首页资源分类嵌入式开发其它 > Introduction to Algorithms 3rd

Introduction to Algorithms 3rd

已有 451217个资源

下载专区

上传者其他资源

    文档信息举报收藏

    标    签:Algorithms

    分    享:

    文档简介

     Introduction to Algorithms 3rd

    文档预览

    T H O M A S H. C O R M E N C H A R L E S E. L E I S E R S O N R O N A L D L. RIVEST C L I F F O R D STEIN INTRODUCTION TO ALGORITHMS THIRD EDITION Introduction to Algorithms Third Edition Thomas H. Cormen Charles E. Leiserson Ronald L. Rivest Clifford Stein Introduction to Algorithms Third Edition The MIT Press Cambridge, Massachusetts London, England c 2009 Massachusetts Institute of Technology All rights reserved. No part of this book may be reproduced in any form or by any electronic or mechanical means (including photocopying, recording, or information storage and retrieval) without permission in writing from the publisher. For information about special quantity discounts, please email special sales@mitpress.mit.edu. This book was set in Times Roman and Mathtime Pro 2 by the authors. Printed and bound in the United States of America. Library of Congress Cataloging-in-Publication Data Introduction to algorithms / Thomas H. Cormen . . . [et al.].—3rd ed. p. cm. Includes bibliographical references and index. ISBN 978-0-262-03384-8 (hardcover : alk. paper)—ISBN 978-0-262-53305-8 (pbk. : alk. paper) 1. Computer programming. 2. Computer algorithms. I. Cormen, Thomas H. QA76.6.I5858 2009 005.1—dc22 2009008593 10 9 8 7 6 5 4 3 2 Contents Preface xiii I Foundations Introduction 3 1 The Role of Algorithms in Computing 5 1.1 Algorithms 5 1.2 Algorithms as a technology 11 2 Getting Started 16 2.1 Insertion sort 16 2.2 Analyzing algorithms 23 2.3 Designing algorithms 29 3 Growth of Functions 43 3.1 Asymptotic notation 43 3.2 Standard notations and common functions 53 4 Divide-and-Conquer 65 4.1 The maximum-subarray problem 68 4.2 Strassen’s algorithm for matrix multiplication 75 4.3 The substitution method for solving recurrences 83 4.4 The recursion-tree method for solving recurrences 88 4.5 The master method for solving recurrences 93 ? 4.6 Proof of the master theorem 97 5 Probabilistic Analysis and Randomized Algorithms 114 5.1 The hiring problem 114 5.2 Indicator random variables 118 5.3 Randomized algorithms 122 ? 5.4 Probabilistic analysis and further uses of indicator random variables 130 vi Contents II Sorting and Order Statistics Introduction 147 6 Heapsort 151 6.1 Heaps 151 6.2 Maintaining the heap property 154 6.3 Building a heap 156 6.4 The heapsort algorithm 159 6.5 Priority queues 162 7 Quicksort 170 7.1 Description of quicksort 170 7.2 Performance of quicksort 174 7.3 A randomized version of quicksort 179 7.4 Analysis of quicksort 180 8 Sorting in Linear Time 191 8.1 Lower bounds for sorting 191 8.2 Counting sort 194 8.3 Radix sort 197 8.4 Bucket sort 200 9 Medians and Order Statistics 213 9.1 Minimum and maximum 214 9.2 Selection in expected linear time 215 9.3 Selection in worst-case linear time 220 III Data Structures Introduction 229 10 Elementary Data Structures 232 10.1 Stacks and queues 232 10.2 Linked lists 236 10.3 Implementing pointers and objects 241 10.4 Representing rooted trees 246 11 Hash Tables 253 11.1 Direct-address tables 254 11.2 Hash tables 256 11.3 Hash functions 262 11.4 Open addressing 269 ? 11.5 Perfect hashing 277 Contents vii 12 Binary Search Trees 286 12.1 What is a binary search tree? 286 12.2 Querying a binary search tree 289 12.3 Insertion and deletion 294 ? 12.4 Randomly built binary search trees 299 13 Red-Black Trees 308 13.1 Properties of red-black trees 308 13.2 Rotations 312 13.3 Insertion 315 13.4 Deletion 323 14 Augmenting Data Structures 339 14.1 Dynamic order statistics 339 14.2 How to augment a data structure 345 14.3 Interval trees 348 IV Advanced Design and Analysis Techniques Introduction 357 15 Dynamic Programming 359 15.1 Rod cutting 360 15.2 Matrix-chain multiplication 370 15.3 Elements of dynamic programming 378 15.4 Longest common subsequence 390 15.5 Optimal binary search trees 397 16 Greedy Algorithms 414 16.1 An activity-selection problem 415 16.2 Elements of the greedy strategy 423 16.3 Huffman codes 428 ? 16.4 Matroids and greedy methods 437 ? 16.5 A task-scheduling problem as a matroid 443 17 Amortized Analysis 451 17.1 Aggregate analysis 452 17.2 The accounting method 456 17.3 The potential method 459 17.4 Dynamic tables 463 viii Contents V Advanced Data Structures Introduction 481 18 B-Trees 484 18.1 Definition of B-trees 488 18.2 Basic operations on B-trees 491 18.3 Deleting a key from a B-tree 499 19 Fibonacci Heaps 505 19.1 Structure of Fibonacci heaps 507 19.2 Mergeable-heap operations 510 19.3 Decreasing a key and deleting a node 518 19.4 Bounding the maximum degree 523 20 van Emde Boas Trees 531 20.1 Preliminary approaches 532 20.2 A recursive structure 536 20.3 The van Emde Boas tree 545 21 Data Structures for Disjoint Sets 561 21.1 Disjoint-set operations 561 21.2 Linked-list representation of disjoint sets 564 21.3 Disjoint-set forests 568 ? 21.4 Analysis of union by rank with path compression 573 VI Graph Algorithms Introduction 587 22 Elementary Graph Algorithms 589 22.1 Representations of graphs 589 22.2 Breadth-first search 594 22.3 Depth-first search 603 22.4 Topological sort 612 22.5 Strongly connected components 615 23 Minimum Spanning Trees 624 23.1 Growing a minimum spanning tree 625 23.2 The algorithms of Kruskal and Prim 631 Contents ix 24 Single-Source Shortest Paths 643 24.1 The Bellman-Ford algorithm 651 24.2 Single-source shortest paths in directed acyclic graphs 655 24.3 Dijkstra’s algorithm 658 24.4 Difference constraints and shortest paths 664 24.5 Proofs of shortest-paths properties 671 25 All-Pairs Shortest Paths 684 25.1 Shortest paths and matrix multiplication 686 25.2 The Floyd-Warshall algorithm 693 25.3 Johnson’s algorithm for sparse graphs 700 26 Maximum Flow 708 26.1 Flow networks 709 26.2 The Ford-Fulkerson method 714 26.3 Maximum bipartite matching 732 ? 26.4 Push-relabel algorithms 736 ? 26.5 The relabel-to-front algorithm 748 VII Selected Topics Introduction 769 27 Multithreaded Algorithms 772 27.1 The basics of dynamic multithreading 774 27.2 Multithreaded matrix multiplication 792 27.3 Multithreaded merge sort 797 28 Matrix Operations 813 28.1 Solving systems of linear equations 813 28.2 Inverting matrices 827 28.3 Symmetric positive-definite matrices and least-squares approximation 832 29 Linear Programming 843 29.1 Standard and slack forms 850 29.2 Formulating problems as linear programs 859 29.3 The simplex algorithm 864 29.4 Duality 879 29.5 The initial basic feasible solution 886 x Contents 30 Polynomials and the FFT 898 30.1 Representing polynomials 900 30.2 The DFT and FFT 906 30.3 Efficient FFT implementations 915 31 Number-Theoretic Algorithms 926 31.1 Elementary number-theoretic notions 927 31.2 Greatest common divisor 933 31.3 Modular arithmetic 939 31.4 Solving modular linear equations 946 31.5 The Chinese remainder theorem 950 31.6 Powers of an element 954 31.7 The RSA public-key cryptosystem 958 ? 31.8 Primality testing 965 ? 31.9 Integer factorization 975 32 String Matching 985 32.1 The naive string-matching algorithm 988 32.2 The Rabin-Karp algorithm 990 32.3 String matching with finite automata 995 ? 32.4 The Knuth-Morris-Pratt algorithm 1002 33 Computational Geometry 1014 33.1 Line-segment properties 1015 33.2 Determining whether any pair of segments intersects 1021 33.3 Finding the convex hull 1029 33.4 Finding the closest pair of points 1039 34 NP-Completeness 1048 34.1 Polynomial time 1053 34.2 Polynomial-time verification 1061 34.3 NP-completeness and reducibility 1067 34.4 NP-completeness proofs 1078 34.5 NP-complete problems 1086 35 Approximation Algorithms 1106 35.1 The vertex-cover problem 1108 35.2 The traveling-salesman problem 1111 35.3 The set-covering problem 1117 35.4 Randomization and linear programming 1123 35.5 The subset-sum problem 1128 Contents xi VIII Appendix: Mathematical Background Introduction 1143 A Summations 1145 A.1 Summation formulas and properties 1145 A.2 Bounding summations 1149 B Sets, Etc. 1158 B.1 Sets 1158 B.2 Relations 1163 B.3 Functions 1166 B.4 Graphs 1168 B.5 Trees 1173 C Counting and Probability 1183 C.1 Counting 1183 C.2 Probability 1189 C.3 Discrete random variables 1196 C.4 The geometric and binomial distributions 1201 ? C.5 The tails of the binomial distribution 1208 D Matrices 1217 D.1 Matrices and matrix operations 1217 D.2 Basic matrix properties 1222 Bibliography 1231 Index 1251 Preface Before there were computers, there were algorithms. But now that there are computers, there are even more algorithms, and algorithms lie at the heart of computing. This book provides a comprehensive introduction to the modern study of computer algorithms. It presents many algorithms and covers them in considerable depth, yet makes their design and analysis accessible to all levels of readers. We have tried to keep explanations elementary without sacrificing depth of coverage or mathematical rigor. Each chapter presents an algorithm, a design technique, an application area, or a related topic. Algorithms are described in English and in a pseudocode designed to be readable by anyone who has done a little programming. The book contains 244 figures—many with multiple parts—illustrating how the algorithms work. Since we emphasize efficiency as a design criterion, we include careful analyses of the running times of all our algorithms. The text is intended primarily for use in undergraduate or graduate courses in algorithms or data structures. Because it discusses engineering issues in algorithm design, as well as mathematical aspects, it is equally well suited for self-study by technical professionals. In this, the third edition, we have once again updated the entire book. The changes cover a broad spectrum, including new chapters, revised pseudocode, and a more active writing style. To the teacher We have designed this book to be both versatile and complete. You should find it useful for a variety of courses, from an undergraduate course in data structures up through a graduate course in algorithms. Because we have provided considerably more material than can fit in a typical one-term course, you can consider this book to be a “buffet” or “smorgasbord” from which you can pick and choose the material that best supports the course you wish to teach. xiv Preface You should find it easy to organize your course around just the chapters you need. We have made chapters relatively self-contained, so that you need not worry about an unexpected and unnecessary dependence of one chapter on another. Each chapter presents the easier material first and the more difficult material later, with section boundaries marking natural stopping points. In an undergraduate course, you might use only the earlier sections from a chapter; in a graduate course, you might cover the entire chapter. We have included 957 exercises and 158 problems. Each section ends with exercises, and each chapter ends with problems. The exercises are generally short questions that test basic mastery of the material. Some are simple self-check thought exercises, whereas others are more substantial and are suitable as assigned homework. The problems are more elaborate case studies that often introduce new material; they often consist of several questions that lead the student through the steps required to arrive at a solution. Departing from our practice in previous editions of this book, we have made publicly available solutions to some, but by no means all, of the problems and exercises. Our Web site, http://mitpress.mit.edu/algorithms/, links to these solutions. You will want to check this site to make sure that it does not contain the solution to an exercise or problem that you plan to assign. We expect the set of solutions that we post to grow slowly over time, so you will need to check it each time you teach the course. We have starred (?) the sections and exercises that are more suitable for graduate students than for undergraduates. A starred section is not necessarily more difficult than an unstarred one, but it may require an understanding of more advanced mathematics. Likewise, starred exercises may require an advanced background or more than average creativity. To the student We hope that this textbook provides you with an enjoyable introduction to the field of algorithms. We have attempted to make every algorithm accessible and interesting. To help you when you encounter unfamiliar or difficult algorithms, we describe each one in a step-by-step manner. We also provide careful explanations of the mathematics needed to understand the analysis of the algorithms. If you already have some familiarity with a topic, you will find the chapters organized so that you can skim introductory sections and proceed quickly to the more advanced material. This is a large book, and your class will probably cover only a portion of its material. We have tried, however, to make this a book that will be useful to you now as a course textbook and also later in your career as a mathematical desk reference or an engineering handbook. Preface xv What are the prerequisites for reading this book? You should have some programming experience. In particular, you should understand recursive procedures and simple data structures such as arrays and linked lists. You should have some facility with mathematical proofs, and especially proofs by mathematical induction. A few portions of the book rely on some knowledge of elementary calculus. Beyond that, Parts I and VIII of this book teach you all the mathematical techniques you will need. We have heard, loud and clear, the call to supply solutions to problems and exercises. Our Web site, http://mitpress.mit.edu/algorithms/, links to solutions for a few of the problems and exercises. Feel free to check your solutions against ours. We ask, however, that you do not send your solutions to us. To the professional The wide range of topics in this book makes it an excellent handbook on algorithms. Because each chapter is relatively self-contained, you can focus in on the topics that most interest you. Most of the algorithms we discuss have great practical utility. We therefore address implementation concerns and other engineering issues. We often provide practical alternatives to the few algorithms that are primarily of theoretical interest. If you wish to implement any of the algorithms, you should find the translation of our pseudocode into your favorite programming language to be a fairly straightforward task. We have designed the pseudocode to present each algorithm clearly and succinctly. Consequently, we do not address error-handling and other software-engineering issues that require specific assumptions about your programming environment. We attempt to present each algorithm simply and directly without allowing the idiosyncrasies of a particular programming language to obscure its essence. We understand that if you are using this book outside of a course, then you might be unable to check your solutions to problems and exercises against solutions provided by an instructor. Our Web site, http://mitpress.mit.edu/algorithms/, links to solutions for some of the problems and exercises so that you can check your work. Please do not send your solutions to us. To our colleagues We have supplied an extensive bibliography and pointers to the current literature. Each chapter ends with a set of chapter notes that give historical details and references. The chapter notes do not provide a complete reference to the whole field xvi Preface of algorithms, however. Though it may be hard to believe for a book of this size, space constraints prevented us from including many interesting algorithms. Despite myriad requests from students for solutions to problems and exercises, we have chosen as a matter of policy not to supply references for problems and exercises, to remove the temptation for students to look up a solution rather than to find it themselves. Changes for the third edition What has changed between the second and third editions of this book? The magnitude of the changes is on a par with the changes between the first and second editions. As we said about the second-edition changes, depending on how you look at it, the book changed either not much or quite a bit. A quick look at the table of contents shows that most of the second-edition chapters and sections appear in the third edition. We removed two chapters and one section, but we have added three new chapters and two new sections apart from these new chapters. We kept the hybrid organization from the first two editions. Rather than organizing chapters by only problem domains or according only to techniques, this book has elements of both. It contains technique-based chapters on divide-and-conquer, dynamic programming, greedy algorithms, amortized analysis, NP-Completeness, and approximation algorithms. But it also has entire parts on sorting, on data structures for dynamic sets, and on algorithms for graph problems. We find that although you need to know how to apply techniques for designing and analyzing algorithms, problems seldom announce to you which techniques are most amenable to solving them. Here is a summary of the most significant changes for the third edition: We added new chapters on van Emde Boas trees and multithreaded algorithms, and we have broken out material on matrix basics into its own appendix chapter. We revised the chapter on recurrences to more broadly cover the divide-andconquer technique, and its first two sections apply divide-and-conquer to solve two problems. The second section of this chapter presents Strassen’s algorithm for matrix multiplication, which we have moved from the chapter on matrix operations. We removed two chapters that were rarely taught: binomial heaps and sorting networks. One key idea in the sorting networks chapter, the 0-1 principle, appears in this edition within Problem 8-7 as the 0-1 sorting lemma for compareexchange algorithms. The treatment of Fibonacci heaps no longer relies on binomial heaps as a precursor. Preface xvii We revised our treatment of dynamic programming and greedy algorithms. Dynamic programming now leads off with a more interesting problem, rod cutting, than the assembly-line scheduling problem from the second edition. Furthermore, we emphasize memoization a bit more than we did in the second edition, and we introduce the notion of the subproblem graph as a way to understand the running time of a dynamic-programming algorithm. In our opening example of greedy algorithms, the activity-selection problem, we get to the greedy algorithm more directly than we did in the second edition. The way we delete a node from binary search trees (which includes red-black trees) now guarantees that the node requested for deletion is the node that is actually deleted. In the first two editions, in certain cases, some other node would be deleted, with its contents moving into the node passed to the deletion procedure. With our new way to delete nodes, if other components of a program maintain pointers to nodes in the tree, they will not mistakenly end up with stale pointers to nodes that have been deleted. The material on flow networks now bases flows entirely on edges. This approach is more intuitive than the net flow used in the first two editions. With the material on matrix basics and Strassen’s algorithm moved to other chapters, the chapter on matrix operations is smaller than in the second edition. We have modified our treatment of the Knuth-Morris-Pratt string-matching algorithm. We corrected several errors. Most of these errors were posted on our Web site of second-edition errata, but a few were not. Based on many requests, we changed the syntax (as it were) of our pseudocode. We now use “D” to indicate assignment and “==” to test for equality, just as C, C++, Java, and Python do. Likewise, we have eliminated the keywords do and then and adopted “//” as our comment-to-end-of-line symbol. We also now use dot-notation to indicate object attributes. Our pseudocode remains procedural, rather than object-oriented. In other words, rather than running methods on objects, we simply call procedures, passing objects as parameters. We added 100 new exercises and 28 new problems. We also updated many bibliography entries and added several new ones. Finally, we went through the entire book and rewrote sentences, paragraphs, and sections to make the writing clearer and more active. xviii Preface Web site You can use our Web site, http://mitpress.mit.edu/algorithms/, to obtain supplementary information and to communicate with us. The Web site links to a list of known errors, solutions to selected exercises and problems, and (of course) a list explaining the corny professor jokes, as well as other content that we might add. The Web site also tells you how to report errors or make suggestions. How we produced this book Like the second edition, the third edition was produced in LATEX 2". We used the Times font with mathematics typeset using the MathTime Pro 2 fonts. We thank Michael Spivak from Publish or Perish, Inc., Lance Carnes from Personal TeX, Inc., and Tim Tregubov from Dartmouth College for technical support. As in the previous two editions, we compiled the index using Windex, a C program that we wrote, and the bibliography was produced with BIBTEX. The PDF files for this book were created on a MacBook running OS 10.5. We drew the illustrations for the third edition using MacDraw Pro, with some of the mathematical expressions in illustrations laid in with the psfrag package for LATEX 2". Unfortunately, MacDraw Pro is legacy software, having not been marketed for over a decade now. Happily, we still have a couple of Macintoshes that can run the Classic environment under OS 10.4, and hence they can run MacDraw Pro—mostly. Even under the Classic environment, we find MacDraw Pro to be far easier to use than any other drawing software for the types of illustrations that accompany computer-science text, and it produces beautiful output.1 Who knows how long our pre-Intel Macs will continue to run, so if anyone from Apple is listening: Please create an OS X-compatible version of MacDraw Pro! Acknowledgments for the third edition We have been working with the MIT Press for over two decades now, and what a terrific relationship it has been! We thank Ellen Faran, Bob Prior, Ada Brunstein, and Mary Reilly for their help and support. We were geographically distributed while producing the third edition, working in the Dartmouth College Department of Computer Science, the MIT Computer 1We investigated several drawing programs that run under Mac OS X, but all had significant shortcomings compared with MacDraw Pro. We briefly attempted to produce the illustrations for this book with a different, well known drawing program. We found that it took at least five times as long to produce each illustration as it took with MacDraw Pro, and the resulting illustrations did not look as good. Hence the decision to revert to MacDraw Pro running on older Macintoshes. Preface xix Science and Artificial Intelligence Laboratory, and the Columbia University Department of Industrial Engineering and Operations Research. We thank our respective universities and colleagues for providing such supportive and stimulating environments. Julie Sussman, P.P.A., once again bailed us out as the technical copyeditor. Time and again, we were amazed at the errors that eluded us, but that Julie caught. She also helped us improve our presentation in several places. If there is a Hall of Fame for technical copyeditors, Julie is a sure-fire, first-ballot inductee. She is nothing short of phenomenal. Thank you, thank you, thank you, Julie! Priya Natarajan also found some errors that we were able to correct before this book went to press. Any errors that remain (and undoubtedly, some do) are the responsibility of the authors (and probably were inserted after Julie read the material). The treatment for van Emde Boas trees derives from Erik Demaine’s notes, which were in turn influenced by Michael Bender. We also incorporated ideas from Javed Aslam, Bradley Kuszmaul, and Hui Zha into this edition. The chapter on multithreading was based on notes originally written jointly with Harald Prokop. The material was influenced by several others working on the Cilk project at MIT, including Bradley Kuszmaul and Matteo Frigo. The design of the multithreaded pseudocode took its inspiration from the MIT Cilk extensions to C and by Cilk Arts’s Cilk++ extensions to C++. We also thank the many readers of the first and second editions who reported errors or submitted suggestions for how to improve this book. We corrected all the bona fide errors that were reported, and we incorporated as many suggestions as we could. We rejoice that the number of such contributors has grown so great that we must regret that it has become impractical to list them all. Finally, we thank our wives—Nicole Cormen, Wendy Leiserson, Gail Rivest, and Rebecca Ivry—and our children—Ricky, Will, Debby, and Katie Leiserson; Alex and Christopher Rivest; and Molly, Noah, and Benjamin Stein—for their love and support while we prepared this book. The patience and encouragement of our families made this project possible. We affectionately dedicate this book to them. THOMAS H. CORMEN CHARLES E. LEISERSON RONALD L. RIVEST CLIFFORD STEIN Lebanon, New Hampshire Cambridge, Massachusetts Cambridge, Massachusetts New York, New York February 2009 Introduction to Algorithms Third Edition I Foundations Introduction This part will start you thinking about designing and analyzing algorithms. It is intended to be a gentle introduction to how we specify algorithms, some of the design strategies we will use throughout this book, and many of the fundamental ideas used in algorithm analysis. Later parts of this book will build upon this base. Chapter 1 provides an overview of algorithms and their place in modern computing systems. This chapter defines what an algorithm is and lists some examples. It also makes a case that we should consider algorithms as a technology, alongside technologies such as fast hardware, graphical user interfaces, object-oriented systems, and networks. In Chapter 2, we see our first algorithms, which solve the problem of sorting a sequence of n numbers. They are written in a pseudocode which, although not directly translatable to any conventional programming language, conveys the structure of the algorithm clearly enough that you should be able to implement it in the language of your choice. The sorting algorithms we examine are insertion sort, which uses an incremental approach, and merge sort, which uses a recursive technique known as “divide-and-conquer.” Although the time each requires increases with the value of n, the rate of increase differs between the two algorithms. We determine these running times in Chapter 2, and we develop a useful notation to express them. Chapter 3 precisely defines this notation, which we call asymptotic notation. It starts by defining several asymptotic notations, which we use for bounding algorithm running times from above and/or below. The rest of Chapter 3 is primarily a presentation of mathematical notation, more to ensure that your use of notation matches that in this book than to teach you new mathematical concepts. 4 Part I Foundations Chapter 4 delves further into the divide-and-conquer method introduced in Chapter 2. It provides additional examples of divide-and-conquer algorithms, including Strassen’s surprising method for multiplying two square matrices. Chapter 4 contains methods for solving recurrences, which are useful for describing the running times of recursive algorithms. One powerful technique is the “master method,” which we often use to solve recurrences that arise from divide-andconquer algorithms. Although much of Chapter 4 is devoted to proving the correctness of the master method, you may skip this proof yet still employ the master method. Chapter 5 introduces probabilistic analysis and randomized algorithms. We typically use probabilistic analysis to determine the running time of an algorithm in cases in which, due to the presence of an inherent probability distribution, the running time may differ on different inputs of the same size. In some cases, we assume that the inputs conform to a known probability distribution, so that we are averaging the running time over all possible inputs. In other cases, the probability distribution comes not from the inputs but from random choices made during the course of the algorithm. An algorithm whose behavior is determined not only by its input but by the values produced by a random-number generator is a randomized algorithm. We can use randomized algorithms to enforce a probability distribution on the inputs—thereby ensuring that no particular input always causes poor performance—or even to bound the error rate of algorithms that are allowed to produce incorrect results on a limited basis. Appendices A–D contain other mathematical material that you will find helpful as you read this book. You are likely to have seen much of the material in the appendix chapters before having read this book (although the specific definitions and notational conventions we use may differ in some cases from what you have seen in the past), and so you should think of the Appendices as reference material. On the other hand, you probably have not already seen most of the material in Part I. All the chapters in Part I and the Appendices are written with a tutorial flavor. 1 The Role of Algorithms in Computing What are algorithms? Why is the study of algorithms worthwhile? What is the role of algorithms relative to other technologies used in computers? In this chapter, we will answer these questions. 1.1 Algorithms Informally, an algorithm is any well-defined computational procedure that takes some value, or set of values, as input and produces some value, or set of values, as output. An algorithm is thus a sequence of computational steps that transform the input into the output. We can also view an algorithm as a tool for solving a well-specified computational problem. The statement of the problem specifies in general terms the desired input/output relationship. The algorithm describes a specific computational procedure for achieving that input/output relationship. For example, we might need to sort a sequence of numbers into nondecreasing order. This problem arises frequently in practice and provides fertile ground for introducing many standard design techniques and analysis tools. Here is how we formally define the sorting problem: Input: A sequence of n numbers ha1; a2; : : : ; ani. Output: A permutation (reordering) ha10 ; a20 ; : : : ; an0 i of the input sequence such that a10 Ä a20 Ä Ä an0 . For example, given the input sequence h31; 41; 59; 26; 41; 58i, a sorting algorithm returns as output the sequence h26; 31; 41; 41; 58; 59i. Such an input sequence is called an instance of the sorting problem. In general, an instance of a problem consists of the input (satisfying whatever constraints are imposed in the problem statement) needed to compute a solution to the problem. 6 Chapter 1 The Role of Algorithms in Computing Because many programs use it as an intermediate step, sorting is a fundamental operation in computer science. As a result, we have a large number of good sorting algorithms at our disposal. Which algorithm is best for a given application depends on—among other factors—the number of items to be sorted, the extent to which the items are already somewhat sorted, possible restrictions on the item values, the architecture of the computer, and the kind of storage devices to be used: main memory, disks, or even tapes. An algorithm is said to be correct if, for every input instance, it halts with the correct output. We say that a correct algorithm solves the given computational problem. An incorrect algorithm might not halt at all on some input instances, or it might halt with an incorrect answer. Contrary to what you might expect, incorrect algorithms can sometimes be useful, if we can control their error rate. We shall see an example of an algorithm with a controllable error rate in Chapter 31 when we study algorithms for finding large prime numbers. Ordinarily, however, we shall be concerned only with correct algorithms. An algorithm can be specified in English, as a computer program, or even as a hardware design. The only requirement is that the specification must provide a precise description of the computational procedure to be followed. What kinds of problems are solved by algorithms? Sorting is by no means the only computational problem for which algorithms have been developed. (You probably suspected as much when you saw the size of this book.) Practical applications of algorithms are ubiquitous and include the following examples: The Human Genome Project has made great progress toward the goals of identifying all the 100,000 genes in human DNA, determining the sequences of the 3 billion chemical base pairs that make up human DNA, storing this information in databases, and developing tools for data analysis. Each of these steps requires sophisticated algorithms. Although the solutions to the various problems involved are beyond the scope of this book, many methods to solve these biological problems use ideas from several of the chapters in this book, thereby enabling scientists to accomplish tasks while using resources efficiently. The savings are in time, both human and machine, and in money, as more information can be extracted from laboratory techniques. The Internet enables people all around the world to quickly access and retrieve large amounts of information. With the aid of clever algorithms, sites on the Internet are able to manage and manipulate this large volume of data. Examples of problems that make essential use of algorithms include finding good routes on which the data will travel (techniques for solving such problems appear in 1.1 Algorithms 7 Chapter 24), and using a search engine to quickly find pages on which particular information resides (related techniques are in Chapters 11 and 32). Electronic commerce enables goods and services to be negotiated and exchanged electronically, and it depends on the privacy of personal information such as credit card numbers, passwords, and bank statements. The core technologies used in electronic commerce include public-key cryptography and digital signatures (covered in Chapter 31), which are based on numerical algorithms and number theory. Manufacturing and other commercial enterprises often need to allocate scarce resources in the most beneficial way. An oil company may wish to know where to place its wells in order to maximize its expected profit. A political candidate may want to determine where to spend money buying campaign advertising in order to maximize the chances of winning an election. An airline may wish to assign crews to flights in the least expensive way possible, making sure that each flight is covered and that government regulations regarding crew scheduling are met. An Internet service provider may wish to determine where to place additional resources in order to serve its customers more effectively. All of these are examples of problems that can be solved using linear programming, which we shall study in Chapter 29. Although some of the details of these examples are beyond the scope of this book, we do give underlying techniques that apply to these problems and problem areas. We also show how to solve many specific problems, including the following: We are given a road map on which the distance between each pair of adjacent intersections is marked, and we wish to determine the shortest route from one intersection to another. The number of possible routes can be huge, even if we disallow routes that cross over themselves. How do we choose which of all possible routes is the shortest? Here, we model the road map (which is itself a model of the actual roads) as a graph (which we will meet in Part VI and Appendix B), and we wish to find the shortest path from one vertex to another in the graph. We shall see how to solve this problem efficiently in Chapter 24. We are given two ordered sequences of symbols, X D hx1; x2; : : : ; xmi and Y D hy1; y2; : : : ; yni, and we wish to find a longest common subsequence of X and Y . A subsequence of X is just X with some (or possibly all or none) of its elements removed. For example, one subsequence of hA; B; C; D; E; F; Gi would be hB; C; E; Gi. The length of a longest common subsequence of X and Y gives one measure of how similar these two sequences are. For example, if the two sequences are base pairs in DNA strands, then we might consider them similar if they have a long common subsequence. If X has m symbols and Y has n symbols, then X and Y have 2m and 2n possible subsequences, 8 Chapter 1 The Role of Algorithms in Computing respectively. Selecting all possible subsequences of X and Y and matching them up could take a prohibitively long time unless m and n are very small. We shall see in Chapter 15 how to use a general technique known as dynamic programming to solve this problem much more efficiently. We are given a mechanical design in terms of a library of parts, where each part may include instances of other parts, and we need to list the parts in order so that each part appears before any part that uses it. If the design comprises n parts, then there are nŠ possible orders, where nŠ denotes the factorial function. Because the factorial function grows faster than even an exponential function, we cannot feasibly generate each possible order and then verify that, within that order, each part appears before the parts using it (unless we have only a few parts). This problem is an instance of topological sorting, and we shall see in Chapter 22 how to solve this problem efficiently. We are given n points in the plane, and we wish to find the convex hull of these points. The convex hull is the smallest convex polygon containing the points. Intuitively, we can think of each point as being represented by a nail sticking out from a board. The convex hull would be represented by a tight rubber band that surrounds all the nails. Each nail around which the rubber band makes a turn is a vertex of the convex hull. (See Figure 33.6 on page 1029 for an example.) Any of the 2n subsets of the points might be the vertices of the convex hull. Knowing which points are vertices of the convex hull is not quite enough, either, since we also need to know the order in which they appear. There are many choices, therefore, for the vertices of the convex hull. Chapter 33 gives two good methods for finding the convex hull. These lists are far from exhaustive (as you again have probably surmised from this book’s heft), but exhibit two characteristics that are common to many interesting algorithmic problems: 1. They have many candidate solutions, the overwhelming majority of which do not solve the problem at hand. Finding one that does, or one that is “best,” can present quite a challenge. 2. They have practical applications. Of the problems in the above list, finding the shortest path provides the easiest examples. A transportation firm, such as a trucking or railroad company, has a financial interest in finding shortest paths through a road or rail network because taking shorter paths results in lower labor and fuel costs. Or a routing node on the Internet may need to find the shortest path through the network in order to route a message quickly. Or a person wishing to drive from New York to Boston may want to find driving directions from an appropriate Web site, or she may use her GPS while driving. 1.1 Algorithms 9 Not every problem solved by algorithms has an easily identified set of candidate solutions. For example, suppose we are given a set of numerical values representing samples of a signal, and we want to compute the discrete Fourier transform of these samples. The discrete Fourier transform converts the time domain to the frequency domain, producing a set of numerical coefficients, so that we can determine the strength of various frequencies in the sampled signal. In addition to lying at the heart of signal processing, discrete Fourier transforms have applications in data compression and multiplying large polynomials and integers. Chapter 30 gives an efficient algorithm, the fast Fourier transform (commonly called the FFT), for this problem, and the chapter also sketches out the design of a hardware circuit to compute the FFT. Data structures This book also contains several data structures. A data structure is a way to store and organize data in order to facilitate access and modifications. No single data structure works well for all purposes, and so it is important to know the strengths and limitations of several of them. Technique Although you can use this book as a “cookbook” for algorithms, you may someday encounter a problem for which you cannot readily find a published algorithm (many of the exercises and problems in this book, for example). This book will teach you techniques of algorithm design and analysis so that you can develop algorithms on your own, show that they give the correct answer, and understand their efficiency. Different chapters address different aspects of algorithmic problem solving. Some chapters address specific problems, such as finding medians and order statistics in Chapter 9, computing minimum spanning trees in Chapter 23, and determining a maximum flow in a network in Chapter 26. Other chapters address techniques, such as divide-and-conquer in Chapter 4, dynamic programming in Chapter 15, and amortized analysis in Chapter 17. Hard problems Most of this book is about efficient algorithms. Our usual measure of efficiency is speed, i.e., how long an algorithm takes to produce its result. There are some problems, however, for which no efficient solution is known. Chapter 34 studies an interesting subset of these problems, which are known as NP-complete. Why are NP-complete problems interesting? First, although no efficient algorithm for an NP-complete problem has ever been found, nobody has ever proven 10 Chapter 1 The Role of Algorithms in Computing that an efficient algorithm for one cannot exist. In other words, no one knows whether or not efficient algorithms exist for NP-complete problems. Second, the set of NP-complete problems has the remarkable property that if an efficient algorithm exists for any one of them, then efficient algorithms exist for all of them. This relationship among the NP-complete problems makes the lack of efficient solutions all the more tantalizing. Third, several NP-complete problems are similar, but not identical, to problems for which we do know of efficient algorithms. Computer scientists are intrigued by how a small change to the problem statement can cause a big change to the efficiency of the best known algorithm. You should know about NP-complete problems because some of them arise surprisingly often in real applications. If you are called upon to produce an efficient algorithm for an NP-complete problem, you are likely to spend a lot of time in a fruitless search. If you can show that the problem is NP-complete, you can instead spend your time developing an efficient algorithm that gives a good, but not the best possible, solution. As a concrete example, consider a delivery company with a central depot. Each day, it loads up each delivery truck at the depot and sends it around to deliver goods to several addresses. At the end of the day, each truck must end up back at the depot so that it is ready to be loaded for the next day. To reduce costs, the company wants to select an order of delivery stops that yields the lowest overall distance traveled by each truck. This problem is the well-known “traveling-salesman problem,” and it is NP-complete. It has no known efficient algorithm. Under certain assumptions, however, we know of efficient algorithms that give an overall distance which is not too far above the smallest possible. Chapter 35 discusses such “approximation algorithms.” Parallelism For many years, we could count on processor clock speeds increasing at a steady rate. Physical limitations present a fundamental roadblock to ever-increasing clock speeds, however: because power density increases superlinearly with clock speed, chips run the risk of melting once their clock speeds become high enough. In order to perform more computations per second, therefore, chips are being designed to contain not just one but several processing “cores.” We can liken these multicore computers to several sequential computers on a single chip; in other words, they are a type of “parallel computer.” In order to elicit the best performance from multicore computers, we need to design algorithms with parallelism in mind. Chapter 27 presents a model for “multithreaded” algorithms, which take advantage of multiple cores. This model has advantages from a theoretical standpoint, and it forms the basis of several successful computer programs, including a championship chess program. 1.2 Algorithms as a technology 11 Exercises 1.1-1 Give a real-world example that requires sorting or a real-world example that requires computing a convex hull. 1.1-2 Other than speed, what other measures of efficiency might one use in a real-world setting? 1.1-3 Select a data structure that you have seen previously, and discuss its strengths and limitations. 1.1-4 How are the shortest-path and traveling-salesman problems given above similar? How are they different? 1.1-5 Come up with a real-world problem in which only the best solution will do. Then come up with one in which a solution that is “approximately” the best is good enough. 1.2 Algorithms as a technology Suppose computers were infinitely fast and computer memory was free. Would you have any reason to study algorithms? The answer is yes, if for no other reason than that you would still like to demonstrate that your solution method terminates and does so with the correct answer. If computers were infinitely fast, any correct method for solving a problem would do. You would probably want your implementation to be within the bounds of good software engineering practice (for example, your implementation should be well designed and documented), but you would most often use whichever method was the easiest to implement. Of course, computers may be fast, but they are not infinitely fast. And memory may be inexpensive, but it is not free. Computing time is therefore a bounded resource, and so is space in memory. You should use these resources wisely, and algorithms that are efficient in terms of time or space will help you do so. 12 Chapter 1 The Role of Algorithms in Computing Efficiency Different algorithms devised to solve the same problem often differ dramatically in their efficiency. These differences can be much more significant than differences due to hardware and software. As an example, in Chapter 2, we will see two algorithms for sorting. The first, known as insertion sort, takes time roughly equal to c1n2 to sort n items, where c1 is a constant that does not depend on n. That is, it takes time roughly proportional to n2. The second, merge sort, takes time roughly equal to c2n lg n, where lg n stands for log2 n and c2 is another constant that also does not depend on n. Insertion sort typically has a smaller constant factor than merge sort, so that c1 < c2. We shall see that the constant factors can have far less of an impact on the running time than the dependence on the input size n. Let’s write insertion sort’s running time as c1n n and merge sort’s running time as c2n lg n. Then we see that where insertion sort has a factor of n in its running time, merge sort has a factor of lg n, which is much smaller. (For example, when n D 1000, lg n is approximately 10, and when n equals one million, lg n is approximately only 20.) Although insertion sort usually runs faster than merge sort for small input sizes, once the input size n becomes large enough, merge sort’s advantage of lg n vs. n will more than compensate for the difference in constant factors. No matter how much smaller c1 is than c2, there will always be a crossover point beyond which merge sort is faster. For a concrete example, let us pit a faster computer (computer A) running insertion sort against a slower computer (computer B) running merge sort. They each must sort an array of 10 million numbers. (Although 10 million numbers might seem like a lot, if the numbers are eight-byte integers, then the input occupies about 80 megabytes, which fits in the memory of even an inexpensive laptop computer many times over.) Suppose that computer A executes 10 billion instructions per second (faster than any single sequential computer at the time of this writing) and computer B executes only 10 million instructions per second, so that computer A is 1000 times faster than computer B in raw computing power. To make the difference even more dramatic, suppose that the world’s craftiest programmer codes insertion sort in machine language for computer A, and the resulting code requires 2n2 instructions to sort n numbers. Suppose further that just an average programmer implements merge sort, using a high-level language with an inefficient compiler, with the resulting code taking 50n lg n instructions. To sort 10 million numbers, computer A takes 2 .107/2 instructions 1010 instructions/second D 20,000 seconds (more than 5.5 hours) ; while computer B takes 1.2 Algorithms as a technology 13 50 107 lg 107 instructions 107 instructions/second 1163 seconds (less than 20 minutes) : By using an algorithm whose running time grows more slowly, even with a poor compiler, computer B runs more than 17 times faster than computer A! The advantage of merge sort is even more pronounced when we sort 100 million numbers: where insertion sort takes more than 23 days, merge sort takes under four hours. In general, as the problem size increases, so does the relative advantage of merge sort. Algorithms and other technologies The example above shows that we should consider algorithms, like computer hardware, as a technology. Total system performance depends on choosing efficient algorithms as much as on choosing fast hardware. Just as rapid advances are being made in other computer technologies, they are being made in algorithms as well. You might wonder whether algorithms are truly that important on contemporary computers in light of other advanced technologies, such as advanced computer architectures and fabrication technologies, easy-to-use, intuitive, graphical user interfaces (GUIs), object-oriented systems, integrated Web technologies, and fast networking, both wired and wireless. The answer is yes. Although some applications do not explicitly require algorithmic content at the application level (such as some simple, Web-based applications), many do. For example, consider a Web-based service that determines how to travel from one location to another. Its implementation would rely on fast hardware, a graphical user interface, wide-area networking, and also possibly on object orientation. However, it would also require algorithms for certain operations, such as finding routes (probably using a shortest-path algorithm), rendering maps, and interpolating addresses. Moreover, even an application that does not require algorithmic content at the application level relies heavily upon algorithms. Does the application rely on fast hardware? The hardware design used algorithms. Does the application rely on graphical user interfaces? The design of any GUI relies on algorithms. Does the application rely on networking? Routing in networks relies heavily on algorithms. Was the application written in a language other than machine code? Then it was processed by a compiler, interpreter, or assembler, all of which make extensive use 14 Chapter 1 The Role of Algorithms in Computing of algorithms. Algorithms are at the core of most technologies used in contemporary computers. Furthermore, with the ever-increasing capacities of computers, we use them to solve larger problems than ever before. As we saw in the above comparison between insertion sort and merge sort, it is at larger problem sizes that the differences in efficiency between algorithms become particularly prominent. Having a solid base of algorithmic knowledge and technique is one characteristic that separates the truly skilled programmers from the novices. With modern computing technology, you can accomplish some tasks without knowing much about algorithms, but with a good background in algorithms, you can do much, much more. Exercises 1.2-1 Give an example of an application that requires algorithmic content at the application level, and discuss the function of the algorithms involved. 1.2-2 Suppose we are comparing implementations of insertion sort and merge sort on the same machine. For inputs of size n, insertion sort runs in 8n2 steps, while merge sort runs in 64n lg n steps. For which values of n does insertion sort beat merge sort? 1.2-3 What is the smallest value of n such that an algorithm whose running time is 100n2 runs faster than an algorithm whose running time is 2n on the same machine? Problems 1-1 Comparison of running times For each function f .n/ and time t in the following table, determine the largest size n of a problem that can be solved in time t, assuming that the algorithm to solve the problem takes f .n/ microseconds. Notes for Chapter 1 1 1 1 second minute hour lg n p n n n lg n n2 n3 2n nŠ 15 1 1 1 1 day month year century Chapter notes There are many excellent texts on the general topic of algorithms, including those by Aho, Hopcroft, and Ullman [5, 6]; Baase and Van Gelder [28]; Brassard and Bratley [54]; Dasgupta, Papadimitriou, and Vazirani [82]; Goodrich and Tamassia [148]; Hofri [175]; Horowitz, Sahni, and Rajasekaran [181]; Johnsonbaugh and Schaefer [193]; Kingston [205]; Kleinberg and Tardos [208]; Knuth [209, 210, 211]; Kozen [220]; Levitin [235]; Manber [242]; Mehlhorn [249, 250, 251]; Purdom and Brown [287]; Reingold, Nievergelt, and Deo [293]; Sedgewick [306]; Sedgewick and Flajolet [307]; Skiena [318]; and Wilf [356]. Some of the more practical aspects of algorithm design are discussed by Bentley [42, 43] and Gonnet [145]. Surveys of the field of algorithms can also be found in the Handbook of Theoretical Computer Science, Volume A [342] and the CRC Algorithms and Theory of Computation Handbook [25]. Overviews of the algorithms used in computational biology can be found in textbooks by Gusfield [156], Pevzner [275], Setubal and Meidanis [310], and Waterman [350]. 2 Getting Started This chapter will familiarize you with the framework we shall use throughout the book to think about the design and analysis of algorithms. It is self-contained, but it does include several references to material that we introduce in Chapters 3 and 4. (It also contains several summations, which Appendix A shows how to solve.) We begin by examining the insertion sort algorithm to solve the sorting problem introduced in Chapter 1. We define a “pseudocode” that should be familiar to you if you have done computer programming, and we use it to show how we shall specify our algorithms. Having specified the insertion sort algorithm, we then argue that it correctly sorts, and we analyze its running time. The analysis introduces a notation that focuses on how that time increases with the number of items to be sorted. Following our discussion of insertion sort, we introduce the divide-and-conquer approach to the design of algorithms and use it to develop an algorithm called merge sort. We end with an analysis of merge sort’s running time. 2.1 Insertion sort Our first algorithm, insertion sort, solves the sorting problem introduced in Chapter 1: Input: A sequence of n numbers ha1; a2; : : : ; ani. Output: A permutation (reordering) ha10 ; a20 ; : : : ; an0 i of the input sequence such that a10 Ä a20 Ä Ä an0 . The numbers that we wish to sort are also known as the keys. Although conceptually we are sorting a sequence, the input comes to us in the form of an array with n elements. In this book, we shall typically describe algorithms as programs written in a pseudocode that is similar in many respects to C, C++, Java, Python, or Pascal. If you have been introduced to any of these languages, you should have little trouble 2.1 Insertion sort 17 ♣♣7♣♣♣♣ 2♣ ♣4♣♣♣5♣♣♣1♣0♣♣♣♣♣ 1♣0♣5♣♣♣4♣♣♣♣2♣♣♣♣♣ ♣7 ♣ ♣ Figure 2.1 Sorting a hand of cards using insertion sort. reading our algorithms. What separates pseudocode from “real” code is that in pseudocode, we employ whatever expressive method is most clear and concise to specify a given algorithm. Sometimes, the clearest method is English, so do not be surprised if you come across an English phrase or sentence embedded within a section of “real” code. Another difference between pseudocode and real code is that pseudocode is not typically concerned with issues of software engineering. Issues of data abstraction, modularity, and error handling are often ignored in order to convey the essence of the algorithm more concisely. We start with insertion sort, which is an efficient algorithm for sorting a small number of elements. Insertion sort works the way many people sort a hand of playing cards. We start with an empty left hand and the cards face down on the table. We then remove one card at a time from the table and insert it into the correct position in the left hand. To find the correct position for a card, we compare it with each of the cards already in the hand, from right to left, as illustrated in Figure 2.1. At all times, the cards held in the left hand are sorted, and these cards were originally the top cards of the pile on the table. We present our pseudocode for insertion sort as a procedure called INSERTIONSORT, which takes as a parameter an array AŒ1 : : n containing a sequence of length n that is to be sorted. (In the code, the number n of elements in A is denoted by A:length.) The algorithm sorts the input numbers in place: it rearranges the numbers within the array A, with at most a constant number of them stored outside the array at any time. The input array A contains the sorted output sequence when the INSERTION-SORT procedure is finished. 18 Chapter 2 Getting Started 123456 (a) 5 2 4 6 1 3 123456 (b) 2 5 4 6 1 3 123456 (c) 2 4 5 6 1 3 123456 (d) 2 4 5 6 1 3 123456 (e) 1 2 4 5 6 3 123456 (f) 1 2 3 4 5 6 Figure 2.2 The operation of INSERTION-SORT on the array A D h5; 2; 4; 6; 1; 3i. Array indices appear above the rectangles, and values stored in the array positions appear within the rectangles. (a)–(e) The iterations of the for loop of lines 1–8. In each iteration, the black rectangle holds the key taken from AŒj , which is compared with the values in shaded rectangles to its left in the test of line 5. Shaded arrows show array values moved one position to the right in line 6, and black arrows indicate where the key moves to in line 8. (f) The final sorted array. IN SERTIO N -S O RT .A/ 1 for j D 2 to A:length 2 key D AŒj  3 // Insert AŒj  into the sorted sequence AŒ1 : : j 1. 4 iDj 1 5 while i > 0 and AŒi > key 6 AŒi C 1 D AŒi 7 iDi 1 8 AŒi C 1 D key Loop invariants and the correctness of insertion sort Figure 2.2 shows how this algorithm works for A D h5; 2; 4; 6; 1; 3i. The index j indicates the “current card” being inserted into the hand. At the beginning of each iteration of the for loop, which is indexed by j , the subarray consisting of elements AŒ1 : : j 1 constitutes the currently sorted hand, and the remaining subarray AŒj C 1 : : n corresponds to the pile of cards still on the table. In fact, elements AŒ1 : : j 1 are the elements originally in positions 1 through j 1, but now in sorted order. We state these properties of AŒ1 : : j 1 formally as a loop invariant: At the start of each iteration of the for loop of lines 1–8, the subarray AŒ1 : : j 1 consists of the elements originally in AŒ1 : : j 1, but in sorted order. We use loop invariants to help us understand why an algorithm is correct. We must show three things about a loop invariant: 2.1 Insertion sort 19 Initialization: It is true prior to the first iteration of the loop. Maintenance: If it is true before an iteration of the loop, it remains true before the next iteration. Termination: When the loop terminates, the invariant gives us a useful property that helps show that the algorithm is correct. When the first two properties hold, the loop invariant is true prior to every iteration of the loop. (Of course, we are free to use established facts other than the loop invariant itself to prove that the loop invariant remains true before each iteration.) Note the similarity to mathematical induction, where to prove that a property holds, you prove a base case and an inductive step. Here, showing that the invariant holds before the first iteration corresponds to the base case, and showing that the invariant holds from iteration to iteration corresponds to the inductive step. The third property is perhaps the most important one, since we are using the loop invariant to show correctness. Typically, we use the loop invariant along with the condition that caused the loop to terminate. The termination property differs from how we usually use mathematical induction, in which we apply the inductive step infinitely; here, we stop the “induction” when the loop terminates. Let us see how these properties hold for insertion sort. Initialization: We start by showing that the loop invariant holds before the first loop iteration, when j D 2.1 The subarray AŒ1 : : j 1, therefore, consists of just the single element AŒ1, which is in fact the original element in AŒ1. Moreover, this subarray is sorted (trivially, of course), which shows that the loop invariant holds prior to the first iteration of the loop. Maintenance: Next, we tackle the second property: showing that each iteration maintains the loop invariant. Informally, the body of the for loop works by moving AŒj 1, AŒj 2, AŒj 3, and so on by one position to the right until it finds the proper position for AŒj  (lines 4–7), at which point it inserts the value of AŒj  (line 8). The subarray AŒ1 : : j  then consists of the elements originally in AŒ1 : : j , but in sorted order. Incrementing j for the next iteration of the for loop then preserves the loop invariant. A more formal treatment of the second property would require us to state and show a loop invariant for the while loop of lines 5–7. At this point, however, 1When the loop is a for loop, the moment at which we check the loop invariant just prior to the first iteration is immediately after the initial assignment to the loop-counter variable and just before the first test in the loop header. In the case of INSERTION-SORT, this time is after assigning 2 to the variable j but before the first test of whether j Ä A: length. 20 Chapter 2 Getting Started we prefer not to get bogged down in such formalism, and so we rely on our informal analysis to show that the second property holds for the outer loop. Termination: Finally, we examine what happens when the loop terminates. The condition causing the for loop to terminate is that j > A:length D n. Because each loop iteration increases j by 1, we must have j D n C 1 at that time. Substituting n C 1 for j in the wording of loop invariant, we have that the subarray AŒ1 : : n consists of the elements originally in AŒ1 : : n, but in sorted order. Observing that the subarray AŒ1 : : n is the entire array, we conclude that the entire array is sorted. Hence, the algorithm is correct. We shall use this method of loop invariants to show correctness later in this chapter and in other chapters as well. Pseudocode conventions We use the following conventions in our pseudocode. Indentation indicates block structure. For example, the body of the for loop that begins on line 1 consists of lines 2–8, and the body of the while loop that begins on line 5 contains lines 6–7 but not line 8. Our indentation style applies to if-else statements2 as well. Using indentation instead of conventional indicators of block structure, such as begin and end statements, greatly reduces clutter while preserving, or even enhancing, clarity.3 The looping constructs while, for, and repeat-until and the if-else conditional construct have interpretations similar to those in C, C++, Java, Python, and Pascal.4 In this book, the loop counter retains its value after exiting the loop, unlike some situations that arise in C++, Java, and Pascal. Thus, immediately after a for loop, the loop counter’s value is the value that first exceeded the for loop bound. We used this property in our correctness argument for insertion sort. The for loop header in line 1 is for j D 2 to A:length, and so when this loop terminates, j D A:length C 1 (or, equivalently, j D n C 1, since n D A:length). We use the keyword to when a for loop increments its loop 2In an if-else statement, we indent else at the same level as its matching if. Although we omit the keyword then, we occasionally refer to the portion executed when the test following if is true as a then clause. For multiway tests, we use elseif for tests after the first one. 3Each pseudocode procedure in this book appears on one page so that you will not have to discern levels of indentation in code that is split across pages. 4Most block-structured languages have equivalent constructs, though the exact syntax may differ. Python lacks repeat-until loops, and its for loops operate a little differently from the for loops in this book. 2.1 Insertion sort 21 counter in each iteration, and we use the keyword downto when a for loop decrements its loop counter. When the loop counter changes by an amount greater than 1, the amount of change follows the optional keyword by. The symbol “//” indicates that the remainder of the line is a comment. A multiple assignment of the form i D j D e assigns to both variables i and j the value of expression e; it should be treated as equivalent to the assignment j D e followed by the assignment i D j . Variables (such as i, j , and key) are local to the given procedure. We shall not use global variables without explicit indication. We access array elements by specifying the array name followed by the index in square brackets. For example, AŒi indicates the ith element of the array A. The notation “: :” is used to indicate a range of values within an array. Thus, AŒ1 : : j  indicates the subarray of A consisting of the j elements AŒ1; AŒ2; : : : ; AŒj . We typically organize compound data into objects, which are composed of attributes. We access a particular attribute using the syntax found in many object-oriented programming languages: the object name, followed by a dot, followed by the attribute name. For example, we treat an array as an object with the attribute length indicating how many elements it contains. To specify the number of elements in an array A, we write A:length. We treat a variable representing an array or object as a pointer to the data representing the array or object. For all attributes f of an object x, setting y D x causes y:f to equal x:f . Moreover, if we now set x:f D 3, then afterward not only does x:f equal 3, but y:f equals 3 as well. In other words, x and y point to the same object after the assignment y D x. Our attribute notation can “cascade.” For example, suppose that the attribute f is itself a pointer to some type of object that has an attribute g. Then the notation x:f :g is implicitly parenthesized as .x:f /:g. In other words, if we had assigned y D x:f , then x:f :g is the same as y:g. Sometimes, a pointer will refer to no object at all. In this case, we give it the special value NIL. We pass parameters to a procedure by value: the called procedure receives its own copy of the parameters, and if it assigns a value to a parameter, the change is not seen by the calling procedure. When objects are passed, the pointer to the data representing the object is copied, but the object’s attributes are not. For example, if x is a parameter of a called procedure, the assignment x D y within the called procedure is not visible to the calling procedure. The assignment x:f D 3, however, is visible. Similarly, arrays are passed by pointer, so that 22 Chapter 2 Getting Started a pointer to the array is passed, rather than the entire array, and changes to individual array elements are visible to the calling procedure. A return statement immediately transfers control back to the point of call in the calling procedure. Most return statements also take a value to pass back to the caller. Our pseudocode differs from many programming languages in that we allow multiple values to be returned in a single return statement. The boolean operators “and” and “or” are short circuiting. That is, when we evaluate the expression “x and y” we first evaluate x. If x evaluates to FALSE, then the entire expression cannot evaluate to TRUE, and so we do not evaluate y. If, on the other hand, x evaluates to TRUE, we must evaluate y to determine the value of the entire expression. Similarly, in the expression “x or y” we evaluate the expression y only if x evaluates to FALSE. Short-circuiting operators allow us to write boolean expressions such as “x ¤ NIL and x:f D y” without worrying about what happens when we try to evaluate x:f when x is NIL. The keyword error indicates that an error occurred because conditions were wrong for the procedure to have been called. The calling procedure is responsible for handling the error, and so we do not specify what action to take. Exercises 2.1-1 Using Figure 2.2 as a model, illustrate the operation of INSERTION-SORT on the array A D h31; 41; 59; 26; 41; 58i. 2.1-2 Rewrite the INSERTION-SORT procedure to sort into nonincreasing instead of nondecreasing order. 2.1-3 Consider the searching problem: Input: A sequence of n numbers A D ha1; a2; : : : ; ani and a value . Output: An index i such that D AŒi  or the special value NIL if appear in A. does not Write pseudocode for linear search, which scans through the sequence, looking for . Using a loop invariant, prove that your algorithm is correct. Make sure that your loop invariant fulfills the three necessary properties. 2.1-4 Consider the problem of adding two n-bit binary integers, stored in two n-element arrays A and B. The sum of the two integers should be stored in binary form in 2.2 Analyzing algorithms 23 an .n C 1/-element array C . State the problem formally and write pseudocode for adding the two integers. 2.2 Analyzing algorithms Analyzing an algorithm has come to mean predicting the resources that the algorithm requires. Occasionally, resources such as memory, communication bandwidth, or computer hardware are of primary concern, but most often it is computational time that we want to measure. Generally, by analyzing several candidate algorithms for a problem, we can identify a most efficient one. Such analysis may indicate more than one viable candidate, but we can often discard several inferior algorithms in the process. Before we can analyze an algorithm, we must have a model of the implementation technology that we will use, including a model for the resources of that technology and their costs. For most of this book, we shall assume a generic oneprocessor, random-access machine (RAM) model of computation as our implementation technology and understand that our algorithms will be implemented as computer programs. In the RAM model, instructions are executed one after another, with no concurrent operations. Strictly speaking, we should precisely define the instructions of the RAM model and their costs. To do so, however, would be tedious and would yield little insight into algorithm design and analysis. Yet we must be careful not to abuse the RAM model. For example, what if a RAM had an instruction that sorts? Then we could sort in just one instruction. Such a RAM would be unrealistic, since real computers do not have such instructions. Our guide, therefore, is how real computers are designed. The RAM model contains instructions commonly found in real computers: arithmetic (such as add, subtract, multiply, divide, remainder, floor, ceiling), data movement (load, store, copy), and control (conditional and unconditional branch, subroutine call and return). Each such instruction takes a constant amount of time. The data types in the RAM model are integer and floating point (for storing real numbers). Although we typically do not concern ourselves with precision in this book, in some applications precision is crucial. We also assume a limit on the size of each word of data. For example, when working with inputs of size n, we typically assume that integers are represented by c lg n bits for some constant c 1. We require c 1 so that each word can hold the value of n, enabling us to index the individual input elements, and we restrict c to be a constant so that the word size does not grow arbitrarily. (If the word size could grow arbitrarily, we could store huge amounts of data in one word and operate on it all in constant time—clearly an unrealistic scenario.) 24 Chapter 2 Getting Started Real computers contain instructions not listed above, and such instructions represent a gray area in the RAM model. For example, is exponentiation a constanttime instruction? In the general case, no; it takes several instructions to compute xy when x and y are real numbers. In restricted situations, however, exponentiation is a constant-time operation. Many computers have a “shift left” instruction, which in constant time shifts the bits of an integer by k positions to the left. In most computers, shifting the bits of an integer by one position to the left is equivalent to multiplication by 2, so that shifting the bits by k positions to the left is equivalent to multiplication by 2k. Therefore, such computers can compute 2k in one constant-time instruction by shifting the integer 1 by k positions to the left, as long as k is no more than the number of bits in a computer word. We will endeavor to avoid such gray areas in the RAM model, but we will treat computation of 2k as a constant-time operation when k is a small enough positive integer. In the RAM model, we do not attempt to model the memory hierarchy that is common in contemporary computers. That is, we do not model caches or virtual memory. Several computational models attempt to account for memory-hierarchy effects, which are sometimes significant in real programs on real machines. A handful of problems in this book examine memory-hierarchy effects, but for the most part, the analyses in this book will not consider them. Models that include the memory hierarchy are quite a bit more complex than the RAM model, and so they can be difficult to work with. Moreover, RAM-model analyses are usually excellent predictors of performance on actual machines. Analyzing even a simple algorithm in the RAM model can be a challenge. The mathematical tools required may include combinatorics, probability theory, algebraic dexterity, and the ability to identify the most significant terms in a formula. Because the behavior of an algorithm may be different for each possible input, we need a means for summarizing that behavior in simple, easily understood formulas. Even though we typically select only one machine model to analyze a given algorithm, we still face many choices in deciding how to express our analysis. We would like a way that is simple to write and manipulate, shows the important characteristics of an algorithm’s resource requirements, and suppresses tedious details. Analysis of insertion sort The time taken by the INSERTION-SORT procedure depends on the input: sorting a thousand numbers takes longer than sorting three numbers. Moreover, INSERTIONSORT can take different amounts of time to sort two input sequences of the same size depending on how nearly sorted they already are. In general, the time taken by an algorithm grows with the size of the input, so it is traditional to describe the running time of a program as a function of the size of its input. To do so, we need to define the terms “running time” and “size of input” more carefully. 2.2 Analyzing algorithms 25 The best notion for input size depends on the problem being studied. For many problems, such as sorting or computing discrete Fourier transforms, the most natural measure is the number of items in the input—for example, the array size n for sorting. For many other problems, such as multiplying two integers, the best measure of input size is the total number of bits needed to represent the input in ordinary binary notation. Sometimes, it is more appropriate to describe the size of the input with two numbers rather than one. For instance, if the input to an algorithm is a graph, the input size can be described by the numbers of vertices and edges in the graph. We shall indicate which input size measure is being used with each problem we study. The running time of an algorithm on a particular input is the number of primitive operations or “steps” executed. It is convenient to define the notion of step so that it is as machine-independent as possible. For the moment, let us adopt the following view. A constant amount of time is required to execute each line of our pseudocode. One line may take a different amount of time than another line, but we shall assume that each execution of the i th line takes time ci , where ci is a constant. This viewpoint is in keeping with the RAM model, and it also reflects how the pseudocode would be implemented on most actual computers.5 In the following discussion, our expression for the running time of INSERTIONSORT will evolve from a messy formula that uses all the statement costs ci to a much simpler notation that is more concise and more easily manipulated. This simpler notation will also make it easy to determine whether one algorithm is more efficient than another. We start by presenting the INSERTION-SORT procedure with the time “cost” of each statement and the number of times each statement is executed. For each j D 2; 3; : : : ; n, where n D A:length, we let tj denote the number of times the while loop test in line 5 is executed for that value of j . When a for or while loop exits in the usual way (i.e., due to the test in the loop header), the test is executed one time more than the loop body. We assume that comments are not executable statements, and so they take no time. 5There are some subtleties here. Computational steps that we specify in English are often variants of a procedure that requires more than just a constant amount of time. For example, later in this book we might say “sort the points by x-coordinate,” which, as we shall see, takes more than a constant amount of time. Also, note that a statement that calls a subroutine takes constant time, though the subroutine, once invoked, may take more. That is, we separate the process of calling the subroutine—passing parameters to it, etc.—from the process of executing the subroutine. 26 Chapter 2 Getting Started IN SERTIO N -S O RT .A/ 1 for j D 2 to A:length 2 key D AŒj  3 // Insert AŒj  into the sorted sequence AŒ1 : : j 1. 4 iDj 1 5 while i > 0 and AŒi > key 6 AŒi C 1 D AŒi 7 iDi 1 8 AŒi C 1 D key cost times c1 n c2 n1 0 n1 c4 c5 c6 c7 P nP PjjjnnnDDD1222..tttjjj 1/ 1/ c8 n1 The running time of the algorithm is the sum of running times for each statement executed; a statement that takes ci steps to execute and executes n times will contribute ci n to the total running time.6 To compute T .n/, the running time of INSERTION-SORT on an input of n values, we sum the products of the cost and times columns, obtaining Xn Xn T .n/ D c1n C c2.n 1/ C c4.n 1/ C c5 tj C c6 .tj 1/ j D2 j D2 Xn C c7 .tj 1/ C c8.n 1/ : j D2 Even for inputs of a given size, an algorithm’s running time may depend on which input of that size is given. For example, in INSERTION-SORT, the best case occurs if the array is already sorted. For each j D 2; 3; : : : ; n, we then find that AŒi Ä key in line 5 when i has its initial value of j 1. Thus tj D 1 for j D 2; 3; : : : ; n, and the best-case running time is T .n/ D c1n C c2.n 1/ C c4.n 1/ C c5.n 1/ C c8.n 1/ D .c1 C c2 C c4 C c5 C c8/n .c2 C c4 C c5 C c8/ : We can express this running time as an C b for constants a and b that depend on the statement costs ci ; it is thus a linear function of n. If the array is in reverse sorted order—that is, in decreasing order—the worst case results. We must compare each element AŒj  with each element in the entire sorted subarray AŒ1 : : j 1, and so tj D j for j D 2; 3; : : : ; n. Noting that 6This characteristic does not necessarily hold for a resource such as memory. A statement that references m words of memory and is executed n times does not necessarily reference mn distinct words of memory. 2.2 Analyzing algorithms 27 Xn n.n C 1/ jD 1 2 j D2 and Xn .j j D2 1/ D n.n 1/ 2 (see Appendix A for a review of how to solve these summations), we find that in the worst case, the running time of INSERTION-SORT is  à n.n C 1/ T .n/ D c1n C c2.n 1/ C c4.n 1/ C c5  Ã à 2 1 n.n 1/ n.n 1/ D C c6 c5 C 2 c6 2 2 C C c7 c7 Á n2 C 2 c1 C 2 c2 C c8.n C c4 C c5 2 1/ c6 2 c7 2 C Á c8 n .c2 C c4 C c5 C c8/ : We can express this worst-case running time as an2 C bn C c for constants a, b, and c that again depend on the statement costs ci ; it is thus a quadratic function of n. Typically, as in insertion sort, the running time of an algorithm is fixed for a given input, although in later chapters we shall see some interesting “randomized” algorithms whose behavior can vary even for a fixed input. Worst-case and average-case analysis In our analysis of insertion sort, we looked at both the best case, in which the input array was already sorted, and the worst case, in which the input array was reverse sorted. For the remainder of this book, though, we shall usually concentrate on finding only the worst-case running time, that is, the longest running time for any input of size n. We give three reasons for this orientation. The worst-case running time of an algorithm gives us an upper bound on the running time for any input. Knowing it provides a guarantee that the algorithm will never take any longer. We need not make some educated guess about the running time and hope that it never gets much worse. For some algorithms, the worst case occurs fairly often. For example, in searching a database for a particular piece of information, the searching algorithm’s worst case will often occur when the information is not present in the database. In some applications, searches for absent information may be frequent. 28 Chapter 2 Getting Started The “average case” is often roughly as bad as the worst case. Suppose that we randomly choose n numbers and apply insertion sort. How long does it take to determine where in subarray AŒ1 : : j 1 to insert element AŒj ? On average, half the elements in AŒ1 : : j 1 are less than AŒj , and half the elements are greater. On average, therefore, we check half of the subarray AŒ1 : : j 1, and so tj is about j=2. The resulting average-case running time turns out to be a quadratic function of the input size, just like the worst-case running time. In some particular cases, we shall be interested in the average-case running time of an algorithm; we shall see the technique of probabilistic analysis applied to various algorithms throughout this book. The scope of average-case analysis is limited, because it may not be apparent what constitutes an “average” input for a particular problem. Often, we shall assume that all inputs of a given size are equally likely. In practice, this assumption may be violated, but we can sometimes use a randomized algorithm, which makes random choices, to allow a probabilistic analysis and yield an expected running time. We explore randomized algorithms more in Chapter 5 and in several other subsequent chapters. Order of growth We used some simplifying abstractions to ease our analysis of the INSERTIONSORT procedure. First, we ignored the actual cost of each statement, using the constants ci to represent these costs. Then, we observed that even these constants give us more detail than we really need: we expressed the worst-case running time as an2 C bn C c for some constants a, b, and c that depend on the statement costs ci . We thus ignored not only the actual statement costs, but also the abstract costs ci . We shall now make one more simplifying abstraction: it is the rate of growth, or order of growth, of the running time that really interests us. We therefore consider only the leading term of a formula (e.g., an2), since the lower-order terms are relatively insignificant for large values of n. We also ignore the leading term’s constant coefficient, since constant factors are less significant than the rate of growth in determining computational efficiency for large inputs. For insertion sort, when we ignore the lower-order terms and the leading term’s constant coefficient, we are left with the factor of n2 from the leading term. We write that insertion sort has a worst-case running time of ‚.n2/ (pronounced “theta of n-squared”). We shall use ‚-notation informally in this chapter, and we will define it precisely in Chapter 3. We usually consider one algorithm to be more efficient than another if its worstcase running time has a lower order of growth. Due to constant factors and lowerorder terms, an algorithm whose running time has a higher order of growth might take less time for small inputs than an algorithm whose running time has a lower 2.3 Designing algorithms 29 order of growth. But for large enough inputs, a ‚.n2/ algorithm, for example, will run more quickly in the worst case than a ‚.n3/ algorithm. Exercises 2.2-1 Express the function n3=1000 100n2 100n C 3 in terms of ‚-notation. 2.2-2 Consider sorting n numbers stored in array A by first finding the smallest element of A and exchanging it with the element in AŒ1. Then find the second smallest element of A, and exchange it with AŒ2. Continue in this manner for the first n 1 elements of A. Write pseudocode for this algorithm, which is known as selection sort. What loop invariant does this algorithm maintain? Why does it need to run for only the first n 1 elements, rather than for all n elements? Give the best-case and worst-case running times of selection sort in ‚-notation. 2.2-3 Consider linear search again (see Exercise 2.1-3). How many elements of the input sequence need to be checked on the average, assuming that the element being searched for is equally likely to be any element in the array? How about in the worst case? What are the average-case and worst-case running times of linear search in ‚-notation? Justify your answers. 2.2-4 How can we modify almost any algorithm to have a good best-case running time? 2.3 Designing algorithms We can choose from a wide range of algorithm design techniques. For insertion sort, we used an incremental approach: having sorted the subarray AŒ1 : : j 1, we inserted the single element AŒj  into its proper place, yielding the sorted subarray AŒ1 : : j . In this section, we examine an alternative design approach, known as “divideand-conquer,” which we shall explore in more detail in Chapter 4. We’ll use divideand-conquer to design a sorting algorithm whose worst-case running time is much less than that of insertion sort. One advantage of divide-and-conquer algorithms is that their running times are often easily determined using techniques that we will see in Chapter 4. 30 Chapter 2 Getting Started 2.3.1 The divide-and-conquer approach Many useful algorithms are recursive in structure: to solve a given problem, they call themselves recursively one or more times to deal with closely related subproblems. These algorithms typically follow a divide-and-conquer approach: they break the problem into several subproblems that are similar to the original problem but smaller in size, solve the subproblems recursively, and then combine these solutions to create a solution to the original problem. The divide-and-conquer paradigm involves three steps at each level of the recursion: Divide the problem into a number of subproblems that are smaller instances of the same problem. Conquer the subproblems by solving them recursively. If the subproblem sizes are small enough, however, just solve the subproblems in a straightforward manner. Combine the solutions to the subproblems into the solution for the original problem. The merge sort algorithm closely follows the divide-and-conquer paradigm. Intuitively, it operates as follows. Divide: Divide the n-element sequence to be sorted into two subsequences of n=2 elements each. Conquer: Sort the two subsequences recursively using merge sort. Combine: Merge the two sorted subsequences to produce the sorted answer. The recursion “bottoms out” when the sequence to be sorted has length 1, in which case there is no work to be done, since every sequence of length 1 is already in sorted order. The key operation of the merge sort algorithm is the merging of two sorted sequences in the “combine” step. We merge by calling an auxiliary procedure MERGE.A; p; q; r/, where A is an array and p, q, and r are indices into the array such that p Ä q < r. The procedure assumes that the subarrays AŒp : : q and AŒq C 1 : : r are in sorted order. It merges them to form a single sorted subarray that replaces the current subarray AŒp : : r. Our MERGE procedure takes time ‚.n/, where n D r p C 1 is the total number of elements being merged, and it works as follows. Returning to our cardplaying motif, suppose we have two piles of cards face up on a table. Each pile is sorted, with the smallest cards on top. We wish to merge the two piles into a single sorted output pile, which is to be face down on the table. Our basic step consists of choosing the smaller of the two cards on top of the face-up piles, removing it from its pile (which exposes a new top card), and placing this card face down onto 2.3 Designing algorithms 31 the output pile. We repeat this step until one input pile is empty, at which time we just take the remaining input pile and place it face down onto the output pile. Computationally, each basic step takes constant time, since we are comparing just the two top cards. Since we perform at most n basic steps, merging takes ‚.n/ time. The following pseudocode implements the above idea, but with an additional twist that avoids having to check whether either pile is empty in each basic step. We place on the bottom of each pile a sentinel card, which contains a special value that we use to simplify our code. Here, we use 1 as the sentinel value, so that whenever a card with 1 is exposed, it cannot be the smaller card unless both piles have their sentinel cards exposed. But once that happens, all the nonsentinel cards have already been placed onto the output pile. Since we know in advance that exactly r p C 1 cards will be placed onto the output pile, we can stop once we have performed that many basic steps. MERGE.A; p; q; r/ 1 n1 D q p C 1 2 n2 D r q 3 let LŒ1 : : n1 C 1 and RŒ1 : : n2 C 1 be new arrays 4 for i D 1 to n1 5 LŒi D AŒp C i 1 6 for j D 1 to n2 7 RŒj  D AŒq C j  8 LŒn1 C 1 D 1 9 RŒn2 C 1 D 1 10 i D 1 11 j D 1 12 for k D p to r 13 if LŒi Ä RŒj  14 AŒk D LŒi 15 i D iC1 16 else AŒk D RŒj  17 j D j C1 In detail, the MERGE procedure works as follows. Line 1 computes the length n1 of the subarray AŒp : : q, and line 2 computes the length n2 of the subarray AŒq C 1 : : r. We create arrays L and R (“left” and “right”), of lengths n1 C 1 and n2 C 1, respectively, in line 3; the extra position in each array will hold the sentinel. The for loop of lines 4–5 copies the subarray AŒp : : q into LŒ1 : : n1, and the for loop of lines 6–7 copies the subarray AŒq C 1 : : r into RŒ1 : : n2. Lines 8–9 put the sentinels at the ends of the arrays L and R. Lines 10–17, illus- 32 Chapter 2 Getting Started 8 9 10 11 12 13 14 15 16 17 A…2 4 5 7 1 2 3 6… k 12345 L 2457∞ 12345 R1 2 3 6∞ i j (a) 8 9 10 11 12 13 14 15 16 17 A…1 2 5 7 1 2 3 6… k 12345 L 2457∞ 12345 R1 2 3 6∞ i j (c) 8 9 10 11 12 13 14 15 16 17 A…1 4 5 7 1 2 3 6… k 12345 L 2 4 5 7∞ 12345 R1 2 3 6∞ i j (b) 8 9 10 11 12 13 14 15 16 17 A…1 2 2 7 1 2 3 6… k 12345 L 2 4 5 7∞ 12345 R1 2 3 6∞ i j (d) Figure 2.3 The operation of lines 10–17 in the call MERGE.A; 9; 12; 16/, when the subarray AŒ9 : : 16 contains the sequence h2; 4; 5; 7; 1; 2; 3; 6i. After copying and inserting sentinels, the array L contains h2; 4; 5; 7; 1i, and the array R contains h1; 2; 3; 6; 1i. Lightly shaded positions in A contain their final values, and lightly shaded positions in L and R contain values that have yet to be copied back into A. Taken together, the lightly shaded positions always comprise the values originally in AŒ9 : : 16, along with the two sentinels. Heavily shaded positions in A contain values that will be copied over, and heavily shaded positions in L and R contain values that have already been copied back into A. (a)–(h) The arrays A, L, and R, and their respective indices k, i, and j prior to each iteration of the loop of lines 12–17. trated in Figure 2.3, perform the r p C 1 basic steps by maintaining the following loop invariant: At the start of each iteration of the for loop of lines 12–17, the subarray AŒp : : k 1 contains the k p smallest elements of LŒ1 : : n1 C 1 and RŒ1 : : n2 C 1, in sorted order. Moreover, LŒi and RŒj  are the smallest elements of their arrays that have not been copied back into A. We must show that this loop invariant holds prior to the first iteration of the for loop of lines 12–17, that each iteration of the loop maintains the invariant, and that the invariant provides a useful property to show correctness when the loop terminates. Initialization: Prior to the first iteration of the loop, we have k D p, so that the subarray AŒp : : k 1 is empty. This empty subarray contains the k p D 0 smallest elements of L and R, and since i D j D 1, both LŒi and RŒj  are the smallest elements of their arrays that have not been copied back into A. 2.3 Designing algorithms 8 9 10 11 12 13 14 15 16 17 A…1 2 2 3 1 2 3 6… k 12345 L 2 4 5 7∞ 12345 R1 2 3 6∞ i j (e) 8 9 10 11 12 13 14 15 16 17 A…1 2 2 3 4 5 3 6… k 12345 L 2 4 5 7∞ 12345 R1 2 3 6∞ i j (g) 8 9 10 11 12 13 14 15 16 17 A…1 2 2 3 4 5 6 7… k 12345 L 2 4 5 7∞ 12345 R1 2 3 6∞ i j (i) 33 8 9 10 11 12 13 14 15 16 17 A…1 2 2 3 4 2 3 6… k 12345 L 2 4 5 7∞ 12345 R1 2 3 6∞ i j (f) 8 9 10 11 12 13 14 15 16 17 A…1 2 2 3 4 5 6 6… k 12345 L 2 4 5 7∞ 12345 R1 2 3 6∞ i j (h) Figure 2.3, continued (i) The arrays and indices at termination. At this point, the subarray in AŒ9 : : 16 is sorted, and the two sentinels in L and R are the only two elements in these arrays that have not been copied into A. Maintenance: To see that each iteration maintains the loop invariant, let us first suppose that LŒi Ä RŒj . Then LŒi is the smallest element not yet copied back into A. Because AŒp : : k 1 contains the k p smallest elements, after line 14 copies LŒi into AŒk, the subarray AŒp : : k will contain the k p C 1 smallest elements. Incrementing k (in the for loop update) and i (in line 15) reestablishes the loop invariant for the next iteration. If instead LŒi > RŒj , then lines 16–17 perform the appropriate action to maintain the loop invariant. Termination: At termination, k D r C 1. By the loop invariant, the subarray AŒp : : k 1, which is AŒp : : r, contains the k p D r p C 1 smallest elements of LŒ1 : : n1 C 1 and RŒ1 : : n2 C 1, in sorted order. The arrays L and R together contain n1 C n2 C 2 D r p C 3 elements. All but the two largest have been copied back into A, and these two largest elements are the sentinels. 34 Chapter 2 Getting Started To see that the MERGE procedure runs in ‚.n/ time, where n D r p C 1, observe that each of lines 1–3 and 8–11 takes constant time, the for loops of lines 4–7 take ‚.n1 C n2/ D ‚.n/ time,7 and there are n iterations of the for loop of lines 12–17, each of which takes constant time. We can now use the MERGE procedure as a subroutine in the merge sort algorithm. The procedure MERGE-SORT.A; p; r/ sorts the elements in the subarray AŒp : : r. If p r, the subarray has at most one element and is therefore already sorted. Otherwise, the divide step simply computes an index q that partitions AŒp : : r into two subarrays: AŒp : : q, containing dn=2e elements, and AŒq C 1 : : r, containing bn=2c elements.8 MERGE-SORT.A; p; r/ 1 if p < r 2 q D b.p C r/=2c 3 MERGE-SORT.A; p; q/ 4 MERGE-SORT.A; q C 1; r/ 5 MERGE.A; p; q; r/ To sort the entire sequence A D hAŒ1; AŒ2; : : : ; AŒni, we make the initial call MERGE-SORT.A; 1; A:length/, where once again A:length D n. Figure 2.4 illustrates the operation of the procedure bottom-up when n is a power of 2. The algorithm consists of merging pairs of 1-item sequences to form sorted sequences of length 2, merging pairs of sequences of length 2 to form sorted sequences of length 4, and so on, until two sequences of length n=2 are merged to form the final sorted sequence of length n. 2.3.2 Analyzing divide-and-conquer algorithms When an algorithm contains a recursive call to itself, we can often describe its running time by a recurrence equation or recurrence, which describes the overall running time on a problem of size n in terms of the running time on smaller inputs. We can then use mathematical tools to solve the recurrence and provide bounds on the performance of the algorithm. 7We shall see in Chapter 3 how to formally interpret equations containing ‚-notation. 8The expression dxe denotes the least integer greater than or equal to x, and bxc denotes the greatest integer less than or equal to x. These notations are defined in Chapter 3. The easiest way to verify that setting q to b.p C r/=2c yields subarrays AŒp : : q and AŒq C 1 : : r of sizes dn=2e and bn=2c, respectively, is to examine the four cases that arise depending on whether each of p and r is odd or even. 2.3 Designing algorithms 35 sorted sequence 12234567 2457 merge 1236 merge 25 47 merge 13 26 merge 5 2 merge merge 4 7 1 3 initial sequence merge 2 6 Figure 2.4 The operation of merge sort on the array A D h5; 2; 4; 7; 1; 3; 2; 6i. The lengths of the sorted sequences being merged increase as the algorithm progresses from bottom to top. A recurrence for the running time of a divide-and-conquer algorithm falls out from the three steps of the basic paradigm. As before, we let T .n/ be the running time on a problem of size n. If the problem size is small enough, say n Ä c for some constant c, the straightforward solution takes constant time, which we write as ‚.1/. Suppose that our division of the problem yields a subproblems, each of which is 1=b the size of the original. (For merge sort, both a and b are 2, but we shall see many divide-and-conquer algorithms in which a ¤ b.) It takes time T .n=b/ to solve one subproblem of size n=b, and so it takes time aT .n=b/ to solve a of them. If we take D.n/ time to divide the problem into subproblems and C.n/ time to combine the solutions to the subproblems into the solution to the original problem, we get the recurrence ( T .n/ D ‚.1/ if n Ä c ; aT .n=b/ C D.n/ C C.n/ otherwise : In Chapter 4, we shall see how to solve common recurrences of this form. Analysis of merge sort Although the pseudocode for MERGE-SORT works correctly when the number of elements is not even, our recurrence-based analysis is simplified if we assume that 36 Chapter 2 Getting Started the original problem size is a power of 2. Each divide step then yields two subsequences of size exactly n=2. In Chapter 4, we shall see that this assumption does not affect the order of growth of the solution to the recurrence. We reason as follows to set up the recurrence for T .n/, the worst-case running time of merge sort on n numbers. Merge sort on just one element takes constant time. When we have n > 1 elements, we break down the running time as follows. Divide: The divide step just computes the middle of the subarray, which takes constant time. Thus, D.n/ D ‚.1/. Conquer: We recursively solve two subproblems, each of size n=2, which contributes 2T .n=2/ to the running time. Combine: We have already noted that the MERGE procedure on an n-element subarray takes time ‚.n/, and so C.n/ D ‚.n/. When we add the functions D.n/ and C.n/ for the merge sort analysis, we are adding a function that is ‚.n/ and a function that is ‚.1/. This sum is a linear function of n, that is, ‚.n/. Adding it to the 2T .n=2/ term from the “conquer” step gives the recurrence for the worst-case running time T .n/ of merge sort: ( ‚.1/ T .n/ D if n D 1 ; (2.1) 2T .n=2/ C ‚.n/ if n > 1 : In Chapter 4, we shall see the “master theorem,” which we can use to show that T .n/ is ‚.n lg n/, where lg n stands for log2 n. Because the logarithm function grows more slowly than any linear function, for large enough inputs, merge sort, with its ‚.n lg n/ running time, outperforms insertion sort, whose running time is ‚.n2/, in the worst case. We do not need the master theorem to intuitively understand why the solution to the recurrence (2.1) is T .n/ D ‚.n lg n/. Let us rewrite recurrence (2.1) as ( T .n/ D c if n D 1 ; (2.2) 2T .n=2/ C cn if n > 1 ; where the constant c represents the time required to solve problems of size 1 as well as the time per array element of the divide and combine steps.9 9It is unlikely that the same constant exactly represents both the time to solve problems of size 1 and the time per array element of the divide and combine steps. We can get around this problem by letting c be the larger of these times and understanding that our recurrence gives an upper bound on the running time, or by letting c be the lesser of these times and understanding that our recurrence gives a lower bound on the running time. Both bounds are on the order of n lg n and, taken together, give a ‚.n lg n/ running time. 2.3 Designing algorithms 37 Figure 2.5 shows how we can solve recurrence (2.2). For convenience, we assume that n is an exact power of 2. Part (a) of the figure shows T .n/, which we expand in part (b) into an equivalent tree representing the recurrence. The cn term is the root (the cost incurred at the top level of recursion), and the two subtrees of the root are the two smaller recurrences T .n=2/. Part (c) shows this process carried one step further by expanding T .n=2/. The cost incurred at each of the two subnodes at the second level of recursion is cn=2. We continue expanding each node in the tree by breaking it into its constituent parts as determined by the recurrence, until the problem sizes get down to 1, each with a cost of c. Part (d) shows the resulting recursion tree. Next, we add the costs across each level of the tree. The top level has total cost cn, the next level down has total cost c.n=2/ C c.n=2/ D cn, the level after that has total cost c.n=4/Cc.n=4/ Cc.n=4/Cc.n=4/ D cn, and so on. In general, the level i below the top has 2i nodes, each contributing a cost of c.n=2i /, so that the i th level below the top has total cost 2i c.n=2i / D cn. The bottom level has n nodes, each contributing a cost of c, for a total cost of cn. The total number of levels of the recursion tree in Figure 2.5 is lg n C 1, where n is the number of leaves, corresponding to the input size. An informal inductive argument justifies this claim. The base case occurs when n D 1, in which case the tree has only one level. Since lg 1 D 0, we have that lg n C 1 gives the correct number of levels. Now assume as an inductive hypothesis that the number of levels of a recursion tree with 2i leaves is lg 2i C 1 D i C 1 (since for any value of i , we have that lg 2i D i). Because we are assuming that the input size is a power of 2, the next input size to consider is 2iC1. A tree with n D 2iC1 leaves has one more level than a tree with 2i leaves, and so the total number of levels is .i C 1/ C 1 D lg 2iC1 C 1. To compute the total cost represented by the recurrence (2.2), we simply add up the costs of all the levels. The recursion tree has lg n C 1 levels, each costing cn, for a total cost of cn.lg n C 1/ D cn lg n C cn. Ignoring the low-order term and the constant c gives the desired result of ‚.n lg n/. Exercises 2.3-1 Using Figure 2.4 as a model, illustrate the operation of merge sort on the array A D h3; 41; 52; 26; 38; 57; 9; 49i. 2.3-2 Rewrite the MERGE procedure so that it does not use sentinels, instead stopping once either array L or R has had all its elements copied back to A and then copying the remainder of the other array back into A. 38 Chapter 2 Getting Started T(n) cn cn T(n/2) T(n/2) cn/2 cn/2 T(n/4) T(n/4) T(n/4) T(n/4) (a) (b) (c) cn cn cn/2 cn/2 cn lg n cn/4 cn/4 cn/4 cn/4 cn … c c c c c …c c cn n (d) Total: cn lg n + cn Figure 2.5 How to construct a recursion tree for the recurrence T .n/ D 2T .n=2/ C cn. Part (a) shows T .n/, which progressively expands in (b)–(d) to form the recursion tree. The fully expanded tree in part (d) has lg n C 1 levels (i.e., it has height lg n, as indicated), and each level contributes a total cost of cn. The total cost, therefore, is cn lg n C cn, which is ‚.n lg n/. Problems for Chapter 2 39 2.3-3 Use mathematical induction to show that when n is an exact power of 2, the solu- tion of th(e recurrence 2 if n D 2 ; T .n/ D 2T .n=2/ C n if n D 2k, for k > 1 is T .n/ D n lg n. 2.3-4 We can express insertion sort as a recursive procedure as follows. In order to sort AŒ1 : : n, we recursively sort AŒ1 : : n 1 and then insert AŒn into the sorted array AŒ1 : : n 1. Write a recurrence for the running time of this recursive version of insertion sort. 2.3-5 Referring back to the searching problem (see Exercise 2.1-3), observe that if the sequence A is sorted, we can check the midpoint of the sequence against and eliminate half of the sequence from further consideration. The binary search algorithm repeats this procedure, halving the size of the remaining portion of the sequence each time. Write pseudocode, either iterative or recursive, for binary search. Argue that the worst-case running time of binary search is ‚.lg n/. 2.3-6 Observe that the while loop of lines 5–7 of the INSERTION-SORT procedure in Section 2.1 uses a linear search to scan (backward) through the sorted subarray AŒ1 : : j 1. Can we use a binary search (see Exercise 2.3-5) instead to improve the overall worst-case running time of insertion sort to ‚.n lg n/? 2.3-7 ? Describe a ‚.n lg n/-time algorithm that, given a set S of n integers and another integer x, determines whether or not there exist two elements in S whose sum is exactly x. Problems 2-1 Insertion sort on small arrays in merge sort Although merge sort runs in ‚.n lg n/ worst-case time and insertion sort runs in ‚.n2/ worst-case time, the constant factors in insertion sort can make it faster in practice for small problem sizes on many machines. Thus, it makes sense to coarsen the leaves of the recursion by using insertion sort within merge sort when 40 Chapter 2 Getting Started subproblems become sufficiently small. Consider a modification to merge sort in which n=k sublists of length k are sorted using insertion sort and then merged using the standard merging mechanism, where k is a value to be determined. a. Show that insertion sort can sort the n=k sublists, each of length k, in ‚.nk/ worst-case time. b. Show how to merge the sublists in ‚.n lg.n=k// worst-case time. c. Given that the modified algorithm runs in ‚.nk C n lg.n=k// worst-case time, what is the largest value of k as a function of n for which the modified algorithm has the same running time as standard merge sort, in terms of ‚-notation? d. How should we choose k in practice? 2-2 Correctness of bubblesort Bubblesort is a popular, but inefficient, sorting algorithm. It works by repeatedly swapping adjacent elements that are out of order. BUBBLESORT.A/ 1 for i D 1 to A:length 1 2 for j D A:length downto i C 1 3 if AŒj  < AŒj 1 4 exchange AŒj  with AŒj 1 a. Let A0 denote the output of BUBBLESORT.A/. To prove that BUBBLESORT is correct, we need to prove that it terminates and that A0Œ1 Ä A0Œ2 Ä Ä A0Œn ; (2.3) where n D A:length. In order to show that BUBBLESORT actually sorts, what else do we need to prove? The next two parts will prove inequality (2.3). b. State precisely a loop invariant for the for loop in lines 2–4, and prove that this loop invariant holds. Your proof should use the structure of the loop invariant proof presented in this chapter. c. Using the termination condition of the loop invariant proved in part (b), state a loop invariant for the for loop in lines 1–4 that will allow you to prove inequality (2.3). Your proof should use the structure of the loop invariant proof presented in this chapter. Problems for Chapter 2 41 d. What is the worst-case running time of bubblesort? How does it compare to the running time of insertion sort? 2-3 Correctness of Horner’s rule The following code fragment implements Horner’s rule for evaluating a polynomial Xn P .x/ D ak x k kD0 D a0 C x.a1 C x.a2 C C x.an 1 C xan/ // ; given the coefficients a0; a1; : : : ; an and a value for x: 1 yD0 2 for i D n downto 0 3 y D ai C x y a. In terms of ‚-notation, what is the running time of this code fragment for Horner’s rule? b. Write pseudocode to implement the naive polynomial-evaluation algorithm that computes each term of the polynomial from scratch. What is the running time of this algorithm? How does it compare to Horner’s rule? c. Consider the following loop invariant: At the start of each iteration of the for loop of lines 2–3, n X .i C1/ yD akCiC1xk : kD0 Interpret a summation with no terms as equaling 0. Following the structure of tthheatl,oaotpteirnmvairniaatniot np,royofDprPesnkeDnt0eadkixnkt.his chapter, use this loop invariant to show d. Conclude by arguing that the given code fragment correctly evaluates a polynomial characterized by the coefficients a0; a1; : : : ; an. 2-4 Inversions Let AŒ1 : : n be an array of n distinct numbers. If i < j and AŒi > AŒj , then the pair .i; j / is called an inversion of A. a. List the five inversions of the array h2; 3; 8; 6; 1i. 42 Chapter 2 Getting Started b. What array with elements from the set f1; 2; : : : ; ng has the most inversions? How many does it have? c. What is the relationship between the running time of insertion sort and the number of inversions in the input array? Justify your answer. d. Give an algorithm that determines the number of inversions in any permutation on n elements in ‚.n lg n/ worst-case time. (Hint: Modify merge sort.) Chapter notes In 1968, Knuth published the first of three volumes with the general title The Art of Computer Programming [209, 210, 211]. The first volume ushered in the modern study of computer algorithms with a focus on the analysis of running time, and the full series remains an engaging and worthwhile reference for many of the topics presented here. According to Knuth, the word “algorithm” is derived from the name “al-Khowaˆrizmˆı,” a ninth-century Persian mathematician. Aho, Hopcroft, and Ullman [5] advocated the asymptotic analysis of algorithms—using notations that Chapter 3 introduces, including ‚-notation—as a means of comparing relative performance. They also popularized the use of recurrence relations to describe the running times of recursive algorithms. Knuth [211] provides an encyclopedic treatment of many sorting algorithms. His comparison of sorting algorithms (page 381) includes exact step-counting analyses, like the one we performed here for insertion sort. Knuth’s discussion of insertion sort encompasses several variations of the algorithm. The most important of these is Shell’s sort, introduced by D. L. Shell, which uses insertion sort on periodic subsequences of the input to produce a faster sorting algorithm. Merge sort is also described by Knuth. He mentions that a mechanical collator capable of merging two decks of punched cards in a single pass was invented in 1938. J. von Neumann, one of the pioneers of computer science, apparently wrote a program for merge sort on the EDVAC computer in 1945. The early history of proving programs correct is described by Gries [153], who credits P. Naur with the first article in this field. Gries attributes loop invariants to R. W. Floyd. The textbook by Mitchell [256] describes more recent progress in proving programs correct. 3 Growth of Functions The order of growth of the running time of an algorithm, defined in Chapter 2, gives a simple characterization of the algorithm’s efficiency and also allows us to compare the relative performance of alternative algorithms. Once the input size n becomes large enough, merge sort, with its ‚.n lg n/ worst-case running time, beats insertion sort, whose worst-case running time is ‚.n2/. Although we can sometimes determine the exact running time of an algorithm, as we did for insertion sort in Chapter 2, the extra precision is not usually worth the effort of computing it. For large enough inputs, the multiplicative constants and lower-order terms of an exact running time are dominated by the effects of the input size itself. When we look at input sizes large enough to make only the order of growth of the running time relevant, we are studying the asymptotic efficiency of algorithms. That is, we are concerned with how the running time of an algorithm increases with the size of the input in the limit, as the size of the input increases without bound. Usually, an algorithm that is asymptotically more efficient will be the best choice for all but very small inputs. This chapter gives several standard methods for simplifying the asymptotic analysis of algorithms. The next section begins by defining several types of “asymptotic notation,” of which we have already seen an example in ‚-notation. We then present several notational conventions used throughout this book, and finally we review the behavior of functions that commonly arise in the analysis of algorithms. 3.1 Asymptotic notation The notations we use to describe the asymptotic running time of an algorithm are defined in terms of functions whose domains are the set of natural numbers N D f0; 1; 2; : : :g. Such notations are convenient for describing the worst-case running-time function T .n/, which usually is defined only on integer input sizes. We sometimes find it convenient, however, to abuse asymptotic notation in a va- 44 Chapter 3 Growth of Functions riety of ways. For example, we might extend the notation to the domain of real numbers or, alternatively, restrict it to a subset of the natural numbers. We should make sure, however, to understand the precise meaning of the notation so that when we abuse, we do not misuse it. This section defines the basic asymptotic notations and also introduces some common abuses. Asymptotic notation, functions, and running times We will use asymptotic notation primarily to describe the running times of algorithms, as when we wrote that insertion sort’s worst-case running time is ‚.n2/. Asymptotic notation actually applies to functions, however. Recall that we characterized insertion sort’s worst-case running time as an2 CbnCc, for some constants a, b, and c. By writing that insertion sort’s running time is ‚.n2/, we abstracted away some details of this function. Because asymptotic notation applies to functions, what we were writing as ‚.n2/ was the function an2 C bn C c, which in that case happened to characterize the worst-case running time of insertion sort. In this book, the functions to which we apply asymptotic notation will usually characterize the running times of algorithms. But asymptotic notation can apply to functions that characterize some other aspect of algorithms (the amount of space they use, for example), or even to functions that have nothing whatsoever to do with algorithms. Even when we use asymptotic notation to apply to the running time of an algorithm, we need to understand which running time we mean. Sometimes we are interested in the worst-case running time. Often, however, we wish to characterize the running time no matter what the input. In other words, we often wish to make a blanket statement that covers all inputs, not just the worst case. We shall see asymptotic notations that are well suited to characterizing running times no matter what the input. ‚-notation In Chapter 2, we found that the worst-case running time of insertion sort is T .n/ D ‚.n2/. Let us define what this notation means. For a given function g.n/, we denote by ‚.g.n// the set of functions ‚.g.n// D ff .n/ W there exist positive constants c1, c2, and n0 such that 0 Ä c1g.n/ Ä f .n/ Ä c2g.n/ for all n n0g :1 1Within set notation, a colon means “such that.” 3.1 Asymptotic notation c2g.n/ f .n/ c1g.n/ cg.n/ f .n/ 45 f .n/ cg.n/ n n0 f .n/ D ‚.g.n// (a) n n0 f .n/ D O.g.n// (b) n n0 f .n/ D .g.n// (c) Figure 3.1 Graphic examples of the ‚, O, and notations. In each part, the value of n0 shown is the minimum possible value; any greater value would also work. (a) ‚-notation bounds a function to within constant factors. We write f .n/ D ‚.g.n// if there exist positive constants n0, c1, and c2 such that at and to the right of n0, the value of f .n/ always lies between c1g.n/ and c2g.n/ inclusive. (b) O-notation gives an upper bound for a function to within a constant factor. We write f .n/ D O.g.n// if there are positive constants n0 and c such that at and to the right of n0, the value of f .n/ always lies on or below cg.n/. (c) -notation gives a lower bound for a function to within a constant factor. We write f .n/ D .g.n// if there are positive constants n0 and c such that at and to the right of n0, the value of f .n/ always lies on or above cg.n/. A function f .n/ belongs to the set ‚.g.n// if there exist positive constants c1 and c2 such that it can be “sandwiched” between c1g.n/ and c2g.n/, for sufficiently large n. Because ‚.g.n// is a set, we could write “f .n/ 2 ‚.g.n//” to indicate that f .n/ is a member of ‚.g.n//. Instead, we will usually write “f .n/ D ‚.g.n//” to express the same notion. You might be confused because we abuse equality in this way, but we shall see later in this section that doing so has its advantages. Figure 3.1(a) gives an intuitive picture of functions f .n/ and g.n/, where f .n/ D ‚.g.n//. For all values of n at and to the right of n0, the value of f .n/ lies at or above c1g.n/ and at or below c2g.n/. In other words, for all n n0, the function f .n/ is equal to g.n/ to within a constant factor. We say that g.n/ is an asymptotically tight bound for f .n/. The definition of ‚.g.n// requires that every member f .n/ 2 ‚.g.n// be asymptotically nonnegative, that is, that f .n/ be nonnegative whenever n is sufficiently large. (An asymptotically positive function is one that is positive for all sufficiently large n.) Consequently, the function g.n/ itself must be asymptotically nonnegative, or else the set ‚.g.n// is empty. We shall therefore assume that every function used within ‚-notation is asymptotically nonnegative. This assumption holds for the other asymptotic notations defined in this chapter as well. 46 Chapter 3 Growth of Functions In Chapter 2, we introduced an informal notion of ‚-notation that amounted to throwing away lower-order terms and ignoring the leading coefficient of the highest-order term. Let us briefly justify this intuition by using the formal defi- nition to show that 1 2 n2 3n D ‚.n2/. To do so, we must determine positive constants c1, c2, and n0 such that c1n2 Ä 1 n2 2 3n Ä c2n2 for all n n0. Dividing by n2 yields 13 c1 Ä 2 n Ä c2 : We can make the right-hand inequality hold for any value of n 1 by choosing any constant c2 1=2. Likewise, we can make the left-hand inequality hold for any value of n 7 by choosing any constant c1 Ä 1=14. Thus, by choosing c1 D 1=14, c2 D 1=2, and n0 D 7, we can verify that 1 2 n2 3n D ‚.n2/. Certainly, other choices for the constants exist, but the important thing is that some choice exists. Note that these constants depend on the function 1 2 n2 3n; a different function belonging to ‚.n2/ would usually require different constants. We can also use the formal definition to verify that 6n3 ¤ ‚.n2/. Suppose for the purpose of contradiction that c2 and n0 exist such that 6n3 Ä c2n2 for all n n0. But then dividing by n2 yields n Ä c2=6, which cannot possibly hold for arbitrarily large n, since c2 is constant. Intuitively, the lower-order terms of an asymptotically positive function can be ignored in determining asymptotically tight bounds because they are insignificant for large n. When n is large, even a tiny fraction of the highest-order term suf- fices to dominate the lower-order terms. Thus, setting c1 to a value that is slightly smaller than the coefficient of the highest-order term and setting c2 to a value that is slightly larger permits the inequalities in the definition of ‚-notation to be sat- isfied. The coefficient of the highest-order term can likewise be ignored, since it only changes c1 and c2 by a constant factor equal to the coefficient. As an example, consider any quadratic function f .n/ D an2 C bn C c, where a, b, and c are constants and a > 0. Throwing away the lower-order terms and ignoring the constant yields f .n/ D ‚.n2/. Formally, to show the spame thing, we take the constants c1 D a=4, c2 D 7a=4, and n0 D 2 max.jbj =a; jcj =a/. You may verify that 0 Ä for any polynomial p .cn1/n2DÄPadi Dn02 C bn C c Ä c2n2 for all n ai ni , where the ai are constants n0. and In ad general, > 0, we have p.n/ D ‚.nd / (see Problem 3-1). Since any constant is a degree-0 polynomial, we can express any constant func- tion as ‚.n0/, or ‚.1/. This latter notation is a minor abuse, however, because the 3.1 Asymptotic notation 47 expression does not indicate what variable is tending to infinity.2 We shall often use the notation ‚.1/ to mean either a constant or a constant function with respect to some variable. O-notation The ‚-notation asymptotically bounds a function from above and below. When we have only an asymptotic upper bound, we use O-notation. For a given function g.n/, we denote by O.g.n// (pronounced “big-oh of g of n” or sometimes just “oh of g of n”) the set of functions O.g.n// D ff .n/ W there exist positive constants c and n0 such that 0 Ä f .n/ Ä cg.n/ for all n n0g : We use O-notation to give an upper bound on a function, to within a constant factor. Figure 3.1(b) shows the intuition behind O-notation. For all values n at and to the right of n0, the value of the function f .n/ is on or below cg.n/. We write f .n/ D O.g.n// to indicate that a function f .n/ is a member of the set O.g.n//. Note that f .n/ D ‚.g.n// implies f .n/ D O.g.n//, since ‚notation is a stronger notion than O-notation. Written set-theoretically, we have ‚.g.n//  O.g.n//. Thus, our proof that any quadratic function an2 C bn C c, where a > 0, is in ‚.n2/ also shows that any such quadratic function is in O.n2/. What may be more surprising is that when a > 0, any linear function an C b is in O.n2/, which is easily verified by taking c D a C jbj and n0 D max.1; b=a/. If you have seen O-notation before, you might find it strange that we should write, for example, n D O.n2/. In the literature, we sometimes find O-notation informally describing asymptotically tight bounds, that is, what we have defined using ‚-notation. In this book, however, when we write f .n/ D O.g.n//, we are merely claiming that some constant multiple of g.n/ is an asymptotic upper bound on f .n/, with no claim about how tight an upper bound it is. Distinguishing asymptotic upper bounds from asymptotically tight bounds is standard in the algorithms literature. Using O-notation, we can often describe the running time of an algorithm merely by inspecting the algorithm’s overall structure. For example, the doubly nested loop structure of the insertion sort algorithm from Chapter 2 immediately yields an O.n2/ upper bound on the worst-case running time: the cost of each iteration of the inner loop is bounded from above by O.1/ (constant), the indices i 2The real problem is that our ordinary notation for functions does not distinguish functions from values. In -calculus, the parameters to a function are clearly specified: the function n2 could be written as n:n2, or even r:r2. Adopting a more rigorous notation, however, would complicate algebraic manipulations, and so we choose to tolerate the abuse. 48 Chapter 3 Growth of Functions and j are both at most n, and the inner loop is executed at most once for each of the n2 pairs of values for i and j . Since O-notation describes an upper bound, when we use it to bound the worstcase running time of an algorithm, we have a bound on the running time of the algorithm on every input—the blanket statement we discussed earlier. Thus, the O.n2/ bound on worst-case running time of insertion sort also applies to its running time on every input. The ‚.n2/ bound on the worst-case running time of insertion sort, however, does not imply a ‚.n2/ bound on the running time of insertion sort on every input. For example, we saw in Chapter 2 that when the input is already sorted, insertion sort runs in ‚.n/ time. Technically, it is an abuse to say that the running time of insertion sort is O.n2/, since for a given n, the actual running time varies, depending on the particular input of size n. When we say “the running time is O.n2/,” we mean that there is a function f .n/ that is O.n2/ such that for any value of n, no matter what particular input of size n is chosen, the running time on that input is bounded from above by the value f .n/. Equivalently, we mean that the worst-case running time is O.n2/. -notation Just as O-notation provides an asymptotic upper bound on a function, -notation provides an asymptotic lower bound. For a given function g.n/, we denote by .g.n// (pronounced “big-omega of g of n” or sometimes just “omega of g of n”) the set of functions .g.n// D ff .n/ W there exist positive constants c and n0 such that 0 Ä cg.n/ Ä f .n/ for all n n0g : Figure 3.1(c) shows the intuition behind -notation. For all values n at or to the right of n0, the value of f .n/ is on or above cg.n/. From the definitions of the asymptotic notations we have seen thus far, it is easy to prove the following important theorem (see Exercise 3.1-5). Theorem 3.1 For any two functions f .n/ and g.n/, we have f .n/ D ‚.g.n// if and only if f .n/ D O.g.n// and f .n/ D .g.n//. As an example of the application of this theorem, our proof that an2 C bn C c D ‚.n2/ for any constants a, b, and c, where a > 0, immediately implies that an2 C bn C c D .n2/ and an2 C bn C c D O.n2/. In practice, rather than using Theorem 3.1 to obtain asymptotic upper and lower bounds from asymptotically tight bounds, as we did for this example, we usually use it to prove asymptotically tight bounds from asymptotic upper and lower bounds. 3.1 Asymptotic notation 49 When we say that the running time (no modifier) of an algorithm is .g.n//, we mean that no matter what particular input of size n is chosen for each value of n, the running time on that input is at least a constant times g.n/, for sufficiently large n. Equivalently, we are giving a lower bound on the best-case running time of an algorithm. For example, the best-case running time of insertion sort is .n/, which implies that the running time of insertion sort is .n/. The running time of insertion sort therefore belongs to both .n/ and O.n2/, since it falls anywhere between a linear function of n and a quadratic function of n. Moreover, these bounds are asymptotically as tight as possible: for instance, the running time of insertion sort is not .n2/, since there exists an input for which insertion sort runs in ‚.n/ time (e.g., when the input is already sorted). It is not contradictory, however, to say that the worst-case running time of insertion sort is .n2/, since there exists an input that causes the algorithm to take .n2/ time. Asymptotic notation in equations and inequalities We have already seen how asymptotic notation can be used within mathematical formulas. For example, in introducing O-notation, we wrote “n D O.n2/.” We might also write 2n2 C 3n C 1 D 2n2 C ‚.n/. How do we interpret such formulas? When the asymptotic notation stands alone (that is, not within a larger formula) on the right-hand side of an equation (or inequality), as in n D O.n2/, we have already defined the equal sign to mean set membership: n 2 O.n2/. In general, however, when asymptotic notation appears in a formula, we interpret it as standing for some anonymous function that we do not care to name. For example, the formula 2n2 C 3n C 1 D 2n2 C ‚.n/ means that 2n2 C 3n C 1 D 2n2 C f .n/, where f .n/ is some function in the set ‚.n/. In this case, we let f .n/ D 3n C 1, which indeed is in ‚.n/. Using asymptotic notation in this manner can help eliminate inessential detail and clutter in an equation. For example, in Chapter 2 we expressed the worst-case running time of merge sort as the recurrence T .n/ D 2T .n=2/ C ‚.n/ : If we are interested only in the asymptotic behavior of T .n/, there is no point in specifying all the lower-order terms exactly; they are all understood to be included in the anonymous function denoted by the term ‚.n/. The number of anonymous functions in an expression is understood to be equal to the number of times the asymptotic notation appears. For example, in the expression Xn O.i/ ; i D1 50 Chapter 3 Growth of Functions there is only a single anonymous function (a function of i). This expression is thus not the same as O.1/ C O.2/ C C O.n/, which doesn’t really have a clean interpretation. In some cases, asymptotic notation appears on the left-hand side of an equation, as in 2n2 C ‚.n/ D ‚.n2/ : We interpret such equations using the following rule: No matter how the anonymous functions are chosen on the left of the equal sign, there is a way to choose the anonymous functions on the right of the equal sign to make the equation valid. Thus, our example means that for any function f .n/ 2 ‚.n/, there is some function g.n/ 2 ‚.n2/ such that 2n2 C f .n/ D g.n/ for all n. In other words, the right-hand side of an equation provides a coarser level of detail than the left-hand side. We can chain together a number of such relationships, as in 2n2 C 3n C 1 D 2n2 C ‚.n/ D ‚.n2/ : We can interpret each equation separately by the rules above. The first equation says that there is some function f .n/ 2 ‚.n/ such that 2n2 C 3n C 1 D 2n2 C f .n/ for all n. The second equation says that for any function g.n/ 2 ‚.n/ (such as the f .n/ just mentioned), there is some function h.n/ 2 ‚.n2/ such that 2n2 C g.n/ D h.n/ for all n. Note that this interpretation implies that 2n2 C 3n C 1 D ‚.n2/, which is what the chaining of equations intuitively gives us. o-notation The asymptotic upper bound provided by O-notation may or may not be asymptotically tight. The bound 2n2 D O.n2/ is asymptotically tight, but the bound 2n D O.n2/ is not. We use o-notation to denote an upper bound that is not asymptotically tight. We formally define o.g.n// (“little-oh of g of n”) as the set o.g.n// D ff .n/ W for any positive constant c > 0, there exists a constant n0 > 0 such that 0 Ä f .n/ < cg.n/ for all n n0g : For example, 2n D o.n2/, but 2n2 ¤ o.n2/. The definitions of O-notation and o-notation are similar. The main difference is that in f .n/ D O.g.n//, the bound 0 Ä f .n/ Ä cg.n/ holds for some constant c > 0, but in f .n/ D o.g.n//, the bound 0 Ä f .n/ < cg.n/ holds for all constants c > 0. Intuitively, in o-notation, the function f .n/ becomes insignificant relative to g.n/ as n approaches infinity; that is, 3.1 Asymptotic notation 51 f .n/ lim D0: n!1 g.n/ (3.1) Some authors use this limit as a definition of the o-notation; the definition in this book also restricts the anonymous functions to be asymptotically nonnegative. !-notation By analogy, !-notation is to -notation as o-notation is to O-notation. We use !-notation to denote a lower bound that is not asymptotically tight. One way to define it is by f .n/ 2 !.g.n// if and only if g.n/ 2 o.f .n// : Formally, however, we define !.g.n// (“little-omega of g of n”) as the set !.g.n// D ff .n/ W for any positive constant c > 0, there exists a constant n0 > 0 such that 0 Ä cg.n/ < f .n/ for all n n0g : For example, n2=2 D !.n/, but n2=2 ¤ !.n2/. The relation f .n/ D !.g.n// implies that f .n/ lim D1; n!1 g.n/ if the limit exists. That is, f .n/ becomes arbitrarily large relative to g.n/ as n approaches infinity. Comparing functions Many of the relational properties of real numbers apply to asymptotic comparisons as well. For the following, assume that f .n/ and g.n/ are asymptotically positive. Transitivity: f .n/ D ‚.g.n// and g.n/ D ‚.h.n// f .n/ D O.g.n// and g.n/ D O.h.n// f .n/ D .g.n// and g.n/ D .h.n// f .n/ D o.g.n// and g.n/ D o.h.n// f .n/ D !.g.n// and g.n/ D !.h.n// imply imply imply imply imply f .n/ D ‚.h.n// ; f .n/ D O.h.n// ; f .n/ D .h.n// ; f .n/ D o.h.n// ; f .n/ D !.h.n// : Reflexivity: f .n/ D ‚.f .n// ; f .n/ D O.f .n// ; f .n/ D .f .n// : 52 Chapter 3 Growth of Functions Symmetry: f .n/ D ‚.g.n// if and only if g.n/ D ‚.f .n// : Transpose symmetry: f .n/ D O.g.n// if and only if g.n/ D .f .n// ; f .n/ D o.g.n// if and only if g.n/ D !.f .n// : Because these properties hold for asymptotic notations, we can draw an analogy between the asymptotic comparison of two functions f and g and the comparison of two real numbers a and b: f .n/ D O.g.n// f .n/ D .g.n// f .n/ D ‚.g.n// f .n/ D o.g.n// f .n/ D !.g.n// is like is like is like is like is like aÄb; a b; aDb; ab: We say that f .n/ is asymptotically smaller than g.n/ if f .n/ D o.g.n//, and f .n/ is asymptotically larger than g.n/ if f .n/ D !.g.n//. One property of real numbers, however, does not carry over to asymptotic notation: Trichotomy: For any two real numbers a and b, exactly one of the following must hold: a < b, a D b, or a > b. Although any two real numbers can be compared, not all functions are asymptot- ically comparable. That is, for two functions f .n/ and g.n/, it may be the case that neither f .n/ D O.g.n// nor f .n/ D .g.n// holds. For example, we cannot compare the functions n and n1Csin n using asymptotic notation, since the value of the exponent in n1Csin n oscillates between 0 and 2, taking on all values in between. Exercises 3.1-1 Let f .n/ and g.n/ be asymptotically nonnegative functions. Using the basic definition of ‚-notation, prove that max.f .n/; g.n// D ‚.f .n/ C g.n//. 3.1-2 Show that for any real constants a and b, where b > 0, .n C a/b D ‚.nb/ : (3.2) 3.2 Standard notations and common functions 53 3.1-3 Explain why the statement, “The running time of algorithm A is at least O.n2/,” is meaningless. 3.1-4 Is 2nC1 D O.2n/? Is 22n D O.2n/? 3.1-5 Prove Theorem 3.1. 3.1-6 Prove that the running time of an algorithm is ‚.g.n// if and only if its worst-case running time is O.g.n// and its best-case running time is .g.n//. 3.1-7 Prove that o.g.n// \ !.g.n// is the empty set. 3.1-8 We can extend our notation to the case of two parameters n and m that can go to infinity independently at different rates. For a given function g.n; m/, we denote by O.g.n; m// the set of functions O.g.n; m// D ff .n; m/ W there exist positive constants c, n0, and m0 such that 0 Ä f .n; m/ Ä cg.n; m/ for all n n0 or m m0g : Give corresponding definitions for .g.n; m// and ‚.g.n; m//. 3.2 Standard notations and common functions This section reviews some standard mathematical functions and notations and explores the relationships among them. It also illustrates the use of the asymptotic notations. Monotonicity A function f .n/ is monotonically increasing if m Ä n implies f .m/ Ä f .n/. Similarly, it is monotonically decreasing if m Ä n implies f .m/ f .n/. A function f .n/ is strictly increasing if m < n implies f .m/ < f .n/ and strictly decreasing if m < n implies f .m/ > f .n/. 54 Chapter 3 Growth of Functions Floors and ceilings For any real number x, we denote the greatest integer less than or equal to x by bxc (read “the floor of x”) and the least integer greater than or equal to x by dxe (read “the ceiling of x”). For all real x, x 1 < bxc Ä x Ä dxe < x C 1 : (3.3) For any integer n, dn=2e C bn=2c D n ; and for any real number x 0 and integers a; b > 0, dx=ae D lxm ; (3.4) b ab bx=ac D jxk ; (3.5) b ab lam a C .b 1/ Ä ; (3.6) j b a k b a .b 1/ : (3.7) b b The floor function f .x/ D bxc is monotonically increasing, as is the ceiling function f .x/ D dxe. Modular arithmetic For any integer a and any positive integer n, the value a mod n is the remainder (or residue) of the quotient a=n: a mod n D a n ba=nc : (3.8) It follows that 0 Ä a mod n < n : (3.9) Given a well-defined notion of the remainder of one integer when divided by another, it is convenient to provide special notation to indicate equality of remainders. If .a mod n/ D .b mod n/, we write a Á b .mod n/ and say that a is equivalent to b, modulo n. In other words, a Á b .mod n/ if a and b have the same remainder when divided by n. Equivalently, a Á b .mod n/ if and only if n is a divisor of b a. We write a 6Á b .mod n/ if a is not equivalent to b, modulo n. 3.2 Standard notations and common functions 55 Polynomials Given a nonnegative integer d , a polynomial in n of degree d is a function p.n/ of the form Xd p.n/ D ai ni ; i D0 where the constants a0; a1; : : : ; ad are the coefficients of the polynomial and ad ¤ 0. A polynomial is asymptotically positive if and only if ad > 0. For an asymptotically positive polynomial p.n/ of degree d , we have p.n/ D ‚.nd /. For any real constant a 0, the function na is monotonically increasing, and for any real constant a Ä 0, the function na is monotonically decreasing. We say that a function f .n/ is polynomially bounded if f .n/ D O.nk/ for some constant k. Exponentials For all real a > 0, m, and n, we have the following identities: a0 D 1 ; a1 D a ; a 1 D 1=a ; .am/n D amn ; .am/n D .an/m ; aman D amCn : For all n and a 1, the function an is monotonically increasing in n. When convenient, we shall assume 00 D 1. We can relate the rates of growth of polynomials and exponentials by the following fact. For all real constants a and b such that a > 1, nb lim D 0 ; n!1 an from which we can conclude that (3.10) nb D o.an/ : Thus, any exponential function with a base strictly greater than 1 grows faster than any polynomial function. Using e to denote 2:71828 : : :, the base of the natural logarithm function, we have for all real x, ex D 1 C x C x2 C x3 C 2Š 3Š X 1 xi D ; iŠ i D0 (3.11) 56 Chapter 3 Growth of Functions where “Š” denotes the factorial function defined later in this section. For all real x, we have the inequality ex 1 C x ; (3.12) where equality holds only when x D 0. When jxj Ä 1, we have the approximation 1 C x Ä ex Ä 1 C x C x2 : (3.13) When x ! 0, the approximation of ex by 1 C x is quite good: ex D 1 C x C ‚.x2/ : (In this equation, the asymptotic notation is used to describe the limiting behavior as x ! 0 rather than as x ! 1.) We have for all x, lim 1 C x Án D ex : n!1 n (3.14) Logarithms We shall use the following notations: lg n D log2 n ln n D loge n lgk n D .lg n/k (binary logarithm) , (natural logarithm) , (exponentiation) , lg lg n D lg.lg n/ (composition) . An important notational convention we shall adopt is that logarithm functions will apply only to the next term in the formula, so that lg n C k will mean .lg n/ C k and not lg.n C k/. If we hold b > 1 constant, then for n > 0, the function logb n is strictly increasing. For all real a > 0, b > 0, c > 0, and n, a D blogb a ; logc.ab/ D logc a C logc b ; logb an D n logb a ; logb a D logc a ; logc b logb.1=a/ D logb a ; 1 logb a D ; loga b alogb c D clogb a ; (3.15) (3.16) where, in each equation above, logarithm bases are not 1. 3.2 Standard notations and common functions 57 By equation (3.15), changing the base of a logarithm from one constant to another changes the value of the logarithm by only a constant factor, and so we shall often use the notation “lg n” when we don’t care about constant factors, such as in O-notation. Computer scientists find 2 to be the most natural base for logarithms because so many algorithms and data structures involve splitting a problem into two parts. There is a simple series expansion for ln.1 C x/ when jxj < 1: x2 x3 x4 x5 ln.1 C x/ D x C C : 2 34 5 We also have the following inequalities for x > 1: x Ä ln.1 C x/ Ä x ; 1Cx (3.17) where equality holds only for x D 0. We say that a function f .n/ is polylogarithmically bounded if f .n/ D O.lgk n/ for some constant k. We can relate the growth of polynomials and polylogarithms by substituting lg n for n and 2a for a in equation (3.10), yielding lgb n lgb n lim n!1 .2a /lg n D lim n!1 na D0: From this limit, we can conclude that lgb n D o.na/ for any constant a > 0. Thus, any positive polynomial function grows faster than any polylogarithmic function. Factorials The notation nŠ (read “n factorial”) is defined for integers n 0 as ( nŠ D 1 if n D 0 ; n .n 1/Š if n > 0 : Thus, nŠ D 1 2 3 n. A weak upper bound on the factorial function is nŠ Ä nn, since each of the n terms in the factorial product is at most n. Stirling’s approximation, p n Án   Ãà 1 nŠ D 2 n 1C‚ ; e n (3.18) 58 Chapter 3 Growth of Functions where e is the base of the natural logarithm, gives us a tighter upper bound, and a lower bound as well. As Exercise 3.2-3 asks you to prove, nŠ D o.nn/ ; nŠ D !.2n/ ; lg.nŠ/ D ‚.n lg n/ ; (3.19) where Stirling’s approximation is helpful in proving equation (3.19). The following equation also holds for all n 1: p nŠ D 2 n n Án e˛n e (3.20) where 1 1 12n C 1 < ˛n < 12n : (3.21) Functional iteration We use the notation f .i/.n/ to denote the function f .n/ iteratively applied i times to an initial value of n. Formally, let f .n/ be a function over the reals. For non- negative integers i, we recursively define ( f .i/.n/ D n if i D 0 ; f .f .i 1/.n// if i > 0 : For example, if f .n/ D 2n, then f .i/.n/ D 2i n. The iterated logarithm function We use the notation lg n (read “log star of n”) to denote the iterated logarithm, defined as follows. Let lg.i/ n be as defined above, with f .n/ D lg n. Because the logarithm of a nonpositive number is undefined, lg.i/ n is defined only if lg.i 1/ n > 0. Be sure to distinguish lg.i/ n (the logarithm function applied i times in succession, starting with argument n) from lgi n (the logarithm of n raised to the ith power). Then we define the iterated logarithm function as ˚ lg n D min i 0 W lg.i/ n Ä « 1 : The iterated logarithm is a very slowly growing function: lg 2 D 1 ; lg 4 D 2 ; lg 16 D 3 ; lg 65536 D 4 ; lg .265536/ D 5 : 3.2 Standard notations and common functions 59 Since the number of atoms in the observable universe is estimated to be about 1080, which is much less than 265536, we rarely encounter an input size n such that lg n > 5. Fibonacci numbers We define the Fibonacci numbers by the following recurrence: F0 D 0 ; F1 D 1 ; Fi D Fi 1 C Fi 2 for i 2 : (3.22) Thus, each Fibonacci number is the sum of the two previous ones, yielding the sequence 0; 1; 1; 2; 3; 5; 8; 13; 21; 34; 55; : : : : Fibonacci numbers are related to the golden ratio are the two roots of the equation x2 D x C 1 and to its conjugate y, which (3.23) and are given by the following formulas (see Exercise 3.2-6): p 1C 5 D 2 D 1:618p03 : : : ; yD 1 5 2 D :61803 : : : : (3.24) Specifically, we have i yi Fi D p ; 5 which we can prove by induction (Exercise 3.2-7). Since ˇˇ yˇˇ < 1, we have ˇˇ yi ˇˇ 1 p

    0, be a degree-d polynomial in n, and let k be a constant. Use the definitions of the asymptotic notations to prove the following properties. a. If k d , then p.n/ D O.nk/. b. If k Ä d , then p.n/ D .nk/. c. If k D d , then p.n/ D ‚.nk/. d. If k > d , then p.n/ D o.nk/. e. If k < d , then p.n/ D !.nk/. 3-2 Relative asymptotic growths Indicate, for each pair of expressions .A; B/ in the table below, whether A is O, o, , !, or ‚ of B. Assume that k 1, > 0, and c > 1 are constants. Your answer should be in the form of the table with “yes” or “no” written in each box. A B Oo a. lgk n n b. nk p c. n cn nsin n d. 2n 2n=2 e. nlg c clg n f. lg.nŠ/ lg.nn/ ! ‚ 3-3 Ordering by asymptotic growth rates a. Rank the following functions by order of growth; that is, find an arrangement g1; g2; : : : ; g30 of the functions satisfying g1 D .g2/, g2 D .g3/, . . . , g29 D .g30/. Partition your list into equivalence classes such that functions f .n/ and g.n/ are in the same class if and only if f .n/ D ‚.g.n//. 62 Chapter 3 Growth of Functions lg.lg n/ . 3 2 /n ln ln n 2lg n lg .lg n/ 2lg n n3 lg n .lg n/lg n p 2 2 lg n p . 2/lg n lg2 n n 2n en n n2 lg.nŠ/ nlg lg n 4lg n 2n nŠ 22n ln n .n C 1/Š n lg n .lg n/Š n1= lg n 1 p lg n 22nC1 b. Give an example of a single nonnegative function f .n/ such that for all functions gi .n/ in part (a), f .n/ is neither O.gi .n// nor .gi .n//. 3-4 Asymptotic notation properties Let f .n/ and g.n/ be asymptotically positive functions. Prove or disprove each of the following conjectures. a. f .n/ D O.g.n// implies g.n/ D O.f .n//. b. f .n/ C g.n/ D ‚.min.f .n/; g.n///. c. f .n/ D O.g.n// implies lg.f .n// D O.lg.g.n///, where lg.g.n// 1 and f .n/ 1 for all sufficiently large n. d. f .n/ D O.g.n// implies 2f .n/ D O 2g.n/ . e. f .n/ D O ..f .n//2/. f. f .n/ D O.g.n// implies g.n/ D .f .n//. g. f .n/ D ‚.f .n=2//. h. f .n/ C o.f .n// D ‚.f .n//. 3-5 Variations on O and ˝ 1 Some authors define in a slightly different way than we do; let’s use (read 1 “omega infinity”) for this alternative definition. We say that f .n/ D .g.n// if there exists a positive constant c such that f .n/ cg.n/ 0 for infinitely many integers n. a. Show that for any two functions f .n/ and g.n/ that are asymptotically nonneg1 ative, either f .n/ D O.g.n// or f .n/ D .g.n// or both, whereas this is not 1 true if we use in place of . Problems for Chapter 3 63 1 b. Describe the potential advantages and disadvantages of using instead of to characterize the running times of programs. Some authors also define O in a slightly different manner; let’s use O0 for the alternative definition. We say that f .n/ D O0.g.n// if and only if jf .n/j D O.g.n//. c. What happens to each direction of the “if and only if” in Theorem 3.1 if we substitute O0 for O but still use ? Some authors define Oe (read “soft-oh”) to mean O with logarithmic factors ignored: Oe.g.n// D ff .n/ W there exist positive constants c, k, and n0 such that 0 Ä f .n/ Ä cg.n/ lgk.n/ for all n n0g : d. Define e and ‚e in a similar manner. Prove the corresponding analog to Theorem 3.1. 3-6 Iterated functions We can apply the iteration operator used in the lg function to any monotonically increasing function f .n/ over the reals. For a given constant c 2 R, we define the iterated function fc by ˚ fc .n/ D min i 0 W f .i /.n/ Ä « c ; which need not be well defined in all cases. In other words, the quantity fc .n/ is the number of iterated applications of the function f required to reduce its argument down to c or less. For each of the following functions f .n/ and constants c, give as tight a bound as possible on fc .n/. f .n/ c a. n 1 0 fc .n/ b. lg n 1 c. n=2 1 d. n=2 2 p e. n 2 p f. n 1 g. n1=3 2 h. n= lg n 2 64 Chapter 3 Growth of Functions Chapter notes Knuth [209] traces the origin of the O-notation to a number-theory text by P. Bachmann in 1892. The o-notation was invented by E. Landau in 1909 for his discussion of the distribution of prime numbers. The and ‚ notations were advocated by Knuth [213] to correct the popular, but technically sloppy, practice in the literature of using O-notation for both upper and lower bounds. Many people continue to use the O-notation where the ‚-notation is more technically precise. Further discussion of the history and development of asymptotic notations appears in works by Knuth [209, 213] and Brassard and Bratley [54]. Not all authors define the asymptotic notations in the same way, although the various definitions agree in most common situations. Some of the alternative definitions encompass functions that are not asymptotically nonnegative, as long as their absolute values are appropriately bounded. Equation (3.20) is due to Robbins [297]. Other properties of elementary mathematical functions can be found in any good mathematical reference, such as Abramowitz and Stegun [1] or Zwillinger [362], or in a calculus book, such as Apostol [18] or Thomas et al. [334]. Knuth [209] and Graham, Knuth, and Patashnik [152] contain a wealth of material on discrete mathematics as used in computer science. 4 Divide-and-Conquer In Section 2.3.1, we saw how merge sort serves as an example of the divide-andconquer paradigm. Recall that in divide-and-conquer, we solve a problem recursively, applying three steps at each level of the recursion: Divide the problem into a number of subproblems that are smaller instances of the same problem. Conquer the subproblems by solving them recursively. If the subproblem sizes are small enough, however, just solve the subproblems in a straightforward manner. Combine the solutions to the subproblems into the solution for the original problem. When the subproblems are large enough to solve recursively, we call that the recursive case. Once the subproblems become small enough that we no longer recurse, we say that the recursion “bottoms out” and that we have gotten down to the base case. Sometimes, in addition to subproblems that are smaller instances of the same problem, we have to solve subproblems that are not quite the same as the original problem. We consider solving such subproblems as part of the combine step. In this chapter, we shall see more algorithms based on divide-and-conquer. The first one solves the maximum-subarray problem: it takes as input an array of numbers, and it determines the contiguous subarray whose values have the greatest sum. Then we shall see two divide-and-conquer algorithms for multiplying n n matrices. One runs in ‚.n3/ time, which is no better than the straightforward method of multiplying square matrices. But the other, Strassen’s algorithm, runs in O.n2:81/ time, which beats the straightforward method asymptotically. Recurrences Recurrences go hand in hand with the divide-and-conquer paradigm, because they give us a natural way to characterize the running times of divide-and-conquer algorithms. A recurrence is an equation or inequality that describes a function in terms 66 Chapter 4 Divide-and-Conquer of its value on smaller inputs. For example, in Section 2.3.2 we described the worst-case running time T .n/ of the MERGE-SORT procedure by the recurrence ( T .n/ D ‚.1/ if n D 1 ; (4.1) 2T .n=2/ C ‚.n/ if n > 1 ; whose solution we claimed to be T .n/ D ‚.n lg n/. Recurrences can take many forms. For example, a recursive algorithm might divide subproblems into unequal sizes, such as a 2=3-to-1=3 split. If the divide and combine steps take linear time, such an algorithm would give rise to the recurrence T .n/ D T .2n=3/ C T .n=3/ C ‚.n/. Subproblems are not necessarily constrained to being a constant fraction of the original problem size. For example, a recursive version of linear search (see Exercise 2.1-3) would create just one subproblem containing only one element fewer than the original problem. Each recursive call would take constant time plus the time for the recursive calls it makes, yielding the recurrence T .n/ D T .n 1/ C ‚.1/. This chapter offers three methods for solving recurrences—that is, for obtaining asymptotic “‚” or “O” bounds on the solution: In the substitution method, we guess a bound and then use mathematical induction to prove our guess correct. The recursion-tree method converts the recurrence into a tree whose nodes represent the costs incurred at various levels of the recursion. We use techniques for bounding summations to solve the recurrence. The master method provides bounds for recurrences of the form T .n/ D aT .n=b/ C f .n/ ; (4.2) where a 1, b > 1, and f .n/ is a given function. Such recurrences arise frequently. A recurrence of the form in equation (4.2) characterizes a divideand-conquer algorithm that creates a subproblems, each of which is 1=b the size of the original problem, and in which the divide and combine steps together take f .n/ time. To use the master method, you will need to memorize three cases, but once you do that, you will easily be able to determine asymptotic bounds for many simple recurrences. We will use the master method to determine the running times of the divide-and-conquer algorithms for the maximum-subarray problem and for matrix multiplication, as well as for other algorithms based on divideand-conquer elsewhere in this book. Chapter 4 Divide-and-Conquer 67 Occasionally, we shall see recurrences that are not equalities but rather inequalities, such as T .n/ Ä 2T .n=2/ C ‚.n/. Because such a recurrence states only an upper bound on T .n/, we will couch its solution using O-notation rather than ‚-notation. Similarly, if the inequality were reversed to T .n/ 2T .n=2/ C ‚.n/, then because the recurrence gives only a lower bound on T .n/, we would use -notation in its solution. Technicalities in recurrences In practice, we neglect certain technical details when we state and solve recur- rences. For example, if we call MERGE-SORT on n elements when n is odd, we end up with subproblems of size bn=2c and dn=2e. Neither size is actually n=2, because n=2 is not an integer when n is odd. Technically, the recurrence describing the worst-case running time of MERGE-SORT is really ( T .n/ D ‚.1/ if n D 1 ; (4.3) T .dn=2e/ C T .bn=2c/ C ‚.n/ if n > 1 : Boundary conditions represent another class of details that we typically ignore. Since the running time of an algorithm on a constant-sized input is a constant, the recurrences that arise from the running times of algorithms generally have T .n/ D ‚.1/ for sufficiently small n. Consequently, for convenience, we shall generally omit statements of the boundary conditions of recurrences and assume that T .n/ is constant for small n. For example, we normally state recurrence (4.1) as T .n/ D 2T .n=2/ C ‚.n/ ; (4.4) without explicitly giving values for small n. The reason is that although changing the value of T .1/ changes the exact solution to the recurrence, the solution typically doesn’t change by more than a constant factor, and so the order of growth is unchanged. When we state and solve recurrences, we often omit floors, ceilings, and boundary conditions. We forge ahead without these details and later determine whether or not they matter. They usually do not, but you should know when they do. Experience helps, and so do some theorems stating that these details do not affect the asymptotic bounds of many recurrences characterizing divide-and-conquer algorithms (see Theorem 4.1). In this chapter, however, we shall address some of these details and illustrate the fine points of recurrence solution methods. 68 Chapter 4 Divide-and-Conquer 4.1 The maximum-subarray problem Suppose that you been offered the opportunity to invest in the Volatile Chemical Corporation. Like the chemicals the company produces, the stock price of the Volatile Chemical Corporation is rather volatile. You are allowed to buy one unit of stock only one time and then sell it at a later date, buying and selling after the close of trading for the day. To compensate for this restriction, you are allowed to learn what the price of the stock will be in the future. Your goal is to maximize your profit. Figure 4.1 shows the price of the stock over a 17-day period. You may buy the stock at any one time, starting after day 0, when the price is $100 per share. Of course, you would want to “buy low, sell high”—buy at the lowest possible price and later on sell at the highest possible price—to maximize your profit. Unfortunately, you might not be able to buy at the lowest price and then sell at the highest price within a given period. In Figure 4.1, the lowest price occurs after day 7, which occurs after the highest price, after day 1. You might think that you can always maximize profit by either buying at the lowest price or selling at the highest price. For example, in Figure 4.1, we would maximize profit by buying at the lowest price, after day 7. If this strategy always worked, then it would be easy to determine how to maximize profit: find the highest and lowest prices, and then work left from the highest price to find the lowest prior price, work right from the lowest price to find the highest later price, and take the pair with the greater difference. Figure 4.2 shows a simple counterexample, 120 110 100 90 80 70 60 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Day 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Price 100 113 110 85 105 102 86 63 81 101 94 106 101 79 94 90 97 Change 13 3 25 20 3 16 23 18 20 7 12 5 22 15 4 7 Figure 4.1 Information about the price of stock in the Volatile Chemical Corporation after the close of trading over a period of 17 days. The horizontal axis of the chart indicates the day, and the vertical axis shows the price. The bottom row of the table gives the change in price from the previous day. 4.1 The maximum-subarray problem 69 11 10 9 8 7 6 01234 Day 01 23 4 Price 10 11 7 10 6 Change 1 43 4 Figure 4.2 An example showing that the maximum profit does not always start at the lowest price or end at the highest price. Again, the horizontal axis indicates the day, and the vertical axis shows the price. Here, the maximum profit of $3 per share would be earned by buying after day 2 and selling after day 3. The price of $7 after day 2 is not the lowest price overall, and the price of $10 after day 3 is not the highest price overall. demonstrating that the maximum profit sometimes comes neither by buying at the lowest price nor by selling at the highest price. A brute-force solution We can easily devise a brute-force solution to this problem: just try every possible pair of buy and sell dates in which the buy date precedes the sell date. A period of n days has n 2 such pairs of dates. Since n 2 is ‚.n2/, and the best we can hope for is to evaluate each pair of dates in constant time, this approach would take .n2/ time. Can we do better? A transformation In order to design an algorithm with an o.n2/ running time, we will look at the input in a slightly different way. We want to find a sequence of days over which the net change from the first day to the last is maximum. Instead of looking at the daily prices, let us instead consider the daily change in price, where the change on day i is the difference between the prices after day i 1 and after day i. The table in Figure 4.1 shows these daily changes in the bottom row. If we treat this row as an array A, shown in Figure 4.3, we now want to find the nonempty, contiguous subarray of A whose values have the largest sum. We call this contiguous subarray the maximum subarray. For example, in the array of Figure 4.3, the maximum subarray of AŒ1 : : 16 is AŒ8 : : 11, with the sum 43. Thus, you would want to buy the stock just before day 8 (that is, after day 7) and sell it after day 11, earning a profit of $43 per share. At first glance, this transformation does not help. We still need to check n1 2 D ‚.n2/ subarrays for a period of n days. Exercise 4.1-2 asks you to show 70 Chapter 4 Divide-and-Conquer 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 A 13 –3 –25 20 –3 –16 –23 18 20 –7 12 –5 –22 15 –4 7 maximum subarray Figure 4.3 The change in stock prices as a maximum-subarray problem. Here, the subarray AŒ8 : : 11, with sum 43, has the greatest sum of any contiguous subarray of array A. that although computing the cost of one subarray might take time proportional to the length of the subarray, when computing all ‚.n2/ subarray sums, we can organize the computation so that each subarray sum takes O.1/ time, given the values of previously computed subarray sums, so that the brute-force solution takes ‚.n2/ time. So let us seek a more efficient solution to the maximum-subarray problem. When doing so, we will usually speak of “a” maximum subarray rather than “the” maximum subarray, since there could be more than one subarray that achieves the maximum sum. The maximum-subarray problem is interesting only when the array contains some negative numbers. If all the array entries were nonnegative, then the maximum-subarray problem would present no challenge, since the entire array would give the greatest sum. A solution using divide-and-conquer Let’s think about how we might solve the maximum-subarray problem using the divide-and-conquer technique. Suppose we want to find a maximum subarray of the subarray AŒlow : : high. Divide-and-conquer suggests that we divide the subarray into two subarrays of as equal size as possible. That is, we find the midpoint, say mid, of the subarray, and consider the subarrays AŒlow : : mid and AŒmid C 1 : : high. As Figure 4.4(a) shows, any contiguous subarray AŒi : : j  of AŒlow : : high must lie in exactly one of the following places: entirely in the subarray AŒlow : : mid, so that low Ä i Ä j Ä mid, entirely in the subarray AŒmid C 1 : : high, so that mid < i Ä j Ä high, or crossing the midpoint, so that low Ä i Ä mid < j Ä high. Therefore, a maximum subarray of AŒlow : : high must lie in exactly one of these places. In fact, a maximum subarray of AŒlow : : high must have the greatest sum over all subarrays entirely in AŒlow : : mid, entirely in AŒmid C 1 : : high, or crossing the midpoint. We can find maximum subarrays of AŒlow : : mid and AŒmidC1 : : high recursively, because these two subproblems are smaller instances of the problem of finding a maximum subarray. Thus, all that is left to do is find a 4.1 The maximum-subarray problem 71 crosses the midpoint low mid high low AŒmid C 1 : : j  i mid high mid C 1 entirely in AŒlow : : mid entirely in AŒmid C 1 : : high (a) mid C 1 j AŒi : : mid (b) Figure 4.4 (a) Possible locations of subarrays of AŒlow : : high: entirely in AŒlow : : mid, entirely in AŒmid C 1 : : high, or crossing the midpoint mid. (b) Any subarray of AŒlow : : high crossing the midpoint comprises two subarrays AŒi : : mid and AŒmid C 1 : : j , where low Ä i Ä mid and mid < j Ä high. maximum subarray that crosses the midpoint, and take a subarray with the largest sum of the three. We can easily find a maximum subarray crossing the midpoint in time linear in the size of the subarray AŒlow : : high. This problem is not a smaller instance of our original problem, because it has the added restriction that the subarray it chooses must cross the midpoint. As Figure 4.4(b) shows, any subarray crossing the midpoint is itself made of two subarrays AŒi : : mid and AŒmid C 1 : : j , where low Ä i Ä mid and mid < j Ä high. Therefore, we just need to find maximum subarrays of the form AŒi : : mid and AŒmid C 1 : : j  and then combine them. The procedure FIND-MAX-CROSSING-SUBARRAY takes as input the array A and the indices low, mid, and high, and it returns a tuple containing the indices demarcating a maximum subarray that crosses the midpoint, along with the sum of the values in a maximum subarray. FIND-MAX-CROSSING-SUBARRAY.A; low; mid; high/ 1 left-sum D 1 2 sum D 0 3 for i D mid downto low 4 sum D sum C AŒi 5 if sum > left-sum 6 left-sum D sum 7 max-left D i 8 right-sum D 1 9 sum D 0 10 for j D mid C 1 to high 11 sum D sum C AŒj  12 if sum > right-sum 13 right-sum D sum 14 max-right D j 15 return .max-left; max-right; left-sum C right-sum/ 72 Chapter 4 Divide-and-Conquer This procedure works as follows. Lines 1–7 find a maximum subarray of the left half, AŒlow : : mid. Since this subarray must contain AŒmid, the for loop of lines 3–7 starts the index i at mid and works down to low, so that every subarray it considers is of the form AŒi : : mid. Lines 1–2 initialize the variables left-sum, which holds the greatest sum found so far, and sum, holding the sum of the entries in AŒi : : mid. Whenever we find, in line 5, a subarray AŒi : : mid with a sum of values greater than left-sum, we update left-sum to this subarray’s sum in line 6, and in line 7 we update the variable max-left to record this index i. Lines 8–14 work analogously for the right half, AŒmid C 1 : : high. Here, the for loop of lines 10–14 starts the index j at midC1 and works up to high, so that every subarray it considers is of the form AŒmid C 1 : : j . Finally, line 15 returns the indices max-left and max-right that demarcate a maximum subarray crossing the midpoint, along with the sum left-sum Cright-sum of the values in the subarray AŒmax-left : : max-right. If the subarray AŒlow : : high contains n entries (so that n D high low C 1), we claim that the call FIND-MAX-CROSSING-SUBARRAY.A; low; mid; high/ takes ‚.n/ time. Since each iteration of each of the two for loops takes ‚.1/ time, we just need to count up how many iterations there are altogether. The for loop of lines 3–7 makes mid low C 1 iterations, and the for loop of lines 10–14 makes high mid iterations, and so the total number of iterations is .mid low C 1/ C .high mid/ D high low C 1 D n: With a linear-time FIND-MAX-CROSSING-SUBARRAY procedure in hand, we can write pseudocode for a divide-and-conquer algorithm to solve the maximumsubarray problem: FIND-MAXIMUM-SUBARRAY.A; low; high/ 1 if high == low 2 return .low; high; AŒlow/ // base case: only one element 3 else mid D b.low C high/=2c 4 .left-low; left-high; left-sum/ D FIND-MAXIMUM-SUBARRAY.A; low; mid/ 5 .right-low; right-high; right-sum/ D FIND-MAXIMUM-SUBARRAY.A; mid C 1; high/ 6 .cross-low; cross-high; cross-sum/ D FIND-MAX-CROSSING-SUBARRAY .A; low; mid; high/ 7 if left-sum right-sum and left-sum cross-sum 8 return .left-low; left-high; left-sum/ 9 elseif right-sum left-sum and right-sum cross-sum 10 return .right-low; right-high; right-sum/ 11 else return .cross-low; cross-high; cross-sum/ 4.1 The maximum-subarray problem 73 The initial call FIND-MAXIMUM-SUBARRAY.A; 1; A:length/ will find a maximum subarray of AŒ1 : : n. Similar to FIND-MAX-CROSSING-SUBARRAY, the recursive procedure FINDMAXIMUM-SUBARRAY returns a tuple containing the indices that demarcate a maximum subarray, along with the sum of the values in a maximum subarray. Line 1 tests for the base case, where the subarray has just one element. A subarray with just one element has only one subarray—itself—and so line 2 returns a tuple with the starting and ending indices of just the one element, along with its value. Lines 3–11 handle the recursive case. Line 3 does the divide part, computing the index mid of the midpoint. Let’s refer to the subarray AŒlow : : mid as the left subarray and to AŒmid C 1 : : high as the right subarray. Because we know that the subarray AŒlow : : high contains at least two elements, each of the left and right subarrays must have at least one element. Lines 4 and 5 conquer by recursively finding maximum subarrays within the left and right subarrays, respectively. Lines 6–11 form the combine part. Line 6 finds a maximum subarray that crosses the midpoint. (Recall that because line 6 solves a subproblem that is not a smaller instance of the original problem, we consider it to be in the combine part.) Line 7 tests whether the left subarray contains a subarray with the maximum sum, and line 8 returns that maximum subarray. Otherwise, line 9 tests whether the right subarray contains a subarray with the maximum sum, and line 10 returns that maximum subarray. If neither the left nor right subarrays contain a subarray achieving the maximum sum, then a maximum subarray must cross the midpoint, and line 11 returns it. Analyzing the divide-and-conquer algorithm Next we set up a recurrence that describes the running time of the recursive FINDMAXIMUM-SUBARRAY procedure. As we did when we analyzed merge sort in Section 2.3.2, we make the simplifying assumption that the original problem size is a power of 2, so that all subproblem sizes are integers. We denote by T .n/ the running time of FIND-MAXIMUM-SUBARRAY on a subarray of n elements. For starters, line 1 takes constant time. The base case, when n D 1, is easy: line 2 takes constant time, and so T .1/ D ‚.1/ : (4.5) The recursive case occurs when n > 1. Lines 1 and 3 take constant time. Each of the subproblems solved in lines 4 and 5 is on a subarray of n=2 elements (our assumption that the original problem size is a power of 2 ensures that n=2 is an integer), and so we spend T .n=2/ time solving each of them. Because we have to solve two subproblems—for the left subarray and for the right subarray—the contribution to the running time from lines 4 and 5 comes to 2T .n=2/. As we have 74 Chapter 4 Divide-and-Conquer already seen, the call to FIND-MAX-CROSSING-SUBARRAY in line 6 takes ‚.n/ time. Lines 7–11 take only ‚.1/ time. For the recursive case, therefore, we have T .n/ D ‚.1/ C 2T .n=2/ C ‚.n/ C ‚.1/ D 2T .n=2/ C ‚.n/ : (4.6) Combining equations (4.5) and (4.6) gives us a recurrence for the running time T .n/ of FIND-MAXIMUM-SUBARRAY: ( T .n/ D ‚.1/ if n D 1 ; 2T .n=2/ C ‚.n/ if n > 1 : (4.7) This recurrence is the same as recurrence (4.1) for merge sort. As we shall see from the master method in Section 4.5, this recurrence has the solution T .n/ D ‚.n lg n/. You might also revisit the recursion tree in Figure 2.5 to understand why the solution should be T .n/ D ‚.n lg n/. Thus, we see that the divide-and-conquer method yields an algorithm that is asymptotically faster than the brute-force method. With merge sort and now the maximum-subarray problem, we begin to get an idea of how powerful the divideand-conquer method can be. Sometimes it will yield the asymptotically fastest algorithm for a problem, and other times we can do even better. As Exercise 4.1-5 shows, there is in fact a linear-time algorithm for the maximum-subarray problem, and it does not use divide-and-conquer. Exercises 4.1-1 What does FIND-MAXIMUM-SUBARRAY return when all elements of A are negative? 4.1-2 Write pseudocode for the brute-force method of solving the maximum-subarray problem. Your procedure should run in ‚.n2/ time. 4.1-3 Implement both the brute-force and recursive algorithms for the maximumsubarray problem on your own computer. What problem size n0 gives the crossover point at which the recursive algorithm beats the brute-force algorithm? Then, change the base case of the recursive algorithm to use the brute-force algorithm whenever the problem size is less than n0. Does that change the crossover point? 4.1-4 Suppose we change the definition of the maximum-subarray problem to allow the result to be an empty subarray, where the sum of the values of an empty subar- 4.2 Strassen’s algorithm for matrix multiplication 75 ray is 0. How would you change any of the algorithms that do not allow empty subarrays to permit an empty subarray to be the result? 4.1-5 Use the following ideas to develop a nonrecursive, linear-time algorithm for the maximum-subarray problem. Start at the left end of the array, and progress toward the right, keeping track of the maximum subarray seen so far. Knowing a maximum subarray of AŒ1 : : j , extend the answer to find a maximum subarray ending at index j C1 by using the following observation: a maximum subarray of AŒ1 : : j C 1 is either a maximum subarray of AŒ1 : : j  or a subarray AŒi : : j C 1, for some 1 Ä i Ä j C 1. Determine a maximum subarray of the form AŒi : : j C 1 in constant time based on knowing a maximum subarray ending at index j . 4.2 Strassen’s algorithm for matrix multiplication If you have seen matrices before, then you probably know how to multiply them. (Otherwise, you should read Section D.1 in Appendix D.) If A D .aij / and B D .bij / are square n n matrices, then in the product C D A B, we define the entry cij , for i; j D 1; 2; : : : ; n, by Xn cij D aik bkj : kD1 (4.8) We must compute n2 matrix entries, and each is the sum of n values. The following procedure takes n n matrices A and B and multiplies them, returning their n n product C . We assume that each matrix has an attribute rows, giving the number of rows in the matrix. SQUARE-MATRIX-MULTIPLY.A; B/ 1 n D A:rows 2 let C be a new n n matrix 3 for i D 1 to n 4 for j D 1 to n 5 cij D 0 6 for k D 1 to n 7 cij D cij C aik bkj 8 return C The SQUARE-MATRIX-MULTIPLY procedure works as follows. The for loop of lines 3–7 computes the entries of each row i, and within a given row i, the 76 Chapter 4 Divide-and-Conquer for loop of lines 4–7 computes each of the entries cij , for each column j . Line 5 initializes cij to 0 as we start computing the sum given in equation (4.8), and each iteration of the for loop of lines 6–7 adds in one more term of equation (4.8). Because each of the triply-nested for loops runs exactly n iterations, and each execution of line 7 takes constant time, the SQUARE-MATRIX-MULTIPLY procedure takes ‚.n3/ time. You might at first think that any matrix multiplication algorithm must take .n3/ time, since the natural definition of matrix multiplication requires that many mul- tiplications. You would be incorrect, however: we have a way to multiply matrices in o.n3/ time. In this section, we shall see Strassen’s remarkable recursive algorithm for multiplying n n matrices. It runs in ‚.nlg 7/ time, which we shall show in Section 4.5. Since lg 7 lies between 2:80 and 2:81, Strassen’s algorithm runs in O.n2:81/ time, which is asymptotically better than the simple SQUARE-MATRIX- MULTIPLY procedure. A simple divide-and-conquer algorithm To keep things simple, when we use a divide-and-conquer algorithm to compute the matrix product C D A B, we assume that n is an exact power of 2 in each of the n n matrices. We make this assumption because in each divide step, we will divide n n matrices into four n=2 n=2 matrices, and by assuming that n is an exact power of 2, we are guaranteed that as long as n 2, the dimension n=2 is an integer. Suppose that we partition each of A, B, and C into four n=2 n=2 matrices  à AD A11 A12 A21 A22 ;  à BD B11 B12 B21 B22 ;  à CD C11 C12 C21 C22 ; (4.9) so that we rewrite the equation C D A B as  C11 C21 àC12 C22 D A11 A21 àA12 B11 A22 B21 à B12 B22 : (4.10) Equation (4.10) corresponds to the four equations C11 D A11 B11 C A12 B21 ; C12 D A11 B12 C A12 B22 ; C21 D A21 B11 C A22 B21 ; C22 D A21 B12 C A22 B22 : (4.11) (4.12) (4.13) (4.14) Each of these four equations specifies two multiplications of n=2 n=2 matrices and the addition of their n=2 n=2 products. We can use these equations to create a straightforward, recursive, divide-and-conquer algorithm: 4.2 Strassen’s algorithm for matrix multiplication 77 SQUARE-MATRIX-MULTIPLY-RECURSIVE .A; B/ 1 n D A:rows 2 let C be a new n n matrix 3 if n == 1 4 c11 D a11 b11 5 else partition A, B, and C as in equations (4.9) 6 C11 D SQUARE-MATRIX-MULTIPLY-RECURSIVE .A11; B11/ C SQUARE-MATRIX-MULTIPLY-RECURSIVE.A12; B21/ 7 C12 D SQUARE-MATRIX-MULTIPLY-RECURSIVE .A11; B12/ C SQUARE-MATRIX-MULTIPLY-RECURSIVE.A12; B22/ 8 C21 D SQUARE-MATRIX-MULTIPLY-RECURSIVE .A21; B11/ C SQUARE-MATRIX-MULTIPLY-RECURSIVE.A22; B21/ 9 C22 D SQUARE-MATRIX-MULTIPLY-RECURSIVE .A21; B12/ C SQUARE-MATRIX-MULTIPLY-RECURSIVE.A22; B22/ 10 return C This pseudocode glosses over one subtle but important implementation detail. How do we partition the matrices in line 5? If we were to create 12 new n=2 n=2 matrices, we would spend ‚.n2/ time copying entries. In fact, we can partition the matrices without copying entries. The trick is to use index calculations. We identify a submatrix by a range of row indices and a range of column indices of the original matrix. We end up representing a submatrix a little differently from how we represent the original matrix, which is the subtlety we are glossing over. The advantage is that, since we can specify submatrices by index calculations, executing line 5 takes only ‚.1/ time (although we shall see that it makes no difference asymptotically to the overall running time whether we copy or partition in place). Now, we derive a recurrence to characterize the running time of SQUAREMATRIX-MULTIPLY-RECURSIVE. Let T .n/ be the time to multiply two n n matrices using this procedure. In the base case, when n D 1, we perform just the one scalar multiplication in line 4, and so T .1/ D ‚.1/ : (4.15) The recursive case occurs when n > 1. As discussed, partitioning the matrices in line 5 takes ‚.1/ time, using index calculations. In lines 6–9, we recursively call SQUARE-MATRIX-MULTIPLY-RECURSIVE a total of eight times. Because each recursive call multiplies two n=2 n=2 matrices, thereby contributing T .n=2/ to the overall running time, the time taken by all eight recursive calls is 8T .n=2/. We also must account for the four matrix additions in lines 6–9. Each of these matrices contains n2=4 entries, and so each of the four matrix additions takes ‚.n2/ time. Since the number of matrix additions is a constant, the total time spent adding ma- 78 Chapter 4 Divide-and-Conquer trices in lines 6–9 is ‚.n2/. (Again, we use index calculations to place the results of the matrix additions into the correct positions of matrix C , with an overhead of ‚.1/ time per entry.) The total time for the recursive case, therefore, is the sum of the partitioning time, the time for all the recursive calls, and the time to add the matrices resulting from the recursive calls: T .n/ D ‚.1/ C 8T .n=2/ C ‚.n2/ D 8T .n=2/ C ‚.n2/ : (4.16) Notice that if we implemented partitioning by copying matrices, which would cost ‚.n2/ time, the recurrence would not change, and hence the overall running time would increase by only a constant factor. Combining equations (4.15) and (4.16) gives us the recurrence for the running time of SQUARE-MATRIX-MULTIPLY-RECURSIVE: ( ‚.1/ if n D 1 ; T .n/ D 8T .n=2/ C ‚.n2/ if n > 1 : (4.17) As we shall see from the master method in Section 4.5, recurrence (4.17) has the solution T .n/ D ‚.n3/. Thus, this simple divide-and-conquer approach is no faster than the straightforward SQUARE-MATRIX-MULTIPLY procedure. Before we continue on to examining Strassen’s algorithm, let us review where the components of equation (4.16) came from. Partitioning each n n matrix by index calculation takes ‚.1/ time, but we have two matrices to partition. Although you could say that partitioning the two matrices takes ‚.2/ time, the constant of 2 is subsumed by the ‚-notation. Adding two matrices, each with, say, k entries, takes ‚.k/ time. Since the matrices we add each have n2=4 entries, you could say that adding each pair takes ‚.n2=4/ time. Again, however, the ‚-notation subsumes the constant factor of 1=4, and we say that adding two n2=4 n2=4 matrices takes ‚.n2/ time. We have four such matrix additions, and once again, instead of saying that they take ‚.4n2/ time, we say that they take ‚.n2/ time. (Of course, you might observe that we could say that the four matrix additions take ‚.4n2=4/ time, and that 4n2=4 D n2, but the point here is that ‚-notation subsumes constant factors, whatever they are.) Thus, we end up with two terms of ‚.n2/, which we can combine into one. When we account for the eight recursive calls, however, we cannot just subsume the constant factor of 8. In other words, we must say that together they take 8T .n=2/ time, rather than just T .n=2/ time. You can get a feel for why by looking back at the recursion tree in Figure 2.5, for recurrence (2.1) (which is identical to recurrence (4.7)), with the recursive case T .n/ D 2T .n=2/C‚.n/. The factor of 2 determined how many children each tree node had, which in turn determined how many terms contributed to the sum at each level of the tree. If we were to ignore 4.2 Strassen’s algorithm for matrix multiplication 79 the factor of 8 in equation (4.16) or the factor of 2 in recurrence (4.1), the recursion tree would just be linear, rather than “bushy,” and each level would contribute only one term to the sum. Bear in mind, therefore, that although asymptotic notation subsumes constant multiplicative factors, recursive notation such as T .n=2/ does not. Strassen’s method The key to Strassen’s method is to make the recursion tree slightly less bushy. That is, instead of performing eight recursive multiplications of n=2 n=2 matrices, it performs only seven. The cost of eliminating one matrix multiplication will be several new additions of n=2 n=2 matrices, but still only a constant number of additions. As before, the constant number of matrix additions will be subsumed by ‚-notation when we set up the recurrence equation to characterize the running time. Strassen’s method is not at all obvious. (This might be the biggest understatement in this book.) It has four steps: 1. Divide the input matrices A and B and output matrix C into n=2 n=2 submatrices, as in equation (4.9). This step takes ‚.1/ time by index calculation, just as in SQUARE-MATRIX-MULTIPLY-RECURSIVE. 2. Create 10 matrices S1; S2; : : : ; S10, each of which is n=2 n=2 and is the sum or difference of two matrices created in step 1. We can create all 10 matrices in ‚.n2/ time. 3. Using the submatrices created in step 1 and the 10 matrices created in step 2, recursively compute seven matrix products P1; P2; : : : ; P7. Each matrix Pi is n=2 n=2. 4. Compute the desired submatrices C11; C12; C21; C22 of the result matrix C by adding and subtracting various combinations of the Pi matrices. We can compute all four submatrices in ‚.n2/ time. We shall see the details of steps 2–4 in a moment, but we already have enough information to set up a recurrence for the running time of Strassen’s method. Let us assume that once the matrix size n gets down to 1, we perform a simple scalar mul- tiplication, just as in line 4 of SQUARE-MATRIX-MULTIPLY-RECURSIVE. When n > 1, steps 1, 2, and 4 take a total of ‚.n2/ time, and step 3 requires us to per- form seven multiplications of n=2 n=2 matrices. Hence, we obtain the following recurrence for the running time T .n/ of Strassen’s algorithm: ( T .n/ D ‚.1/ 7T .n=2/ C ‚.n2/ if n D 1 ; if n > 1 : (4.18) 80 Chapter 4 Divide-and-Conquer We have traded off one matrix multiplication for a constant number of matrix additions. Once we understand recurrences and their solutions, we shall see that this tradeoff actually leads to a lower asymptotic running time. By the master method in Section 4.5, recurrence (4.18) has the solution T .n/ D ‚.nlg 7/. We now proceed to describe the details. In step 2, we create the following 10 matrices: S1 D B12 B22 ; S2 D A11 C A12 ; S3 D A21 C A22 ; S4 D B21 B11 ; S5 D A11 C A22 ; S6 D B11 C B22 ; S7 D A12 A22 ; S8 D B21 C B22 ; S9 D A11 A21 ; S10 D B11 C B12 : Since we must add or subtract n=2 n=2 matrices 10 times, this step does indeed take ‚.n2/ time. In step 3, we recursively multiply n=2 n=2 matrices seven times to compute the following n=2 n=2 matrices, each of which is the sum or difference of products of A and B submatrices: P1 D A11 S1 D A11 B12 A11 B22 ; P2 D S2 B22 D A11 B22 C A12 B22 ; P3 D S3 B11 D A21 B11 C A22 B11 ; P4 D A22 S4 D A22 B21 A22 B11 ; P5 D S5 S6 D A11 B11 C A11 B22 C A22 B11 C A22 B22 ; P6 D S7 S8 D A12 B21 C A12 B22 A22 B21 A22 B22 ; P7 D S9 S10 D A11 B11 C A11 B12 A21 B11 A21 B12 : Note that the only multiplications we need to perform are those in the middle column of the above equations. The right-hand column just shows what these products equal in terms of the original submatrices created in step 1. Step 4 adds and subtracts the Pi matrices created in step 3 to construct the four n=2 n=2 submatrices of the product C . We start with C11 D P5 C P4 P2 C P6 : 4.2 Strassen’s algorithm for matrix multiplication 81 Expanding out the right-hand side, with the expansion of each Pi on its own line and vertically aligning terms that cancel out, we see that C11 equals A11 B11 C A11 B22 C A22 B11 C A22 B22 A22 B11 C A22 B21 A11 B22 A12 B22 A22 B22 A22 B21 C A12 B22 C A12 B21 A11 B11 which corresponds to equation (4.11). Similarly, we set C12 D P1 C P2 ; and so C12 equals A11 B12 A11 B22 C A11 B22 C A12 B22 C A12 B21 ; A11 B12 C A12 B22 ; corresponding to equation (4.12). Setting C21 D P3 C P4 makes C21 equal A21 B11 C A22 B11 A22 B11 C A22 B21 A21 B11 C A22 B21 ; corresponding to equation (4.13). Finally, we set C22 D P5 C P1 P3 P7 ; so that C22 equals A11 B11 C A11 B22 C A22 B11 C A22 B22 A11 B22 C A11 B12 A22 B11 A21 B11 A11 B11 A11 B12 C A21 B11 C A21 B12 A22 B22 C A21 B12 ; 82 Chapter 4 Divide-and-Conquer which corresponds to equation (4.14). Altogether, we add or subtract n=2 n=2 matrices eight times in step 4, and so this step indeed takes ‚.n2/ time. Thus, we see that Strassen’s algorithm, comprising steps 1–4, produces the correct matrix product and that recurrence (4.18) characterizes its running time. Since we shall see in Section 4.5 that this recurrence has the solution T .n/ D ‚.nlg7/, Strassen’s method is asymptotically faster than the straightforward SQUAREMATRIX-MULTIPLY procedure. The notes at the end of this chapter discuss some of the practical aspects of Strassen’s algorithm. Exercises Note: Although Exercises 4.2-3, 4.2-4, and 4.2-5 are about variants on Strassen’s algorithm, you should read Section 4.5 before trying to solve them. 4.2-1 Use Strassen’s algorithm to compute the matrix product  Ã à 13 75 68 42 : Show your work. 4.2-2 Write pseudocode for Strassen’s algorithm. 4.2-3 How would you modify Strassen’s algorithm to multiply n n matrices in which n is not an exact power of 2? Show that the resulting algorithm runs in time ‚.nlg7/. 4.2-4 What is the largest k such that if you can multiply 3 3 matrices using k multiplications (not assuming commutativity of multiplication), then you can multiply n n matrices in time o.nlg 7/? What would the running time of this algorithm be? 4.2-5 V. Pan has discovered a way of multiplying 68 68 matrices using 132,464 multiplications, a way of multiplying 70 70 matrices using 143,640 multiplications, and a way of multiplying 72 72 matrices using 155,424 multiplications. Which method yields the best asymptotic running time when used in a divide-and-conquer matrix-multiplication algorithm? How does it compare to Strassen’s algorithm? 4.3 The substitution method for solving recurrences 83 4.2-6 How quickly can you multiply a k n n matrix by an n k n matrix, using Strassen’s algorithm as a subroutine? Answer the same question with the order of the input matrices reversed. 4.2-7 Show how to multiply the complex numbers a C bi and c C d i using only three multiplications of real numbers. The algorithm should take a, b, c, and d as input and produce the real component ac bd and the imaginary component ad C bc separately. 4.3 The substitution method for solving recurrences Now that we have seen how recurrences characterize the running times of divideand-conquer algorithms, we will learn how to solve recurrences. We start in this section with the “substitution” method. The substitution method for solving recurrences comprises two steps: 1. Guess the form of the solution. 2. Use mathematical induction to find the constants and show that the solution works. We substitute the guessed solution for the function when applying the inductive hypothesis to smaller values; hence the name “substitution method.” This method is powerful, but we must be able to guess the form of the answer in order to apply it. We can use the substitution method to establish either upper or lower bounds on a recurrence. As an example, let us determine an upper bound on the recurrence T .n/ D 2T .bn=2c/ C n ; (4.19) which is similar to recurrences (4.3) and (4.4). We guess that the solution is T .n/ D O.n lg n/. The substitution method requires us to prove that T .n/ Ä cn lg n for an appropriate choice of the constant c > 0. We start by assuming that this bound holds for all positive m < n, in particular for m D bn=2c, yielding T .bn=2c/ Ä c bn=2c lg.bn=2c/. Substituting into the recurrence yields T .n/ Ä 2.c bn=2c lg.bn=2c// C n Ä cn lg.n=2/ C n D cn lg n cn lg 2 C n D cn lg n cn C n Ä cn lg n ; 84 Chapter 4 Divide-and-Conquer where the last step holds as long as c 1. Mathematical induction now requires us to show that our solution holds for the boundary conditions. Typically, we do so by showing that the boundary conditions are suitable as base cases for the inductive proof. For the recurrence (4.19), we must show that we can choose the constant c large enough so that the bound T .n/ Ä cn lg n works for the boundary conditions as well. This requirement can sometimes lead to problems. Let us assume, for the sake of argument, that T .1/ D 1 is the sole boundary condition of the recurrence. Then for n D 1, the bound T .n/ Ä cn lg n yields T .1/ Ä c1 lg 1 D 0, which is at odds with T .1/ D 1. Consequently, the base case of our inductive proof fails to hold. We can overcome this obstacle in proving an inductive hypothesis for a specific boundary condition with only a little more effort. In the recurrence (4.19), for example, we take advantage of asymptotic notation requiring us only to prove T .n/ Ä cn lg n for n n0, where n0 is a constant that we get to choose. We keep the troublesome boundary condition T .1/ D 1, but remove it from consideration in the inductive proof. We do so by first observing that for n > 3, the recurrence does not depend directly on T .1/. Thus, we can replace T .1/ by T .2/ and T .3/ as the base cases in the inductive proof, letting n0 D 2. Note that we make a distinction between the base case of the recurrence (n D 1) and the base cases of the inductive proof (n D 2 and n D 3). With T .1/ D 1, we derive from the recurrence that T .2/ D 4 and T .3/ D 5. Now we can complete the inductive proof that T .n/ Ä cn lg n for some constant c 1 by choosing c large enough so that T .2/ Ä c2 lg 2 and T .3/ Ä c3 lg 3. As it turns out, any choice of c 2 suffices for the base cases of n D 2 and n D 3 to hold. For most of the recurrences we shall examine, it is straightforward to extend boundary conditions to make the inductive assumption work for small n, and we shall not always explicitly work out the details. Making a good guess Unfortunately, there is no general way to guess the correct solutions to recurrences. Guessing a solution takes experience and, occasionally, creativity. Fortunately, though, you can use some heuristics to help you become a good guesser. You can also use recursion trees, which we shall see in Section 4.4, to generate good guesses. If a recurrence is similar to one you have seen before, then guessing a similar solution is reasonable. As an example, consider the recurrence T .n/ D 2T .bn=2c C 17/ C n ; which looks difficult because of the added “17” in the argument to T on the righthand side. Intuitively, however, this additional term cannot substantially affect the 4.3 The substitution method for solving recurrences 85 solution to the recurrence. When n is large, the difference between bn=2c and bn=2c C 17 is not that large: both cut n nearly evenly in half. Consequently, we make the guess that T .n/ D O.n lg n/, which you can verify as correct by using the substitution method (see Exercise 4.3-6). Another way to make a good guess is to prove loose upper and lower bounds on the recurrence and then reduce the range of uncertainty. For example, we might start with a lower bound of T .n/ D .n/ for the recurrence (4.19), since we have the term n in the recurrence, and we can prove an initial upper bound of T .n/ D O.n2/. Then, we can gradually lower the upper bound and raise the lower bound until we converge on the correct, asymptotically tight solution of T .n/ D ‚.n lg n/. Subtleties Sometimes you might correctly guess an asymptotic bound on the solution of a recurrence, but somehow the math fails to work out in the induction. The problem frequently turns out to be that the inductive assumption is not strong enough to prove the detailed bound. If you revise the guess by subtracting a lower-order term when you hit such a snag, the math often goes through. Consider the recurrence T .n/ D T .bn=2c/ C T .dn=2e/ C 1 : We guess that the solution is T .n/ D O.n/, and we try to show that T .n/ Ä cn for an appropriate choice of the constant c. Substituting our guess in the recurrence, we obtain T .n/ Ä c bn=2c C c dn=2e C 1 D cn C 1 ; which does not imply T .n/ Ä cn for any choice of c. We might be tempted to try a larger guess, say T .n/ D O.n2/. Although we can make this larger guess work, our original guess of T .n/ D O.n/ is correct. In order to show that it is correct, however, we must make a stronger inductive hypothesis. Intuitively, our guess is nearly right: we are off only by the constant 1, a lower-order term. Nevertheless, mathematical induction does not work unless we prove the exact form of the inductive hypothesis. We overcome our difficulty by subtracting a lower-order term from our previous guess. Our new guess is T .n/ Ä cn d , where d 0 is a constant. We now have T .n/ Ä .c bn=2c d / C .c dn=2e D cn 2d C 1 Ä cn d ; d/ C1 86 Chapter 4 Divide-and-Conquer as long as d 1. As before, we must choose the constant c large enough to handle the boundary conditions. You might find the idea of subtracting a lower-order term counterintuitive. After all, if the math does not work out, we should increase our guess, right? Not necessarily! When proving an upper bound by induction, it may actually be more difficult to prove that a weaker upper bound holds, because in order to prove the weaker bound, we must use the same weaker bound inductively in the proof. In our current example, when the recurrence has more than one recursive term, we get to subtract out the lower-order term of the proposed bound once per recursive term. In the above example, we subtracted out the constant d twice, once for the T .bn=2c/ term and once for the T .dn=2e/ term. We ended up with the inequality T .n/ Ä cn 2d C 1, and it was easy to find values of d to make cn 2d C 1 be less than or equal to cn d . Avoiding pitfalls It is easy to err in the use of asymptotic notation. For example, in the recurrence (4.19) we can falsely “prove” T .n/ D O.n/ by guessing T .n/ Ä cn and then arguing T .n/ Ä 2.c bn=2c/ C n Ä cn C n D O.n/ ;  wrong!! since c is a constant. The error is that we have not proved the exact form of the inductive hypothesis, that is, that T .n/ Ä cn. We therefore will explicitly prove that T .n/ Ä cn when we want to show that T .n/ D O.n/. Changing variables Sometimes, a little algebraic manipulation can make an unknown recurrence similar to one you have seen before. As an example, consider the recurrence p˘ T .n/ D 2T n C lg n ; which looks difficult. We can simplify this recurrence, though, with a change of varpiables. For convenience, we shall not worry about rounding off values, such as n, to be integers. Renaming m D lg n yields T .2m/ D 2T .2m=2/ C m : We can now rename S.m/ D T .2m/ to produce the new recurrence S.m/ D 2S.m=2/ C m ; 4.3 The substitution method for solving recurrences 87 which is very much like recurrence (4.19). Indeed, this new recurrence has the same solution: S.m/ D O.m lg m/. Changing back from S.m/ to T .n/, we obtain T .n/ D T .2m/ D S.m/ D O.m lg m/ D O.lg n lg lg n/ : Exercises 4.3-1 Show that the solution of T .n/ D T .n 1/ C n is O.n2/. 4.3-2 Show that the solution of T .n/ D T .dn=2e/ C 1 is O.lg n/. 4.3-3 We saw that the solution of T .n/ D 2T .bn=2c/ C n is O.n lg n/. Show that the solution of this recurrence is also .n lg n/. Conclude that the solution is ‚.n lg n/. 4.3-4 Show that by making a different inductive hypothesis, we can overcome the difficulty with the boundary condition T .1/ D 1 for recurrence (4.19) without adjusting the boundary conditions for the inductive proof. 4.3-5 Show that ‚.n lg n/ is the solution to the “exact” recurrence (4.3) for merge sort. 4.3-6 Show that the solution to T .n/ D 2T .bn=2c C 17/ C n is O.n lg n/. 4.3-7 Using the master method in Section 4.5, you can show that the solution to the recurrence T .n/ D 4T .n=3/ C n is T .n/ D ‚.nlog3 4/. Show that a substitution proof with the assumption T .n/ Ä cnlog3 4 fails. Then show how to subtract off a lower-order term to make a substitution proof work. 4.3-8 Using the master method in Section 4.5, you can show that the solution to the recurrence T .n/ D 4T .n=2/ C n2 is T .n/ D ‚.n2/. Show that a substitution proof with the assumption T .n/ Ä cn2 fails. Then show how to subtract off a lower-order term to make a substitution proof work. 88 Chapter 4 Divide-and-Conquer 4.3-9 p Solve the recurrence T .n/ D 3T . n/ C log n by making a change of variables. Your solution should be asymptotically tight. Do not worry about whether values are integral. 4.4 The recursion-tree method for solving recurrences Although you can use the substitution method to provide a succinct proof that a solution to a recurrence is correct, you might have trouble coming up with a good guess. Drawing out a recursion tree, as we did in our analysis of the merge sort recurrence in Section 2.3.2, serves as a straightforward way to devise a good guess. In a recursion tree, each node represents the cost of a single subproblem somewhere in the set of recursive function invocations. We sum the costs within each level of the tree to obtain a set of per-level costs, and then we sum all the per-level costs to determine the total cost of all levels of the recursion. A recursion tree is best used to generate a good guess, which you can then verify by the substitution method. When using a recursion tree to generate a good guess, you can often tolerate a small amount of “sloppiness,” since you will be verifying your guess later on. If you are very careful when drawing out a recursion tree and summing the costs, however, you can use a recursion tree as a direct proof of a solution to a recurrence. In this section, we will use recursion trees to generate good guesses, and in Section 4.6, we will use recursion trees directly to prove the theorem that forms the basis of the master method. For example, let us see how a recursion tree would provide a good guess for the recurrence T .n/ D 3T .bn=4c/ C ‚.n2/. We start by focusing on finding an upper bound for the solution. Because we know that floors and ceilings usually do not matter when solving recurrences (here’s an example of sloppiness that we can tolerate), we create a recursion tree for the recurrence T .n/ D 3T .n=4/ C cn2, having written out the implied constant coefficient c > 0. Figure 4.5 shows how we derive the recursion tree for T .n/ D 3T .n=4/ C cn2. For convenience, we assume that n is an exact power of 4 (another example of tolerable sloppiness) so that all subproblem sizes are integers. Part (a) of the figure shows T .n/, which we expand in part (b) into an equivalent tree representing the recurrence. The cn2 term at the root represents the cost at the top level of recursion, and the three subtrees of the root represent the costs incurred by the subproblems of size n=4. Part (c) shows this process carried one step further by expanding each node with cost T .n=4/ from part (b). The cost for each of the three children of the root is c.n=4/2. We continue expanding each node in the tree by breaking it into its constituent parts as determined by the recurrence. 4.4 The recursion-tree method for solving recurrences 89 T .n/ cn2 cn2 T n 4 T n 4 T n 4 c n2 4 c n2 4 c n2 4 T n 16 T n 16 T n 16 T n 16 T n 16 T n 16 T n 16 T n 16 T n 16 (a) (b) (c) cn2 c n2 4 c n2 4 c n2 4 log4 n c n2 16 c n2 16 c n2 16 c n2 16 c n2 16 c n2 16 c n2 16 c n2 16 c n2 16 cn2 3 16 cn2 3 16 2 cn2 … T .1/ T .1/ T .1/ T .1/ T .1/ T .1/ T .1/ T .1/ T .1/ T .1/ … T .1/ T .1/ T .1/ nlog4 3 (d) ‚.nlog4 3/ Total: O.n2/ Figure 4.5 Constructing a recursion tree for the recurrence T .n/ D 3T .n=4/ C cn2. Part (a) shows T .n/, which progressively expands in (b)–(d) to form the recursion tree. The fully expanded tree in part (d) has height log4 n (it has log4 n C 1 levels). 90 Chapter 4 Divide-and-Conquer Because subproblem sizes decrease by a factor of 4 each time we go down one level, we eventually must reach a boundary condition. How far from the root do we reach one? The subproblem size for a node at depth i is n=4i . Thus, the subproblem size hits n D 1 when n=4i D 1 or, equivalently, when i D log4 n. Thus, the tree has log4 n C 1 levels (at depths 0; 1; 2; : : : ; log4 n). Next we determine the cost at each level of the tree. Each level has three times more nodes than the level above, and so the number of nodes at depth i is 3i . Because subproblem sizes reduce by a factor of 4 for each level we go down from the root, each node at depth i, for i D 0; 1; 2; : : : ; log4 n 1, has a cost of c.n=4i /2. Multiplying, we see that the total cost over all nodes at depth i , for i D 0; 1; 2; : : : ; log4 n 1, is 3i c.n=4i /2 D .3=16/i cn2. The bottom level, at depth log4 n, has 3log4 n D nlog4 3 nodes, each contributing cost T .1/, for a total cost of nlog4 3T .1/, which is ‚.nlog4 3/, since we assume that T .1/ is a constant. Now we add up the costs over all levels to determine the cost for the entire tree: T .n/ D cn2 C 3 cn2 C  3 Ã2 cn2 C 16 16 D logX 4 n 1  3 Ãi cn2 C ‚.nlog4 3/ 16 i D0  3 C Ãlog4 n 1 cn2 C ‚.nlog4 3/ 16 .3=16/log4 n D 1 cn2 C ‚.nlog4 3/ .3=16/ 1 (by equation (A.5)) : This last formula looks somewhat messy until we realize that we can again take advantage of small amounts of sloppiness and use an infinite decreasing geometric series as an upper bound. Backing up one step and applying equation (A.6), we have T .n/ D logX 4 n 1  3 Ãi cn2 C ‚.nlog4 3/ 16 i D0 < X 1  3 Ãi cn2 C ‚.nlog4 3/ 16 i D0 D 1 cn2 C ‚.nlog4 3/ 1 .3=16/ D 16 cn2 C ‚.nlog4 3/ 13 D O.n2/ : Thus, we have derived a guess of T .n/ D O.n2/ for our original recurrence T .n/ D 3T .bn=4c/ C ‚.n2/. In this example, the coefficients of cn2 form a decreasing geometric series and, by equation (A.6), the sum of these coefficients 4.4 The recursion-tree method for solving recurrences 91 cn cn c n 3 c 2n 3 cn log3=2 n c n 9 c 2n 9 c 2n 9 c 4n 9 cn … … Total: O.n lg n/ Figure 4.6 A recursion tree for the recurrence T .n/ D T .n=3/ C T .2n=3/ C cn. is bounded from above by the constant 16=13. Since the root’s contribution to the total cost is cn2, the root contributes a constant fraction of the total cost. In other words, the cost of the root dominates the total cost of the tree. In fact, if O.n2/ is indeed an upper bound for the recurrence (as we shall verify in a moment), then it must be a tight bound. Why? The first recursive call contributes a cost of ‚.n2/, and so .n2/ must be a lower bound for the recurrence. Now we can use the substitution method to verify that our guess was correct, that is, T .n/ D O.n2/ is an upper bound for the recurrence T .n/ D 3T .bn=4c/ C ‚.n2/. We want to show that T .n/ Ä d n2 for some constant d > 0. Using the same constant c > 0 as before, we have T .n/ Ä 3T .bn=4c/ C cn2 Ä 3d bn=4c2 C cn2 Ä 3d.n=4/2 C cn2 D 3 d n2 C cn2 16 Ä d n2 ; where the last step holds as long as d .16=13/c. In another, more intricate, example, Figure 4.6 shows the recursion tree for T .n/ D T .n=3/ C T .2n=3/ C O.n/ : (Again, we omit floor and ceiling functions for simplicity.) As before, we let c represent the constant factor in the O.n/ term. When we add the values across the levels of the recursion tree shown in the figure, we get a value of cn for every level. 92 Chapter 4 Divide-and-Conquer The longest simple path from the root to a leaf is n ! .2=3/n ! .2=3/2n ! ! 1. Since .2=3/kn D 1 when k D log3=2 n, the height of the tree is log3=2 n. Intuitively, we expect the solution to the recurrence to be at most the number of levels times the cost of each level, or O.cn log3=2 n/ D O.n lg n/. Figure 4.6 shows only the top levels of the recursion tree, however, and not every level in the tree contributes a cost of cn. Consider the cost of the leaves. If this recursion tree were a complete binary tree of height log3=2 n, there would be 2log3=2 n D nlog3=2 2 leaves. Since the cost of each leaf is a constant, the total cost of all leaves would then be ‚.nlog3=2 2/ which, since log3=2 2 is a constant strictly greater than 1, is !.n lg n/. This recursion tree is not a complete binary tree, however, and so it has fewer than nlog3=2 2 leaves. Moreover, as we go down from the root, more and more internal nodes are absent. Consequently, levels toward the bottom of the recursion tree contribute less than cn to the total cost. We could work out an accurate accounting of all costs, but remember that we are just trying to come up with a guess to use in the substitution method. Let us tolerate the sloppiness and attempt to show that a guess of O.n lg n/ for the upper bound is correct. Indeed, we can use the substitution method to verify that O.n lg n/ is an upper bound for the solution to the recurrence. We show that T .n/ Ä d n lg n, where d is a suitable positive constant. We have T .n/ Ä T .n=3/ C T .2n=3/ C cn Ä d.n=3/ lg.n=3/ C d.2n=3/ lg.2n=3/ C cn D .d.n=3/ lg n d.n=3/ lg 3/ C .d.2n=3/ lg n d.2n=3/ lg.3=2// C cn D d n lg n d..n=3/ lg 3 C .2n=3/ lg.3=2// C cn D d n lg n d..n=3/ lg 3 C .2n=3/ lg 3 .2n=3/ lg 2/ C cn D d n lg n d n.lg 3 2=3/ C cn Ä d n lg n ; as long as d c=.lg 3 .2=3//. Thus, we did not need to perform a more accurate accounting of costs in the recursion tree. Exercises 4.4-1 Use a recursion tree to determine a good asymptotic upper bound on the recurrence T .n/ D 3T .bn=2c/ C n. Use the substitution method to verify your answer. 4.4-2 Use a recursion tree to determine a good asymptotic upper bound on the recurrence T .n/ D T .n=2/ C n2. Use the substitution method to verify your answer. 4.5 The master method for solving recurrences 93 4.4-3 Use a recursion tree to determine a good asymptotic upper bound on the recurrence T .n/ D 4T .n=2 C 2/ C n. Use the substitution method to verify your answer. 4.4-4 Use a recursion tree to determine a good asymptotic upper bound on the recurrence T .n/ D 2T .n 1/ C 1. Use the substitution method to verify your answer. 4.4-5 Use a recursion tree to determine a good asymptotic upper bound on the recurrence T .n/ D T .n 1/CT .n=2/Cn. Use the substitution method to verify your answer. 4.4-6 Argue that the solution to the recurrence T .n/ D T .n=3/CT .2n=3/Ccn, where c is a constant, is .n lg n/ by appealing to a recursion tree. 4.4-7 Draw the recursion tree for T .n/ D 4T .bn=2c/ C cn, where c is a constant, and provide a tight asymptotic bound on its solution. Verify your bound by the substitution method. 4.4-8 Use a recursion tree to give an asymptotically tight solution to the recurrence T .n/ D T .n a/ C T .a/ C cn, where a 1 and c > 0 are constants. 4.4-9 Use a recursion tree to give an asymptotically tight solution to the recurrence T .n/ D T .˛ n/ C T ..1 ˛/n/ C cn, where ˛ is a constant in the range 0 < ˛ < 1 and c > 0 is also a constant. 4.5 The master method for solving recurrences The master method provides a “cookbook” method for solving recurrences of the form T .n/ D aT .n=b/ C f .n/ ; (4.20) where a 1 and b > 1 are constants and f .n/ is an asymptotically positive function. To use the master method, you will need to memorize three cases, but then you will be able to solve many recurrences quite easily, often without pencil and paper. 94 Chapter 4 Divide-and-Conquer The recurrence (4.20) describes the running time of an algorithm that divides a problem of size n into a subproblems, each of size n=b, where a and b are positive constants. The a subproblems are solved recursively, each in time T .n=b/. The function f .n/ encompasses the cost of dividing the problem and combining the results of the subproblems. For example, the recurrence arising from Strassen’s algorithm has a D 7, b D 2, and f .n/ D ‚.n2/. As a matter of technical correctness, the recurrence is not actually well defined, because n=b might not be an integer. Replacing each of the a terms T .n=b/ with either T .bn=bc/ or T .dn=be/ will not affect the asymptotic behavior of the recurrence, however. (We will prove this assertion in the next section.) We normally find it convenient, therefore, to omit the floor and ceiling functions when writing divide-and-conquer recurrences of this form. The master theorem The master method depends on the following theorem. Theorem 4.1 (Master theorem) Let a 1 and b > 1 be constants, let f .n/ be a function, and let T .n/ be defined on the nonnegative integers by the recurrence T .n/ D aT .n=b/ C f .n/ ; where we interpret n=b to mean either bn=bc or dn=be. Then T .n/ has the following asymptotic bounds: 1. If f .n/ D O.nlogb a / for some constant > 0, then T .n/ D ‚.nlogb a/. 2. If f .n/ D ‚.nlogb a/, then T .n/ D ‚.nlogb a lg n/. 3. If f .n/ D .nlogb aC / for some constant > 0, and if af .n=b/ Ä cf .n/ for some constant c < 1 and all sufficiently large n, then T .n/ D ‚.f .n//. Before applying the master theorem to some examples, let’s spend a moment trying to understand what it says. In each of the three cases, we compare the function f .n/ with the function nlogb a. Intuitively, the larger of the two functions determines the solution to the recurrence. If, as in case 1, the function nlogb a is the larger, then the solution is T .n/ D ‚.nlogb a/. If, as in case 3, the function f .n/ is the larger, then the solution is T .n/ D ‚.f .n//. If, as in case 2, the two func- tions are the same size, we multiply by a logarithmic factor, and the solution is T .n/ D ‚.nlogb a lg n/ D ‚.f .n/ lg n/. Beyond this intuition, you need to be aware of some technicalities. In the first case, not only must f .n/ be smaller than nlogb a, it must be polynomially smaller. 4.5 The master method for solving recurrences 95 That is, f .n/ must be asymptotically smaller than nlogb a by a factor of n for some constant > 0. In the third case, not only must f .n/ be larger than nlogb a, it also must be polynomially larger and in addition satisfy the “regularity” condition that af .n=b/ Ä cf .n/. This condition is satisfied by most of the polynomially bounded functions that we shall encounter. Note that the three cases do not cover all the possibilities for f .n/. There is a gap between cases 1 and 2 when f .n/ is smaller than nlogb a but not polynomially smaller. Similarly, there is a gap between cases 2 and 3 when f .n/ is larger than nlogb a but not polynomially larger. If the function f .n/ falls into one of these gaps, or if the regularity condition in case 3 fails to hold, you cannot use the master method to solve the recurrence. Using the master method To use the master method, we simply determine which case (if any) of the master theorem applies and write down the answer. As a first example, consider T .n/ D 9T .n=3/ C n : For this recurrence, we have a D 9, b D 3, f .n/ D n, and thus we have that nlogb a D nlog3 9 D ‚.n2). Since f .n/ D O.nlog3 9 /, where D 1, we can apply case 1 of the master theorem and conclude that the solution is T .n/ D ‚.n2/. Now consider T .n/ D T .2n=3/ C 1; in which a D 1, b D 3=2, f .n/ D 1, and nlogb a D nlog3=2 1 D n0 D 1. Case 2 applies, since f .n/ D ‚.nlogb a/ D ‚.1/, and thus the solution to the recurrence is T .n/ D ‚.lg n/. For the recurrence T .n/ D 3T .n=4/ C n lg n ; we have a D 3, b D 4, f .n/ D n lg n, and nlogb a D nlog4 3 D O.n0:793/. Since f .n/ D .nlog4 3C /, where 0:2, case 3 applies if we can show that the regularity condition holds for f .n/. For sufficiently large n, we have that af .n=b/ D 3.n=4/ lg.n=4/ Ä .3=4/n lg n D cf .n/ for c D 3=4. Consequently, by case 3, the solution to the recurrence is T .n/ D ‚.n lg n/. The master method does not apply to the recurrence T .n/ D 2T .n=2/ C n lg n ; even though it appears to have the proper form: a D 2, b D 2, f .n/ D n lg n, and nlogb a D n. You might mistakenly think that case 3 should apply, since 96 Chapter 4 Divide-and-Conquer f .n/ D n lg n is asymptotically larger than nlogb a D n. The problem is that it is not polynomially larger. The ratio f .n/=nlogb a D .n lg n/=n D lg n is asymptotically less than n for any positive constant . Consequently, the recurrence falls into the gap between case 2 and case 3. (See Exercise 4.6-2 for a solution.) Let’s use the master method to solve the recurrences we saw in Sections 4.1 and 4.2. Recurrence (4.7), T .n/ D 2T .n=2/ C ‚.n/ ; characterizes the running times of the divide-and-conquer algorithm for both the maximum-subarray problem and merge sort. (As is our practice, we omit stating the base case in the recurrence.) Here, we have a D 2, b D 2, f .n/ D ‚.n/, and thus we have that nlogb a D nlog2 2 D n. Case 2 applies, since f .n/ D ‚.n/, and so we have the solution T .n/ D ‚.n lg n/. Recurrence (4.17), T .n/ D 8T .n=2/ C ‚.n2/ ; describes the running time of the first divide-and-conquer algorithm that we saw for matrix multiplication. Now we have a D 8, b D 2, and f .n/ D ‚.n2/, and so nlogb a D nlog2 8 D n3. Since n3 is polynomially larger than f .n/ (that is, f .n/ D O.n3 / for D 1), case 1 applies, and T .n/ D ‚.n3/. Finally, consider recurrence (4.18), T .n/ D 7T .n=2/ C ‚.n2/ ; which describes the running time of Strassen’s algorithm. Here, we have a D 7, b D 2, f .n/ D ‚.n2/, and thus nlogb a D nlog2 7. Rewriting log2 7 as lg 7 and recalling that 2:80 < lg 7 < 2:81, we see that f .n/ D O.nlg 7 / for D 0:8. Again, case 1 applies, and we have the solution T .n/ D ‚.nlg 7/. Exercises 4.5-1 Use the master method to give tight asymptotic bounds for the following recurrences. a. T .n/ D 2T .n=4/ C 1. p b. T .n/ D 2T .n=4/ C n. c. T .n/ D 2T .n=4/ C n. d. T .n/ D 2T .n=4/ C n2. 4.6 Proof of the master theorem 97 4.5-2 Professor Caesar wishes to develop a matrix-multiplication algorithm that is asymptotically faster than Strassen’s algorithm. His algorithm will use the divideand-conquer method, dividing each matrix into pieces of size n=4 n=4, and the divide and combine steps together will take ‚.n2/ time. He needs to determine how many subproblems his algorithm has to create in order to beat Strassen’s algorithm. If his algorithm creates a subproblems, then the recurrence for the running time T .n/ becomes T .n/ D aT .n=4/ C ‚.n2/. What is the largest integer value of a for which Professor Caesar’s algorithm would be asymptotically faster than Strassen’s algorithm? 4.5-3 Use the master method to show that the solution to the binary-search recurrence T .n/ D T .n=2/ C ‚.1/ is T .n/ D ‚.lg n/. (See Exercise 2.3-5 for a description of binary search.) 4.5-4 Can the master method be applied to the recurrence T .n/ D 4T .n=2/ C n2 lg n? Why or why not? Give an asymptotic upper bound for this recurrence. 4.5-5 ? Consider the regularity condition af .n=b/ Ä cf .n/ for some constant c < 1, which is part of case 3 of the master theorem. Give an example of constants a 1 and b > 1 and a function f .n/ that satisfies all the conditions in case 3 of the master theorem except the regularity condition. ? 4.6 Proof of the master theorem This section contains a proof of the master theorem (Theorem 4.1). You do not need to understand the proof in order to apply the master theorem. The proof appears in two parts. The first part analyzes the master recurrence (4.20), under the simplifying assumption that T .n/ is defined only on exact powers of b > 1, that is, for n D 1; b; b2; : : :. This part gives all the intuition needed to understand why the master theorem is true. The second part shows how to extend the analysis to all positive integers n; it applies mathematical technique to the problem of handling floors and ceilings. In this section, we shall sometimes abuse our asymptotic notation slightly by using it to describe the behavior of functions that are defined only over exact powers of b. Recall that the definitions of asymptotic notations require that 98 Chapter 4 Divide-and-Conquer bounds be proved for all sufficiently large numbers, not just those that are powers of b. Since we could make new asymptotic notations that apply only to the set fbi W i D 0; 1; 2; : : :g, instead of to the nonnegative numbers, this abuse is minor. Nevertheless, we must always be on guard when we use asymptotic notation over a limited domain lest we draw improper conclusions. For example, proving that T .n/ D O.n/ when n is an exact power of 2 does not guarantee that T .n/ D O.n/. The function T .n/ could be defined as ( n if n D 1; 2; 4; 8; : : : ; T .n/ D n2 otherwise ; in which case the best upper bound that applies to all values of n is T .n/ D O.n2/. Because of this sort of drastic consequence, we shall never use asymptotic notation over a limited domain without making it absolutely clear from the context that we are doing so. 4.6.1 The proof for exact powers The first part of the proof of the master theorem analyzes the recurrence (4.20) T .n/ D aT .n=b/ C f .n/ ; for the master method, under the assumption that n is an exact power of b > 1, where b need not be an integer. We break the analysis into three lemmas. The first reduces the problem of solving the master recurrence to the problem of evaluating an expression that contains a summation. The second determines bounds on this summation. The third lemma puts the first two together to prove a version of the master theorem for the case in which n is an exact power of b. Lemma 4.2 Let a 1 and b > 1 be constants, and let f .n/ be a nonnegative function defined on exact powers of b. Define T .n/ on exact powers of b by the recurrence ( ‚.1/ if n D 1 ; T .n/ D aT .n=b/ C f .n/ if n D bi ; where i is a positive integer. Then logX b n 1 T .n/ D ‚.nlogb a/ C aj f .n=bj / : j D0 (4.21) Proof We use the recursion tree in Figure 4.7. The root of the tree has cost f .n/, and it has a children, each with cost f .n=b/. (It is convenient to think of a as being 4.6 Proof of the master theorem f .n/ a f .n=b/ f .n=b/ … f .n=b/ a a logb n f .n=b2/f .n=b2/…f .n=b2/ f .n=b2/f .n=b2/…f .n=b2/ a f .n=b2/f .n=b2/…f .n=b2/ a a a …… … a a a …… … a a a …… … 99 f .n/ af .n=b/ a2f .n=b2/ … ‚.1/ ‚.1/ ‚.1/ ‚.1/ ‚.1/ ‚.1/ ‚.1/ ‚.1/ ‚.1/ ‚.1/ … ‚.1/ ‚.1/ ‚.1/ ‚.nlogb a/ nlogb a logX b n 1 Total: ‚.nlogb a/ C aj f .n=bj / j D0 Figure 4.7 The recursion tree generated by T .n/ D aT .n=b/ C f .n/. The tree is a complete a-ary tree with nlogb a leaves and height logb n. The cost of the nodes at each depth is shown at the right, and their sum is given in equation (4.21). an integer, especially when visualizing the recursion tree, but the mathematics does not require it.) Each of these children has a children, making a2 nodes at depth 2, and each of the a children has cost f .n=b2/. In general, there are aj nodes at depth j , and each has cost f .n=bj /. The cost of each leaf is T .1/ D ‚.1/, and each leaf is at depth logb n, since n=blogb n D 1. There are alogb n D nlogb a leaves in the tree. We can obtain equation (4.21) by summing the costs of the nodes at each depth in the tree, as shown in the figure. The cost for all internal nodes at depth j is aj f .n=bj /, and so the total cost of all internal nodes is logX b n 1 aj f .n=bj / : j D0 In the underlying divide-and-conquer algorithm, this sum represents the costs of dividing problems into subproblems and then recombining the subproblems. The 100 Chapter 4 Divide-and-Conquer cost of all the leaves, which is the cost of doing all nlogb a subproblems of size 1, is ‚.nlogb a/. In terms of the recursion tree, the three cases of the master theorem correspond to cases in which the total cost of the tree is (1) dominated by the costs in the leaves, (2) evenly distributed among the levels of the tree, or (3) dominated by the cost of the root. The summation in equation (4.21) describes the cost of the dividing and combining steps in the underlying divide-and-conquer algorithm. The next lemma provides asymptotic bounds on the summation’s growth. Lemma 4.3 Let a 1 and b > 1 be constants, and let f .n/ be a nonnegative function defined on exact powers of b. A function g.n/ defined over exact powers of b by logX b n 1 g.n/ D aj f .n=bj / j D0 (4.22) has the following asymptotic bounds for exact powers of b: 1. If f .n/ D O.nlogb a / for some constant > 0, then g.n/ D O.nlogb a/. 2. If f .n/ D ‚.nlogb a/, then g.n/ D ‚.nlogb a lg n/. 3. If af .n=b/ Ä cf .n/ for some constant c < 1 and for all sufficiently large n, then g.n/ D ‚.f .n//. Proof For case 1, we have f .n/ D O.nlogb a /, which implies that f .n=bj / D O..n=bj /logb a /. Substituting into equation (4.22) yields logX b n 1 g.n/ D O aj n Álogb a ! : bj j D0 (4.23) We bound the summation within the O-notation by factoring out terms and simpli- fying, which leaves an increasing geometric series: logX b n 1 aj j D0 n Álogb a bj D nlogb a logX b n 1  ab Ãj j D0 blogb a logX b n 1 D nlogb a .b /j D nlogb a ÂjbD0logb n à 1 b1 4.6 Proof of the master theorem 101  D nlogb a n b à 1 : 1 Since b and are constants, we can rewrite the last expression as nlogb a O.n / D O.nlogb a/. Substituting this expression for the summation in equation (4.23) yields g.n/ D O.nlogb a/ ; thereby proving case 1. Because case 2 assumes that f .n/ D ‚.nlogb a/, we have that f .n=bj / D ‚..n=bj /logb a/. Substituting into equation (4.22) yields g.n/ D ‚ logX b n 1 aj j D0 n Álogb ! a : bj (4.24) We bound the summation within the ‚-notation as in case 1, but this time we do not obtain a geometric series. Instead, we discover that every term of the summation is the same: logX b n 1 aj n Álogb a D logX b n 1 nlogb a a Áj bj blogb a j D0 j D0 logX b n 1 D nlogb a 1 j D0 D nlogb a logb n : Substituting this expression for the summation in equation (4.24) yields g.n/ D ‚.nlogb a logb n/ D ‚.nlogb a lg n/ ; proving case 2. We prove case 3 similarly. Since f .n/ appears in the definition (4.22) of g.n/ and all terms of g.n/ are nonnegative, we can conclude that g.n/ D .f .n// for exact powers of b. We assume in the statement of the lemma that af .n=b/ Ä cf .n/ for some constant c < 1 and all sufficiently large n. We rewrite this assumption as f .n=b/ Ä .c=a/f .n/ and iterate j times, yielding f .n=bj / Ä .c=a/j f .n/ or, equivalently, aj f .n=bj / Ä cj f .n/, where we assume that the values we iterate on are sufficiently large. Since the last, and smallest, such value is n=bj 1, it is enough to assume that n=bj 1 is sufficiently large. Substituting into equation (4.22) and simplifying yields a geometric series, but unlike the series in case 1, this one has decreasing terms. We use an O.1/ term to 102 Chapter 4 Divide-and-Conquer capture the terms that are not covered by our assumption that n is sufficiently large: logX b n 1 g.n/ D aj f .n=bj / j D0 logX b n 1 Ä cj f .n/ C O.1/ j D0 X 1 Ä f .n/ cj C O.1/ jÂD0 à 1 D f .n/ C O.1/ 1c D O.f .n// ; since c is a constant. Thus, we can conclude that g.n/ D ‚.f .n// for exact powers of b. With case 3 proved, the proof of the lemma is complete. We can now prove a version of the master theorem for the case in which n is an exact power of b. Lemma 4.4 Let a 1 and b > 1 be constants, and let f .n/ be a nonnegative function defined on exact powers of b. Define T .n/ on exact powers of b by the recurrence ( T .n/ D ‚.1/ if n D 1 ; aT .n=b/ C f .n/ if n D bi ; where i is a positive integer. Then T .n/ has the following asymptotic bounds for exact powers of b: 1. If f .n/ D O.nlogb a / for some constant > 0, then T .n/ D ‚.nlogb a/. 2. If f .n/ D ‚.nlogb a/, then T .n/ D ‚.nlogb a lg n/. 3. If f .n/ D .nlogb aC / for some constant > 0, and if af .n=b/ Ä cf .n/ for some constant c < 1 and all sufficiently large n, then T .n/ D ‚.f .n//. Proof We use the bounds in Lemma 4.3 to evaluate the summation (4.21) from Lemma 4.2. For case 1, we have T .n/ D ‚.nlogb a/ C O.nlogb a/ D ‚.nlogb a/ ; 4.6 Proof of the master theorem 103 and for case 2, T .n/ D ‚.nlogb a/ C ‚.nlogb a lg n/ D ‚.nlogb a lg n/ : For case 3, T .n/ D ‚.nlogb a/ C ‚.f .n// D ‚.f .n// ; because f .n/ D .nlogb aC /. 4.6.2 Floors and ceilings To complete the proof of the master theorem, we must now extend our analysis to the situation in which floors and ceilings appear in the master recurrence, so that the recurrence is defined for all integers, not for just exact powers of b. Obtaining a lower bound on T .n/ D aT .dn=be/ C f .n/ (4.25) and an upper bound on T .n/ D aT .bn=bc/ C f .n/ (4.26) is routine, since we can push through the bound dn=be n=b in the first case to yield the desired result, and we can push through the bound bn=bc Ä n=b in the second case. We use much the same technique to lower-bound the recurrence (4.26) as to upper-bound the recurrence (4.25), and so we shall present only this latter bound. We modify the recursion tree of Figure 4.7 to produce the recursion tree in Figure 4.8. As we go down in the recursion tree, we obtain a sequence of recursive invocations on the arguments n; dn=be ; ddn=be =be ; dddn=be =be =be ; ::: Let us denote the j th element in the sequence by nj , where ( nj D n if j D 0 ; dnj 1=be if j > 0 : (4.27) 104 Chapter 4 Divide-and-Conquer f .n/ a f .n1/ a blogb nc f .n2/ f .n2/ … f .n2/ a a a …… … f .n1/ … f .n1/ a a f .n2/ f .n2/ … f .n2/ a a a …… … f .n2/ f .n2/ … f .n2/ a a a …… … f .n/ af .n1/ a2f .n2/ … ‚.1/ ‚.1/ ‚.1/ ‚.1/ ‚.1/ ‚.1/ ‚.1/ ‚.1/ ‚.1/ ‚.1/ … ‚.1/ ‚.1/ ‚.1/ ‚.nlogb a/ ‚.nlogb a/ X blogb nc 1 Total: ‚.nlogb a/ C aj f .nj / j D0 Figure 4.8 The recursion tree generated by T .n/ D aT .dn=be/Cf .n/. The recursive argument nj is given by equation (4.27). Our first goal is to determine the depth k such that nk is a constant. Using the inequality dxe Ä x C 1, we obtain n0 Ä n ; n n1 Ä C1; b n1 n2 Ä C C1; b2 b n 11 n3 Ä C C C1; b3 b2 b ::: In general, we have 4.6 Proof of the master theorem 105 n X j 1 1 nj Ä C bj bi i D0 n X 1 1 bCb=.b 1/, where c < 1 is a constant, then it follows that aj f .nj / Ä cj f .n/. Therefore, we can evaluate the sum in equation (4.29) just as in Lemma 4.3. For case 2, we have f .n/ D ‚.nlogb a/. If we can show that f .nj / D O.nlogb a=aj / D O..n=bj /logb a/, then the proof for case 2 of Lemma 4.3 will go through. Observe that j Ä blogb nc implies bj =n Ä 1. The bound f .n/ D O.nlogb a/ implies that there exists a constant c > 0 such that for all sufficiently large nj , 106 Chapter 4 Divide-and-Conquer  n b Ãlogb a f .nj / Ä c C bj b 1  n bj b ÃÃlogb a Dc 1C bj nb 1  nlogb a à   bj b ÃÃlogb a Dc 1C aj nb 1  nlogb a à  b Ãlogb a Äc 1C aj b1  nlogb a à DO ; aj since c.1 C b=.b 1//logb a is a constant. Thus, we have proved case 2. The proof of case 1 is almost identical. The key is to prove the bound f .nj / D O.nlogb a /, which is similar to the corresponding proof of case 2, though the algebra is more intricate. We have now proved the upper bounds in the master theorem for all integers n. The proof of the lower bounds is similar. Exercises 4.6-1 ? Give a simple and exact expression for nj in equation (4.27) for the case in which b is a positive integer instead of an arbitrary real number. 4.6-2 ? Show that if f .n/ D ‚.nlogb a lgk n/, where k 0, then the master recurrence has solution T .n/ D ‚.nlogb a lgkC1 n/. For simplicity, confine your analysis to exact powers of b. 4.6-3 ? Show that case 3 of the master theorem is overstated, in the sense that the regularity condition af .n=b/ Ä cf .n/ for some constant c < 1 implies that there exists a constant > 0 such that f .n/ D .nlogb aC /. Problems for Chapter 4 107 Problems 4-1 Recurrence examples Give asymptotic upper and lower bounds for T .n/ in each of the following recurrences. Assume that T .n/ is constant for n Ä 2. Make your bounds as tight as possible, and justify your answers. a. T .n/ D 2T .n=2/ C n4. b. T .n/ D T .7n=10/ C n. c. T .n/ D 16T .n=4/ C n2. d. T .n/ D 7T .n=3/ C n2. e. T .n/ D 7T .n=2/ C n2. p f. T .n/ D 2T .n=4/ C n. g. T .n/ D T .n 2/ C n2. 4-2 Parameter-passing costs Throughout this book, we assume that parameter passing during procedure calls takes constant time, even if an N -element array is being passed. This assumption is valid in most systems because a pointer to the array is passed, not the array itself. This problem examines the implications of three parameter-passing strategies: 1. An array is passed by pointer. Time D ‚.1/. 2. An array is passed by copying. Time D ‚.N /, where N is the size of the array. 3. An array is passed by copying only the subrange that might be accessed by the called procedure. Time D ‚.q p C 1/ if the subarray AŒp : : q is passed. a. Consider the recursive binary search algorithm for finding a number in a sorted array (see Exercise 2.3-5). Give recurrences for the worst-case running times of binary search when arrays are passed using each of the three methods above, and give good upper bounds on the solutions of the recurrences. Let N be the size of the original problem and n be the size of a subproblem. b. Redo part (a) for the MERGE-SORT algorithm from Section 2.3.1. 108 Chapter 4 Divide-and-Conquer 4-3 More recurrence examples Give asymptotic upper and lower bounds for T .n/ in each of the following recurrences. Assume that T .n/ is constant for sufficiently small n. Make your bounds as tight as possible, and justify your answers. a. T .n/ D 4T .n=3/ C n lg n. b. T .n/ D 3T .n=3/ C n= lg n. c. T .n/ D 4T .n=2/ C n2pn. d. T .n/ D 3T .n=3 2/ C n=2. e. T .n/ D 2T .n=2/ C n= lg n. f. T .n/ D T .n=2/ C T .n=4/ C T .n=8/ C n. g. T .n/ D T .n 1/ C 1=n. h. T .n/ D T .n 1/ C lg n. i. T .n/ D T .n 2/ C 1= lg n. pp j. T .n/ D nT . n/ C n. 4-4 Fibonacci numbers This problem develops properties of the Fibonacci numbers, which are defined by recurrence (3.22). We shall use the technique of generating functions to solve the Fibonacci recurrence. Define the generating function (or formal power series) F as X 1 F .´/ D Fi ´i i D0 D 0 C ´ C ´2 C 2´3 C 3´4 C 5´5 C 8´6 C 13´7 C 21´8 C ; where Fi is the i th Fibonacci number. a. Show that F .´/ D ´ C ´F .´/ C ´2F .´/. Problems for Chapter 4 109 b. Show that F .´/ D D D ´ 1 ´ ´2 ´ .1  ´/.1 y´/ à 1 1 1 p 51 ´ 1 y´ ; where p 1C 5 D D 1:61803 : : : 2 and p yD 1 5 D 2 0:61803 : : : : c. Show that X 1 1 F .´/ D p . i iD0 5 yi /´i : d. (UHsienpt:aOrt b(cse)rtvoeptrhoavteˇˇtyhˇˇat 0, rounded to the nearest integer. 4-5 Chip testing Professor Diogenes has n supposedly identical integrated-circuit chips that in principle are capable of testing each other. The professor’s test jig accommodates two chips at a time. When the jig is loaded, each chip tests the other and reports whether it is good or bad. A good chip always reports accurately whether the other chip is good or bad, but the professor cannot trust the answer of a bad chip. Thus, the four possible outcomes of a test are as follows: Chip A says B is good B is good B is bad B is bad Chip B says A is good A is bad A is good A is bad Conclusion both are good, or both are bad at least one is bad at least one is bad at least one is bad a. Show that if more than n=2 chips are bad, the professor cannot necessarily determine which chips are good using any strategy based on this kind of pairwise test. Assume that the bad chips can conspire to fool the professor. 110 Chapter 4 Divide-and-Conquer b. Consider the problem of finding a single good chip from among n chips, assuming that more than n=2 of the chips are good. Show that bn=2c pairwise tests are sufficient to reduce the problem to one of nearly half the size. c. Show that the good chips can be identified with ‚.n/ pairwise tests, assuming that more than n=2 of the chips are good. Give and solve the recurrence that describes the number of tests. 4-6 Monge arrays An m n array A of real numbers is a Monge array if for all i, j , k, and l such that 1 Ä i < k Ä m and 1 Ä j < l Ä n, we have AŒi; j  C AŒk; l Ä AŒi; l C AŒk; j  : In other words, whenever we pick two rows and two columns of a Monge array and consider the four elements at the intersections of the rows and the columns, the sum of the upper-left and lower-right elements is less than or equal to the sum of the lower-left and upper-right elements. For example, the following array is Monge: 10 17 13 28 23 17 22 16 29 23 24 28 22 34 24 11 13 6 17 7 45 44 32 37 23 36 33 19 21 6 75 66 51 53 34 a. Prove that an array is Monge if and only if for all i D 1; 2; :::; m 1 and j D 1; 2; :::; n 1, we have AŒi; j  C AŒi C 1; j C 1 Ä AŒi; j C 1 C AŒi C 1; j  : (Hint: For the “if” part, use induction separately on rows and columns.) b. The following array is not Monge. Change one element in order to make it Monge. (Hint: Use part (a).) 37 23 22 32 21 6 7 10 53 34 30 31 32 13 9 6 43 21 15 8 Notes for Chapter 4 111 c. Let f .i/ be the index of the column containing the leftmost minimum element of row i. Prove that f .1/ Ä f .2/ Ä Ä f .m/ for any m n Monge array. d. Here is a description of a divide-and-conquer algorithm that computes the leftmost minimum element in each row of an m n Monge array A: Construct a submatrix A0 of A consisting of the even-numbered rows of A. Recursively determine the leftmost minimum for each row of A0. Then compute the leftmost minimum in the odd-numbered rows of A. Explain how to compute the leftmost minimum in the odd-numbered rows of A (given that the leftmost minimum of the even-numbered rows is known) in O.m C n/ time. e. Write the recurrence describing the running time of the algorithm described in part (d). Show that its solution is O.m C n log m/. Chapter notes Divide-and-conquer as a technique for designing algorithms dates back to at least 1962 in an article by Karatsuba and Ofman [194]. It might have been used well before then, however; according to Heideman, Johnson, and Burrus [163], C. F. Gauss devised the first fast Fourier transform algorithm in 1805, and Gauss’s formulation breaks the problem into smaller subproblems whose solutions are combined. The maximum-subarray problem in Section 4.1 is a minor variation on a problem studied by Bentley [43, Chapter 7]. Strassen’s algorithm [325] caused much excitement when it was published in 1969. Before then, few imagined the possibility of an algorithm asymptotically faster than the basic SQUARE-MATRIX-MULTIPLY procedure. The asymptotic upper bound for matrix multiplication has been improved since then. The most asymptotically efficient algorithm for multiplying n n matrices to date, due to Coppersmith and Winograd [78], has a running time of O.n2:376/. The best lower bound known is just the obvious .n2/ bound (obvious because we must fill in n2 elements of the product matrix). From a practical point of view, Strassen’s algorithm is often not the method of choice for matrix multiplication, for four reasons: 1. The constant factor hidden in the ‚.nlg 7/ running time of Strassen’s algorithm is larger than the constant factor in the ‚.n3/-time SQUARE-MATRIXMULTIPLY procedure. 2. When the matrices are sparse, methods tailored for sparse matrices are faster. 112 Chapter 4 Divide-and-Conquer 3. Strassen’s algorithm is not quite as numerically stable as SQUARE-MATRIXMULTIPLY. In other words, because of the limited precision of computer arithmetic on noninteger values, larger errors accumulate in Strassen’s algorithm than in SQUARE-MATRIX-MULTIPLY. 4. The submatrices formed at the levels of recursion consume space. The latter two reasons were mitigated around 1990. Higham [167] demonstrated that the difference in numerical stability had been overemphasized; although Strassen’s algorithm is too numerically unstable for some applications, it is within acceptable limits for others. Bailey, Lee, and Simon [32] discuss techniques for reducing the memory requirements for Strassen’s algorithm. In practice, fast matrix-multiplication implementations for dense matrices use Strassen’s algorithm for matrix sizes above a “crossover point,” and they switch to a simpler method once the subproblem size reduces to below the crossover point. The exact value of the crossover point is highly system dependent. Analyses that count operations but ignore effects from caches and pipelining have produced crossover points as low as n D 8 (by Higham [167]) or n D 12 (by Huss-Lederman et al. [186]). D’Alberto and Nicolau [81] developed an adaptive scheme, which determines the crossover point by benchmarking when their software package is installed. They found crossover points on various systems ranging from n D 400 to n D 2150, and they could not find a crossover point on a couple of systems. Recurrences were studied as early as 1202 by L. Fibonacci, for whom the Fi- bonacci numbers are named. A. De Moivre introduced the method of generating functions (see Problem 4-4) for solving recurrences. The master method is adapted from Bentley, Haken, and Saxe [44], which provides the extended method justified by Exercise 4.6-2. Knuth [209] and Liu [237] show how to solve linear recurrences using the method of generating functions. Purdom and Brown [287] and Graham, Knuth, and Patashnik [152] contain extended discussions of recurrence solving. Several researchers, including Akra and Bazzi [13], Roura [299], Verma [346], and Yap [360], have given methods for solving more general divide-and-conquer recurrences than are solved by the master method. We describe the result of Akra and Bazzi here, as modified by Leighton [228]. The Akra-Bazzi method works for recurrences of the form ( T .x/ D P ‚.k1/ i D1 ai T .bi x / C f .x / if 1 Ä x Ä x0 ; if x > x0 ; (4.30) where x 1 is a real number, x0 is a constant such that x0 1=bi and x0 1=.1 bi / for i D 1; 2; : : : ; k, ai is a positive constant for i D 1; 2; : : : ; k, Notes for Chapter 4 113 bi is a constant in the range 0 < bi < 1 for i D 1; 2; : : : ; k, k 1 is an integer constant, and f .x/ is a nonnegative function that satisfies the polynomial-growth condition: there exist positive constants c1 and c2 such that for all x 1, for i D 1; 2; : : : ; k, and for all u such that bi x Ä u Ä x, we have c1f .x/ Ä f .u/ Ä c2f .x/. (If jf 0.x/j is upper-bounded by some polynomial in x, then f .x/ satisfies the polynomial-growth condition. For example, f .x/ D x˛ lgˇ x satisfies this condition for any real constants ˛ and ˇ.) Although the master method does not apply to a recurrence such as T .n/ D T .bn=3c/ C T .b2n=3c/ currence (4.30), we first C O.n/, the Akra-Bazzi method does. find the unique real number p such that PTokiDs1olaviebipthDe re1-. (Such a p always exists.) The solution to the recurrence is then T .n/ D  ‚ xp 1 Zx C 1 f .u/ upC1 Ãà du : The Akra-Bazzi method can be somewhat difficult to use, but it serves in solving recurrences that model division of the problem into substantially unequally sized subproblems. The master method is simpler to use, but it applies only when subproblem sizes are equal. 5 Probabilistic Analysis and Randomized Algorithms This chapter introduces probabilistic analysis and randomized algorithms. If you are unfamiliar with the basics of probability theory, you should read Appendix C, which reviews this material. We shall revisit probabilistic analysis and randomized algorithms several times throughout this book. 5.1 The hiring problem Suppose that you need to hire a new office assistant. Your previous attempts at hiring have been unsuccessful, and you decide to use an employment agency. The employment agency sends you one candidate each day. You interview that person and then decide either to hire that person or not. You must pay the employment agency a small fee to interview an applicant. To actually hire an applicant is more costly, however, since you must fire your current office assistant and pay a substantial hiring fee to the employment agency. You are committed to having, at all times, the best possible person for the job. Therefore, you decide that, after interviewing each applicant, if that applicant is better qualified than the current office assistant, you will fire the current office assistant and hire the new applicant. You are willing to pay the resulting price of this strategy, but you wish to estimate what that price will be. The procedure HIRE-ASSISTANT, given below, expresses this strategy for hiring in pseudocode. It assumes that the candidates for the office assistant job are numbered 1 through n. The procedure assumes that you are able to, after interviewing candidate i, determine whether candidate i is the best candidate you have seen so far. To initialize, the procedure creates a dummy candidate, numbered 0, who is less qualified than each of the other candidates. 5.1 The hiring problem 115 HIRE-ASSISTANT .n/ 1 best D 0 // candidate 0 is a least-qualified dummy candidate 2 for i D 1 to n 3 interview candidate i 4 if candidate i is better than candidate best 5 best D i 6 hire candidate i The cost model for this problem differs from the model described in Chapter 2. We focus not on the running time of HIRE-ASSISTANT, but instead on the costs incurred by interviewing and hiring. On the surface, analyzing the cost of this algorithm may seem very different from analyzing the running time of, say, merge sort. The analytical techniques used, however, are identical whether we are analyzing cost or running time. In either case, we are counting the number of times certain basic operations are executed. Interviewing has a low cost, say ci , whereas hiring is expensive, costing ch. Letting m be the number of people hired, the total cost associated with this algorithm is O.ci n C chm/. No matter how many people we hire, we always interview n candidates and thus always incur the cost ci n associated with interviewing. We therefore concentrate on analyzing chm, the hiring cost. This quantity varies with each run of the algorithm. This scenario serves as a model for a common computational paradigm. We often need to find the maximum or minimum value in a sequence by examining each element of the sequence and maintaining a current “winner.” The hiring problem models how often we update our notion of which element is currently winning. Worst-case analysis In the worst case, we actually hire every candidate that we interview. This situation occurs if the candidates come in strictly increasing order of quality, in which case we hire n times, for a total hiring cost of O.chn/. Of course, the candidates do not always come in increasing order of quality. In fact, we have no idea about the order in which they arrive, nor do we have any control over this order. Therefore, it is natural to ask what we expect to happen in a typical or average case. Probabilistic analysis Probabilistic analysis is the use of probability in the analysis of problems. Most commonly, we use probabilistic analysis to analyze the running time of an algorithm. Sometimes we use it to analyze other quantities, such as the hiring cost 116 Chapter 5 Probabilistic Analysis and Randomized Algorithms in procedure HIRE-ASSISTANT. In order to perform a probabilistic analysis, we must use knowledge of, or make assumptions about, the distribution of the inputs. Then we analyze our algorithm, computing an average-case running time, where we take the average over the distribution of the possible inputs. Thus we are, in effect, averaging the running time over all possible inputs. When reporting such a running time, we will refer to it as the average-case running time. We must be very careful in deciding on the distribution of inputs. For some problems, we may reasonably assume something about the set of all possible inputs, and then we can use probabilistic analysis as a technique for designing an efficient algorithm and as a means for gaining insight into a problem. For other problems, we cannot describe a reasonable input distribution, and in these cases we cannot use probabilistic analysis. For the hiring problem, we can assume that the applicants come in a random order. What does that mean for this problem? We assume that we can compare any two candidates and decide which one is better qualified; that is, there is a total order on the candidates. (See Appendix B for the definition of a total order.) Thus, we can rank each candidate with a unique number from 1 through n, using rank.i/ to denote the rank of applicant i, and adopt the convention that a higher rank corresponds to a better qualified applicant. The ordered list hrank.1/; rank.2/; : : : ; rank.n/i is a permutation of the list h1; 2; : : : ; ni. Saying that the applicants come in a random order is equivalent to saying that this list of ranks is equally likely to be any one of the nŠ permutations of the numbers 1 through n. Alternatively, we say that the ranks form a uniform random permutation; that is, each of the possible nŠ permutations appears with equal probability. Section 5.2 contains a probabilistic analysis of the hiring problem. Randomized algorithms In order to use probabilistic analysis, we need to know something about the distribution of the inputs. In many cases, we know very little about the input distribution. Even if we do know something about the distribution, we may not be able to model this knowledge computationally. Yet we often can use probability and randomness as a tool for algorithm design and analysis, by making the behavior of part of the algorithm random. In the hiring problem, it may seem as if the candidates are being presented to us in a random order, but we have no way of knowing whether or not they really are. Thus, in order to develop a randomized algorithm for the hiring problem, we must have greater control over the order in which we interview the candidates. We will, therefore, change the model slightly. We say that the employment agency has n candidates, and they send us a list of the candidates in advance. On each day, we choose, randomly, which candidate to interview. Although we know nothing about 5.1 The hiring problem 117 the candidates (besides their names), we have made a significant change. Instead of relying on a guess that the candidates come to us in a random order, we have instead gained control of the process and enforced a random order. More generally, we call an algorithm randomized if its behavior is determined not only by its input but also by values produced by a random-number generator. We shall assume that we have at our disposal a random-number generator RANDOM. A call to RANDOM.a; b/ returns an integer between a and b, inclusive, with each such integer being equally likely. For example, RANDOM.0; 1/ produces 0 with probability 1=2, and it produces 1 with probability 1=2. A call to RANDOM.3; 7/ returns either 3, 4, 5, 6, or 7, each with probability 1=5. Each integer returned by RANDOM is independent of the integers returned on previous calls. You may imagine RANDOM as rolling a .b a C 1/-sided die to obtain its output. (In practice, most programming environments offer a pseudorandom-number generator: a deterministic algorithm returning numbers that “look” statistically random.) When analyzing the running time of a randomized algorithm, we take the expectation of the running time over the distribution of values returned by the random number generator. We distinguish these algorithms from those in which the input is random by referring to the running time of a randomized algorithm as an expected running time. In general, we discuss the average-case running time when the probability distribution is over the inputs to the algorithm, and we discuss the expected running time when the algorithm itself makes random choices. Exercises 5.1-1 Show that the assumption that we are always able to determine which candidate is best, in line 4 of procedure HIRE-ASSISTANT, implies that we know a total order on the ranks of the candidates. 5.1-2 ? Describe an implementation of the procedure RANDOM.a; b/ that only makes calls to RANDOM.0; 1/. What is the expected running time of your procedure, as a function of a and b? 5.1-3 ? Suppose that you want to output 0 with probability 1=2 and 1 with probability 1=2. At your disposal is a procedure BIASED-RANDOM, that outputs either 0 or 1. It outputs 1 with some probability p and 0 with probability 1 p, where 0 < p < 1, but you do not know what p is. Give an algorithm that uses BIASED-RANDOM as a subroutine, and returns an unbiased answer, returning 0 with probability 1=2 118 Chapter 5 Probabilistic Analysis and Randomized Algorithms and 1 with probability 1=2. What is the expected running time of your algorithm as a function of p? 5.2 Indicator random variables In order to analyze many algorithms, including the hiring problem, we use indicator random variables. Indicator random variables provide a convenient method for converting between probabilities and expectations. Suppose we are given a sample space S and an event A. Then the indicator random variable I fAg associated with event A is defined as ( I fAg D 1 if A occurs ; (5.1) 0 if A does not occur : As a simple example, let us determine the expected number of heads that we obtain when flipping a fair coin. Our sample space is S D fH; T g, with Pr fH g D Pr fT g D 1=2. We can then define an indicator random variable XH , associated with the coin coming up heads, which is the event H . This variable counts the number of heads obtained in this flip, and it is 1 if the coin comes up heads and 0 otherwise. We write XH D I(fH g D 1 if H occurs ; 0 if T occurs : The expected number of heads obtained in one flip of the coin is simply the expected value of our indicator variable XH : E ŒXH  D E ŒI fH g D 1 Pr fH g C 0 Pr fT g D 1 .1=2/ C 0 .1=2/ D 1=2 : Thus the expected number of heads obtained by one flip of a fair coin is 1=2. As the following lemma shows, the expected value of an indicator random variable associated with an event A is equal to the probability that A occurs. Lemma 5.1 Given a sample space S and an event A in the sample space S, let XA D I fAg. Then E ŒXA D Pr fAg. 5.2 Indicator random variables 119 Proof By the definition of an indicator random variable from equation (5.1) and the definition of expected value, we have E ŒXA D E ŒI fAg ˚« D 1 Pr fAg C 0 Pr A D Pr fAg ; where A denotes S A, the complement of A. Although indicator random variables may seem cumbersome for an application such as counting the expected number of heads on a flip of a single coin, they are useful for analyzing situations in which we perform repeated random trials. For example, indicator random variables give us a simple way to arrive at the result of equation (C.37). In this equation, we compute the number of heads in n coin flips by considering separately the probability of obtaining 0 heads, 1 head, 2 heads, etc. The simpler method proposed in equation (C.38) instead uses indicator random variables implicitly. Making this argument more explicit, we let Xi be the indicator random variable associated with the event in which the ith flip comes up heads: Xi D I fthe ith flip results in the event H g. Let X be the random variable denoting the total number of heads in the n coin flips, so that Xn X D Xi : i D1 We wish to compute the expected number of heads, and so we take the expectation of both sides of the above equation to obtain " # Xn E ŒX  D E Xi : i D1 The above equation gives the expectation of the sum of n indicator random variables. By Lemma 5.1, we can easily compute the expectation of each of the random variables. By equation (C.21)—linearity of expectation—it is easy to compute the expectation of the sum: it equals the sum of the expectations of the n random variables. Linearity of expectation makes the use of indicator random variables a powerful analytical technique; it applies even when there is dependence among the random variables. We now can easily compute the expected number of heads: 120 Chapter 5 Probabilistic Analysis and Randomized Algorithms " # Xn E ŒX  D E Xi i D1 Xn D E ŒXi  i D1 Xn D 1=2 i D1 D n=2 : Thus, compared to the method used in equation (C.37), indicator random variables greatly simplify the calculation. We shall use indicator random variables throughout this book. Analysis of the hiring problem using indicator random variables Returning to the hiring problem, we now wish to compute the expected number of times that we hire a new office assistant. In order to use a probabilistic analysis, we assume that the candidates arrive in a random order, as discussed in the previous section. (We shall see in Section 5.3 how to remove this assumption.) Let X be the random variable whose value equals the number of times we hire a new office assistant. We could then apply the definition of expected value from equation (C.20) to obtain Xn E ŒX  D x Pr fX D xg ; xD1 but this calculation would be cumbersome. We shall instead use indicator random variables to greatly simplify the calculation. To use indicator random variables, instead of computing E ŒX  by defining one variable associated with the number of times we hire a new office assistant, we define n variables related to whether or not each particular candidate is hired. In particular, we let Xi be the indicator random variable associated with the event in which the ith candidate is hired. Thus, Xi D (I fcandidate i is hiredg D 1 if candidate i is hired ; 0 if candidate i is not hired ; and X D X1 C X2 C C Xn : (5.2) 5.2 Indicator random variables 121 By Lemma 5.1, we have that E ŒXi  D Pr fcandidate i is hiredg ; and we must therefore compute the probability that lines 5–6 of HIRE-ASSISTANT are executed. Candidate i is hired, in line 6, exactly when candidate i is better than each of candidates 1 through i 1. Because we have assumed that the candidates arrive in a random order, the first i candidates have appeared in a random order. Any one of these first i candidates is equally likely to be the best-qualified so far. Candidate i has a probability of 1=i of being better qualified than candidates 1 through i 1 and thus a probability of 1=i of being hired. By Lemma 5.1, we conclude that E ŒXi  D 1=i : (5.3) Now we can compute E ŒX : " # Xn E ŒX  D E Xi (by equation (5.2)) (5.4) i D1 Xn D E ŒXi  (by linearity of expectation) i D1 Xn D 1=i (by equation (5.3)) i D1 D ln n C O.1/ (by equation (A.7)) . (5.5) Even though we interview n people, we actually hire only approximately ln n of them, on average. We summarize this result in the following lemma. Lemma 5.2 Assuming that the candidates are presented in a random order, algorithm HIREASSISTANT has an average-case total hiring cost of O.ch ln n/. Proof The bound follows immediately from our definition of the hiring cost and equation (5.5), which shows that the expected number of hires is approximately ln n. The average-case hiring cost is a significant improvement over the worst-case hiring cost of O.chn/. 122 Chapter 5 Probabilistic Analysis and Randomized Algorithms Exercises 5.2-1 In HIRE-ASSISTANT, assuming that the candidates are presented in a random order, what is the probability that you hire exactly one time? What is the probability that you hire exactly n times? 5.2-2 In HIRE-ASSISTANT, assuming that the candidates are presented in a random order, what is the probability that you hire exactly twice? 5.2-3 Use indicator random variables to compute the expected value of the sum of n dice. 5.2-4 Use indicator random variables to solve the following problem, which is known as the hat-check problem. Each of n customers gives a hat to a hat-check person at a restaurant. The hat-check person gives the hats back to the customers in a random order. What is the expected number of customers who get back their own hat? 5.2-5 Let AŒ1 : : n be an array of n distinct numbers. If i < j and AŒi > AŒj , then the pair .i; j / is called an inversion of A. (See Problem 2-4 for more on inversions.) Suppose that the elements of A form a uniform random permutation of h1; 2; : : : ; ni. Use indicator random variables to compute the expected number of inversions. 5.3 Randomized algorithms In the previous section, we showed how knowing a distribution on the inputs can help us to analyze the average-case behavior of an algorithm. Many times, we do not have such knowledge, thus precluding an average-case analysis. As mentioned in Section 5.1, we may be able to use a randomized algorithm. For a problem such as the hiring problem, in which it is helpful to assume that all permutations of the input are equally likely, a probabilistic analysis can guide the development of a randomized algorithm. Instead of assuming a distribution of inputs, we impose a distribution. In particular, before running the algorithm, we randomly permute the candidates in order to enforce the property that every permutation is equally likely. Although we have modified the algorithm, we still expect to hire a new office assistant approximately ln n times. But now we expect 5.3 Randomized algorithms 123 this to be the case for any input, rather than for inputs drawn from a particular distribution. Let us further explore the distinction between probabilistic analysis and randomized algorithms. In Section 5.2, we claimed that, assuming that the candidates arrive in a random order, the expected number of times we hire a new office assistant is about ln n. Note that the algorithm here is deterministic; for any particular input, the number of times a new office assistant is hired is always the same. Furthermore, the number of times we hire a new office assistant differs for different inputs, and it depends on the ranks of the various candidates. Since this number depends only on the ranks of the candidates, we can represent a particular input by listing, in order, the ranks of the candidates, i.e., hrank.1/; rank.2/; : : : ; rank.n/i. Given the rank list A1 D h1; 2; 3; 4; 5; 6; 7; 8; 9; 10i, a new office assistant is always hired 10 times, since each successive candidate is better than the previous one, and lines 5–6 are executed in each iteration. Given the list of ranks A2 D h10; 9; 8; 7; 6; 5; 4; 3; 2; 1i, a new office assistant is hired only once, in the first iteration. Given a list of ranks A3 D h5; 2; 1; 8; 4; 7; 10; 9; 3; 6i, a new office assistant is hired three times, upon interviewing the candidates with ranks 5, 8, and 10. Recalling that the cost of our algorithm depends on how many times we hire a new office assistant, we see that there are expensive inputs such as A1, inexpensive inputs such as A2, and moderately expensive inputs such as A3. Consider, on the other hand, the randomized algorithm that first permutes the candidates and then determines the best candidate. In this case, we randomize in the algorithm, not in the input distribution. Given a particular input, say A3 above, we cannot say how many times the maximum is updated, because this quantity differs with each run of the algorithm. The first time we run the algorithm on A3, it may produce the permutation A1 and perform 10 updates; but the second time we run the algorithm, we may produce the permutation A2 and perform only one update. The third time we run it, we may perform some other number of updates. Each time we run the algorithm, the execution depends on the random choices made and is likely to differ from the previous execution of the algorithm. For this algorithm and many other randomized algorithms, no particular input elicits its worst-case behavior. Even your worst enemy cannot produce a bad input array, since the random permutation makes the input order irrelevant. The randomized algorithm performs badly only if the random-number generator produces an “unlucky” permutation. For the hiring problem, the only change needed in the code is to randomly permute the array. 124 Chapter 5 Probabilistic Analysis and Randomized Algorithms R A N D O M IZED -H IR E- A S S I S TA N T .n/ 1 randomly permute the list of candidates 2 best D 0 // candidate 0 is a least-qualified dummy candidate 3 for i D 1 to n 4 interview candidate i 5 if candidate i is better than candidate best 6 best D i 7 hire candidate i With this simple change, we have created a randomized algorithm whose performance matches that obtained by assuming that the candidates were presented in a random order. Lemma 5.3 The expected hiring cost of the procedure RANDOMIZED-HIRE-ASSISTANT is O.ch ln n/. Proof After permuting the input array, we have achieved a situation identical to that of the probabilistic analysis of HIRE-ASSISTANT. Comparing Lemmas 5.2 and 5.3 highlights the difference between probabilistic analysis and randomized algorithms. In Lemma 5.2, we make an assumption about the input. In Lemma 5.3, we make no such assumption, although randomizing the input takes some additional time. To remain consistent with our terminology, we couched Lemma 5.2 in terms of the average-case hiring cost and Lemma 5.3 in terms of the expected hiring cost. In the remainder of this section, we discuss some issues involved in randomly permuting inputs. Randomly permuting arrays Many randomized algorithms randomize the input by permuting the given input array. (There are other ways to use randomization.) Here, we shall discuss two methods for doing so. We assume that we are given an array A which, without loss of generality, contains the elements 1 through n. Our goal is to produce a random permutation of the array. One common method is to assign each element AŒi of the array a random priority P Œi, and then sort the elements of A according to these priorities. For example, if our initial array is A D h1; 2; 3; 4i and we choose random priorities P D h36; 3; 62; 19i, we would produce an array B D h2; 4; 1; 3i, since the second priority is the smallest, followed by the fourth, then the first, and finally the third. We call this procedure PERMUTE-BY-SORTING: 5.3 Randomized algorithms 125 PERMUTE-BY-SORTING .A/ 1 n D A:length 2 let P Œ1 : : n be a new array 3 for i D 1 to n 4 P Œi  D RANDOM.1; n3/ 5 sort A, using P as sort keys Line 4 chooses a random number between 1 and n3. We use a range of 1 to n3 to make it likely that all the priorities in P are unique. (Exercise 5.3-5 asks you to prove that the probability that all entries are unique is at least 1 1=n, and Exercise 5.3-6 asks how to implement the algorithm even if two or more priorities are identical.) Let us assume that all the priorities are unique. The time-consuming step in this procedure is the sorting in line 5. As we shall see in Chapter 8, if we use a comparison sort, sorting takes .n lg n/ time. We can achieve this lower bound, since we have seen that merge sort takes ‚.n lg n/ time. (We shall see other comparison sorts that take ‚.n lg n/ time in Part II. Exercise 8.3-4 asks you to solve the very similar problem of sorting numbers in the range 0 to n3 1 in O.n/ time.) After sorting, if P Œi is the j th smallest priority, then AŒi lies in position j of the output. In this manner we obtain a permutation. It remains to prove that the procedure produces a uniform random permutation, that is, that the procedure is equally likely to produce every permutation of the numbers 1 through n. Lemma 5.4 Procedure PERMUTE-BY-SORTING produces a uniform random permutation of the input, assuming that all priorities are distinct. Proof We start by considering the particular permutation in which each element AŒi receives the ith smallest priority. We shall show that this permutation occurs with probability exactly 1=nŠ. For i D 1; 2; : : : ; n, let Ei be the event that element AŒi receives the ith smallest priority. Then we wish to compute the probability that for all i, event Ei occurs, which is Pr fE1 \ E2 \ E3 \ \ En 1 \ Eng : Using Exercise C.2-5, this probability is equal to Pr fE1g Pr fE2 j E1g Pr fE3 j E2 \ E1g Pr fE4 j E3 \ E2 \ E1g Pr fEi j Ei 1 \ Ei 2 \ \ E1g Pr fEn j En 1 \ \ E1g : We have that Pr fE1g D 1=n because it is the probability that one priority chosen randomly out of a set of n is the smallest priority. Next, we observe 126 Chapter 5 Probabilistic Analysis and Randomized Algorithms that Pr fE2 j E1g D 1=.n 1/ because given that element AŒ1 has the smallest priority, each of the remaining n 1 elements has an equal chance of hav- ing the second smallest priority. In general, for i D 2; 3; : : : ; n, we have that Pr fEi j Ei 1 \ Ei 2 \ \ E1g D 1=.n i C 1/, since, given that elements AŒ1 through AŒi 1 have the i 1 smallest priorities (in order), each of the remaining n .i 1/ elements has an equal chance of having the ith smallest priority. Thus, we have Pr fE1 \ E2 \ E3 \  Ã à 1 1 \ En 1 \ Eng D n n 1  Ã à 11 21 1 D; nŠ and we have shown that the probability of obtaining the identity permutation is 1=nŠ. We can extend this proof to work for any permutation of priorities. Consider any fixed permutation D h .1/; .2/; : : : ; .n/i of the set f1; 2; : : : ; ng. Let us denote by ri the rank of the priority assigned to element AŒi, where the element with the j th smallest priority has rank j . If we define Ei as the event in which element AŒi receives the .i /th smallest priority, or ri D .i /, the same proof still applies. Therefore, if we calculate the probability of obtaining any particular permutation, the calculation is identical to the one above, so that the probability of obtaining this permutation is also 1=nŠ. You might think that to prove that a permutation is a uniform random permutation, it suffices to show that, for each element AŒi, the probability that the element winds up in position j is 1=n. Exercise 5.3-4 shows that this weaker condition is, in fact, insufficient. A better method for generating a random permutation is to permute the given array in place. The procedure RANDOMIZE-IN-PLACE does so in O.n/ time. In its ith iteration, it chooses the element AŒi randomly from among elements AŒi through AŒn. Subsequent to the ith iteration, AŒi is never altered. RANDOMIZE-IN-PLACE .A/ 1 n D A:length 2 for i D 1 to n 3 swap AŒi  with AŒRANDOM.i; n/ We shall use a loop invariant to show that procedure RANDOMIZE-IN-PLACE produces a uniform random permutation. A k-permutation on a set of n elements is a sequence containing k of the n elements, with no repetitions. (See Appendix C.) There are nŠ=.n k/Š such possible k-permutations. 5.3 Randomized algorithms 127 Lemma 5.5 Procedure RANDOMIZE-IN-PLACE computes a uniform random permutation. Proof We use the following loop invariant: Just prior to the ith iteration of the for loop of lines 2–3, for each possible .i 1/-permutation of the n elements, the subarray AŒ1 : : i 1 contains this .i 1/-permutation with probability .n i C 1/Š=nŠ. We need to show that this invariant is true prior to the first loop iteration, that each iteration of the loop maintains the invariant, and that the invariant provides a useful property to show correctness when the loop terminates. Initialization: Consider the situation just before the first loop iteration, so that i D 1. The loop invariant says that for each possible 0-permutation, the subarray AŒ1 : : 0 contains this 0-permutation with probability .n i C 1/Š=nŠ D nŠ=nŠ D 1. The subarray AŒ1 : : 0 is an empty subarray, and a 0-permutation has no elements. Thus, AŒ1 : : 0 contains any 0-permutation with probability 1, and the loop invariant holds prior to the first iteration. Maintenance: We assume that just before the ith iteration, each possible .i 1/-permutation appears in the subarray AŒ1 : : i 1 with probability .n i C 1/Š=nŠ, and we shall show that after the ith iteration, each possible i-permutation appears in the subarray AŒ1 : : i with probability .n i/Š=nŠ. Incrementing i for the next iteration then maintains the loop invariant. Let us examine the ith iteration. Consider a particular i-permutation, and denote the elements in it by hx1; x2; : : : ; xi i. This permutation consists of an .i 1/-permutation hx1; : : : ; xi 1i followed by the value xi that the algorithm places in AŒi. Let E1 denote the event in which the first i 1 iterations have created the particular .i 1/-permutation hx1; : : : ; xi 1i in AŒ1 : : i 1. By the loop invariant, Pr fE1g D .n i C 1/Š=nŠ. Let E2 be the event that i th iteration puts xi in position AŒi . The i -permutation hx1; : : : ; xi i appears in AŒ1 : : i  precisely when both E1 and E2 occur, and so we wish to compute Pr fE2 \ E1g. Using equation (C.14), we have Pr fE2 \ E1g D Pr fE2 j E1g Pr fE1g : The probability Pr fE2 j E1g equals 1=.n i C1/ because in line 3 the algorithm chooses xi randomly from the n i C 1 values in positions AŒi : : n. Thus, we have 128 Chapter 5 Probabilistic Analysis and Randomized Algorithms Pr fE2 \ E1g D Pr fE2 j E1g Pr fE1g D 1 .n i C 1/Š n i C1 nŠ .n i/Š D : nŠ Termination: At termination, i D n C 1, and we have that the subarray AŒ1 : : n is a given n-permutation with probability .n .nC1/C1/=nŠ D 0Š=nŠ D 1=nŠ. Thus, RANDOMIZE-IN-PLACE produces a uniform random permutation. A randomized algorithm is often the simplest and most efficient way to solve a problem. We shall use randomized algorithms occasionally throughout this book. Exercises 5.3-1 Professor Marceau objects to the loop invariant used in the proof of Lemma 5.5. He questions whether it is true prior to the first iteration. He reasons that we could just as easily declare that an empty subarray contains no 0-permutations. Therefore, the probability that an empty subarray contains a 0-permutation should be 0, thus invalidating the loop invariant prior to the first iteration. Rewrite the procedure RANDOMIZE-IN-PLACE so that its associated loop invariant applies to a nonempty subarray prior to the first iteration, and modify the proof of Lemma 5.5 for your procedure. 5.3-2 Professor Kelp decides to write a procedure that produces at random any permutation besides the identity permutation. He proposes the following procedure: PERMUTE-WITHOUT-IDENTITY .A/ 1 n D A:length 2 for i D 1 to n 1 3 swap AŒi  with AŒRANDOM.i C 1; n/ Does this code do what Professor Kelp intends? 5.3-3 Suppose that instead of swapping element AŒi with a random element from the subarray AŒi : : n, we swapped it with a random element from anywhere in the array: 5.3 Randomized algorithms 129 P ER M U TE-W ITH - A L L .A/ 1 n D A:length 2 for i D 1 to n 3 swap AŒi  with AŒRANDOM.1; n/ Does this code produce a uniform random permutation? Why or why not? 5.3-4 Professor Armstrong suggests the following procedure for generating a uniform random permutation: P ER M U TE- B Y- C Y C LI C .A/ 1 n D A:length 2 let BŒ1 : : n be a new array 3 offset D RANDOM.1; n/ 4 for i D 1 to n 5 dest D i C offset 6 if dest > n 7 dest D dest n 8 BŒdest D AŒi 9 return B Show that each element AŒi has a 1=n probability of winding up in any particular position in B. Then show that Professor Armstrong is mistaken by showing that the resulting permutation is not uniformly random. 5.3-5 ? Prove that in the array P in procedure PERMUTE-BY-SORTING, the probability that all elements are unique is at least 1 1=n. 5.3-6 Explain how to implement the algorithm PERMUTE-BY-SORTING to handle the case in which two or more priorities are identical. That is, your algorithm should produce a uniform random permutation, even if two or more priorities are identical. 5.3-7 Suppose we want to create a random sample of the set f1; 2; 3; : : : ; ng, that is, an m-element subset S, where 0 Ä m Ä n, such that each m-subset is equally likely to be created. One way would be to set AŒi D i for i D 1; 2; 3; : : : ; n, call RANDOMIZE-IN-PLACE.A/, and then take just the first m array elements. This method would make n calls to the RANDOM procedure. If n is much larger than m, we can create a random sample with fewer calls to RANDOM. Show that 130 Chapter 5 Probabilistic Analysis and Randomized Algorithms the following recursive procedure returns a random m-subset S of f1; 2; 3; : : : ; ng, in which each m-subset is equally likely, while making only m calls to RANDOM: RANDOM-SAMPLE.m; n/ 1 if m == 0 2 return ; 3 else S D RANDOM-SAMPLE.m 1; n 1/ 4 i D RANDOM.1; n/ 5 if i 2 S 6 S D S [ fng 7 else S D S [ fig 8 return S ? 5.4 Probabilistic analysis and further uses of indicator random variables This advanced section further illustrates probabilistic analysis by way of four examples. The first determines the probability that in a room of k people, two of them share the same birthday. The second example examines what happens when we randomly toss balls into bins. The third investigates “streaks” of consecutive heads when we flip coins. The final example analyzes a variant of the hiring problem in which you have to make decisions without actually interviewing all the candidates. 5.4.1 The birthday paradox Our first example is the birthday paradox. How many people must there be in a room before there is a 50% chance that two of them were born on the same day of the year? The answer is surprisingly few. The paradox is that it is in fact far fewer than the number of days in a year, or even half the number of days in a year, as we shall see. To answer this question, we index the people in the room with the integers 1; 2; : : : ; k, where k is the number of people in the room. We ignore the issue of leap years and assume that all years have n D 365 days. For i D 1; 2; : : : ; k, let bi be the day of the year on which person i ’s birthday falls, where 1 Ä bi Ä n. We also assume that birthdays are uniformly distributed across the n days of the year, so that Pr fbi D rg D 1=n for i D 1; 2; : : : ; k and r D 1; 2; : : : ; n. The probability that two given people, say i and j , have matching birthdays depends on whether the random selection of birthdays is independent. We assume from now on that birthdays are independent, so that the probability that i’s birthday 5.4 Probabilistic analysis and further uses of indicator random variables 131 and j ’s birthday both fall on day r is Pr fbi D r and bj D rg D Pr fbi D rg Pr fbj D rg D 1=n2 : Thus, the probability that they both fall on the same day is Xn Pr fbi D bj g D Pr fbi D r and bj D rg r D1 Xn D .1=n2/ r D1 D 1=n : (5.6) More intuitively, once bi is chosen, the probability that bj is chosen to be the same day is 1=n. Thus, the probability that i and j have the same birthday is the same as the probability that the birthday of one of them falls on a given day. Notice, however, that this coincidence depends on the assumption that the birthdays are independent. We can analyze the probability of at least 2 out of k people having matching birthdays by looking at the complementary event. The probability that at least two of the birthdays match is 1 minus the probability that all the birthdays are different. The event that k people have distinct birthdays is \k Bk D Ai ; i D1 where Ai is the event that person i’s birthday is different from person j ’s for all j < i . Since we can write Bk D Ak \ Bk 1, we obtain from equation (C.16) the recurrence Pr fBkg D Pr fBk 1g Pr fAk j Bk 1g ; (5.7) where we take Pr fB1g D Pr fA1g D 1 as an initial condition. In other words, the probability that b1; b2; : : : ; bk are distinct birthdays is the probability that b1; b2; : : : ; bk 1 are distinct birthdays times the probability that bk ¤ bi for i D 1; 2; : : : ; k 1, given that b1; b2; : : : ; bk 1 are distinct. If b1; b2; : : : ; bk 1 are distinct, the conditional probability that bk ¤ bi for i D 1; 2; : : : ; k 1 is Pr fAk j Bk 1g D .n k C 1/=n, since out of the n days, n .k 1/ days are not taken. We iteratively apply the recurrence (5.7) to obtain 132 Chapter 5 Probabilistic Analysis and Randomized Algorithms Pr fBkg D Pr fBk 1g Pr fAk j Bk 1g D Pr fBk 2g Pr fAk 1 j Bk 2g Pr fAk j Bk 1g ::: D Pr fÂB1g Pr fÃAÂ2 j B1g ÃPr fAÂ3 j B2g PrÃfAk j Bk 1g n1 n2 n kC1 D1  n àn à  n à 1 2 k1 D1 1 1 1 : n n n Inequality (3.12), 1 C x Ä ex, gives us Pr fBkg Ä e 1=ne 2=n D e Pk 1 i D1 i=n D e k.k 1/=2n Ä 1=2 e .k 1/=n when k.k 1/=2n Ä ln.1=2/. The probability that all k birthdays are distinct is at most 1p=2 when k.k 1/ 2n ln 2 or, solving the quadratic equation, when k .1 C 1 C .8 ln 2/n/=2. For n D 365, we must have k 23. Thus, if at least 23 people are in a room, the probability is at least 1=2 that at least two people have the same birthday. On Mars, a year is 669 Martian days long; it therefore takes 31 Martians to get the same effect. An analysis using indicator random variables We can use indicator random variables to provide a simpler but approximate analysis of the birthday paradox. For each pair .i; j / of the k people in the room, we define the indicator random variable Xij , for 1 Ä i < j Ä k, by Xij D I(fperson i and person j have the same birthdayg D 1 if person i and person j have the same birthday ; 0 otherwise : By equation (5.6), the probability that two people have matching birthdays is 1=n, and thus by Lemma 5.1, we have E ŒXij  D Pr fperson i and person j have the same birthdayg D 1=n : Letting X be the random variable that counts the number of pairs of individuals having the same birthday, we have 5.4 Probabilistic analysis and further uses of indicator random variables 133 Xk Xk XD Xij : i D1 j Di C1 Taking expectations of both sides and applying linearity of expectation, we obtain " # Xk Xk E ŒX  D E Xij i D1 j Di C1 Xk Xk D E ŒXij  i D1!j Di C1 k1 D 2n k.k 1/ D : 2n When k.k 1/ 2n, therefore, the expected numpber of pairs of people with the same birthday is at least 1. Thus, if we have at least 2nC1 individuals in a room, we can expect at least two to have the same birthday. For n D 365, if k D 28, the expected number of pairs with the same birthday is .28 27/=.2 365/ 1:0356. Thus, with at least 28 people, we expect to find at least one matching pair of birthdays. On Mars, where a year is 669 Martian days long, we need at least 38 Martians. The first analysis, which used only probabilities, determined the number of people required for the probability to exceed 1=2 that a matching pair of birthdays exists, and the second analysis, which used indicator random variables, determined the number such that the expected number of matching birthdays is 1. Although the exact numpbers of people differ for the two situations, they are the same asymptotically: ‚. n/. 5.4.2 Balls and bins Consider a process in which we randomly toss identical balls into b bins, numbered 1; 2; : : : ; b. The tosses are independent, and on each toss the ball is equally likely to end up in any bin. The probability that a tossed ball lands in any given bin is 1=b. Thus, the ball-tossing process is a sequence of Bernoulli trials (see Appendix C.4) with a probability 1=b of success, where success means that the ball falls in the given bin. This model is particularly useful for analyzing hashing (see Chapter 11), and we can answer a variety of interesting questions about the ball-tossing process. (Problem C-1 asks additional questions about balls and bins.) 134 Chapter 5 Probabilistic Analysis and Randomized Algorithms How many balls fall in a given bin? The number of balls that fall in a given bin follows the binomial distribution b.kI n; 1=b/. If we toss n balls, equation (C.37) tells us that the expected number of balls that fall in the given bin is n=b. How many balls must we toss, on the average, until a given bin contains a ball? The number of tosses until the given bin receives a ball follows the geometric distribution with probability 1=b and, by equation (C.32), the expected number of tosses until success is 1=.1=b/ D b. How many balls must we toss until every bin contains at least one ball? Let us call a toss in which a ball falls into an empty bin a “hit.” We want to know the expected number n of tosses required to get b hits. Using the hits, we can partition the n tosses into stages. The ith stage consists of the tosses after the .i 1/st hit until the ith hit. The first stage consists of the first toss, since we are guaranteed to have a hit when all bins are empty. For each toss during the ith stage, i 1 bins contain balls and b i C 1 bins are empty. Thus, for each toss in the ith stage, the probability of obtaining a hit is .b i C 1/=b. Let ni required tdoengoettebthheitsnuismnbeDr oP f tobiDss1ensii.nEthaechi th stage. random Thus, the number of tosses variable ni has a geometric distribution with probability of success .b i C 1/=b and thus, by equation (C.32), we have b E Œni  D b : i C1 By linearity of expectation, we have "# Xb E Œn D E ni i D1 Xb D E Œni  i D1 Xb D b i D1 b i C1 Xb 1 Db i i D1 D b.ln b C O.1// (by equation (A.7)) . It therefore takes approximately b ln b tosses before we can expect that every bin has a ball. This problem is also known as the coupon collector’s problem, which says that a person trying to collect each of b different coupons expects to acquire approximately b ln b randomly obtained coupons in order to succeed. 5.4 Probabilistic analysis and further uses of indicator random variables 135 5.4.3 Streaks Suppose you flip a fair coin n times. What is the longest streak of consecutive heads that you expect to see? The answer is ‚.lg n/, as the following analysis shows. We first prove that the expected length of the longest streak of heads is O.lg n/. The probability that each coin flip is a head is 1=2. Let Aik be the event that a streak of heads of length at least k begins with the ith coin flip or, more precisely, the event that the k consecutive coin flips i; i C 1; : : : ; i C k 1 yield only heads, where 1 Ä k Ä n and 1 Ä i Ä n k C1. Since coin flips are mutually independent, for any given event Aik, the probability that all k flips are heads is Pr fAikg D 1=2k : (5.8) For k D 2 dlg ne, Pr fAi;2dlg neg D 1=22dlg ne Ä 1=22 lg n D 1=n2 ; and thus the probability that a streak of heads of length at least 2 dlg ne begins in position i is quite small. There are at most n 2 dlg ne C 1 positions where such a streak can begin. The probability that a streak of heads of length at least 2 dlg ne begins anywhere is therefore (n 2d[lg neC1 ) n 2X dlg neC1 Pr Ai;2dlg ne Ä 1=n2 i D1 i D1 Xn < 1=n2 i D1 D 1=n ; (5.9) since by Boole’s inequality (C.19), the probability of a union of events is at most the sum of the probabilities of the individual events. (Note that Boole’s inequality holds even for events such as these that are not independent.) We now use inequality (5.9) to bound the length of the longest streak. For j D 0; 1; 2; : : : ; n, let Lj be the event that the longest streak of heads has length exactly j , and let L be the length of the longest streak. By the definition of expected value, we have Xn E ŒL D j Pr fLj g : j D0 (5.10) 136 Chapter 5 Probabilistic Analysis and Randomized Algorithms We could try to evaluate this sum using upper bounds on each Pr fLj g similar to those computed in inequality (5.9). Unfortunately, this method would yield weak bounds. We can use some intuition gained by the above analysis to obtain a good bound, however. Informally, we observe that for no individual term in the summation in equation (5.10) are both the factors j and Pr fLj g large. Why? When j 2 dlg ne, then Pr fLj g is very small, and when j < 2 dlg ne, then j is fairly small. More formally, we note that the events Lj for j D 0; 1; : : : ; n are disjoint, Awanlhdseors,eonitsohtPeinpjgnDrot2hbdalagtbnP ielPitjnryDft0LhPjargt .faLBsjytrgeinaDekqo1uf,alwhiteyead(h5sa.v9oe)f,tlwhenaetghtPhavja2eDdtlPg0lenaejnDs1t2d2PlgrdnflegLPnjrgefLbÄejgg1i.n bestscore 4 bestscore D score.i/ 5 for i D k C 1 to n 6 if score.i/ > bestscore 7 return i 8 return n We wish to determine, for each possible value of k, the probability that we hire the most qualified applicant. We then choose the best possible k, and implement the strategy with that value. For the moment, assume that k is fixed. Let M.j / D max1ÄiÄj fscore.i /g denote the maximum score among ap- plicants 1 through j . Let S be the event that we succeed in choosing the best- qualified applicant, and let Si be the event that we succeed when the best-qualified applicant that Pr fS gisDthPe iniDth1 one interviewed. Since the various Si are Pr fSi g. Noting that we never succeed when disjoint, we have the best-qualified applicant is one of the first k, we have that Pr fSi g D 0 for i D 1; 2; : : : ; k. Thus, we obtain Xn Pr fSg D Pr fSi g : i DkC1 (5.12) We now compute Pr fSi g. In order to succeed when the best-qualified applicant is the ith one, two things must happen. First, the best-qualified applicant must be in position i, an event which we denote by Bi . Second, the algorithm must not select any of the applicants in positions k C 1 through i 1, which happens only if, for each j such that k C 1 Ä j Ä i 1, we find that score.j / < bestscore in line 6. (Because scores are unique, we can ignore the possibility of score.j / D bestscore.) In other words, all of the values score.k C 1/ through score.i 1/ must be less than M.k/; if any are greater than M.k/, we instead return the index of the first one that is greater. We use Oi to denote the event that none of the applicants in position k C 1 through i 1 are chosen. Fortunately, the two events Bi and Oi are independent. The event Oi depends only on the relative ordering of the values in positions 1 through i 1, whereas Bi depends only on whether the value in position i is greater than the values in all other positions. The ordering of the values in positions 1 through i 1 does not affect whether the value in position i is greater than all of them, and the value in position i does not affect the ordering of the values in positions 1 through i 1. Thus we can apply equation (C.15) to obtain 5.4 Probabilistic analysis and further uses of indicator random variables 141 Pr fSi g D Pr fBi \ Oi g D Pr fBi g Pr fOi g : The probability Pr fBi g is clearly 1=n, since the maximum is equally likely to be in any one of the n positions. For event Oi to occur, the maximum value in positions 1 through i 1, which is equally likely to be in any of these i 1 positions, must be in one of the first k positions. Consequently, Pr fOi g D k=.i 1/ and Pr fSi g D k=.n.i 1//. Using equation (5.12), we have Xn Pr fSg D Pr fSi g i DkC1 Xn k D n.i 1/ i DkC1 k Xn 1 D n i1 i DkC1 k X n 1 1 D : ni i Dk We approximate by integrals to bound this summation from above and below. By the inequalities (A.12), we have Z n1 X n 1 1 Z n 1 1 dx Ä Ä dx : kx i i Dk k1 x Evaluating these definite integrals gives us the bounds k .ln n ln k/ Ä Pr fSg Ä k .ln.n 1/ ln.k 1// ; n n which provide a rather tight bound for Pr fSg. Because we wish to maximize our probability of success, let us focus on choosing the value of k that maximizes the lower bound on Pr fSg. (Besides, the lower-bound expression is easier to maximize than the upper-bound expression.) Differentiating the expression .k=n/.ln n ln k/ with respect to k, we obtain 1 .ln n ln k 1/ : n Setting this derivative equal to 0, we see that we maximize the lower bound on the probability when ln k D ln n 1 D ln.n=e/ or, equivalently, when k D n=e. Thus, if we implement our strategy with k D n=e, we succeed in hiring our best-qualified applicant with probability at least 1=e. 142 Chapter 5 Probabilistic Analysis and Randomized Algorithms Exercises 5.4-1 How many people must there be in a room before the probability that someone has the same birthday as you do is at least 1=2? How many people must there be before the probability that at least two people have a birthday on July 4 is greater than 1=2? 5.4-2 Suppose that we toss balls into b bins until some bin contains two balls. Each toss is independent, and each ball is equally likely to end up in any bin. What is the expected number of ball tosses? 5.4-3 ? For the analysis of the birthday paradox, is it important that the birthdays be mutually independent, or is pairwise independence sufficient? Justify your answer. 5.4-4 ? How many people should be invited to a party in order to make it likely that there are three people with the same birthday? 5.4-5 ? What is the probability that a k-string over a set of size n forms a k-permutation? How does this question relate to the birthday paradox? 5.4-6 ? Suppose that n balls are tossed into n bins, where each toss is independent and the ball is equally likely to end up in any bin. What is the expected number of empty bins? What is the expected number of bins with exactly one ball? 5.4-7 ? Sharpen the lower bound on streak length by showing that in n flips of a fair coin, the probability is less than 1=n that no streak longer than lg n 2 lg lg n consecutive heads occurs. Problems for Chapter 5 143 Problems 5-1 Probabilistic counting With a b-bit counter, we can ordinarily only count up to 2b 1. With R. Morris’s probabilistic counting, we can count up to a much larger value at the expense of some loss of precision. We let a counter value of i represent a count of ni for i D 0; 1; : : : ; 2b 1, where the ni form an increasing sequence of nonnegative values. We assume that the initial value of the counter is 0, representing a count of n0 D 0. The INCREMENT operation works on a counter containing the value i in a probabilistic manner. If i D 2b 1, then the operation reports an overflow error. Otherwise, the INCREMENT operation increases the counter by 1 with probability 1=.niC1 ni /, and it leaves the counter unchanged with probability 1 1=.niC1 ni /. If we select ni D i for all i 0, then the counter is an ordinary one. More interesting situations arise if we select, say, ni D 2i 1 for i > 0 or ni D Fi (the ith Fibonacci number—see Section 3.2). For this problem, assume that n2b 1 is large enough that the probability of an overflow error is negligible. a. Show that the expected value represented by the counter after n INCREMENT operations have been performed is exactly n. b. The analysis of the variance of the count represented by the counter depends on the sequence of the ni . Let us consider a simple case: ni D 100i for all i 0. Estimate the variance in the value represented by the register after n INCREMENT operations have been performed. 5-2 Searching an unsorted array This problem examines three algorithms for searching for a value x in an unsorted array A consisting of n elements. Consider the following randomized strategy: pick a random index i into A. If AŒi D x, then we terminate; otherwise, we continue the search by picking a new random index into A. We continue picking random indices into A until we find an index j such that AŒj  D x or until we have checked every element of A. Note that we pick from the whole set of indices each time, so that we may examine a given element more than once. a. Write pseudocode for a procedure RANDOM-SEARCH to implement the strategy above. Be sure that your algorithm terminates when all indices into A have been picked. 144 Chapter 5 Probabilistic Analysis and Randomized Algorithms b. Suppose that there is exactly one index i such that AŒi D x. What is the expected number of indices into A that we must pick before we find x and RANDOM-SEARCH terminates? c. Generalizing your solution to part (b), suppose that there are k 1 indices i such that AŒi D x. What is the expected number of indices into A that we must pick before we find x and RANDOM-SEARCH terminates? Your answer should be a function of n and k. d. Suppose that there are no indices i such that AŒi D x. What is the expected number of indices into A that we must pick before we have checked all elements of A and RANDOM-SEARCH terminates? Now consider a deterministic linear search algorithm, which we refer to as DETERMINISTIC-SEARCH. Specifically, the algorithm searches A for x in order, considering AŒ1; AŒ2; AŒ3; : : : ; AŒn until either it finds AŒi D x or it reaches the end of the array. Assume that all possible permutations of the input array are equally likely. e. Suppose that there is exactly one index i such that AŒi D x. What is the average-case running time of DETERMINISTIC-SEARCH? What is the worstcase running time of DETERMINISTIC-SEARCH? f. Generalizing your solution to part (e), suppose that there are k 1 indices i such that AŒi  D x. What is the average-case running time of DETERMINISTICSEARCH? What is the worst-case running time of DETERMINISTIC-SEARCH? Your answer should be a function of n and k. g. Suppose that there are no indices i such that AŒi D x. What is the average-case running time of DETERMINISTIC-SEARCH? What is the worst-case running time of DETERMINISTIC-SEARCH? Finally, consider a randomized algorithm SCRAMBLE-SEARCH that works by first randomly permuting the input array and then running the deterministic linear search given above on the resulting permuted array. h. Letting k be the number of indices i such that AŒi D x, give the worst-case and expected running times of SCRAMBLE-SEARCH for the cases in which k D 0 and k D 1. Generalize your solution to handle the case in which k 1. i. Which of the three searching algorithms would you use? Explain your answer. Notes for Chapter 5 145 Chapter notes Bolloba´s [53], Hofri [174], and Spencer [321] contain a wealth of advanced probabilistic techniques. The advantages of randomized algorithms are discussed and surveyed by Karp [200] and Rabin [288]. The textbook by Motwani and Raghavan [262] gives an extensive treatment of randomized algorithms. Several variants of the hiring problem have been widely studied. These problems are more commonly referred to as “secretary problems.” An example of work in this area is the paper by Ajtai, Meggido, and Waarts [11]. II Sorting and Order Statistics Introduction This part presents several algorithms that solve the following sorting problem: Input: A sequence of n numbers ha1; a2; : : : ; ani. Output: A permutation (reordering) ha10 ; a20 ; : : : ; an0 i of the input sequence such that a10 Ä a20 Ä Ä an0 . The input sequence is usually an n-element array, although it may be represented in some other fashion, such as a linked list. The structure of the data In practice, the numbers to be sorted are rarely isolated values. Each is usually part of a collection of data called a record. Each record contains a key, which is the value to be sorted. The remainder of the record consists of satellite data, which are usually carried around with the key. In practice, when a sorting algorithm permutes the keys, it must permute the satellite data as well. If each record includes a large amount of satellite data, we often permute an array of pointers to the records rather than the records themselves in order to minimize data movement. In a sense, it is these implementation details that distinguish an algorithm from a full-blown program. A sorting algorithm describes the method by which we determine the sorted order, regardless of whether we are sorting individual numbers or large records containing many bytes of satellite data. Thus, when focusing on the problem of sorting, we typically assume that the input consists only of numbers. Translating an algorithm for sorting numbers into a program for sorting records 148 Part II Sorting and Order Statistics is conceptually straightforward, although in a given engineering situation other subtleties may make the actual programming task a challenge. Why sorting? Many computer scientists consider sorting to be the most fundamental problem in the study of algorithms. There are several reasons: Sometimes an application inherently needs to sort information. For example, in order to prepare customer statements, banks need to sort checks by check number. Algorithms often use sorting as a key subroutine. For example, a program that renders graphical objects which are layered on top of each other might have to sort the objects according to an “above” relation so that it can draw these objects from bottom to top. We shall see numerous algorithms in this text that use sorting as a subroutine. We can draw from among a wide variety of sorting algorithms, and they employ a rich set of techniques. In fact, many important techniques used throughout algorithm design appear in the body of sorting algorithms that have been developed over the years. In this way, sorting is also a problem of historical interest. We can prove a nontrivial lower bound for sorting (as we shall do in Chapter 8). Our best upper bounds match the lower bound asymptotically, and so we know that our sorting algorithms are asymptotically optimal. Moreover, we can use the lower bound for sorting to prove lower bounds for certain other problems. Many engineering issues come to the fore when implementing sorting algorithms. The fastest sorting program for a particular situation may depend on many factors, such as prior knowledge about the keys and satellite data, the memory hierarchy (caches and virtual memory) of the host computer, and the software environment. Many of these issues are best dealt with at the algorithmic level, rather than by “tweaking” the code. Sorting algorithms We introduced two algorithms that sort n real numbers in Chapter 2. Insertion sort takes ‚.n2/ time in the worst case. Because its inner loops are tight, however, it is a fast in-place sorting algorithm for small input sizes. (Recall that a sorting algorithm sorts in place if only a constant number of elements of the input array are ever stored outside the array.) Merge sort has a better asymptotic running time, ‚.n lg n/, but the MERGE procedure it uses does not operate in place. Part II Sorting and Order Statistics 149 In this part, we shall introduce two more algorithms that sort arbitrary real numbers. Heapsort, presented in Chapter 6, sorts n numbers in place in O.n lg n/ time. It uses an important data structure, called a heap, with which we can also implement a priority queue. Quicksort, in Chapter 7, also sorts n numbers in place, but its worst-case running time is ‚.n2/. Its expected running time is ‚.n lg n/, however, and it generally outperforms heapsort in practice. Like insertion sort, quicksort has tight code, and so the hidden constant factor in its running time is small. It is a popular algorithm for sorting large input arrays. Insertion sort, merge sort, heapsort, and quicksort are all comparison sorts: they determine the sorted order of an input array by comparing elements. Chapter 8 begins by introducing the decision-tree model in order to study the performance limitations of comparison sorts. Using this model, we prove a lower bound of .n lg n/ on the worst-case running time of any comparison sort on n inputs, thus showing that heapsort and merge sort are asymptotically optimal comparison sorts. Chapter 8 then goes on to show that we can beat this lower bound of .n lg n/ if we can gather information about the sorted order of the input by means other than comparing elements. The counting sort algorithm, for example, assumes that the input numbers are in the set f0; 1; : : : ; kg. By using array indexing as a tool for determining relative order, counting sort can sort n numbers in ‚.k C n/ time. Thus, when k D O.n/, counting sort runs in time that is linear in the size of the input array. A related algorithm, radix sort, can be used to extend the range of counting sort. If there are n integers to sort, each integer has d digits, and each digit can take on up to k possible values, then radix sort can sort the numbers in ‚.d.n C k// time. When d is a constant and k is O.n/, radix sort runs in linear time. A third algorithm, bucket sort, requires knowledge of the probabilistic distribution of numbers in the input array. It can sort n real numbers uniformly distributed in the half-open interval Œ0; 1/ in average-case O.n/ time. The following table summarizes the running times of the sorting algorithms from Chapters 2 and 6–8. As usual, n denotes the number of items to sort. For counting sort, the items to sort are integers in the set f0; 1; : : : ; kg. For radix sort, each item is a d -digit number, where each digit takes on k possible values. For bucket sort, we assume that the keys are real numbers uniformly distributed in the half-open interval Œ0; 1/. The rightmost column gives the average-case or expected running time, indicating which it gives when it differs from the worst-case running time. We omit the average-case running time of heapsort because we do not analyze it in this book. 150 Part II Sorting and Order Statistics Algorithm Insertion sort Merge sort Heapsort Quicksort Counting sort Radix sort Bucket sort Worst-case running time ‚.n2/ ‚.n lg n/ O.n lg n/ ‚.n2/ ‚.k C n/ ‚.d.n C k// ‚.n2/ Average-case/expected running time ‚.n2/ ‚.n lg n/ — ‚.n lg n/ (expected) ‚.k C n/ ‚.d.n C k// ‚.n/ (average-case) Order statistics The ith order statistic of a set of n numbers is the ith smallest number in the set. We can, of course, select the ith order statistic by sorting the input and indexing the ith element of the output. With no assumptions about the input distribution, this method runs in .n lg n/ time, as the lower bound proved in Chapter 8 shows. In Chapter 9, we show that we can find the ith smallest element in O.n/ time, even when the elements are arbitrary real numbers. We present a randomized algorithm with tight pseudocode that runs in ‚.n2/ time in the worst case, but whose expected running time is O.n/. We also give a more complicated algorithm that runs in O.n/ worst-case time. Background Although most of this part does not rely on difficult mathematics, some sections do require mathematical sophistication. In particular, analyses of quicksort, bucket sort, and the order-statistic algorithm use probability, which is reviewed in Appendix C, and the material on probabilistic analysis and randomized algorithms in Chapter 5. The analysis of the worst-case linear-time algorithm for order statistics involves somewhat more sophisticated mathematics than the other worst-case analyses in this part. 6 Heapsort In this chapter, we introduce another sorting algorithm: heapsort. Like merge sort, but unlike insertion sort, heapsort’s running time is O.n lg n/. Like insertion sort, but unlike merge sort, heapsort sorts in place: only a constant number of array elements are stored outside the input array at any time. Thus, heapsort combines the better attributes of the two sorting algorithms we have already discussed. Heapsort also introduces another algorithm design technique: using a data structure, in this case one we call a “heap,” to manage information. Not only is the heap data structure useful for heapsort, but it also makes an efficient priority queue. The heap data structure will reappear in algorithms in later chapters. The term “heap” was originally coined in the context of heapsort, but it has since come to refer to “garbage-collected storage,” such as the programming languages Java and Lisp provide. Our heap data structure is not garbage-collected storage, and whenever we refer to heaps in this book, we shall mean a data structure rather than an aspect of garbage collection. 6.1 Heaps The (binary) heap data structure is an array object that we can view as a nearly complete binary tree (see Section B.5.3), as shown in Figure 6.1. Each node of the tree corresponds to an element of the array. The tree is completely filled on all levels except possibly the lowest, which is filled from the left up to a point. An array A that represents a heap is an object with two attributes: A:length, which (as usual) gives the number of elements in the array, and A:heap-size, which represents how many elements in the heap are stored within array A. That is, although AŒ1 : : A:length may contain numbers, only the elements in AŒ1 : : A:heap-size, where 0 Ä A:heap-size Ä A:length, are valid elements of the heap. The root of the tree is AŒ1, and given the index i of a node, we can easily compute the indices of its parent, left child, and right child: 152 Chapter 6 Heapsort 1 16 2 3 14 10 4 5 6 7 8 79 3 8 9 10 241 (a) 1 2 3 4 5 6 7 8 9 10 16 14 10 8 7 9 3 2 4 1 (b) Figure 6.1 A max-heap viewed as (a) a binary tree and (b) an array. The number within the circle at each node in the tree is the value stored at that node. The number above a node is the corresponding index in the array. Above and below the array are lines showing parent-child relationships; parents are always to the left of their children. The tree has height three; the node at index 4 (with value 8) has height one. PARENT.i / 1 return bi=2c LEFT.i / 1 return 2i RIGHT.i / 1 return 2i C 1 On most computers, the LEFT procedure can compute 2i in one instruction by simply shifting the binary representation of i left by one bit position. Similarly, the RIGHT procedure can quickly compute 2i C 1 by shifting the binary representation of i left by one bit position and then adding in a 1 as the low-order bit. The PARENT procedure can compute bi=2c by shifting i right one bit position. Good implementations of heapsort often implement these procedures as “macros” or “inline” procedures. There are two kinds of binary heaps: max-heaps and min-heaps. In both kinds, the values in the nodes satisfy a heap property, the specifics of which depend on the kind of heap. In a max-heap, the max-heap property is that for every node i other than the root, AŒPARENT.i / AŒi  ; that is, the value of a node is at most the value of its parent. Thus, the largest element in a max-heap is stored at the root, and the subtree rooted at a node contains 6.1 Heaps 153 values no larger than that contained at the node itself. A min-heap is organized in the opposite way; the min-heap property is that for every node i other than the root, AŒPARENT.i / Ä AŒi  : The smallest element in a min-heap is at the root. For the heapsort algorithm, we use max-heaps. Min-heaps commonly imple- ment priority queues, which we discuss in Section 6.5. We shall be precise in specifying whether we need a max-heap or a min-heap for any particular application, and when properties apply to either max-heaps or min-heaps, we just use the term “heap.” Viewing a heap as a tree, we define the height of a node in a heap to be the number of edges on the longest simple downward path from the node to a leaf, and we define the height of the heap to be the height of its root. Since a heap of n elements is based on a complete binary tree, its height is ‚.lg n/ (see Exercise 6.1-2). We shall see that the basic operations on heaps run in time at most proportional to the height of the tree and thus take O.lg n/ time. The remainder of this chapter presents some basic procedures and shows how they are used in a sorting algorithm and a priority-queue data structure. The MAX-HEAPIFY procedure, which runs in O.lg n/ time, is the key to maintaining the max-heap property. The BUILD-MAX-HEAP procedure, which runs in linear time, produces a maxheap from an unordered input array. The HEAPSORT procedure, which runs in O.n lg n/ time, sorts an array in place. The MAX-HEAP-INSERT, HEAP-EXTRACT-MAX, HEAP-INCREASE-KEY, and HEAP-MAXIMUM procedures, which run in O.lg n/ time, allow the heap data structure to implement a priority queue. Exercises 6.1-1 What are the minimum and maximum numbers of elements in a heap of height h? 6.1-2 Show that an n-element heap has height blg nc. 6.1-3 Show that in any subtree of a max-heap, the root of the subtree contains the largest value occurring anywhere in that subtree. 154 Chapter 6 Heapsort 6.1-4 Where in a max-heap might the smallest element reside, assuming that all elements are distinct? 6.1-5 Is an array that is in sorted order a min-heap? 6.1-6 Is the array with values h23; 17; 14; 6; 13; 10; 1; 5; 7; 12i a max-heap? 6.1-7 Show that, with the array representation for storing an n-element heap, the leaves are the nodes indexed by bn=2c C 1; bn=2c C 2; : : : ; n. 6.2 Maintaining the heap property In order to maintain the max-heap property, we call the procedure MAX-HEAPIFY. Its inputs are an array A and an index i into the array. When it is called, MAXHEAPIFY assumes that the binary trees rooted at LEFT.i / and RIGHT.i / are maxheaps, but that AŒi might be smaller than its children, thus violating the max-heap property. MAX-HEAPIFY lets the value at AŒi  “float down” in the max-heap so that the subtree rooted at index i obeys the max-heap property. MAX-HEAPIFY.A; i / 1 l D LEFT.i / 2 r D RIGHT.i / 3 if l Ä A:heap-size and AŒl > AŒi 4 largest D l 5 else largest D i 6 if r Ä A:heap-size and AŒr > AŒlargest 7 largest D r 8 if largest ¤ i 9 exchange AŒi with AŒlargest 10 MAX-HEAPIFY.A; largest/ Figure 6.2 illustrates the action of MAX-HEAPIFY. At each step, the largest of the elements AŒi , AŒLEFT.i /, and AŒRIGHT.i / is determined, and its index is stored in largest. If AŒi is largest, then the subtree rooted at node i is already a max-heap and the procedure terminates. Otherwise, one of the two children has the largest element, and AŒi is swapped with AŒlargest, which causes node i and its 6.2 Maintaining the heap property 1 16 2 3 i4 10 4 5 6 7 14 79 3 8 9 10 281 (a) 1 16 2 3 14 10 4 5 6 7 8 79 3 8 9 i 10 241 (c) 155 1 16 2 3 14 10 4 5 6 7 i4 79 3 8 9 10 281 (b) Figure 6.2 The action of MAX-HEAPIFY.A; 2/, where A: heap-size D 10. (a) The initial configuration, with AŒ2 at node i D 2 violating the max-heap property since it is not larger than both children. The max-heap property is restored for node 2 in (b) by exchanging AŒ2 with AŒ4, which destroys the max-heap property for node 4. The recursive call MAX-HEAPIFY.A; 4/ now has i D 4. After swapping AŒ4 with AŒ9, as shown in (c), node 4 is fixed up, and the recursive call MAX-HEAPIFY.A; 9/ yields no further change to the data structure. children to satisfy the max-heap property. The node indexed by largest, however, now has the original value AŒi, and thus the subtree rooted at largest might violate the max-heap property. Consequently, we call MAX-HEAPIFY recursively on that subtree. The running time of MAX-HEAPIFY on a subtree of size n rooted at a given node i is the ‚.1/ time to fix up the relationships among the elements AŒi, AŒLEFT.i /, and AŒRIGHT.i /, plus the time to run MAX-HEAPIFY on a subtree rooted at one of the children of node i (assuming that the recursive call occurs). The children’s subtrees each have size at most 2n=3—the worst case occurs when the bottom level of the tree is exactly half full—and therefore we can describe the running time of MAX-HEAPIFY by the recurrence T .n/ Ä T .2n=3/ C ‚.1/ : 156 Chapter 6 Heapsort The solution to this recurrence, by case 2 of the master theorem (Theorem 4.1), is T .n/ D O.lg n/. Alternatively, we can characterize the running time of MAXHEAPIFY on a node of height h as O.h/. Exercises 6.2-1 Using Figure 6.2 as a model, illustrate the operation of MAX-HEAPIFY.A; 3/ on the array A D h27; 17; 3; 16; 13; 10; 1; 5; 7; 12; 4; 8; 9; 0i. 6.2-2 Starting with the procedure MAX-HEAPIFY, write pseudocode for the procedure MIN-HEAPIFY.A; i /, which performs the corresponding manipulation on a minheap. How does the running time of MIN-HEAPIFY compare to that of MAXHEAPIFY? 6.2-3 What is the effect of calling MAX-HEAPIFY.A; i / when the element AŒi  is larger than its children? 6.2-4 What is the effect of calling MAX-HEAPIFY.A; i / for i > A:heap-size=2? 6.2-5 The code for MAX-HEAPIFY is quite efficient in terms of constant factors, except possibly for the recursive call in line 10, which might cause some compilers to produce inefficient code. Write an efficient MAX-HEAPIFY that uses an iterative control construct (a loop) instead of recursion. 6.2-6 Show that the worst-case running time of MAX-HEAPIFY on a heap of size n is .lg n/. (Hint: For a heap with n nodes, give node values that cause MAXHEAPIFY to be called recursively at every node on a simple path from the root down to a leaf.) 6.3 Building a heap We can use the procedure MAX-HEAPIFY in a bottom-up manner to convert an array AŒ1 : : n, where n D A:length, into a max-heap. By Exercise 6.1-7, the elements in the subarray AŒ.bn=2cC1/ : : n are all leaves of the tree, and so each is 6.3 Building a heap 157 a 1-element heap to begin with. The procedure BUILD-MAX-HEAP goes through the remaining nodes of the tree and runs MAX-HEAPIFY on each one. BU ILD -M A X -H EA P .A/ 1 A:heap-size D A:length 2 for i D bA:length=2c downto 1 3 MAX-HEAPIFY.A; i / Figure 6.3 shows an example of the action of BUILD-MAX-HEAP. To show why BUILD-MAX-HEAP works correctly, we use the following loop invariant: At the start of each iteration of the for loop of lines 2–3, each node i C 1; i C 2; : : : ; n is the root of a max-heap. We need to show that this invariant is true prior to the first loop iteration, that each iteration of the loop maintains the invariant, and that the invariant provides a useful property to show correctness when the loop terminates. Initialization: Prior to the first iteration of the loop, i D bn=2c. Each node bn=2c C 1; bn=2c C 2; : : : ; n is a leaf and is thus the root of a trivial max-heap. Maintenance: To see that each iteration maintains the loop invariant, observe that the children of node i are numbered higher than i. By the loop invariant, therefore, they are both roots of max-heaps. This is precisely the condition required for the call MAX-HEAPIFY.A; i / to make node i a max-heap root. Moreover, the MAX-HEAPIFY call preserves the property that nodes i C 1; i C 2; : : : ; n are all roots of max-heaps. Decrementing i in the for loop update reestablishes the loop invariant for the next iteration. Termination: At termination, i D 0. By the loop invariant, each node 1; 2; : : : ; n is the root of a max-heap. In particular, node 1 is. We can compute a simple upper bound on the running time of BUILD-MAX- HEAP as follows. Each call to MAX-HEAPIFY costs O.lg n/ time, and BUILD- MAX-HEAP makes O.n/ such calls. Thus, the running time is O.n lg n/. This upper bound, though correct, is not asymptotically tight. We can derive a tighter bound by observing that the time for MAX-HEAPIFY to run at a node varies with the height of the node in the tree, and the heights of most nhaosdehseiagrhetsbmlgalnl.cO(sueretiEgxheterrciasnea6ly.1s-is2)realniedsaotnmthoestp˙ronp=e2rhtCie1s that an nodes n-element heap of any height h (see Exercise 6.3-3). The time required by MAX-HEAPIFY when called on a node of height h is O.h/, and so we can express the total cost of BUILD-MAX-HEAP as being bounded from above by 158 Chapter 6 Heapsort A 4 1 3 2 16 9 10 14 8 7 1 4 2 1 4 5 6 2 i 16 9 8 9 10 14 8 7 (a) 3 3 7 10 4 14 8 9 28 1 4 2 1 5 6 16 9 10 7 (c) 3 3i 7 10 1 i4 2 3 16 10 4 5 6 7 14 79 3 8 9 10 281 (e) 1 4 2 1 4 5 6 i2 16 9 8 9 10 14 8 7 (b) 3 3 7 10 1 4 2 3 i1 10 4 5 6 7 14 16 9 3 8 9 10 287 (d) 1 16 2 3 14 10 4 5 6 7 8 79 3 8 9 10 241 (f) Figure 6.3 The operation of BUILD-MAX-HEAP, showing the data structure before the call to MAX-HEAPIFY in line 3 of BUILD-MAX-HEAP. (a) A 10-element input array A and the binary tree it represents. The figure shows that the loop index i refers to node 5 before the call MAX-HEAPIFY.A; i/. (b) The data structure that results. The loop index i for the next iteration refers to node 4. (c)–(e) Subsequent iterations of the for loop in BUILD-MAX-HEAP. Observe that whenever MAX-HEAPIFY is called on a node, the two subtrees of that node are both max-heaps. (f) The max-heap after BUILD-MAX-HEAP finishes. 6.4 The heapsort algorithm 159 bX lg nc l n m O.h/ D O 2hC1 hD0 bX lg nc ! h n : 2h hD0 We evalaute the last summation by substituting x D 1=2 in the formula (A.8), yielding X 1 h 1=2 D 2h .1 1=2/2 hD0 D 2: Thus, we can bound the running time of BUILD-MAX-HEAP as bX lg nc ! h ! X 1 h On DO n 2h 2h hD0 hD0 D O.n/ : Hence, we can build a max-heap from an unordered array in linear time. We can build a min-heap by the procedure BUILD-MIN-HEAP, which is the same as BUILD-MAX-HEAP but with the call to MAX-HEAPIFY in line 3 replaced by a call to MIN-HEAPIFY (see Exercise 6.2-2). BUILD-MIN-HEAP produces a min-heap from an unordered linear array in linear time. Exercises 6.3-1 Using Figure 6.3 as a model, illustrate the operation of BUILD-MAX-HEAP on the array A D h5; 3; 17; 10; 84; 19; 6; 22; 9i. 6.3-2 Why do we want the loop index i in line 2 of BUILD-MAX-HEAP to decrease from bA:length=2c to 1 rather than increase from 1 to bA:length=2c? 6.3-3 Show that there are at most ˙n=2hC1 nodes of height h in any n-element heap. 6.4 The heapsort algorithm The heapsort algorithm starts by using BUILD-MAX-HEAP to build a max-heap on the input array AŒ1 : : n, where n D A:length. Since the maximum element of the array is stored at the root AŒ1, we can put it into its correct final position 160 Chapter 6 Heapsort by exchanging it with AŒn. If we now discard node n from the heap—and we can do so by simply decrementing A:heap-size—we observe that the children of the root remain max-heaps, but the new root element might violate the max-heap property. All we need to do to restore the max-heap property, however, is call MAX-HEAPIFY.A; 1/, which leaves a max-heap in AŒ1 : : n 1. The heapsort algorithm then repeats this process for the max-heap of size n 1 down to a heap of size 2. (See Exercise 6.4-2 for a precise loop invariant.) HEAPSORT.A/ 1 BUILD-MAX-HEAP.A/ 2 for i D A:length downto 2 3 exchange AŒ1 with AŒi 4 A:heap-size D A:heap-size 1 5 MAX-HEAPIFY.A; 1/ Figure 6.4 shows an example of the operation of HEAPSORT after line 1 has built the initial max-heap. The figure shows the max-heap before the first iteration of the for loop of lines 2–5 and after each iteration. The HEAPSORT procedure takes time O.n lg n/, since the call to BUILD-MAXHEAP takes time O.n/ and each of the n 1 calls to MAX-HEAPIFY takes time O.lg n/. Exercises 6.4-1 Using Figure 6.4 as a model, illustrate the operation of HEAPSORT on the array A D h5; 13; 2; 25; 7; 17; 20; 8; 4i. 6.4-2 Argue the correctness of HEAPSORT using the following loop invariant: At the start of each iteration of the for loop of lines 2–5, the subarray AŒ1 : : i is a max-heap containing the i smallest elements of AŒ1 : : n, and the subarray AŒi C 1 : : n contains the n i largest elements of AŒ1 : : n, sorted. 6.4-3 What is the running time of HEAPSORT on an array A of length n that is already sorted in increasing order? What about decreasing order? 6.4-4 Show that the worst-case running time of HEAPSORT is .n lg n/. 6.4 The heapsort algorithm 161 16 14 10 8 79 3 241 (a) 14 8 10 4 79 3 2 1 16 i (b) 10 8 9 4 71 3 i 2 14 16 (c) 9 8 3 4 71 2 i 10 14 16 (d) 8 7 4 21 10 14 16 (e) 3 i9 7 4 3 1 2 8i 9 10 14 16 (f) 4 2 3 1 i7 8 9 10 14 16 (g) 3 2 1 i4 78 9 10 14 16 (h) 2 1 4 78 10 14 16 (i) 3i 9 1 i2 3 4 78 9 10 14 16 (j) A 1 2 3 4 7 8 9 10 14 16 (k) Figure 6.4 The operation of HEAPSORT. (a) The max-heap data structure just after BUILD-MAXHEAP has built it in line 1. (b)–(j) The max-heap just after each call of MAX-HEAPIFY in line 5, showing the value of i at that time. Only lightly shaded nodes remain in the heap. (k) The resulting sorted array A. 162 Chapter 6 Heapsort 6.4-5 ? Show that when all elements are distinct, the best-case running time of HEAPSORT is .n lg n/. 6.5 Priority queues Heapsort is an excellent algorithm, but a good implementation of quicksort, presented in Chapter 7, usually beats it in practice. Nevertheless, the heap data structure itself has many uses. In this section, we present one of the most popular applications of a heap: as an efficient priority queue. As with heaps, priority queues come in two forms: max-priority queues and min-priority queues. We will focus here on how to implement max-priority queues, which are in turn based on maxheaps; Exercise 6.5-3 asks you to write the procedures for min-priority queues. A priority queue is a data structure for maintaining a set S of elements, each with an associated value called a key. A max-priority queue supports the following operations: INSERT.S; x/ inserts the element x into the set S, which is equivalent to the operation S D S [ fxg. MAXIMUM.S/ returns the element of S with the largest key. EXTRACT-MAX.S/ removes and returns the element of S with the largest key. INCREASE-KEY.S; x; k/ increases the value of element x’s key to the new value k, which is assumed to be at least as large as x’s current key value. Among their other applications, we can use max-priority queues to schedule jobs on a shared computer. The max-priority queue keeps track of the jobs to be performed and their relative priorities. When a job is finished or interrupted, the scheduler selects the highest-priority job from among those pending by calling EXTRACT-MAX. The scheduler can add a new job to the queue at any time by calling INSERT. Alternatively, a min-priority queue supports the operations INSERT, MINIMUM, EXTRACT-MIN, and DECREASE-KEY. A min-priority queue can be used in an event-driven simulator. The items in the queue are events to be simulated, each with an associated time of occurrence that serves as its key. The events must be simulated in order of their time of occurrence, because the simulation of an event can cause other events to be simulated in the future. The simulation program calls EXTRACT-MIN at each step to choose the next event to simulate. As new events are produced, the simulator inserts them into the min-priority queue by calling INSERT. 6.5 Priority queues 163 We shall see other uses for min-priority queues, highlighting the DECREASE-KEY operation, in Chapters 23 and 24. Not surprisingly, we can use a heap to implement a priority queue. In a given application, such as job scheduling or event-driven simulation, elements of a priority queue correspond to objects in the application. We often need to determine which application object corresponds to a given priority-queue element, and vice versa. When we use a heap to implement a priority queue, therefore, we often need to store a handle to the corresponding application object in each heap element. The exact makeup of the handle (such as a pointer or an integer) depends on the application. Similarly, we need to store a handle to the corresponding heap element in each application object. Here, the handle would typically be an array index. Because heap elements change locations within the array during heap operations, an actual implementation, upon relocating a heap element, would also have to update the array index in the corresponding application object. Because the details of accessing application objects depend heavily on the application and its implementation, we shall not pursue them here, other than noting that in practice, these handles do need to be correctly maintained. Now we discuss how to implement the operations of a max-priority queue. The procedure HEAP-MAXIMUM implements the MAXIMUM operation in ‚.1/ time. H E A P - M A X I M U M .A/ 1 return AŒ1 The procedure HEAP-EXTRACT-MAX implements the EXTRACT-MAX operation. It is similar to the for loop body (lines 3–5) of the HEAPSORT procedure. H EA P-E X TR AC T-M A X .A/ 1 if A:heap-size < 1 2 error “heap underflow” 3 max D AŒ1 4 AŒ1 D AŒA:heap-size 5 A:heap-size D A:heap-size 1 6 MAX-HEAPIFY.A; 1/ 7 return max The running time of HEAP-EXTRACT-MAX is O.lg n/, since it performs only a constant amount of work on top of the O.lg n/ time for MAX-HEAPIFY. The procedure HEAP-INCREASE-KEY implements the INCREASE-KEY operation. An index i into the array identifies the priority-queue element whose key we wish to increase. The procedure first updates the key of element AŒi to its new value. Because increasing the key of AŒi might violate the max-heap property, 164 Chapter 6 Heapsort the procedure then, in a manner reminiscent of the insertion loop (lines 5–7) of INSERTION-SORT from Section 2.1, traverses a simple path from this node toward the root to find a proper place for the newly increased key. As HEAP-INCREASEKEY traverses this path, it repeatedly compares an element to its parent, exchanging their keys and continuing if the element’s key is larger, and terminating if the element’s key is smaller, since the max-heap property now holds. (See Exercise 6.5-5 for a precise loop invariant.) HEAP-INCREASE-KEY .A; i; key/ 1 if key < AŒi 2 error “new key is smaller than current key” 3 AŒi D key 4 while i > 1 and AŒPARENT.i / < AŒi  5 exchange AŒi  with AŒPARENT.i / 6 i D PARENT.i / Figure 6.5 shows an example of a HEAP-INCREASE-KEY operation. The running time of HEAP-INCREASE-KEY on an n-element heap is O.lg n/, since the path traced from the node updated in line 3 to the root has length O.lg n/. The procedure MAX-HEAP-INSERT implements the INSERT operation. It takes as an input the key of the new element to be inserted into max-heap A. The procedure first expands the max-heap by adding to the tree a new leaf whose key is 1. Then it calls HEAP-INCREASE-KEY to set the key of this new node to its correct value and maintain the max-heap property. MAX-HEAP-INSERT.A; key/ 1 A:heap-size D A:heap-size C 1 2 AŒA:heap-size D 1 3 HEAP-INCREASE-KEY.A; A:heap-size; key/ The running time of MAX-HEAP-INSERT on an n-element heap is O.lg n/. In summary, a heap can support any priority-queue operation on a set of size n in O.lg n/ time. Exercises 6.5-1 Illustrate the operation of HEAP-EXTRACT-MAX on the heap A D h15; 13; 9; 5; 12; 8; 7; 4; 0; 6; 2; 1i. 6.5 Priority queues 16 14 10 8 79 3 i 241 (a) 165 16 14 10 8 79 3 i 2 15 1 (b) 16 14 i 15 7 10 9 3 281 (c) 16 i 15 10 14 79 3 281 (d) Figure 6.5 The operation of HEAP-INCREASE-KEY. (a) The max-heap of Figure 6.4(a) with a node whose index is i heavily shaded. (b) This node has its key increased to 15. (c) After one iteration of the while loop of lines 4–6, the node and its parent have exchanged keys, and the index i moves up to the parent. (d) The max-heap after one more iteration of the while loop. At this point, AŒPARENT.i/ AŒi. The max-heap property now holds and the procedure terminates. 6.5-2 Illustrate the operation of MAX-HEAP-INSERT.A; 10/ on the heap A D h15; 13; 9; 5; 12; 8; 7; 4; 0; 6; 2; 1i. 6.5-3 Write pseudocode for the procedures HEAP-MINIMUM, HEAP-EXTRACT-MIN, HEAP-DECREASE-KEY, and MIN-HEAP-INSERT that implement a min-priority queue with a min-heap. 6.5-4 Why do we bother setting the key of the inserted node to 1 in line 2 of MAXHEAP-INSERT when the next thing we do is increase its key to the desired value? 166 Chapter 6 Heapsort 6.5-5 Argue the correctness of HEAP-INCREASE-KEY using the following loop invariant: At the start of each iteration of the while loop of lines 4–6, the subarray AŒ1 : : A:heap-size satisfies the max-heap property, except that there may be one violation: AŒi  may be larger than AŒPARENT.i /. You may assume that the subarray AŒ1 : : A:heap-size satisfies the max-heap property at the time HEAP-INCREASE-KEY is called. 6.5-6 Each exchange operation on line 5 of HEAP-INCREASE-KEY typically requires three assignments. Show how to use the idea of the inner loop of INSERTIONSORT to reduce the three assignments down to just one assignment. 6.5-7 Show how to implement a first-in, first-out queue with a priority queue. Show how to implement a stack with a priority queue. (Queues and stacks are defined in Section 10.1.) 6.5-8 The operation HEAP-DELETE.A; i / deletes the item in node i from heap A. Give an implementation of HEAP-DELETE that runs in O.lg n/ time for an n-element max-heap. 6.5-9 Give an O.n lg k/-time algorithm to merge k sorted lists into one sorted list, where n is the total number of elements in all the input lists. (Hint: Use a minheap for k-way merging.) Problems 6-1 Building a heap using insertion We can build a heap by repeatedly calling MAX-HEAP-INSERT to insert the elements into the heap. Consider the following variation on the BUILD-MAX-HEAP procedure: Problems for Chapter 6 167 BUILD-MAX-HEAP0 .A/ 1 A:heap-size D 1 2 for i D 2 to A:length 3 MAX-HEAP-INSERT.A; AŒi / a. Do the procedures BUILD-MAX-HEAP and BUILD-MAX-HEAP0 always create the same heap when run on the same input array? Prove that they do, or provide a counterexample. b. Show that in the worst case, BUILD-MAX-HEAP0 requires ‚.n lg n/ time to build an n-element heap. 6-2 Analysis of d -ary heaps A d -ary heap is like a binary heap, but (with one possible exception) non-leaf nodes have d children instead of 2 children. a. How would you represent a d -ary heap in an array? b. What is the height of a d -ary heap of n elements in terms of n and d ? c. Give an efficient implementation of EXTRACT-MAX in a d -ary max-heap. Analyze its running time in terms of d and n. d. Give an efficient implementation of INSERT in a d -ary max-heap. Analyze its running time in terms of d and n. e. Give an efficient implementation of INCREASE-KEY.A; i; k/, which flags an error if k < AŒi, but otherwise sets AŒi D k and then updates the d -ary maxheap structure appropriately. Analyze its running time in terms of d and n. 6-3 Young tableaus An m n Young tableau is an m n matrix such that the entries of each row are in sorted order from left to right and the entries of each column are in sorted order from top to bottom. Some of the entries of a Young tableau may be 1, which we treat as nonexistent elements. Thus, a Young tableau can be used to hold r Ä mn finite numbers. a. Draw a 4 4 Young tableau containing the elements f9; 16; 3; 2; 4; 8; 5; 14; 12g. b. Argue that an m n Young tableau Y is empty if Y Œ1; 1 D 1. Argue that Y is full (contains mn elements) if Y Œm; n < 1. 168 Chapter 6 Heapsort c. Give an algorithm to implement EXTRACT-MIN on a nonempty m n Young tableau that runs in O.m C n/ time. Your algorithm should use a recursive subroutine that solves an m n problem by recursively solving either an .m 1/ n or an m .n 1/ subproblem. (Hint: Think about MAXHEAPIFY.) Define T .p/, where p D m C n, to be the maximum running time of EXTRACT-MIN on any m n Young tableau. Give and solve a recurrence for T .p/ that yields the O.m C n/ time bound. d. Show how to insert a new element into a nonfull m n Young tableau in O.m C n/ time. e. Using no other sorting method as a subroutine, show how to use an n n Young tableau to sort n2 numbers in O.n3/ time. f. Give an O.m C n/-time algorithm to determine whether a given number is stored in a given m n Young tableau. Chapter notes The heapsort algorithm was invented by Williams [357], who also described how to implement a priority queue with a heap. The BUILD-MAX-HEAP procedure was suggested by Floyd [106]. We use min-heaps to implement min-priority queues in Chapters 16, 23, and 24. We also give an implementation with improved time bounds for certain operations in Chapter 19 and, assuming that the keys are drawn from a bounded set of nonnegative integers, Chapter 20. If the data are b-bit integers, and the computer memory consists of addressable b-bit words, Fredman and Willard [115] showed hpow to implement MINIMUM in O.1/ time and INpSERT and EXTRACT-MIN in O. lg n/ time. Thorup [337] has improved the O. lg n/ bound to O.lg lg n/ time. This bound uses an amount of space unbounded in n, but it can be implemented in linear space by using randomized hashing. An important special case of priority queues occurs when the sequence of EXTRACT-MIN operations is monotone, that is, the values returned by successive EXTRACT-MIN operations are monotonically increasing over time. This case arises in several important applications, such as Dijkstra’s single-source shortestpaths algorithm, which we discuss in Chapter 24, and in discrete-event simulation. For Dijkstra’s algorithm it is particularly important that the DECREASE-KEY operation be implemented efficiently. For the monotone case, if the data are integers in the range 1; 2; : : : ; C , Ahuja, Mehlhorn, Orlin, and Tarjan [8] describe Notes for Chapter 6 169 how to implement EXTRACT-MIN and INSERT in O.lg C / amortized time (see Chapter 17 for more on amortized analysis) and DECREASE-KEY in O.1/ time, using pa data structure called a radix heap. The O.lg C / bound can be improved to O. lg C / using Fibonacci heaps (see Chapter 19) in conjunction with radix heaps. Cherkassky, Goldberg, and Silverstein [65] further improved the bound to O.lg1=3C C / expected time by combining the multilevel bucketing structure of Denardo and Fox [85] with the heap of Thorup mentioned earlier. Raman [291] further improved these results to obtain a bound of O.min.lg1=4C C; lg1=3C n//, for any fixed > 0. 7 Quicksort The quicksort algorithm has a worst-case running time of ‚.n2/ on an input array of n numbers. Despite this slow worst-case running time, quicksort is often the best practical choice for sorting because it is remarkably efficient on the average: its expected running time is ‚.n lg n/, and the constant factors hidden in the ‚.n lg n/ notation are quite small. It also has the advantage of sorting in place (see page 17), and it works well even in virtual-memory environments. Section 7.1 describes the algorithm and an important subroutine used by quicksort for partitioning. Because the behavior of quicksort is complex, we start with an intuitive discussion of its performance in Section 7.2 and postpone its precise analysis to the end of the chapter. Section 7.3 presents a version of quicksort that uses random sampling. This algorithm has a good expected running time, and no particular input elicits its worst-case behavior. Section 7.4 analyzes the randomized algorithm, showing that it runs in ‚.n2/ time in the worst case and, assuming distinct elements, in expected O.n lg n/ time. 7.1 Description of quicksort Quicksort, like merge sort, applies the divide-and-conquer paradigm introduced in Section 2.3.1. Here is the three-step divide-and-conquer process for sorting a typical subarray AŒp : : r: Divide: Partition (rearrange) the array AŒp : : r into two (possibly empty) subarrays AŒp : : q 1 and AŒq C 1 : : r such that each element of AŒp : : q 1 is less than or equal to AŒq, which is, in turn, less than or equal to each element of AŒq C 1 : : r. Compute the index q as part of this partitioning procedure. Conquer: Sort the two subarrays AŒp : : q 1 and AŒq C 1 : : r by recursive calls to quicksort. 7.1 Description of quicksort 171 Combine: Because the subarrays are already sorted, no work is needed to combine them: the entire array AŒp : : r is now sorted. The following procedure implements quicksort: QUICKSORT.A; p; r/ 1 if p < r 2 q D PARTITION.A; p; r/ 3 QUICKSORT.A; p; q 1/ 4 QUICKSORT.A; q C 1; r/ To sort an entire array A, the initial call is QUICKSORT.A; 1; A:length/. Partitioning the array The key to the algorithm is the PARTITION procedure, which rearranges the subarray AŒp : : r in place. PARTITION.A; p; r/ 1 x D AŒr 2 i Dp 1 3 for j D p to r 1 4 if AŒj  Ä x 5 i D i C1 6 exchange AŒi with AŒj  7 exchange AŒi C 1 with AŒr 8 return i C 1 Figure 7.1 shows how PARTITION works on an 8-element array. PARTITION always selects an element x D AŒr as a pivot element around which to partition the subarray AŒp : : r. As the procedure runs, it partitions the array into four (possibly empty) regions. At the start of each iteration of the for loop in lines 3–6, the regions satisfy certain properties, shown in Figure 7.2. We state these properties as a loop invariant: At the beginning of each iteration of the loop of lines 3–6, for any array index k, 1. If p Ä k Ä i, then AŒk Ä x. 2. If i C 1 Ä k Ä j 1, then AŒk > x. 3. If k D r, then AŒk D x. 172 Chapter 7 Quicksort i p,j r (a) 28713564 p,i j r (b) 28713564 p,i j r (c) 28713564 p,i j r (d) 28713564 pi j r (e) 21783564 pi jr (f) 21387564 pi jr (g) 21387564 pi r (h) 21387564 pi r (i) 21347568 Figure 7.1 The operation of PARTITION on a sample array. Array entry AŒr becomes the pivot element x. Lightly shaded array elements are all in the first partition with values no greater than x. Heavily shaded elements are in the second partition with values greater than x. The unshaded elements have not yet been put in one of the first two partitions, and the final white element is the pivot x. (a) The initial array and variable settings. None of the elements have been placed in either of the first two partitions. (b) The value 2 is “swapped with itself” and put in the partition of smaller values. (c)–(d) The values 8 and 7 are added to the partition of larger values. (e) The values 1 and 8 are swapped, and the smaller partition grows. (f) The values 3 and 7 are swapped, and the smaller partition grows. (g)–(h) The larger partition grows to include 5 and 6, and the loop terminates. (i) In lines 7–8, the pivot element is swapped so that it lies between the two partitions. The indices between j and r 1 are not covered by any of the three cases, and the values in these entries have no particular relationship to the pivot x. We need to show that this loop invariant is true prior to the first iteration, that each iteration of the loop maintains the invariant, and that the invariant provides a useful property to show correctness when the loop terminates. 7.1 Description of quicksort 173 p i j r x ≤x >x unrestricted Figure 7.2 The four regions maintained by the procedure PARTITION on a subarray AŒp : : r. The values in AŒp : : i are all less than or equal to x, the values in AŒi C 1 : : j 1 are all greater than x, and AŒr D x. The subarray AŒj : : r 1 can take on any values. Initialization: Prior to the first iteration of the loop, i D p 1 and j D p. Because no values lie between p and i and no values lie between i C 1 and j 1, the first two conditions of the loop invariant are trivially satisfied. The assignment in line 1 satisfies the third condition. Maintenance: As Figure 7.3 shows, we consider two cases, depending on the outcome of the test in line 4. Figure 7.3(a) shows what happens when AŒj  > x; the only action in the loop is to increment j . After j is incremented, condition 2 holds for AŒj 1 and all other entries remain unchanged. Figure 7.3(b) shows what happens when AŒj  Ä x; the loop increments i, swaps AŒi and AŒj , and then increments j . Because of the swap, we now have that AŒi Ä x, and condition 1 is satisfied. Similarly, we also have that AŒj 1 > x, since the item that was swapped into AŒj 1 is, by the loop invariant, greater than x. Termination: At termination, j D r. Therefore, every entry in the array is in one of the three sets described by the invariant, and we have partitioned the values in the array into three sets: those less than or equal to x, those greater than x, and a singleton set containing x. The final two lines of PARTITION finish up by swapping the pivot element with the leftmost element greater than x, thereby moving the pivot into its correct place in the partitioned array, and then returning the pivot’s new index. The output of PARTITION now satisfies the specifications given for the divide step. In fact, it satisfies a slightly stronger condition: after line 2 of QUICKSORT, AŒq is strictly less than every element of AŒq C 1 : : r. The running time of PARTITION on the subarray AŒp : : r is ‚.n/, where n D r p C 1 (see Exercise 7.1-3). Exercises 7.1-1 Using Figure 7.1 as a model, illustrate the operation of PARTITION on the array A D h13; 19; 9; 5; 12; 8; 7; 4; 21; 2; 6; 11i. 174 Chapter 7 Quicksort p i (a) j r >x x ≤x >x p i j r x ≤x >x p i (b) j r ≤x x ≤x >x p i j r x ≤x >x Figure 7.3 The two cases for one iteration of procedure PARTITION. (a) If AŒj  > x, the only action is to increment j , which maintains the loop invariant. (b) If AŒj  Ä x, index i is incremented, AŒi and AŒj  are swapped, and then j is incremented. Again, the loop invariant is maintained. 7.1-2 What value of q does PARTITION return when all elements in the array AŒp : : r have the same value? Modify PARTITION so that q D b.p C r/=2c when all elements in the array AŒp : : r have the same value. 7.1-3 Give a brief argument that the running time of PARTITION on a subarray of size n is ‚.n/. 7.1-4 How would you modify QUICKSORT to sort into nonincreasing order? 7.2 Performance of quicksort The running time of quicksort depends on whether the partitioning is balanced or unbalanced, which in turn depends on which elements are used for partitioning. If the partitioning is balanced, the algorithm runs asymptotically as fast as merge 7.2 Performance of quicksort 175 sort. If the partitioning is unbalanced, however, it can run asymptotically as slowly as insertion sort. In this section, we shall informally investigate how quicksort performs under the assumptions of balanced versus unbalanced partitioning. Worst-case partitioning The worst-case behavior for quicksort occurs when the partitioning routine produces one subproblem with n 1 elements and one with 0 elements. (We prove this claim in Section 7.4.1.) Let us assume that this unbalanced partitioning arises in each recursive call. The partitioning costs ‚.n/ time. Since the recursive call on an array of size 0 just returns, T .0/ D ‚.1/, and the recurrence for the running time is T .n/ D T .n 1/ C T .0/ C ‚.n/ D T .n 1/ C ‚.n/ : Intuitively, if we sum the costs incurred at each level of the recursion, we get an arithmetic series (equation (A.2)), which evaluates to ‚.n2/. Indeed, it is straightforward to use the substitution method to prove that the recurrence T .n/ D T .n 1/ C ‚.n/ has the solution T .n/ D ‚.n2/. (See Exercise 7.2-1.) Thus, if the partitioning is maximally unbalanced at every recursive level of the algorithm, the running time is ‚.n2/. Therefore the worst-case running time of quicksort is no better than that of insertion sort. Moreover, the ‚.n2/ running time occurs when the input array is already completely sorted—a common situation in which insertion sort runs in O.n/ time. Best-case partitioning In the most even possible split, PARTITION produces two subproblems, each of size no more than n=2, since one is of size bn=2c and one of size dn=2e 1. In this case, quicksort runs much faster. The recurrence for the running time is then T .n/ D 2T .n=2/ C ‚.n/ ; where we tolerate the sloppiness from ignoring the floor and ceiling and from subtracting 1. By case 2 of the master theorem (Theorem 4.1), this recurrence has the solution T .n/ D ‚.n lg n/. By equally balancing the two sides of the partition at every level of the recursion, we get an asymptotically faster algorithm. Balanced partitioning The average-case running time of quicksort is much closer to the best case than to the worst case, as the analyses in Section 7.4 will show. The key to understand- 176 Chapter 7 Quicksort n 1 10 n log10 n 1 100 n 9 100 n log10=9 n 1 9 10 n 9 100 n 81 100 n 81 1000 n 729 1000 n cn cn cn cn Ä cn 1 Ä cn O.n lg n/ Figure 7.4 A recursion tree for QUICKSORT in which PARTITION always produces a 9-to-1 split, yielding a running time of O.n lg n/. Nodes show subproblem sizes, with per-level costs on the right. The per-level costs include the constant c implicit in the ‚.n/ term. ing why is to understand how the balance of the partitioning is reflected in the recurrence that describes the running time. Suppose, for example, that the partitioning algorithm always produces a 9-to-1 proportional split, which at first blush seems quite unbalanced. We then obtain the recurrence T .n/ D T .9n=10/ C T .n=10/ C cn ; on the running time of quicksort, where we have explicitly included the constant c hidden in the ‚.n/ term. Figure 7.4 shows the recursion tree for this recurrence. Notice that every level of the tree has cost cn, until the recursion reaches a boundary condition at depth log10 n D ‚.lg n/, and then the levels have cost at most cn. The recursion terminates at depth log10=9 n D ‚.lg n/. The total cost of quicksort is therefore O.n lg n/. Thus, with a 9-to-1 proportional split at every level of recursion, which intuitively seems quite unbalanced, quicksort runs in O.n lg n/ time—asymptotically the same as if the split were right down the middle. Indeed, even a 99-to-1 split yields an O.n lg n/ running time. In fact, any split of constant proportionality yields a recursion tree of depth ‚.lg n/, where the cost at each level is O.n/. The running time is therefore O.n lg n/ whenever the split has constant proportionality. 7.2 Performance of quicksort 177 n Θ(n) 0 n–1 (n–1)/2 – 1 (a) (n–1)/2 n Θ(n) (n–1)/2 (n–1)/2 (b) Figure 7.5 (a) Two levels of a recursion tree for quicksort. The partitioning at the root costs n and produces a “bad” split: two subarrays of sizes 0 and n 1. The partitioning of the subarray of size n 1 costs n 1 and produces a “good” split: subarrays of size .n 1/=2 1 and .n 1/=2. (b) A single level of a recursion tree that is very well balanced. In both parts, the partitioning cost for the subproblems shown with elliptical shading is ‚.n/. Yet the subproblems remaining to be solved in (a), shown with square shading, are no larger than the corresponding subproblems remaining to be solved in (b). Intuition for the average case To develop a clear notion of the randomized behavior of quicksort, we must make an assumption about how frequently we expect to encounter the various inputs. The behavior of quicksort depends on the relative ordering of the values in the array elements given as the input, and not by the particular values in the array. As in our probabilistic analysis of the hiring problem in Section 5.2, we will assume for now that all permutations of the input numbers are equally likely. When we run quicksort on a random input array, the partitioning is highly unlikely to happen in the same way at every level, as our informal analysis has assumed. We expect that some of the splits will be reasonably well balanced and that some will be fairly unbalanced. For example, Exercise 7.2-6 asks you to show that about 80 percent of the time PARTITION produces a split that is more balanced than 9 to 1, and about 20 percent of the time it produces a split that is less balanced than 9 to 1. In the average case, PARTITION produces a mix of “good” and “bad” splits. In a recursion tree for an average-case execution of PARTITION, the good and bad splits are distributed randomly throughout the tree. Suppose, for the sake of intuition, that the good and bad splits alternate levels in the tree, and that the good splits are best-case splits and the bad splits are worst-case splits. Figure 7.5(a) shows the splits at two consecutive levels in the recursion tree. At the root of the tree, the cost is n for partitioning, and the subarrays produced have sizes n 1 and 0: the worst case. At the next level, the subarray of size n 1 undergoes best-case partitioning into subarrays of size .n 1/=2 1 and .n 1/=2. Let’s assume that the boundary-condition cost is 1 for the subarray of size 0. 178 Chapter 7 Quicksort The combination of the bad split followed by the good split produces three subarrays of sizes 0, .n 1/=2 1, and .n 1/=2 at a combined partitioning cost of ‚.n/ C ‚.n 1/ D ‚.n/. Certainly, this situation is no worse than that in Figure 7.5(b), namely a single level of partitioning that produces two subarrays of size .n 1/=2, at a cost of ‚.n/. Yet this latter situation is balanced! Intuitively, the ‚.n 1/ cost of the bad split can be absorbed into the ‚.n/ cost of the good split, and the resulting split is good. Thus, the running time of quicksort, when levels alternate between good and bad splits, is like the running time for good splits alone: still O.n lg n/, but with a slightly larger constant hidden by the O-notation. We shall give a rigorous analysis of the expected running time of a randomized version of quicksort in Section 7.4.2. Exercises 7.2-1 Use the substitution method to prove that the recurrence T .n/ D T .n 1/ C ‚.n/ has the solution T .n/ D ‚.n2/, as claimed at the beginning of Section 7.2. 7.2-2 What is the running time of QUICKSORT when all elements of array A have the same value? 7.2-3 Show that the running time of QUICKSORT is ‚.n2/ when the array A contains distinct elements and is sorted in decreasing order. 7.2-4 Banks often record transactions on an account in order of the times of the transactions, but many people like to receive their bank statements with checks listed in order by check number. People usually write checks in order by check number, and merchants usually cash them with reasonable dispatch. The problem of converting time-of-transaction ordering to check-number ordering is therefore the problem of sorting almost-sorted input. Argue that the procedure INSERTION-SORT would tend to beat the procedure QUICKSORT on this problem. 7.2-5 Suppose that the splits at every level of quicksort are in the proportion 1 ˛ to ˛, where 0 < ˛ Ä 1=2 is a constant. Show that the minimum depth of a leaf in the recursion tree is approximately lg n= lg ˛ and the maximum depth is approximately lg n= lg.1 ˛/. (Don’t worry about integer round-off.) 7.3 A randomized version of quicksort 179 7.2-6 ? Argue that for any constant 0 < ˛ Ä 1=2, the probability is approximately 1 2˛ that on a random input array, PARTITION produces a split more balanced than 1 ˛ to ˛. 7.3 A randomized version of quicksort In exploring the average-case behavior of quicksort, we have made an assumption that all permutations of the input numbers are equally likely. In an engineering situation, however, we cannot always expect this assumption to hold. (See Exercise 7.2-4.) As we saw in Section 5.3, we can sometimes add randomization to an algorithm in order to obtain good expected performance over all inputs. Many people regard the resulting randomized version of quicksort as the sorting algorithm of choice for large enough inputs. In Section 5.3, we randomized our algorithm by explicitly permuting the input. We could do so for quicksort also, but a different randomization technique, called random sampling, yields a simpler analysis. Instead of always using AŒr as the pivot, we will select a randomly chosen element from the subarray AŒp : : r. We do so by first exchanging element AŒr with an element chosen at random from AŒp : : r. By randomly sampling the range p; : : : ; r, we ensure that the pivot element x D AŒr is equally likely to be any of the r p C 1 elements in the subarray. Because we randomly choose the pivot element, we expect the split of the input array to be reasonably well balanced on average. The changes to PARTITION and QUICKSORT are small. In the new partition procedure, we simply implement the swap before actually partitioning: RANDOMIZED-PARTITION .A; p; r / 1 i D RANDOM.p; r/ 2 exchange AŒr with AŒi 3 return PARTITION.A; p; r/ The new quicksort calls RANDOMIZED-PARTITION in place of PARTITION: RANDOMIZED-QUICKSORT.A; p; r / 1 if p < r 2 q D RANDOMIZED-PARTITION .A; p; r / 3 RANDOMIZED-QUICKSORT.A; p; q 1/ 4 RANDOMIZED-QUICKSORT.A; q C 1; r / We analyze this algorithm in the next section. 180 Chapter 7 Quicksort Exercises 7.3-1 Why do we analyze the expected running time of a randomized algorithm and not its worst-case running time? 7.3-2 When RANDOMIZED-QUICKSORT runs, how many calls are made to the randomnumber generator RANDOM in the worst case? How about in the best case? Give your answer in terms of ‚-notation. 7.4 Analysis of quicksort Section 7.2 gave some intuition for the worst-case behavior of quicksort and for why we expect it to run quickly. In this section, we analyze the behavior of quicksort more rigorously. We begin with a worst-case analysis, which applies to either QUICKSORT or RANDOMIZED-QUICKSORT, and conclude with an analysis of the expected running time of RANDOMIZED-QUICKSORT. 7.4.1 Worst-case analysis We saw in Section 7.2 that a worst-case split at every level of recursion in quicksort produces a ‚.n2/ running time, which, intuitively, is the worst-case running time of the algorithm. We now prove this assertion. Using the substitution method (see Section 4.3), we can show that the running time of quicksort is O.n2/. Let T .n/ be the worst-case time for the procedure QUICKSORT on an input of size n. We have the recurrence T .n/ D max .T .q/ C T .n q 1// C ‚.n/ ; 0ÄqÄn 1 (7.1) where the parameter q ranges from 0 to n 1 because the procedure PARTITION produces two subproblems with total size n 1. We guess that T .n/ Ä cn2 for some constant c. Substituting this guess into recurrence (7.1), we obtain T .n/ Ä max .cq2 C c.n q 1/2/ C ‚.n/ 0ÄqÄn 1 D c max .q2 C .n q 1/2/ C ‚.n/ : 0ÄqÄn 1 The expression q2 C .n q 1/2 achieves a maximum over the parameter’s range 0 Ä q Ä n 1 at either endpoint. To verify this claim, note that the second derivative of the expression with respect to q is positive (see Exercise 7.4-3). This 7.4 Analysis of quicksort 181 observation gives us the bound max0ÄqÄn 1.q2 C .n q 1/2/ Ä .n n2 2n C 1. Continuing with our bounding of T .n/, we obtain T .n/ Ä cn2 c.2n 1/ C ‚.n/ Ä cn2 ; 1/2 D since we can pick the constant c large enough so that the c.2n 1/ term dominates the ‚.n/ term. Thus, T .n/ D O.n2/. We saw in Section 7.2 a specific case in which quicksort takes .n2/ time: when partitioning is unbalanced. Al- ternatively, Exercise 7.4-1 asks you to show that recurrence (7.1) has a solution of T .n/ D .n2/. Thus, the (worst-case) running time of quicksort is ‚.n2/. 7.4.2 Expected running time We have already seen the intuition behind why the expected running time of RANDOMIZED-QUICKSORT is O.n lg n/: if, in each level of recursion, the split induced by RANDOMIZED-PARTITION puts any constant fraction of the elements on one side of the partition, then the recursion tree has depth ‚.lg n/, and O.n/ work is performed at each level. Even if we add a few new levels with the most unbalanced split possible between these levels, the total time remains O.n lg n/. We can analyze the expected running time of RANDOMIZED-QUICKSORT precisely by first understanding how the partitioning procedure operates and then using this understanding to derive an O.n lg n/ bound on the expected running time. This upper bound on the expected running time, combined with the ‚.n lg n/ best-case bound we saw in Section 7.2, yields a ‚.n lg n/ expected running time. We assume throughout that the values of the elements being sorted are distinct. Running time and comparisons The QUICKSORT and RANDOMIZED-QUICKSORT procedures differ only in how they select pivot elements; they are the same in all other respects. We can therefore couch our analysis of RANDOMIZED-QUICKSORT by discussing the QUICKSORT and PARTITION procedures, but with the assumption that pivot elements are selected randomly from the subarray passed to RANDOMIZED-PARTITION. The running time of QUICKSORT is dominated by the time spent in the PARTITION procedure. Each time the PARTITION procedure is called, it selects a pivot element, and this element is never included in any future recursive calls to QUICKSORT and PARTITION. Thus, there can be at most n calls to PARTITION over the entire execution of the quicksort algorithm. One call to PARTITION takes O.1/ time plus an amount of time that is proportional to the number of iterations of the for loop in lines 3–6. Each iteration of this for loop performs a comparison in line 4, comparing the pivot element to another element of the array A. Therefore, 182 Chapter 7 Quicksort if we can count the total number of times that line 4 is executed, we can bound the total time spent in the for loop during the entire execution of QUICKSORT. Lemma 7.1 Let X be the number of comparisons performed in line 4 of PARTITION over the entire execution of QUICKSORT on an n-element array. Then the running time of QUICKSORT is O.n C X /. Proof By the discussion above, the algorithm makes at most n calls to PARTITION, each of which does a constant amount of work and then executes the for loop some number of times. Each iteration of the for loop executes line 4. Our goal, therefore, is to compute X , the total number of comparisons performed in all calls to PARTITION. We will not attempt to analyze how many comparisons are made in each call to PARTITION. Rather, we will derive an overall bound on the total number of comparisons. To do so, we must understand when the algorithm compares two elements of the array and when it does not. For ease of analysis, we rename the elements of the array A as ´1; ´2; : : : ; ´n, with ´i being the i th smallest element. We also define the set Zij D f´i ; ´iC1; : : : ; ´j g to be the set of elements between ´i and ´j , inclusive. When does the algorithm compare ´i and ´j ? To answer this question, we first observe that each pair of elements is compared at most once. Why? Elements are compared only to the pivot element and, after a particular call of PARTITION finishes, the pivot element used in that call is never again compared to any other elements. Our analysis uses indicator random variables (see Section 5.2). We define Xij D I f´i is compared to ´j g ; where we are considering whether the comparison takes place at any time during the execution of the algorithm, not just during one iteration or one call of PARTITION. Since each pair is compared at most once, we can easily characterize the total number of comparisons performed by the algorithm: X n 1 Xn XD Xij : i D1 j Di C1 Taking expectations of both sides, and then using linearity of expectation and Lemma 5.1, we obtain " # X n 1 Xn E ŒX  D E Xij i D1 j Di C1 7.4 Analysis of quicksort 183 X n 1 Xn D E ŒXij  i D1 j Di C1 X n 1 Xn D Pr f´i is compared to ´j g : i D1 j Di C1 (7.2) It remains to compute Pr f´i is compared to ´j g. Our analysis assumes that the RANDOMIZED-PARTITION procedure chooses each pivot randomly and independently. Let us think about when two items are not compared. Consider an input to quicksort of the numbers 1 through 10 (in any order), and suppose that the first pivot element is 7. Then the first call to PARTITION separates the numbers into two sets: f1; 2; 3; 4; 5; 6g and f8; 9; 10g. In doing so, the pivot element 7 is compared to all other elements, but no number from the first set (e.g., 2) is or ever will be compared to any number from the second set (e.g., 9). In general, because we assume that element values are distinct, once a pivot x is chosen with ´i < x < ´j , we know that ´i and ´j cannot be compared at any subsequent time. If, on the other hand, ´i is chosen as a pivot before any other item in Zij , then ´i will be compared to each item in Zij , except for itself. Similarly, if ´j is chosen as a pivot before any other item in Zij , then ´j will be compared to each item in Zij , except for itself. In our example, the values 7 and 9 are compared because 7 is the first item from Z7;9 to be chosen as a pivot. In contrast, 2 and 9 will never be compared because the first pivot element chosen from Z2;9 is 7. Thus, ´i and ´j are compared if and only if the first element to be chosen as a pivot from Zij is either ´i or ´j . We now compute the probability that this event occurs. Prior to the point at which an element from Zij has been chosen as a pivot, the whole set Zij is together in the same partition. Therefore, any element of Zij is equally likely to be the first one chosen as a pivot. Because the set Zij has j i C1 elements, and because pivots are chosen randomly and independently, the probability that any given element is the first one chosen as a pivot is 1=.j i C 1/. Thus, we have Pr f´i is compared to ´j g D Pr f´i or ´j is first pivot chosen from Zij g D Pr f´i is first pivot chosen from Zij g C Pr f´j is first pivot chosen from Zij g 1 1 D C j i C1 j iC1 2 D j i C1 : (7.3) 184 Chapter 7 Quicksort The second line follows because the two events are mutually exclusive. Combining equations (7.2) and (7.3), we get that X n 1 Xn E ŒX  D j i D1 j Di C1 2 : i C1 We can evaluate this sum using a change of variables (k D j on the harmonic series in equation (A.7): i) and the bound X n 1 Xn E ŒX  D j i D1 j Di C1 2 i C1 X n 1 X n i 2 D kC1 i D1 kD1 X n 1 Xn 2 < k i D1 kD1 X n 1 D O.lg n/ i D1 D O.n lg n/ : (7.4) Thus we conclude that, using RANDOMIZED-PARTITION, the expected running time of quicksort is O.n lg n/ when element values are distinct. Exercises 7.4-1 Show that in the recurrence T .n/ D max .T .q/ C T .n q 1// C ‚.n/ ; 0ÄqÄn 1 T .n/ D .n2/. 7.4-2 Show that quicksort’s best-case running time is .n lg n/. 7.4-3 Show that the expression q2 C .n q 1/2 achieves a maximum over q D 0; 1; : : : ; n 1 when q D 0 or q D n 1. 7.4-4 Show that RANDOMIZED-QUICKSORT’s expected running time is .n lg n/. Problems for Chapter 7 185 7.4-5 We can improve the running time of quicksort in practice by taking advantage of the fast running time of insertion sort when its input is “nearly” sorted. Upon calling quicksort on a subarray with fewer than k elements, let it simply return without sorting the subarray. After the top-level call to quicksort returns, run insertion sort on the entire array to finish the sorting process. Argue that this sorting algorithm runs in O.nk C n lg.n=k// expected time. How should we pick k, both in theory and in practice? 7.4-6 ? Consider modifying the PARTITION procedure by randomly picking three elements from array A and partitioning about their median (the middle value of the three elements). Approximate the probability of getting at worst an ˛-to-.1 ˛/ split, as a function of ˛ in the range 0 < ˛ < 1. Problems 7-1 Hoare partition correctness The version of PARTITION given in this chapter is not the original partitioning algorithm. Here is the original partition algorithm, which is due to C. A. R. Hoare: HOARE-PARTITION.A; p; r / 1 x D AŒp 2 iDp 1 3 j D r C1 4 while TRUE 5 repeat 6 j Dj 1 7 until AŒj  Ä x 8 repeat 9 i D iC1 10 until AŒi x 11 if i < j 12 exchange AŒi with AŒj  13 else return j a. Demonstrate the operation of HOARE-PARTITION on the array A D h13; 19; 9; 5; 12; 8; 7; 4; 11; 2; 6; 21i, showing the values of the array and auxiliary values after each iteration of the while loop in lines 4–13. 186 Chapter 7 Quicksort The next three questions ask you to give a careful argument that the procedure HOARE-PARTITION is correct. Assuming that the subarray AŒp : : r contains at least two elements, prove the following: b. The indices i and j are such that we never access an element of A outside the subarray AŒp : : r. c. When HOARE-PARTITION terminates, it returns a value j such that p Ä j < r. d. Every element of AŒp : : j  is less than or equal to every element of AŒj C1 : : r when HOARE-PARTITION terminates. The PARTITION procedure in Section 7.1 separates the pivot value (originally in AŒr) from the two partitions it forms. The HOARE-PARTITION procedure, on the other hand, always places the pivot value (originally in AŒp) into one of the two partitions AŒp : : j  and AŒj C 1 : : r. Since p Ä j < r, this split is always nontrivial. e. Rewrite the QUICKSORT procedure to use HOARE-PARTITION. 7-2 Quicksort with equal element values The analysis of the expected running time of randomized quicksort in Section 7.4.2 assumes that all element values are distinct. In this problem, we examine what happens when they are not. a. Suppose that all element values are equal. What would be randomized quicksort’s running time in this case? b. The PARTITION procedure returns an index q such that each element of AŒp : : q 1 is less than or equal to AŒq and each element of AŒq C 1 : : r is greater than AŒq. Modify the PARTITION procedure to produce a procedure PARTITION0.A; p; r/, which permutes the elements of AŒp : : r and returns two indices q and t, where p Ä q Ä t Ä r, such that all elements of AŒq : : t are equal, each element of AŒp : : q 1 is less than AŒq, and each element of AŒt C 1 : : r is greater than AŒq. Like PARTITION, your PARTITION0 procedure should take ‚.r p/ time. c. Modify the RANDOMIZED-QUICKSORT procedure to call PARTITION0, and name the new procedure RANDOMIZED-QUICKSORT0. Then modify the QUICKSORT procedure to produce a procedure QUICKSORT0.p; r/ that calls Problems for Chapter 7 187 RANDOMIZED-PARTITION0 and recurses only on partitions of elements not known to be equal to each other. d. Using QUICKSORT0, how would you adjust the analysis in Section 7.4.2 to avoid the assumption that all elements are distinct? 7-3 Alternative quicksort analysis An alternative analysis of the running time of randomized quicksort focuses on the expected running time of each individual recursive call to RANDOMIZEDQUICKSORT, rather than on the number of comparisons performed. a. Argue that, given an array of size n, the probability that any particular element is chosen as the pivot is 1=n. Use this to define indicator random variables Xi D I fi th smallest element is chosen as the pivotg. What is E ŒXi ? b. Let T .n/ be a random variable denoting the running time of quicksort on an array of size n. Argue that " # Xn E ŒT .n/ D E Xq .T .q 1/ C T .n q/ C ‚.n// : (7.5) qD1 c. Show that we can rewrite equation (7.5) as 2 X n 1 E ŒT .n/ D E ŒT .q/ C ‚.n/ : n qD2 (7.6) d. Show that X n 1 k lg k Ä 1 n2 lg n 2 kD2 1 n2 : 8 (7.7) (Hint: Split the summation into two parts, one for k D 2; 3; : : : ; dn=2e 1 and one for k D dn=2e ; : : : ; n 1.) e. Using the bound from equation (7.7), show that the recurrence in equation (7.6) has the solution E ŒT .n/ D ‚.n lg n/. (Hint: Show, by substitution, that E ŒT .n/ Ä an lg n for sufficiently large n and for some positive constant a.) 188 Chapter 7 Quicksort 7-4 Stack depth for quicksort The QUICKSORT algorithm of Section 7.1 contains two recursive calls to itself. After QUICKSORT calls PARTITION, it recursively sorts the left subarray and then it recursively sorts the right subarray. The second recursive call in QUICKSORT is not really necessary; we can avoid it by using an iterative control structure. This technique, called tail recursion, is provided automatically by good compilers. Consider the following version of quicksort, which simulates tail recursion: TAIL-RECURSIVE-QUICKSORT.A; p; r / 1 while p < r 2 // Partition and sort left subarray. 3 q D PARTITION.A; p; r/ 4 TAIL-RECURSIVE-QUICKSORT .A; p; q 1/ 5 p D qC1 a. Argue that TAIL-RECURSIVE-QUICKSORT.A; 1; A:length/ correctly sorts the array A. Compilers usually execute recursive procedures by using a stack that contains pertinent information, including the parameter values, for each recursive call. The information for the most recent call is at the top of the stack, and the information for the initial call is at the bottom. Upon calling a procedure, its information is pushed onto the stack; when it terminates, its information is popped. Since we assume that array parameters are represented by pointers, the information for each procedure call on the stack requires O.1/ stack space. The stack depth is the maximum amount of stack space used at any time during a computation. b. Describe a scenario in which TAIL-RECURSIVE-QUICKSORT’s stack depth is ‚.n/ on an n-element input array. c. Modify the code for TAIL-RECURSIVE-QUICKSORT so that the worst-case stack depth is ‚.lg n/. Maintain the O.n lg n/ expected running time of the algorithm. 7-5 Median-of-3 partition One way to improve the RANDOMIZED-QUICKSORT procedure is to partition around a pivot that is chosen more carefully than by picking a random element from the subarray. One common approach is the median-of-3 method: choose the pivot as the median (middle element) of a set of 3 elements randomly selected from the subarray. (See Exercise 7.4-6.) For this problem, let us assume that the elements in the input array AŒ1 : : n are distinct and that n 3. We denote the Problems for Chapter 7 189 sorted output array by A0Œ1 : : n. Using the median-of-3 method to choose the pivot element x, define pi D Pr fx D A0Œi g. a. Give an exact formula for pi as a function of n and i for i D 2; 3; : : : ; n 1. (Note that p1 D pn D 0.) b. By what amount have we increased the likelihood of choosing the pivot as x D A0Œb.n C 1/=2c, the median of AŒ1 : : n, compared with the ordinary implementation? Assume that n ! 1, and give the limiting ratio of these probabilities. c. If we define a “good” split to mean choosing the pivot as x D A0Œi, where n=3 Ä i Ä 2n=3, by what amount have we increased the likelihood of getting a good split compared with the ordinary implementation? (Hint: Approximate the sum by an integral.) d. Argue that in the .n lg n/ running time of quicksort, the median-of-3 method affects only the constant factor. 7-6 Fuzzy sorting of intervals Consider a sorting problem in which we do not know the numbers exactly. Instead, for each number, we know an interval on the real line to which it belongs. That is, we are given n closed intervals of the form Œai ; bi , where ai Ä bi . We wish to fuzzy-sort these intervals, i.e., to produce a permutation hi1; i2; : : : ; ini of the intervals such that for j D 1; 2; : : : ; n, there exist cj 2 Œaij ; bij  satisfying c1 Ä c2 Ä Ä cn. a. Design a randomized algorithm for fuzzy-sorting n intervals. Your algorithm should have the general structure of an algorithm that quicksorts the left endpoints (the ai values), but it should take advantage of overlapping intervals to improve the running time. (As the intervals overlap more and more, the problem of fuzzy-sorting the intervals becomes progressively easier. Your algorithm should take advantage of such overlapping, to the extent that it exists.) b. Argue that your algorithm runs in expected time ‚.n lg n/ in general, but runs in expected time ‚.n/ when all of the intervals overlap (i.e., when there exists a value x such that x 2 Œai ; bi  for all i ). Your algorithm should not be checking for this case explicitly; rather, its performance should naturally improve as the amount of overlap increases. 190 Chapter 7 Quicksort Chapter notes The quicksort procedure was invented by Hoare [170]; Hoare’s version appears in Problem 7-1. The PARTITION procedure given in Section 7.1 is due to N. Lomuto. The analysis in Section 7.4 is due to Avrim Blum. Sedgewick [305] and Bentley [43] provide a good reference on the details of implementation and how they matter. McIlroy [248] showed how to engineer a “killer adversary” that produces an array on which virtually any implementation of quicksort takes ‚.n2/ time. If the implementation is randomized, the adversary produces the array after seeing the random choices of the quicksort algorithm. 8 Sorting in Linear Time We have now introduced several algorithms that can sort n numbers in O.n lg n/ time. Merge sort and heapsort achieve this upper bound in the worst case; quicksort achieves it on average. Moreover, for each of these algorithms, we can produce a sequence of n input numbers that causes the algorithm to run in .n lg n/ time. These algorithms share an interesting property: the sorted order they determine is based only on comparisons between the input elements. We call such sorting algorithms comparison sorts. All the sorting algorithms introduced thus far are comparison sorts. In Section 8.1, we shall prove that any comparison sort must make .n lg n/ comparisons in the worst case to sort n elements. Thus, merge sort and heapsort are asymptotically optimal, and no comparison sort exists that is faster by more than a constant factor. Sections 8.2, 8.3, and 8.4 examine three sorting algorithms—counting sort, radix sort, and bucket sort—that run in linear time. Of course, these algorithms use operations other than comparisons to determine the sorted order. Consequently, the .n lg n/ lower bound does not apply to them. 8.1 Lower bounds for sorting In a comparison sort, we use only comparisons between elements to gain order information about an input sequence ha1; a2; : : : ; ani. That is, given two elements ai and aj , we perform one of the tests ai < aj , ai Ä aj , ai D aj , ai aj , or ai > aj to determine their relative order. We may not inspect the values of the elements or gain order information about them in any other way. In this section, we assume without loss of generality that all the input elements are distinct. Given this assumption, comparisons of the form ai D aj are useless, so we can assume that no comparisons of this form are made. We also note that the comparisons ai Ä aj , ai aj , ai > aj , and ai < aj are all equivalent in that 192 Chapter 8 Sorting in Linear Time 1:2 ≤ > 2:3 ≤ > 1:3 ≤ > 〈1,2,3〉 1:3 ≤ > 〈2,1,3〉 2:3 ≤ > 〈1,3,2〉 〈3,1,2〉 〈2,3,1〉 〈3,2,1〉 Figure 8.1 The decision tree for insertion sort operating on three elements. An internal node annotated by i:j indicates a comparison between ai and aj . A leaf annotated by the permutation h .1/; .2/; : : : ; .n/i indicates the ordering a .1/ Ä a .2/ Ä Ä a .n/. The shaded path indicates the decisions made when sorting the input sequence ha1 D 6; a2 D 8; a3 D 5i; the permutation h3; 1; 2i at the leaf indicates that the sorted ordering is a3 D 5 Ä a1 D 6 Ä a2 D 8. There are 3Š D 6 possible permutations of the input elements, and so the decision tree must have at least 6 leaves. they yield identical information about the relative order of ai and aj . We therefore assume that all comparisons have the form ai Ä aj . The decision-tree model We can view comparison sorts abstractly in terms of decision trees. A decision tree is a full binary tree that represents the comparisons between elements that are performed by a particular sorting algorithm operating on an input of a given size. Control, data movement, and all other aspects of the algorithm are ignored. Figure 8.1 shows the decision tree corresponding to the insertion sort algorithm from Section 2.1 operating on an input sequence of three elements. In a decision tree, we annotate each internal node by i:j for some i and j in the range 1 Ä i; j Ä n, where n is the number of elements in the input sequence. We also annotate each leaf by a permutation h .1/; .2/; : : : ; .n/i. (See Section C.1 for background on permutations.) The execution of the sorting algorithm corresponds to tracing a simple path from the root of the decision tree down to a leaf. Each internal node indicates a comparison ai Ä aj . The left subtree then dictates subsequent comparisons once we know that ai Ä aj , and the right subtree dictates subsequent comparisons knowing that ai > aj . When we come to a leaf, the sorting algorithm has established the ordering a .1/ Ä a .2/ Ä Ä a .n/. Because any correct sorting algorithm must be able to produce each permutation of its input, each of the nŠ permutations on n elements must appear as one of the leaves of the decision tree for a comparison sort to be correct. Furthermore, each of these leaves must be reachable from the root by a downward path corresponding to an actual 8.1 Lower bounds for sorting 193 execution of the comparison sort. (We shall refer to such leaves as “reachable.”) Thus, we shall consider only decision trees in which each permutation appears as a reachable leaf. A lower bound for the worst case The length of the longest simple path from the root of a decision tree to any of its reachable leaves represents the worst-case number of comparisons that the corresponding sorting algorithm performs. Consequently, the worst-case number of comparisons for a given comparison sort algorithm equals the height of its decision tree. A lower bound on the heights of all decision trees in which each permutation appears as a reachable leaf is therefore a lower bound on the running time of any comparison sort algorithm. The following theorem establishes such a lower bound. Theorem 8.1 Any comparison sort algorithm requires .n lg n/ comparisons in the worst case. Proof From the preceding discussion, it suffices to determine the height of a decision tree in which each permutation appears as a reachable leaf. Consider a decision tree of height h with l reachable leaves corresponding to a comparison sort on n elements. Because each of the nŠ permutations of the input appears as some leaf, we have nŠ Ä l. Since a binary tree of height h has no more than 2h leaves, we have nŠ Ä l Ä 2h ; which, by taking logarithms, implies h lg.nŠ/ (since the lg function is monotonically increasing) D .n lg n/ (by equation (3.19)) . Corollary 8.2 Heapsort and merge sort are asymptotically optimal comparison sorts. Proof The O.n lg n/ upper bounds on the running times for heapsort and merge sort match the .n lg n/ worst-case lower bound from Theorem 8.1. Exercises 8.1-1 What is the smallest possible depth of a leaf in a decision tree for a comparison sort? 194 Chapter 8 Sorting in Linear Time 8.1-2 Obtain mation. asInysmtepatdo,ticeavlallyuattieghtthebosuunmdms aotinonlgP.nnkŠ/D1wligthkouutsiunsgintgechSntiirqluinegs’sfroapmprSoexci-- tion A.2. 8.1-3 Show that there is no comparison sort whose running time is linear for at least half of the nŠ inputs of length n. What about a fraction of 1=n of the inputs of length n? What about a fraction 1=2n? 8.1-4 Suppose that you are given a sequence of n elements to sort. The input sequence consists of n=k subsequences, each containing k elements. The elements in a given subsequence are all smaller than the elements in the succeeding subsequence and larger than the elements in the preceding subsequence. Thus, all that is needed to sort the whole sequence of length n is to sort the k elements in each of the n=k subsequences. Show an .n lg k/ lower bound on the number of comparisons needed to solve this variant of the sorting problem. (Hint: It is not rigorous to simply combine the lower bounds for the individual subsequences.) 8.2 Counting sort Counting sort assumes that each of the n input elements is an integer in the range 0 to k, for some integer k. When k D O.n/, the sort runs in ‚.n/ time. Counting sort determines, for each input element x, the number of elements less than x. It uses this information to place element x directly into its position in the output array. For example, if 17 elements are less than x, then x belongs in output position 18. We must modify this scheme slightly to handle the situation in which several elements have the same value, since we do not want to put them all in the same position. In the code for counting sort, we assume that the input is an array AŒ1 : : n, and thus A:length D n. We require two other arrays: the array BŒ1 : : n holds the sorted output, and the array C Œ0 : : k provides temporary working storage. 8.2 Counting sort 12345678 A25302303 012345 C202301 (a) 12345678 B0 3 012345 C124678 (d) 012345 C224778 (b) 12345678 B0 33 012345 C124578 (e) 195 12345678 B 3 012345 C224678 (c) 12345678 B00223335 (f) Figure 8.2 The operation of COUNTING-SORT on an input array AŒ1 : : 8, where each element of A is a nonnegative integer no larger than k D 5. (a) The array A and the auxiliary array C after line 5. (b) The array C after line 8. (c)–(e) The output array B and the auxiliary array C after one, two, and three iterations of the loop in lines 10–12, respectively. Only the lightly shaded elements of array B have been filled in. (f) The final sorted output array B. COUNTING-SORT.A; B; k/ 1 let C Œ0 : : k be a new array 2 for i D 0 to k 3 C Œi D 0 4 for j D 1 to A:length 5 C ŒAŒj  D C ŒAŒj  C 1 6 // C Œi now contains the number of elements equal to i. 7 for i D 1 to k 8 C Œi D C Œi C C Œi 1 9 // C Œi now contains the number of elements less than or equal to i. 10 for j D A:length downto 1 11 BŒC ŒAŒj  D AŒj  12 C ŒAŒj  D C ŒAŒj  1 Figure 8.2 illustrates counting sort. After the for loop of lines 2–3 initializes the array C to all zeros, the for loop of lines 4–5 inspects each input element. If the value of an input element is i, we increment C Œi. Thus, after line 5, C Œi holds the number of input elements equal to i for each integer i D 0; 1; : : : ; k. Lines 7–8 determine for each i D 0; 1; : : : ; k how many input elements are less than or equal to i by keeping a running sum of the array C . 196 Chapter 8 Sorting in Linear Time Finally, the for loop of lines 10–12 places each element AŒj  into its correct sorted position in the output array B. If all n elements are distinct, then when we first enter line 10, for each AŒj , the value C ŒAŒj  is the correct final position of AŒj  in the output array, since there are C ŒAŒj  elements less than or equal to AŒj . Because the elements might not be distinct, we decrement C ŒAŒj  each time we place a value AŒj  into the B array. Decrementing C ŒAŒj  causes the next input element with a value equal to AŒj , if one exists, to go to the position immediately before AŒj  in the output array. How much time does counting sort require? The for loop of lines 2–3 takes time ‚.k/, the for loop of lines 4–5 takes time ‚.n/, the for loop of lines 7–8 takes time ‚.k/, and the for loop of lines 10–12 takes time ‚.n/. Thus, the overall time is ‚.k C n/. In practice, we usually use counting sort when we have k D O.n/, in which case the running time is ‚.n/. Counting sort beats the lower bound of .n lg n/ proved in Section 8.1 because it is not a comparison sort. In fact, no comparisons between input elements occur anywhere in the code. Instead, counting sort uses the actual values of the elements to index into an array. The .n lg n/ lower bound for sorting does not apply when we depart from the comparison sort model. An important property of counting sort is that it is stable: numbers with the same value appear in the output array in the same order as they do in the input array. That is, it breaks ties between two numbers by the rule that whichever number appears first in the input array appears first in the output array. Normally, the property of stability is important only when satellite data are carried around with the element being sorted. Counting sort’s stability is important for another reason: counting sort is often used as a subroutine in radix sort. As we shall see in the next section, in order for radix sort to work correctly, counting sort must be stable. Exercises 8.2-1 Using Figure 8.2 as a model, illustrate the operation of COUNTING-SORT on the array A D h6; 0; 2; 0; 1; 3; 4; 6; 1; 3; 2i. 8.2-2 Prove that COUNTING-SORT is stable. 8.2-3 Suppose that we were to rewrite the for loop header in line 10 of the COUNTINGSORT as 10 for j D 1 to A:length Show that the algorithm still works properly. Is the modified algorithm stable? 8.3 Radix sort 197 8.2-4 Describe an algorithm that, given n integers in the range 0 to k, preprocesses its input and then answers any query about how many of the n integers fall into a range Œa : : b in O.1/ time. Your algorithm should use ‚.n C k/ preprocessing time. 8.3 Radix sort Radix sort is the algorithm used by the card-sorting machines you now find only in computer museums. The cards have 80 columns, and in each column a machine can punch a hole in one of 12 places. The sorter can be mechanically “programmed” to examine a given column of each card in a deck and distribute the card into one of 12 bins depending on which place has been punched. An operator can then gather the cards bin by bin, so that cards with the first place punched are on top of cards with the second place punched, and so on. For decimal digits, each column uses only 10 places. (The other two places are reserved for encoding nonnumeric characters.) A d -digit number would then occupy a field of d columns. Since the card sorter can look at only one column at a time, the problem of sorting n cards on a d -digit number requires a sorting algorithm. Intuitively, you might sort numbers on their most significant digit, sort each of the resulting bins recursively, and then combine the decks in order. Unfortunately, since the cards in 9 of the 10 bins must be put aside to sort each of the bins, this procedure generates many intermediate piles of cards that you would have to keep track of. (See Exercise 8.3-5.) Radix sort solves the problem of card sorting—counterintuitively—by sorting on the least significant digit first. The algorithm then combines the cards into a single deck, with the cards in the 0 bin preceding the cards in the 1 bin preceding the cards in the 2 bin, and so on. Then it sorts the entire deck again on the second-least significant digit and recombines the deck in a like manner. The process continues until the cards have been sorted on all d digits. Remarkably, at that point the cards are fully sorted on the d -digit number. Thus, only d passes through the deck are required to sort. Figure 8.3 shows how radix sort operates on a “deck” of seven 3-digit numbers. In order for radix sort to work correctly, the digit sorts must be stable. The sort performed by a card sorter is stable, but the operator has to be wary about not changing the order of the cards as they come out of a bin, even though all the cards in a bin have the same digit in the chosen column. 198 Chapter 8 Sorting in Linear Time 329 720 720 329 457 355 329 355 657 436 436 436 839 457 839 457 436 657 355 657 720 329 457 720 355 839 657 839 Figure 8.3 The operation of radix sort on a list of seven 3-digit numbers. The leftmost column is the input. The remaining columns show the list after successive sorts on increasingly significant digit positions. Shading indicates the digit position sorted on to produce each list from the previous one. In a typical computer, which is a sequential random-access machine, we sometimes use radix sort to sort records of information that are keyed by multiple fields. For example, we might wish to sort dates by three keys: year, month, and day. We could run a sorting algorithm with a comparison function that, given two dates, compares years, and if there is a tie, compares months, and if another tie occurs, compares days. Alternatively, we could sort the information three times with a stable sort: first on day, next on month, and finally on year. The code for radix sort is straightforward. The following procedure assumes that each element in the n-element array A has d digits, where digit 1 is the lowest-order digit and digit d is the highest-order digit. RADIX-SORT.A; d / 1 for i D 1 to d 2 use a stable sort to sort array A on digit i Lemma 8.3 Given n d -digit numbers in which each digit can take on up to k possible values, RADIX-SORT correctly sorts these numbers in ‚.d.n C k// time if the stable sort it uses takes ‚.n C k/ time. Proof The correctness of radix sort follows by induction on the column being sorted (see Exercise 8.3-3). The analysis of the running time depends on the stable sort used as the intermediate sorting algorithm. When each digit is in the range 0 to k 1 (so that it can take on k possible values), and k is not too large, counting sort is the obvious choice. Each pass over n d -digit numbers then takes time ‚.n C k/. There are d passes, and so the total time for radix sort is ‚.d.n C k//. When d is constant and k D O.n/, we can make radix sort run in linear time. More generally, we have some flexibility in how to break each key into digits. 8.3 Radix sort 199 Lemma 8.4 Given n b-bit numbers and any positive integer r Ä b, RADIX-SORT correctly sorts these numbers in ‚..b=r/.n C 2r // time if the stable sort it uses takes ‚.n C k/ time for inputs in the range 0 to k. Proof For a value r Ä b, we view each key as having d D db=re digits of r bits each. Each digit is an integer in the range 0 to 2r 1, so that we can use counting sort with k D 2r 1. (For example, we can view a 32-bit word as having four 8-bit digits, so that b D 32, r D 8, k D 2r 1 D 255, and d D b=r D 4.) Each pass of counting sort takes time ‚.n C k/ D ‚.n C 2r / and there are d passes, for a total running time of ‚.d.n C 2r // D ‚..b=r/.n C 2r //. For given values of n and b, we wish to choose the value of r, with r Ä b, that minimizes the expression .b=r/.n C 2r /. If b < blg nc, then for any value of r Ä b, we have that .n C 2r / D ‚.n/. Thus, choosing r D b yields a running time of .b=b/.n C 2b/ D ‚.n/, which is asymptotically optimal. If b blg nc, then choosing r D blg nc gives the best time to within a constant factor, which we can see as follows. Choosing r D blg nc yields a running time of ‚.bn= lg n/. As we increase r above blg nc, the 2r term in the numerator increases faster than the r term in the denominator, and so increasing r above blg nc yields a running time of .bn= lg n/. If instead we were to decrease r below blg nc, then the b=r term increases and the n C 2r term remains at ‚.n/. Is radix sort preferable to a comparison-based sorting algorithm, such as quicksort? If b D O.lg n/, as is often the case, and we choose r lg n, then radix sort’s running time is ‚.n/, which appears to be better than quicksort’s expected running time of ‚.n lg n/. The constant factors hidden in the ‚-notation differ, however. Although radix sort may make fewer passes than quicksort over the n keys, each pass of radix sort may take significantly longer. Which sorting algorithm we prefer depends on the characteristics of the implementations, of the underlying machine (e.g., quicksort often uses hardware caches more effectively than radix sort), and of the input data. Moreover, the version of radix sort that uses counting sort as the intermediate stable sort does not sort in place, which many of the ‚.n lg n/-time comparison sorts do. Thus, when primary memory storage is at a premium, we might prefer an in-place algorithm such as quicksort. Exercises 8.3-1 Using Figure 8.3 as a model, illustrate the operation of RADIX-SORT on the following list of English words: COW, DOG, SEA, RUG, ROW, MOB, BOX, TAB, BAR, EAR, TAR, DIG, BIG, TEA, NOW, FOX. 200 Chapter 8 Sorting in Linear Time 8.3-2 Which of the following sorting algorithms are stable: insertion sort, merge sort, heapsort, and quicksort? Give a simple scheme that makes any sorting algorithm stable. How much additional time and space does your scheme entail? 8.3-3 Use induction to prove that radix sort works. Where does your proof need the assumption that the intermediate sort is stable? 8.3-4 Show how to sort n integers in the range 0 to n3 1 in O.n/ time. 8.3-5 ? In the first card-sorting algorithm in this section, exactly how many sorting passes are needed to sort d -digit decimal numbers in the worst case? How many piles of cards would an operator need to keep track of in the worst case? 8.4 Bucket sort Bucket sort assumes that the input is drawn from a uniform distribution and has an average-case running time of O.n/. Like counting sort, bucket sort is fast because it assumes something about the input. Whereas counting sort assumes that the input consists of integers in a small range, bucket sort assumes that the input is generated by a random process that distributes elements uniformly and independently over the interval Œ0; 1/. (See Section C.2 for a definition of uniform distribution.) Bucket sort divides the interval Œ0; 1/ into n equal-sized subintervals, or buckets, and then distributes the n input numbers into the buckets. Since the inputs are uniformly and independently distributed over Œ0; 1/, we do not expect many numbers to fall into each bucket. To produce the output, we simply sort the numbers in each bucket and then go through the buckets in order, listing the elements in each. Our code for bucket sort assumes that the input is an n-element array A and that each element AŒi in the array satisfies 0 Ä AŒi < 1. The code requires an auxiliary array BŒ0 : : n 1 of linked lists (buckets) and assumes that there is a mechanism for maintaining such lists. (Section 10.2 describes how to implement basic operations on linked lists.) 8.4 Bucket sort 201 A 1 .78 B 0 2 .17 1 .12 .17 3 .39 2 .21 .23 .26 4 .26 3 .39 5 .72 4 6 .94 5 7 .21 6 .68 8 .12 7 .72 .78 9 .23 8 10 .68 9 .94 (a) (b) Figure 8.4 The operation of BUCKET-SORT for n D 10. (a) The input array AŒ1 : : 10. (b) The array BŒ0 : : 9 of sorted lists (buckets) after line 8 of the algorithm. Bucket i holds values in the half-open interval Œi=10; .i C 1/=10/. The sorted output consists of a concatenation in order of the lists BŒ0; BŒ1; : : : ; BŒ9. BUCKET-SORT.A/ 1 let BŒ0 : : n 1 be a new array 2 n D A:length 3 for i D 0 to n 1 4 make BŒi an empty list 5 for i D 1 to n 6 insert AŒi into list BŒbnAŒic 7 for i D 0 to n 1 8 sort list BŒi with insertion sort 9 concatenate the lists BŒ0; BŒ1; : : : ; BŒn 1 together in order Figure 8.4 shows the operation of bucket sort on an input array of 10 numbers. To see that this algorithm works, consider two elements AŒi and AŒj . Assume without loss of generality that AŒi Ä AŒj . Since bnAŒic Ä bnAŒj c, either element AŒi goes into the same bucket as AŒj  or it goes into a bucket with a lower index. If AŒi and AŒj  go into the same bucket, then the for loop of lines 7–8 puts them into the proper order. If AŒi and AŒj  go into different buckets, then line 9 puts them into the proper order. Therefore, bucket sort works correctly. To analyze the running time, observe that all lines except line 8 take O.n/ time in the worst case. We need to analyze the total time taken by the n calls to insertion sort in line 8. 202 Chapter 8 Sorting in Linear Time To analyze the cost of the calls to insertion sort, let ni be the random variable denoting the number of elements placed in bucket BŒi. Since insertion sort runs in quadratic time (see Section 2.2), the running time of bucket sort is X n 1 T .n/ D ‚.n/ C O.n2i / : i D0 We now analyze the average-case running time of bucket sort, by computing the expected value of the running time, where we take the expectation over the input distribution. Taking expectations of both sides and using linearity of expectation, we have " # X n 1 E ŒT .n/ D E ‚.n/ C O.n2i / i D0 X n 1 D ‚.n/ C E O.n2i / (by linearity of expectation) i D0 X n 1 D ‚.n/ C O E n2i (by equation (C.22)) . (8.1) i D0 We claim that E n2i D 2 1=n (8.2) for i D 0; 1; : : : ; n 1. It is no surprise that each bucket i has the same value of E Œn2i , since each value in the input array A is equally likely to fall in any bucket. To prove equation (8.2), we define indicator random variables Xij D I fAŒj  falls in bucket i g for i D 0; 1; : : : ; n 1 and j D 1; 2; : : : ; n. Thus, Xn ni D Xij : j D1 To compute E Œn2i , we expand the square and regroup terms: 8.4 Bucket sort 203 E n2i " Xn !2# DE Xij " j D1 # Xn Xn DE Xij Xik 2X j Dn 1 kD1 X D E4 Xi2j C X 3 Xij Xik5 j D1 1Äj Än 1ÄkÄn k¤j Xn XX D E Xi2j C E ŒXij Xik ; j D1 1Äj Än 1ÄkÄn k¤j (8.3) where the last line follows by linearity of expectation. We evaluate the two sum- mations separately. Indicator random variable Xij is 1 with probability 1=n and 0 otherwise, and therefore E Xi2j Âà D 12 1 C 02 1 1 n n 1 D: n When k ¤ j , the variables Xij and Xik are independent, and hence E ŒXij Xik D E ŒXij  E ŒXik D 11 nn 1 D: n2 Substituting these two expected values in equation (8.3), we obtain E n2i Xn 1 X X 1 D C n n2 j D1 1Äj Än 1ÄkÄn k¤j 1 D n C n.n n n1 D 1C n 1 D2 ; n 1 1/ n2 which proves equation (8.2). 204 Chapter 8 Sorting in Linear Time Using this expected value in equation (8.1), we conclude that the average-case running time for bucket sort is ‚.n/ C n O.2 1=n/ D ‚.n/. Even if the input is not drawn from a uniform distribution, bucket sort may still run in linear time. As long as the input has the property that the sum of the squares of the bucket sizes is linear in the total number of elements, equation (8.1) tells us that bucket sort will run in linear time. Exercises 8.4-1 Using Figure 8.4 as a model, illustrate the operation of BUCKET-SORT on the array A D h:79; :13; :16; :64; :39; :20; :89; :53; :71; :42i. 8.4-2 Explain why the worst-case running time for bucket sort is ‚.n2/. What simple change to the algorithm preserves its linear average-case running time and makes its worst-case running time O.n lg n/? 8.4-3 Let X be a random variable that is equal to the number of heads in two flips of a fair coin. What is E ŒX 2? What is E2 ŒX ? 8.4-4 ? We are given n points in the unit circle, pi D .xi ; yi /, such that 0 < xi2 C yi2 Ä 1 for i D 1; 2; : : : ; n. Suppose that the points are uniformly distributed; that is, the probability of finding a point in any region of the circle is proportional to the area of that region. Design an algorithm withpan average-case running time of ‚.n/ to sort the n points by their distances di D xi2 C yi2 from the origin. (Hint: Design the bucket sizes in BUCKET-SORT to reflect the uniform distribution of the points in the unit circle.) 8.4-5 ? A probability distribution function P .x/ for a random variable X is defined by P .x/ D Pr fX Ä xg. Suppose that we draw a list of n random variables X1; X2; : : : ; Xn from a continuous probability distribution function P that is computable in O.1/ time. Give an algorithm that sorts these numbers in linear averagecase time. Problems for Chapter 8 205 Problems 8-1 Probabilistic lower bounds on comparison sorting In this problem, we prove a probabilistic .n lg n/ lower bound on the running time of any deterministic or randomized comparison sort on n distinct input elements. We begin by examining a deterministic comparison sort A with decision tree TA. We assume that every permutation of A’s inputs is equally likely. a. Suppose that each leaf of TA is labeled with the probability that it is reached given a random input. Prove that exactly nŠ leaves are labeled 1=nŠ and that the rest are labeled 0. b. Let D.T / denote the external path length of a decision tree T ; that is, D.T / is the sum of the depths of all the leaves of T . Let T be a decision tree with k > 1 leaves, and let LT and RT be the left and right subtrees of T . Show that D.T / D D.LT/ C D.RT/ C k. c. Let d.k/ be the minimum value of D.T / over all decision trees T with k > 1 leaves. Show that d.k/ D min1ÄiÄk 1 fd.i / C d.k i / C kg. (Hint: Consider a decision tree T with k leaves that achieves the minimum. Let i0 be the number of leaves in LT and k i0 the number of leaves in RT.) d. Prove that for a given value of k > 1 and i in the range 1 Ä i Ä k 1, the function i lg i C .k i/ lg.k i/ is minimized at i D k=2. Conclude that d.k/ D .k lg k/. e. Prove that D.TA/ D .nŠ lg.nŠ//, and conclude that the average-case time to sort n elements is .n lg n/. Now, consider a randomized comparison sort B. We can extend the decisiontree model to handle randomization by incorporating two kinds of nodes: ordinary comparison nodes and “randomization” nodes. A randomization node models a random choice of the form RANDOM.1; r/ made by algorithm B; the node has r children, each of which is equally likely to be chosen during an execution of the algorithm. f. Show that for any randomized comparison sort B, there exists a deterministic comparison sort A whose expected number of comparisons is no more than those made by B. 206 Chapter 8 Sorting in Linear Time 8-2 Sorting in place in linear time Suppose that we have an array of n data records to sort and that the key of each record has the value 0 or 1. An algorithm for sorting such a set of records might possess some subset of the following three desirable characteristics: 1. The algorithm runs in O.n/ time. 2. The algorithm is stable. 3. The algorithm sorts in place, using no more than a constant amount of storage space in addition to the original array. a. Give an algorithm that satisfies criteria 1 and 2 above. b. Give an algorithm that satisfies criteria 1 and 3 above. c. Give an algorithm that satisfies criteria 2 and 3 above. d. Can you use any of your sorting algorithms from parts (a)–(c) as the sorting method used in line 2 of RADIX-SORT, so that RADIX-SORT sorts n records with b-bit keys in O.bn/ time? Explain how or why not. e. Suppose that the n records have keys in the range from 1 to k. Show how to modify counting sort so that it sorts the records in place in O.n C k/ time. You may use O.k/ storage outside the input array. Is your algorithm stable? (Hint: How would you do it for k D 3?) 8-3 Sorting variable-length items a. You are given an array of integers, where different integers may have different numbers of digits, but the total number of digits over all the integers in the array is n. Show how to sort the array in O.n/ time. b. You are given an array of strings, where different strings may have different numbers of characters, but the total number of characters over all the strings is n. Show how to sort the strings in O.n/ time. (Note that the desired order here is the standard alphabetical order; for example, a < ab < b.) 8-4 Water jugs Suppose that you are given n red and n blue water jugs, all of different shapes and sizes. All red jugs hold different amounts of water, as do the blue ones. Moreover, for every red jug, there is a blue jug that holds the same amount of water, and vice versa. Problems for Chapter 8 207 Your task is to find a grouping of the jugs into pairs of red and blue jugs that hold the same amount of water. To do so, you may perform the following operation: pick a pair of jugs in which one is red and one is blue, fill the red jug with water, and then pour the water into the blue jug. This operation will tell you whether the red or the blue jug can hold more water, or that they have the same volume. Assume that such a comparison takes one time unit. Your goal is to find an algorithm that makes a minimum number of comparisons to determine the grouping. Remember that you may not directly compare two red jugs or two blue jugs. a. Describe a deterministic algorithm that uses ‚.n2/ comparisons to group the jugs into pairs. b. Prove a lower bound of .n lg n/ for the number of comparisons that an algorithm solving this problem must make. c. Give a randomized algorithm whose expected number of comparisons is O.n lg n/, and prove that this bound is correct. What is the worst-case number of comparisons for your algorithm? 8-5 Average sorting Suppose that, instead of sorting an array, we just require that the elements increase on average. More precisely, we call an n-element array A k-sorted if, for all i D 1; 2; : : : ; n k, the following holds: Pi Ck j Di 1 AŒj  Ä Pi j Ck Di C1 AŒj  : k k a. What does it mean for an array to be 1-sorted? b. Give a permutation of the numbers 1; 2; : : : ; 10 that is 2-sorted, but not sorted. c. Prove that an n-element array is k-sorted if and only if AŒi Ä AŒi C k for all i D 1; 2; : : : ; n k. d. Give an algorithm that k-sorts an n-element array in O.n lg.n=k// time. We can also show a lower bound on the time to produce a k-sorted array, when k is a constant. e. Show that we can sort a k-sorted array of length n in O.n lg k/ time. (Hint: Use the solution to Exercise 6.5-9. ) f. Show that when k is a constant, k-sorting an n-element array requires .n lg n/ time. (Hint: Use the solution to the previous part along with the lower bound on comparison sorts.) 208 Chapter 8 Sorting in Linear Time 8-6 Lower bound on merging sorted lists The problem of merging two sorted lists arises frequently. We have seen a procedure for it as the subroutine MERGE in Section 2.3.1. In this problem, we will prove a lower bound of 2n 1 on the worst-case number of comparisons required to merge two sorted lists, each containing n items. First we will show a lower bound of 2n o.n/ comparisons by using a decision tree. a. Given 2n numbers, compute the number of possible ways to divide them into two sorted lists, each with n numbers. b. Using a decision tree and your answer to part (a), show that any algorithm that correctly merges two sorted lists must perform at least 2n o.n/ comparisons. Now we will show a slightly tighter 2n 1 bound. c. Show that if two elements are consecutive in the sorted order and from different lists, then they must be compared. d. Use your answer to the previous part to show a lower bound of 2n 1 comparisons for merging two sorted lists. 8-7 The 0-1 sorting lemma and columnsort A compare-exchange operation on two array elements AŒi and AŒj , where i < j , has the form COMPARE-EXCHANGE.A; i; j / 1 if AŒi > AŒj  2 exchange AŒi with AŒj  After the compare-exchange operation, we know that AŒi Ä AŒj . An oblivious compare-exchange algorithm operates solely by a sequence of prespecified compare-exchange operations. The indices of the positions compared in the sequence must be determined in advance, and although they can depend on the number of elements being sorted, they cannot depend on the values being sorted, nor can they depend on the result of any prior compare-exchange operation. For example, here is insertion sort expressed as an oblivious compare-exchange algorithm: IN SERTIO N -S O RT .A/ 1 for j D 2 to A:length 2 for i D j 1 downto 1 3 COMPARE-EXCHANGE.A; i; i C 1/ Problems for Chapter 8 209 The 0-1 sorting lemma provides a powerful way to prove that an oblivious compare-exchange algorithm produces a sorted result. It states that if an oblivious compare-exchange algorithm correctly sorts all input sequences consisting of only 0s and 1s, then it correctly sorts all inputs containing arbitrary values. You will prove the 0-1 sorting lemma by proving its contrapositive: if an oblivious compare-exchange algorithm fails to sort an input containing arbitrary values, then it fails to sort some 0-1 input. Assume that an oblivious compare-exchange algorithm X fails to correctly sort the array AŒ1 : : n. Let AŒp be the smallest value in A that algorithm X puts into the wrong location, and let AŒq be the value that algorithm X moves to the location into which AŒp should have gone. Define an array BŒ1 : : n of 0s and 1s as follows: ( BŒi D 0 if AŒi Ä AŒp ; 1 if AŒi > AŒp : a. Argue that AŒq > AŒp, so that BŒp D 0 and BŒq D 1. b. To complete the proof of the 0-1 sorting lemma, prove that algorithm X fails to sort array B correctly. Now you will use the 0-1 sorting lemma to prove that a particular sorting algorithm works correctly. The algorithm, columnsort, works on a rectangular array of n elements. The array has r rows and s columns (so that n D rs), subject to three restrictions: r must be even, s must be a divisor of r, and r 2s2. When columnsort completes, the array is sorted in column-major order: reading down the columns, from left to right, the elements monotonically increase. Columnsort operates in eight steps, regardless of the value of n. The odd steps are all the same: sort each column individually. Each even step is a fixed permutation. Here are the steps: 1. Sort each column. 2. Transpose the array, but reshape it back to r rows and s columns. In other words, turn the leftmost column into the top r=s rows, in order; turn the next column into the next r=s rows, in order; and so on. 3. Sort each column. 4. Perform the inverse of the permutation performed in step 2. 210 Chapter 8 Sorting in Linear Time 10 14 5 8 7 17 12 1 6 16 9 11 4 15 2 18 3 13 (a) 412 835 10 7 6 12 9 11 16 14 13 18 15 17 (b) 4 8 10 12 16 18 137 9 14 15 256 11 13 17 (c) 136 257 4 8 10 9 13 15 11 14 17 12 16 18 (d) 1 4 11 3 8 14 6 10 17 2 9 12 5 13 16 7 15 18 (e) 1 4 11 2 8 12 3 9 14 5 10 16 6 13 17 7 15 18 (f) 5 10 16 6 13 17 7 15 18 1 4 11 2 8 12 3 9 14 (g) 4 10 16 5 11 17 6 12 18 1 7 13 2 8 14 3 9 15 (h) 1 7 13 2 8 14 3 9 15 4 10 16 5 11 17 6 12 18 (i) Figure 8.5 The steps of columnsort. (a) The input array with 6 rows and 3 columns. (b) After sorting each column in step 1. (c) After transposing and reshaping in step 2. (d) After sorting each column in step 3. (e) After performing step 4, which inverts the permutation from step 2. (f) After sorting each column in step 5. (g) After shifting by half a column in step 6. (h) After sorting each column in step 7. (i) After performing step 8, which inverts the permutation from step 6. The array is now sorted in column-major order. 5. Sort each column. 6. Shift the top half of each column into the bottom half of the same column, and shift the bottom half of each column into the top half of the next column to the right. Leave the top half of the leftmost column empty. Shift the bottom half of the last column into the top half of a new rightmost column, and leave the bottom half of this new column empty. 7. Sort each column. 8. Perform the inverse of the permutation performed in step 6. Figure 8.5 shows an example of the steps of columnsort with r D 6 and s D 3. (Even though this example violates the requirement that r 2s2, it happens to work.) c. Argue that we can treat columnsort as an oblivious compare-exchange algorithm, even if we do not know what sorting method the odd steps use. Although it might seem hard to believe that columnsort actually sorts, you will use the 0-1 sorting lemma to prove that it does. The 0-1 sorting lemma applies because we can treat columnsort as an oblivious compare-exchange algorithm. A Notes for Chapter 8 211 couple of definitions will help you apply the 0-1 sorting lemma. We say that an area of an array is clean if we know that it contains either all 0s or all 1s. Otherwise, the area might contain mixed 0s and 1s, and it is dirty. From here on, assume that the input array contains only 0s and 1s, and that we can treat it as an array with r rows and s columns. d. Prove that after steps 1–3, the array consists of some clean rows of 0s at the top, some clean rows of 1s at the bottom, and at most s dirty rows between them. e. Prove that after step 4, the array, read in column-major order, starts with a clean area of 0s, ends with a clean area of 1s, and has a dirty area of at most s2 elements in the middle. f. Prove that steps 5–8 produce a fully sorted 0-1 output. Conclude that columnsort correctly sorts all inputs containing arbitrary values. g. Now suppose that s does not divide r. Prove that after steps 1–3, the array consists of some clean rows of 0s at the top, some clean rows of 1s at the bottom, and at most 2s 1 dirty rows between them. How large must r be, compared with s, for columnsort to correctly sort when s does not divide r? h. Suggest a simple change to step 1 that allows us to maintain the requirement that r 2s2 even when s does not divide r, and prove that with your change, columnsort correctly sorts. Chapter notes The decision-tree model for studying comparison sorts was introduced by Ford and Johnson [110]. Knuth’s comprehensive treatise on sorting [211] covers many variations on the sorting problem, including the information-theoretic lower bound on the complexity of sorting given here. Ben-Or [39] studied lower bounds for sorting using generalizations of the decision-tree model. Knuth credits H. H. Seward with inventing counting sort in 1954, as well as with the idea of combining counting sort with radix sort. Radix sorting starting with the least significant digit appears to be a folk algorithm widely used by operators of mechanical card-sorting machines. According to Knuth, the first published reference to the method is a 1929 document by L. J. Comrie describing punched-card equipment. Bucket sorting has been in use since 1956, when the basic idea was proposed by E. J. Isaac and R. C. Singleton [188]. Munro and Raman [263] give a stable sorting algorithm that performs O.n1C / comparisons in the worst case, where 0 < Ä 1 is any fixed constant. Although 212 Chapter 8 Sorting in Linear Time any of the O.n lg n/-time algorithms make fewer comparisons, the algorithm by Munro and Raman moves data only O.n/ times and operates in place. The case of sorting n b-bit integers in o.n lg n/ time has been considered by many researchers. Several positive results have been obtained, each under slightly different assumptions about the model of computation and the restrictions placed on the algorithm. All the results assume that the computer memory is divided into addressable b-bit words. Fredman and Willard [115] introduced the fusion tree data structure and used it topsort n integers in O.n lg n= lg lg n/ time. This bound was later improved to O.n lg n/ time by Andersson [16]. These algorithms require the use of multiplication and several precomputed constants. Andersson, Hagerup, Nilsson, and Raman [17] have shown how to sort n integers in O.n lg lg n/ time without using multiplication, but their method requires storage that can be unbounded in terms of n. Using multiplicative hashing, we can reduce the storage needed to O.n/, but then the O.n lg lg n/ worst-case bound on the running time becomes an expected-time bound. Generalizing the exponential search trees of Andersson [16], Thorup [335] gave an O.n.lg lg n/2/-time sorting algorithm that does not use multiplication or randomization, and it uses linear space. Combining these techniques with some new ideas, Han [158] improved the bound for sorting to O.n lg lg n lg lg lg n/ time. Although these algorithms are important theoretical breakthroughs, they are all fairly complicated and at the present time seem unlikely to compete with existing sorting algorithms in practice. The columnsort algorithm in Problem 8-7 is by Leighton [227]. 9 Medians and Order Statistics The ith order statistic of a set of n elements is the ith smallest element. For example, the minimum of a set of elements is the first order statistic (i D 1), and the maximum is the nth order statistic (i D n). A median, informally, is the “halfway point” of the set. When n is odd, the median is unique, occurring at i D .n C 1/=2. When n is even, there are two medians, occurring at i D n=2 and i D n=2C1. Thus, regardless of the parity of n, medians occur at i D b.n C 1/=2c (the lower median) and i D d.n C 1/=2e (the upper median). For simplicity in this text, however, we consistently use the phrase “the median” to refer to the lower median. This chapter addresses the problem of selecting the ith order statistic from a set of n distinct numbers. We assume for convenience that the set contains distinct numbers, although virtually everything that we do extends to the situation in which a set contains repeated values. We formally specify the selection problem as follows: Input: A set A of n (distinct) numbers and an integer i, with 1 Ä i Ä n. Output: The element x 2 A that is larger than exactly i 1 other elements of A. We can solve the selection problem in O.n lg n/ time, since we can sort the numbers using heapsort or merge sort and then simply index the ith element in the output array. This chapter presents faster algorithms. In Section 9.1, we examine the problem of selecting the minimum and maximum of a set of elements. More interesting is the general selection problem, which we investigate in the subsequent two sections. Section 9.2 analyzes a practical randomized algorithm that achieves an O.n/ expected running time, assuming distinct elements. Section 9.3 contains an algorithm of more theoretical interest that achieves the O.n/ running time in the worst case. 214 Chapter 9 Medians and Order Statistics 9.1 Minimum and maximum How many comparisons are necessary to determine the minimum of a set of n elements? We can easily obtain an upper bound of n 1 comparisons: examine each element of the set in turn and keep track of the smallest element seen so far. In the following procedure, we assume that the set resides in array A, where A:length D n. MINIMUM.A/ 1 min D AŒ1 2 for i D 2 to A:length 3 if min > AŒi 4 min D AŒi 5 return min We can, of course, find the maximum with n 1 comparisons as well. Is this the best we can do? Yes, since we can obtain a lower bound of n 1 comparisons for the problem of determining the minimum. Think of any algorithm that determines the minimum as a tournament among the elements. Each comparison is a match in the tournament in which the smaller of the two elements wins. Observing that every element except the winner must lose at least one match, we conclude that n 1 comparisons are necessary to determine the minimum. Hence, the algorithm MINIMUM is optimal with respect to the number of comparisons performed. Simultaneous minimum and maximum In some applications, we must find both the minimum and the maximum of a set of n elements. For example, a graphics program may need to scale a set of .x; y/ data to fit onto a rectangular display screen or other graphical output device. To do so, the program must first determine the minimum and maximum value of each coordinate. At this point, it should be obvious how to determine both the minimum and the maximum of n elements using ‚.n/ comparisons, which is asymptotically optimal: simply find the minimum and maximum independently, using n 1 comparisons for each, for a total of 2n 2 comparisons. In fact, we can find both the minimum and the maximum using at most 3 bn=2c comparisons. We do so by maintaining both the minimum and maximum elements seen thus far. Rather than processing each element of the input by comparing it against the current minimum and maximum, at a cost of 2 comparisons per element, 9.2 Selection in expected linear time 215 we process elements in pairs. We compare pairs of elements from the input first with each other, and then we compare the smaller with the current minimum and the larger to the current maximum, at a cost of 3 comparisons for every 2 elements. How we set up initial values for the current minimum and maximum depends on whether n is odd or even. If n is odd, we set both the minimum and maximum to the value of the first element, and then we process the rest of the elements in pairs. If n is even, we perform 1 comparison on the first 2 elements to determine the initial values of the minimum and maximum, and then process the rest of the elements in pairs as in the case for odd n. Let us analyze the total number of comparisons. If n is odd, then we perform 3 bn=2c comparisons. If n is even, we perform 1 initial comparison followed by 3.n 2/=2 comparisons, for a total of 3n=2 2. Thus, in either case, the total number of comparisons is at most 3 bn=2c. Exercises 9.1-1 Show that the second smallest of n elements can be found with n C dlg ne 2 comparisons in the worst case. (Hint: Also find the smallest element.) 9.1-2 ? Prove the lower bound of d3n=2e 2 comparisons in the worst case to find both the maximum and minimum of n numbers. (Hint: Consider how many numbers are potentially either the maximum or minimum, and investigate how a comparison affects these counts.) 9.2 Selection in expected linear time The general selection problem appears more difficult than the simple problem of finding a minimum. Yet, surprisingly, the asymptotic running time for both problems is the same: ‚.n/. In this section, we present a divide-and-conquer algorithm for the selection problem. The algorithm RANDOMIZED-SELECT is modeled after the quicksort algorithm of Chapter 7. As in quicksort, we partition the input array recursively. But unlike quicksort, which recursively processes both sides of the partition, RANDOMIZED-SELECT works on only one side of the partition. This difference shows up in the analysis: whereas quicksort has an expected running time of ‚.n lg n/, the expected running time of RANDOMIZED-SELECT is ‚.n/, assuming that the elements are distinct. 216 Chapter 9 Medians and Order Statistics RANDOMIZED-SELECT uses the procedure RANDOMIZED-PARTITION introduced in Section 7.3. Thus, like RANDOMIZED-QUICKSORT, it is a randomized algorithm, since its behavior is determined in part by the output of a random-number generator. The following code for RANDOMIZED-SELECT returns the i th smallest element of the array AŒp : : r. RANDOMIZED-SELECT .A; p; r; i / 1 if p == r 2 return AŒp 3 q D RANDOMIZED-PARTITION.A; p; r / 4 k D q pC1 5 if i == k // the pivot value is the answer 6 return AŒq 7 elseif i < k 8 return RANDOMIZED-SELECT.A; p; q 1; i / 9 else return RANDOMIZED-SELECT.A; q C 1; r; i k/ The RANDOMIZED-SELECT procedure works as follows. Line 1 checks for the base case of the recursion, in which the subarray AŒp : : r consists of just one element. In this case, i must equal 1, and we simply return AŒp in line 2 as the i th smallest element. Otherwise, the call to RANDOMIZED-PARTITION in line 3 partitions the array AŒp : : r into two (possibly empty) subarrays AŒp : : q 1 and AŒq C 1 : : r such that each element of AŒp : : q 1 is less than or equal to AŒq, which in turn is less than each element of AŒq C 1 : : r. As in quicksort, we will refer to AŒq as the pivot element. Line 4 computes the number k of elements in the subarray AŒp : : q, that is, the number of elements in the low side of the partition, plus one for the pivot element. Line 5 then checks whether AŒq is the ith smallest element. If it is, then line 6 returns AŒq. Otherwise, the algorithm determines in which of the two subarrays AŒp : : q 1 and AŒq C 1 : : r the ith smallest element lies. If i < k, then the desired element lies on the low side of the partition, and line 8 recursively selects it from the subarray. If i > k, however, then the desired element lies on the high side of the partition. Since we already know k values that are smaller than the ith smallest element of AŒp : : r—namely, the elements of AŒp : : q—the desired element is the .i k/th smallest element of AŒq C 1 : : r, which line 9 finds recursively. The code appears to allow recursive calls to subarrays with 0 elements, but Exercise 9.2-1 asks you to show that this situation cannot happen. The worst-case running time for RANDOMIZED-SELECT is ‚.n2/, even to find the minimum, because we could be extremely unlucky and always partition around the largest remaining element, and partitioning takes ‚.n/ time. We will see that 9.2 Selection in expected linear time 217 the algorithm has a linear expected running time, though, and because it is randomized, no particular input elicits the worst-case behavior. To analyze the expected running time of RANDOMIZED-SELECT, we let the running time on an input array AŒp : : r of n elements be a random variable that we denote by T .n/, and we obtain an upper bound on E ŒT .n/ as follows. The procedure RANDOMIZED-PARTITION is equally likely to return any element as the pivot. Therefore, for each k such that 1 Ä k Ä n, the subarray AŒp : : q has k elements (all less than or equal to the pivot) with probability 1=n. For k D 1; 2; : : : ; n, we define indicator random variables Xk where Xk D I fthe subarray AŒp : : q has exactly k elementsg ; and so, assuming that the elements are distinct, we have E ŒXk D 1=n : (9.1) When we call RANDOMIZED-SELECT and choose AŒq as the pivot element, we do not know, a priori, if we will terminate immediately with the correct answer, recurse on the subarray AŒp : : q 1, or recurse on the subarray AŒq C 1 : : r. This decision depends on where the ith smallest element falls relative to AŒq. Assuming that T .n/ is monotonically increasing, we can upper-bound the time needed for the recursive call by the time needed for the recursive call on the largest possible input. In other words, to obtain an upper bound, we assume that the ith element is always on the side of the partition with the greater number of elements. For a given call of RANDOMIZED-SELECT, the indicator random variable Xk has the value 1 for exactly one value of k, and it is 0 for all other k. When Xk D 1, the two subarrays on which we might recurse have sizes k 1 and n k. Hence, we have the recurrence Xn T .n/ Ä Xk .T .max.k 1; n k// C O.n// kD1 Xn D Xk T .max.k 1; n k// C O.n/ : kD1 218 Chapter 9 Medians and Order Statistics Taking expected values, we have E ŒT .n/" # Xn Ä E Xk T .max.k 1; n k// C O.n/ kD1 Xn D E ŒXk T .max.k 1; n k// C O.n/ (by linearity of expectation) kD1 Xn D E ŒXk E ŒT .max.k 1; n k// C O.n/ (by equation (C.24)) kD1 Xn 1 D E ŒT .max.k 1; n k// C O.n/ n kD1 (by equation (9.1)) . In order to apply equation (C.24), we rely on Xk and T .max.k 1; n k// being independent random variables. Exercise 9.2-2 asks you to justify this assertion. Let us consider the expression max.k 1; n k/. We have ( max.k 1; n k/ D k 1 if k > dn=2e ; n k if k Ä dn=2e : If n is even, each term from T .dn=2e/ up to T .n 1/ appears exactly twice in the summation, and if n is odd, all these terms appear twice and T .bn=2c/ appears once. Thus, we have 2 X n 1 E ŒT .n/ Ä E ŒT .k/ C O.n/ : n kDbn=2c We show that E ŒT .n/ D O.n/ by substitution. Assume that E ŒT .n/ Ä cn for some constant c that satisfies the initial conditions of the recurrence. We assume that T .n/ D O.1/ for n less than some constant; we shall pick this constant later. We also pick a constant a such that the function described by the O.n/ term above (which describes the non-recursive component of the running time of the algorithm) is bounded from above by an for all n > 0. Using this inductive hypothesis, we have 2 X n 1 E ŒT .n/ Ä ck C an n kDbn=2c 2c X n 1 bnX =2c 1 ! D k k C an n kD1 kD1 9.2 Selection in expected linear time 219  à 2c .n 1/n .bn=2c 1/ bn=2c D C an n 2 2 à 2c .n 1/n .n=2 2/.n=2 1/ Ä C an n 2 2 2c  n2 n n2=4 à 3n=2 C 2 D C an n c 3n2 2 n à 2 D C 2 C an n 4 2 à 3n 1 2 Dc C C an 4 2n 3cn c Ä C C an 4 D cn 2 cn c Á an : 42 In order to complete the proof, we need to show that for sufficiently large n, this last expression is at most cn or, equivalently, that cn=4 c=2 an 0. If we add c=2 to both sides and factor out n, we get n.c=4 a/ c=2. As long as we choose the constant c so that c=4 a > 0, i.e., c > 4a, we can divide both sides by c=4 a, giving c=2 2c n D : c=4 a c 4a Thus, if we assume that T .n/ D O.1/ for n < 2c=.c 4a/, then E ŒT .n/ D O.n/. We conclude that we can find any order statistic, and in particular the median, in expected linear time, assuming that the elements are distinct. Exercises 9.2-1 Show that RANDOMIZED-SELECT never makes a recursive call to a 0-length array. 9.2-2 Argue that the indicator random variable Xk and the value T .max.k 1; n k// are independent. 9.2-3 Write an iterative version of RANDOMIZED-SELECT. 220 Chapter 9 Medians and Order Statistics 9.2-4 Suppose we use RANDOMIZED-SELECT to select the minimum element of the array A D h3; 2; 9; 0; 7; 5; 4; 8; 6; 1i. Describe a sequence of partitions that results in a worst-case performance of RANDOMIZED-SELECT. 9.3 Selection in worst-case linear time We now examine a selection algorithm whose running time is O.n/ in the worst case. Like RANDOMIZED-SELECT, the algorithm SELECT finds the desired element by recursively partitioning the input array. Here, however, we guarantee a good split upon partitioning the array. SELECT uses the deterministic partitioning algorithm PARTITION from quicksort (see Section 7.1), but modified to take the element to partition around as an input parameter. The SELECT algorithm determines the i th smallest of an input array of n > 1 distinct elements by executing the following steps. (If n D 1, then SELECT merely returns its only input value as the ith smallest.) 1. Divide the n elements of the input array into bn=5c groups of 5 elements each and at most one group made up of the remaining n mod 5 elements. 2. Find the median of each of the dn=5e groups by first insertion-sorting the elements of each group (of which there are at most 5) and then picking the median from the sorted list of group elements. 3. Use SELECT recursively to find the median x of the dn=5e medians found in step 2. (If there are an even number of medians, then by our convention, x is the lower median.) 4. Partition the input array around the median-of-medians x using the modified version of PARTITION. Let k be one more than the number of elements on the low side of the partition, so that x is the kth smallest element and there are n k elements on the high side of the partition. 5. If i D k, then return x. Otherwise, use SELECT recursively to find the i th smallest element on the low side if i < k, or the .i k/th smallest element on the high side if i > k. To analyze the running time of SELECT, we first determine a lower bound on the number of elements that are greater than the partitioning element x. Figure 9.1 helps us to visualize this bookkeeping. At least half of the medians found in 9.3 Selection in worst-case linear time 221 x Figure 9.1 Analysis of the algorithm SELECT. The n elements are represented by small circles, and each group of 5 elements occupies a column. The medians of the groups are whitened, and the median-of-medians x is labeled. (When finding the median of an even number of elements, we use the lower median.) Arrows go from larger elements to smaller, from which we can see that 3 out of every full group of 5 elements to the right of x are greater than x, and 3 out of every group of 5 elements to the left of x are less than x. The elements known to be greater than x appear on a shaded background. step 2 are greater than or equal to the median-of-medians x.1 Thus, at least half of the dn=5e groups contribute at least 3 elements that are greater than x, except for the one group that has fewer than 5 elements if 5 does not divide n exactly, and the one group containing x itself. Discounting these two groups, it follows that the number of elements greater than x is at least  1 lnm à 3 2 25 3n 6: 10 Similarly, at least 3n=10 6 elements are less than x. Thus, in the worst case, step 5 calls SELECT recursively on at most 7n=10 C 6 elements. We can now develop a recurrence for the worst-case running time T .n/ of the algorithm SELECT. Steps 1, 2, and 4 take O.n/ time. (Step 2 consists of O.n/ calls of insertion sort on sets of size O.1/.) Step 3 takes time T .dn=5e/, and step 5 takes time at most T .7n=10 C 6/, assuming that T is monotonically increasing. We make the assumption, which seems unmotivated at first, that any input of fewer than 140 elements requires O.1/ time; the origin of the magic constant 140 will be clear shortly. We can therefore obtain the recurrence 1Because of our assumption that the numbers are distinct, all medians except x are either greater than or less than x. 222 Chapter 9 Medians and Order Statistics ( T .n/ Ä O.1/ if n < 140 ; T .dn=5e/ C T .7n=10 C 6/ C O.n/ if n 140 : We show that the running time is linear by substitution. More specifically, we will show that T .n/ Ä cn for some suitably large constant c and all n > 0. We begin by assuming that T .n/ Ä cn for some suitably large constant c and all n < 140; this assumption holds if c is large enough. We also pick a constant a such that the function described by the O.n/ term above (which describes the non-recursive component of the running time of the algorithm) is bounded above by an for all n > 0. Substituting this inductive hypothesis into the right-hand side of the recurrence yields T .n/ Ä c dn=5e C c.7n=10 C 6/ C an Ä cn=5 C c C 7cn=10 C 6c C an D 9cn=10 C 7c C an D cn C . cn=10 C 7c C an/ ; which is at most cn if cn=10 C 7c C an Ä 0 : (9.2) Inequality (9.2) is equivalent to the inequality c 10a.n=.n 70// when n > 70. Because we assume that n 140, we have n=.n 70/ Ä 2, and so choosing c 20a will satisfy inequality (9.2). (Note that there is nothing special about the constant 140; we could replace it by any integer strictly greater than 70 and then choose c accordingly.) The worst-case running time of SELECT is therefore linear. As in a comparison sort (see Section 8.1), SELECT and RANDOMIZED-SELECT determine information about the relative order of elements only by comparing elements. Recall from Chapter 8 that sorting requires .n lg n/ time in the comparison model, even on average (see Problem 8-1). The linear-time sorting algorithms in Chapter 8 make assumptions about the input. In contrast, the linear-time selection algorithms in this chapter do not require any assumptions about the input. They are not subject to the .n lg n/ lower bound because they manage to solve the selection problem without sorting. Thus, solving the selection problem by sorting and indexing, as presented in the introduction to this chapter, is asymptotically inefficient. 9.3 Selection in worst-case linear time 223 Exercises 9.3-1 In the algorithm SELECT, the input elements are divided into groups of 5. Will the algorithm work in linear time if they are divided into groups of 7? Argue that SELECT does not run in linear time if groups of 3 are used. 9.3-2 Analyze SELECT to show that if n 140, then at least dn=4e elements are greater than the median-of-medians x and at least dn=4e elements are less than x. 9.3-3 Show how quicksort can be made to run in O.n lg n/ time in the worst case, assuming that all elements are distinct. 9.3-4 ? Suppose that an algorithm uses only comparisons to find the ith smallest element in a set of n elements. Show that it can also find the i 1 smaller elements and the n i larger elements without performing any additional comparisons. 9.3-5 Suppose that you have a “black-box” worst-case linear-time median subroutine. Give a simple, linear-time algorithm that solves the selection problem for an arbitrary order statistic. 9.3-6 The kth quantiles of an n-element set are the k 1 order statistics that divide the sorted set into k equal-sized sets (to within 1). Give an O.n lg k/-time algorithm to list the kth quantiles of a set. 9.3-7 Describe an O.n/-time algorithm that, given a set S of n distinct numbers and a positive integer k Ä n, determines the k numbers in S that are closest to the median of S. 9.3-8 Let X Œ1 : : n and Y Œ1 : : n be two arrays, each containing n numbers already in sorted order. Give an O.lg n/-time algorithm to find the median of all 2n elements in arrays X and Y . 9.3-9 Professor Olay is consulting for an oil company, which is planning a large pipeline running east to west through an oil field of n wells. The company wants to connect 224 Chapter 9 Medians and Order Statistics Figure 9.2 Professor Olay needs to determine the position of the east-west oil pipeline that minimizes the total length of the north-south spurs. a spur pipeline from each well directly to the main pipeline along a shortest route (either north or south), as shown in Figure 9.2. Given the x- and y-coordinates of the wells, how should the professor pick the optimal location of the main pipeline, which would be the one that minimizes the total length of the spurs? Show how to determine the optimal location in linear time. Problems 9-1 Largest i numbers in sorted order Given a set of n numbers, we wish to find the i largest in sorted order using a comparison-based algorithm. Find the algorithm that implements each of the following methods with the best asymptotic worst-case running time, and analyze the running times of the algorithms in terms of n and i. a. Sort the numbers, and list the i largest. b. Build a max-priority queue from the numbers, and call EXTRACT-MAX i times. c. Use an order-statistic algorithm to find the ith largest number, partition around that number, and sort the i largest numbers. Problems for Chapter 9 225 9-2 Weighted median For that nPdniiDs1tiwncit elements x1; x2; : D 1, the weighted : : ; xn with positive weights w1; w2; : : : ; wn (lower) median is the element xk satisfying such X 1 wi < 2 xi xk For example, if the elements are 0:1; 0:35; 0:05; 0:1; 0:15; 0:05; 0:2 and each element equals its weight (that is, wi D xi for i D 1; 2; : : : ; 7), then the median is 0:1, but the weighted median is 0:2. a. Argue that the median of x1; x2; : : : ; xn is the weighted median of the xi with weights wi D 1=n for i D 1; 2; : : : ; n. b. Show how to compute the weighted median of n elements in O.n lg n/ worstcase time using sorting. c. Show how to compute the weighted median in ‚.n/ worst-case time using a linear-time median algorithm such as SELECT from Section 9.3. The post-office location problem is defined as follows. We are given n points p1; p2; : : : ; pn with (not necessarily one associated weights of the input points) w1; w2; : : : ; wn that minimizes .thWe esuwmisPh tnioDfi1 nwdi a point d.p; pi p /, where d.a; b/ is the distance between points a and b. d. Argue that the weighted median is a best solution for the 1-dimensional postoffice location problem, in which points are simply real numbers and the distance between points a and b is d.a; b/ D ja bj. e. Find the best solution for the 2-dimensional post-office location problem, in which the points are .x; y/ coordinate pairs and the distance between points a D .x1; y1/ and b D .x2; y2/ is the Manhattan distance given by d.a; b/ D jx1 x2j C jy1 y2j. 9-3 Small order statistics We showed that the worst-case number T .n/ of comparisons used by SELECT to select the ith order statistic from n numbers satisfies T .n/ D ‚.n/, but the constant hidden by the ‚-notation is rather large. When i is small relative to n, we can implement a different procedure that uses SELECT as a subroutine but makes fewer comparisons in the worst case. 226 Chapter 9 Medians and Order Statistics a. Describe an algorithm that uses Ui .n/ comparisons to find the ith smallest of n elements, where ( Ui .n/ D T .n/ if i n=2 ; bn=2c C Ui .dn=2e/ C T .2i / otherwise : (Hint: Begin with bn=2c disjoint pairwise comparisons, and recurse on the set containing the smaller element from each pair.) b. Show that, if i < n=2, then Ui .n/ D n C O.T .2i/ lg.n=i//. c. Show that if i is a constant less than n=2, then Ui .n/ D n C O.lg n/. d. Show that if i D n=k for k 2, then Ui .n/ D n C O.T .2n=k/ lg k/. 9-4 Alternative analysis of randomized selection In this problem, we use indicator random variables to analyze the RANDOMIZEDSELECT procedure in a manner akin to our analysis of RANDOMIZED-QUICKSORT in Section 7.4.2. As in the quicksort analysis, we assume that all elements are distinct, and we rename the elements of the input array A as ´1; ´2; : : : ; ´n, where ´i is the i th smallest element. Thus, the call RANDOMIZED-SELECT.A; 1; n; k/ returns ´k. For 1 Ä i < j Ä n, let Xijk D I f ´i is compared with ´j sometime during the execution of the algorithm to find ´kg : a. Give an exact expression for E ŒXijk. (Hint: Your expression may have different values, depending on the values of i, j , and k.) b. Let Xk denote the total number of comparisons between elements of array A when finding ´k. Show that Xk Xn E ŒXk Ä 2 j i D1 j Dk 1 Xn j C i C1 j j DkC1 k 1 X k 2 k C kC1 k i D1 ! i1 : i C1 c. Show that E ŒXk Ä 4n. d. Conclude that, assuming all elements of array A are distinct, RANDOMIZEDSELECT runs in expected time O.n/. Notes for Chapter 9 227 Chapter notes The worst-case linear-time median-finding algorithm was devised by Blum, Floyd, Pratt, Rivest, and Tarjan [50]. The fast randomized version is due to Hoare [169]. Floyd and Rivest [108] have developed an improved randomized version that partitions around an element recursively selected from a small sample of the elements. It is still unknown exactly how many comparisons are needed to determine the median. Bent and John [41] gave a lower bound of 2n comparisons for median finding, and Scho¨nhage, Paterson, and Pippenger [302] gave an upper bound of 3n. Dor and Zwick have improved on both of these bounds. Their upper bound [93] is slightly less than 2:95n, and their lower bound [94] is .2 C /n, for a small positive constant , thereby improving slightly on related work by Dor et al. [92]. Paterson [272] describes some of these results along with other related work. III Data Structures Introduction Sets are as fundamental to computer science as they are to mathematics. Whereas mathematical sets are unchanging, the sets manipulated by algorithms can grow, shrink, or otherwise change over time. We call such sets dynamic. The next five chapters present some basic techniques for representing finite dynamic sets and manipulating them on a computer. Algorithms may require several different types of operations to be performed on sets. For example, many algorithms need only the ability to insert elements into, delete elements from, and test membership in a set. We call a dynamic set that supports these operations a dictionary. Other algorithms require more complicated operations. For example, min-priority queues, which Chapter 6 introduced in the context of the heap data structure, support the operations of inserting an element into and extracting the smallest element from a set. The best way to implement a dynamic set depends upon the operations that must be supported. Elements of a dynamic set In a typical implementation of a dynamic set, each element is represented by an object whose attributes can be examined and manipulated if we have a pointer to the object. (Section 10.3 discusses the implementation of objects and pointers in programming environments that do not contain them as basic data types.) Some kinds of dynamic sets assume that one of the object’s attributes is an identifying key. If the keys are all different, we can think of the dynamic set as being a set of key values. The object may contain satellite data, which are carried around in other object attributes but are otherwise unused by the set implementation. It may 230 Part III Data Structures also have attributes that are manipulated by the set operations; these attributes may contain data or pointers to other objects in the set. Some dynamic sets presuppose that the keys are drawn from a totally ordered set, such as the real numbers, or the set of all words under the usual alphabetic ordering. A total ordering allows us to define the minimum element of the set, for example, or to speak of the next element larger than a given element in a set. Operations on dynamic sets Operations on a dynamic set can be grouped into two categories: queries, which simply return information about the set, and modifying operations, which change the set. Here is a list of typical operations. Any specific application will usually require only a few of these to be implemented. SEARCH.S; k/ A query that, given a set S and a key value k, returns a pointer x to an element in S such that x:key D k, or NIL if no such element belongs to S. INSERT.S; x/ A modifying operation that augments the set S with the element pointed to by x. We usually assume that any attributes in element x needed by the set implementation have already been initialized. DELETE.S; x/ A modifying operation that, given a pointer x to an element in the set S, removes x from S. (Note that this operation takes a pointer to an element x, not a key value.) MINIMUM.S / A query on a totally ordered set S that returns a pointer to the element of S with the smallest key. MAXIMUM.S / A query on a totally ordered set S that returns a pointer to the element of S with the largest key. SUCCESSOR.S; x/ A query that, given an element x whose key is from a totally ordered set S, returns a pointer to the next larger element in S, or NIL if x is the maximum element. PREDECESSOR.S; x/ A query that, given an element x whose key is from a totally ordered set S, returns a pointer to the next smaller element in S, or NIL if x is the minimum element. Part III Data Structures 231 In some situations, we can extend the queries SUCCESSOR and PREDECESSOR so that they apply to sets with nondistinct keys. For a set on n keys, the normal presumption is that a call to MINIMUM followed by n 1 calls to SUCCESSOR enumerates the elements in the set in sorted order. We usually measure the time taken to execute a set operation in terms of the size of the set. For example, Chapter 13 describes a data structure that can support any of the operations listed above on a set of size n in time O.lg n/. Overview of Part III Chapters 10–14 describe several data structures that we can use to implement dynamic sets; we shall use many of these later to construct efficient algorithms for a variety of problems. We already saw another important data structure—the heap—in Chapter 6. Chapter 10 presents the essentials of working with simple data structures such as stacks, queues, linked lists, and rooted trees. It also shows how to implement objects and pointers in programming environments that do not support them as primitives. If you have taken an introductory programming course, then much of this material should be familiar to you. Chapter 11 introduces hash tables, which support the dictionary operations INSERT, DELETE, and SEARCH. In the worst case, hashing requires ‚.n/ time to perform a SEARCH operation, but the expected time for hash-table operations is O.1/. The analysis of hashing relies on probability, but most of the chapter requires no background in the subject. Binary search trees, which are covered in Chapter 12, support all the dynamicset operations listed above. In the worst case, each operation takes ‚.n/ time on a tree with n elements, but on a randomly built binary search tree, the expected time for each operation is O.lg n/. Binary search trees serve as the basis for many other data structures. Chapter 13 introduces red-black trees, which are a variant of binary search trees. Unlike ordinary binary search trees, red-black trees are guaranteed to perform well: operations take O.lg n/ time in the worst case. A red-black tree is a balanced search tree; Chapter 18 in Part V presents another kind of balanced search tree, called a B-tree. Although the mechanics of red-black trees are somewhat intricate, you can glean most of their properties from the chapter without studying the mechanics in detail. Nevertheless, you probably will find walking through the code to be quite instructive. In Chapter 14, we show how to augment red-black trees to support operations other than the basic ones listed above. First, we augment them so that we can dynamically maintain order statistics for a set of keys. Then, we augment them in a different way to maintain intervals of real numbers. 10 Elementary Data Structures In this chapter, we examine the representation of dynamic sets by simple data structures that use pointers. Although we can construct many complex data structures using pointers, we present only the rudimentary ones: stacks, queues, linked lists, and rooted trees. We also show ways to synthesize objects and pointers from arrays. 10.1 Stacks and queues Stacks and queues are dynamic sets in which the element removed from the set by the DELETE operation is prespecified. In a stack, the element deleted from the set is the one most recently inserted: the stack implements a last-in, first-out, or LIFO, policy. Similarly, in a queue, the element deleted is always the one that has been in the set for the longest time: the queue implements a first-in, first-out, or FIFO, policy. There are several efficient ways to implement stacks and queues on a computer. In this section we show how to use a simple array to implement each. Stacks The INSERT operation on a stack is often called PUSH, and the DELETE operation, which does not take an element argument, is often called POP. These names are allusions to physical stacks, such as the spring-loaded stacks of plates used in cafeterias. The order in which plates are popped from the stack is the reverse of the order in which they were pushed onto the stack, since only the top plate is accessible. As Figure 10.1 shows, we can implement a stack of at most n elements with an array SŒ1 : : n. The array has an attribute S:top that indexes the most recently 10.1 Stacks and queues 1234567 S 15 6 2 9 S:top D 4 (a) 1234567 S 15 6 2 9 17 3 S:top D 6 (b) 233 1234567 S 15 6 2 9 17 3 S:top D 5 (c) Figure 10.1 An array implementation of a stack S. Stack elements appear only in the lightly shaded positions. (a) Stack S has 4 elements. The top element is 9. (b) Stack S after the calls PUSH.S; 17/ and PUSH.S; 3/. (c) Stack S after the call POP.S / has returned the element 3, which is the one most recently pushed. Although element 3 still appears in the array, it is no longer in the stack; the top is element 17. inserted element. The stack consists of elements SŒ1 : : S:top, where SŒ1 is the element at the bottom of the stack and SŒS:top is the element at the top. When S:top D 0, the stack contains no elements and is empty. We can test to see whether the stack is empty by query operation STACK-EMPTY. If we attempt to pop an empty stack, we say the stack underflows, which is normally an error. If S:top exceeds n, the stack overflows. (In our pseudocode implementation, we don’t worry about stack overflow.) We can implement each of the stack operations with just a few lines of code: STACK-EMPTY.S / 1 if S:top == 0 2 return TRUE 3 else return FALSE PUSH.S; x/ 1 S:top D S:top C 1 2 SŒS:top D x POP.S / 1 if STACK-EMPTY.S / 2 error “underflow” 3 else S:top D S:top 1 4 return SŒS:top C 1 Figure 10.1 shows the effects of the modifying operations PUSH and POP. Each of the three stack operations takes O.1/ time. 234 Chapter 10 Elementary Data Structures 1 2 3 4 5 6 7 8 9 10 11 12 (a) Q 15 6 9 8 4 Q:head D 7 Q:tail D 12 1 2 3 4 5 6 7 8 9 10 11 12 (b) Q 3 5 15 6 9 8 4 17 Q:tail D 3 Q:head D 7 1 2 3 4 5 6 7 8 9 10 11 12 (c) Q 3 5 15 6 9 8 4 17 Q:tail D 3 Q:head D 8 Figure 10.2 A queue implemented using an array QŒ1 : : 12. Queue elements appear only in the lightly shaded positions. (a) The queue has 5 elements, in locations QŒ7 : : 11. (b) The configuration of the queue after the calls ENQUEUE.Q; 17/, ENQUEUE.Q; 3/, and ENQUEUE.Q; 5/. (c) The configuration of the queue after the call DEQUEUE.Q/ returns the key value 15 formerly at the head of the queue. The new head has key 6. Queues We call the INSERT operation on a queue ENQUEUE, and we call the DELETE operation DEQUEUE; like the stack operation POP, DEQUEUE takes no element argument. The FIFO property of a queue causes it to operate like a line of customers waiting to pay a cashier. The queue has a head and a tail. When an element is enqueued, it takes its place at the tail of the queue, just as a newly arriving customer takes a place at the end of the line. The element dequeued is always the one at the head of the queue, like the customer at the head of the line who has waited the longest. Figure 10.2 shows one way to implement a queue of at most n 1 elements using an array QŒ1 : : n. The queue has an attribute Q:head that indexes, or points to, its head. The attribute Q:tail indexes the next location at which a newly arriving element will be inserted into the queue. The elements in the queue reside in locations Q:head; Q:head C 1; : : : ; Q:tail 1, where we “wrap around” in the sense that location 1 immediately follows location n in a circular order. When Q:head D Q:tail, the queue is empty. Initially, we have Q:head D Q:tail D 1. If we attempt to dequeue an element from an empty queue, the queue underflows. 10.1 Stacks and queues 235 When Q:head D Q:tail C 1, the queue is full, and if we attempt to enqueue an element, then the queue overflows. In our procedures ENQUEUE and DEQUEUE, we have omitted the error checking for underflow and overflow. (Exercise 10.1-4 asks you to supply code that checks for these two error conditions.) The pseudocode assumes that n D Q:length. ENQUEUE.Q; x/ 1 QŒQ:tail D x 2 if Q:tail == Q:length 3 Q:tail D 1 4 else Q:tail D Q:tail C 1 DEQUEUE.Q/ 1 x D QŒQ:head 2 if Q:head == Q:length 3 Q:head D 1 4 else Q:head D Q:head C 1 5 return x Figure 10.2 shows the effects of the ENQUEUE and DEQUEUE operations. Each operation takes O.1/ time. Exercises 10.1-1 Using Figure 10.1 as a model, illustrate the result of each operation in the sequence PUSH.S; 4/, PUSH.S; 1/, PUSH.S; 3/, POP.S /, PUSH.S; 8/, and POP.S / on an initially empty stack S stored in array SŒ1 : : 6. 10.1-2 Explain how to implement two stacks in one array AŒ1 : : n in such a way that neither stack overflows unless the total number of elements in both stacks together is n. The PUSH and POP operations should run in O.1/ time. 10.1-3 Using Figure 10.2 as a model, illustrate the result of each operation in the sequence ENQUEUE.Q; 4/, ENQUEUE.Q; 1/, ENQUEUE.Q; 3/, DEQUEUE.Q/, ENQUEUE.Q; 8/, and DEQUEUE.Q/ on an initially empty queue Q stored in array QŒ1 : : 6. 10.1-4 Rewrite ENQUEUE and DEQUEUE to detect underflow and overflow of a queue. 236 Chapter 10 Elementary Data Structures 10.1-5 Whereas a stack allows insertion and deletion of elements at only one end, and a queue allows insertion at one end and deletion at the other end, a deque (doubleended queue) allows insertion and deletion at both ends. Write four O.1/-time procedures to insert elements into and delete elements from both ends of a deque implemented by an array. 10.1-6 Show how to implement a queue using two stacks. Analyze the running time of the queue operations. 10.1-7 Show how to implement a stack using two queues. Analyze the running time of the stack operations. 10.2 Linked lists A linked list is a data structure in which the objects are arranged in a linear order. Unlike an array, however, in which the linear order is determined by the array indices, the order in a linked list is determined by a pointer in each object. Linked lists provide a simple, flexible representation for dynamic sets, supporting (though not necessarily efficiently) all the operations listed on page 230. As shown in Figure 10.3, each element of a doubly linked list L is an object with an attribute key and two other pointer attributes: next and pre . The object may also contain other satellite data. Given an element x in the list, x:next points to its successor in the linked list, and x:pre points to its predecessor. If x:pre D NIL, the element x has no predecessor and is therefore the first element, or head, of the list. If x:next D NIL, the element x has no successor and is therefore the last element, or tail, of the list. An attribute L:head points to the first element of the list. If L:head D NIL, the list is empty. A list may have one of several forms. It may be either singly linked or doubly linked, it may be sorted or not, and it may be circular or not. If a list is singly linked, we omit the pre pointer in each element. If a list is sorted, the linear order of the list corresponds to the linear order of keys stored in elements of the list; the minimum element is then the head of the list, and the maximum element is the tail. If the list is unsorted, the elements can appear in any order. In a circular list, the pre pointer of the head of the list points to the tail, and the next pointer of the tail of the list points to the head. We can think of a circular list as a ring of 10.2 Linked lists prev key next (a) L:head 9 16 4 1 (b) L:head 25 9 16 4 (c) L:head 25 9 16 1 237 1 Figure 10.3 (a) A doubly linked list L representing the dynamic set f1; 4; 9; 16g. Each element in the list is an object with attributes for the key and pointers (shown by arrows) to the next and previous objects. The next attribute of the tail and the pre attribute of the head are NIL, indicated by a diagonal slash. The attribute L: head points to the head. (b) Following the execution of LIST-INSERT.L; x/, where x: key D 25, the linked list has a new object with key 25 as the new head. This new object points to the old head with key 9. (c) The result of the subsequent call LIST-DELETE.L; x/, where x points to the object with key 4. elements. In the remainder of this section, we assume that the lists with which we are working are unsorted and doubly linked. Searching a linked list The procedure LIST-SEARCH.L; k/ finds the first element with key k in list L by a simple linear search, returning a pointer to this element. If no object with key k appears in the list, then the procedure returns NIL. For the linked list in Figure 10.3(a), the call LIST-SEARCH.L; 4/ returns a pointer to the third element, and the call LIST-SEARCH.L; 7/ returns NIL. LIST-SEARCH.L; k/ 1 x D L:head 2 while x ¤ NIL and x:key ¤ k 3 x D x:next 4 return x To search a list of n objects, the LIST-SEARCH procedure takes ‚.n/ time in the worst case, since it may have to search the entire list. Inserting into a linked list Given an element x whose key attribute has already been set, the LIST-INSERT procedure “splices” x onto the front of the linked list, as shown in Figure 10.3(b). 238 Chapter 10 Elementary Data Structures LIST-INSERT.L; x/ 1 x:next D L:head 2 if L:head ¤ NIL 3 L:head:pre D x 4 L:head D x 5 x:pre D NIL (Recall that our attribute notation can cascade, so that L:head:pre denotes the pre attribute of the object that L:head points to.) The running time for LISTINSERT on a list of n elements is O.1/. Deleting from a linked list The procedure LIST-DELETE removes an element x from a linked list L. It must be given a pointer to x, and it then “splices” x out of the list by updating pointers. If we wish to delete an element with a given key, we must first call LIST-SEARCH to retrieve a pointer to the element. LIST-DELETE .L; x/ 1 if x:pre ¤ NIL 2 x:pre :next D x:next 3 else L:head D x:next 4 if x:next ¤ NIL 5 x:next:pre D x:pre Figure 10.3(c) shows how an element is deleted from a linked list. LIST-DELETE runs in O.1/ time, but if we wish to delete an element with a given key, ‚.n/ time is required in the worst case because we must first call LIST-SEARCH to find the element. Sentinels The code for LIST-DELETE would be simpler if we could ignore the boundary conditions at the head and tail of the list: LIST-DELETE0 .L; x/ 1 x:pre :next D x:next 2 x:next:pre D x:pre A sentinel is a dummy object that allows us to simplify boundary conditions. For example, suppose that we provide with list L an object L:nil that represents NIL 10.2 Linked lists 239 (a) L:nil (b) L:nil (c) L:nil (d) L:nil 9 16 4 1 25 9 16 4 1 25 9 16 4 Figure 10.4 A circular, doubly linked list with a sentinel. The sentinel L: nil appears between the head and tail. The attribute L: head is no longer needed, since we can access the head of the list by L: nil: next. (a) An empty list. (b) The linked list from Figure 10.3(a), with key 9 at the head and key 1 at the tail. (c) The list after executing LIST-INSERT0.L; x/, where x: key D 25. The new object becomes the head of the list. (d) The list after deleting the object with key 1. The new tail is the object with key 4. but has all the attributes of the other objects in the list. Wherever we have a reference to NIL in list code, we replace it by a reference to the sentinel L:nil. As shown in Figure 10.4, this change turns a regular doubly linked list into a circular, doubly linked list with a sentinel, in which the sentinel L:nil lies between the head and tail. The attribute L:nil:next points to the head of the list, and L:nil:pre points to the tail. Similarly, both the next attribute of the tail and the pre attribute of the head point to L:nil. Since L:nil:next points to the head, we can eliminate the attribute L:head altogether, replacing references to it by references to L:nil:next. Figure 10.4(a) shows that an empty list consists of just the sentinel, and both L:nil:next and L:nil:pre point to L:nil. The code for LIST-SEARCH remains the same as before, but with the references to NIL and L:head changed as specified above: LIST-SEARCH0.L; k/ 1 x D L:nil:next 2 while x ¤ L:nil and x:key ¤ k 3 x D x:next 4 return x We use the two-line procedure LIST-DELETE0 from before to delete an element from the list. The following procedure inserts an element into the list: 240 Chapter 10 Elementary Data Structures LIST-INSERT0 .L; x/ 1 x:next D L:nil:next 2 L:nil:next:pre D x 3 L:nil:next D x 4 x:pre D L:nil Figure 10.4 shows the effects of LIST-INSERT0 and LIST-DELETE0 on a sample list. Sentinels rarely reduce the asymptotic time bounds of data structure operations, but they can reduce constant factors. The gain from using sentinels within loops is usually a matter of clarity of code rather than speed; the linked list code, for example, becomes simpler when we use sentinels, but we save only O.1/ time in the LIST-INSERT0 and LIST-DELETE0 procedures. In other situations, however, the use of sentinels helps to tighten the code in a loop, thus reducing the coefficient of, say, n or n2 in the running time. We should use sentinels judiciously. When there are many small lists, the extra storage used by their sentinels can represent significant wasted memory. In this book, we use sentinels only when they truly simplify the code. Exercises 10.2-1 Can you implement the dynamic-set operation INSERT on a singly linked list in O.1/ time? How about DELETE? 10.2-2 Implement a stack using a singly linked list L. The operations PUSH and POP should still take O.1/ time. 10.2-3 Implement a queue by a singly linked list L. The operations ENQUEUE and DEQUEUE should still take O.1/ time. 10.2-4 As written, each loop iteration in the LIST-SEARCH0 procedure requires two tests: one for x ¤ L:nil and one for x:key ¤ k. Show how to eliminate the test for x ¤ L:nil in each iteration. 10.2-5 Implement the dictionary operations INSERT, DELETE, and SEARCH using singly linked, circular lists. What are the running times of your procedures? 10.3 Implementing pointers and objects 241 10.2-6 The dynamic-set operation UNION takes two disjoint sets S1 and S2 as input, and it returns a set S D S1 [ S2 consisting of all the elements of S1 and S2. The sets S1 and S2 are usually destroyed by the operation. Show how to support UNION in O.1/ time using a suitable list data structure. 10.2-7 Give a ‚.n/-time nonrecursive procedure that reverses a singly linked list of n elements. The procedure should use no more than constant storage beyond that needed for the list itself. 10.2-8 ? Explain how to implement doubly linked lists using only one pointer value x:np per item instead of the usual two (next and pre ). Assume that all pointer values can be interpreted as k-bit integers, and define x:np to be x:np D x:next XOR x:pre , the k-bit “exclusive-or” of x:next and x:pre . (The value NIL is represented by 0.) Be sure to describe what information you need to access the head of the list. Show how to implement the SEARCH, INSERT, and DELETE operations on such a list. Also show how to reverse such a list in O.1/ time. 10.3 Implementing pointers and objects How do we implement pointers and objects in languages that do not provide them? In this section, we shall see two ways of implementing linked data structures without an explicit pointer data type. We shall synthesize objects and pointers from arrays and array indices. A multiple-array representation of objects We can represent a collection of objects that have the same attributes by using an array for each attribute. As an example, Figure 10.5 shows how we can implement the linked list of Figure 10.3(a) with three arrays. The array key holds the values of the keys currently in the dynamic set, and the pointers reside in the arrays next and pre . For a given array index x, the array entries keyŒx, nextŒx, and pre Œx represent an object in the linked list. Under this interpretation, a pointer x is simply a common index into the key, next, and pre arrays. In Figure 10.3(a), the object with key 4 follows the object with key 16 in the linked list. In Figure 10.5, key 4 appears in keyŒ2, and key 16 appears in keyŒ5, and so nextŒ5 D 2 and pre Œ2 D 5. Although the constant NIL appears in the next 242 Chapter 10 Elementary Data Structures L7 12345678 next 3 25 key 4 1 16 9 prev 52 7 Figure 10.5 The linked list of Figure 10.3(a) represented by the arrays key, next, and pre . Each vertical slice of the arrays represents a single object. Stored pointers correspond to the array indices shown at the top; the arrows show how to interpret them. Lightly shaded object positions contain list elements. The variable L keeps the index of the head. attribute of the tail and the pre attribute of the head, we usually use an integer (such as 0 or 1) that cannot possibly represent an actual index into the arrays. A variable L holds the index of the head of the list. A single-array representation of objects The words in a computer memory are typically addressed by integers from 0 to M 1, where M is a suitably large integer. In many programming languages, an object occupies a contiguous set of locations in the computer memory. A pointer is simply the address of the first memory location of the object, and we can address other memory locations within the object by adding an offset to the pointer. We can use the same strategy for implementing objects in programming environments that do not provide explicit pointer data types. For example, Figure 10.6 shows how to use a single array A to store the linked list from Figures 10.3(a) and 10.5. An object occupies a contiguous subarray AŒj : : k. Each attribute of the object corresponds to an offset in the range from 0 to k j , and a pointer to the object is the index j . In Figure 10.6, the offsets corresponding to key, next, and pre are 0, 1, and 2, respectively. To read the value of i:pre , given a pointer i, we add the value i of the pointer to the offset 2, thus reading AŒi C 2. The single-array representation is flexible in that it permits objects of different lengths to be stored in the same array. The problem of managing such a heterogeneous collection of objects is more difficult than the problem of managing a homogeneous collection, where all objects have the same attributes. Since most of the data structures we shall consider are composed of homogeneous elements, it will be sufficient for our purposes to use the multiple-array representation of objects. 10.3 Implementing pointers and objects 243 L 19 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 A 4 7 13 1 4 16 4 19 9 13 key prev next Figure 10.6 The linked list of Figures 10.3(a) and 10.5 represented in a single array A. Each list element is an object that occupies a contiguous subarray of length 3 within the array. The three attributes key, next, and pre correspond to the offsets 0, 1, and 2, respectively, within each object. A pointer to an object is the index of the first element of the object. Objects containing list elements are lightly shaded, and arrows show the list ordering. Allocating and freeing objects To insert a key into a dynamic set represented by a doubly linked list, we must allocate a pointer to a currently unused object in the linked-list representation. Thus, it is useful to manage the storage of objects not currently used in the linked-list representation so that one can be allocated. In some systems, a garbage collector is responsible for determining which objects are unused. Many applications, however, are simple enough that they can bear responsibility for returning an unused object to a storage manager. We shall now explore the problem of allocating and freeing (or deallocating) homogeneous objects using the example of a doubly linked list represented by multiple arrays. Suppose that the arrays in the multiple-array representation have length m and that at some moment the dynamic set contains n Ä m elements. Then n objects represent elements currently in the dynamic set, and the remaining m n objects are free; the free objects are available to represent elements inserted into the dynamic set in the future. We keep the free objects in a singly linked list, which we call the free list. The free list uses only the next array, which stores the next pointers within the list. The head of the free list is held in the global variable free. When the dynamic set represented by linked list L is nonempty, the free list may be intertwined with list L, as shown in Figure 10.7. Note that each object in the representation is either in list L or in the free list, but not in both. The free list acts like a stack: the next object allocated is the last one freed. We can use a list implementation of the stack operations PUSH and POP to implement the procedures for allocating and freeing objects, respectively. We assume that the global variable free used in the following procedures points to the first element of the free list. 244 Chapter 10 Elementary Data Structures free 4 L7 12345678 next 3 8 2 1 5 6 key 4 1 16 9 prev 5 2 7 (a) free 5 L4 12345678 next 3 7 8 1 2 6 key 4 1 25 9 prev 7 2 4 (c) free 8 L4 12345678 next 3 7 2 1 5 6 key 4 1 25 16 9 prev 5 2 7 4 (b) Figure 10.7 The effect of the ALLOCATE-OBJECT and FREE-OBJECT procedures. (a) The list of Figure 10.5 (lightly shaded) and a free list (heavily shaded). Arrows show the free-list structure. (b) The result of calling ALLOCATE-OBJECT./ (which returns index 4), setting keyŒ4 to 25, and calling LIST-INSERT.L; 4/. The new free-list head is object 8, which had been nextŒ4 on the free list. (c) After executing LIST-DELETE.L; 5/, we call FREE-OBJECT.5/. Object 5 becomes the new free-list head, with object 8 following it on the free list. ALLOCATE-OBJECT ./ 1 if free == NIL 2 error “out of space” 3 else x D free 4 free D x:next 5 return x F R EE-O B JEC T .x / 1 x:next D free 2 free D x The free list initially contains all n unallocated objects. Once the free list has been exhausted, running the ALLOCATE-OBJECT procedure signals an error. We can even service several linked lists with just a single free list. Figure 10.8 shows two linked lists and a free list intertwined through key, next, and pre arrays. The two procedures run in O.1/ time, which makes them quite practical. We can modify them to work for any homogeneous collection of objects by letting any one of the attributes in the object act like a next attribute in the free list. 10.3 Implementing pointers and objects 245 free 10 L2 9 L1 3 1 2 3 4 5 6 7 8 9 10 next 5 6 8 2 1 7 4 key k1 k2 k3 k5 k6 k7 k9 prev 7 6 139 Figure 10.8 Two linked lists, L1 (lightly shaded) and L2 (heavily shaded), and a free list (darkened) intertwined. Exercises 10.3-1 Draw a picture of the sequence h13; 4; 8; 19; 5; 11i stored as a doubly linked list using the multiple-array representation. Do the same for the single-array representation. 10.3-2 Write the procedures ALLOCATE-OBJECT and FREE-OBJECT for a homogeneous collection of objects implemented by the single-array representation. 10.3-3 Why don’t we need to set or reset the pre attributes of objects in the implementation of the ALLOCATE-OBJECT and FREE-OBJECT procedures? 10.3-4 It is often desirable to keep all elements of a doubly linked list compact in storage, using, for example, the first m index locations in the multiple-array representation. (This is the case in a paged, virtual-memory computing environment.) Explain how to implement the procedures ALLOCATE-OBJECT and FREE-OBJECT so that the representation is compact. Assume that there are no pointers to elements of the linked list outside the list itself. (Hint: Use the array implementation of a stack.) 10.3-5 Let L be a doubly linked list of length n stored in arrays key, pre , and next of length m. Suppose that these arrays are managed by ALLOCATE-OBJECT and FREE-OBJECT procedures that keep a doubly linked free list F . Suppose further that of the m items, exactly n are on list L and m n are on the free list. Write a procedure COMPACTIFY-LIST.L; F / that, given the list L and the free list F , moves the items in L so that they occupy array positions 1; 2; : : : ; n and adjusts the free list F so that it remains correct, occupying array positions n C1; n C2; : : : ; m. The running time of your procedure should be ‚.n/, and it should use only a constant amount of extra space. Argue that your procedure is correct. 246 Chapter 10 Elementary Data Structures 10.4 Representing rooted trees The methods for representing lists given in the previous section extend to any homogeneous data structure. In this section, we look specifically at the problem of representing rooted trees by linked data structures. We first look at binary trees, and then we present a method for rooted trees in which nodes can have an arbitrary number of children. We represent each node of a tree by an object. As with linked lists, we assume that each node contains a key attribute. The remaining attributes of interest are pointers to other nodes, and they vary according to the type of tree. Binary trees Figure 10.9 shows how we use the attributes p, left, and right to store pointers to the parent, left child, and right child of each node in a binary tree T . If x:p D NIL, then x is the root. If node x has no left child, then x:left D NIL, and similarly for the right child. The root of the entire tree T is pointed to by the attribute T:root. If T:root D NIL, then the tree is empty. Rooted trees with unbounded branching We can extend the scheme for representing a binary tree to any class of trees in which the number of children of each node is at most some constant k: we replace the left and right attributes by child1; child2; : : : ; childk. This scheme no longer works when the number of children of a node is unbounded, since we do not know how many attributes (arrays in the multiple-array representation) to allocate in advance. Moreover, even if the number of children k is bounded by a large constant but most nodes have a small number of children, we may waste a lot of memory. Fortunately, there is a clever scheme to represent trees with arbitrary numbers of children. It has the advantage of using only O.n/ space for any n-node rooted tree. The left-child, right-sibling representation appears in Figure 10.10. As before, each node contains a parent pointer p, and T:root points to the root of tree T . Instead of having a pointer to each of its children, however, each node x has only two pointers: 1. x:left-child points to the leftmost child of node x, and 2. x:right-sibling points to the sibling of x immediately to its right. If node x has no children, then x:left-child D NIL, and if node x is the rightmost child of its parent, then x:right-sibling D NIL. 10.4 Representing rooted trees 247 T: root Figure 10.9 The representation of a binary tree T . Each node x has the attributes x: p (top), x: left (lower left), and x: right (lower right). The key attributes are not shown. T: root Figure 10.10 The left-child, right-sibling representation of a tree T . Each node x has attributes x: p (top), x: left-child (lower left), and x: right-sibling (lower right). The key attributes are not shown. 248 Chapter 10 Elementary Data Structures Other tree representations We sometimes represent rooted trees in other ways. In Chapter 6, for example, we represented a heap, which is based on a complete binary tree, by a single array plus the index of the last node in the heap. The trees that appear in Chapter 21 are traversed only toward the root, and so only the parent pointers are present; there are no pointers to children. Many other schemes are possible. Which scheme is best depends on the application. Exercises 10.4-1 Draw the binary tree rooted at index 6 that is represented by the following attributes: index key left right 1 12 7 3 2 15 8 NIL 3 4 10 NIL 4 10 5 9 5 2 NIL NIL 6 18 1 4 7 7 NIL NIL 8 14 6 2 9 21 NIL NIL 10 5 NIL NIL 10.4-2 Write an O.n/-time recursive procedure that, given an n-node binary tree, prints out the key of each node in the tree. 10.4-3 Write an O.n/-time nonrecursive procedure that, given an n-node binary tree, prints out the key of each node in the tree. Use a stack as an auxiliary data structure. 10.4-4 Write an O.n/-time procedure that prints all the keys of an arbitrary rooted tree with n nodes, where the tree is stored using the left-child, right-sibling representation. 10.4-5 ? Write an O.n/-time nonrecursive procedure that, given an n-node binary tree, prints out the key of each node. Use no more than constant extra space outside Problems for Chapter 10 249 of the tree itself and do not modify the tree, even temporarily, during the procedure. 10.4-6 ? The left-child, right-sibling representation of an arbitrary rooted tree uses three pointers in each node: left-child, right-sibling, and parent. From any node, its parent can be reached and identified in constant time and all its children can be reached and identified in time linear in the number of children. Show how to use only two pointers and one boolean value in each node so that the parent of a node or all of its children can be reached and identified in time linear in the number of children. Problems 10-1 Comparisons among lists For each of the four types of lists in the following table, what is the asymptotic worst-case running time for each dynamic-set operation listed? SEARCH.L; k/ INSERT.L; x/ DELETE.L; x/ SUCCESSOR.L; x/ PREDECESSOR.L; x/ MINIMUM.L/ MAXIMUM.L/ unsorted, singly linked sorted, singly linked unsorted, doubly linked sorted, doubly linked 250 Chapter 10 Elementary Data Structures 10-2 Mergeable heaps using linked lists A mergeable heap supports the following operations: MAKE-HEAP (which creates an empty mergeable heap), INSERT, MINIMUM, EXTRACT-MIN, and UNION.1 Show how to implement mergeable heaps using linked lists in each of the following cases. Try to make each operation as efficient as possible. Analyze the running time of each operation in terms of the size of the dynamic set(s) being operated on. a. Lists are sorted. b. Lists are unsorted. c. Lists are unsorted, and dynamic sets to be merged are disjoint. 10-3 Searching a sorted compact list Exercise 10.3-4 asked how we might maintain an n-element list compactly in the first n positions of an array. We shall assume that all keys are distinct and that the compact list is also sorted, that is, keyŒi < keyŒnextŒi for all i D 1; 2; : : : ; n such that nextŒi ¤ NIL. We will also assume that we have a variable L that contains the index of the first element on the list. Under these assumptions, you will sphow that we can use the following randomized algorithm to search the list in O. n/ expected time. COMPACT-LIST-SEARCH.L; n; k/ 1 i DL 2 while i ¤ NIL and keyŒi  < k 3 j D RANDOM.1; n/ 4 if keyŒi < keyŒj  and keyŒj  Ä k 5 i Dj 6 if keyŒi == k 7 return i 8 i D nextŒi 9 if i == NIL or keyŒi  > k 10 return NIL 11 else return i If we ignore lines 3–7 of the procedure, we have an ordinary algorithm for searching a sorted linked list, in which index i points to each position of the list in 1Because we have defined a mergeable heap to support MINIMUM and EXTRACT-MIN, we can also refer to it as a mergeable min-heap. Alternatively, if it supported MAXIMUM and EXTRACT-MAX, it would be a mergeable max-heap. Problems for Chapter 10 251 turn. The search terminates once the index i “falls off” the end of the list or once keyŒi k. In the latter case, if keyŒi D k, clearly we have found a key with the value k. If, however, keyŒi > k, then we will never find a key with the value k, and so terminating the search was the right thing to do. Lines 3–7 attempt to skip ahead to a randomly chosen position j . Such a skip benefits us if keyŒj  is larger than keyŒi and no larger than k; in such a case, j marks a position in the list that i would have to reach during an ordinary list search. Because the list is compact, we know that any choice of j between 1 and n indexes some object in the list rather than a slot on the free list. Instead of analyzing the performance of COMPACT-LIST-SEARCH directly, we shall analyze a related algorithm, COMPACT-LIST-SEARCH0, which executes two separate loops. This algorithm takes an additional parameter t which determines an upper bound on the number of iterations of the first loop. COMPACT-LIST-SEARCH0 .L; n; k; t / 1 iDL 2 for q D 1 to t 3 j D RANDOM.1; n/ 4 if keyŒi < keyŒj  and keyŒj  Ä k 5 i Dj 6 if keyŒi == k 7 return i 8 while i ¤ NIL and keyŒi  < k 9 i D nextŒi 10 if i == NIL or keyŒi  > k 11 return NIL 12 else return i To compare the execution of the algorithms COMPACT-LIST-SEARCH.L; n; k/ and COMPACT-LIST-SEARCH0.L; n; k; t/, assume that the sequence of integers returned by the calls of RANDOM.1; n/ is the same for both algorithms. a. Suppose that COMPACT-LIST-SEARCH.L; n; k/ takes t iterations of the while loop of lines 2–8. Argue that COMPACT-LIST-SEARCH0.L; n; k; t/ returns the same answer and that the total number of iterations of both the for and while loops within COMPACT-LIST-SEARCH 0 is at least t . In the call COMPACT-LIST-SEARCH0.L; n; k; t/, let Xt be the random variable that describes the distance in the linked list (that is, through the chain of next pointers) from position i to the desired key k after t iterations of the for loop of lines 2–7 have occurred. 252 Chapter 10 Elementary Data Structures b. Argue that the expected running time of COMPACT-LIST-SEARCH0.L; n; k; t/ is O.t C E ŒXt /. c. Show that E ŒXt  Ä Pn r D1 .1 r=n/t . (Hint: Use equation (C.25).) d. Show that Pn 1 r D0 r t Ä nt C1=.t C 1/. e. Prove that E ŒXt  Ä n=.t C 1/. f. Show that COMPACT-LIST-SEARCH0.L; n; k; t/ runs in O.t C n=t/ expected time. p g. Conclude that COMPACT-LIST-SEARCH runs in O. n/ expected time. h. Why do we assume that all keys are distinct in COMPACT-LIST-SEARCH? Argue that random skips do not necessarily help asymptotically when the list contains repeated key values. Chapter notes Aho, Hopcroft, and Ullman [6] and Knuth [209] are excellent references for elementary data structures. Many other texts cover both basic data structures and their implementation in a particular programming language. Examples of these types of textbooks include Goodrich and Tamassia [147], Main [241], Shaffer [311], and Weiss [352, 353, 354]. Gonnet [145] provides experimental data on the performance of many data-structure operations. The origin of stacks and queues as data structures in computer science is unclear, since corresponding notions already existed in mathematics and paper-based business practices before the introduction of digital computers. Knuth [209] cites A. M. Turing for the development of stacks for subroutine linkage in 1947. Pointer-based data structures also seem to be a folk invention. According to Knuth, pointers were apparently used in early computers with drum memories. The A-1 language developed by G. M. Hopper in 1951 represented algebraic formulas as binary trees. Knuth credits the IPL-II language, developed in 1956 by A. Newell, J. C. Shaw, and H. A. Simon, for recognizing the importance and promoting the use of pointers. Their IPL-III language, developed in 1957, included explicit stack operations. 11 Hash Tables Many applications require a dynamic set that supports only the dictionary operations INSERT, SEARCH, and DELETE. For example, a compiler that translates a programming language maintains a symbol table, in which the keys of elements are arbitrary character strings corresponding to identifiers in the language. A hash table is an effective data structure for implementing dictionaries. Although searching for an element in a hash table can take as long as searching for an element in a linked list—‚.n/ time in the worst case—in practice, hashing performs extremely well. Under reasonable assumptions, the average time to search for an element in a hash table is O.1/. A hash table generalizes the simpler notion of an ordinary array. Directly addressing into an ordinary array makes effective use of our ability to examine an arbitrary position in an array in O.1/ time. Section 11.1 discusses direct addressing in more detail. We can take advantage of direct addressing when we can afford to allocate an array that has one position for every possible key. When the number of keys actually stored is small relative to the total number of possible keys, hash tables become an effective alternative to directly addressing an array, since a hash table typically uses an array of size proportional to the number of keys actually stored. Instead of using the key as an array index directly, the array index is computed from the key. Section 11.2 presents the main ideas, focusing on “chaining” as a way to handle “collisions,” in which more than one key maps to the same array index. Section 11.3 describes how we can compute array indices from keys using hash functions. We present and analyze several variations on the basic theme. Section 11.4 looks at “open addressing,” which is another way to deal with collisions. The bottom line is that hashing is an extremely effective and practical technique: the basic dictionary operations require only O.1/ time on the average. Section 11.5 explains how “perfect hashing” can support searches in O.1/ worstcase time, when the set of keys being stored is static (that is, when the set of keys never changes once stored). 254 Chapter 11 Hash Tables 11.1 Direct-address tables Direct addressing is a simple technique that works well when the universe U of keys is reasonably small. Suppose that an application needs a dynamic set in which each element has a key drawn from the universe U D f0; 1; : : : ; m 1g, where m is not too large. We shall assume that no two elements have the same key. To represent the dynamic set, we use an array, or direct-address table, denoted by T Œ0 : : m 1, in which each position, or slot, corresponds to a key in the universe U . Figure 11.1 illustrates the approach; slot k points to an element in the set with key k. If the set contains no element with key k, then T Œk D NIL. The dictionary operations are trivial to implement: DIRECT-ADDRESS-SEARCH .T; k/ 1 return T Œk DIRECT-ADDRESS-INSERT.T; x/ 1 T Œx:key D x DIRECT-ADDRESS-DELETE .T; x/ 1 T Œx:key D NIL Each of these operations takes only O.1/ time. U (universe of keys) 0 6 9 4 7 1 K2 (actual 3 keys) 5 8 T 0 1 2 3 4 5 6 7 8 9 key satellite data 2 3 5 8 Figure 11.1 How to implement a dynamic set by a direct-address table T . Each key in the universe U D f0; 1; : : : ; 9g corresponds to an index in the table. The set K D f2; 3; 5; 8g of actual keys determines the slots in the table that contain pointers to elements. The other slots, heavily shaded, contain NIL. 11.1 Direct-address tables 255 For some applications, the direct-address table itself can hold the elements in the dynamic set. That is, rather than storing an element’s key and satellite data in an object external to the direct-address table, with a pointer from a slot in the table to the object, we can store the object in the slot itself, thus saving space. We would use a special key within an object to indicate an empty slot. Moreover, it is often unnecessary to store the key of the object, since if we have the index of an object in the table, we have its key. If keys are not stored, however, we must have some way to tell whether the slot is empty. Exercises 11.1-1 Suppose that a dynamic set S is represented by a direct-address table T of length m. Describe a procedure that finds the maximum element of S. What is the worst-case performance of your procedure? 11.1-2 A bit vector is simply an array of bits (0s and 1s). A bit vector of length m takes much less space than an array of m pointers. Describe how to use a bit vector to represent a dynamic set of distinct elements with no satellite data. Dictionary operations should run in O.1/ time. 11.1-3 Suggest how to implement a direct-address table in which the keys of stored elements do not need to be distinct and the elements can have satellite data. All three dictionary operations (INSERT, DELETE, and SEARCH) should run in O.1/ time. (Don’t forget that DELETE takes as an argument a pointer to an object to be deleted, not a key.) 11.1-4 ? We wish to implement a dictionary by using direct addressing on a huge array. At the start, the array entries may contain garbage, and initializing the entire array is impractical because of its size. Describe a scheme for implementing a directaddress dictionary on a huge array. Each stored object should use O.1/ space; the operations SEARCH, INSERT, and DELETE should take O.1/ time each; and initializing the data structure should take O.1/ time. (Hint: Use an additional array, treated somewhat like a stack whose size is the number of keys actually stored in the dictionary, to help determine whether a given entry in the huge array is valid or not.) 256 Chapter 11 Hash Tables 11.2 Hash tables The downside of direct addressing is obvious: if the universe U is large, storing a table T of size jU j may be impractical, or even impossible, given the memory available on a typical computer. Furthermore, the set K of keys actually stored may be so small relative to U that most of the space allocated for T would be wasted. When the set K of keys stored in a dictionary is much smaller than the universe U of all possible keys, a hash table requires much less storage than a directaddress table. Specifically, we can reduce the storage requirement to ‚.jKj/ while we maintain the benefit that searching for an element in the hash table still requires only O.1/ time. The catch is that this bound is for the average-case time, whereas for direct addressing it holds for the worst-case time. With direct addressing, an element with key k is stored in slot k. With hashing, this element is stored in slot h.k/; that is, we use a hash function h to compute the slot from the key k. Here, h maps the universe U of keys into the slots of a hash table T Œ0 : : m 1: h W U ! f0; 1; : : : ; m 1g ; where the size m of the hash table is typically much less than jU j. We say that an element with key k hashes to slot h.k/; we also say that h.k/ is the hash value of key k. Figure 11.2 illustrates the basic idea. The hash function reduces the range of array indices and hence the size of the array. Instead of a size of jU j, the array can have size m. U (universe of keys) k1 K (actual k4 k5 keys) k2 k3 T 0 h(k1) h(k4) h(k2) = h(k5) h(k3) m–1 Figure 11.2 Using a hash function h to map keys to hash-table slots. Because keys k2 and k5 map to the same slot, they collide. 11.2 Hash tables U (universe of keys) k1 K (actual keys) k4 k5 k7 k2 k6 k8 k3 257 T k1 k4 k5 k2 k7 k3 k8 k6 Figure 11.3 Collision resolution by chaining. Each hash-table slot T Œj  contains a linked list of all the keys whose hash value is j . For example, h.k1/ D h.k4/ and h.k5/ D h.k7/ D h.k2/. The linked list can be either singly or doubly linked; we show it as doubly linked because deletion is faster that way. There is one hitch: two keys may hash to the same slot. We call this situation a collision. Fortunately, we have effective techniques for resolving the conflict created by collisions. Of course, the ideal solution would be to avoid collisions altogether. We might try to achieve this goal by choosing a suitable hash function h. One idea is to make h appear to be “random,” thus avoiding collisions or at least minimizing their number. The very term “to hash,” evoking images of random mixing and chopping, captures the spirit of this approach. (Of course, a hash function h must be deterministic in that a given input k should always produce the same output h.k/.) Because jU j > m, however, there must be at least two keys that have the same hash value; avoiding collisions altogether is therefore impossible. Thus, while a welldesigned, “random”-looking hash function can minimize the number of collisions, we still need a method for resolving the collisions that do occur. The remainder of this section presents the simplest collision resolution technique, called chaining. Section 11.4 introduces an alternative method for resolving collisions, called open addressing. Collision resolution by chaining In chaining, we place all the elements that hash to the same slot into the same linked list, as Figure 11.3 shows. Slot j contains a pointer to the head of the list of all stored elements that hash to j ; if there are no such elements, slot j contains NIL. 258 Chapter 11 Hash Tables The dictionary operations on a hash table T are easy to implement when collisions are resolved by chaining: CHAINED-HASH-INSERT.T; x/ 1 insert x at the head of list T Œh.x:key/ CHAINED-HASH-SEARCH .T; k/ 1 search for an element with key k in list T Œh.k/ CHAINED-HASH-DELETE .T; x/ 1 delete x from the list T Œh.x:key/ The worst-case running time for insertion is O.1/. The insertion procedure is fast in part because it assumes that the element x being inserted is not already present in the table; if necessary, we can check this assumption (at additional cost) by searching for an element whose key is x:key before we insert. For searching, the worstcase running time is proportional to the length of the list; we shall analyze this operation more closely below. We can delete an element in O.1/ time if the lists are doubly linked, as Figure 11.3 depicts. (Note that CHAINED-HASH-DELETE takes as input an element x and not its key k, so that we don’t have to search for x first. If the hash table supports deletion, then its linked lists should be doubly linked so that we can delete an item quickly. If the lists were only singly linked, then to delete element x, we would first have to find x in the list T Œh.x:key/ so that we could update the next attribute of x’s predecessor. With singly linked lists, both deletion and searching would have the same asymptotic running times.) Analysis of hashing with chaining How well does hashing with chaining perform? In particular, how long does it take to search for an element with a given key? Given a hash table T with m slots that stores n elements, we define the load factor ˛ for T as n=m, that is, the average number of elements stored in a chain. Our analysis will be in terms of ˛, which can be less than, equal to, or greater than 1. The worst-case behavior of hashing with chaining is terrible: all n keys hash to the same slot, creating a list of length n. The worst-case time for searching is thus ‚.n/ plus the time to compute the hash function—no better than if we used one linked list for all the elements. Clearly, we do not use hash tables for their worst-case performance. (Perfect hashing, described in Section 11.5, does provide good worst-case performance when the set of keys is static, however.) The average-case performance of hashing depends on how well the hash function h distributes the set of keys to be stored among the m slots, on the average. 11.2 Hash tables 259 Section 11.3 discusses these issues, but for now we shall assume that any given element is equally likely to hash into any of the m slots, independently of where any other element has hashed to. We call this the assumption of simple uniform hashing. For j D 0; 1; : : : ; m 1, let us denote the length of the list T Œj  by nj , so that n D n0 C n1 C C nm 1 ; (11.1) and the expected value of nj is E Œnj  D ˛ D n=m. We assume that O.1/ time suffices to compute the hash value h.k/, so that the time required to search for an element with key k depends linearly on the length nh.k/ of the list T Œh.k/. Setting aside the O.1/ time required to compute the hash function and to access slot h.k/, let us consider the expected number of elements examined by the search algorithm, that is, the number of elements in the list T Œh.k/ that the algorithm checks to see whether any have a key equal to k. We shall consider two cases. In the first, the search is unsuccessful: no element in the table has key k. In the second, the search successfully finds an element with key k. Theorem 11.1 In a hash table in which collisions are resolved by chaining, an unsuccessful search takes average-case time ‚.1C˛/, under the assumption of simple uniform hashing. Proof Under the assumption of simple uniform hashing, any key k not already stored in the table is equally likely to hash to any of the m slots. The expected time to search unsuccessfully for a key k is the expected time to search to the end of list T Œh.k/, which has expected length E Œnh.k/ D ˛. Thus, the expected number of elements examined in an unsuccessful search is ˛, and the total time required (including the time for computing h.k/) is ‚.1 C ˛/. The situation for a successful search is slightly different, since each list is not equally likely to be searched. Instead, the probability that a list is searched is proportional to the number of elements it contains. Nonetheless, the expected search time still turns out to be ‚.1 C ˛/. Theorem 11.2 In a hash table in which collisions are resolved by chaining, a successful search takes average-case time ‚.1C˛/, under the assumption of simple uniform hashing. Proof We assume that the element being searched for is equally likely to be any of the n elements stored in the table. The number of elements examined during a successful search for an element x is one more than the number of elements that 260 Chapter 11 Hash Tables appear before x in x’s list. Because new elements are placed at the front of the list, elements before x in the list were all inserted after x was inserted. To find the expected number of elements examined, we take the average, over the n ele- ments x in the table, of 1 plus the expected number of elements added to x’s list after x was added to the list. Let xi denote the ith element inserted into the ta- ble, for i D 1; 2; : : : ; n, and let ki D xi :key. For keys ki and kj , we define the indicator random variable Xij D I fh.ki / D h.kj /g. Under the assumption of sim- ple uniform hashing, we have Pr fh.ki / D h.kj /g D 1=m, and so by Lemma 5.1, E ŒXij  D 1=m. Thus, the expected number of elements examined in a successful search is " !# 1 Xn Xn E n 1C Xij i D1 j Di C1 ! 1 Xn Xn D n 1C E ŒXij  i D1 j DiC1 ! 1 Xn Xn 1 D 1C n m i D1 j Di C1 (by linearity of expectation) 1 Xn D 1C .n i/ nm i D1 ! 1 Xn Xn D 1C n i nm  iD1 iD1 à 1 D 1C n2 n.n C 1/ (by equation (A.1)) nm 2 n1 D 1C 2m D 1C ˛ ˛ : 2 2n Thus, the total time required for a successful search (including the time for computing the hash function) is ‚.2 C ˛=2 ˛=2n/ D ‚.1 C ˛/. What does this analysis mean? If the number of hash-table slots is at least proportional to the number of elements in the table, we have n D O.m/ and, consequently, ˛ D n=m D O.m/=m D O.1/. Thus, searching takes constant time on average. Since insertion takes O.1/ worst-case time and deletion takes O.1/ worst-case time when the lists are doubly linked, we can support all dictionary operations in O.1/ time on average. 11.2 Hash tables 261 Exercises 11.2-1 Suppose we use a hash function h to hash n distinct keys into an array T of length m. Assuming simple uniform hashing, what is the expected number of collisions? More precisely, what is the expected cardinality of ffk; lg W k ¤ l and h.k/ D h.l/g? 11.2-2 Demonstrate what happens when we insert the keys 5; 28; 19; 15; 20; 33; 12; 17; 10 into a hash table with collisions resolved by chaining. Let the table have 9 slots, and let the hash function be h.k/ D k mod 9. 11.2-3 Professor Marley hypothesizes that he can obtain substantial performance gains by modifying the chaining scheme to keep each list in sorted order. How does the professor’s modification affect the running time for successful searches, unsuccessful searches, insertions, and deletions? 11.2-4 Suggest how to allocate and deallocate storage for elements within the hash table itself by linking all unused slots into a free list. Assume that one slot can store a flag and either one element plus a pointer or two pointers. All dictionary and free-list operations should run in O.1/ expected time. Does the free list need to be doubly linked, or does a singly linked free list suffice? 11.2-5 Suppose that we are storing a set of n keys into a hash table of size m. Show that if the keys are drawn from a universe U with jU j > nm, then U has a subset of size n consisting of keys that all hash to the same slot, so that the worst-case searching time for hashing with chaining is ‚.n/. 11.2-6 Suppose we have stored n keys in a hash table of size m, with collisions resolved by chaining, and that we know the length of each chain, including the length L of the longest chain. Describe a procedure that selects a key uniformly at random from among the keys in the hash table and returns it in expected time O.L .1 C 1=˛//. 262 Chapter 11 Hash Tables 11.3 Hash functions In this section, we discuss some issues regarding the design of good hash functions and then present three schemes for their creation. Two of the schemes, hashing by division and hashing by multiplication, are heuristic in nature, whereas the third scheme, universal hashing, uses randomization to provide provably good performance. What makes a good hash function? A good hash function satisfies (approximately) the assumption of simple uniform hashing: each key is equally likely to hash to any of the m slots, independently of where any other key has hashed to. Unfortunately, we typically have no way to check this condition, since we rarely know the probability distribution from which the keys are drawn. Moreover, the keys might not be drawn independently. Occasionally we do know the distribution. For example, if we know that the keys are random real numbers k independently and uniformly distributed in the range 0 Ä k < 1, then the hash function h.k/ D bkmc satisfies the condition of simple uniform hashing. In practice, we can often employ heuristic techniques to create a hash function that performs well. Qualitative information about the distribution of keys may be useful in this design process. For example, consider a compiler’s symbol table, in which the keys are character strings representing identifiers in a program. Closely related symbols, such as pt and pts, often occur in the same program. A good hash function would minimize the chance that such variants hash to the same slot. A good approach derives the hash value in a way that we expect to be independent of any patterns that might exist in the data. For example, the “division method” (discussed in Section 11.3.1) computes the hash value as the remainder when the key is divided by a specified prime number. This method frequently gives good results, assuming that we choose a prime number that is unrelated to any patterns in the distribution of keys. Finally, we note that some applications of hash functions might require stronger properties than are provided by simple uniform hashing. For example, we might want keys that are “close” in some sense to yield hash values that are far apart. (This property is especially desirable when we are using linear probing, defined in Section 11.4.) Universal hashing, described in Section 11.3.3, often provides the desired properties. 11.3 Hash functions 263 Interpreting keys as natural numbers Most hash functions assume that the universe of keys is the set N D f0; 1; 2; : : :g of natural numbers. Thus, if the keys are not natural numbers, we find a way to interpret them as natural numbers. For example, we can interpret a character string as an integer expressed in suitable radix notation. Thus, we might interpret the identifier pt as the pair of decimal integers .112; 116/, since p D 112 and t D 116 in the ASCII character set; then, expressed as a radix-128 integer, pt becomes .112 128/ C 116 D 14452. In the context of a given application, we can usually devise some such method for interpreting each key as a (possibly large) natural number. In what follows, we assume that the keys are natural numbers. 11.3.1 The division method In the division method for creating hash functions, we map a key k into one of m slots by taking the remainder of k divided by m. That is, the hash function is h.k/ D k mod m : For example, if the hash table has size m D 12 and the key is k D 100, then h.k/ D 4. Since it requires only a single division operation, hashing by division is quite fast. When using the division method, we usually avoid certain values of m. For example, m should not be a power of 2, since if m D 2p, then h.k/ is just the p lowest-order bits of k. Unless we know that all low-order p-bit patterns are equally likely, we are better off designing the hash function to depend on all the bits of the key. As Exercise 11.3-3 asks you to show, choosing m D 2p 1 when k is a character string interpreted in radix 2p may be a poor choice, because permuting the characters of k does not change its hash value. A prime not too close to an exact power of 2 is often a good choice for m. For example, suppose we wish to allocate a hash table, with collisions resolved by chaining, to hold roughly n D 2000 character strings, where a character has 8 bits. We don’t mind examining an average of 3 elements in an unsuccessful search, and so we allocate a hash table of size m D 701. We could choose m D 701 because it is a prime near 2000=3 but not near any power of 2. Treating each key k as an integer, our hash function would be h.k/ D k mod 701 : 11.3.2 The multiplication method The multiplication method for creating hash functions operates in two steps. First, we multiply the key k by a constant A in the range 0 < A < 1 and extract the 264 Chapter 11 Hash Tables w bits k × s D A 2w r1 r0 extract p bits h.k/ Figure 11.4 The multiplication method of hashing. The w-bit representation of the key k is multiplied by the w-bit value s D A 2w . The p highest-order bits of the lower w-bit half of the product form the desired hash value h.k/. fractional part of kA. Then, we multiply this value by m and take the floor of the result. In short, the hash function is h.k/ D bm .kA mod 1/c ; where “kA mod 1” means the fractional part of kA, that is, kA bkAc. An advantage of the multiplication method is that the value of m is not critical. We typically choose it to be a power of 2 (m D 2p for some integer p), since we can then easily implement the function on most computers as follows. Suppose that the word size of the machine is w bits and that k fits into a single word. We restrict A to be a fraction of the form s=2w , where s is an integer in the range 0 < s < 2w . Referring to Figure 11.4, we first multiply k by the w-bit integer s D A 2w . The result is a 2w-bit value r12w C r0, where r1 is the high-order word of the product and r0 is the low-order word of the product. The desired p-bit hash value consists of the p most significant bits of r0. Although this method works with any value of the constant A, it works better with some values than with others. The optimal choice depends on the character- istics of the data being hashed. Knuth [211] suggests that p A . 5 1/=2 D 0:6180339887 : : : (11.2) is likely to work reasonably well. As an example, suppose we have k D 123456, p D 14, m D 214 D 16384, and w D 32. Adapting Knuth’ps suggestion, we choose A to be the fraction of the form s=232 that is closest to . 5 1/=2, so that A D 2654435769=232 . Then k s D 327706022297664 D .76300 232/ C 17612864, and so r1 D 76300 and r0 D 17612864. The 14 most significant bits of r0 yield the value h.k/ D 67. 11.3 Hash functions 265 ? 11.3.3 Universal hashing If a malicious adversary chooses the keys to be hashed by some fixed hash function, then the adversary can choose n keys that all hash to the same slot, yielding an average retrieval time of ‚.n/. Any fixed hash function is vulnerable to such terrible worst-case behavior; the only effective way to improve the situation is to choose the hash function randomly in a way that is independent of the keys that are actually going to be stored. This approach, called universal hashing, can yield provably good performance on average, no matter which keys the adversary chooses. In universal hashing, at the beginning of execution we select the hash function at random from a carefully designed class of functions. As in the case of quicksort, randomization guarantees that no single input will always evoke worst-case behavior. Because we randomly select the hash function, the algorithm can behave differently on each execution, even for the same input, guaranteeing good average-case performance for any input. Returning to the example of a compiler’s symbol table, we find that the programmer’s choice of identifiers cannot now cause consistently poor hashing performance. Poor performance occurs only when the compiler chooses a random hash function that causes the set of identifiers to hash poorly, but the probability of this situation occurring is small and is the same for any set of identifiers of the same size. Let H be a finite collection of hash functions that map a given universe U of keys into the range f0; 1; : : : ; m 1g. Such a collection is said to be universal if for each pair of distinct keys k; l 2 U , the number of hash functions h 2 H for which h.k/ D h.l/ is at most jH j =m. In other words, with a hash function randomly chosen from H , the chance of a collision between distinct keys k and l is no more than the chance 1=m of a collision if h.k/ and h.l/ were randomly and independently chosen from the set f0; 1; : : : ; m 1g. The following theorem shows that a universal class of hash functions gives good average-case behavior. Recall that ni denotes the length of list T Œi. Theorem 11.3 Suppose that a hash function h is chosen randomly from a universal collection of hash functions and has been used to hash n keys into a table T of size m, using chaining to resolve collisions. If key k is not in the table, then the expected length E Œnh.k/ of the list that key k hashes to is at most the load factor ˛ D n=m. If key k is in the table, then the expected length E Œnh.k/ of the list containing key k is at most 1 C ˛. Proof We note that the expectations here are over the choice of the hash function and do not depend on any assumptions about the distribution of the keys. For each pair k and l of distinct keys, define the indicator random variable 266 Chapter 11 Hash Tables Xkl D I fh.k/ D h.l/g. Since by the definition of a universal collection of hash functions, a single pair of keys collides with probability at most 1=m, we have Pr fh.k/ D h.l/g Ä 1=m. By Lemma 5.1, therefore, we have E ŒXkl  Ä 1=m. Next we define, for each key k, the random variable Yk that equals the number of keys other than k that hash to the same slot as k, so that X Yk D Xkl : l 2T l ¤k Thus we have 2X 3 E ŒYk D E4 Xkl 5 l 2T l ¤k X D E ŒXkl  l 2T l ¤k X1 Ä : m l 2T l ¤k (by linearity of expectation) The remainder of the proof depends on whether key k is in table T . If k 62 T , then nh.k/ D Yk and jfl W l 2 T and l ¤ kgj D n. Thus E Œnh.k/ D E ŒYk Ä n=m D ˛. If k 2 T , then because key k appears in list T Œh.k/ and the count Yk does not include key k, we have nh.k/ D Yk C 1 and jfl W l 2 T and l ¤ kgj D n 1. Thus E Œnh.k/ D E ŒYk C 1 Ä .n 1/=m C 1 D 1 C ˛ 1=m < 1 C ˛. The following corollary says universal hashing provides the desired payoff: it has now become impossible for an adversary to pick a sequence of operations that forces the worst-case running time. By cleverly randomizing the choice of hash function at run time, we guarantee that we can process every sequence of operations with a good average-case running time. Corollary 11.4 Using universal hashing and collision resolution by chaining in an initially empty table with m slots, it takes expected time ‚.n/ to handle any sequence of n INSERT, SEARCH, and DELETE operations containing O.m/ INSERT operations. Proof Since the number of insertions is O.m/, we have n D O.m/ and so ˛ D O.1/. The INSERT and DELETE operations take constant time and, by Theorem 11.3, the expected time for each SEARCH operation is O.1/. By linearity of 11.3 Hash functions 267 expectation, therefore, the expected time for the entire sequence of n operations is O.n/. Since each operation takes .1/ time, the ‚.n/ bound follows. Designing a universal class of hash functions It is quite easy to design a universal class of hash functions, as a little number theory will help us prove. You may wish to consult Chapter 31 first if you are unfamiliar with number theory. We begin by choosing a prime number p large enough so that every possible key k is in the range 0 to p 1, inclusive. Let Zp denote the set f0; 1; : : : ; p 1g, and let Zp denote the set f1; 2; : : : ; p 1g. Since p is prime, we can solve equations modulo p with the methods given in Chapter 31. Because we assume that the size of the universe of keys is greater than the number of slots in the hash table, we have p > m. We now define the hash function hab for any a 2 Zp and any b 2 Zp using a linear transformation followed by reductions modulo p and then modulo m: hab.k/ D ..ak C b/ mod p/ mod m : (11.3) For example, with p D 17 and m D 6, we have h3;4.8/ D 5. The family of all such hash functions is ˚ « Hpm D hab W a 2 Zp and b 2 Zp : (11.4) Each hash function hab maps Zp to Zm. This class of hash functions has the nice property that the size m of the output range is arbitrary—not necessarily prime—a feature which we shall use in Section 11.5. Since we have p 1 choices for a and p choices for b, the collection Hpm contains p.p 1/ hash functions. Theorem 11.5 The class Hpm of hash functions defined by equations (11.3) and (11.4) is universal. Proof Consider two distinct keys k and l from Zp, so that k ¤ l. For a given hash function hab we let r D .ak C b/ mod p ; s D .al C b/ mod p : We first note that r ¤ s. Why? Observe that r s Á a.k l/ .mod p/ : It follows that r ¤ s because p is prime and both a and .k l/ are nonzero modulo p, and so their product must also be nonzero modulo p by Theorem 31.6. Therefore, when computing any hab 2 Hpm, distinct inputs k and l map to distinct 268 Chapter 11 Hash Tables values r and s modulo p; there are no collisions yet at the “mod p level.” Moreover, each of the possible p.p 1/ choices for the pair .a; b/ with a ¤ 0 yields a different resulting pair .r; s/ with r ¤ s, since we can solve for a and b given r and s: a D .r s/..k l/ 1 mod p/ mod p ; b D .r ak/ mod p ; where ..k l/ 1 mod p/ denotes the unique multiplicative inverse, modulo p, of k l. Since there are only p.p 1/ possible pairs .r; s/ with r ¤ s, there is a one-to-one correspondence between pairs .a; b/ with a ¤ 0 and pairs .r; s/ with r ¤ s. Thus, for any given pair of inputs k and l, if we pick .a; b/ uniformly at random from Zp Zp, the resulting pair .r; s/ is equally likely to be any pair of distinct values modulo p. Therefore, the probability that distinct keys k and l collide is equal to the probability that r Á s .mod m/ when r and s are randomly chosen as distinct values modulo p. For a given value of r, of the p 1 possible remaining values for s, the number of values s such that s ¤ r and s Á r .mod m/ is at most dp=me 1 Ä ..p C m 1/=m/ 1 (by inequality (3.6)) D .p 1/=m : The probability that s collides with r when reduced modulo m is at most ..p 1/=m/=.p 1/ D 1=m. Therefore, for any pair of distinct values k; l 2 Zp, Pr fhab.k/ D hab.l/g Ä 1=m ; so that Hpm is indeed universal. Exercises 11.3-1 Suppose we wish to search a linked list of length n, where each element contains a key k along with a hash value h.k/. Each key is a long character string. How might we take advantage of the hash values when searching the list for an element with a given key? 11.3-2 Suppose that we hash a string of r characters into m slots by treating it as a radix-128 number and then using the division method. We can easily represent the number m as a 32-bit computer word, but the string of r characters, treated as a radix-128 number, takes many words. How can we apply the division method to compute the hash value of the character string without using more than a constant number of words of storage outside the string itself? 11.4 Open addressing 269 11.3-3 Consider a version of the division method in which h.k/ D k mod m, where m D 2p 1 and k is a character string interpreted in radix 2p. Show that if we can derive string x from string y by permuting its characters, then x and y hash to the same value. Give an example of an application in which this property would be undesirable in a hash function. 11.3-4 Consider a hash table of sizepm D 1000 and a corresponding hash function h.k/ D bm .kA mod 1/c for A D . 5 1/=2. Compute the locations to which the keys 61, 62, 63, 64, and 65 are mapped. 11.3-5 ? Define a family H of hash functions from a finite set U to a finite set B to be -universal if for all pairs of distinct elements k and l in U , Pr fh.k/ D h.l/g Ä ; where the probability is over the choice of the hash function h drawn at random from the family H . Show that an -universal family of hash functions must have 11 : jBj jU j 11.3-6 ? Let U be the set of n-tuples of values drawn from Zp, and let B D Zp, where p is prime. Define the hash function hb W U ! B for b 2 Zp on an input n-tuple ha0; a1; : : : ; an 1i from U as ! X n 1 hb.ha0; a1; : : : ; an 1i/ D aj bj mod p ; j D0 and let H D fhb W b 2 Zpg. Argue that H is ..n 1/=p/-universal according to the definition of -universal in Exercise 11.3-5. (Hint: See Exercise 31.4-4.) 11.4 Open addressing In open addressing, all elements occupy the hash table itself. That is, each table entry contains either an element of the dynamic set or NIL. When searching for an element, we systematically examine table slots until either we find the desired element or we have ascertained that the element is not in the table. No lists and 270 Chapter 11 Hash Tables no elements are stored outside the table, unlike in chaining. Thus, in open addressing, the hash table can “fill up” so that no further insertions can be made; one consequence is that the load factor ˛ can never exceed 1. Of course, we could store the linked lists for chaining inside the hash table, in the otherwise unused hash-table slots (see Exercise 11.2-4), but the advantage of open addressing is that it avoids pointers altogether. Instead of following pointers, we compute the sequence of slots to be examined. The extra memory freed by not storing pointers provides the hash table with a larger number of slots for the same amount of memory, potentially yielding fewer collisions and faster retrieval. To perform insertion using open addressing, we successively examine, or probe, the hash table until we find an empty slot in which to put the key. Instead of being fixed in the order 0; 1; : : : ; m 1 (which requires ‚.n/ search time), the sequence of positions probed depends upon the key being inserted. To determine which slots to probe, we extend the hash function to include the probe number (starting from 0) as a second input. Thus, the hash function becomes h W U f0; 1; : : : ; m 1g ! f0; 1; : : : ; m 1g : With open addressing, we require that for every key k, the probe sequence hh.k; 0/; h.k; 1/; : : : ; h.k; m 1/i be a permutation of h0; 1; : : : ; m 1i, so that every hash-table position is eventually considered as a slot for a new key as the table fills up. In the following pseudocode, we assume that the elements in the hash table T are keys with no satellite information; the key k is identical to the element containing key k. Each slot contains either a key or NIL (if the slot is empty). The HASH-INSERT procedure takes as input a hash table T and a key k. It either returns the slot number where it stores key k or flags an error because the hash table is already full. HASH-INSERT.T; k/ 1 iD0 2 repeat 3 j D h.k; i/ 4 if T Œj  == NIL 5 T Œj  D k 6 return j 7 else i D i C 1 8 until i == m 9 error “hash table overflow” The algorithm for searching for key k probes the same sequence of slots that the insertion algorithm examined when key k was inserted. Therefore, the search can 11.4 Open addressing 271 terminate (unsuccessfully) when it finds an empty slot, since k would have been inserted there and not later in its probe sequence. (This argument assumes that keys are not deleted from the hash table.) The procedure HASH-SEARCH takes as input a hash table T and a key k, returning j if it finds that slot j contains key k, or NIL if key k is not present in table T . HASH-SEARCH.T; k/ 1 i D0 2 repeat 3 j D h.k; i/ 4 if T Œj  == k 5 return j 6 i D i C1 7 until T Œj  = = NIL or i == m 8 return NIL Deletion from an open-address hash table is difficult. When we delete a key from slot i, we cannot simply mark that slot as empty by storing NIL in it. If we did, we might be unable to retrieve any key k during whose insertion we had probed slot i and found it occupied. We can solve this problem by marking the slot, storing in it the special value DELETED instead of NIL. We would then modify the procedure HASH-INSERT to treat such a slot as if it were empty so that we can insert a new key there. We do not need to modify HASH-SEARCH, since it will pass over DELETED values while searching. When we use the special value DELETED, however, search times no longer depend on the load factor ˛, and for this reason chaining is more commonly selected as a collision resolution technique when keys must be deleted. In our analysis, we assume uniform hashing: the probe sequence of each key is equally likely to be any of the mŠ permutations of h0; 1; : : : ; m 1i. Uniform hashing generalizes the notion of simple uniform hashing defined earlier to a hash function that produces not just a single number, but a whole probe sequence. True uniform hashing is difficult to implement, however, and in practice suitable approximations (such as double hashing, defined below) are used. We will examine three commonly used techniques to compute the probe sequences required for open addressing: linear probing, quadratic probing, and double hashing. These techniques all guarantee that hh.k; 0/; h.k; 1/; : : : ; h.k; m 1/i is a permutation of h0; 1; : : : ; m 1i for each key k. None of these techniques fulfills the assumption of uniform hashing, however, since none of them is capable of generating more than m2 different probe sequences (instead of the mŠ that uniform hashing requires). Double hashing has the greatest number of probe sequences and, as one might expect, seems to give the best results. 272 Chapter 11 Hash Tables Linear probing Given an ordinary hash function h0 W U ! f0; 1; : : : ; m 1g, which we refer to as an auxiliary hash function, the method of linear probing uses the hash function h.k; i / D .h0.k/ C i / mod m for i D 0; 1; : : : ; m 1. Given key k, we first probe T Œh0.k/, i.e., the slot given by the auxiliary hash function. We next probe slot T Œh0.k/ C 1, and so on up to slot T Œm 1. Then we wrap around to slots T Œ0; T Œ1; : : : until we finally probe slot T Œh0.k/ 1. Because the initial probe determines the entire probe sequence, there are only m distinct probe sequences. Linear probing is easy to implement, but it suffers from a problem known as primary clustering. Long runs of occupied slots build up, increasing the average search time. Clusters arise because an empty slot preceded by i full slots gets filled next with probability .i C 1/=m. Long runs of occupied slots tend to get longer, and the average search time increases. Quadratic probing Quadratic probing uses a hash function of the form h.k; i / D .h0.k/ C c1i C c2i 2/ mod m ; (11.5) where h0 is an auxiliary hash function, c1 and c2 are positive auxiliary constants, and i D 0; 1; : : : ; m 1. The initial position probed is T Œh0.k/; later positions probed are offset by amounts that depend in a quadratic manner on the probe number i. This method works much better than linear probing, but to make full use of the hash table, the values of c1, c2, and m are constrained. Problem 11-3 shows one way to select these parameters. Also, if two keys have the same initial probe position, then their probe sequences are the same, since h.k1; 0/ D h.k2; 0/ implies h.k1; i / D h.k2; i /. This property leads to a milder form of clustering, called secondary clustering. As in linear probing, the initial probe determines the entire sequence, and so only m distinct probe sequences are used. Double hashing Double hashing offers one of the best methods available for open addressing because the permutations produced have many of the characteristics of randomly chosen permutations. Double hashing uses a hash function of the form h.k; i / D .h1.k/ C ih2.k// mod m ; where both h1 and h2 are auxiliary hash functions. The initial probe goes to position T Œh1.k/; successive probe positions are offset from previous positions by the 11.4 Open addressing 273 0 1 79 2 3 4 69 5 98 6 7 72 8 9 14 10 11 50 12 Figure 11.5 Insertion by double hashing. Here we have a hash table of size 13 with h1.k/ D k mod 13 and h2.k/ D 1 C .k mod 11/. Since 14 Á 1 .mod 13/ and 14 Á 3 .mod 11/, we insert the key 14 into empty slot 9, after examining slots 1 and 5 and finding them to be occupied. amount h2.k/, modulo m. Thus, unlike the case of linear or quadratic probing, the probe sequence here depends in two ways upon the key k, since the initial probe position, the offset, or both, may vary. Figure 11.5 gives an example of insertion by double hashing. The value h2.k/ must be relatively prime to the hash-table size m for the entire hash table to be searched. (See Exercise 11.4-4.) A convenient way to ensure this condition is to let m be a power of 2 and to design h2 so that it always produces an odd number. Another way is to let m be prime and to design h2 so that it always returns a positive integer less than m. For example, we could choose m prime and let h1.k/ D k mod m ; h2.k/ D 1 C .k mod m0/ ; where m0 is chosen to be slightly less than m (say, m 1). For example, if k D 123456, m D 701, and m0 D 700, we have h1.k/ D 80 and h2.k/ D 257, so that we first probe position 80, and then we examine every 257th slot (modulo m) until we find the key or have examined every slot. When m is prime or a power of 2, double hashing improves over linear or quadratic probing in that ‚.m2/ probe sequences are used, rather than ‚.m/, since each possible .h1.k/; h2.k// pair yields a distinct probe sequence. As a result, for 274 Chapter 11 Hash Tables such values of m, the performance of double hashing appears to be very close to the performance of the “ideal” scheme of uniform hashing. Although values of m other than primes or powers of 2 could in principle be used with double hashing, in practice it becomes more difficult to efficiently generate h2.k/ in a way that ensures that it is relatively prime to m, in part because the relative density .m/=m of such numbers may be small (see equation (31.24)). Analysis of open-address hashing As in our analysis of chaining, we express our analysis of open addressing in terms of the load factor ˛ D n=m of the hash table. Of course, with open addressing, at most one element occupies each slot, and thus n Ä m, which implies ˛ Ä 1. We assume that we are using uniform hashing. In this idealized scheme, the probe sequence hh.k; 0/; h.k; 1/; : : : ; h.k; m 1/i used to insert or search for each key k is equally likely to be any permutation of h0; 1; : : : ; m 1i. Of course, a given key has a unique fixed probe sequence associated with it; what we mean here is that, considering the probability distribution on the space of keys and the operation of the hash function on the keys, each possible probe sequence is equally likely. We now analyze the expected number of probes for hashing with open addressing under the assumption of uniform hashing, beginning with an analysis of the number of probes made in an unsuccessful search. Theorem 11.6 Given an open-address hash table with load factor ˛ D n=m < 1, the expected number of probes in an unsuccessful search is at most 1=.1 ˛/, assuming uniform hashing. Proof In an unsuccessful search, every probe but the last accesses an occupied slot that does not contain the desired key, and the last slot probed is empty. Let us define the random variable X to be the number of probes made in an unsuccessful search, and let us also define the event Ai , for i D 1; 2; : : :, to be the event that an ith probe occurs and it is to an occupied slot. Then the event fX ig is the intersection of events A1 \ A2 \ \ Ai 1. We will bound Pr fX i g by bounding Pr fA1 \ A2 \ \ Ai 1g. By Exercise C.2-5, Pr fA1 \ A2 \ \ Ai 1g D Pr fA1g Pr fA2 j A1g Pr fA3 j A1 \ A2g Pr fAi 1 j A1 \ A2 \ \ Ai 2g : Since there are n elements and m slots, Pr fA1g D n=m. For j > 1, the probability that there is a j th probe and it is to an occupied slot, given that the first j 1 probes were to occupied slots, is .n j C 1/=.m j C 1/. This probability follows 11.4 Open addressing 275 because we would be finding one of the remaining .n .j 1// elements in one of the .m .j 1// unexamined slots, and by the assumption of uniform hashing, the probability is the ratio of these quantities. Observing that n < m implies that .n j /=.m j / Ä n=m for all j such that 0 Ä j < m, we have for all i such that 1 Ä i Ä m, Pr fX nn1n2 ig D mm 1m 2 n Ái 1 Ä m D ˛i 1 : n iC2 m iC2 Now, we use equation (C.25) to bound the expected number of probes: X 1 E ŒX  D Pr fX ig i D1 X 1 Ä ˛i 1 i D1 X 1 D ˛i i D0 1 D : 1˛ This bound of 1=.1 ˛/ D 1 C ˛ C ˛2 C ˛3 C has an intuitive interpretation. We always make the first probe. With probability approximately ˛, the first probe finds an occupied slot, so that we need to probe a second time. With probability approximately ˛2, the first two slots are occupied so that we make a third probe, and so on. If ˛ is a constant, Theorem 11.6 predicts that an unsuccessful search runs in O.1/ time. For example, if the hash table is half full, the average number of probes in an unsuccessful search is at most 1=.1 :5/ D 2. If it is 90 percent full, the average number of probes is at most 1=.1 :9/ D 10. Theorem 11.6 gives us the performance of the HASH-INSERT procedure almost immediately. Corollary 11.7 Inserting an element into an open-address hash table with load factor ˛ requires at most 1=.1 ˛/ probes on average, assuming uniform hashing. 276 Chapter 11 Hash Tables Proof An element is inserted only if there is room in the table, and thus ˛ < 1. Inserting a key requires an unsuccessful search followed by placing the key into the first empty slot found. Thus, the expected number of probes is at most 1=.1 ˛/. We have to do a little more work to compute the expected number of probes for a successful search. Theorem 11.8 Given an open-address hash table with load factor ˛ < 1, the expected number of probes in a successful search is at most 11 ln ; ˛ 1˛ assuming uniform hashing and assuming that each key in the table is equally likely to be searched for. Proof A search for a key k reproduces the same probe sequence as when the element with key k was inserted. By Corollary 11.7, if k was the .i C 1/st key inserted into the hash table, the expected number of probes made in a search for k is at most 1=.1 i=m/ D m=.m i/. Averaging over all n keys in the hash table gives us the expected number of probes in a successful search: 1 X n 1 m m X n 1 1 D n mi n mi i D0 i D0 1 Xm 1 D ˛ k 1 kZDmm nC1 Ä .1=x/ dx (by inequality (A.12)) ˛ mn D 1 ln m ˛ mn D 1 ln 1 : ˛ 1˛ If the hash table is half full, the expected number of probes in a successful search is less than 1:387. If the hash table is 90 percent full, the expected number of probes is less than 2:559. 11.5 Perfect hashing 277 Exercises 11.4-1 Consider inserting the keys 10; 22; 31; 4; 15; 28; 17; 88; 59 into a hash table of length m D 11 using open addressing with the auxiliary hash function h0.k/ D k. Illustrate the result of inserting these keys using linear probing, using quadratic probing with c1 D 1 and c2 D 3, and using double hashing with h1.k/ D k and h2.k/ D 1 C .k mod .m 1//. 11.4-2 Write pseudocode for HASH-DELETE as outlined in the text, and modify HASHINSERT to handle the special value DELETED. 11.4-3 Consider an open-address hash table with uniform hashing. Give upper bounds on the expected number of probes in an unsuccessful search and on the expected number of probes in a successful search when the load factor is 3=4 and when it is 7=8. 11.4-4 ? Suppose that we use double hashing to resolve collisions—that is, we use the hash function h.k; i / D .h1.k/ C ih2.k// mod m. Show that if m and h2.k/ have greatest common divisor d 1 for some key k, then an unsuccessful search for key k examines .1=d /th of the hash table before returning to slot h1.k/. Thus, when d D 1, so that m and h2.k/ are relatively prime, the search may examine the entire hash table. (Hint: See Chapter 31.) 11.4-5 ? Consider an open-address hash table with a load factor ˛. Find the nonzero value ˛ for which the expected number of probes in an unsuccessful search equals twice the expected number of probes in a successful search. Use the upper bounds given by Theorems 11.6 and 11.8 for these expected numbers of probes. ? 11.5 Perfect hashing Although hashing is often a good choice for its excellent average-case performance, hashing can also provide excellent worst-case performance when the set of keys is static: once the keys are stored in the table, the set of keys never changes. Some applications naturally have static sets of keys: consider the set of reserved words in a programming language, or the set of file names on a CD-ROM. We 278 Chapter 11 Hash Tables T 0 1 2 3 4 5 6 7 8 m0 a0 b0 S0 1 0 0 10 m2 a2 b2 0 9 10 18 01 m5 a5 b5 S5 1 0 0 70 m7 a7 b7 0 16 23 88 01 S2 60 72 2345 2345 75 678 S7 40 52 22 37 6 7 8 9 10 11 12 13 14 15 Figure 11.6 Using perfect hashing to store the set K D f10; 22; 37; 40; 52; 60; 70; 72; 75g. The outer hash function is h.k/ D ..ak C b/ mod p/ mod m, where a D 3, b D 42, p D 101, and m D 9. For example, h.75/ D 2, and so key 75 hashes to slot 2 of table T . A secondary hash table Sj stores all keys hashing to slot j . The size of hash table Sj is mj D nj2, and the associated hash function is hj .k/ D ..aj k C bj / mod p/ mod mj . Since h2.75/ D 7, key 75 is stored in slot 7 of secondary hash table S2. No collisions occur in any of the secondary hash tables, and so searching takes constant time in the worst case. call a hashing technique perfect hashing if O.1/ memory accesses are required to perform a search in the worst case. To create a perfect hashing scheme, we use two levels of hashing, with universal hashing at each level. Figure 11.6 illustrates the approach. The first level is essentially the same as for hashing with chaining: we hash the n keys into m slots using a hash function h carefully selected from a family of universal hash functions. Instead of making a linked list of the keys hashing to slot j , however, we use a small secondary hash table Sj with an associated hash function hj . By choosing the hash functions hj carefully, we can guarantee that there are no collisions at the secondary level. In order to guarantee that there are no collisions at the secondary level, however, we will need to let the size mj of hash table Sj be the square of the number nj of keys hashing to slot j . Although you might think that the quadratic dependence of mj on nj may seem likely to cause the overall storage requirement to be excessive, we shall show that by choosing the first-level hash function well, we can limit the expected total amount of space used to O.n/. We use hash functions chosen from the universal classes of hash functions of Section 11.3.3. The first-level hash function comes from the class Hpm, where as in Section 11.3.3, p is a prime number greater than any key value. Those keys 11.5 Perfect hashing 279 hashing to slot j are re-hashed into a secondary hash table Sj of size mj using a hash function hj chosen from the class Hp;mj .1 We shall proceed in two steps. First, we shall determine how to ensure that the secondary tables have no collisions. Second, we shall show that the expected amount of memory used overall—for the primary hash table and all the secondary hash tables—is O.n/. Theorem 11.9 Suppose that we store n keys in a hash table of size m D n2 using a hash function h randomly chosen from a universal class of hash functions. Then, the probability is less than 1=2 that there are any collisions. Proof There are n 2 pairs of keys that may collide; each pair collides with prob- ability 1=m if h is chosen at random from a universal family H of hash functions. Let X be a random variable that counts the number of collisions. When m D n2, the expected number of collisions is ! n1 E ŒX  D 2 n2 n2 n 1 D 2 n2 < 1=2 : (This analysis is similar to the analysis of the birthday paradox in Section 5.4.1.) Applying Markov’s inequality (C.30), Pr fX tg Ä E ŒX  =t, with t D 1, completes the proof. In the situation described in Theorem 11.9, where m D n2, it follows that a hash function h chosen at random from H is more likely than not to have no collisions. Given the set K of n keys to be hashed (remember that K is static), it is thus easy to find a collision-free hash function h with a few random trials. When n is large, however, a hash table of size m D n2 is excessive. Therefore, we adopt the two-level hashing approach, and we use the approach of Theorem 11.9 only to hash the entries within each slot. We use an outer, or first-level, hash function h to hash the keys into m D n slots. Then, if nj keys hash to slot j , we use a secondary hash table Sj of size mj D nj2 to provide collision-free constanttime lookup. 1When nj D mj D 1, we don’t really need a hash function for slot j ; when we choose a hash function hab .k/ D ..ak C b/ mod p/ mod mj for such a slot, we just use a D b D 0. 280 Chapter 11 Hash Tables We now turn to the issue of ensuring that the overall memory used is O.n/. Since the size mj of the j th secondary hash table grows quadratically with the number nj of keys stored, we run the risk that the overall amount of storage could be excessive. If the first-level table size is m D n, then the amount of memory used is O.n/ for the primary hash table, for the storage of the sizes mj of the secondary hash tables, and for the storage of the parameters aj and bj defining the secondary hash functions hj drawn from the class Hp;mj of Section 11.3.3 (except when nj D 1 and we use a D b D 0). The following theorem and a corollary provide a bound on the expected combined sizes of all the secondary hash tables. A second corollary bounds the probability that the combined size of all the secondary hash tables is superlinear (actually, that it equals or exceeds 4n). Theorem 11.10 Suppose that we store n keys in a hash table of size m D n using a hash function h randomly chosen from a universal class of hash functions. Then, we have " # mX1 E nj2 < 2n ; j D0 where nj is the number of keys hashing to slot j . Proof We start with the following identity, which holds for any nonnegative inte- ger a: ! a2 D a C 2 a : 2 (11.6) We have " # mX1 E nj2 j D0 D D D " !!# E mX1 nj C 2 nj 2 "j D0 # " !# mX1 E nj C2E mX1 nj 2 j D0 " j D!0# E Œn C 2 E mX1 nj 2 j D0 (by equation (11.6)) (by linearity of expectation) (by equation (11.1)) 11.5 Perfect hashing 281 " !# D n C 2 E mX1 nj 2 j D0 (since n is not a random variable) . To evaluate the summation Pm 1 j D0 nj 2 , we observe that it is just the total number of pairs of keys in the hash table that collide. By the properties of universal hashing, the expected value of this summation is at most ! n1 n.n 1/ D 2m 2m n1 D ; 2 since m D n. Thus, " # mX1 E nj2 Ä nC2 n 1 2 j D0 D 2n 1 < 2n : Corollary 11.11 Suppose that we store n keys in a hash table of size m D n using a hash func- tion h randomly chosen from a universal class of hash functions, and we set the size of each secondary hash table to mj D nj2 for j D 0; 1; : : : ; m 1. Then, the expected amount of storage required for all secondary hash tables in a perfect hashing scheme is less than 2n. Proof Since mj D nj2 for j D 0; 1; : : : ; m " # " # mX1 mX1 E mj D E nj2 j D0 j D0 < 2n ; 1, Theorem 11.10 gives which completes the proof. (11.7) Corollary 11.12 Suppose that we store n keys in a hash table of size m D n using a hash function h randomly chosen from a universal class of hash functions, and we set the size of each secondary hash table to mj D nj2 for j D 0; 1; : : : ; m 1. Then, the probability is less than 1=2 that the total storage used for secondary hash tables equals or exceeds 4n. 282 Chapter 11 Hash Tables Proof Again we apply time to inequality (11.7), ( ) mX1 Pr mj 4n Ä MwiatrhkoXv’Ds inPeqjmuDa01limtyj E Pm 1 j D0 mj 4n (C.30), Pr fX and t D 4n: j D0 2n < 4n D 1=2 : tg Ä E ŒX  =t, this From Corollary 11.12, we see that if we test a few randomly chosen hash functions from the universal family, we will quickly find one that uses a reasonable amount of storage. Exercises 11.5-1 ? Suppose that we insert n keys into a hash table of size m using open addressing and uniform hashing. Let p.n; m/ be the probability that no collisions occur. Show that pp.n; m/ Ä e n.n 1/=2m. (Hint: See equation (3.12).) Argue that when n exceeds m, the probability of avoiding collisions goes rapidly to zero. Problems 11-1 Longest-probe bound for hashing Suppose that we use an open-addressed hash table of size m to store n Ä m=2 items. a. Assuming uniform hashing, show that for i D 1; 2; : : : ; n, the probability is at most 2 k that the ith insertion requires strictly more than k probes. b. Show that for i D 1; 2; : : : ; n, the probability is O.1=n2/ that the ith insertion requires more than 2 lg n probes. Let the random variable Xi denote the number of probes required by the ith insertion. You have shown in part (b) that Pr fXi > 2 lg ng D O.1=n2/. Let the random variable X D max1ÄiÄn Xi denote the maximum number of probes required by any of the n insertions. c. Show that Pr fX > 2 lg ng D O.1=n/. d. Show that the expected length E ŒX  of the longest probe sequence is O.lg n/. Problems for Chapter 11 283 11-2 Slot-size bound for chaining Suppose that we have a hash table with n slots, with collisions resolved by chaining, and suppose that n keys are inserted into the table. Each key is equally likely to be hashed to each slot. Let M be the maximum number of keys in any slot after all the keys have been inserted. Your mission is to prove an O.lg n= lg lg n/ upper bound on E ŒM , the expected value of M . a. Argue that the probability Qk that exactly k keys hash to a particular slot is given by  1 Ãk Qk D n 1 1 Ãn k ! n : n k b. Let Pk be the probability that M D k, that is, the probability that the slot containing the most keys contains k keys. Show that Pk Ä nQk. c. Use Stirling’s approximation, equation (3.18), to show that Qk < ek=kk. d. Show that there exists a constant c > 1 such that Qk0 < 1=n3 for k0 D c lg n= lg lg n. Conclude that Pk < 1=n2 for k k0 D c lg n= lg lg n. e. Argue that c lg n c lg n E ŒM  Ä Pr M > n C Pr M Ä lg lg n lg lg n Conclude that E ŒM  D O.lg n= lg lg n/. c lg n : lg lg n 11-3 Quadratic probing Suppose that we are given a key k to search for in a hash table with positions 0; 1; : : : ; m 1, and suppose that we have a hash function h mapping the key space into the set f0; 1; : : : ; m 1g. The search scheme is as follows: 1. Compute the value j D h.k/, and set i D 0. 2. Probe in position j for the desired key k. If you find it, or if this position is empty, terminate the search. 3. Set i D i C 1. If i now equals m, the table is full, so terminate the search. Otherwise, set j D .i C j / mod m, and return to step 2. Assume that m is a power of 2. a. Show that this scheme is an instance of the general “quadratic probing” scheme by exhibiting the appropriate constants c1 and c2 for equation (11.5). b. Prove that this algorithm examines every table position in the worst case. 284 Chapter 11 Hash Tables 11-4 Hashing and authentication Let H be a class of hash functions in which each hash function h 2 H maps the universe U of keys to f0; 1; : : : ; m 1g. We say that H is k-universal if, for every fixed sequence of k distinct keys hx.1/; x.2/; : : : ; x.k/i and for any h chosen at random from H , the sequence hh.x.1//; h.x.2//; : : : ; h.x.k//i is equally likely to be any of the mk sequences of length k with elements drawn from f0; 1; : : : ; m 1g. a. Show that if the family H of hash functions is 2-universal, then it is universal. b. Suppose that the universe U is the set of n-tuples of values drawn from Zp D f0; 1; : : : ; p 1g, where p is prime. Consider an element x D hx0; x1; : : : ; xn 1i 2 U . For any n-tuple a D ha0; a1; : : : ; an 1i 2 U , define the hash function ha by ! X n 1 ha.x/ D aj xj mod p : j D0 Let H D fhag. Show that H is universal, but not 2-universal. (Hint: Find a key for which all hash functions in H produce the same value.) c. Suppose that we modify H slightly from part (b): for any a 2 U and for any b 2 Zp, define h0ab.x/ D ! X n 1 aj xj C b mod p j D0 and H 0 D fh0abg. Argue that H 0 is 2-universal. (Hint: Consider fixed n-tuples x 2 U and y 2 U , with xi ¤ yi for some i . What happens to h0ab.x/ and h0ab.y/ as ai and b range over Zp?) d. Suppose that Alice and Bob secretly agree on a hash function h from a 2-universal family H of hash functions. Each h 2 H maps from a universe of keys U to Zp, where p is prime. Later, Alice sends a message m to Bob over the Internet, where m 2 U . She authenticates this message to Bob by also sending an authentication tag t D h.m/, and Bob checks that the pair .m; t/ he receives indeed satisfies t D h.m/. Suppose that an adversary intercepts .m; t/ en route and tries to fool Bob by replacing the pair .m; t/ with a different pair .m0; t0/. Argue that the probability that the adversary succeeds in fooling Bob into accepting .m0; t0/ is at most 1=p, no matter how much computing power the adversary has, and even if the adversary knows the family H of hash functions used. Notes for Chapter 11 285 Chapter notes Knuth [211] and Gonnet [145] are excellent references for the analysis of hashing algorithms. Knuth credits H. P. Luhn (1953) for inventing hash tables, along with the chaining method for resolving collisions. At about the same time, G. M. Amdahl originated the idea of open addressing. Carter and Wegman introduced the notion of universal classes of hash functions in 1979 [58]. Fredman, Komlo´s, and Szemere´di [112] developed the perfect hashing scheme for static sets presented in Section 11.5. An extension of their method to dynamic sets, handling insertions and deletions in amortized expected time O.1/, has been given by Dietzfelbinger et al. [86]. 12 Binary Search Trees The search tree data structure supports many dynamic-set operations, including SEARCH, MINIMUM, MAXIMUM, PREDECESSOR, SUCCESSOR, INSERT, and DELETE. Thus, we can use a search tree both as a dictionary and as a priority queue. Basic operations on a binary search tree take time proportional to the height of the tree. For a complete binary tree with n nodes, such operations run in ‚.lg n/ worst-case time. If the tree is a linear chain of n nodes, however, the same operations take ‚.n/ worst-case time. We shall see in Section 12.4 that the expected height of a randomly built binary search tree is O.lg n/, so that basic dynamic-set operations on such a tree take ‚.lg n/ time on average. In practice, we can’t always guarantee that binary search trees are built randomly, but we can design variations of binary search trees with good guaranteed worst-case performance on basic operations. Chapter 13 presents one such variation, red-black trees, which have height O.lg n/. Chapter 18 introduces B-trees, which are particularly good for maintaining databases on secondary (disk) storage. After presenting the basic properties of binary search trees, the following sections show how to walk a binary search tree to print its values in sorted order, how to search for a value in a binary search tree, how to find the minimum or maximum element, how to find the predecessor or successor of an element, and how to insert into or delete from a binary search tree. The basic mathematical properties of trees appear in Appendix B. 12.1 What is a binary search tree? A binary search tree is organized, as the name suggests, in a binary tree, as shown in Figure 12.1. We can represent such a tree by a linked data structure in which each node is an object. In addition to a key and satellite data, each node contains attributes left, right, and p that point to the nodes corresponding to its left child, 12.1 What is a binary search tree? 6 5 2 5 7 8 (a) 287 2 5 7 68 5 (b) Figure 12.1 Binary search trees. For any node x, the keys in the left subtree of x are at most x: key, and the keys in the right subtree of x are at least x: key. Different binary search trees can represent the same set of values. The worst-case running time for most search-tree operations is proportional to the height of the tree. (a) A binary search tree on 6 nodes with height 2. (b) A less efficient binary search tree with height 4 that contains the same keys. its right child, and its parent, respectively. If a child or the parent is missing, the appropriate attribute contains the value NIL. The root node is the only node in the tree whose parent is NIL. The keys in a binary search tree are always stored in such a way as to satisfy the binary-search-tree property: Let x be a node in a binary search tree. If y is a node in the left subtree of x, then y:key Ä x:key. If y is a node in the right subtree of x, then y:key x:key. Thus, in Figure 12.1(a), the key of the root is 6, the keys 2, 5, and 5 in its left subtree are no larger than 6, and the keys 7 and 8 in its right subtree are no smaller than 6. The same property holds for every node in the tree. For example, the key 5 in the root’s left child is no smaller than the key 2 in that node’s left subtree and no larger than the key 5 in the right subtree. The binary-search-tree property allows us to print out all the keys in a binary search tree in sorted order by a simple recursive algorithm, called an inorder tree walk. This algorithm is so named because it prints the key of the root of a subtree between printing the values in its left subtree and printing those in its right subtree. (Similarly, a preorder tree walk prints the root before the values in either subtree, and a postorder tree walk prints the root after the values in its subtrees.) To use the following procedure to print all the elements in a binary search tree T , we call INORDER-TREE-WALK.T: root/. 288 Chapter 12 Binary Search Trees INORDER-TREE-WALK .x/ 1 if x ¤ NIL 2 INORDER-TREE-WALK.x: left/ 3 print x:key 4 INORDER-TREE-WALK.x: right/ As an example, the inorder tree walk prints the keys in each of the two binary search trees from Figure 12.1 in the order 2; 5; 5; 6; 7; 8. The correctness of the algorithm follows by induction directly from the binary-search-tree property. It takes ‚.n/ time to walk an n-node binary search tree, since after the initial call, the procedure calls itself recursively exactly twice for each node in the tree—once for its left child and once for its right child. The following theorem gives a formal proof that it takes linear time to perform an inorder tree walk. Theorem 12.1 If x is the root of an n-node subtree, then the call INORDER-TREE-WALK.x/ takes ‚.n/ time. Proof Let T .n/ denote the time taken by INORDER-TREE-WALK when it is called on the root of an n-node subtree. Since INORDER-TREE-WALK visits all n nodes of the subtree, we have T .n/ D .n/. It remains to show that T .n/ D O.n/. Since INORDER-TREE-WALK takes a small, constant amount of time on an empty subtree (for the test x ¤ NIL), we have T .0/ D c for some constant c > 0. For n > 0, suppose that INORDER-TREE-WALK is called on a node x whose left subtree has k nodes and whose right subtree has n k 1 nodes. The time to perform INORDER-TREE-WALK.x/ is bounded by T .n/ Ä T .k/CT .n k 1/Cd for some constant d > 0 that reflects an upper bound on the time to execute the body of INORDER-TREE-WALK.x/, exclusive of the time spent in recursive calls. We use the substitution method to show that T .n/ D O.n/ by proving that T .n/ Ä .c C d /n C c. For n D 0, we have .c C d / 0 C c D c D T .0/. For n > 0, we have T .n/ Ä T .k/ C T .n k 1/ C d D ..c C d /k C c/ C ..c C d /.n k D .c C d /n C c .c C d / C c C d D .c C d /n C c ; 1/ C c/ C d which completes the proof. 12.2 Querying a binary search tree 289 Exercises 12.1-1 For the set of f1; 4; 5; 10; 16; 17; 21g of keys, draw binary search trees of heights 2, 3, 4, 5, and 6. 12.1-2 What is the difference between the binary-search-tree property and the min-heap property (see page 153)? Can the min-heap property be used to print out the keys of an n-node tree in sorted order in O.n/ time? Show how, or explain why not. 12.1-3 Give a nonrecursive algorithm that performs an inorder tree walk. (Hint: An easy solution uses a stack as an auxiliary data structure. A more complicated, but elegant, solution uses no stack but assumes that we can test two pointers for equality.) 12.1-4 Give recursive algorithms that perform preorder and postorder tree walks in ‚.n/ time on a tree of n nodes. 12.1-5 Argue that since sorting n elements takes .n lg n/ time in the worst case in the comparison model, any comparison-based algorithm for constructing a binary search tree from an arbitrary list of n elements takes .n lg n/ time in the worst case. 12.2 Querying a binary search tree We often need to search for a key stored in a binary search tree. Besides the SEARCH operation, binary search trees can support such queries as MINIMUM, MAXIMUM, SUCCESSOR, and PREDECESSOR. In this section, we shall examine these operations and show how to support each one in time O.h/ on any binary search tree of height h. Searching We use the following procedure to search for a node with a given key in a binary search tree. Given a pointer to the root of the tree and a key k, TREE-SEARCH returns a pointer to a node with key k if one exists; otherwise, it returns NIL. 290 Chapter 12 Binary Search Trees 6 3 24 15 18 7 17 20 13 9 Figure 12.2 Queries on a binary search tree. To search for the key 13 in the tree, we follow the path 15 ! 6 ! 7 ! 13 from the root. The minimum key in the tree is 2, which is found by following left pointers from the root. The maximum key 20 is found by following right pointers from the root. The successor of the node with key 15 is the node with key 17, since it is the minimum key in the right subtree of 15. The node with key 13 has no right subtree, and thus its successor is its lowest ancestor whose left child is also an ancestor. In this case, the node with key 15 is its successor. TREE-SEARCH.x; k/ 1 if x = = NIL or k == x:key 2 return x 3 if k < x:key 4 return TREE-SEARCH.x:left; k/ 5 else return TREE-SEARCH.x:right; k/ The procedure begins its search at the root and traces a simple path downward in the tree, as shown in Figure 12.2. For each node x it encounters, it compares the key k with x:key. If the two keys are equal, the search terminates. If k is smaller than x:key, the search continues in the left subtree of x, since the binary-searchtree property implies that k could not be stored in the right subtree. Symmetrically, if k is larger than x:key, the search continues in the right subtree. The nodes encountered during the recursion form a simple path downward from the root of the tree, and thus the running time of TREE-SEARCH is O.h/, where h is the height of the tree. We can rewrite this procedure in an iterative fashion by “unrolling” the recursion into a while loop. On most computers, the iterative version is more efficient. 12.2 Querying a binary search tree 291 ITERATIVE-TREE-SEARCH.x; k/ 1 while x ¤ NIL and k ¤ x:key 2 if k < x:key 3 x D x:left 4 else x D x:right 5 return x Minimum and maximum We can always find an element in a binary search tree whose key is a minimum by following left child pointers from the root until we encounter a NIL, as shown in Figure 12.2. The following procedure returns a pointer to the minimum element in the subtree rooted at a given node x, which we assume to be non-NIL: T R EE- M I N I M U M .x / 1 while x:left ¤ NIL 2 x D x:left 3 return x The binary-search-tree property guarantees that TREE-MINIMUM is correct. If a node x has no left subtree, then since every key in the right subtree of x is at least as large as x:key, the minimum key in the subtree rooted at x is x:key. If node x has a left subtree, then since no key in the right subtree is smaller than x:key and every key in the left subtree is not larger than x:key, the minimum key in the subtree rooted at x resides in the subtree rooted at x:left. The pseudocode for TREE-MAXIMUM is symmetric: T R E E - M A X I M U M .x / 1 while x:right ¤ NIL 2 x D x:right 3 return x Both of these procedures run in O.h/ time on a tree of height h since, as in TREESEARCH, the sequence of nodes encountered forms a simple path downward from the root. Successor and predecessor Given a node in a binary search tree, sometimes we need to find its successor in the sorted order determined by an inorder tree walk. If all keys are distinct, the 292 Chapter 12 Binary Search Trees successor of a node x is the node with the smallest key greater than x:key. The structure of a binary search tree allows us to determine the successor of a node without ever comparing keys. The following procedure returns the successor of a node x in a binary search tree if it exists, and NIL if x has the largest key in the tree: TREE-SUCCESSOR .x/ 1 if x:right ¤ NIL 2 return TREE-MINIMUM.x:right/ 3 y D x:p 4 while y ¤ NIL and x == y:right 5 xDy 6 y D y:p 7 return y We break the code for TREE-SUCCESSOR into two cases. If the right subtree of node x is nonempty, then the successor of x is just the leftmost node in x’s right subtree, which we find in line 2 by calling TREE-MINIMUM.x:right/. For example, the successor of the node with key 15 in Figure 12.2 is the node with key 17. On the other hand, as Exercise 12.2-6 asks you to show, if the right subtree of node x is empty and x has a successor y, then y is the lowest ancestor of x whose left child is also an ancestor of x. In Figure 12.2, the successor of the node with key 13 is the node with key 15. To find y, we simply go up the tree from x until we encounter a node that is the left child of its parent; lines 3–7 of TREE-SUCCESSOR handle this case. The running time of TREE-SUCCESSOR on a tree of height h is O.h/, since we either follow a simple path up the tree or follow a simple path down the tree. The procedure TREE-PREDECESSOR, which is symmetric to TREE-SUCCESSOR, also runs in time O.h/. Even if keys are not distinct, we define the successor and predecessor of any node x as the node returned by calls made to TREE-SUCCESSOR.x/ and TREEPREDECESSOR.x/, respectively. In summary, we have proved the following theorem. Theorem 12.2 We can implement the dynamic-set operations SEARCH, MINIMUM, MAXIMUM, SUCCESSOR, and PREDECESSOR so that each one runs in O.h/ time on a binary search tree of height h. 12.2 Querying a binary search tree 293 Exercises 12.2-1 Suppose that we have numbers between 1 and 1000 in a binary search tree, and we want to search for the number 363. Which of the following sequences could not be the sequence of nodes examined? a. 2, 252, 401, 398, 330, 344, 397, 363. b. 924, 220, 911, 244, 898, 258, 362, 363. c. 925, 202, 911, 240, 912, 245, 363. d. 2, 399, 387, 219, 266, 382, 381, 278, 363. e. 935, 278, 347, 621, 299, 392, 358, 363. 12.2-2 Write recursive versions of TREE-MINIMUM and TREE-MAXIMUM. 12.2-3 Write the TREE-PREDECESSOR procedure. 12.2-4 Professor Bunyan thinks he has discovered a remarkable property of binary search trees. Suppose that the search for key k in a binary search tree ends up in a leaf. Consider three sets: A, the keys to the left of the search path; B, the keys on the search path; and C , the keys to the right of the search path. Professor Bunyan claims that any three keys a 2 A, b 2 B, and c 2 C must satisfy a Ä b Ä c. Give a smallest possible counterexample to the professor’s claim. 12.2-5 Show that if a node in a binary search tree has two children, then its successor has no left child and its predecessor has no right child. 12.2-6 Consider a binary search tree T whose keys are distinct. Show that if the right subtree of a node x in T is empty and x has a successor y, then y is the lowest ancestor of x whose left child is also an ancestor of x. (Recall that every node is its own ancestor.) 12.2-7 An alternative method of performing an inorder tree walk of an n-node binary search tree finds the minimum element in the tree by calling TREE-MINIMUM and then making n 1 calls to TREE-SUCCESSOR. Prove that this algorithm runs in ‚.n/ time. 294 Chapter 12 Binary Search Trees 12.2-8 Prove that no matter what node we start at in a height-h binary search tree, k successive calls to TREE-SUCCESSOR take O.k C h/ time. 12.2-9 Let T be a binary search tree whose keys are distinct, let x be a leaf node, and let y be its parent. Show that y:key is either the smallest key in T larger than x:key or the largest key in T smaller than x:key. 12.3 Insertion and deletion The operations of insertion and deletion cause the dynamic set represented by a binary search tree to change. The data structure must be modified to reflect this change, but in such a way that the binary-search-tree property continues to hold. As we shall see, modifying the tree to insert a new element is relatively straightforward, but handling deletion is somewhat more intricate. Insertion To insert a new value into a binary search tree T , we use the procedure TREEINSERT. The procedure takes a node ´ for which ´:key D , ´:left D NIL, and ´:right D NIL. It modifies T and some of the attributes of ´ in such a way that it inserts ´ into an appropriate position in the tree. TREE-INSERT.T; ´/ 1 y D NIL 2 x D T:root 3 while x ¤ NIL 4 y Dx 5 if ´:key < x:key 6 x D x:left 7 else x D x:right 8 ´:p D y 9 if y == NIL 10 T:root D ´ // tree T was empty 11 elseif ´:key < y:key 12 y:left D ´ 13 else y:right D ´ 12.3 Insertion and deletion 295 12 5 18 2 9 15 19 13 17 Figure 12.3 Inserting an item with key 13 into a binary search tree. Lightly shaded nodes indicate the simple path from the root down to the position where the item is inserted. The dashed line indicates the link in the tree that is added to insert the item. Figure 12.3 shows how TREE-INSERT works. Just like the procedures TREESEARCH and ITERATIVE-TREE-SEARCH, TREE-INSERT begins at the root of the tree and the pointer x traces a simple path downward looking for a NIL to replace with the input item ´. The procedure maintains the trailing pointer y as the parent of x. After initialization, the while loop in lines 3–7 causes these two pointers to move down the tree, going left or right depending on the comparison of ´:key with x:key, until x becomes NIL. This NIL occupies the position where we wish to place the input item ´. We need the trailing pointer y, because by the time we find the NIL where ´ belongs, the search has proceeded one step beyond the node that needs to be changed. Lines 8–13 set the pointers that cause ´ to be inserted. Like the other primitive operations on search trees, the procedure TREE-INSERT runs in O.h/ time on a tree of height h. Deletion The overall strategy for deleting a node ´ from a binary search tree T has three basic cases but, as we shall see, one of the cases is a bit tricky. If ´ has no children, then we simply remove it by modifying its parent to replace ´ with NIL as its child. If ´ has just one child, then we elevate that child to take ´’s position in the tree by modifying ´’s parent to replace ´ by ´’s child. If ´ has two children, then we find ´’s successor y—which must be in ´’s right subtree—and have y take ´’s position in the tree. The rest of ´’s original right subtree becomes y’s new right subtree, and ´’s left subtree becomes y’s new left subtree. This case is the tricky one because, as we shall see, it matters whether y is ´’s right child. 296 Chapter 12 Binary Search Trees The procedure for deleting a given node ´ from a binary search tree T takes as arguments pointers to T and ´. It organizes its cases a bit differently from the three cases outlined previously by considering the four cases shown in Figure 12.4. If ´ has no left child (part (a) of the figure), then we replace ´ by its right child, which may or may not be NIL. When ´’s right child is NIL, this case deals with the situation in which ´ has no children. When ´’s right child is non-NIL, this case handles the situation in which ´ has just one child, which is its right child. If ´ has just one child, which is its left child (part (b) of the figure), then we replace ´ by its left child. Otherwise, ´ has both a left and a right child. We find ´’s successor y, which lies in ´’s right subtree and has no left child (see Exercise 12.2-5). We want to splice y out of its current location and have it replace ´ in the tree. If y is ´’s right child (part (c)), then we replace ´ by y, leaving y’s right child alone. Otherwise, y lies within ´’s right subtree but is not ´’s right child (part (d)). In this case, we first replace y by its own right child, and then we replace ´ by y. In order to move subtrees around within the binary search tree, we define a subroutine TRANSPLANT, which replaces one subtree as a child of its parent with another subtree. When TRANSPLANT replaces the subtree rooted at node u with the subtree rooted at node , node u’s parent becomes node ’s parent, and u’s parent ends up having as its appropriate child. TRANSPLANT.T; u; / 1 if u:p == NIL 2 T:root D 3 elseif u == u:p:left 4 u:p:left D 5 else u:p:right D 6 if ¤ NIL 7 :p D u:p Lines 1–2 handle the case in which u is the root of T . Otherwise, u is either a left child or a right child of its parent. Lines 3–4 take care of updating u:p:left if u is a left child, and line 5 updates u:p:right if u is a right child. We allow to be NIL, and lines 6–7 update :p if is non-NIL. Note that TRANSPLANT does not attempt to update :left and :right; doing so, or not doing so, is the responsibility of TRANSPLANT’s caller. 12.3 Insertion and deletion q (a) z NIL r (b) l q z NIL (c) l q z y NIL x (d) l q z r y NIL x q r q l q y l x q z y l NIL r x 297 q y l r x Figure 12.4 Deleting a node ´ from a binary search tree. Node ´ may be the root, a left child of node q, or a right child of q. (a) Node ´ has no left child. We replace ´ by its right child r, which may or may not be NIL. (b) Node ´ has a left child l but no right child. We replace ´ by l. (c) Node ´ has two children; its left child is node l, its right child is its successor y, and y’s right child is node x. We replace ´ by y, updating y’s left child to become l, but leaving x as y’s right child. (d) Node ´ has two children (left child l and right child r), and its successor y ¤ r lies within the subtree rooted at r. We replace y by its own right child x, and we set y to be r’s parent. Then, we set y to be q’s child and the parent of l. 298 Chapter 12 Binary Search Trees With the TRANSPLANT procedure in hand, here is the procedure that deletes node ´ from binary search tree T : TREE-DELETE .T; ´/ 1 if ´:left = = NIL 2 TRANSPLANT.T; ´; ´:right/ 3 elseif ´:right = = NIL 4 TRANSPLANT.T; ´; ´: left/ 5 else y D TREE-MINIMUM.´:right/ 6 if y:p ¤ ´ 7 TRANSPLANT.T; y; y:right/ 8 y:right D ´:right 9 y:right:p D y 10 TRANSPLANT.T; ´; y/ 11 y:left D ´:left 12 y:left:p D y The TREE-DELETE procedure executes the four cases as follows. Lines 1–2 handle the case in which node ´ has no left child, and lines 3–4 handle the case in which ´ has a left child but no right child. Lines 5–12 deal with the remaining two cases, in which ´ has two children. Line 5 finds node y, which is the successor of ´. Because ´ has a nonempty right subtree, its successor must be the node in that subtree with the smallest key; hence the call to TREE-MINIMUM.´:right/. As we noted before, y has no left child. We want to splice y out of its current location, and it should replace ´ in the tree. If y is ´’s right child, then lines 10–12 replace ´ as a child of its parent by y and replace y’s left child by ´’s left child. If y is not ´’s left child, lines 7–9 replace y as a child of its parent by y’s right child and turn ´’s right child into y’s right child, and then lines 10–12 replace ´ as a child of its parent by y and replace y’s left child by ´’s left child. Each line of TREE-DELETE, including the calls to TRANSPLANT, takes constant time, except for the call to TREE-MINIMUM in line 5. Thus, TREE-DELETE runs in O.h/ time on a tree of height h. In summary, we have proved the following theorem. Theorem 12.3 We can implement the dynamic-set operations INSERT and DELETE so that each one runs in O.h/ time on a binary search tree of height h. 12.4 Randomly built binary search trees 299 Exercises 12.3-1 Give a recursive version of the TREE-INSERT procedure. 12.3-2 Suppose that we construct a binary search tree by repeatedly inserting distinct values into the tree. Argue that the number of nodes examined in searching for a value in the tree is one plus the number of nodes examined when the value was first inserted into the tree. 12.3-3 We can sort a given set of n numbers by first building a binary search tree containing these numbers (using TREE-INSERT repeatedly to insert the numbers one by one) and then printing the numbers by an inorder tree walk. What are the worstcase and best-case running times for this sorting algorithm? 12.3-4 Is the operation of deletion “commutative” in the sense that deleting x and then y from a binary search tree leaves the same tree as deleting y and then x? Argue why it is or give a counterexample. 12.3-5 Suppose that instead of each node x keeping the attribute x:p, pointing to x’s parent, it keeps x:succ, pointing to x’s successor. Give pseudocode for SEARCH, INSERT, and DELETE on a binary search tree T using this representation. These procedures should operate in time O.h/, where h is the height of the tree T . (Hint: You may wish to implement a subroutine that returns the parent of a node.) 12.3-6 When node ´ in TREE-DELETE has two children, we could choose node y as its predecessor rather than its successor. What other changes to TREE-DELETE would be necessary if we did so? Some have argued that a fair strategy, giving equal priority to predecessor and successor, yields better empirical performance. How might TREE-DELETE be changed to implement such a fair strategy? ? 12.4 Randomly built binary search trees We have shown that each of the basic operations on a binary search tree runs in O.h/ time, where h is the height of the tree. The height of a binary search 300 Chapter 12 Binary Search Trees tree varies, however, as items are inserted and deleted. If, for example, the n items are inserted in strictly increasing order, the tree will be a chain with height n 1. On the other hand, Exercise B.5-4 shows that h blg nc. As with quicksort, we can show that the behavior of the average case is much closer to the best case than to the worst case. Unfortunately, little is known about the average height of a binary search tree when both insertion and deletion are used to create it. When the tree is created by insertion alone, the analysis becomes more tractable. Let us therefore define a randomly built binary search tree on n keys as one that arises from inserting the keys in random order into an initially empty tree, where each of the nŠ permutations of the input keys is equally likely. (Exercise 12.4-3 asks you to show that this notion is different from assuming that every binary search tree on n keys is equally likely.) In this section, we shall prove the following theorem. Theorem 12.4 The expected height of a randomly built binary search tree on n distinct keys is O.lg n/. Proof We start by defining three random variables that help measure the height of a randomly built binary search tree. We denote the height of a randomly built binary search on n keys by Xn, and we define the exponential height Yn D 2Xn. When we build a binary search tree on n keys, we choose one key as that of the root, and we let Rn denote the random variable that holds this key’s rank within the set of n keys; that is, Rn holds the position that this key would occupy if the set of keys were sorted. The value of Rn is equally likely to be any element of the set f1; 2; : : : ; ng. If Rn D i, then the left subtree of the root is a randomly built binary search tree on i 1 keys, and the right subtree is a randomly built binary search tree on n i keys. Because the height of a binary tree is 1 more than the larger of the heights of the two subtrees of the root, the exponential height of a binary tree is twice the larger of the exponential heights of the two subtrees of the root. If we know that Rn D i , it follows that Yn D 2 max.Yi 1; Yn i / : As base cases, we have that Y1 D 1, because the exponential height of a tree with 1 node is 20 D 1 and, for convenience, we define Y0 D 0. Next, define indicator random variables Zn;1; Zn;2; : : : ; Zn;n, where Zn;i D I fRn D i g : Because Rn is equally likely to be any element of f1; 2; : : : ; ng, it follows that Pr fRn D ig D 1=n for i D 1; 2; : : : ; n, and hence, by Lemma 5.1, we have E ŒZn;i  D 1=n ; (12.1) 12.4 Randomly built binary search trees 301 for i D 1; 2; : : : ; n. Because exactly one value of Zn;i is 1 and all others are 0, we also have Xn Yn D Zn;i .2 max.Yi 1; Yn i // : i D1 We shall show that E ŒYn is polynomial in n, which will ultimately imply that E ŒXn D O.lg n/. We claim that the indicator random variable Zn;i D I fRn D i g is independent of the values of Yi 1 and Yn i . Having chosen Rn D i , the left subtree (whose exponential height is Yi 1) is randomly built on the i 1 keys whose ranks are less than i. This subtree is just like any other randomly built binary search tree on i 1 keys. Other than the number of keys it contains, this subtree’s structure is not affected at all by the choice of Rn D i, and hence the random variables Yi 1 and Zn;i are independent. Likewise, the right subtree, whose exponential height is Yn i , is randomly built on the n i keys whose ranks are greater than i . Its structure is independent of the value of Rn, and so the random variables Yn i and Zn;i are independent. Hence, we have " # Xn E ŒYn D E Zn;i .2 max.Yi 1; Yn i // i D1 Xn D E ŒZn;i .2 max.Yi 1; Yn i // (by linearity of expectation) i D1 Xn D E ŒZn;i  E Œ2 max.Yi 1; Yn i / (by independence) i D1 Xn 1 D n E Œ2 max.Yi 1; Yn i / i D1 (by equation (12.1)) 2 Xn D n E Œmax.Yi 1; Yn i / i D1 (by equation (C.22)) 2 Xn Ä n .E ŒYi 1 C E ŒYn i / i D1 (by Exercise C.3-4) . Since each term E ŒY0 ; E ŒY1 ; : : : ; E ŒYn 1 appears twice in the last summation, once as E ŒYi 1 and once as E ŒYn i , we have the recurrence 4 X n 1 E ŒYn Ä n E ŒYi  : i D0 (12.2) 302 Chapter 12 Binary Search Trees Using the substitution method, we shall show that for all positive integers n, the recurrence (12.2) h!as the solution 1 nC3 E ŒYn Ä 4 3 : In doing so, we shall use the identity ! ! X n 1 i C 3 nC3 D : 3 4 i D0 (12.3) (Exercise 12.4-1 asks you to prove this identity.) For the base cases, we note that the bounds 0 D Y0 D E ŒY0 Ä .1=4/ 3 3 D 1=4 and 1 D Y1 D E ŒY1 Ä .1=4/ 1C3 3 D 1 hold. For the inductive case, we have that 4 X n 1 E ŒYn Ä n E ŒYi  i D0 ! 4 X n 1 1 i C 3 Ä (by the inductive hypothesis) n 43 i D0 ! 1 X n 1 i C 3 D n 3 iD0 ! 1 nC3 D n4 (by equation (12.3)) D 1 .n C 3/Š n 4Š .n 1/Š 1 .n C 3/Š D 4 3Š n!Š D 1 nC3 : 43 We have bounded E ŒYn, but our ultimate goal is to bound E ŒXn. As Exercise 12.4-4 asks you to show, the function f .x/ D 2x is convex (see page 1199). Therefore, we can employ Jensen’s inequality (C.26), which says that 2EŒXn Ä E 2Xn D E ŒYn ; as follows: ! 2EŒXn Ä 1 n C 3 43 Problems for Chapter 12 303 D 1 .n C 3/.n C 2/.n C 1/ 4 6 n3 C 6n2 C 11n C 6 D : 24 Taking logarithms of both sides gives E ŒXn D O.lg n/. Exercises 12.4-1 Prove equation (12.3). 12.4-2 Describe a binary search tree on n nodes such that the average depth of a node in the tree is ‚.lg n/ but the height of the tree is !.lg n/. Give an asymptotic upper bound on the height of an n-node binary search tree in which the average depth of a node is ‚.lg n/. 12.4-3 Show that the notion of a randomly chosen binary search tree on n keys, where each binary search tree of n keys is equally likely to be chosen, is different from the notion of a randomly built binary search tree given in this section. (Hint: List the possibilities when n D 3.) 12.4-4 Show that the function f .x/ D 2x is convex. 12.4-5 ? Consider RANDOMIZED-QUICKSORT operating on a sequence of n distinct input numbers. Prove that for any constant k > 0, all but O.1=nk/ of the nŠ input permutations yield an O.n lg n/ running time. Problems 12-1 Binary search trees with equal keys Equal keys pose a problem for the implementation of binary search trees. a. What is the asymptotic performance of TREE-INSERT when used to insert n items with identical keys into an initially empty binary search tree? We propose to improve TREE-INSERT by testing before line 5 to determine whether ´:key D x:key and by testing before line 11 to determine whether ´:key D y:key. 304 Chapter 12 Binary Search Trees If equality holds, we implement one of the following strategies. For each strategy, find the asymptotic performance of inserting n items with identical keys into an initially empty binary search tree. (The strategies are described for line 5, in which we compare the keys of ´ and x. Substitute y for x to arrive at the strategies for line 11.) b. Keep a boolean flag x:b at node x, and set x to either x:left or x:right based on the value of x:b, which alternates between FALSE and TRUE each time we visit x while inserting a node with the same key as x. c. Keep a list of nodes with equal keys at x, and insert ´ into the list. d. Randomly set x to either x:left or x:right. (Give the worst-case performance and informally derive the expected running time.) 12-2 Radix trees Given two strings a D a0a1 : : : ap and b D b0b1 : : : bq, where each ai and each bj is in some ordered set of characters, we say that string a is lexicographically less than string b if either 1. there exists an integer j , where 0 Ä j Ä min.p; q/, such that ai D bi for all i D 0; 1; : : : ; j 1 and aj < bj , or 2. p < q and ai D bi for all i D 0; 1; : : : ; p. For example, if a and b are bit strings, then 10100 < 10110 by rule 1 (letting j D 3) and 10100 < 101000 by rule 2. This ordering is similar to that used in English-language dictionaries. The radix tree data structure shown in Figure 12.5 stores the bit strings 1011, 10, 011, 100, and 0. When searching for a key a D a0a1 : : : ap, we go left at a node of depth i if ai D 0 and right if ai D 1. Let S be a set of distinct bit strings whose lengths sum to n. Show how to use a radix tree to sort S lexicographically in ‚.n/ time. For the example in Figure 12.5, the output of the sort should be the sequence 0, 011, 10, 100, 1011. 12-3 Average node depth in a randomly built binary search tree In this problem, we prove that the average depth of a node in a randomly built binary search tree with n nodes is O.lg n/. Although this result is weaker than that of Theorem 12.4, the technique we shall use reveals a surprising similarity between the building of a binary search tree and the execution of RANDOMIZEDQUICKSORT from Section 7.3. We define the total path length P .T / of a binary tree T as the sum, over all nodes x in T , of the depth of node x, which we denote by d.x; T /. Problems for Chapter 12 305 0 0 1 1 011 1 0 10 0 1 100 1 1011 Figure 12.5 A radix tree storing the bit strings 1011, 10, 011, 100, and 0. We can determine each node’s key by traversing the simple path from the root to that node. There is no need, therefore, to store the keys in the nodes; the keys appear here for illustrative purposes only. Nodes are heavily shaded if the keys corresponding to them are not in the tree; such nodes are present only to establish a path to other nodes. a. Argue that the average depth of a node in T is 1X 1 d.x; T / D P .T / : n n x2T Thus, we wish to show that the expected value of P .T / is O.n lg n/. b. Let TL and TR denote the left and right subtrees of tree T , respectively. Argue that if T has n nodes, then P .T / D P .TL/ C P .TR/ C n 1 : c. Let P .n/ denote the average total path length of a randomly built binary search tree with n nodes. Show that 1 X n 1 P .n/ D .P .i/ C P .n i 1/ C n 1/ : n i D0 d. Show how to rewrite P .n/ as 2 X n 1 P .n/ D P .k/ C ‚.n/ : n kD1 e. Recalling the alternative analysis of the randomized version of quicksort given in Problem 7-3, conclude that P .n/ D O.n lg n/. 306 Chapter 12 Binary Search Trees At each recursive invocation of quicksort, we choose a random pivot element to partition the set of elements being sorted. Each node of a binary search tree partitions the set of elements that fall into the subtree rooted at that node. f. Describe an implementation of quicksort in which the comparisons to sort a set of elements are exactly the same as the comparisons to insert the elements into a binary search tree. (The order in which comparisons are made may differ, but the same comparisons must occur.) 12-4 Number of different binary trees Let bn denote the number of different binary trees with n nodes. In this problem, you will find a formula for bn, as well as an asymptotic estimate. a. Show that b0 D 1 and that, for n 1, X n 1 bn D bkbn 1 k : kD0 b. Referring to Problem 4-4 for the definition of a generating function, let B.x/ be the generating function X 1 B.x/ D bnxn : nD0 Show that B.x/ D xB.x/2 C 1, and hence one way to express B.x/ in closed form is 1 p B.x/ D 1 1 4x : 2x The Taylor expansion of f .x/ around the point x D a is given by X 1 f .k/.a/ f .x/ D .x kŠ kD0 a/k ; where f .k/.x/ is the kth derivative of f evaluated at x. c. Show that ! 1 2n bn D n C 1 n Notes for Chapter 12 307 p (the nth Catalan number) by using the Taylor expansion of 1 4x around x D 0. (If you wish, instead of using the Taylor expansion, you may use the generalization of the binomial expansion (C.4) to nonintegral exponents n, where for any real number n and for any integer k, we interpret n k to be n.n 1/ .n k C 1/=kŠ if k 0, and 0 otherwise.) d. Show that 4n bn D p .1 C O.1=n// : n3=2 Chapter notes Knuth [211] contains a good discussion of simple binary search trees as well as many variations. Binary search trees seem to have been independently discovered by a number of people in the late 1950s. Radix trees are often called “tries,” which comes from the middle letters in the word retrieval. Knuth [211] also discusses them. Many texts, including the first two editions of this book, have a somewhat simpler method of deleting a node from a binary search tree when both of its children are present. Instead of replacing node ´ by its successor y, we delete node y but copy its key and satellite data into node ´. The downside of this approach is that the node actually deleted might not be the node passed to the delete procedure. If other components of a program maintain pointers to nodes in the tree, they could mistakenly end up with “stale” pointers to nodes that have been deleted. Although the deletion method presented in this edition of this book is a bit more complicated, it guarantees that a call to delete node ´ deletes node ´ and only node ´. Section 15.5 will show how to construct an optimal binary search tree when we know the search frequencies before constructing the tree. That is, given the frequencies of searching for each key and the frequencies of searching for values that fall between keys in the tree, we construct a binary search tree for which a set of searches that follows these frequencies examines the minimum number of nodes. The proof in Section 12.4 that bounds the expected height of a randomly built binary search tree is due to Aslam [24]. Mart´ınez and Roura [243] give randomized algorithms for insertion into and deletion from binary search trees in which the result of either operation is a random binary search tree. Their definition of a random binary search tree differs—only slightly—from that of a randomly built binary search tree in this chapter, however. 13 Red-Black Trees Chapter 12 showed that a binary search tree of height h can support any of the basic dynamic-set operations—such as SEARCH, PREDECESSOR, SUCCESSOR, MINIMUM, MAXIMUM, INSERT, and DELETE—in O.h/ time. Thus, the set operations are fast if the height of the search tree is small. If its height is large, however, the set operations may run no faster than with a linked list. Red-black trees are one of many search-tree schemes that are “balanced” in order to guarantee that basic dynamic-set operations take O.lg n/ time in the worst case. 13.1 Properties of red-black trees A red-black tree is a binary search tree with one extra bit of storage per node: its color, which can be either RED or BLACK. By constraining the node colors on any simple path from the root to a leaf, red-black trees ensure that no such path is more than twice as long as any other, so that the tree is approximately balanced. Each node of the tree now contains the attributes color, key, left, right, and p. If a child or the parent of a node does not exist, the corresponding pointer attribute of the node contains the value NIL. We shall regard these NILs as being pointers to leaves (external nodes) of the binary search tree and the normal, key-bearing nodes as being internal nodes of the tree. A red-black tree is a binary tree that satisfies the following red-black properties: 1. Every node is either red or black. 2. The root is black. 3. Every leaf (NIL) is black. 4. If a node is red, then both its children are black. 5. For each node, all simple paths from the node to descendant leaves contain the same number of black nodes. 13.1 Properties of red-black trees 309 Figure 13.1(a) shows an example of a red-black tree. As a matter of convenience in dealing with boundary conditions in red-black tree code, we use a single sentinel to represent NIL (see page 238). For a red-black tree T , the sentinel T:nil is an object with the same attributes as an ordinary node in the tree. Its color attribute is BLACK, and its other attributes—p, left, right, and key—can take on arbitrary values. As Figure 13.1(b) shows, all pointers to NIL are replaced by pointers to the sentinel T:nil. We use the sentinel so that we can treat a NIL child of a node x as an ordinary node whose parent is x. Although we instead could add a distinct sentinel node for each NIL in the tree, so that the parent of each NIL is well defined, that approach would waste space. Instead, we use the one sentinel T:nil to represent all the NILs—all leaves and the root’s parent. The values of the attributes p, left, right, and key of the sentinel are immaterial, although we may set them during the course of a procedure for our convenience. We generally confine our interest to the internal nodes of a red-black tree, since they hold the key values. In the remainder of this chapter, we omit the leaves when we draw red-black trees, as shown in Figure 13.1(c). We call the number of black nodes on any simple path from, but not including, a node x down to a leaf the black-height of the node, denoted bh.x/. By property 5, the notion of black-height is well defined, since all descending simple paths from the node have the same number of black nodes. We define the black-height of a red-black tree to be the black-height of its root. The following lemma shows why red-black trees make good search trees. Lemma 13.1 A red-black tree with n internal nodes has height at most 2 lg.n C 1/. Proof We start by showing that the subtree rooted at any node x contains at least 2bh.x/ 1 internal nodes. We prove this claim by induction on the height of x. If the height of x is 0, then x must be a leaf (T:nil), and the subtree rooted at x indeed contains at least 2bh.x/ 1 D 20 1 D 0 internal nodes. For the inductive step, consider a node x that has positive height and is an internal node with two children. Each child has a black-height of either bh.x/ or bh.x/ 1, depending on whether its color is red or black, respectively. Since the height of a child of x is less than the height of x itself, we can apply the inductive hypothesis to conclude that each child has at least 2bh.x/ 1 1 internal nodes. Thus, the subtree rooted at x contains at least .2bh.x/ 1 1/ C .2bh.x/ 1 1/ C 1 D 2bh.x/ 1 internal nodes, which proves the claim. To complete the proof of the lemma, let h be the height of the tree. According to property 4, at least half the nodes on any simple path from the root to a leaf, not 310 Chapter 13 Red-Black Trees 2 14 3 17 2 21 2 10 1 7 1 12 1 16 1 19 1 15 NIL NIL 1 20 1 3 NIL NIL NIL NIL NIL NIL NIL NIL NIL 3 26 41 2 1 23 1 28 2 30 1 38 NIL NIL NIL NIL 1 35 1 39 NIL NIL NIL NIL (a) 1 47 NIL NIL 26 17 41 14 21 30 47 10 16 19 23 28 38 7 12 15 20 35 39 3 T: nil (b) 26 17 41 14 21 30 47 10 16 19 23 28 38 7 12 15 20 35 39 3 (c) Figure 13.1 A red-black tree with black nodes darkened and red nodes shaded. Every node in a red-black tree is either red or black, the children of a red node are both black, and every simple path from a node to a descendant leaf contains the same number of black nodes. (a) Every leaf, shown as a NIL, is black. Each non-NIL node is marked with its black-height; NILs have black-height 0. (b) The same red-black tree but with each NIL replaced by the single sentinel T: nil, which is always black, and with black-heights omitted. The root’s parent is also the sentinel. (c) The same red-black tree but with leaves and the root’s parent omitted entirely. We shall use this drawing style in the remainder of this chapter. 13.1 Properties of red-black trees 311 including the root, must be black. Consequently, the black-height of the root must be at least h=2; thus, n 2h=2 1 : Moving the 1 to the left-hand side and taking logarithms on both sides yields lg.n C 1/ h=2, or h Ä 2 lg.n C 1/. As an immediate consequence of this lemma, we can implement the dynamic-set operations SEARCH, MINIMUM, MAXIMUM, SUCCESSOR, and PREDECESSOR in O.lg n/ time on red-black trees, since each can run in O.h/ time on a binary search tree of height h (as shown in Chapter 12) and any red-black tree on n nodes is a binary search tree with height O.lg n/. (Of course, references to NIL in the algorithms of Chapter 12 would have to be replaced by T:nil.) Although the algorithms TREE-INSERT and TREE-DELETE from Chapter 12 run in O.lg n/ time when given a red-black tree as input, they do not directly support the dynamic-set operations INSERT and DELETE, since they do not guarantee that the modified binary search tree will be a red-black tree. We shall see in Sections 13.3 and 13.4, however, how to support these two operations in O.lg n/ time. Exercises 13.1-1 In the style of Figure 13.1(a), draw the complete binary search tree of height 3 on the keys f1; 2; : : : ; 15g. Add the NIL leaves and color the nodes in three different ways such that the black-heights of the resulting red-black trees are 2, 3, and 4. 13.1-2 Draw the red-black tree that results after TREE-INSERT is called on the tree in Figure 13.1 with key 36. If the inserted node is colored red, is the resulting tree a red-black tree? What if it is colored black? 13.1-3 Let us define a relaxed red-black tree as a binary search tree that satisfies redblack properties 1, 3, 4, and 5. In other words, the root may be either red or black. Consider a relaxed red-black tree T whose root is red. If we color the root of T black but make no other changes to T , is the resulting tree a red-black tree? 13.1-4 Suppose that we “absorb” every red node in a red-black tree into its black parent, so that the children of the red node become children of the black parent. (Ignore what happens to the keys.) What are the possible degrees of a black node after all 312 Chapter 13 Red-Black Trees its red children are absorbed? What can you say about the depths of the leaves of the resulting tree? 13.1-5 Show that the longest simple path from a node x in a red-black tree to a descendant leaf has length at most twice that of the shortest simple path from node x to a descendant leaf. 13.1-6 What is the largest possible number of internal nodes in a red-black tree with blackheight k? What is the smallest possible number? 13.1-7 Describe a red-black tree on n keys that realizes the largest possible ratio of red internal nodes to black internal nodes. What is this ratio? What tree has the smallest possible ratio, and what is the ratio? 13.2 Rotations The search-tree operations TREE-INSERT and TREE-DELETE, when run on a redblack tree with n keys, take O.lg n/ time. Because they modify the tree, the result may violate the red-black properties enumerated in Section 13.1. To restore these properties, we must change the colors of some of the nodes in the tree and also change the pointer structure. We change the pointer structure through rotation, which is a local operation in a search tree that preserves the binary-search-tree property. Figure 13.2 shows the two kinds of rotations: left rotations and right rotations. When we do a left rotation on a node x, we assume that its right child y is not T:nil; x may be any node in the tree whose right child is not T:nil. The left rotation “pivots” around the link from x to y. It makes y the new root of the subtree, with x as y’s left child and y’s left child as x’s right child. The pseudocode for LEFT-ROTATE assumes that x:right ¤ T:nil and that the root’s parent is T:nil. 13.2 Rotations 313 y x γ α β LEFT-ROTATE(T, x) RIGHT-ROTATE(T, y) x α y β γ Figure 13.2 The rotation operations on a binary search tree. The operation LEFT-ROTATE.T; x/ transforms the configuration of the two nodes on the right into the configuration on the left by changing a constant number of pointers. The inverse operation RIGHT-ROTATE.T; y/ transforms the configuration on the left into the configuration on the right. The letters ˛, ˇ, and represent arbitrary subtrees. A rotation operation preserves the binary-search-tree property: the keys in ˛ precede x: key, which precedes the keys in ˇ, which precede y: key, which precedes the keys in . LEFT-ROTATE.T; x/ 1 y D x:right 2 x:right D y:left 3 if y:left ¤ T:nil 4 y:left:p D x 5 y:p D x:p 6 if x:p == T:nil 7 T:root D y 8 elseif x == x:p:left 9 x:p:left D y 10 else x:p:right D y 11 y:left D x 12 x:p D y // set y // turn y’s left subtree into x’s right subtree // link x’s parent to y // put x on y’s left Figure 13.3 shows an example of how LEFT-ROTATE modifies a binary search tree. The code for RIGHT-ROTATE is symmetric. Both LEFT-ROTATE and RIGHTROTATE run in O.1/ time. Only pointers are changed by a rotation; all other attributes in a node remain the same. Exercises 13.2-1 Write pseudocode for RIGHT-ROTATE. 13.2-2 Argue that in every n-node binary search tree, there are exactly n 1 possible rotations. 314 Chapter 13 Red-Black Trees 3 2 3 2 7 4 6 LEFT-ROTATE(T, x) 11 x 9 14 12 18 y 19 17 22 20 7 4 18 y 6 x 11 19 9 14 22 12 17 20 Figure 13.3 An example of how the procedure LEFT-ROTATE.T; x/ modifies a binary search tree. Inorder tree walks of the input tree and the modified tree produce the same listing of key values. 13.2-3 Let a, b, and c be arbitrary nodes in subtrees ˛, ˇ, and , respectively, in the left tree of Figure 13.2. How do the depths of a, b, and c change when a left rotation is performed on node x in the figure? 13.2-4 Show that any arbitrary n-node binary search tree can be transformed into any other arbitrary n-node binary search tree using O.n/ rotations. (Hint: First show that at most n 1 right rotations suffice to transform the tree into a right-going chain.) 13.2-5 ? We say that a binary search tree T1 can be right-converted to binary search tree T2 if it is possible to obtain T2 from T1 via a series of calls to RIGHT-ROTATE. Give an example of two trees T1 and T2 such that T1 cannot be right-converted to T2. Then, show that if a tree T1 can be right-converted to T2, it can be right-converted using O.n2/ calls to RIGHT-ROTATE. 13.3 Insertion 315 13.3 Insertion We can insert a node into an n-node red-black tree in O.lg n/ time. To do so, we use a slightly modified version of the TREE-INSERT procedure (Section 12.3) to insert node ´ into the tree T as if it were an ordinary binary search tree, and then we color ´ red. (Exercise 13.3-1 asks you to explain why we choose to make node ´ red rather than black.) To guarantee that the red-black properties are preserved, we then call an auxiliary procedure RB-INSERT-FIXUP to recolor nodes and perform rotations. The call RB-INSERT.T; ´/ inserts node ´, whose key is assumed to have already been filled in, into the red-black tree T . RB-INSERT.T; ´/ 1 y D T:nil 2 x D T:root 3 while x ¤ T:nil 4 yDx 5 if ´:key < x:key 6 x D x:left 7 else x D x:right 8 ´:p D y 9 if y == T:nil 10 T:root D ´ 11 elseif ´:key < y:key 12 y:left D ´ 13 else y:right D ´ 14 ´:left D T:nil 15 ´:right D T:nil 16 ´:color D RED 17 RB-INSERT-FIXUP.T; ´/ The procedures TREE-INSERT and RB-INSERT differ in four ways. First, all instances of NIL in TREE-INSERT are replaced by T:nil. Second, we set ´:left and ´:right to T:nil in lines 14–15 of RB-INSERT, in order to maintain the proper tree structure. Third, we color ´ red in line 16. Fourth, because coloring ´ red may cause a violation of one of the red-black properties, we call RB-INSERT-FIXUP.T; ´/ in line 17 of RB-INSERT to restore the red-black properties. 316 Chapter 13 Red-Black Trees RB-INSERT-FIXUP .T; ´/ 1 while ´:p:color == RED 2 if ´:p == ´:p:p:left 3 y D ´:p:p:right 4 if y:color == RED 5 ´:p:color D BLACK 6 y:color D BLACK 7 ´:p:p:color D RED 8 ´ D ´:p:p 9 else if ´ == ´:p:right 10 ´ D ´:p 11 LEFT-ROTATE .T; ´/ 12 ´:p:color D BLACK 13 ´:p:p:color D RED 14 RIGHT-ROTATE.T; ´: p: p/ 15 else (same as then clause with “right” and “left” exchanged) 16 T:root:color D BLACK // case 1 // case 1 // case 1 // case 1 // case 2 // case 2 // case 3 // case 3 // case 3 To understand how RB-INSERT-FIXUP works, we shall break our examination of the code into three major steps. First, we shall determine what violations of the red-black properties are introduced in RB-INSERT when node ´ is inserted and colored red. Second, we shall examine the overall goal of the while loop in lines 1–15. Finally, we shall explore each of the three cases1 within the while loop’s body and see how they accomplish the goal. Figure 13.4 shows how RBINSERT-FIXUP operates on a sample red-black tree. Which of the red-black properties might be violated upon the call to RBINSERT-FIXUP? Property 1 certainly continues to hold, as does property 3, since both children of the newly inserted red node are the sentinel T:nil. Property 5, which says that the number of black nodes is the same on every simple path from a given node, is satisfied as well, because node ´ replaces the (black) sentinel, and node ´ is red with sentinel children. Thus, the only properties that might be violated are property 2, which requires the root to be black, and property 4, which says that a red node cannot have a red child. Both possible violations are due to ´ being colored red. Property 2 is violated if ´ is the root, and property 4 is violated if ´’s parent is red. Figure 13.4(a) shows a violation of property 4 after the node ´ has been inserted. 1Case 2 falls through into case 3, and so these two cases are not mutually exclusive. 13.3 Insertion 317 11 2 14 (a) 1 7 15 5 8y z4 Case 1 11 2 14 y (b) 1 7z 15 5 8 4 Case 2 7 (c) z2 1 5 4 11 8 Case 3 14 y 15 7 z2 11 (d) 1 5 8 14 4 15 Figure 13.4 The operation of RB-INSERT-FIXUP. (a) A node ´ after insertion. Because both ´ and its parent ´: p are red, a violation of property 4 occurs. Since ´’s uncle y is red, case 1 in the code applies. We recolor nodes and move the pointer ´ up the tree, resulting in the tree shown in (b). Once again, ´ and its parent are both red, but ´’s uncle y is black. Since ´ is the right child of ´: p, case 2 applies. We perform a left rotation, and the tree that results is shown in (c). Now, ´ is the left child of its parent, and case 3 applies. Recoloring and right rotation yield the tree in (d), which is a legal red-black tree. 318 Chapter 13 Red-Black Trees The while loop in lines 1–15 maintains the following three-part invariant at the start of each iteration of the loop: a. Node ´ is red. b. If ´:p is the root, then ´:p is black. c. If the tree violates any of the red-black properties, then it violates at most one of them, and the violation is of either property 2 or property 4. If the tree violates property 2, it is because ´ is the root and is red. If the tree violates property 4, it is because both ´ and ´:p are red. Part (c), which deals with violations of red-black properties, is more central to showing that RB-INSERT-FIXUP restores the red-black properties than parts (a) and (b), which we use along the way to understand situations in the code. Because we’ll be focusing on node ´ and nodes near it in the tree, it helps to know from part (a) that ´ is red. We shall use part (b) to show that the node ´:p:p exists when we reference it in lines 2, 3, 7, 8, 13, and 14. Recall that we need to show that a loop invariant is true prior to the first iteration of the loop, that each iteration maintains the loop invariant, and that the loop invariant gives us a useful property at loop termination. We start with the initialization and termination arguments. Then, as we examine how the body of the loop works in more detail, we shall argue that the loop maintains the invariant upon each iteration. Along the way, we shall also demonstrate that each iteration of the loop has two possible outcomes: either the pointer ´ moves up the tree, or we perform some rotations and then the loop terminates. Initialization: Prior to the first iteration of the loop, we started with a red-black tree with no violations, and we added a red node ´. We show that each part of the invariant holds at the time RB-INSERT-FIXUP is called: a. When RB-INSERT-FIXUP is called, ´ is the red node that was added. b. If ´:p is the root, then ´:p started out black and did not change prior to the call of RB-INSERT-FIXUP. c. We have already seen that properties 1, 3, and 5 hold when RB-INSERTFIXUP is called. If the tree violates property 2, then the red root must be the newly added node ´, which is the only internal node in the tree. Because the parent and both children of ´ are the sentinel, which is black, the tree does not also violate property 4. Thus, this violation of property 2 is the only violation of red-black properties in the entire tree. If the tree violates property 4, then, because the children of node ´ are black sentinels and the tree had no other violations prior to ´ being added, the 13.3 Insertion 319 violation must be because both ´ and ´:p are red. Moreover, the tree violates no other red-black properties. Termination: When the loop terminates, it does so because ´:p is black. (If ´ is the root, then ´:p is the sentinel T:nil, which is black.) Thus, the tree does not violate property 4 at loop termination. By the loop invariant, the only property that might fail to hold is property 2. Line 16 restores this property, too, so that when RB-INSERT-FIXUP terminates, all the red-black properties hold. Maintenance: We actually need to consider six cases in the while loop, but three of them are symmetric to the other three, depending on whether line 2 determines ´’s parent ´:p to be a left child or a right child of ´’s grandparent ´:p:p. We have given the code only for the situation in which ´:p is a left child. The node ´:p:p exists, since by part (b) of the loop invariant, if ´:p is the root, then ´:p is black. Since we enter a loop iteration only if ´:p is red, we know that ´:p cannot be the root. Hence, ´:p:p exists. We distinguish case 1 from cases 2 and 3 by the color of ´’s parent’s sibling, or “uncle.” Line 3 makes y point to ´’s uncle ´:p:p:right, and line 4 tests y’s color. If y is red, then we execute case 1. Otherwise, control passes to cases 2 and 3. In all three cases, ´’s grandparent ´:p:p is black, since its parent ´:p is red, and property 4 is violated only between ´ and ´:p. Case 1: ´’s uncle y is red Figure 13.5 shows the situation for case 1 (lines 5–8), which occurs when both ´:p and y are red. Because ´:p:p is black, we can color both ´:p and y black, thereby fixing the problem of ´ and ´:p both being red, and we can color ´:p:p red, thereby maintaining property 5. We then repeat the while loop with ´:p:p as the new node ´. The pointer ´ moves up two levels in the tree. Now, we show that case 1 maintains the loop invariant at the start of the next iteration. We use ´ to denote node ´ in the current iteration, and ´0 D ´:p:p to denote the node that will be called node ´ at the test in line 1 upon the next iteration. a. Because this iteration colors ´:p:p red, node ´0 is red at the start of the next iteration. b. The node ´0:p is ´:p:p:p in this iteration, and the color of this node does not change. If this node is the root, it was black prior to this iteration, and it remains black at the start of the next iteration. c. We have already argued that case 1 maintains property 5, and it does not introduce a violation of properties 1 or 3. 320 Chapter 13 Red-Black Trees C (a) A α Bz β γ Dy δ ε new z C A α B β γ D δ ε C (b) B Dy zA γ δ ε α β new z C B A γ α β D δ ε Figure 13.5 Case 1 of the procedure RB-INSERT-FIXUP. Property 4 is violated, since ´ and its parent ´: p are both red. We take the same action whether (a) ´ is a right child or (b) ´ is a left child. Each of the subtrees ˛, ˇ, , ı, and " has a black root, and each has the same black-height. The code for case 1 changes the colors of some nodes, preserving property 5: all downward simple paths from a node to a leaf have the same number of blacks. The while loop continues with node ´’s grandparent ´: p: p as the new ´. Any violation of property 4 can now occur only between the new ´, which is red, and its parent, if it is red as well. If node ´0 is the root at the start of the next iteration, then case 1 corrected the lone violation of property 4 in this iteration. Since ´0 is red and it is the root, property 2 becomes the only one that is violated, and this violation is due to ´0. If node ´0 is not the root at the start of the next iteration, then case 1 has not created a violation of property 2. Case 1 corrected the lone violation of property 4 that existed at the start of this iteration. It then made ´0 red and left ´0:p alone. If ´0:p was black, there is no violation of property 4. If ´0:p was red, coloring ´0 red created one violation of property 4 between ´0 and ´0:p. Case 2: ´’s uncle y is black and ´ is a right child Case 3: ´’s uncle y is black and ´ is a left child In cases 2 and 3, the color of ´’s uncle y is black. We distinguish the two cases according to whether ´ is a right or left child of ´:p. Lines 10–11 constitute case 2, which is shown in Figure 13.6 together with case 3. In case 2, node ´ is a right child of its parent. We immediately use a left rotation to transform the situation into case 3 (lines 12–14), in which node ´ is a left child. Because 13.3 Insertion 321 C A δy α Bz β γ Case 2 B zA γ α β Case 3 C δy B zA C α β γ δ Figure 13.6 Cases 2 and 3 of the procedure RB-INSERT-FIXUP. As in case 1, property 4 is violated in either case 2 or case 3 because ´ and its parent ´: p are both red. Each of the subtrees ˛, ˇ, , and ı has a black root (˛, ˇ, and from property 4, and ı because otherwise we would be in case 1), and each has the same black-height. We transform case 2 into case 3 by a left rotation, which preserves property 5: all downward simple paths from a node to a leaf have the same number of blacks. Case 3 causes some color changes and a right rotation, which also preserve property 5. The while loop then terminates, because property 4 is satisfied: there are no longer two red nodes in a row. both ´ and ´:p are red, the rotation affects neither the black-height of nodes nor property 5. Whether we enter case 3 directly or through case 2, ´’s uncle y is black, since otherwise we would have executed case 1. Additionally, the node ´:p:p exists, since we have argued that this node existed at the time that lines 2 and 3 were executed, and after moving ´ up one level in line 10 and then down one level in line 11, the identity of ´:p:p remains unchanged. In case 3, we execute some color changes and a right rotation, which preserve property 5, and then, since we no longer have two red nodes in a row, we are done. The while loop does not iterate another time, since ´:p is now black. We now show that cases 2 and 3 maintain the loop invariant. (As we have just argued, ´:p will be black upon the next test in line 1, and the loop body will not execute again.) a. Case 2 makes ´ point to ´:p, which is red. No further change to ´ or its color occurs in cases 2 and 3. b. Case 3 makes ´:p black, so that if ´:p is the root at the start of the next iteration, it is black. c. As in case 1, properties 1, 3, and 5 are maintained in cases 2 and 3. Since node ´ is not the root in cases 2 and 3, we know that there is no violation of property 2. Cases 2 and 3 do not introduce a violation of property 2, since the only node that is made red becomes a child of a black node by the rotation in case 3. Cases 2 and 3 correct the lone violation of property 4, and they do not introduce another violation. 322 Chapter 13 Red-Black Trees Having shown that each iteration of the loop maintains the invariant, we have shown that RB-INSERT-FIXUP correctly restores the red-black properties. Analysis What is the running time of RB-INSERT? Since the height of a red-black tree on n nodes is O.lg n/, lines 1–16 of RB-INSERT take O.lg n/ time. In RB-INSERTFIXUP, the while loop repeats only if case 1 occurs, and then the pointer ´ moves two levels up the tree. The total number of times the while loop can be executed is therefore O.lg n/. Thus, RB-INSERT takes a total of O.lg n/ time. Moreover, it never performs more than two rotations, since the while loop terminates if case 2 or case 3 is executed. Exercises 13.3-1 In line 16 of RB-INSERT, we set the color of the newly inserted node ´ to red. Observe that if we had chosen to set ´’s color to black, then property 4 of a redblack tree would not be violated. Why didn’t we choose to set ´’s color to black? 13.3-2 Show the red-black trees that result after successively inserting the keys 41; 38; 31; 12; 19; 8 into an initially empty red-black tree. 13.3-3 Suppose that the black-height of each of the subtrees ˛; ˇ; ; ı; " in Figures 13.5 and 13.6 is k. Label each node in each figure with its black-height to verify that the indicated transformation preserves property 5. 13.3-4 Professor Teach is concerned that RB-INSERT-FIXUP might set T:nil:color to RED, in which case the test in line 1 would not cause the loop to terminate when ´ is the root. Show that the professor’s concern is unfounded by arguing that RBINSERT-FIXUP never sets T:nil:color to RED. 13.3-5 Consider a red-black tree formed by inserting n nodes with RB-INSERT. Argue that if n > 1, the tree has at least one red node. 13.3-6 Suggest how to implement RB-INSERT efficiently if the representation for redblack trees includes no storage for parent pointers. 13.4 Deletion 323 13.4 Deletion Like the other basic operations on an n-node red-black tree, deletion of a node takes time O.lg n/. Deleting a node from a red-black tree is a bit more complicated than inserting a node. The procedure for deleting a node from a red-black tree is based on the TREEDELETE procedure (Section 12.3). First, we need to customize the TRANSPLANT subroutine that TREE-DELETE calls so that it applies to a red-black tree: RB-TRANSPLANT.T; u; / 1 if u:p == T:nil 2 T:root D 3 elseif u == u:p:left 4 u:p:left D 5 else u:p:right D 6 :p D u:p The procedure RB-TRANSPLANT differs from TRANSPLANT in two ways. First, line 1 references the sentinel T:nil instead of NIL. Second, the assignment to :p in line 6 occurs unconditionally: we can assign to :p even if points to the sentinel. In fact, we shall exploit the ability to assign to :p when D T:nil. The procedure RB-DELETE is like the TREE-DELETE procedure, but with additional lines of pseudocode. Some of the additional lines keep track of a node y that might cause violations of the red-black properties. When we want to delete node ´ and ´ has fewer than two children, then ´ is removed from the tree, and we want y to be ´. When ´ has two children, then y should be ´’s successor, and y moves into ´’s position in the tree. We also remember y’s color before it is removed from or moved within the tree, and we keep track of the node x that moves into y’s original position in the tree, because node x might also cause violations of the red-black properties. After deleting node ´, RB-DELETE calls an auxiliary procedure RB-DELETE-FIXUP, which changes colors and performs rotations to restore the red-black properties. 324 Chapter 13 Red-Black Trees RB-DELETE.T; ´/ 1 yD´ 2 y-original-color D y:color 3 if ´:left == T:nil 4 x D ´:right 5 RB-TRANSPLANT.T; ´; ´:right/ 6 elseif ´:right == T:nil 7 x D ´:left 8 RB-TRANSPLANT.T; ´; ´:left/ 9 else y D TREE-MINIMUM.´:right/ 10 y-original-color D y:color 11 x D y:right 12 if y:p == ´ 13 x:p D y 14 else RB-TRANSPLANT.T; y; y:right/ 15 y:right D ´:right 16 y:right:p D y 17 RB-TRANSPLANT.T; ´; y/ 18 y:left D ´:left 19 y:left:p D y 20 y:color D ´:color 21 if y-original-color == BLACK 22 RB-DELETE-FIXUP .T; x/ Although RB-DELETE contains almost twice as many lines of pseudocode as TREE-DELETE, the two procedures have the same basic structure. You can find each line of TREE-DELETE within RB-DELETE (with the changes of replacing NIL by T:nil and replacing calls to TRANSPLANT by calls to RB-TRANSPLANT), executed under the same conditions. Here are the other differences between the two procedures: We maintain node y as the node either removed from the tree or moved within the tree. Line 1 sets y to point to node ´ when ´ has fewer than two children and is therefore removed. When ´ has two children, line 9 sets y to point to ´’s successor, just as in TREE-DELETE, and y will move into ´’s position in the tree. Because node y’s color might change, the variable y-original-color stores y’s color before any changes occur. Lines 2 and 10 set this variable immediately after assignments to y. When ´ has two children, then y ¤ ´ and node y moves into node ´’s original position in the red-black tree; line 20 gives y the same color as ´. We need to save y’s original color in order to test it at the 13.4 Deletion 325 end of RB-DELETE; if it was black, then removing or moving y could cause violations of the red-black properties. As discussed, we keep track of the node x that moves into node y’s original position. The assignments in lines 4, 7, and 11 set x to point to either y’s only child or, if y has no children, the sentinel T:nil. (Recall from Section 12.3 that y has no left child.) Since node x moves into node y’s original position, the attribute x:p is always set to point to the original position in the tree of y’s parent, even if x is, in fact, the sentinel T:nil. Unless ´ is y’s original parent (which occurs only when ´ has two children and its successor y is ´’s right child), the assignment to x:p takes place in line 6 of RB-TRANSPLANT. (Observe that when RB-TRANSPLANT is called in lines 5, 8, or 14, the second parameter passed is the same as x.) When y’s original parent is ´, however, we do not want x:p to point to y’s original parent, since we are removing that node from the tree. Because node y will move up to take ´’s position in the tree, setting x:p to y in line 13 causes x:p to point to the original position of y’s parent, even if x D T:nil. Finally, if node y was black, we might have introduced one or more violations of the red-black properties, and so we call RB-DELETE-FIXUP in line 22 to restore the red-black properties. If y was red, the red-black properties still hold when y is removed or moved, for the following reasons: 1. No black-heights in the tree have changed. 2. No red nodes have been made adjacent. Because y takes ´’s place in the tree, along with ´’s color, we cannot have two adjacent red nodes at y’s new position in the tree. In addition, if y was not ´’s right child, then y’s original right child x replaces y in the tree. If y is red, then x must be black, and so replacing y by x cannot cause two red nodes to become adjacent. 3. Since y could not have been the root if it was red, the root remains black. If node y was black, three problems may arise, which the call of RB-DELETEFIXUP will remedy. First, if y had been the root and a red child of y becomes the new root, we have violated property 2. Second, if both x and x:p are red, then we have violated property 4. Third, moving y within the tree causes any simple path that previously contained y to have one fewer black node. Thus, property 5 is now violated by any ancestor of y in the tree. We can correct the violation of property 5 by saying that node x, now occupying y’s original position, has an “extra” black. That is, if we add 1 to the count of black nodes on any simple path that contains x, then under this interpretation, property 5 holds. When we remove or move the black node y, we “push” its blackness onto node x. The problem is that now node x is neither red nor black, thereby violating property 1. Instead, 326 Chapter 13 Red-Black Trees node x is either “doubly black” or “red-and-black,” and it contributes either 2 or 1, respectively, to the count of black nodes on simple paths containing x. The color attribute of x will still be either RED (if x is red-and-black) or BLACK (if x is doubly black). In other words, the extra black on a node is reflected in x’s pointing to the node rather than in the color attribute. We can now see the procedure RB-DELETE-FIXUP and examine how it restores the red-black properties to the search tree. RB-DELETE-FIXUP .T; x/ 1 while x ¤ T:root and x:color == BLACK 2 if x == x:p:left 3 w D x:p:right 4 if w:color = = RED 5 w:color D BLACK 6 x:p:color D RED 7 LEFT-ROTATE.T; x: p/ 8 w D x:p:right 9 if w:left:color == BLACK and w:right:color = = BLACK 10 w:color D RED 11 x D x:p 12 else if w:right:color == BLACK 13 w:left:color D BLACK 14 w:color D RED 15 RIGHT-ROTATE.T; w/ 16 w D x:p:right 17 w:color D x:p:color 18 x:p:color D BLACK 19 w:right:color D BLACK 20 LEFT-ROTATE.T; x: p/ 21 x D T:root 22 else (same as then clause with “right” and “left” exchanged) 23 x:color D BLACK // case 1 // case 1 // case 1 // case 1 // case 2 // case 2 // case 3 // case 3 // case 3 // case 3 // case 4 // case 4 // case 4 // case 4 // case 4 The procedure RB-DELETE-FIXUP restores properties 1, 2, and 4. Exercises 13.4-1 and 13.4-2 ask you to show that the procedure restores properties 2 and 4, and so in the remainder of this section, we shall focus on property 1. The goal of the while loop in lines 1–22 is to move the extra black up the tree until 1. x points to a red-and-black node, in which case we color x (singly) black in line 23; 2. x points to the root, in which case we simply “remove” the extra black; or 3. having performed suitable rotations and recolorings, we exit the loop. 13.4 Deletion 327 Within the while loop, x always points to a nonroot doubly black node. We determine in line 2 whether x is a left child or a right child of its parent x:p. (We have given the code for the situation in which x is a left child; the situation in which x is a right child—line 22—is symmetric.) We maintain a pointer w to the sibling of x. Since node x is doubly black, node w cannot be T:nil, because otherwise, the number of blacks on the simple path from x:p to the (singly black) leaf w would be smaller than the number on the simple path from x:p to x. The four cases2 in the code appear in Figure 13.7. Before examining each case in detail, let’s look more generally at how we can verify that the transformation in each of the cases preserves property 5. The key idea is that in each case, the transformation applied preserves the number of black nodes (including x’s extra black) from (and including) the root of the subtree shown to each of the subtrees ˛; ˇ; : : : ; . Thus, if property 5 holds prior to the transformation, it continues to hold afterward. For example, in Figure 13.7(a), which illustrates case 1, the number of black nodes from the root to either subtree ˛ or ˇ is 3, both before and after the transformation. (Again, remember that node x adds an extra black.) Similarly, the number of black nodes from the root to any of , ı, ", and is 2, both before and after the transformation. In Figure 13.7(b), the counting must involve the value c of the color attribute of the root of the subtree shown, which can be either RED or BLACK. If we define count.RED/ D 0 and count.BLACK/ D 1, then the number of black nodes from the root to ˛ is 2 C count.c/, both before and after the transformation. In this case, after the transformation, the new node x has color attribute c, but this node is really either red-and-black (if c D RED) or doubly black (if c D BLACK). You can verify the other cases similarly (see Exercise 13.4-5). Case 1: x’s sibling w is red Case 1 (lines 5–8 of RB-DELETE-FIXUP and Figure 13.7(a)) occurs when node w, the sibling of node x, is red. Since w must have black children, we can switch the colors of w and x:p and then perform a left-rotation on x:p without violating any of the red-black properties. The new sibling of x, which is one of w’s children prior to the rotation, is now black, and thus we have converted case 1 into case 2, 3, or 4. Cases 2, 3, and 4 occur when node w is black; they are distinguished by the colors of w’s children. 2As in RB-INSERT-FIXUP, the cases in RB-DELETE-FIXUP are not mutually exclusive. 328 Chapter 13 Red-Black Trees Case 2: x’s sibling w is black, and both of w’s children are black In case 2 (lines 10–11 of RB-DELETE-FIXUP and Figure 13.7(b)), both of w’s children are black. Since w is also black, we take one black off both x and w, leaving x with only one black and leaving w red. To compensate for removing one black from x and w, we would like to add an extra black to x:p, which was originally either red or black. We do so by repeating the while loop with x:p as the new node x. Observe that if we enter case 2 through case 1, the new node x is red-and-black, since the original x:p was red. Hence, the value c of the color attribute of the new node x is RED, and the loop terminates when it tests the loop condition. We then color the new node x (singly) black in line 23. Case 3: x’s sibling w is black, w’s left child is red, and w’s right child is black Case 3 (lines 13–16 and Figure 13.7(c)) occurs when w is black, its left child is red, and its right child is black. We can switch the colors of w and its left child w:left and then perform a right rotation on w without violating any of the red-black properties. The new sibling w of x is now a black node with a red right child, and thus we have transformed case 3 into case 4. Case 4: x’s sibling w is black, and w’s right child is red Case 4 (lines 17–21 and Figure 13.7(d)) occurs when node x’s sibling w is black and w’s right child is red. By making some color changes and performing a left rotation on x:p, we can remove the extra black on x, making it singly black, without violating any of the red-black properties. Setting x to be the root causes the while loop to terminate when it tests the loop condition. Analysis What is the running time of RB-DELETE? Since the height of a red-black tree of n nodes is O.lg n/, the total cost of the procedure without the call to RB-DELETEFIXUP takes O.lg n/ time. Within RB-DELETE-FIXUP, each of cases 1, 3, and 4 lead to termination after performing a constant number of color changes and at most three rotations. Case 2 is the only case in which the while loop can be repeated, and then the pointer x moves up the tree at most O.lg n/ times, performing no rotations. Thus, the procedure RB-DELETE-FIXUP takes O.lg n/ time and performs at most three rotations, and the overall time for RB-DELETE is therefore also O.lg n/. 13.4 Deletion 329 B (a) x A Dw α β C E γ δε ζ Case 1 D B E x A new w C ε ζ α βγ δ Bc (b) x A Dw α β C E γ δε ζ Case 2 new x B c A D α β C E γ δε ζ Bc (c) x A Dw α β C E γ δε ζ Bc (d) x A Dw α β C c′ E γ δε ζ Case 3 Case 4 Bc xA C new w α β γ D δ E ε ζ Dc B E A C c′ ε ζ α βγ δ new x D T:root Figure 13.7 The cases in the while loop of the procedure RB-DELETE-FIXUP. Darkened nodes have color attributes BLACK, heavily shaded nodes have color attributes RED, and lightly shaded nodes have color attributes represented by c and c0, which may be either RED or BLACK. The letters ˛; ˇ; : : : ; represent arbitrary subtrees. Each case transforms the configuration on the left into the configuration on the right by changing some colors and/or performing a rotation. Any node pointed to by x has an extra black and is either doubly black or red-and-black. Only case 2 causes the loop to repeat. (a) Case 1 is transformed to case 2, 3, or 4 by exchanging the colors of nodes B and D and performing a left rotation. (b) In case 2, the extra black represented by the pointer x moves up the tree by coloring node D red and setting x to point to node B. If we enter case 2 through case 1, the while loop terminates because the new node x is red-and-black, and therefore the value c of its color attribute is RED. (c) Case 3 is transformed to case 4 by exchanging the colors of nodes C and D and performing a right rotation. (d) Case 4 removes the extra black represented by x by changing some colors and performing a left rotation (without violating the red-black properties), and then the loop terminates. 330 Chapter 13 Red-Black Trees Exercises 13.4-1 Argue that after executing RB-DELETE-FIXUP, the root of the tree must be black. 13.4-2 Argue that if in RB-DELETE both x and x:p are red, then property 4 is restored by the call to RB-DELETE-FIXUP .T; x/. 13.4-3 In Exercise 13.3-2, you found the red-black tree that results from successively inserting the keys 41; 38; 31; 12; 19; 8 into an initially empty tree. Now show the red-black trees that result from the successive deletion of the keys in the order 8; 12; 19; 31; 38; 41. 13.4-4 In which lines of the code for RB-DELETE-FIXUP might we examine or modify the sentinel T:nil? 13.4-5 In each of the cases of Figure 13.7, give the count of black nodes from the root of the subtree shown to each of the subtrees ˛; ˇ; : : : ; , and verify that each count remains the same after the transformation. When a node has a color attribute c or c0, use the notation count.c/ or count.c0/ symbolically in your count. 13.4-6 Professors Skelton and Baron are concerned that at the start of case 1 of RBDELETE-FIXUP, the node x:p might not be black. If the professors are correct, then lines 5–6 are wrong. Show that x:p must be black at the start of case 1, so that the professors have nothing to worry about. 13.4-7 Suppose that a node x is inserted into a red-black tree with RB-INSERT and then is immediately deleted with RB-DELETE. Is the resulting red-black tree the same as the initial red-black tree? Justify your answer. Problems for Chapter 13 331 Problems 13-1 Persistent dynamic sets During the course of an algorithm, we sometimes find that we need to maintain past versions of a dynamic set as it is updated. We call such a set persistent. One way to implement a persistent set is to copy the entire set whenever it is modified, but this approach can slow down a program and also consume much space. Sometimes, we can do much better. Consider a persistent set S with the operations INSERT, DELETE, and SEARCH, which we implement using binary search trees as shown in Figure 13.8(a). We maintain a separate root for every version of the set. In order to insert the key 5 into the set, we create a new node with key 5. This node becomes the left child of a new node with key 7, since we cannot modify the existing node with key 7. Similarly, the new node with key 7 becomes the left child of a new node with key 8 whose right child is the existing node with key 10. The new node with key 8 becomes, in turn, the right child of a new root r0 with key 4 whose left child is the existing node with key 3. We thus copy only part of the tree and share some of the nodes with the original tree, as shown in Figure 13.8(b). Assume that each tree node has the attributes key, left, and right but no parent. (See also Exercise 13.3-6.) 4r r 4 4 r′ 3 8 3 88 2 7 10 2 7 7 10 5 (a) (b) Figure 13.8 (a) A binary search tree with keys 2; 3; 4; 7; 8; 10. (b) The persistent binary search tree that results from the insertion of key 5. The most recent version of the set consists of the nodes reachable from the root r0, and the previous version consists of the nodes reachable from r. Heavily shaded nodes are added when key 5 is inserted. 332 Chapter 13 Red-Black Trees a. For a general persistent binary search tree, identify the nodes that we need to change to insert a key k or delete a node y. b. Write a procedure PERSISTENT-TREE-INSERT that, given a persistent tree T and a key k to insert, returns a new persistent tree T 0 that is the result of inserting k into T . c. If the height of the persistent binary search tree T is h, what are the time and space requirements of your implementation of PERSISTENT-TREE-INSERT? (The space requirement is proportional to the number of new nodes allocated.) d. Suppose that we had included the parent attribute in each node. In this case, PERSISTENT-TREE-INSERT would need to perform additional copying. Prove that PERSISTENT-TREE-INSERT would then require .n/ time and space, where n is the number of nodes in the tree. e. Show how to use red-black trees to guarantee that the worst-case running time and space are O.lg n/ per insertion or deletion. 13-2 Join operation on red-black trees The join operation takes two dynamic sets S1 and S2 and an element x such that for any x1 2 S1 and x2 2 S2, we have x1:key Ä x:key Ä x2:key. It returns a set S D S1 [ fxg [ S2. In this problem, we investigate how to implement the join operation on red-black trees. a. Given a red-black tree T , let us store its black-height as the new attribute T:bh. Argue that RB-INSERT and RB-DELETE can maintain the bh attribute without requiring extra storage in the nodes of the tree and without increasing the asymptotic running times. Show that while descending through T , we can determine the black-height of each node we visit in O.1/ time per node visited. We wish to implement the operation RB-JOIN.T1; x; T2/, which destroys T1 and T2 and returns a red-black tree T D T1 [ fxg [ T2. Let n be the total number of nodes in T1 and T2. b. Assume that T1:bh T2:bh. Describe an O.lg n/-time algorithm that finds a black node y in T1 with the largest key from among those nodes whose blackheight is T2:bh. c. Let Ty be the subtree rooted at y. Describe how Ty [ fxg [ T2 can replace Ty in O.1/ time without destroying the binary-search-tree property. d. What color should we make x so that red-black properties 1, 3, and 5 are maintained? Describe how to enforce properties 2 and 4 in O.lg n/ time. Problems for Chapter 13 333 e. Argue that no generality is lost by making the assumption in part (b). Describe the symmetric situation that arises when T1:bh Ä T2:bh. f. Argue that the running time of RB-JOIN is O.lg n/. 13-3 AVL trees An AVL tree is a binary search tree that is height balanced: for each node x, the heights of the left and right subtrees of x differ by at most 1. To implement an AVL tree, we maintain an extra attribute in each node: x:h is the height of node x. As for any other binary search tree T , we assume that T:root points to the root node. a. Prove that an AVL tree with n nodes has height O.lg n/. (Hint: Prove that an AVL tree of height h has at least Fh nodes, where Fh is the hth Fibonacci number.) b. To insert into an AVL tree, we first place a node into the appropriate place in binary search tree order. Afterward, the tree might no longer be height balanced. Specifically, the heights of the left and right children of some node might differ by 2. Describe a procedure BALANCE.x/, which takes a subtree rooted at x whose left and right children are height balanced and have heights that differ by at most 2, i.e., jx:right:h x:left:hj Ä 2, and alters the subtree rooted at x to be height balanced. (Hint: Use rotations.) c. Using part (b), describe a recursive procedure AVL-INSERT.x; ´/ that takes a node x within an AVL tree and a newly created node ´ (whose key has already been filled in), and adds ´ to the subtree rooted at x, maintaining the property that x is the root of an AVL tree. As in TREE-INSERT from Section 12.3, assume that ´:key has already been filled in and that ´:left D NIL and ´:right D NIL; also assume that ´:h D 0. Thus, to insert the node ´ into the AVL tree T , we call AVL-INSERT.T:root; ´/. d. Show that AVL-INSERT, run on an n-node AVL tree, takes O.lg n/ time and performs O.1/ rotations. 13-4 Treaps If we insert a set of n items into a binary search tree, the resulting tree may be horribly unbalanced, leading to long search times. As we saw in Section 12.4, however, randomly built binary search trees tend to be balanced. Therefore, one strategy that, on average, builds a balanced tree for a fixed set of items would be to randomly permute the items and then insert them in that order into the tree. What if we do not have all the items at once? If we receive the items one at a time, can we still randomly build a binary search tree out of them? 334 Chapter 13 Red-Black Trees G: 4 B: 7 A: 10 E: 23 H: 5 K: 65 I: 73 Figure 13.9 A treap. Each node x is labeled with x: key : x: priority. For example, the root has key G and priority 4. We will examine a data structure that answers this question in the affirmative. A treap is a binary search tree with a modified way of ordering the nodes. Figure 13.9 shows an example. As usual, each node x in the tree has a key value x:key. In addition, we assign x:priority, which is a random number chosen independently for each node. We assume that all priorities are distinct and also that all keys are distinct. The nodes of the treap are ordered so that the keys obey the binary-searchtree property and the priorities obey the min-heap order property: If is a left child of u, then :key < u:key. If is a right child of u, then :key > u:key. If is a child of u, then :priority > u:priority. (This combination of properties is why the tree is called a “treap”: it has features of both a binary search tree and a heap.) It helps to think of treaps in the following way. Suppose that we insert nodes x1; x2; : : : ; xn, with associated keys, into a treap. Then the resulting treap is the tree that would have been formed if the nodes had been inserted into a normal binary search tree in the order given by their (randomly chosen) priorities, i.e., xi :priority < xj :priority means that we had inserted xi before xj . a. Show that given a set of nodes x1; x2; : : : ; xn, with associated keys and priorities, all distinct, the treap associated with these nodes is unique. b. Show that the expected height of a treap is ‚.lg n/, and hence the expected time to search for a value in the treap is ‚.lg n/. Let us see how to insert a new node into an existing treap. The first thing we do is assign to the new node a random priority. Then we call the insertion algorithm, which we call TREAP-INSERT, whose operation is illustrated in Figure 13.10. Problems for Chapter 13 335 G: 4 B: 7 A: 10 E: 23 (a) G: 4 H: 5 C: 25 B: 7 K: 65 A: 10 E: 23 I: 73 C: 25 (b) H: 5 K: 65 I: 73 G: 4 D: 9 B: 7 A: 10 E: 23 C: 25 D: 9 (c) H: 5 K: 65 I: 73 G: 4 B: 7 A: 10 E: 23 D: 9 C: 25 (d) H: 5 K: 65 I: 73 G: 4 B: 7 H: 5 A: 10 D: 9 K: 65 C: 25 E: 23 I: 73 (e) F: 2 … F: 2 B: 7 A: 10 D: 9 G: 4 H: 5 C: 25 E: 23 K: 65 I: 73 (f) Figure 13.10 The operation of TREAP-INSERT. (a) The original treap, prior to insertion. (b) The treap after inserting a node with key C and priority 25. (c)–(d) Intermediate stages when inserting a node with key D and priority 9. (e) The treap after the insertion of parts (c) and (d) is done. (f) The treap after inserting a node with key F and priority 2. 336 Chapter 13 Red-Black Trees 15 9 3 12 6 (a) 18 25 21 15 9 3 12 6 (b) 18 25 21 Figure 13.11 Spines of a binary search tree. The left spine is shaded in (a), and the right spine is shaded in (b). c. Explain how TREAP-INSERT works. Explain the idea in English and give pseudocode. (Hint: Execute the usual binary-search-tree insertion procedure and then perform rotations to restore the min-heap order property.) d. Show that the expected running time of TREAP-INSERT is ‚.lg n/. TREAP-INSERT performs a search and then a sequence of rotations. Although these two operations have the same expected running time, they have different costs in practice. A search reads information from the treap without modifying it. In contrast, a rotation changes parent and child pointers within the treap. On most computers, read operations are much faster than write operations. Thus we would like TREAP-INSERT to perform few rotations. We will show that the expected number of rotations performed is bounded by a constant. In order to do so, we will need some definitions, which Figure 13.11 depicts. The left spine of a binary search tree T is the simple path from the root to the node with the smallest key. In other words, the left spine is the simple path from the root that consists of only left edges. Symmetrically, the right spine of T is the simple path from the root consisting of only right edges. The length of a spine is the number of nodes it contains. e. Consider the treap T immediately after TREAP-INSERT has inserted node x. Let C be the length of the right spine of the left subtree of x. Let D be the length of the left spine of the right subtree of x. Prove that the total number of rotations that were performed during the insertion of x is equal to C C D. We will now calculate the expected values of C and D. Without loss of generality, we assume that the keys are 1; 2; : : : ; n, since we are comparing them only to one another. Notes for Chapter 13 337 For nodes x and y in treap T , where y ¤ x, let k D x:key and i D y:key. We define indicator random variables Xik D I fy is in the right spine of the left subtree of xg : f. Show that Xik D 1 if and only if y:priority > x:priority, y:key < x:key, and, for every ´ such that y:key < ´:key < x:key, we have y:priority < ´:priority. g. Show that .k i 1/Š Pr fXik D 1g D .k i C 1/Š D .k 1 i C 1/.k : i/ h. Show that X k 1 1 E ŒC  D j.j C 1/ j D1 D1 1 : k i. Use a symmetry argument to show that E ŒD D 1 1 : n kC1 j. Conclude that the expected number of rotations performed when inserting a node into a treap is less than 2. Chapter notes The idea of balancing a search tree is due to Adel’son-Vel’ski˘ı and Landis [2], who introduced a class of balanced search trees called “AVL trees” in 1962, described in Problem 13-3. Another class of search trees, called “2-3 trees,” was introduced by J. E. Hopcroft (unpublished) in 1970. A 2-3 tree maintains balance by manipulating the degrees of nodes in the tree. Chapter 18 covers a generalization of 2-3 trees introduced by Bayer and McCreight [35], called “B-trees.” Red-black trees were invented by Bayer [34] under the name “symmetric binary B-trees.” Guibas and Sedgewick [155] studied their properties at length and introduced the red/black color convention. Andersson [15] gives a simpler-to-code 338 Chapter 13 Red-Black Trees variant of red-black trees. Weiss [351] calls this variant AA-trees. An AA-tree is similar to a red-black tree except that left children may never be red. Treaps, the subject of Problem 13-4, were proposed by Seidel and Aragon [309]. They are the default implementation of a dictionary in LEDA [253], which is a well-implemented collection of data structures and algorithms. There are many other variations on balanced binary trees, including weightbalanced trees [264], k-neighbor trees [245], and scapegoat trees [127]. Perhaps the most intriguing are the “splay trees” introduced by Sleator and Tarjan [320], which are “self-adjusting.” (See Tarjan [330] for a good description of splay trees.) Splay trees maintain balance without any explicit balance condition such as color. Instead, “splay operations” (which involve rotations) are performed within the tree every time an access is made. The amortized cost (see Chapter 17) of each operation on an n-node tree is O.lg n/. Skip lists [286] provide an alternative to balanced binary trees. A skip list is a linked list that is augmented with a number of additional pointers. Each dictionary operation runs in expected time O.lg n/ on a skip list of n items. 14 Augmenting Data Structures Some engineering situations require no more than a “textbook” data structure—such as a doubly linked list, a hash table, or a binary search tree—but many others require a dash of creativity. Only in rare situations will you need to create an entirely new type of data structure, though. More often, it will suffice to augment a textbook data structure by storing additional information in it. You can then program new operations for the data structure to support the desired application. Augmenting a data structure is not always straightforward, however, since the added information must be updated and maintained by the ordinary operations on the data structure. This chapter discusses two data structures that we construct by augmenting redblack trees. Section 14.1 describes a data structure that supports general orderstatistic operations on a dynamic set. We can then quickly find the ith smallest number in a set or the rank of a given element in the total ordering of the set. Section 14.2 abstracts the process of augmenting a data structure and provides a theorem that can simplify the process of augmenting red-black trees. Section 14.3 uses this theorem to help design a data structure for maintaining a dynamic set of intervals, such as time intervals. Given a query interval, we can then quickly find an interval in the set that overlaps it. 14.1 Dynamic order statistics Chapter 9 introduced the notion of an order statistic. Specifically, the ith order statistic of a set of n elements, where i 2 f1; 2; : : : ; ng, is simply the element in the set with the ith smallest key. We saw how to determine any order statistic in O.n/ time from an unordered set. In this section, we shall see how to modify red-black trees so that we can determine any order statistic for a dynamic set in O.lg n/ time. We shall also see how to compute the rank of an element—its position in the linear order of the set—in O.lg n/ time. 340 Chapter 14 Augmenting Data Structures 26 20 17 41 12 7 14 21 30 47 7 4 5 1 10 16 19 21 28 38 4 2 2 1 1 3 7 12 14 20 key 35 39 2 1 1 3 1 size 1 1 1 Figure 14.1 An order-statistic tree, which is an augmented red-black tree. Shaded nodes are red, and darkened nodes are black. In addition to its usual attributes, each node x has an attribute x: size, which is the number of nodes, other than the sentinel, in the subtree rooted at x. Figure 14.1 shows a data structure that can support fast order-statistic operations. An order-statistic tree T is simply a red-black tree with additional information stored in each node. Besides the usual red-black tree attributes x:key, x:color, x:p, x:left, and x:right in a node x, we have another attribute, x:size. This attribute contains the number of (internal) nodes in the subtree rooted at x (including x itself), that is, the size of the subtree. If we define the sentinel’s size to be 0—that is, we set T:nil:size to be 0—then we have the identity x:size D x:left:size C x:right:size C 1 : We do not require keys to be distinct in an order-statistic tree. (For example, the tree in Figure 14.1 has two keys with value 14 and two keys with value 21.) In the presence of equal keys, the above notion of rank is not well defined. We remove this ambiguity for an order-statistic tree by defining the rank of an element as the position at which it would be printed in an inorder walk of the tree. In Figure 14.1, for example, the key 14 stored in a black node has rank 5, and the key 14 stored in a red node has rank 6. Retrieving an element with a given rank Before we show how to maintain this size information during insertion and deletion, let us examine the implementation of two order-statistic queries that use this additional information. We begin with an operation that retrieves an element with a given rank. The procedure OS-SELECT.x; i / returns a pointer to the node containing the ith smallest key in the subtree rooted at x. To find the node with the ith smallest key in an order-statistic tree T , we call OS-SELECT.T:root; i /. 14.1 Dynamic order statistics 341 OS-SELECT.x; i / 1 r D x:left:size C 1 2 if i == r 3 return x 4 elseif i < r 5 return OS-SELECT.x:left; i / 6 else return OS-SELECT.x:right; i r/ In line 1 of OS-SELECT, we compute r, the rank of node x within the subtree rooted at x. The value of x:left:size is the number of nodes that come before x in an inorder tree walk of the subtree rooted at x. Thus, x:left:size C 1 is the rank of x within the subtree rooted at x. If i D r, then node x is the ith smallest element, and so we return x in line 3. If i < r, then the ith smallest element resides in x’s left subtree, and so we recurse on x:left in line 5. If i > r, then the ith smallest element resides in x’s right subtree. Since the subtree rooted at x contains r elements that come before x’s right subtree in an inorder tree walk, the ith smallest element in the subtree rooted at x is the .i r/th smallest element in the subtree rooted at x:right. Line 6 determines this element recursively. To see how OS-SELECT operates, consider a search for the 17th smallest element in the order-statistic tree of Figure 14.1. We begin with x as the root, whose key is 26, and with i D 17. Since the size of 26’s left subtree is 12, its rank is 13. Thus, we know that the node with rank 17 is the 17 13 D 4th smallest element in 26’s right subtree. After the recursive call, x is the node with key 41, and i D 4. Since the size of 41’s left subtree is 5, its rank within its subtree is 6. Thus, we know that the node with rank 4 is the 4th smallest element in 41’s left subtree. After the recursive call, x is the node with key 30, and its rank within its subtree is 2. Thus, we recurse once again to find the 4 2 D 2nd smallest element in the subtree rooted at the node with key 38. We now find that its left subtree has size 1, which means it is the second smallest element. Thus, the procedure returns a pointer to the node with key 38. Because each recursive call goes down one level in the order-statistic tree, the total time for OS-SELECT is at worst proportional to the height of the tree. Since the tree is a red-black tree, its height is O.lg n/, where n is the number of nodes. Thus, the running time of OS-SELECT is O.lg n/ for a dynamic set of n elements. Determining the rank of an element Given a pointer to a node x in an order-statistic tree T , the procedure OS-RANK returns the position of x in the linear order determined by an inorder tree walk of T . 342 Chapter 14 Augmenting Data Structures OS-RANK.T; x/ 1 r D x:left:size C 1 2 yDx 3 while y ¤ T:root 4 if y == y:p:right 5 r D r C y:p:left:size C 1 6 y D y:p 7 return r The procedure works as follows. We can think of node x’s rank as the number of nodes preceding x in an inorder tree walk, plus 1 for x itself. OS-RANK maintains the following loop invariant: At the start of each iteration of the while loop of lines 3–6, r is the rank of x:key in the subtree rooted at node y. We use this loop invariant to show that OS-RANK works correctly as follows: Initialization: Prior to the first iteration, line 1 sets r to be the rank of x:key within the subtree rooted at x. Setting y D x in line 2 makes the invariant true the first time the test in line 3 executes. Maintenance: At the end of each iteration of the while loop, we set y D y:p. Thus we must show that if r is the rank of x:key in the subtree rooted at y at the start of the loop body, then r is the rank of x:key in the subtree rooted at y:p at the end of the loop body. In each iteration of the while loop, we consider the subtree rooted at y:p. We have already counted the number of nodes in the subtree rooted at node y that precede x in an inorder walk, and so we must add the nodes in the subtree rooted at y’s sibling that precede x in an inorder walk, plus 1 for y:p if it, too, precedes x. If y is a left child, then neither y:p nor any node in y:p’s right subtree precedes x, and so we leave r alone. Otherwise, y is a right child and all the nodes in y:p’s left subtree precede x, as does y:p itself. Thus, in line 5, we add y:p:left:size C 1 to the current value of r. Termination: The loop terminates when y D T:root, so that the subtree rooted at y is the entire tree. Thus, the value of r is the rank of x:key in the entire tree. As an example, when we run OS-RANK on the order-statistic tree of Figure 14.1 to find the rank of the node with key 38, we get the following sequence of values of y:key and r at the top of the while loop: iteration y:key r 1 38 2 2 30 4 3 41 4 4 26 17 14.1 Dynamic order statistics 343 The procedure returns the rank 17. Since each iteration of the while loop takes O.1/ time, and y goes up one level in the tree with each iteration, the running time of OS-RANK is at worst proportional to the height of the tree: O.lg n/ on an n-node order-statistic tree. Maintaining subtree sizes Given the size attribute in each node, OS-SELECT and OS-RANK can quickly compute order-statistic information. But unless we can efficiently maintain these attributes within the basic modifying operations on red-black trees, our work will have been for naught. We shall now show how to maintain subtree sizes for both insertion and deletion without affecting the asymptotic running time of either operation. We noted in Section 13.3 that insertion into a red-black tree consists of two phases. The first phase goes down the tree from the root, inserting the new node as a child of an existing node. The second phase goes up the tree, changing colors and performing rotations to maintain the red-black properties. To maintain the subtree sizes in the first phase, we simply increment x:size for each node x on the simple path traversed from the root down toward the leaves. The new node added gets a size of 1. Since there are O.lg n/ nodes on the traversed path, the additional cost of maintaining the size attributes is O.lg n/. In the second phase, the only structural changes to the underlying red-black tree are caused by rotations, of which there are at most two. Moreover, a rotation is a local operation: only two nodes have their size attributes invalidated. The link around which the rotation is performed is incident on these two nodes. Referring to the code for LEFT-ROTATE.T; x/ in Section 13.2, we add the following lines: 13 y:size D x:size 14 x:size D x:left:size C x:right:size C 1 Figure 14.2 illustrates how the attributes are updated. The change to RIGHTROTATE is symmetric. Since at most two rotations are performed during insertion into a red-black tree, we spend only O.1/ additional time updating size attributes in the second phase. Thus, the total time for insertion into an n-node order-statistic tree is O.lg n/, which is asymptotically the same as for an ordinary red-black tree. Deletion from a red-black tree also consists of two phases: the first operates on the underlying search tree, and the second causes at most three rotations and otherwise performs no structural changes. (See Section 13.4.) The first phase either removes one node y from the tree or moves upward it within the tree. To update the subtree sizes, we simply traverse a simple path from node y (starting from its original position within the tree) up to the root, decrementing the size 344 Chapter 14 Augmenting Data Structures 93 19 y LEFT-ROTATE(T, x) 42 11 x 7 RIGHT-ROTATE(T, y) 6 6 4 42 19 x 93 12 y 4 7 Figure 14.2 Updating subtree sizes during rotations. The link around which we rotate is incident on the two nodes whose size attributes need to be updated. The updates are local, requiring only the size information stored in x, y, and the roots of the subtrees shown as triangles. attribute of each node on the path. Since this path has length O.lg n/ in an nnode red-black tree, the additional time spent maintaining size attributes in the first phase is O.lg n/. We handle the O.1/ rotations in the second phase of deletion in the same manner as for insertion. Thus, both insertion and deletion, including maintaining the size attributes, take O.lg n/ time for an n-node order-statistic tree. Exercises 14.1-1 Show how OS-SELECT.T:root; 10/ operates on the red-black tree T of Figure 14.1. 14.1-2 Show how OS-RANK.T; x/ operates on the red-black tree T of Figure 14.1 and the node x with x:key D 35. 14.1-3 Write a nonrecursive version of OS-SELECT. 14.1-4 Write a recursive procedure OS-KEY-RANK.T; k/ that takes as input an orderstatistic tree T and a key k and returns the rank of k in the dynamic set represented by T . Assume that the keys of T are distinct. 14.1-5 Given an element x in an n-node order-statistic tree and a natural number i, how can we determine the ith successor of x in the linear order of the tree in O.lg n/ time? 14.2 How to augment a data structure 345 14.1-6 Observe that whenever we reference the size attribute of a node in either OSSELECT or OS-RANK, we use it only to compute a rank. Accordingly, suppose we store in each node its rank in the subtree of which it is the root. Show how to maintain this information during insertion and deletion. (Remember that these two operations can cause rotations.) 14.1-7 Show how to use an order-statistic tree to count the number of inversions (see Problem 2-4) in an array of size n in time O.n lg n/. 14.1-8 ? Consider n chords on a circle, each defined by its endpoints. Describe an O.n lg n/- time algorithm to determine the number of pairs of chords that intersect inside the circle. (For example, if the n chords are all diameters that meet at the center, then the correct answer is n 2 .) Assume that no two chords share an endpoint. 14.2 How to augment a data structure The process of augmenting a basic data structure to support additional functionality occurs quite frequently in algorithm design. We shall use it again in the next section to design a data structure that supports operations on intervals. In this section, we examine the steps involved in such augmentation. We shall also prove a theorem that allows us to augment red-black trees easily in many cases. We can break the process of augmenting a data structure into four steps: 1. Choose an underlying data structure. 2. Determine additional information to maintain in the underlying data structure. 3. Verify that we can maintain the additional information for the basic modifying operations on the underlying data structure. 4. Develop new operations. As with any prescriptive design method, you should not blindly follow the steps in the order given. Most design work contains an element of trial and error, and progress on all steps usually proceeds in parallel. There is no point, for example, in determining additional information and developing new operations (steps 2 and 4) if we will not be able to maintain the additional information efficiently. Nevertheless, this four-step method provides a good focus for your efforts in augmenting a data structure, and it is also a good way to organize the documentation of an augmented data structure. 346 Chapter 14 Augmenting Data Structures We followed these steps in Section 14.1 to design our order-statistic trees. For step 1, we chose red-black trees as the underlying data structure. A clue to the suitability of red-black trees comes from their efficient support of other dynamicset operations on a total order, such as MINIMUM, MAXIMUM, SUCCESSOR, and PREDECESSOR. For step 2, we added the size attribute, in which each node x stores the size of the subtree rooted at x. Generally, the additional information makes operations more efficient. For example, we could have implemented OS-SELECT and OS-RANK using just the keys stored in the tree, but they would not have run in O.lg n/ time. Sometimes, the additional information is pointer information rather than data, as in Exercise 14.2-1. For step 3, we ensured that insertion and deletion could maintain the size attributes while still running in O.lg n/ time. Ideally, we should need to update only a few elements of the data structure in order to maintain the additional information. For example, if we simply stored in each node its rank in the tree, the OS-SELECT and OS-RANK procedures would run quickly, but inserting a new minimum element would cause a change to this information in every node of the tree. When we store subtree sizes instead, inserting a new element causes information to change in only O.lg n/ nodes. For step 4, we developed the operations OS-SELECT and OS-RANK. After all, the need for new operations is why we bother to augment a data structure in the first place. Occasionally, rather than developing new operations, we use the additional information to expedite existing ones, as in Exercise 14.2-1. Augmenting red-black trees When red-black trees underlie an augmented data structure, we can prove that insertion and deletion can always efficiently maintain certain kinds of additional information, thereby making step 3 very easy. The proof of the following theorem is similar to the argument from Section 14.1 that we can maintain the size attribute for order-statistic trees. Theorem 14.1 (Augmenting a red-black tree) Let f be an attribute that augments a red-black tree T of n nodes, and suppose that the value of f for each node x depends on only the information in nodes x, x:left, and x:right, possibly including x:left:f and x:right:f . Then, we can maintain the values of f in all nodes of T during insertion and deletion without asymptotically affecting the O.lg n/ performance of these operations. Proof The main idea of the proof is that a change to an f attribute in a node x propagates only to ancestors of x in the tree. That is, changing x:f may re- 14.2 How to augment a data structure 347 quire x:p:f to be updated, but nothing else; updating x:p:f may require x:p:p:f to be updated, but nothing else; and so on up the tree. Once we have updated T:root:f , no other node will depend on the new value, and so the process terminates. Since the height of a red-black tree is O.lg n/, changing an f attribute in a node costs O.lg n/ time in updating all nodes that depend on the change. Insertion of a node x into T consists of two phases. (See Section 13.3.) The first phase inserts x as a child of an existing node x:p. We can compute the value of x:f in O.1/ time since, by supposition, it depends only on information in the other attributes of x itself and the information in x’s children, but x’s children are both the sentinel T:nil. Once we have computed x:f , the change propagates up the tree. Thus, the total time for the first phase of insertion is O.lg n/. During the second phase, the only structural changes to the tree come from rotations. Since only two nodes change in a rotation, the total time for updating the f attributes is O.lg n/ per rotation. Since the number of rotations during insertion is at most two, the total time for insertion is O.lg n/. Like insertion, deletion has two phases. (See Section 13.4.) In the first phase, changes to the tree occur when the deleted node is removed from the tree. If the deleted node had two children at the time, then its successor moves into the position of the deleted node. Propagating the updates to f caused by these changes costs at most O.lg n/, since the changes modify the tree locally. Fixing up the red-black tree during the second phase requires at most three rotations, and each rotation requires at most O.lg n/ time to propagate the updates to f . Thus, like insertion, the total time for deletion is O.lg n/. In many cases, such as maintaining the size attributes in order-statistic trees, the cost of updating after a rotation is O.1/, rather than the O.lg n/ derived in the proof of Theorem 14.1. Exercise 14.2-3 gives an example. Exercises 14.2-1 Show, by adding pointers to the nodes, how to support each of the dynamic-set queries MINIMUM, MAXIMUM, SUCCESSOR, and PREDECESSOR in O.1/ worstcase time on an augmented order-statistic tree. The asymptotic performance of other operations on order-statistic trees should not be affected. 14.2-2 Can we maintain the black-heights of nodes in a red-black tree as attributes in the nodes of the tree without affecting the asymptotic performance of any of the redblack tree operations? Show how, or argue why not. How about maintaining the depths of nodes? 348 Chapter 14 Augmenting Data Structures 14.2-3 ? Let ˝ be an associative binary operator, and let a be an attribute maintained in each node of a red-black tree. Suppose that we want to include in each node x an additional attribute f such that x:f D x1:a ˝ x2:a ˝ ˝ xm:a, where x1; x2; : : : ; xm is the inorder listing of nodes in the subtree rooted at x. Show how to update the f attributes in O.1/ time after a rotation. Modify your argument slightly to apply it to the size attributes in order-statistic trees. 14.2-4 ? We wish to augment red-black trees with an operation RB-ENUMERATE.x; a; b/ that outputs all the keys k such that a Ä k Ä b in a red-black tree rooted at x. Describe how to implement RB-ENUMERATE in ‚.m C lg n/ time, where m is the number of keys that are output and n is the number of internal nodes in the tree. (Hint: You do not need to add new attributes to the red-black tree.) 14.3 Interval trees In this section, we shall augment red-black trees to support operations on dynamic sets of intervals. A closed interval is an ordered pair of real numbers Œt1; t2, with t1 Ä t2. The interval Œt1; t2 represents the set ft 2 R W t1 Ä t Ä t2g. Open and half-open intervals omit both or one of the endpoints from the set, respectively. In this section, we shall assume that intervals are closed; extending the results to open and half-open intervals is conceptually straightforward. Intervals are convenient for representing events that each occupy a continuous period of time. We might, for example, wish to query a database of time intervals to find out what events occurred during a given interval. The data structure in this section provides an efficient means for maintaining such an interval database. We can represent an interval Œt1; t2 as an object i , with attributes i:low D t1 (the low endpoint) and i:high D t2 (the high endpoint). We say that intervals i and i 0 overlap if i \ i 0 ¤ ;, that is, if i:low Ä i 0:high and i 0:low Ä i:high. As Figure 14.3 shows, any two intervals i and i0 satisfy the interval trichotomy; that is, exactly one of the following three properties holds: a. i and i 0 overlap, b. i is to the left of i 0 (i.e., i:high < i 0:low), c. i is to the right of i 0 (i.e., i 0:high < i:low). An interval tree is a red-black tree that maintains a dynamic set of elements, with each element x containing an interval x:int. Interval trees support the following operations: 14.3 Interval trees 349 i i i i i′ i′ i′ i′ (a) i i′ i′ i (b) (c) Figure 14.3 The interval trichotomy for two closed intervals i and i0. (a) If i and i0 overlap, there are four situations; in each, i: low Ä i0: high and i0: low Ä i: high. (b) The intervals do not overlap, and i: high < i0: low. (c) The intervals do not overlap, and i0: high < i: low. INTERVAL-INSERT.T; x/ adds the element x, whose int attribute is assumed to contain an interval, to the interval tree T . INTERVAL-DELETE.T; x/ removes the element x from the interval tree T . INTERVAL-SEARCH.T; i / returns a pointer to an element x in the interval tree T such that x:int overlaps interval i, or a pointer to the sentinel T:nil if no such element is in the set. Figure 14.4 shows how an interval tree represents a set of intervals. We shall track the four-step method from Section 14.2 as we review the design of an interval tree and the operations that run on it. Step 1: Underlying data structure We choose a red-black tree in which each node x contains an interval x:int and the key of x is the low endpoint, x:int:low, of the interval. Thus, an inorder tree walk of the data structure lists the intervals in sorted order by low endpoint. Step 2: Additional information In addition to the intervals themselves, each node x contains a value x:max, which is the maximum value of any interval endpoint stored in the subtree rooted at x. Step 3: Maintaining the information We must verify that insertion and deletion take O.lg n/ time on an interval tree of n nodes. We can determine x:max given interval x:int and the max values of node x’s children: 350 Chapter 14 Augmenting Data Structures (a) 0 0 6 5 3 5 89 10 8 10 26 26 25 30 19 20 17 19 16 21 15 23 15 20 25 30 [8,9] (b) 23 [5,8] 10 [0,3] 3 [6,10] 10 [16,21] 30 [25,30] int 30 max [15,23] 23 [17,19] 20 [26,26] 26 [19,20] 20 Figure 14.4 An interval tree. (a) A set of 10 intervals, shown sorted bottom to top by left endpoint. (b) The interval tree that represents them. Each node x contains an interval, shown above the dashed line, and the maximum value of any interval endpoint in the subtree rooted at x, shown below the dashed line. An inorder tree walk of the tree lists the nodes in sorted order by left endpoint. x:max D max.x:int:high; x:left:max; x:right:max/ : Thus, by Theorem 14.1, insertion and deletion run in O.lg n/ time. In fact, we can update the max attributes after a rotation in O.1/ time, as Exercises 14.2-3 and 14.3-1 show. Step 4: Developing new operations The only new operation we need is INTERVAL-SEARCH.T; i /, which finds a node in tree T whose interval overlaps interval i. If there is no interval that overlaps i in the tree, the procedure returns a pointer to the sentinel T:nil. 14.3 Interval trees 351 INTERVAL-SEARCH.T; i / 1 x D T:root 2 while x ¤ T:nil and i does not overlap x:int 3 if x:left ¤ T:nil and x:left:max i:low 4 x D x:left 5 else x D x:right 6 return x The search for an interval that overlaps i starts with x at the root of the tree and proceeds downward. It terminates when either it finds an overlapping interval or x points to the sentinel T:nil. Since each iteration of the basic loop takes O.1/ time, and since the height of an n-node red-black tree is O.lg n/, the INTERVAL-SEARCH procedure takes O.lg n/ time. Before we see why INTERVAL-SEARCH is correct, let’s examine how it works on the interval tree in Figure 14.4. Suppose we wish to find an interval that overlaps the interval i D Œ22; 25. We begin with x as the root, which contains Œ16; 21 and does not overlap i. Since x:left:max D 23 is greater than i:low D 22, the loop continues with x as the left child of the root—the node containing Œ8; 9, which also does not overlap i. This time, x:left:max D 10 is less than i:low D 22, and so the loop continues with the right child of x as the new x. Because the interval Œ15; 23 stored in this node overlaps i, the procedure returns this node. As an example of an unsuccessful search, suppose we wish to find an interval that overlaps i D Œ11; 14 in the interval tree of Figure 14.4. We once again begin with x as the root. Since the root’s interval Œ16; 21 does not overlap i, and since x:left:max D 23 is greater than i:low D 11, we go left to the node containing Œ8; 9. Interval Œ8; 9 does not overlap i, and x:left:max D 10 is less than i:low D 11, and so we go right. (Note that no interval in the left subtree overlaps i.) Interval Œ15; 23 does not overlap i, and its left child is T:nil, so again we go right, the loop terminates, and we return the sentinel T:nil. To see why INTERVAL-SEARCH is correct, we must understand why it suffices to examine a single path from the root. The basic idea is that at any node x, if x:int does not overlap i, the search always proceeds in a safe direction: the search will definitely find an overlapping interval if the tree contains one. The following theorem states this property more precisely. Theorem 14.2 Any execution of INTERVAL-SEARCH.T; i / either returns a node whose interval overlaps i, or it returns T:nil and the tree T contains no node whose interval overlaps i. 352 Chapter 14 Augmenting Data Structures i′ i′ i (a) i′′ i′′ i′′ i i′ (b) Figure 14.5 Intervals in the proof of Theorem 14.2. The value of x: left: max is shown in each case as a dashed line. (a) The search goes right. No interval i0 in x’s left subtree can overlap i. (b) The search goes left. The left subtree of x contains an interval that overlaps i (situation not shown), or x’s left subtree contains an interval i0 such that i0: high D x: left: max. Since i does not overlap i0, neither does it overlap any interval i00 in x’s right subtree, since i0: low Ä i00: low. Proof The while loop of lines 2–5 terminates either when x D T:nil or i overlaps x:int. In the latter case, it is certainly correct to return x. Therefore, we focus on the former case, in which the while loop terminates because x D T:nil. We use the following invariant for the while loop of lines 2–5: If tree T contains an interval that overlaps i, then the subtree rooted at x contains such an interval. We use this loop invariant as follows: Initialization: Prior to the first iteration, line 1 sets x to be the root of T , so that the invariant holds. Maintenance: Each iteration of the while loop executes either line 4 or line 5. We shall show that both cases maintain the loop invariant. If line 5 is executed, then because of the branch condition in line 3, we have x:left D T:nil, or x:left:max < i:low. If x:left D T:nil, the subtree rooted at x:left clearly contains no interval that overlaps i, and so setting x to x:right maintains the invariant. Suppose, therefore, that x:left ¤ T:nil and x:left:max < i:low. As Figure 14.5(a) shows, for each interval i0 in x’s left subtree, we have i 0:high Ä x:left:max < i:low : By the interval trichotomy, therefore, i0 and i do not overlap. Thus, the left subtree of x contains no intervals that overlap i, so that setting x to x:right maintains the invariant. 14.3 Interval trees 353 If, on the other hand, line 4 is executed, then we will show that the contrapositive of the loop invariant holds. That is, if the subtree rooted at x:left contains no interval overlapping i, then no interval anywhere in the tree overlaps i. Since line 4 is executed, then because of the branch condition in line 3, we have x:left:max i:low. Moreover, by definition of the max attribute, x’s left subtree must contain some interval i0 such that i 0:high D x:left:max i:low : (Figure 14.5(b) illustrates the situation.) Since i and i0 do not overlap, and since it is not true that i0:high < i:low, it follows by the interval trichotomy that i:high < i0:low. Interval trees are keyed on the low endpoints of intervals, and thus the search-tree property implies that for any interval i00 in x’s right subtree, i:high < i 0:low Ä i 00: low : By the interval trichotomy, i and i 00 do not overlap. We conclude that whether or not any interval in x’s left subtree overlaps i, setting x to x:left maintains the invariant. Termination: If the loop terminates when x D T:nil, then the subtree rooted at x contains no interval overlapping i. The contrapositive of the loop invariant implies that T contains no interval that overlaps i. Hence it is correct to return x D T:nil. Thus, the INTERVAL-SEARCH procedure works correctly. Exercises 14.3-1 Write pseudocode for LEFT-ROTATE that operates on nodes in an interval tree and updates the max attributes in O.1/ time. 14.3-2 Rewrite the code for INTERVAL-SEARCH so that it works properly when all intervals are open. 14.3-3 Describe an efficient algorithm that, given an interval i, returns an interval overlapping i that has the minimum low endpoint, or T:nil if no such interval exists. 354 Chapter 14 Augmenting Data Structures 14.3-4 Given an interval tree T and an interval i, describe how to list all intervals in T that overlap i in O.min.n; k lg n// time, where k is the number of intervals in the output list. (Hint: One simple method makes several queries, modifying the tree between queries. A slightly more complicated method does not modify the tree.) 14.3-5 Suggest modifications to the interval-tree procedures to support the new operation INTERVAL-SEARCH-EXACTLY.T; i /, where T is an interval tree and i is an interval. The operation should return a pointer to a node x in T such that x:int:low D i:low and x:int:high D i:high, or T:nil if T contains no such node. All operations, including INTERVAL-SEARCH-EXACTLY, should run in O.lg n/ time on an n-node interval tree. 14.3-6 Show how to maintain a dynamic set Q of numbers that supports the operation MIN-GAP, which gives the magnitude of the difference of the two closest numbers in Q. For example, if Q D f1; 5; 9; 15; 18; 22g, then MIN-GAP.Q/ returns 18 15 D 3, since 15 and 18 are the two closest numbers in Q. Make the operations INSERT, DELETE, SEARCH, and MIN-GAP as efficient as possible, and analyze their running times. 14.3-7 ? VLSI databases commonly represent an integrated circuit as a list of rectangles. Assume that each rectangle is rectilinearly oriented (sides parallel to the x- and y-axes), so that we represent a rectangle by its minimum and maximum xand y-coordinates. Give an O.n lg n/-time algorithm to decide whether or not a set of n rectangles so represented contains two rectangles that overlap. Your algorithm need not report all intersecting pairs, but it must report that an overlap exists if one rectangle entirely covers another, even if the boundary lines do not intersect. (Hint: Move a “sweep” line across the set of rectangles.) Problems 14-1 Point of maximum overlap Suppose that we wish to keep track of a point of maximum overlap in a set of intervals—a point with the largest number of intervals in the set that overlap it. a. Show that there will always be a point of maximum overlap that is an endpoint of one of the segments. Notes for Chapter 14 355 b. Design a data structure that efficiently supports the operations INTERVALINSERT, INTERVAL-DELETE, and FIND-POM, which returns a point of maximum overlap. (Hint: Keep a red-black tree of all the endpoints. Associate a value of C1 with each left endpoint, and associate a value of 1 with each right endpoint. Augment each node of the tree with some extra information to maintain the point of maximum overlap.) 14-2 Josephus permutation We define the Josephus problem as follows. Suppose that n people form a circle and that we are given a positive integer m Ä n. Beginning with a designated first person, we proceed around the circle, removing every mth person. After each person is removed, counting continues around the circle that remains. This process continues until we have removed all n people. The order in which the people are removed from the circle defines the .n; m/-Josephus permutation of the integers 1; 2; : : : ; n. For example, the .7; 3/-Josephus permutation is h3; 6; 2; 7; 5; 1; 4i. a. Suppose that m is a constant. Describe an O.n/-time algorithm that, given an integer n, outputs the .n; m/-Josephus permutation. b. Suppose that m is not a constant. Describe an O.n lg n/-time algorithm that, given integers n and m, outputs the .n; m/-Josephus permutation. Chapter notes In their book, Preparata and Shamos [282] describe several of the interval trees that appear in the literature, citing work by H. Edelsbrunner (1980) and E. M. McCreight (1981). The book details an interval tree that, given a static database of n intervals, allows us to enumerate all k intervals that overlap a given query interval in O.k C lg n/ time. IV Advanced Design and Analysis Techniques Introduction This part covers three important techniques used in designing and analyzing efficient algorithms: dynamic programming (Chapter 15), greedy algorithms (Chapter 16), and amortized analysis (Chapter 17). Earlier parts have presented other widely applicable techniques, such as divide-and-conquer, randomization, and how to solve recurrences. The techniques in this part are somewhat more sophisticated, but they help us to attack many computational problems. The themes introduced in this part will recur later in this book. Dynamic programming typically applies to optimization problems in which we make a set of choices in order to arrive at an optimal solution. As we make each choice, subproblems of the same form often arise. Dynamic programming is effective when a given subproblem may arise from more than one partial set of choices; the key technique is to store the solution to each such subproblem in case it should reappear. Chapter 15 shows how this simple idea can sometimes transform exponential-time algorithms into polynomial-time algorithms. Like dynamic-programming algorithms, greedy algorithms typically apply to optimization problems in which we make a set of choices in order to arrive at an optimal solution. The idea of a greedy algorithm is to make each choice in a locally optimal manner. A simple example is coin-changing: to minimize the number of U.S. coins needed to make change for a given amount, we can repeatedly select the largest-denomination coin that is not larger than the amount that remains. A greedy approach provides an optimal solution for many such problems much more quickly than would a dynamic-programming approach. We cannot always easily tell whether a greedy approach will be effective, however. Chapter 16 introduces 358 Part IV Advanced Design and Analysis Techniques matroid theory, which provides a mathematical basis that can help us to show that a greedy algorithm yields an optimal solution. We use amortized analysis to analyze certain algorithms that perform a sequence of similar operations. Instead of bounding the cost of the sequence of operations by bounding the actual cost of each operation separately, an amortized analysis provides a bound on the actual cost of the entire sequence. One advantage of this approach is that although some operations might be expensive, many others might be cheap. In other words, many of the operations might run in well under the worstcase time. Amortized analysis is not just an analysis tool, however; it is also a way of thinking about the design of algorithms, since the design of an algorithm and the analysis of its running time are often closely intertwined. Chapter 17 introduces three ways to perform an amortized analysis of an algorithm. 15 Dynamic Programming Dynamic programming, like the divide-and-conquer method, solves problems by combining the solutions to subproblems. (“Programming” in this context refers to a tabular method, not to writing computer code.) As we saw in Chapters 2 and 4, divide-and-conquer algorithms partition the problem into disjoint subproblems, solve the subproblems recursively, and then combine their solutions to solve the original problem. In contrast, dynamic programming applies when the subproblems overlap—that is, when subproblems share subsubproblems. In this context, a divide-and-conquer algorithm does more work than necessary, repeatedly solving the common subsubproblems. A dynamic-programming algorithm solves each subsubproblem just once and then saves its answer in a table, thereby avoiding the work of recomputing the answer every time it solves each subsubproblem. We typically apply dynamic programming to optimization problems. Such problems can have many possible solutions. Each solution has a value, and we wish to find a solution with the optimal (minimum or maximum) value. We call such a solution an optimal solution to the problem, as opposed to the optimal solution, since there may be several solutions that achieve the optimal value. When developing a dynamic-programming algorithm, we follow a sequence of four steps: 1. Characterize the structure of an optimal solution. 2. Recursively define the value of an optimal solution. 3. Compute the value of an optimal solution, typically in a bottom-up fashion. 4. Construct an optimal solution from computed information. Steps 1–3 form the basis of a dynamic-programming solution to a problem. If we need only the value of an optimal solution, and not the solution itself, then we can omit step 4. When we do perform step 4, we sometimes maintain additional information during step 3 so that we can easily construct an optimal solution. The sections that follow use the dynamic-programming method to solve some optimization problems. Section 15.1 examines the problem of cutting a rod into 360 Chapter 15 Dynamic Programming rods of smaller length in way that maximizes their total value. Section 15.2 asks how we can multiply a chain of matrices while performing the fewest total scalar multiplications. Given these examples of dynamic programming, Section 15.3 discusses two key characteristics that a problem must have for dynamic programming to be a viable solution technique. Section 15.4 then shows how to find the longest common subsequence of two sequences via dynamic programming. Finally, Section 15.5 uses dynamic programming to construct binary search trees that are optimal, given a known distribution of keys to be looked up. 15.1 Rod cutting Our first example uses dynamic programming to solve a simple problem in deciding where to cut steel rods. Serling Enterprises buys long steel rods and cuts them into shorter rods, which it then sells. Each cut is free. The management of Serling Enterprises wants to know the best way to cut up the rods. We assume that we know, for i D 1; 2; : : :, the price pi in dollars that Serling Enterprises charges for a rod of length i inches. Rod lengths are always an integral number of inches. Figure 15.1 gives a sample price table. The rod-cutting problem is the following. Given a rod of length n inches and a table of prices pi for i D 1; 2; : : : ; n, determine the maximum revenue rn obtainable by cutting up the rod and selling the pieces. Note that if the price pn for a rod of length n is large enough, an optimal solution may require no cutting at all. Consider the case when n D 4. Figure 15.2 shows all the ways to cut up a rod of 4 inches in length, including the way with no cuts at all. We see that cutting a 4-inch rod into two 2-inch pieces produces revenue p2 C p2 D 5 C 5 D 10, which is optimal. We can cut up a rod of length n in 2n 1 different ways, since we have an independent option of cutting, or not cutting, at distance i inches from the left end, length i 1 2 3 4 5 6 7 8 9 10 price pi 1 5 8 9 10 17 17 20 24 30 Figure 15.1 A sample price table for rods. Each rod of length i inches earns the company pi dollars of revenue. 15.1 Rod cutting 9 1 8 5 5 361 8 1 (a) 11 5 (b) 1 5 1 (c) 5 11 (d) 1111 (e) (f) (g) (h) Figure 15.2 The 8 possible ways of cutting up a rod of length 4. Above each piece is the value of that piece, according to the sample price chart of Figure 15.1. The optimal strategy is part (c)—cutting the rod into two pieces of length 2—which has total value 10. for i D 1; 2; : : : ; n 1.1 We denote a decomposition into pieces using ordinary additive notation, so that 7 D 2 C 2 C 3 indicates that a rod of length 7 is cut into three pieces—two of length 2 and one of length 3. If an optimal solution cuts the rod into k pieces, for some 1 Ä k Ä n, then an optimal decomposition n D i1 C i2 C C ik of the rod into pieces of lengths i1, i2, . . . , ik provides maximum corresponding revenue rn D pi1 C pi2 C C pik : For our sample problem, we can determine the optimal revenue figures ri , for i D 1; 2; : : : ; 10, by inspection, with the corresponding optimal decompositions 1If we required the pieces to be cut in order of nondecreasing size, there would be fewer ways to consider. For n D 4, we would consider only 5 such ways: parts (a), (b), (c), (e), and (h) in e pFi2gnu=r3e=145n.2p. 3T. hTehnisumqubaenrtiotyf ways is called is less than 2n the 1, partition function; it but still much greater is approximately equal than any polynomial in to n. We shall not pursue this line of inquiry further, however. 362 Chapter 15 Dynamic Programming r1 D 1 from solution 1 D 1 (no cuts) ; r2 D 5 from solution 2 D 2 (no cuts) ; r3 D 8 from solution 3 D 3 (no cuts) ; r4 D 10 from solution 4 D 2 C 2 ; r5 D 13 from solution 5 D 2 C 3 ; r6 D 17 from solution 6 D 6 (no cuts) ; r7 D 18 from solution 7 D 1 C 6 or 7 D 2 C 2 C 3 ; r8 D 22 from solution 8 D 2 C 6 ; r9 D 25 from solution 9 D 3 C 6 ; r10 D 30 from solution 10 D 10 (no cuts) : More generally, we can frame the values rn for n 1 in terms of optimal revenues from shorter rods: rn D max .pn; r1 C rn 1; r2 C rn 2; : : : ; rn 1 C r1/ : (15.1) The first argument, pn, corresponds to making no cuts at all and selling the rod of length n as is. The other n 1 arguments to max correspond to the maximum revenue obtained by making an initial cut of the rod into two pieces of size i and n i, for each i D 1; 2; : : : ; n 1, and then optimally cutting up those pieces further, obtaining revenues ri and rn i from those two pieces. Since we don’t know ahead of time which value of i optimizes revenue, we have to consider all possible values for i and pick the one that maximizes revenue. We also have the option of picking no i at all if we can obtain more revenue by selling the rod uncut. Note that to solve the original problem of size n, we solve smaller problems of the same type, but of smaller sizes. Once we make the first cut, we may consider the two pieces as independent instances of the rod-cutting problem. The overall optimal solution incorporates optimal solutions to the two related subproblems, maximizing revenue from each of those two pieces. We say that the rod-cutting problem exhibits optimal substructure: optimal solutions to a problem incorporate optimal solutions to related subproblems, which we may solve independently. In a related, but slightly simpler, way to arrange a recursive structure for the rodcutting problem, we view a decomposition as consisting of a first piece of length i cut off the left-hand end, and then a right-hand remainder of length n i. Only the remainder, and not the first piece, may be further divided. We may view every decomposition of a length-n rod in this way: as a first piece followed by some decomposition of the remainder. When doing so, we can couch the solution with no cuts at all as saying that the first piece has size i D n and revenue pn and that the remainder has size 0 with corresponding revenue r0 D 0. We thus obtain the following simpler version of equation (15.1): rn D max .pi C rn i / : 1Äi Än (15.2) 15.1 Rod cutting 363 In this formulation, an optimal solution embodies the solution to only one related subproblem—the remainder—rather than two. Recursive top-down implementation The following procedure implements the computation implicit in equation (15.2) in a straightforward, top-down, recursive manner. CUT-ROD.p; n/ 1 if n == 0 2 return 0 3 qD 1 4 for i D 1 to n 5 q D max.q; pŒi  C CUT-ROD.p; n i // 6 return q Procedure CUT-ROD takes as input an array pŒ1 : : n of prices and an integer n, and it returns the maximum revenue possible for a rod of length n. If n D 0, no revenue is possible, and so CUT-ROD returns 0 in line 2. Line 3 initializes the maximum revenue q to 1, so that the for loop in lines 4–5 correctly computes q D max1ÄiÄn.pi C CUT-ROD.p; n i //; line 6 then returns this value. A simple induction on n proves that this answer is equal to the desired answer rn, using equation (15.2). If you were to code up CUT-ROD in your favorite programming language and run it on your computer, you would find that once the input size becomes moderately large, your program would take a long time to run. For n D 40, you would find that your program takes at least several minutes, and most likely more than an hour. In fact, you would find that each time you increase n by 1, your program’s running time would approximately double. Why is CUT-ROD so inefficient? The problem is that CUT-ROD calls itself recursively over and over again with the same parameter values; it solves the same subproblems repeatedly. Figure 15.3 illustrates what happens for n D 4: CUT-ROD.p; n/ calls CUT-ROD.p; n i / for i D 1; 2; : : : ; n. Equivalently, CUT-ROD.p; n/ calls CUT-ROD.p; j / for each j D 0; 1; : : : ; n 1. When this process unfolds recursively, the amount of work done, as a function of n, grows explosively. To analyze the running time of CUT-ROD, let T .n/ denote the total number of calls made to CUT-ROD when called with its second parameter equal to n. This expression equals the number of nodes in a subtree whose root is labeled n in the recursion tree. The count includes the initial call at its root. Thus, T .0/ D 1 and 364 Chapter 15 Dynamic Programming 4 3 2 10 2 10100 100 0 0 Figure 15.3 The recursion tree showing recursive calls resulting from a call CUT-ROD.p; n/ for n D 4. Each node label gives the size n of the corresponding subproblem, so that an edge from a parent with label s to a child with label t corresponds to cutting off an initial piece of size s t and leaving a remaining subproblem of size t. A path from the root to a leaf corresponds to one of the 2n 1 ways of cutting up a rod of length n. In general, this recursion tree has 2n nodes and 2n 1 leaves. X n 1 T .n/ D 1 C T .j / : j D0 (15.3) The initial 1 is for the call at the root, and the term T .j / counts the number of calls (including recursive calls) due to the call CUT-ROD.p; n i /, where j D n i . As Exercise 15.1-1 asks you to show, T .n/ D 2n ; (15.4) and so the running time of CUT-ROD is exponential in n. In retrospect, this exponential running time is not so surprising. CUT-ROD ex- plicitly considers all the 2n 1 possible ways of cutting up a rod of length n. The tree of recursive calls has 2n 1 leaves, one for each possible way of cutting up the rod. The labels on the simple path from the root to a leaf give the sizes of each remaining right-hand piece before making each cut. That is, the labels give the corresponding cut points, measured from the right-hand end of the rod. Using dynamic programming for optimal rod cutting We now show how to convert CUT-ROD into an efficient algorithm, using dynamic programming. The dynamic-programming method works as follows. Having observed that a naive recursive solution is inefficient because it solves the same subproblems repeatedly, we arrange for each subproblem to be solved only once, saving its solution. If we need to refer to this subproblem’s solution again later, we can just look it 15.1 Rod cutting 365 up, rather than recompute it. Dynamic programming thus uses additional memory to save computation time; it serves an example of a time-memory trade-off. The savings may be dramatic: an exponential-time solution may be transformed into a polynomial-time solution. A dynamic-programming approach runs in polynomial time when the number of distinct subproblems involved is polynomial in the input size and we can solve each such subproblem in polynomial time. There are usually two equivalent ways to implement a dynamic-programming approach. We shall illustrate both of them with our rod-cutting example. The first approach is top-down with memoization.2 In this approach, we write the procedure recursively in a natural manner, but modified to save the result of each subproblem (usually in an array or hash table). The procedure now first checks to see whether it has previously solved this subproblem. If so, it returns the saved value, saving further computation at this level; if not, the procedure computes the value in the usual manner. We say that the recursive procedure has been memoized; it “remembers” what results it has computed previously. The second approach is the bottom-up method. This approach typically depends on some natural notion of the “size” of a subproblem, such that solving any particular subproblem depends only on solving “smaller” subproblems. We sort the subproblems by size and solve them in size order, smallest first. When solving a particular subproblem, we have already solved all of the smaller subproblems its solution depends upon, and we have saved their solutions. We solve each subproblem only once, and when we first see it, we have already solved all of its prerequisite subproblems. These two approaches yield algorithms with the same asymptotic running time, except in unusual circumstances where the top-down approach does not actually recurse to examine all possible subproblems. The bottom-up approach often has much better constant factors, since it has less overhead for procedure calls. Here is the the pseudocode for the top-down CUT-ROD procedure, with memoization added: MEMOIZED-CUT-ROD .p; n/ 1 let rŒ0 : : n be a new array 2 for i D 0 to n 3 rŒi D 1 4 return MEMOIZED-CUT-ROD-AUX.p; n; r/ 2This is not a misspelling. The word really is memoization, not memorization. Memoization comes from memo, since the technique consists of recording a value so that we can look it up later. 366 Chapter 15 Dynamic Programming MEMOIZED-CUT-ROD-AUX.p; n; r/ 1 if rŒn 0 2 return rŒn 3 if n == 0 4 qD0 5 else q D 1 6 for i D 1 to n 7 q D max.q; pŒi  C MEMOIZED-CUT-ROD-AUX.p; n 8 rŒn D q 9 return q i; r// Here, the main procedure MEMOIZED-CUT-ROD initializes a new auxiliary array rŒ0 : : n with the value 1, a convenient choice with which to denote “unknown.” (Known revenue values are always nonnegative.) It then calls its helper routine, MEMOIZED-CUT-ROD-AUX. The procedure MEMOIZED-CUT-ROD-AUX is just the memoized version of our previous procedure, CUT-ROD. It first checks in line 1 to see whether the desired value is already known and, if it is, then line 2 returns it. Otherwise, lines 3–7 compute the desired value q in the usual manner, line 8 saves it in rŒn, and line 9 returns it. The bottom-up version is even simpler: BOTTOM-UP-CUT-ROD.p; n/ 1 let rŒ0 : : n be a new array 2 rŒ0 D 0 3 for j D 1 to n 4 qD 1 5 for i D 1 to j 6 q D max.q; pŒi C rŒj i/ 7 rŒj  D q 8 return rŒn For the bottom-up dynamic-programming approach, BOTTOM-UP-CUT-ROD uses the natural ordering of the subproblems: a problem of size i is “smaller” than a subproblem of size j if i < j . Thus, the procedure solves subproblems of sizes j D 0; 1; : : : ; n, in that order. Line 1 of procedure BOTTOM-UP-CUT-ROD creates a new array rŒ0 : : n in which to save the results of the subproblems, and line 2 initializes rŒ0 to 0, since a rod of length 0 earns no revenue. Lines 3–6 solve each subproblem of size j , for j D 1; 2; : : : ; n, in order of increasing size. The approach used to solve a problem of a particular size j is the same as that used by CUT-ROD, except that line 6 now 15.1 Rod cutting 367 4 3 2 1 0 Figure 15.4 The subproblem graph for the rod-cutting problem with n D 4. The vertex labels give the sizes of the corresponding subproblems. A directed edge .x; y/ indicates that we need a solution to subproblem y when solving subproblem x. This graph is a reduced version of the tree of Figure 15.3, in which all nodes with the same label are collapsed into a single vertex and all edges go from parent to child. directly references array entry rŒj i instead of making a recursive call to solve the subproblem of size j i. Line 7 saves in rŒj  the solution to the subproblem of size j . Finally, line 8 returns rŒn, which equals the optimal value rn. The bottom-up and top-down versions have the same asymptotic running time. The running time of procedure BOTTOM-UP-CUT-ROD is ‚.n2/, due to its doubly-nested loop structure. The number of iterations of its inner for loop, in lines 5–6, forms an arithmetic series. The running time of its top-down counterpart, MEMOIZED-CUT-ROD, is also ‚.n2/, although this running time may be a little harder to see. Because a recursive call to solve a previously solved subproblem returns immediately, MEMOIZED-CUT-ROD solves each subproblem just once. It solves subproblems for sizes 0; 1; : : : ; n. To solve a subproblem of size n, the for loop of lines 6–7 iterates n times. Thus, the total number of iterations of this for loop, over all recursive calls of MEMOIZED-CUT-ROD, forms an arithmetic series, giving a total of ‚.n2/ iterations, just like the inner for loop of BOTTOM-UPCUT-ROD. (We actually are using a form of aggregate analysis here. We shall see aggregate analysis in detail in Section 17.1.) Subproblem graphs When we think about a dynamic-programming problem, we should understand the set of subproblems involved and how subproblems depend on one another. The subproblem graph for the problem embodies exactly this information. Figure 15.4 shows the subproblem graph for the rod-cutting problem with n D 4. It is a directed graph, containing one vertex for each distinct subproblem. The sub- 368 Chapter 15 Dynamic Programming problem graph has a directed edge from the vertex for subproblem x to the vertex for subproblem y if determining an optimal solution for subproblem x involves directly considering an optimal solution for subproblem y. For example, the subproblem graph contains an edge from x to y if a top-down recursive procedure for solving x directly calls itself to solve y. We can think of the subproblem graph as a “reduced” or “collapsed” version of the recursion tree for the top-down recursive method, in which we coalesce all nodes for the same subproblem into a single vertex and direct all edges from parent to child. The bottom-up method for dynamic programming considers the vertices of the subproblem graph in such an order that we solve the subproblems y adjacent to a given subproblem x before we solve subproblem x. (Recall from Section B.4 that the adjacency relation is not necessarily symmetric.) Using the terminology from Chapter 22, in a bottom-up dynamic-programming algorithm, we consider the vertices of the subproblem graph in an order that is a “reverse topological sort,” or a “topological sort of the transpose” (see Section 22.4) of the subproblem graph. In other words, no subproblem is considered until all of the subproblems it depends upon have been solved. Similarly, using notions from the same chapter, we can view the top-down method (with memoization) for dynamic programming as a “depth-first search” of the subproblem graph (see Section 22.3). The size of the subproblem graph G D .V; E/ can help us determine the running time of the dynamic programming algorithm. Since we solve each subproblem just once, the running time is the sum of the times needed to solve each subproblem. Typically, the time to compute the solution to a subproblem is proportional to the degree (number of outgoing edges) of the corresponding vertex in the subproblem graph, and the number of subproblems is equal to the number of vertices in the subproblem graph. In this common case, the running time of dynamic programming is linear in the number of vertices and edges. Reconstructing a solution Our dynamic-programming solutions to the rod-cutting problem return the value of an optimal solution, but they do not return an actual solution: a list of piece sizes. We can extend the dynamic-programming approach to record not only the optimal value computed for each subproblem, but also a choice that led to the optimal value. With this information, we can readily print an optimal solution. Here is an extended version of BOTTOM-UP-CUT-ROD that computes, for each rod size j , not only the maximum revenue rj , but also sj , the optimal size of the first piece to cut off: 15.1 Rod cutting 369 EXTENDED-BOTTOM-UP-CUT-ROD.p; n/ 1 let rŒ0 : : n and sŒ0 : : n be new arrays 2 rŒ0 D 0 3 for j D 1 to n 4 qD 1 5 for i D 1 to j 6 if q < pŒi C rŒj i 7 q D pŒi C rŒj i 8 sŒj  D i 9 rŒj  D q 10 return r and s This procedure is similar to BOTTOM-UP-CUT-ROD, except that it creates the array s in line 1, and it updates sŒj  in line 8 to hold the optimal size i of the first piece to cut off when solving a subproblem of size j . The following procedure takes a price table p and a rod size n, and it calls EXTENDED-BOTTOM-UP-CUT-ROD to compute the array sŒ1 : : n of optimal first-piece sizes and then prints out the complete list of piece sizes in an optimal decomposition of a rod of length n: PRINT-CUT-ROD-SOLUTION .p; n/ 1 .r; s/ D EXTENDED-BOTTOM-UP-CUT-ROD .p; n/ 2 while n > 0 3 print sŒn 4 n D n sŒn In our rod-cutting example, the call EXTENDED-BOTTOM-UP-CUT-ROD.p; 10/ would return the following arrays: i 0 1 2 3 4 5 6 7 8 9 10 rŒi 0 1 5 8 10 13 17 18 22 25 30 sŒi 0 1 2 3 2 2 6 1 2 3 10 A call to PRINT-CUT-ROD-SOLUTION.p; 10/ would print just 10, but a call with n D 7 would print the cuts 1 and 6, corresponding to the first optimal decomposition for r7 given earlier. Exercises 15.1-1 Show that equation (15.4) follows from equation (15.3) and the initial condition T .0/ D 1. 370 Chapter 15 Dynamic Programming 15.1-2 Show, by means of a counterexample, that the following “greedy” strategy does not always determine an optimal way to cut rods. Define the density of a rod of length i to be pi =i, that is, its value per inch. The greedy strategy for a rod of length n cuts off a first piece of length i, where 1 Ä i Ä n, having maximum density. It then continues by applying the greedy strategy to the remaining piece of length n i. 15.1-3 Consider a modification of the rod-cutting problem in which, in addition to a price pi for each rod, each cut incurs a fixed cost of c. The revenue associated with a solution is now the sum of the prices of the pieces minus the costs of making the cuts. Give a dynamic-programming algorithm to solve this modified problem. 15.1-4 Modify MEMOIZED-CUT-ROD to return not only the value but the actual solution, too. 15.1-5 The Fibonacci numbers are defined by recurrence (3.22). Give an O.n/-time dynamic-programming algorithm to compute the nth Fibonacci number. Draw the subproblem graph. How many vertices and edges are in the graph? 15.2 Matrix-chain multiplication Our next example of dynamic programming is an algorithm that solves the problem of matrix-chain multiplication. We are given a sequence (chain) hA1; A2; : : : ; Ani of n matrices to be multiplied, and we wish to compute the product A1A2 An : (15.5) We can evaluate the expression (15.5) using the standard algorithm for multiplying pairs of matrices as a subroutine once we have parenthesized it to resolve all ambiguities in how the matrices are multiplied together. Matrix multiplication is associative, and so all parenthesizations yield the same product. A product of matrices is fully parenthesized if it is either a single matrix or the product of two fully parenthesized matrix products, surrounded by parentheses. For example, if the chain of matrices is hA1; A2; A3; A4i, then we can fully parenthesize the product A1A2A3A4 in five distinct ways: 15.2 Matrix-chain multiplication 371 .A1.A2.A3A4/// ; .A1..A2A3/A4// ; ..A1A2/.A3A4// ; ..A1.A2A3//A4/ ; ...A1A2/A3/A4/ : How we parenthesize a chain of matrices can have a dramatic impact on the cost of evaluating the product. Consider first the cost of multiplying two matrices. The standard algorithm is given by the following pseudocode, which generalizes the SQUARE-MATRIX-MULTIPLY procedure from Section 4.2. The attributes rows and columns are the numbers of rows and columns in a matrix. MATRIX-MULTIPLY.A; B/ 1 if A:columns ¤ B:rows 2 error “incompatible dimensions” 3 else let C be a new A:rows B:columns matrix 4 for i D 1 to A:rows 5 for j D 1 to B:columns 6 cij D 0 7 for k D 1 to A:columns 8 cij D cij C aik bkj 9 return C We can multiply two matrices A and B only if they are compatible: the number of columns of A must equal the number of rows of B. If A is a p q matrix and B is a q r matrix, the resulting matrix C is a p r matrix. The time to compute C is dominated by the number of scalar multiplications in line 8, which is pqr. In what follows, we shall express costs in terms of the number of scalar multiplications. To illustrate the different costs incurred by different parenthesizations of a matrix product, consider the problem of a chain hA1; A2; A3i of three matrices. Suppose that the dimensions of the matrices are 10 100, 100 5, and 5 50, respectively. If we multiply according to the parenthesization ..A1A2/A3/, we perform 10 100 5 D 5000 scalar multiplications to compute the 10 5 matrix product A1A2, plus another 10 5 50 D 2500 scalar multiplications to multiply this matrix by A3, for a total of 7500 scalar multiplications. If instead we multiply according to the parenthesization .A1.A2A3//, we perform 100 5 50 D 25,000 scalar multiplications to compute the 100 50 matrix product A2A3, plus another 10 100 50 D 50,000 scalar multiplications to multiply A1 by this matrix, for a total of 75,000 scalar multiplications. Thus, computing the product according to the first parenthesization is 10 times faster. We state the matrix-chain multiplication problem as follows: given a chain hA1; A2; : : : ; Ani of n matrices, where for i D 1; 2; : : : ; n, matrix Ai has dimension 372 Chapter 15 Dynamic Programming pi 1 pi , fully parenthesize the product A1A2 An in a way that minimizes the number of scalar multiplications. Note that in the matrix-chain multiplication problem, we are not actually multiplying matrices. Our goal is only to determine an order for multiplying matrices that has the lowest cost. Typically, the time invested in determining this optimal order is more than paid for by the time saved later on when actually performing the matrix multiplications (such as performing only 7500 scalar multiplications instead of 75,000). Counting the number of parenthesizations Before solving the matrix-chain multiplication problem by dynamic programming, let us convince ourselves that exhaustively checking all possible parenthesizations does not yield an efficient algorithm. Denote the number of alternative parenthesizations of a sequence of n matrices by P .n/. When n D 1, we have just one matrix and therefore only one way to fully parenthesize the matrix product. When n 2, a fully parenthesized matrix product is the product of two fully parenthe- sized matrix subproducts, and the split between the two subproducts may occur between the kth and .k C 1/st matrices for any k D 1; 2; : : : ; n 1. Thus, we € obtain the recurrence 1 if n D 1 ; P .n/ D X n 1 P .k/P .n k/ if n 2 : (15.6) kD1 Problem 12-4 asked you to show that the solution to a similar recurrence is the sequence of Catalan numbers, which grows as .4n=n3=2/. A simpler exercise (see Exercise 15.2-3) is to show that the solution to the recurrence (15.6) is .2n/. The number of solutions is thus exponential in n, and the brute-force method of exhaustive search makes for a poor strategy when determining how to optimally parenthesize a matrix chain. Applying dynamic programming We shall use the dynamic-programming method to determine how to optimally parenthesize a matrix chain. In so doing, we shall follow the four-step sequence that we stated at the beginning of this chapter: 1. Characterize the structure of an optimal solution. 2. Recursively define the value of an optimal solution. 3. Compute the value of an optimal solution. 15.2 Matrix-chain multiplication 373 4. Construct an optimal solution from computed information. We shall go through these steps in order, demonstrating clearly how we apply each step to the problem. Step 1: The structure of an optimal parenthesization For our first step in the dynamic-programming paradigm, we find the optimal substructure and then use it to construct an optimal solution to the problem from optimal solutions to subproblems. In the matrix-chain multiplication problem, we can perform this step as follows. For convenience, let us adopt the notation Ai::j , where i Ä j , for the matrix that results from evaluating the product Ai AiC1 Aj . Observe that if the problem is nontrivial, i.e., i < j , then to parenthesize the product Ai AiC1 Aj , we must split the product between Ak and AkC1 for some integer k in the range i Ä k < j . That is, for some value of k, we first compute the matrices Ai::k and AkC1::j and then multiply them together to produce the final product Ai::j . The cost of parenthesizing this way is the cost of computing the matrix Ai::k, plus the cost of computing AkC1::j , plus the cost of multiplying them together. The optimal substructure of this problem is as follows. Suppose that to optimally parenthesize Ai AiC1 Aj , we split the product between Ak and AkC1. Then the way we parenthesize the “prefix” subchain Ai AiC1 Ak within this optimal parenthesization of Ai AiC1 Aj must be an optimal parenthesization of Ai AiC1 Ak. Why? If there were a less costly way to parenthesize Ai AiC1 Ak, then we could substitute that parenthesization in the optimal parenthesization of Ai AiC1 Aj to produce another way to parenthesize Ai AiC1 Aj whose cost was lower than the optimum: a contradiction. A similar observation holds for how we parenthesize the subchain AkC1AkC2 Aj in the optimal parenthesization of Ai AiC1 Aj : it must be an optimal parenthesization of AkC1AkC2 Aj . Now we use our optimal substructure to show that we can construct an optimal solution to the problem from optimal solutions to subproblems. We have seen that any solution to a nontrivial instance of the matrix-chain multiplication problem requires us to split the product, and that any optimal solution contains within it optimal solutions to subproblem instances. Thus, we can build an optimal solution to an instance of the matrix-chain multiplication problem by splitting the problem into two subproblems (optimally parenthesizing Ai AiC1 Ak and AkC1AkC2 Aj ), finding optimal solutions to subproblem instances, and then combining these optimal subproblem solutions. We must ensure that when we search for the correct place to split the product, we have considered all possible places, so that we are sure of having examined the optimal one. 374 Chapter 15 Dynamic Programming Step 2: A recursive solution Next, we define the cost of an optimal solution recursively in terms of the optimal solutions to subproblems. For the matrix-chain multiplication problem, we pick as our subproblems the problems of determining the minimum cost of parenthesizing Ai AiC1 Aj for 1 Ä i Ä j Ä n. Let mŒi; j  be the minimum number of scalar multiplications needed to compute the matrix Ai::j ; for the full problem, the lowestcost way to compute A1::n would thus be mŒ1; n. We can define mŒi; j  recursively as follows. If i D j , the problem is trivial; the chain consists of just one matrix Ai::i D Ai , so that no scalar multiplications are necessary to compute the product. Thus, mŒi; i D 0 for i D 1; 2; : : : ; n. To compute mŒi; j  when i < j , we take advantage of the structure of an optimal solution from step 1. Let us assume that to optimally parenthesize, we split the product Ai AiC1 Aj between Ak and AkC1, where i Ä k < j . Then, mŒi; j  equals the minimum cost for computing the subproducts Ai::k and AkC1::j , plus the cost of multiplying these two matrices together. Recalling that each matrix Ai is pi 1 pi , we see that computing the matrix product Ai::kAkC1::j takes pi 1pkpj scalar multiplications. Thus, we obtain mŒi; j  D mŒi; k C mŒk C 1; j  C pi 1pkpj : This recursive equation assumes that we know the value of k, which we do not. There are only j i possible values for k, however, namely k D i; i C1; : : : ; j 1. Since the optimal parenthesization must use one of these values for k, we need only check them all to find the best. Thus, our recursive definition for the minimum cost of parenthesizing the product Ai AiC1 Aj becomes ( 0 mŒi; j  D min fmŒi; k C mŒk C 1; j  C pi 1pkpj g i Äk 1 : kD1 Noting that for i D 1; 2; : : : ; n 1, each term T .i/ appears once as T .k/ and once as T .n k/, and collecting the n 1 1s in the summation together with the 1 out front, we can rewrite the recurrence as X n 1 T .n/ 2 T .i/ C n : (15.8) i D1 We shall prove that T .n/ D .2n/ using the substitution method. Specifically, we shall show that T .n/ 2n 1 for all n 1. The basis is easy, since T .1/ 1 D 20. Inductively, for n 2 we have T .n/ X n 1 2 2i 1 C n i D1 X n 2 D 2 2i C n i D0 D 2.2n 1 1/ C n (by equation (A.5)) D 2n 2 C n 2n 1 ; which completes the proof. Thus, the total amount of work performed by the call RECURSIVE-MATRIX-CHAIN.p; 1; n/ is at least exponential in n. Compare this top-down, recursive algorithm (without memoization) with the bottom-up dynamic-programming algorithm. The latter is more efficient because it takes advantage of the overlapping-subproblems property. Matrix-chain multiplication has only ‚.n2/ distinct subproblems, and the dynamic-programming algorithm solves each exactly once. The recursive algorithm, on the other hand, must again solve each subproblem every time it reappears in the recursion tree. Whenever a recursion tree for the natural recursive solution to a problem contains the same subproblem repeatedly, and the total number of distinct subproblems is small, dynamic programming can improve efficiency, sometimes dramatically. 15.3 Elements of dynamic programming 387 Reconstructing an optimal solution As a practical matter, we often store which choice we made in each subproblem in a table so that we do not have to reconstruct this information from the costs that we stored. For matrix-chain multiplication, the table sŒi; j  saves us a significant amount of work when reconstructing an optimal solution. Suppose that we did not maintain the sŒi; j  table, having filled in only the table mŒi; j  containing optimal subproblem costs. We choose from among j i possibilities when we determine which subproblems to use in an optimal solution to parenthesizing Ai AiC1 Aj , and j i is not a constant. Therefore, it would take ‚.j i/ D !.1/ time to reconstruct which subproblems we chose for a solution to a given problem. By storing in sŒi; j  the index of the matrix at which we split the product Ai AiC1 Aj , we can reconstruct each choice in O.1/ time. Memoization As we saw for the rod-cutting problem, there is an alternative approach to dynamic programming that often offers the efficiency of the bottom-up dynamicprogramming approach while maintaining a top-down strategy. The idea is to memoize the natural, but inefficient, recursive algorithm. As in the bottom-up approach, we maintain a table with subproblem solutions, but the control structure for filling in the table is more like the recursive algorithm. A memoized recursive algorithm maintains an entry in a table for the solution to each subproblem. Each table entry initially contains a special value to indicate that the entry has yet to be filled in. When the subproblem is first encountered as the recursive algorithm unfolds, its solution is computed and then stored in the table. Each subsequent time that we encounter this subproblem, we simply look up the value stored in the table and return it.5 Here is a memoized version of RECURSIVE-MATRIX-CHAIN. Note where it resembles the memoized top-down method for the rod-cutting problem. 5This approach presupposes that we know the set of all possible subproblem parameters and that we have established the relationship between table positions and subproblems. Another, more general, approach is to memoize by using hashing with the subproblem parameters as keys. 388 Chapter 15 Dynamic Programming MEMOIZED-MATRIX-CHAIN .p/ 1 n D p:length 1 2 let mŒ1 : : n; 1 : : n be a new table 3 for i D 1 to n 4 for j D i to n 5 mŒi; j  D 1 6 return LOOKUP-CHAIN.m; p; 1; n/ LOOKUP-CHAIN.m; p; i; j / 1 if mŒi; j  < 1 2 return mŒi; j  3 if i == j 4 mŒi; j  D 0 5 else for k D i to j 1 6 q D LOOKUP-CHAIN.m; p; i; k/ C LOOKUP-CHAIN.m; p; k C 1; j / C pi 1pkpj 7 if q < mŒi; j  8 mŒi; j  D q 9 return mŒi; j  The MEMOIZED-MATRIX-CHAIN procedure, like MATRIX-CHAIN-ORDER, maintains a table mŒ1 : : n; 1 : : n of computed values of mŒi; j , the minimum number of scalar multiplications needed to compute the matrix Ai::j . Each table entry initially contains the value 1 to indicate that the entry has yet to be filled in. Upon calling LOOKUP-CHAIN.m; p; i; j /, if line 1 finds that mŒi; j  < 1, then the procedure simply returns the previously computed cost mŒi; j  in line 2. Otherwise, the cost is computed as in RECURSIVE-MATRIX-CHAIN, stored in mŒi; j , and returned. Thus, LOOKUP-CHAIN.m; p; i; j / always returns the value of mŒi; j , but it computes it only upon the first call of LOOKUP-CHAIN with these specific values of i and j . Figure 15.7 illustrates how MEMOIZED-MATRIX-CHAIN saves time compared with RECURSIVE-MATRIX-CHAIN. Shaded subtrees represent values that it looks up rather than recomputes. Like the bottom-up dynamic-programming algorithm MATRIX-CHAIN-ORDER, the procedure MEMOIZED-MATRIX-CHAIN runs in O.n3/ time. Line 5 of MEMOIZED-MATRIX-CHAIN executes ‚.n2/ times. We can categorize the calls of LOOKUP-CHAIN into two types: 1. calls in which mŒi; j  D 1, so that lines 3–9 execute, and 2. calls in which mŒi; j  < 1, so that LOOKUP-CHAIN simply returns in line 2. 15.3 Elements of dynamic programming 389 There are ‚.n2/ calls of the first type, one per table entry. All calls of the second type are made as recursive calls by calls of the first type. Whenever a given call of LOOKUP-CHAIN makes recursive calls, it makes O.n/ of them. Therefore, there are O.n3/ calls of the second type in all. Each call of the second type takes O.1/ time, and each call of the first type takes O.n/ time plus the time spent in its recursive calls. The total time, therefore, is O.n3/. Memoization thus turns an .2n/-time algorithm into an O.n3/-time algorithm. In summary, we can solve the matrix-chain multiplication problem by either a top-down, memoized dynamic-programming algorithm or a bottom-up dynamicprogramming algorithm in O.n3/ time. Both methods take advantage of the overlapping-subproblems property. There are only ‚.n2/ distinct subproblems in total, and either of these methods computes the solution to each subproblem only once. Without memoization, the natural recursive algorithm runs in exponential time, since solved subproblems are repeatedly solved. In general practice, if all subproblems must be solved at least once, a bottom-up dynamic-programming algorithm usually outperforms the corresponding top-down memoized algorithm by a constant factor, because the bottom-up algorithm has no overhead for recursion and less overhead for maintaining the table. Moreover, for some problems we can exploit the regular pattern of table accesses in the dynamicprogramming algorithm to reduce time or space requirements even further. Alternatively, if some subproblems in the subproblem space need not be solved at all, the memoized solution has the advantage of solving only those subproblems that are definitely required. Exercises 15.3-1 Which is a more efficient way to determine the optimal number of multiplications in a matrix-chain multiplication problem: enumerating all the ways of parenthesizing the product and computing the number of multiplications for each, or running RECURSIVE-MATRIX-CHAIN? Justify your answer. 15.3-2 Draw the recursion tree for the MERGE-SORT procedure from Section 2.3.1 on an array of 16 elements. Explain why memoization fails to speed up a good divideand-conquer algorithm such as MERGE-SORT. 15.3-3 Consider a variant of the matrix-chain multiplication problem in which the goal is to parenthesize the sequence of matrices so as to maximize, rather than minimize, 390 Chapter 15 Dynamic Programming the number of scalar multiplications. Does this problem exhibit optimal substructure? 15.3-4 As stated, in dynamic programming we first solve the subproblems and then choose which of them to use in an optimal solution to the problem. Professor Capulet claims that we do not always need to solve all the subproblems in order to find an optimal solution. She suggests that we can find an optimal solution to the matrixchain multiplication problem by always choosing the matrix Ak at which to split the subproduct Ai AiC1 Aj (by selecting k to minimize the quantity pi 1pkpj ) before solving the subproblems. Find an instance of the matrix-chain multiplication problem for which this greedy approach yields a suboptimal solution. 15.3-5 Suppose that in the rod-cutting problem of Section 15.1, we also had limit li on the number of pieces of length i that we are allowed to produce, for i D 1; 2; : : : ; n. Show that the optimal-substructure property described in Section 15.1 no longer holds. 15.3-6 Imagine that you wish to exchange one currency for another. You realize that instead of directly exchanging one currency for another, you might be better off making a series of trades through other currencies, winding up with the currency you want. Suppose that you can trade n different currencies, numbered 1; 2; : : : ; n, where you start with currency 1 and wish to wind up with currency n. You are given, for each pair of currencies i and j , an exchange rate rij , meaning that if you start with d units of currency i , you can trade for drij units of currency j . A sequence of trades may entail a commission, which depends on the number of trades you make. Let ck be the commission that you are charged when you make k trades. Show that, if ck D 0 for all k D 1; 2; : : : ; n, then the problem of finding the best sequence of exchanges from currency 1 to currency n exhibits optimal substructure. Then show that if commissions ck are arbitrary values, then the problem of finding the best sequence of exchanges from currency 1 to currency n does not necessarily exhibit optimal substructure. 15.4 Longest common subsequence Biological applications often need to compare the DNA of two (or more) different organisms. A strand of DNA consists of a string of molecules called 15.4 Longest common subsequence 391 bases, where the possible bases are adenine, guanine, cytosine, and thymine. Representing each of these bases by its initial letter, we can express a strand of DNA as a string over the finite set fA; C; G; Tg. (See Appendix C for the definition of a string.) For example, the DNA of one organism may be S1 D ACCGGTCGAGTGCGCGGAAGCCGGCCGAA, and the DNA of another organism may be S2 D GTCGTTCGGAATGCCGTTGCTCTGTAAA. One reason to compare two strands of DNA is to determine how “similar” the two strands are, as some measure of how closely related the two organisms are. We can, and do, define similarity in many different ways. For example, we can say that two DNA strands are similar if one is a substring of the other. (Chapter 32 explores algorithms to solve this problem.) In our example, neither S1 nor S2 is a substring of the other. Alternatively, we could say that two strands are similar if the number of changes needed to turn one into the other is small. (Problem 15-5 looks at this notion.) Yet another way to measure the similarity of strands S1 and S2 is by finding a third strand S3 in which the bases in S3 appear in each of S1 and S2; these bases must appear in the same order, but not necessarily consecutively. The longer the strand S3 we can find, the more similar S1 and S2 are. In our example, the longest strand S3 is GTCGTCGGAAGCCGGCCGAA. We formalize this last notion of similarity as the longest-common-subsequence problem. A subsequence of a given sequence is just the given sequence with zero or more elements left out. Formally, given a sequence X D hx1; x2; : : : ; xmi, another sequence Z D h´1; ´2; : : : ; ´ki is a subsequence of X if there exists a strictly increasing sequence hi1; i2; : : : ; iki of indices of X such that for all j D 1; 2; : : : ; k, we have xij D ´j . For example, Z D hB; C; D; Bi is a subsequence of X D hA; B; C; B; D; A; Bi with corresponding index sequence h2; 3; 5; 7i. Given two sequences X and Y , we say that a sequence Z is a common subsequence of X and Y if Z is a subsequence of both X and Y . For example, if X D hA; B; C; B; D; A; Bi and Y D hB; D; C; A; B; Ai, the sequence hB; C; Ai is a common subsequence of both X and Y . The sequence hB; C; Ai is not a longest common subsequence (LCS) of X and Y , however, since it has length 3 and the sequence hB; C; B; Ai, which is also common to both X and Y , has length 4. The sequence hB; C; B; Ai is an LCS of X and Y , as is the sequence hB; D; A; Bi, since X and Y have no common subsequence of length 5 or greater. In the longest-common-subsequence problem, we are given two sequences X D hx1; x2; : : : ; xmi and Y D hy1; y2; : : : ; yni and wish to find a maximumlength common subsequence of X and Y . This section shows how to efficiently solve the LCS problem using dynamic programming. 392 Chapter 15 Dynamic Programming Step 1: Characterizing a longest common subsequence In a brute-force approach to solving the LCS problem, we would enumerate all subsequences of X and check each subsequence to see whether it is also a subsequence of Y , keeping track of the longest subsequence we find. Each subsequence of X corresponds to a subset of the indices f1; 2; : : : ; mg of X . Because X has 2m subsequences, this approach requires exponential time, making it impractical for long sequences. The LCS problem has an optimal-substructure property, however, as the following theorem shows. As we shall see, the natural classes of subproblems correspond to pairs of “prefixes” of the two input sequences. To be precise, given a sequence X D hx1; x2; : : : ; xmi, we define the i th prefix of X , for i D 0; 1; : : : ; m, as Xi D hx1; x2; : : : ; xi i. For example, if X D hA; B; C; B; D; A; Bi, then X4 D hA; B; C; Bi and X0 is the empty sequence. Theorem 15.1 (Optimal substructure of an LCS) Let X D hx1; x2; : : : ; xmi and Y D hy1; y2; : : : ; yni be sequences, and let Z D h´1; ´2; : : : ; ´ki be any LCS of X and Y . 1. If xm D yn, then ´k D xm D yn and Zk 1 is an LCS of Xm 1 and Yn 1. 2. If xm ¤ yn, then ´k ¤ xm implies that Z is an LCS of Xm 1 and Y . 3. If xm ¤ yn, then ´k ¤ yn implies that Z is an LCS of X and Yn 1. Proof (1) If ´k ¤ xm, then we could append xm D yn to Z to obtain a common subsequence of X and Y of length k C 1, contradicting the supposition that Z is a longest common subsequence of X and Y . Thus, we must have ´k D xm D yn. Now, the prefix Zk 1 is a length-.k 1/ common subsequence of Xm 1 and Yn 1. We wish to show that it is an LCS. Suppose for the purpose of contradiction that there exists a common subsequence W of Xm 1 and Yn 1 with length greater than k 1. Then, appending xm D yn to W produces a common subsequence of X and Y whose length is greater than k, which is a contradiction. (2) If ´k ¤ xm, then Z is a common subsequence of Xm 1 and Y . If there were a common subsequence W of Xm 1 and Y with length greater than k, then W would also be a common subsequence of Xm and Y , contradicting the assumption that Z is an LCS of X and Y . (3) The proof is symmetric to (2). The way that Theorem 15.1 characterizes longest common subsequences tells us that an LCS of two sequences contains within it an LCS of prefixes of the two sequences. Thus, the LCS problem has an optimal-substructure property. A recur- 15.4 Longest common subsequence 393 sive solution also has the overlapping-subproblems property, as we shall see in a moment. Step 2: A recursive solution Theorem 15.1 implies that we should examine either one or two subproblems when finding an LCS of X D hx1; x2; : : : ; xmi and Y D hy1; y2; : : : ; yni. If xm D yn, we must find an LCS of Xm 1 and Yn 1. Appending xm D yn to this LCS yields an LCS of X and Y . If xm ¤ yn, then we must solve two subproblems: finding an LCS of Xm 1 and Y and finding an LCS of X and Yn 1. Whichever of these two LCSs is longer is an LCS of X and Y . Because these cases exhaust all possibilities, we know that one of the optimal subproblem solutions must appear within an LCS of X and Y . We can readily see the overlapping-subproblems property in the LCS problem. To find an LCS of X and Y , we may need to find the LCSs of X and Yn 1 and of Xm 1 and Y . But each of these subproblems has the subsubproblem of finding an LCS of Xm 1 and Yn 1. Many other subproblems share subsubproblems. As in the matrix-chain multiplication problem, our recursive solution to the LCS problem involves establishing a recurrence for the value of an optimal solution. Let us define cŒi; j  to be the length of an LCS of the sequences Xi and Yj . If either i D 0 or j D 0, one of the sequences has length 0, and so the LCS has € length 0. The optimal substructure of the LCS problem gives the recursive formula 0 if i D 0 or j D 0 ; cŒi; j  D cŒi 1; j 1 C 1 if i; j > 0 and xi D yj ; (15.9) max.cŒi; j 1; cŒi 1; j / if i; j > 0 and xi ¤ yj : Observe that in this recursive formulation, a condition in the problem restricts which subproblems we may consider. When xi D yj , we can and should consider the subproblem of finding an LCS of Xi 1 and Yj 1. Otherwise, we instead consider the two subproblems of finding an LCS of Xi and Yj 1 and of Xi 1 and Yj . In the previous dynamic-programming algorithms we have examined—for rod cutting and matrix-chain multiplication—we ruled out no subproblems due to conditions in the problem. Finding an LCS is not the only dynamic-programming algorithm that rules out subproblems based on conditions in the problem. For example, the edit-distance problem (see Problem 15-5) has this characteristic. Step 3: Computing the length of an LCS Based on equation (15.9), we could easily write an exponential-time recursive algorithm to compute the length of an LCS of two sequences. Since the LCS problem 394 Chapter 15 Dynamic Programming has only ‚.mn/ distinct subproblems, however, we can use dynamic programming to compute the solutions bottom up. Procedure LCS-LENGTH takes two sequences X D hx1; x2; : : : ; xmi and Y D hy1; y2; : : : ; yni as inputs. It stores the cŒi; j  values in a table cŒ0 : : m; 0 : : n, and it computes the entries in row-major order. (That is, the procedure fills in the first row of c from left to right, then the second row, and so on.) The procedure also maintains the table bŒ1 : : m; 1 : : n to help us construct an optimal solution. Intuitively, bŒi; j  points to the table entry corresponding to the optimal subproblem solution chosen when computing cŒi; j . The procedure returns the b and c tables; cŒm; n contains the length of an LCS of X and Y . LCS-LENGTH.X; Y / 1 m D X:length 2 n D Y:length 3 let bŒ1 : : m; 1 : : n and cŒ0 : : m; 0 : : n be new tables 4 for i D 1 to m 5 cŒi; 0 D 0 6 for j D 0 to n 7 cŒ0; j  D 0 8 for i D 1 to m 9 for j D 1 to n 10 if xi == yj 11 cŒi; j  D cŒi 1; j 1 C 1 12 bŒi; j  D “-” 13 elseif cŒi 1; j  cŒi; j 1 14 cŒi; j  D cŒi 1; j  15 bŒi; j  D “"” 16 else cŒi; j  D cŒi; j 1 17 bŒi; j  D “ ” 18 return c and b Figure 15.8 shows the tables produced by LCS-LENGTH on the sequences X D hA; B; C; B; D; A; Bi and Y D hB; D; C; A; B; Ai. The running time of the procedure is ‚.mn/, since each table entry takes ‚.1/ time to compute. Step 4: Constructing an LCS The b table returned by LCS-LENGTH enables us to quickly construct an LCS of X D hx1; x2; : : : ; xmi and Y D hy1; y2; : : : ; yni. We simply begin at bŒm; n and trace through the table by following the arrows. Whenever we encounter a “-” in entry bŒi; j , it implies that xi D yj is an element of the LCS that LCS-LENGTH 15.4 Longest common subsequence 395 j 0123456 i yj B D C A B A 0 xi 0 0 0 0 0 0 0 1A 0 0 0 0 1 1 1 2B 3C 0111122 0112222 4B 0 1 1 2 2 3 3 5D 0 1 2 2 2 3 3 6A 0 1 2 2 3 3 4 7B 0 1 2 2 3 4 4 Figure 15.8 The c and b tables computed by LCS-LENGTH on the sequences X D hA; B; C; B; D; A; Bi and Y D hB; D; C; A; B; Ai. The square in row i and column j contains the value of cŒi; j  and the appropriate arrow for the value of bŒi; j . The entry 4 in cŒ7; 6—the lower right-hand corner of the table—is the length of an LCS hB; C; B; Ai of X and Y . For i; j > 0, entry cŒi; j  depends only on whether xi D yj and the values in entries cŒi 1; j , cŒi; j 1, and cŒi 1; j 1, which are computed before cŒi; j . To reconstruct the elements of an LCS, follow the bŒi; j  arrows from the lower right-hand corner; the sequence is shaded. Each “-” on the shaded sequence corresponds to an entry (highlighted) for which xi D yj is a member of an LCS. found. With this method, we encounter the elements of this LCS in reverse order. The following recursive procedure prints out an LCS of X and Y in the proper, forward order. The initial call is PRINT-LCS.b; X; X:length; Y:length/. PRINT-LCS.b; X; i; j / 1 if i == 0 or j == 0 2 return 3 if bŒi; j  == “-” 4 PRINT-LCS.b; X; i 1; j 1/ 5 print xi 6 elseif bŒi; j  == “"” 7 PRINT-LCS.b; X; i 1; j / 8 else PRINT-LCS.b; X; i; j 1/ For the b table in Figure 15.8, this procedure prints BCBA. The procedure takes time O.m C n/, since it decrements at least one of i and j in each recursive call. 396 Chapter 15 Dynamic Programming Improving the code Once you have developed an algorithm, you will often find that you can improve on the time or space it uses. Some changes can simplify the code and improve constant factors but otherwise yield no asymptotic improvement in performance. Others can yield substantial asymptotic savings in time and space. In the LCS algorithm, for example, we can eliminate the b table altogether. Each cŒi; j  entry depends on only three other c table entries: cŒi 1; j 1, cŒi 1; j , and cŒi; j 1. Given the value of cŒi; j , we can determine in O.1/ time which of these three values was used to compute cŒi; j , without inspecting table b. Thus, we can reconstruct an LCS in O.mCn/ time using a procedure similar to PRINT-LCS. (Exercise 15.4-2 asks you to give the pseudocode.) Although we save ‚.mn/ space by this method, the auxiliary space requirement for computing an LCS does not asymptotically decrease, since we need ‚.mn/ space for the c table anyway. We can, however, reduce the asymptotic space requirements for LCS-LENGTH, since it needs only two rows of table c at a time: the row being computed and the previous row. (In fact, as Exercise 15.4-4 asks you to show, we can use only slightly more than the space for one row of c to compute the length of an LCS.) This improvement works if we need only the length of an LCS; if we need to reconstruct the elements of an LCS, the smaller table does not keep enough information to retrace our steps in O.m C n/ time. Exercises 15.4-1 Determine an LCS of h1; 0; 0; 1; 0; 1; 0; 1i and h0; 1; 0; 1; 1; 0; 1; 1; 0i. 15.4-2 Give pseudocode to reconstruct an LCS from the completed c table and the original sequences X D hx1; x2; : : : ; xmi and Y D hy1; y2; : : : ; yni in O.m C n/ time, without using the b table. 15.4-3 Give a memoized version of LCS-LENGTH that runs in O.mn/ time. 15.4-4 Show how to compute the length of an LCS using only 2 min.m; n/ entries in the c table plus O.1/ additional space. Then show how to do the same thing, but using min.m; n/ entries plus O.1/ additional space. 15.5 Optimal binary search trees 397 15.4-5 Give an O.n2/-time algorithm to find the longest monotonically increasing subsequence of a sequence of n numbers. 15.4-6 ? Give an O.n lg n/-time algorithm to find the longest monotonically increasing subsequence of a sequence of n numbers. (Hint: Observe that the last element of a candidate subsequence of length i is at least as large as the last element of a candidate subsequence of length i 1. Maintain candidate subsequences by linking them through the input sequence.) 15.5 Optimal binary search trees Suppose that we are designing a program to translate text from English to French. For each occurrence of each English word in the text, we need to look up its French equivalent. We could perform these lookup operations by building a binary search tree with n English words as keys and their French equivalents as satellite data. Because we will search the tree for each individual word in the text, we want the total time spent searching to be as low as possible. We could ensure an O.lg n/ search time per occurrence by using a red-black tree or any other balanced binary search tree. Words appear with different frequencies, however, and a frequently used word such as the may appear far from the root while a rarely used word such as machicolation appears near the root. Such an organization would slow down the translation, since the number of nodes visited when searching for a key in a binary search tree equals one plus the depth of the node containing the key. We want words that occur frequently in the text to be placed nearer the root.6 Moreover, some words in the text might have no French translation,7 and such words would not appear in the binary search tree at all. How do we organize a binary search tree so as to minimize the number of nodes visited in all searches, given that we know how often each word occurs? What we need is known as an optimal binary search tree. Formally, we are given a sequence K D hk1; k2; : : : ; kni of n distinct keys in sorted order (so that k1 < k2 < < kn), and we wish to build a binary search tree from these keys. For each key ki , we have a probability pi that a search will be for ki . Some searches may be for values not in K, and so we also have n C 1 “dummy keys” 6If the subject of the text is castle architecture, we might want machicolation to appear near the root. 7Yes, machicolation has a French counterpart: maˆchicoulis. 398 Chapter 15 Dynamic Programming k2 k1 k4 d0 d1 k3 k5 d2 d3 d4 d5 (a) k2 k1 k5 d0 d1 k4 d5 k3 d4 d2 d3 (b) Figure 15.9 Two binary search trees for a set of n D 5 keys with the following probabilities: i0 1 2 3 4 5 pi 0.15 0.10 0.05 0.10 0.20 qi 0.05 0.10 0.05 0.05 0.05 0.10 (a) A binary search tree with expected search cost 2.80. (b) A binary search tree with expected search cost 2.75. This tree is optimal. d0; d1; d2; : : : ; dn representing values not in K. In particular, d0 represents all values less than k1, dn represents all values greater than kn, and for i D 1; 2; : : : ; n 1, the dummy key di represents all values between ki and kiC1. For each dummy key di , we have a probability qi that a search will correspond to di . Figure 15.9 shows two binary search trees for a set of n D 5 keys. Each key ki is an internal node, and each dummy key di is a leaf. Every search is either successful (finding some key ki ) or unsuccessful (finding some dummy key di ), and so we have Xn Xn pi C qi D 1 : (15.10) i D1 i D0 Because we have probabilities of searches for each key and each dummy key, we can determine the expected cost of a search in a given binary search tree T . Let us assume that the actual cost of a search equals the number of nodes examined, i.e., the depth of the node found by the search in T , plus 1. Then the expected cost of a search in T is Xn Xn E Œsearch cost in T  D .depthT .ki / C 1/ pi C .depthT .di / C 1/ qi i D1 i D0 Xn Xn D 1 C depthT .ki / pi C depthT .di / qi ; (15.11) i D1 i D0 15.5 Optimal binary search trees 399 where depthT denotes a node’s depth in the tree T . The last equality follows from equation (15.10). In Figure 15.9(a), we can calculate the expected search cost node by node: node k1 k2 k3 k4 k5 d0 d1 d2 d3 d4 d5 Total depth 1 0 2 1 2 2 2 3 3 3 3 probability 0.15 0.10 0.05 0.10 0.20 0.05 0.10 0.05 0.05 0.05 0.10 contribution 0.30 0.10 0.15 0.20 0.60 0.15 0.30 0.20 0.20 0.20 0.40 2.80 For a given set of probabilities, we wish to construct a binary search tree whose expected search cost is smallest. We call such a tree an optimal binary search tree. Figure 15.9(b) shows an optimal binary search tree for the probabilities given in the figure caption; its expected cost is 2.75. This example shows that an optimal binary search tree is not necessarily a tree whose overall height is smallest. Nor can we necessarily construct an optimal binary search tree by always putting the key with the greatest probability at the root. Here, key k5 has the greatest search probability of any key, yet the root of the optimal binary search tree shown is k2. (The lowest expected cost of any binary search tree with k5 at the root is 2.85.) As with matrix-chain multiplication, exhaustive checking of all possibilities fails to yield an efficient algorithm. We can label the nodes of any n-node binary tree with the keys k1; k2; : : : ; kn to construct a binary search tree, and then add in the dummy keys as leaves. In Problem 12-4, we saw that the number of binary trees with n nodes is .4n=n3=2/, and so we would have to examine an exponential number of binary search trees in an exhaustive search. Not surprisingly, we shall solve this problem with dynamic programming. Step 1: The structure of an optimal binary search tree To characterize the optimal substructure of optimal binary search trees, we start with an observation about subtrees. Consider any subtree of a binary search tree. It must contain keys in a contiguous range ki ; : : : ; kj , for some 1 Ä i Ä j Ä n. In addition, a subtree that contains keys ki ; : : : ; kj must also have as its leaves the dummy keys di 1; : : : ; dj . Now we can state the optimal substructure: if an optimal binary search tree T has a subtree T 0 containing keys ki ; : : : ; kj , then this subtree T 0 must be optimal as 400 Chapter 15 Dynamic Programming well for the subproblem with keys ki ; : : : ; kj and dummy keys di 1; : : : ; dj . The usual cut-and-paste argument applies. If there were a subtree T 00 whose expected cost is lower than that of T 0, then we could cut T 0 out of T and paste in T 00, resulting in a binary search tree of lower expected cost than T , thus contradicting the optimality of T . We need to use the optimal substructure to show that we can construct an optimal solution to the problem from optimal solutions to subproblems. Given keys ki ; : : : ; kj , one of these keys, say kr (i Ä r Ä j ), is the root of an optimal subtree containing these keys. The left subtree of the root kr contains the keys ki ; : : : ; kr 1 (and dummy keys di 1; : : : ; dr 1), and the right subtree contains the keys krC1; : : : ; kj (and dummy keys dr ; : : : ; dj ). As long as we examine all candidate roots kr , where i Ä r Ä j , and we determine all optimal binary search trees containing ki ; : : : ; kr 1 and those containing krC1; : : : ; kj , we are guaranteed that we will find an optimal binary search tree. There is one detail worth noting about “empty” subtrees. Suppose that in a subtree with keys ki ; : : : ; kj , we select ki as the root. By the above argument, ki ’s left subtree contains the keys ki ; : : : ; ki 1. We interpret this sequence as containing no keys. Bear in mind, however, that subtrees also contain dummy keys. We adopt the convention that a subtree containing keys ki ; : : : ; ki 1 has no actual keys but does contain the single dummy key di 1. Symmetrically, if we select kj as the root, then kj ’s right subtree contains the keys kjC1; : : : ; kj ; this right subtree contains no actual keys, but it does contain the dummy key dj . Step 2: A recursive solution We are ready to define the value of an optimal solution recursively. We pick our subproblem domain as finding an optimal binary search tree containing the keys ki ; : : : ; kj , where i 1, j Ä n, and j i 1. (When j D i 1, there are no actual keys; we have just the dummy key di 1.) Let us define eŒi; j  as the expected cost of searching an optimal binary search tree containing the keys ki ; : : : ; kj . Ultimately, we wish to compute eŒ1; n. The easy case occurs when j D i 1. Then we have just the dummy key di 1. The expected search cost is eŒi; i 1 D qi 1. When j i , we need to select a root kr from among ki ; : : : ; kj and then make an optimal binary search tree with keys ki ; : : : ; kr 1 as its left subtree and an optimal binary search tree with keys krC1; : : : ; kj as its right subtree. What happens to the expected search cost of a subtree when it becomes a subtree of a node? The depth of each node in the subtree increases by 1. By equation (15.11), the expected search cost of this subtree increases by the sum of all the probabilities in the subtree. For a subtree with keys ki ; : : : ; kj , let us denote this sum of probabilities as 15.5 Optimal binary search trees 401 Xj Xj w.i; j / D pl C ql : l Di lDi 1 (15.12) Thus, if kr is the root of an optimal subtree containing keys ki ; : : : ; kj , we have eŒi; j  D pr C .eŒi; r 1 C w.i; r 1// C .eŒr C 1; j  C w.r C 1; j // : Noting that w.i; j / D w.i; r 1/ C pr C w.r C 1; j / ; we rewrite eŒi; j  as eŒi; j  D eŒi; r 1 C eŒr C 1; j  C w.i; j / : (15.13) The recursive equation (15.13) assumes that we know which node kr to use as the root. We choose the root that gives the lowest expected search cost, giving us our final recursive formulation: ( qi 1 if j D i 1 ; eŒi; j  D min feŒi; r 1 C eŒr C 1; j  C w.i; j /g if i Ä j : i ÄrÄj (15.14) The eŒi; j  values give the expected search costs in optimal binary search trees. To help us keep track of the structure of optimal binary search trees, we define rootŒi; j , for 1 Ä i Ä j Ä n, to be the index r for which kr is the root of an optimal binary search tree containing keys ki ; : : : ; kj . Although we will see how to compute the values of rootŒi; j , we leave the construction of an optimal binary search tree from these values as Exercise 15.5-1. Step 3: Computing the expected search cost of an optimal binary search tree At this point, you may have noticed some similarities between our characterizations of optimal binary search trees and matrix-chain multiplication. For both problem domains, our subproblems consist of contiguous index subranges. A direct, recursive implementation of equation (15.14) would be as inefficient as a direct, recursive matrix-chain multiplication algorithm. Instead, we store the eŒi; j  values in a table eŒ1 : : n C 1; 0 : : n. The first index needs to run to n C 1 rather than n because in order to have a subtree containing only the dummy key dn, we need to compute and store eŒn C 1; n. The second index needs to start from 0 because in order to have a subtree containing only the dummy key d0, we need to compute and store eŒ1; 0. We use only the entries eŒi; j  for which j i 1. We also use a table rootŒi; j , for recording the root of the subtree containing keys ki ; : : : ; kj . This table uses only the entries for which 1 Ä i Ä j Ä n. We will need one other table for efficiency. Rather than compute the value of w.i; j / from scratch every time we are computing eŒi; j —which would take 402 Chapter 15 Dynamic Programming ‚.j i/ additions—we store these values in a table wŒ1 : : n C 1; 0 : : n. For the base case, we compute wŒi; i 1 D qi 1 for 1 Ä i Ä n C 1. For j i , we compute wŒi; j  D wŒi; j 1 C pj C qj : (15.15) Thus, we can compute the ‚.n2/ values of wŒi; j  in ‚.1/ time each. The pseudocode that follows takes as inputs the probabilities p1; : : : ; pn and q0; : : : ; qn and the size n, and it returns the tables e and root. OPTIMAL-BST.p; q; n/ 1 let eŒ1 : : n C 1; 0 : : n, wŒ1 : : n C 1; 0 : : n, and rootŒ1 : : n; 1 : : n be new tables 2 for i D 1 to n C 1 3 eŒi; i 1 D qi 1 4 wŒi; i 1 D qi 1 5 for l D 1 to n 6 for i D 1 to n l C 1 7 j D i Cl 1 8 eŒi; j  D 1 9 wŒi; j  D wŒi; j 1 C pj C qj 10 for r D i to j 11 t D eŒi; r 1 C eŒr C 1; j  C wŒi; j  12 if t < eŒi; j  13 eŒi; j  D t 14 rootŒi; j  D r 15 return e and root From the description above and the similarity to the MATRIX-CHAIN-ORDER procedure in Section 15.2, you should find the operation of this procedure to be fairly straightforward. The for loop of lines 2–4 initializes the values of eŒi; i 1 and wŒi; i 1. The for loop of lines 5–14 then uses the recurrences (15.14) and (15.15) to compute eŒi; j  and wŒi; j  for all 1 Ä i Ä j Ä n. In the first iteration, when l D 1, the loop computes eŒi; i and wŒi; i for i D 1; 2; : : : ; n. The second iteration, with l D 2, computes eŒi; i C1 and wŒi; i C1 for i D 1; 2; : : : ; n 1, and so forth. The innermost for loop, in lines 10–14, tries each candidate index r to determine which key kr to use as the root of an optimal binary search tree containing keys ki ; : : : ; kj . This for loop saves the current value of the index r in rootŒi; j  whenever it finds a better key to use as the root. Figure 15.10 shows the tables eŒi; j , wŒi; j , and rootŒi; j  computed by the procedure OPTIMAL-BST on the key distribution shown in Figure 15.9. As in the matrix-chain multiplication example of Figure 15.5, the tables are rotated to make 15.5 Optimal binary search trees e 51 j 4 2.75 2 i 3 1.75 2.00 3 2 1.25 1.20 1.30 4 1 0.90 0.70 0.60 0.90 5 0 0.45 0.40 0.25 0.30 0.50 6 0.05 0.10 0.05 0.05 0.05 0.10 403 w 51 j 4 1.00 2 i 3 0.70 0.80 3 2 0.55 0.50 0.60 4 1 0.45 0.35 0.30 0.50 5 0 0.30 0.25 0.15 0.20 0.35 6 0.05 0.10 0.05 0.05 0.05 0.10 root 51 j 422 i 32 43 22 2 54 11 2 4 55 12345 Figure 15.10 The tables eŒi; j , wŒi; j , and rootŒi; j  computed by OPTIMAL-BST on the key distribution shown in Figure 15.9. The tables are rotated so that the diagonals run horizontally. the diagonals run horizontally. OPTIMAL-BST computes the rows from bottom to top and from left to right within each row. The OPTIMAL-BST procedure takes ‚.n3/ time, just like MATRIX-CHAINORDER. We can easily see that its running time is O.n3/, since its for loops are nested three deep and each loop index takes on at most n values. The loop indices in OPTIMAL-BST do not have exactly the same bounds as those in MATRIX-CHAINORDER, but they are within at most 1 in all directions. Thus, like MATRIX-CHAINORDER, the OPTIMAL-BST procedure takes .n3/ time. Exercises 15.5-1 Write pseudocode for the procedure CONSTRUCT-OPTIMAL-BST.root/ which, given the table root, outputs the structure of an optimal binary search tree. For the example in Figure 15.10, your procedure should print out the structure 404 Chapter 15 Dynamic Programming k2 is the root k1 is the left child of k2 d0 is the left child of k1 d1 is the right child of k1 k5 is the right child of k2 k4 is the left child of k5 k3 is the left child of k4 d2 is the left child of k3 d3 is the right child of k3 d4 is the right child of k4 d5 is the right child of k5 corresponding to the optimal binary search tree shown in Figure 15.9(b). 15.5-2 Determine the cost and structure of an optimal binary search tree for a set of n D 7 keys with the following probabilities: i0 1 2 3 4 5 6 7 pi 0.04 0.06 0.08 0.02 0.10 0.12 0.14 qi 0.06 0.06 0.06 0.06 0.05 0.05 0.05 0.05 15.5-3 Suppose that instead of maintaining the table wŒi; j , we computed the value of w.i; j / directly from equation (15.12) in line 9 of OPTIMAL-BST and used this computed value in line 11. How would this change affect the asymptotic running time of OPTIMAL-BST? 15.5-4 ? Knuth [212] has shown that there are always roots of optimal subtrees such that rootŒi; j 1 Ä rootŒi; j  Ä rootŒi C 1; j  for all 1 Ä i < j Ä n. Use this fact to modify the OPTIMAL-BST procedure to run in ‚.n2/ time. Problems 15-1 Longest simple path in a directed acyclic graph Suppose that we are given a directed acyclic graph G D .V; E/ with realvalued edge weights and two distinguished vertices s and t. Describe a dynamicprogramming approach for finding a longest weighted simple path from s to t. What does the subproblem graph look like? What is the efficiency of your algorithm? Problems for Chapter 15 405 (a) (b) Figure 15.11 Seven points in the plane, shown on a unit grid. (a) The shortest closed tour, with length approximately 24:89. This tour is not bitonic. (b) The shortest bitonic tour for the same set of points. Its length is approximately 25:58. 15-2 Longest palindrome subsequence A palindrome is a nonempty string over some alphabet that reads the same forward and backward. Examples of palindromes are all strings of length 1, civic, racecar, and aibohphobia (fear of palindromes). Give an efficient algorithm to find the longest palindrome that is a subsequence of a given input string. For example, given the input character, your algorithm should return carac. What is the running time of your algorithm? 15-3 Bitonic euclidean traveling-salesman problem In the euclidean traveling-salesman problem, we are given a set of n points in the plane, and we wish to find the shortest closed tour that connects all n points. Figure 15.11(a) shows the solution to a 7-point problem. The general problem is NP-hard, and its solution is therefore believed to require more than polynomial time (see Chapter 34). J. L. Bentley has suggested that we simplify the problem by restricting our attention to bitonic tours, that is, tours that start at the leftmost point, go strictly rightward to the rightmost point, and then go strictly leftward back to the starting point. Figure 15.11(b) shows the shortest bitonic tour of the same 7 points. In this case, a polynomial-time algorithm is possible. Describe an O.n2/-time algorithm for determining an optimal bitonic tour. You may assume that no two points have the same x-coordinate and that all operations on real numbers take unit time. (Hint: Scan left to right, maintaining optimal possibilities for the two parts of the tour.) 15-4 Printing neatly Consider the problem of neatly printing a paragraph with a monospaced font (all characters having the same width) on a printer. The input text is a sequence of n 406 Chapter 15 Dynamic Programming words of lengths l1; l2; : : : ; ln, measured in characters. We want to print this para- graph neatly on a number of lines that hold a maximum of M characters each. Our criterion of “neatness” is as follows. If a given line contains words i through j , where space i Ä j , and we leave characters at the end exactly one of the line space is M between j Ci woPrdjksD, tihlek number , which of extra must be nonnegative so that the words fit on the line. We wish to minimize the sum, over all lines except the last, of the cubes of the numbers of extra space characters at the ends of lines. Give a dynamic-programming algorithm to print a paragraph of n words neatly on a printer. Analyze the running time and space requirements of your algorithm. 15-5 Edit distance In order to transform one source string of text xŒ1 : : m to a target string yŒ1 : : n, we can perform various transformation operations. Our goal is, given x and y, to produce a series of transformations that change x to y. We use an array ´—assumed to be large enough to hold all the characters it will need—to hold the intermediate results. Initially, ´ is empty, and at termination, we should have ´Œj  D yŒj  for j D 1; 2; : : : ; n. We maintain current indices i into x and j into ´, and the operations are allowed to alter ´ and these indices. Initially, i D j D 1. We are required to examine every character in x during the transformation, which means that at the end of the sequence of transformation operations, we must have i D m C 1. We may choose from among six transformation operations: Copy a character from x to ´ by setting ´Œj  D xŒi and then incrementing both i and j . This operation examines xŒi. Replace a character from x by another character c, by setting ´Œj  D c, and then incrementing both i and j . This operation examines xŒi. Delete a character from x by incrementing i but leaving j alone. This operation examines xŒi. Insert the character c into ´ by setting ´Œj  D c and then incrementing j , but leaving i alone. This operation examines no characters of x. Twiddle (i.e., exchange) the next two characters by copying them from x to ´ but in the opposite order; we do so by setting ´Œj  D xŒi C 1 and ´Œj C 1 D xŒi and then setting i D i C 2 and j D j C 2. This operation examines xŒi and xŒi C 1. Kill the remainder of x by setting i D m C 1. This operation examines all characters in x that have not yet been examined. This operation, if performed, must be the final operation. Problems for Chapter 15 407 As an example, one way to transform the source string algorithm to the target string altruistic is to use the following sequence of operations, where the underlined characters are xŒi and ´Œj  after the operation: Operation initial strings copy copy replace by t delete copy insert u insert i insert s twiddle insert c kill x algorithm algorithm algorithm algorithm algorithm algorithm algorithm algorithm algorithm algorithm algorithm algorithm ´ a al alt alt altr altru altrui altruis altruisti altruistic altruistic Note that there are several other sequences of transformation operations that transform algorithm to altruistic. Each of the transformation operations has an associated cost. The cost of an operation depends on the specific application, but we assume that each operation’s cost is a constant that is known to us. We also assume that the individual costs of the copy and replace operations are less than the combined costs of the delete and insert operations; otherwise, the copy and replace operations would not be used. The cost of a given sequence of transformation operations is the sum of the costs of the individual operations in the sequence. For the sequence above, the cost of transforming algorithm to altruistic is .3 cost.copy// C cost.replace/ C cost.delete/ C .4 cost.insert// C cost.twiddle/ C cost.kill/ : a. Given two sequences xŒ1 : : m and yŒ1 : : n and set of transformation-operation costs, the edit distance from x to y is the cost of the least expensive operation sequence that transforms x to y. Describe a dynamic-programming algorithm that finds the edit distance from xŒ1 : : m to yŒ1 : : n and prints an optimal operation sequence. Analyze the running time and space requirements of your algorithm. The edit-distance problem generalizes the problem of aligning two DNA sequences (see, for example, Setubal and Meidanis [310, Section 3.2]). There are several methods for measuring the similarity of two DNA sequences by aligning them. One such method to align two sequences x and y consists of inserting spaces at 408 Chapter 15 Dynamic Programming arbitrary locations in the two sequences (including at either end) so that the resulting sequences x0 and y0 have the same length but do not have a space in the same position (i.e., for no position j are both x0Œj  and y0Œj  a space). Then we assign a “score” to each position. Position j receives a score as follows: C1 if x0Œj  D y0Œj  and neither is a space, 1 if x0Œj  ¤ y0Œj  and neither is a space, 2 if either x0Œj  or y0Œj  is a space. The score for the alignment is the sum of the scores of the individual positions. For example, given the sequences x D GATCGGCAT and y D CAATGTGAATC, one alignment is G ATCG GCAT CAAT GTGAATC -*++*+*+-++* A + under a position indicates a score of C1 for that position, a - indicates a score of 1, and a * indicates a score of 2, so that this alignment has a total score of 6 1 2 1 4 2 D 4. b. Explain how to cast the problem of finding an optimal alignment as an edit distance problem using a subset of the transformation operations copy, replace, delete, insert, twiddle, and kill. 15-6 Planning a company party Professor Stewart is consulting for the president of a corporation that is planning a company party. The company has a hierarchical structure; that is, the supervisor relation forms a tree rooted at the president. The personnel office has ranked each employee with a conviviality rating, which is a real number. In order to make the party fun for all attendees, the president does not want both an employee and his or her immediate supervisor to attend. Professor Stewart is given the tree that describes the structure of the corporation, using the left-child, right-sibling representation described in Section 10.4. Each node of the tree holds, in addition to the pointers, the name of an employee and that employee’s conviviality ranking. Describe an algorithm to make up a guest list that maximizes the sum of the conviviality ratings of the guests. Analyze the running time of your algorithm. 15-7 Viterbi algorithm We can use dynamic programming on a directed graph G D .V; E/ for speech recognition. Each edge .u; / 2 E is labeled with a sound .u; / from a finite set † of sounds. The labeled graph is a formal model of a person speaking Problems for Chapter 15 409 a restricted language. Each path in the graph starting from a distinguished vertex 0 2 V corresponds to a possible sequence of sounds produced by the model. We define the label of a directed path to be the concatenation of the labels of the edges on that path. a. Describe an efficient algorithm that, given an edge-labeled graph G with distinguished vertex 0 and a sequence s D h 1; 2; : : : ; ki of sounds from †, returns a path in G that begins at 0 and has s as its label, if any such path exists. Otherwise, the algorithm should return NO-SUCH-PATH. Analyze the running time of your algorithm. (Hint: You may find concepts from Chapter 22 useful.) Now, suppose that every edge .u; / 2 E has an associated nonnegative probability p.u; / of traversing the edge .u; / from vertex u and thus producing the corresponding sound. The sum of the probabilities of the edges leaving any vertex equals 1. The probability of a path is defined to be the product of the probabilities of its edges. We can view the probability of a path beginning at 0 as the probability that a “random walk” beginning at 0 will follow the specified path, where we randomly choose which edge to take leaving a vertex u according to the probabilities of the available edges leaving u. b. Extend your answer to part (a) so that if a path is returned, it is a most probable path starting at 0 and having label s. Analyze the running time of your algorithm. 15-8 Image compression by seam carving We are given a color picture consisting of an m n array AŒ1 : : m; 1 : : n of pixels, where each pixel specifies a triple of red, green, and blue (RGB) intensities. Suppose that we wish to compress this picture slightly. Specifically, we wish to remove one pixel from each of the m rows, so that the whole picture becomes one pixel narrower. To avoid disturbing visual effects, however, we require that the pixels removed in two adjacent rows be in the same or adjacent columns; the pixels removed form a “seam” from the top row to the bottom row where successive pixels in the seam are adjacent vertically or diagonally. a. Show that the number of such possible seams grows at least exponentially in m, assuming that n > 1. b. Suppose now that along with each pixel AŒi; j , we have calculated a realvalued disruption measure d Œi; j , indicating how disruptive it would be to remove pixel AŒi; j . Intuitively, the lower a pixel’s disruption measure, the more similar the pixel is to its neighbors. Suppose further that we define the disruption measure of a seam to be the sum of the disruption measures of its pixels. 410 Chapter 15 Dynamic Programming Give an algorithm to find a seam with the lowest disruption measure. How efficient is your algorithm? 15-9 Breaking a string A certain string-processing language allows a programmer to break a string into two pieces. Because this operation copies the string, it costs n time units to break a string of n characters into two pieces. Suppose a programmer wants to break a string into many pieces. The order in which the breaks occur can affect the total amount of time used. For example, suppose that the programmer wants to break a 20-character string after characters 2, 8, and 10 (numbering the characters in ascending order from the left-hand end, starting from 1). If she programs the breaks to occur in left-to-right order, then the first break costs 20 time units, the second break costs 18 time units (breaking the string from characters 3 to 20 at character 8), and the third break costs 12 time units, totaling 50 time units. If she programs the breaks to occur in right-to-left order, however, then the first break costs 20 time units, the second break costs 10 time units, and the third break costs 8 time units, totaling 38 time units. In yet another order, she could break first at 8 (costing 20), then break the left piece at 2 (costing 8), and finally the right piece at 10 (costing 12), for a total cost of 40. Design an algorithm that, given the numbers of characters after which to break, determines a least-cost way to sequence those breaks. More formally, given a string S with n characters and an array LŒ1 : : m containing the break points, compute the lowest cost for a sequence of breaks, along with a sequence of breaks that achieves this cost. 15-10 Planning an investment strategy Your knowledge of algorithms helps you obtain an exciting job with the Acme Computer Company, along with a $10,000 signing bonus. You decide to invest this money with the goal of maximizing your return at the end of 10 years. You decide to use the Amalgamated Investment Company to manage your investments. Amalgamated Investments requires you to observe the following rules. It offers n different investments, numbered 1 through n. In each year j , investment i provides a return rate of rij . In other words, if you invest d dollars in investment i in year j , then at the end of year j , you have drij dollars. The return rates are guaranteed, that is, you are given all the return rates for the next 10 years for each investment. You make investment decisions only once per year. At the end of each year, you can leave the money made in the previous year in the same investments, or you can shift money to other investments, by either shifting money between existing investments or moving money to a new investement. If you do not move your money between two consecutive years, you pay a fee of f1 dollars, whereas if you switch your money, you pay a fee of f2 dollars, where f2 > f1. Problems for Chapter 15 411 a. The problem, as stated, allows you to invest your money in multiple investments in each year. Prove that there exists an optimal investment strategy that, in each year, puts all the money into a single investment. (Recall that an optimal investment strategy maximizes the amount of money after 10 years and is not concerned with any other objectives, such as minimizing risk.) b. Prove that the problem of planning your optimal investment strategy exhibits optimal substructure. c. Design an algorithm that plans your optimal investment strategy. What is the running time of your algorithm? d. Suppose that Amalgamated Investments imposed the additional restriction that, at any point, you can have no more than $15,000 in any one investment. Show that the problem of maximizing your income at the end of 10 years no longer exhibits optimal substructure. 15-11 Inventory planning The Rinky Dink Company makes machines that resurface ice rinks. The demand for such products varies from month to month, and so the company needs to de- velop a strategy to plan its manufacturing given the fluctuating, but predictable, demand. The company wishes to design a plan for the next n months. For each month it will i, the company sell. Let D D PknnioDw1sdtihbeedtehmeatnodtadl id,etmhaatnids,otvheerntuhme bneerxtofnmmaocnhtihnse.s that The company keeps a full-time staff who provide labor to manufacture up to m ma- chines per month. If the company needs to make more than m machines in a given month, it can hire additional, part-time labor, at a cost that works out to c dollars per machine. Furthermore, if, at the end of a month, the company is holding any unsold machines, it must pay inventory costs. The cost for holding j machines is given as a function h.j / for j D 1; 2; : : : ; D, where h.j / 0 for 1 Ä j Ä D and h.j / Ä h.j C 1/ for 1 Ä j Ä D 1. Give an algorithm that calculates a plan for the company that minimizes its costs while fulfilling all the demand. The running time should be polyomial in n and D. 15-12 Signing free-agent baseball players Suppose that you are the general manager for a major-league baseball team. During the off-season, you need to sign some free-agent players for your team. The team owner has given you a budget of $X to spend on free agents. You are allowed to spend less than $X altogether, but the owner will fire you if you spend any more than $X . 412 Chapter 15 Dynamic Programming You are considering N different positions, and for each position, P free-agent players who play that position are available.8 Because you do not want to overload your roster with too many players at any position, for each position you may sign at most one free agent who plays that position. (If you do not sign any players at a particular position, then you plan to stick with the players you already have at that position.) To determine how valuable a player is going to be, you decide to use a sabermetric statistic9 known as “VORP,” or “value over replacement player.” A player with a higher VORP is more valuable than a player with a lower VORP. A player with a higher VORP is not necessarily more expensive to sign than a player with a lower VORP, because factors other than a player’s value determine how much it costs to sign him. For each available free-agent player, you have three pieces of information: the player’s position, the amount of money it will cost to sign the player, and the player’s VORP. Devise an algorithm that maximizes the total VORP of the players you sign while spending no more than $X altogether. You may assume that each player signs for a multiple of $100,000. Your algorithm should output the total VORP of the players you sign, the total amount of money you spend, and a list of which players you sign. Analyze the running time and space requirement of your algorithm. Chapter notes R. Bellman began the systematic study of dynamic programming in 1955. The word “programming,” both here and in linear programming, refers to using a tabular solution method. Although optimization techniques incorporating elements of dynamic programming were known earlier, Bellman provided the area with a solid mathematical basis [37]. 8Although there are nine positions on a baseball team, N is not necesarily equal to 9 because some general managers have particular ways of thinking about positions. For example, a general manager might consider right-handed pitchers and left-handed pitchers to be separate “positions,” as well as starting pitchers, long relief pitchers (relief pitchers who can pitch several innings), and short relief pitchers (relief pitchers who normally pitch at most only one inning). 9Sabermetrics is the application of statistical analysis to baseball records. It provides several ways to compare the relative values of individual players. Notes for Chapter 15 413 Galil and Park [125] classify dynamic-programming algorithms according to the size of the table and the number of other table entries each entry depends on. They call a dynamic-programming algorithm tD=eD if its table size is O.nt / and each entry depends on O.ne/ other entries. For example, the matrix-chain multiplication algorithm in Section 15.2 would be 2D=1D, and the longest-common-subsequence algorithm in Section 15.4 would be 2D=0D. Hu and Shing [182, 183] give an O.n lg n/-time algorithm for the matrix-chain multiplication problem. The O.mn/-time algorithm for the longest-common-subsequence problem appears to be a folk algorithm. Knuth [70] posed the question of whether subquadratic algorithms for the LCS problem exist. Masek and Paterson [244] answered this question in the affirmative by giving an algorithm that runs in O.mn= lg n/ time, where n Ä m and the sequences are drawn from a set of bounded size. For the special case in which no element appears more than once in an input sequence, Szymanski [326] shows how to solve the problem in O..n C m/ lg.n C m// time. Many of these results extend to the problem of computing string edit distances (Problem 15-5). An early paper on variable-length binary encodings by Gilbert and Moore [133] had applications to constructing optimal binary search trees for the case in which all probabilities pi are 0; this paper contains an O.n3/-time algorithm. Aho, Hopcroft, and Ullman [5] present the algorithm from Section 15.5. Exercise 15.5-4 is due to Knuth [212]. Hu and Tucker [184] devised an algorithm for the case in which all probabilities pi are 0 that uses O.n2/ time and O.n/ space; subsequently, Knuth [211] reduced the time to O.n lg n/. Problem 15-8 is due to Avidan and Shamir [27], who have posted on the Web a wonderful video illustrating this image-compression technique. 16 Greedy Algorithms Algorithms for optimization problems typically go through a sequence of steps, with a set of choices at each step. For many optimization problems, using dynamic programming to determine the best choices is overkill; simpler, more efficient algorithms will do. A greedy algorithm always makes the choice that looks best at the moment. That is, it makes a locally optimal choice in the hope that this choice will lead to a globally optimal solution. This chapter explores optimization problems for which greedy algorithms provide optimal solutions. Before reading this chapter, you should read about dynamic programming in Chapter 15, particularly Section 15.3. Greedy algorithms do not always yield optimal solutions, but for many problems they do. We shall first examine, in Section 16.1, a simple but nontrivial problem, the activity-selection problem, for which a greedy algorithm efficiently computes an optimal solution. We shall arrive at the greedy algorithm by first considering a dynamic-programming approach and then showing that we can always make greedy choices to arrive at an optimal solution. Section 16.2 reviews the basic elements of the greedy approach, giving a direct approach for proving greedy algorithms correct. Section 16.3 presents an important application of greedy techniques: designing data-compression (Huffman) codes. In Section 16.4, we investigate some of the theory underlying combinatorial structures called “matroids,” for which a greedy algorithm always produces an optimal solution. Finally, Section 16.5 applies matroids to solve a problem of scheduling unit-time tasks with deadlines and penalties. The greedy method is quite powerful and works well for a wide range of problems. Later chapters will present many algorithms that we can view as applications of the greedy method, including minimum-spanning-tree algorithms (Chapter 23), Dijkstra’s algorithm for shortest paths from a single source (Chapter 24), and Chva´tal’s greedy set-covering heuristic (Chapter 35). Minimum-spanning-tree algorithms furnish a classic example of the greedy method. Although you can read 16.1 An activity-selection problem 415 this chapter and Chapter 23 independently of each other, you might find it useful to read them together. 16.1 An activity-selection problem Our first example is the problem of scheduling several competing activities that require exclusive use of a common resource, with a goal of selecting a maximum-size set of mutually compatible activities. Suppose we have a set S D fa1; a2; : : : ; ang of n proposed activities that wish to use a resource, such as a lecture hall, which can serve only one activity at a time. Each activity ai has a start time si and a finish time fi , where 0 Ä si < fi < 1. If selected, activity ai takes place during the half-open time interval Œsi ; fi /. Activities ai and aj are compatible if the intervals Œsi ; fi / and Œsj ; fj / do not overlap. That is, ai and aj are compatible if si fj or sj fi . In the activity-selection problem, we wish to select a maximum-size subset of mutually compatible activities. We assume that the activities are sorted in monotonically increasing order of finish time: f1 Ä f2 Ä f3 Ä Ä fn 1 Ä fn : (16.1) (We shall see later the advantage that this assumption provides.) For example, consider the following set S of activities: i 1 2 3 4 5 6 7 8 9 10 11 si 1 3 0 5 3 5 6 8 8 2 12 fi 4 5 6 7 9 9 10 11 12 14 16 For this example, the subset fa3; a9; a11g consists of mutually compatible activities. It is not a maximum subset, however, since the subset fa1; a4; a8; a11g is larger. In fact, fa1; a4; a8; a11g is a largest subset of mutually compatible activities; another largest subset is fa2; a4; a9; a11g. We shall solve this problem in several steps. We start by thinking about a dynamic-programming solution, in which we consider several choices when determining which subproblems to use in an optimal solution. We shall then observe that we need to consider only one choice—the greedy choice—and that when we make the greedy choice, only one subproblem remains. Based on these observations, we shall develop a recursive greedy algorithm to solve the activity-scheduling problem. We shall complete the process of developing a greedy solution by converting the recursive algorithm to an iterative one. Although the steps we shall go through in this section are slightly more involved than is typical when developing a greedy algorithm, they illustrate the relationship between greedy algorithms and dynamic programming. 416 Chapter 16 Greedy Algorithms The optimal substructure of the activity-selection problem We can easily verify that the activity-selection problem exhibits optimal substructure. Let us denote by Sij the set of activities that start after activity ai finishes and that finish before activity aj starts. Suppose that we wish to find a maximum set of mutually compatible activities in Sij , and suppose further that such a maximum set is Aij , which includes some activity ak. By including ak in an optimal solution, we are left with two subproblems: finding mutually compatible activities in the set Sik (activities that start after activity ai finishes and that finish before activity ak starts) and finding mutually compatible activities in the set Skj (activities that start after activity ak finishes and that finish before activity aj starts). Let Aik D Aij \ Sik and Akj D Aij \ Skj , so that Aik contains the activities in Aij that finish before ak starts and Akj contains the activities in Aij that start after ak finishes. Thus, we have Aij D Aik [ fakg [ Akj , and so the maximum-size set Aij of mutually compatible activities in Sij consists of jAij j D jAikj C jAkj j C 1 activities. The usual cut-and-paste argument shows that the optimal solution Aij must also include optimal solutions to the two subproblems for Sik and Skj . If we could find a set A0kj of mutually compatible activities in Skj where jA0kj j > jAkj j, then we could use A0kj , rather than Akj , in a solution to the subproblem for Sij . We would have constructed a set of jAikj C jA0kj j C 1 > jAikj C jAkj j C 1 D jAij j mutually compatible activities, which contradicts the assumption that Aij is an optimal solution. A symmetric argument applies to the activities in Sik. This way of characterizing optimal substructure suggests that we might solve the activity-selection problem by dynamic programming. If we denote the size of an optimal solution for the set Sij by cŒi; j , then we would have the recurrence cŒi; j  D cŒi; k C cŒk; j  C 1 : Of course, if we did not know that an optimal solution for the set Sij includes activity ak, we would have to examine all activities in Sij to find which one to choose, so that ( 0 cŒi; j  D max fcŒi; k C cŒk; j  C 1g ak 2Sij if Sij D ; ; if Sij ¤ ; : (16.2) We could then develop a recursive algorithm and memoize it, or we could work bottom-up and fill in table entries as we go along. But we would be overlooking another important characteristic of the activity-selection problem that we can use to great advantage. 16.1 An activity-selection problem 417 Making the greedy choice What if we could choose an activity to add to our optimal solution without having to first solve all the subproblems? That could save us from having to consider all the choices inherent in recurrence (16.2). In fact, for the activity-selection problem, we need consider only one choice: the greedy choice. What do we mean by the greedy choice for the activity-selection problem? Intuition suggests that we should choose an activity that leaves the resource available for as many other activities as possible. Now, of the activities we end up choosing, one of them must be the first one to finish. Our intuition tells us, therefore, to choose the activity in S with the earliest finish time, since that would leave the resource available for as many of the activities that follow it as possible. (If more than one activity in S has the earliest finish time, then we can choose any such activity.) In other words, since the activities are sorted in monotonically increasing order by finish time, the greedy choice is activity a1. Choosing the first activity to finish is not the only way to think of making a greedy choice for this problem; Exercise 16.1-3 asks you to explore other possibilities. If we make the greedy choice, we have only one remaining subproblem to solve: finding activities that start after a1 finishes. Why don’t we have to consider activities that finish before a1 starts? We have that s1 < f1, and f1 is the earliest finish time of any activity, and therefore no activity can have a finish time less than or equal to s1. Thus, all activities that are compatible with activity a1 must start after a1 finishes. Furthermore, we have already established that the activity-selection problem exhibits optimal substructure. Let Sk D fai 2 S W si fkg be the set of activities that start after activity ak finishes. If we make the greedy choice of activity a1, then S1 remains as the only subproblem to solve.1 Optimal substructure tells us that if a1 is in the optimal solution, then an optimal solution to the original problem consists of activity a1 and all the activities in an optimal solution to the subproblem S1. One big question remains: is our intuition correct? Is the greedy choice—in which we choose the first activity to finish—always part of some optimal solution? The following theorem shows that it is. 1We sometimes refer to the sets Sk as subproblems rather than as just sets of activities. It will always be clear from the context whether we are referring to Sk as a set of activities or as a subproblem whose input is that set. 418 Chapter 16 Greedy Algorithms Theorem 16.1 Consider any nonempty subproblem Sk, and let am be an activity in Sk with the earliest finish time. Then am is included in some maximum-size subset of mutually compatible activities of Sk. Proof Let Ak be a maximum-size subset of mutually compatible activities in Sk, and let aj be the activity in Ak with the earliest finish time. If aj D am, we are done, since we have shown that am is in some maximum-size subset of mutually compatible activities of Sk. If aj ¤ am, let the set A0k D Ak faj g [ famg be Ak but substituting am for aj . The activities in A0k are disjoint, which follows because the activities in Ak are disjoint, aj is the first activity in Ak to finish, and fm Ä fj . Since jA0kj D jAkj, we conclude that A0k is a maximum-size subset of mutually compatible activities of Sk, and it includes am. Thus, we see that although we might be able to solve the activity-selection problem with dynamic programming, we don’t need to. (Besides, we have not yet examined whether the activity-selection problem even has overlapping subproblems.) Instead, we can repeatedly choose the activity that finishes first, keep only the activities compatible with this activity, and repeat until no activities remain. Moreover, because we always choose the activity with the earliest finish time, the finish times of the activities we choose must strictly increase. We can consider each activity just once overall, in monotonically increasing order of finish times. An algorithm to solve the activity-selection problem does not need to work bottom-up, like a table-based dynamic-programming algorithm. Instead, it can work top-down, choosing an activity to put into the optimal solution and then solving the subproblem of choosing activities from those that are compatible with those already chosen. Greedy algorithms typically have this top-down design: make a choice and then solve a subproblem, rather than the bottom-up technique of solving subproblems before making a choice. A recursive greedy algorithm Now that we have seen how to bypass the dynamic-programming approach and instead use a top-down, greedy algorithm, we can write a straightforward, recursive procedure to solve the activity-selection problem. The procedure RECURSIVEACTIVITY-SELECTOR takes the start and finish times of the activities, represented as arrays s and f ,2 the index k that defines the subproblem Sk it is to solve, and 2Because the pseudocode takes s and f as arrays, it indexes into them with square brackets rather than subscripts. 16.1 An activity-selection problem 419 the size n of the original problem. It returns a maximum-size set of mutually compatible activities in Sk. We assume that the n input activities are already ordered by monotonically increasing finish time, according to equation (16.1). If not, we can sort them into this order in O.n lg n/ time, breaking ties arbitrarily. In order to start, we add the fictitious activity a0 with f0 D 0, so that subproblem S0 is the entire set of activities S. The initial call, which solves the entire problem, is RECURSIVE-ACTIVITY-SELECTOR .s; f; 0; n/. RECURSIVE-ACTIVITY-SELECTOR .s; f; k; n/ 1 m D kC1 2 while m Ä n and sŒm < f Œk // find the first activity in Sk to finish 3 m D mC1 4 if m Ä n 5 return famg [ RECURSIVE-ACTIVITY-SELECTOR .s; f; m; n/ 6 else return ; Figure 16.1 shows the operation of the algorithm. In a given recursive call RECURSIVE-ACTIVITY-SELECTOR .s; f; k; n/, the while loop of lines 2–3 looks for the first activity in Sk to finish. The loop examines akC1; akC2; : : : ; an, until it finds the first activity am that is compatible with ak; such an activity has sm fk. If the loop terminates because it finds such an activity, line 5 returns the union of famg and the maximum-size subset of Sm returned by the recursive call RECURSIVE-ACTIVITY-SELECTOR.s; f; m; n/. Alternatively, the loop may terminate because m > n, in which case we have examined all activities in Sk without finding one that is compatible with ak. In this case, Sk D ;, and so the procedure returns ; in line 6. Assuming that the activities have already been sorted by finish times, the running time of the call RECURSIVE-ACTIVITY-SELECTOR .s; f; 0; n/ is ‚.n/, which we can see as follows. Over all recursive calls, each activity is examined exactly once in the while loop test of line 2. In particular, activity ai is examined in the last call made in which k < i. An iterative greedy algorithm We easily can convert our recursive procedure to an iterative one. The procedure RECURSIVE-ACTIVITY-SELECTOR is almost “tail recursive” (see Problem 7-4): it ends with a recursive call to itself followed by a union operation. It is usually a straightforward task to transform a tail-recursive procedure to an iterative form; in fact, some compilers for certain programming languages perform this task automatically. As written, RECURSIVE-ACTIVITY-SELECTOR works for subproblems Sk, i.e., subproblems that consist of the last activities to finish. 420 Chapter 16 Greedy Algorithms k sk fk 0–0 114 235 306 457 539 659 7 6 10 8 8 11 9 8 12 10 2 14 11 12 16 a0 a1 a0 m=1 RECURSIVE-ACTIVITY-SELECTOR(s, f, 0, 11) a2 a1 RECURSIVE-ACTIVITY-SELECTOR(s, f, 1, 11) a3 a1 a4 a1 m=4 RECURSIVE-ACTIVITY-SELECTOR(s, f, 4, 11) a5 a1 a4 a6 a1 a4 a7 a1 a4 a8 a1 a4 m=8 RECURSIVE-ACTIVITY-SELECTOR(s, f, 8, 11) a1 a4 a9 a8 a10 a1 a4 a8 a1 a4 a8 a11 m = 11 RECURSIVE-ACTIVITY-SELECTOR(s, f, 11, 11) a1 a4 a8 a11 time 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Figure 16.1 The operation of RECURSIVE-ACTIVITY-SELECTOR on the 11 activities given earlier. Activities considered in each recursive call appear between horizontal lines. The fictitious activity a0 finishes at time 0, and the initial call RECURSIVE-ACTIVITY-SELECTOR.s; f; 0; 11/, selects activity a1. In each recursive call, the activities that have already been selected are shaded, and the activity shown in white is being considered. If the starting time of an activity occurs before the finish time of the most recently added activity (the arrow between them points left), it is rejected. Otherwise (the arrow points directly up or to the right), it is selected. The last recursive call, RECURSIVE-ACTIVITY-SELECTOR.s; f; 11; 11/, returns ;. The resulting set of selected activities is fa1; a4; a8; a11g. 16.1 An activity-selection problem 421 The procedure GREEDY-ACTIVITY-SELECTOR is an iterative version of the procedure RECURSIVE-ACTIVITY-SELECTOR. It also assumes that the input activities are ordered by monotonically increasing finish time. It collects selected activities into a set A and returns this set when it is done. GREEDY-ACTIVITY-SELECTOR .s; f / 1 n D s:length 2 A D fa1g 3 kD1 4 for m D 2 to n 5 if sŒm f Œk 6 A D A [ famg 7 kDm 8 return A The procedure works as follows. The variable k indexes the most recent addition to A, corresponding to the activity ak in the recursive version. Since we consider the activities in order of monotonically increasing finish time, fk is always the maximum finish time of any activity in A. That is, fk D max ffi W ai 2 Ag : (16.3) Lines 2–3 select activity a1, initialize A to contain just this activity, and initialize k to index this activity. The for loop of lines 4–7 finds the earliest activity in Sk to finish. The loop considers each activity am in turn and adds am to A if it is compatible with all previously selected activities; such an activity is the earliest in Sk to finish. To see whether activity am is compatible with every activity currently in A, it suffices by equation (16.3) to check (in line 5) that its start time sm is not earlier than the finish time fk of the activity most recently added to A. If activity am is compatible, then lines 6–7 add activity am to A and set k to m. The set A returned by the call GREEDY-ACTIVITY-SELECTOR.s; f / is precisely the set returned by the call RECURSIVE-ACTIVITY-SELECTOR.s; f; 0; n/. Like the recursive version, GREEDY-ACTIVITY-SELECTOR schedules a set of n activities in ‚.n/ time, assuming that the activities were already sorted initially by their finish times. Exercises 16.1-1 Give a dynamic-programming algorithm for the activity-selection problem, based on recurrence (16.2). Have your algorithm compute the sizes cŒi; j  as defined above and also produce the maximum-size subset of mutually compatible activities. 422 Chapter 16 Greedy Algorithms Assume that the inputs have been sorted as in equation (16.1). Compare the running time of your solution to the running time of GREEDY-ACTIVITY-SELECTOR. 16.1-2 Suppose that instead of always selecting the first activity to finish, we instead select the last activity to start that is compatible with all previously selected activities. Describe how this approach is a greedy algorithm, and prove that it yields an optimal solution. 16.1-3 Not just any greedy approach to the activity-selection problem produces a maximum-size set of mutually compatible activities. Give an example to show that the approach of selecting the activity of least duration from among those that are compatible with previously selected activities does not work. Do the same for the approaches of always selecting the compatible activity that overlaps the fewest other remaining activities and always selecting the compatible remaining activity with the earliest start time. 16.1-4 Suppose that we have a set of activities to schedule among a large number of lecture halls, where any activity can take place in any lecture hall. We wish to schedule all the activities using as few lecture halls as possible. Give an efficient greedy algorithm to determine which activity should use which lecture hall. (This problem is also known as the interval-graph coloring problem. We can create an interval graph whose vertices are the given activities and whose edges connect incompatible activities. The smallest number of colors required to color every vertex so that no two adjacent vertices have the same color corresponds to finding the fewest lecture halls needed to schedule all of the given activities.) 16.1-5 Consider a modification to the activity-selection problem in which each activity ai has, in addition to a start and finish time, a value i . The objective is no longer to maximize the number of activities scheduled, but instead to maximize the total value of the activitiP es scheduled. That is, we wish to choose a set A of compatible activities such that ak2A k is maximized. Give a polynomial-time algorithm for this problem. 16.2 Elements of the greedy strategy 423 16.2 Elements of the greedy strategy A greedy algorithm obtains an optimal solution to a problem by making a sequence of choices. At each decision point, the algorithm makes choice that seems best at the moment. This heuristic strategy does not always produce an optimal solution, but as we saw in the activity-selection problem, sometimes it does. This section discusses some of the general properties of greedy methods. The process that we followed in Section 16.1 to develop a greedy algorithm was a bit more involved than is typical. We went through the following steps: 1. Determine the optimal substructure of the problem. 2. Develop a recursive solution. (For the activity-selection problem, we formulated recurrence (16.2), but we bypassed developing a recursive algorithm based on this recurrence.) 3. Show that if we make the greedy choice, then only one subproblem remains. 4. Prove that it is always safe to make the greedy choice. (Steps 3 and 4 can occur in either order.) 5. Develop a recursive algorithm that implements the greedy strategy. 6. Convert the recursive algorithm to an iterative algorithm. In going through these steps, we saw in great detail the dynamic-programming underpinnings of a greedy algorithm. For example, in the activity-selection problem, we first defined the subproblems Sij , where both i and j varied. We then found that if we always made the greedy choice, we could restrict the subproblems to be of the form Sk. Alternatively, we could have fashioned our optimal substructure with a greedy choice in mind, so that the choice leaves just one subproblem to solve. In the activity-selection problem, we could have started by dropping the second subscript and defining subproblems of the form Sk. Then, we could have proven that a greedy choice (the first activity am to finish in Sk), combined with an optimal solution to the remaining set Sm of compatible activities, yields an optimal solution to Sk. More generally, we design greedy algorithms according to the following sequence of steps: 1. Cast the optimization problem as one in which we make a choice and are left with one subproblem to solve. 2. Prove that there is always an optimal solution to the original problem that makes the greedy choice, so that the greedy choice is always safe. 424 Chapter 16 Greedy Algorithms 3. Demonstrate optimal substructure by showing that, having made the greedy choice, what remains is a subproblem with the property that if we combine an optimal solution to the subproblem with the greedy choice we have made, we arrive at an optimal solution to the original problem. We shall use this more direct process in later sections of this chapter. Nevertheless, beneath every greedy algorithm, there is almost always a more cumbersome dynamic-programming solution. How can we tell whether a greedy algorithm will solve a particular optimization problem? No way works all the time, but the greedy-choice property and optimal substructure are the two key ingredients. If we can demonstrate that the problem has these properties, then we are well on the way to developing a greedy algorithm for it. Greedy-choice property The first key ingredient is the greedy-choice property: we can assemble a globally optimal solution by making locally optimal (greedy) choices. In other words, when we are considering which choice to make, we make the choice that looks best in the current problem, without considering results from subproblems. Here is where greedy algorithms differ from dynamic programming. In dynamic programming, we make a choice at each step, but the choice usually depends on the solutions to subproblems. Consequently, we typically solve dynamic-programming problems in a bottom-up manner, progressing from smaller subproblems to larger subproblems. (Alternatively, we can solve them top down, but memoizing. Of course, even though the code works top down, we still must solve the subproblems before making a choice.) In a greedy algorithm, we make whatever choice seems best at the moment and then solve the subproblem that remains. The choice made by a greedy algorithm may depend on choices so far, but it cannot depend on any future choices or on the solutions to subproblems. Thus, unlike dynamic programming, which solves the subproblems before making the first choice, a greedy algorithm makes its first choice before solving any subproblems. A dynamicprogramming algorithm proceeds bottom up, whereas a greedy strategy usually progresses in a top-down fashion, making one greedy choice after another, reducing each given problem instance to a smaller one. Of course, we must prove that a greedy choice at each step yields a globally optimal solution. Typically, as in the case of Theorem 16.1, the proof examines a globally optimal solution to some subproblem. It then shows how to modify the solution to substitute the greedy choice for some other choice, resulting in one similar, but smaller, subproblem. We can usually make the greedy choice more efficiently than when we have to consider a wider set of choices. For example, in the activity-selection problem, as- 16.2 Elements of the greedy strategy 425 suming that we had already sorted the activities in monotonically increasing order of finish times, we needed to examine each activity just once. By preprocessing the input or by using an appropriate data structure (often a priority queue), we often can make greedy choices quickly, thus yielding an efficient algorithm. Optimal substructure A problem exhibits optimal substructure if an optimal solution to the problem contains within it optimal solutions to subproblems. This property is a key ingredient of assessing the applicability of dynamic programming as well as greedy algorithms. As an example of optimal substructure, recall how we demonstrated in Section 16.1 that if an optimal solution to subproblem Sij includes an activity ak, then it must also contain optimal solutions to the subproblems Sik and Skj . Given this optimal substructure, we argued that if we knew which activity to use as ak, we could construct an optimal solution to Sij by selecting ak along with all activities in optimal solutions to the subproblems Sik and Skj . Based on this observation of optimal substructure, we were able to devise the recurrence (16.2) that described the value of an optimal solution. We usually use a more direct approach regarding optimal substructure when applying it to greedy algorithms. As mentioned above, we have the luxury of assuming that we arrived at a subproblem by having made the greedy choice in the original problem. All we really need to do is argue that an optimal solution to the subproblem, combined with the greedy choice already made, yields an optimal solution to the original problem. This scheme implicitly uses induction on the subproblems to prove that making the greedy choice at every step produces an optimal solution. Greedy versus dynamic programming Because both the greedy and dynamic-programming strategies exploit optimal substructure, you might be tempted to generate a dynamic-programming solution to a problem when a greedy solution suffices or, conversely, you might mistakenly think that a greedy solution works when in fact a dynamic-programming solution is required. To illustrate the subtleties between the two techniques, let us investigate two variants of a classical optimization problem. The 0-1 knapsack problem is the following. A thief robbing a store finds n items. The i th item is worth i dollars and weighs wi pounds, where i and wi are integers. The thief wants to take as valuable a load as possible, but he can carry at most W pounds in his knapsack, for some integer W . Which items should he take? (We call this the 0-1 knapsack problem because for each item, the thief must either 426 Chapter 16 Greedy Algorithms take it or leave it behind; he cannot take a fractional amount of an item or take an item more than once.) In the fractional knapsack problem, the setup is the same, but the thief can take fractions of items, rather than having to make a binary (0-1) choice for each item. You can think of an item in the 0-1 knapsack problem as being like a gold ingot and an item in the fractional knapsack problem as more like gold dust. Both knapsack problems exhibit the optimal-substructure property. For the 0-1 problem, consider the most valuable load that weighs at most W pounds. If we remove item j from this load, the remaining load must be the most valuable load weighing at most W wj that the thief can take from the n 1 original items excluding j . For the comparable fractional problem, consider that if we remove a weight w of one item j from the optimal load, the remaining load must be the most valuable load weighing at most W w that the thief can take from the n 1 original items plus wj w pounds of item j . Although the problems are similar, we can solve the fractional knapsack problem by a greedy strategy, but we cannot solve the 0-1 problem by such a strategy. To solve the fractional problem, we first compute the value per pound i =wi for each item. Obeying a greedy strategy, the thief begins by taking as much as possible of the item with the greatest value per pound. If the supply of that item is exhausted and he can still carry more, he takes as much as possible of the item with the next greatest value per pound, and so forth, until he reaches his weight limit W . Thus, by sorting the items by value per pound, the greedy algorithm runs in O.n lg n/ time. We leave the proof that the fractional knapsack problem has the greedychoice property as Exercise 16.2-1. To see that this greedy strategy does not work for the 0-1 knapsack problem, consider the problem instance illustrated in Figure 16.2(a). This example has 3 items and a knapsack that can hold 50 pounds. Item 1 weighs 10 pounds and is worth 60 dollars. Item 2 weighs 20 pounds and is worth 100 dollars. Item 3 weighs 30 pounds and is worth 120 dollars. Thus, the value per pound of item 1 is 6 dollars per pound, which is greater than the value per pound of either item 2 (5 dollars per pound) or item 3 (4 dollars per pound). The greedy strategy, therefore, would take item 1 first. As you can see from the case analysis in Figure 16.2(b), however, the optimal solution takes items 2 and 3, leaving item 1 behind. The two possible solutions that take item 1 are both suboptimal. For the comparable fractional problem, however, the greedy strategy, which takes item 1 first, does yield an optimal solution, as shown in Figure 16.2(c). Taking item 1 doesn’t work in the 0-1 problem because the thief is unable to fill his knapsack to capacity, and the empty space lowers the effective value per pound of his load. In the 0-1 problem, when we consider whether to include an item in the knapsack, we must compare the solution to the subproblem that includes the item with the solution to the subproblem that excludes the item before we can make the 16.2 Elements of the greedy strategy 427 item 3 item 2 50 item 1 30 20 10 $60 $100 $120 knapsack (a) 30 $120 + 20 $100 = $220 20 $100 + 10 $60 = $160 (b) 30 $120 + 10 $60 = $180 20 30 $80 + 20 $100 + 10 $60 = $240 (c) Figure 16.2 An example showing that the greedy strategy does not work for the 0-1 knapsack problem. (a) The thief must select a subset of the three items shown whose weight must not exceed 50 pounds. (b) The optimal subset includes items 2 and 3. Any solution with item 1 is suboptimal, even though item 1 has the greatest value per pound. (c) For the fractional knapsack problem, taking the items in order of greatest value per pound yields an optimal solution. choice. The problem formulated in this way gives rise to many overlapping subproblems—a hallmark of dynamic programming, and indeed, as Exercise 16.2-2 asks you to show, we can use dynamic programming to solve the 0-1 problem. Exercises 16.2-1 Prove that the fractional knapsack problem has the greedy-choice property. 16.2-2 Give a dynamic-programming solution to the 0-1 knapsack problem that runs in O.n W / time, where n is the number of items and W is the maximum weight of items that the thief can put in his knapsack. 16.2-3 Suppose that in a 0-1 knapsack problem, the order of the items when sorted by increasing weight is the same as their order when sorted by decreasing value. Give an efficient algorithm to find an optimal solution to this variant of the knapsack problem, and argue that your algorithm is correct. 16.2-4 Professor Gekko has always dreamed of inline skating across North Dakota. He plans to cross the state on highway U.S. 2, which runs from Grand Forks, on the eastern border with Minnesota, to Williston, near the western border with Montana. 428 Chapter 16 Greedy Algorithms The professor can carry two liters of water, and he can skate m miles before running out of water. (Because North Dakota is relatively flat, the professor does not have to worry about drinking water at a greater rate on uphill sections than on flat or downhill sections.) The professor will start in Grand Forks with two full liters of water. His official North Dakota state map shows all the places along U.S. 2 at which he can refill his water and the distances between these locations. The professor’s goal is to minimize the number of water stops along his route across the state. Give an efficient method by which he can determine which water stops he should make. Prove that your strategy yields an optimal solution, and give its running time. 16.2-5 Describe an efficient algorithm that, given a set fx1; x2; : : : ; xng of points on the real line, determines the smallest set of unit-length closed intervals that contains all of the given points. Argue that your algorithm is correct. 16.2-6 ? Show how to solve the fractional knapsack problem in O.n/ time. 16.2-7 Suppose you are given two sets A and B, each containing n positive integers. You can choose to reorder each set however you like. After reordering, let ai be the ith oelfeQmeni Dnt1 of ai set bi . A, and let bi be the i th Give an algorithm that element of set will maximize B. You then your payoff. receive a payoff Prove that your algorithm maximizes the payoff, and state its running time. 16.3 Huffman codes Huffman codes compress data very effectively: savings of 20% to 90% are typical, depending on the characteristics of the data being compressed. We consider the data to be a sequence of characters. Huffman’s greedy algorithm uses a table giving how often each character occurs (i.e., its frequency) to build up an optimal way of representing each character as a binary string. Suppose we have a 100,000-character data file that we wish to store compactly. We observe that the characters in the file occur with the frequencies given by Figure 16.3. That is, only 6 different characters appear, and the character a occurs 45,000 times. We have many options for how to represent such a file of information. Here, we consider the problem of designing a binary character code (or code for short) 16.3 Huffman codes 429 abcd e f Frequency (in thousands) 45 13 12 16 9 5 Fixed-length codeword 000 001 010 011 100 101 Variable-length codeword 0 101 100 111 1101 1100 Figure 16.3 A character-coding problem. A data file of 100,000 characters contains only the characters a–f, with the frequencies indicated. If we assign each character a 3-bit codeword, we can encode the file in 300,000 bits. Using the variable-length code shown, we can encode the file in only 224,000 bits. in which each character is represented by a unique binary string, which we call a codeword. If we use a fixed-length code, we need 3 bits to represent 6 characters: a = 000, b = 001, . . . , f = 101. This method requires 300,000 bits to code the entire file. Can we do better? A variable-length code can do considerably better than a fixed-length code, by giving frequent characters short codewords and infrequent characters long codewords. Figure 16.3 shows such a code; here the 1-bit string 0 represents a, and the 4-bit string 1100 represents f. This code requires .45 1 C 13 3 C 12 3 C 16 3 C 9 4 C 5 4/ 1,000 D 224,000 bits to represent the file, a savings of approximately 25%. In fact, this is an optimal character code for this file, as we shall see. Prefix codes We consider here only codes in which no codeword is also a prefix of some other codeword. Such codes are called prefix codes.3 Although we won’t prove it here, a prefix code can always achieve the optimal data compression among any character code, and so we suffer no loss of generality by restricting our attention to prefix codes. Encoding is always simple for any binary character code; we just concatenate the codewords representing each character of the file. For example, with the variablelength prefix code of Figure 16.3, we code the 3-character file abc as 0 101 100 D 0101100, where “ ” denotes concatenation. Prefix codes are desirable because they simplify decoding. Since no codeword is a prefix of any other, the codeword that begins an encoded file is unambiguous. We can simply identify the initial codeword, translate it back to the original char- 3Perhaps “prefix-free codes” would be a better name, but the term “prefix codes” is standard in the literature. 430 Chapter 16 Greedy Algorithms 100 0 1 86 0 1 58 28 14 0 14 01 01 01 a:45 b:13 c:12 d:16 e:9 f:5 (a) 100 0 1 a:45 55 0 1 25 30 01 01 c:12 b:13 14 d:16 01 f:5 e:9 (b) Figure 16.4 Trees corresponding to the coding schemes in Figure 16.3. Each leaf is labeled with a character and its frequency of occurrence. Each internal node is labeled with the sum of the frequencies of the leaves in its subtree. (a) The tree corresponding to the fixed-length code a = 000, . . . , f = 101. (b) The tree corresponding to the optimal prefix code a = 0, b = 101, . . . , f = 1100. acter, and repeat the decoding process on the remainder of the encoded file. In our example, the string 001011101 parses uniquely as 0 0 101 1101, which decodes to aabe. The decoding process needs a convenient representation for the prefix code so that we can easily pick off the initial codeword. A binary tree whose leaves are the given characters provides one such representation. We interpret the binary codeword for a character as the simple path from the root to that character, where 0 means “go to the left child” and 1 means “go to the right child.” Figure 16.4 shows the trees for the two codes of our example. Note that these are not binary search trees, since the leaves need not appear in sorted order and internal nodes do not contain character keys. An optimal code for a file is always represented by a full binary tree, in which every nonleaf node has two children (see Exercise 16.3-2). The fixed-length code in our example is not optimal since its tree, shown in Figure 16.4(a), is not a full binary tree: it contains codewords beginning 10. . . , but none beginning 11. . . . Since we can now restrict our attention to full binary trees, we can say that if C is the alphabet from which the characters are drawn and all character frequencies are positive, then the tree for an optimal prefix code has exactly jC j leaves, one for each letter of the alphabet, and exactly jC j 1 internal nodes (see Exercise B.5-3). Given a tree T corresponding to a prefix code, we can easily compute the number of bits required to encode a file. For each character c in the alphabet C , let the attribute c:freq denote the frequency of c in the file and let dT .c/ denote the depth 16.3 Huffman codes 431 of c’s leaf in the tree. Note that dT .c/ is also the length of the codeword for character c. The number of bits required to encode a file is thus X B.T / D c:freq dT .c/ ; (16.4) c2C which we define as the cost of the tree T . Constructing a Huffman code Huffman invented a greedy algorithm that constructs an optimal prefix code called a Huffman code. In line with our observations in Section 16.2, its proof of correctness relies on the greedy-choice property and optimal substructure. Rather than demonstrating that these properties hold and then developing pseudocode, we present the pseudocode first. Doing so will help clarify how the algorithm makes greedy choices. In the pseudocode that follows, we assume that C is a set of n characters and that each character c 2 C is an object with an attribute c:freq giving its frequency. The algorithm builds the tree T corresponding to the optimal code in a bottom-up manner. It begins with a set of jC j leaves and performs a sequence of jC j 1 “merging” operations to create the final tree. The algorithm uses a min-priority queue Q, keyed on the freq attribute, to identify the two least-frequent objects to merge together. When we merge two objects, the result is a new object whose frequency is the sum of the frequencies of the two objects that were merged. HUFFMAN.C / 1 n D jC j 2 QDC 3 for i D 1 to n 1 4 allocate a new node ´ 5 ´:left D x D EXTRACT-MIN.Q/ 6 ´:right D y D EXTRACT-MIN.Q/ 7 ´:freq D x:freq C y:freq 8 INSERT.Q; ´/ 9 return EXTRACT-MIN.Q/ // return the root of the tree For our example, Huffman’s algorithm proceeds as shown in Figure 16.5. Since the alphabet contains 6 letters, the initial queue size is n D 6, and 5 merge steps build the tree. The final tree represents the optimal prefix code. The codeword for a letter is the sequence of edge labels on the simple path from the root to the letter. Line 2 initializes the min-priority queue Q with the characters in C . The for loop in lines 3–8 repeatedly extracts the two nodes x and y of lowest frequency 432 Chapter 16 Greedy Algorithms (a) f:5 e:9 c:12 b:13 d:16 a:45 (b) c:12 b:13 14 d:16 a:45 01 f:5 e:9 (c) 14 d:16 25 a:45 01 01 f:5 e:9 c:12 b:13 (d) 25 01 c:12 b:13 30 a:45 01 14 d:16 01 f:5 e:9 (e) a:45 55 0 1 25 30 01 01 c:12 b:13 14 d:16 01 f:5 e:9 (f) 100 0 1 a:45 55 0 1 25 30 01 01 c:12 b:13 14 d:16 01 f:5 e:9 Figure 16.5 The steps of Huffman’s algorithm for the frequencies given in Figure 16.3. Each part shows the contents of the queue sorted into increasing order by frequency. At each step, the two trees with lowest frequencies are merged. Leaves are shown as rectangles containing a character and its frequency. Internal nodes are shown as circles containing the sum of the frequencies of their children. An edge connecting an internal node with its children is labeled 0 if it is an edge to a left child and 1 if it is an edge to a right child. The codeword for a letter is the sequence of labels on the edges connecting the root to the leaf for that letter. (a) The initial set of n D 6 nodes, one for each letter. (b)–(e) Intermediate stages. (f) The final tree. from the queue, replacing them in the queue with a new node ´ representing their merger. The frequency of ´ is computed as the sum of the frequencies of x and y in line 7. The node ´ has x as its left child and y as its right child. (This order is arbitrary; switching the left and right child of any node yields a different code of the same cost.) After n 1 mergers, line 9 returns the one node left in the queue, which is the root of the code tree. Although the algorithm would produce the same result if we were to excise the variables x and y—assigning directly to ´:left and ´:right in lines 5 and 6, and changing line 7 to ´:freq D ´:left:freq C ´:right:freq—we shall use the node 16.3 Huffman codes 433 names x and y in the proof of correctness. Therefore, we find it convenient to leave them in. To analyze the running time of Huffman’s algorithm, we assume that Q is implemented as a binary min-heap (see Chapter 6). For a set C of n characters, we can initialize Q in line 2 in O.n/ time using the BUILD-MIN-HEAP procedure discussed in Section 6.3. The for loop in lines 3–8 executes exactly n 1 times, and since each heap operation requires time O.lg n/, the loop contributes O.n lg n/ to the running time. Thus, the total running time of HUFFMAN on a set of n characters is O.n lg n/. We can reduce the running time to O.n lg lg n/ by replacing the binary min-heap with a van Emde Boas tree (see Chapter 20). Correctness of Huffman’s algorithm To prove that the greedy algorithm HUFFMAN is correct, we show that the problem of determining an optimal prefix code exhibits the greedy-choice and optimalsubstructure properties. The next lemma shows that the greedy-choice property holds. Lemma 16.2 Let C be an alphabet in which each character c 2 C has frequency c:freq. Let x and y be two characters in C having the lowest frequencies. Then there exists an optimal prefix code for C in which the codewords for x and y have the same length and differ only in the last bit. Proof The idea of the proof is to take the tree T representing an arbitrary optimal prefix code and modify it to make a tree representing another optimal prefix code such that the characters x and y appear as sibling leaves of maximum depth in the new tree. If we can construct such a tree, then the codewords for x and y will have the same length and differ only in the last bit. Let a and b be two characters that are sibling leaves of maximum depth in T . Without loss of generality, we assume that a:freq Ä b:freq and x:freq Ä y:freq. Since x:freq and y:freq are the two lowest leaf frequencies, in order, and a:freq and b:freq are two arbitrary frequencies, in order, we have x:freq Ä a:freq and y:freq Ä b:freq. In the remainder of the proof, it is possible that we could have x:freq D a:freq or y:freq D b:freq. However, if we had x:freq D b:freq, then we would also have a:freq D b:freq D x:freq D y:freq (see Exercise 16.3-1), and the lemma would be trivially true. Thus, we will assume that x:freq ¤ b:freq, which means that x ¤ b. As Figure 16.6 shows, we exchange the positions in T of a and x to produce a tree T 0, and then we exchange the positions in T 0 of b and y to produce a tree T 00 434 Chapter 16 Greedy Algorithms T x y ab T′ a y xb T′′ a b xy Figure 16.6 An illustration of the key step in the proof of Lemma 16.2. In the optimal tree T , leaves a and b are two siblings of maximum depth. Leaves x and y are the two characters with the lowest frequencies; they appear in arbitrary positions in T . Assuming that x ¤ b, swapping leaves a and x produces tree T 0, and then swapping leaves b and y produces tree T 00. Since each swap does not increase the cost, the resulting tree T 00 is also an optimal tree. in which x and y are sibling leaves of maximum depth. (Note that if x D b but y ¤ a, then tree T 00 does not have x and y as sibling leaves of maximum depth. Because we assume that x ¤ b, this situation cannot occur.) By equation (16.4), the difference in cost between T and T 0 is B.T / B.TX0/ X D c:freq dT .c/ c:freq dT 0.c/ c2C c2C D x:freq dT .x/ C a:freq dT .a/ x:freq dT 0.x/ D x:freq dT .x/ C a:freq dT .a/ x:freq dT .a/ D .a:freq x:freq/.dT .a/ dT .x// 0; a:freq dT 0 .a/ a:freq dT .x/ because both a:freq x:freq and dT .a/ dT .x/ are nonnegative. More specifi- cally, a:freq x:freq is nonnegative because x is a minimum-frequency leaf, and dT .a/ dT .x/ is nonnegative because a is a leaf of maximum depth in T . Similarly, exchanging y and b does not increase the cost, and so B.T 0/ B.T 00/ is nonnegative. Therefore, B.T 00/ Ä B.T /, and since T is optimal, we have B.T / Ä B.T 00/, which implies B.T 00/ D B.T /. Thus, T 00 is an optimal tree in which x and y appear as sibling leaves of maximum depth, from which the lemma follows. Lemma 16.2 implies that the process of building up an optimal tree by mergers can, without loss of generality, begin with the greedy choice of merging together those two characters of lowest frequency. Why is this a greedy choice? We can view the cost of a single merger as being the sum of the frequencies of the two items being merged. Exercise 16.3-4 shows that the total cost of the tree constructed equals the sum of the costs of its mergers. Of all possible mergers at each step, HUFFMAN chooses the one that incurs the least cost. 16.3 Huffman codes 435 The next lemma shows that the problem of constructing optimal prefix codes has the optimal-substructure property. Lemma 16.3 Let C be a given alphabet with frequency c:freq defined for each character c 2 C . Let x and y be two characters in C with minimum frequency. Let C 0 be the alphabet C with the characters x and y removed and a new character ´ added, so that C 0 D C fx; yg [ f´g. Define f for C 0 as for C , except that ´:freq D x:freq C y:freq. Let T 0 be any tree representing an optimal prefix code for the alphabet C 0. Then the tree T , obtained from T 0 by replacing the leaf node for ´ with an internal node having x and y as children, represents an optimal prefix code for the alphabet C . Proof We first show how to express the cost B.T / of tree T in terms of the cost B.T 0/ of tree T 0, by considering the component costs in equation (16.4). For each character c 2 C fx; yg, we have that dT .c/ D dT 0 .c/, and hence c:freq dT .c/ D c:freq dT 0 .c/. Since dT .x/ D dT .y/ D dT 0 .´/ C 1, we have x:freq dT .x/ C y:freq dT .y/ D .x:freq C y:freq/.dT 0.´/ C 1/ D ´:freq dT 0.´/ C .x:freq C y:freq/ ; from which we conclude that B.T / D B.T 0/ C x:freq C y:freq or, equivalently, B.T 0/ D B.T / x:freq y:freq : We now prove the lemma by contradiction. Suppose that T does not represent an optimal prefix code for C . Then there exists an optimal tree T 00 such that B.T 00/ < B.T /. Without loss of generality (by Lemma 16.2), T 00 has x and y as siblings. Let T 000 be the tree T 00 with the common parent of x and y replaced by a leaf ´ with frequency ´:freq D x:freq C y:freq. Then B.T 000/ D B.T 00/ x:freq y:freq < B.T / x:freq y:freq D B.T 0/ ; yielding a contradiction to the assumption that T 0 represents an optimal prefix code for C 0. Thus, T must represent an optimal prefix code for the alphabet C . Theorem 16.4 Procedure HUFFMAN produces an optimal prefix code. Proof Immediate from Lemmas 16.2 and 16.3. 436 Chapter 16 Greedy Algorithms Exercises 16.3-1 Explain why, in the proof of Lemma 16.2, if x:freq D b:freq, then we must have a:freq D b:freq D x:freq D y:freq. 16.3-2 Prove that a binary tree that is not full cannot correspond to an optimal prefix code. 16.3-3 What is an optimal Huffman code for the following set of frequencies, based on the first 8 Fibonacci numbers? a:1 b:1 c:2 d:3 e:5 f:8 g:13 h:21 Can you generalize your answer to find the optimal code when the frequencies are the first n Fibonacci numbers? 16.3-4 Prove that we can also express the total cost of a tree for a code as the sum, over all internal nodes, of the combined frequencies of the two children of the node. 16.3-5 Prove that if we order the characters in an alphabet so that their frequencies are monotonically decreasing, then there exists an optimal code whose codeword lengths are monotonically increasing. 16.3-6 Suppose we have an optimal prefix code on a set C D f0; 1; : : : ; n 1g of characters and we wish to transmit this code using as few bits as possible. Show how to represent any optimal prefix code on C using only 2n 1 C n dlg ne bits. (Hint: Use 2n 1 bits to specify the structure of the tree, as discovered by a walk of the tree.) 16.3-7 Generalize Huffman’s algorithm to ternary codewords (i.e., codewords using the symbols 0, 1, and 2), and prove that it yields optimal ternary codes. 16.3-8 Suppose that a data file contains a sequence of 8-bit characters such that all 256 characters are about equally common: the maximum character frequency is less than twice the minimum character frequency. Prove that Huffman coding in this case is no more efficient than using an ordinary 8-bit fixed-length code. 16.4 Matroids and greedy methods 437 16.3-9 Show that no compression scheme can expect to compress a file of randomly chosen 8-bit characters by even a single bit. (Hint: Compare the number of possible files with the number of possible encoded files.) ? 16.4 Matroids and greedy methods In this section, we sketch a beautiful theory about greedy algorithms. This theory describes many situations in which the greedy method yields optimal solutions. It involves combinatorial structures known as “matroids.” Although this theory does not cover all cases for which a greedy method applies (for example, it does not cover the activity-selection problem of Section 16.1 or the Huffman-coding problem of Section 16.3), it does cover many cases of practical interest. Furthermore, this theory has been extended to cover many applications; see the notes at the end of this chapter for references. Matroids A matroid is an ordered pair M D .S; « / satisfying the following conditions. 1. S is a finite set. 2. « is a nonempty family of subsets of S, called the independent subsets of S, such that if B 2 « and A  B, then A 2 « . We say that « is hereditary if it satisfies this property. Note that the empty set ; is necessarily a member of « . 3. If A 2 « , B 2 « , and jAj < jBj, then there exists some element x 2 B A such that A [ fxg 2 « . We say that M satisfies the exchange property. The word “matroid” is due to Hassler Whitney. He was studying matric matroids, in which the elements of S are the rows of a given matrix and a set of rows is independent if they are linearly independent in the usual sense. As Exercise 16.4-2 asks you to show, this structure defines a matroid. As another example of matroids, consider the graphic matroid MG D .SG; «G/ defined in terms of a given undirected graph G D .V; E/ as follows: The set SG is defined to be E, the set of edges of G. If A is a subset of E, then A 2 «G if and only if A is acyclic. That is, a set of edges A is independent if and only if the subgraph GA D .V; A/ forms a forest. The graphic matroid MG is closely related to the minimum-spanning-tree problem, which Chapter 23 covers in detail. 438 Chapter 16 Greedy Algorithms Theorem 16.5 If G D .V; E/ is an undirected graph, then MG D .SG; «G/ is a matroid. Proof Clearly, SG D E is a finite set. Furthermore, «G is hereditary, since a subset of a forest is a forest. Putting it another way, removing edges from an acyclic set of edges cannot create cycles. Thus, it remains to show that MG satisfies the exchange property. Suppose that GA D .V; A/ and GB D .V; B/ are forests of G and that jBj > jAj. That is, A and B are acyclic sets of edges, and B contains more edges than A does. We claim that a forest F D .VF ; EF / contains exactly jVF j jEF j trees. To see why, suppose that F consists of t trees, where the ith tree contains i vertices and ei edges. Then, we have Xt jEF j D ei i D1 Xt D . i 1/ (by Theorem B.2) i D1 Xt D it i D1 D jVF j t ; which implies that t D jVF j jEF j. Thus, forest GA contains jV j jAj trees, and forest GB contains jV j jBj trees. Since forest GB has fewer trees than forest GA does, forest GB must contain some tree T whose vertices are in two different trees in forest GA. Moreover, since T is connected, it must contain an edge .u; / such that vertices u and are in different trees in forest GA. Since the edge .u; / connects vertices in two different trees in forest GA, we can add the edge .u; / to forest GA without creating a cycle. Therefore, MG satisfies the exchange property, completing the proof that MG is a matroid. Given a matroid M D .S; « /, we call an element x … A an extension of A 2 « if we can add x to A while preserving independence; that is, x is an extension of A if A [ fxg 2 « . As an example, consider a graphic matroid MG. If A is an independent set of edges, then edge e is an extension of A if and only if e is not in A and the addition of e to A does not create a cycle. If A is an independent subset in a matroid M , we say that A is maximal if it has no extensions. That is, A is maximal if it is not contained in any larger independent subset of M . The following property is often useful. 16.4 Matroids and greedy methods 439 Theorem 16.6 All maximal independent subsets in a matroid have the same size. Proof Suppose to the contrary that A is a maximal independent subset of M and there exists another larger maximal independent subset B of M . Then, the exchange property implies that for some x 2 B A, we can extend A to a larger independent set A [ fxg, contradicting the assumption that A is maximal. As an illustration of this theorem, consider a graphic matroid MG for a connected, undirected graph G. Every maximal independent subset of MG must be a free tree with exactly jV j 1 edges that connects all the vertices of G. Such a tree is called a spanning tree of G. We say that a matroid M D .S; « / is weighted if it is associated with a weight function w that assigns a strictly positive weight w.x/ to each element x 2 S. The weight function w extends to subsets of S by summation: X w.A/ D w.x/ x2A for any A  S. For example, if we let w.e/ denote the weight of an edge e in a graphic matroid MG, then w.A/ is the total weight of the edges in edge set A. Greedy algorithms on a weighted matroid Many problems for which a greedy approach provides optimal solutions can be formulated in terms of finding a maximum-weight independent subset in a weighted matroid. That is, we are given a weighted matroid M D .S; « /, and we wish to find an independent set A 2 « such that w.A/ is maximized. We call such a subset that is independent and has maximum possible weight an optimal subset of the matroid. Because the weight w.x/ of any element x 2 S is positive, an optimal subset is always a maximal independent subset—it always helps to make A as large as possible. For example, in the minimum-spanning-tree problem, we are given a connected undirected graph G D .V; E/ and a length function w such that w.e/ is the (positive) length of edge e. (We use the term “length” here to refer to the original edge weights for the graph, reserving the term “weight” to refer to the weights in the associated matroid.) We wish to find a subset of the edges that connects all of the vertices together and has minimum total length. To view this as a problem of finding an optimal subset of a matroid, consider the weighted matroid MG with weight function w0, where w0.e/ D w0 w.e/ and w0 is larger than the maximum length of any edge. In this weighted matroid, all weights are positive and an optimal subset is a spanning tree of minimum total length in the original graph. More specifically, each maximal independent subset A corresponds to a spanning tree 440 Chapter 16 Greedy Algorithms with jV j 1 edges, and since X w0.A/ D w0.e/ X e2A D .w0 w.e// e2A X D .jV j 1/w0 w.e/ e2A D .jV j 1/w0 w.A/ for any maximal independent subset A, an independent subset that maximizes the quantity w0.A/ must minimize w.A/. Thus, any algorithm that can find an optimal subset A in an arbitrary matroid can solve the minimum-spanning-tree problem. Chapter 23 gives algorithms for the minimum-spanning-tree problem, but here we give a greedy algorithm that works for any weighted matroid. The algorithm takes as input a weighted matroid M D .S; « / with an associated positive weight function w, and it returns an optimal subset A. In our pseudocode, we denote the components of M by M:S and M:« and the weight function by w. The algorithm is greedy because it considers in turn each element x 2 S, in order of monotonically decreasing weight, and immediately adds it to the set A being accumulated if A [ fxg is independent. GREEDY.M; w/ 1 AD; 2 sort M:S into monotonically decreasing order by weight w 3 for each x 2 M:S, taken in monotonically decreasing order by weight w.x/ 4 if A [ fxg 2 M:« 5 A D A [ fxg 6 return A Line 4 checks whether adding each element x to A would maintain A as an independent set. If A would remain independent, then line 5 adds x to A. Otherwise, x is discarded. Since the empty set is independent, and since each iteration of the for loop maintains A’s independence, the subset A is always independent, by induction. Therefore, GREEDY always returns an independent subset A. We shall see in a moment that A is a subset of maximum possible weight, so that A is an optimal subset. The running time of GREEDY is easy to analyze. Let n denote jSj. The sorting phase of GREEDY takes time O.n lg n/. Line 4 executes exactly n times, once for each element of S. Each execution of line 4 requires a check on whether or not the set A [ fxg is independent. If each such check takes time O.f .n//, the entire algorithm runs in time O.n lg n C nf .n//. 16.4 Matroids and greedy methods 441 We now prove that GREEDY returns an optimal subset. Lemma 16.7 (Matroids exhibit the greedy-choice property) Suppose that M D .S; « / is a weighted matroid with weight function w and that S is sorted into monotonically decreasing order by weight. Let x be the first element of S such that fxg is independent, if any such x exists. If x exists, then there exists an optimal subset A of S that contains x. Proof If no such x exists, then the only independent subset is the empty set and the lemma is vacuously true. Otherwise, let B be any nonempty optimal subset. Assume that x … B; otherwise, letting A D B gives an optimal subset of S that contains x. No element of B has weight greater than w.x/. To see why, observe that y 2 B implies that fyg is independent, since B 2 « and « is hereditary. Our choice of x therefore ensures that w.x/ w.y/ for any y 2 B. Construct the set A as follows. Begin with A D fxg. By the choice of x, set A is independent. Using the exchange property, repeatedly find a new element of B that we can add to A until jAj D jBj, while preserving the independence of A. At that point, A and B are the same except that A has x and B has some other element y. That is, A D B fyg [ fxg for some y 2 B, and so w.A/ D w.B/ w.y/ C w.x/ w.B/ : Because set B is optimal, set A, which contains x, must also be optimal. We next show that if an element is not an option initially, then it cannot be an option later. Lemma 16.8 Let M D .S; « / be any matroid. If x is an element of S that is an extension of some independent subset A of S, then x is also an extension of ;. Proof Since x is an extension of A, we have that A [ fxg is independent. Since « is hereditary, fxg must be independent. Thus, x is an extension of ;. Corollary 16.9 Let M D .S; « / be any matroid. If x is an element of S such that x is not an extension of ;, then x is not an extension of any independent subset A of S. Proof This corollary is simply the contrapositive of Lemma 16.8. 442 Chapter 16 Greedy Algorithms Corollary 16.9 says that any element that cannot be used immediately can never be used. Therefore, GREEDY cannot make an error by passing over any initial elements in S that are not an extension of ;, since they can never be used. Lemma 16.10 (Matroids exhibit the optimal-substructure property) Let x be the first element of S chosen by GREEDY for the weighted matroid M D .S; « /. The remaining problem of finding a maximum-weight independent subset containing x reduces to finding a maximum-weight independent subset of the weighted matroid M 0 D .S0; « 0/, where S0 D fy 2 S W fx; yg 2 « g ; « 0 D fB  S fxg W B [ fxg 2 « g ; and the weight function for M 0 is the weight function for M , restricted to S0. (We call M 0 the contraction of M by the element x.) Proof If A is any maximum-weight independent subset of M containing x, then A0 D A fxg is an independent subset of M 0. Conversely, any independent subset A0 of M 0 yields an independent subset A D A0 [ fxg of M . Since we have in both cases that w.A/ D w.A0/ C w.x/, a maximum-weight solution in M containing x yields a maximum-weight solution in M 0, and vice versa. Theorem 16.11 (Correctness of the greedy algorithm on matroids) If M D .S; « / is a weighted matroid with weight function w, then GREEDY.M; w/ returns an optimal subset. Proof By Corollary 16.9, any elements that GREEDY passes over initially because they are not extensions of ; can be forgotten about, since they can never be useful. Once GREEDY selects the first element x, Lemma 16.7 implies that the algorithm does not err by adding x to A, since there exists an optimal subset containing x. Finally, Lemma 16.10 implies that the remaining problem is one of finding an optimal subset in the matroid M 0 that is the contraction of M by x. After the procedure GREEDY sets A to fxg, we can interpret all of its remaining steps as acting in the matroid M 0 D .S0; « 0/, because B is independent in M 0 if and only if B [ fxg is independent in M , for all sets B 2 « 0. Thus, the subsequent operation of GREEDY will find a maximum-weight independent subset for M 0, and the overall operation of GREEDY will find a maximum-weight independent subset for M . 16.5 A task-scheduling problem as a matroid 443 Exercises 16.4-1 Show that .S; «k/ is a matroid, where S is any finite set and «k is the set of all subsets of S of size at most k, where k Ä jSj. 16.4-2 ? Given an m n matrix T over some field (such as the reals), show that .S; « / is a matroid, where S is the set of columns of T and A 2 « if and only if the columns in A are linearly independent. 16.4-3 ? Show that if .S; « / is a matroid, then .S; « 0/ is a matroid, where « 0 D fA0 W S A0 contains some maximal A 2 « g : That is, the maximal independent sets of .S; « 0/ are just the complements of the maximal independent sets of .S; « /. 16.4-4 ? Let S be a finite set and let S1; S2; : : : ; Sk be a partition of S into nonempty disjoint subsets. Define the structure .S; « / by the condition that « D fA W jA \ Si j Ä 1 for i D 1; 2; : : : ; kg. Show that .S; « / is a matroid. That is, the set of all sets A that contain at most one member of each subset in the partition determines the independent sets of a matroid. 16.4-5 Show how to transform the weight function of a weighted matroid problem, where the desired optimal solution is a minimum-weight maximal independent subset, to make it a standard weighted-matroid problem. Argue carefully that your transformation is correct. ? 16.5 A task-scheduling problem as a matroid An interesting problem that we can solve using matroids is the problem of optimally scheduling unit-time tasks on a single processor, where each task has a deadline, along with a penalty paid if the task misses its deadline. The problem looks complicated, but we can solve it in a surprisingly simple manner by casting it as a matroid and using a greedy algorithm. A unit-time task is a job, such as a program to be run on a computer, that requires exactly one unit of time to complete. Given a finite set S of unit-time tasks, a 444 Chapter 16 Greedy Algorithms schedule for S is a permutation of S specifying the order in which to perform these tasks. The first task in the schedule begins at time 0 and finishes at time 1, the second task begins at time 1 and finishes at time 2, and so on. The problem of scheduling unit-time tasks with deadlines and penalties for a single processor has the following inputs: a set S D fa1; a2; : : : ; ang of n unit-time tasks; a set of n integer deadlines d1; d2; : : : ; dn, such that each di satisfies 1 Ä di Ä n and task ai is supposed to finish by time di ; and a set of n nonnegative weights or penalties w1; w2; : : : ; wn, such that we incur a penalty of wi if task ai is not finished by time di , and we incur no penalty if a task finishes by its deadline. We wish to find a schedule for S that minimizes the total penalty incurred for missed deadlines. Consider a given schedule. We say that a task is late in this schedule if it finishes after its deadline. Otherwise, the task is early in the schedule. We can always transform an arbitrary schedule into early-first form, in which the early tasks precede the late tasks. To see why, note that if some early task ai follows some late task aj , then we can switch the positions of ai and aj , and ai will still be early and aj will still be late. Furthermore, we claim that we can always transform an arbitrary schedule into canonical form, in which the early tasks precede the late tasks and we schedule the early tasks in order of monotonically increasing deadlines. To do so, we put the schedule into early-first form. Then, as long as there exist two early tasks ai and aj finishing at respective times k and k C 1 in the schedule such that dj < di , we swap the positions of ai and aj . Since aj is early before the swap, k C 1 Ä dj . Therefore, k C 1 < di , and so ai is still early after the swap. Because task aj is moved earlier in the schedule, it remains early after the swap. The search for an optimal schedule thus reduces to finding a set A of tasks that we assign to be early in the optimal schedule. Having determined A, we can create the actual schedule by listing the elements of A in order of monotonically increasing deadlines, then listing the late tasks (i.e., S A) in any order, producing a canonical ordering of the optimal schedule. We say that a set A of tasks is independent if there exists a schedule for these tasks such that no tasks are late. Clearly, the set of early tasks for a schedule forms an independent set of tasks. Let « denote the set of all independent sets of tasks. Consider the problem of determining whether a given set A of tasks is independent. For t D 0; 1; 2; : : : ; n, let Nt .A/ denote the number of tasks in A whose deadline is t or earlier. Note that N0.A/ D 0 for any set A. 16.5 A task-scheduling problem as a matroid 445 Lemma 16.12 For any set of tasks A, the following statements are equivalent. 1. The set A is independent. 2. For t D 0; 1; 2; : : : ; n, we have Nt .A/ Ä t. 3. If the tasks in A are scheduled in order of monotonically increasing deadlines, then no task is late. Proof To show that (1) implies (2), we prove the contrapositive: if Nt .A/ > t for some t, then there is no way to make a schedule with no late tasks for set A, because more than t tasks must finish before time t. Therefore, (1) implies (2). If (2) holds, then (3) must follow: there is no way to “get stuck” when scheduling the tasks in order of monotonically increasing deadlines, since (2) implies that the ith largest deadline is at least i. Finally, (3) trivially implies (1). Using property 2 of Lemma 16.12, we can easily compute whether or not a given set of tasks is independent (see Exercise 16.5-2). The problem of minimizing the sum of the penalties of the late tasks is the same as the problem of maximizing the sum of the penalties of the early tasks. The following theorem thus ensures that we can use the greedy algorithm to find an independent set A of tasks with the maximum total penalty. Theorem 16.13 If S is a set of unit-time tasks with deadlines, and « is the set of all independent sets of tasks, then the corresponding system .S; « / is a matroid. Proof Every subset of an independent set of tasks is certainly independent. To prove the exchange property, suppose that B and A are independent sets of tasks and that jBj > jAj. Let k be the largest t such that Nt .B/ Ä Nt .A/. (Such a value of t exists, since N0.A/ D N0.B/ D 0.) Since Nn.B/ D jBj and Nn.A/ D jAj, but jBj > jAj, we must have that k < n and that Nj .B/ > Nj .A/ for all j in the range k C 1 Ä j Ä n. Therefore, B contains more tasks with deadline k C 1 than A does. Let ai be a task in B A with deadline k C 1. Let A0 D A [ fai g. We now show that A0 must be independent by using property 2 of Lemma 16.12. For 0 Ä t Ä k, we have Nt .A0/ D Nt .A/ Ä t, since A is independent. For k < t Ä n, we have Nt .A0/ Ä Nt .B/ Ä t, since B is independent. Therefore, A0 is independent, completing our proof that .S; « / is a matroid. By Theorem 16.11, we can use a greedy algorithm to find a maximum-weight independent set of tasks A. We can then create an optimal schedule having the tasks in A as its early tasks. This method is an efficient algorithm for scheduling 446 Chapter 16 Greedy Algorithms Task ai 1 2 3 4 5 6 7 di 4 2 4 3 1 4 6 wi 70 60 50 40 30 20 10 Figure 16.7 An instance of the problem of scheduling unit-time tasks with deadlines and penalties for a single processor. unit-time tasks with deadlines and penalties for a single processor. The running time is O.n2/ using GREEDY, since each of the O.n/ independence checks made by that algorithm takes time O.n/ (see Exercise 16.5-2). Problem 16-4 gives a faster implementation. Figure 16.7 demonstrates an example of the problem of scheduling unit-time tasks with deadlines and penalties for a single processor. In this example, the greedy algorithm selects, in order, tasks a1, a2, a3, and a4, then rejects a5 (because N4.fa1; a2; a3; a4; a5g/ D 5) and a6 (because N4.fa1; a2; a3; a4; a6g/ D 5), and finally accepts a7. The final optimal schedule is ha2; a4; a1; a3; a7; a5; a6i ; which has a total penalty incurred of w5 C w6 D 50. Exercises 16.5-1 Solve the instance of the scheduling problem given in Figure 16.7, but with each penalty wi replaced by 80 wi . 16.5-2 Show how to use property 2 of Lemma 16.12 to determine in time O.jAj/ whether or not a given set A of tasks is independent. Problems 16-1 Coin changing Consider the problem of making change for n cents using the fewest number of coins. Assume that each coin’s value is an integer. a. Describe a greedy algorithm to make change consisting of quarters, dimes, nickels, and pennies. Prove that your algorithm yields an optimal solution. Problems for Chapter 16 447 b. Suppose that the available coins are in the denominations that are powers of c, i.e., the denominations are c0; c1; : : : ; ck for some integers c > 1 and k 1. Show that the greedy algorithm always yields an optimal solution. c. Give a set of coin denominations for which the greedy algorithm does not yield an optimal solution. Your set should include a penny so that there is a solution for every value of n. d. Give an O.nk/-time algorithm that makes change for any set of k different coin denominations, assuming that one of the coins is a penny. 16-2 Scheduling to minimize average completion time Suppose you are given a set S D fa1; a2; : : : ; ang of tasks, where task ai re- quires pi units of processing time to complete, once it has started. You have one computer on which to run these tasks, and the computer can run only one task at a time. Let ci be the completion time of task ai , that is, the time at which task ai com- pletes processing. to minimize .1=n/ P YoniuDr1gcoi a. lFiosrtoexmaminpimle,izseupthpeosaevethraegree completion time, are two tasks, a1 that is, and a2, with p1 D 3 and p2 D 5, and consider the schedule in which a2 runs first, followed by a1. Then c2 D 5, c1 D 8, and the average completion time is .5 C 8/=2 D 6:5. If task a1 runs first, however, then c1 D 3, c2 D 8, and the average completion time is .3 C 8/=2 D 5:5. a. Give an algorithm that schedules the tasks so as to minimize the average completion time. Each task must run non-preemptively, that is, once task ai starts, it must run continuously for pi units of time. Prove that your algorithm minimizes the average completion time, and state the running time of your algorithm. b. Suppose now that the tasks are not all available at once. That is, each task cannot start until its release time ri . Suppose also that we allow preemption, so that a task can be suspended and restarted at a later time. For example, a task ai with processing time pi D 6 and release time ri D 1 might start running at time 1 and be preempted at time 4. It might then resume at time 10 but be preempted at time 11, and it might finally resume at time 13 and complete at time 15. Task ai has run for a total of 6 time units, but its running time has been divided into three pieces. In this scenario, ai ’s completion time is 15. Give an algorithm that schedules the tasks so as to minimize the average completion time in this new scenario. Prove that your algorithm minimizes the average completion time, and state the running time of your algorithm. 448 Chapter 16 Greedy Algorithms 16-3 Acyclic subgraphs a. The incidence matrix for an undirected graph G D .V; E/ is a jV j jEj ma- trix M such that M e D 1 if edge e is incident on vertex , and M e D 0 otherwise. Argue that a set of columns of M is linearly independent over the field of integers modulo 2 if and only if the corresponding set of edges is acyclic. Then, use the result of Exercise 16.4-2 to provide an alternate proof that .E; « / of part (a) is a matroid. b. Suppose that we associate a nonnegative weight w.e/ with each edge in an undirected graph G D .V; E/. Give an efficient algorithm to find an acyclic subset of E of maximum total weight. c. Let G.V; E/ be an arbitrary directed graph, and let .E; « / be defined so that A 2 « if and only if A does not contain any directed cycles. Give an example of a directed graph G such that the associated system .E; « / is not a matroid. Specify which defining condition for a matroid fails to hold. d. The incidence matrix for a directed graph G D .V; E/ with no self-loops is a jV j jEj matrix M such that M e D 1 if edge e leaves vertex , M e D 1 if edge e enters vertex , and M e D 0 otherwise. Argue that if a set of columns of M is linearly independent, then the corresponding set of edges does not contain a directed cycle. e. Exercise 16.4-2 tells us that the set of linearly independent sets of columns of any matrix M forms a matroid. Explain carefully why the results of parts (d) and (e) are not contradictory. How can there fail to be a perfect correspondence between the notion of a set of edges being acyclic and the notion of the associated set of columns of the incidence matrix being linearly independent? 16-4 Scheduling variations Consider the following algorithm for the problem from Section 16.5 of scheduling unit-time tasks with deadlines and penalties. Let all n time slots be initially empty, where time slot i is the unit-length slot of time that finishes at time i. We consider the tasks in order of monotonically decreasing penalty. When considering task aj , if there exists a time slot at or before aj ’s deadline dj that is still empty, assign aj to the latest such slot, filling it. If there is no such slot, assign task aj to the latest of the as yet unfilled slots. a. Argue that this algorithm always gives an optimal answer. b. Use the fast disjoint-set forest presented in Section 21.3 to implement the algorithm efficiently. Assume that the set of input tasks has already been sorted into Problems for Chapter 16 449 monotonically decreasing order by penalty. Analyze the running time of your implementation. 16-5 Off-line caching Modern computers use a cache to store a small amount of data in a fast memory. Even though a program may access large amounts of data, by storing a small subset of the main memory in the cache—a small but faster memory—overall access time can greatly decrease. When a computer program executes, it makes a sequence hr1; r2; : : : ; rni of n memory requests, where each request is for a particular data element. For example, a program that accesses 4 distinct elements fa; b; c; d g might make the sequence of requests hd; b; d; b; d; a; c; d; b; a; c; bi. Let k be the size of the cache. When the cache contains k elements and the program requests the .k C 1/st element, the system must decide, for this and each subsequent request, which k elements to keep in the cache. More precisely, for each request ri , the cache-management algorithm checks whether element ri is already in the cache. If it is, then we have a cache hit; otherwise, we have a cache miss. Upon a cache miss, the system retrieves ri from the main memory, and the cache-management algorithm must decide whether to keep ri in the cache. If it decides to keep ri and the cache already holds k elements, then it must evict one element to make room for ri . The cache-management algorithm evicts data with the goal of minimizing the number of cache misses over the entire sequence of requests. Typically, caching is an on-line problem. That is, we have to make decisions about which data to keep in the cache without knowing the future requests. Here, however, we consider the off-line version of this problem, in which we are given in advance the entire sequence of n requests and the cache size k, and we wish to minimize the total number of cache misses. We can solve this off-line problem by a greedy strategy called furthest-in-future, which chooses to evict the item in the cache whose next access in the request sequence comes furthest in the future. a. Write pseudocode for a cache manager that uses the furthest-in-future strategy. The input should be a sequence hr1; r2; : : : ; rni of requests and a cache size k, and the output should be a sequence of decisions about which data element (if any) to evict upon each request. What is the running time of your algorithm? b. Show that the off-line caching problem exhibits optimal substructure. c. Prove that furthest-in-future produces the minimum possible number of cache misses. 450 Chapter 16 Greedy Algorithms Chapter notes Much more material on greedy algorithms and matroids can be found in Lawler [224] and Papadimitriou and Steiglitz [271]. The greedy algorithm first appeared in the combinatorial optimization literature in a 1971 article by Edmonds [101], though the theory of matroids dates back to a 1935 article by Whitney [355]. Our proof of the correctness of the greedy algorithm for the activity-selection problem is based on that of Gavril [131]. The task-scheduling problem is studied in Lawler [224]; Horowitz, Sahni, and Rajasekaran [181]; and Brassard and Bratley [54]. Huffman codes were invented in 1952 [185]; Lelewer and Hirschberg [231] surveys data-compression techniques known as of 1987. An extension of matroid theory to greedoid theory was pioneered by Korte and Lova´sz [216, 217, 218, 219], who greatly generalize the theory presented here. 17 Amortized Analysis In an amortized analysis, we average the time required to perform a sequence of data-structure operations over all the operations performed. With amortized analysis, we can show that the average cost of an operation is small, if we average over a sequence of operations, even though a single operation within the sequence might be expensive. Amortized analysis differs from average-case analysis in that probability is not involved; an amortized analysis guarantees the average performance of each operation in the worst case. The first three sections of this chapter cover the three most common techniques used in amortized analysis. Section 17.1 starts with aggregate analysis, in which we determine an upper bound T .n/ on the total cost of a sequence of n operations. The average cost per operation is then T .n/=n. We take the average cost as the amortized cost of each operation, so that all operations have the same amortized cost. Section 17.2 covers the accounting method, in which we determine an amortized cost of each operation. When there is more than one type of operation, each type of operation may have a different amortized cost. The accounting method overcharges some operations early in the sequence, storing the overcharge as “prepaid credit” on specific objects in the data structure. Later in the sequence, the credit pays for operations that are charged less than they actually cost. Section 17.3 discusses the potential method, which is like the accounting method in that we determine the amortized cost of each operation and may overcharge operations early on to compensate for undercharges later. The potential method maintains the credit as the “potential energy” of the data structure as a whole instead of associating the credit with individual objects within the data structure. We shall use two examples to examine these three methods. One is a stack with the additional operation MULTIPOP, which pops several objects at once. The other is a binary counter that counts up from 0 by means of the single operation INCREMENT. 452 Chapter 17 Amortized Analysis While reading this chapter, bear in mind that the charges assigned during an amortized analysis are for analysis purposes only. They need not—and should not—appear in the code. If, for example, we assign a credit to an object x when using the accounting method, we have no need to assign an appropriate amount to some attribute, such as x:credit, in the code. When we perform an amortized analysis, we often gain insight into a particular data structure, and this insight can help us optimize the design. In Section 17.4, for example, we shall use the potential method to analyze a dynamically expanding and contracting table. 17.1 Aggregate analysis In aggregate analysis, we show that for all n, a sequence of n operations takes worst-case time T .n/ in total. In the worst case, the average cost, or amortized cost, per operation is therefore T .n/=n. Note that this amortized cost applies to each operation, even when there are several types of operations in the sequence. The other two methods we shall study in this chapter, the accounting method and the potential method, may assign different amortized costs to different types of operations. Stack operations In our first example of aggregate analysis, we analyze stacks that have been augmented with a new operation. Section 10.1 presented the two fundamental stack operations, each of which takes O.1/ time: PUSH.S; x/ pushes object x onto stack S. POP.S/ pops the top of stack S and returns the popped object. Calling POP on an empty stack generates an error. Since each of these operations runs in O.1/ time, let us consider the cost of each to be 1. The total cost of a sequence of n PUSH and POP operations is therefore n, and the actual running time for n operations is therefore ‚.n/. Now we add the stack operation MULTIPOP.S; k/, which removes the k top objects of stack S, popping the entire stack if the stack contains fewer than k objects. Of course, we assume that k is positive; otherwise the MULTIPOP operation leaves the stack unchanged. In the following pseudocode, the operation STACK-EMPTY returns TRUE if there are no objects currently on the stack, and FALSE otherwise. 17.1 Aggregate analysis 453 top 23 17 6 39 10 47 top 10 47 (a) (b) (c) Figure 17.1 The action of MULTIPOP on a stack S , shown initially in (a). The top 4 objects are popped by MULTIPOP.S; 4/, whose result is shown in (b). The next operation is MULTIPOP.S; 7/, which empties the stack—shown in (c)—since there were fewer than 7 objects remaining. MULTIPOP.S; k/ 1 while not STACK-EMPTY.S / and k > 0 2 POP.S / 3 kDk 1 Figure 17.1 shows an example of MULTIPOP. What is the running time of MULTIPOP.S; k/ on a stack of s objects? The actual running time is linear in the number of POP operations actually executed, and thus we can analyze MULTIPOP in terms of the abstract costs of 1 each for PUSH and POP. The number of iterations of the while loop is the number min.s; k/ of objects popped off the stack. Each iteration of the loop makes one call to POP in line 2. Thus, the total cost of MULTIPOP is min.s; k/, and the actual running time is a linear function of this cost. Let us analyze a sequence of n PUSH, POP, and MULTIPOP operations on an initially empty stack. The worst-case cost of a MULTIPOP operation in the sequence is O.n/, since the stack size is at most n. The worst-case time of any stack operation is therefore O.n/, and hence a sequence of n operations costs O.n2/, since we may have O.n/ MULTIPOP operations costing O.n/ each. Although this analysis is correct, the O.n2/ result, which we obtained by considering the worst-case cost of each operation individually, is not tight. Using aggregate analysis, we can obtain a better upper bound that considers the entire sequence of n operations. In fact, although a single MULTIPOP operation can be expensive, any sequence of n PUSH, POP, and MULTIPOP operations on an initially empty stack can cost at most O.n/. Why? We can pop each object from the stack at most once for each time we have pushed it onto the stack. Therefore, the number of times that POP can be called on a nonempty stack, including calls within MULTIPOP, is at most the number of PUSH operations, which is at most n. For any value of n, any sequence of n PUSH, POP, and MULTIPOP operations takes a total of O.n/ time. The average cost of an operation is O.n/=n D O.1/. In aggregate 454 Chapter 17 Amortized Analysis analysis, we assign the amortized cost of each operation to be the average cost. In this example, therefore, all three stack operations have an amortized cost of O.1/. We emphasize again that although we have just shown that the average cost, and hence the running time, of a stack operation is O.1/, we did not use probabilistic reasoning. We actually showed a worst-case bound of O.n/ on a sequence of n operations. Dividing this total cost by n yielded the average cost per operation, or the amortized cost. Incrementing a binary counter As another example of aggregate analysis, consider the problem of implementing a k-bit binary counter that counts upward from 0. We use an array AŒ0 : : k 1 of bits, where A:length D k, as the counter. A binary number x that is stored in the cxoDuntPerkihDa01sAitŒsi lowest-order bit  2i . Initially, x in AŒ0 and its D 0, and thus highest-order AŒi D 0 for i bit D in 0; AŒk 1; : : : ; 1, k so that 1. To add 1 (modulo 2k) to the value in the counter, we use the following procedure. INCREMENT.A/ 1 iD0 2 while i < A:length and AŒi == 1 3 AŒi D 0 4 i D iC1 5 if i < A:length 6 AŒi D 1 Figure 17.2 shows what happens to a binary counter as we increment it 16 times, starting with the initial value 0 and ending with the value 16. At the start of each iteration of the while loop in lines 2–4, we wish to add a 1 into position i. If AŒi D 1, then adding 1 flips the bit to 0 in position i and yields a carry of 1, to be added into position i C 1 on the next iteration of the loop. Otherwise, the loop ends, and then, if i < k, we know that AŒi D 0, so that line 6 adds a 1 into position i , flipping the 0 to a 1. The cost of each INCREMENT operation is linear in the number of bits flipped. As with the stack example, a cursory analysis yields a bound that is correct but not tight. A single execution of INCREMENT takes time ‚.k/ in the worst case, in which array A contains all 1s. Thus, a sequence of n INCREMENT operations on an initially zero counter takes time O.nk/ in the worst case. We can tighten our analysis to yield a worst-case cost of O.n/ for a sequence of n INCREMENT operations by observing that not all bits flip each time INCREMENT is called. As Figure 17.2 shows, AŒ0 does flip each time INCREMENT is called. The next bit up, AŒ1, flips only every other time: a sequence of n INCREMENT 17.1 Aggregate analysis 455 Counter value 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 A[7] A[6] A[5] A[4] A[3] A[2] A[1] A[0] 00000000 00000001 00000010 00000011 00000100 00000101 00000110 00000111 00001000 00001001 00001010 00001011 00001100 00001101 00001110 00001111 00010000 Total cost 0 1 3 4 7 8 10 11 15 16 18 19 22 23 25 26 31 Figure 17.2 An 8-bit binary counter as its value goes from 0 to 16 by a sequence of 16 INCREMENT operations. Bits that flip to achieve the next value are shaded. The running cost for flipping bits is shown at the right. Notice that the total cost is always less than twice the total number of INCREMENT operations. operations on an initially zero counter causes AŒ1 to flip bn=2c times. Similarly, bit AŒ2 flips only every fourth time, or bn=4c times in a sequence of n INCREMENT operations. In general, for i D 0; 1; : : : ; k 1, bit AŒi flips bn=2i c times in a sequence of n INCREMENT operations on an initially zero counter. For i k, bit AŒi does not exist, and so it cannot flip. The total number of flips in the sequence is thus X k 1 j n k X 1 1 0, then bi D bi 1 ti C 1. In either case, bi Ä bi 1 ti C 1, and the potential difference is ˆ.Di / ˆ.Di 1/ Ä .bi 1 ti C 1/ bi 1 D 1 ti : The amortized cost is therefore cyi D ci C ˆ.Di / ˆ.Di 1/ Ä .ti C 1/ C .1 ti / D 2: If the counter starts at zero, then ˆ.D0/ D 0. Since ˆ.Di / 0 for all i , the total amortized cost of a sequence of n INCREMENT operations is an upper bound on the total actual cost, and so the worst-case cost of n INCREMENT operations is O.n/. The potential method gives us an easy way to analyze the counter even when it does not start at zero. The counter starts with b0 1s, and after n INCREMENT 462 Chapter 17 Amortized Analysis operations it has bn 1s, where 0 Ä b0; bn Ä k. (Recall that k is the number of bits in the counter.) We can rewrite equation (17.3) as Xn Xn ci D cyi i D1 i D1 ˆ.Dn/ C ˆ.D0/ : (17.4) We have cyi Ä 2 for all 1 Ä i Ä n. Since ˆ.D0/ D b0 and ˆ.Dn/ D bn, the total actual cost of n INCREMENT operations is Xn Xn ci Ä 2 bn C b0 i D1 i D1 D 2n bn C b0 : Note in particular that since b0 Ä k, as long as k D O.n/, the total actual cost is O.n/. In other words, if we execute at least n D .k/ INCREMENT operations, the total actual cost is O.n/, no matter what initial value the counter contains. Exercises 17.3-1 Suppose we have a potential function ˆ such that ˆ.Di / ˆ.D0/ for all i , but ˆ.D0/ ¤ 0. Show that there exists a potential function ˆ0 such that ˆ0.D0/ D 0, ˆ0.Di / 0 for all i 1, and the amortized costs using ˆ0 are the same as the amortized costs using ˆ. 17.3-2 Redo Exercise 17.1-3 using a potential method of analysis. 17.3-3 Consider an ordinary binary min-heap data structure with n elements supporting the instructions INSERT and EXTRACT-MIN in O.lg n/ worst-case time. Give a potential function ˆ such that the amortized cost of INSERT is O.lg n/ and the amortized cost of EXTRACT-MIN is O.1/, and show that it works. 17.3-4 What is the total cost of executing n of the stack operations PUSH, POP, and MULTIPOP, assuming that the stack begins with s0 objects and finishes with sn objects? 17.3-5 Suppose that a counter begins at a number with b 1s in its binary representation, rather than at 0. Show that the cost of performing n INCREMENT operations is O.n/ if n D .b/. (Do not assume that b is constant.) 17.4 Dynamic tables 463 17.3-6 Show how to implement a queue with two ordinary stacks (Exercise 10.1-6) so that the amortized cost of each ENQUEUE and each DEQUEUE operation is O.1/. 17.3-7 Design a data structure to support the following two operations for a dynamic multiset S of integers, which allows duplicate values: INSERT.S; x/ inserts x into S . DELETE-LARGER-HALF.S / deletes the largest djS j =2e elements from S . Explain how to implement this data structure so that any sequence of m INSERT and DELETE-LARGER-HALF operations runs in O.m/ time. Your implementation should also include a way to output the elements of S in O.jSj/ time. 17.4 Dynamic tables We do not always know in advance how many objects some applications will store in a table. We might allocate space for a table, only to find out later that it is not enough. We must then reallocate the table with a larger size and copy all objects stored in the original table over into the new, larger table. Similarly, if many objects have been deleted from the table, it may be worthwhile to reallocate the table with a smaller size. In this section, we study this problem of dynamically expanding and contracting a table. Using amortized analysis, we shall show that the amortized cost of insertion and deletion is only O.1/, even though the actual cost of an operation is large when it triggers an expansion or a contraction. Moreover, we shall see how to guarantee that the unused space in a dynamic table never exceeds a constant fraction of the total space. We assume that the dynamic table supports the operations TABLE-INSERT and TABLE-DELETE. TABLE-INSERT inserts into the table an item that occupies a single slot, that is, a space for one item. Likewise, TABLE-DELETE removes an item from the table, thereby freeing a slot. The details of the data-structuring method used to organize the table are unimportant; we might use a stack (Section 10.1), a heap (Chapter 6), or a hash table (Chapter 11). We might also use an array or collection of arrays to implement object storage, as we did in Section 10.3. We shall find it convenient to use a concept introduced in our analysis of hashing (Chapter 11). We define the load factor ˛.T / of a nonempty table T to be the number of items stored in the table divided by the size (number of slots) of the table. We assign an empty table (one with no items) size 0, and we define its load factor to be 1. If the load factor of a dynamic table is bounded below by a constant, 464 Chapter 17 Amortized Analysis the unused space in the table is never more than a constant fraction of the total amount of space. We start by analyzing a dynamic table in which we only insert items. We then consider the more general case in which we both insert and delete items. 17.4.1 Table expansion Let us assume that storage for a table is allocated as an array of slots. A table fills up when all slots have been used or, equivalently, when its load factor is 1.1 In some software environments, upon attempting to insert an item into a full table, the only alternative is to abort with an error. We shall assume, however, that our software environment, like many modern ones, provides a memory-management system that can allocate and free blocks of storage on request. Thus, upon inserting an item into a full table, we can expand the table by allocating a new table with more slots than the old table had. Because we always need the table to reside in contiguous memory, we must allocate a new array for the larger table and then copy items from the old table into the new table. A common heuristic allocates a new table with twice as many slots as the old one. If the only table operations are insertions, then the load factor of the table is always at least 1=2, and thus the amount of wasted space never exceeds half the total space in the table. In the following pseudocode, we assume that T is an object representing the table. The attribute T:table contains a pointer to the block of storage representing the table, T:num contains the number of items in the table, and T:size gives the total number of slots in the table. Initially, the table is empty: T:num D T:size D 0. TABLE-INSERT .T; x/ 1 if T:size == 0 2 allocate T:table with 1 slot 3 T:size D 1 4 if T:num == T:size 5 allocate new-table with 2 T:size slots 6 insert all items in T:table into new-table 7 free T:table 8 T:table D new-table 9 T:size D 2 T:size 10 insert x into T:table 11 T:num D T:num C 1 1In some situations, such as an open-address hash table, we may wish to consider a table to be full if its load factor equals some constant strictly less than 1. (See Exercise 17.4-1.) 17.4 Dynamic tables 465 Notice that we have two “insertion” procedures here: the TABLE-INSERT proce- dure itself and the elementary insertion into a table in lines 6 and 10. We can analyze the running time of TABLE-INSERT in terms of the number of elementary insertions by assigning a cost of 1 to each elementary insertion. We assume that the actual running time of TABLE-INSERT is linear in the time to insert individual items, so that the overhead for allocating an initial table in line 2 is constant and the overhead for allocating and freeing storage in lines 5 and 7 is dominated by the cost of transferring items in line 6. We call the event in which lines 5–9 are executed an expansion. Let us analyze a sequence of n TABLE-INSERT operations on an initially empty table. What is the cost ci of the ith operation? If the current table has room for the new item (or if this is the first operation), then ci D 1, since we need only perform the one elementary insertion in line 10. If the current table is full, however, and an expansion occurs, then ci D i: the cost is 1 for the elementary insertion in line 10 plus i 1 for the items that we must copy from the old table to the new table in line 6. If we perform n operations, the worst-case cost of an operation is O.n/, which leads to an upper bound of O.n2/ on the total running time for n operations. This bound is not tight, because we rarely expand the table in the course of n TABLE-INSERT operations. Specifically, the i th operation causes an expansion only when i 1 is an exact power of 2. The amortized cost of an operation is in fact O.1/, as we can show using aggregate analysis. The cost of the ith operation is ( ci D i if i 1 is an exact power of 2 ; 1 otherwise : The total cost of n TABLE-INSERT operations is therefore Xn bX lg nc ci Ä n C 2j i D1 j D0 < n C 2n D 3n ; because at most n operations cost 1 and the costs of the remaining operations form a geometric series. Since the total cost of n TABLE-INSERT operations is bounded by 3n, the amortized cost of a single operation is at most 3. By using the accounting method, we can gain some feeling for why the amortized cost of a TABLE-INSERT operation should be 3. Intuitively, each item pays for 3 elementary insertions: inserting itself into the current table, moving itself when the table expands, and moving another item that has already been moved once when the table expands. For example, suppose that the size of the table is m immediately after an expansion. Then the table holds m=2 items, and it contains 466 Chapter 17 Amortized Analysis no credit. We charge 3 dollars for each insertion. The elementary insertion that occurs immediately costs 1 dollar. We place another dollar as credit on the item inserted. We place the third dollar as credit on one of the m=2 items already in the table. The table will not fill again until we have inserted another m=2 1 items, and thus, by the time the table contains m items and is full, we will have placed a dollar on each item to pay to reinsert it during the expansion. We can use the potential method to analyze a sequence of n TABLE-INSERT operations, and we shall use it in Section 17.4.2 to design a TABLE-DELETE operation that has an O.1/ amortized cost as well. We start by defining a potential function ˆ that is 0 immediately after an expansion but builds to the table size by the time the table is full, so that we can pay for the next expansion by the potential. The function ˆ.T / D 2 T:num T:size (17.5) is one possibility. Immediately after an expansion, we have T:num D T:size=2, and thus ˆ.T / D 0, as desired. Immediately before an expansion, we have T:num D T:size, and thus ˆ.T / D T:num, as desired. The initial value of the potential is 0, and since the table is always at least half full, T:num T:size=2, which implies that ˆ.T / is always nonnegative. Thus, the sum of the amortized costs of n TABLE-INSERT operations gives an upper bound on the sum of the actual costs. To analyze the amortized cost of the i th TABLE-INSERT operation, we let numi denote the number of items stored in the table after the ith operation, sizei denote the total size of the table after the ith operation, and ˆi denote the potential after the i th operation. Initially, we have num0 D 0, size0 D 0, and ˆ0 D 0. If the i th TABLE-INSERT operation does not trigger an expansion, then we have sizei D sizei 1 and the amortized cost of the operation is cyi D ci C ˆi ˆi 1 D 1 C .2 numi sizei / .2 numi 1 sizei 1/ D 1 C .2 numi sizei / .2.numi 1/ sizei / D 3: If the i th operation does trigger an expansion, then we have sizei D 2 sizei 1 and sizei 1 D numi 1 D numi 1, which implies that sizei D 2 .numi 1/. Thus, the amortized cost of the operation is cyi D ci C ˆi ˆi 1 D numi C .2 numi sizei / .2 numi 1 sizei 1/ D numi C .2 numi 2 .numi 1// .2.numi 1/ .numi 1// D numi C 2 .numi 1/ D 3: 17.4 Dynamic tables 467 32 sizei numi 24 16 Φi 8 0 i 0 8 16 24 32 Figure 17.3 The effect of a sequence of n TABLE-INSERT operations on the number numi of items in the table, the number sizei of slots in the table, and the potential ˆi D 2 numi sizei , each being measured after the ith operation. The thin line shows numi , the dashed line shows sizei , and the thick line shows ˆi . Notice that immediately before an expansion, the potential has built up to the number of items in the table, and therefore it can pay for moving all the items to the new table. Afterwards, the potential drops to 0, but it is immediately increased by 2 upon inserting the item that caused the expansion. Figure 17.3 plots the values of numi , sizei , and ˆi against i . Notice how the potential builds to pay for expanding the table. 17.4.2 Table expansion and contraction To implement a TABLE-DELETE operation, it is simple enough to remove the specified item from the table. In order to limit the amount of wasted space, however, we might wish to contract the table when the load factor becomes too small. Table contraction is analogous to table expansion: when the number of items in the table drops too low, we allocate a new, smaller table and then copy the items from the old table into the new one. We can then free the storage for the old table by returning it to the memory-management system. Ideally, we would like to preserve two properties: the load factor of the dynamic table is bounded below by a positive constant, and the amortized cost of a table operation is bounded above by a constant. 468 Chapter 17 Amortized Analysis We assume that we measure the cost in terms of elementary insertions and deletions. You might think that we should double the table size upon inserting an item into a full table and halve the size when a deleting an item would cause the table to become less than half full. This strategy would guarantee that the load factor of the table never drops below 1=2, but unfortunately, it can cause the amortized cost of an operation to be quite large. Consider the following scenario. We perform n operations on a table T , where n is an exact power of 2. The first n=2 operations are insertions, which by our previous analysis cost a total of ‚.n/. At the end of this sequence of insertions, T:num D T:size D n=2. For the second n=2 operations, we perform the following sequence: insert, delete, delete, insert, insert, delete, delete, insert, insert, . . . . The first insertion causes the table to expand to size n. The two following deletions cause the table to contract back to size n=2. Two further insertions cause another expansion, and so forth. The cost of each expansion and contraction is ‚.n/, and there are ‚.n/ of them. Thus, the total cost of the n operations is ‚.n2/, making the amortized cost of an operation ‚.n/. The downside of this strategy is obvious: after expanding the table, we do not delete enough items to pay for a contraction. Likewise, after contracting the table, we do not insert enough items to pay for an expansion. We can improve upon this strategy by allowing the load factor of the table to drop below 1=2. Specifically, we continue to double the table size upon inserting an item into a full table, but we halve the table size when deleting an item causes the table to become less than 1=4 full, rather than 1=2 full as before. The load factor of the table is therefore bounded below by the constant 1=4. Intuitively, we would consider a load factor of 1=2 to be ideal, and the table’s potential would then be 0. As the load factor deviates from 1=2, the potential increases so that by the time we expand or contract the table, the table has garnered sufficient potential to pay for copying all the items into the newly allocated table. Thus, we will need a potential function that has grown to T:num by the time that the load factor has either increased to 1 or decreased to 1=4. After either expanding or contracting the table, the load factor goes back to 1=2 and the table’s potential reduces back to 0. We omit the code for TABLE-DELETE, since it is analogous to TABLE-INSERT. For our analysis, we shall assume that whenever the number of items in the table drops to 0, we free the storage for the table. That is, if T:num D 0, then T:size D 0. We can now use the potential method to analyze the cost of a sequence of n TABLE-INSERT and TABLE-DELETE operations. We start by defining a potential function ˆ that is 0 immediately after an expansion or contraction and builds as the load factor increases to 1 or decreases to 1=4. Let us denote the load fac- 17.4 Dynamic tables 469 32 24 sizei 16 numi 8 Φi 0 i 0 8 16 24 32 40 48 Figure 17.4 The effect of a sequence of n TABLE-INSERT and TABLE-DELETE operations on the number numi of items in the table, the number sizei of slots in the table, and the potential ˆi D 2 numi sizei sizei =2 numi if ˛i 1=2 ; if ˛i < 1=2 ; each measured after the ith operation. The thin line shows numi , the dashed line shows sizei , and the thick line shows ˆi . Notice that immediately before an expansion, the potential has built up to the number of items in the table, and therefore it can pay for moving all the items to the new table. Likewise, immediately before a contraction, the potential has built up to the number of items in the table. tor of a nonempty table T by ˛.T / D T:num=T:size. Since for an empty table, T:num D T:size D 0 and ˛.T / D 1, we always have T:num D ˛.T / T:size, whether the table is empty or not. We shall use as our potential function ( ˆ.T / D 2 T:num T:size if ˛.T / 1=2 ; T:size=2 T:num if ˛.T / < 1=2 : (17.6) Observe that the potential of an empty table is 0 and that the potential is never negative. Thus, the total amortized cost of a sequence of operations with respect to ˆ provides an upper bound on the actual cost of the sequence. Before proceeding with a precise analysis, we pause to observe some properties of the potential function, as illustrated in Figure 17.4. Notice that when the load factor is 1=2, the potential is 0. When the load factor is 1, we have T:size D T:num, which implies ˆ.T / D T:num, and thus the potential can pay for an expansion if an item is inserted. When the load factor is 1=4, we have T:size D 4 T:num, which 470 Chapter 17 Amortized Analysis implies ˆ.T / D T:num, and thus the potential can pay for a contraction if an item is deleted. To analyze a sequence of n TABLE-INSERT and TABLE-DELETE operations, we let ci denote the actual cost of the i th operation, cyi denote its amortized cost with respect to ˆ, numi denote the number of items stored in the table after the ith operation, sizei denote the total size of the table after the i th operation, ˛i denote the load factor of the table after the ith operation, and ˆi denote the potential after the i th operation. Initially, num0 D 0, size0 D 0, ˛0 D 1, and ˆ0 D 0. We start with the case in which the i th operation is TABLE-INSERT. The analysis is identical to that for table expansion in Section 17.4.1 if ˛i 1 1=2. Whether the table expands or not, the amortized cost cyi of the operation is at most 3. If ˛i 1 < 1=2, the table cannot expand as a result of the operation, since the table expands only when ˛i 1 D 1. If ˛i < 1=2 as well, then the amortized cost of the ith operation is cyi D ci C ˆi ˆi 1 D 1 C .sizei =2 numi / .sizei 1=2 numi 1/ D 1 C .sizei =2 numi / .sizei =2 .numi 1// D 0: If ˛i 1 < 1=2 but ˛i 1=2, then cyi D ci C ˆi ˆi 1 D 1 C .2 numi sizei / .sizei 1=2 numi 1/ D 1 C .2.numi 1 C 1/ sizei 1/ 3 D 3 numi 1 2 sizei 1 C 3 3 D 3˛i 1sizei 1 2 sizei 1 C 3 3 3 < 2 sizei 1 2 sizei 1 C 3 D 3: .sizei 1=2 numi 1/ Thus, the amortized cost of a TABLE-INSERT operation is at most 3. We now turn to the case in which the i th operation is TABLE-DELETE. In this case, numi D numi 1 1. If ˛i 1 < 1=2, then we must consider whether the operation causes the table to contract. If it does not, then sizei D sizei 1 and the amortized cost of the operation is cyi D ci C ˆi ˆi 1 D 1 C .sizei =2 numi / .sizei 1=2 numi 1/ D 1 C .sizei =2 numi / .sizei =2 .numi C 1// D 2: 17.4 Dynamic tables 471 If ˛i 1 < 1=2 and the i th operation does trigger a contraction, then the actual cost of the operation is ci D numi C 1, since we delete one item and move numi items. We have sizei =2 D sizei 1=4 D numi 1 D numi C 1, and the amortized cost of the operation is cyi D ci C ˆi ˆi 1 D .numi C 1/ C .sizei =2 numi / .sizei 1=2 numi 1/ D .numi C 1/ C ..numi C 1/ numi / ..2 numi C 2/ .numi C 1// D 1: When the i th operation is a TABLE-DELETE and ˛i 1 1=2, the amortized cost is also bounded above by a constant. We leave the analysis as Exercise 17.4-2. In summary, since the amortized cost of each operation is bounded above by a constant, the actual time for any sequence of n operations on a dynamic table is O.n/. Exercises 17.4-1 Suppose that we wish to implement a dynamic, open-address hash table. Why might we consider the table to be full when its load factor reaches some value ˛ that is strictly less than 1? Describe briefly how to make insertion into a dynamic, open-address hash table run in such a way that the expected value of the amortized cost per insertion is O.1/. Why is the expected value of the actual cost per insertion not necessarily O.1/ for all insertions? 17.4-2 Show that if ˛i 1 1=2 and the i th operation on a dynamic table is TABLEDELETE, then the amortized cost of the operation with respect to the potential function (17.6) is bounded above by a constant. 17.4-3 Suppose that instead of contracting a table by halving its size when its load factor drops below 1=4, we contract it by multiplying its size by 2=3 when its load factor drops below 1=3. Using the potential function ˆ.T / D j2 T:num T:sizej ; show that the amortized cost of a TABLE-DELETE that uses this strategy is bounded above by a constant. 472 Chapter 17 Amortized Analysis Problems 17-1 Bit-reversed binary counter Chapter 30 examines an important algorithm called the fast Fourier transform, or FFT. The first step of the FFT algorithm performs a bit-reversal permutation on an input array AŒ0 : : n 1 whose length is n D 2k for some nonnegative integer k. This permutation swaps elements whose indices have binary representations that are the reverse of each other. a DWePcikaDn01 express each index ai 2i . We define a as a k-bit sequence hak 1; ak 2; : : : ; a0i, where revk.hak 1; ak 2; : : : ; a0i/ D ha0; a1; : : : ; ak 1i I thus, X k 1 revk.a/ D ak i 12i : i D0 For example, if n D 16 (or, equivalently, k D 4), then revk.3/ D 12, since the 4-bit representation of 3 is 0011, which when reversed gives 1100, the 4-bit representation of 12. a. Given a function revk that runs in ‚.k/ time, write an algorithm to perform the bit-reversal permutation on an array of length n D 2k in O.nk/ time. We can use an algorithm based on an amortized analysis to improve the running time of the bit-reversal permutation. We maintain a “bit-reversed counter” and a procedure BIT-REVERSED-INCREMENT that, when given a bit-reversed-counter value a, produces revk.revk.a/ C 1/. If k D 4, for example, and the bit-reversed counter starts at 0, then successive calls to BIT-REVERSED-INCREMENT produce the sequence 0000; 1000; 0100; 1100; 0010; 1010; : : : D 0; 8; 4; 12; 2; 10; : : : : b. Assume that the words in your computer store k-bit values and that in unit time, your computer can manipulate the binary values with operations such as shifting left or right by arbitrary amounts, bitwise-AND, bitwise-OR, etc. Describe an implementation of the BIT-REVERSED-INCREMENT procedure that allows the bit-reversal permutation on an n-element array to be performed in a total of O.n/ time. c. Suppose that you can shift a word left or right by only one bit in unit time. Is it still possible to implement an O.n/-time bit-reversal permutation? Problems for Chapter 17 473 17-2 Making binary search dynamic Binary search of a sorted array takes logarithmic search time, but the time to insert a new element is linear in the size of the array. We can improve the time for insertion by keeping several sorted arrays. Specifically, suppose that we wish to support SEARCH and INSERT on a set of n elements. Let k D dlg.n C 1/e, and let the binary representation of n be hnk 1; nk 2; : : : ; n0i. We have k sorted arrays A0; A1; : : : ; Ak 1, where for i D 0; 1; : : : ; k 1, the length of array Ai is 2i . Each array is either full or empty, depending on ments held in whether ni all k arrays D is th1eorerfnorieDP0kiD, 01rensipe2ci tiDvelny.. The total Although number of eleeach individual array is sorted, elements in different arrays bear no particular relationship to each other. a. Describe how to perform the SEARCH operation for this data structure. Analyze its worst-case running time. b. Describe how to perform the INSERT operation. Analyze its worst-case and amortized running times. c. Discuss how to implement DELETE. 17-3 Amortized weight-balanced trees Consider an ordinary binary search tree augmented by adding to each node x the attribute x:size giving the number of keys stored in the subtree rooted at x. Let ˛ be a constant in the range 1=2 Ä ˛ < 1. We say that a given node x is ˛-balanced if x:left:size Ä ˛ x:size and x:right:size Ä ˛ x:size. The tree as a whole is ˛-balanced if every node in the tree is ˛-balanced. The following amortized approach to maintaining weight-balanced trees was suggested by G. Varghese. a. A 1=2-balanced tree is, in a sense, as balanced as it can be. Given a node x in an arbitrary binary search tree, show how to rebuild the subtree rooted at x so that it becomes 1=2-balanced. Your algorithm should run in time ‚.x:size/, and it can use O.x:size/ auxiliary storage. b. Show that performing a search in an n-node ˛-balanced binary search tree takes O.lg n/ worst-case time. For the remainder of this problem, assume that the constant ˛ is strictly greater than 1=2. Suppose that we implement INSERT and DELETE as usual for an n-node binary search tree, except that after every such operation, if any node in the tree is no longer ˛-balanced, then we “rebuild” the subtree rooted at the highest such node in the tree so that it becomes 1=2-balanced. 474 Chapter 17 Amortized Analysis We shall analyze this rebuilding scheme using the potential method. For a node x in a binary search tree T , we define .x/ D jx:left:size x:right:sizej ; and we define the potential of T as X ˆ.T / D c .x/ ; x2T W.x/ 2 where c is a sufficiently large constant that depends on ˛. c. Argue that any binary search tree has nonnegative potential and that a 1=2balanced tree has potential 0. d. Suppose that m units of potential can pay for rebuilding an m-node subtree. How large must c be in terms of ˛ in order for it to take O.1/ amortized time to rebuild a subtree that is not ˛-balanced? e. Show that inserting a node into or deleting a node from an n-node ˛-balanced tree costs O.lg n/ amortized time. 17-4 The cost of restructuring red-black trees There are four basic operations on red-black trees that perform structural modifications: node insertions, node deletions, rotations, and color changes. We have seen that RB-INSERT and RB-DELETE use only O.1/ rotations, node insertions, and node deletions to maintain the red-black properties, but they may make many more color changes. a. Describe a legal red-black tree with n nodes such that calling RB-INSERT to add the .n C 1/st node causes .lg n/ color changes. Then describe a legal red-black tree with n nodes for which calling RB-DELETE on a particular node causes .lg n/ color changes. Although the worst-case number of color changes per operation can be logarithmic, we shall prove that any sequence of m RB-INSERT and RB-DELETE operations on an initially empty red-black tree causes O.m/ structural modifications in the worst case. Note that we count each color change as a structural modification. b. Some of the cases handled by the main loop of the code of both RB-INSERTFIXUP and RB-DELETE-FIXUP are terminating: once encountered, they cause the loop to terminate after a constant number of additional operations. For each of the cases of RB-INSERT-FIXUP and RB-DELETE-FIXUP, specify which are terminating and which are not. (Hint: Look at Figures 13.5, 13.6, and 13.7.) Problems for Chapter 17 475 We shall first analyze the structural modifications when only insertions are performed. Let T be a red-black tree, and define ˆ.T / to be the number of red nodes in T . Assume that 1 unit of potential can pay for the structural modifications performed by any of the three cases of RB-INSERT-FIXUP. c. Let T 0 be the result of applying Case 1 of RB-INSERT-FIXUP to T . Argue that ˆ.T 0/ D ˆ.T / 1. d. When we insert a node into a red-black tree using RB-INSERT, we can break the operation into three parts. List the structural modifications and potential changes resulting from lines 1–16 of RB-INSERT, from nonterminating cases of RB-INSERT-FIXUP, and from terminating cases of RB-INSERT-FIXUP. e. Using part (d), argue that the amortized number of structural modifications performed by any call of RB-INSERT is O.1/. We now wish to prove that there are O.m/ structural modifications when there are „ both insertions and deletions. Let us define, for each node x, 0 if x is red ; w.x/ D 1 if x is black and has no red children ; 0 if x is black and has one red child ; 2 if x is black and has two red children : Now we redefine the potential of a red-black tree T as X ˆ.T / D w.x/ ; x2T and let T 0 be the tree that results from applying any nonterminating case of RBINSERT-FIXUP or RB-DELETE-FIXUP to T . f. Show that ˆ.T 0/ Ä ˆ.T / 1 for all nonterminating cases of RB-INSERTFIXUP. Argue that the amortized number of structural modifications performed by any call of RB-INSERT-FIXUP is O.1/. g. Show that ˆ.T 0/ Ä ˆ.T / 1 for all nonterminating cases of RB-DELETEFIXUP. Argue that the amortized number of structural modifications performed by any call of RB-DELETE-FIXUP is O.1/. h. Complete the proof that in the worst case, any sequence of m RB-INSERT and RB-DELETE operations performs O.m/ structural modifications. 476 Chapter 17 Amortized Analysis 17-5 Competitive analysis of self-organizing lists with move-to-front A self-organizing list is a linked list of n elements, in which each element has a unique key. When we search for an element in the list, we are given a key, and we want to find an element with that key. A self-organizing list has two important properties: 1. To find an element in the list, given its key, we must traverse the list from the beginning until we encounter the element with the given key. If that element is the kth element from the start of the list, then the cost to find the element is k. 2. We may reorder the list elements after any operation, according to a given rule with a given cost. We may choose any heuristic we like to decide how to reorder the list. Assume that we start with a given list of n elements, and we are given an access sequence D h 1; 2; : : : ; mi of keys to find, in order. The cost of the sequence is the sum of the costs of the individual accesses in the sequence. Out of the various possible ways to reorder the list after an operation, this problem focuses on transposing adjacent list elements—switching their positions in the list—with a unit cost for each transpose operation. You will show, by means of a potential function, that a particular heuristic for reordering the list, move-to-front, entails a total cost no worse than 4 times that of any other heuristic for maintaining the list order—even if the other heuristic knows the access sequence in advance! We call this type of analysis a competitive analysis. For a heuristic H and a given initial ordering of the list, denote the access cost of sequence by CH . /. Let m be the number of accesses in . a. Argue that if heuristic H does not know the access sequence in advance, then the worst-case cost for H on an access sequence is CH . / D .mn/. With the move-to-front heuristic, immediately after searching for an element x, we move x to the first position on the list (i.e., the front of the list). Let rankL.x/ denote the rank of element x in list L, that is, the position of x in list L. For example, if x is the fourth element in L, then rankL.x/ D 4. Let ci denote the cost of access i using the move-to-front heuristic, which includes the cost of finding the element in the list and the cost of moving it to the front of the list by a series of transpositions of adjacent list elements. b. Show that if i accesses element x in list L using the move-to-front heuristic, then ci D 2 rankL.x/ 1. Now we compare move-to-front with any other heuristic H that processes an access sequence according to the two properties above. Heuristic H may transpose Problems for Chapter 17 477 elements in the list in any way it wants, and it might even know the entire access sequence in advance. Let Li be the list after access i using move-to-front, and let Li be the list after access i using heuristic H. We denote the cost of access i by ci for move-tofront and by ci for heuristic H. Suppose that heuristic H performs ti transpositions during access i . c. In part (b), you showed that ci D 2 rankLi 1.x/ 1. Now show that ci D rankLi .x/ 1 C ti . We define an inversion in list Li as a pair of elements y and ´ such that y precedes ´ in Li and ´ precedes y in list Li . Suppose that list Li has qi inversions after processing the access sequence h 1; 2; : : : ; i i. Then, we define a potential function ˆ that maps Li to a real number by ˆ.Li / D 2qi . For example, if Li has the elements he; c; a; d; bi and Li has the elements hc; a; b; d; ei, then Li has 5 inversions (.e; c/; .e; a/; .e; d /; .e; b/; .d; b/), and so ˆ.Li / D 10. Observe that ˆ.Li / 0 for all i and that, if move-to-front and heuristic H start with the same list L0, then ˆ.L0/ D 0. d. Argue that a transposition either increases the potential by 2 or decreases the potential by 2. Suppose that access i finds the element x. To understand how the potential changes due to i , let us partition the elements other than x into four sets, depending on where they are in the lists just before the ith access: Set A consists of elements that precede x in both Li 1 and Li 1. Set B consists of elements that precede x in Li 1 and follow x in Li 1. Set C consists of elements that follow x in Li 1 and precede x in Li 1. Set D consists of elements that follow x in both Li 1 and Li 1. e. Argue that rankLi 1.x/ D jAj C jBj C 1 and rankLi .x/ D jAj C jC j C 1. 1 f. Show that access i causes a change in potential of ˆ.Li / ˆ.Li 1/ Ä 2.jAj jBj C ti / ; where, as before, heuristic H performs ti transpositions during access i . Define the amortized cost cyi of access i by cyi D ci C ˆ.Li / ˆ.Li 1/. g. Show that the amortized cost cyi of access i is bounded from above by 4ci . h. Conclude that the cost CMTF. / of access sequence with move-to-front is at most 4 times the cost CH . / of with any other heuristic H, assuming that both heuristics start with the same list. 478 Chapter 17 Amortized Analysis Chapter notes Aho, Hopcroft, and Ullman [5] used aggregate analysis to determine the running time of operations on a disjoint-set forest; we shall analyze this data structure using the potential method in Chapter 21. Tarjan [331] surveys the accounting and potential methods of amortized analysis and presents several applications. He attributes the accounting method to several authors, including M. R. Brown, R. E. Tarjan, S. Huddleston, and K. Mehlhorn. He attributes the potential method to D. D. Sleator. The term “amortized” is due to D. D. Sleator and R. E. Tarjan. Potential functions are also useful for proving lower bounds for certain types of problems. For each configuration of the problem, we define a potential function that maps the configuration to a real number. Then we determine the potential ˆinit of the initial configuration, the potential ˆfinal of the final configuration, and the maximum change in potential ˆmax due to any step. The number of steps must therefore be at least jˆfinal ˆinitj = jˆmaxj. Examples of potential functions to prove lower bounds in I/O complexity appear in works by Cormen, Sundquist, and Wisniewski [79]; Floyd [107]; and Aggarwal and Vitter [3]. Krumme, Cybenko, and Venkataraman [221] applied potential functions to prove lower bounds on gossiping: communicating a unique item from each vertex in a graph to every other vertex. The move-to-front heuristic from Problem 17-5 works quite well in practice. Moreover, if we recognize that when we find an element, we can splice it out of its position in the list and relocate it to the front of the list in constant time, we can show that the cost of move-to-front is at most twice the cost of any other heuristic including, again, one that knows the entire access sequence in advance. V Advanced Data Structures Introduction This part returns to studying data structures that support operations on dynamic sets, but at a more advanced level than Part III. Two of the chapters, for example, make extensive use of the amortized analysis techniques we saw in Chapter 17. Chapter 18 presents B-trees, which are balanced search trees specifically designed to be stored on disks. Because disks operate much more slowly than random-access memory, we measure the performance of B-trees not only by how much computing time the dynamic-set operations consume but also by how many disk accesses they perform. For each B-tree operation, the number of disk accesses increases with the height of the B-tree, but B-tree operations keep the height low. Chapter 19 gives an implementation of a mergeable heap, which supports the operations INSERT, MINIMUM, EXTRACT-MIN, and UNION.1 The UNION operation unites, or merges, two heaps. Fibonacci heaps—the data structure in Chapter 19—also support the operations DELETE and DECREASE-KEY. We use amortized time bounds to measure the performance of Fibonacci heaps. The operations INSERT, MINIMUM, and UNION take only O.1/ actual and amortized time on Fibonacci heaps, and the operations EXTRACT-MIN and DELETE take O.lg n/ amortized time. The most significant advantage of Fibonacci heaps, however, is that DECREASE-KEY takes only O.1/ amortized time. Because the DECREASE- 1As in Problem 10-2, we have defined a mergeable heap to support MINIMUM and EXTRACT-MIN, and so we can also refer to it as a mergeable min-heap. Alternatively, if it supported MAXIMUM and EXTRACT-MAX, it would be a mergeable max-heap. Unless we specify otherwise, mergeable heaps will be by default mergeable min-heaps. 482 Part V Advanced Data Structures KEY operation takes constant amortized time, Fibonacci heaps are key components of some of the asymptotically fastest algorithms to date for graph problems. Noting that we can beat the .n lg n/ lower bound for sorting when the keys are integers in a restricted range, Chapter 20 asks whether we can design a data structure that supports the dynamic-set operations SEARCH, INSERT, DELETE, MINIMUM, MAXIMUM, SUCCESSOR, and PREDECESSOR in o.lg n/ time when the keys are integers in a restricted range. The answer turns out to be that we can, by using a recursive data structure known as a van Emde Boas tree. If the keys are unique integers drawn from the set f0; 1; 2; : : : ; u 1g, where u is an exact power of 2, then van Emde Boas trees support each of the above operations in O.lg lg u/ time. Finally, Chapter 21 presents data structures for disjoint sets. We have a universe of n elements that are partitioned into dynamic sets. Initially, each element belongs to its own singleton set. The operation UNION unites two sets, and the query FINDSET identifies the unique set that contains a given element at the moment. By representing each set as a simple rooted tree, we obtain surprisingly fast operations: a sequence of m operations runs in O.m ˛.n// time, where ˛.n/ is an incredibly slowly growing function—˛.n/ is at most 4 in any conceivable application. The amortized analysis that proves this time bound is as complex as the data structure is simple. The topics covered in this part are by no means the only examples of “advanced” data structures. Other advanced data structures include the following: Dynamic trees, introduced by Sleator and Tarjan [319] and discussed by Tarjan [330], maintain a forest of disjoint rooted trees. Each edge in each tree has a real-valued cost. Dynamic trees support queries to find parents, roots, edge costs, and the minimum edge cost on a simple path from a node up to a root. Trees may be manipulated by cutting edges, updating all edge costs on a simple path from a node up to a root, linking a root into another tree, and making a node the root of the tree it appears in. One implementation of dynamic trees gives an O.lg n/ amortized time bound for each operation; a more complicated implementation yields O.lg n/ worst-case time bounds. Dynamic trees are used in some of the asymptotically fastest network-flow algorithms. Splay trees, developed by Sleator and Tarjan [320] and, again, discussed by Tarjan [330], are a form of binary search tree on which the standard searchtree operations run in O.lg n/ amortized time. One application of splay trees simplifies dynamic trees. Persistent data structures allow queries, and sometimes updates as well, on past versions of a data structure. Driscoll, Sarnak, Sleator, and Tarjan [97] present techniques for making linked data structures persistent with only a small time Part V Advanced Data Structures 483 and space cost. Problem 13-1 gives a simple example of a persistent dynamic set. As in Chapter 20, several data structures allow a faster implementation of dictionary operations (INSERT, DELETE, and SEARCH) for a restricted universe of keys. By taking advantage of these restrictions, they are able to achieve better worst-case asymptotic running times than comparison-based data structures. Fredman and Willard introduced fusion trees [115], which were the first data structure to allow faster dictionary operations when the universe is restricted to integers. They showed how to implement these operations in O.lg n= lg lg n/ time. Several subsequent data structures, including exponential search trees [16], have also given improved bounds on some or all of the dictionary operations and are mentioned in the chapter notes throughout this book. Dynamic graph data structures support various queries while allowing the structure of a graph to change through operations that insert or delete vertices or edges. Examples of the queries that they support include vertex connectivity [166], edge connectivity, minimum spanning trees [165], biconnectivity, and transitive closure [164]. Chapter notes throughout this book mention additional data structures. 18 B-Trees B-trees are balanced search trees designed to work well on disks or other directaccess secondary storage devices. B-trees are similar to red-black trees (Chapter 13), but they are better at minimizing disk I/O operations. Many database systems use B-trees, or variants of B-trees, to store information. B-trees differ from red-black trees in that B-tree nodes may have many children, from a few to thousands. That is, the “branching factor” of a B-tree can be quite large, although it usually depends on characteristics of the disk unit used. B-trees are similar to red-black trees in that every n-node B-tree has height O.lg n/. The exact height of a B-tree can be considerably less than that of a red-black tree, however, because its branching factor, and hence the base of the logarithm that expresses its height, can be much larger. Therefore, we can also use B-trees to implement many dynamic-set operations in time O.lg n/. B-trees generalize binary search trees in a natural manner. Figure 18.1 shows a simple B-tree. If an internal B-tree node x contains x:n keys, then x has x:n C 1 children. The keys in node x serve as dividing points separating the range of keys handled by x into x:n C 1 subranges, each handled by one child of x. When searching for a key in a B-tree, we make an .x:n C 1/-way decision based on comparisons with the x:n keys stored at node x. The structure of leaf nodes differs from that of internal nodes; we will examine these differences in Section 18.1. Section 18.1 gives a precise definition of B-trees and proves that the height of a B-tree grows only logarithmically with the number of nodes it contains. Section 18.2 describes how to search for a key and insert a key into a B-tree, and Section 18.3 discusses deletion. Before proceeding, however, we need to ask why we evaluate data structures designed to work on a disk differently from data structures designed to work in main random-access memory. Data structures on secondary storage Computer systems take advantage of various technologies that provide memory capacity. The primary memory (or main memory) of a computer system normally Chapter 18 B-Trees 485 T: root M DH QTX BC FG J KL N P R S VW Y Z Figure 18.1 A B-tree whose keys are the consonants of English. An internal node x containing x: n keys has x: n C 1 children. All leaves are at the same depth in the tree. The lightly shaded nodes are examined in a search for the letter R. spindle platter track read/write head arms Figure 18.2 A typical disk drive. It comprises one or more platters (two platters are shown here) that rotate around a spindle. Each platter is read and written with a head at the end of an arm. Arms rotate around a common pivot axis. A track is the surface that passes beneath the read/write head when the head is stationary. consists of silicon memory chips. This technology is typically more than an order of magnitude more expensive per bit stored than magnetic storage technology, such as tapes or disks. Most computer systems also have secondary storage based on magnetic disks; the amount of such secondary storage often exceeds the amount of primary memory by at least two orders of magnitude. Figure 18.2 shows a typical disk drive. The drive consists of one or more platters, which rotate at a constant speed around a common spindle. A magnetizable material covers the surface of each platter. The drive reads and writes each platter by a head at the end of an arm. The arms can move their heads toward or away 486 Chapter 18 B-Trees from the spindle. When a given head is stationary, the surface that passes underneath it is called a track. Multiple platters increase only the disk drive’s capacity and not its performance. Although disks are cheaper and have higher capacity than main memory, they are much, much slower because they have moving mechanical parts.1 The mechanical motion has two components: platter rotation and arm movement. As of this writing, commodity disks rotate at speeds of 5400–15,000 revolutions per minute (RPM). We typically see 15,000 RPM speeds in server-grade drives, 7200 RPM speeds in drives for desktops, and 5400 RPM speeds in drives for laptops. Although 7200 RPM may seem fast, one rotation takes 8.33 milliseconds, which is over 5 orders of magnitude longer than the 50 nanosecond access times (more or less) commonly found for silicon memory. In other words, if we have to wait a full rotation for a particular item to come under the read/write head, we could access main memory more than 100,000 times during that span. On average we have to wait for only half a rotation, but still, the difference in access times for silicon memory compared with disks is enormous. Moving the arms also takes some time. As of this writing, average access times for commodity disks are in the range of 8 to 11 milliseconds. In order to amortize the time spent waiting for mechanical movements, disks access not just one item but several at a time. Information is divided into a number of equal-sized pages of bits that appear consecutively within tracks, and each disk read or write is of one or more entire pages. For a typical disk, a page might be 211 to 214 bytes in length. Once the read/write head is positioned correctly and the disk has rotated to the beginning of the desired page, reading or writing a magnetic disk is entirely electronic (aside from the rotation of the disk), and the disk can quickly read or write large amounts of data. Often, accessing a page of information and reading it from a disk takes longer than examining all the information read. For this reason, in this chapter we shall look separately at the two principal components of the running time: the number of disk accesses, and the CPU (computing) time. We measure the number of disk accesses in terms of the number of pages of information that need to be read from or written to the disk. We note that disk-access time is not constant—it depends on the distance between the current track and the desired track and also on the initial rotational position of the disk. We shall 1As of this writing, solid-state drives have recently come onto the consumer market. Although they are faster than mechanical disk drives, they cost more per gigabyte and have lower capacities than mechanical disk drives. Chapter 18 B-Trees 487 nonetheless use the number of pages read or written as a first-order approximation of the total time spent accessing the disk. In a typical B-tree application, the amount of data handled is so large that all the data do not fit into main memory at once. The B-tree algorithms copy selected pages from disk into main memory as needed and write back onto disk the pages that have changed. B-tree algorithms keep only a constant number of pages in main memory at any time; thus, the size of main memory does not limit the size of B-trees that can be handled. We model disk operations in our pseudocode as follows. Let x be a pointer to an object. If the object is currently in the computer’s main memory, then we can refer to the attributes of the object as usual: x:key, for example. If the object referred to by x resides on disk, however, then we must perform the operation DISK-READ.x/ to read object x into main memory before we can refer to its attributes. (We assume that if x is already in main memory, then DISK-READ.x/ requires no disk accesses; it is a “no-op.”) Similarly, the operation DISK-WRITE.x/ is used to save any changes that have been made to the attributes of object x. That is, the typical pattern for working with an object is as follows: x D a pointer to some object D I S K - R E A D .x / operations that access and/or modify the attributes of x D ISK -W R ITE .x / // omitted if no attributes of x were changed other operations that access but do not modify attributes of x The system can keep only a limited number of pages in main memory at any one time. We shall assume that the system flushes from main memory pages no longer in use; our B-tree algorithms will ignore this issue. Since in most systems the running time of a B-tree algorithm depends primarily on the number of DISK-READ and DISK-WRITE operations it performs, we typically want each of these operations to read or write as much information as possible. Thus, a B-tree node is usually as large as a whole disk page, and this size limits the number of children a B-tree node can have. For a large B-tree stored on a disk, we often see branching factors between 50 and 2000, depending on the size of a key relative to the size of a page. A large branching factor dramatically reduces both the height of the tree and the number of disk accesses required to find any key. Figure 18.3 shows a B-tree with a branching factor of 1001 and height 2 that can store over one billion keys; nevertheless, since we can keep the root node permanently in main memory, we can find any key in this tree by making at most only two disk accesses. 488 Chapter 18 B-Trees T: root 1000 1001 1000 1000 … 1001 1000 1001 1000 … 1000 1000 1001 1 node, 1000 keys 1001 nodes, 1,001,000 keys 1,002,001 nodes, 1,002,001,000 keys Figure 18.3 A B-tree of height 2 containing over one billion keys. Shown inside each node x is x: n, the number of keys in x. Each internal node and leaf contains 1000 keys. This B-tree has 1001 nodes at depth 1 and over one million leaves at depth 2. 18.1 Definition of B-trees To keep things simple, we assume, as we have for binary search trees and red-black trees, that any “satellite information” associated with a key resides in the same node as the key. In practice, one might actually store with each key just a pointer to another disk page containing the satellite information for that key. The pseudocode in this chapter implicitly assumes that the satellite information associated with a key, or the pointer to such satellite information, travels with the key whenever the key is moved from node to node. A common variant on a B-tree, known as a BC-tree, stores all the satellite information in the leaves and stores only keys and child pointers in the internal nodes, thus maximizing the branching factor of the internal nodes. A B-tree T is a rooted tree (whose root is T:root) having the following properties: 1. Every node x has the following attributes: a. x:n, the number of keys currently stored in node x, b. the x:n keys themselves, x:key1; x:key2; : : : ; x:keyx:n, stored in nondecreas- ing order, so that x:key1 Ä x:key2 Ä Ä x:keyx:n, c. x:leaf , a boolean value that is TRUE if x is a leaf and FALSE if x is an internal node. 2. Each internal node x also contains x:n C 1 pointers x:c1; x:c2; : : : ; x:cx:nC1 to its children. Leaf nodes have no children, and so their ci attributes are undefined. 18.1 Definition of B-trees 489 3. The keys x:keyi separate the ranges of keys stored in each subtree: if ki is any key stored in the subtree with root x:ci, then k1 Ä x: key1 Ä k2 Ä x: key2 Ä Ä x: keyx: n Ä kx: nC1 : 4. All leaves have the same depth, which is the tree’s height h. 5. Nodes have lower and upper bounds on the number of keys they can contain. We express these bounds in terms of a fixed integer t 2 called the minimum degree of the B-tree: a. Every node other than the root must have at least t 1 keys. Every internal node other than the root thus has at least t children. If the tree is nonempty, the root must have at least one key. b. Every node may contain at most 2t 1 keys. Therefore, an internal node may have at most 2t children. We say that a node is full if it contains exactly 2t 1 keys.2 The simplest B-tree occurs when t D 2. Every internal node then has either 2, 3, or 4 children, and we have a 2-3-4 tree. In practice, however, much larger values of t yield B-trees with smaller height. The height of a B-tree The number of disk accesses required for most operations on a B-tree is proportional to the height of the B-tree. We now analyze the worst-case height of a B-tree. Theorem 18.1 If n 1, then for any n-key B-tree T of height h and minimum degree t 2, nC1 h Ä logt 2 : Proof The root of a B-tree T contains at least one key, and all other nodes contain at least t 1 keys. Thus, T , whose height is h, has at least 2 nodes at depth 1, at least 2t nodes at depth 2, at least 2t2 nodes at depth 3, and so on, until at depth h it has at least 2th 1 nodes. Figure 18.4 illustrates such a tree for h D 3. Thus, the 2Another common variant on a B-tree, known as a B -tree, requires each internal node to be at least 2=3 full, rather than at least half full, as a B-tree requires. 490 Chapter 18 B-Trees T: root 1 t–1 t t–1 … t–1 t t–1 … t–1 t t–1 … t–1 t–1 t t–1 … t–1 t t–1 … t–1 t t–1 … t–1 depth 0 number of nodes 1 1 2 2 2t 3 2t2 Figure 18.4 A B-tree of height 3 containing a minimum possible number of keys. Shown inside each node x is x: n. number n of keys satisfies the inequality Xh n 1 C .t 1/ 2t i 1 D 1 C 2.t i DÂ1 t h à 1 1/ t1 D 2th 1 : By simple algebra, we get th Ä .n C 1/=2. Taking base-t logarithms of both sides proves the theorem. Here we see the power of B-trees, as compared with red-black trees. Although the height of the tree grows as O.lg n/ in both cases (recall that t is a constant), for B-trees the base of the logarithm can be many times larger. Thus, B-trees save a factor of about lg t over red-black trees in the number of nodes examined for most tree operations. Because we usually have to access the disk to examine an arbitrary node in a tree, B-trees avoid a substantial number of disk accesses. Exercises 18.1-1 Why don’t we allow a minimum degree of t D 1? 18.1-2 For what values of t is the tree of Figure 18.1 a legal B-tree? 18.2 Basic operations on B-trees 491 18.1-3 Show all legal B-trees of minimum degree 2 that represent f1; 2; 3; 4; 5g. 18.1-4 As a function of the minimum degree t, what is the maximum number of keys that can be stored in a B-tree of height h? 18.1-5 Describe the data structure that would result if each black node in a red-black tree were to absorb its red children, incorporating their children with its own. 18.2 Basic operations on B-trees In this section, we present the details of the operations B-TREE-SEARCH, BTREE-CREATE, and B-TREE-INSERT. In these procedures, we adopt two conventions: The root of the B-tree is always in main memory, so that we never need to perform a DISK-READ on the root; we do have to perform a DISK-WRITE of the root, however, whenever the root node is changed. Any nodes that are passed as parameters must already have had a DISK-READ operation performed on them. The procedures we present are all “one-pass” algorithms that proceed downward from the root of the tree, without having to back up. Searching a B-tree Searching a B-tree is much like searching a binary search tree, except that instead of making a binary, or “two-way,” branching decision at each node, we make a multiway branching decision according to the number of the node’s children. More precisely, at each internal node x, we make an .x:n C 1/-way branching decision. B-TREE-SEARCH is a straightforward generalization of the TREE-SEARCH procedure defined for binary search trees. B-TREE-SEARCH takes as input a pointer to the root node x of a subtree and a key k to be searched for in that subtree. The top-level call is thus of the form B-TREE-SEARCH.T:root; k/. If k is in the B-tree, B-TREE-SEARCH returns the ordered pair .y; i / consisting of a node y and an index i such that y:keyi D k. Otherwise, the procedure returns NIL. 492 Chapter 18 B-Trees B-TREE-SEARCH.x; k/ 1 iD1 2 while i Ä x:n and k > x:keyi 3 i D iC1 4 if i Ä x:n and k == x:keyi 5 return .x; i/ 6 elseif x:leaf 7 return NIL 8 else DISK-READ.x:ci / 9 return B-TREE-SEARCH.x:ci; k/ Using a linear-search procedure, lines 1–3 find the smallest index i such that k Ä x:keyi , or else they set i to x:n C 1. Lines 4–5 check to see whether we have now discovered the key, returning if we have. Otherwise, lines 6–9 either terminate the search unsuccessfully (if x is a leaf) or recurse to search the appropriate subtree of x, after performing the necessary DISK-READ on that child. Figure 18.1 illustrates the operation of B-TREE-SEARCH. The procedure examines the lightly shaded nodes during a search for the key R. As in the TREE-SEARCH procedure for binary search trees, the nodes encountered during the recursion form a simple path downward from the root of the tree. The B-TREE-SEARCH procedure therefore accesses O.h/ D O.logt n/ disk pages, where h is the height of the B-tree and n is the number of keys in the B-tree. Since x:n < 2t, the while loop of lines 2–3 takes O.t/ time within each node, and the total CPU time is O.th/ D O.t logt n/. Creating an empty B-tree To build a B-tree T , we first use B-TREE-CREATE to create an empty root node and then call B-TREE-INSERT to add new keys. Both of these procedures use an auxiliary procedure ALLOCATE-NODE, which allocates one disk page to be used as a new node in O.1/ time. We can assume that a node created by ALLOCATENODE requires no DISK-READ, since there is as yet no useful information stored on the disk for that node. B-TREE-CREATE .T / 1 x D ALLOCATE-NODE./ 2 x:leaf D TRUE 3 x:n D 0 4 DISK-WRITE.x/ 5 T:root D x B-TREE-CREATE requires O.1/ disk operations and O.1/ CPU time. 18.2 Basic operations on B-trees 493 Inserting a key into a B-tree Inserting a key into a B-tree is significantly more complicated than inserting a key into a binary search tree. As with binary search trees, we search for the leaf position at which to insert the new key. With a B-tree, however, we cannot simply create a new leaf node and insert it, as the resulting tree would fail to be a valid B-tree. Instead, we insert the new key into an existing leaf node. Since we cannot insert a key into a leaf node that is full, we introduce an operation that splits a full node y (having 2t 1 keys) around its median key y:keyt into two nodes having only t 1 keys each. The median key moves up into y’s parent to identify the dividing point between the two new trees. But if y’s parent is also full, we must split it before we can insert the new key, and thus we could end up splitting full nodes all the way up the tree. As with a binary search tree, we can insert a key into a B-tree in a single pass down the tree from the root to a leaf. To do so, we do not wait to find out whether we will actually need to split a full node in order to do the insertion. Instead, as we travel down the tree searching for the position where the new key belongs, we split each full node we come to along the way (including the leaf itself). Thus whenever we want to split a full node y, we are assured that its parent is not full. Splitting a node in a B-tree The procedure B-TREE-SPLIT-CHILD takes as input a nonfull internal node x (assumed to be in main memory) and an index i such that x:ci (also assumed to be in main memory) is a full child of x. The procedure then splits this child in two and adjusts x so that it has an additional child. To split a full root, we will first make the root a child of a new empty root node, so that we can use B-TREE-SPLIT-CHILD. The tree thus grows in height by one; splitting is the only means by which the tree grows. Figure 18.5 illustrates this process. We split the full node y D x:ci about its median key S, which moves up into y’s parent node x. Those keys in y that are greater than the median key move into a new node ´, which becomes a new child of x. 494 Chapter 18 B-Trees 1 x x: key i x: key i …N W… y D x:ci PQRS TUV T1 T2 T3 T4 T5 T6 T7 T8 x x: key i x 1 : key i x : key i C1 …N S W… y D x:ci PQR ´ D x: ciC1 TUV T1 T2 T3 T4 T5 T6 T7 T8 Figure 18.5 Splitting a node with t D 4. Node y D x: ci splits into two nodes, y and ´, and the median key S of y moves up into y’s parent. B-TREE-SPLIT-CHILD .x; i / 1 ´ D ALLOCATE-NODE./ 2 y D x:ci 3 ´:leaf D y:leaf 4 ´:n D t 1 5 for j D 1 to t 1 6 ´: keyj D y: keyj Ct 7 if not y:leaf 8 for j D 1 to t 9 ´: cj D y: cj Ct 10 y:n D t 1 11 for j D x:n C 1 downto i C 1 12 x: cj C1 D x: cj 13 x: ciC1 D ´ 14 for j D x:n downto i 15 x: keyj C1 D x: keyj 16 x:keyi D y:keyt 17 x:n D x:n C 1 18 DISK-WRITE.y/ 19 DISK-WRITE.´/ 20 DISK-WRITE.x/ B-TREE-SPLIT-CHILD works by straightforward “cutting and pasting.” Here, x is the node being split, and y is x’s ith child (set in line 2). Node y originally has 2t children (2t 1 keys) but is reduced to t children (t 1 keys) by this operation. Node ´ takes the t largest children (t 1 keys) from y, and ´ becomes a new child 18.2 Basic operations on B-trees 495 of x, positioned just after y in x’s table of children. The median key of y moves up to become the key in x that separates y and ´. Lines 1–9 create node ´ and give it the largest t 1 keys and corresponding t children of y. Line 10 adjusts the key count for y. Finally, lines 11–17 insert ´ as a child of x, move the median key from y up to x in order to separate y from ´, and adjust x’s key count. Lines 18–20 write out all modified disk pages. The CPU time used by B-TREE-SPLIT-CHILD is ‚.t/, due to the loops on lines 5–6 and 8–9. (The other loops run for O.t/ iterations.) The procedure performs O.1/ disk operations. Inserting a key into a B-tree in a single pass down the tree We insert a key k into a B-tree T of height h in a single pass down the tree, requiring O.h/ disk accesses. The CPU time required is O.th/ D O.t logt n/. The B-TREE-INSERT procedure uses B-TREE-SPLIT-CHILD to guarantee that the recursion never descends to a full node. B-TREE-INSERT .T; k/ 1 r D T:root 2 if r:n == 2t 1 3 s D ALLOCATE-NODE./ 4 T:root D s 5 s: leaf D FALSE 6 s:n D 0 7 s:c1 D r 8 B-TREE-SPLIT-CHILD .s; 1/ 9 B-TREE-INSERT-NONFULL .s; k/ 10 else B-TREE-INSERT-NONFULL .r; k/ Lines 3–9 handle the case in which the root node r is full: the root splits and a new node s (having two children) becomes the root. Splitting the root is the only way to increase the height of a B-tree. Figure 18.6 illustrates this case. Unlike a binary search tree, a B-tree increases in height at the top instead of at the bottom. The procedure finishes by calling B-TREE-INSERT-NONFULL to insert key k into the tree rooted at the nonfull root node. B-TREE-INSERT-NONFULL recurses as necessary down the tree, at all times guaranteeing that the node to which it recurses is not full by calling B-TREE-SPLIT-CHILD as necessary. The auxiliary recursive procedure B-TREE-INSERT-NONFULL inserts key k into node x, which is assumed to be nonfull when the procedure is called. The operation of B-TREE-INSERT and the recursive operation of B-TREE-INSERT-NONFULL guarantee that this assumption is true. 496 Chapter 18 B-Trees T: root r ADFHLNP T1 T2 T3 T4 T5 T6 T7 T8 T: root s H r ADF LNP T1 T2 T3 T4 T5 T6 T7 T8 Figure 18.6 Splitting the root with t D 4. Root node r splits in two, and a new root node s is created. The new root contains the median key of r and has the two halves of r as children. The B-tree grows in height by one when the root is split. B-TREE-INSERT-NONFULL.x; k/ 1 i D x:n 2 if x:leaf 3 while i 1 and k < x:keyi 4 x: keyiC1 D x: keyi 5 i Di 1 6 x: keyiC1 D k 7 x:n D x:n C 1 8 D ISK -W R ITE .x / 9 else while i 1 and k < x:keyi 10 i Di 1 11 i D i C1 12 DISK-READ.x: ci / 13 if x:ci :n == 2t 1 14 B-TREE-SPLIT-CHILD .x; i / 15 if k > x:keyi 16 i D iC1 17 B-TREE-INSERT-NONFULL.x: ci ; k/ The B-TREE-INSERT-NONFULL procedure works as follows. Lines 3–8 handle the case in which x is a leaf node by inserting key k into x. If x is not a leaf node, then we must insert k into the appropriate leaf node in the subtree rooted at internal node x. In this case, lines 9–11 determine the child of x to which the recursion descends. Line 13 detects whether the recursion would descend to a full child, in which case line 14 uses B-TREE-SPLIT-CHILD to split that child into two nonfull children, and lines 15–16 determine which of the two children is now the 18.2 Basic operations on B-trees 497 correct one to descend to. (Note that there is no need for a DISK-READ.x:ci/ after line 16 increments i, since the recursion will descend in this case to a child that was just created by B-TREE-SPLIT-CHILD.) The net effect of lines 13–16 is thus to guarantee that the procedure never recurses to a full node. Line 17 then recurses to insert k into the appropriate subtree. Figure 18.7 illustrates the various cases of inserting into a B-tree. For a B-tree of height h, B-TREE-INSERT performs O.h/ disk accesses, since only O.1/ DISK-READ and DISK-WRITE operations occur between calls to B-TREE-INSERT-NONFULL. The total CPU time used is O.th/ D O.t logt n/. Since B-TREE-INSERT-NONFULL is tail-recursive, we can alternatively implement it as a while loop, thereby demonstrating that the number of pages that need to be in main memory at any time is O.1/. Exercises 18.2-1 Show the results of inserting the keys F; S; Q; K; C; L; H; T; V; W; M; R; N; P; A; B; X; Y; D; Z; E in order into an empty B-tree with minimum degree 2. Draw only the configurations of the tree just before some node must split, and also draw the final configuration. 18.2-2 Explain under what circumstances, if any, redundant DISK-READ or DISK-WRITE operations occur during the course of executing a call to B-TREE-INSERT. (A redundant DISK-READ is a DISK-READ for a page that is already in memory. A redundant DISK-WRITE writes to disk a page of information that is identical to what is already stored there.) 18.2-3 Explain how to find the minimum key stored in a B-tree and how to find the predecessor of a given key stored in a B-tree. 18.2-4 ? Suppose that we insert the keys f1; 2; : : : ; ng into an empty B-tree with minimum degree 2. How many nodes does the final B-tree have? 18.2-5 Since leaf nodes require no pointers to children, they could conceivably use a different (larger) t value than internal nodes for the same disk page size. Show how to modify the procedures for creating and inserting into a B-tree to handle this variation. 498 Chapter 18 B-Trees (a) initial tree GMP X ACDE J K NO R S TUV Y Z (b) B inserted ABCDE GMP X J K NO R S TUV YZ (c) Q inserted ABCDE GMP T X J K NO QR S UV YZ (d) L inserted P GM TX ABCDE J K L NO QRS UV Y Z (e) F inserted P CGM AB DE F J K L NO TX QRS UV Y Z Figure 18.7 Inserting keys into a B-tree. The minimum degree t for this B-tree is 3, so a node can hold at most 5 keys. Nodes that are modified by the insertion process are lightly shaded. (a) The initial tree for this example. (b) The result of inserting B into the initial tree; this is a simple insertion into a leaf node. (c) The result of inserting Q into the previous tree. The node RST U V splits into two nodes containing RS and U V , the key T moves up to the root, and Q is inserted in the leftmost of the two halves (the RS node). (d) The result of inserting L into the previous tree. The root splits right away, since it is full, and the B-tree grows in height by one. Then L is inserted into the leaf containing JK. (e) The result of inserting F into the previous tree. The node ABCDE splits before F is inserted into the rightmost of the two halves (the DE node). 18.3 Deleting a key from a B-tree 499 18.2-6 Suppose that we were to implement B-TREE-SEARCH to use binary search rather than linear search within each node. Show that this change makes the CPU time required O.lg n/, independently of how t might be chosen as a function of n. 18.2-7 Suppose that disk hardware allows us to choose the size of a disk page arbitrarily, but that the time it takes to read the disk page is a C bt, where a and b are specified constants and t is the minimum degree for a B-tree using pages of the selected size. Describe how to choose t so as to minimize (approximately) the B-tree search time. Suggest an optimal value of t for the case in which a D 5 milliseconds and b D 10 microseconds. 18.3 Deleting a key from a B-tree Deletion from a B-tree is analogous to insertion but a little more complicated, because we can delete a key from any node—not just a leaf—and when we delete a key from an internal node, we will have to rearrange the node’s children. As in insertion, we must guard against deletion producing a tree whose structure violates the B-tree properties. Just as we had to ensure that a node didn’t get too big due to insertion, we must ensure that a node doesn’t get too small during deletion (except that the root is allowed to have fewer than the minimum number t 1 of keys). Just as a simple insertion algorithm might have to back up if a node on the path to where the key was to be inserted was full, a simple approach to deletion might have to back up if a node (other than the root) along the path to where the key is to be deleted has the minimum number of keys. The procedure B-TREE-DELETE deletes the key k from the subtree rooted at x. We design this procedure to guarantee that whenever it calls itself recursively on a node x, the number of keys in x is at least the minimum degree t. Note that this condition requires one more key than the minimum required by the usual B-tree conditions, so that sometimes a key may have to be moved into a child node before recursion descends to that child. This strengthened condition allows us to delete a key from the tree in one downward pass without having to “back up” (with one exception, which we’ll explain). You should interpret the following specification for deletion from a B-tree with the understanding that if the root node x ever becomes an internal node having no keys (this situation can occur in cases 2c and 3b on pages 501–502), then we delete x, and x’s only child x:c1 becomes the new root of the tree, decreasing the height of the tree by one and preserving the property that the root of the tree contains at least one key (unless the tree is empty). 500 Chapter 18 B-Trees (a) initial tree P CGM AB DE F J K L NO TX QRS UV Y Z (b) F deleted: case 1 P C GM AB DE J K L NO TX QRS UV Y Z (c) M deleted: case 2a P CG L TX AB DE J K NO QRS UV Y Z (d) G deleted: case 2c P CL TX AB DE J K NO QRS UV Y Z Figure 18.8 Deleting keys from a B-tree. The minimum degree for this B-tree is t D 3, so a node (other than the root) cannot have fewer than 2 keys. Nodes that are modified are lightly shaded. (a) The B-tree of Figure 18.7(e). (b) Deletion of F . This is case 1: simple deletion from a leaf. (c) Deletion of M . This is case 2a: the predecessor L of M moves up to take M ’s position. (d) Deletion of G. This is case 2c: we push G down to make node DEGJK and then delete G from this leaf (case 1). We sketch how deletion works instead of presenting the pseudocode. Figure 18.8 illustrates the various cases of deleting keys from a B-tree. 1. If the key k is in node x and x is a leaf, delete the key k from x. 2. If the key k is in node x and x is an internal node, do the following: 18.3 Deleting a key from a B-tree 501 (e) D deleted: case 3b CLPT X A B E J K NO QR S UV Y Z (e′) tree shrinks in height AB EJK CLPT X NO QR S UV YZ (f) B deleted: case 3a E L P T X AC J K NO QR S UV Y Z Figure 18.8, continued (e) Deletion of D. This is case 3b: the recursion cannot descend to node CL because it has only 2 keys, so we push P down and merge it with CL and TX to form CLP TX; then we delete D from a leaf (case 1). (e0) After (e), we delete the root and the tree shrinks in height by one. (f) Deletion of B. This is case 3a: C moves to fill B’s position and E moves to fill C ’s position. a. If the child y that precedes k in node x has at least t keys, then find the predecessor k0 of k in the subtree rooted at y. Recursively delete k0, and replace k by k0 in x. (We can find k0 and delete it in a single downward pass.) b. If y has fewer than t keys, then, symmetrically, examine the child ´ that follows k in node x. If ´ has at least t keys, then find the successor k0 of k in the subtree rooted at ´. Recursively delete k0, and replace k by k0 in x. (We can find k0 and delete it in a single downward pass.) c. Otherwise, if both y and ´ have only t 1 keys, merge k and all of ´ into y, so that x loses both k and the pointer to ´, and y now contains 2t 1 keys. Then free ´ and recursively delete k from y. 3. If the key k is not present in internal node x, determine the root x:ci of the appropriate subtree that must contain k, if k is in the tree at all. If x:ci has only t 1 keys, execute step 3a or 3b as necessary to guarantee that we descend to a node containing at least t keys. Then finish by recursing on the appropriate child of x. 502 Chapter 18 B-Trees a. If x:ci has only t 1 keys but has an immediate sibling with at least t keys, give x:ci an extra key by moving a key from x down into x:ci , moving a key from x:ci ’s immediate left or right sibling up into x, and moving the appropriate child pointer from the sibling into x:ci. b. If x:ci and both of x:ci ’s immediate siblings have t 1 keys, merge x:ci with one sibling, which involves moving a key from x down into the new merged node to become the median key for that node. Since most of the keys in a B-tree are in the leaves, we may expect that in practice, deletion operations are most often used to delete keys from leaves. The B-TREE-DELETE procedure then acts in one downward pass through the tree, without having to back up. When deleting a key in an internal node, however, the procedure makes a downward pass through the tree but may have to return to the node from which the key was deleted to replace the key with its predecessor or successor (cases 2a and 2b). Although this procedure seems complicated, it involves only O.h/ disk operations for a B-tree of height h, since only O.1/ calls to DISK-READ and DISKWRITE are made between recursive invocations of the procedure. The CPU time required is O.th/ D O.t logt n/. Exercises 18.3-1 Show the results of deleting C , P , and V , in order, from the tree of Figure 18.8(f). 18.3-2 Write pseudocode for B-TREE-DELETE. Problems 18-1 Stacks on secondary storage Consider implementing a stack in a computer that has a relatively small amount of fast primary memory and a relatively large amount of slower disk storage. The operations PUSH and POP work on single-word values. The stack we wish to support can grow to be much larger than can fit in memory, and thus most of it must be stored on disk. A simple, but inefficient, stack implementation keeps the entire stack on disk. We maintain in memory a stack pointer, which is the disk address of the top element on the stack. If the pointer has value p, the top element is the .p mod m/th word on page bp=mc of the disk, where m is the number of words per page. Problems for Chapter 18 503 To im