an optimal arborescence
by (4.36). Otherwise, we consider the problem with the modied 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
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 identied 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 signicant 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 x1, 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 rst 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 trail. Thus
a set of stopping points is valid if one could camp only at these places and
183
184
Chapter 4 Greedy Algorithms
still 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 friends greedy
algorithmhiking as long as possible each dayoptimal, in the sense that it
nds a valid set whose size is as small as possible?
Solution Often a greedy algorithm looks correct when you rst 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, given 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 = {xp1 , . . . , 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 = {xq1 , . . . , xqm }, with
m < k.
To obtain a contradiction, we rst show that the stopping point reached by
the greedy algorithm on each day j is farther than the stopping point reached
under the alternate solution. That is,
(4.40)
For each j = 1, 2, . . . , m, we have xpj xqj .
Proof. We prove this by induction on j . The case j = 1 follows directly from
the denition of the greedy algorithm: your friends travel as long as possible
Solved Exercises
on the rst day before stopping. Now let j > 1 and assume that the claim is
true for all i < j . Then
xqj xqj1 d ,
since S is a valid set of stopping points, and
xqj xpj1 xqj xqj1
since xpj1 xqj1 by the induction hypothesis. Combining these two inequalities, we have
xq j x p j 1 d .
This means that your friends have the option of hiking all the way from
xpj1 to xqj in one day; and hence the location xpj at which they nally stop
can only be farther along than xqj . (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.)
Statement (4.40) implies in particular that xqm xpm . Now, if m < k, then
we must have xpm < L d, for otherwise your friends would never have needed
to stop at the location xpm+1. 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 friends 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 $100. However, they are
all becoming more expensive according to exponential growth curves: in
particular, the cost of license j increases by a factor of rj > 1 each month, where
rj is a given parameter. This means that if license j is purchased t months from
now, it will cost 100 rjt . We will assume that all the price growth rates are
distinct; that is, ri = rj for licenses i = j (even though they start at the same
price of $100).
185
186
Chapter 4 Greedy Algorithms
The question is: Given that the company can only buy at most one license
a month, in which order should it buy the licenses so that the total amount of
money it spends is as small as possible?
Give an algorithm that takes the n rates of price growth r1, 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 should 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 r1 = 2, r2 = 3, and
r3 = 4. Buying the licenses in increasing order results in a total cost of
100(2 + 32 + 43) = 7,500,
while buying them in decreasing order results in a total cost of
100(4 + 32 + 23) = 2,100.
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 inversionthat
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 rt ) is less
than that bought in month t + 1 (similarly, we use rt +1 to denote this). That
is, we have rt < rt +1.
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 successfully show that our
algorithm is indeed the correct one.
Notice that if we swap these two purchases, the rest of the purchases
are identically priced. In O, the amount paid during the two months involved
+1
in the swap is 100(rtt + rtt+1 ). On the other hand, if we swapped these two
purchases, we would pay 100(rtt+1 + rtt +1). Since the constant 100 is common
Solved Exercises
to both expressions, we want to show that the second term is less than the
rst one. So we want to show that
+1
rtt+1 + rtt +1 < rtt + rtt+1
+1
rtt +1 rtt < rtt+1 rtt+1
rtt (rt 1) < rtt+1(rt +1 1).
But this last inequality is true simply because ri > 1 for all i and since rt < rt +1.
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 < 1 per month, starting from
$100, so if you sell it t months from now you will receive 100 rit . (In 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
31
1
4 , 2 , 100 ).
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 specied. 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
188
Chapter 4 Greedy Algorithms
long way around C (avoiding e ) can be viewed as an alternate route between
the endpoints of e that only uses cheaper edges.
Putting these two observations together suggests that we should try proving the following statement.
(4.41) Edge e = (v , w) does not belong to a minimum spanning tree of G if
and only if v and w can be joined by a path consisting entirely of edges that
are cheaper than e.
Proof. First suppose that P is a v-w path consisting entirely of edges cheaper
than e . If we add e to P , we get a cycle on which e is the most expensive edge.
Thus, by the Cycle Property, e does not belong to a minimum spanning tree
of G .
On the other hand, suppose that v and w cannot be joined by a path
consisting entirely of edges cheaper than e . We will now identify a set S for
which e is the cheapest edge with one end in S and the other in V S; if we can
do this, the Cut Property will imply that e belongs to every minimum spanning
tree. Our set S will be the set of all nodes that are reachable from v using a path
consisting only of edges that are cheaper than e . By our assumption, we have
w V S. Also, by the denition of S, there cannot be an edge f = (x , y) that
is cheaper than e , and for which one end x lies in S and the other end y lies in
V S. Indeed, if there were such an edge f , then since the node x is reachable
from v using only edges cheaper than e , the node y would be reachable as
well. Hence e is the cheapest edge with one end in S and the other in V S,
as desired, and so we are done.
Given this fact, our algorithm is now simply the following. We form a graph
G by deleting from G all edges of weight greater than ce (as well as deleting
e itself). We then use one of the connectivity algorithms from Chapter 3 to
determine whether there is a path from v to w in G . Statement (4.41) says that
e belongs to a minimum spanning tree if and only if there is no such path.
The running time of this algorithm is O(m + n) to build G , and O(m + n)
to test for a path from v to w.
Exercises
1. 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.
Let G be an arbitrary connected, undirected graph with a distinct cost c (e ) on
every edge e. Suppose e is the cheapest edge in G; that is, c (e ) < c (e ) for every
Exercises
edge e = e . Then there is a minimum spanning tree T of G that contains the
edge e .
2. For each of the following two statements, decide whether it is true or false.
If it is true, give a short explanation. If it is false, give a counterexample.
(a)
Suppose we are given an instance of the Minimum Spanning Tree
Problem on a graph G , with edge costs that are all positive and
distinct. Let T be a minimum spanning tree for this instance. Now
2
suppose we replace each edge cost ce by its square, ce , thereby
creating a new instance of the problem with the same graph but
different costs.
True or false? T must still be a minimum spanning tree for this
new instance.
(b) Suppose we are given an instance of the Shortest s-t Path Problem
on a directed graph G . We assume that all edge costs are positive
and distinct. Let P be a minimum-cost s-t path for this instance.
2
Now suppose we replace each edge cost ce by its square, ce , thereby
creating a new instance of the problem with the same graph but
different costs.
True or false? P must still be a minimum-cost s-t path for this
new instance.
3. You are consulting for a trucking company that does a large amount of
business shipping packages between New York and Boston. The volume is
high enough that they have to send a number of trucks each day between
the two locations. Trucks have a fixed limit W on the maximum amount
of weight they are allowed to carry. Boxes arrive at the New York station
one by one, and each package i has a weight wi . The trucking station
is quite small, so at most one truck can be at the station at any time.
Company policy requires that boxes are shipped in the order they arrive;
otherwise, a customer might get upset upon seeing a box that arrived
after his make it to Boston faster. At the moment, the company is using
a simple greedy algorithm for packing: they pack boxes in the order they
arrive, and whenever the next box does not fit, they send the truck on its
way.
But they wonder if they might be using too many trucks, and they
want your opinion on whether the situation can be improved. Here is
how they are thinking. Maybe one could decrease the number of trucks
needed by sometimes sending off a truck that was less full, and in this
way allow the next few trucks to be better packed.
189
190
Chapter 4 Greedy Algorithms
Prove that, for a given set of boxes with specified weights, the greedy
algorithm currently in use actually minimizes the number of trucks that
are needed. Your proof should follow the type of analysis we used for
the Interval Scheduling Problem: it should establish the optimality of this
greedy packing algorithm by identifying a measure under which it stays
ahead of all other solutions.
4. 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 exchangeswhats 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 themfor example, they may want to know if the
four events
buy Yahoo, buy eBay, buy Yahoo, buy Oracle
occur in this sequence S, in order but not necessarily consecutively.
They begin with a collection of possible events (e.g., the possible
transactions) and a sequence S of n of these events. A given event may
occur multiple times in S (e.g., Yahoo stock may be bought many times
in a single sequence S). We will say that a sequence S is a subsequence
of S if there is a way to delete certain of the events from S so that the
remaining events, in order, are equal to the sequence S . So, for example,
the sequence of four events above is a subsequence of the sequence
buy Amazon, buy Yahoo, buy eBay, buy Yahoo, buy Yahoo,
buy Oracle
Their goal is to be able to dream up short sequences and quickly
detect whether they are subsequences of S. So this is the problem they
pose to you: Give an algorithm that takes two sequences of eventsS of
length m and S of length n, each possibly containing an event more than
onceand decides in time O(m + n) whether S is a subsequence of S.
5. 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
6. 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 time. 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 schedule 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.
7. The wildly popular Spanish-language search engine El 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 essentially unlimited supply of high-end PCs.
Theyve broken the overall computation into n distinct jobs, labeled
J1, 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 Ji needs pi seconds of time on the supercomputer,
followed by fi 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 parallelall the jobs can be processed at the same time. However, the supercomputer can only work on
a single job at a time, so the system managers need to work out an order
in which to feed the jobs to the supercomputer. As soon as the first job
191
192
Chapter 4 Greedy Algorithms
in order is done on the supercomputer, it can be handed off to a PC for
finishing; at that point in time a second job can be fed to the supercomputer; 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 El 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 unique minimum spanning tree.
9. One of the basic motivations behind the Minimum Spanning Tree Problem
is the goal of designing a spanning network for a set of nodes with
minimum total cost. Here we 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 if 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 , w 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(|E |). Can you do it in O(|V |) time?
Please note any assumptions you make about what data structure is
used to represent the tree T and the graph G .
Exercises
(b) Suppose T is no longer the minimum-cost spanning tree. Give a
linear-time algorithm (time O(|E |)) to update the tree T to the new
minimum-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 Kruskals Algorithm sorted the edges in order of increasing cost, then greedily processed edges one by one, adding an edge e as
long as it did not form a cycle. When some edges have the same cost, the
phrase in order of increasing cost has to be specified a little more carefully: well say that an ordering of the edges is valid if the corresponding
sequence of edge costs is nondecreasing. Well say that a valid execution
of Kruskals Algorithm is one that begins with a valid ordering of the
edges of G .
For any graph G , and any minimum spanning tree T of G , is there a
valid execution of Kruskals Algorithm on G that produces T as output?
Give a proof or a counterexample.
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 n 1 ti , whichever order you choose). We assume that all the values
i=
bi and ti 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 > 0, the total number of bits you send over the
time interval from 0 to t cannot exceed rt.
Note that this constraint is only imposed for time intervals that start at
0, not for time intervals that start at any other value.
We say that a schedule is valid if it satisfies the constraint () imposed
by the link.
193
194
Chapter 4 Greedy Algorithms
The Problem. Given a set of n streams, each specified by its number of
bits bi and its time duration ti , as well as the link parameter r , determine
whether there exists a valid schedule.
Example. Suppose we have n = 3 streams, with
(b1, t1) = (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 rst stream has been sent, and 2000 < 5000 1
t = 2: half of the second stream has also been sent,
and 2000 + 3000 < 5000 2
Similar calculations hold for t = 3 and t = 4.
(a)
Consider the following claim:
Claim: There exists a valid schedule if and only if each stream i satises
bi rti .
Decide whether you think the claim is true or false, and give a proof
of either the claim or its negation.
(b) 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 running
time of your algorithm should be polynomial in n.
13. A small businesssay, a photocopying service with a single large
machinefaces 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 done, we would have Cj = tj ; and if job j is done right after job
i, we would have Cj = Ci + tj . Each customer i also has a given weight wi
that represents his or her importance to the business. The happiness of
customer i is expected to be dependent on the finishing time of is job.
So the company decides that they want to order the jobs to minimize the
weighted sum of the completion times, n 1 wi Ci .
i=
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 wi for each
job. You want to order the jobs so as to minimize the weighted sum of
the completion times, n 1 wi Ci .
i=
Example. Suppose there are two jobs: the first takes time t1 = 1 and has
weight w1 = 10, 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 runs continuously between
these times; the consultants have a list of the planned start and finish
times of all sensitive processes that will be run that day.
As a simple first step, theyve written a program called status_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 status_check, subject to the requirement
that status_check is invoked at least once during each sensitive
process P .
(b) While 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 status_check at least k times: no one invocation
of status_check can handle more than one of these processes.
This is true, of course, and after some further discussion, you all
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 case that there must
be a set of k times at which you can run status_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 status_
check.) Decide whether you think this claim is true or false, and give
a proof or a counterexample.
195
196
Chapter 4 Greedy Algorithms
15. The manager of a large student union on campus comes to you with the
following problem. Shes in charge of a group of n students, each of whom
is scheduled to work one shift during the week. There are different jobs
associated with these shifts (tending the main desk, helping with package
delivery, rebooting cranky information kiosks, etc.), but we can view each
shift as a single contiguous interval of time. There can be multiple shifts
going on at once.
Shes trying to choose a subset of these n students to form a supervising committee that she can meet with once a week. She considers such
a committee to be complete if, for every student not on the committee,
that students shift overlaps (at least partially) the shift of some student
who is on the committee. In this way, each students performance can be
observed by at least one person whos serving on the committee.
Give an efficient algorithm that takes the schedule of n shifts and
produces a complete supervising committee containing as few students
as possible.
Example. Suppose n = 3, and the shifts are
Monday 4 P.M.Monday 8 P.M.,
Monday 6 P.M.Monday 10 P.M.,
Monday 9 P.M.Monday 11 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 working in the financial domain are currently advising a client who is investigating a potential money-laundering
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 identity of the account, the amounts of
the transactions, or the exact times 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 ei , for some
margin of error ei . (In other words, it took place sometime between ti ei
and ti + ei .) 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 x1, x2 , . . . , xn . To see whether its plausible
that this really is the account theyre looking for, theyre wondering
Exercises
whether its possible to associate each of the accounts n events with
a distinct one of the n suspicious transactions in such a way that, if the
account event at time xi is associated with the suspicious transaction that
occurred approximately at time tj , then |tj xi | ej . (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 continuously, 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 following 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
7 P.M.), which can be scheduled without overlapping.
A.M.)
and (1
P.M.,
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 travel along the
roads; however, the speed of travel depends on the time of year. More
precisely, the Web site answers queries of the following form: given an
197
198
Chapter 4 Greedy Algorithms
edge e = (v , w) connecting two sites v and w, and given a proposed starting
time t from location v, the site will return a value fe (t ), the predicted
arrival time at w. The Web site guarantees that fe (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 + e , where e is the time needed to travel from the
beginning to the end of edge e .
Your friends want to use the Web site to determine the fastest way
to travel through the directed graph from their starting point to their
intended destination. (You should assume that they start at time 0, and
that all predictions made by the Web site are completely correct.) Give a
polynomial-time algorithm to do this, where we treat a single query to
the Web site (based on a specific edge e and a time t ) as taking a single
computational step.
19. A group of network designers at the communications company CluNet
find themselves facing the following problem. They have a connected
graph G = (V , E ), in which the nodes represent sites that want to communicate. Each edge e is a communication link, with a given available
bandwidth be .
For each pair of nodes u , v V , they want to select a single u-v path P
on which this pair will communicate. The bottleneck rate b(P ) of this path P
is the minimum bandwidth of any edge it contains; that is, b(P ) = mineP be .
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-v path
in T .)
This idea is roundly heckled in the offices of CluNet for a few days,
and theres a natural reason for the skepticism: each pair of nodes
might want a very different-looking path to maximize its bottleneck rate;
why should there be a single tree that simultaneously makes everybody
happy? But after some failed attempts to rule out the idea, people begin
to suspect it could be possible.
Exercises
Show that such a tree exists, and give an efficient algorithm to find
one. That is, give an algorithm constructing a spanning tree T in which,
for each u , v V , the bottleneck rate of the u-v path in T is equal to the
best achievable bottleneck rate for the pair u , v in G .
20. Every September, somewhere in a far-away mountainous part of the
world, the county highway crews get together and decide which roads to
keep clear through the coming 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 roadeach edge e in the graphis annotated with a number
ae 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 .
Finally, a path between towns i and j is declared to be winter-optimal if it
achieves the minimum possible height over all paths from i to j .
The highway crews are going to select a set E E of the roads to keep
clear 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 property 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
full graph G = (V , E ). Well say that (V , E ) is a minimum-altitude connected
subgraph if it has this property.
Given that theyre going to maintain this key property, however, they
otherwise want to keep as few roads clear as possible. One year, they hit
upon the following conjecture:
The minimum spanning tree of G, with respect to the edge weights ae , is a
minimum-altitude connected subgraph.
(In an earlier problem, we claimed that there is a unique minimum spanning tree when the edge weights are distinct. Thus, thanks to the assumption that all ae are distinct, it is okay for us to speak of the minimum
spanning tree.)
Initially, this conjecture is somewhat counterintuitive, since the minimum spanning tree is trying to minimize the sum of the values ae , while
the goal of minimizing altitude seems to be asking for a fairly different
thing. But lacking an argument to the contrary, they begin considering an
even bolder second conjecture:
199
200
Chapter 4 Greedy Algorithms
A subgraph (V , E ) is a minimum-altitude connected subgraph if and only if
it contains the edges of the minimum spanning tree.
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
ae ? Give a proof or a counterexample with explanation.
(b) Is the second conjecture true, for all choices of G and distinct altitudes ae ? Give a proof or a counterexample 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 = |V |. Give an algorithm with running time O(n)
that takes a near-tree G with costs on its edges, and returns a minimum
spanning tree of G . You may assume that all the edge costs are distinct.
22. Consider the Minimum Spanning Tree Problem on an undirected graph
G = (V , E ), with a cost ce 0 on each edge, where the costs may not all
be different. If the costs are not all distinct, there can in general be
many distinct minimum-cost solutions. Suppose we are given a spanning
tree T E with the guarantee that for every e T , e belongs to some
minimum-cost spanning tree in G . Can we conclude that T itself must
be a minimum-cost spanning tree in G ? Give a proof or a counterexample
with explanation.
23. Recall the problem of computing a minimum-cost arborescence in a
directed graph G = (V , E ), with a cost ce 0 on each edge. Here we will
consider the case in which G is a directed acyclic graphthat 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 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. Timing 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.
Exercises
v
2
1
v
v
2
a
1
2
b
c
1
d
Figure 4.20 An instance of the zero-skew problem, described in Exercise 23.
The root generates a clock signal which is propagated along the edges
to the leaves. Well 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 small 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 this 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 = {p1, p2 , . . . , pn }, together with a
distance function d on the set P ; d is simply a function on pairs of points in
P with the properties that d(pi , pj ) = d(pj , pi ) > 0 if i = j , and that d(pi , pi ) = 0
for each i.
We define a hierarchical metric on P to be any distance function that
can be constructed as follows. 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 hv . These heights must satisfy the properties that h(v) = 0 for each
201
202
Chapter 4 Greedy Algorithms
leaf v, and if 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 pi and pj , their
distance (pi , pj ) is defined as follows. We determine the least common
ancestor v in T of the leaves containing pi and pj , and define (pi , pj ) = hv .
We say that a hierarchical metric is consistent with our distance
function d if, for all pairs i , j , we have (pi , pj ) d(pi , pj ).
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) if is any other hierarchical metric consistent with d, then (pi , pj )
(pi , pj ) for each pair of points pi and pj .
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 achievedrather than a continuum of
possibilitiesand 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 :RR. Thus, at time t , it has cost
fe (t ). Well 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 cG (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: fe (t ) = ae t 2 +
be t + 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 , ce )} can
be done in constant time per operation.
27. In trying to understand the combinatorial structure 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.
Exercises
Here is one way to do this. Let G be a connected graph, and T and T
two different spanning trees of G . We say that T and T are neighbors if
T contains exactly one edge that is not in T , and T contains exactly one
edge that is not in T .
Now, from any graph G , we can build a (large) graph H as follows.
The nodes of H are the spanning trees of G , and there is an edge between
two nodes of H if the corresponding spanning trees are neighbors.
Is it true that, for any connected graph G , the resulting graph H
is connected? Give a proof that H is always connected, or provide an
example (with explanation) of a connected graph G for which H 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 exactly k edges
labeled X , or (ii) reports correctly that no such tree exists.
29. Given a list of n natural numbers d1, d2 , . . . , dn , show how to decide
in polynomial time whether there exists an undirected graph G = (V , E )
whose node degrees are precisely the numbers d1, d2 , . . . , dn . (That is, if
V = {v1, v2 , . . . , vn }, then the degree of vi should be exactly di .) G should not
contain multiple edges between the same pair of nodes, or loop 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 wij on each edge (i , j ); and
we will assume these weights satisfy the triangle inequality wik wij + wjk .
For a subset V V , we will use G [V ] to denote the subgraph (with edge
weights) induced on the nodes in V .
203
204
Chapter 4 Greedy Algorithms
We are given a set X V of k terminals that must be connected by
edges. We say that a Steiner tree on X 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 O(nO(k)).
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 willing to settle for a subgraph H = (V , F ) that is
denser than a tree, and we are interested in guaranteeing that, for each
pair of vertices u , v V , the length of the shortest u-v path in H is not
much longer than the length of the shortest u-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.)
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 H if there
is currently no u-v path in H . (This is what Kruskals Algorithm would
do as well.) On the other hand, if there is a u-v path in H , we let duv
denote the length of the shortest such path; again, length is with
respect to the values { e }. We add e to H if 3 e < duv .
In other words, we add an edge even when u 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 every pair of nodes u , v V , the length of the shortest
u-v path in H is at most three times the length of the shortest u-v
path in G .
(b) Despite its ability to approximately preserve shortest-path distances,
the subgraph H produced by the algorithm cannot be too dense.
Let f (n) denote the maximum number of edges that can possibly
be produced as the output of this algorithm, over all n-node input
graphs with edge lengths. Prove that
lim
n
f (n)
= 0.
n2
Notes and Further Reading
32. Consider a directed graph G = (V , E ) with a root r V and nonnegative
costs on the edges. In this problem we consider variants of the minimumcost 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 v, and we modified the costs of all edges e
entering node v to be ce = ce yv . Suppose we instead use the following modified cost: ce = max(0, ce 2yv ). 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 all 0cost strong components and recursively apply the same procedure
on the resulting graph until an arborescence is found. Prove that this
T has cost at most twice the cost of the minimum-cost arborescence
in the original graph.
33. Suppose you are given a directed graph G = (V , E ) in which each edge has
a cost of either 0 or 1. Also suppose that G has a node r such that there is a
path from r to every other node in G . You are also given an integer k. Give a
polynomial-time algorithm that either constructs an arborescence rooted
at r of cost exactly k, or reports (correctly) that no such arborescence
exists.
Notes and Further Reading
Due to their conceptual cleanness and intuitive appeal, greedy algorithms have
a long history and many applications throughout computer science. In this
chapter we focused on cases in which greedy algorithms nd the optimal
solution. Greedy algorithms are also often used as simple heuristics even when
they are not guaranteed to nd the optimal solution. In Chapter 11 we will
discuss greedy algorithms that nd 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
205
206
Chapter 4 Greedy Algorithms
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 Lawler, 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 Jain and Dubes (1981) survey a variety of clustering
techniques.
The algorithm for optimal prex 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 eld
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 nding 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. It is 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 programming. 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
nd a precise denition of what constitutes a greedy algorithm. In the search
for such a denition, it is not even clear that one can apply the analogue
Notes and Further Reading
of U.S. Supreme Court Justice Potter Stewarts famous test for obscenity
I know it when I see itsince one nds 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 inuential
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 question.
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, Gautam Das, David Dobkin, and Deborah Joseph.
207
This page intentionally left blank
Chapter 5
Divide and Conquer
Divide and conquer refers to a class of algorithmic techniques in which one
breaks the input into several parts, solves the problem in 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; nding 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-efcient solution to a fundamental sequence comparison problem, and
in Chapter 13 we will see it combined with randomization to yield a simple
and efcient 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 lower 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 5 Divide and Conquer
Chapter 2 that the natural brute-force algorithm for nding the closest pair
among n points in the plane would simply measure all (n2) distances, for
a (polynomial) running time of (n2). Using divide and conquer, we will
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
efciently, 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 polynomial.
5.1 A First Recurrence: The Mergesort Algorithm
To motivate the general approach to analyzing divide-and-conquer algorithms,
we begin with the Mergesort Algorithm. We discussed the Mergesort Algorithm
briey in Chapter 2, when we surveyed common running times for algorithms.
Mergesort sorts a given list of numbers by rst dividing them into two equal
halves, sorting each half separately by recursion, and then combining the
results of these recursive callsin the form of the two sorted halvesusing
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.
() 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 nal recombining.
In Mergesort, as in any algorithm that ts 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 ts the pattern in (), 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 running time for an input of size n/2); and nally it spends O(n) time
to combine the solutions from the two recursive calls. Thus the running time
T (n) satises the following recurrence relation.
5.1 A First Recurrence: The Mergesort Algorithm
(5.1)
For some constant c,
T (n) 2T (n/2) + cn
when n > 2, and
T (2) c .
The structure of (5.1) is typical of what recurrences will look like: theres an
inequality or equation that bounds T (n) in terms of an expression involving
T (k) for smaller 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 n/2 and
n/2 , and the recurrence relation would say that
T (n) T ( n/2 ) + T ( n/2 ) + 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 oors and ceilings, and it makes the symbolic manipulation
much cleaner.
Now (5.1) does not explicitly provide an asymptotic bound on the growth
rate of the function T ; rather, it species 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
right-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 efcient divideand-conquer algorithm is heavily intertwined with an understanding of how
a recurrence relation determines a running time.
Approaches to Solving Recurrences
There are two basic ways one can go about solving a recurrence, each of which
we describe in more detail below.
211
212
Chapter 5 Divide and Conquer
.
The most intuitively natural way to search for a solution to a recurrence is
to unroll the recursion, accounting for the running time across the rst
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 justies
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 unspecied in the substitution, one can often work them out
as needed.
We now discuss each of these approaches, using the recurrence in (5.1) as an
example.
Unrolling the Mergesort Recurrence
Lets start with the rst approach to solving the recurrence in (5.1). The basic
argument is depicted in Figure 5.1.
.
Analyzing the rst few levels: At the rst 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 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
cn
cn/2
cn/4
cn/2
cn/4
cn/4
Level 1: cn/2 + cn/2 = cn total
cn/4 Level 2: 4(cn/4) = cn total
Figure 5.1 Unrolling the recurrence T (n) 2T (n/2) + O(n).
5.1 A First Recurrence: The Mergesort Algorithm
.
Identifying a pattern: Whats going on in general? At level j of the
recursion, the number of subproblems has doubled j times, so there are
now a total of 2j . Each has correspondingly shrunk in size by a factor
of two j times, and so each has size n/2j , and hence each takes time at
most cn/2j . Thus level j contributes a total of at most 2j (cn/2j ) = 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 > 1.
Substituting a Solution into the Mergesort Recurrence
The argument establishing (5.2) can be used to determine that the function
T (n) is bounded by O(n log n). If, on the other hand, we have a guess for
the running time that we want to verify, we can do so by plugging it into the
recurrence as follows.
Suppose we believe that T (n) cn log2 n for all n 2, and we want to
check whether this is indeed true. This clearly holds for n = 2, since in this
case cn log2 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 full
calculation.
T (n) 2T (n/2) + cn
2c (n/2) log2(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
Chapter 5 Divide and Conquer
An Approach Using Partial Substitution
There is a somewhat weaker kind of substitution one can do, in which one
guesses the overall form of the solution without pinning down the exact values
of all the constants and other parameters at the outset.
Specically, 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 rst 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 will
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 will let us apply the simplication log2(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
= kn[(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 just need to choose any
k that is at least as large as c , and we get
T (n) (kn log2 n) kn + cn kn log2 n ,
which completes the induction.
Thus the substitution method can actually be useful in working out the
exact constants when one has some guess of the general form of the solution.
5.2 Further Recurrence Relations
Weve just worked out the solution to a recurrence relation, (5.1), that will
come up in the design of several divide-and-conquer algorithms later in this
chapter. As a way to explore this issue further, we now consider a class
of recurrence relations that generalizes (5.1), and show how to solve the
recurrences in this class. Other members of this class will arise in the design
of algorithms both in this and in later chapters.
5.2 Further Recurrence Relations
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 nd 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 nding in Chapter 13.
If T (n) denotes the running time of an algorithm designed in this style,
then T (n) obeys the following recurrence relation, which directly generalizes
(5.1) by replacing 2 with q:
(5.3)
For some constant c,
T (n) qT (n/2) + cn
when n > 2, and
T (2) c .
We now describe how to solve (5.3) by the methods weve seen above:
unrolling, substitution, and partial substitution. We treat the cases q > 2 and
q = 1 separately, since they are qualitatively different from each otherand
different from the case q = 2 as well.
The Case of q > 2 Subproblems
We begin by unrolling (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 rst few levels: We show an example of this for the case
q = 3 in Figure 5.2. At the rst 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 a pattern: At an arbitrary level j , we have qj distinct instances,
each of size n/2j . Thus the total work performed at level j is qj (cn/2j ) =
(q/2)j cn.
215
216
Chapter 5 Divide and Conquer
cn time, plus
recursive calls
cn/2
cn/4
cn/4
Level 0: cn total
cn/2
cn/4
cn/4
cn/4
Level 1: cn/2 + cn/2 + cn/2 = (3/2)cn total
cn/2
cn/4
cn/4
cn/4
cn/4
Level 2: 9(cn/4) = (9/4)cn total
Figure 5.2 Unrolling the recurrence T (n) 3T (n/2) + O(n).
.
Summing over all levels of recursion: As before, there are log2 n levels of
recursion, and the total amount of work performed is the sum over all
these:
log2 n1
T (n)
j =0
q
2
log2 n1
j
cn = cn
j =0
q
2
j
.
This is a geometric sum, consisting of powers of r = q/2. We can use the
formula for a geometric sum when r > 1, which gives us the formula
T (n) cn
r log2 n 1
cn
r 1
r log2 n
r 1
.
Since were aiming for an asymptotic upper bound, it is useful to gure
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)
c
nr log2 n .
r 1
Finally, we need to gure out what r log2 n is. Here we use a very handy
identity, which says that, for any a > 1 and b > 1, we have a log b = blog a .
Thus
r log2 n = nlog2 r = nlog2(q/2) = n(log2 q)1.
Thus we have
T (n)
c
n n(log2 q)1
r 1
We sum this up as follows.
c
nlog2 q = O(nlog2 q ).
r 1
5.2 Further Recurrence Relations
(5.4)
Any function T () satisfying (5.3) with q > 2 is bounded by O(nlog2 q ).
So we nd that the running time is more than linear, since log2 q > 1,
but still polynomial in n. Plugging in specic values of q, the running time
is O(nlog2 3) = O(n1.59) when q = 3; and the running time is O(nlog2 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.
Applying Partial Substitution The appearance of log2 q in the exponent
followed naturally from our solution to (5.3), but its not necessarily an
expression one would have guessed at the outset. We now consider how an
approach based on partial substitution into the recurrence yields a different
way of discovering this exponent.
Suppose we guess that the solution to (5.3), when q > 2, has the form
T (n) knd 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
T (n) qk
=
n
2
d
+ cn
qd
kn + cn .
2d
This is remarkably close to something that works: if we choose d so that
q/2d = 1, then we have T (n) knd + cn, which is almost right except for the
extra term cn. So lets deal with these two issues: rst, how to choose d so we
get q/2d = 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 decide 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) knd n, where weve now decided that d = log2 q but we
havent xed the constants k or . Applying the new formula to T (n/2), this
expands to
217
218
Chapter 5 Divide and Conquer
T (n) qk
n
2
d
q
n
2
+ cn
qdq
kn n + cn
2d
2
q
= knd n + cn
2
q
= knd ( c )n .
2
=
This now works completely, if we simply choose so that ( q2 c ) = : in other
words, = 2c /(q 2). This completes the inductive step for n. We also need
to handle the base case n = 2, and this we do using the fact that the value of
k has not yet been xed: we choose k large enough so that the formula is a
valid upper bound for the case n = 2.
The Case of One Subproblem
We now consider the case of q = 1 in (5.3), since this illustrates an outcome
of yet another avor. 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 rst few levels: We show the rst few levels of the recursion
in Figure 5.3. At the rst 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 level when q = 1 is actually decreasing as we proceed through
the recursion.
.
Identifying a pattern: At an arbitrary level j , we still have just one
instance; it has size n/2j and contributes cn/2j 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:
log2 n1
T (n)
j =0
cn
= cn
2j
log2 n1
j =0
1
2j
.
This geometric sum is very easy to work out; even if we continued it to
innity, it would converge to 2. Thus we have
T (n) 2cn = O(n).
5.2 Further Recurrence Relations
cn time, plus
recursive calls
Level 0: cn total
cn/2
Level 1: cn/2 total
cn/4
Level 2: cn/4 total
Figure 5.3 Unrolling the recurrence T (n) T (n/2) + O(n).
We sum this up as follows.
(5.5)
Any function T () satisfying (5.3) with q = 1 is bounded by O(n).
This is counterintuitive when you rst 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 powerful 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) knd . 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
k
=
n
2
d
+ cn
kd
n + cn .
2d
If we now simply choose d = 1 and k = 2c , we have
T (n)
k
k
n + cn = ( + c )n = kn ,
2
2
which completes the induction.
The Effect of the Parameter q. It is worth reecting briey on the role of the
parameter q in the class of recurrences T (n) qT (n/2) + O(n) dened by (5.3).
When q = 1, the resulting 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 1 that
grows with q. The reason for this range of different running times lies in where
219
220
Chapter 5 Divide and Conquer
most of the work is spent in the recursion: when q = 1, the total running time
is dominated by the top level, whereas when q > 2 its dominated by the work
done on constant-size subproblems at the bottom of the recursion. Viewed this
way, we can appreciate that the recurrence for q = 2 really represents a knifeedgethe 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 nal 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
will 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 nal recombining.
For our purposes here, we note that this style of algorithm has a running time
T (n) that satises the following recurrence.
(5.6)
For some constant c,
T (n) 2T (n/2) + cn2
when n > 2, and
T (2) c .
Ones rst 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.
.
Analyzing the rst few levels: At the rst level of recursion, we have a
single problem of size n, which takes time at most cn2 plus the time spent
in all 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
5.3 Counting Inversions
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
(5.1); 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
n
bounded by 2j c ( 2j )2 = cn2/2j .
Summing over all levels of recursion: 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
log2 n1
T (n)
j =0
cn2
= cn2
2j
log2 n1
j =0
1
2j
2cn2 = O(n2) ,
where the second inequality follows from the fact that we have a convergent geometric sum.
In retrospect, our initial guess of T (n) = O(n2 log n), based on the analogy
to (5.1), was an overestimate because of how quickly n2 decreases as we
replace it with ( n )2, ( n )2, ( n )2, and so forth in the unrolling of the recurrence.
2
4
8
This means that we get a geometric sum, rather than one that grows by a xed
amount over all n levels (as in the solution to (5.1)).
5.3 Counting Inversions
Weve spent some time discussing approaches to solving a number of common
recurrences. The remainder of the chapter will illustrate the application of
divide-and-conquer to problems from a number of different domains; we will
use what weve seen in the previous sections to bound the running times
of these algorithms. We begin by showing how a variant of the Mergesort
technique can be used to solve a problem that is not directly related to sorting
numbers.
The Problem
We will consider a problem that arises in the analysis of rankings, which
are becoming important to a number of current applications. For example, a
number of sites on the Web make use of a technique known as collaborative
ltering , 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 identied people with similar tastes to yoursbased on a comparison
221
222
Chapter 5 Divide and Conquer
of how you and they rate various thingsit can recommend new things that
these other people have liked. Another application arises in meta-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 n movies, and then a collaborative ltering system
consults its database to look for other people who had similar rankings. But
whats a good way to measure, numerically, how similar two peoples rankings
are? Clearly an identical ranking is very similar, and a completely reversed
ranking is very different; we want something that interpolates through the
middle region.
Lets consider comparing your ranking and a strangers ranking of the
same set of n movies. A natural method would be to label the movies from
1 to n according to your ranking, then order these labels according to the
strangers ranking, and see how many pairs are out of order. More concretely,
we will consider the following problem. We are given a sequence of n numbers
a1, . . . , an ; we will assume that all the numbers are distinct. We want to dene
a measure that tells us how far this list is from being in ascending order; the
value of the measure should be 0 if a1 < a2 < . . . < an , and should increase as
the numbers become more scrambled.
2
1
4
2
1
3
3
4
5
5
Figure 5.4 Counting the
number of inversions in the
sequence 2, 4, 1, 3, 5. Each
crossing pair of line segments
corresponds to one pair that
is in the opposite order in
the input list and the ascending listin other words, an
inversion.
A natural way to quantify this notion is by counting the number of
inversions. We say that two indices i < j form an inversion if ai > aj , that is,
if the two elements ai and aj are out of order. We will seek to determine the
number of inversions in the sequence a1, . . . , an .
Just to pin down this denition, 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
order theyre provided, and below that in ascending order. We then draw a
line segment between each number in the top list and its copy in the lower
list. Each crossing pair of line segments corresponds to one pair that is in the
opposite order in the two listsin other words, an inversion.
Note how the number of inversions is a measure that smoothly interpolates
between complete agreement (when the sequence is in ascending order, then
there are no inversions) and complete disagreement (if the sequence is in
descending order, then every pair forms an inversion, and so there are n of
2
them).
5.3 Counting Inversions
Designing and Analyzing the Algorithm
What is the simplest algorithm to count inversions? Clearly, we could look
at every 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 () dened in Section 5.1. We set m = n/2 and divide the list into
the two pieces a1, . . . , am and am+1, . . . , an . We rst count the number of
inversions in each of these two halves separately. Then we count the number
of inversions (ai , 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 rst-half/second-half inversions have a particularly nice form: they
are precisely the pairs (ai , aj ), where ai is in the rst 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 well as counting
inversions) will make the combining portion of the algorithm easier.
So the crucial routine in this process is Merge-and-Count. Suppose we
have recursively sorted the rst and second halves of the list and counted the
inversions in each. We now have two sorted lists A and B, containing the rst
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 rst-half/second-half
inversions.
This is closely related to the simpler problem we discussed in Chapter 2,
which formed the corresponding combining step for Mergesort: 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
Chapter 5 Divide and Conquer
Elements inverted
with bj < ai
////// ai
A
/// bj
B
Merged result
Figure 5.5 Merging two sorted lists while also counting the number of inversions
between them.
at elements ai and bj . In one step, we compare the elements ai and bj 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 ai is appended to
C , no new inversions are encountered, since ai is smaller than everything left
in list B, and it comes before all of them. On the other hand, if bj 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 Current 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 bj be the elements pointed to by the Current pointer
Append the smaller of these two to the output list
If bj is the smaller element then
Increment Count by the number of elements remaining in A
Endif
Advance the Current pointer in the list from which the
smaller element was selected.
EndWhile
5.4 Finding the Closest Pair of Points
Once one list is empty, append the remainder of the other list
to the output
Return Count and the merged list
The running time of Merge-and-Count can be bounded by the analogue
of the argument we used for the original merging algorithm at the heart of
Mergesort: each iteration of the While loop takes constant time, and in each
iteration we add some element to the output that will never be seen again.
Thus the number of iterations can be at most the sum of the initial lengths of
A and B, and so the total running time is O(n).
We use this Merge-and-Count routine in a recursive procedure that
simultaneously sorts and counts the number of inversions in a list L.
Sort-and-Count(L)
If the list has one element then
there are no inversions
Else
Divide the list into two halves:
A contains the first n/2
elements
B contains the remaining n/2 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 running time
T (n) of the full Sort-and-Count procedure satises the recurrence (5.1). By
(5.2), we have
(5.7) The Sort-and-Count algorithm correctly sorts the input list and counts
the number of inversions; it runs in O(n log n) time for a list with n elements.
5.4 Finding the Closest Pair of Points
We now describe another problem that can be solved by an algorithm in the
style weve been discussing; but nding the right way to merge the solutions
to the two subproblems it generates requires quite a bit of ingenuity.
225
226
Chapter 5 Divide and Conquer
The Problem
The problem we consider is very simple to state: Given n points in the plane,
nd 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 project to work out efcient algorithms for basic computational primitives in geometry. These algorithms formed the foundations
of the then-edgling eld of computational 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 surprisingly
hard to nd an efcient algorithm for it. It is immediately clear that there is an
O(n2) solutioncompute the distance between each pair of points and take
the minimumand 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 will see that it is possible to further improve the running time
to O(n) using randomization.
Designing the Algorithm
We begin with a bit of notation. Let us denote the set of points by P =
{p1, . . . , pn }, where pi has coordinates (xi , yi ); and for two points pi , pj P ,
we use d(pi , pj ) to denote the standard Euclidean distance between them. Our
goal is to nd a pair of points pi , pj that minimizes d(pi , pj ).
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 true, 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 nd the closest pair of points on a line? Wed rst 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 nd the closest pair among the points in the left half of
5.4 Finding the Closest Pair of Points
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 (n2) such distances, yet we need to nd 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 rst.
It will be very useful if every recursive call, on a set P P , begins with two
lists: a list Px in which all the points in P have been sorted by increasing xcoordinate, and a list Py 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 rst level of recursion will work as follows, with all further levels
working in a completely analogous way. We dene Q to be the set of points
in the rst n/2 positions of the list Px (the left half) and R to be the set of
points in the nal n/2 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
Q
Line L
R
Figure 5.6 The first level of recursion: The point set P is divided evenly into Q and R by
the line L, and the closest pair is found on each side recursively.
227
228
Chapter 5 Divide and Conquer
following four lists: Qx , consisting of the points in Q sorted by increasing xcoordinate; Qy , consisting of the points in Q sorted by increasing y-coordinate;
and analogous lists Rx and Ry . For each entry of each of these lists, as before,
we record the position of the point in both lists it belongs to.
We now recursively determine a closest pair of points in Q (with access
to the lists Qx and Qy ). Suppose that q0 and q1 are (correctly) returned as a
closest pair of points in Q. Similarly, we determine a closest pair of points in
R, obtaining r0 and r1 .
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?
Let be the minimum of d(q0 ,q1 ) and d(r0 ,r1 ). The real question is: Are
there points q Q and r R for which d(q , r ) < ? 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 denition of x, we know that qx x rx . Then we have
x qx rx qx d(q , r ) <
and
rx x rx qx d(q , r ) < ,
so each of q and r has an x-coordinate within of x and hence lies within
distance of the line L.
So if we want to nd a close q and r , we can restrict our search to the
narrow band consisting only of points in P within of L. Let S 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.
229
5.4 Finding the Closest Pair of Points
(5.9) There exist q Q and r R for which d(q , r ) < if and only if there
exist s , s S for which d(s , s ) < .
Each box can
contain at most
one input point.
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 actually
far from true, as the following amazing fact shows.
(5.10) If s , s S have the property that d(s , s ) < , 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 /2. One row of Z will consist of four boxes whose horizontal
sides have the 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 Q or both belong
to R. But any two points in the same box are within distance 2/2 < ,
which contradicts our denition 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 ) < , 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 3/2
aparta 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 from each other. So having done this, we can compare the smallest
such distance to , and we can report one of two things: (i) the closest pair
of points in S, if their distance is less than ; or (ii) 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 (ii), 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
Line L
/2
/2
Boxes
Figure 5.7 The portion of the
plane close to the dividing
line L, as analyzed in the
proof of (5.10).
230
Chapter 5 Divide and Conquer
of y-coordinate. The reason such an approach works now is due to the extra knowledge (the value of ) 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 , 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)
(p0 , p1 ) = Closest-Pair-Rec(Px ,Py )
Closest-Pair-Rec(Px , Py )
If |P | 3 then
find closest pair by measuring all pairwise distances
Endif
Construct Qx , Qy , Rx , Ry (O(n) time)
(q0 ,q1 ) = Closest-Pair-Rec(Qx , Qy )
,r ) = Closest-Pair-Rec(R , R )
(r0 1
x
y
= min(d(q0 ,q1 ) , d(r0 ,r1 ))
x = maximum x-coordinate of a point in set Q
L = {(x,y) : x = x }
S = points in P within distance of L.
Construct Sy (O(n) time)
For each point s Sy , compute distance from s
to each of next 15 points in Sy
Let s, s be pair achieving minimum of these distances
(O(n) time)
If d(s,s ) < then
Return (s,s )
Else if d(q0 ,q1 ) < d(r0 ,r1 ) then
,q )
Return (q0 1
5.5 Integer Multiplication
Else
Return (r0 ,r1 )
Endif
Analyzing the Algorithm
We rst prove that the algorithm produces a correct answer, using the facts
weve established in the process of designing it.
(5.11)
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 t 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 (5.10) and (5.9), the remainder of the algorithm
correctly determines whether any pair of points in S is at distance less than
, 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 , and it is correctly found by the remainder
of the algorithm.
We now bound the running time as well, using (5.2).
(5.12)
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 satises the recurrence
(5.1), and hence is O(n log n) by (5.2).
5.5 Integer Multiplication
We now discuss a different application of divide and conquer, in which the
default quadratic algorithm is improved by means of a different recurrence.
The analysis of the faster algorithm will exploit one of the recurrences considered in Section 5.2, in which more than two recursive calls are spawned at
each level.
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
Chapter 5 Divide and Conquer
12
13
36
12
156
1100
1101
1100
0000
1100
1100
10011100
(a)
(b)
Figure 5.8 The elementary-school algorithm for multiplying two integers, in (a) decimal
and (b) binary representation.
even as an algorithmic question. But, in fact, elementary schoolers are taught a
concrete (and quite efcient) algorithm to multiply two n-digit numbers x and
y. You rst compute a partial product by multiplying each digit of y 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 single
operation on a pair of bits as one primitive step in this computation, it takes
O(n) time to compute each partial product, and O(n) time to combine it in
with the running sum of all partial products so far. Since there are n partial
products, this is a total running time of O(n2).
If you havent thought about this much since elementary school, theres
something initially striking about the prospect of improving on this algorithm.
Arent all those partial products necessary in some way? But, in fact, it
is possible to improve on O(n2) time using a different, recursive way of
performing the multiplication.
Designing the Algorithm
The improved algorithm is based on a more clever way to break up the product
into partial sums. Lets assume were in base-2 (it doesnt really matter), and
start by writing x as x1 2n/2 + x0. In other words, x1 corresponds to the highorder n/2 bits, and x0 corresponds to the low-order n/2 bits. Similarly, we
write y = y1 2n/2 + y0. Thus, we have
xy = (x1 2n/2 + x0)(y1 2n/2 + y0)
= x1y1 2n + (x1y0 + x0y1) 2n/2 + x0y0 .
(5.1)
Equation (5.1) reduces the problem of solving a single n-bit instance
(multiplying the two n-bit numbers x and y) to the problem of solving four n/2bit instances (computing the products x1y1, x1y0 , x0y1, and x0y0). So we have
a rst candidate for a divide-and-conquer solution: recursively compute the
results for these four n/2-bit instances, and then combine them using Equation
5.5 Integer Multiplication
(5.1). The combining of the solution requires a constant number of additions
of O(n)-bit numbers, so it takes time O(n); thus, the running time T (n) is
bounded by the recurrence
T (n) 4T (n/2) + cn
for a constant c . Is this good enough to give us a subquadratic running time?
We can work out the answer by observing that this is just the case q = 4 of
the class of recurrences in (5.3). As we saw earlier in the chapter, the solution
to this is T (n) O(nlog2 q ) = 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(nlog2 q ) = O(n1.59).
Recall that our goal is to compute the expression x1y1 2n + (x1y0 + x0y1)
+ x0y0 in Equation (5.1). It turns out there is a simple trick that lets us
determine all of the terms in this expression using just three recursive calls. The
trick is to consider the result of the single multiplication (x1 + x0)(y1 + y0) =
x1y1 + x1y0 + x0y1 + x0y0. This has the four products above added together, at
the cost of a single recursive multiplication. If we now also determine x1y1 and
x0y0 by recursion, then we get the outermost terms explicitly, and we get the
middle term by subtracting x1y1 and x0y0 away from (x1 + x0)(y1 + y0).
2n/2
Thus, in full, our algorithm is
Recursive-Multiply(x,y):
Write x = x1 2n/2 + x0
y = y1 2n/2 + y0
Compute x1 + x0 and y1 + y0
p = Recursive-Multiply(x1 + x0 , y1 + y0)
x1y1 = Recursive-Multiply(x1, y1)
x0y0 = Recursive-Multiply(x0 , y0)
Return x1y1 2n + (p x1y1 x0y0) 2n/2 + x0y0
Analyzing the Algorithm
We can determine the running time of this algorithm as follows. Given two nbit numbers, it performs a constant number of additions on O(n)-bit numbers,
in addition to the three recursive calls. Ignoring for now the issue that x1 + x0
and y1 + y0 may have n/2 + 1 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
Chapter 5 Divide and Conquer
a three-way branching one, with a running time that satises
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
(5.13) The running time of Recursive-Multiply on two n-bit factors is
O(nlog2 3) = O(n1.59).
5.6 Convolutions and the Fast Fourier Transform
As a nal 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 = (a0 , a1, . . . , an1) and b = (b0 , b1, . . . , bn1), there are
a number of common ways of combining them. For example, one can compute
the sum, producing the vector a + b = (a0 + b0 , a1 + b1, . . . , an1 + bn1);
or one can compute the inner product, producing the real number a b =
a0b0 + a1b1 + . . . + an1bn1. (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
ai bj .
(i , j ):i+j =k
i , j <n
In other words,
a b = (a0b0 , a0b1 + a1b0 , a0b2 + a1b1 + a2b0 , . . . ,
an2bn1 + an1bn2 , an1bn1).
This denition is a bit hard to absorb when you rst see it. Another way to
think about the convolution is to picture an n n table whose (i , j ) entry is
ai bj , like this,
5.6 Convolutions and the Fast Fourier Transform
a0 b 0
a0 b1
a1b0
a1b1
a2b0
a2b1
...
...
an1b0 an1b1
. . . a0bn2
a0bn1
. . . a1bn2
a1bn1
. . . a2bn2
a2bn1
...
...
...
. . . an1bn2 an1bn1
and then to compute the coordinates in the convolution vector by summing
along the diagonals.
Its worth mentioning that, unlike the vector sum and inner product,
the convolution can be easily generalized to vectors of different lengths,
a = (a0 , a1, . . . , am1) and b = (b0 , b1, . . . , bn1). In this more general case,
we dene a b to be a vector with m + n 1 coordinates, where coordinate
k is equal to
a i bj .
(i , j ):i+j =k
i<m , j <n
We can picture this using the table of products ai bj 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 dened.)
Its not just the denition of a convolution that is a bit hard to absorb at
rst; the motivation for the denition 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 rst example (which also proves that the convolution is something that
we all saw implicitly in high school) is polynomial multiplication. Any
polynomial A(x) = a0 + a1x + a2x2 + . . . am1xm1 can be represented
just as naturally using its vector of coefcients, a = (a0 , a1, . . . , am1).
Now, given two polynomials A(x) = a0 + a1x + a2x2 + . . . am1xm1 and
B(x) = b0 + b1x + b2x2 + . . . bn1xn1, consider the polynomial C (x) =
A(x)B(x) that is equal to their product. In this polynomial C (x), the
coefcient on the xk term is equal to
ck =
ai bj .
(i , j ):i+j =k
In other words, the coefcient vector c of C (x) is the convolution of the
coefcient 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 ll an entire course, so
235
236
Chapter 5 Divide and Conquer
well just give a simple example here to suggest one way in which the
convolution arises.
Suppose we have a vector a = (a0 , a1, . . . , am1) 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 uctuations, 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
ai =
1
Z
i+k
aj e (ji) ,
2
j =ik
for some width parameter k, and with Z chosen simply to normalize
the weights in the average to add up to 1. (There are some issues with
boundary conditionswhat do we do when i k < 0 or i + k > m?but
we could deal with these, for example, by discarding the rst 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 rst dene a mask
w = (wk , w(k1) , . . . , w1, w0 , w1, . . . , wk1, wk )
consisting of the weights we want to use for averaging each point with
2
2
1
its neighbors. (For example, w = Z (e k , e (k1) , . . . , e 1, 1, e 1, . . . ,
2
2
e (k1) , e k ) in the Gaussian case above.) We then iteratively position
this mask so it is centered at each possible point in the sequence a ; and
for each positioning, we compute the weighted average. In other words,
we replace ai with ai = k=k wsai+s.
s
This last expression is essentially a convolution; we just have to
warp the notation a bit so that this becomes clear. Lets dene b =
(b0 , b1, . . . , b2k ) by setting b = wk . Then its not hard to check that
with this denition we have the smoothed value
ai =
aj b .
(j , ):j + =i+k
In other words, the smoothed sequence is just the convolution of the
original signal and the reverse of the mask (with some meaningless
coordinates at the beginning and end).
5.6 Convolutions and the Fast Fourier Transform
.
We mention one nal application: the problem of combining histograms.
Suppose were studying a population of people, and we have the following two histograms: One shows the annual income of all the men in the
population, and one shows the annual income of all the women. Wed
now like to produce a new histogram, showing for each k the number of
pairs (M , W ) for which man M and woman W have a combined income
of k.
This is precisely a convolution. We can write the rst histogram as a
vector a = (a0 , . . . , am1), to indicate that there are ai men with annual
income equal to i. We can similarly write the second histogram as a
vector b = (b0 , . . . , bn1). Now, let ck 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 bj , for any pair (i , j ) where
i + j = k. In other words,
ck =
ai bj .
(i , j ):i+j =k
so the combined histogram c = (c0 , . . . , cm+n2) is simply the convolution of a and b.
(Using terminology from probability that we will develop in Chapter 13, one can view this example as showing how convolution is the
underlying means for computing the distribution of the sum of two independent random variables.)
Computing the Convolution Having now motivated the notion of convolution, lets discuss the problem of computing it efciently. 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 rst
appear. The denition of convolution, after all, gives us a perfectly valid way
to compute it: for each k, we just calculate the sum
ai bj
(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 ai 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 denition involves O(n2) multiplications
ai bj . 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).
237
238
Chapter 5 Divide and Conquer
Could one design an algorithm that bypasses the quadratic-size denition of
convolution and computes it in some smarter way?
In fact, quite surprisingly, this is possible. We now describe a method
that computes the convolution of two vectors using only O(n log n) arithmetic
operations. The crux of this method is a powerful technique known as the Fast
Fourier Transform (FFT). The FFT has a wide range of further applications in
analyzing sequences of numerical values; computing convolutions quickly,
which we focus on here, is just one of these applications.
Designing and Analyzing the Algorithm
To break through the quadratic time barrier for convolutions, we are going
to exploit the connection between the convolution and the multiplication of
two polynomials, as illustrated in the rst 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 = (a0 , a1, . . . , an1) and b = (b0 ,
b1, . . . , bn1). We will view them as the polynomials A(x) = a0 + a1x + a2x2 +
. . . an1xn1 and B(x) = b0 + b1x + b2x2 + . . . bn1xn1, and well seek to compute their product C (x) = A(x)B(x) in O(n log n) time. If c = (c0 , c1, . . . , c2n2)
is the vector of coefcients 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 coefcients 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 x1, x2 , . . . , x2n and evaluate A(xj ) and B(xj ) for
each of j = 1, 2, . . . , 2n.
(ii) We can now compute C (xj ) for each j 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 x1, 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 1, their product C has degree at most 2n 2, and so it can be
reconstructed from the values C (x1) , C (x2) , . . . , C (x2n ) that we computed
in step (ii).
This approach to multiplying polynomials has some promising aspects
and some problematic ones. First, the good news: step (ii) requires only
5.6 Convolutions and the Fast Fourier Transform
O(n) arithmetic operations, since it simply involves the multiplication of O(n)
numbers. But the situation doesnt look as hopeful with steps (i) and (iii). In
particular, evaluating the polynomials A and B on a single value takes (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 nd a set of 2n values
x1, 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 will turn out to work very well is the complex roots of unity.
The Complex Roots of Unity At this point, were going to need to recall 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 re i ,
where e i = 1 (and e 2 i = 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 j , k = e 2 ji/k (for j = 0, 1, 2, . . . , k 1) satises
the equation, since
(e 2 ji/k )k = e 2 ji = (e 2 i )j = 1j = 1,
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 x1, . . . , 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
i
1
1
i
Figure 5.9 The 8th roots of unity in the complex plane.
239
240
Chapter 5 Divide and Conquer
polynomial P by its values on the (d + 1)st roots of unity is sometimes referred
to as the discrete Fourier transform of P ; and the heart of our procedure is a
method for making this computation fast.
A Recursive Procedure for Polynomial Evaluation 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 n is a
power of 2.
How does one break the evaluation of a polynomial into two equal-sized
subproblems? A useful trick is to dene two polynomials, Aeven (x) and Aodd (x),
that consist of the even and odd coefcients of A, respectively. That is,
Aeven (x) = a0 + a2x + a4x2 + . . . + an2x(n2)/2 ,
and
Aodd (x) = a1 + a3x + a5x2 + . . . + a(n1)x(n2)/2.
Simple algebra shows us that
A(x) = Aeven (x2) + xAodd (x2) ,
and so this gives us a way to compute A(x) in a constant number of operations,
given the evaluation of the two constituent polynomials that each have half
the degree of A.
Now suppose that we evaluate each of Aeven and Aodd on the nth 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 n roots of unity rather than 2n. Thus we can
perform these evaluations in time T (n/2) for each of Aeven and Aodd , 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 Aeven and Aodd . Consider
one of these roots of unity j ,2n = e 2 ji/2n . The quantity j2,2n is equal to
(e 2 ji/2n )2 = e 2 ji/n , and hence j2,2n is an nth root of unity. So when we go
to compute
A(j ,2n ) = Aeven (j2,2n ) + j ,2n Aodd (j2,2n ) ,
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(j ,2n ) using a
5.6 Convolutions and the Fast Fourier Transform
constant number of operations. Doing this for all 2n roots of unity is therefore
O(n) additional operations after the two recursive calls, and so the bound
T (n) on the number of operations indeed satises 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 (j , n ) = A(j ,2n )B(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 dening 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.
n
Consider a polynomial C (x) = 2=1 csxs that we want to reconstruct
s0
from its values C (s ,2n ) at the (2n)th roots of unity. Dene a new polynomial
n
D(x) = 2=1 dsxs, where ds = C (s ,2n ). We now consider the values of D(x)
s0
at the (2n)th roots of unity.
2n1
C (s ,2n )js,2n
D(j ,2n ) =
s=0
2n1 2n1
=
t
ct s ,2n )js,2n
(
s=0
t =0
2n1
=
2n1
t
s ,2n js,2n ) ,
ct (
t =0
s=0
by denition. Now recall that s ,2n = (e 2 i/2n )s. Using this fact and extending
the notation to s ,2n = (e 2 i/2n )s even when s 2n, we get that
2n1
D(j ,2n ) =
2n1
ct (
t =0
s=0
2n1
e (2 i)(st +js)/2n )
2n1
ct (
=
t =0
s=0
s
t +j ,2n ).
241
242
Chapter 5 Divide and Conquer
To analyze the last line, we use the fact that for any (2n)th root of unity = 1,
2n
we have s=1 s = 0. This is simply because is by denition a root of
0
2n
x2n 1 = 0; since x2n 1 = (x 1)( t =1 xt ) and = 1, it follows that is
0
2n1 t
also a root of ( t =0 x ).
Thus the only term of the last lines outer sum that is not equal to 0 is
for ct such that t +j ,2n = 1; and this happens if t + j is a multiple of 2n, that
n
n
s
is, if t = 2n j . For this value, 2=1 t +j ,2n = 2=1 1 = 2n. So we get that
s0
s0
D(j ,2n ) = 2nc2nj . Evaluating the polynomial D(x) at the (2n)th roots of unity
thus gives us the coefents of the polynomial C (x) in reverse order (multiplied
by 2n each). We sum this up as follows.
(5.14)
D(x) =
n
For any polynomial C (x) = 2=1 csxs, and corresponding polynomial
s0
2n1
1
s
s=0 C (s ,2n )x , we have that cs = 2n D(2ns ,2n ).
We can do all the evaluations of the values D(2ns ,2n ) in O(n log 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 coefcients of C are the
coordinates in the convolution vector c = a b that we were originally seeking.
In summary, we have shown the following.
(5.15) Using the Fast Fourier Transform to determine the product polynomial
C (x), we can compute the convolution of the original vectors a and b in
O(n log n) time.
Solved Exercises
Solved Exercise 1
Suppose you are given an array A with n entries, with each entry holding a
distinct number. You are told that the sequence of values A[1], A[2], . . . , 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
until 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 plotted points would
rise until x-value p, where theyd achieve their maximum, and then fall from
there on.)
Youd like to nd the peak entry p without having to read the entire
arrayin fact, by reading as few entries of A as possible. Show how to nd
the entry p by reading at most O(log n) entries of A.
Solved Exercises
Solution Lets start with a general discussion on how to achieve a running
time of O(log n) and then come back to the specic 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 most n/2. As in the chapter, we will assume that the recursion
bottoms out when n = 2, performing at most c operations to nish the
computation. If T (n) denotes the running time on an input of size n, then
we have the recurrence
(5.16)
T (n) T (n/2) + c
when n > 2, and
T (2) c .
It is not hard to solve this recurrence by unrolling it, as follows.
.
Analyzing the rst few levels: At the rst 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 a pattern: 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 j .
.
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 n).
We can also do this by partial substitution. Suppose we guess that T (n)
k logb 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 logb(n/2) + c
= k logb n k logb 2 + c .
243
244
Chapter 5 Divide and Conquer
The rst term on the right 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 recursion involves at most c operations.
Now lets get back to the problem at hand. If we wanted to set ourselves
up to use (5.16), 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[n/2 1] < A[n/2] < A[n/2 + 1], then entry n/2 must come strictly
before p, and so we can continue recursively on entries n/2 + 1 through n.
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 n/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 xed 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
Solved Exercises
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 nd the correct numbers i and j 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 rst n/2 days and the nal
n/2 days separately, solving the problem recursively on each of these two
sets, and then gure out how to get an overall solution from this in O(n) time.
This would give us the usual recurrence T (n) 2T n + O(n), and hence
2
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 O() notation).
Now, let S be the set of days 1, . . . , n/2, and S be the set of days n/2 +
1, . . . , n. Our divide-and-conquer algorithm will be based on the following
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.
.
The optimal solution on S.
.
The optimal solution on S .
.
The maximum of p(j ) p(i), over i S and j S .
The rst two alternatives are computed in time T (n/2), each by recursion,
and the third alternative is computed by nding the minimum in S and the
245
246
Chapter 5 Divide and Conquer
maximum in S , which takes time O(n). Thus the running time T (n) satises
T (n) 2T
n
2
+ O(n) ,
as desired.
We note that this is not the best running time achievable for this problem.
In fact, one can nd the optimal pair of days in O(n) time 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 valuesso there are
2n values totaland 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. In a single query, you can specify a value k to one of the
two databases, and the chosen database will return the kth 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.
2. Recall the problem of finding the number of inversions. As in the text,
we are given a sequence of n numbers a1, . . . , an , which we assume are all
distinct, and we define an inversion to be a pair i < j such that ai > aj .
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 if i < j and
ai > 2aj . Give an O(n log n) algorithm to count the number of significant
inversions between two orderings.
3. 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
Exercises
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.
4. 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
as consisting of the points {1, 2, 3, . . . , 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 computational prediction. This computational
part is where they need your help. The total net force on particle j , by
Coulombs Law, is equal to
Cqi qj
Fj =
i <j
(j i)2
Cqi qj
i>j
( j i )2
Theyve written the following simple program to compute Fj for all j :
For j = 1, 2, . . . , n
Initialize Fj to 0
For i = 1, 2, . . . , n
If i < j then
C q qj
Add (j i i)2 to Fj
Else if i > j then
C q qj
Add (j i i)2 to Fj
Endif
Endfor
Output Fj
Endfor
247
248
Chapter 5 Divide and Conquer
Its not hard to analyze the running time of this program: each
invocation of the inner loop, over i, takes O(n) time, and this inner loop
is invoked O(n) times total, so the overall running time is O(n2).
The trouble is, for the large values of n theyre working with, the program takes several minutes to run. On the other hand, their experimental
setup is optimized so that they can throw down n particles, perform the
measurements, and be ready to handle n more particles within a few seconds. So theyd really like it if there were a way to compute all the forces
Fj 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
Fj in O(n log n) time.
5. 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 lines in the plane, labeled L1, . . . , Ln , with the ith line specified
by the equation y = ai x + bi . We will make the assumption that no three of
the lines all meet at a single point. We say line Li is uppermost at a given
x-coordinate x0 if its y-coordinate at x0 is greater than the y-coordinates
of all the other lines at x0: ai x0 + bi > aj x0 + bj for all j = i. We say line Li is
visible if there is some x-coordinate at which it is uppermostintuitively,
some portion of it can be seen if you look down from y = .
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.
6. Consider an n-node complete binary tree T , where n = 2d 1 for 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 if 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 xv by probing the node v. Show how to find a local minimum of
T using only O(log n) probes to the nodes of T .
7. Suppose now that youre given an n n grid graph G . (An n n grid graph
is just the adjacency graph of an n n chessboard. To be completely
precise, it is a graph whose node set is the set of all ordered pairs of
Notes and Further Reading
1
5
4
2
3
Figure 5.10 An instance of hidden surface removal with five lines (labeled 15 in the
figure). All the lines except for 2 are visible.
natural numbers (i , j ), where 1 i n and 1 j n; the nodes (i , j ) and
(k , ) are joined by an edge if and only if |i k| + |j | = 1.)
We use some of the terminology of the previous question. Again,
each node v is labeled by a real number xv ; you may assume that all these
labels are distinct. Show how to find a local minimum of G using only
O(n) probes to the nodes of G . (Note that G has n2 nodes.)
Notes and Further Reading
The militaristic coinage divide and conquer was introduced somewhat after
the technique itself. Knuth (1998) credits John von 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 eld
of computational geometry; the survey paper by Smid (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, Smid also makes the interesting historical observation that researchers originally suspected quadratic
time might be the best one could do for nding the closest pair of points in
the plane.) More generally, the divide-and-conquer approach has proved very
useful in computational geometry, and the books by Preparata and Shamos
249
250
Chapter 5 Divide and Conquer
(1985) and de Berg et al. (1997) give many further examples of this technique
in the design of geometric algorithms.
The algorithm for multiplying two n-bit integers in subquadratic time is
due to Karatsuba and Ofman (1962). Further background on asymptotically fast
multiplication algorithms is given by Knuth (1997b). Of course, the number
of bits in the input must be sufciently large for any of these subquadratic
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 6
Dynamic Programming
We began our study of algorithmic techniques with greedy algorithms, which
in some sense form the most natural approach to algorithm design. Faced with
a new computational problem, weve seen that its not hard to propose multiple
possible greedy algorithms; the challenge is then to determine whether any of
these algorithms provides a correct solution to the problem in all cases.
The problems we saw in Chapter 4 were all unied 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 difculty 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
of divide and conquer that we saw in the previous chapter are often not strong
enough to reduce exponential brute-force search down to polynomial time.
Rather, as we noted in Chapter 5, the applications there tended to reduce a
running time that was unnecessarily large, but already polynomial, down to a
faster running 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
Chapter 6 Dynamic Programming
brute-force search: although its systematically working through the exponentially large set of possible solutions to the problem, it does this without ever
examining them all explicitly. It is because of this careful balancing act that
dynamic programming can be a tricky technique to get used to; it typically
takes a reasonable amount of practice before one is fully comfortable with it.
With this in mind, we now turn to a rst example of dynamic programming: the Weighted Interval Scheduling Problem that we dened back in
Section 1.2. We are going to develop a dynamic programming algorithm for
this problem in two stages: rst as a recursive procedure that closely resembles
brute-force search; and then, by reinterpreting this procedure, as an iterative
algorithm that works by building up solutions to larger and larger subproblems.
6.1 Weighted Interval Scheduling:
A Recursive Procedure
We have seen that a particular greedy algorithm produces an optimal solution
to the Interval Scheduling Problem, where the goal is to accept as large a
set of nonoverlapping intervals as possible. The Weighted Interval Scheduling
Problem is a strictly more general version, in which each interval has a certain
value (or weight ), and we want to accept a set of maximum value.
Designing a Recursive Algorithm
Since the original Interval Scheduling Problem is simply the special case in
which all values are equal to 1, we know already that most greedy algorithms
will not solve this problem optimally. But even the algorithm that worked
before (repeatedly choosing the interval that ends earliest) is no longer optimal
in this more general setting, as the simple example in Figure 6.1 shows.
Indeed, no natural greedy algorithm is known for this problem, which is
what motivates our switch to dynamic programming. As discussed above, we
will 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.
Index
1
2
3
Value = 1
Value = 3
Value = 1
Figure 6.1 A simple instance of weighted interval scheduling.
6.1 Weighted Interval Scheduling: A Recursive Procedure
We use the notation from our discussion of Interval Scheduling in Section 1.2. We have n requests labeled 1, . . . , n, with each request i specifying a
start time si and a nish time fi . Each interval i now also has a value , or weight
vi . Two intervals are compatible if they do not overlap. The goal of our current
problem is to select a subset S {1, . . . , n} of mutually compatible intervals,
so as to maximize the sum of the values of the selected intervals, iS vi .
Lets suppose that the requests are sorted in order of nondecreasing nish
time: f1 f2 . . . fn . Well say a request i comes before a request j if i < j .
This will be the natural left-to-right order in which well consider intervals.
To help in talking about this order, we dene p(j ), for an interval j , to be the
largest index i < j such that intervals i and j are disjoint. In other words, i
is the leftmost interval that ends before j begins. We dene p(j ) = 0 if no
request i < j is disjoint from j . An example of the denition of p(j ) is shown
in Figure 6.2.
Now, given an instance of the Weighted Interval Scheduling Problem, lets
consider an optimal solution O, ignoring for now that we have no idea what
it is. Heres something completely obvious that we can say about O: either
interval n (the last one) belongs to O, or it doesnt. Suppose we explore both
sides of this dichotomy a little further. If n O, then clearly no interval indexed
strictly between p(n) and n can belong to O, because by the denition of p(n),
we know that intervals p(n) + 1, p(n) + 2, . . . , n 1 all overlap interval n.
Moreover, if n O, then O must include an optimal solution to the problem
consisting of requests {1, . . . , p(n)}for if it didnt, we could replace Os
choice of requests from {1, . . . , p(n)} with a better one, with no danger of
overlapping request n.
Index
1
2
3
4
5
6
v1 = 2
p(1) = 0
v2 = 4
p(2) = 0
v3 = 4
p(3) = 1
v4 = 7
p(4) = 0
v5 = 2
v6 = 1
p(5) = 3
p(6) = 3
Figure 6.2 An instance of weighted interval scheduling with the functions p(j ) defined
for each interval j .
253
254
Chapter 6 Dynamic Programming
On the other hand, if n O, then O is simply equal to the optimal solution
to the problem consisting of requests {1, . . . , n 1}. This is by completely
analogous reasoning: were assuming that O does not include request n; so if
it does not choose the optimal set of requests from {1, . . . , n 1}, we could
replace it with a better one.
All this suggests that nding the optimal solution on intervals {1, 2, . . . , n}
involves looking at the optimal solutions of smaller problems of the form
{1, 2, . . . , j }. Thus, for any value of j between 1 and n, let Oj denote the optimal
solution to the problem consisting of requests {1, . . . , j }, and let OPT(j ) denote
the value of this solution. (We dene 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 On , with value OPT(n). For the optimal solution Oj
on {1, 2, . . . , j }, our reasoning above (generalizing from the case in which
j = n) says that either j Oj , in which case OPT(j ) = vj + OPT(p(j )), or j Oj ,
in which case OPT(j ) = OPT(j 1). Since these are precisely the two possible
choices (j Oj or j Oj ), we can further say that
(6.1)
OPT(j ) = max(vj
+ OPT(p(j )) , OPT(j 1)).
And how do we decide whether n belongs to the optimal solution Oj ? This
too is easy: it belongs to the optimal solution if and only if the rst of the
options above is at least as good as the second; in other words,
(6.2) Request j belongs to an optimal solution on the set {1, 2, . . . , j } if and
only if
vj + OPT(p(j )) OPT(j 1).
These facts form the rst crucial component on which a dynamic programming solution is based: a recurrence equation that expresses the 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
signicant development. It directly gives us a recursive algorithm to compute
OPT(n), assuming that we have already sorted the requests by nishing time
and computed the values of p(j ) for each j .
Compute-Opt(j )
If j = 0 then
Return 0
Else
Return max(vj +Compute-Opt(p(j)), Compute-Opt(j 1))
Endif
6.1 Weighted Interval Scheduling: A Recursive Procedure
The correctness of the algorithm follows directly by induction on j :
(6.3)
Compute-Opt(j ) correctly computes OPT(j ) for each j = 1, 2, . . . , n.
Proof. By denition OPT(0) = 0. Now, take some j > 0, and suppose by way
of induction that Compute-Opt(i) correctly computes OPT(i) for all i < j . By
the induction hypothesis, we know that Compute-Opt(p(j )) = OPT(p(j )) and
Compute-Opt(j 1) = OPT(j 1); and hence from (6.1) it follows that
OPT(j ) = max(vj
+ Compute-Opt(p(j )) , Compute-Opt(j 1))
= Compute-Opt(j ).
Unfortunately, if we really implemented the algorithm Compute-Opt as
just written, it would take exponential time to run in the worst case. For
example, see Figure 6.3 for the tree of calls issued for the instance of Figure 6.2:
the tree widens very quickly due to the recursive branching. To take a more
extreme example, on a nicely layered instance like the one in Figure 6.4, where
p(j ) = j 2 for each j = 2, 3, 4, . . . , 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)
OPT(5)
OPT(4)
OPT(3)
OPT(3)
OPT(2)
OPT(1)
OPT(3)
OPT(2)
OPT(2)
OPT(1)
OPT(1)
OPT(1)
OPT(1)
The tree of subproblems
grows 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
Chapter 6 Dynamic Programming
Figure 6.4 An instance of weighted interval scheduling on which the simple ComputeOpt recursion will take exponential time. The values of all intervals in this instance
are 1.
like the Fibonacci numbers, which increase exponentially. Thus we have not
achieved a polynomial-time solution.
Memoizing the Recursion
In fact, though, were not so far from having a polynomial-time algorithm.
A fundamental observation, which forms the second crucial component of a
dynamic programming solution, is that our recursive algorithm ComputeOpt is really only solving n + 1 different subproblems: Compute-Opt(0),
Compute-Opt(1) , . . . , Compute-Opt(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-Opt in a globally accessible place the rst 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-Opt. This procedure will make use of an array M [0 . . . n]; M [j ] will
start with the value empty, but will hold the value of Compute-Opt(j ) as
soon as it is rst determined. To determine OPT(n), we invoke M-ComputeOpt(n).
M-Compute-Opt(j )
If j = 0 then
Return 0
Else if M [j ] is not empty then
Return M [j ]
Else
6.1 Weighted Interval Scheduling: A Recursive Procedure
Define M [j ] = max(vj +M-Compute-Opt(p(j )), M-Compute-Opt(j 1))
Return M [j ]
Endif
Analyzing the Memoized Version
Clearly, this looks very similar to our previous implementation of the algorithm; however, memoization has brought the running time way down.
(6.4) The running time of M-Compute-Opt(n) is O(n) (assuming the input
intervals are sorted by their nish times).
Proof. The time spent in a single call to M-Compute-Opt is O(1), excluding the
time spent in recursive calls it generates. So the running time is bounded by a
constant times the number of calls ever issued to M-Compute-Opt. Since the
implementation itself gives no explicit upper bound on this number of calls,
we try to nd 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-Opt, it lls in a new
entry, and hence increases the number of lled-in entries by 1. Since M has
only n + 1 entries, it follows that there can be at most O(n) calls to M-ComputeOpt, and hence the running time of M-Compute-Opt(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 full optimal set of intervals as well. It would be easy to extend
M-Compute-Opt 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
set of intervals among {1, 2, . . . , i}. Naively enhancing the code to maintain
the solutions in the array S, however, would blow up the running 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 S array takes O(n) time. We can avoid
this O(n) blow-up by not explicitly maintaining S, but rather 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
of intervals {1, . . . , j } if and only if vj + OPT(p(j )) OPT(j 1). Using this
observation, we get the following simple procedure, which traces back
through the array M to nd the set of intervals in an optimal solution.
257
258
Chapter 6 Dynamic Programming
Find-Solution(j )
If j = 0 then
Output nothing
Else
If vj + M [p(j )] M [j 1] then
Output j together with the result of Find-Solution(p(j ))
Else
Output the result of Find-Solution(j 1)
Endif
Endif
Since Find-Solution calls itself recursively only on strictly smaller values, it makes a total of O(n) recursive calls; and since it spends constant time
per call, we have
(6.5) Given the array M of the optimal values of the sub-problems, FindSolution returns an optimal solution in O(n) time.
6.2 Principles of Dynamic Programming:
Memoization or Iteration over Subproblems
We now use the algorithm for the Weighted Interval Scheduling Problem
developed in the previous section to summarize the basic principles of dynamic
programming, and also to 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 rst designing an exponential-time
recursive algorithm and then converting it (by memoization) to an efcient
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.
Designing the Algorithm
The key to the efcient algorithm is really the array M . It encodes the notion
that we are using the value of optimal solutions to the subproblems on intervals
{1, 2, . . . , j } for each j , and it uses (6.1) to dene the value of M [j ] based on
6.2 Principles of Dynamic Programming
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-Solution can be used to trace back through M efciently 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 [0] = 0 and keep incrementing j ; each time we need to determine
a value M [j ], the answer is provided by (6.1). The algorithm looks as follows.
Iterative-Compute-Opt
M [0] = 0
For j = 1, 2, . . . , n
M [j ] = max(vj + M [p(j )], M [j 1])
Endfor
Analyzing the Algorithm
By exact analogy with the proof of (6.3), we can prove by induction on j that
this algorithm writes OPT(j ) in array entry M [j ]; (6.1) provides the induction
step. Also, as before, we can pass the lled-in array M to Find-Solution to
get an optimal solution in addition to the value. Finally, the running time
of Iterative-Compute-Opt 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-Opt is depicted in
Figure 6.5. In each iteration, the algorithm lls in one additional entry of the
array M , by comparing the value of vj + M [p(j )] to the value of M [j 1].
A Basic Outline of Dynamic Programming
This, then, provides a second efcient 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 will develop dynamic
programming algorithms using the second type of approachiterative building up of subproblemsbecause 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 generally 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 satises a few basic properties.
259
260
Chapter 6 Dynamic Programming
0
1
2
p(1) = 0
w2 = 4
p(2) = 0
w3 = 4
3
p(4) = 0
w5 = 2
5
w6 = 1
6
(a)
2
2
3
4
5
0
2
4
0
2
4
6
0
2
4
6
7
2
4
6
7
8
2
4
6
7
8
6
p(3) = 1
w4 = 7
4
1
0
w1 = 2
M= 0
0
Index
p(5) = 3
p(6) = 3
8
(b)
Figure 6.5 Part (b) shows the iterations of Iterative-Compute-Opt on the sample
instance of Weighted Interval Scheduling depicted in part (a).
(i) There are only a polynomial number of subproblems.
(ii) The solution to the original problem can be easily computed from the
solutions to the subproblems. (For example, the original problem may
actually be one of the subproblems.)
(iii) There is a natural ordering on subproblems from smallest to largest,
together with an easy-to-compute recurrence (as in (6.1) and (6.2)) that
allows one to determine the solution to a subproblem from the solutions
to some number of smaller subproblems.
Naturally, these are informal guidelines. In particular, the notion of smaller
in part (iii) will depend on the type of recurrence one has.
We will 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 guring out a recurrence that links them together; but often (as happened
in the case of weighted interval scheduling), it can be useful to rst dene 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 useful until one nds a recurrence linking them
together; but it can be difcult 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.
6.3 Segmented Least Squares: Multi-way Choices
6.3 Segmented Least Squares: Multi-way Choices
We now discuss a different type of problem, which illustrates a slightly
more complicated style of dynamic programming. In the previous section,
we developed a recurrence based on a fundamentally binary choice: either
the interval n belonged to an optimal solution or it didnt. In the problem
we consider here, the recurrence will involve what might be called multiway choices: at each step, we have a polynomial number of possibilities to
consider for the structure 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 denition can formalize a notion that
initially seems too fuzzy and nonintuitive to work with mathematically.
The Problem
Often when looking at scientic or statistical data, plotted on a twodimensional set of axes, one tries to pass a line of best t through the
data, as in Figure 6.6.
This is a foundational problem in statistics and numerical analysis, formulated as follows. Suppose our data consists of a set P of n points in the plane,
denoted (x1, y1) , (x2 , y2) , . . . , (xn , yn ); and suppose x1 < x2 < . . . < xn . Given
a line L dened by the equation y = ax + b, we say that the error of L with
respect to P is the sum of its squared distances to the points in P :
n
(yi axi b)2.
Error(L , P ) =
i =1
Figure 6.6 A line of best fit.
261
262
Chapter 6 Dynamic Programming
Figure 6.7 A set of points that lie approximately on two lines.
A natural goal is then to nd 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 minimum
error is y = ax + b, where
a=
n
i
n
x i yi
i
2
xi
i
xi
i
i yi
xi
2
and b =
i yi
a
n
i
xi
.
Now, heres a kind of issue that these formulas werent designed to cover.
Often 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 gure 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 t, 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 t the points with an arbitrarily large set of lines, we could t the points
perfectly by having a different line pass through each pair of consecutive points
in P .
At the other extreme, we could try hard-coding the number two into the
problem; we could seek the best t 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.
6.3 Segmented Least Squares: Multi-way Choices
Figure 6.8 A set of points that lie approximately on three lines.
Thus, intuitively, we need a problem formulation that requires us to t
the points well, using as few lines as possible. We now formulate a problem
the Segmented Least Squares Problemthat captures these issues quite cleanly.
The problem is a fundamental instance of an issue in data mining and statistics
known as change detection: Given a sequence of data points, we want to
identify a few points in the sequence at which a discrete change occurs (in
this case, a change from one linear approximation to another).
Formulating the Problem As in the discussion above, we are given a set of
points P = {(x1, y1) , (x2 , y2) , . . . , (xn , yn )}, with x1 < x2 < . . . < xn . We will use
pi to denote the point (xi , yi ). We must rst partition P into some number
of segments. Each segment is a subset of P that represents a contiguous set
of x-coordinates; that is, it is a subset of the form {pi , pi+1, . . . , pj1, pj } for
some indices i j . 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 dened to be a sum of the following terms.
(i) The number of segments into which we partition P , times a xed, given
multiplier C > 0.
(ii) For each segment, the error value of the optimal line through that
segment.
Our goal in the Segmented Least Squares Problem is to nd 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 denition, but we increase the term in part (i). (The multiplier C is provided
263
264
Chapter 6 Dynamic Programming
with the input, and by tuning C , we can penalize the use of additional lines
to a greater or lesser extent.)
There are exponentially many possible partitions of P , and initially it is not
clear that we should be able to nd the optimal one efciently. We now show
how to use dynamic programming to nd a partition of minimum penalty in
time polynomial in n.
Designing the Algorithm
To begin with, we should recall the ingredients we need for a dynamic programming algorithm, as outlined at the end of Section 6.2.We want a polynomial
number of subproblems, the solutions of which should yield a solution to the
original problem; and we should be able to build up solutions to these subproblems using a recurrence. As with the Weighted Interval Scheduling Problem,
it helps to think about some simple properties of the optimal solution. Note,
however, that there is not really a direct analogy to weighted interval scheduling: there we were looking for a subset of n objects, whereas here we are
seeking to partition n objects.
For segmented least squares, the following observation is very useful:
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
last segment pi , . . . , pn (see Figure 6.9), then we could remove those points
from consideration and recursively solve the problem on the remaining points
p1, . . . , pi1.
OPT(i 1)
i
n
Figure 6.9 A possible solution: a single line segment fits points pi , pi+1, . . . , pn , and then
an optimal solution is found for the remaining points p1, p2 , . . . , pi1.
6.3 Segmented Least Squares: Multi-way Choices
Suppose we let OPT(i) denote the optimum solution for the points
p1, . . . , pi , and we let ei , j denote the minimum error of any line with respect to pi , pi+1, . . . , pj . (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 , . . . , pn , then the value
of the optimal solution is OPT(n) = ei , n + C + OPT(i 1).
Using the same observation for the subproblem consisting of the points
p1, . . . , pj , we see that to get OPT(j ) we should nd the best way to produce a
nal segment pi , . . . , pj 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 justied the following recurrence.
(6.7)
For the subproblem on the points p1, . . . , pj ,
OPT(j ) =
min (ei , j + C + OPT(i 1)) ,
1ij
and the segment pi , . . . , pj 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 [0 . . . n]
Set M [0] = 0
For all pairs i j
Compute the least squares error ei , j for the segment pi , . . . , pj
Endfor
For j = 1, 2, . . . , n
Use the recurrence (6.7) to compute M [j ]
Endfor
Return M [n]
By analogy with the arguments for weighted interval scheduling, the
correctness of this algorithm can be proved directly by induction, with (6.7)
providing the induction step.
And as in our algorithm for weighted interval scheduling, we can trace
back through the array M to compute an optimum partition.
265
266
Chapter 6 Dynamic Programming
Find-Segments(j )
If j = 0 then
Output nothing
Else
Find an i that minimizes ei , j + C + M [i 1]
Output the segment {pi , . . . , pj } and the result of
Find-Segments(i 1)
Endif
Analyzing the Algorithm
Finally, 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(n2)
pairs (i , j ) for which this computation is needed; and for each pair (i , j ), 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 ei , j values is O(n3).
Following this, the algorithm has n iterations, for values j = 1, . . . , n. For
each value of j , we have to determine the minimum in the recurrence (6.7) to
ll in the array entry M [j ]; this takes time O(n) for each j , for a total of O(n2).
Thus the running time is O(n2) once all the ei , j values have been determined.1
6.4 Subset Sums and Knapsacks: Adding a Variable
Were seeing more and more that issues in scheduling provide a rich source of
practically motivated algorithmic problems. So far weve considered problems
in which requests are specied 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 the second type of problem,
with durations and deadlines, which is difcult 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 will turn out not
to be enough, and so we end up creating a richer collection of subproblems. As
1
In this analysis, the running time is dominated by the O(n3) needed to compute all ei , j values. But,
in fact, it is possible to compute all these values in O(n2) time, which brings the running time of the
full algorithm down to O(n2). The idea, whose details we will leave as an exercise for the reader, is to
rst compute ei , j for all pairs (i , j ) where j 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 ei , j value, we can use the ingredients of the calculation
for ei , j1 to determine ei , j in constant time.
6.4 Subset Sums and Knapsacks: Adding a Variable
we will 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 jobs, 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 request corresponds to a job that requires time wi 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 jobs 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 iS wi W and, subject
to this restriction, iS wi 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 vi and a weight wi .
The goal in this more general problem is to select a subset of maximum total
value, subject 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 lling a knapsack of capacity W as
full 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
wi and W .
Since this resembles other scheduling problems weve seen before, its
natural to ask whether a greedy algorithm can nd the optimal solution. It
appears that the answer is noat least, no efcient greedy rule is known that
always constructs an optimal solution. One natural greedy approach to try
would be to sort the items by decreasing weightor at least to do this for all
items of weight at most W and then start selecting items in this order as long
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 /2}.
The goal of this section is to show how to use dynamic programming to
solve this problem. Recall the main principles of dynamic programming: We
have to come up with a small number of subproblems so that each subproblem
can be solved easily from smaller subproblems, and the solution to the
original problem can be obtained easily once we know the solutions to all
267
268
Chapter 6 Dynamic Programming
the subproblems. The tricky issue here lies in guring out a good set of
subproblems.
Designing the Algorithm
A False Start One general strategy, which worked for us in the case of
Weighted Interval Scheduling, is to consider subproblems involving only the
rst 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
solution using a subset of the requests {1, . . . , i}. The key to our method for
the Weighted Interval Scheduling Problem was to concentrate on an optimal
solution O 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 rst part, which follows immediately from the
denition of OPT(i).
.
If n O, then OPT(n) = OPT(n 1).
Next we have to consider the case in which n O. 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 conicted 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,
it means that for the subset of requests S {1, . . . , 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 nd 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 rst n 1 items
and total allowed weight W wn . We are therefore going to use many more
subproblems: one for each initial set {1, . . . , i} of the items, and each possible
W
wn
Figure 6.10 After item n is included in the solution, a weight of wn is used up and there
is W wn available weight left.
6.4 Subset Sums and Knapsacks: Adding a Variable
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 = 0, 1, . . . , n and each integer 0 w W . We will use OPT(i , w) to
denote the value of the optimal solution using a subset of the items {1, . . . , i}
with maximum allowed weight w, that is,
OPT(i , w) = max
S
wj ,
j S
where the maximum is over subsets S {1, . . . , i} that satisfy jS wj w.
Using this new set of subproblems, we will be able to express the value
OPT(i , w ) as a simple expression in terms of values from smaller problems.
Moreover, OPT(n , W ) is the quantity were looking for in the end. As before,
let O denote an optimum solution for the original problem.
.
.
If n O, then OPT(n , W ) = OPT(n 1, W ), since we can simply ignore
item n.
If n O, 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
1, 2, . . . , n 1.
When the nth item is too big, that is, W < wn , then we must have OPT(n , W ) =
OPT(n 1, W ). Otherwise, we get the optimum solution allowing all n requests
by taking the better of these two options. Using the same line of argument for
the subproblem for items {1, . . . , i}, and maximum allowed weight w, gives
us the following recurrence.
(6.8)
If w < wi then OPT(i , w) = OPT(i 1, w). Otherwise
OPT(i , w ) = max(OPT(i 1, w) , wi
+ OPT(i 1, w wi )).
As before, we want to design an algorithm that builds up a table of all
values while computing each of them at most once.
OPT(i , w )
Subset-Sum(n , W )
Array M [0 . . . n , 0 . . . W ]
Initialize M [0, w] = 0 for each w = 0, 1, . . . , W
For i = 1, 2, . . . , n
For w = 0, . . . , W
Use the recurrence (6.8) to compute M [i , w]
Endfor
Endfor
Return M [n , W ]
269
270
Chapter 6 Dynamic Programming
n0
0
0
0
i0
i1 0
0
0
0
20
10
00
0
0
0
1
2
0
0
0
0
w wi
0
0
w
0
0
0
0
0
0
W
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 1, w) and OPT(i 1, w wi ), as indicated by the arrows.
Using (6.8) one can immediately prove by induction that the returned
value M [n , W ] is the optimum solution value for the requests 1, . . . , n and
available weight W .
Analyzing the Algorithm
Recall the tabular picture we considered in Figure 6.5, associated with
weighted interval scheduling, where we also showed the way in which the array M for that algorithm was iteratively lled in. For the algorithm weve
just designed, we can use a similar representation, but we need a twodimensional table, reecting 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, w 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 nd
that the optimal value OPT(3, 6) = 5 (which we get by using the third item and
one of the rst two items). Figure 6.12 illustrates the way the algorithm lls
in the two-dimensional table of OPT values row by row.
Next we will 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.
6.4 Subset Sums and Knapsacks: Adding a Variable
Knapsack size W = 6, items w1 = 2, w2 = 2, w3 = 3
3
3
2
2
1
00
0
0
0
0
0
0
0
1
2
3
4
5
10
00
0
0
2
0
2
0
2
0
2
0
2
0
0
1
2
3
4
5
6
6
Initial values
Filling in values for i = 1
3
30
20
10
0
2
2
4
4
4
00
0
0
2
0
2
0
2
0
2
0
2
0
0
1
2
3
4
5
20
10
6
Filling in values for i = 2
0
2
3
4
5
5
0
2
2
4
4
4
00
0
0
2
0
2
0
2
0
2
0
2
0
0
1
2
3
4
5
6
Filling in values for i = 3
Figure 6.12 The iterations of the algorithm on a sample instance of the Subset Sum
Problem.
(6.9) The Subset-Sum(n , W) Algorithm correctly computes the optimal
value of the problem, and runs in O(nW ) time.
Note that this method is not as efcient 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 dening the problem. We call such algorithms
pseudo-polynomial. Pseudo-polynomial algorithms can be reasonably efcient
when the numbers {wi } involved in the input are reasonably small; however,
they become less practical as these numbers grow large.
To recover an optimal set S of items, we can trace back through the array
M by a procedure similar to those we developed in the previous sections.
(6.10) Given a table M of the optimal values of the subproblems, the optimal
set S can be found in O(n) time.
Extension: The Knapsack Problem
The Knapsack Problem is a bit more complex than the scheduling problem we
discussed earlier. Consider a situation in which each item i has a nonnegative
weight wi as before, and also a distinct value vi . Our goal is now to nd a
271
272
Chapter 6 Dynamic Programming
subset S of maximum value iS vi , subject to the restriction that the total
weight of the set should not exceed W : iS 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 , w), 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 O, and identify
two cases depending on whether or not n O.
.
If n O, then OPT(n , W ) = OPT(n 1, W ).
.
If n O, then OPT(n , W ) = vn + OPT(n 1, W wn ).
Using this line of argument for the subproblems implies the following analogue
of (6.8).
(6.11)
If w < wi then OPT(i , w) = OPT(i 1, w). Otherwise
OPT(i , w) = max(OPT(i 1, w) , vi
+ OPT(i 1, w wi )).
Using this recurrence, we can write down a completely analogous dynamic
programming algorithm, and this implies the following fact.
(6.12)
The Knapsack Problem can be solved in O(nW ) time.
6.5 RNA Secondary Structure: Dynamic
Programming over Intervals
In the Knapsack Problem, we were able to formulate a dynamic programming
algorithm by adding a new variable. A different but very common way by
which one ends up adding a variable to a dynamic program is through
the following scenario. We start by thinking about the set of subproblems
on {1, 2, . . . , j }, for all choices of j , and nd 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 nd 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 t this prole; 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.
6.5 RNA Secondary Structure: Dynamic Programming over Intervals
G
A
C
A
U
A
U
A
A
G
C
G
G
C
G
C
U
C
G
C
A
G
A
U
U
G
C
A
U
U
G
C
G
A
C
G
U
C
G
A
A
A
A
G
C
A
Figure 6.13 An RNA secondary structure. Thick lines connect adjacent elements of the
sequence; thin lines indicate pairs of elements that are matched.
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) formed 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.
2
Adenine, cytosine, guanine, and thymine, the four basic units of DNA.
273
274
Chapter 6 Dynamic Programming
For our purposes, a single-stranded RNA molecule can be viewed as a
sequence of n symbols (bases) drawn from the alphabet {A, C , G , U }.3 Let B =
b1b2 . . . bn be a single-stranded RNA molecule, where each bi {A, C , G , U }.
To a rst 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 basein other words, the set
of base pairs forms a matching . It also turns out that secondary structures are
(again, to a rst approximation) knot-free, 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
S = {(i , j )}, where i , j {1, 2, . . . , n}, that satises the following conditions.
(i) (No sharp turns.) 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 , ) are two pairs in S, then
we cannot have i < k < j < . (See Figure 6.14 for an illustration.)
Note that the RNA secondary structure in Figure 6.13 satises 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
on real RNA secondary structures.
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 will
form the secondary structure with the optimum total free energy. The correct
model for the free energy of a secondary structure is a subject of much debate;
but a rst approximation here is to assume that the free 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 efcient 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.
6.5 RNA Secondary Structure: Dynamic Programming over Intervals
G
G
U
C
A
C
G
A
U
U
A
G
C
U
A
AC
AU
GAU
(a)
G
G
C
C
AU
G
U
(b)
Figure 6.14 Two views of an RNA secondary structure. In the second view, (b), the
string has been stretched lengthwise, and edges connecting matched pairs appear as
noncrossing bubbles over the string.
a single-stranded RNA molecule B = b1b2 . . . bn and determines a secondary
structure S with the maximum possible number of base pairs.
Designing and Analyzing the Algorithm
A First Attempt at Dynamic Programming The natural rst 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 b1b2 . . . 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
in terms of the solutions to smaller subproblems. We can get partway
there: in the optimal secondary structure on b1b2 . . . bj , its the case that either
OPT(j )
.
j is not involved in a pair; or
.
j pairs with t for some t < j 4.
In the rst 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 + 1 and j 1. Weve therefore effectively isolated two
new subproblems: one on the bases b1b2 . . . bt 1, and the other on the bases
bt +1 . . . bj1. The rst is solved by OPT(t 1), but the second is not on our list
of subproblems, because it does not begin with b1.
275
276
Chapter 6 Dynamic Programming
Including the pair (t, j ) results in
two independent subproblems.
1
2
t1
t t+1
j1
j
j1
j
(a)
i
t1
t t+1
(b)
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 b1; in other
words, we need to consider subproblems on bi bi+1 . . . bj for all choices of i j .
Dynamic Programming over Intervals Once we make this decision, our
previous reasoning leads straight to a successful recurrence. Let OPT(i , j ) denote
the maximum number of base pairs in a secondary structure on bi bi+1 . . . bj .
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 bi bi+1 . . . bj , we have the same
alternatives as before:
.
j is not involved in a pair; or
.
j pairs with t for some t < j 4.
In the rst 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 justied the following recurrence.
(6.13) OPT(i , j ) = max(OPT(i , j 1) , max(1 + OPT(i , t 1) + OPT(t + 1, j 1))),
where the max is taken over t such that bt and bj are an allowable base pair
(under conditions (i) and (ii) from the denition 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
6.5 RNA Secondary Structure: Dynamic Programming over Intervals
RNA sequence ACCGGUAGU
40
0
30
0
0
40
0
0
1
20
20
0
30
0
0
j=6
7
8
9
Initial values
j=6
8
9
40
0
0
0
40
0
0
0
1
1
30
0
1
0
1
1
20
0
1
1
1
i=1 1
1
1
7
8
j=6
7
8
j=6
7
8
9
Filling in the values
for k = 6
2
j=6
1
1
i=1 1
1
1
20
1
0
0
30
0
1
i=1 1
7
0
0
20
Filling in the values
for k = 5
0
30
i=1 1
i=1
40
9
9
Filling in the values
for k = 7
Filling in the values
for k = 8
Figure 6.16 The iterations of the algorithm on a sample instance of the RNA Secondary
Structure Prediction Problem.
for which k = j 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
For k = 5, 6, . . . , n 1
For i = 1, 2, . . . n k
Set j = i + k
Compute OPT(i , j ) using the recurrence in (6.13)
Endfor
Endfor
Return OPT(1, n)
As an example of this algorithm executing, we consider the input
ACCGGUAGU , a subsequence of the sequence in Figure 6.14. As with the
Knapsack Problem, we need two dimensions to depict the array M : one for
the left endpoint of the interval being considered, and one for the right endpoint. In the gure, we only show entries corresponding to [i , j ] 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(n3).
277
278
Chapter 6 Dynamic Programming
As always, we can recover the secondary structure itself (not just its value)
by recording how the minima in (6.13) are achieved and tracing back through
the computation.
6.6 Sequence Alignment
For the remainder of this chapter, we consider two further dynamic programming algorithms that each have a wide range of applications. In the next two
sections we discuss sequence 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 denition and type in a
word it doesnt containsay, 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 rst 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 dene 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 rst 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:
o-currance
occurrence
The hyphen (-) indicates a gap where we had to add a letter to the second
word to get it to line up with the rst. 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
6.6 Sequence Alignment
o-curr-ance
occurre-nce
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 look like. When the two strings dont look like
English wordsfor example, abbbaabbbbaab and ababaaabbbbbabit may
take a little work to decide whether they can be lined up nicely or not:
abbbaa--bbbbaab
ababaaabbbbba-b
Dictionary interfaces and spell-checkers are not the most computationally
intensive application for this type of problem. In fact, determining similarities
among strings is one of the central computational problems facing molecular
biologists today.
Strings arise very naturally in biology: an organisms genome its full set
of genetic materialis 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 rst 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 evolutionarily. 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 eld of computational biology.
All this leaves us with the same question we asked initially, while typing
badly spelled words into our online dictionary: How should we dene the
notion of similarity between two strings?
In the early 1970s, the two molecular biologists Needleman and Wunsch
proposed a denition of similarity, which, basically unchanged, has become
279
280
Chapter 6 Dynamic Programming
the standard denition in use today. Its position as a standard was reinforced by
its simplicity and intuitive appeal, as well as through its independent discovery
by several other researchers around the same time. Moreover, this denition of
similarity came with an efcient 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 rst articulated it.
The denition 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 x1x2 . . . xm and Y
consists of the sequence of symbols y1y2 . . . yn . Consider the sets {1, 2, . . . , m}
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 , j ) , (i , j ) M and i < i , then j < j . Intuitively, an alignment gives a way
of lining up the two strings, by telling us which pairs of positions will be lined
up with one another. Thus, for example,
stop-tops
corresponds to the alignment {(2, 1) , (3, 2) , (4, 3)}.
Our denition of similarity will be based on nding the optimal alignment
between X and Y , according to the following criteria. Suppose M is a given
alignment between X and Y .
.
First, there is a parameter > 0 that denes a gap penalty. For each
position of X or Y that is not matched in M it is a gapwe incur a
cost of .
.
Second, for each pair of letters p , q in our alphabet, there is a mismatch
cost of pq for lining up p with q. Thus, for each (i , j ) M , we pay the
appropriate mismatch cost xi yj for lining up xi with yj . One generally
assumes that pp = 0 for each letter pthere is no mismatch cost to line
up a letter with another copy of itselfalthough this will not be necessary
in anything that follows.
.
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 alignment
in the biology literature. The quantities and {pq } are external parameters
that must be plugged into software for sequence alignment; indeed, a lot of
work goes into choosing the settings for these parameters. From our point of
6.6 Sequence Alignment
view, in designing an algorithm for sequence alignment, we will take them as
given. To go back to our rst example, notice how these parameters determine
which alignment of ocurrance and occurrence we should prefer: the rst is
strictly better if and only if + ae < 3 .
Designing the Algorithm
We now have a concrete numerical denition for the similarity between
strings X and Y : it is the minimum 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 programming, and we are motivated by the following basic dichotomy.
.
In the optimal alignment M , either (m , n) M or (m , n) M . (That is,
either the last symbols in the two strings are matched to each other, or
they arent.)
By itself, this fact would be too weak to provide us with a dynamic programming solution. Suppose, however, that we compound it with the following
basic fact.
(6.14) Let M be any alignment of X and Y. If (m , n) M, then either the
mth position of X or the nth position of Y is not matched in M.
Proof. Suppose by way of contradiction that (m , n) M , and there are numbers i < m and j < n so that (m , j ) M and (i , n) M . But this contradicts our
denition of alignment : we have (i , n) , (m , j ) M with i < m, but n > i so the
pairs (i , n) and (m , j ) cross.
There is an equivalent way to write (6.14) that exposes three alternative
possibilities, and leads directly to the formulation of a recurrence.
(6.15)
In an optimal alignment M, at least one of the following is true:
(i) (m , n) M; or
(ii) the mth position of X is not matched; or
(iii) the nth position of Y is not matched.
Now, let OPT(i , j ) denote the minimum cost of an alignment between
x1x2 . . . xi and y1y2 . . . yj . If case (i) of (6.15) holds, we pay xm yn and then
align x1x2 . . . xm1 as well as possible with y1y2 . . . yn1; we get OPT(m , n) =
xm yn + OPT(m 1, n 1). If case (ii) holds, we pay a gap cost of since the
mth position of X is not matched, and then we align x1x2 . . . xm1 as well as
281
282
Chapter 6 Dynamic Programming
possible with y1y2 . . . yn . In this way, we get OPT(m , n) = + OPT(m 1, n).
Similarly, if case (iii) holds, we get OPT(m , n) = + OPT(m , n 1).
Using the same argument for the subproblem of nding the minimum-cost
alignment between x1x2 . . . xi and y1y2 . . . yj , we get the following fact.
(6.16) The minimum alignment costs satisfy the following recurrence for i 1
and j 1:
OPT(i , j ) = min[xi yj
+ OPT(i 1, j 1) , + OPT(i 1, j ) , + OPT(i , j 1)].
Moreover, (i , j ) is in an optimal alignment M for this subproblem if and only
if the minimum is achieved by the rst of these values.
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 seeking.
We now specify the algorithm to compute the value of the optimal alignment. For purposes of initialization, we note that OPT(i , 0) = OPT(0, i) = i 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] = i for each i
Initialize A[0, j ] = j for each j
For j = 1, . . . , n
For i = 1, . . . , m
Use the recurrence (6.16) to compute A[i , j ]
Endfor
Endfor
Return A[m , n]
As in previous dynamic programming algorithms, we can trace back
through the array A, using the second part of fact (6.16), to construct the
alignment itself.
Analyzing the Algorithm
The correctness of the algorithm follows directly from (6.16). The running time
is O(mn), since the array A has O(mn) entries, and at worst we spend constant
time on each.
6.6 Sequence Alignment
x3
x2
x1
y1
y2
y3
y4
Figure 6.17 A graph-based picture of sequence alignment.
There is an appealing pictorial way in which people think about this
sequence alignment algorithm. Suppose we build a two-dimensional m n
grid graph GXY , with the rows labeled by symbols in the string X , the columns
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 j th column by the label (i , j ). We put costs on
the edges of GXY : the cost of each horizontal and vertical edge is , and the
cost of the diagonal edge from (i 1, j 1) to (i , j ) is xi yj .
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 , j ) denote the minimum cost of a path from (0, 0) to (i , j ) in
GXY . Then for all i , j, we have f (i , j ) = OPT(i , j ).
Proof. We can easily prove this by induction on i + j . When i + j = 0, we have
i = j = 0, and indeed f (i , j ) = 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 , j ) is either from (i 1, j 1), (i 1, j ), or (i , j 1). Thus we have
f (i , j ) = min[xi yj + f (i 1, j 1) , + f (i 1, j ) , + f (i , j 1)]
= min[xi yj + OPT(i 1, j 1) , + OPT(i 1, j ) , + OPT(i , j 1)]
= OPT(i , j ) ,
where we pass from the rst line to the second using the induction hypothesis,
and we pass from the second to the third using (6.16).
283
284
Chapter 6 Dynamic Programming
n
8
6
5
4
6
a
6
5
3
5
5
e
4
3
2
4
4
m
2
1
3
4
6
0
2
4
6
8
n
a
m
e
Figure 6.18 The OPT values
for the problem of aligning
the words mean to name .
Thus the value of the optimal alignment is the length of the shortest path
in GXY from (0, 0) to (m , n). (Well call any path in GXY from (0, 0) to (m , n)
a corner-to-corner 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 GXY 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 mean and name . For
the purpose of this example, we assume that = 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 cell in
the table (representing the corresponding node), the arrow indicates the last
step of the shortest path leading to that nodein other words, the way that
the minimum is achieved in (6.16). Thus, by following arrows backward from
node (4, 4), we can trace back to construct the alignment.
6.7 Sequence Alignment in Linear Space via
Divide and Conquer
In the previous section, we showed how to compute the optimal alignment
between two strings X and Y of lengths m and n, respectively. Building up the
two-dimensional m-by-n array of optimal solutions to subproblems, OPT( , ),
turned out to be equivalent to constructing a graph GXY with mn 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(mn), because it takes constant time to determine the value
in each of the mn cells of the array OPT; and the space requirement is O(mn)
as well, since it was dominated by the cost of storing the array (or the graph
GXY ).
The Problem
The question we ask in this section is: Should we be happy with O(mn)
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 (mn) space requirement can potentially be a more
severe problem than the (mn) 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 10 billion primitive
6.7 Sequence Alignment in Linear Space via Divide and Conquer
operations might be less cause for worry than the prospect of working with a
single 10-gigabyte array.
Fortunately, this is not the end of the story. In this section we describe a
very clever enhancement of the sequence alignment algorithm that makes it
work in O(mn) time using only O(m + 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 GXY , 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 GXY .
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 the computation can be
reused from one call to the next. The way in which this idea is used, however,
is fairly subtle.
Designing the Algorithm
We rst 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 ll 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 m 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-Efficient-Alignment(X ,Y )
Array B[0 . . . m , 0 . . . 1]
Initialize B[i , 0] = i for each i (just as in column 0 of A)
For j = 1, . . . , n
B[0, 1] = j (since this corresponds to entry A[0, j ])
For i = 1, . . . , m
B[i , 1] = min[xi yj + B[i 1, 0],
+ B[i 1, 1], + B[i , 0]]
Endfor
Move column 1 of B to column 0 to make room for next iteration:
Update B[i , 0] = B[i , 1] for each i
Endfor
285
286
Chapter 6 Dynamic Programming
It is easy to verify that when this algorithm completes, the array entry
B[i , 1] holds the value of OPT(i , n) for i = 0, 1, . . . , m. Moreover, it uses O(mn)
time and O(m) space. The problem is: where is the alignment itself? 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 just these two
columns. We could imagine getting around this difculty by trying to predict
what the alignment is going to be in the process of running our space-efcient
procedure. In particular, as we compute the values in the j th 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 problemwe will be able to recover
the alignment itself using O(m + n) spacebut 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 GXY .
(As we showed in the initial sequence alignment algorithm, f (i , j ) has the
same value as OPT(i , j ).) Now lets dene g (i , j ) to be the length of the shortest
path from (i , j ) to (m , n) in GXY . 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 < m and j < n we have
g (i , j ) = min[xi+1yj+1 + g (i + 1, j + 1) , + g (i , j + 1) , + g (i + 1, j )].
This is just the recurrence one obtains by taking the graph GXY , 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-efcient version of this backward dynamic programming algorithm, analogous to Space-Efficient-Alignment, which computes the
value of the optimal alignment using only O(m + n) space. We will refer to
6.7 Sequence Alignment in Linear Space via Divide and Conquer
this backward version, naturally enough, as Backward-Space-EfficientAlignment.
Combining the Forward and Backward Formulations So now we have
symmetric algorithms which build up the values of the functions f and g .
The idea will be to use these two algorithms in concert to nd the optimal
alignment. First, here are two basic facts summarizing some relationships
between the functions f and g .
(6.19) The length of the shortest corner-to-corner path in GXY that passes
through (i , j ) is f (i , j ) + g (i , j ).
Proof. Let ij denote the length of the shortest corner-to-corner path in GXY
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 f (i , j ) + g (i , j ), and so
we have ij 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 , j ) + g (i , j ),
and so we have ij f (i , j ) + g (i , j ). It follows that ij = f (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 f (q , k) + g (q , k). Then there is a corner-to-corner 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 x a value of k {0, . . . , n}. The shortest corner-to-corner path must use
some node in the kth column of GXY lets suppose it is node (p , k)and thus
by (6.19)
= f (p , k) + g (p , k) min f (q , k) + g (q , k).
q
Now consider the index q that achieves the minimum in the right-hand side
of this expression; we have
f (q , k) + g (q , k).
By (6.19) again, the shortest corner-to-corner path using the node (q , k) has
length f (q , k) + g (q , k), and since is the minimum length of any corner-tocorner path, we have
f (q , k) + g (q , k).
It follows that = f (q , k) + g (q , k). Thus the shortest corner-to-corner path
using the node (q , k) has length , and this proves (6.20).
287
288
Chapter 6 Dynamic Programming
Using (6.20) and our space-efcient 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 (i , n/2) and g (i , n/2) for each value of i,
using our two space-efcient algorithms. We can then determine the minimum
value of f (i , n/2) + g (i , n/2), and conclude via (6.20) that there is a shortest
corner-to-corner path passing through the node (i , 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(mn).
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 m + 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 i j m, denotes the substring of X consisting of xi xi+1 . . . xj ;
and we dene 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 n 2 then
Compute optimal alignment using Alignment(X ,Y )
Call Space-Efficient-Alignment(X ,Y [1 : n/2])
Call Backward-Space-Efficient-Alignment(X ,Y [n/2 + 1 : n])
Let q be the index minimizing f (q , n/2) + g (q , n/2)
Add (q , n/2) to global list P
Divide-and-Conquer-Alignment(X [1 : q],Y [1 : n/2])
Divide-and-Conquer-Alignment(X [q + 1 : n],Y [n/2 + 1 : n])
Return P
As an example of the rst level of recursion, consider Figure 6.19. If the
minimizing index q turns out to be 1, we get the two subproblems pictured.
Analyzing the Algorithm
The previous arguments already establish that the algorithm returns the correct
answer and that it uses O(m + n) space. Thus, we need only verify the
following fact.
6.7 Sequence Alignment in Linear Space via Divide and Conquer
Second recursive call
x3
x2
x1
y1
y2
y3
y4
First recursive call
Figure 6.19 The first level of recurrence for the space-efficient Divide-and-ConquerAlignment. The two boxed regions indicate the input to the two recursive cells.
(6.21) The running time of Divide-and-Conquer-Alignment on strings of
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 (m , n) cmn + 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 ll out the details of this guess. Specically,
suppose that we were in a case in which m = 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
Chapter 6 Dynamic Programming
This is a useful expression, since its something that we solved in our earlier
discussion of recurrences at the outset of Chapter 5. Specically, 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. Specically, 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 ) km n
holds for pairs (m , n ) with a smaller product, we have
T (m , n) cmn + T (q , n/2) + T (m q , n/2)
cmn + kqn/2 + k(m q)n/2
= cmn + kqn/2 + kmn/2 kqn/2
= (c + k/2)mn .
Thus the inductive step will work if we choose k = 2c , and this completes the
proof.
6.8 Shortest Paths in a Graph
For the nal three sections, we focus on the problem of nding 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 cij represents
a cost for going directly from node i to node j in the graph.
Earlier we discussed Dijkstras Algorithm for nding 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 nancial setting, and cij 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 prots. Second, the algorithm
that we develop for dealing with edges of negative cost turns out, in certain
crucial ways, to be more exible and decentralized than Dijkstras Algorithm.
As a consequence, it has important applications for the design of distributed
6.8 Shortest Paths in a Graph
routing algorithms that determine the most efcient path in a communication
network.
In this section and the next two, we will consider the following two related
problems.
.
Given a graph G with weights, as described above, decide if G has a
negative cyclethat is, a directed cycle C such that
cij < 0.
ij C
.
If the graph has no negative cycles, nd a path P from an origin node s
to a destination node t with minimum total cost:
cij
ij P
should be as small as possible for any s-t path. This is generally called
both the Minimum-Cost Path Problem and the Shortest-Path Problem.
In terms of our nancial motivation above, a negative cycle corresponds to a
protable 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, nally arriving back
at i1 with a net prot. 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 rst 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 .
Designing and Analyzing the Algorithm
A Few False Starts Lets begin by recalling Dijkstras Algorithm for the
Shortest-Path Problem when there are no negative costs. That method
s
2
Ps
1
1
2
C
1
1
2
t
Pt
1
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
Chapter 6 Dynamic Programming
2
u
3
s
w
1
6
v
(a)
2
2
s
t
3
3
3
(b)
Figure 6.21 (a) With negative
edge costs, Dijkstras Algorithm can give the wrong
answer for the Shortest-Path
Problem. (b) Adding 3 to the
cost of each edge will 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 edgesand we add elements greedily to this set S.
As our rst greedy step, we consider the minimum-cost edge leaving node s,
that is, miniV 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 v
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 , v).
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 cheaper than a path that starts on a cheap edge. This suggests that the
Dijkstra-style greedy approach will not work here.
Another natural idea is to rst modify the costs cij by adding some large
constant M to each; that is, we let cij = cij + M for each edge (i , j ) E . If the
constant M is large enough, then all modied costs are nonnegative, and we
can use Dijkstras Algorithm to nd the minimum-cost path subject to costs
c . However, this approach fails to nd 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 .
A Dynamic Programming Approach We will try to use dynamic programming to solve the problem of nding 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 nd a shortest path using only
the rst 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
efcient 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 1950s; and the Bellman-Ford Shortest-Path Algorithm was
one of the rst applications.
The dynamic programming solution we develop will be based on the
following crucial observation.
6.8 Shortest Paths in a Graph
v
w
t
P
Figure 6.22 The minimum-cost path P from v to t using at most i edges.
(6.22) If G has no negative cycles, then there is a shortest path from s to t
that is simple (i.e., does not repeat nodes), and hence has at most n 1 edges.
Proof. Since every cycle has nonnegative cost, the shortest path P from s to
t with the fewest number of edges does not repeat any vertex v. For if P did
repeat a vertex v, we could remove the portion of P between consecutive visits
to v, resulting in a path of no greater cost and fewer edges.
Lets use OPT(i , v) 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 x an optimal path P representing OPT(i , v) as depicted in Figure 6.22.
.
.
If the path P uses at most i 1 edges, then OPT(i , v) = OPT(i 1, v).
If the path P uses i edges, and the rst 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 1, w) + cvw )).
wV
Using this recurrence, we get the following dynamic programming algorithm to compute the value OPT(n 1, s).
293
294
Chapter 6 Dynamic Programming
a
3
6
4
d
1
b
4
t
2
8
2
e
3
c
3
Shortest-Path(G , s , t )
n = number of nodes in G
Array M [0 . . . n 1, V ]
Define M [0, t ] = 0 and M [0, v] = for all other v V
For i = 1, . . . , n 1
For v V in any order
Compute M [i , v] using the recurrence (6.23)
Endfor
Endfor
Return M [n 1, s]
(a)
0
t
1
2
3
4
5
0
0
0
0
0
0
a 3 3 4 6 6
b 0 2 2 2
c3
3
3
3
3
d4
e2
3
0
3
0
2
0
0
0
(b)
Figure 6.23 For the directed
graph in (a), the ShortestPath Algorithm constructs
the dynamic programming
table in (b).
The correctness of the method follows directly by induction from (6.23).
We can bound the running time as follows. The table M has n2 entries; and
each entry can take O(n) time to compute, as there are at most n nodes w V
we have to consider.
(6.24) The Shortest-Path method correctly computes the minimum cost of
an s-t path in any graph that has no negative cycles, and runs in 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
nd 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 nally to d-a -b-e -c -t .
Extensions: Some Basic Improvements to the Algorithm
An Improved Running-Time Analysis We can actually provide a better
running-time analysis for the case in which the graph G does not have too
many edges. A directed graph with n nodes can have close to n2 edges, since
there could potentially be an edge between each pair of nodes, but many
graphs are much sparser than this. When we work with a graph for which
the number of edges m is signicantly less than n2, 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.
6.8 Shortest Paths in a Graph
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 signicantly changing the
algorithm itself.
(6.25)
The Shortest-Path 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] + cvw )).
wV
We assumed it could take up to O(n) time to compute this minimum, since
there are n possible nodes w. But, of course, we need only compute this
minimum over all nodes w for which v has an edge to w; let us use nv to denote
this number. Then it takes time O(nv ) 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
nv .
On
vV
In Chapter 3, we performed exactly this kind of analysis for other graph
algorithms, and used (3.9) from that chapter to bound the expression vV nv
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 vV nv 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 vV nv = m. Plugging this into our expression
nv
On
vV
for the running time, we get a running-time bound of O(mn).
Improving the Memory Requirements We can also signicantly 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
Chapter 6 Dynamic Programming
iterations i = 1, 2, . . . , n 1, but the role of i will now simply be as a counter;
in each iteration, and for each node v, we perform the update
M [v] = min(M [v], min(cvw + M [w])).
wV
We now observe the following fact.
(6.26) Throughout the algorithm M [v] is the length of some path from v to
t, and after i rounds of updates the value M [v] is no larger than the length of
the shortest path from v to t using at most i edges.
Given (6.26), we can then use (6.22) as before to show that we are done after
n 1 iterations. Since we are only storing an M array that indexes over the
nodes, this requires only O(n) working memory.
Finding the Shortest Paths One issue to be concerned about is whether this
space-efcient 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 method
to recover the solution from a similar space-efcient 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 rst node (after itself) on its path to the
destination t ; we will denote this rst node by rst [v]. To maintain rst [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 min(cvw + M [w]), we set
rst [v] to the node w that attains this minimum.
wV
Now let P denote the directed pointer graph whose nodes are V , and
whose edges are {(v , rst [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 rst [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 rst [v] equal to w; and since M [w] may decrease, this equation may
turn into an inequality.
Let v1, v2 , . . . , vk be the nodes along the cycle C in the pointer graph,
and assume that (vk , v1) is the last edge to have been added. Now, consider
the values right before this last update. At this time we have M [vi ] cvi vi+1 +
M [vi+1] for all i = 1, . . . , k 1, and we also have M [vk ] > cvk v1 + M [v1] since
we are about to update M [vk ] and change rst [vk ] to v1. Adding all these
1
inequalities, the M [vi ] values cancel, and we get 0 > k=1 cvi vi+1 + cvk v1: a
i
negative cycle, as claimed.
6.9 Shortest Paths and Distance Vector Protocols
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 rst [v] = v1, to rst [v1] = 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 w = rst [v]. Since the algorithm terminated,
we must have M [v] = cvw + M [w]. The value M [t ] = 0, and hence the length
of the path traced out by the pointer graph is exactly M [v], which we know is
the shortest-path distance.
Note that in the more space-efcient version of Bellman-Ford, the path
whose length is M [v] after i iterations can have substantially more edges than
i. For example, if the graph is a single path from s to t , and we perform updates
in the reverse of the order the edges appear on the path, then we get the nal
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 opportunistically to speed up the algorithm on instances
where it does happen. In order to do this, we need a stopping signal in the
algorithmsomething that tells us its safe to terminate before iteration n 1
is reached.
Such a stopping signal is a simple consequence of the following observation: If we ever execute a complete iteration i in which no M [v] value changes,
then no M [v] value will ever change again, since future iterations will begin
with exactly the same set of array entries. Thus it is safe to stop the algorithm.
Note that it is not enough for a particular M [v] value to remain the same; in
order to safely terminate, we need for all these values to remain the same for
a single iteration.
6.9 Shortest Paths and Distance Vector Protocols
One important application of the Shortest-Path Problem is for routers in a
communication network to determine the most efcient 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 w if the two routers are connected
by a direct communication link. We dene a cost cvw representing the delay on
the link (v , w); the Shortest-Path Problem with these costs is to determine the
path with minimum delay from a source node s to a destination t . Delays are
297
298
Chapter 6 Dynamic Programming
naturally nonnegative, so one could use Dijkstras Algorithm to compute the
shortest path. However, Dijkstras shortest-path computation requires global
knowledge of the network: it needs to maintain a set S of nodes for which
shortest paths have been determined, and make a global decision about which
node to add next to S. While routers 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 exible 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, v needs only obtain the value M [w]
from each neighbor w, and compute
min(cvw + M [w])
wV
based on the information obtained.
We now discuss an improvement to the Bellman-Ford Algorithm that
makes it better suited for routers 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 v to get the value again;
however, v has no way of knowing this fact, and so it must execute the pull
anyway.
This wastefulness suggests a symmetric push-based implementation,
where values are only transmitted when they change. Specically, 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.
Push-Based-Shortest-Path(G , s , t )
n = number of nodes in G
Array M [V ]
Initialize M [t ] = 0 and M [v] = for all other v V
For i = 1, . . . , n 1
For w V in any order
If M [w] has been updated in the previous iteration then
6.9 Shortest Paths and Distance Vector Protocols
For all edges (v , w) in any order
M [v] = min(M [v], cvw + M [w])
If this changes the value of M [v], then rst [v] = w
Endfor
Endfor
If no value changed in this iteration, then end the algorithm
Endfor
Return M [s]
In this algorithm, nodes are sent updates of their neighbors distance
values in rounds, and each node sends out an update in each iteration in which
it has changed. However, if the nodes correspond to routers in a network, then
we do not expect everything to run in lockstep like this; some routers 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 routers 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 noties its neighbors of the new value. If we were
to watch the behavior of all routers 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 [v] = for all other v V
Declare t to be active and all other nodes inactive
While there exists an active node
Choose an active node w
For all edges (v , w) in any order
M [v] = min(M [v], cvw + M [w])
If this changes the value of M [v], then
rst [v] = w
v becomes active
Endfor
w becomes inactive
EndWhile
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
Chapter 6 Dynamic Programming
presumably interested in nding 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 vector protocol, since each node maintains a vector of distances to
every other node in the network.
Problems with the Distance Vector Protocol
One of the major problems with the distributed implementation of BellmanFord on routers (the protocol we have been discussing above) is that its derived
from an initial dynamic programming algorithm that assumes edge costs will
remain constant during the execution of the algorithm. Thus far weve been
designing algorithms with the tacit understanding that a program executing
the algorithm will be running on a single computer (or a centrally managed
set of computers), processing some specied input. In this context, its a rather
benign assumption to require that the input not change while the program 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 cvw effectively increases to .
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 (v , t ), each of cost 1.
Now suppose the edge (v , t ) in Figure 6.24 is deleted. How does 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
The deleted edge causes an unbounded
sequence of updates by s and v.
1
s
v
1
1
t
Deleted
Figure 6.24 When the edge (v , t ) is deleted, the distributed Bellman-Ford Algorithm
will begin counting to infinity.
6.10 Negative Cycles in a Graph
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 nds an
alternate route; in the case, as here, that the network is truly disconnected,
these updates will continue indenitelya behavior known as the problem of
counting to innity.
To avoid this problem and related difculties 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 rst 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 signicantly more storage to keep track of the full
paths. In the history of the Internet, there has been a shift from distance vector
protocols to path vector protocols; currently, the path vector approach is used
in the Border Gateway Protocol (BGP) in the Internet core.
* 6.10 Negative Cycles in a Graph
So far in our consideration of the Bellman-Ford Algorithm, we have assumed
that the underlying graph has negative edge costs but no negative cycles. We
now consider the more general case of a graph that may contain negative
cycles.
The Problem
There are two natural questions we will consider.
.
How do we decide if a graph contains a negative cycle?
.
How do we actually nd a negative cycle in a graph that contains one?
The algorithm developed for nding 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 nd negative
cycles that have a path reaching a sink t . Before we develop the details of this,
lets compare the problem of nding a negative cycle that can reach a given t
with the seemingly more natural problem of nding a negative cycle anywhere
in the graph, regardless of its position related to a sink. It turns out that if we
301
302
Chapter 6 Dynamic Programming
Any negative cycle in G will be able to reach t.
G
t
Figure 6.25 The augmented graph.
develop a solution to the rst problem, well be able to obtain a solution to
the second problem as well, in the following way. Suppose we start with a
graph G , add a new node t to it, and connect each other node v in the graph
to node t via an edge of cost 0, as shown in Figure 6.25. Let us call the new
augmented graph G .
(6.29) The augmented graph G has a negative cycle C such that there is a
path from C to the sink t if and only if the original graph has a negative cycle.
Proof. Assume G has a negative cycle. Then this cycle C clearly has an edge
to t in G , since all nodes have an edge to t .
Now suppose G has a negative cycle with a path to t . Since no edge leaves
t in G , this cycle cannot contain t . Since G is the same as G aside from the
node t , it follows that this cycle is also a negative cycle of G .
So it is really enough to solve the problem of deciding whether G has a
negative cycle that has a path to a given sink node t , and we do this now.
Designing and Analyzing the Algorithm
To get started thinking about the algorithm, we begin by adopting the original
version of the Bellman-Ford Algorithm, which was less efcient in its use
of space. We rst extend the denitions of OPT(i , v) from the Bellman-Ford
Algorithm, dening 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
6.10 Negative Cycles in a Graph
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) = .
i
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 1, v) for
all nodes v and all i n.
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) = OPT(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 cycle with a path to t if and only if OPT(n , v) =
OPT(n 1, v ) for all nodes v.
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. Specically, 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 OPT(n
1, v). It follows that we will have OPT(n + 1, v) = OPT(n 1, 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 all nodes v and all i n. Thus there
cannot be a negative cycle C that has a path to t ; for any node w on this cycle
C , (6.30) implies that the values OPT(i , w) would have to become arbitrarily
negative as i increased.
Statement (6.32) gives an O(mn) method to decide if G has a negative
cycle that can reach t . We compute values of OPT(i , v) for nodes of G and for
values of i up to n. By (6.32), there is no negative cycle if and only if there is
some value of i n at which OPT(i , v) = OPT(i 1, v) for 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
nd a negative cycle, we consider a node v such that OPT(n , v) = 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 nd 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 1
303
304
Chapter 6 Dynamic Programming
edges, so P must contain a cycle C . We claim that this cycle C has negative
cost.
(6.33) If G has n nodes and OPT(n , v) = OPT(n 1, v), then a path P from v
to t of cost OPT(n , v) contains a cycle C, and C has negative cost.
Proof. First observe that the path P must have n edges, as OPT(n , v) = OPT(n
1, v), and so every path using n 1 edges has cost greater than that of the
path P . In a graph with n nodes, a path consisting of n edges must repeat
a node somewhere; let w be a node that occurs on P more than once. Let C
be the cycle on P between two consecutive occurrences of node w. If C were
not a negative cycle, then deleting C from P would give us a v-t path with
fewer than n edges and no greater cost. This contradicts our assumption that
OPT(n , v ) = OPT(n 1, v ), and hence C must be a negative cycle.
(6.34) The algorithm above nds a negative cycle in G, if such a cycle exists,
and runs in O(mn) time.
Extensions: Improved Shortest Paths and Negative Cycle
Detection Algorithms
At the end of Section 6.8 we discussed a space-efcient implementation of the
Bellman-Ford algorithm for graphs with no negative cycles. Here we implement
the detection of negative cycles in a comparably space-efcient 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 rst edges (v , rst [v]) that we
used for the space-efcient 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 rst [v] = 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 will be an analogous early termination rule for graphs that
have negative cycles.
6.10 Negative Cycles in a Graph
Consider a new edge (v , w), with rst [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 w 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 rst [v] is being updated from u to w; 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 nd 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 , w) creates
a cycle is to consider all nodes in the subtree directed toward v. If w is in this
subtree, then (v , w) forms a cycle; otherwise it does not. (Again, consider the
two sample cases in Figure 6.26.) To be able to nd 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 nd
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 M [x] 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
w
t
t
w
v
u
v
u
Update to
first[v] = w
Update to
f irst[v] = w
(a)
(b)
Figure 6.26 Changing the pointer graph P when rst [v] is updated from u to w. In (b),
this creates a (negative) cycle, whereas in (a) it does not.
305
306
Chapter 6 Dynamic Programming
edge (x , rst [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 dened 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 = (v , 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 problemhowever, in this case, the path through edge (v , 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 rst 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 M [v]
is the distance on all nodes v where the shortest path from v to t uses at most
i edges. Note that nodes v where M [v] is the actual shortest-path distance
cannot be dormant, as the value M [v] will be updated in the next iteration for
all dormant nodes.
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 1 iterations that update values in the array M without nding
Solved Exercises
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 nds a negative cycle in G if
such a cycle exists. It terminates immediately if the pointer graph P of rst [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
x1, 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
{r1, r2 , r3, r4} = {5, 6, 5, 1}.
Then the optimal solution would be to place billboards at x1 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 x1, . . . , xn is really the same as the optimal solution
307
308
Chapter 6 Dynamic Programming
on sites x1, . . . , xn1; if we do, then we should eliminate xn and all other sites
that are within 5 miles of it, and nd an optimal solution on whats left. The
same reasoning applies when were looking at the problem dened by just the
rst j sites, x1, . . . , xj : we either include xj in the optimal solution or we dont,
with the same consequences.
Lets dene 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 x1, x2 , . . . , xe(j) are still
valid options once weve chosen to place a billboard at xj , but the sites
xe(j)+1, . . . , xj1 are not.
Now, our reasoning above justies the following recurrence. If we let OPT(j )
denote the revenue from the optimal subset of sites among x1, . . . , xj , then we
have
OPT(j ) = max(rj
+ OPT(e (j )) , OPT(j 1)).
We now have most of the ingredients we need for a dynamic programming
algorithm. First, we have a set of n subproblems, consisting of the rst j sites
for j = 0, 1, 2, . . . , n. Second, we have a recurrence that lets us build up the
solutions to subproblems, given by OPT(j ) = max(rj + OPT(e (j )) , OPT(j 1)).
To turn this into an algorithm, we just need to dene 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 .
Initialize M [0] = 0 and M [1] = r1
For j = 2, 3, . . . , n:
Compute M [j ] using the recurrence
Endfor
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 (j ) 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 all e (j )
values in O(n) time as follows. For each site location xi , we dene xi = xi 5.
We then merge the sorted list x1, . . . , xn with the sorted list x1, . . . , xn 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 xj , we know that anything from this point onward
to xj cannot be chosen together with xj (since its within 5 miles), and so we
Solved Exercises
simply dene e (j ) to be the largest value of i for which weve seen xi in our
scan.
Heres a nal observation on this problem. Clearly, the solution looks
very much like 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 dene an interval with endpoints
[xi 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 nonoverlapping intervals correspond precisely to the set of valid billboard
placements, and so dropping the set of intervals weve just dened (with their
weights) into an algorithm for Weighted Interval Scheduling will yield the
desired solution.
Solved Exercise 2
Through some friends of friends, you end up on a consulting visit to the
cutting-edge biotech rm Clones R Us (CRU). At rst youre not sure how
your algorithmic background will be of any help to them, but you soon nd
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 xed
alphabet S , then XY denotes the string obtained by concatenating them
writing X followed by Y . CRU has identied a target sequence A of genetic
material, consisting of m 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 Bi and Bj could be the same string in L , for different values of i and j .)
The problem is to nd a concatenation over {Bi } 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 and
a mismatch cost pq for each pair p , q S.)
Give a polynomial-time algorithm for this problem.
309
310
Chapter 6 Dynamic Programming
Solution This problem is vaguely reminiscent of Segmented Least Squares:
we have a long sequence of data (the string A) that we want to t 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 . . . B 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 rst
position in A that is matched with some symbol in B , and let A denote the
substring of A from position t to the end. (See Figure 6.27 for an illustration
of this with = 3.) Now, the point is that in this optimal alignment M , the
substring A is optimally aligned with B ; indeed, if there were a way to better
align A with B , we could substitute it for the portion of M that aligns A with
B 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 nal piece of A that is aligned with one of the strings in L , and for this
piece all were doing is nding the string in L that aligns with it as well as
possible. Having found this optimal alignment for A , we can break it off and
continue to nd the optimal solution for the remainder of A.
Thinking about the problem this way doesnt tell us exactly how to
proceedwe dont know how long A is supposed to be, or which string 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 nal subsequence of the input points, t 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 nd the one that
t
A
A3
B1
B2
B3
Figure 6.27 In the optimal concatentation of strings to align with A, there is a final
string (B3 in the figure) that aligns with a substring of A (A3 in the figure) that extends
from some position t to the end.
Solved Exercises
aligns best with A[x : y].) Let OPT(j ) denote the alignment cost of the optimal
solution on the string A[1 : j ].
The argument above says that an optimal solution on A[1 : j ] consists of
identifying a nal segment boundary t < j , nding the optimal alignment
of A[t : j ] with a single string in L , and iterating on A[1 : t 1]. 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 t together very nicely,
and it justies the following recurrence.
(6.37)
OPT(j ) = mint <j
c (t , j ) + OPT(t 1) for j 1, and OPT(0) = 0.
The full algorithm consists of rst 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] = 0
For all pairs 1 t j m
Compute the cost c (t , j ) as follows:
For each string B L
Compute the optimal alignment of B with A[t : j ]
Endfor
Choose the B that achieves the best alignment, and use
this alignment cost as c (t , j )
Endfor
For j = 1, 2, . . . , n
Use the recurrence (6.37) to compute M [j ]
Endfor
Return M [n]
As usual, we can get a concatentation that achieves it by tracing back over
the array of OPT values.
Lets consider the running time of this algorithm. First, there are O(m2)
values c (t , j ) that need to be computed. For each, we try each string of the
k strings B L , and compute the optimal alignment of B with A[t : j ] in
time O(n(j t )) = O(mn). Thus the total time to compute all c (t , j ) values
is O(km3n).
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
Chapter 6 Dynamic Programming
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 if its nodes can be written as v1, v2 , . . . , vn ,
with an edge between vi and vj if and only if the numbers i and j differ by
exactly 1. With each node vi , we associate a positive integer weight wi .
Consider, for example, the five-node path drawn in Figure 6.28. The
weights are the numbers drawn 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 vi of maximum weight
Add vi to S
Delete vi 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 vi where i is an odd number
Let S2 be the set of all vi where i is an even number
(Note that S1 and S2 are both independent sets)
Determine which of S1 or S2 has greater total weight,
and return this one
1
8
6
3
6
Figure 6.28 A paths with weights on the nodes. The maximum weight of an independent
set is 14.
Exercises
(c)
Give an algorithm that takes an n-node path G with weights and
returns an independent set of maximum total weight. The running
time should be polynomial in n, independent of the values of the
weights.
2. Suppose youre managing a consulting team of expert computer hackers,
and each week you have to choose a job for them to undertake. Now, as
you can well imagine, the set of possible jobs is divided 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 Cornell 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.
If you select a low-stress job for your team in week i, then you get a
revenue of i > 0 dollars; if you select a high-stress job, you get a revenue
of hi > 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
i 1.
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 i to the value if you choose low-stress in week i, and you add hi to
the value if you choose high-stress in week i. (You add 0 if you choose
none in week i.)
The problem. Given sets of values 1, 2 , . . . , n and h1, h2 , . . . , hn , find a
plan of maximum value. (Such a plan will be called optimal .)
Example. Suppose n = 4, and the values of i and hi 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 low-stress jobs in weeks
3 and 4. The value of this plan would be 0 + 50 + 10 + 10 = 70.
Week 1
Week 3
Week 4
10
h
Week 2
1
10
10
5
50
5
1
313
314
Chapter 6 Dynamic Programming
(a)
Show that the following algorithm does not correctly solve this
problem, by giving an instance on which it does not return the correct
answer.
For iterations i = 1 to n
If hi+1 > i + i+1 then
Output "Choose no job in week i"
Output "Choose a high-stress job in week i + 1"
Continue with iteration i + 2
Else
Output "Choose a low-stress job in week i"
Continue with iteration i + 1
Endif
End
hi =
To avoid problems with overflowing array bounds, we define
i = 0 when i > n.
In your example, say what the correct answer is and also what
the above algorithm finds.
(b) Give an efficient algorithm that takes values for 1, 2 , . . . ,
h1, h2 , . . . , hn and returns the value of an optimal plan.
n
and
3. Let G = (V , E ) be a directed graph with nodes v1, . . . , vn . We say that G is
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 , vj ) 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 form (vi , vj ).
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, nd the length of the longest path that begins at
v1 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 w = v1
Set L = 0
Exercises
v1
v2
v3
v4
v5
Figure 6.29 The correct answer for this ordered graph is 3: The longest path from v1 to
vn uses the three edges (v1, v2),(v2 , v4 ), and (v4 , v5).
While there is an edge out of the node w
Choose the edge (w , vj )
for which j is as small as possible
Set w = vj
Increase L by 1
end while
Return L as the length of the longest path
In your example, say what the correct answer is and also what the
algorithm above finds.
(b) Give an efficient algorithm that takes an ordered graph G and returns
the length of the longest path that begins at v1 and ends at vn . (Again,
the length of a path is the number of edges in the path.)
4. Suppose youre running a lightweight consulting businessjust you, two
associates, and some rented equipment. Your clients are distributed
between the East Coast and the West Coast, and this leads to the following
question.
Each month, you can either run your business from an office in New
York (NY) or from an office in San Francisco (SF). In month i, youll incur
an operating cost of Ni if 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, if 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 SFsuch 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
Chapter 6 Dynamic Programming
The problem. 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
NY
1
3
20
30
SF
50
20
2
4
Then the plan of minimum cost would be the sequence of locations
[NY , NY , 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 = 1 to n
If Ni < Si then
Output "NY in Month i"
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 S1, . . . , 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 likely boundaries between consecutive words in the
Exercises
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 are given a black box that,
for any string of letters x = x1x2 . . . xk , will return a number quality(x). This
number can be either positive or negative; larger numbers correspond to
more plausible English words. (So quality(me ) would be positive, while
quality(ght ) would be negative.)
Given a long string of letters y = y1y2 . . . 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 quality(meet ) + 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 problemfor 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.)
6. 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
Chapter 6 Dynamic Programming
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.
and turn it into text whose right margin is as even as possible, like this.
Call me Ishmael. Some years ago, never
mind how long precisely, having little
or no money in my purse, and nothing
particular to interest me on shore, I
thought I would sail about a little
and see the watery part of the world.
To make this precise enough for us to start thinking about how to
write a pretty-printer for text, we need to figure out what it means for the
right margins to be even. So suppose our text consists of a sequence of
words , W = {w1, w2 , . . . , 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
word except the last; and so if wj , wj+1, . . . , wk are assigned to one line,
then we should have
k1
(ci + 1) + ck L.
i=j
We will call an assignment of words to a line valid if it satisfies this
inequality. The difference between the left-hand side and the right-hand
side will be called the slack of the linethat 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 minimized.
7. As a solved exercise in Chapter 5, we gave an algorithm with O(n log n)
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
Exercises
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).
8. The residents of the underground city of Zion defend themselves through
a combination of kung fu, heavy artillery, and efficient algorithms. Recently they have become interested in automated methods that can help
fend off attacks by swarms of robots.
Heres what one of these robot attacks looks like.
.
.
.
.
A swarm of robots arrives over the course of n seconds; in the ith
second, xi robots arrive. Based on remote sensing data, you know
this sequence x1, x2 , . . . , xn 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 if j seconds have passed since the EMP
was last used, then it is capable of destroying up to f (j ) robots.
So specifically, if it is used in the kth second, and it has been j seconds
since it was previously used, then it will destroy min(xk , f (j )) robots.
(After this use, it will be completely drained.)
We will also assume that the EMP starts off completely drained, so
if it is used for the first time in the j th second, then it is capable of
destroying up to f (j ) robots.
The problem. Given the data on robot arrivals x1, 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.
i
1
2
3
4
xi
1
10
10
1
f (i )
1
2
4
8
The best solution would be to activate the EMP in the 3rd and the 4th
seconds. In the 3rd 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
320
Chapter 6 Dynamic Programming
(a)
Show that the following algorithm does not correctly solve this
problem, by giving an instance on which it does not return the correct
answer.
Schedule-EMP(x1, . . . , xn )
Let j be the smallest number for which f (j ) xn
(If no such j exists then set j = n)
Activate the EMP in the nth second
If n j 1 then
Continue recursively on the input x1, . . . , xnj
(i.e., invoke Schedule-EMP(x1, . . . , xnj ))
In your example, say what the correct answer is and also what the
algorithm above finds.
(b) Give an efficient algorithm that takes the data on robot arrivals
x1, x2 , . . . , xn , and the recharging function f (), and returns the maximum number of robots that can be destroyed by a sequence of EMP
activations.
9. Youre helping to run a high-performance computing system capable of
processing several terabytes of data per day. For each of n days, youre
presented with a quantity of data; on day i, 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 constrained 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 s1 terabytes, on the second day after
a reboot, you can process s2 terabytes, and so on, up to sn ; we assume
s1 > s2 > s3 > . . . > sn > 0. (Of course, on day i you can only process up to xi
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 x1, x2 , . . . , xn for the
next n days, and given the profile of your system as expressed by
s1, s2 , . . . , sn (and starting from a freshly rebooted system on day 1), choose
Exercises
the days on which youre going to reboot so as to maximize the total
amount of data you process.
Example. Suppose n = 4, and the values of xi and si are given by the
following table.
Day 1
Day 2
Day 3
Day 4
x
10
1
7
7
s
8
4
2
1
The best solution would be to reboot on day 2 only; this way, you process
8 terabytes 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
well.)
(a)
Give an example of an instance with the following properties.
There is a surplus of data in the sense that xi > s1 for every i.
The optimal solution reboots the system at least twice.
In addition to the example, you should 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 x1, x2 , . . . , xn and
s1, s2 , . . . , sn and returns the total number of terabytes processed by
an optimal solution.
10. Youre trying 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 supercomputers, 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 ai > 0 steps of the simulation if
your job is on machine A, and bi > 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
Chapter 6 Dynamic Programming
B cannot appear in consecutive minutes. For example, if your job is on
machine A in minute i, and you want to switch to machine B, then your
choice for minute i + 1 must be move , and then your choice for minute i + 2
can be 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.
The problem. Given values a1, a2 , . . . , an and b1, b2 , . . . , bn , find a plan of
maximum value. (Such a strategy will be called optimal .) 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 ai and bi are given by the
following table.
Minute 1
Minute 2
Minute 3
Minute 4
A
10
1
1
10
B
5
1
20
20
Then the plan of maximum value would be to choose A for minute 1,
then move for minute 2, and then B for minutes 3 and 4. The value of this
plan would be 10 + 0 + 20 + 20 = 50.
(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.
In minute 1, choose the machine achieving the larger of a1, b1
Set i = 2
While i n
What was the choice in minute i 1?
If A:
If bi+1 > ai + ai+1 then
Choose move in minute i and B in minute i + 1
Proceed to iteration i + 2
Else
Choose A in minute i
Proceed to iteration i + 1
Endif
If B: behave as above with roles of A and B reversed
EndWhile
Exercises
In your example, say what the correct answer is and also what the
algorithm above finds.
(b) Give an efficient algorithm that takes values for a1, a2 , . . . , an and
b1, b2 , . . . , bn and returns the value of an optimal plan.
11. Suppose youre consulting for a company that manufactures PC equipment and ships it to distributors all over the country. For each of the
next n weeks, they have a projected supply si of equipment (measured in
pounds), which has to be shipped by an air freight carrier.
Each weeks supply can be carried by one of two air freight companies,
A or B.
.
.
Company A charges a fixed rate r per pound (so it costs r si 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 time. 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
values s1, s2 , . . . , 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.
Then the optimal schedule would be to choose company A for the first
three weeks, then company B for a block of four consecutive weeks, and
then company A for the final three weeks.
12. Suppose we want to replicate a file over a collection of n servers, labeled
S1, S2 , . . . , Sn . To place a copy of the file at server Si results in a placement
cost of ci , for an integer ci > 0.
Now, if a user requests the file from server Si , and no copy of the file is
present at Si , then the servers Si+1, Si+2 , Si+3 . . . are searched in order until
a copy of the file is finally found, say at server Sj , where j > i. This results
in an access cost of j i. (Note that the lower-indexed servers Si1, Si2 , . . .
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 will terminate, at the latest, at Sn .
323
324
Chapter 6 Dynamic Programming
Wed like to place copies of the files 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 polynomial-time algorithm to find a configuration of minimum
total cost.
13. The problem of searching for cycles in graphs arises naturally in financial
trading applications. Consider a firm that trades shares in n different
companies. For each pair i = j , they maintain a trade ratio rij , meaning
that one share of i trades for rij shares of j . Here we allow the rate r to be
fractional; that is, rij = 2 means that you can trade three shares of i to get
3
two shares of j .
A trading cycle for a sequence of shares i1, i2 , . . . , ik consists of
successively trading shares in company i1 for shares in company i2, then
shares in company i2 for shares i3, and so on, finally trading shares in ik
back to shares in company i1. After such a sequence of trades, one ends up
with shares in the same company i1 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
dynamic 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
Exercises
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 E1, then to E2, then to E3, and so on, to an
edge set Eb. For i = 0, 1, 2, . . . , b, let Gi denote the graph (V , Ei ). 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 G0 , G1, G2 , . . . , Gb1, Gb.
We will assume that each of these graphs Gi 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 (P ). Our goal is to produce a sequence of
paths P0 , P1, . . . , Pb so that for each i, Pi is an s-t path in Gi . 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(P0 , P1, . . . , Pb) to be the number of indices i (0 i b 1)
for which Pi = Pi+1.
Fix a constant K > 0. We define the cost of the sequence of paths
P0 , P1, . . . , Pb to be
b
cost (P0 , P1, . . . , Pb) =
(Pi ) + K changes(P0 , P1, . . . , Pb).
i=0
(a)
Suppose it is possible to choose a single path P that is an s-t path in
each of the graphs G0 , G1, . . . , 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 , P1, . . . , Pb of minimum cost, where Pi is an s-t path in Gi 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.
.
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
Chapter 6 Dynamic Programming
.
.
The sky is mapped according to a one-dimensional coordinate system
(measured in degrees from some central baseline); event j will be
taking 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 if it is possible to
observe each event j S at its appointed time j , and the telescope has
adequate time (moving at its maximum of one degree per minute) to move
between consecutive events in S.
The problem. Given the coordinates of each of the n events, find a
viewable subset of maximum size, subject to the requirement that it
should contain event n. Such a solution will be called optimal .
Example. Suppose the one-dimensional coordinates of the events are as
shown here.
Event
1
2
3
4
5
6
7
8
9
Coordinate
1
4
1
4
5
4
6
7
2
Then the optimal solution is to observe events 1, 3, 6, 9. Note that the
telescope has time to move from one event in this set to the next, even
moving at one degree per minute.
(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.
Mark all events j with |dn dj | > n j as illegal (as
observing them would prevent you from observing event n)
Mark all other events as legal
Initialize current position to coordinate 0 at minute 0
While not at end of event sequence
Find the earliest legal event j that can be reached without
exceeding the maximum movement rate of the telescope
Add j to the set S
327
Exercises
Update current position to be coord.~dj at minute j
Endwhile
Output the set S
In your example, say what the correct answer is and also what
the algorithm above finds.
(b) Give an efficient algorithm that takes values for the coordinates
d1, d2 , . . . , dn of the events and returns the size of an optimal solution.
A should call B before D.
A
16. There are many sunny days in Ithaca, New York; but this year, as it
happens, the spring ROTC picnic at Cornell has fallen on a rainy day. The
ranking officer decides to postpone the picnic and must notify everyone
by phone. Here is the mechanism she uses to do this.
Each ROTC person on campus except the ranking officer reports to
a unique superior officer . Thus the reporting hierarchy can be described
by a tree T , rooted at the ranking officer, in which each other node v
has a parent node u equal to his or her superior officer. Conversely, we
will call v a direct subordinate of u. See Figure 6.30, in which A is the
ranking officer, B and D are the direct subordinates of A, and C is the
direct subordinate of B.
To notify everyone of the postponement, the ranking officer first
calls each of her direct subordinates, one at a time. As soon as each
subordinate gets the phone call, he or she must notify each of his or
her direct subordinates, one at a time. The process continues this way
until everyone has been notified. Note that each person in this process
can only call direct subordinates on the phone; for example, in Figure
6.30, A would not be allowed to call C.
We can picture this process as being divided into rounds . In one
round, each person who has already learned of the postponement can
call one of his or her direct subordinates on the phone. The number of
rounds it takes for everyone to be notified depends on the sequence in
which each person calls their direct subordinates. For example, in Figure
6.30, it will take only two rounds if A starts by calling B, but it will take
three rounds if A starts by calling D.
Give an efficient algorithm that determines the minimum number of
rounds needed for everyone to be notified, and outputs a sequence of
phone calls that achieves this minimum number of rounds.
17. Your friends have been studying the closing prices of tech stocks, looking
for interesting patterns. Theyve defined something called a rising trend ,
as follows.
B
D
C
Figure 6.30 A hierarchy with
four people. The fastest
broadcast scheme is for A
to call B in the first round.
In the second round, A calls
D and B calls C. If A were to
call D first, then C could not
learn the news until the third
round.
328
Chapter 6 Dynamic Programming
They have the closing price for a given stock recorded for n days in
succession; let these prices be denoted P [1], P [2], . . . , P [n]. A rising trend
in these prices is a subsequence of the prices P [i1], P [i2], . . . , P [ik ], for days
i1 < i2 < . . . < ik , so that
.
i1 = 1, and
.
P [ij ] < P [ij+1] for each j = 1, 2, . . . , k 1.
Thus a rising trend is a subsequence of the daysbeginning on the first
day and not necessarily contiguousso that the price strictly increases
over the days in this subsequence.
They are interested in finding the longest rising trend in a given
sequence of prices.
Example. Suppose n = 7, and the sequence of prices is
10, 1, 2, 11, 3, 4, 12.
Then the longest rising trend is given by the prices on days 1, 4, and 7.
Note that days 2, 3, 5, and 6 consist of increasing prices; but because this
subsequence does not begin on day 1, it does not fit the definition of a
rising trend.
(a)
Show that the following algorithm does not correctly return the
length of the longest rising trend, by giving an instance on which
it fails to return the correct answer.
Define i = 1
L=1
For j = 2 to n
If P [j ] > P [i] then
Set i = j .
Add 1 to L
Endif
Endfor
In your example, give the actual length of the longest rising trend,
and say what the algorithm above returns.
(b) Give an efficient algorithm that takes a sequence of prices P [1],
P [2], . . . , P [n] and returns the length of the longest rising trend.
18. Consider the sequence alignment problem over a four-letter alphabet
{z1, z2 , z3 , z4}, with a given gap cost and given mismatch costs. Assume
that each of these parameters is a positive integer.
Exercises
Suppose you are given two strings A = a1a2 . . . am and B = b1b2 . . . bn
and a proposed alignment between them. Give an O(mn) algorithm to
decide whether this alignment is the unique minimum-cost alignment
between A and B.
19. Youre consulting for a group of people (who would prefer not to be
mentioned here by name) whose jobs consist of monitoring and analyzing
electronic signals coming from ships in coastal Atlantic waters. They want
a fast algorithm for a basic primitive that arises frequently: untangling
a superposition of two known signals. Specifically, theyre picturing a
situation in which each of two ships is emitting a short sequence of 0s
and 1s over and over, and they want to make sure that the signal theyre
hearing is simply an interleaving of these two emissions, with nothing
extra added in.
This describes the whole problem; we can make it a little more explicit
as follows. Given a string x consisting of 0s and 1s, we write xk to denote k
copies of x concatenated together. We say that a string x is a repetition
of x if it is a prefix of xk for some number k. So x = 10110110110 is a repetition
of x = 101.
We say that a string s is an interleaving of x and y if its symbols can be
partitioned into two (not necessarily contiguous) subsequences s and s ,
so that s is a repetition of x and s is a repetition of y. (So each symbol in
s must belong to exactly one of s or s .) For example, if x = 101 and y = 00,
then s = 100010101 is an interleaving of x and y, since characters 1,2,5,7,8,9
form 101101a repetition of xand the remaining characters 3,4,6 form
000a repetition of y.
In terms of our application, x and y are the repeating sequences from
the two ships, and s is the signal were listening to: We want to make sure
s unravels into simple repetitions of x and y. Give an efficient algorithm
that takes strings s, x, and y and decides if s is an interleaving of x and y.
20. Suppose its nearing the end of the semester and youre taking n courses,
each with a final project that still has to be done. Each project will be
graded on the following scale: It will be assigned an integer number on
a scale of 1 to g > 1, higher numbers being better grades. Your goal, of
course, is to maximize your average grade on the n projects.
You have a total of H > n hours in which to work on the n projects
cumulatively, and you want to decide how to divide up this time. For
simplicity, assume H is a positive integer, and youll spend an integer
number of hours on each project. To figure out how best to divide up
your time, youve come up with a set of functions {fi : i = 1, 2, . . . , n} (rough
329
330
Chapter 6 Dynamic Programming
estimates, of course) for each of your n courses; if you spend h H hours
on the project for course i, youll get a grade of fi (h). (You may assume
that the functions fi are nondecreasing : if h < h , then fi (h) fi (h ).)
So the problem is: Given these functions {fi }, decide how many hours
to spend on each project (in integer values only) so that your average
grade, as computed according to the fi , is as large as possible. In order
to be efficient, the running time of your algorithm should be polynomial
in n, g , and H ; none of these quantities should appear as an exponent in
your running time.
21. Some time back, you helped a group of friends who were doing simulations for a computation-intensive investment company, and theyve
come back to you with a new problem. Theyre looking at n consecutive
days of a given stock, at some point in the past. The days are numbered
i = 1, 2, . . . , n; for each day i, they have a price p(i) per share for the stock
on that day.
For certain (possibly large) values of k, they want to study what they
call k-shot strategies . A k-shot strategy is a collection of m pairs of days
(b1, s1) , . . . , (bm , sm ), where 0 m k and
1 b1 < s1 < b2 < s2 . . . < bm < sm n .
We view these as a set of up to k nonoverlapping intervals, during each
of which the investors buy 1,000 shares of the stock (on day bi ) and then
sell it (on day si ). The return of a given k-shot strategy is simply the profit
obtained from the m buy-sell transactions, namely,
m
p(si ) p(bi ).
1,000
i=1
The investors want to assess the value of k-shot strategies by running
simulations on their n-day trace of the stock price. Your goal is to design
an efficient algorithm that determines, given the sequence of prices, the kshot strategy with the maximum possible return. Since k may be relatively
large in these simulations, your running time should be polynomial in
both n and k; it should not contain k in the exponent.
22. To assess how well-connected two nodes in a directed graph are, one
can not only look at the length of the shortest path between them, but
can also count the number of shortest paths.
This turns out to be a problem that can be solved efficiently, subject
to some restrictions on the edge costs. Suppose we are given a directed
graph G = (V , E ), with costs on the edges; the costs may be positive or
Exercises
negative, but every cycle in the graph has strictly positive cost. We are
also given two nodes v , w V . Give an efficient algorithm that computes
the number of shortest v-w paths in G . (The algorithm should not list all
the paths; just the number suffices.)
23. Suppose you are given a directed graph G = (V , E ) with costs on the edges
ce for e E and a sink t (costs may be negative). Assume that you also have
finite values d(v) for v V . Someone claims that, for each node v V , the
quantity d(v) is the cost of the minimum-cost path from node v to the
sink t .
(a)
Give a linear-time algorithm (time O(m) if the graph has m edges) that
verifies whether this claim is correct.
(b) Assume that the distances are correct, and d(v) is finite for all v V .
Now you need to compute distances to a different sink t . Give an
O(m log n) algorithm for computing distances d (v) for all nodes v V
to the sink node t . (Hint: It is useful to consider a new cost function
defined as follows: for edge e = (v , w), let ce = ce d(v) + d(w). Is there
a relation between costs of paths for the two different costs c and c ?)
24. Gerrymandering is the practice of carving up electoral districts in very
careful ways so as to lead to outcomes that favor a particular political
party. Recent court challenges to the practice have argued that through
this calculated redistricting, large numbers of voters are being effectively
(and intentionally) disenfranchised.
Computers, it turns out, have been implicated as the source of some
of the villainy in the news coverage on this topic: Thanks to powerful
software, gerrymandering has changed from an activity carried out by a
bunch of people with maps, pencil, and paper into the industrial-strength
process that it is today. Why is gerrymandering a computational problem?
There are database issues involved in tracking voter demographics down
to the level of individual streets and houses; and there are algorithmic
issues involved in grouping voters into districts. Lets think a bit about
what these latter issues look like.
Suppose we have a set of n precincts P1, P2 , . . . , Pn , each containing
m registered voters. Were supposed to divide these precincts into two
districts , each consisting of n/2 of the precincts. Now, for each precinct,
we have information on how many voters are registered to each of two
political parties. (Suppose, for simplicity, that every voter is registered
to one of these two.) Well say that the set of precincts is susceptible to
gerrymandering if it is possible to perform the division into two districts
in such a way that the same party holds a majority in both districts.
331
332
Chapter 6 Dynamic Programming
Give an algorithm to determine whether a given set of precincts
is susceptible to gerrymandering; the running time of your algorithm
should be polynomial in n and m.
Example. Suppose we have n = 4 precincts, and the following information
on registered voters.
Precinct
1
2
3
4
Number registered for party A
55
43
60
47
Number registered for party B
45
57
40
53
This set of precincts is susceptible since, if we grouped precincts 1
and 4 into one district, and precincts 2 and 3 into the other, then party
A would have a majority in both districts. (Presumably, the we who are
doing the grouping here are members of party A.) This example is a quick
illustration of the basic unfairness in gerrymandering: Although party A
holds only a slim majority in the overall population (205 to 195), it ends
up with a majority in not one but both districts.
25. Consider the problem faced by a stockbroker trying to sell a large number
of shares of stock in a company whose stock price has been steadily
falling in value. It is always hard to predict the right moment to sell stock,
but owning a lot of shares in a single company adds an extra complication:
the mere act of selling many shares in a single day will have an adverse
effect on the price.
Since future market prices, and the effect of large sales on these
prices, are very hard to predict, brokerage firms use models of the market
to help them make such decisions. In this problem, we will consider the
following simple model. Suppose we need to sell x shares of stock in a
company, and suppose that we have an accurate model of the market:
it predicts that the stock price will take the values p1, p2 , . . . , pn over the
next n days. Moreover, there is a function f () that predicts the effect
of large sales: if we sell y shares on a single day, it will permanently
decrease the price by f (y) from that day onward. So, if we sell y1 shares
on day 1, we obtain a price per share of p1 f (y1), for a total income of
y1 (p1 f (y1)). Having sold y1 shares on day 1, we can then sell y2 shares
on day 2 for a price per share of p2 f (y1) f (y2); this yields an additional
income of y2 (p2 f (y1) f (y2)). This process continues over all n days.
(Note, as in our calculation for day 2, that the decreases from earlier days
are absorbed into the prices for all later days.)
Design an efficient algorithm that takes the prices p1, . . . , pn and the
function f () (written as a list of values f (1) , f (2) , . . . , f (x)) and determines
Exercises
the best way to sell x shares by day n. In other words, find natural numbers
y1, y2 , . . . , yn so that x = y1 + . . . + yn , and selling yi shares on day i for
i = 1, 2, . . . , n maximizes the total income achievable. You should assume
that the share value pi is monotone decreasing, and f () is monotone
increasing; that is, selling a larger number of shares causes a larger
drop in the price. Your algorithms running time can have a polynomial
dependence on n (the number of days), x (the number of shares), and p1
(the peak price of the stock).
Example Consider the case when n = 3; the prices for the three days are
90, 80, 40; and f (y) = 1 for y 40,000 and f (y) = 20 for y > 40, 000. Assume
you start with x = 100, 000 shares. Selling all of them on day 1 would yield
a price of 70 per share, for a total income of 7,000,000. On the other hand,
selling 40,000 shares on day 1 yields a price of 89 per share, and selling
the remaining 60,000 shares on day 2 results in a price of 59 per share,
for a total income of 7,100,000.
26. Consider the following inventory problem. You are running a company
that sells some large product (lets assume you sell trucks), and predictions tell you the quantity of sales to expect over the next n months. Let
di denote the number of sales you expect in month i. Well assume that
all sales happen at the beginning of the month, and trucks that are not
sold are stored until the beginning of the next month. You can store at
most S trucks, and it costs C to store a single truck for a month. You
receive shipments of trucks by placing orders for them, and there is a
fixed ordering fee of K each time you place an order (regardless of the
number of trucks you order). You start out with no trucks. The problem
is to design an algorithm that decides how to place orders so that you
satisfy all the demands {di }, and minimize the costs. In summary:
.
.
There are two parts to the cost: (1) storageit costs C for every truck
on hand that is not needed that month; (2) ordering feesit costs K
for every order placed.
In each month you need enough trucks to satisfy the demand di ,
but the number left over after satisfying the demand for the month
should not exceed the inventory limit S.
Give an algorithm that solves this problem in time that is polynomial in
n and S.
27. The owners of an independently operated gas station are faced with the
following situation. They have a large underground tank in which they
store gas; the tank can hold up to L gallons at one time. Ordering gas is
quite expensive, so they want to order relatively rarely. For each order,
333
334
Chapter 6 Dynamic Programming
they need to pay a fixed price P for delivery in addition to the cost of the
gas ordered. However, it costs c to store a gallon of gas for an extra day,
so ordering too much ahead increases the storage cost.
They are planning to close for a week in the winter, and they want
their tank to be empty by the time they close. Luckily, based on years of
experience, they have accurate projections for how much gas they will
need each day until this point in time. Assume that there are n days left
until they close, and they need gi gallons of gas for each of the days
i = 1, . . . , n. Assume that the tank is empty at the end of day 0. Give an
algorithm to decide on which days they should place orders, and how
much to order so as to minimize their total cost.
28. Recall the scheduling problem from Section 4.2 in which we sought to
minimize the maximum lateness. There are n jobs, each with a deadline
di and a required processing time ti , and all jobs are available to be
scheduled starting at time s. For a job i to be done, it needs to be assigned
a period from si s to fi = si + ti , and different jobs should be assigned
nonoverlapping intervals. As usual, such an assignment of times will be
called a schedule .
In this problem, we consider the same setup, but want to optimize a
different objective. In particular, we consider the case in which each job
must either be done by its deadline or not at all. Well say that a subset J of
the jobs is schedulable if there is a schedule for the jobs in J so that each
of them finishes by its deadline. Your problem is to select a schedulable
subset of maximum possible size and give a schedule for this subset that
allows each job to finish by its deadline.
(a)
Prove that there is an optimal solution J (i.e., a schedulable set of
maximum size) in which the jobs in J are scheduled in increasing
order of their deadlines.
(b) Assume that all deadlines di and required times ti are integers. Give
an algorithm to find an optimal solution. Your algorithm should
run in time polynomial in the number of jobs n, and the maximum
deadline D = maxi di .
29. 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 wij on each edge (i , j ); and
we will assume these weights satisfy the triangle inequality wik wij + wjk .
For a subset V V , we will use G [V ] to denote the subgraph (with edge
weights) induced on the nodes in V .
We are given a set X V of k terminals that must be connected by
edges. We say that a Steiner tree on X is a set Z so that X Z V , together
Notes and Further Reading
with a spanning subtree T of G [Z ]. The weight of the Steiner tree is the
weight of the tree T .
Show that there is function f () and a polynomial function p() so that
the problem of finding a minimum-weight Steiner tree on X can be solved
in time O(f (k) p(n)).
Notes and Further Reading
Richard Bellman is credited with pioneering the systematic study of dynamic
programming (Bellman 1957); the algorithm in this chapter for segmented least
squares is based on Bellmans work from this early period (Bellman 1961).
Dynamic programming has since grown into a technique that is widely used
across computer science, operations research, control theory, and a number
of other areas. Much of the recent work on this topic has been concerned with
stochastic dynamic programming : Whereas our problem formulations tended
to tacitly assume that all input is known at the outset, many problems in
scheduling, production and inventory planning, and other domains involve
uncertainty, and dynamic programming algorithms for these problems encode
this uncertainty using a probabilistic formulation. The book by Ross (1983)
provides an introduction to stochastic dynamic programming.
Many extensions and variations of the Knapsack Problem have been
studied in the area of combinatorial optimization. As we discussed in the
chapter, the pseudo-polynomial bound arising from dynamic programming
can become prohibitive when the input numbers get large; in these cases,
dynamic programming is often combined with other heuristics to solve large
instances of Knapsack Problems in practice. The book by Martello and Toth
(1990) is devoted to computational approaches to versions of the Knapsack
Problem.
Dynamic programming emerged as a basic technique in computational biology in the early 1970s, in a urry of activity on the problem of sequence
comparison. Sankoff (2000) gives an interesting historical account of the early
work in this period. The books by Waterman (1995) and Guseld (1997) provide extensive coverage of sequence alignment algorithms (as well as many
related algorithms in computational biology); Mathews and Zuker (2004) discuss further approaches to the problem of RNA secondary structure prediction.
The space-efcient algorithm for sequence alignment is due to Hirschberg
(1975).
The algorithm for the Shortest-Path Problem described in this chapter is
based originally on the work of Bellman (1958) and Ford (1956). Many optimizations, motivated both by theoretical and experimental considerations,
335
336
Chapter 6 Dynamic Programming
have been added to this basic approach to shortest paths; a Web site maintained by Andrew Goldberg contains state-of-the-art code that he has developed for this problem (among a number of others), based on work by
Cherkassky, Goldberg and Radzik (1994). The applications of shortest-path
methods to Internet routing, and the trade-offs among the different algorithms
for networking applications, are covered in books by Bertsekas and Gallager
(1992), Keshav (1997), and Stewart (1998).
Notes on the Exercises Exercise 5 is based on discussions with Lillian Lee;
Exercise 6 is based on a result of Donald Knuth; Exercise 25 is based on results
of Dimitris Bertsimas and Andrew Lo; and Exercise 29 is based on a result of
S. Dreyfus and R. Wagner.
Chapter 7
Network Flow
In this chapter, we focus on a rich set of algorithmic problems that grow, in a
sense, out of one of the original problems we formulated at the beginning of
the course: Bipartite Matching .
Recall the set-up of the Bipartite Matching Problem. A bipartite graph
G = (V , E ) is an undirected graph whose node set can be partitioned as
V = X Y , with the property that every edge e E has one end in X and
the other end in Y . We often draw bipartite graphs as in Figure 7.1, with the
nodes in X in a column on the left, the nodes in Y in a column on the right,
and each edge crossing from the left column to the right column.
Now, weve already seen the notion of a matching at several points in
the course: Weve used the term to describe collections of pairs over a set,
with the property that no element of the set appears in more than one pair.
(Think of men (X ) matched to women (Y ) in the Stable Matching Problem,
or characters in the Sequence Alignment Problem.) In the case of a graph, the
edges constitute pairs of nodes, and we consequently say that a matching in
a graph G = (V , E ) is a set of edges M E with the property that each node
appears in at most one edge of M . A set of edges M is a perfect matching if
every node appears in exactly one edge of M .
Matchings in bipartite graphs can model situations in which objects are
being assigned to other objects. We have seen a number of such situations in
our earlier discussions of graphs and bipartite graphs. One natural example
arises when the nodes in X represent jobs, the nodes in Y represent machines,
and an edge (xi , yj ) indicates that machine yj 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.
Bipartite graphs can represent many other relations that arise between two
x1
y1
x2
y2
x3
y3
x4
y4
x5
y5
Figure 7.1 A bipartite graph.
338
Chapter 7 Network Flow
distinct sets of objects, such as the relation between customers and stores; or
houses and nearby re stations; and so forth.
One of the oldest problems in combinatorial algorithms is that of determining the size of the largest matching in a bipartite graph G . (As a special
case, note that G has a perfect matching if and only if |X | = |Y | and it has a
matching of size |X |.) This problem turns out to be solvable by an algorithm
that runs in polynomial time, but the development of this algorithm needs
ideas fundamentally different from the techniques that weve seen so far.
Rather than developing the algorithm directly, we begin by formulating a
general class of problemsnetwork ow problemsthat includes the Bipartite
Matching Problem as a special case. We then develop a polynomial-time
algorithm for a general problem, the Maximum-Flow Problem, and show how
this provides an efcient algorithm for Bipartite Matching as well. While the
initial motivation for network ow problems comes from the issue of trafc in
a network, we will see that they have applications in a surprisingly diverse set
of areas and lead to efcient algorithms not just for Bipartite Matching, but
for a host of other problems as well.
7.1 The Maximum-Flow Problem and the
Ford-Fulkerson Algorithm
The Problem
One often uses graphs to model transportation networksnetworks whose
edges carry some sort of trafc and whose nodes act as switches passing
trafc between different edges. Consider, for example, a highway system in
which the edges are highways and the nodes are interchanges; or a computer
network in which the edges are links that can carry packets and the nodes are
switches; or a uid network in which edges are pipes that carry liquid, and
the nodes are junctures where pipes are plugged together. Network models
of this type have several ingredients: capacities on the edges, indicating how
much they can carry; source nodes in the graph, which generate trafc; sink
(or destination) nodes in the graph, which can absorb trafc as it arrives;
and nally, the trafc itself, which is transmitted across the edges.
Flow Networks Well be considering graphs of this form, and we refer to the
trafc as owan abstract entity that is generated at source nodes, transmitted
across edges, and absorbed at sink nodes. Formally, well say that a ow
network is a directed graph G = (V , E ) with the following features.
.
Associated with each edge e is a capacity, which is a nonnegative number
that we denote ce .
339
7.1 The Maximum-Flow Problem and the Ford-Fulkerson Algorithm
.
There is a single source node s V .
.
There is a single sink node t V .
u
Nodes other than s and t will be called internal nodes.
We will make two assumptions about the ow networks we deal with: rst,
that no edge enters the source s and no edge leaves the sink t ; second, that
there is at least one edge incident to each node; and third, that all capacities
are integers. These assumptions make things cleaner to think about, and while
they eliminate a few pathologies, they preserve essentially all the issues we
want to think about.
Figure 7.2 illustrates a ow network with four nodes and ve edges, and
capacity values given next to each edge.
Dening Flow Next we dene what it means for our network to carry trafc,
or ow. We say that an s-t ow is a function f that maps each edge e to a
nonnegative real number, f : E R+; the value f (e ) intuitively represents the
amount of ow carried by edge e . A ow f must satisfy the following two
properties.1
(i) (Capacity conditions) For each e E , we have 0 f (e ) ce .
(ii) (Conservation conditions) For each node v other than s and t , we have
f (e ) =
e into v
Here
while
f (e ).
e out of v
f (e ) sums the ow value f (e ) over all edges entering node v,
e out of v f (e ) is the sum of ow values over all edges leaving node v .
e into v
Thus the ow on an edge cannot exceed the capacity of the edge. For
every node other than the source and the sink, the amount of ow entering
must equal the amount of ow leaving. The source has no entering edges (by
our assumption), but it is allowed to have ow going out; in other words, it
can generate ow. Symmetrically, the sink is allowed to have ow coming in,
even though it has no edges leaving it. The value of a ow f , denoted (f ), is
dened to be the amount of ow generated at the source:
f (e ).
(f ) =
e out of s
To make the notation more compact, we dene f out (v) = e out of v f (e )
and f in (v) = e into v f (e ). We can extend this to sets of vertices; if S V , we
1
Our notion of ow models trac as it goes through the network at a steady rate. We have a single
variable f (e ) to denote the amount of ow on edge e . We do not model bursty trac, where the ow
uctuates over time.
20
10
30
s
10
t
20
v
Figure 7.2 A flow network,
with source s and sink t . The
numbers next to the edges
are the capacities.
340
Chapter 7 Network Flow
dene f out (S) = e out of S f (e ) and f in (S) = e into S f (e ). In this terminology,
the conservation condition for nodes v = s , t becomes f in (v) = f out (v); and we
can write (f ) = f out (s).
The Maximum-Flow Problem Given a ow network, a natural goal is to
arrange the trafc so as to make as efcient use as possible of the available
capacity. Thus the basic algorithmic problem we will consider is the following:
Given a ow network, nd a ow of maximum possible value.
As we think about designing algorithms for this problem, its useful to
consider how the structure of the ow network places upper bounds on the
maximum value of an s-t ow. Here is a basic obstacle to the existence of
large ows: Suppose we divide the nodes of the graph into two sets, A and
B, so that s A and t B. Then, intuitively, any ow that goes from s to t
must cross from A into B at some point, and thereby use up some of the edge
capacity from A to B. This suggests that each such cut of the graph puts a
bound on the maximum possible ow value. The maximum-ow algorithm
that we develop here will be intertwined with a proof that the maximum-ow
value equals the minimum capacity of any such division, called the minimum
cut . As a bonus, our algorithm will also compute the minimum cut. We will
see that the problem of nding cuts of minimum capacity in a ow network
turns out to be as valuable, from the point of view of applications, as that of
nding a maximum ow.
Designing the Algorithm
Suppose we wanted to nd a maximum ow in a network. How should we
go about doing this? It takes some testing out to decide that an approach
such as dynamic programming doesnt seem to workat least, there is no
algorithm known for the Maximum-Flow Problem that could really be viewed
as naturally belonging to the dynamic programming paradigm. In the absence
of other ideas, we could go back and think about simple greedy approaches,
to see where they break down.
Suppose we start with zero ow: f (e ) = 0 for all e . Clearly this respects the
capacity and conservation conditions; the problem is that its value is 0. We
now try to increase the value of f by pushing ow along a path from s to t ,
up to the limits imposed by the edge capacities. Thus, in Figure 7.3, we might
choose the path consisting of the edges {(s , u) , (u , v) , (v , t )} and increase the
ow on each of these edges to 20, and leave f (e ) = 0 for the other two. In this
way, we still respect the capacity conditionssince we only set the ow as
high as the edge capacities would allowand the conservation conditions
since when we increase ow on an edge entering an internal node, we also
increase it on an edge leaving the node. Now, the value of our ow is 20, and
we can ask: Is this the maximum possible for the graph in the gure? If we
341
7.1 The Maximum-Flow Problem and the Ford-Fulkerson Algorithm
u
20
u
10
30
s
10
20
t
20
u
10
30
s
10
t
20
10
20
30
s
20
10
v
v
v
(a)
(b)
(c)
Figure 7.3 (a) The network of Figure 7.2. (b) Pushing 20 units of flow along the path
s , u , v , t . (c) The new kind of augmenting path using the edge (u , v) backward.
think about it, we see that the answer is no, since it is possible to construct
a ow of value 30. The problem is that were now stuckthere is no s-t path
on which we can directly push ow without exceeding some capacityand
yet we do not have a maximum ow. What we need is a more general way of
pushing ow from s to t , so that in a situation such as this, we have a way to
increase the value of the current ow.
Essentially, wed like to perform the following operation denoted by a
dotted line in Figure 7.3(c). We push 10 units of ow along (s , v); this now
results in too much ow coming into v. So we undo 10 units of ow on
(u , v); this restores the conservation condition at v but results in too little
ow leaving u. So, nally, we push 10 units of ow along (u , t ), restoring the
conservation condition at u. We now have a valid ow, and its value is 30. See
Figure 7.3, where the dark edges are carrying ow before the operation, and
the dashed edges form the new kind of augmentation.
This is a more general way of pushing ow: We can push forward on
edges with leftover capacity, and we can push backward on edges that are
already carrying ow, to divert it in a different direction. We now dene
the residual graph, which provides a systematic way to search for forwardbackward operations such as this.
The Residual Graph Given a ow network G , and a ow f on G , we dene
the residual graph Gf of G with respect to f as follows. (See Figure 7.4 for the
residual graph of the ow on Figure 7.3 after pushing 20 units of ow along
the path s , u , v , t .)
.
.
t
The node set of Gf is the same as that of G .
For each edge e = (u , v) of G on which f (e ) < ce , there are ce f (e )
leftover units of capacity on which we could try pushing ow forward.
342
Chapter 7 Network Flow
u
20
10
30
s
10
u
u
20
t
20
s
20
10
10
10
t
20
10
20
s
10
10
20
t
20
v
v
v
(a)
(b)
(c)
Figure 7.4 (a) The graph G with the path s , u , v , t used to push the first 20 units of flow.
(b) The residual graph of the resulting flow f , with the residual capacity next to each
edge. The dotted line is the new augmenting path. (c) The residual graph after pushing
an additional 10 units of flow along the new augmenting path s , v , u , t .
So we include the edge e = (u , v) in Gf , with a capacity of ce f (e ). We
will call edges included this way forward edges.
.
For each edge e = (u , v) of G on which f (e ) > 0, there are f (e ) units of
ow that we can undo if we want to, by pushing ow backward. So
we include the edge e = (v , u) in Gf , with a capacity of f (e ). Note that
e has the same ends as e , but its direction is reversed; we will call edges
included this way backward edges.
This completes the denition of the residual graph Gf . Note that each edge e
in G can give rise to one or two edges in Gf : If 0 < f (e ) < ce it results in both
a forward edge and a backward edge being included in Gf . Thus Gf has at
most twice as many edges as G . We will sometimes refer to the capacity of an
edge in the residual graph as a residual capacity, to help distinguish it from
the capacity of the corresponding edge in the original ow network G .
Augmenting Paths in a Residual Graph Now we want to make precise the
way in which we push ow from s to t in Gf . Let P be a simple s-t path in Gf
that is, P does not visit any node more than once. We dene bottleneck(P , f )
to be the minimum residual capacity of any edge on P , with respect to the
ow f . We now dene the following operation augment(f , P ), which yields a
new ow f in G .
augment(f , P )
Let b = bottleneck(P , f )
For each edge (u , v) P
If e = (u , v) is a forward edge then
increase f (e ) in G by b
7.1 The Maximum-Flow Problem and the Ford-Fulkerson Algorithm
Else ((u , v) is a backward edge, and let e = (v , u))
decrease f (e ) in G by b
Endif
Endfor
Return(f )
It was purely to be able to perform this operation that we dened the residual
graph; to reect the importance of augment, one often refers to any s-t path
in the residual graph as an augmenting path.
The result of augment(f , P ) is a new ow f in G , obtained by increasing
and decreasing the ow values on edges of P . Let us rst verify that f is indeed
a ow.
(7.1)
f is a ow in G.
Proof. We must verify the capacity and conservation conditions.
Since f differs from f only on edges of P , we need to check the capacity
conditions only on these edges. Thus, let (u , v) be an edge of P . Informally,
the capacity condition continues to hold because if e = (u , v) is a forward
edge, we specically avoided increasing the ow on e above ce ; and if (u , v)
is a backward edge arising from edge e = (v , u) E , we specically avoided
decreasing the ow on e below 0. More concretely, note that bottleneck(P , f )
is no larger than the residual capacity of (u , v). If e = (u , v) is a forward edge,
then its residual capacity is ce f (e ); thus we have
0 f (e ) f (e ) = f (e ) + bottleneck(P , f ) f (e ) + (ce f (e )) = ce ,
so the capacity condition holds. If (u , v) is a backward edge arising from edge
e = (v , u) E , then its residual capacity is f (e ), so we have
ce f (e ) f (e ) = f (e ) bottleneck(P , f ) f (e ) f (e ) = 0,
and again the capacity condition holds.
We need to check the conservation condition at each internal node that
lies on the path P . Let v be such a node; we can verify that the change in
the amount of ow entering v is the same as the change in the amount of
ow exiting v; since f satised the conservation condition at v, so must f .
Technically, there are four cases to check, depending on whether the edge of
P that enters v is a forward or backward edge, and whether the edge of P that
exits v is a forward or backward edge. However, each of these cases is easily
worked out, and we leave them to the reader.
343
344
Chapter 7 Network Flow
This augmentation operation captures the type of forward and backward
pushing of ow that we discussed earlier. Lets now consider the following
algorithm to compute an s-t ow in G .
Max-Flow
Initially f (e ) = 0 for all e in G
While there is an s-t path in the residual graph Gf
Let P be a simple s-t path in Gf
f = augment(f , P )
Update f to be f
Update the residual graph Gf to be Gf
Endwhile
Return f
Well call this the Ford-Fulkerson Algorithm, after the two researchers who
developed it in 1956. See Figure 7.4 for a run of the algorithm. The FordFulkerson Algorithm is really quite simple. What is not at all clear is whether
its central While loop terminates, and whether the ow returned is a maximum
ow. The answers to both of these questions turn out to be fairly subtle.
Analyzing the Algorithm: Termination and Running Time
First we consider some properties that the algorithm maintains by induction
on the number of iterations of the While loop, relying on our assumption that
all capacities are integers.
(7.2) At every intermediate stage of the Ford-Fulkerson Algorithm, the ow
values {f (e )} and the residual capacities in Gf are integers.
Proof. The statement is clearly true before any iterations of the While loop.
Now suppose it is true after j iterations. Then, since all residual capacities in
Gf are integers, the value bottleneck(P , f ) for the augmenting path found in
iteration j + 1 will be an integer. Thus the ow f will have integer values, and
hence so will the capacities of the new residual graph.
We can use this property to prove that the Ford-Fulkerson Algorithm
terminates. As at previous points in the book we will look for a measure of
progress that will imply termination.
First we show that the ow value strictly increases when we apply an
augmentation.
(7.3) Let f be a ow in G, and let P be a simple s-t path in Gf . Then
(f ) = (f ) + bottleneck(P , f ); and since bottleneck(P , f ) > 0, we have
(f ) > (f ).
7.1 The Maximum-Flow Problem and the Ford-Fulkerson Algorithm
Proof. The rst edge e of P must be an edge out of s in the residual graph
Gf ; and since the path is simple, it does not visit s again. Since G has no
edges entering s, the edge e must be a forward edge. We increase the ow
on this edge by bottleneck(P , f ), and we do not change the ow on any
other edge incident to s. Therefore the value of f exceeds the value of f by
bottleneck(P , f ).
We need one more observation to prove termination: We need to be able
to bound the maximum possible ow value. Heres one upper bound: If all the
edges out of s could be completely saturated with ow, the value of the ow
would be e out of s ce . Let C denote this sum. Thus we have (f ) C for all
s-t ows f . (C may be a huge overestimate of the maximum value of a ow
in G , but its handy for us as a nite, simply stated bound.) Using statement
(7.3), we can now prove termination.
(7.4) Suppose, as above, that all capacities in the ow network G are integers.
Then the Ford-Fulkerson Algorithm terminates in at most C iterations of the
While loop.
Proof. We noted above that no ow in G can have value greater than C , due to
the capacity condition on the edges leaving s. Now, by (7.3), the value of the
ow maintained by the Ford-Fulkerson Algorithm increases in each iteration;
so by (7.2), it increases by at least 1 in each iteration. Since it starts with the
value 0, and cannot go higher than C , the While loop in the Ford-Fulkerson
Algorithm can run for at most C iterations.
Next we consider the running time of the Ford-Fulkerson Algorithm. Let n
denote the number of nodes in G , and m denote the number of edges in G . We
have assumed that all nodes have at least one incident edge, hence m n/2,
and so we can use O(m + n) = O(m) to simplify the bounds.
(7.5) Suppose, as above, that all capacities in the ow network G are integers.
Then the Ford-Fulkerson Algorithm can be implemented to run in O(mC ) time.
Proof. We know from (7.4) that the algorithm terminates in at most C iterations of the While loop. We therefore consider the amount of work involved
in one iteration when the current ow is f .
The residual graph Gf has at most 2m edges, since each edge of G gives
rise to at most two edges in the residual graph. We will maintain Gf using an
adjacency list representation; we will have two linked lists for each node v,
one containing the edges entering v, and one containing the edges leaving v.
To nd an s-t path in Gf , we can use breadth-rst search or depth-rst search,
345
346
Chapter 7 Network Flow
which run in O(m + n) time; by our assumption that m n/2, O(m + n) is the
same as O(m). The procedure augment(f , P ) takes time O(n), as the path P
has at most n 1 edges. Given the new ow f , we can build the new residual
graph in O(m) time: For each edge e of G , we construct the correct forward
and backward edges in Gf .
A somewhat more efcient version of the algorithm would maintain the
linked lists of edges in the residual graph Gf as part of the augment procedure
that changes the ow f via augmentation.
7.2 Maximum Flows and Minimum Cuts in a
Network
We now continue with the analysis of the Ford-Fulkerson Algorithm, an activity
that will occupy this whole section. In the process, we will not only learn a
lot about the algorithm, but also nd that analyzing the algorithm provides us
with considerable insight into the Maximum-Flow Problem itself.
Analyzing the Algorithm: Flows and Cuts
Our next goal is to show that the ow that is returned by the Ford-Fulkerson
Algorithm has the maximum possible value of any ow in G . To make progress
toward this goal, we return to an issue that we raised in Section 7.1: the way in
which the structure of the ow network places upper bounds on the maximum
value of an s-t ow. We have already seen one upper bound: the value (f ) of
any s-t -ow f is at most C = e out of s ce . Sometimes this bound is useful, but
sometimes it is very weak. We now use the notion of a cut to develop a much
more general means of placing upper bounds on the maximum-ow value.
Consider dividing the nodes of the graph into two sets, A and B, so that
s A and t B. As in our discussion in Section 7.1, any such division places
an upper bound on the maximum possible ow value, since all the ow must
cross from A to B somewhere. Formally, we say that an s-t cut is a partition
(A, B) of the vertex set V , so that s A and t B. The capacity of a cut (A, B),
which we will denote c (A, B), is simply the sum of the capacities of all edges
out of A: c (A, B) = e out of A ce .
Cuts turn out to provide very natural upper bounds on the values of ows,
as expressed by our intuition above. We make this precise via a sequence of
facts.
(7.6) Let f be any s-t ow, and (A, B) any s-t cut. Then (f ) = f out (A)
f in (A).
7.2 Maximum Flows and Minimum Cuts in a Network
This statement is actually much stronger than a simple upper bound. It
says that by watching the amount of ow f sends across a cut, we can exactly
measure the ow value: It is the total amount that leaves A, minus the amount
that swirls back into A. This makes sense intuitively, although the proof
requires a little manipulation of sums.
Proof. By denition (f ) = f out (s). By assumption we have f in (s) = 0, as the
source s has no entering edges, so we can write (f ) = f out (s) f in (s). Since
every node v in A other than s is internal, we know that f out (v) f in (v) = 0
for all such nodes. Thus
(f out (v) f in (v)) ,
(f ) =
vA
since the only term in this sum that is nonzero is the one in which v is set to s.
Lets try to rewrite the sum on the right as follows. If an edge e has both
ends in A, then f (e ) appears once in the sum with a + and once with a ,
and hence these two terms cancel out. If e has only its tail in A, then f (e )
appears just once in the sum, with a +. If e has only its head in A, then f (e )
also appears just once in the sum, with a . Finally, if e has neither end in
A, then f (e ) doesnt appear in the sum at all. In view of this, we have
f out (v) f in (v) =
v A
f (e ) = f out (A) f in (A).
f (e )
e out of A
e into A
Putting together these two equations, we have the statement of (7.6).
If A = {s}, then f out (A) = f out (s), and f in (A) = 0 as there are no edges
entering the source by assumption. So the statement for this set A = {s} is
exactly the denition of the ow value (f ).
Note that if (A, B) is a cut, then the edges into B are precisely the edges
out of A. Similarly, the edges out of B are precisely the edges into A. Thus we
have f out (A) = f in (B) and f in (A) = f out (B), just by comparing the denitions
for these two expressions. So we can rephrase (7.6) in the following way.
(7.7)
Let f be any s-t ow, and (A, B) any s-t cut. Then (f ) = f in (B) f out (B).
If we set A = V {t } and B = {t } in (7.7), we have (f ) = f in (B) f out (B) =
By our assumption the sink t has no leaving edges, so we have
f out (t ) = 0. This says that we could have originally dened the value of a ow
equally well in terms of the sink t : It is f in (t ), the amount of ow arriving at
the sink.
f in (t ) f out (t ).
A very useful consequence of (7.6) is the following upper bound.
(7.8)
Let f be any s-t ow, and (A, B) any s-t cut. Then (f ) c (A, B).
347
348
Chapter 7 Network Flow
Proof.
(f ) = f out (A) f in (A)
f out (A)
f (e )
=
e out of A
ce
e out of A
= c (A, B).
Here the rst line is simply (7.6); we pass from the rst to the second since
f in (A) 0, and we pass from the third to the fourth by applying the capacity
conditions to each term of the sum.
In a sense, (7.8) looks weaker than (7.6), since it is only an inequality
rather than an equality. However, it will be extremely useful for us, since its
right-hand side is independent of any particular ow f . What (7.8) says is that
the value of every ow is upper-bounded by the capacity of every cut . In other
words, if we exhibit any s-t cut in G of some value c , we know immediately by
(7.8) that there cannot be an s-t ow in G of value greater than c . Conversely,
if we exhibit any s-t ow in G of some value , we know immediately by (7.8)
that there cannot be an s-t cut in G of value less than .
Analyzing the Algorithm: Max-Flow Equals Min-Cut
Let f denote the ow that is returned by the Ford-Fulkerson Algorithm. We
want to show that f has the maximum possible value of any ow in G , and
we do this by the method discussed above: We exhibit an s-t cut (A , B) for
which (f ) = c (A , B). This immediately establishes that f has the maximum
value of any ow, and that (A , B) has the minimum capacity of any s-t cut.
The Ford-Fulkerson Algorithm terminates when the ow f has no s-t path
in the residual graph Gf . This turns out to be the only property needed for
proving its maximality.
(7.9) If f is an s-t-ow such that there is no s-t path in the residual graph Gf ,
then there is an s-t cut (A , B) in G for which (f ) = c (A , B). Consequently,
f has the maximum value of any ow in G, and (A , B) has the minimum
capacity of any s-t cut in G.
Proof. The statement claims the existence of a cut satisfying a certain desirable
property; thus we must now identify such a cut. To this end, let A denote the
set of all nodes v in G for which there is an s-v path in Gf . Let B denote the
set of all other nodes: B = V A.
7.2 Maximum Flows and Minimum Cuts in a Network
(u, v) is saturated
with flow.
Residual graph
v
u
t
s
v
u
(u , v ) carries
no flow.
A*
B*
Figure 7.5 The (A , B ) cut in the proof of (7.9).
First we establish that (A , B) is indeed an s-t cut. It is clearly a partition
of V . The source s belongs to A since there is always a path from s to s.
Moreover, t A by the assumption that there is no s-t path in the residual
graph; hence t B as desired.
Next, suppose that e = (u , v) is an edge in G for which u A and v B, as
shown in Figure 7.5. We claim that f (e ) = ce . For if not, e would be a forward
edge in the residual graph Gf , and since u A, there is an s-u path in Gf ;
appending e to this path, we would obtain an s-v path in Gf , contradicting our
assumption that v B.
Now suppose that e = (u , v ) is an edge in G for which u B and v A.
We claim that f (e ) = 0. For if not, e would give rise to a backward edge
e = (v , u ) in the residual graph Gf , and since v A, there is an s-v path in
Gf ; appending e to this path, we would obtain an s-u path in Gf , contradicting
our assumption that u B.
So all edges out of A are completely saturated with ow, while all edges
into A are completely unused. We can now use (7.6) to reach the desired
conclusion:
(f ) = f out (A) f in (A)
f (e )
=
ce 0
=
e out of
f (e )
e into A
e out of A
A
= c (A , B).
349
350
Chapter 7 Network Flow
Note how, in retrospect, we can see why the two types of residual edges
forward and backwardare crucial in analyzing the two terms in the expression from (7.6).
Given that the Ford-Fulkerson Algorithm terminates when there is no s-t
in the residual graph, (7.6) immediately implies its optimality.
(7.10)
ow.
The ow f returned by the Ford-Fulkerson Algorithm is a maximum
We also observe that our algorithm can easily be extended to compute a
minimum s-t cut (A , B), as follows.
(7.11) Given a ow f of maximum value, we can compute an s-t cut of
minimum capacity in O(m) time.
Proof. We simply follow the construction in the proof of (7.9). We construct
the residual graph Gf , and perform breadth-rst search or depth-rst search to
determine the set A of all nodes that s can reach. We then dene B = V A,
and return the cut (A , B).
Note that there can be many minimum-capacity cuts in a graph G ; the
procedure in the proof of (7.11) is simply nding a particular one of these
cuts, starting from a maximum ow f .
As a bonus, we have obtained the following striking fact through the
analysis of the algorithm.
(7.12) In every ow network, there is a ow f and a cut (A, B) so that
(f ) = c (A, B).
The point is that f in (7.12) must be a maximum s-t ow; for if there were
a ow f of greater value, the value of f would exceed the capacity of (A, B),
and this would contradict (7.8). Similarly, it follows that (A, B) in (7.12) is
a minimum cut no other cut can have smaller capacityfor if there were a
cut (A , B ) of smaller capacity, it would be less than the value of f , and this
again would contradict (7.8). Due to these implications, (7.12) is often called
the Max-Flow Min-Cut Theorem, and is phrased as follows.
(7.13) In every ow network, the maximum value of an s-t ow is equal to
the minimum capacity of an s-t cut.
7.2 Maximum Flows and Minimum Cuts in a Network
Further Analysis: Integer-Valued Flows
Among the many corollaries emerging from our analysis of the Ford-Fulkerson
Algorithm, here is another extremely important one. By (7.2), we maintain an
integer-valued ow at all times, and by (7.9), we conclude with a maximum
ow. Thus we have
(7.14) If all capacities in the ow network are integers, then there is a
maximum ow f for which every ow value f (e ) is an integer.
Note that (7.14) does not claim that every maximum ow is integer-valued,
only that some maximum ow has this property. Curiously, although (7.14)
makes no reference to the Ford-Fulkerson Algorithm, our algorithmic approach
here provides what is probably the easiest way to prove it.
Real Numbers as Capacities? Finally, before moving on, we can ask how
crucial our assumption of integer capacities was (ignoring (7.4), (7.5) and
(7.14), which clearly needed it). First we notice that allowing capacities to be
rational numbers does not make the situation any more general, since we can
determine the least common multiple of all capacities, and multiply them all
by this value to obtain an equivalent problem with integer capacities.
But what if we have real numbers as capacities? Where in the proof did we
rely on the capacities being integers? In fact, we relied on it quite crucially: We
used (7.2) to establish, in (7.4), that the value of the ow increased by at least 1
in every step. With real numbers as capacities, we should be concerned that the
value of our ow keeps increasing, but in increments that become arbitrarily
smaller and smaller; and hence we have no guarantee that the number of
iterations of the loop is nite. And this turns out to be an extremely real worry,
for the following reason: With pathological choices for the augmenting path,
the Ford-Fulkerson Algorithm with real-valued capacities can run forever.
However, one can still prove that the Max-Flow Min-Cut Theorem (7.12) is
true even if the capacities may be real numbers. Note that (7.9) assumed only
that the ow f has no s-t path in its residual graph Gf , in order to conclude that
there is an s-t cut of equal value. Clearly, for any ow f of maximum value, the
residual graph has no s-t -path; otherwise there would be a way to increase the
value of the ow. So one can prove (7.12) in the case of real-valued capacities
by simply establishing that for every ow network, there exists a maximum
ow.
Of course, the capacities in any practical application of network ow would
be integers or rational numbers. However, the problem of pathological choices
for the augmenting paths can manifest itself even with integer capacities: It
can make the Ford-Fulkerson Algorithm take a gigantic number of iterations.
351
352
Chapter 7 Network Flow
In the next section, we discuss how to select augmenting paths so as to avoid
the potential bad behavior of the algorithm.
7.3 Choosing Good Augmenting Paths
In the previous section, we saw that any way of choosing an augmenting path
increases the value of the ow, and this led to a bound of C on the number of
augmentations, where C = e out of s ce . When C is not very large, this can be
a reasonable bound; however, it is very weak when C is large.
To get a sense for how bad this bound can be, consider the example graph
in Figure 7.2; but this time assume the capacities are as follows: The edges
(s , v), (s , u), (v , t ) and (u , t ) have capacity 100, and the edge (u , v) has capacity
1, as shown in Figure 7.6. It is easy to see that the maximum ow has value 200,
and has f (e ) = 100 for the edges (s , v) , (s , u) , (v , t ) and (u , t ) and value 0 on the
edge (u , v). This ow can be obtained by a sequence of two augmentations,
using the paths of nodes s , u , t and path s , v , t . But consider how bad the
Ford-Fulkerson Algorithm can be with pathological choices for the augmenting
paths. Suppose we start with augmenting path P1 of nodes s , u , v , t in this
order (as shown in Figure 7.6). This path has bottleneck(P1, f ) = 1. After
this augmentation, we have f (e ) = 1 on the edge e = (u , v), so the reverse
edge is in the residual graph. For the next augmenting path, we choose the
path P2 of the nodes s , v , u , t in this order. In this second augmentation, we
get bottleneck(P2 , f ) = 1 as well. After this second augmentation, we have
f (e ) = 0 for the edge e = (u , v), so the edge is again in the residual graph.
Suppose we alternate between choosing P1 and P2 for augmentation. In this
case, each augmentation will have 1 as the bottleneck capacity, and it will
take 200 augmentations to get the desired ow of value 200. This is exactly
the bound we proved in (7.4), since C = 200 in this example.
Designing a Faster Flow Algorithm
The goal of this section is to show that with a better choice of paths, we can
improve this bound signicantly. A large amount of work has been devoted
to nding good ways of choosing augmenting paths in the Maximum-Flow
Problem so as to minimize the number of iterations. We focus here on one
of the most natural approaches and will mention other approaches at the end
of the section. Recall that augmentation increases the value of the maximum
ow by the bottleneck capacity of the selected path; so if we choose paths
with large bottleneck capacity, we will be making a lot of progress. A natural
idea is to select the path that has the largest bottleneck capacity. Having to
nd such paths can slow down each individual iteration by quite a bit. We will
avoid this slowdown by not worrying about selecting the path that has exactly
7.3 Choosing Good Augmenting Paths
u
u
99
100
100
P1
100
1
1
s
t
1
s
t
P2
100
100
100
1
v
v
(a)
(b)
u
u
99
99
98
1
P1
99
2
1
1
1
s
1
t
99
v
(c)
1
s
1
P2
99
1
t
98
99
v
2
(d)
Figure 7.6 Parts (a) through (d) depict four iterations of the Ford-Fulkerson Algorithm
using a bad choice of augmenting paths: The augmentations alternate between the path
P1 through the nodes s, u, v, t in order and the path P2 through the nodes s, v, u, t in
order.
the largest bottleneck capacity. Instead, we will maintain a so-called scaling
parameter , and we will look for paths that have bottleneck capacity of at
least .
Let Gf ( ) be the subset of the residual graph consisting only of edges with
residual capacity of at least . We will work with values of that are powers
of 2. The algorithm is as follows.
Scaling Max-Flow
Initially f (e ) = 0 for all e in G
Initially set
to be the largest power of 2 that is no larger
than the maximum capacity out of s: maxe out of s ce
While 1
While there is an s-t path in the graph Gf ( )
Let P be a simple s-t path in Gf ( )
353
354
Chapter 7 Network Flow
f = augment(f , P )
Update f to be f and update Gf ( )
Endwhile
=
/2
Endwhile
Return f
Analyzing the Algorithm
First observe that the new Scaling Max-Flow Algorithm is really just an
implementation of the original Ford-Fulkerson Algorithm. The new loops, the
value , and the restricted residual graph Gf ( ) are only used to guide the
selection of residual pathwith the goal of using edges with large residual
capacity for as long as possible. Hence all the properties that we proved
about the original Max-Flow Algorithm are also true for this new version: the
ow remains integer-valued throughout the algorithm, and hence all residual
capacities are integer-valued.
(7.15) If the capacities are integer-valued, then throughout the Scaling MaxFlow Algorithm the ow and the residual capacities remain integer-valued. This
implies that when = 1, Gf ( ) is the same as Gf , and hence when the algorithm
terminates the ow, f is of maximum value.
Next we consider the running time. We call an iteration of the outside
While loopwith a xed value of the -scaling phase . It is easy to give
an upper bound on the number of different -scaling phases, in terms of the
value C = e out of s ce that we also used in the previous section. The initial
value of
is at most C , it drops by factors of 2, and it never gets below 1.
Thus,
(7.16) The number of iterations of the outer While loop is at most 1 +
log2 C .
The harder part is to bound the number of augmentations done in each
scaling phase. The idea here is that we are using paths that augment the ow
by a lot, and so there should be relatively few augmentations. During the scaling phase, we only use edges with residual capacity of at least . Using
(7.3), we have
(7.17) During the
value by at least .
-scaling phase, each augmentation increases the ow
7.3 Choosing Good Augmenting Paths
The key insight is that at the end of the -scaling phase, the ow f cannot be
too far from the maximum possible value.
(7.18) Let f be the ow at the end of the -scaling phase. There is an s-t
cut (A, B) in G for which c (A, B) (f ) + m , where m is the number of edges
in the graph G. Consequently, the maximum ow in the network has value at
most (f ) + m .
Proof. This proof is analogous to our proof of (7.9), which established that
the ow returned by the original Max-Flow Algorithm is of maximum value.
As in that proof, we must identify a cut (A, B) with the desired property.
Let A denote the set of all nodes v in G for which there is an s-v path in Gf ( ).
Let B denote the set of all other nodes: B = V A. We can see that (A, B) is
indeed an s-t cut as otherwise the phase would not have ended.
Now consider an edge e = (u , v) in G for which u A and v B. We claim
that ce < f (e ) + . For if this were not the case, then e would be a forward
edge in the graph Gf ( ), and since u A, there is an s-u path in Gf ( );
appending e to this path, we would obtain an s-v path in Gf ( ), contradicting
our assumption that v B. Similarly, we claim that for any edge e = (u , v ) in
G for which u B and v A, we have f (e ) < . Indeed, if f (e ) , then e
would give rise to a backward edge e = (v , u ) in the graph Gf ( ), and since
v A, there is an s-v path in Gf ( ); appending e to this path, we would
obtain an s-u path in Gf ( ), contradicting our assumption that u B.
So all edges e out of A are almost saturatedthey satisfy ce < f (e ) +
and all edges into A are almost emptythey satisfy f (e ) < . We can now use
(7.6) to reach the desired conclusion:
f (e )
(f ) =
e out of A
f (e )
e into A
(ce
)
e into A
e out of A
ce
=
e out of A
e out of A
e into A
c (A, B) m .
Here the rst inequality follows from our bounds on the ow values of edges
across the cut, and the second inequality follows from the simple fact that the
graph only contains m edges total.
The maximum-ow value is bounded by the capacity of any cut by (7.8).
We use the cut (A, B) to obtain the bound claimed in the second statement.
355
356
Chapter 7 Network Flow
(7.19)
The number of augmentations in a scaling phase is at most 2m.
Proof. The statement is clearly true in the rst scaling phase: we can use
each of the edges out of s only for at most one augmentation in that phase.
Now consider a later scaling phase , and let fp be the ow at the end of the
= 2 as our parameter. By
previous scaling phase. In that phase, we used
(7.18), the maximum ow f is at most (f ) (fp) + m = (fp) + 2m . In
the -scaling phase, each augmentation increases the ow by at least , and
hence there can be at most 2m augmentations.
An augmentation takes O(m) time, including the time required to set up
the graph and nd the appropriate path. We have at most 1 + log2 C scaling
phases and at most 2m augmentations in each scaling phase. Thus we have
the following result.
(7.20) The Scaling Max-Flow Algorithm in a graph with m edges and integer
capacities nds a maximum ow in at most 2m(1 + log2 C ) augmentations.
It can be implemented to run in at most O(m2 log2 C ) time.
When C is large, this time bound is much better than the O(mC ) bound
that applied to an arbitrary implementation of the Ford-Fulkerson Algorithm.
In our example at the beginning of this section, we had capacities of size
100, but we could just as well have used capacities of size 2100; in this case,
the generic Ford-Fulkerson Algorithm could take time proportional to 2100,
while the scaling algorithm will take time proportional to log2(2100) = 100.
One way to view this distinction is as follows: The generic Ford-Fulkerson
Algorithm requires time proportional to the magnitude of the capacities, while
the scaling algorithm only requires time proportional to the number of bits
needed to specify the capacities in the input to the problem. As a result, the
scaling algorithm is running in time polynomial in the size of the input (i.e., the
number of edges and the numerical representation of the capacities), and so
it meets our traditional goal of achieving a polynomial-time algorithm. Bad
implementations of the Ford-Fulkerson Algorithm, which can require close
to C iterations, do not meet this standard of polynomiality. (Recall that in
Section 6.4 we used the term pseudo-polynomial to describe such algorithms,
which are polynomial in the magnitudes of the input numbers but not in the
number of bits needed to represent them.)
Extensions: Strongly Polynomial Algorithms
Could we ask for something qualitatively better than what the scaling algorithm guarantees? Here is one thing we could hope for: Our example graph
(Figure 7.6) had four nodes and ve edges; so it would be nice to use a
7.4 The Preow-Push Maximum-Flow Algorithm
number of iterations that is polynomial in the numbers 4 and 5, completely
independently of the values of the capacities. Such an algorithm, which is
polynomial in |V | and |E | only, and works with numbers having a polynomial number of bits, is called a strongly polynomial algorithm. In fact, there
is a simple and natural implementation of the Ford-Fulkerson Algorithm that
leads to such a strongly polynomial bound: each iteration chooses the augmenting path with the fewest number of edges. Dinitz, and independently
Edmonds and Karp, proved that with this choice the algorithm terminates in
at most O(mn) iterations. In fact, these were the rst polynomial algorithms
for the Maximum-Flow Problem. There has since been a huge amount of work
devoted to improving the running times of maximum-ow algorithms. There
are currently algorithms that achieve running times of O(mn log n), O(n3), and
O(min(n2/3, m1/2)m log n log U ), where the last bound assumes that all capacities are integral and at most U . In the next section, well discuss a strongly
polynomial maximum-ow algorithm based on a different principle.
* 7.4 The Preow-Push Maximum-Flow Algorithm
From the very beginning, our discussion of the Maximum-Flow Problem has
been centered around the idea of an augmenting path in the residual graph.
However, there are some very powerful techniques for maximum ow that are
not explicitly based on augmenting paths. In this section we study one such
technique, the Preow-Push Algorithm.
Designing the Algorithm
Algorithms based on augmenting paths maintain a ow f , and use the augment
procedure to increase the value of the ow. By way of contrast, the PreowPush Algorithm will, in essence, increase the ow on an edge-by-edge basis.
Changing the ow on a single edge will typically violate the conservation condition, and so the algorithm will have to maintain something less well behaved
than a owsomething that does not obey conservationas it operates.
Preows We say that an s-t preow (preow, for short) is a function f that
maps each edge e to a nonnegative real number, f : E R+. A preow f must
satisfy the capacity conditions:
(i) For each e E , we have 0 f (e ) ce .
In place of the conservation conditions, we require only inequalities: Each
node other than s must have at least as much ow entering as leaving.
(ii) For each node v other than the source s, we have
f (e )
e into v
f (e ) .
e out of v
357
358
Chapter 7 Network Flow
We will call the difference
ef (v) =
f (e )
e into v
f (e )
e out of v
the excess of the preow at node v. Notice that a preow where all nodes
other than s and t have zero excess is a ow, and the value of the ow is
exactly ef (t ) = ef (s). We can still dene the concept of a residual graph Gf
for a preow f , just as we did for a ow. The algorithm will push ow along
edges of the residual graph (using both forward and backward edges).
Preows and Labelings The Preow-Push Algorithm will maintain a preow
and work on converting the preow into a ow. The algorithm is based on the
physical intuition that ow naturally nds its way downhill. The heights
for this intuition will be labels h(v) for each node v that the algorithm will
dene and maintain, as shown in Figure 7.7. We will push ow from nodes
with higher labels to those with lower labels, following the intuition that uid
ows downhill. To make this precise, a labeling is a function h : V Z0 from
the nodes to the nonnegative integers. We will also refer to the labels as heights
of the nodes. We will say that a labeling h and an s-t preow f are compatible if
(i) (Source and sink conditions) h(t ) = 0 and h(s) = n,
(ii) (Steepness conditions) For all edges (v , w) Ef in the residual graph, we
have h(v) h(w) + 1.
Heights
Edges in the residual graph
may not be too steep.
4
3
2
1
0
t
Nodes
Figure 7.7 A residual graph and a compatible labeling. No edge in the residual graph
can be too steepits tail can be at most one unit above its head in height. The source
node s must have h(s) = n and is not drawn in the figure.
7.4 The Preow-Push Maximum-Flow Algorithm
Intuitively, the height difference n between the source and the sink is meant to
ensure that the ow starts high enough to ow from s toward the sink t , while
the steepness condition will help by making the descent of the ow gradual
enough to make it to the sink.
The key property of a compatible preow and labeling is that there can be
no s-t path in the residual graph.
(7.21) If s-t preow f is compatible with a labeling h, then there is no s-t
path in the residual graph Gf .
Proof. We prove the statement by contradiction. Let P be a simple s-t path in
the residual graph G . Assume that the nodes along P are s , v1, . . . , vk = t . By
denition of a labeling compatible with preow f , we have that h(s) = n. The
edge (s , v1) is in the residual graph, and hence h(v1) h(s) 1 = n 1. Using
induction on i and the steepness condition for the edge (vi1, vi ), we get that
for all nodes vi in path P the height is at least h(vi ) n i. Notice that the last
node of the path is vk = t ; hence we get that h(t ) n k. However, h(t ) = 0
by denition; and k < n as the path P is simple. This contradiction proves the
claim.
Recall from (7.9) that if there is no s-t path in the residual graph Gf of a
ow f , then the ow has maximum value. This implies the following corollary.
(7.22) If s-t ow f is compatible with a labeling h, then f is a ow of
maximum value.
Note that (7.21) applies to preows, while (7.22) is more restrictive in
that it applies only to ows. Thus the Preow-Push Algorithm will maintain a
preow f and a labeling h compatible with f , and it will work on modifying f
and h so as to move f toward being a ow. Once f actually becomes a ow, we
can invoke (7.22) to conclude that it is a maximum ow. In light of this, we
can view the Preow-Push Algorithm as being in a way orthogonal to the FordFulkerson Algorithm. The Ford-Fulkerson Algorithm maintains a feasible ow
while changing it gradually toward optimality. The Preow-Push Algorithm,
on the other hand, maintains a condition that would imply the optimality of a
preow f , if it were to be a feasible ow, and the algorithm gradually transforms
the preow f into a ow.
To start the algorithm, we will need to dene an initial preow f and
labeling h that are compatible. We will use h(v) = 0 for all v = s, and h(s) = n,
as our initial labeling. To make a preow f compatible with this labeling, we
need to make sure that no edges leaving s are in the residual graph (as these
edges do not satisfy the steepness condition). To this end, we dene the initial
359
360
Chapter 7 Network Flow
preow as f (e ) = ce for all edges e = (s , v) leaving the source, and f (e ) = 0 for
all other edges.
(7.23)
The initial preow f and labeling h are compatible.
Pushing and Relabeling Next we will discuss the steps the algorithm makes
toward turning the preow f into a feasible ow, while keeping it compatible
with some labeling h. Consider any node v that has excessthat is, ef (v) > 0.
If there is any edge e in the residual graph Gf that leaves v and goes to a node
w at a lower height (note that h(w) is at most 1 less than h(v) due to the
steepness condition), then we can modify f by pushing some of the excess
ow from v to w. We will call this a push operation.
push(f , h , v , w)
Applicable if ef (v) > 0, h(w) < h(v) and (v , w) Ef
If e = (v , w) is a forward edge then
let = min(ef (v) , ce f (e )) and
increase f (e ) by
If (v , w) is a backward edge then
let e = (w , v), = min(ef (v) , f (e )) and
decrease f (e ) by
Return(f , h)
If we cannot push the excess of v along any edge leaving v, then we will
need to raise vs height. We will call this a relabel operation.
relabel(f , h , v)
Applicable if ef (v) > 0, and
for all edges (v , w) Ef we have h(w) h(v)
Increase h(v) by 1
Return(f , h)
The Full Preow-Push Algorithm So, in summary, the Preow-Push Algorithm is as follows.
Preflow-Push
Initially h(v) = 0 for all v = s and h(s) = n and
f (e ) = ce for all e = (s , v) and f (e ) = 0 for all other edges
While there is a node v = t with excess ef (v) > 0
Let v be a node with excess
If there is w such that push(f , h , v , w) can be applied then
push(f , h , v , w)
7.4 The Preow-Push Maximum-Flow Algorithm
Else
relabel(f , h , v)
Endwhile
Return(f)
Analyzing the Algorithm
As usual, this algorithm is somewhat underspecied. For an implementation
of the algorithm, we will have to specify which node with excess to choose,
and how to efciently select an edge on which to push. However, it is clear
that each iteration of this algorithm can be implemented in polynomial time.
(Well discuss later how to implement it reasonably efciently.) Further, it is
not hard to see that the preow f and the labeling h are compatible throughout
the algorithm. If the algorithm terminatessomething that is far from obvious
based on its descriptionthen there are no nodes other than t with positive
excess, and hence the preow f is in fact a ow. It then follows from (7.22)
that f would be a maximum ow at termination.
We summarize a few simple observations about the algorithm.
(7.24)
Throughout the Preow-Push Algorithm:
(i) the labels are nonnegative integers;
(ii) f is a preow, and if the capacities are integral, then the preow f is
integral; and
(iii) the preow f and labeling h are compatible.
If the algorithm returns a preow f , then f is a ow of maximum value.
Proof. By (7.23) the initial preow f and labeling h are compatible. We will
show using induction on the number of push and relabel operations that
f and h satisfy the properties of the statement. The push operation modies
the preow f , but the bounds on guarantee that the f returned satises
the capacity constraints, and that excesses all remain nonnegative, so f is a
preow. To see that the preow f and the labeling h are compatible, note that
push(f , h , v , w) can add one edge to the residual graph, the reverse edge (v , w),
and this edge does satisfy the steepness condition. The relabel operation
increases the label of v, and hence increases the steepness of all edges leaving
v. However, it only applies when no edge leaving v in the residual graph is
going downward, and hence the preow f and the labeling h are compatible
after relabeling.
The algorithm terminates if no node other than s or t has excess. In this
case, f is a ow by denition; and since the preow f and the labeling h
361
362
Chapter 7 Network Flow
remain compatible throughout the algorithm, (7.22) implies that f is a ow of
maximum value.
Next we will consider the number of push and relabel operations. First
we will prove a limit on the relabel operations, and this will help prove a
limit on the maximum number of push operations possible. The algorithm
never changes the label of s (as the source never has positive excess). Each
other node v starts with h(v) = 0, and its label increases by 1 every time it
changes. So we simply need to give a limit on how high a label can get. We
only consider a node v for relabel when v has excess. The only source of ow
in the network is the source s; hence, intuitively, the excess at v must have
originated at s. The following consequence of this fact will be key to bounding
the labels.
(7.25) Let f be a preow. If the node v has excess, then there is a path in Gf
from v to the source s.
Proof. Let A denote all the nodes w such that there is a path from w to s in
the residual graph Gf , and let B = V A. We need to show that all nodes with
excess are in A.
Notice that s A. Further, no edges e = (x , y) leaving A can have positive
ow, as an edge with f (e ) > 0 would give rise to a reverse edge (y , x) in the
residual graph, and then y would have been in A. Now consider the sum of
excesses in the set B, and recall that each node in B has nonnegative excess,
as s B.
(f in (v) f out (v))
ef (v) =
0
vB
vB
Lets rewrite the sum on the right as follows. If an edge e has both ends
in B, then f (e ) appears once in the sum with a + and once with a , and
hence these two terms cancel out. If e has only its head in B, then e leaves A,
and we saw above that all edges leaving A have f (e ) = 0. If e has only its tail
in B, then f (e ) appears just once in the sum, with a . So we get
ef (v) = f out (B).
0
vB
Since ows are nonnegative, we see that the sum of the excesses in B is zero;
since each individual excess in B is nonnegative, they must therefore all be 0.
Now we are ready to prove that the labels do not change too much. Recall
that n denotes the number of nodes in V .
7.4 The Preow-Push Maximum-Flow Algorithm
(7.26)
Throughout the algorithm, all nodes have h(v) 2n 1.
Proof. The initial labels h(t ) = 0 and h(s) = n do not change during the
algorithm. Consider some other node v = s , t . The algorithm changes vs label
only when applying the relabel operation, so let f and h be the preow and
labeling returned by a relabel(f , h , v) operation. By (7.25) there is a path P in
the residual graph Gf from v to s. Let |P | denote the number of edges in P , and
note that |P | n 1. The steepness condition implies that heights of the nodes
can decrease by at most 1 along each edge in P , and hence h(v) h(s) |P |,
which proves the statement.
Labels are monotone increasing throughout the algorithm, so this statement immediately implies a limit on the number of relabeling operations.
(7.27) Throughout the algorithm, each node is relabeled at most 2n 1 times,
and the total number of relabeling operations is less than 2n2.
Next we will bound the number of push operations. We will distinguish two
kinds of push operations. A push(f , h , v , w) operation is saturating if either
e = (v , w) is a forward edge in Ef and = ce f (e ), or (v , w) is a backward
edge with e = (w , v) and = f (e ). In other words, the push is saturating if,
after the push, the edge (v , w) is no longer in the residual graph. All other
push operations will be referred to as nonsaturating .
(7.28) Throughout the algorithm, the number of saturating push operations
is at most 2nm.
Proof. Consider an edge (v , w) in the residual graph. After a saturating
push(f , h , v , w) operation, we have h(v) = h(w) + 1, and the edge (v , w) is no
longer in the residual graph Gf , as shown in Figure 7.8. Before we can push
again along this edge, rst we have to push from w to v to make the edge
(v , w) appear in the residual graph. However, in order to push from w to v,
we rst need for ws label to increase by at least 2 (so that w is above v). The
label of w can increase by 2 at most n 1 times, so a saturating push from v
to w can occur at most n times. Each edge e E can give rise to two edges in
the residual graph, so overall we can have at most 2nm saturating pushes.
The hardest part of the analysis is proving a bound on the number of
nonsaturating pushes, and this also will be the bottleneck for the theoretical
bound on the running time.
(7.29) Throughout the algorithm, the number of nonsaturating push operations is at most 2n2m.
363
364
Chapter 7 Network Flow
Heights
The height of node w has to
increase by 2 before it can
push flow back to node v.
4
3
2
1
v
w
t
0
Nodes
Figure 7.8 After a saturating push(f , h , v , w), the height of v exceeds the height of w
by 1.
Proof. For this proof, we will use a so-called potential function method. For a
preow f and a compatible labeling h, we dene
(f , h ) =
h(v)
v:ef (v)>0
to be the sum of the heights of all nodes with positive excess. ( is often called
a potential since it resembles the potential energy of all nodes with positive
excess.)
In the initial preow and labeling, all nodes with positive excess are at
height 0, so (f , h) = 0. (f , h) remains nonnegative throughout the algorithm. A nonsaturating push(f , h , v , w) operation decreases (f , h) by at least
1, since after the push the node v will have no excess, and w, the only node
that gets new excess from the operation, is at a height 1 less than v. However, each saturating push and each relabel operation can increase (f , h).
A relabel operation increases (f , h) by exactly 1. There are at most 2n2
relabel operations, so the total increase in (f , h) due to relabel operations is 2n2. A saturating push(f , h , v , w) operation does not change labels,
but it can increase (f , h), since the node w may suddenly acquire positive
excess after the push. This would increase (f , h) by the height of w, which
is at most 2n 1. There are at most 2nm saturating push operations, so the
total increase in (f , h) due to push operations is at most 2mn(2n 1). So,
between the two causes, (f , h) can increase by at most 4mn2 during the
algorithm.
7.4 The Preow-Push Maximum-Flow Algorithm
But since remains nonnegative throughout, and it decreases by at least
1 on each nonsaturating push operation, it follows that there can be at most
4mn2 nonsaturating push operations.
Extensions: An Improved Version of the Algorithm
There has been a lot of work devoted to choosing node selection rules for
the Preow-Push Algorithm to improve the worst-case running time. Here we
consider a simple rule that leads to an improved O(n3) bound on the number
of nonsaturating push operations.
(7.30) If at each step we choose the node with excess at maximum height,
then the number of nonsaturating push operations throughout the algorithm is
at most 4n3.
Proof. Consider the maximum height H = maxv:ef (v)>0 h(v) of any node with
excess as the algorithm proceeds. The analysis will use this maximum height
H in place of the potential function in the previous O(n2m) bound.
This maximum height H can only increase due to relabeling (as ow
is always pushed to nodes at lower height), and so the total increase in H
throughout the algorithm is at most 2n2 by (7.26). H starts out 0 and remains
nonnegative, so the number of times H changes is at most 4n2.
Now consider the behavior of the algorithm over a phase of time in
which H remains constant. We claim that each node can have at most one
nonsaturating push operation during this phase. Indeed, during this phase,
ow is being pushed from nodes at height H to nodes at height H 1; and
after a nonsaturating push operation from v, it must receive ow from a node
at height H + 1 before we can push from it again.
Since there are at most n nonsaturating push operations between each
change to H , and H changes at most 4n2 times, the total number of nonsaturating push operations is at most 4n3.
As a follow-up to (7.30), it is interesting to note that experimentally the
computational bottleneck of the method is the number of relabeling operations,
and a better experimental running time is obtained by variants that work on
increasing labels faster than one by one. This is a point that we pursue further
in some of the exercises.
Implementing the Preow-Push Algorithm
Finally, we need to briey discuss how to implement this algorithm efciently.
Maintaining a few simple data structures will allow us to effectively implement
365
366
Chapter 7 Network Flow
the operations of the algorithm in constant time each, and overall to implement the algorithm in time O(mn) plus the number of nonsaturating push
operations. Hence the generic algorithm will run in O(mn2) time, while the
version that always selects the node at maximum height will run in O(n3) time.
We can maintain all nodes with excess on a simple list, and so we will be
able to select a node with excess in constant time. One has to be a bit more
careful to be able to select a node with maximum height H in constant time. In
order to do this, we will maintain a linked list of all nodes with excess at every
possible height. Note that whenever a node v gets relabeled, or continues to
have positive excess after a push, it remains a node with maximum height H .
Thus we only have to select a new node after a push when the current node v
no longer has positive excess. If node v was at height H , then the new node at
maximum height will also be at height H or, if no node at height H has excess,
then the maximum height will be H 1, since the previous push operation
out of v pushed ow to a node at height H 1.
Now assume we have selected a node v, and we need to select an edge
(v , w) on which to apply push(f , h , v , w) (or relabel(f , h , v) if no such w
exists). To be able to select an edge quickly, we will use the adjacency list
representation of the graph. More precisely, we will maintain, for each node v,
all possible edges leaving v in the residual graph (both forward and backward
edges) in a linked list, and with each edge we keep its capacity and ow value.
Note that this way we have two copies of each edge in our data structure: a
forward and a backward copy. These two copies will have pointers to each
other, so that updates done at one copy can be carried over to the other one
in O(1) time. We will select edges leaving a node v for push operations in the
order they appear on node vs list. To facilitate this selection, we will maintain
a pointer current(v) for each node v to the last edge on the list that has been
considered for a push operation. So, if node v no longer has excess after a
nonsaturating push operation out of node v, the pointer current(v) will stay
at this edge, and we will use the same edge for the next push operation out of
v. After a saturating push operation out of node v, we advance current(v) to
the next edge on the list.
The key observation is that, after advancing the pointer current(v) from
an edge (v , w), we will not want to apply push to this edge again until we
relabel v.
(7.31) After the current(v) pointer is advanced from an edge (v , w), we
cannot apply push to this edge until v gets relabeled.
Proof. At the moment current(v) is advanced from the edge (v , w), there is
some reason push cannot be applied to this edge. Either h(w) h(v), or the
7.5 A First Application: The Bipartite Matching Problem
edge is not in the residual graph. In the rst case, we clearly need to relabel v
before applying a push on this edge. In the latter case, one needs to apply push
to the reverse edge (w , v) to make (v , w) reenter the residual graph. However,
when we apply push to edge (w , v), then w is above v, and so v needs to be
relabeled before one can push ow from v to w again.
Since edges do not have to be considered again for push before relabeling,
we get the following.
(7.32) When the current(v) pointer reaches the end of the edge list for v,
the relabel operation can be applied to node v.
After relabeling node v, we reset current(v) to the rst edge on the list and
start considering edges again in the order they appear on vs list.
(7.33) The running time of the Preow-Push Algorithm, implemented using
the above data structures, is O(mn) plus O(1) for each nonsaturating push
operation. In particular, the generic Preow-Push Algorithm runs in O(n2m)
time, while the version where we always select the node at maximum height
runs in O(n3) time.
Proof. The initial ow and relabeling is set up in O(m) time. Both push and
relabel operations can be implemented in O(1) time, once the operation
has been selected. Consider a node v. We know that v can be relabeled at
most 2n times throughout the algorithm. We will consider the total time the
algorithm spends on nding the right edge on which to push ow out of node v,
between two times that node v gets relabeled. If node v has dv adjacent edges,
then by (7.32) we spend O(dv ) time on advancing the current(v) pointer
between consecutive relabelings of v. Thus the total time spent on advancing
the current pointers throughout the algorithm is O( vV ndv ) = O(mn), as
claimed.
7.5 A First Application: The Bipartite Matching
Problem
Having developed a set of powerful algorithms for the Maximum-Flow Problem, we now turn to the task of developing applications of maximum ows
and minimum cuts in graphs. We begin with two very basic applications. First,
in this section, we discuss the Bipartite Matching Problem mentioned at the
beginning of this chapter. In the next section, we discuss the more general
Disjoint Paths Problem.
367
368
Chapter 7 Network Flow
The Problem
One of our original goals in developing the Maximum-Flow Problem was to
be able to solve the Bipartite Matching Problem, and we now show how to
do this. Recall that a bipartite graph G = (V , E ) is an undirected graph whose
node set can be partitioned as V = X Y , with the property that every edge
e E has one end in X and the other end in Y . A matching M in G is a subset
of the edges M E such that each node appears in at most one edge in M .
The Bipartite Matching Problem is that of nding a matching in G of largest
possible size.
Designing the Algorithm
The graph dening a matching problem is undirected, while ow networks are
directed; but it is actually not difcult to use an algorithm for the MaximumFlow Problem to nd a maximum matching.
Beginning with the graph G in an instance of the Bipartite Matching
Problem, we construct a ow network G as shown in Figure 7.9. First we
direct all edges in G from X to Y . We then add a node s, and an edge (s , x)
from s to each node in X . We add a node t , and an edge (y , t ) from each node
in Y to t . Finally, we give each edge in G a capacity of 1.
We now compute a maximum s-t ow in this network G . We will discover
that the value of this maximum is equal to the size of the maximum matching
in G . Moreover, our analysis will show how one can use the ow itself to
recover the matching.
s
(a)
t
(b)
Figure 7.9 (a) A bipartite graph. (b) The corresponding flow network, with all capacities
equal to 1.
7.5 A First Application: The Bipartite Matching Problem
Analyzing the Algorithm
The analysis is based on showing that integer-valued ows in G encode
matchings in G in a fairly transparent fashion. First, suppose there is a
matching in G consisting of k edges (xi1 , yi1) , . . . , (xik , yik ). Then consider the
ow f that sends one unit along each path of the form s , xij , yij , t that is,
f (e ) = 1 for each edge on one of these paths. One can verify easily that the
capacity and conservation conditions are indeed met and that f is an s-t ow
of value k.
Conversely, suppose there is a ow f in G of value k. By the integrality
theorem for maximum ows (7.14), we know there is an integer-valued ow f
of value k; and since all capacities are 1, this means that f (e ) is equal to either
0 or 1 for each edge e . Now, consider the set M of edges of the form (x , y) on
which the ow value is 1.
Here are three simple facts about the set M .
(7.34)
M contains k edges.
Proof. To prove this, consider the cut (A, B) in G with A = {s} X . The value
of the ow is the total ow leaving A, minus the total ow entering A. The
rst of these terms is simply the cardinality of M , since these are the edges
leaving A that carry ow, and each carries exactly one unit of ow. The second
of these terms is 0, since there are no edges entering A. Thus, M contains k
edges.
(7.35)
Each node in X is the tail of at most one edge in M .
Proof. To prove this, suppose x X were the tail of at least two edges in M .
Since our ow is integer-valued, this means that at least two units of ow
leave from x. By conservation of ow, at least two units of ow would have
to come into xbut this is not possible, since only a single edge of capacity 1
enters x. Thus x is the tail of at most one edge in M .
By the same reasoning, we can show
(7.36)
Each node in Y is the head of at most one edge in M .
Combining these facts, we see that if we view M as a set of edges in the
original bipartite graph G , we get a matching of size k. In summary, we have
proved the following fact.
(7.37) The size of the maximum matching in G is equal to the value of the
maximum ow in G ; and the edges in such a matching in G are the edges that
carry ow from X to Y in G .
369
370
Chapter 7 Network Flow
Note the crucial way in which the integrality theorem (7.14) gured in
this construction: we needed to know if there is a maximum ow in G that
takes only the values 0 and 1.
Bounding the Running Time Now lets consider how quickly we can compute a maximum matching in G . Let n = |X | = |Y |, and let m be the number
of edges of G . Well tacitly assume that there is at least one edge incident to
each node in the original problem, and hence m n/2. The time to compute
a maximum matching is dominated by the time to compute an integer-valued
maximum ow in G , since converting this to a matching in G is simple. For
this ow problem, we have that C = e out of s ce = |X | = n, as s has an edge
of capacity 1 to each node of X . Thus, by using the O(mC ) bound in (7.5), we
get the following.
(7.38) The Ford-Fulkerson Algorithm can be used to nd a maximum matching in a bipartite graph in O(mn) time.
Its interesting that if we were to use the better bounds of O(m2 log2 C ) or
wed get the inferior running
times of O(m2 log n) or O(n3) for this problem. There is nothing contradictory
in this. These bounds were designed to be good for all instances, even when C
is very large relative to m and n. But C = n for the Bipartite Matching Problem,
and so the cost of this extra sophistication is not needed.
O(n3) that we developed in the previous sections,
It is worthwhile to consider what the augmenting paths mean in the
network G . Consider the matching M consisting of edges (x2 , y2), (x3, y3),
and (x5 , y5) in the bipartite graph in Figure 7.1; see also Figure 7.10. Let f
be the corresponding ow in G . This matching is not maximum, so f is not
a maximum s-t ow, and hence there is an augmenting path in the residual
graph Gf . One such augmenting path is marked in Figure 7.10(b). Note that
the edges (x2 , y2) and (x3, y3) are used backward, and all other edges are used
forward. All augmenting paths must alternate between edges used backward
and forward, as all edges of the graph G go from X to Y . Augmenting paths
are therefore also called alternating paths in the context of nding a maximum
matching. The effect of this augmentation is to take the edges used backward
out of the matching, and replace them with the edges going forward. Because
the augmenting path goes from s to t , there is one more forward edge than
backward edge; thus the size of the matching increases by one.
7.5 A First Application: The Bipartite Matching Problem
x1
y1
x1
y1
x1
y1
x2
y2
x2
y2
x2
y2
x3
y3
x3
y3
x3
y3
x4
y4
x4
y4
x4
y4
x5
y5
x5
y5
x5
y5
(a)
(b)
(c)
Figure 7.10 (a) A bipartite graph, with a matching M . (b) The augmenting path in the
corresponding residual graph. (c) The matching obtained by the augmentation.
Extensions: The Structure of Bipartite Graphs with
No Perfect Matching
Algorithmically, weve seen how to nd perfect matchings: We use the algorithm above to nd a maximum matching and then check to see if this matching
is perfect.
But lets ask a slightly less algorithmic question. Not all bipartite graphs
have perfect matchings. What does a bipartite graph without a perfect matching look like? Is there an easy way to see that a bipartite graph does not have a
perfect matchingor at least an easy way to convince someone the graph has
no perfect matching, after we run the algorithm? More concretely, it would be
nice if the algorithm, upon concluding that there is no perfect matching, could
produce a short certicate of this fact. The certicate could allow someone
to be quickly convinced that there is no perfect matching, without having to
look over a trace of the entire execution of the algorithm.
One way to understand the idea of such a certicate is as follows. We can
decide if the graph G has a perfect matching by checking if the maximum ow
in a related graph G has value at least n. By the Max-Flow Min-Cut Theorem,
there will be an s-t cut of capacity less than n if the maximum-ow value in
G has value less than n. So, in a way, a cut with capacity less than n provides
such a certicate. However, we want a certicate that has a natural meaning
in terms of the original graph G .
What might such a certicate look like? For example, if there are nodes
x1, x2 X that have only one incident edge each, and the other end of each
edge is the same node y, then clearly the graph has no perfect matching: both
x1 and x2 would need to get matched to the same node y. More generally,
consider a subset of nodes A X , and let (A) Y denote the set of all nodes
371
372
Chapter 7 Network Flow
that are adjacent to nodes in A. If the graph has a perfect matching, then each
node in A has to be matched to a different node in (A), so (A) has to be at
least as large as A. This gives us the following fact.
(7.39) If a bipartite graph G = (V , E ) with two sides X and Y has a perfect
matching, then for all A X we must have | (A)| |A|.
This statement suggests a type of certicate demonstrating that a graph
does not have a perfect matching: a set A X such that | (A)| < |A|. But is the
converse of (7.39) also true? Is it the case that whenever there is no perfect
matching, there is a set A like this that proves it? The answer turns out to
be yes, provided we add the obvious condition that |X | = |Y | (without which
there could certainly not be a perfect matching). This statement is known
in the literature as Halls Theorem, though versions of it were discovered
independently by a number of different peopleperhaps rst by K nigin
o
the early 1900s. The proof of the statement also provides a way to nd such a
subset A in polynomial time.
(7.40) Assume that the bipartite graph G = (V , E ) has two sides X and Y
such that |X | = |Y |. Then the graph G either has a perfect matching or there is
a subset A X such that | (A)| < |A|. A perfect matching or an appropriate
subset A can be found in O(mn) time.
Proof. We will use the same graph G as in (7.37). Assume that |X | = |Y | = n.
By (7.37) the graph G has a maximum matching if and only if the value of the
maximum ow in G is n.
We need to show that if the value of the maximum ow is less than n,
then there is a subset A such that | (A)| < |A|, as claimed in the statement.
By the Max-Flow Min-Cut Theorem (7.12), if the maximum-ow value is less
than n, then there is a cut (A , B ) with capacity less than n in G . Now the
set A contains s, and may contain nodes from both X and Y as shown in
Figure 7.11. We claim that the set A = X A has the claimed property. This
will prove both parts of the statement, as weve seen in (7.11) that a minimum
cut (A , B ) can also be found by running the Ford-Fulkerson Algorithm.
First we claim that one can modify the minimum cut (A , B ) so as to
ensure that (A) A , where A = X A as before. To do this, consider a node
y (A) that belongs to B as shown in Figure 7.11(a). We claim that by moving
y from B to A , we do not increase the capacity of the cut. For what happens
when we move y from B to A ? The edge (y , t ) now crosses the cut, increasing
the capacity by one. But previously there was at least one edge (x , y) with
x A, since y (A); all edges from A and y used to cross the cut, and dont
anymore. Thus, overall, the capacity of the cut cannot increase. (Note that we
373
7.6 Disjoint Paths in Directed and Undirected Graphs
Node y can be moved
to the s-side of the cut.
A
A
A
s
x
y
t
s
(a)
x
y
(b)
Figure 7.11 (a) A minimum cut in proof of (7.40). (b) The same cut after moving node
y to the A side. The edges crossing the cut are dark.
dont have to be concerned about nodes x X that are not in A. The two ends
of the edge (x , y) will be on different sides of the cut, but this edge does not
add to the capacity of the cut, as it goes from B to A .)
Next consider the capacity of this minimum cut (A , B ) that has (A) A
as shown in Figure 7.11(b). Since all neighbors of A belong to A , we see that
the only edges out of A are either edges that leave the source s or that enter
the sink t . Thus the capacity of the cut is exactly
c (A , B ) = |X B | + |Y A |.
Notice that |X B | = n |A|, and |Y A | | (A)|. Now the assumption that
c (A , B ) < n implies that
n |A| + | (A)| |X B | + |Y A | = c (A , B ) < n .
Comparing the rst and the last terms, we get the claimed inequality |A| >
| (A)|.
7.6 Disjoint Paths in Directed and
Undirected Graphs
In Section 7.1, we described a ow f as a kind of trafc in the network.
But our actual denition of a ow has a much more static feel to it: For each
edge e , we simply specify a number f (e ) saying the amount of ow crossing e .
Lets see if we can revive the more dynamic, trafc-oriented picture a bit, and
try formalizing the sense in which units of ow travel from the source to
t
374
Chapter 7 Network Flow
the sink. From this more dynamic view of ows, we will arrive at something
called the s-t Disjoint Paths Problem.
The Problem
In dening this problem precisely, we will deal with two issues. First, we will
make precise this intuitive correspondence between units of ow traveling
along paths, and the notion of ow weve studied so far. Second, we will
extend the Disjoint Paths Problem to undirected graphs. Well see that, despite
the fact that the Maximum-Flow Problem was dened for a directed graph, it
can naturally be used also to handle related problems on undirected graphs.
We say that a set of paths is edge-disjoint if their edge sets are disjoint, that
is, no two paths share an edge, though multiple paths may go through some
of the same nodes. Given a directed graph G = (V , E ) with two distinguished
nodes s , t V , the Directed Edge-Disjoint Paths Problem is to nd the maximum
number of edge-disjoint s-t paths in G . The Undirected Edge-Disjoint Paths
Problem is to nd the maximum number of edge-disjoint s-t paths in an
undirected graph G . The related question of nding paths that are not only
edge-disjoint, but also node-disjoint (of course, other than at nodes s and t )
will be considered in the exercises to this chapter.
Designing the Algorithm
Both the directed and the undirected versions of the problem can be solved
very naturally using ows. Lets start with the directed problem. Given the
graph G = (V , E ), with its two distinguished nodes s and t , we dene a ow
network in which s and t are the source and sink, respectively, and with a
capacity of 1 on each edge. Now suppose there are k edge-disjoint s-t paths.
We can make each of these paths carry one unit of ow: We set the ow to be
f (e ) = 1 for each edge e on any of the paths, and f (e ) = 0 on all other edges,
and this denes a feasible ow of value k.
(7.41) If there are k edge-disjoint paths in a directed graph G from s to t, then
the value of the maximum s-t ow in G is at least k.
Suppose we could show the converse to (7.41) as well: If there is a ow
of value k, then there exist k edge-disjoint s-t paths. Then we could simply
compute a maximum s-t ow in G and declare (correctly) this to be the
maximum number of edge-disjoint s-t paths.
We now proceed to prove this converse statement, conrming that this
approach using ow indeed gives us the correct answer. Our analysis will
also provide a way to extract k edge-disjoint paths from an integer-valued
ow sending k units from s to t . Thus computing a maximum ow in G will
7.6 Disjoint Paths in Directed and Undirected Graphs
not only give us the maximum number of edge-disjoint paths, but the paths
as well.
Analyzing the Algorithm
Proving the converse direction of (7.41) is the heart of the analysis, since it
will immediately establish the optimality of the ow-based algorithm to nd
disjoint paths.
To prove this, we will consider a ow of value at least k, and construct k
edge-disjoint paths. By (7.14), we know that there is a maximum ow f with
integer ow values. Since all edges have a capacity bound of 1, and the ow
is integer-valued, each edge that carries ow under f has exactly one unit of
ow on it. Thus we just need to show the following.
(7.42) If f is a 0-1 valued ow of value , then the set of edges with ow
value f (e ) = 1 contains a set of edge-disjoint paths.
Proof. We prove this by induction on the number of edges in f that carry ow.
If = 0, there is nothing to prove. Otherwise, there must be an edge (s , u) that
carries one unit of ow. We now trace out a path of edges that must also
carry ow: Since (s , u) carries a unit of ow, it follows by conservation that
there is some edge (u , v) that carries one unit of ow, and then there must be
an edge (v , w) that carries one unit of ow, and so forth. If we continue in this
way, one of two things will eventually happen: Either we will reach t , or we
will reach a node v for the second time.
If the rst case happenswe nd a path P from s to t then well use this
path as one of our paths. Let f be the ow obtained by decreasing the ow
values on the edges along P to 0. This new ow f has value 1, and it has
fewer edges that carry ow. Applying the induction hypothesis for f , we get
1 edge-disjoint paths, which, along with path P , form the paths claimed.
If P reaches a node v for the second time, then we have a situation like
the one pictured in Figure 7.12. (The edges in the gure all carry one unit of
ow, and the dashed edges indicate the path traversed so far, which has just
reached a node v for the second time.) In this case, we can make progress in
a different way.
Consider the cycle C of edges visited between the rst and second appearances of v. We obtain a new ow f from f by decreasing the ow values on
the edges along C to 0. This new ow f has value , but it has fewer edges that
carry ow. Applying the induction hypothesis for f , we get the edge-disjoint
paths as claimed.
375
376
Chapter 7 Network Flow
Flow around a cycle
can be zeroed out.
v
P
t
s
Figure 7.12 The edges in the figure all carry one unit of flow. The path P of dashed
edges is one possible path in the proof of (7.42).
We can summarize (7.41) and (7.42) in the following result.
(7.43) There are k edge-disjoint paths in a directed graph G from s to t if and
only if the value of the maximum value of an s-t ow in G is at least k.
Notice also how the proof of (7.42) provides an actual procedure for
constructing the k paths, given an integer-valued maximum ow in G . This
procedure is sometimes referred to as a path decomposition of the ow, since it
decomposes the ow into a constituent set of paths. Hence we have shown
that our ow-based algorithm nds the maximum number of edge-disjoint s-t
paths and also gives us a way to construct the actual paths.
Bounding the Running Time For this ow problem, C = e out of s ce
|V | = n, as there are at most |V | edges out of s, each of which has capacity 1. Thus, by using the O(mC ) bound in (7.5), we get an integer maximum
ow in O(mn) time.
The path decomposition procedure in the proof of (7.42), which produces
the paths themselves, can also be made to run in O(mn) time. To see this, note
that this procedure, with a little care, can produce a single path from s to t
using at most constant work per edge in the graph, and hence in O(m) time.
Since there can be at most n 1 edge-disjoint paths from s to t (each must
use a different edge out of s), it therefore takes time O(mn) to produce all the
paths.
In summary, we have shown
(7.44) The Ford-Fulkerson Algorithm can be used to nd a maximum set of
edge-disjoint s-t paths in a directed graph G in O(mn) time.
A Version of the Max-Flow Min-Cut Theorem for Disjoint Paths The MaxFlow Min-Cut Theorem (7.13) can be used to give the following characteri-
7.6 Disjoint Paths in Directed and Undirected Graphs
zation of the maximum number of edge-disjoint s-t paths. We say that a set
F E of edges separates s from t if, after removing the edges F from the graph
G , no s-t paths remain in the graph.
(7.45) In every directed graph with nodes s and t, the maximum number of
edge-disjoint s-t paths is equal to the minimum number of edges whose removal
separates s from t.
Proof. If the removal of a set F E of edges separates s from t , then each s-t
path must use at least one edge from F , and hence the number of edge-disjoint
s-t paths is at most |F |.
To prove the other direction, we will use the Max-Flow Min-Cut Theorem
(7.13). By (7.43) the maximum number of edge-disjoint paths is the value
of the maximum s-t ow. Now (7.13) states that there is an s-t cut (A, B) with
capacity . Let F be the set of edges that go from A to B. Each edge has capacity
1, so |F | = and, by the denition of an s-t cut, removing these edges from
G separates s from t .
This result, then, can be viewed as the natural special case of the MaxFlow Min-Cut Theorem in which all edge capacities are equal to 1. In fact,
this special case was proved by Menger in 1927, much before the full MaxFlow Min-Cut Theorem was formulated and proved; for this reason, (7.45)
is often called Mengers Theorem. If we think about it, the proof of Halls
Theorem (7.40) for bipartite matchings involves a reduction to a graph with
unit-capacity edges, and so it can be proved using Mengers Theorem rather
than the general Max-Flow Min-Cut Theorem. In other words, Halls Theorem
is really a special case of Mengers Theorem, which in turn is a special case
of the Max-Flow Min-Cut Theorem. And the history follows this progression,
since they were discovered in this order, a few decades apart.2
Extensions: Disjoint Paths in Undirected Graphs
Finally, we consider the disjoint paths problem in an undirected graph G .
Despite the fact that our graph G is now undirected, we can use the maximumow algorithm to obtain edge-disjoint paths in G . The idea is quite simple: We
replace each undirected edge (u , v) in G by two directed edges (u , v) and
2
In fact, in an interesting retrospective written in 1981, Menger relates his version of the story of how
he rst explained his theorem to Knig, one of the independent discoverers of Halls Theorem. You
o
might think that Knig, having thought a lot about these problems, would have immediately grasped
o
why Mengers generalization of his theorem was true, and perhaps even considered it obvious. But, in
fact, the opposite happened; Knig didnt believe it could be right and stayed up all night searching
o
for a counterexample. The next day, exhausted, he sought out Menger and asked him for the proof.
377
378
Chapter 7 Network Flow
(v , u), and in this way create a directed version G of G . (We may delete the
edges into s and out of t , since they are not useful.) Now we want to use the
Ford-Fulkerson Algorithm in the resulting directed graph. However, there is an
important issue we need to deal with rst. Notice that two paths P1 and P2 may
be edge-disjoint in the directed graph and yet share an edge in the undirected
graph G : This happens if P1 uses directed edge (u , v) while P2 uses edge (v , u).
However, it is not hard to see that there always exists a maximum ow in any
network that uses at most one out of each pair of oppositely directed edges.
(7.46) In any ow network, there is a maximum ow f where for all opposite
directed edges e = (u , v) and e = (v , u), either f (e ) = 0 or f (e ) = 0. If the
capacities of the ow network are integral, then there also is such an integral
maximum ow.
Proof. We consider any maximum ow f , and we modify it to satisfy the
claimed condition. Assume e = (u , v) and e = (v , u) are opposite directed
edges, and f (e ) = 0, f (e ) = 0. Let be the smaller of these values, and modify
f by decreasing the ow value on both e and e by . The resulting ow f is
feasible, has the same value as f , and its value on one of e and e is 0.
Now we can use the Ford-Fulkerson Algorithm and the path decomposition
procedure from (7.42) to obtain edge-disjoint paths in the undirected graph G .
(7.47) There are k edge-disjoint paths in an undirected graph G from s to t
if and only if the maximum value of an s-t ow in the directed version G of G
is at least k. Furthermore, the Ford-Fulkerson Algorithm can be used to nd a
maximum set of disjoint s-t paths in an undirected graph G in O(mn) time.
The undirected analogue of (7.45) is also true, as in any s-t cut, at most
one of the two oppositely directed edges can cross from the s-side to the t side of the cut (for if one crosses, then the other must go from the t -side to
the s-side).
(7.48) In every undirected graph with nodes s and t, the maximum number of
edge-disjoint s-t paths is equal to the minimum number of edges whose removal
separates s from t.
7.7 Extensions to the Maximum-Flow Problem
Much of the power of the Maximum-Flow Problem has essentially nothing to
do with the fact that it models trafc in a network. Rather, it lies in the fact
that many problems with a nontrivial combinatorial search component can
7.7 Extensions to the Maximum-Flow Problem
be solved in polynomial time because they can be reduced to the problem of
nding a maximum ow or a minimum cut in a directed graph.
Bipartite Matching is a natural rst application in this vein; in the coming
sections, we investigate a range of further applications. To begin with, we
stay with the picture of ow as an abstract kind of trafc, and look for
more general conditions we might impose on this trafc. These more general
conditions will turn out to be useful for some of our further applications.
In particular, we focus on two generalizations of maximum ow. We will
see that both can be reduced to the basic Maximum-Flow Problem.
The Problem: Circulations with Demands
One simplifying aspect of our initial formulation of the Maximum-Flow Problem is that we had only a single source s and a single sink t . Now suppose
that there can be a set S of sources generating ow, and a set T of sinks that
can absorb ow. As before, there is an integer capacity on each edge.
With multiple sources and sinks, it is a bit unclear how to decide which
source or sink to favor in a maximization problem. So instead of maximizing
the ow value, we will consider a problem where sources have xed supply
values and sinks have xed demand values, and our goal is to ship ow
from nodes with available supply to those with given demands. Imagine, for
example, that the network represents a system of highways or railway lines in
which we want to ship products from factories (which have supply) to retail
outlets (which have demand). In this type of problem, we will not be seeking to
maximize a particular value; rather, we simply want to satisfy all the demand
using the available supply.
Thus we are given a ow network G = (V , E ) with capacities on the edges.
Now, associated with each node v V is a demand dv . If dv > 0, this indicates
that the node v has a demand of dv for ow; the node is a sink, and it wishes
to receive dv units more ow than it sends out. If dv < 0, this indicates that v
has a supply of dv ; the node is a source, and it wishes to send out dv units
more ow than it receives. If dv = 0, then the node v is neither a source nor a
sink. We will assume that all capacities and demands are integers.
We use S to denote the set of all nodes with negative demand and T to
denote the set of all nodes with positive demand. Although a node v in S wants
to send out more ow than it receives, it will be okay for it to have ow that
enters on incoming edges; it should just be more than compensated by the ow
that leaves v on outgoing edges. The same applies (in the opposite direction)
to the set T .
379
380
Chapter 7 Network Flow
3
3
3
3
s*
1
1
2
3
3
3
3
3
3
3
2
2
2
3
2
2
2
2
2
2
2
2
2
2
2
2
2
2
4
4
t*
4
(a)
4
(b)
Figure 7.13 (a) An instance of the Circulation Problem together with a solution: Numbers
inside the nodes are demands; numbers labeling the edges are capacities and flow
values, with the flow values inside boxes. (b) The result of reducing this instance to an
equivalent instance of the Maximum-Flow Problem.
In this setting, we say that a circulation with demands {dv } is a function f
that assigns a nonnegative real number to each edge and satises the following
two conditions.
(i) (Capacity conditions) For each e E , we have 0 f (e ) ce .
(ii) (Demand conditions) For each v V , we have v, f in (v) f out (v) = dv .
Now, instead of considering a maximization problem, we are concerned with
a feasibility problem: We want to know whether there exists a circulation that
meets conditions (i) and (ii).
For example, consider the instance in Figure 7.13(a). Two of the nodes
are sources, with demands 3 and 3; and two of the nodes are sinks, with
demands 2 and 4. The ow values in the gure constitute a feasible circulation,
indicating how all demands can be satised while respecting the capacities.
If we consider an arbitrary instance of the Circulation Problem, here is a
simple condition that must hold in order for a feasible circulation to exist: The
total supply must equal the total demand.
(7.49)
If there exists a feasible circulation with demands {dv }, then
v
dv = 0.
Proof. Suppose there exists a feasible circulation f in this setting. Then
in
out (v ). Now, in this latter expression, the value f (e ) for
v dv =
v f (v ) f
each edge e = (u , v) is counted exactly twice: once in f out (u) and once in f in (v).
These two terms cancel out; and since this holds for all values f (e ), the overall
sum is 0.
7.7 Extensions to the Maximum-Flow Problem
s* supplies sources
with flow.
t* siphons flow
out of sinks.
T
S
u
v
s*
t*
Figure 7.14 Reducing the Circulation Problem to the Maximum-Flow Problem.
Thanks to (7.49), we know that
dv =
v:dv >0
dv .
v:dv <0
Let D denote this common value.
Designing and Analyzing an Algorithm for Circulations
It turns out that we can reduce the problem of nding a feasible circulation
with demands {dv } to the problem of nding a maximum s-t ow in a different
network, as shown in Figure 7.14.
The reduction looks very much like the one we used for Bipartite Matching:
we attach a super-source s to each node in S, and a super-sink t to each
node in T . More specically, we create a graph G from G by adding new nodes
s and t to G . For each node v T that is, each node v with dv > 0we add
an edge (v , t ) with capacity dv . For each node u Sthat is, each node with
du < 0we add an edge (s , u) with capacity du . We carry the remaining
structure of G over to G unchanged.
In this graph G , we will be seeking a maximum s-t ow. Intuitively,
we can think of this reduction as introducing a node s that supplies all the
sources with their extra ow, and a node t that siphons the extra ow out
of the sinks. For example, part (b) of Figure 7.13 shows the result of applying
this reduction to the instance in part (a).
Note that there cannot be an s-t ow in G of value greater than D, since
the cut (A, B) with A = {s} only has capacity D. Now, if there is a feasible
circulation f with demands {dv } in G , then by sending a ow value of dv on
each edge (s , v), and a ow value of dv on each edge (v , t ), we obtain an st ow in G of value D, and so this is a maximum ow. Conversely, suppose
there is a (maximum) s-t ow in G of value D. It must be that every edge
381
382
Chapter 7 Network Flow
out of s, and every edge into t , is completely saturated with ow. Thus, if
we delete these edges, we obtain a circulation f in G with f in (v) f out (v) = dv
for each node v. Further, if there is a ow of value D in G , then there is such
a ow that takes integer values.
In summary, we have proved the following.
(7.50) There is a feasible circulation with demands {dv } in G if and only if the
maximum s-t ow in G has value D. If all capacities and demands in G are
integers, and there is a feasible circulation, then there is a feasible circulation
that is integer-valued.
At the end of Section 7.5, we used the Max-Flow Min-Cut Theorem to
derive the characterization (7.40) of bipartite graphs that do not have perfect
matchings. We can give an analogous characterization for graphs that do not
have a feasible circulation. The characterization uses the notion of a cut ,
adapted to the present setting. In the context of circulation problems with
demands, a cut (A, B) is any partition of the node set V into two sets, with no
restriction on which side of the partition the sources and sinks fall. We include
the characterization here without a proof.
(7.51) The graph G has a feasible circulation with demands {dv } if and only
if for all cuts (A, B),
dv c (A, B).
vB
It is important to note that our network has only a single kind of ow.
Although the ow is supplied from multiple sources, and absorbed at multiple
sinks, we cannot place restrictions on which source will supply the ow to
which sink; we have to let our algorithm decide this. A harder problem is
the Multicommodity Flow Problem; here sink ti must be supplied with ow
that originated at source si , for each i. We will discuss this issue further in
Chapter 11.
The Problem: Circulations with Demands and
Lower Bounds
Finally, let us generalize the previous problem a little. In many applications, we
not only want to satisfy demands at various nodes; we also want to force the
ow to make use of certain edges. This can be enforced by placing lower bounds
on edges, as well as the usual upper bounds imposed by edge capacities.
Consider a ow network G = (V , E ) with a capacity ce and a lower bound
on each edge e . We will assume 0 e ce for each e . As before, each node
v will also have a demand dv , which can be either positive or negative. We
will assume that all demands, capacities, and lower bounds are integers.
e
7.7 Extensions to the Maximum-Flow Problem
The given quantities have the same meaning as before, and now a lower
bound e means that the ow value on e must be at least e . Thus a circulation
in our ow network must satisfy the following two conditions.
(i) (Capacity conditions) For each e E , we have
(ii) (Demand conditions) For every v V , we have
f (e ) ce .
in (v ) f out (v ) = d
f
e
v.
As before, we wish to decide whether there exists a feasible circulationone
that satises these conditions.
Designing and Analyzing an Algorithm with
Lower Bounds
Our strategy will be to reduce this to the problem of nding a circulation
with demands but no lower bounds. (Weve seen that this latter problem, in
turn, can be reduced to the standard Maximum-Flow Problem.) The idea is
as follows. We know that on each edge e , we need to send at least e units of
ow. So suppose that we dene an initial circulation f0 simply by f0(e ) = e .
f0 satises all the capacity conditions (both lower and upper bounds); but it
presumably does not satisfy all the demand conditions. In particular,
f0in (v) f0out (v) =
e
e into v
e.
e out of v
Let us denote this quantity by Lv . If Lv = dv , then we have satised the
demand condition at v; but if not, then we need to superimpose a circulation
f1 on top of f0 that will clear the remaining imbalance at v. So we need
f1in (v) f1out (v) = dv Lv . And how much capacity do we have with which to
do this? Having already sent e units of ow on each edge e , we have ce e
more units to work with.
These considerations directly motivate the following construction. Let the
graph G have the same nodes and edges, with capacities and demands, but
no lower bounds. The capacity of edge e will be ce e . The demand of node
v will be dv Lv .
For example, consider the instance in Figure 7.15(a). This is the same as
the instance we saw in Figure 7.13, except that we have now given one of the
edges a lower bound of 2. In part (b) of the gure, we eliminate this lower
bound by sending two units of ow across the edge. This reduces the upper
bound on the edge and changes the demands at the two ends of the edge. In
the process, it becomes clear that there is no feasible circulation, since after
applying the construction there is a node with a demand of 5, and a total of
only four units of capacity on its outgoing edges.
We now claim that our general construction produces an equivalent instance with demands but no lower bounds; we can therefore use our algorithm
for this latter problem.
383
384
Chapter 7 Network Flow
Eliminating a lower
bound from an edge
Lower bound of 2
3
3
1
3
1
2
3
2
2
2
5
2
3
2
2
2
4
4
(a)
(b)
Figure 7.15 (a) An instance of the Circulation Problem with lower bounds: Numbers
inside the nodes are demands, and numbers labeling the edges are capacities. We also
assign a lower bound of 2 to one of the edges. (b) The result of reducing this instance
to an equivalent instance of the Circulation Problem without lower bounds.
(7.52) There is a feasible circulation in G if and only if there is a feasible
circulation in G . If all demands, capacities, and lower bounds in G are integers,
and there is a feasible circulation, then there is a feasible circulation that is
integer-valued.
Proof. First suppose there is a circulation f in G . Dene a circulation f in G
by f (e ) = f (e ) + e . Then f satises the capacity conditions in G , and
f in (v) f out (v) =
(
e
+ f (e ))
e into v
(
e
+ f (e )) = Lv + (dv Lv ) = dv ,
e out of v
so it satises the demand conditions in G as well.
Conversely, suppose there is a circulation f in G , and dene a circulation
f in G by f (e ) = f (e ) e . Then f satises the capacity conditions in G , and
in
out
(f ) (v) (f )
(v) =
(f (e )
e into v
e)
(f (e )
e ) = dv
Lv ,
e out of v
so it satises the demand conditions in G as well.
7.8 Survey Design
Many problems that arise in applications can, in fact, be solved efciently by
a reduction to Maximum Flow, but it is often difcult to discover when such
a reduction is possible. In the next few sections, we give several paradigmatic
examples of such problems. The goal is to indicate what such reductions tend
7.8 Survey Design
to look like and to illustrate some of the most common uses of ows and cuts
in the design of efcient combinatorial algorithms. One point that will emerge
is the following: Sometimes the solution one wants involves the computation
of a maximum ow, and sometimes it involves the computation of a minimum
cut; both ows and cuts are very useful algorithmic tools.
We begin with a basic application that we call survey design, a simple
version of a task faced by many companies wanting to measure customer
satisfaction. More generally, the problem illustrates how the construction used
to solve the Bipartite Matching Problem arises naturally in any setting where
we want to carefully balance decisions across a set of optionsin this case,
designing questionnaires by balancing relevant questions across a population
of consumers.
The Problem
A major issue in the burgeoning eld of data mining is the study of consumer
preference patterns. Consider a company that sells k products and has a
database containing the purchase histories of a large number of customers.
(Those of you with Shoppers Club cards may be able to guess how this data
gets collected.) The company wishes to conduct a survey, sending customized
questionnaires to a particular group of n of its customers, to try determining
which products people like overall.
Here are the guidelines for designing the survey.
.
Each customer will receive questions about a certain subset of the
products.
.
A customer can only be asked about products that he or she has purchased.
.
To make each questionnaire informative, but not too long so as to discourage participation, each customer i should be asked about a number
of products between ci and ci .
.
Finally, to collect sufcient data about each product, there must be
between pj and pj distinct customers asked about each product j .
More formally, the input to the Survey Design Problem consists of a bipartite
graph G whose nodes are the customers and the products, and there is an edge
between customer i and product j if he or she has ever purchased product j .
Further, for each customer i = 1, . . . , n, we have limits ci ci on the number
of products he or she can be asked about; for each product j = 1, . . . , k, we
have limits pj pj on the number of distinct customers that have to be asked
about it. The problem is to decide if there is a way to design a questionnaire
for each customer so as to satisfy all these conditions.
385
386
Chapter 7 Network Flow
Customers
i
c i ,c i
s
Products
0,1
j
p j, p j
t
Figure 7.16 The Survey Design Problem can be reduced to the problem of finding a
feasible circulation: Flow passes from customers (with capacity bounds indicating how
many questions they can be asked) to products (with capacity bounds indicating how
many questions should be asked about each product).
Designing the Algorithm
We will solve this problem by reducing it to a circulation problem on a ow
network G with demands and lower bounds as shown in Figure 7.16. To obtain
the graph G from G , we orient the edges of G from customers to products, add
nodes s and t with edges (s , i) for each customer i = 1, . . . , n, edges (j , t ) for
each product j = 1, . . . , k, and an edge (t , s). The circulation in this network
will correspond to the way in which questions are asked. The ow on the edge
(s , i) is the number of products included on the questionnaire for customer i,
so this edge will have a capacity of ci and a lower bound of ci . The ow on the
edge (j , t ) will correspond to the number of customers who were asked about
product j , so this edge will have a capacity of pj and a lower bound of pj . Each
edge (i , j ) going from a customer to a product he or she bought has capacity
1, and 0 as the lower bound. The ow carried by the edge (t , s) corresponds
to the overall number of questions asked. We can give this edge a capacity of
i ci and a lower bound of
i ci . All nodes have demand 0.
Our algorithm is simply to construct this network G and check whether
it has a feasible circulation. We now formulate a claim that establishes the
correctness of this algorithm.
Analyzing the Algorithm
(7.53) The graph G just constructed has a feasible circulation if and only if
there is a feasible way to design the survey.
7.9 Airline Scheduling
Proof. The construction above immediately suggests a way to turn a survey
design into the corresponding ow. The edge (i , j ) will carry one unit of ow
if customer i is asked about product j in the survey, and will carry no ow
otherwise. The ow on the edges (s , i) is the number of questions asked
from customer i, the ow on the edge (j , t ) is the number of customers who
were asked about product j , and nally, the ow on edge (t , s) is the overall
number of questions asked. This ow satises the 0 demand, that is, there is
ow conservation at every node. If the survey satises these rules, then the
corresponding ow satises the capacities and lower bounds.
Conversely, if the Circulation Problem is feasible, then by (7.52) there
is a feasible circulation that is integer-valued, and such an integer-valued
circulation naturally corresponds to a feasible survey design. Customer i will
be surveyed about product j if and only if the edge (i , j ) carries a unit of ow.
7.9 Airline Scheduling
The computational problems faced by the nations large airline carriers are
almost too complex to even imagine. They have to produce schedules for thousands of routes each day that are efcient in terms of equipment usage, crew
allocation, customer satisfaction, and a host of other factorsall in the face
of unpredictable issues like weather and breakdowns. Its not surprising that
theyre among the largest consumers of high-powered algorithmic techniques.
Covering these computational problems in any realistic level of detail
would take us much too far aeld. Instead, well discuss a toy problem that
captures, in a very clean way, some of the resource allocation issues that arise
in a context such as this. And, as is common in this book, the toy problem will
be much more useful for our purposes than the real problem, for the solution
to the toy problem involves a very general technique that can be applied in a
wide range of situations.
The Problem
Suppose youre in charge of managing a eet of airplanes and youd like to
create a ight schedule for them. Heres a very simple model for this. Your
market research has identied a set of m particular ight segments that would
be very lucrative if you could serve them; ight segment j is specied by four
parameters: its origin airport, its destination airport, its departure time, and
its arrival time. Figure 7.17(a) shows a simple example, consisting of six ight
segments youd like to serve with your planes over the course of a single day:
(1) Boston (depart 6 A.M.) Washington DC (arrive 7 A.M.)
(2) Philadelphia (depart 7 A.M.) Pittsburgh (arrive 8 A.M.)
387
388
BOS 6
Chapter 7 Network Flow
DCA 7
DCA 8
LAX 11
PHL 7
PIT 8
PHL 11
LAS 5
SFO 2
(a)
BOS 6
DCA 7
PHL 7
DCA 8
PIT 8
SFO
2:15
LAX 11
PHL 11
SEA
3:15
LAS 5
SFO 2
SFO
2:15
SEA 6
SEA 6
SEA
3:15
(b)
Figure 7.17 (a) A small instance of our simple Airline Scheduling Problem. (b) An
expanded graph showing which flights are reachable from which others.
(3) Washington DC (depart 8 A.M.) Los Angeles (arrive 11 A.M.)
(4) Philadelphia (depart 11 A.M.) San Francisco (arrive 2 P.M.)
(5) San Francisco (depart 2:15 P.M.) Seattle (arrive 3:15 P.M.)
(6) Las Vegas (depart 5 P.M.) Seattle (arrive 6 P.M.)
Note that each segment includes the times you want the ight to serve as well
as the airports.
It is possible to use a single plane for a ight segment i, and then later for
a ight segment j , provided that
(a) the destination of i is the same as the origin of j , and theres enough time
to perform maintenance on the plane in between; or
(b) you can add a ight segment in between that gets the plane from the
destination of i to the origin of j with adequate time in between.
For example, assuming an hour for intermediate maintenance time, you could
use a single plane for ights (1), (3), and (6) by having the plane sit in
Washington, DC, between ights (1) and (3), and then inserting the ight
7.9 Airline Scheduling
Los Angeles (depart 12 noon) Las Vegas (1 P.M.)
in between ights (3) and (6).
Formulating the Problem We can model this situation in a very general
way as follows, abstracting away from specic rules about maintenance times
and intermediate ight segments: We will simply say that ight j is reachable
from ight i if it is possible to use the same plane for ight i, and then later
for ight j as well. So under our specic rules (a) and (b) above, we can
easily determine for each pair i , j whether ight j is reachable from ight
i. (Of course, one can easily imagine more complex rules for reachability.
For example, the length of maintenance time needed in (a) might depend on
the airport; or in (b) we might require that the ight segment you insert be
sufciently protable on its own.) But the point is that we can handle any
set of rules with our denition: The input to the problem will include not just
the ight segments, but also a specication of the pairs (i , j ) for which a later
ight j is reachable from an earlier ight i. These pairs can form an arbitrary
directed acyclic graph.
The goal in this problem is to determine whether its possible to serve all
m ights on your original list, using at most k planes total. In order to do this,
you need to nd a way of efciently reusing planes for multiple ights.
For example, lets go back to the instance in Figure 7.17 and assume we
have k = 2 planes. If we use one of the planes for ights (1), (3), and (6)
as proposed above, we wouldnt be able to serve all of ights (2), (4), and
(5) with the other (since there wouldnt be enough maintenance time in San
Francisco between ights (4) and (5)). However, there is a way to serve all six
ights using two planes, via a different solution: One plane serves ights (1),
(3), and (5) (splicing in an LAXSFO ight), while the other serves (2), (4),
and (6) (splicing in PITPHL and SFOLAS).
Designing the Algorithm
We now discuss an efcient algorithm that can solve arbitrary instances of
the Airline Scheduling Problem, based on network ow. We will see that ow
techniques adapt very naturally to this problem.
The solution is based on the following idea. Units of ow will correspond
to airplanes. We will have an edge for each ight, and upper and lower capacity
bounds of 1 on these edges to require that exactly one unit of ow crosses this
edge. In other words, each ight must be served by one of the planes. If (ui , vi )
is the edge representing ight i, and (uj , vj ) is the edge representing ight j ,
and ight j is reachable from ight i, then we will have an edge from vi to uj
389
390
Chapter 7 Network Flow
with capacity 1; in this way, a unit of ow can traverse (ui , vi ) and then move
directly to (uj , vj ). Such a construction of edges is shown in Figure 7.17(b).
We extend this to a ow network by including a source and sink; we now
give the full construction in detail. The node set of the underlying graph G is
dened as follows.
.
For each ight i, the graph G will have the two nodes ui and vi .
.
G will also have a distinct source node s and sink node t .
The edge set of G is dened as follows.
.
.
.
.
.
For each i, there is an edge (ui , vi ) with a lower bound of 1 and a capacity
of 1. (Each ight on the list must be served.)
For each i and j so that ight j is reachable from ight i, there is an edge
(vi , uj ) with a lower bound of 0 and a capacity of 1. (The same plane can
perform ights i and j.)
For each i, there is an edge (s , ui ) with a lower bound of 0 and a capacity
of 1. (Any plane can begin the day with ight i.)
For each j , there is an edge (vj , t ) with a lower bound of 0 and a capacity
of 1. (Any plane can end the day with ight j.)
There is an edge (s , t ) with lower bound 0 and capacity k. (If we have
extra planes, we dont need to use them for any of the ights.)
Finally, the node s will have a demand of k, and the node t will have a
demand of k. All other nodes will have a demand of 0.
Our algorithm is to construct the network G and search for a feasible
circulation in it. We now prove the correctness of this algorithm.
Analyzing the Algorithm
(7.54) There is a way to perform all ights using at most k planes if and only
if there is a feasible circulation in the network G.
Proof. First, suppose there is a way to perform all ights using k k planes.
The set of ights performed by each individual plane denes a path P in the
network G , and we send one unit of ow on each such path P . To satisfy the full
demands at s and t , we send k k units of ow on the edge (s , t ). The resulting
circulation satises all demand, capacity, and lower bound conditions.
Conversely, consider a feasible circulation in the network G . By (7.52),
we know that there is a feasible circulation with integer ow values. Suppose
that k units of ow are sent on edges other than (s , t ). Since all other edges
have a capacity bound of 1, and the circulation is integer-valued, each such
edge that carries ow has exactly one unit of ow on it.
7.10 Image Segmentation
We now convert this to a schedule using the same kind of construction we
saw in the proof of (7.42), where we converted a ow to a collection of paths.
In fact, the situation is easier here since the graph has no cycles. Consider an
edge (s , ui ) that carries one unit of ow. It follows by conservation that (ui , vi )
carries one unit of ow, and that there is a unique edge out of vi that carries
one unit of ow. If we continue in this way, we construct a path P from s to
t , so that each edge on this path carries one unit of ow. We can apply this
construction to each edge of the form (s , uj ) carrying one unit of ow; in this
way, we produce k paths from s to t , each consisting of edges that carry one
unit of ow. Now, for each path P we create in this way, we can assign a single
plane to perform all the ights contained in this path.
Extensions: Modeling Other Aspects of the Problem
Airline scheduling consumes countless hours of CPU time in real life. We
mentioned at the beginning, however, that our formulation here is really a
toy problem; it ignores several obvious factors that would have to be taken
into account in these applications. First of all, it ignores the fact that a given
plane can only y a certain number of hours before it needs to be temporarily
taken out of service for more signicant maintenance. Second, we are making
up an optimal schedule for a single day (or at least for a single span of time) as
though there were no yesterday or tomorrow; in fact we also need the planes
to be optimally positioned for the start of day N + 1 at the end of day N . Third,
all these planes need to be staffed by ight crews, and while crews are also
reused across multiple ights, a whole different set of constraints operates here,
since human beings and airplanes experience fatigue at different rates. And
these issues dont even begin to cover the fact that serving any particular ight
segment is not a hard constraint; rather, the real goal is to optimize revenue,
and so we can pick and choose among many possible ights to include in our
schedule (not to mention designing a good fare structure for passengers) in
order to achieve this goal.
Ultimately, the message is probably this: Flow techniques are useful for
solving problems of this type, and they are genuinely used in practice. Indeed,
our solution above is a general approach to the efcient reuse of a limited set
of resources in many settings. At the same time, running an airline efciently
in real life is a very difcult problem.
7.10 Image Segmentation
A central problem in image processing is the segmentation of an image into
various coherent regions. For example, you may have an image representing
a picture of three people standing in front of a complex background scene. A
391
392
Chapter 7 Network Flow
natural but difcult goal is to identify each of the three people as coherent
objects in the scene.
The Problem
One of the most basic problems to be considered along these lines is that
of foreground/background segmentation: We wish to label each pixel in an
image as belonging to either the foreground of the scene or the background. It
turns out that a very natural model here leads to a problem that can be solved
efciently by a minimum cut computation.
Let V be the set of pixels in the underlying image that were analyzing.
We will declare certain pairs of pixels to be neighbors, and use E to denote
the set of all pairs of neighboring pixels. In this way, we obtain an undirected
graph G = (V , E ). We will be deliberately vague on what exactly we mean by
a pixel, or what we mean by the neighbor relation. In fact, any graph
G will yield an efciently solvable problem, so we are free to dene these
notions in any way that we want. Of course, it is natural to picture the pixels
as constituting a grid of dots, and the neighbors of a pixel to be those that are
directly adjacent to it in this grid, as shown in Figure 7.18(a).
s
(a)
t
(b)
Figure 7.18 (a) A pixel graph. (b) A sketch of the corresponding flow graph. Not all
edges from the source or to the sink are drawn.
7.10 Image Segmentation
For each pixel i, we have a likelihood ai that it belongs to the foreground,
and a likelihood bi that it belongs to the background. For our purposes, we
will assume that these likelihood values are arbitrary nonnegative numbers
provided as part of the problem, and that they specify how desirable it is to
have pixel i in the background or foreground. Beyond this, it is not crucial
precisely what physical properties of the image they are measuring, or how
they were determined.
In isolation, we would want to label pixel i as belonging to the foreground
if ai > bi , and to the background otherwise. However, decisions that we
make about the neighbors of i should affect our decision about i. If many
of is neighbors are labeled background, for example, we should be more
inclined to label i as background too; this makes the labeling smoother by
minimizing the amount of foreground/background boundary. Thus, for each
pair (i , j ) of neighboring pixels, there is a separation penalty pij 0 for placing
one of i or j in the foreground and the other in the background.
We can now specify our Segmentation Problem precisely, in terms of the
likelihood and separation parameters: It is to nd a partition of the set of pixels
into sets A and B (foreground and background, respectively) so as to maximize
ai +
q(A, B) =
iA
bj
j B
pij .
( i , j ) E
|A{i , j }|=1
Thus we are rewarded for having high likelihood values and penalized for
having neighboring pairs (i , j ) with one pixel in A and the other in B. The
problem, then, is to compute an optimal labeling a partition (A, B) that
maximizes q(A, B).
Designing and Analyzing the Algorithm
We notice right away that there is clearly a resemblance between the minimumcut problem and the problem of nding an optimal labeling. However, there
are a few signicant differences. First, we are seeking to maximize an objective
function rather than minimizing one. Second, there is no source and sink in the
labeling problem; and, moreover, we need to deal with values ai and bi on the
nodes. Third, we have an undirected graph G , whereas for the minimum-cut
problem we want to work with a directed graph. Lets address these problems
in order.
We deal with the fact that our Segmentation Problem is a maximization
problem through the following observation. Let Q = i (ai + bi ). The sum
i A a i +
j B bj is the same as the sum Q
iA bi
j B aj , so we can
write
393
394
Chapter 7 Network Flow
q(A, B) = Q
bi
iA
aj
j B
pij .
(i , j )E
|A{i , j }|=1
Thus we see that the maximization of q(A, B) is the same problem as the
minimization of the quantity
q (A, B) =
bi +
iA
aj +
j B
pij .
(i , j )E
|A{i , j }|=1
As for the missing source and the sink, we work by analogy with our constructions in previous sections: We create a new super-source s to represent
the foreground, and a new super-sink t to represent the background. This
also gives us a way to deal with the values ai and bi that reside at the nodes
(whereas minimum cuts can only handle numbers associated with edges).
Specically, we will attach each of s and t to every pixel, and use ai and bi to
dene appropriate capacities on the edges between pixel i and the source and
sink respectively.
Finally, to take care of the undirected edges, we model each neighboring
pair (i , j ) with two directed edges, (i , j ) and (j , i), as we did in the undirected
Disjoint Paths Problem. We will see that this works very well here too, since
in any s-t cut, at most one of these two oppositely directed edges can cross
from the s-side to the t -side of the cut (for if one does, then the other must go
from the t -side to the s-side).
Specically, we dene the following ow network G = (V , E ) shown in
Figure 7.18(b). The node set V consists of the set V of pixels, together with
two additional nodes s and t . For each neighboring pair of pixels i and j , we
add directed edges (i , j ) and (j , i), each with capacity pij . For each pixel i, we
add an edge (s , i) with capacity ai and an edge (i , t ) with capacity bi .
Now, an s-t cut (A, B) corresponds to a partition of the pixels into sets A
and B. Lets consider how the capacity of the cut c (A, B) relates to the quantity
q (A, B) that we are trying to minimize. We can group the edges that cross the
cut (A, B) into three natural categories.
.
.
.
Edges (s , j ), where j B; this edge contributes aj to the capacity of the
cut.
Edges (i , t ), where i A; this edge contributes bi to the capacity of the
cut.
Edges (i , j ) where i A and j B; this edge contributes pij to the capacity
of the cut.
Figure 7.19 illustrates what each of these three kinds of edges looks like relative
to a cut, on an example with four pixels.
7.10 Image Segmentation
s
aw
ax
u
puw
w
bu
v
pvx
x
bv
t
Figure 7.19 An s-t cut on a graph constructed from four pixels. Note how the three
types of terms in the expression for q (A, B) are captured by the cut.
If we add up the contributions of these three kinds of edges, we get
bi +
c (A, B) =
iA
aj +
j B
pij
(i , j )E
|A{i , j }|=1
= q (A, B).
So everything ts together perfectly. The ow network is set up so that the
capacity of the cut (A, B) exactly measures the quantity q (A, B): The three
kinds of edges crossing the cut (A, B), as we have just dened them (edges
from the source, edges to the sink, and edges involving neither the source nor
the sink), correspond to the three kinds of terms in the expression for q (A, B).
Thus, if we want to minimize q (A, B) (since we have argued earlier that
this is equivalent to maximizing q(A, B)), we just have to nd a cut of minimum
capacity. And this latter problem, of course, is something that we know how
to solve efciently.
Thus, through solving this minimum-cut problem, we have an optimal
algorithm in our model of foreground/background segmentation.
(7.55) The solution to the Segmentation Problem can be obtained by a
minimum-cut algorithm in the graph G constructed above. For a minimum
cut (A , B ), the partition (A, B) obtained by deleting s and t maximizes the
segmentation value q(A, B).
395
396
Chapter 7 Network Flow
7.11 Project Selection
Large (and small) companies are constantly faced with a balancing act between
projects that can yield revenue, and the expenses needed for activities that can
support these projects. Suppose, for example, that the telecommunications
giant CluNet is assessing the pros and cons of a project to offer some new
type of high-speed access service to residential customers. Marketing research
shows that the service will yield a good amount of revenue, but it must be
weighed against some costly preliminary projects that would be needed in
order to make this service possible: increasing the ber-optic capacity in the
core of their network, and buying a newer generation of high-speed routers.
What makes these types of decisions particularly tricky is that they interact
in complex ways: in isolation, the revenue from the high-speed access service
might not be enough to justify modernizing the routers; however , once the
company has modernized the routers, theyll also be in a position to pursue
a lucrative additional project with their corporate customers; and maybe this
additional project will tip the balance. And these interactions chain together:
the corporate project actually would require another expense, but this in
turn would enable two other lucrative projectsand so forth. In the end, the
question is: Which projects should be pursued, and which should be passed
up? Its a basic issue of balancing costs incurred with protable opportunities
that are made possible.
The Problem
Heres a very general framework for modeling a set of decisions such as this.
There is an underlying set P of projects, and each project i P has an associated
revenue pi , which can either be positive or negative. (In other words, each
of the lucrative opportunities and costly infrastructure-building steps in our
example above will be referred to as a separate project.) Certain projects are
prerequisites for other projects, and we model this by an underlying directed
acyclic graph G = (P , E ). The nodes of G are the projects, and there is an edge
(i , j ) to indicate that project i can only be selected if project j is selected as
well. Note that a project i can have many prerequisites, and there can be many
projects that have project j as one of their prerequisites. A set of projects A P
is feasible if the prerequisite of every project in A also belongs to A: for each
i A, and each edge (i , j ) E , we also have j A. We will refer to requirements
of this form as precedence constraints. The prot of a set of projects is dened
to be
profit(A) =
pi .
iA
7.11 Project Selection
The Project Selection Problem is to select a feasible set of projects with maximum prot.
This problem also became a hot topic of study in the mining literature,
starting in the early 1960s; here it was called the Open-Pit Mining Problem.3
Open-pit mining is a surface mining operation in which blocks of earth are
extracted from the surface to retrieve the ore contained in them. Before the
mining operation begins, the entire area is divided into a set P of blocks,
and the net value pi of each block is estimated: This is the value of the ore
minus the processing costs, for this block considered in isolation. Some of
these net values will be positive, others negative. The full set of blocks has
precedence constraints that essentially prevent blocks from being extracted
before others on top of them are extracted. The Open-Pit Mining Problem is to
determine the most protable set of blocks to extract, subject to the precedence
constraints. This problem falls into the framework of project selectioneach
block corresponds to a separate project.
Designing the Algorithm
Here we will show that the Project Selection Problem can be solved by reducing
it to a minimum-cut computation on an extended graph G , dened analogously
to the graph we used in Section 7.10 for image segmentation. The idea is to
construct G from G in such a way that the source side of a minimum cut in
G will correspond to an optimal set of projects to select.
To form the graph G , we add a new source s and a new sink t to the graph
G as shown in Figure 7.20. For each node i P with pi > 0, we add an edge
(s , i) with capacity pi . For each node i P with pi < 0, we add an edge (i , t )
with capacity pi . We will set the capacities on the edges in G later. However,
we can already see that the capacity of the cut ({s} , P {t }) is C = iP :pi >0 pi ,
so the maximum-ow value in this network is at most C .
We want to ensure that if (A , B ) is a minimum cut in this graph, then
A = A {s} obeys the precedence constraints; that is, if the node i A has
an edge (i , j ) E , then we must have j A. The conceptually cleanest way
to ensure this is to give each of the edges in G capacity of . We havent
previously formalized what an innite capacity would mean, but there is no
problem in doing this: it is simply an edge for which the capacity condition
imposes no upper bound at all. The algorithms of the previous sections, as well
as the Max-Flow Min-Cut Theorem, carry over to handle innite capacities.
However, we can also avoid bringing in the notion of innite capacities by
3
In contrast to the eld of data mining, which has motivated several of the problems we considered
earlier, were talking here about actual mining, where you dig things out of the ground.
397
398
Chapter 7 Network Flow
t
t
Projects with
negative value
Projects
s
Projects with
positive value
Projects
An
optimal
subset of
projects
s
Figure 7.20 The flow graph used to solve the Project Selection Problem. A possible
minimum-capacity cut is shown on the right.
simply assigning each of these edges a capacity that is effectively innite. In
our context, giving each of these edges a capacity of C + 1 would accomplish
this: The maximum possible ow value in G is at most C , and so no minimum
cut can contain an edge with capacity above C . In the description below, it
will not matter which of these options we choose.
We can now state the algorithm: We compute a minimum cut (A , B ) in
G , and we declare A {s} to be the optimal set of projects. We now turn to
proving that this algorithm indeed gives the optimal solution.
Analyzing the Algorithm
First consider a set of projects A that satises the precedence constraints. Let
A = A {s} and B = (P A) {t }, and consider the s-t cut (A , B ). If the set
A satises the precedence constraints, then no edge (i , j ) E crosses this cut,
as shown in Figure 7.20. The capacity of the cut can be expressed as follows.
7.11 Project Selection
(7.56) The capacity of the cut (A , B ), as dened from a project set A
satisfying the precedence constraints, is c (A , B ) = C iA pi .
Proof. Edges of G can be divided into three categories: those corresponding
to the edge set E of G , those leaving the source s, and those entering the sink
t . Because A satises the precedence constraints, the edges in E do not cross
the cut (A , B ), and hence do not contribute to its capacity. The edges entering
the sink t contribute
pi
iA and pi <0
to the capacity of the cut, and the edges leaving the source s contribute
pi .
i A and pi >0
Using the denition of C , we can rewrite this latter quantity as C
iA and pi >0 pi . The capacity of the cut (A , B ) is the sum of these two terms,
which is
pi = C
(pi ) + C
iA and pi <0
iA and pi >0
pi ,
iA
as claimed.
Next, recall that edges of G have capacity more than C = iP :pi >0 pi , and
so these edges cannot cross a cut of capacity at most C . This implies that such
cuts dene feasible sets of projects.
(7.57) If (A , B ) is a cut with capacity at most C, then the set A = A {s}
satises the precedence constraints.
Now we can prove the main goal of our construction, that the minimum
cut in G determines the optimum set of projects. Putting the previous two
claims together, we see that the cuts (A , B ) of capacity at most C are in oneto-one correspondence with feasible sets of project A = A {s}. The capacity
of such a cut (A , B ) is
c (A , B ) = C profit(A).
The capacity value C is a constant, independent of the cut (A , B ), so the cut
with minimum capacity corresponds to the set of projects A with maximum
prot. We have therefore proved the following.
(7.58) If (A , B ) is a minimum cut in G then the set A = A {s} is an
optimum solution to the Project Selection Problem.
399
400
Chapter 7 Network Flow
7.12 Baseball Elimination
Over on the radio side the producers saying, See that thing in the
paper last week about Einstein? . . . Some reporter asked him to gure
out the mathematics of the pennant race. You know, one team wins so
many of their remaining games, the other teams win this number or
that number. What are the myriad possibilities? Whos got the edge?
The hell does he know?
Apparently not much. He picked the Dodgers to eliminate the
Giants last Friday.
Don DeLillo, Underworld
The Problem
Suppose youre a reporter for the Algorithmic Sporting News, and the following
situation arises late one September. There are four baseball teams trying to
nish in rst place in the American League Eastern Division; lets call them
New York, Baltimore, Toronto, and Boston. Currently, each team has the
following number of wins:
New York: 92,
Baltimore: 91,
Toronto: 91,
Boston: 90.
There are ve games left in the season: These consist of all possible pairings
of the four teams above, except for New York and Boston.
The question is: Can Boston nish with at least as many wins as every
other team in the division (that is, nish in rst place, possibly in a tie)?
If you think about it, you realize that the answer is no. One argument is
the following. Clearly, Boston must win both its remaining games and New
York must lose both its remaining games. But this means that Baltimore and
Toronto will both beat New York; so then the winner of the Baltimore-Toronto
game will end up with the most wins.
Heres an argument that avoids this kind of cases analysis. Boston can
nish with at most 92 wins. Cumulatively, the other three teams have 274
wins currently, and their three games against each other will produce exactly
three more wins, for a nal total of 277. But 277 wins over three teams means
that one of them must have ended up with more than 92 wins.
So now you might start wondering: (i) Is there an efcient algorithm
to determine whether a team has been eliminated from rst place? And (ii)
whenever a team has been eliminated from rst place, is there an averaging
argument like this that proves it?
In more concrete notation, suppose we have a set S of teams, and for each
x S, its current number of wins is wx . Also, for two teams x , y S, they still
7.12 Baseball Elimination
have to play gxy games against one another. Finally, we are given a specic
team z .
We will use maximum-ow techniques to achieve the following two things.
First, we give an efcient algorithm to decide whether z has been eliminated
from rst placeor, to put it in positive terms, whether it is possible to choose
outcomes for all the remaining games in such a way that the team z ends with
at least as many wins as every other team in S. Second, we prove the following
clean characterization theorem for baseball eliminationessentially, that there
is always a short proof when a team has been eliminated.
(7.59) Suppose that team z has indeed been eliminated. Then there exists a
proof of this fact of the following form:
.
z can nish with at most m wins.
.
There is a set of teams T S so that
wx +
xT
gxy > m|T |.
x , yT
(And hence one of the teams in T must end with strictly more than m
wins.)
As a second, more complex illustration of how the averaging argument in
(7.59) works, consider the following example. Suppose we have the same four
teams as before, but now the current number of wins is
New York: 90,
Baltimore: 88,
Toronto: 87,
Boston: 79.
The remaining games are as follows. Boston still has four games against each
of the other three teams. Baltimore has one more game against each of New
York and Toronto. And nally, New York and Toronto still have six games left
to play against each other. Clearly, things dont look good for Boston, but is it
actually eliminated?
The answer is yes; Boston has been eliminated. To see this, rst note
that Boston can end with at most 91 wins; and now consider the set of teams
T = {New York, Toronto}. Together New York and Toronto already have 177
wins; their six remaining games will result in a total of 183; and 183 > 91.
2
This means that one of them must end up with more than 91 wins, and so
Boston cant nish in rst. Interestingly, in this instance the set of all three
teams ahead of Boston cannot constitute a similar proof: All three teams taken
togeher have a total of 265 wins with 8 games left among them; this is a total
of 273, and 273 = 91 not enough by itself to prove that Boston couldnt end
3
up in a multi-way tie for rst. So its crucial for the averaging argument that we
choose the set T consisting just of New York and Toronto, and omit Baltimore.
401
402
Chapter 7 Network Flow
Designing and Analyzing the Algorithm
We begin by constructing a ow network that provides an efcient algorithm
for determining whether z has been eliminated. Then, by examining the
minimum cut in this network, we will prove (7.59).
Clearly, if theres any way for z to end up in rst place, we should have
z win all its remaining games. Lets suppose that this leaves it with m wins.
We now want to carefully allocate the wins from all remaining games so that
no other team ends with more than m wins. Allocating wins in this way can
be solved by a maximum-ow computation, via the following basic idea. We
have a source s from which all wins emanate. The ith win can pass through
one of the two teams involved in the ith game. We then impose a capacity
constraint saying that at most m wx wins can pass through team x.
More concretely, we construct the following ow network G , as shown in
Figure 7.21. First, let S = S {z }, and let g = x , yS gxy the total number
of games left between all pairs of teams in S . We include nodes s and t , a
node vx for each team x S , and a node uxy for each pair of teams x , y S
with a nonzero number of games left to play against each other. We have the
following edges.
.
.
.
Edges (s , uxy ) (wins emanate from s);
Edges (uxy , vx ) and (uxy , vy ) (only x or y can win a game that they play
against each other ); and
Edges (vx , t ) (wins are absorbed at t).
Lets consider what capacities we want to place on these edges. We want gxy
wins to ow from s to uxy at saturation, so we give (s , uxy ) a capacity of gxy .
We want to ensure that team x cannot win more than m wx games, so we
The set T={NY, Toronto}
proves Boston is
eliminated.
NYTor
6
NY
1
s
1
4
NYBalt
t
Tor
1
3
BaltTor
Balt
Figure 7.21 The flow network for the second example. As the minimum cut indicates,
there is no flow of value g , and so Boston has been eliminated.
7.12 Baseball Elimination
give the edge (vx , t ) a capacity of m wx . Finally, an edge of the form (uxy , vy )
should have at least gxy units of capacity, so that it has the ability to transport
all the wins from uxy on to vx ; in fact, our analysis will be the cleanest if we
give it innite capacity. (We note that the construction still works even if this
edge is given only gxy units of capacity, but the proof of (7.59) will become a
little more complicated.)
Now, if there is a ow of value g , then it is possible for the outcomes
of all remaining games to yield a situation where no team has more than m
wins; and hence, if team z wins all its remaining games, it can still achieve at
least a tie for rst place. Conversely, if there are outcomes for the remaining
games in which z achieves at least a tie, we can use these outcomes to dene
a ow of value g . For example, in Figure 7.21, which is based on our second
example, the indicated cut shows that the maximum ow has value at most
7, whereas g = 6 + 1 + 1 = 8.
In summary, we have shown
(7.60) Team z has been eliminated if and only if the maximum ow in G
has value strictly less than g . Thus we can test in polynomial time whether z
has been eliminated.
Characterizing When a Team Has Been Eliminated
Our network ow construction can also be used to prove (7.59). The idea is that
the Max-Flow Min-Cut Theorem gives a nice if and only if characterization
for the existence of ow, and if we interpret this characterization in terms
of our application, we get the comparably nice characterization here. This
illustrates a general way in which one can generate characterization theorems
for problems that are reducible to network ow.
Proof of (7.59). Suppose that z has been eliminated from rst place. Then
the maximum s-t ow in G has value g < g ; so there is an s-t cut (A, B) of
capacity g , and (A, B) is a minimum cut. Let T be the set of teams x for which
vx A. We will now prove that T can be used in the averaging argument in
(7.59).
First, consider the node uxy , and suppose one of x or y is not in T , but
uxy A. Then the edge (uxy , vx ) would cross from A into B, and hence the
cut (A, B) would have innite capacity. This contradicts the assumption that
(A, B) is a minimum cut of capacity less than g . So if one of x or y is not in
T , then uxy B. On the other hand, suppose both x and y belong to T , but
uxy B. Consider the cut (A , B ) that we would obtain by adding uxy to the set
A and deleting it from the set B. The capacity of (A , B ) is simply the capacity
of (A, B), minus the capacity gxy of the edge (s , uxy )for this edge (s , uxy ) used
403
404
Chapter 7 Network Flow
to cross from A to B, and now it does not cross from A to B . But since gxy > 0,
this means that (A , B ) has smaller capacity than (A, B), again contradicting
our assumption that (A, B) is a minimum cut. So, if both x and y belong to T ,
then uxy A.
Thus we have established the following conclusion, based on the fact that
(A, B) is a minimum cut: uxy A if and only if both x , y T .
Now we just need to work out the minimum-cut capacity c (A, B) in terms
of its constituent edge capacities. By the conclusion in the previous paragraph,
we know that edges crossing from A to B have one of the following two forms:
.
.
edges of the form (vx , t ), where x T , and
edges of the form (s , uxy ), where at least one of x or y does not belong
to T (in other words, {x , y} T ).
Thus we have
(m wx ) +
c (A, B) =
xT
gxy
{x , y}T
wx + (g
= m|T |
xT
gxy ).
x , yT
Since we know that c (A, B) = g < g , this last inequality implies
wx
m|T |
xT
gxy < 0,
x , yT
and hence
wx +
xT
gxy > m|T |.
x , yT
For example, applying the argument in the proof of (7.59) to the instance
in Figure 7.21, we see that the nodes for New York and Toronto are on the
source side of the minimum cut, and, as we saw earlier, these two teams
indeed constitute a proof that Boston has been eliminated.
* 7.13 A Further Direction: Adding Costs to the
Matching Problem
Lets go back to the rst problem we discussed in this chapter, Bipartite
Matching. Perfect matchings in a bipartite graph formed a way to model the
problem of pairing one kind of object with anotherjobs with machines, for
example. But in many settings, there are a large number of possible perfect
matchings on the same set of objects, and wed like a way to express the idea
that some perfect matchings may be better than others.
7.13 A Further Direction: Adding Costs to the Matching Problem
The Problem
A natural way to formulate a problem based on this notion is to introduce
costs. It may be that we incur a certain cost to perform a given job on a given
machine, and wed like to match jobs with machines in a way that minimizes
the total cost. Or there may be n re trucks that must be sent to n distinct
houses; each house is at a given distance from each re station, and wed
like a matching that minimizes the average distance each truck drives to its
associated house. In short, it is very useful to have an algorithm that nds a
perfect matching of minimum total cost .
Formally, we consider a bipartite graph G = (V , E ) whose node set, as
usual, is partitioned as V = X Y so that every edge e E has one end in X
and the other end in Y . Furthermore, each edge e has a nonnegative cost ce .
For a matching M , we say that the cost of the matching is the total cost of all
edges in M , that is, cost (M ) = eM ce . The Minimum-Cost Perfect Matching
Problem assumes that |X | = |Y | = n, and the goal is to nd a perfect matching
of minimum cost.
Designing and Analyzing the Algorithm
We now describe an efcient algorithm to solve this problem, based on the
idea of augmenting paths but adapted to take the costs into account. Thus, the
algorithm will iteratively construct matchings using i edges, for each value of i
from 1 to n. We will show that when the algorithm concludes with a matching
of size n, it is a minimum-cost perfect matching. The high-level structure of the
algorithm is quite simple. If we have a minimum-cost matching of size i, then
we seek an augmenting path to produce a matching of size i + 1; and rather
than looking for any augmenting path (as was sufcient in the case without
costs), we use the cheapest augmenting path so that the larger matching will
also have minimum cost.
Recall the construction of the residual graph used for nding augmenting
paths. Let M be a matching. We add two new nodes s and t to the graph. We
add edges (s , x) for all nodes x X that are unmatched and edges (y , t ) for all
nodes y Y that are unmatched. An edge e = (x , y) E is oriented from x to
y if e is not in the matching M and from y to x if e M . We will use GM to
denote this residual graph. Note that all edges going from Y to X are in the
matching M , while the edges going from X to Y are not. Any directed s-t path
P in the graph GM corresponds to a matching one larger than M by swapping
edges along P , that is, the edges in P from X to Y are added to M and all edges
in P that go from Y to X are deleted from M . As before, we will call a path P in
GM an augmenting path, and we say that we augment the matching M using
the path P .
405
406
Chapter 7 Network Flow
Now we would like the resulting matching to have as small a cost as
possible. To achieve this, we will search for a cheap augmenting path with
respect to the following natural costs. The edges leaving s and entering t will
have cost 0; an edge e oriented from X to Y will have cost ce (as including this
edge in the path means that we add the edge to M ); and an edge e oriented
from Y to X will have cost ce (as including this edge in the path means that
we delete the edge from M ). We will use cost (P ) to denote the cost of a path
P in GM . The following statement summarizes this construction.
(7.61) Let M be a matching and P be a path in GM from s to t. Let M be the
matching obtained from M by augmenting along P. Then |M | = |M | + 1 and
cost (M ) = cost (M ) + cost (P ).
Given this statement, it is natural to suggest an algorithm to nd a
minimum-cost perfect matching: We iteratively nd minimum-cost paths in
GM , and use the paths to augment the matchings. But how can we be sure
that the perfect matching we nd is of minimum cost? Or even worse, is this
algorithm even meaningful? We can only nd minimum-cost paths if we know
that the graph GM has no negative cycles.
Analyzing Negative Cycles In fact, understanding the role of negative cycles
in GM is the key to analyzing the algorithm. First consider the case in which M
is a perfect matching. Note that in this case the node s has no leaving edges,
and t has no entering edges in GM (as our matching is perfect), and hence no
cycle in GM contains s or t .
(7.62) Let M be a perfect matching. If there is a negative-cost directed cycle
C in GM , then M is not minimum cost.
Proof. To see this, we use the cycle C for augmentation, just the same way
we used directed paths to obtain larger matchings. Augmenting M along C
involves swapping edges along C in and out of M . The resulting new perfect
matching M has cost cost (M ) = cost (M ) + cost (C ); but cost (C ) < 0, and hence
M is not of minimum cost.
More importantly, the converse of this statement is true as well; so in fact
a perfect matching M has minimum cost precisely when there is no negative
cycle in GM .
(7.63) Let M be a perfect matching. If there are no negative-cost directed
cycles C in GM , then M is a minimum-cost perfect matching.
Proof. Suppose the statement is not true, and let M be a perfect matching of
smaller cost. Consider the set of edges in one of M and M but not in both.
7.13 A Further Direction: Adding Costs to the Matching Problem
Observe that this set of edges corresponds to a set of node-disjoint directed
cycles in GM . The cost of the set of directed cycles is exactly cost (M ) cost (M ).
Assuming M has smaller cost than M , it must be that at least one of these
cycles has negative cost.
Our plan is thus to iterate through matchings of larger and larger size,
maintaining the property that the graph GM has no negative cycles in any
iteration. In this way, our computation of a minimum-cost path will always
be well dened; and when we terminate with a perfect matching, we can use
(7.63) to conclude that it has minimum cost.
Maintaining Prices on the Nodes It will help to think about a numerical price
p(v) associated with each node v. These prices will help both in understanding
how the algorithm runs, and they will also help speed up the implementation.
One issue we have to deal with is to maintain the property that the graph
GM has no negative cycles in any iteration. How do we know that after an
augmentation, the new residual graph still has no negative cycles? The prices
will turn out to serve as a compact proof to show this.
To understand prices, it helps to keep in mind an economic interpretation
of them. For this purpose, consider the following scenario. Assume that the
set X represents people who need to be assigned to do a set of jobs Y . For an
edge e = (x , y), the cost ce is a cost associated with having person x doing job
y. Now we will think of the price p(x) as an extra bonus we pay for person x to
participate in this system, like a signing bonus. With this in mind, the cost
for assigning person x to do job y will become p(x) + ce . On the other hand,
we will think of the price p(y) for nodes y Y as a reward, or value gained by
taking care of job y (no matter which person in X takes care of it). This way
the net cost of assigning person x to do job y becomes p(x) + ce p(y): this
is the cost of hiring x for a bonus of p(x), having him do job y for a cost of ce ,
and then cashing in on the reward p(y). We will call this the reduced cost of an
p
edge e = (x , y) and denote it by ce = p(x) + ce p(y). However, it is important
to keep in mind that only the costs ce are part of the problem description; the
prices (bonuses and rewards) will be a way to think about our solution.
Specically, we say that a set of numbers {p(v) : v V } forms a set of
compatible prices with respect to a matching M if
(i) for all unmatched nodes x X we have p(x) = 0 (that is, people not asked
to do any job do not need to be paid);
(ii) for all edges e = (x , y) we have p(x) + ce p(y) (that is, every edge has
a nonnegative reduced cost); and
(iii) for all edges e = (x , y) M we have p(x) + ce = p(y) (every edge used in
the assignment has a reduced cost of 0).
407
408
Chapter 7 Network Flow
Why are such prices useful? Intuitively, compatible prices suggest that the
matching is cheap: Along the matched edges reward equals cost, while on
all other edges the reward is no bigger than the cost. For a partial matching,
this may not imply that the matching has the smallest possible cost for its
size (it may be taking care of expensive jobs). However, we claim that if M
is any matching for which there exists a set of compatible prices, then GM
has no negative cycles. For a perfect matching M , this will imply that M is of
minimum cost by (7.63).
To see why GM can have no negative cycles, we extend the denition of
reduced cost to edges in the residual graph by using the same expression
p
ce = p(v) + ce p(w) for any edge e = (v , w). Observe that the denition
of compatible prices implies that all edges in the residual graph GM have
nonnegative reduced costs. Now, note that for any cycle C , we have
cost (C ) =
p
ce ,
ce =
e C
e C
since all the terms on the right-hand side corresponding to prices cancel out.
We know that each term on the right-hand side is nonnegative, and so clearly
cost (C ) is nonnegative.
There is a second, algorithmic reason why it is useful to have prices on
the nodes. When you have a graph with negative-cost edges but no negative
cycles, you can compute shortest paths using the Bellman-Ford Algorithm in
O(mn) time. But if the graph in fact has no negative-cost edges, then you can
use Dijkstras Algorithm instead, which only requires time O(m log n)almost
a full factor of n faster.
In our case, having the prices around allows us to compute shortest paths
p
with respect to the nonnegative reduced costs ce , arriving at an equivalent
answer. Indeed, suppose we use Dijkstras Algorithm to nd the minimum
cost dp , M (v) of a directed path from s to every node v X Y subject to the
p
costs ce . Given the minimum costs dp , M (y) for an unmatched node y Y , the
(nonreduced) cost of the path from s to t through y is dp , M (y) + p(y), and so
we nd the minimum cost in O(n) additional time. In summary, we have the
following fact.
(7.64) Let M be a matching, and p be compatible prices. We can use one
run of Dijkstras Algorithm and O(n) extra time to nd the minimum-cost path
from s to t.
Updating the Node Prices We took advantage of the prices to improve one
iteration of the algorithm. In order to be ready for the next iteration, we need
not only the minimum-cost path (to get the next matching), but also a way to
produce a set of compatible prices with respect to the new matching.
7.13 A Further Direction: Adding Costs to the Matching Problem
x
s
x
e
e
y
t
y
Figure 7.22 A matching M (the dark edges), and a residual graph used to increase the
size of the matching.
To get some intuition on how to do this, consider an unmatched node x
with respect to a matching M , and an edge e = (x , y), as shown in Figure 7.22.
If the new matching M includes edge e (that is, if e is on the augmenting
path we use to update the matching), then we will want to have the reduced
cost of this edge to be zero. However, the prices p we used with matching M
p
may result in a reduced cost ce > 0 that is, the assignment of person x to
job y, in our economic interpretation, may not be viewed as cheap enough.
We can arrange the zero reduced cost by either increasing the price p(y) (ys
p
reward) by ce , or by decreasing the price p(x) by the same amount. To keep
prices nonnegative, we will increase the price p(y). However, node y may be
matched in the matching M to some other node x via an edge e = (x , y), as
shown in Figure 7.22. Increasing the reward p(y) decreases the reduced cost
of edge e to negative, and hence the prices are no longer compatible. To keep
things compatible, we can increase p(x ) by the same amount. However, this
change might cause problems on other edges. Can we update all prices and
keep the matching and the prices compatible on all edges? Surprisingly, this
can be done quite simply by using the distances from s to all other nodes
computed by Dijkstras Algorithm.
(7.65) Let M be a matching, let p be compatible prices, and let M be a
matching obtained by augmenting along the minimum-cost path from s to t.
Then p (v) = dp , M (v) + p(v) is a compatible set of prices for M .
Proof. To prove compatibility, consider rst an edge e = (x , y) M . The only
edge entering x is the directed edge (y , x ), and hence dp , M (x ) = dp , M (y)
p
p
ce , where ce = p(y) + ce p(x ), and thus we get the desired equation on
such edges. Next consider edges (x , y) in M M . These edges are along the
p
minimum-cost path from s to t , and hence they satisfy dp , M (y) = dp , M (x) + ce
as desired. Finally, we get the required inequality for all other edges since all
p
edges e = (x , y) M must satisfy dp , M (y) dp , M (x) + ce .
409
410
Chapter 7 Network Flow
Finally, we have to consider how to initialize the algorithm, so as to get it
underway. We initialize M to be the empty set, dene p(x) = 0 for all x X ,
and dene p(y), for y Y , to be the minimum cost of an edge entering y. Note
that these prices are compatible with respect to M = .
We summarize the algorithm below.
Start with M equal to the empty set
Define p(x) = 0 for x X , and p(y) = min ce for y Y
e into y
While M is not a perfect matching
Find a minimum-cost s-t path P in GM using (7.64) with prices p
Augment along P to produce a new matching M
Find a set of compatible prices with respect to M via (7.65)
Endwhile
The nal set of compatible prices yields a proof that GM has no negative
cycles; and by (7.63), this implies that M has minimum cost.
(7.66) The minimum-cost perfect matching can be found in the time required
for n shortest-path computations with nonegative edge lengths.
Extensions: An Economic Interpretation of the Prices
To conclude our discussion of the Minimum-Cost Perfect Matching Problem,
we develop the economic interpretation of the prices a bit further. We consider
the following scenario. Assume X is a set of n people each looking to buy a
house, and Y is a set of n houses that they are all considering. Let v(x , y) denote
the value of house y to buyer x. Since each buyer wants one of the houses,
one could argue that the best arrangement would be to nd a perfect matching
M that maximizes (x , y)M v(x , y). We can nd such a perfect matching by
using our minimum-cost perfect matching algorithm with costs ce = v(x , y)
if e = (x , y).
The question we will ask now is this: Can we convince these buyers to
buy the house they are allocated? On her own, each buyer x would want to
buy the house y that has maximum value v(x , y) to her. How can we convince
her to buy instead the house that our matching M allocated? We will use prices
to change the incentives of the buyers. Suppose we set a price P (y) for each
house y, that is, the person buying the house y must pay P (y). With these
prices in mind, a buyer will be interested in buying the house with maximum
net value, that is, the house y that maximizes v(x , y) P (y). We say that a
Solved Exercises
perfect matching M and house prices P are in equilibrium if, for all edges
(x , y) M and all other houses y , we have
v(x , y) P (y) v(x , y ) P (y ).
But can we nd a perfect matching and a set of prices so as to achieve this
state of affairs, with every buyer ending up happy? In fact, the minimum-cost
perfect matching and an associated set of compatible prices provide exactly
what were looking for.
(7.67) Let M be a perfect matching of minimum cost, where ce = v(x , y) for
each edge e = (x , y), and let p be a compatible set of prices. Then the matching
M and the set of prices {P (y) = p(y) : y Y } are in equilibrium.
Proof. Consider an edge e = (x , y) M , and let e = (x , y ). Since M and p are
compatible, we have p(x) + ce = p(y) and p(x) + ce p(y ). Subtracting these
two inequalities to cancel p(x), and substituting the values of p and c , we get
the desired inequality in the denition of equilibrium.
Solved Exercises
Solved Exercise 1
Suppose you are given a directed graph G = (V , E ), with a positive integer
capacity ce on each edge e , a designated source s V , and a designated sink
t V . You are also given an integer maximum s-t ow in G , dened by a ow
value fe on each edge e .
Now suppose we pick a specic edge e E and increase its capacity by
one unit. Show how to nd a maximum ow in the resulting capacitated graph
in time O(m + n), where m is the number of edges in G and n is the number
of nodes.
Solution The point here is that O(m + n) is not enough time to compute a
new maximum ow from scratch, so we need to gure out how to use the ow
f that we are given. Intuitively, even after we add 1 to the capacity of edge e ,
the ow f cant be that far from maximum; after all, we havent changed the
network very much.
In fact, its not hard to show that the maximum ow value can go up by
at most 1.
(7.68) Consider the ow network G obtained by adding 1 to the capacity of
e. The value of the maximum ow in G is either (f ) or (f ) + 1.
411
412
Chapter 7 Network Flow
Proof. The value of the maximum ow in G is at least (f ), since f is still a
feasible ow in this network. It is also integer-valued. So it is enough to show
that the maximum-ow value in G is at most (f ) + 1.
By the Max-Flow Min-Cut Theorem, there is some s-t cut (A, B) in the
original ow network G of capacity (f ). Now we ask: What is the capacity of
(A, B) in the new ow network G ? All the edges crossing (A, B) have the same
capacity in G that they did in G , with the possible exception of e (in case e
crosses (A, B)). But ce only increased by 1, and so the capacity of (A, B) in the
new ow network G is at most (f ) + 1.
Statement (7.68) suggests a natural algorithm. Starting with the feasible
ow f in G , we try to nd a single augmenting path from s to t in the residual
graph Gf . This takes time O(m + n). Now one of two things will happen. Either
we will fail to nd an augmenting path, and in this case we know that f is
a maximum ow. Otherwise the augmentation succeeds, producing a ow f
of value at least (f ) + 1. In this case, we know by (7.68) that f must be a
maximum ow. So either way, we produce a maximum ow after a single
augmenting path computation.
Solved Exercise 2
You are helping the medical consulting rm Doctors Without Weekends set up
the work schedules of doctors in a large hospital. Theyve got the regular daily
schedules mainly worked out. Now, however, they need to deal with all the
special cases and, in particular, make sure that they have at least one doctor
covering each vacation day.
Heres how this works. There are k vacation periods (e.g., the week of
Christmas, the July 4th weekend, the Thanksgiving weekend, . . . ), each
spanning several contiguous days. Let Dj be the set of days included in the
j th vacation period; we will refer to the union of all these days, j Dj , as the set
of all vacation days.
There are n doctors at the hospital, and doctor i has a set of vacation days
Si when he or she is available to work. (This may include certain days from a
given vacation period but not others; so, for example, a doctor may be able to
work the Friday, Saturday, or Sunday of Thanksgiving weekend, but not the
Thursday.)
Give a polynomial-time algorithm that takes this information and determines whether it is possible to select a single doctor to work on each vacation
day, subject to the following constraints.
413
Solved Exercises
.
For a given parameter c , each doctor should be assigned to work at most
c vacation days total, and only days when he or she is available.
.
For each vacation period j , each doctor should be assigned to work at
most one of the days in the set Dj . (In other words, although a particular
doctor may work on several vacation days over the course of a year, he or
she should not be assigned to work two or more days of the Thanksgiving
weekend, or two or more days of the July 4th weekend, etc.)
The algorithm should either return an assignment of doctors satisfying these
constraints or report (correctly) that no such assignment exists.
Solution This is a very natural setting in which to apply network ow, since
at a high level were trying to match one set (the doctors) with another set
(the vacation days). The complication comes from the requirement that each
doctor can work at most one day in each vacation period.
So to begin, lets see how wed solve the problem without that requirement, in the simpler case where each doctor i has a set Si of days when he or
she can work, and each doctor should be scheduled for at most c days total.
The construction is pictured in Figure 7.23(a). We have a node ui representing
each doctor attached to a node v representing each day when he or she can
Holidays
Holidays
Gadgets
Doctors
Doctors
Source
Sink
(a)
Source
Sink
(b)
Figure 7.23 (a) Doctors are assigned to holiday days without restricting how many
days in one holiday a doctor can work. (b) The flow network is expanded with gadgets
that prevent a doctor from working more than one day from each vacation period. The
shaded sets correspond to the different vacation periods.
414
Chapter 7 Network Flow
work; this edge has a capacity of 1. We attach a super-source s to each doctor
node ui by an edge of capacity c , and we attach each day node v to a supersink t by an edge with upper and lower bounds of 1. This way, assigned days
can ow through doctors to days when they can work, and the lower bounds
on the edges from the days to the sink guarantee that each day is covered. Finally, suppose there are d vacation days total; we put a demand of +d on the
sink and d on the source, and we look for a feasible circulation. (Recall that
once weve introduced lower bounds on some edges, the algorithms in the text
are phrased in terms of circulations with demands, not maximum ow.)
But now we have to handle the extra requirement, that each doctor can
work at most one day from each vacation period. To do this, we take each pair
(i , j ) consisting of a doctor i and a vacation period j , and we add a vacation
gadget as follows. We include a new node wij with an incoming edge of
capacity 1 from the doctor node ui , and with outgoing edges of capacity 1 to
each day in vacation period j when doctor i is available to work. This gadget
serves to choke off the ow from ui into the days associated with vacation
period j , so that at most one unit of ow can go to them collectively. The
construction is pictured in Figure 7.23(b). As before, we put a demand of +d
on the sink and d on the source, and we look for a feasible circulation. The
total running time is the time to construct the graph, which is O(nd), plus the
time to check for a single feasible circulation in this graph.
The correctness of the algorithm is a consequence of the following claim.
(7.69) There is a way to assign doctors to vacation days in a way that respects
all constraints if and only if there is a feasible circulation in the ow network
we have constructed.
Proof. First, if there is a way to assign doctors to vacation days in a way
that respects all constraints, then we can construct the following circulation.
If doctor i works on day of vacation period j , then we send one unit of
ow along the path s, ui , wij , v , t ; we do this for all such (i , ) pairs. Since
the assignment of doctors satised all the constraints, the resulting circulation
respects all capacities; and it sends d units of ow out of s and into t , so it
meets the demands.
Conversely, suppose there is a feasible circulation. For this direction of
the proof, we will show how to use the circulation to construct a schedule
for all the doctors. First, by (7.52), there is a feasible circulation in which all
ow values are integers. We now construct the following schedule: If the edge
(wij , v ) carries a unit of ow, then we have doctor i work on day . Because
of the capacities, the resulting schedule has each doctor work at most c days,
at most one in each vacation period, and each day is covered by one doctor.
415
Exercises
Exercises
1. (a)
u
List all the minimum s-t cuts in the flow network pictured in Figure 7.24. The capacity of each edge appears as a label next to the
edge.
(b) What is the minimum capacity of an s-t cut in the flow network in
Figure 7.25? Again, the capacity of each edge appears as a label next
to the edge.
2. Figure 7.26 shows a flow network on which an s-t flow has been computed.
The capacity of each edge appears as a label next to the edge, and the
numbers in boxes give the amount of flow sent on each edge. (Edges
without boxed numbersspecifically, the four edges of capacity 3have
no flow being sent on them.)
(a)
What is the value of this flow? Is this a maximum (s,t ) flow in this
graph?
(b) Find a minimum s-t cut in the flow network pictured in Figure 7.26,
and also say what its capacity is.
3. Figure 7.27 shows a flow network on which an s-t flow has been computed.
The capacity of each edge appears as a label next to the edge, and the
numbers in boxes give the amount of flow sent on each edge. (Edges
without boxed numbers have no flow being sent on them.)
(a)
What is the value of this flow? Is this a maximum (s,t ) flow in this
graph?
5
10
5
3
88
3
10 8
5
88
s
t
3
5
3
10
5
5
d
Figure 7.26 What is the value of the depicted flow? Is it a maximum flow? What is the
minimum cut?
1
1
s
t
1
1
1
v
Figure 7.24 What are the
minimum s-t cuts in this flow
network?
u
4
2
s
t
6
2
4
v
Figure 7.25 What is the minimum capacity of an s-t cut in
this flow network?
416
Chapter 7 Network Flow
a
6
5
10
1
2
5
1
s
3
6
b
3
c
5
1
1
3
5
t
5
3
10
1
d
Figure 7.27 What is the value of the depicted flow? Is it a maximum flow? What is the
minimum cut?
(b) Find a minimum s-t cut in the flow network pictured in Figure 7.27,
and also say what its capacity is.
4. 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.
Let G be an arbitrary ow network, with a source s, a sink t, and a positive
integer capacity ce on every edge e. If f is a maximum s-t ow in G, then f
saturates every edge out of s with ow (i.e., for all edges e out of s, we have
f (e ) = ce ).
5. 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.
Let G be an arbitrary ow network, with a source s, a sink t, and a positive
integer capacity ce on every edge e; and let (A, B) be a mimimum s-t cut with
respect to these capacities {ce : e E }. Now suppose we add 1 to every capacity;
then (A, B) is still a minimum s-t cut with respect to these new capacities
{1 + ce : e E }.
6. Suppose youre a consultant for the Ergonomic Architecture Commission,
and they come to you with the following problem.
Theyre really concerned about designing houses that are userfriendly, and theyve been having a lot of trouble with the setup of light
fixtures and switches in newly designed houses. Consider, for example,
a one-floor house with n light fixtures and n locations for light switches
mounted in the wall. Youd like to be able to wire up one switch to control
each light fixture, in such a way that a person at the switch can see the
light fixture being controlled.
Exercises
1
1
a
a
2
b
b
3
c
(a) Ergonomic
23
c
(b) Not ergonomic
Figure 7.28 The floor plan in (a) is ergonomic, because we can wire switches to fixtures
in such a way that each fixture is visible from the switch that controls it. (This can be
done by wiring switch 1 to a , switch 2 to b, and switch 3 to c .) The floor plan in (b) is not
ergonomic, because no such wiring is possible.
Sometimes this is possible and sometimes it isnt. Consider the two
simple floor plans for houses in Figure 7.28. There are three light fixtures
(labeled a , b, c ) and three switches (labeled 1, 2, 3). It is possible to wire
switches to fixtures in Figure 7.28(a) so that every switch has a line of
sight to the fixture, but this is not possible in Figure 7.28(b).
Lets call a floor plan, together with n light fixture locations and n
switch locations, ergonomic if its possible to wire one switch to each
fixture so that every fixture is visible from the switch that controls it.
A floor plan will be represented by a set of m horizontal or vertical
line segments in the plane (the walls), where the ith wall has endpoints
(xi , yi ) , (xi , yi ). Each of the n switches and each of the n fixtures is given by
its coordinates in the plane. A fixture is visible from a switch if the line
segment joining them does not cross any of the walls.
Give an algorithm to decide if a given floor plan is ergonomic. The
running time should be polynomial in m and n. You may assume that you
have a subroutine with O(1) running time that takes two line segments as
input and decides whether or not they cross in the plane.
7. Consider a set of mobile computing clients in a certain town who each
need to be connected to one of several possible base stations . Well
suppose there are n clients, with the position of each client specified
by its (x , y) coordinates in the plane. There are also k base stations; the
position of each of these is specified by (x , y) coordinates as well.
For each client, we wish to connect it to exactly one of the base
stations. Our choice of connections is constrained in the following ways.
417
418
Chapter 7 Network Flow
There is a range parameter r a client can only be connected to a base
station that is within distance r . There is also a load parameter Lno
more than L clients can be connected to any single base station.
Your goal is to design a polynomial-time algorithm for the following
problem. Given the positions of a set of clients and a set of base stations,
as well as the range and load parameters, decide whether every client can
be connected simultaneously to a base station, subject to the range and
load conditions in the previous paragraph.
8. Statistically, the arrival of spring typically results in increased accidents
and increased need for emergency medical treatment, which often requires blood transfusions. Consider the problem faced by a hospital that
is trying to evaluate whether its blood supply is sufficient.
The basic rule for blood donation is the following. A persons own
blood supply has certain antigens present (we can think of antigens as a
kind of molecular signature); and a person cannot receive blood with a
particular antigen if their own blood does not have this antigen present.
Concretely, this principle underpins the division of blood into four types :
A, B, AB, and O. Blood of type A has the A antigen, blood of type B has the B
antigen, blood of type AB has both, and blood of type O has neither. Thus,
patients with type A can receive only blood types A or O in a transfusion,
patients with type B can receive only B or O, patients with type O can
receive only O, and patients with type AB can receive any of the four
types.4
(a)
Let sO , sA, sB , and sAB denote the supply in whole units of the different
blood types on hand. Assume that the hospital knows the projected
demand for each blood type dO , dA, dB , and dAB for the coming week.
Give a polynomial-time algorithm to evaluate if the blood on hand
would suffice for the projected need.
(b) Consider the following example. Over the next week, they expect to
need at most 100 units of blood. The typical distribution of blood
types in U.S. patients is roughly 45 percent type O, 42 percent type
A, 10 percent type B, and 3 percent type AB. The hospital wants to
know if the blood supply it has on hand would be enough if 100
patients arrive with the expected type distribution. There is a total
of 105 units of blood on hand. The table below gives these demands,
and the supply on hand.
4
The Austrian scientist Karl Landsteiner received the Nobel Prize in 1930 for his discovery of the
blood types A, B, O, and AB.
Exercises
blood type
supply
demand
O
50
45
A
36
42
B
11
8
AB
8
3
Is the 105 units of blood on hand enough to satisfy the 100 units
of demand? Find an allocation that satisfies the maximum possible
number of patients. Use an argument based on a minimum-capacity
cut to show why not all patients can receive blood. Also, provide an
explanation for this fact that would be understandable to the clinic
administrators, who have not taken a course on algorithms. (So, for
example, this explanation should not involve the words flow , cut , or
graph in the sense we use them in this book.)
9. Network flow issues come up in dealing with natural disasters and other
crises, since major unexpected events often require the movement and
evacuation of large numbers of people in a short amount of time.
Consider the following scenario. Due to large-scale flooding in a region, paramedics have identified a set of n injured people distributed
across the region who need to be rushed to hospitals. There are k hospitals in the region, and each of the n people needs to be brought to a
hospital that is within a half-hours driving time of their current location
(so different people will have different options for hospitals, depending
on where they are right now).
At the same time, one doesnt want to overload any one of the
hospitals by sending too many patients its way. The paramedics are in
touch by cell phone, and they want to collectively work out whether they
can choose a hospital for each of the injured people in such a way that
the load on the hospitals is balanced : Each hospital receives at most n/k
people.
Give a polynomial-time algorithm that takes the given information
about the peoples locations and determines whether this is possible.
10. Suppose you are given a directed graph G = (V , E ), with a positive integer
capacity ce on each edge e , a source s V , and a sink t V . You are also
given a maximum s-t flow in G , defined by a flow value fe on each edge
e . The flow f is acyclic: There is no cycle in G on which all edges carry
positive flow. The flow f is also integer-valued.
419
420
Chapter 7 Network Flow
Now suppose we pick a specific edge e E and reduce its capacity
by 1 unit. Show how to find a maximum flow in the resulting capacitated
graph in time O(m + n), where m is the number of edges in G and n is the
number of nodes.
11. Your friends have written a very fast piece of maximum-flow code based
on repeatedly finding augmenting paths as in Section 7.1. However, after
youve looked at a bit of output from it, you realize that its not always
finding a flow of maximum value. The bug turns out to be pretty easy
to find; your friends hadnt really gotten into the whole backward-edge
thing when writing the code, and so their implementation builds a variant
of the residual graph that only includes the forward edges . In other words,
it searches for s-t paths in a graph Gf consisting only of edges e for which
f (e ) < ce , and it terminates when there is no augmenting path consisting
entirely of such edges. Well call this the Forward-Edge-Only Algorithm.
(Note that we do not try to prescribe how this algorithm chooses its
forward-edge paths; it may choose them in any fashion it wants, provided
that it terminates only when there are no forward-edge paths.)
Its hard to convince your friends they need to reimplement the
code. In addition to its blazing speed, they claim, in fact, that it never
returns a flow whose value is less than a fixed fraction of optimal. Do you
believe this? The crux of their claim can be made precise in the following
statement.
There is an absolute constant b > 1 (independent of the particular input
ow network), so that on every instance of the Maximum-Flow Problem, the
Forward-Edge-Only Algorithm is guaranteed to nd a ow of value at least 1/b
times the maximum-ow value (regardless of how it chooses its forward-edge
paths).
Decide whether you think this statement is true or false, and give a proof
of either the statement or its negation.
12. Consider the following problem. You are given a flow network with unitcapacity edges: It consists of a directed graph G = (V , E ), a source s V ,
and a sink t V ; and ce = 1 for every e E . You are also given a parameter k.
The goal is to delete k edges so as to reduce the maximum s-t flow in
G by as much as possible. In other words, you should find a set of edges
F E so that |F | = k and the maximum s-t flow in G = (V , E F ) is as small
as possible subject to this.
Give a polynomial-time algorithm to solve this problem.
13. In a standard s-t Maximum-Flow Problem, we assume edges have capacities, and there is no limit on how much flow is allowed to pass through a
Exercises
node. In this problem, we consider the variant of the Maximum-Flow and
Minimum-Cut problems with node capacities.
Let G = (V , E ) be a directed graph, with source s V , sink t V , and
nonnegative node capacities {cv 0} for each v V . Given a flow f in this
graph, the flow though a node v is defined as f in (v). We say that a flow
is feasible if it satisfies the usual flow-conservation constraints and the
node-capacity constraints: f in (v) cv for all nodes.
Give a polynomial-time algorithm to find an s-t maximum flow in
such a node-capacitated network. Define an s-t cut for node-capacitated
networks, and show that the analogue of the Max-Flow Min-Cut Theorem
holds true.
14. We define the Escape Problem as follows. We are given a directed graph
G = (V , E ) (picture a network of roads). A certain collection of nodes X V
are designated as populated nodes , and a certain other collection S V
are designated as safe nodes . (Assume that X and S are disjoint.) In case
of an emergency, we want evacuation routes from the populated nodes
to the safe nodes. A set of evacuation routes is defined as a set of paths
in G so that (i) each node in X is the tail of one path, (ii) the last node on
each path lies in S, and (iii) the paths do not share any edges. Such a set of
paths gives a way for the occupants of the populated nodes to escape
to S, without overly congesting any edge in G .
(a)
Given G , X , and S, show how to decide in polynomial time whether
such a set of evacuation routes exists.
(b) Suppose we have exactly the same problem as in (a), but we want to
enforce an even stronger version of the no congestion condition
(iii). Thus we change (iii) to say the paths do not share any nodes .
With this new condition, show how to decide in polynomial time
whether such a set of evacuation routes exists.
Also, provide an example with the same G , X , and S, in which the
answer is yes to the question in (a) but no to the question in (b).
15. Suppose you and your friend Alanis live, together with n 2 other people,
at a popular off-campus cooperative apartment, the Upson Collective.
Over the next n nights, each of you is supposed to cook dinner for the
co-op exactly once, so that someone cooks on each of the nights.
Of course, everyone has scheduling conflicts with some of the nights
(e.g., exams, concerts, etc.), so deciding who should cook on which night
becomes a tricky task. For concreteness, lets label the people
{p1, . . . , pn } ,
421
422
Chapter 7 Network Flow
the nights
{d1, . . . , dn };
and for person pi , theres a set of nights Si {d1, . . . , dn } when they are
not able to cook.
A feasible dinner schedule is an assignment of each person in the coop to a different night, so that each person cooks on exactly one night,
there is someone cooking on each night, and if pi cooks on night dj , then
dj Si .
(a)
Describe a bipartite graph G so that G has a perfect matching if and
only if there is a feasible dinner schedule for the co-op.
(b) Your friend Alanis takes on the task of trying to construct a feasible
dinner schedule. After great effort, she constructs what she claims
is a feasible schedule and then heads off to class for the day.
Unfortunately, when you look at the schedule she created, you
notice a big problem. n 2 of the people at the co-op are assigned to
different nights on which they are available: no problem there. But
for the other two people, pi and pj , and the other two days, dk and
d , you discover that she has accidentally assigned both pi and pj to
cook on night dk , and assigned no one to cook on night d .
You want to fix Alaniss mistake but without having to recompute everything from scratch. Show that its possible, using her almost correct schedule, to decide in only O(n2) time whether there
exists a feasible dinner schedule for the co-op. (If one exists, you
should also output it.)
16. Back in the euphoric early days of the Web, people liked to claim that
much of the enormous potential in a company like Yahoo! was in the
eyeballsthe simple fact that millions of people look at its pages every
day. Further, by convincing people to register personal data with the site,
a site like Yahoo! can show each user an extremely targeted advertisement
whenever he or she visits the site, in a way that TV networks or magazines
couldnt hope to match. So if a user has told Yahoo! that he or she is a
20-year-old computer science major from Cornell University, the site can
present a banner ad for apartments in Ithaca, New York; on the other
hand, if he or she is a 50-year-old investment banker from Greenwich,
Connecticut, the site can display a banner ad pitching Lincoln Town Cars
instead.
But deciding on which ads to show to which people involves some
serious computation behind the scenes. Suppose that the managers
of a popular Web site have identified k distinct demographic groups
Exercises
G1, G2 , . . . , Gk . (These groups can overlap; for example, G1 can be equal to
all residents of New York State, and G2 can be equal to all people with
a degree in computer science.) The site has contracts with m different
advertisers , to show a certain number of copies of their ads to users of
the site. Heres what the contract with the ith advertiser looks like.
.
.
For a subset Xi {G1, . . . , Gk } of the demographic groups, advertiser
i wants its ads shown only to users who belong to at least one of the
demographic groups in the set Xi .
For a number ri , advertiser i wants its ads shown to at least ri users
each minute.
Now consider the problem of designing a good advertising policy
a way to show a single ad to each user of the site. Suppose at a given
minute, there are n users visiting the site. Because we have registration
information on each of these users, we know that user j (for j = 1, 2, . . . , n)
belongs to a subset Uj {G1, . . . , Gk } of the demographic groups. The
problem is: Is there a way to show a single ad to each user so that the sites
contracts with each of the m advertisers is satisfied for minute? this (That
is, for each i = 1, 2, . . . , m, can at least ri of the n users, each belonging
to at least one demographic group in Xi , be shown an ad provided by
advertiser i?)
Give an efficient algorithm to decide if this is possible, and if so, to
actually choose an ad to show each user.
17. Youve been called in to help some network administrators diagnose the
extent of a failure in their network. The network is designed to carry
traffic from a designated source node s to a designated target node t , so
we will model the network as a directed graph G = (V , E ), in which the
capacity of each edge is 1 and in which each node lies on at least one path
from s to t .
Now, when everything is running smoothly in the network, the maximum s-t flow in G has value k. However, the current situation (and the
reason youre here) is that an attacker has destroyed some of the edges in
the network, so that there is now no path from s to t using the remaining
(surviving) edges. For reasons that we wont go into here, they believe
the attacker has destroyed only k edges, the minimum number needed
to separate s from t (i.e., the size of a minimum s-t cut); and well assume
theyre correct in believing this.
The network administrators are running a monitoring tool on node s,
which has the following behavior. If you issue the command ping (v), for
a given node v, it will tell you whether there is currently a path from s
to v. (So ping (t ) reports that no path currently exists; on the other hand,
423
424
Chapter 7 Network Flow
ping (s) always reports a path from s to itself.) Since its not practical to go
out and inspect every edge of the network, theyd like to determine the
extent of the failure using this monitoring tool, through judicious use of
the ping command.
So heres the problem you face: Give an algorithm that issues a
sequence of ping commands to various nodes in the network and then
reports the full set of nodes that are not currently reachable from s. You
could do this by pinging every node in the network, of course, but youd
like to do it using many fewer pings (given the assumption that only
k edges have been deleted). In issuing this sequence, your algorithm is
allowed to decide which node to ping next based on the outcome of earlier
ping operations.
Give an algorithm that accomplishes this task using only O(k log n)
pings.
18. We consider the Bipartite Matching Problem on a bipartite graph G =
(V , E ). As usual, we say that V is partitioned into sets X and Y , and each
edge has one end in X and the other in Y .
If M is a matching in G , we say that a node y Y is covered by M if y
is an end of one of the edges in M .
y1
x1
y2
y3
x2
y4
y5
Figure 7.29 An instance of
Coverage Expansion.
(a) Consider the following problem. We are given G and a matching M in
G . For a given number k, we want to decide if there is a matching M
in G so that
(i) M has k more edges than M does, and
(ii) every node y Y that is covered by M is also covered by M .
We call this the Coverage Expansion Problem , with input G , M , and k.
and we will say that M is a solution to the instance.
Give a polynomial-time algorithm that takes an instance of Coverage Expansion and either returns a solution M or reports (correctly)
that there is no solution. (You should include an analysis of the running time and a brief proof of why it is correct.)
Note: You may wish to also look at part (b) to help in thinking about
this.
Example. Consider Figure 7.29, and suppose M is the matching consisting of the edge (x1, y2). Suppose we are asked the above question
with k = 1.
Then the answer to this instance of Coverage Expansion is yes.
We can let M be the matching consisting (for example) of the two
edges (x1, y2) and (x2 , y4); M has one more edge than M , and y2 is still
covered by M .
Exercises
(b) Give an example of an instance of Coverage Expansion, specified by
G , M , and k, so that the following situation happens.
The instance has a solution; but in any solution M , the edges of M do
not form a subset of the edges of M .
(c) Let G be a bipartite graph, and let M be any matching in G . Consider
the following two quantities.
K1 is the size of the largest matching M so that every node y that
is covered by M is also covered by M .
K2 is the size of the largest matching M in G .
Clearly K1 K2, since K2 is obtained by considering all possible matchings in G .
Prove that in fact K1 = K2; that is, we can obtain a maximum
matching even if were constrained to cover all the nodes covered
by our initial matching M .
19. Youve periodically helped the medical consulting firm Doctors Without
Weekends on various hospital scheduling issues, and theyve just come
to you with a new problem. For each of the next n days, the hospital has
determined the number of doctors they want on hand; thus, on day i,
they have a requirement that exactly pi doctors be present.
There are k doctors, and each is asked to provide a list of days on
which he or she is willing to work. Thus doctor j provides a set Lj of days
on which he or she is willing to work.
The system produced by the consulting firm should take these lists
and try to return to each doctor j a list Lj with the following properties.
(A) Lj is a subset of Lj , so that doctor j only works on days he or she finds
acceptable.
(B) If we consider the whole set of lists L1, . . . , Lk , it causes exactly pi
doctors to be present on day i, for i = 1, 2, . . . , n.
(a)
Describe a polynomial-time algorithm that implements this system.
Specifically, give a polynomial-time algorithm that takes the numbers p1, p2 , . . . , pn , and the lists L1, . . . , Lk , and does one of the following two things.
Return lists L1, L2 , . . . , Lk satisfying properties (A) and (B); or
Report (correctly) that there is no set of lists L1, L2 , . . . , Lk that
satisfies both properties (A) and (B).
(b) The hospital finds that the doctors tend to submit lists that are much
too restrictive, and so it often happens that the system reports (correctly, but unfortunately) that no acceptable set of lists L1, L2 , . . . , Lk
exists.
425
426
Chapter 7 Network Flow
Thus the hospital relaxes the requirements as follows. They add
a new parameter c > 0, and the system now should try to return to
each doctor j a list Lj with the following properties.
(A*) Lj contains at most c days that do not appear on the list Lj .
(B) (Same as before) If we consider the whole set of lists L1, . . . , Lk ,
it causes exactly pi doctors to be present on day i, for i = 1, 2, . . . , n.
Describe a polynomial-time algorithm that implements this revised system. It should take the numbers p1, p2 , . . . , pn , the lists
L1, . . . , Lk , and the parameter c > 0, and do one of the following two
things.
Return lists L1, L2 , . . . , Lk satisfying properties (A) and (B); or
Report (correctly) that there is no set of lists L1, L2 , . . . , Lk that
satisfies both properties (A) and (B).
20. Your friends are involved in a large-scale atmospheric science experiment. They need to get good measurements on a set S of n different
conditions in the atmosphere (such as the ozone level at various places),
and they have a set of m balloons that they plan to send up to make these
measurements. Each balloon can make at most two measurements.
Unfortunately, not all balloons are capable of measuring all conditions, so for each balloon i = 1, . . . , m, they have a set Si of conditions
that balloon i can measure. Finally, to make the results more reliable, they
plan to take each measurement from at least k different balloons. (Note
that a single balloon should not measure the same condition twice.) They
are having trouble figuring out which conditions to measure on which
balloon.
Example. Suppose that k = 2, there are n = 4 conditions labeled c1, c2 , c3 , c4,
and there are m = 4 balloons that can measure conditions, subject to
the limitation that S1 = S2 = {c1, c2 , c3}, and S3 = S4 = {c1, c3 , c4}. Then one
possible way to make sure that each condition is measured at least k = 2
times is to have
.
balloon 1 measure conditions c1, c2,
.
balloon 2 measure conditions c2 , c3,
.
balloon 3 measure conditions c3 , c4, and
.
balloon 4 measure conditions c1, c4.
(a)
Give a polynomial-time algorithm that takes the input to an instance
of this problem (the n conditions, the sets Si for each of the m
balloons, and the parameter k) and decides whether there is a way to
measure each condition by k different balloons, while each balloon
only measures at most two conditions.
Exercises
(b) You show your friends a solution computed by your algorithm from
(a), and to your surprise they reply, This wont do at allone of the
conditions is only being measured by balloons from a single subcontractor. You hadnt heard anything about subcontractors before; it
turns out theres an extra wrinkle they forgot to mention. . . .
Each of the balloons is produced by one of three different subcontractors involved in the experiment. A requirement of the experiment is that there be no condition for which all k measurements
come from balloons produced by a single subcontractor.
For example, suppose balloon 1 comes from the first subcontractor, balloons 2 and 3 come from the second subcontractor, and
balloon 4 comes from the third subcontractor. Then our previous solution no longer works, as both of the measurements for condition
c3 were done by balloons from the second subcontractor. However,
we could use balloons 1 and 2 to each measure conditions c1, c2, and
use balloons 3 and 4 to each measure conditions c3 , c4.
Explain how to modify your polynomial-time algorithm for part
(a) into a new algorithm that decides whether there exists a solution
satisfying all the conditions from (a), plus the new requirement about
subcontractors.
21. Youre helping to organize a class on campus that has decided to give all
its students wireless laptops for the semester. Thus there is a collection
of n wireless laptops; there is also have a collection of n wireless access
points , to which a laptop can connect when it is in range.
The laptops are currently scattered across campus; laptop is within
range of a set S of access points. We will assume that each laptop is within
range of at least one access point (so the sets S are nonempty); we will
also assume that every access point p has at least one laptop within range
of it.
To make sure that all the wireless connectivity software is working
correctly, you need to try having laptops make contact with access points
in such a way that each laptop and each access point is involved in at
least one connection. Thus we will say that a test set T is a collection of
ordered pairs of the form ( , p), for a laptop and access point p, with
the properties that
(i)
If ( , p) T , then
is within range of p (i.e., p S ).
(ii) Each laptop appears in at least one ordered pair in T .
(iii) Each access point appears in at least one ordered pair in T .
427
428
Chapter 7 Network Flow
This way, by trying out all the connections specified by the pairs in T ,
we can be sure that each laptop and each access point have correctly
functioning software.
The problem is: Given the sets S for each laptop (i.e., which laptops
are within range of which access points), and a number k, decide whether
there is a test set of size at most k.
Example. Suppose that n = 3; laptop 1 is within range of access points 1
and 2; laptop 2 is within range of access point 2; and laptop 3 is within
range of access points 2 and 3. Then the set of pairs
(laptop 1, access point 1), (laptop 2, access point 2),
(laptop 3, access point 3)
would form a test set of size 3.
(a)
Give an example of an instance of this problem for which there is no
test set of size n. (Recall that we assume each laptop is within range
of at least one access point, and each access point p has at least one
laptop within range of it.)
(b) Give a polynomial-time algorithm that takes the input to an instance
of this problem (including the parameter k) and decides whether
there is a test set of size at most k.
22. Let M be an n n matrix with each entry equal to either 0 or 1. Let mij
denote the entry in row i and column j . A diagonal entry is one of the
form mii for some i.
Swapping rows i and j of the matrix M denotes the following action:
we swap the values mik and mjk for k = 1, 2, . . . , n. Swapping two columns
is defined analogously.
We say that M is rearrangeable if it is possible to swap some of the
pairs of rows and some of the pairs of columns (in any sequence) so that,
after all the swapping, all the diagonal entries of M are equal to 1.
(a)
Give an example of a matrix M that is not rearrangeable, but for
which at least one entry in each row and each column is equal to 1.
(b) Give a polynomial-time algorithm that determines whether a matrix
M with 0-1 entries is rearrangeable.
23. Suppose youre looking at a flow network G with source s and sink t , and
you want to be able to express something like the following intuitive notion: Some nodes are clearly on the source side of the main bottlenecks;
some nodes are clearly on the sink side of the main bottlenecks; and
some nodes are in the middle. However, G can have many minimum cuts,
so we have to be careful in how we try making this idea precise.
Exercises
Heres one way to divide the nodes of G into three categories of this
sort.
.
.
.
We say a node v is upstream if, for all minimum s-t cuts (A, B), we
have v Athat is, v lies on the source side of every minimum cut.
We say a node v is downstream if, for all minimum s-t cuts (A, B), we
have v Bthat is, v lies on the sink side of every minimum cut.
We say a node v is central if it is neither upstream nor downstream;
there is at least one minimum s-t cut (A, B) for which v A, and at
least one minimum s-t cut (A , B ) for which v B .
Give an algorithm that takes a flow network G and classifies each of
its nodes as being upstream, downstream, or central. The running time
of your algorithm should be within a constant factor of the time required
to compute a single maximum flow.
24. Let G = (V , E ) be a directed graph, with source s V , sink t V , and
nonnegative edge capacities {ce }. Give a polynomial-time algorithm to
decide whether G has a unique minimum s-t cut (i.e., an s-t of capacity
strictly less than that of all other s-t cuts).
25. Suppose you live in a big apartment with a lot of friends. Over the
course of a year, there are many occasions when one of you pays for an
expense shared by some subset of the apartment, with the expectation
that everything will get balanced out fairly at the end of the year. For
example, one of you may pay the whole phone bill in a given month,
another will occasionally make communal grocery runs to the nearby
organic food emporium, and a third might sometimes use a credit card
to cover the whole bill at the local Italian-Indian restaurant, Little Idli.
In any case, its now the end of the year and time to settle up. There
are n people in the apartment; and for each ordered pair (i , j ) theres an
amount aij 0 that i owes j , accumulated over the course of the year. We
will require that for any two people i and j , at least one of the quantities
aij or aji is equal to 0. This can be easily made to happen as follows: If it
turns out that i owes j a positive amount x, and j owes i a positive amount
y < x, then we will subtract off y from both sides and declare aij = x y
while aji = 0. In terms of all these quantities, we now define the imbalance
of a person i to be the sum of the amounts that i is owed by everyone
else, minus the sum of the amounts that i owes everyone else. (Note that
an imbalance can be positive, negative, or zero.)
In order to restore all imbalances to 0, so that everyone departs on
good terms, certain people will write checks to others; in other words, for
certain ordered pairs (i , j ), i will write a check to j for an amount bij > 0.
429
430
Chapter 7 Network Flow
We will say that a set of checks constitutes a reconciliation if, for each
person i, the total value of the checks received by i, minus the total value
of the checks written by i, is equal to the imbalance of i. Finally, you and
your friends feel it is bad form for i to write j a check if i did not actually
owe j money, so we say that a reconciliation is consistent if, whenever i
writes a check to j , it is the case that aij > 0.
Show that, for any set of amounts aij , there is always a consistent
reconciliation in which at most n 1 checks get written, by giving a
polynomial-time algorithm to compute such a reconciliation.
26. You can tell that cellular phones are at work in rural communities, from
the giant microwave towers you sometimes see sprouting out of corn
fields and cow pastures. Lets consider a very simplified model of a
cellular phone network in a sparsely populated area.
We are given the locations of n base stations , specified as points
b1, . . . , bn in the plane. We are also given the locations of n cellular phones,
specified as points p1, . . . , pn in the plane. Finally, we are given a range
parameter
> 0. We call the set of cell phones fully connected if it is
possible to assign each phone to a base station in such a way that
.
.
Each phone is assigned to a different base station, and
If a phone at pi is assigned to a base station at bj , then the straight-line
distance between the points pi and bj is at most .
Suppose that the owner of the cell phone at point p1 decides to go
for a drive, traveling continuously for a total of z units of distance due
east. As this cell phone moves, we may have to update the assignment of
phones to base stations (possibly several times) in order to keep the set
of phones fully connected.
Give a polynomial-time algorithm to decide whether it is possible to
keep the set of phones fully connected at all times during the travel of
this one cell phone. (You should assume that all other phones remain stationary during this travel.) If it is possible, you should report a sequence
of assignments of phones to base stations that will be sufficient in order
to maintain full connectivity; if it is not possible, you should report a
point on the traveling phones path at which full connectivity cannot be
maintained.
You should try to make your algorithm run in O(n3) time if possible.
Example. Suppose we have phones at p1 = (0, 0) and p2 = (2, 1); we have
base stations at b1 = (1, 1) and b2 = (3, 1); and = 2. Now consider the case
in which the phone at p1 moves due east a distance of 4 units, ending at
(4, 0). Then it is possible to keep the phones fully connected during this
Exercises
motion: We begin by assigning p1 to b1 and p2 to b2, and we reassign p1 to
b2 and p2 to b1 during the motion (for example, when p1 passes the point
(2, 0)).
27. Some of your friends with jobs out West decide they really need some
extra time each day to sit in front of their laptops, and the morning
commute from Woodside to Palo Alto seems like the only option. So they
decide to carpool to work.
Unfortunately, they all hate to drive, so they want to make sure that
any carpool arrangement they agree upon is fair and doesnt overload
any individual with too much driving. Some sort of simple round-robin
scheme is out, because none of them goes to work every day, and so the
subset of them in the car varies from day to day.
Heres one way to define fairness . Let the people be labeled S =
{p1, . . . , pk }. We say that the total driving obligation of pj over a set of
days is the expected number of times that pj would have driven, had a
driver been chosen uniformly at random from among the people going
to work each day. More concretely, suppose the carpool plan lasts for d
days, and on the ith day a subset Si S of the people go to work. Then the
above definition of the total driving obligation j for pj can be written as
1
j=
i:pj Si |Si | . Ideally, wed like to require that pj drives at most j times;
unfortunately, j may not be an integer.
So lets say that a driving schedule is a choice of a driver for each
daythat is, a sequence pi1 , pi2 , . . . , pid with pit St and that a fair driving
schedule is one in which each pj is chosen as the driver on at most
j
days.
(a)
Prove that for any sequence of sets S1, . . . , Sd , there exists a fair
driving schedule.
(b) Give an algorithm to compute a fair driving schedule with running
time polynomial in k and d.
28. A group of students has decided to add some features to Cornells on-line
Course Management System (CMS), to handle aspects of course planning
that are not currently covered by the software. Theyre beginning with a
module that helps schedule office hours at the start of the semester.
Their initial prototype works as follows. The office hour schedule will
be the same from one week to the next, so its enough to focus on the
scheduling problem for a single week. The course administrator enters
a collection of nonoverlapping one-hour time intervals I1, I2 , . . . , Ik when
it would be possible for teaching assistants (TAs) to hold office hours;
the eventual office-hour schedule will consist of a subset of some, but
431
432
Chapter 7 Network Flow
generally not all, of these time slots. Then each of the TAs enters his or
her weekly schedule, showing the times when he or she would be available
to hold office hours.
Finally, the course administrator specifies, for parameters a , b, and
c , that they would like each TA to hold between a and b office hours per
week, and they would like a total of exactly c office hours to be held over
the course of the week.
The problem, then, is how to assign each TA to some of the officehour time slots, so that each TA is available for each of his or her officehour slots, and so that the right number of office hours gets held. (There
should be only one TA at each office hour.)
Example. Suppose there are five possible time slots for office hours:
I1 = Mon 34 P.M.; I2 = Tue 12 P.M.; I3 = Wed 1011 A.M.; I4 = Wed 34
P.M.; and I5 = Thu 1011 A.M..
There are two TAs; the first would be able to hold office hours at any
time on Monday or Wednesday afternoons, and the second would be able
to hold office hours at any time on Tuesday, Wednesday, or Thursday.
(In general, TA availability might be more complicated to specify than
this, but were keeping this example simple.) Finally, each TA should hold
between a = 1 and b = 2 office hours, and we want exactly c = 3 office hours
per week total.
One possible solution would be to have the first TA hold office hours
in time slot I1, and the second TA to hold office hours in time slots I2
and I5.
(a)
Give a polynomial-time algorithm that takes the input to an instance
of this problem (the time slots, the TA schedules, and the parameters
a , b, and c ) and does one of the following two things:
Constructs a valid schedule for office hours, specifying which
TA will cover which time slots, or
Reports (correctly) that there is no valid way to schedule office
hours.
(b) This office-hour scheduling feature becomes very popular, and so
course staffs begin to demand more. In particular, they observe that
its good to have a greater density of office hours closer to the due
date of a homework assignment.
So what they want to be able to do is to specify an office-hour
density parameter for each day of the week: The number di specifies
that they want to have at least di office hours on a given day i of the
week.
Exercises
For example, suppose that in our previous example, we add the
constraint that we want at least one office hour on Wednesday and at
least one office hour on Thursday. Then the previous solution does
not work; but there is a possible solution in which we have the first
TA hold office hours in time slot I1, and the second TA hold office
hours in time slots I3 and I5. (Another solution would be to have the
first TA hold office hours in time slots I1 and I4, and the second TA
hold office hours in time slot I5.)
Give a polynomial-time algorithm that computes office-hour
schedules under this more complex set of constraints. The algorithm should either construct a schedule or report (correctly) that
none exists.
29. Some of your friends have recently graduated and started a small company, which they are currently running out of their parents garages in
Santa Clara. Theyre in the process of porting all their software from an
old system to a new, revved-up system; and theyre facing the following
problem.
They have a collection of n software applications, {1, 2, . . . , n}, running on their old system; and theyd like to port some of these to the new
system. If they move application i to the new system, they expect a net
(monetary) benefit of bi 0. The different software applications interact
with one another; if applications i and j have extensive interaction, then
the company will incur an expense if they move one of i or j to the new
system but not both; lets denote this expense by xij 0.
So, if the situation were really this simple, your friends would just
port all n applications, achieving a total benefit of i bi . Unfortunately,
theres a problem. . . .
Due to small but fundamental incompatibilities between the two
systems, theres no way to port application 1 to the new system; it will
have to remain on the old system. Nevertheless, it might still pay off to
port some of the other applications, accruing the associated benefit and
incurring the expense of the interaction between applications on different
systems.
So this is the question they pose to you: Which of the remaining
applications, if any, should be moved? Give a polynomial-time algorithm
to find a set S {2, 3, . . . , n} for which the sum of the benefits minus the
expenses of moving the applications in S to the new system is maximized.
30. Consider a variation on the previous problem. In the new scenario, any
application can potentially be moved, but now some of the benefits bi for
433
434
Chapter 7 Network Flow
moving to the new system are in fact negative : If bi < 0, then it is preferable
(by an amount quantified in bi ) to keep i on the old system. Again, give
a polynomial-time algorithm to find a set S {1, 2, . . . , n} for which the
sum of the benefits minus the expenses of moving the applications in S
to the new system is maximized.
31. Some of your friends are interning at the small high-tech company WebExodus. A running joke among the employees there is that the back room
has less space devoted to high-end servers than it does to empty boxes
of computer equipment, piled up in case something needs to be shipped
back to the supplier for maintainence.
A few days ago, a large shipment of computer monitors arrived, each
in its own large box; and since there are many different kinds of monitors
in the shipment, the boxes do not all have the same dimensions. A bunch
of people spent some time in the morning trying to figure out how to
store all these things, realizing of course that less space would be taken
up if some of the boxes could be nested inside others.
Suppose each box i is a rectangular parallelepiped with side lengths
equal to (i1, i2 , i3); and suppose each side length is strictly between half a
meter and one meter. Geometrically, you know what it means for one box
to nest inside another: Its possible if you can rotate the smaller so that
it fits inside the larger in each dimension. Formally, we can say that box
i with dimensions (i1, i2 , i3) nests inside box j with dimensions (j1, j2 , j3) if
there is a permutation a , b , c of the dimensions {1, 2, 3} so that ia < j1, and
ib < j2, and ic < j3. Of course, nesting is recursive: If i nests in j , and j nests
in k, then by putting i inside j inside k, only box k is visible. We say that
a nesting arrangement for a set of n boxes is a sequence of operations
in which a box i is put inside another box j in which it nests; and if there
were already boxes nested inside i, then these end up inside j as well.
(Also notice the following: Since the side lengths of i are more than half
a meter each, and since the side lengths of j are less than a meter each,
box i will take up more than half of each dimension of j , and so after i is
put inside j , nothing else can be put inside j .) We say that a box k is visible
in a nesting arrangement if the sequence of operations does not result in
its ever being put inside another box.
Here is the problem faced by the people at WebExodus: Since only the
visible boxes are taking up any space, how should a nesting arrangement
be chosen so as to minimize the number of visible boxes?
Give a polynomial-time algorithm to solve this problem.
Example. Suppose there are three boxes with dimensions (.6, .6, .6),
(.75, .75, .75), and (.9, .7, .7). The first box can be put into either of the
Exercises
second or third boxes; but in any nesting arrangement, both the second
and third boxes will be visible. So the minimum possible number of visible boxes is two, and one solution that achieves this is to nest the first
box inside the second.
32. Given a graph G = (V , E ), and a natural number k, we can define a relation
G, k
G, k
on pairs of vertices of G as follows. If x , y V , we say that x y if
there exist k mutually edge-disjoint paths from x to y in G .
G, k
Is it true that for every G and every k 0, the relation is transitive?
G, k
G, k
G, k
That is, is it always the case that if x y and y z , then we have x z ?
Give a proof or a counterexample.
33. Let G = (V , E ) be a directed graph, and suppose that for each node v, the
number of edges into v is equal to the number of edges out of v. That is,
for all v,
|{(u , v) : (u , v) E }| = |{(v , w) : (v , w) E }|.
Let x , y be two nodes of G , and suppose that there exist k mutually edgedisjoint paths from x to y. Under these conditions, does it follow that
there exist k mutually edge-disjoint paths from y to x? Give a proof or a
counterexample with explanation.
34. Ad hoc networks , made up of low-powered wireless devices, have been
proposed for situations like natural disasters in which the coordinators
of a rescue effort might want to monitor conditions in a hard-to-reach
area. The idea is that a large collection of these wireless devices could be
dropped into such an area from an airplane and then configured into a
functioning network.
Note that were talking about (a) relatively inexpensive devices that
are (b) being dropped from an airplane into (c) dangerous territory; and
for the combination of reasons (a), (b), and (c), it becomes necessary to
include provisions for dealing with the failure of a reasonable number of
the nodes.
Wed like it to be the case that if one of the devices v detects that it is in
danger of failing, it should transmit a representation of its current state to
some other device in the network. Each device has a limited transmitting
rangesay it can communicate with other devices that lie within d meters
of it. Moreover, since we dont want it to try transmitting its state to a
device that has already failed, we should include some redundancy: A
device v should have a set of k other devices that it can potentially contact,
each within d meters of it. Well call this a back-up set for device v.
435
436
Chapter 7 Network Flow
(a)
Suppose youre given a set of n wireless devices, with positions
represented by an (x , y) coordinate pair for each. Design an algorithm
that determines whether it is possible to choose a back-up set for
each device (i.e., k other devices, each within d meters), with the
further property that, for some parameter b, no device appears in
the back-up set of more than b other devices. The algorithm should
output the back-up sets themselves, provided they can be found.
(b) The idea that, for each pair of devices v and w, theres a strict
dichotomy between being in range or out of range is a simplified
abstraction. More accurately, theres a power decay function f () that
specifies, for a pair of devices at distance , the signal strength f ()
that theyll be able to achieve on their wireless connection. (Well
assume that f () decreases with increasing .)
We might want to build this into our notion of back-up sets as
follows: among the k devices in the back-up set of v, there should
be at least one that can be reached with very high signal strength,
at least one other that can be reached with moderately high signal
strength, and so forth. More concretely, we have values p1 p2 . . .
pk , so that if the back-up set for v consists of devices at distances
d1 d2 . . . dk , then we should have f (dj ) pj for each j .
Give an algorithm that determines whether it is possible to
choose a back-up set for each device subject to this more detailed
condition, still requiring that no device should appear in the back-up
set of more than b other devices. Again, the algorithm should output
the back-up sets themselves, provided they can be found.
35. Youre designing an interactive image segmentation tool that works as
follows. You start with the image segmentation setup described in Section
7.10, with n pixels, a set of neighboring pairs, and parameters {ai }, {bi },
and {pij }. We will make two assumptions about this instance. First, we will
suppose that each of the parameters {ai }, {bi }, and {pij } is a nonnegative
integer between 0 and d, for some number d. Second, we will suppose that
the neighbor relation among the pixels has the property that each pixel
is a neighbor of at most four other pixels (so in the resulting graph, there
are at most four edges out of each node).
You first perform an initial segmentation (A0 , B0) so as to maximize
the quantity q(A0 , B0). Now, this might result in certain pixels being
assigned to the background when the user knows that they ought to be
in the foreground. So, when presented with the segmentation, the user
has the option of mouse-clicking on a particular pixel v1, thereby bringing
it to the foreground. But the tool should not simply bring this pixel into
Exercises
the foreground; rather, it should compute a segmentation (A1, B1) that
maximizes the quantity q(A1, B1) subject to the condition that v1 is in the
foreground. (In practice, this is useful for the following kind of operation:
In segmenting a photo of a group of people, perhaps someone is holding
a bag that has been accidentally labeled as part of the background. By
clicking on a single pixel belonging to the bag, and recomputing an
optimal segmentation subject to the new condition, the whole bag will
often become part of the foreground.)
In fact, the system should allow the user to perform a sequence
of such mouse-clicks v1, v2 , . . . , vt ; and after mouse-click vi , the system should produce a segmentation (Ai , Bi ) that maximizes the quantity
q(Ai , Bi ) subject to the condition that all of v1, v2 , . . . , vi are in the foreground.
Give an algorithm that performs these operations so that the initial
segmentation is computed within a constant factor of the time for a single
maximum flow, and then the interaction with the user is handled in O(dn)
time per mouse-click.
(Note: Solved Exercise 1 from this chapter is a useful primitive for
doing this. Also, the symmetric operation of forcing a pixel to belong to
the background can be handled by analogous means, but you do not have
to work this out here.)
36. We now consider a different variation of the image segmentation problem
in Section 7.10. We will develop a solution to an image labeling problem,
where the goal is to label each pixel with a rough estimate of its distance
from the camera (rather than the simple foreground/background labeling
used in the text). The possible labels for each pixel will be 0, 1, 2, . . . , M
for some integer M .
Let G = (V , E ) denote the graph whose nodes are pixels, and edges
indicate neighboring pairs of pixels. A labeling of the pixels is a partition
of V into sets A0 , A1, . . . , AM , where Ak is the set of pixels that is labeled
with distance k for k = 0, . . . , M . We will seek a labeling of minimum cost ;
the cost will come from two types of terms. By analogy with the foreground/background segmentation problem, we will have an assignment
cost : for each pixel i and label k, the cost ai , k is the cost of assigning label
k to pixel i. Next, if two neighboring pixels (i , j ) E are assigned different
labels, there will be a separation cost. In Section 7.10, we used a separation penalty pij . In our current problem, the separation cost will also
depend on how far the two pixels are separated; specifically, it will be
proportional to the difference in value between their two labels.
Thus the overall cost q of a labeling is defined as follows:
437
438
Chapter 7 Network Flow
ai,1
ai,0
vi,1
s
L
ai,2
vi,2
L
ai,3
ai,4
vi,3
L
vi,4
L
ai,5
vi,5
L
t
L
Figure 7.30 The set of nodes corresponding to a single pixel i in Exercise 36 (shown
together with the source s and sink t ).
M
q (A0 , . . . , AM ) =
ai , k +
k=0 iAi
( k)pij .
k<
(i , j )E
iAk , j A
The goal of this problem is to develop a polynomial-time algorithm
that finds the optimal labeling given the graph G and the penalty parameters ai , k and pij . The algorithm will be based on constructing a flow
network, and we will start you off on designing the algorithm by providing
a portion of the construction.
The flow network will have a source s and a sink t . In addition, for
each pixel i V we will have nodes vi , k in the flow network for k = 1, . . . , M ,
as shown in Figure 7.30. (M = 5 in the example in the figure.)
For notational convenience, the nodes vi ,0 and vi , M +1 will refer to s
and t , respectively, for any choice of i V .
We now add edges (vi , k , vi , k+1) with capacity ai , k for k = 0, . . . , M ; and
edges (vi , k+1, vi , k ) in the opposite direction with very large capacity L. We
will refer to this collection of nodes and edges as the chain associated
with pixel i.
Notice that if we make this very large capacity L large enough, then
there will be no minimum cut (A, B) so that an edge of capacity L leaves
the set A. (How large do we have to make it for this to happen?). Hence, for
any minimum cut (A, B), and each pixel i, there will be exactly one lowcapacity edge in the chain associated with i that leaves the set A. (You
should check that if there were two such edges, then a large-capacity
edge would also have to leave the set A.)
Finally, heres the question: Use the nodes and edges defined so far
to complete the construction of a flow network with the property that a
minimum-cost labeling can be efficiently computed from a minimum s-t
cut. You should prove that your construction has the desired property,
and show how to recover the minimum-cost labeling from the cut.
37. In a standard minimum s-t cut problem, we assume that all capacities are
nonnegative; allowing an arbitrary set of positive and negative capacities
results in a problem that is computationally much more difficult. How-
Exercises
ever, as well see here, it is possible to relax the nonnegativity requirement
a little and still have a problem that can be solved in polynomial time.
Let G = (V , E ) be a directed graph, with source s V , sink t V , and
edge capacities {ce }. Suppose that for every edge e that has neither s nor t
as an endpoint, we have ce 0. Thus ce can be negative for edges e that have
at least one end equal to either s or t . Give a polynomial-time algorithm
to find an s-t cut of minimum value in such a graph. (Despite the new
nonnegativity requirements, we still define the value of an s-t cut (A, B)
to be the sum of the capacities of all edges e for which the tail of e is in
A and the head of e is in B.)
38. Youre working with a large database of employee records. For the purposes of this question, well picture the database as a two-dimensional
table T with a set R of m rows and a set C of n columns; the rows correspond to individual employees, and the columns correspond to different
attributes.
To take a simple example, we may have four columns labeled
name,
phone number ,
start date,
manager s name
and a table with five employees as shown here.
name
phone number
start date
managers name
Alanis
3-4563
6/13/95
Chelsea
Chelsea
3-2341
1/20/93
Lou
Elrond
3-2345
12/19/01
Chelsea
Hal
3-9000
1/12/97
Chelsea
Raj
3-3453
7/1/96
Chelsea
Given a subset S of the columns, we can obtain a new, smaller table
by keeping only the entries that involve columns from S. We will call this
new table the projection of T onto S, and denote it by T [S]. For example,
if S = {name, start date}, then the projection T [S] would be the table
consisting of just the first and third columns.
Theres a different operation on tables that is also useful, which is
to permute the columns. Given a permutation p of the columns, we can
obtain a new table of the same size as T by simply reordering the columns
according to p. We will call this new table the permutation of T by p, and
denote it by Tp .
All of this comes into play for your particular application, as follows.
You have k different subsets of the columns S1, S2 , . . . , Sk that youre
439
440
Chapter 7 Network Flow
going to be working with a lot, so youd like to have them available in a
readily accessible format. One choice would be to store the k projections
T [S1], T [S2], . . . , T [Sk ], but this would take up a lot of space. In considering
alternatives to this, you learn that you may not need to explicitly project
onto each subset, because the underlying database system can deal with
a subset of the columns particularly efficiently if (in some order) the
members of the subset constitute a prefix of the columns in left-to-right
order. So, in our example, the subsets {name, phone number} and {name,
start date, phone number,} constitute prefixes (theyre the first two and
first three columns from the left, respectively); and as such, they can
be processed much more efficiently in this table than a subset such as
{name, start date}, which does not constitute a prefix. (Again, note that
a given subset Si does not come with a specified order, and so we are
interested in whether there is some order under which it forms a prefix
of the columns.)
So heres the question: Given a parameter < k, can you find permutations of the columns p1, p2 , . . . , p so that for every one of the given
subsets Si (for i = 1, 2, . . . , k), its the case that the columns in Si constitute a prefix of at least one of the permuted tables Tp1 , Tp2 , . . . , Tp ? Well
say that such a set of permutations constitutes a valid solution to the
problem; if a valid solution exists, it means you only need to store the
permuted tables rather than all k projections. Give a polynomial-time
algorithm to solve this problem; for instances on which there is a valid
solution, your algorithm should return an appropriate set of permutations.
Example. Suppose the table is as above, the given subsets are
S1 = {name, phone number} ,
S2 = {name, start date} ,
S3 = {name, manager s name, start date} ,
and = 2. Then there is a valid solution to the instance, and it could be
achieved by the two permutations
p1 = {name, phone number , start date, manager s name} ,
p2 = {name, start date, manager s name, phone number}.
This way, S1 constitutes a prefix of the permuted table Tp1, and both S2
and S3 constitute prefixes of the permuted table Tp2 .
39. You are consulting for an environmental statistics firm. They collect
statistics and publish the collected data in a book. The statistics are
about populations of different regions in the world and are recorded in
Exercises
multiples of one million. Examples of such statistics would look like the
following table.
Country
A
B
C
Total
grown-up men
11.998
9.083
2.919
24.000
grown-up women
12.983
10.872
3.145
27.000
1.019
2.045
0.936
4.000
26.000
22.000
7.000
55.000
children
Total
We will assume here for simplicity that our data is such that all
row and column sums are integers. The Census Rounding Problem is to
round all data to integers without changing any row or column sum. Each
fractional number can be rounded either up or down. For example, a good
rounding for our table data would be as follows.
Country
A
B
C
Total
grown-up men
10.000
3.000
24.000
13.000
10.000
4.000
27.000
children
2.000
2.000
0.000
4.000
Total
(a)
11.000
grown-up women
26.000
22.000
7.000
55.000
Consider first the special case when all data are between 0 and 1.
So you have a matrix of fractional numbers between 0 and 1, and
your problem is to round each fraction that is between 0 and 1 to
either 0 or 1 without changing the row or column sums. Use a flow
computation to check if the desired rounding is possible.
(b) Consider the Census Rounding Problem as defined above, where row
and column sums are integers, and you want to round each fractional
number to either or . Use a flow computation to check if the
desired rounding is possible.
(c)
Prove that the rounding we are looking for in (a) and (b) always exists.
40. In a lot of numerical computations, we can ask about the stability
or robustness of the answer. This kind of question can be asked for
combinatorial problems as well; heres one way of phrasing the question
for the Minimum Spanning Tree Problem.
Suppose you are given a graph G = (V , E ), with a cost ce on each edge e .
We view the costs as quantities that have been measured experimentally,
subject to possible errors in measurement. Thus, the minimum spanning
441
442
Chapter 7 Network Flow
tree one computes for G may not in fact be the real minimum spanning
tree.
Given error parameters > 0 and k > 0, and a specific edge e = (u , v),
you would like to be able to make a claim of the following form.
() Even if the cost of each edge were to be changed by at most (either
increased or decreased), and the costs of k of the edges other than e were
further changed to arbitrarily different values, the edge e would still not belong
to any minimum spanning tree of G.
Such a property provides a type of guarantee that e is not likely to belong
to the minimum spanning tree, even assuming significant measurement
error.
Give a polynomial-time algorithm that takes G , e , , and k, and decides
whether or not property () holds for e .
41. Suppose youre managing a collection of processors and must schedule
a sequence of jobs over time.
The jobs have the following characteristics. Each job j has an arrival
time aj when it is first available for processing, a length j which indicates
how much processing time it needs, and a deadline dj by which it must
be finished. (Well assume 0 < j dj aj .) Each job can be run on any
of the processors, but only on one at a time; it can also be preempted
and resumed from where it left off (possibly after a delay) on another
processor.
Moreover, the collection of processors is not entirely static either:
You have an overall pool of k po