# The Algorithm Design Manual By Steven S. Skiena

### The%20Algorithm%20Design%20Manual%20by%20Steven%20S.%20Skiena

### The%20Algorithm%20Design%20Manual%20-%20Second%20Edition

### The%20Algorithm%20Design%20Manual%20-%20Second%20Edition

### The%20Algorithm%20Design%20Manual%20-%20Second%20Edition

### The%20Algorithm%20Design%20Manual%20by%20Steven%20S.%20Skiena

User Manual: Pdf

Open the PDF directly: View PDF .

Page Count: 739

Download | |

Open PDF In Browser | View PDF |

The Algorithm Design Manual Second Edition Steven S. Skiena The Algorithm Design Manual Second Edition 123 Steven S. Skiena Department of Computer Science State University of New York at Stony Brook New York, USA skiena@cs.sunysb.edu ISBN: 978-1-84800-069-8 DOI: 10.1007/978-1-84800-070-4 e-ISBN: 978-1-84800-070-4 British Library Cataloguing in Publication Data A catalogue record for this book is available from the British Library Library of Congress Control Number: 2008931136 c Springer-Verlag London Limited 2008 Apart from any fair dealing for the purposes of research or private study, or criticism or review, as permitted under the Copyright, Designs and Patents Act 1988, this publication may only be reproduced, stored or transmitted, in any form or by any means, with the prior permission in writing of the publishers, or in the case of reprographic reproduction in accordance with the terms of licenses issued by the Copyright Licensing Agency. Enquiries concerning reproduction outside those terms should be sent to the publishers. The use of registered names, trademarks, etc., in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant laws and regulations and therefore free for general use. The publisher makes no representation, express or implied, with regard to the accuracy of the information contained in this book and cannot accept any legal responsibility or liability for any errors or omissions that may be made. Printed on acid-free paper Springer Science+Business Media springer.com Preface Most professional programmers that I’ve encountered are not well prepared to tackle algorithm design problems. This is a pity, because the techniques of algorithm design form one of the core practical technologies of computer science. Designing correct, eﬃcient, and implementable algorithms for real-world problems requires access to two distinct bodies of knowledge: • Techniques – Good algorithm designers understand several fundamental algorithm design techniques, including data structures, dynamic programming, depth-ﬁrst search, backtracking, and heuristics. Perhaps the single most important design technique is modeling, the art of abstracting a messy real-world application into a clean problem suitable for algorithmic attack. • Resources – Good algorithm designers stand on the shoulders of giants. Rather than laboring from scratch to produce a new algorithm for every task, they can ﬁgure out what is known about a particular problem. Rather than re-implementing popular algorithms from scratch, they seek existing implementations to serve as a starting point. They are familiar with many classic algorithmic problems, which provide suﬃcient source material to model most any application. This book is intended as a manual on algorithm design, providing access to combinatorial algorithm technology for both students and computer professionals. It is divided into two parts: Techniques and Resources. The former is a general guide to techniques for the design and analysis of computer algorithms. The Resources section is intended for browsing and reference, and comprises the catalog of algorithmic resources, implementations, and an extensive bibliography. vi PREFACE To the Reader I have been gratiﬁed by the warm reception the ﬁrst edition of The Algorithm Design Manual has received since its initial publication in 1997. It has been recognized as a unique guide to using algorithmic techniques to solve problems that often arise in practice. But much has changed in the world since the The Algorithm Design Manual was ﬁrst published over ten years ago. Indeed, if we date the origins of modern algorithm design and analysis to about 1970, then roughly 30% of modern algorithmic history has happened since the ﬁrst coming of The Algorithm Design Manual. Three aspects of The Algorithm Design Manual have been particularly beloved: (1) the catalog of algorithmic problems, (2) the war stories, and (3) the electronic component of the book. These features have been preserved and strengthened in this edition: • The Catalog of Algorithmic Problems – Since ﬁnding out what is known about an algorithmic problem can be a diﬃcult task, I provide a catalog of the 75 most important problems arising in practice. By browsing through this catalog, the student or practitioner can quickly identify what their problem is called, what is known about it, and how they should proceed to solve it. To aid in problem identiﬁcation, we include a pair of “before” and “after” pictures for each problem, illustrating the required input and output speciﬁcations. One perceptive reviewer called my book “The Hitchhiker’s Guide to Algorithms” on the strength of this catalog. The catalog is the most important part of this book. To update the catalog for this edition, I have solicited feedback from the world’s leading experts on each associated problem. Particular attention has been paid to updating the discussion of available software implementations for each problem. • War Stories – In practice, algorithm problems do not arise at the beginning of a large project. Rather, they typically arise as subproblems when it becomes clear that the programmer does not know how to proceed or that the current solution is inadequate. To provide a better perspective on how algorithm problems arise in the real world, we include a collection of “war stories,” or tales from our experience with real problems. The moral of these stories is that algorithm design and analysis is not just theory, but an important tool to be pulled out and used as needed. This edition retains all the original war stories (with updates as appropriate) plus additional new war stories covering external sorting, graph algorithms, simulated annealing, and other topics. • Electronic Component – Since the practical person is usually looking for a program more than an algorithm, we provide pointers to solid implementations whenever they are available. We have collected these implementations PREFACE at one central website site (http://www.cs.sunysb.edu/∼algorith) for easy retrieval. We have been the number one “Algorithm” site on Google pretty much since the initial publication of the book. Further, we provide recommendations to make it easier to identify the correct code for the job. With these implementations available, the critical issue in algorithm design becomes properly modeling your application, more so than becoming intimate with the details of the actual algorithm. This focus permeates the entire book. Equally important is what we do not do in this book. We do not stress the mathematical analysis of algorithms, leaving most of the analysis as informal arguments. You will not ﬁnd a single theorem anywhere in this book. When more details are needed, the reader should study the cited programs or references. The goal of this manual is to get you going in the right direction as quickly as possible. To the Instructor This book covers enough material for a standard Introduction to Algorithms course. We assume the reader has completed the equivalent of a second programming course, typically titled Data Structures or Computer Science II. A full set of lecture slides for teaching this course is available online at http://www.algorist.com . Further, I make available online audio and video lectures using these slides to teach a full-semester algorithm course. Let me help teach your course, by the magic of the Internet! This book stresses design over analysis. It is suitable for both traditional lecture courses and the new “active learning” method, where the professor does not lecture but instead guides student groups to solve real problems. The “war stories” provide an appropriate introduction to the active learning method. I have made several pedagogical improvements throughout the book. Textbookoriented features include: • More Leisurely Discussion – The tutorial material in the ﬁrst part of the book has been doubled over the previous edition. The pages have been devoted to more thorough and careful exposition of fundamental material, instead of adding more specialized topics. • False Starts – Algorithms textbooks generally present important algorithms as a fait accompli, obscuring the ideas involved in designing them and the subtle reasons why other approaches fail. The war stories illustrate such development on certain applied problems, but I have expanded such coverage into classical algorithm design material as well. • Stop and Think – Here I illustrate my thought process as I solve a topicspeciﬁc homework problem—false starts and all. I have interspersed such vii viii PREFACE problem blocks throughout the text to increase the problem-solving activity of my readers. Answers appear immediately following each problem. • More and Improved Homework Problems – This edition of The Algorithm Design Manual has twice as many homework exercises as the previous one. Exercises that proved confusing or ambiguous have been improved or replaced. Degree of diﬃculty ratings (from 1 to 10) have been assigned to all problems. • Self-Motivating Exam Design – In my algorithms courses, I promise the students that all midterm and ﬁnal exam questions will be taken directly from homework problems in this book. This provides a “student-motivated exam,” so students know exactly how to study to do well on the exam. I have carefully picked the quantity, variety, and diﬃculty of homework exercises to make this work; ensuring there are neither too few or too many candidate problems. • Take-Home Lessons – Highlighted “take-home” lesson boxes scattered throughout the text emphasize the big-picture concepts to be gained from the chapter. • Links to Programming Challenge Problems – Each chapter’s exercises will contain links to 3-5 relevant “Programming Challenge” problems from http://www.programming-challenges.com. These can be used to add a programming component to paper-and-pencil algorithms courses. • More Code, Less Pseudo-code – More algorithms in this book appear as code (written in C) instead of pseudo-code. I believe the concreteness and reliability of actual tested implementations provides a big win over less formal presentations for simple algorithms. Full implementations are available for study at http://www.algorist.com . • Chapter Notes – Each tutorial chapter concludes with a brief notes section, pointing readers to primary sources and additional references. Acknowledgments Updating a book dedication after ten years focuses attention on the eﬀects of time. Since the ﬁrst edition, Renee has become my wife and then the mother of our two children, Bonnie and Abby. My father has left this world, but Mom and my brothers Len and Rob remain a vital presence in my life. I dedicate this book to my family, new and old, here and departed. I would like to thank several people for their concrete contributions to this new edition. Andrew Gaun and Betson Thomas helped in many ways, particularly in building the infrastructure for the new http://www.cs.sunysb.edu/∼algorithand dealing with a variety of manuscript preparation issues. David Gries oﬀered valuable feedback well beyond the call of duty. Himanshu Gupta and Bin Tang bravely PREFACE taught courses using a manuscript version of this edition. Thanks also to my Springer-Verlag editors, Wayne Wheeler and Allan Wylde. A select group of algorithmic sages reviewed sections of the Hitchhiker’s guide, sharing their wisdom and alerting me to new developments. Thanks to: Ami Amir, Herve Bronnimann, Bernard Chazelle, Chris Chu, Scott Cotton, Yeﬁm Dinitz, Komei Fukuda, Michael Goodrich, Lenny Heath, Cihat Imamoglu, Tao Jiang, David Karger, Giuseppe Liotta, Albert Mao, Silvano Martello, Catherine McGeoch, Kurt Mehlhorn, Scott A. Mitchell, Naceur Meskini, Gene Myers, Gonzalo Navarro, Stephen North, Joe O’Rourke, Mike Paterson, Theo Pavlidis, Seth Pettie, Michel Pocchiola, Bart Preneel, Tomasz Radzik, Edward Reingold, Frank Ruskey, Peter Sanders, Joao Setubal, Jonathan Shewchuk, Robert Skeel, Jens Stoye, Torsten Suel, Bruce Watson, and Uri Zwick. Several exercises were originated by colleagues or inspired by other texts. Reconstructing the original sources years later can be challenging, but credits for each problem (to the best of my recollection) appear on the website. It would be rude not to thank important contributors to the original edition. Ricky Bradley and Dario Vlah built up the substantial infrastructure required for the WWW site in a logical and extensible manner. Zhong Li drew most of the catalog ﬁgures using xﬁg. Richard Crandall, Ron Danielson, Takis Metaxas, Dave Miller, Giri Narasimhan, and Joe Zachary all reviewed preliminary versions of the ﬁrst edition; their thoughtful feedback helped to shape what you see here. Much of what I know about algorithms I learned along with my graduate students. Several of them (Yaw-Ling Lin, Sundaram Gopalakrishnan, Ting Chen, Francine Evans, Harald Rau, Ricky Bradley, and Dimitris Margaritis) are the real heroes of the war stories related within. My Stony Brook friends and algorithm colleagues Estie Arkin, Michael Bender, Jie Gao, and Joe Mitchell have always been a pleasure to work and be with. Finally, thanks to Michael Brochstein and the rest of the city contingent for revealing a proper life well beyond Stony Brook. Caveat It is traditional for the author to magnanimously accept the blame for whatever deﬁciencies remain. I don’t. Any errors, deﬁciencies, or problems in this book are somebody else’s fault, but I would appreciate knowing about them so as to determine who is to blame. Steven S. Skiena Department of Computer Science Stony Brook University Stony Brook, NY 11794-4400 http://www.cs.sunysb.edu/∼skiena April 2008 ix Contents I Practical Algorithm Design 1 Introduction to Algorithm Design 1.1 Robot Tour Optimization . . 1.2 Selecting the Right Jobs . . . 1.3 Reasoning about Correctness 1.4 Modeling the Problem . . . . 1.5 About the War Stories . . . . 1.6 War Story: Psychic Modeling 1.7 Exercises . . . . . . . . . . . . 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 5 9 11 19 22 23 27 2 Algorithm Analysis 2.1 The RAM Model of Computation . . . . 2.2 The Big Oh Notation . . . . . . . . . . . 2.3 Growth Rates and Dominance Relations 2.4 Working with the Big Oh . . . . . . . . 2.5 Reasoning About Eﬃciency . . . . . . . 2.6 Logarithms and Their Applications . . . 2.7 Properties of Logarithms . . . . . . . . . 2.8 War Story: Mystery of the Pyramids . . 2.9 Advanced Analysis (*) . . . . . . . . . . 2.10 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 31 34 37 40 41 46 50 51 54 57 3 Data Structures 3.1 Contiguous vs. Linked Data Structures . . . . . . . . . . . . . . . 65 66 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii CONTENTS 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 3.10 4 Sorting 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 4.10 4.11 Stacks and Queues . . . . . . . . . . Dictionaries . . . . . . . . . . . . . . Binary Search Trees . . . . . . . . . . Priority Queues . . . . . . . . . . . . War Story: Stripping Triangulations Hashing and Strings . . . . . . . . . Specialized Data Structures . . . . . War Story: String ’em Up . . . . . . Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 72 77 83 85 89 93 94 98 and Searching Applications of Sorting . . . . . . . . . . . . . Pragmatics of Sorting . . . . . . . . . . . . . . Heapsort: Fast Sorting via Data Structures . War Story: Give me a Ticket on an Airplane Mergesort: Sorting by Divide-and-Conquer . Quicksort: Sorting by Randomization . . . . Distribution Sort: Sorting via Bucketing . . . War Story: Skiena for the Defense . . . . . . Binary Search and Related Algorithms . . . . Divide-and-Conquer . . . . . . . . . . . . . . . Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 104 107 108 118 120 123 129 131 132 135 139 5 Graph Traversal 5.1 Flavors of Graphs . . . . . . . . . . . . . . . 5.2 Data Structures for Graphs . . . . . . . . . 5.3 War Story: I was a Victim of Moore’s Law 5.4 War Story: Getting the Graph . . . . . . . . 5.5 Traversing a Graph . . . . . . . . . . . . . . 5.6 Breadth-First Search . . . . . . . . . . . . . 5.7 Applications of Breadth-First Search . . . . 5.8 Depth-First Search . . . . . . . . . . . . . . 5.9 Applications of Depth-First Search . . . . . 5.10 Depth-First Search on Directed Graphs . . 5.11 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 146 151 155 158 161 162 166 169 172 178 184 6 Weighted Graph Algorithms 6.1 Minimum Spanning Trees . . . . . . . . 6.2 War Story: Nothing but Nets . . . . . . 6.3 Shortest Paths . . . . . . . . . . . . . . . 6.4 War Story: Dialing for Documents . . . 6.5 Network Flows and Bipartite Matching 6.6 Design Graphs, Not Algorithms . . . . . 6.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191 192 202 205 212 217 222 225 . . . . . . . . . . . . . . CONTENTS 7 Combinatorial Search and Heuristic Methods 7.1 Backtracking . . . . . . . . . . . . . . . . . . 7.2 Search Pruning . . . . . . . . . . . . . . . . 7.3 Sudoku . . . . . . . . . . . . . . . . . . . . . 7.4 War Story: Covering Chessboards . . . . . . 7.5 Heuristic Search Methods . . . . . . . . . . 7.6 War Story: Only it is Not a Radio . . . . . 7.7 War Story: Annealing Arrays . . . . . . . . 7.8 Other Heuristic Search Methods . . . . . . 7.9 Parallel Algorithms . . . . . . . . . . . . . . 7.10 War Story: Going Nowhere Fast . . . . . . . 7.11 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230 231 238 239 244 247 260 263 266 267 268 270 8 Dynamic Programming 8.1 Caching vs. Computation . . . . . . . . . . . 8.2 Approximate String Matching . . . . . . . . . 8.3 Longest Increasing Sequence . . . . . . . . . . 8.4 War Story: Evolution of the Lobster . . . . . 8.5 The Partition Problem . . . . . . . . . . . . . 8.6 Parsing Context-Free Grammars . . . . . . . 8.7 Limitations of Dynamic Programming: TSP . 8.8 War Story: What’s Past is Prolog . . . . . . . 8.9 War Story: Text Compression for Bar Codes 8.10 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273 274 280 289 291 294 298 301 304 307 310 9 Intractable Problems and Approximation Algorithms 9.1 Problems and Reductions . . . . . . . . . . . . . . . 9.2 Reductions for Algorithms . . . . . . . . . . . . . . . 9.3 Elementary Hardness Reductions . . . . . . . . . . . 9.4 Satisﬁability . . . . . . . . . . . . . . . . . . . . . . . 9.5 Creative Reductions . . . . . . . . . . . . . . . . . . 9.6 The Art of Proving Hardness . . . . . . . . . . . . . 9.7 War Story: Hard Against the Clock . . . . . . . . . . 9.8 War Story: And Then I Failed . . . . . . . . . . . . . 9.9 P vs. NP . . . . . . . . . . . . . . . . . . . . . . . . . 9.10 Dealing with NP-complete Problems . . . . . . . . . 9.11 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316 317 319 323 328 330 334 337 339 341 344 350 10 How to Design Algorithms II The Hitchhiker’s Guide to Algorithms 11 A Catalog of Algorithmic Problems 356 361 363 xiii xiv CONTENTS 12 Data Structures 12.1 Dictionaries . . . . . . . 12.2 Priority Queues . . . . . 12.3 Suﬃx Trees and Arrays . 12.4 Graph Data Structures . 12.5 Set Data Structures . . . 12.6 Kd-Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366 367 373 377 381 385 389 13 Numerical Problems 13.1 Solving Linear Equations . . . . . . . . . . . . 13.2 Bandwidth Reduction . . . . . . . . . . . . . 13.3 Matrix Multiplication . . . . . . . . . . . . . . 13.4 Determinants and Permanents . . . . . . . . . 13.5 Constrained and Unconstrained Optimization 13.6 Linear Programming . . . . . . . . . . . . . . 13.7 Random Number Generation . . . . . . . . . 13.8 Factoring and Primality Testing . . . . . . . . 13.9 Arbitrary-Precision Arithmetic . . . . . . . . 13.10 Knapsack Problem . . . . . . . . . . . . . . . 13.11 Discrete Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393 395 398 401 404 407 411 415 420 423 427 431 14 Combinatorial Problems 14.1 Sorting . . . . . . . . . . 14.2 Searching . . . . . . . . . 14.3 Median and Selection . . 14.4 Generating Permutations 14.5 Generating Subsets . . . 14.6 Generating Partitions . . 14.7 Generating Graphs . . . 14.8 Calendrical Calculations 14.9 Job Scheduling . . . . . . 14.10 Satisﬁability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 434 436 441 445 448 452 456 460 465 468 472 15 Graph Problems: Polynomial-Time 15.1 Connected Components . . . . . 15.2 Topological Sorting . . . . . . . . 15.3 Minimum Spanning Tree . . . . . 15.4 Shortest Path . . . . . . . . . . . 15.5 Transitive Closure and Reduction 15.6 Matching . . . . . . . . . . . . . . 15.7 Eulerian Cycle/Chinese Postman 15.8 Edge and Vertex Connectivity . . 15.9 Network Flow . . . . . . . . . . . 15.10 Drawing Graphs Nicely . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 475 477 481 484 489 495 498 502 505 509 513 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . CONTENTS 15.11 Drawing Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.12 Planarity Detection and Embedding . . . . . . . . . . . . . . . . 16 Graph Problems: Hard Problems 16.1 Clique . . . . . . . . . . . . . . 16.2 Independent Set . . . . . . . . 16.3 Vertex Cover . . . . . . . . . . 16.4 Traveling Salesman Problem . 16.5 Hamiltonian Cycle . . . . . . 16.6 Graph Partition . . . . . . . . 16.7 Vertex Coloring . . . . . . . . 16.8 Edge Coloring . . . . . . . . . 16.9 Graph Isomorphism . . . . . . 16.10 Steiner Tree . . . . . . . . . . 16.11 Feedback Edge/Vertex Set . . 517 520 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523 525 528 530 533 538 541 544 548 550 555 559 17 Computational Geometry 17.1 Robust Geometric Primitives . 17.2 Convex Hull . . . . . . . . . . . 17.3 Triangulation . . . . . . . . . . 17.4 Voronoi Diagrams . . . . . . . . 17.5 Nearest Neighbor Search . . . . 17.6 Range Search . . . . . . . . . . 17.7 Point Location . . . . . . . . . . 17.8 Intersection Detection . . . . . 17.9 Bin Packing . . . . . . . . . . . 17.10 Medial-Axis Transform . . . . . 17.11 Polygon Partitioning . . . . . . 17.12 Simplifying Polygons . . . . . . 17.13 Shape Similarity . . . . . . . . . 17.14 Motion Planning . . . . . . . . 17.15 Maintaining Line Arrangements 17.16 Minkowski Sum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 562 564 568 572 576 580 584 587 591 595 598 601 604 607 610 614 617 String Problems Set Cover . . . . . . . . . . . . . . . . . . . Set Packing . . . . . . . . . . . . . . . . . String Matching . . . . . . . . . . . . . . . Approximate String Matching . . . . . . . Text Compression . . . . . . . . . . . . . . Cryptography . . . . . . . . . . . . . . . . Finite State Machine Minimization . . . . Longest Common Substring/Subsequence Shortest Common Superstring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 620 621 625 628 631 637 641 646 650 654 18 Set and 18.1 18.2 18.3 18.4 18.5 18.6 18.7 18.8 18.9 xv xvi CONTENTS 19 Algorithmic Resources 19.1 Software Systems . . . . . . . . . 19.2 Data Sources . . . . . . . . . . . . 19.3 Online Bibliographic Resources . 19.4 Professional Consulting Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 657 657 663 663 664 Bibliography 665 Index 709 1 Introduction to Algorithm Design What is an algorithm? An algorithm is a procedure to accomplish a speciﬁc task. An algorithm is the idea behind any reasonable computer program. To be interesting, an algorithm must solve a general, well-speciﬁed problem. An algorithmic problem is speciﬁed by describing the complete set of instances it must work on and of its output after running on one of these instances. This distinction, between a problem and an instance of a problem, is fundamental. For example, the algorithmic problem known as sorting is deﬁned as follows: Problem: Sorting Input: A sequence of n keys a1 , . . . , an . Output: The permutation (reordering) of the input sequence such that a1 ≤ a2 ≤ · · · ≤ an−1 ≤ an . An instance of sorting might be an array of names, like {Mike, Bob, Sally, Jill, Jan}, or a list of numbers like {154, 245, 568, 324, 654, 324}. Determining that you are dealing with a general problem is your ﬁrst step towards solving it. An algorithm is a procedure that takes any of the possible input instances and transforms it to the desired output. There are many diﬀerent algorithms for solving the problem of sorting. For example, insertion sort is a method for sorting that starts with a single element (thus forming a trivially sorted list) and then incrementally inserts the remaining elements so that the list stays sorted. This algorithm, implemented in C, is described below: S.S. Skiena, The Algorithm Design Manual, 2nd ed., DOI: 10.1007/978-1-84800-070-4 1, c Springer-Verlag London Limited 2008 4 1. INTRODUCTION TO ALGORITHM DESIGN I N I N I N E I E I E I E I E I E I E I E I E I E I S E R T I O N S O S E R T I O N S O S E R T I O N S O N S R T I O N S O N R S T I O N S O N R S T I O N S O I N R S T O N S O I N O R S T N S O I N N O R S T S O I N N O R S S T O I N N O O R S S T I N N O O R R S S I N N O O R R S S R R R R R R R R R R R T T T T T T T T T T T T T T T Figure 1.1: Animation of insertion sort in action (time ﬂows down) insertion_sort(item s[], int n) { int i,j; /* counters */ for (i=1; i0) && (s[j] < s[j-1])) { swap(&s[j],&s[j-1]); j = j-1; } } } An animation of the logical ﬂow of this algorithm on a particular instance (the letters in the word “INSERTIONSORT”) is given in Figure 1.1 Note the generality of this algorithm. It works just as well on names as it does on numbers, given the appropriate comparison operation (<) to test which of the two keys should appear ﬁrst in sorted order. It can be readily veriﬁed that this algorithm correctly orders every possible input instance according to our deﬁnition of the sorting problem. There are three desirable properties for a good algorithm. We seek algorithms that are correct and eﬃcient, while being easy to implement. These goals may not be simultaneously achievable. In industrial settings, any program that seems to give good enough answers without slowing the application down is often acceptable, regardless of whether a better algorithm exists. The issue of ﬁnding the best possible answer or achieving maximum eﬃciency usually arises in industry only after serious performance or legal troubles. In this chapter, we will focus on the issues of algorithm correctness, and defer a discussion of eﬃciency concerns to Chapter 2. It is seldom obvious whether a given 1.1 ROBOT TOUR OPTIMIZATION 0 0 8 1 7 2 6 3 5 4 Figure 1.2: A good instance for the nearest-neighbor heuristic algorithm correctly solves a given problem. Correct algorithms usually come with a proof of correctness, which is an explanation of why we know that the algorithm must take every instance of the problem to the desired result. However, before we go further we demonstrate why “it’s obvious” never suﬃces as a proof of correctness, and is usually ﬂat-out wrong. 1.1 Robot Tour Optimization Let’s consider a problem that arises often in manufacturing, transportation, and testing applications. Suppose we are given a robot arm equipped with a tool, say a soldering iron. In manufacturing circuit boards, all the chips and other components must be fastened onto the substrate. More speciﬁcally, each chip has a set of contact points (or wires) that must be soldered to the board. To program the robot arm for this job, we must ﬁrst construct an ordering of the contact points so the robot visits (and solders) the ﬁrst contact point, then the second point, third, and so forth until the job is done. The robot arm then proceeds back to the ﬁrst contact point to prepare for the next board, thus turning the tool-path into a closed tour, or cycle. Robots are expensive devices, so we want the tour that minimizes the time it takes to assemble the circuit board. A reasonable assumption is that the robot arm moves with ﬁxed speed, so the time to travel between two points is proportional to their distance. In short, we must solve the following algorithm problem: Problem: Robot Tour Optimization Input: A set S of n points in the plane. Output: What is the shortest cycle tour that visits each point in the set S? You are given the job of programming the robot arm. Stop right now and think up an algorithm to solve this problem. I’ll be happy to wait until you ﬁnd one. . . 5 6 1. INTRODUCTION TO ALGORITHM DESIGN Several algorithms might come to mind to solve this problem. Perhaps the most popular idea is the nearest-neighbor heuristic. Starting from some point p0 , we walk ﬁrst to its nearest neighbor p1 . From p1 , we walk to its nearest unvisited neighbor, thus excluding only p0 as a candidate. We now repeat this process until we run out of unvisited points, after which we return to p0 to close oﬀ the tour. Written in pseudo-code, the nearest-neighbor heuristic looks like this: NearestNeighbor(P ) Pick and visit an initial point p0 from P p = p0 i=0 While there are still unvisited points i=i+1 Select pi to be the closest unvisited point to pi−1 Visit pi Return to p0 from pn−1 This algorithm has a lot to recommend it. It is simple to understand and implement. It makes sense to visit nearby points before we visit faraway points to reduce the total travel time. The algorithm works perfectly on the example in Figure 1.2. The nearest-neighbor rule is reasonably eﬃcient, for it looks at each pair of points (pi , pj ) at most twice: once when adding pi to the tour, the other when adding pj . Against all these positives there is only one problem. This algorithm is completely wrong. Wrong? How can it be wrong? The algorithm always ﬁnds a tour, but it doesn’t necessarily ﬁnd the shortest possible tour. It doesn’t necessarily even come close. Consider the set of points in Figure 1.3, all of which lie spaced along a line. The numbers describe the distance that each point lies to the left or right of the point labeled ‘0’. When we start from the point ‘0’ and repeatedly walk to the nearest unvisited neighbor, we might keep jumping left-right-left-right over ‘0’ as the algorithm oﬀers no advice on how to break ties. A much better (indeed optimal) tour for these points starts from the leftmost point and visits each point as we walk right before returning at the rightmost point. Try now to imagine your boss’s delight as she watches a demo of your robot arm hopscotching left-right-left-right during the assembly of such a simple board. “But wait,” you might be saying. “The problem was in starting at point ‘0’. Instead, why don’t we start the nearest-neighbor rule using the leftmost point as the initial point p0 ? By doing this, we will ﬁnd the optimal solution on this instance.” That is 100% true, at least until we rotate our example 90 degrees. Now all points are equally leftmost. If the point ‘0’ were moved just slightly to the left, it would be picked as the starting point. Now the robot arm will hopscotch up-downup-down instead of left-right-left-right, but the travel time will be just as bad as before. No matter what you do to pick the ﬁrst point, the nearest-neighbor rule is doomed to work incorrectly on certain point sets. 1.1 -21 -21 -5 -5 ROBOT TOUR OPTIMIZATION -1 -1 0 0 1 1 3 3 11 11 Figure 1.3: A bad instance for the nearest-neighbor heuristic, with the optimal solution Maybe what we need is a diﬀerent approach. Always walking to the closest point is too restrictive, since it seems to trap us into making moves we didn’t want. A diﬀerent idea might be to repeatedly connect the closest pair of endpoints whose connection will not create a problem, such as premature termination of the cycle. Each vertex begins as its own single vertex chain. After merging everything together, we will end up with a single chain containing all the points in it. Connecting the ﬁnal two endpoints gives us a cycle. At any step during the execution of this closest-pair heuristic, we will have a set of single vertices and vertex-disjoint chains available to merge. In pseudocode: ClosestPair(P ) Let n be the number of points in set P . For i = 1 to n − 1 do d=∞ For each pair of endpoints (s, t) from distinct vertex chains if dist(s, t) ≤ d then sm = s, tm = t, and d = dist(s, t) Connect (sm , tm ) by an edge Connect the two endpoints by an edge This closest-pair rule does the right thing in the example in Figure 1.3. It starts by connecting ‘0’ to its immediate neighbors, the points 1 and −1. Subsequently, the next closest pair will alternate left-right, growing the central path by one link at a time. The closest-pair heuristic is somewhat more complicated and less eﬃcient than the previous one, but at least it gives the right answer in this example. But this is not true in all examples. Consider what this algorithm does on the point set in Figure 1.4(l). It consists of two rows of equally spaced points, with the rows slightly closer together (distance 1 − e) than the neighboring points are spaced within each row (distance 1 + e). Thus the closest pairs of points stretch across the gap, not around the boundary. After we pair oﬀ these points, the closest 7 8 1. INTRODUCTION TO ALGORITHM DESIGN 1+e 1+e 1−e 1−e 1+e (l) 1−e 1−e 1+e (r) Figure 1.4: A bad instance for the closest-pair heuristic, with the optimal solution remaining pairs will connect these pairs alternately around the boundary. The total path length of the closest-pair tour is 3(1 − e) + 2(1 + e) + (1 − e)2 + (2 + 2e)2 . Compared to the tour shown in Figure 1.4(r), we travel over 20% farther than necessary when e ≈ 0. Examples exist where the penalty is considerably worse than this. Thus this second algorithm is also wrong. Which one of these algorithms performs better? You can’t tell just by looking at them. Clearly, both heuristics can end up with very bad tours on very innocent-looking input. At this point, you might wonder what a correct algorithm for our problem looks like. Well, we could try enumerating all possible orderings of the set of points, and then select the ordering that minimizes the total length: OptimalTSP(P) d=∞ For each of the n! permutations Pi of point set P If (cost(Pi ) ≤ d) then d = cost(Pi ) and Pmin = Pi Return Pmin Since all possible orderings are considered, we are guaranteed to end up with the shortest possible tour. This algorithm is correct, since we pick the best of all the possibilities. But it is also extremely slow. The fastest computer in the world couldn’t hope to enumerate all the 20! =2,432,902,008,176,640,000 orderings of 20 points within a day. For real circuit boards, where n ≈ 1, 000, forget about it. All of the world’s computers working full time wouldn’t come close to ﬁnishing the problem before the end of the universe, at which point it presumably becomes moot. The quest for an eﬃcient algorithm to solve this problem, called the traveling salesman problem (TSP), will take us through much of this book. If you need to know how the story ends, check out the catalog entry for the traveling salesman problem in Section 16.4 (page 533). 1.2 Tarjan of the Jungle The President’s Algorist SELECTING THE RIGHT JOBS The Four Volume Problem Steiner’s Tree Halting State "Discrete" Mathematics Process Terminated Programming Challenges Calculated Bets Figure 1.5: An instance of the non-overlapping movie scheduling problem Take-Home Lesson: There is a fundamental diﬀerence between algorithms, which always produce a correct result, and heuristics, which may usually do a good job but without providing any guarantee. 1.2 Selecting the Right Jobs Now consider the following scheduling problem. Imagine you are a highly-indemand actor, who has been presented with oﬀers to star in n diﬀerent movie projects under development. Each oﬀer comes speciﬁed with the ﬁrst and last day of ﬁlming. To take the job, you must commit to being available throughout this entire period. Thus you cannot simultaneously accept two jobs whose intervals overlap. For an artist such as yourself, the criteria for job acceptance is clear: you want to make as much money as possible. Because each of these ﬁlms pays the same fee per ﬁlm, this implies you seek the largest possible set of jobs (intervals) such that no two of them conﬂict with each other. For example, consider the available projects in Figure 1.5. We can star in at most four ﬁlms, namely “Discrete” Mathematics, Programming Challenges, Calculated Bets, and one of either Halting State or Steiner’s Tree. You (or your agent) must solve the following algorithmic scheduling problem: Problem: Movie Scheduling Problem Input: A set I of n intervals on the line. Output: What is the largest subset of mutually non-overlapping intervals which can be selected from I? You are given the job of developing a scheduling algorithm for this task. Stop right now and try to ﬁnd one. Again, I’ll be happy to wait. There are several ideas that may come to mind. One is based on the notion that it is best to work whenever work is available. This implies that you should start with the job with the earliest start date – after all, there is no other job you can work on, then at least during the begining of this period. 9 10 1. INTRODUCTION TO ALGORITHM DESIGN War and Peace (l) (r) Figure 1.6: Bad instances for the (l) earliest job ﬁrst and (r) shortest job ﬁrst heuristics. EarliestJobFirst(I) Accept the earlest starting job j from I which does not overlap any previously accepted job, and repeat until no more such jobs remain. This idea makes sense, at least until we realize that accepting the earliest job might block us from taking many other jobs if that ﬁrst job is long. Check out Figure 1.6(l), where the epic “War and Peace” is both the ﬁrst job available and long enough to kill oﬀ all other prospects. This bad example naturally suggests another idea. The problem with “War and Peace” is that it is too long. Perhaps we should start by taking the shortest job, and keep seeking the shortest available job at every turn. Maximizing the number of jobs we do in a given period is clearly connected to banging them out as quickly as possible. This yields the heuristic: ShortestJobFirst(I) While (I = ∅) do Accept the shortest possible job j from I. Delete j, and any interval which intersects j from I. Again this idea makes sense, at least until we realize that accepting the shortest job might block us from taking two other jobs, as shown in Figure 1.6(r). While the potential loss here seems smaller than with the previous heuristic, it can readily limit us to half the optimal payoﬀ. At this point, an algorithm where we try all possibilities may start to look good, because we can be certain it is correct. If we ignore the details of testing whether a set of intervals are in fact disjoint, it looks something like this: ExhaustiveScheduling(I) j=0 Smax = ∅ For each of the 2n subsets Si of intervals I If (Si is mutally non-overlapping) and (size(Si ) > j) then j = size(Si ) and Smax = Si . Return Smax But how slow is it? The key limitation is enumerating the 2n subsets of n things. The good news is that this is much better than enumerating all n! orders 1.3 REASONING ABOUT CORRECTNESS of n things, as proposed for the robot tour optimization problem. There are only about one million subsets when n = 20, which could be exhaustively counted within seconds on a decent computer. However, when fed n = 100 movies, 2100 is much much greater than the 20! which made our robot cry “uncle” in the previous problem. The diﬀerence between our scheduling and robotics problems are that there is an algorithm which solves movie scheduling both correctly and eﬃciently. Think about the ﬁrst job to terminate—i.e. the interval x which contains the rightmost point which is leftmost among all intervals. This role is played by “Discrete” Mathematics in Figure 1.5. Other jobs may well have started before x, but all of these must at least partially overlap each other, so we can select at most one from the group. The ﬁrst of these jobs to terminate is x, so any of the overlapping jobs potentially block out other opportunities to the right of it. Clearly we can never lose by picking x. This suggests the following correct, eﬃcient algorithm: OptimalScheduling(I) While (I = ∅) do Accept the job j from I with the earliest completion date. Delete j, and any interval which intersects j from I. Ensuring the optimal answer over all possible inputs is a diﬃcult but often achievable goal. Seeking counterexamples that break pretender algorithms is an important part of the algorithm design process. Eﬃcient algorithms are often lurking out there; this book seeks to develop your skills to help you ﬁnd them. Take-Home Lesson: Reasonable-looking algorithms can easily be incorrect. Algorithm correctness is a property that must be carefully demonstrated. 1.3 Reasoning about Correctness Hopefully, the previous examples have opened your eyes to the subtleties of algorithm correctness. We need tools to distinguish correct algorithms from incorrect ones, the primary one of which is called a proof. A proper mathematical proof consists of several parts. First, there is a clear, precise statement of what you are trying to prove. Second, there is a set of assumptions of things which are taken to be true and hence used as part of the proof. Third, there is a chain of reasoning which takes you from these assumptions to the statement you are trying to prove. Finally, there is a little square ( ) or QED at the bottom to denote that you have ﬁnished, representing the Latin phrase for “thus it is demonstrated.” This book is not going to emphasize formal proofs of correctness, because they are very diﬃcult to do right and quite misleading when you do them wrong. A proof is indeed a demonstration. Proofs are useful only when they are honest; crisp arguments explaining why an algorithm satisﬁes a nontrivial correctness property. 11 12 1. INTRODUCTION TO ALGORITHM DESIGN Correct algorithms require careful exposition, and eﬀorts to show both correctness and not incorrectness. We develop tools for doing so in the subsections below. 1.3.1 Expressing Algorithms Reasoning about an algorithm is impossible without a careful description of the sequence of steps to be performed. The three most common forms of algorithmic notation are (1) English, (2) pseudocode, or (3) a real programming language. We will use all three in this book. Pseudocode is perhaps the most mysterious of the bunch, but it is best deﬁned as a programming language that never complains about syntax errors. All three methods are useful because there is a natural tradeoﬀ between greater ease of expression and precision. English is the most natural but least precise programming language, while Java and C/C++ are precise but diﬃcult to write and understand. Pseudocode is generally useful because it represents a happy medium. The choice of which notation is best depends upon which method you are most comfortable with. I usually prefer to describe the ideas of an algorithm in English, moving to a more formal, programming-language-like pseudocode or even real code to clarify suﬃciently tricky details. A common mistake my students make is to use pseudocode to dress up an illdeﬁned idea so that it looks more formal. Clarity should be the goal. For example, the ExhaustiveScheduling algorithm on page 10 could have been better written in English as: ExhaustiveScheduling(I) Test all 2n subsets of intervals from I, and return the largest subset consisting of mutually non-overlapping intervals. Take-Home Lesson: The heart of any algorithm is an idea. If your idea is not clearly revealed when you express an algorithm, then you are using too low-level a notation to describe it. 1.3.2 Problems and Properties We need more than just an algorithm description in order to demonstrate correctness. We also need a careful description of the problem that it is intended to solve. Problem speciﬁcations have two parts: (1) the set of allowed input instances, and (2) the required properties of the algorithm’s output. It is impossible to prove the correctness of an algorithm for a fuzzily-stated problem. Put another way, ask the wrong problem and you will get the wrong answer. Some problem speciﬁcations allow too broad a class of input instances. Suppose we had allowed ﬁlm projects in our movie scheduling problem to have gaps in 1.3 REASONING ABOUT CORRECTNESS production (i.e. , ﬁlming in September and November but a hiatus in October). Then the schedule associated with any particular ﬁlm would consist of a given set of intervals. Our star would be free to take on two interleaving but not overlapping projects (such as the ﬁlm above nested with one ﬁlming in August and October). The earliest completion algorithm would not work for such a generalized scheduling problem. Indeed, no eﬃcient algorithm exists for this generalized problem. Take-Home Lesson: An important and honorable technique in algorithm design is to narrow the set of allowable instances until there is a correct and eﬃcient algorithm. For example, we can restrict a graph problem from general graphs down to trees, or a geometric problem from two dimensions down to one. There are two common traps in specifying the output requirements of a problem. One is asking an ill-deﬁned question. Asking for the best route between two places on a map is a silly question unless you deﬁne what best means. Do you mean the shortest route in total distance, or the fastest route, or the one minimizing the number of turns? The second trap is creating compound goals. The three path-planning criteria mentioned above are all well-deﬁned goals that lead to correct, eﬃcient optimization algorithms. However, you must pick a single criteria. A goal like Find the shortest path from a to b that doesn’t use more than twice as many turns as necessary is perfectly well deﬁned, but complicated to reason and solve. I encourage you to check out the problem statements for each of the 75 catalog problems in the second part of this book. Finding the right formulation for your problem is an important part of solving it. And studying the deﬁnition of all these classic algorithm problems will help you recognize when someone else has thought about similar problems before you. 1.3.3 Demonstrating Incorrectness The best way to prove that an algorithm is incorrect is to produce an instance in which it yields an incorrect answer. Such instances are called counter-examples. No rational person will ever leap to the defense of an algorithm after a counterexample has been identiﬁed. Very simple instances can instantly kill reasonablelooking heuristics with a quick touché. Good counter-examples have two important properties: • Veriﬁability – To demonstrate that a particular instance is a counter-example to a particular algorithm, you must be able to (1) calculate what answer your algorithm will give in this instance, and (2) display a better answer so as to prove the algorithm didn’t ﬁnd it. Since you must hold the given instance in your head to reason about it, an important part of veriﬁability is. . . 13 14 1. INTRODUCTION TO ALGORITHM DESIGN • Simplicity – Good counter-examples have all unnecessary details boiled away. They make clear exactly why the proposed algorithm fails. Once a counterexample has been found, it is worth simplifying it down to its essence. For example, the counter-example of Figure 1.6(l) could be made simpler and better by reducing the number of overlapped segments from four to two. Hunting for counter-examples is a skill worth developing. It bears some similarity to the task of developing test sets for computer programs, but relies more on inspiration than exhaustion. Here are some techniques to aid your quest: • Think small – Note that the robot tour counter-examples I presented boiled down to six points or less, and the scheduling counter-examples to only three intervals. This is indicative of the fact that when algorithms fail, there is usually a very simple example on which they fail. Amateur algorists tend to draw a big messy instance and then stare at it helplessly. The pros look carefully at several small examples, because they are easier to verify and reason about. • Think exhaustively – There are only a small number of possibilities for the smallest nontrivial value of n. For example, there are only three interesting ways two intervals on the line can occur: (1) as disjoint intervals, (2) as overlapping intervals, and (3) as properly nesting intervals, one within the other. All cases of three intervals (including counter-examples to both movie heuristics) can be systematically constructed by adding a third segment in each possible way to these three instances. • Hunt for the weakness – If a proposed algorithm is of the form “always take the biggest” (better known as the greedy algorithm), think about why that might prove to be the wrong thing to do. In particular, . . . • Go for a tie – A devious way to break a greedy heuristic is to provide instances where everything is the same size. Suddenly the heuristic has nothing to base its decision on, and perhaps has the freedom to return something suboptimal as the answer. • Seek extremes – Many counter-examples are mixtures of huge and tiny, left and right, few and many, near and far. It is usually easier to verify or reason about extreme examples than more muddled ones. Consider two tightly bunched clouds of points separated by a much larger distance d. The optimal TSP tour will be essentially 2d regardless of the number of points, because what happens within each cloud doesn’t really matter. Take-Home Lesson: Searching for counterexamples is the best way to disprove the correctness of a heuristic. 1.3 1.3.4 REASONING ABOUT CORRECTNESS Induction and Recursion Failure to ﬁnd a counterexample to a given algorithm does not mean “it is obvious” that the algorithm is correct. A proof or demonstration of correctness is needed. Often mathematical induction is the method of choice. When I ﬁrst learned about mathematical induction it seemed like complete n magic. You proved a formula like i=1 i = n(n + 1)/2 for some basis case like 1 or 2, then assumed it was true all the way to n − 1 before proving it was true for general n using the assumption. That was a proof? Ridiculous! When I ﬁrst learned the programming technique of recursion it also seemed like complete magic. The program tested whether the input argument was some basis case like 1 or 2. If not, you solved the bigger case by breaking it into pieces and calling the subprogram itself to solve these pieces. That was a program? Ridiculous! The reason both seemed like magic is because recursion is mathematical induction. In both, we have general and boundary conditions, with the general condition breaking the problem into smaller and smaller pieces. The initial or boundary condition terminates the recursion. Once you understand either recursion or induction, you should be able to see why the other one also works. I’ve heard it said that a computer scientist is a mathematician who only knows how to prove things by induction. This is partially true because computer scientists are lousy at proving things, but primarily because so many of the algorithms we study are either recursive or incremental. Consider the correctness of insertion sort, which we introduced at the beginning of this chapter. The reason it is correct can be shown inductively: • The basis case consists of a single element, and by deﬁnition a one-element array is completely sorted. • In general, we can assume that the ﬁrst n − 1 elements of array A are completely sorted after n − 1 iterations of insertion sort. • To insert one last element x to A, we ﬁnd where it goes, namely the unique spot between the biggest element less than or equal to x and the smallest element greater than x. This is done by moving all the greater elements back by one position, creating room for x in the desired location. One must be suspicious of inductive proofs, however, because very subtle reasoning errors can creep in. The ﬁrst are boundary errors. For example, our insertion sort correctness proof above boldly stated that there was a unique place to insert x between two elements, when our basis case was a single-element array. Greater care is needed to properly deal with the special cases of inserting the minimum or maximum elements. The second and more common class of inductive proof errors concerns cavallier extension claims. Adding one extra item to a given problem instance might cause the entire optimal solution to change. This was the case in our scheduling problem (see Figure 1.7). The optimal schedule after inserting a new segment may contain 15 16 1. INTRODUCTION TO ALGORITHM DESIGN Figure 1.7: Large-scale changes in the optimal solution (boxes) after inserting a single interval (dashed) into the instance none of the segments of any particular optimal solution prior to insertion. Boldly ignoring such diﬃculties can lead to very convincing inductive proofs of incorrect algorithms. Take-Home Lesson: Mathematical induction is usually the right way to verify the correctness of a recursive or incremental insertion algorithm. Stop and Think: Incremental Correctness Problem: Prove the correctness of the following recursive algorithm for incrementing natural numbers, i.e. y → y + 1: Increment(y) if y = 0 then return(1) else if (y mod 2) = 1 then return(2 · Increment( y/2 )) else return(y + 1) Solution: The correctness of this algorithm is certainly not obvious to me. But as it is recursive and I am a computer scientist, my natural instinct is to try to prove it by induction. The basis case of y = 0 is obviously correctly handled. Clearly the value 1 is returned, and 0 + 1 = 1. Now assume the function works correctly for the general case of y = n−1. Given this, we must demonstrate the truth for the case of y = n. Half of the cases are easy, namely the even numbers (For which (y mod 2) = 0), since y + 1 is explicitly returned. For the odd numbers, the answer depends upon what is returned by Increment( y/2 ). Here we want to use our inductive assumption, but it isn’t quite right. We have assumed that increment worked correctly for y = n − 1, but not for a value which is about half of it. We can ﬁx this problem by strengthening our assumption to declare that the general case holds for all y ≤ n−1. This costs us nothing in principle, but is necessary to establish the correctness of the algorithm. 1.3 REASONING ABOUT CORRECTNESS Now, the case of odd y (i.e. y = 2m + 1 for some integer m) can be dealt with as: 2 · Increment( (2m + 1)/2 ) = 2 · Increment( m + 1/2 ) = 2 · Increment(m) = 2(m + 1) = 2m + 2 = y + 1 and the general case is resolved. 1.3.5 Summations Mathematical summation formulae arise often in algorithm analysis, which we will study in Chapter 2. Further, proving the correctness of summation formulae is a classic application of induction. Several exercises on inductive proofs of summations appear as exercises at the end this chapter. To make these more accessible, I review the basics of summations here. Summation formula are concise expressions describing the addition of an arbitrarily large set of numbers, in particular the formula n f (i) = f (1) + f (2) + . . . + f (n) i=1 There are simple closed forms for summations of many algebraic functions. For example, since n ones is n, n 1=n i=1 The sum of the ﬁrst n integers can be seen by pairing up the ith and (n − i + 1)th integers: n i=1 i= n/2 (i + (n − i + 1)) = n(n + 1)/2 i=1 Recognizing two basic classes of summation formulae will get you a long way in algorithm analysis: • Arithmetic progressions– We already encountered arithmetic progressions n when we saw S(n) = i i = n(n + 1)/2 in the analysis of selection sort. From the big picture perspective, the important thing is that the sum is quadratic, not that the constant is 1/2. In general, S(n, p) = n i ip = Θ(np+1 ) 17 18 1. INTRODUCTION TO ALGORITHM DESIGN for p ≥ 1. Thus the sum of squares is cubic, and the sum of cubes is quartic (if you use such a word). The “big Theta” notation (Θ(x)) will be properly explained in Section 2.2. For p < −1, this sum always converges to a constant, even as n → ∞. The interesting case is between results in . . . • Geometric series – In geometric progressions, the index of the loop eﬀects the exponent, i.e. G(n, a) = n ai = a(an+1 − 1)/(a − 1) i=0 How we interpret this sum depends upon the base of the progression, i.e. a. When a < 1, this converges to a constant even as n → ∞. This series convergence proves to be the great “free lunch” of algorithm analysis. It means that the sum of a linear number of things can be constant, not linear. For example, 1 + 1/2 + 1/4 + 1/8 + . . . ≤ 2 no matter how many terms we add up. When a > 1, the sum grows rapidly with each new term, as in 1 + 2 + 4 + 8 + 16 + 32 = 63. Indeed, G(n, a) = Θ(an+1 ) for a > 1. Stop and Think: Factorial Formulae Problem: Prove that n i=1 i × i! = (n + 1)! − 1 by induction. Solution: The inductive paradigm is straightforward. First verify the basis case (here we do n = 1, although n = 0 would be even more general): 1 i × i! = 1 = (1 + 1)! − 1 = 2 − 1 = 1 i=1 Now assume the statement is true up to n. To prove the general case of n + 1, observe that rolling out the largest term n+1 i × i! = (n + 1) × (n + 1)! + i=1 n i × i! i=1 reveals the left side of our inductive assumption. Substituting the right side gives us n+1 i=1 i × i! = (n + 1) × (n + 1)! + (n + 1)! − 1 1.4 = = MODELING THE PROBLEM (n + 1)! × ((n + 1) + 1) − 1 (n + 2)! − 1 This general trick of separating out the largest term from the summation to reveal an instance of the inductive assumption lies at the heart of all such proofs. 1.4 Modeling the Problem Modeling is the art of formulating your application in terms of precisely described, well-understood problems. Proper modeling is the key to applying algorithmic design techniques to real-world problems. Indeed, proper modeling can eliminate the need to design or even implement algorithms, by relating your application to what has been done before. Proper modeling is the key to eﬀectively using the “Hitchhiker’s Guide” in Part II of this book. Real-world applications involve real-world objects. You might be working on a system to route traﬃc in a network, to ﬁnd the best way to schedule classrooms in a university, or to search for patterns in a corporate database. Most algorithms, however, are designed to work on rigorously deﬁned abstract structures such as permutations, graphs, and sets. To exploit the algorithms literature, you must learn to describe your problem abstractly, in terms of procedures on fundamental structures. 1.4.1 Combinatorial Objects Odds are very good that others have stumbled upon your algorithmic problem before you, perhaps in substantially diﬀerent contexts. But to ﬁnd out what is known about your particular “widget optimization problem,” you can’t hope to look in a book under widget. You must formulate widget optimization in terms of computing properties of common structures such as: • Permutations – which are arrangements, or orderings, of items. For example, {1, 4, 3, 2} and {4, 3, 2, 1} are two distinct permutations of the same set of four integers. We have already seen permutations in the robot optimization problem, and in sorting. Permutations are likely the object in question whenever your problem seeks an “arrangement,” “tour,” “ordering,” or “sequence.” • Subsets – which represent selections from a set of items. For example, {1, 3, 4} and {2} are two distinct subsets of the ﬁrst four integers. Order does not matter in subsets the way it does with permutations, so the subsets {1, 3, 4} and {4, 3, 1} would be considered identical. We saw subsets arise in the movie scheduling problem. Subsets are likely the object in question whenever your problem seeks a “cluster,” “collection,” “committee,” “group,” “packaging,” or “selection.” 19 20 1. INTRODUCTION TO ALGORITHM DESIGN Sol Stony Brook Greenport Orient Point Riverhead Eve Morris Sid Shelter Island Montauk Steve Len Rob Richard Laurie Jim Lisa Jeff Islip Sag Harbor Figure 1.8: Modeling real-world structures with trees and graphs • Trees – which represent hierarchical relationships between items. Figure 1.8(a) shows part of the family tree of the Skiena clan. Trees are likely the object in question whenever your problem seeks a “hierarchy,” “dominance relationship,” “ancestor/descendant relationship,” or “taxonomy.” • Graphs – which represent relationships between arbitrary pairs of objects. Figure 1.8(b) models a network of roads as a graph, where the vertices are cities and the edges are roads connecting pairs of cities. Graphs are likely the object in question whenever you seek a “network,” “circuit,” “web,” or “relationship.” • Points – which represent locations in some geometric space. For example, the locations of McDonald’s restaurants can be described by points on a map/plane. Points are likely the object in question whenever your problems work on “sites,” “positions,” “data records,” or “locations.” • Polygons – which represent regions in some geometric spaces. For example, the borders of a country can be described by a polygon on a map/plane. Polygons and polyhedra are likely the object in question whenever you are working on “shapes,” “regions,” “conﬁgurations,” or “boundaries.” • Strings – which represent sequences of characters or patterns. For example, the names of students in a class can be represented by strings. Strings are likely the object in question whenever you are dealing with “text,” “characters,” “patterns,” or “labels.” These fundamental structures all have associated algorithm problems, which are presented in the catalog of Part II. Familiarity with these problems is important, because they provide the language we use to model applications. To become ﬂuent in this vocabulary, browse through the catalog and study the input and output pictures for each problem. Understanding these problems, even at a cartoon/deﬁnition level, will enable you to know where to look later when the problem arises in your application. 1.4 {4,1,5,2,3} 4+{1,4,2,3} {1,2,7,9} 9+{1,2,7} MODELING THE PROBLEM ALGORITHM 21 A LGORITHM Figure 1.9: Recursive decompositions of combinatorial objects. (left column) Permutations, subsets, trees, and graphs. (right column) Point sets, polygons, and strings Examples of successful application modeling will be presented in the war stories spaced throughout this book. However, some words of caution are in order. The act of modeling reduces your application to one of a small number of existing problems and structures. Such a process is inherently constraining, and certain details might not ﬁt easily into the given target problem. Also, certain problems can be modeled in several diﬀerent ways, some much better than others. Modeling is only the ﬁrst step in designing an algorithm for a problem. Be alert for how the details of your applications diﬀer from a candidate model, but don’t be too quick to say that your problem is unique and special. Temporarily ignoring details that don’t ﬁt can free the mind to ask whether they really were fundamental in the ﬁrst place. Take-Home Lesson: Modeling your application in terms of well-deﬁned structures and algorithms is the most important single step towards a solution. 1.4.2 Recursive Objects Learning to think recursively is learning to look for big things that are made from smaller things of exactly the same type as the big thing. If you think of houses as sets of rooms, then adding or deleting a room still leaves a house behind. Recursive structures occur everywhere in the algorithmic world. Indeed, each of the abstract structures described above can be thought about recursively. You just have to see how you can break them down, as shown in Figure 1.9: • Permutations – Delete the ﬁrst element of a permutation of {1, . . . , n} things and you get a permutation of the remaining n − 1 things. Permutations are recursive objects. 22 1. INTRODUCTION TO ALGORITHM DESIGN • Subsets – Every subset of the elements {1, . . . , n} contains a subset of {1, . . . , n − 1} made visible by deleting element n if it is present. Subsets are recursive objects. • Trees – Delete the root of a tree and what do you get? A collection of smaller trees. Delete any leaf of a tree and what do you get? A slightly smaller tree. Trees are recursive objects. • Graphs – Delete any vertex from a graph, and you get a smaller graph. Now divide the vertices of a graph into two groups, left and right. Cut through all edges which span from left to right, and what do you get? Two smaller graphs, and a bunch of broken edges. Graphs are recursive objects. • Points – Take a cloud of points, and separate them into two groups by drawing a line. Now you have two smaller clouds of points. Point sets are recursive objects. • Polygons – Inserting any internal chord between two nonadjacent vertices of a simple polygon on n vertices cuts it into two smaller polygons. Polygons are recursive objects. • Strings – Delete the ﬁrst character from a string, and what do you get? A shorter string. Strings are recursive objects. Recursive descriptions of objects require both decomposition rules and basis cases, namely the speciﬁcation of the smallest and simplest objects where the decomposition stops. These basis cases are usually easily deﬁned. Permutations and subsets of zero things presumably look like {}. The smallest interesting tree or graph consists of a single vertex, while the smallest interesting point cloud consists of a single point. Polygons are a little trickier; the smallest genuine simple polygon is a triangle. Finally, the empty string has zero characters in it. The decision of whether the basis case contains zero or one element is more a question of taste and convenience than any fundamental principle. Such recursive decompositions will come to deﬁne many of the algorithms we will see in this book. Keep your eyes open for them. 1.5 About the War Stories The best way to learn how careful algorithm design can have a huge impact on performance is to look at real-world case studies. By carefully studying other people’s experiences, we learn how they might apply to our work. Scattered throughout this text are several of my own algorithmic war stories, presenting our successful (and occasionally unsuccessful) algorithm design eﬀorts on real applications. I hope that you will be able to internalize these experiences so that they will serve as models for your own attacks on problems. 1.6 WAR STORY: PSYCHIC MODELING Every one of the war stories is true. Of course, the stories improve somewhat in the retelling, and the dialogue has been punched up to make them more interesting to read. However, I have tried to honestly trace the process of going from a raw problem to a solution, so you can watch how this process unfolded. The Oxford English Dictionary deﬁnes an algorist as “one skillful in reckonings or ﬁguring.” In these stories, I have tried to capture some of the mindset of the algorist in action as they attack a problem. The various war stories usually involve at least one, and often several, problems from the problem catalog in Part II. I reference the appropriate section of the catalog when such a problem occurs. This emphasizes the beneﬁts of modeling your application in terms of standard algorithm problems. By using the catalog, you will be able to pull out what is known about any given problem whenever it is needed. 1.6 War Story: Psychic Modeling The call came for me out of the blue as I sat in my oﬃce. “Professor Skiena, I hope you can help me. I’m the President of Lotto Systems Group Inc., and we need an algorithm for a problem arising in our latest product.” “Sure,” I replied. After all, the dean of my engineering school is always encouraging our faculty to interact more with industry. “At Lotto Systems Group, we market a program designed to improve our customers’ psychic ability to predict winning lottery numbers.1 In a standard lottery, each ticket consists of six numbers selected from, say, 1 to 44. Thus, any given ticket has only a very small chance of winning. However, after proper training, our clients can visualize, say, 15 numbers out of the 44 and be certain that at least four of them will be on the winning ticket. Are you with me so far?” “Probably not,” I replied. But then I recalled how my dean encourages us to interact with industry. “Our problem is this. After the psychic has narrowed the choices down to 15 numbers and is certain that at least 4 of them will be on the winning ticket, we must ﬁnd the most eﬃcient way to exploit this information. Suppose a cash prize is awarded whenever you pick at least three of the correct numbers on your ticket. We need an algorithm to construct the smallest set of tickets that we must buy in order to guarantee that we win at least one prize.” “Assuming the psychic is correct?” “Yes, assuming the psychic is correct. We need a program that prints out a list of all the tickets that the psychic should buy in order to minimize their investment. Can you help us?” Maybe they did have psychic ability, for they had come to the right place. Identifying the best subset of tickets to buy was very much a combinatorial algorithm 1 Yes, this is a true story. 23 24 1. INTRODUCTION TO ALGORITHM DESIGN 12 13 45 35 14 15 34 25 23 24 Figure 1.10: Covering all pairs of {1, 2, 3, 4, 5} with tickets {1, 2, 3}, {1, 4, 5}, {2, 4, 5}, {3, 4, 5} problem. It was going to be some type of covering problem, where each ticket we buy was going to “cover” some of the possible 4-element subsets of the psychic’s set. Finding the absolute smallest set of tickets to cover everything was a special instance of the NP-complete problem set cover (discussed in Section 18.1 (page 621)), and presumably computationally intractable. It was indeed a special instance of set cover, completely speciﬁed by only four numbers: the size n of the candidate set S (typically n ≈ 15), the number of slots k for numbers on each ticket (typically k ≈ 6), the number of psychically-promised correct numbers j from S (say j = 4), and ﬁnally, the number of matching numbers l necessary to win a prize (say l = 3). Figure 1.10 illustrates a covering of a smaller instance, where n = 5, j = k = 3, and l = 2. “Although it will be hard to ﬁnd the exact minimum set of tickets to buy, with heuristics I should be able to get you pretty close to the cheapest covering ticket set,” I told him. “Will that be good enough?” “So long as it generates better ticket sets than my competitor’s program, that will be ﬁne. His system doesn’t always guarantee a win. I really appreciate your help on this, Professor Skiena.” “One last thing. If your program can train people to pick lottery winners, why don’t you use it to win the lottery yourself?” “I look forward to talking to you again real soon, Professor Skiena. Thanks for the help.” I hung up the phone and got back to thinking. It seemed like the perfect project to give to a bright undergraduate. After modeling it in terms of sets and subsets, the basic components of a solution seemed fairly straightforward: 1.6 WAR STORY: PSYCHIC MODELING • We needed the ability to generate all subsets of k numbers from the candidate set S. Algorithms for generating and ranking/unranking subsets of sets are presented in Section 14.5 (page 452). • We needed the right formulation of what it meant to have a covering set of purchased tickets. The obvious criteria would be to pick a small set of tickets such that we have purchased at least one ticket containing each of the nl l-subsets of S that might pay oﬀ with the prize. • We needed to keep track of which prize combinations we have thus far covered. We seek tickets to cover as many thus-far-uncovered prize combinations as possible. The currently covered combinations are a subset of all possible combinations. Data structures for subsets are discussed in Section 12.5 (page 385). The best candidate seemed to be a bit vector, which would answer in constant time “is this combination already covered?” • We needed a search mechanism to decide which ticket to buy next. For small enough set sizes, we could do an exhaustive search over all possible subsets of tickets and pick the smallest one. For larger problems, a randomized search process like simulated annealing (see Section 7.5.3 (page 254)) would select tickets-to-buy to cover as many uncovered combinations as possible. By repeating this randomized procedure several times and picking the best solution, we would be likely to come up with a good set of tickets. Excluding the details of the search mechanism, the pseudocode for the bookkeeping looked something like this: LottoTicketSet(n, k,l) Initialize the nl -element bit-vector V to all false While there exists a false entry in V Select a k-subset T of {1, . . . , n} as the next ticket to buy For each of the l-subsets Ti of T , V [rank(Ti )] = true Report the set of tickets bought The bright undergraduate, Fayyaz Younas, rose to the challenge. Based on this framework, he implemented a brute-force search algorithm and found optimal solutions for problems with n ≤ 5 in a reasonable time. He implemented a random search procedure to solve larger problems, tweaking it for a while before settling on the best variant. Finally, the day arrived when we could call Lotto Systems Group and announce that we had solved the problem. “Our program found an optimal solution for n = 15, k = 6, j = 6, l = 3 meant buying 28 tickets.” “Twenty-eight tickets!” complained the president. “You must have a bug. Look, these ﬁve tickets will suﬃce to cover everything twice over: {2, 4, 8, 10, 13, 14}, {4, 5, 7, 8, 12, 15}, {1, 2, 3, 6, 11, 13}, {3, 5, 6, 9, 10, 15}, {1, 7, 9, 11, 12, 14}.” 25 26 1. INTRODUCTION TO ALGORITHM DESIGN 12 13 45 35 14 15 34 25 23 24 Figure 1.11: Guaranteeing a winning pair from {1, 2, 3, 4, 5} using only tickets {1, 2, 3} and {1, 4, 5} We ﬁddled with this example for a while before admitting that he was right. We hadn’t modeled the problem correctly! In fact, we didn’t need to explicitly cover all possible winning combinations. Figure 1.11 illustrates the principle by giving a twoticket solution to our previous four-ticket example. Such unpromising outcomes as {2, 3, 4} and {3, 4, 5} each agree in one matching pair with tickets from Figure 1.11. We were trying to cover too many combinations, and the penny-pinching psychics were unwilling to pay for such extravagance. Fortunately, this story has a happy ending. The general outline of our searchbased solution still holds for the real problem. All we must ﬁx is which subsets we get credit for covering with a given set of tickets. After this modiﬁcation, we obtained the kind of results they were hoping for. Lotto Systems Group gratefully accepted our program to incorporate into their product, and hopefully hit the jackpot with it. The moral of this story is to make sure that you model the problem correctly before trying to solve it. In our case, we came up with a reasonable model, but didn’t work hard enough to validate it before we started to program. Our misinterpretation would have become obvious had we worked out a small example by hand and bounced it oﬀ our sponsor before beginning work. Our success in recovering from this error is a tribute to the basic correctness of our initial formulation, and our use of well-deﬁned abstractions for such tasks as (1) ranking/unranking k-subsets, (2) the set data structure, and (3) combinatorial search. 1.7 EXERCISES Chapter Notes Every decent algorithm book reﬂects the design philosophy of its author. For students seeking alternative presentations and viewpoints, we particularly recommend the books of Corman, et. al [CLRS01], Kleinberg/Tardos [KT06], and Manber [Man89]. Formal proofs of algorithm correctness are important, and deserve a fuller discussion than we are able to provide in this chapter. See Gries [Gri89] for a thorough introduction to the techniques of program veriﬁcation. The movie scheduling problem represents a very special case of the general independent set problem, which is discussed in Section 16.2 (page 528). The restriction limits the allowable input instances to interval graphs, where the vertices of the graph G can be represented by intervals on the line and (i, j) is an edge of G iﬀ the intervals overlap. Golumbic [Gol04] provides a full treatment of this interesting and important class of graphs. Jon Bentley’s Programming Pearls columns are probably the best known collection of algorithmic “war stories.” Originally published in the Communications of the ACM, they have been collected in two books [Ben90, Ben99]. Brooks’s The Mythical Man Month [Bro95] is another wonderful collection of war stories, focused more on software engineering than algorithm design, but they remain a source of considerable wisdom. Every programmer should read all these books, for pleasure as well as insight. Our solution to the lotto ticket set covering problem is presented in more detail in [YS96]. 1.7 Exercises Finding Counterexamples 1-1. [3] Show that a + b can be less than min(a, b). 1-2. [3] Show that a × b can be less than min(a, b). 1-3. [5] Design/draw a road network with two points a and b such that the fastest route between a and b is not the shortest route. 1-4. [5] Design/draw a road network with two points a and b such that the shortest route between a and b is not the route with the fewest turns. 1-5. [4] The knapsack problem is as follows: given a set of integers S = {s1 , s2 , . . . , sn }, and a target number T , ﬁnd a subset of S which adds up exactly to T . For example, there exists a subset within S = {1, 2, 5, 9, 10} that adds up to T = 22 but not T = 23. Find counterexamples to each of the following algorithms for the knapsack problem. That is, giving an S and T such that the subset is selected using the algorithm does not leave the knapsack completely full, even though such a solution exists. 27 28 1. INTRODUCTION TO ALGORITHM DESIGN (a) Put the elements of S in the knapsack in left to right order if they ﬁt, i.e. the ﬁrst-ﬁt algorithm. (b) Put the elements of S in the knapsack from smallest to largest, i.e. the best-ﬁt algorithm. (c) Put the elements of S in the knapsack from largest to smallest. 1-6. [5] The set cover problem is as follows: given a set of subsets S1 , ..., Sm of the universal set U = {1, ..., n}, ﬁnd the smallest subset of subsets T ⊂ S such that ∪ti ∈T ti = U . For example, there are the following subsets, S1 = {1, 3, 5}, S2 = {2, 4}, S3 = {1, 4}, and S4 = {2, 5} The set cover would then be S1 and S2 . Find a counterexample for the following algorithm: Select the largest subset for the cover, and then delete all its elements from the universal set. Repeat by adding the subset containing the largest number of uncovered elements until all are covered. Proofs of Correctness 1-7. [3] Prove the correctness of the following recursive algorithm to multiply two natural numbers, for all integer constants c ≥ 2. function multiply(y, z) comment Return the product yz. 1. if z = 0 then return(0) else 2. return(multiply(cy, z/c) + y · (z mod c)) 1-8. [3] Prove the correctness of the following algorithm for evaluating a polynomial. P(x) = an xn + an−1 xn−1 + . . . + a1 x + a0 function horner(A, x) p = An for i from n − 1 to 0 p = p ∗ x + Ai return p 1-9. [3] Prove the correctness of the following sorting algorithm. function bubblesort (A : list[1 . . . n]) var int i, j for i from n to 1 for j from 1 to i − 1 if (A[j] > A[j + 1]) swap the values of A[j] and A[j + 1] Induction n i=n(n + 1)/2 for n ≥ 0, by induction. i=1 n 1-11. [3] Prove that i2 =n(n + 1)(2n + 1)/6 for n ≥ 0, by induction. i=1 n 3 2 2 1-10. [3] Prove that 1-12. [3] Prove that i=1 i =n (n + 1) /4 for n ≥ 0, by induction. 1-13. [3] Prove that n i=1 i(i + 1)(i + 2) = n(n + 1)(n + 2)(n + 3)/4 1.7 EXERCISES 1-14. [5] Prove by induction on n ≥ 1 that for every a = 1, n i=0 ai = an+1 − 1 a−1 1-15. [3] Prove by induction that for n ≥ 1, n i=1 1 n = i(i + 1) n+1 1-16. [3] Prove by induction that n3 + 2n is divisible by 3 for all n ≥ 0. 1-17. [3] Prove by induction that a tree with n vertices has exactly n − 1 edges. 1-18. [3] Prove by mathematical induction that the sum of the cubes of the ﬁrst n positive integers is equal to the square of the sum of these integers, i.e. n i=1 n i3 = ( i)2 i=1 Estimation 1-19. [3] Do all the books you own total at least one million pages? How many total pages are stored in your school library? 1-20. [3] How many words are there in this textbook? 1-21. [3] How many hours are one million seconds? How many days? Answer these questions by doing all arithmetic in your head. 1-22. [3] Estimate how many cities and towns there are in the United States. 1-23. [3] Estimate how many cubic miles of water ﬂow out of the mouth of the Mississippi River each day. Do not look up any supplemental facts. Describe all assumptions you made in arriving at your answer. 1-24. [3] Is disk drive access time normally measured in milliseconds (thousandths of a second) or microseconds (millionths of a second)? Does your RAM memory access a word in more or less than a microsecond? How many instructions can your CPU execute in one year if the machine is left running all the time? 1-25. [4] A sorting algorithm takes 1 second to sort 1,000 items on your local machine. How long will it take to sort 10,000 items. . . (a) if you believe that the algorithm takes time proportional to n2 , and (b) if you believe that the algorithm takes time roughly proportional to n log n? Implementation Projects 1-26. [5] Implement the two TSP heuristics of Section 1.1 (page 5). Which of them gives better-quality solutions in practice? Can you devise a heuristic that works better than both of them? 1-27. [5] Describe how to test whether a given set of tickets establishes suﬃcient coverage in the Lotto problem of Section 1.6 (page 23). Write a program to ﬁnd good ticket sets. 29 30 1. INTRODUCTION TO ALGORITHM DESIGN Interview Problems 1-28. [5] Write a function to perform integer division without using either the / or * operators. Find a fast way to do it. 1-29. [5] There are 25 horses. At most, 5 horses can race together at a time. You must determine the fastest, second fastest, and third fastest horses. Find the minimum number of races in which this can be done. 1-30. [3] How many piano tuners are there in the entire world? 1-31. [3] How many gas stations are there in the United States? 1-32. [3] How much does the ice in a hockey rink weigh? 1-33. [3] How many miles of road are there in the United States? 1-34. [3] On average, how many times would you have to ﬂip open the Manhattan phone book at random in order to ﬁnd a speciﬁc name? Programming Challenges These programming challenge problems with robot judging are available at http://www.programming-challenges.com or http://online-judge.uva.es. 1-1. “The 3n + 1 Problem” – Programming Challenges 110101, UVA Judge 100. 1-2. “The Trip” – Programming Challenges 110103, UVA Judge 10137. 1-3. “Australian Voting” – Programming Challenges 110108, UVA Judge 10142. 2 Algorithm Analysis Algorithms are the most important and durable part of computer science because they can be studied in a language- and machine-independent way. This means that we need techniques that enable us to compare the eﬃciency of algorithms without implementing them. Our two most important tools are (1) the RAM model of computation and (2) the asymptotic analysis of worst-case complexity. Assessing algorithmic performance makes use of the “big Oh” notation that, proves essential to compare algorithms and design more eﬃcient ones. While the hopelessly practical person may blanch at the notion of theoretical analysis, we present the material because it really is useful in thinking about algorithms. This method of keeping score will be the most mathematically demanding part of this book. But once you understand the intuition behind these ideas, the formalism becomes a lot easier to deal with. 2.1 The RAM Model of Computation Machine-independent algorithm design depends upon a hypothetical computer called the Random Access Machine or RAM. Under this model of computation, we are confronted with a computer where: • Each simple operation (+, *, –, =, if, call) takes exactly one time step. • Loops and subroutines are not considered simple operations. Instead, they are the composition of many single-step operations. It makes no sense for sort to be a single-step operation, since sorting 1,000,000 items will certainly take much longer than sorting 10 items. The time it takes to run through a loop or execute a subprogram depends upon the number of loop iterations or the speciﬁc nature of the subprogram. S.S. Skiena, The Algorithm Design Manual, 2nd ed., DOI: 10.1007/978-1-84800-070-4 2, c Springer-Verlag London Limited 2008 32 2. ALGORITHM ANALYSIS • Each memory access takes exactly one time step. Further, we have as much memory as we need. The RAM model takes no notice of whether an item is in cache or on the disk. Under the RAM model, we measure run time by counting up the number of steps an algorithm takes on a given problem instance. If we assume that our RAM executes a given number of steps per second, this operation count converts naturally to the actual running time. The RAM is a simple model of how computers perform. Perhaps it sounds too simple. After all, multiplying two numbers takes more time than adding two numbers on most processors, which violates the ﬁrst assumption of the model. Fancy compiler loop unrolling and hyperthreading may well violate the second assumption. And certainly memory access times diﬀer greatly depending on whether data sits in cache or on the disk. This makes us zero for three on the truth of our basic assumptions. And yet, despite these complaints, the RAM proves an excellent model for understanding how an algorithm will perform on a real computer. It strikes a ﬁne balance by capturing the essential behavior of computers while being simple to work with. We use the RAM model because it is useful in practice. Every model has a size range over which it is useful. Take, for example, the model that the Earth is ﬂat. You might argue that this is a bad model, since it has been fairly well established that the Earth is in fact round. But, when laying the foundation of a house, the ﬂat Earth model is suﬃciently accurate that it can be reliably used. It is so much easier to manipulate a ﬂat-Earth model that it is inconceivable that you would try to think spherically when you don’t have to.1 The same situation is true with the RAM model of computation. We make an abstraction that is generally very useful. It is quite diﬃcult to design an algorithm such that the RAM model gives you substantially misleading results. The robustness of the RAM enables us to analyze algorithms in a machine-independent way. Take-Home Lesson: Algorithms can be understood and studied in a languageand machine-independent manner. 2.1.1 Best, Worst, and Average-Case Complexity Using the RAM model of computation, we can count how many steps our algorithm takes on any given input instance by executing it. However, to understand how good or bad an algorithm is in general, we must know how it works over all instances. To understand the notions of the best, worst, and average-case complexity, think about running an algorithm over all possible instances of data that can be 1 The Earth is not completely spherical either, but a spherical Earth provides a useful model for such things as longitude and latitude. 2.1 THE RAM MODEL OF COMPUTATION Number of Steps Worst Case Average Case . . Best Case 1 2 3 4 ...... N Problem Size Figure 2.1: Best, worst, and average-case complexity fed to it. For the problem of sorting, the set of possible input instances consists of all possible arrangements of n keys, over all possible values of n. We can represent each input instance as a point on a graph (shown in Figure 2.1) where the x-axis represents the size of the input problem (for sorting, the number of items to sort), and the y-axis denotes the number of steps taken by the algorithm in this instance. These points naturally align themselves into columns, because only integers represent possible input size (e.g., it makes no sense to sort 10.57 items). We can deﬁne three interesting functions over the plot of these points: • The worst-case complexity of the algorithm is the function deﬁned by the maximum number of steps taken in any instance of size n. This represents the curve passing through the highest point in each column. • The best-case complexity of the algorithm is the function deﬁned by the minimum number of steps taken in any instance of size n. This represents the curve passing through the lowest point of each column. • The average-case complexity of the algorithm, which is the function deﬁned by the average number of steps over all instances of size n. The worst-case complexity proves to be most useful of these three measures in practice. Many people ﬁnd this counterintuitive. To illustrate why, try to project what will happen if you bring n dollars into a casino to gamble. The best case, that you walk out owning the place, is possible but so unlikely that you should not even think about it. The worst case, that you lose all n dollars, is easy to calculate and distressingly likely to happen. The average case, that the typical bettor loses 87.32% of the money that he brings to the casino, is diﬃcult to establish and its meaning subject to debate. What exactly does average mean? Stupid people lose 33 34 2. ALGORITHM ANALYSIS more than smart people, so are you smarter or stupider than the average person, and by how much? Card counters at blackjack do better on average than customers who accept three or more free drinks. We avoid all these complexities and obtain a very useful result by just considering the worst case. The important thing to realize is that each of these time complexities deﬁne a numerical function, representing time versus problem size. These functions are as well deﬁned as any other numerical function, be it y = x2 − 2x + 1 or the price of Google stock as a function of time. But time complexities are such complicated functions that we must simplify them to work with them. For this, we need the “Big Oh” notation. 2.2 The Big Oh Notation The best, worst, and average-case time complexities for any given algorithm are numerical functions over the size of possible problem instances. However, it is very diﬃcult to work precisely with these functions, because they tend to: • Have too many bumps – An algorithm such as binary search typically runs a bit faster for arrays of size exactly n = 2k − 1 (where k is an integer), because the array partitions work out nicely. This detail is not particularly signiﬁcant, but it warns us that the exact time complexity function for any algorithm is liable to be very complicated, with little up and down bumps as shown in Figure 2.2. • Require too much detail to specify precisely – Counting the exact number of RAM instructions executed in the worst case requires the algorithm be speciﬁed to the detail of a complete computer program. Further, the precise answer depends upon uninteresting coding details (e.g., did he use a case statement or nested ifs?). Performing a precise worst-case analysis like T (n) = 12754n2 + 4353n + 834 lg2 n + 13546 would clearly be very diﬃcult work, but provides us little extra information than the observation that “the time grows quadratically with n.” It proves to be much easier to talk in terms of simple upper and lower bounds of time-complexity functions using the Big Oh notation. The Big Oh simpliﬁes our analysis by ignoring levels of detail that do not impact our comparison of algorithms. The Big Oh notation ignores the diﬀerence between multiplicative constants. The functions f (n) = 2n and g(n) = n are identical in Big Oh analysis. This makes sense given our application. Suppose a given algorithm in (say) C language ran twice as fast as one with the same algorithm written in Java. This multiplicative 2.2 f(n) n THE BIG OH NOTATION 35 upper bound 0 lower bound 1 2 3 4 ....... n Figure 2.2: Upper and lower bounds valid for n > n0 smooth out the behavior of complex functions factor of two tells us nothing about the algorithm itself, since both programs implement exactly the same algorithm. We ignore such constant factors when comparing two algorithms. The formal deﬁnitions associated with the Big Oh notation are as follows: • f (n) = O(g(n)) means c · g(n) is an upper bound on f (n). Thus there exists some constant c such that f (n) is always ≤ c · g(n), for large enough n (i.e. , n ≥ n0 for some constant n0 ). • f (n) = Ω(g(n)) means c · g(n) is a lower bound on f (n). Thus there exists some constant c such that f (n) is always ≥ c · g(n), for all n ≥ n0 . • f (n) = Θ(g(n)) means c1 · g(n) is an upper bound on f (n) and c2 · g(n) is a lower bound on f (n), for all n ≥ n0 . Thus there exist constants c1 and c2 such that f (n) ≤ c1 · g(n) and f (n) ≥ c2 · g(n). This means that g(n) provides a nice, tight bound on f (n). Got it? These deﬁnitions are illustrated in Figure 2.3. Each of these deﬁnitions assumes a constant n0 beyond which they are always satisﬁed. We are not concerned about small values of n (i.e. , anything to the left of n0 ). After all, we don’t really care whether one sorting algorithm sorts six items faster than another, but seek which algorithm proves faster when sorting 10,000 or 1,000,000 items. The Big Oh notation enables us to ignore details and focus on the big picture. Take-Home Lesson: The Big Oh notation and worst-case analysis are tools that greatly simplify our ability to compare the eﬃciency of algorithms. Make sure you understand this notation by working through the following examples. We choose certain constants (c and n0 ) in the explanations below because 36 2. ALGORITHM ANALYSIS c1*g(n) c*g(n) f(n) f(n) c*g(n) n n0 c2*g(n) n n0 (a) f(n) n n0 (b) (c) Figure 2.3: Illustrating the big (a) O, (b) Ω, and (c) Θ notations they work and make a point, but other pairs of constants will do exactly the same job. You are free to choose any constants that maintain the same inequality—ideally constants that make it obvious that the inequality holds: 3n2 − 100n + 6 = O(n2 ), because I choose c = 3 and 3n2 > 3n2 − 100n + 6; 3n2 − 100n + 6 = O(n3 ), because I choose c = 1 and n3 > 3n2 − 100n + 6 when n > 3; 3n2 − 100n + 6 = O(n), because for any c I choose c × n < 3n2 when n > c; 3n2 − 100n + 6 = Ω(n2 ), because I choose c = 2 and 2n2 < 3n2 − 100n + 6 when n > 100; 3n2 − 100n + 6 = Ω(n3 ), because I choose c = 3 and 3n2 − 100n + 6 < n3 when n > 3; 3n2 − 100n + 6 = Ω(n), because for any c I choose cn < 3n2 − 100n + 6 when n > 100c; 3n2 − 100n + 6 = Θ(n2 ), because both O and Ω apply; 3n2 − 100n + 6 = Θ(n3 ), because only O applies; 3n2 − 100n + 6 = Θ(n), because only Ω applies. The Big Oh notation provides for a rough notion of equality when comparing functions. It is somewhat jarring to see an expression like n2 = O(n3 ), but its meaning can always be resolved by going back to the deﬁnitions in terms of upper and lower bounds. It is perhaps most instructive to read the “=” here as meaning one of the functions that are. Clearly, n2 is one of functions that are O(n3 ). 2.3 GROWTH RATES AND DOMINANCE RELATIONS Stop and Think: Back to the Deﬁnition Problem: Is 2n+1 = Θ(2n )? Solution: Designing novel algorithms requires cleverness and inspiration. However, applying the Big Oh notation is best done by swallowing any creative instincts you may have. All Big Oh problems can be correctly solved by going back to the deﬁnition and working with that. • Is 2n+1 = O(2n )? Well, f (n) = O(g(n)) iﬀ (if and only if) there exists a constant c such that for all suﬃciently large n f (n) ≤ c · g(n). Is there? The key observation is that 2n+1 = 2 · 2n , so 2 · 2n ≤ c · 2n for any c ≥ 2. • Is 2n+1 = Ω(2n )? Go back to the deﬁnition. f (n) = Ω(g(n)) iﬀ there exists a constant c > 0 such that for all suﬃciently large n f (n) ≥ c · g(n). This would be satisﬁed for any 0 < c ≤ 2. Together the Big Oh and Ω bounds imply 2n+1 = Θ(2n ) Stop and Think: Hip to the Squares? Problem: Is (x + y)2 = O(x2 + y 2 ). Solution: Working with the Big Oh means going back to the deﬁnition at the slightest sign of confusion. By deﬁnition, this expression is valid iﬀ we can ﬁnd some c such that (x + y)2 ≤ c(x2 + y 2 ). My ﬁrst move would be to expand the left side of the equation, i.e. (x + y)2 = 2 x +2xy +y 2 . If the middle 2xy term wasn’t there, the inequality would clearly hold for any c > 1. But it is there, so we need to relate the 2xy to x2 +y 2 . What if x ≤ y? Then 2xy ≤ 2y 2 ≤ 2(x2 + y 2 ). What if x ≥ y? Then 2xy ≤ 2x2 ≤ 2(x2 + y 2 ). Either way, we now can bound this middle term by two times the right-side function. This means that (x + y)2 ≤ 3(x2 + y 2 ), and so the result holds. 2.3 Growth Rates and Dominance Relations With the Big Oh notation, we cavalierly discard the multiplicative constants. Thus, the functions f (n) = 0.001n2 and g(n) = 1000n2 are treated identically, even though g(n) is a million times larger than f (n) for all values of n. 37 38 2. ALGORITHM ANALYSIS n f (n) 10 20 30 40 50 100 1,000 10,000 100,000 1,000,000 10,000,000 100,000,000 1,000,000,000 lg n 0.003 0.004 0.005 0.005 0.006 0.007 0.010 0.013 0.017 0.020 0.023 0.027 0.030 μs μs μs μs μs μs μs μs μs μs μs μs μs n 0.01 μs 0.02 μs 0.03 μs 0.04 μs 0.05 μs 0.1 μs 1.00 μs 10 μs 0.10 ms 1 ms 0.01 sec 0.10 sec 1 sec n lg n 0.033 μs 0.086 μs 0.147 μs 0.213 μs 0.282 μs 0.644 μs 9.966 μs 130 μs 1.67 ms 19.93 ms 0.23 sec 2.66 sec 29.90 sec n2 0.1 μs 0.4 μs 0.9 μs 1.6 μs 2.5 μs 10 μs 1 ms 100 ms 10 sec 16.7 min 1.16 days 115.7 days 31.7 years 2n 1 μs 1 ms 1 sec 18.3 min 13 days 4 × 1013 yrs n! 3.63 ms 77.1 years 8.4 × 1015 yrs Figure 2.4: Growth rates of common functions measured in nanoseconds The reason why we are content with coarse Big Oh analysis is provided by Figure 2.4, which shows the growth rate of several common time analysis functions. In particular, it shows how long algorithms that use f (n) operations take to run on a fast computer, where each operation takes one nanosecond (10−9 seconds). The following conclusions can be drawn from this table: • All such algorithms take roughly the same time for n = 10. • Any algorithm with n! running time becomes useless for n ≥ 20. • Algorithms whose running time is 2n have a greater operating range, but become impractical for n > 40. • Quadratic-time algorithms whose running time is n2 remain usable up to about n = 10, 000, but quickly deteriorate with larger inputs. They are likely to be hopeless for n > 1,000,000. • Linear-time and n lg n algorithms remain practical on inputs of one billion items. • An O(lg n) algorithm hardly breaks a sweat for any imaginable value of n. The bottom line is that even ignoring constant factors, we get an excellent idea of whether a given algorithm is appropriate for a problem of a given size. An algorithm whose running time is f (n) = n3 seconds will beat one whose running time is g(n) = 1,000,000 · n2 seconds only when n < 1,000,000. Such enormous diﬀerences in constant factors between algorithms occur far less frequently in practice than large problems do. 2.3 2.3.1 GROWTH RATES AND DOMINANCE RELATIONS Dominance Relations The Big Oh notation groups functions into a set of classes, such that all the functions in a particular class are equivalent with respect to the Big Oh. Functions f (n) = 0.34n and g(n) = 234,234n belong in the same class, namely those that are order Θ(n). Further, when two functions f and g belong to diﬀerent classes, they are diﬀerent with respect to our notation. Either f (n) = O(g(n)) or g(n) = O(f (n)), but not both. We say that a faster-growing function dominates a slower-growing one, just as a faster-growing country eventually comes to dominate the laggard. When f and g belong to diﬀerent classes (i.e., f (n) = Θ(g(n))), we say g dominates f when f (n) = O(g(n)), sometimes written g f. The good news is that only a few function classes tend to occur in the course of basic algorithm analysis. These suﬃce to cover almost all the algorithms we will discuss in this text, and are listed in order of increasing dominance: • Constant functions, f (n) = 1 – Such functions might measure the cost of adding two numbers, printing out “The Star Spangled Banner,” or the growth realized by functions such as f (n) = min(n, 100). In the big picture, there is no dependence on the parameter n. • Logarithmic functions, f (n) = log n – Logarithmic time-complexity shows up in algorithms such as binary search. Such functions grow quite slowly as n gets big, but faster than the constant function (which is standing still, after all). Logarithms will be discussed in more detail in Section 2.6 (page 46) • Linear functions, f (n) = n – Such functions measure the cost of looking at each item once (or twice, or ten times) in an n-element array, say to identify the biggest item, the smallest item, or compute the average value. • Superlinear functions, f (n) = n lg n – This important class of functions arises in such algorithms as Quicksort and Mergesort. They grow just a little faster than linear (see Figure 2.4), just enough to be a diﬀerent dominance class. • Quadratic functions, f (n) = n2 – Such functions measure the cost of looking at most or all pairs of items in an n-element universe. This arises in algorithms such as insertion sort and selection sort. • Cubic functions, f (n) = n3 – Such functions enumerate through all triples of items in an n-element universe. These also arise in certain dynamic programming algorithms developed in Chapter 8. • Exponential functions, f (n) = cn for a given constant c > 1 – Functions like 2n arise when enumerating all subsets of n items. As we have seen, exponential algorithms become useless fast, but not as fast as. . . • Factorial functions, f (n) = n! – Functions like n! arise when generating all permutations or orderings of n items. 39 40 2. ALGORITHM ANALYSIS The intricacies of dominance relations will be futher discussed in Section 2.9.2 (page 56). However, all you really need to understand is that: n! 2n n3 n2 n log n n log n 1 Take-Home Lesson: Although esoteric functions arise in advanced algorithm analysis, a small variety of time complexities suﬃce and account for most algorithms that are widely used in practice. 2.4 Working with the Big Oh You learned how to do simpliﬁcations of algebraic expressions back in high school. Working with the Big Oh requires dusting oﬀ these tools. Most of what you learned there still holds in working with the Big Oh, but not everything. 2.4.1 Adding Functions The sum of two functions is governed by the dominant one, namely: O(f (n)) + O(g(n)) → O(max(f (n), g(n))) Ω(f (n)) + Ω(g(n)) → Ω(max(f (n), g(n))) Θ(f (n)) + Θ(g(n)) → Θ(max(f (n), g(n))) This is very useful in simplifying expressions, since it implies that n3 + n2 + n + 1 = O(n3 ). Everything is small potatoes besides the dominant term. The intuition is as follows. At least half the bulk of f (n) + g(n) must come from the larger value. The dominant function will, by deﬁnition, provide the larger value as n → ∞. Thus, dropping the smaller function from consideration reduces the value by at most a factor of 1/2, which is just a multiplicative constant. Suppose f (n) = O(n2 ) and g(n) = O(n2 ). This implies that f (n) + g(n) = O(n2 ) as well. 2.4.2 Multiplying Functions Multiplication is like repeated addition. Consider multiplication by any constant c > 0, be it 1.02 or 1,000,000. Multiplying a function by a constant can not aﬀect its asymptotic behavior, because we can multiply the bounding constants in the Big Oh analysis of c · f (n) by 1/c to give appropriate constants for the Big Oh analysis of f (n). Thus: O(c · f (n)) → O(f (n)) Ω(c · f (n)) → Ω(f (n)) 2.5 REASONING ABOUT EFFICIENCY Θ(c · f (n)) → Θ(f (n)) Of course, c must be strictly positive (i.e., c > 0) to avoid any funny business, since we can wipe out even the fastest growing function by multiplying it by zero. On the other hand, when two functions in a product are increasing, both are important. The function O(n! log n) dominates n! just as much as log n dominates 1. In general, O(f (n)) ∗ O(g(n)) → O(f (n) ∗ g(n)) Ω(f (n)) ∗ Ω(g(n)) → Ω(f (n) ∗ g(n)) Θ(f (n)) ∗ Θ(g(n)) → Θ(f (n) ∗ g(n)) Stop and Think: Transitive Experience Problem: Show that Big Oh relationships are transitive. That is, if f (n) = O(g(n)) and g(n) = O(h(n)), then f (n) = O(h(n)). Solution: We always go back to the deﬁnition when working with the Big Oh. What we need to show here is that f (n) ≤ c3 h(n) for n > n3 given that f (n) ≤ c1 g(n) and g(n) ≤ c2 h(n), for n > n1 and n > n2 , respectively. Cascading these inequalities, we get that f (n) ≤ c1 g(n) ≤ c1 c2 h(n) for n > n3 = max(n1 , n2 ). 2.5 Reasoning About Eﬃciency Gross reasoning about an algorithm’s running time of is usually easy given a precise written description of the algorithm. In this section, I will work through several examples, perhaps in greater detail than necessary. 2.5.1 Selection Sort Here we analyze the selection sort algorithm, which repeatedly identiﬁes the smallest remaining unsorted element and puts it at the end of the sorted portion of the array. An animation of selection sort in action appears in Figure 2.5, and the code is shown below: 41 42 2. ALGORITHM ANALYSIS S C C C C C C C C C C C C C E E E E E E E E E E E E E E L L L E E E E E E E E E E E E C T I O N S O E S T I O N S O E S T I O N S O L S T I O N S O I S T L O N S O I L T S O N S O I L N S O T S O I L N O S T S O I L N O O T S S I L N O O R S S I L N O O R S S I L N O O R S S I L N O O R S S I L N O O R S S R R R R R R R R R T T T T T T T T T T T T T T T T T T T Figure 2.5: Animation of selection sort in action selection_sort(int s[], int n) { int i,j; int min; /* counters */ /* index of minimum */ for (i=0; i 0) && (s[j] < s[j-1])) { swap(&s[j],&s[j-1]); j = j-1; } } How often does the inner while loop iterate? This is tricky because there are two diﬀerent stopping conditions: one to prevent us from running oﬀ the bounds of the array (j > 0) and the other to mark when the element ﬁnds its proper place in sorted order (s[j] < s[j − 1]). Since worst-case analysis seeks an upper bound on the running time, we ignore the early termination and assume that this loop always goes around i times. In fact, we can assume it always goes around n times since i < n. Since the outer loop goes around n times, insertion sort must be a quadratic-time algorithm, i.e. O(n2 ). This crude “round it up” analysis always does the job, in that the Big Oh running time bound you get will always be correct. Occasionally, it might be too generous, meaning the actual worst case time might be of a lower order than implied by such analysis. Still, I strongly encourage this kind of reasoning as a basis for simple algorithm analysis. 2.5.3 String Pattern Matching Pattern matching is the most fundamental algorithmic operation on text strings. This algorithm implements the ﬁnd command available in any web browser or text editor: Problem: Substring Pattern Matching Input: A text string t and a pattern string p. Output: Does t contain the pattern p as a substring, and if so where? 43 44 2. ALGORITHM ANALYSIS a b a b b a a b b a a a b a b b a Figure 2.6: Searching for the substring abba in the text aababba. Perhaps you are interested ﬁnding where “Skiena” appears in a given news article (well, I would be interested in such a thing). This is an instance of string pattern matching with t as the news article and p=“Skiena.” There is a fairly straightforward algorithm for string pattern matching that considers the possibility that p may start at each possible position in t and then tests if this is so. int findmatch(char *p, char *t) { int i,j; int m, n; /* counters */ /* string lengths */ m = strlen(p); n = strlen(t); for (i=0; i<=(n-m); i=i+1) { j=0; while ((j g(n) for all n, but it does not dominate since lim g(n)/f (n) = lim n2 /2n2 = lim 1/2 = 0 n→∞ n→∞ n→∞ This is to be expected because both functions are in the class Θ(n2 ). What about f (n) = n3 and g(n) = n2 ? Since lim g(n)/f (n) = lim n2 /n3 = lim 1/n = 0 n→∞ n→∞ n→∞ the higher-degree polynomial dominates. This is true for any two polynomials, namely that na dominates nb if a > b since lim nb /na = lim nb−a → 0 n→∞ n→∞ 1.2 Thus n dominates n1.1999999 . Now consider two exponential functions, say f (n) = 3n and g(n) = 2n . Since lim g(n)/f (n) = 2n /3n = lim (2/3)n = 0 n→∞ n→∞ the exponential with the higher base dominates. Our ability to prove dominance relations from scratch depends upon our ability to prove limits. Let’s look at one important pair of functions. Any polynomial (say f (n) = n ) dominates logarithmic functions (say g(n) = lg n). Since n = 2lg n , f (n) = (2lg n ) = 2 lg n Now consider lim g(n)/f (n) = lg n/2 lg n n→∞ In fact, this does go to 0 as n → ∞. Take-Home Lesson: By interleaving the functions here with those of Section 2.3.1 (page 39), we see where everything ﬁts into the dominance pecking order: √ n3 n2 n1+ n log n n n n! cn log2 n log n log n/ log log n log log n α(n) 1 Chapter Notes Most other algorithm texts devote considerably more eﬀorts to the formal analysis of algorithms than we have here, and so we refer the theoretically-inclined reader 2.10 EXERCISES elsewhere for more depth. Algorithm texts more heavily stressing analysis include [CLRS01, KT06]. The book Concrete Mathematics by Knuth, Graham, and Patashnik [GKP89] oﬀers an interesting and thorough presentation of mathematics for the analysis of algorithms. Niven and Zuckerman [NZ80] is an nice introduction to number theory, including Waring’s problem, discussed in the war story. The notion of dominance also gives rise to the “Little Oh” notation. We say that f (n) = o(g(n)) iﬀ g(n) dominates f (n). Among other things, the Little Oh proves useful for asking questions. Asking for an o(n2 ) algorithm means you want one that is better than quadratic in the worst case—and means you would be willing to settle for O(n1.999 log2 n). 2.10 Exercises Program Analysis 2-1. [3] What value is returned by the following function? Express your answer as a function of n. Give the worst-case running time using the Big Oh notation. function mystery(n) r := 0 for i := 1 to n − 1 do for j := i + 1 to n do for k := 1 to j do r := r + 1 return(r) 2-2. [3] What value is returned by the following function? Express your answer as a function of n. Give the worst-case running time using Big Oh notation. function pesky(n) r := 0 for i := 1 to n do for j := 1 to i do for k := j to i + j do r := r + 1 return(r) 2-3. [5] What value is returned by the following function? Express your answer as a function of n. Give the worst-case running time using Big Oh notation. function prestiferous(n) r := 0 for i := 1 to n do for j := 1 to i do for k := j to i + j do for l := 1 to i + j − k do 57 58 2. ALGORITHM ANALYSIS r := r + 1 return(r) 2-4. [8] What value is returned by the following function? Express your answer as a function of n. Give the worst-case running time using Big Oh notation. function conundrum(n) r := 0 for i := 1 to n do for j := i + 1 to n do for k := i + j − 1 to n do r := r + 1 return(r) 2-5. [5] Suppose the following algorithm is used to evaluate the polynomial p(x) = an xn + an−1 xn−1 + . . . + a1 x + a0 p := a0 ; xpower := 1; for i := 1 to n do xpower := x ∗ xpower; p := p + ai ∗ xpower end (a) How many multiplications are done in the worst-case? How many additions? (b) How many multiplications are done on the average? (c) Can you improve this algorithm? 2-6. [3] Prove that the following algorithm for computing the maximum value in an array A[1..n] is correct. function max(A) m := A[1] for i := 2 to n do if A[i] > m then m := A[i] return (m) Big Oh 2-7. [3] True or False? (a) Is 2n+1 = O(2n )? (b) Is 22n = O(2n )? 2-8. [3] For each of the following pairs of functions, either f (n) is in O(g(n)), f (n) is in Ω(g(n)), or f (n) = Θ(g(n)). Determine which relationship is correct and brieﬂy explain why. (a) f (n) = log n2 ; g(n) = log n + 5 2.10 (b) f (n) = √ EXERCISES n; g(n) = log n2 (c) f (n) = log2 n; g(n) = log n (d) f (n) = n; g(n) = log2 n (e) f (n) = n log n + n; g(n) = log n (f) f (n) = 10; g(n) = log 10 (g) f (n) = 2n ; g(n) = 10n2 (h) f (n) = 2n ; g(n) = 3n 2-9. [3] For each of the following pairs of functions f (n) and g(n), determine whether f (n) = O(g(n)), g(n) = O(f (n)), or both. (a) f (n) = (n2 − n)/2, g(n) = 6n √ (b) f (n) = n + 2 n, g(n) = n2 √ (c) f (n) = n log n, g(n) = n n/2 √ (d) f (n) = n + log n, g(n) = n (e) f (n) = 2(log n)2 , g(n) = log n + 1 (f) f (n) = 4n log n + n, g(n) = (n2 − n)/2 2-10. [3] Prove that n3 − 3n2 − n + 1 = Θ(n3 ). 2-11. [3] Prove that n2 = O(2n ). 2-12. [3] For each of the following pairs of functions f (n) and g(n), give an appropriate positive constant c such that f (n) ≤ c · g(n) for all n > 1. (a) f (n) = n2 + n + 1, g(n) = 2n3 √ (b) f (n) = n n + n2 , g(n) = n2 (c) f (n) = n2 − n + 1, g(n) = n2 /2 2-13. [3] Prove that if f1 (n) = O(g1 (n)) and f2 (n) = O(g2 (n)), then f1 (n) + f2 (n) = O(g1 (n) + g2 (n)). 2-14. [3] Prove that if f1 (N ) = Ω(g1 (n)) and f2 (n) = Ω(g2 (n)), then f1 (n) + f2 (n) = Ω(g1 (n) + g2 (n)). 2-15. [3] Prove that if f1 (n) = O(g1 (n)) and f2 (n) = O(g2 (n)), then f1 (n) · f2 (n) = O(g1 (n) · g2 (n)) 2-16. [5] Prove for all k ≥ 1 and all sets of constants {ak , ak−1 , . . . , a1 , a0 } ∈ R, ak nk + ak−1 nk−1 + .... + a1 n + a0 = O(nk ) 2-17. [5] Show that for any real constants a and b, b > 0 (n + a)b = Θ(nb ) 2-18. [5] List the functions below from the lowest to the highest order. If any two or more are of the same order, indicate which. 59 60 2. ALGORITHM ANALYSIS n n − n3 + 7n5 n2 + lg n n3 2n lg n n2 (lg n)2 n √lg n n 2n−1 n! ln n en lg lg n n1+ε where 0 < ε < 1 2-19. [5] List the functions below from the lowest to the more are of the same order, indicate which. √ n n n log n n − n3 + 7n5 n3 n2 1 n 3 + log n (log n)2 n ln n log n n (3/2)n (1/3) highest order. If any two or 2n n2 + log n log n n! log log n 6 2-20. [5] Find two functions f (n) and g(n) that satisfy the following relationship. If no such f and g exist, write “None.” (a) f (n) = o(g(n)) and f (n) = Θ(g(n)) (b) f (n) = Θ(g(n)) and f (n) = o(g(n)) (c) f (n) = Θ(g(n)) and f (n) = O(g(n)) (d) f (n) = Ω(g(n)) and f (n) = O(g(n)) 2-21. [5] True or False? (a) 2n2 + 1 = O(n2 ) √ (b) n = O(log n) √ (c) log n = O( n) √ (d) n2 (1 + n) = O(n2 log n) √ (e) 3n2 + n = O(n2 ) √ (f) n log n = O(n) (g) log n = O(n−1/2 ) 2-22. [5] For each of the following pairs of functions f (n) and g(n), state whether f (n) = O(g(n)), f (n) = Ω(g(n)), f (n) = Θ(g(n)), or none of the above. (a) f (n) = n2 + 3n + 4, g(n) = 6n + 7 √ (b) f (n) = n n, g(n) = n2 − n (c) f (n) = 2n − n2 , g(n) = n4 + n2 2-23. [3] For each of these questions, brieﬂy explain your answer. (a) If I prove that an algorithm takes O(n2 ) worst-case time, is it possible that it takes O(n) on some inputs? (b) If I prove that an algorithm takes O(n2 ) worst-case time, is it possible that it takes O(n) on all inputs? (c) If I prove that an algorithm takes Θ(n2 ) worst-case time, is it possible that it takes O(n) on some inputs? 2.10 EXERCISES (d) If I prove that an algorithm takes Θ(n2 ) worst-case time, is it possible that it takes O(n) on all inputs? (e) Is the function f (n) = Θ(n2 ), where f (n) = 100n2 for even n and f (n) = 20n2 − n log2 n for odd n? 2-24. [3] For each of the following, answer yes, no, or can’t tell. Explain your reasoning. (a) Is 3n = O(2n )? (b) Is log 3n = O(log 2n )? (c) Is 3n = Ω(2n )? (d) Is log 3n = Ω(log 2n )? 2-25. [5] For each of the following expressions f (n) ﬁnd a simple g(n) such that f (n) = Θ(g(n)). (a) f (n) = (b) f (n) = (c) f (n) = n 1 . i 1 . i=1 i n log i. i=1 i=1 n (d) f (n) = log(n!). 2-26. [5] Place the following functions into increasing asymptotic order. n n f1 (n) = n2 log2 n, f2 (n) = n(log2 n)2 , f3 (n) = i=0 2i , f4 (n) = log2 ( i=0 2i ). 2-27. [5] Place the following functions into increasing asymptotic order. If two or more of the functions are of the same asymptotic order then indicate this. n √ √ √ 3 f1 (n) = i=1 i, f2 (n) = ( n) log n, f3 (n) = n log n, f4 (n) = 12n 2 + 4n, 2-28. [5] For each of the following expressions f (n) ﬁnd a simple g(n) such that f (n) = Θ(g(n)). (You should be able to prove your result by exhibiting the relevant parameters, but this is not required for the homework.) (a) f (n) = (b) f (n) = (c) f (n) = n i=1 3i4 + 2i3 − 19i + 20. i=1 3(4i ) + 2(3i ) − i19 + 20. i=1 5i + 32i . n n 2-29. [5] Which of the following are true? (a) (b) (c) n i=1 3i = Θ(3n−1 ). i=1 3i = Θ(3n ). i=1 3i = Θ(3n+1 ). n n 2-30. [5] For each of the following functions f ﬁnd a simple function g such that f (n) = Θ(g(n)). (a) f1 (n) = (1000)2n + 4n . √ (b) f2 (n) = n + n log n + n. (c) f3 (n) = log(n20 ) + (log n)10 . (d) f4 (n) = (0.99)n + n100 . 61 62 2. ALGORITHM ANALYSIS 2-31. [5] For each pair of expressions (A, B) below, indicate whether A is O, o, Ω, ω, or Θ of B. Note that zero, one or more of these relations may hold for a given pair; list all correct ones. A B 2n (a) n100 √ (b) (lg n)12 n √ cos(πn/8) (c) n n (d) 10n 100n lg n (e) n (lg n)n (f ) lg (n!) n lg n Summations 2-32. [5] Prove that: 12 − 22 + 32 − 42 + . . . + (−1)k−1 k2 = (−1)k−1 k(k + 1)/2 2-33. [5] Find an expression for the sum of the ith row of the following triangle, and prove its correctness. Each entry is the sum of the three entries directly above it. All non existing entries are considered 0. 1 1 4 1 3 10 1 2 6 16 1 1 3 7 19 1 2 6 16 1 3 10 1 4 1 2-34. [3] Assume that Christmas has n days. Exactly how many presents did my “true love” send me? (Do some research if you do not understand this question.) 2-35. [5] Consider the following code fragment. for i=1 to n do for j=i to 2*i do output ‘‘foobar’’ Let T (n) denote the number of times ‘foobar’ is printed as a function of n. a. Express T (n) as a summation (actually two nested summations). b. Simplify the summation. Show your work. 2-36. [5] Consider the following code fragment. for i=1 to n/2 do for j=i to n-i do for k=1 to j do output ‘‘foobar’’ Assume n is even. Let T (n) denote the number of times ‘foobar’ is printed as a function of n. (a) Express T (n) as three nested summations. (b) Simplify the summation. Show your work. 2.10 EXERCISES 2-37. [6] When you ﬁrst learned to multiply numbers, you were told that x × y means adding x a total of y times, so 5×4 = 5+5+5+5 = 20. What is the time complexity of multiplying two n-digit numbers in base b (people work in base 10, of course, while computers work in base 2) using the repeated addition method, as a function of n and b. Assume that single-digit by single-digit addition or multiplication takes O(1) time. (Hint: how big can y be as a function of n and b?) 2-38. [6] In grade school, you learned to multiply long numbers on a digit-by-digit basis, so that 127 × 211 = 127 × 1 + 127 × 10 + 127 × 200 = 26, 397. Analyze the time complexity of multiplying two n-digit numbers with this method as a function of n (assume constant base size). Assume that single-digit by single-digit addition or multiplication takes O(1) time. Logarithms 2-39. [5] Prove the following identities on logarithms: (a) Prove that loga (xy) = loga x + loga y (b) Prove that loga xy = y loga x (c) Prove that loga x = logb x logb a (d) Prove that xlogb y = y logb x 2-40. [3] Show that lg(n + 1) = lg n + 1 2-41. [3] Prove that that the binary representation of n ≥ 1 has lg2 n + 1 bits. 2-42. [5] In one of my√research papers I give a comparison-based sorting algorithm that runs in O(n log( n)). Given the existence of an Ω(n log n) lower bound for sorting, how can this be possible? Interview Problems 2-43. [5] You are given a set S of n numbers. You must pick a subset S of k numbers from S such that the probability of each element of S occurring in S is equal (i.e., each is selected with probability k/n). You may make only one pass over the numbers. What if n is unknown? 2-44. [5] We have 1,000 data items to store on 1,000 nodes. Each node can store copies of exactly three diﬀerent items. Propose a replication scheme to minimize data loss as nodes fail. What is the expected number of data entries that get lost when three random nodes fail? 2-45. [5] Consider the following algorithm to ﬁnd the minimum element in an array of numbers A[0, . . . , n]. One extra variable tmp is allocated to hold the current minimum value. Start from A[0]; ”tmp” is compared against A[1], A[2], . . . , A[N ] in order. When A[i] < tmp, tmp = A[i]. What is the expected number of times that the assignment operation tmp = A[i] is performed? 2-46. [5] You have a 100-story building and a couple of marbles. You must identify the lowest ﬂoor for which a marble will break if you drop it from this ﬂoor. How fast can you ﬁnd this ﬂoor if you are given an inﬁnite supply of marbles? What if you have only two marbles? 63 64 2. ALGORITHM ANALYSIS 2-47. [5] You are given 10 bags of gold coins. Nine bags contain coins that each weigh 10 grams. One bag contains all false coins that weigh one gram less. You must identify this bag in just one weighing. You have a digital balance that reports the weight of what is placed on it. 2-48. [5] You have eight balls all of the same size. Seven of them weigh the same, and one of them weighs slightly more. How can you ﬁnd the ball that is heavier by using a balance and only two weighings? 2-49. [5] Suppose we start with n companies that eventually merge into one big company. How many diﬀerent ways are there for them to merge? 2-50. [5] A Ramanujam number can be written two diﬀerent ways as the sum of two cubes—i.e. , there exist distinct a, b, c, and d such that a3 + b3 = c3 + d3 . Generate all Ramanujam numbers where a, b, c, d < n. 2-51. [7] Six pirates must divide $300 dollars among themselves. The division is to proceed as follows. The senior pirate proposes a way to divide the money. Then the pirates vote. If the senior pirate gets at least half the votes he wins, and that division remains. If he doesn’t, he is killed and then the next senior-most pirate gets a chance to do the division. Now you have to tell what will happen and why (i.e. , how many pirates survive and how the division is done)? All the pirates are intelligent and the ﬁrst priority is to stay alive and the next priority is to get as much money as possible. 2-52. [7] Reconsider the pirate problem above, where only one indivisible dollar is to be divided. Who gets the dollar and how many are killed? Programming Challenges These programming challenge problems with robot judging are available at http://www.programming-challenges.com or http://online-judge.uva.es. 2-1. “Primary Arithmetic” – Programming Challenges 110501, UVA Judge 10035. 2-2. “A Multiplication Game” – Programming Challenges 110505, UVA Judge 847. 2-3. “Light, More Light” – Programming Challenges 110701, UVA Judge 10110. 3 Data Structures Changing a data structure in a slow program can work the same way an organ transplant does in a sick patient. Important classes of abstract data types such as containers, dictionaries, and priority queues, have many diﬀerent but functionally equivalent data structures that implement them. Changing the data structure does not change the correctness of the program, since we presumably replace a correct implementation with a diﬀerent correct implementation. However, the new implementation of the data type realizes diﬀerent tradeoﬀs in the time to execute various operations, so the total performance can improve dramatically. Like a patient in need of a transplant, only one part might need to be replaced in order to ﬁx the problem. But it is better to be born with a good heart than have to wait for a replacement. The maximum beneﬁt from good data structures results from designing your program around them in the ﬁrst place. We assume that the reader has had some previous exposure to elementary data structures and pointer manipulation. Still, data structure (CS II) courses these days focus more on data abstraction and object orientation than the nitty-gritty of how structures should be represented in memory. We will review this material to make sure you have it down. In data structures, as with most subjects, it is more important to really understand the basic material than have exposure to more advanced concepts. We will focus on each of the three fundamental abstract data types (containers, dictionaries, and priority queues) and see how they can be implemented with arrays and lists. Detailed discussion of the tradeoﬀs between more sophisticated implementations is deferred to the relevant catalog entry for each of these data types. S.S. Skiena, The Algorithm Design Manual, 2nd ed., DOI: 10.1007/978-1-84800-070-4 3, c Springer-Verlag London Limited 2008 66 3. DATA STRUCTURES 3.1 Contiguous vs. Linked Data Structures Data structures can be neatly classiﬁed as either contiguous or linked, depending upon whether they are based on arrays or pointers: • Contiguously-allocated structures are composed of single slabs of memory, and include arrays, matrices, heaps, and hash tables. • Linked data structures are composed of distinct chunks of memory bound together by pointers, and include lists, trees, and graph adjacency lists. In this section, we review the relative advantages of contiguous and linked data structures. These tradeoﬀs are more subtle than they appear at ﬁrst glance, so I encourage readers to stick with me here even if you may be familiar with both types of structures. 3.1.1 Arrays The array is the fundamental contiguously-allocated data structure. Arrays are structures of ﬁxed-size data records such that each element can be eﬃciently located by its index or (equivalently) address. A good analogy likens an array to a street full of houses, where each array element is equivalent to a house, and the index is equivalent to the house number. Assuming all the houses are equal size and numbered sequentially from 1 to n, we can compute the exact position of each house immediately from its address.1 Advantages of contiguously-allocated arrays include: • Constant-time access given the index – Because the index of each element maps directly to a particular memory address, we can access arbitrary data items instantly provided we know the index. • Space eﬃciency – Arrays consist purely of data, so no space is wasted with links or other formatting information. Further, end-of-record information is not needed because arrays are built from ﬁxed-size records. • Memory locality – A common programming idiom involves iterating through all the elements of a data structure. Arrays are good for this because they exhibit excellent memory locality. Physical continuity between successive data accesses helps exploit the high-speed cache memory on modern computer architectures. The downside of arrays is that we cannot adjust their size in the middle of a program’s execution. Our program will fail soon as we try to add the (n + 1 Houses in Japanese cities are traditionally numbered in the order they were built, not by their physical location. This makes it extremely diﬃcult to locate a Japanese address without a detailed map. 3.1 CONTIGUOUS VS. LINKED DATA STRUCTURES 1)st customer, if we only allocate room for n records. We can compensate by allocating extremely large arrays, but this can waste space, again restricting what our programs can do. Actually, we can eﬃciently enlarge arrays as we need them, through the miracle of dynamic arrays. Suppose we start with an array of size 1, and double its size from m to 2m each time we run out of space. This doubling process involves allocating a new contiguous array of size 2m, copying the contents of the old array to the lower half of the new one, and returning the space used by the old array to the storage allocation system. The apparent waste in this procedure involves the recopying of the old contents on each expansion. How many times might an element have to be recopied after a total of n insertions? Well, the ﬁrst inserted element will have been recopied when the array expands after the ﬁrst, second, fourth, eighth, . . . insertions. It will take log2 n doublings until the array gets to have n positions. However, most elements do not suﬀer much upheaval. Indeed, the (n/2 + 1)st through nth elements will move at most once and might never have to move at all. If half the elements move once, a quarter of the elements twice, and so on, the total number of movements M is given by M= lg n i=1 i · n/2i = n lg n i=1 i/2i ≤ n ∞ i/2i = 2n i=1 Thus, each of the n elements move only two times on average, and the total work of managing the dynamic array is the same O(n) as it would have been if a single array of suﬃcient size had been allocated in advance! The primary thing lost using dynamic arrays is the guarantee that each array access takes constant time in the worst case. Now all the queries will be fast, except for those relatively few queries triggering array doubling. What we get instead is a promise that the nth array access will be completed quickly enough that the total eﬀort expended so far will still be O(n). Such amortized guarantees arise frequently in the analysis of data structures. 3.1.2 Pointers and Linked Structures Pointers are the connections that hold the pieces of linked structures together. Pointers represent the address of a location in memory. A variable storing a pointer to a given data item can provide more freedom than storing a copy of the item itself. A cell-phone number can be thought of as a pointer to its owner as they move about the planet. Pointer syntax and power diﬀer signiﬁcantly across programming languages, so we begin with a quick review of pointers in C language. A pointer p is assumed to 67 68 3. DATA STRUCTURES Lincoln Jefferson Clinton NIL Figure 3.1: Linked list example showing data and pointer ﬁelds give the address in memory where a particular chunk of data is located.2 Pointers in C have types declared at compiler time, denoting the data type of the items they can point to. We use *p to denote the item that is pointed to by pointer p, and &x to denote the address (i.e. , pointer) of a particular variable x. A special NULL pointer value is used to denote structure-terminating or unassigned pointers. All linked data structures share certain properties, as revealed by the following linked list type declaration: typedef struct list { item_type item; struct list *next; } list; /* data item */ /* point to successor */ In particular: • Each node in our data structure (here list) contains one or more data ﬁelds (here item) that retain the data that we need to store. • Each node contains a pointer ﬁeld to at least one other node (here next). This means that much of the space used in linked data structures has to be devoted to pointers, not data. • Finally, we need a pointer to the head of the structure, so we know where to access it. The list is the simplest linked structure. The three basic operations supported by lists are searching, insertion, and deletion. In doubly-linked lists, each node points both to its predecessor and its successor element. This simpliﬁes certain operations at a cost of an extra pointer ﬁeld per node. Searching a List Searching for item x in a linked list can be done iteratively or recursively. We opt for recursively in the implementation below. If x is in the list, it is either the ﬁrst element or located in the smaller rest of the list. Eventually, we reduce the problem to searching in an empty list, which clearly cannot contain x. 2 C permits direct manipulation of memory addresses in ways which may horrify Java programmers, but we will avoid doing any such tricks. 3.1 CONTIGUOUS VS. LINKED DATA STRUCTURES list *search_list(list *l, item_type x) { if (l == NULL) return(NULL); if (l->item == x) return(l); else return( search_list(l->next, x) ); } Insertion into a List Insertion into a singly-linked list is a nice exercise in pointer manipulation, as shown below. Since we have no need to maintain the list in any particular order, we might as well insert each new item in the simplest place. Insertion at the beginning of the list avoids any need to traverse the list, but does require us to update the pointer (denoted l) to the head of the data structure. void insert_list(list **l, item_type x) { list *p; /* temporary pointer */ p = malloc( sizeof(list) ); p->item = x; p->next = *l; *l = p; } Two C-isms to note. First, the malloc function allocates a chunk of memory of suﬃcient size for a new node to contain x. Second, the funny double star (**l) denotes that l is a pointer to a pointer to a list node. Thus the last line, *l=p; copies p to the place pointed to l, which is the external variable maintaining access to the head of the list. Deletion From a List Deletion from a linked list is somewhat more complicated. First, we must ﬁnd a pointer to the predecessor of the item to be deleted. We do this recursively: 69 70 3. DATA STRUCTURES list *predecessor_list(list *l, item_type x) { if ((l == NULL) || (l->next == NULL)) { printf("Error: predecessor sought on null list.\n"); return(NULL); } if ((l->next)->item == x) return(l); else return( predecessor_list(l->next, x) ); } The predecessor is needed because it points to the doomed node, so its next pointer must be changed. The actual deletion operation is simple, once ruling out the case that the to-be-deleted element does not exist. Special care must be taken to reset the pointer to the head of the list (l) when the ﬁrst element is deleted: delete_list(list **l, item_type x) { list *p; /* item pointer */ list *pred; /* predecessor pointer */ list *search_list(), *predecessor_list(); p = search_list(*l,x); if (p != NULL) { pred = predecessor_list(*l,x); if (pred == NULL) /* splice out out list */ *l = p->next; else pred->next = p->next; free(p); /* free memory used by node */ } } C language requires explicit deallocation of memory, so we must free the deleted node after we are ﬁnished with it to return the memory to the system. 3.1.3 Comparison The relative advantages of linked lists over static arrays include: • Overﬂow on linked structures can never occur unless the memory is actually full. 3.2 STACKS AND QUEUES • Insertions and deletions are simpler than for contiguous (array) lists. • With large records, moving pointers is easier and faster than moving the items themselves. while the relative advantages of arrays include: • Linked structures require extra space for storing pointer ﬁelds. • Linked lists do not allow eﬃcient random access to items. • Arrays allow better memory locality and cache performance than random pointer jumping. Take-Home Lesson: Dynamic memory allocation provides us with ﬂexibility on how and where we use our limited storage resources. One ﬁnal thought about these fundamental structures is that they can be thought of as recursive objects: • Lists – Chopping the ﬁrst element oﬀ a linked list leaves a smaller linked list. This same argument works for strings, since removing characters from string leaves a string. Lists are recursive objects. • Arrays – Splitting the ﬁrst k elements oﬀ of an n element array gives two smaller arrays, of size k and n − k, respectively. Arrays are recursive objects. This insight leads to simpler list processing, and eﬃcient divide-and-conquer algorithms such as quicksort and binary search. 3.2 Stacks and Queues We use the term container to denote a data structure that permits storage and retrieval of data items independent of content. By contrast, dictionaries are abstract data types that retrieve based on key values or content, and will be discussed in Section 3.3 (page 72). Containers are distinguished by the particular retrieval order they support. In the two most important types of containers, this retrieval order depends on the insertion order: • Stacks – Support retrieval by last-in, ﬁrst-out (LIFO) order. Stacks are simple to implement and very eﬃcient. For this reason, stacks are probably the right container to use when retrieval order doesn’t matter at all, such as when processing batch jobs. The put and get operations for stacks are usually called push and pop: 71 72 3. DATA STRUCTURES – Push(x,s): Insert item x at the top of stack s. – Pop(s): Return (and remove) the top item of stack s. LIFO order arises in many real-world contexts. People crammed into a subway car exit in LIFO order. Food inserted into my refrigerator usually exits the same way, despite the incentive of expiration dates. Algorithmically, LIFO tends to happen in the course of executing recursive algorithms. • Queues – Support retrieval in ﬁrst in, ﬁrst out (FIFO) order. This is surely the fairest way to control waiting times for services. You want the container holding jobs to be processed in FIFO order to minimize the maximum time spent waiting. Note that the average waiting time will be the same regardless of whether FIFO or LIFO is used. Many computing applications involve data items with inﬁnite patience, which renders the question of maximum waiting time moot. Queues are somewhat trickier to implement than stacks and thus are most appropriate for applications (like certain simulations) where the order is important. The put and get operations for queues are usually called enqueue and dequeue. – Enqueue(x,q): Insert item x at the back of queue q. – Dequeue(q): Return (and remove) the front item from queue q. We will see queues later as the fundamental data structure controlling breadth-ﬁrst searches in graphs. Stacks and queues can be eﬀectively implemented using either arrays or linked lists. The key issue is whether an upper bound on the size of the container is known in advance, thus permitting the use of a statically-allocated array. 3.3 Dictionaries The dictionary data type permits access to data items by content. You stick an item into a dictionary so you can ﬁnd it when you need it. The primary operations of dictionary support are: • Search(D,k) – Given a search key k, return a pointer to the element in dictionary D whose key value is k, if one exists. • Insert(D,x) – Given a data item x, add it to the set in the dictionary D. • Delete(D,x) – Given a pointer to a given data item x in the dictionary D, remove it from D. 3.3 DICTIONARIES Certain dictionary data structures also eﬃciently support other useful operations: • Max(D) or Min(D) – Retrieve the item with the largest (or smallest) key from D. This enables the dictionary to serve as a priority queue, to be discussed in Section 3.5 (page 83). • Predecessor(D,k) or Successor(D,k) – Retrieve the item from D whose key is immediately before (or after) k in sorted order. These enable us to iterate through the elements of the data structure. Many common data processing tasks can be handled using these dictionary operations. For example, suppose we want to remove all duplicate names from a mailing list, and print the results in sorted order. Initialize an empty dictionary D, whose search key will be the record name. Now read through the mailing list, and for each record search to see if the name is already in D. If not, insert it into D. Once ﬁnished, we must extract the remaining names out of the dictionary. By starting from the ﬁrst item Min(D) and repeatedly calling Successor until we obtain Max(D), we traverse all elements in sorted order. By deﬁning such problems in terms of abstract dictionary operations, we avoid the details of the data structure’s representation and focus on the task at hand. In the rest of this section, we will carefully investigate simple dictionary implementations based on arrays and linked lists. More powerful dictionary implementations such as binary search trees (see Section 3.4 (page 77)) and hash tables (see Section 3.7 (page 89)) are also attractive options in practice. A complete discussion of diﬀerent dictionary data structures is presented in the catalog in Section 12.1 (page 367). We encourage the reader to browse through the data structures section of the catalog to better learn what your options are. Stop and Think: Comparing Dictionary Implementations (I) Problem: What are the asymptotic worst-case running times for each of the seven fundamental dictionary operations (search, insert, delete, successor, predecessor, minimum, and maximum) when the data structure is implemented as: • An unsorted array. • A sorted array. Solution: This problem (and the one following it) reveal some of the inherent tradeoﬀs of data structure design. A given data representation may permit eﬃcient implementation of certain operations at the cost that other operations are expensive. 73 74 3. DATA STRUCTURES In addition to the array in question, we will assume access to a few extra variables such as n—the number of elements currently in the array. Note that we must maintain the value of these variables in the operations where they change (e.g., insert and delete), and charge these operations the cost of this maintenance. The basic dictionary operations can be implemented with the following costs on unsorted and sorted arrays, respectively: Dictionary operation Search(L, k) Insert(L, x) Delete(L, x) Successor(L, x) Predecessor(L, x) Minimum(L) Maximum(L) Unsorted array O(n) O(1) O(1)∗ O(n) O(n) O(n) O(n) Sorted array O(log n) O(n) O(n) O(1) O(1) O(1) O(1) We must understand the implementation of each operation to see why. First, we discuss the operations when maintaining an unsorted array A. • Search is implemented by testing the search key k against (potentially) each element of an unsorted array. Thus, search takes linear time in the worst case, which is when key k is not found in A. • Insertion is implemented by incrementing n and then copying item x to the nth cell in the array, A[n]. The bulk of the array is untouched, so this operation takes constant time. • Deletion is somewhat trickier, hence the superscript(∗ ) in the table. The deﬁnition states that we are given a pointer x to the element to delete, so we need not spend any time searching for the element. But removing the xth element from the array A leaves a hole that must be ﬁlled. We could ﬁll the hole by moving each of the elements A[x + 1] to A[n] up one position, but this requires Θ(n) time when the ﬁrst element is deleted. The following idea is better: just write over A[x] with A[n], and decrement n. This only takes constant time. • The deﬁnition of the traversal operations, Predecessor and Successor, refer to the item appearing before/after x in sorted order. Thus, the answer is not simply A[x − 1] (or A[x + 1]), because in an unsorted array an element’s physical predecessor (successor) is not necessarily its logical predecessor (successor). Instead, the predecessor of A[x] is the biggest element smaller than A[x]. Similarly, the successor of A[x] is the smallest element larger than A[x]. Both require a sweep through all n elements of A to determine the winner. • Minimum and Maximum are similarly deﬁned with respect to sorted order, and so require linear sweeps to identify in an unsorted array. 3.3 DICTIONARIES Implementing a dictionary using a sorted array completely reverses our notions of what is easy and what is hard. Searches can now be done in O(log n) time, using binary search, because we know the median element sits in A[n/2]. Since the upper and lower portions of the array are also sorted, the search can continue recursively on the appropriate portion. The number of halvings of n until we get to a single element is lg n. The sorted order also beneﬁts us with respect to the other dictionary retrieval operations. The minimum and maximum elements sit in A[1] and A[n], while the predecessor and successor to A[x] are A[x − 1] and A[x + 1], respectively. Insertion and deletion become more expensive, however, because making room for a new item or ﬁlling a hole may require moving many items arbitrarily. Thus both become linear-time operations. Take-Home Lesson: Data structure design must balance all the diﬀerent operations it supports. The fastest data structure to support both operations A and B may well not be the fastest structure to support either operation A or B. Stop and Think: Comparing Dictionary Implementations (II) Problem: What is the asymptotic worst-case running times for each of the seven fundamental dictionary operations when the data structure is implemented as • A singly-linked unsorted list. • A doubly-linked unsorted list. • A singly-linked sorted list. • A doubly-linked sorted list. Solution: Two diﬀerent issues must be considered in evaluating these implementations: singly- vs. doubly-linked lists and sorted vs. unsorted order. Subtle operations are denoted with a superscript: Dictionary operation Search(L, k) Insert(L, x) Delete(L, x) Successor(L, x) Predecessor(L, x) Minimum(L) Maximum(L) Singly unsorted O(n) O(1) O(n)∗ O(n) O(n) O(n) O(n) Double unsorted O(n) O(1) O(1) O(n) O(n) O(n) O(n) Singly sorted O(n) O(n) O(n)∗ O(1) O(n)∗ O(1) O(1)∗ Doubly sorted O(n) O(n) O(1) O(1) O(1) O(1) O(1) 75 76 3. DATA STRUCTURES As with unsorted arrays, search operations are destined to be slow while maintenance operations are fast. • Insertion/Deletion – The complication here is deletion from a singly-linked list. The deﬁnition of the Delete operation states we are given a pointer x to the item to be deleted. But what we really need is a pointer to the element pointing to x in the list, because that is the node that needs to be changed. We can do nothing without this list predecessor, and so must spend linear time searching for it on a singly-linked list. Doubly-linked lists avoid this problem, since we can immediately retrieve the list predecessor of x. Deletion is faster for sorted doubly-linked lists than sorted arrays, because splicing out the deleted element from the list is more eﬃcient than ﬁlling the hole by moving array elements. The predecessor pointer problem again complicates deletion from singly-linked sorted lists. • Search – Sorting provides less beneﬁt for linked lists than it did for arrays. Binary search is no longer possible, because we can’t access the median element without traversing all the elements before it. What sorted lists do provide is quick termination of unsuccessful searches, for if we have not found Abbott by the time we hit Costello we can deduce that he doesn’t exist. Still, searching takes linear time in the worst case. • Traversal operations – The predecessor pointer problem again complicates implementing Predecessor. The logical successor is equivalent to the node successor for both types of sorted lists, and hence can be implemented in constant time. • Maximum – The maximum element sits at the tail of the list, which would normally require Θ(n) time to reach in either singly- or doubly-linked lists. However, we can maintain a separate pointer to the list tail, provided we pay the maintenance costs for this pointer on every insertion and deletion. The tail pointer can be updated in constant time on doubly-linked lists: on insertion check whether last->next still equals NULL, and on deletion set last to point to the list predecessor of last if the last element is deleted. We have no eﬃcient way to ﬁnd this predecessor for singly-linked lists. So why can we implement maximum in Θ(1) on singly-linked lists? The trick is to charge the cost to each deletion, which already took linear time. Adding an extra linear sweep to update the pointer does not harm the asymptotic complexity of Delete, while gaining us Maximum in constant time as a reward for clear thinking. 3.4 3 2 3 1 2 1 1 1 3 1 BINARY SEARCH TREES 2 3 2 2 3 Figure 3.2: The ﬁve distinct binary search trees on three nodes 3.4 Binary Search Trees We have seen data structures that allow fast search or ﬂexible update, but not fast search and ﬂexible update. Unsorted, doubly-linked lists supported insertion and deletion in O(1) time but search took linear time in the worse case. Sorted arrays support binary search and logarithmic query times, but at the cost of linear-time update. Binary search requires that we have fast access to two elements—speciﬁcally the median elements above and below the given node. To combine these ideas, we need a “linked list” with two pointers per node. This is the basic idea behind binary search trees. A rooted binary tree is recursively deﬁned as either being (1) empty, or (2) consisting of a node called the root, together with two rooted binary trees called the left and right subtrees, respectively. The order among “brother” nodes matters in rooted trees, so left is diﬀerent from right. Figure 3.2 gives the shapes of the ﬁve distinct binary trees that can be formed on three nodes. A binary search tree labels each node in a binary tree with a single key such that for any node labeled x, all nodes in the left subtree of x have keys < x while all nodes in the right subtree of x have keys > x. This search tree labeling scheme is very special. For any binary tree on n nodes, and any set of n keys, there is exactly one labeling that makes it a binary search tree. The allowable labelings for three-node trees are given in Figure 3.2. 3.4.1 Implementing Binary Search Trees Binary tree nodes have left and right pointer ﬁelds, an (optional) parent pointer, and a data ﬁeld. These relationships are shown in Figure 3.3; a type declaration for the tree structure is given below: 77 78 3. DATA STRUCTURES parent right left Figure 3.3: Relationships in a binary search tree (left). Finding the minimum (center) and maximum (right) elements in a binary search tree typedef struct tree { item_type item; struct tree *parent; struct tree *left; struct tree *right; } tree; /* /* /* /* data item */ pointer to parent */ pointer to left child */ pointer to right child */ The basic operations supported by binary trees are searching, traversal, insertion, and deletion. Searching in a Tree The binary search tree labeling uniquely identities where each key is located. Start at the root. Unless it contains the query key x, proceed either left or right depending upon whether x occurs before or after the root key. This algorithm works because both the left and right subtrees of a binary search tree are themselves binary search trees. This recursive structure yields the recursive search algorithm below: tree *search_tree(tree *l, item_type x) { if (l == NULL) return(NULL); if (l->item == x) return(l); if (x < l->item) return( search_tree(l->left, x) ); else return( search_tree(l->right, x) ); } 3.4 BINARY SEARCH TREES This search algorithm runs in O(h) time, where h denotes the height of the tree. Finding Minimum and Maximum Elements in a Tree Implementing the ﬁnd-minimum operation requires knowing where the minimum element is in the tree. By deﬁnition, the smallest key must reside in the left subtree of the root, since all keys in the left subtree have values less than that of the root. Therefore, as shown in Figure 3.3, the minimum element must be the leftmost descendent of the root. Similarly, the maximum element must be the rightmost descendent of the root. tree *find_minimum(tree *t) { tree *min; /* pointer to minimum */ if (t == NULL) return(NULL); min = t; while (min->left != NULL) min = min->left; return(min); } Traversal in a Tree Visiting all the nodes in a rooted binary tree proves to be an important component of many algorithms. It is a special case of traversing all the nodes and edges in a graph, which will be the foundation of Chapter 5. A prime application of tree traversal is listing the labels of the tree nodes. Binary search trees make it easy to report the labels in sorted order. By deﬁnition, all the keys smaller than the root must lie in the left subtree of the root, and all keys bigger than the root in the right subtree. Thus, visiting the nodes recursively in accord with such a policy produces an in-order traversal of the search tree: void traverse_tree(tree *l) { if (l != NULL) { traverse_tree(l->left); process_item(l->item); traverse_tree(l->right); } } 79 80 3. DATA STRUCTURES Each item is processed once during the course of traversal, which runs in O(n) time, where n denotes the number of nodes in the tree. Alternate traversal orders come from changing the position of process item relative to the traversals of the left and right subtrees. Processing the item ﬁrst yields a pre-order traversal, while processing it last gives a post-order traversal. These make relatively little sense with search trees, but prove useful when the rooted tree represents arithmetic or logical expressions. Insertion in a Tree There is only one place to insert an item x into a binary search tree T where we know we can ﬁnd it again. We must replace the NULL pointer found in T after an unsuccessful query for the key k. This implementation uses recursion to combine the search and node insertion stages of key insertion. The three arguments to insert tree are (1) a pointer l to the pointer linking the search subtree to the rest of the tree, (2) the key x to be inserted, and (3) a parent pointer to the parent node containing l. The node is allocated and linked in on hitting the NULL pointer. Note that we pass the pointer to the appropriate left/right pointer in the node during the search, so the assignment *l = p; links the new node into the tree: insert_tree(tree **l, item_type x, tree *parent) { tree *p; /* temporary pointer */ if (*l == NULL) { p = malloc(sizeof(tree)); /* allocate new node */ p->item = x; p->left = p->right = NULL; p->parent = parent; *l = p; /* link into parent’s record */ return; } if (x < (*l)->item) insert_tree(&((*l)->left), x, *l); else insert_tree(&((*l)->right), x, *l); } Allocating the node and linking it in to the tree is a constant-time operation after the search has been performed in O(h) time. 3.4 2 2 1 1 7 4 3 8 2 7 4 1 8 6 6 BINARY SEARCH TREES 2 7 4 3 1 8 5 7 5 3 8 6 5 5 initial tree delete node with zero children (3) delete node with 1 child (6) delete node with 2 children (4) Figure 3.4: Deleting tree nodes with 0, 1, and 2 children Deletion from a Tree Deletion is somewhat trickier than insertion, because removing a node means appropriately linking its two descendant subtrees back into the tree somewhere else. There are three cases, illustrated in Figure 3.4. Leaf nodes have no children, and so may be deleted by simply clearing the pointer to the given node. The case of the doomed node having one child is also straightforward. There is one parent and one grandchild, and we can link the grandchild directly to the parent without violating the in-order labeling property of the tree. But what of a to-be-deleted node with two children? Our solution is to relabel this node with the key of its immediate successor in sorted order. This successor must be the smallest value in the right subtree, speciﬁcally the leftmost descendant in the right subtree (p). Moving this to the point of deletion results in a properlylabeled binary search tree, and reduces our deletion problem to physically removing a node with at most one child—a case that has been resolved above. The full implementation has been omitted here because it looks a little ghastly, but the code follows logically from the description above. The worst-case complexity analysis is as follows. Every deletion requires the cost of at most two search operations, each taking O(h) time where h is the height of the tree, plus a constant amount of pointer manipulation. 3.4.2 How Good Are Binary Search Trees? When implemented using binary search trees, all three dictionary operations take O(h) time, where h is the height of the tree. The smallest height we can hope for occurs when the tree is perfectly balanced, where h = log n. This is very good, but the tree must be perfectly balanced. 81 82 3. DATA STRUCTURES Our insertion algorithm puts each new item at a leaf node where it should have been found. This makes the shape (and more importantly height) of the tree a function of the order in which we insert the keys. Unfortunately, bad things can happen when building trees through insertion. The data structure has no control over the order of insertion. Consider what happens if the user inserts the keys in sorted order. The operations insert(a), followed by insert(b), insert(c), insert(d), . . . will produce a skinny linear height tree where only right pointers are used. Thus binary trees can have heights ranging from lg n to n. But how tall are they on average? The average case analysis of algorithms can be tricky because we must carefully specify what we mean by average. The question is well deﬁned if we consider each of the n! possible insertion orderings equally likely and average over those. If so, we are in luck, because with high probability the resulting tree will have O(log n) height. This will be shown in Section 4.6 (page 123). This argument is an important example of the power of randomization. We can often develop simple algorithms that oﬀer good performance with high probability. We will see that a similar idea underlies the fastest known sorting algorithm, quicksort. 3.4.3 Balanced Search Trees Random search trees are usually good. But if we get unlucky with our order of insertion, we can end up with a linear-height tree in the worst case. This worst case is outside of our direct control, since we must build the tree in response to the requests given by our potentially nasty user. What would be better is an insertion/deletion procedure which adjusts the tree a little after each insertion, keeping it close enough to be balanced so the maximum height is logarithmic. Sophisticated balanced binary search tree data structures have been developed that guarantee the height of the tree always to be O(log n). Therefore, all dictionary operations (insert, delete, query) take O(log n) time each. Implementations of balanced tree data structures such as red-black trees and splay trees are discussed in Section 12.1 (page 367). From an algorithm design viewpoint, it is important to know that these trees exist and that they can be used as black boxes to provide an eﬃcient dictionary implementation. When ﬁguring the costs of dictionary operations for algorithm analysis, we can assume the worst-case complexities of balanced binary trees to be a fair measure. Take-Home Lesson: Picking the wrong data structure for the job can be disastrous in terms of performance. Identifying the very best data structure is usually not as critical, because there can be several choices that perform similarly. 3.5 PRIORITY QUEUES Stop and Think: Exploiting Balanced Search Trees Problem: You are given the task of reading n numbers and then printing them out in sorted order. Suppose you have access to a balanced dictionary data structure, which supports the operations search, insert, delete, minimum, maximum, successor, and predecessor each in O(log n) time. 1. How can you sort in O(n log n) time using only insert and in-order traversal? 2. How can you sort in O(n log n) time using only minimum, successor, and insert? 3. How can you sort in O(n log n) time using only minimum, insert, delete, search? Solution: The ﬁrst problem allows us to do insertion and inorder-traversal. We can build a search tree by inserting all n elements, then do a traversal to access the items in sorted order: Sort2() initialize-tree(t) Sort1() While (not EOF) initialize-tree(t) read(x); While (not EOF) insert(x,t); read(x); y = Minimum(t) insert(x,t) While (y = NULL) do Traverse(t) print(y → item) y = Successor(y,t) Sort3() initialize-tree(t) While (not EOF) read(x); insert(x,t); y = Minimum(t) While (y = NULL) do print(y→item) Delete(y,t) y = Minimum(t) The second problem allows us to use the minimum and successor operations after constructing the tree. We can start from the minimum element, and then repeatedly ﬁnd the successor to traverse the elements in sorted order. The third problem does not give us successor, but does allow us delete. We can repeatedly ﬁnd and delete the minimum element to once again traverse all the elements in sorted order. Each of these algorithms does a linear number of logarithmic-time operations, and hence runs in O(n log n) time. The key to exploiting balanced binary search trees is using them as black boxes. 3.5 Priority Queues Many algorithms process items in a speciﬁc order. For example, suppose you must schedule jobs according to their importance relative to other jobs. Scheduling the 83 84 3. DATA STRUCTURES jobs requires sorting them by importance, and then evaluating them in this sorted order. Priority queues are data structures that provide more ﬂexibility than simple sorting, because they allow new elements to enter a system at arbitrary intervals. It is much more cost-eﬀective to insert a new job into a priority queue than to re-sort everything on each such arrival. The basic priority queue supports three primary operations: • Insert(Q,x)– Given an item x with key k, insert it into the priority queue Q. • Find-Minimum(Q) or Find-Maximum(Q)– Return a pointer to the item whose key value is smaller (larger) than any other key in the priority queue Q. • Delete-Minimum(Q) or Delete-Maximum(Q)– Remove the item from the priority queue Q whose key is minimum (maximum). Many naturally occurring processes are accurately modeled by priority queues. Single people maintain a priority queue of potential dating candidates—mentally if not explicitly. One’s impression on meeting a new person maps directly to an attractiveness or desirability score. Desirability serves as the key ﬁeld for inserting this new entry into the “little black book” priority queue data structure. Dating is the process of extracting the most desirable person from the data structure (FindMaximum), spending an evening to evaluate them better, and then reinserting them into the priority queue with a possibly revised score. Take-Home Lesson: Building algorithms around data structures such as dictionaries and priority queues leads to both clean structure and good performance. Stop and Think: Basic Priority Queue Implementations Problem: What is the worst-case time complexity of the three basic priority queue operations (insert, ﬁnd-minimum, and delete-minimum) when the basic data structure is • An unsorted array. • A sorted array. • A balanced binary search tree. Solution: There is surprising subtlety in implementing these three operations, even when using a data structure as simple as an unsorted array. The unsorted array 3.6 WAR STORY: STRIPPING TRIANGULATIONS dictionary (discussed on page 73) implemented insertion and deletion in constant time, and search and minimum in linear time. A linear time implementation of delete-minimum can be composed from ﬁnd-minimum, followed by search, followed by delete. For sorted arrays, we can implement insert and delete in linear time, and minimum in constant time. However, all priority queue deletions involve only the minimum element. By storing the sorted array in reverse order (largest value on top), the minimum element will be the last one in the array. Deleting the tail element requires no movement of any items, just decrementing the number of remaining items n, and so delete-minimum can be implemented in constant time. All this is ﬁne, yet the following table claims we can implement ﬁnd-minimum in constant time for each data structure: Insert(Q, x) Find-Minimum(Q) Delete-Minimum(Q) Unsorted array O(1) O(1) O(n) Sorted array O(n) O(1) O(1) Balanced tree O(log n) O(1) O(log n) The trick is using an extra variable to store a pointer/index to the minimum entry in each of these structures, so we can simply return this value whenever we are asked to ﬁnd-minimum. Updating this pointer on each insertion is easy—we update it if and only if the newly inserted value is less than the current minimum. But what happens on a delete-minimum? We can delete the minimum entry have, then do an honest ﬁnd-minimum to restore our canned value. The honest ﬁndminimum takes linear time on an unsorted array and logarithmic time on a tree, and hence can be folded into the cost of each deletion. Priority queues are very useful data structures. Indeed, they will be the hero of two of our war stories, including the next one. A particularly nice priority queue implementation (the heap) will be discussed in the context of sorting in Section 4.3 (page 108). Further, a complete set of priority queue implementations is presented in Section 12.2 (page 373) of the catalog. 3.6 War Story: Stripping Triangulations Geometric models used in computer graphics are commonly represented as a triangulated surface, as shown in Figure 3.5(l). High-performance rendering engines have special hardware for rendering and shading triangles. This hardware is so fast that the bottleneck of rendering is the cost of feeding the triangulation structure into the hardware engine. Although each triangle can be described by specifying its three endpoints, an alternative representation is more eﬃcient. Instead of specifying each triangle in isolation, suppose that we partition the triangles into strips of adjacent triangles 85 86 3. DATA STRUCTURES Figure 3.5: (l) A triangulated model of a dinosaur (r) Several triangle strips in the model i-2 i-4 i i-1 i-3 a) i-2 i i-4 i-1 i-3 b) Figure 3.6: Partitioning a triangular mesh into strips: (a) with left-right turns (b) with the ﬂexibility of arbitrary turns and walk along the strip. Since each triangle shares two vertices in common with its neighbors, we save the cost of retransmitting the two extra vertices and any associated information. To make the description of the triangles unambiguous, the OpenGL triangular-mesh renderer assumes that all turns alternate from left to right (as shown in Figure 3.6). The task of ﬁnding a small number of strips that cover each triangle in a mesh can be thought of as a graph problem. The graph of interest has a vertex for every triangle of the mesh, and an edge between every pair of vertices representing adjacent triangles. This dual graph representation captures all the information about the triangulation (see Section 15.12 (page 520)) needed to partition it into triangular strips. Once we had the dual graph available, the project could begin in earnest. We sought to partition the vertices into as few paths or strips as possible. Partitioning it into one path implied that we had discovered a Hamiltonian path, which by deﬁnition visits each vertex exactly once. Since ﬁnding a Hamiltonian path is NP-complete (see Section 16.5 (page 538)), we knew not to look for an optimal algorithm, but concentrate instead on heuristics. The simplest heuristic for strip cover would start from an arbitrary triangle and then do a left-right walk until the walk ends, either by hitting the boundary of 3.6 WAR STORY: STRIPPING TRIANGULATIONS top 1 2 3 4 253 254 255 256 Figure 3.7: A bounded height priority queue for triangle strips the object or a previously visited triangle. This heuristic had the advantage that it would be fast and simple, although there is no reason why it should ﬁnd the smallest possible set of left-right strips for a given triangulation. The greedy heuristic would be more likely to result in a small number of strips however. Greedy heuristics always try to grab the best possible thing ﬁrst. In the case of the triangulation, the natural greedy heuristic would identify the starting triangle that yields the longest left-right strip, and peel that one oﬀ ﬁrst. Being greedy does not guarantee you the best possible solution either, since the ﬁrst strip you peel oﬀ might break apart a lot of potential strips we might have wanted to use later. Still, being greedy is a good rule of thumb if you want to get rich. Since removing the longest strip would leave the fewest number of triangles for later strips, the greedy heuristic should outperform the naive heuristic. But how much time does it take to ﬁnd the largest strip to peel oﬀ next? Let k be the length of the walk possible from an average vertex. Using the simplest possible implementation, we could walk from each of the n vertices to ﬁnd the largest remaining strip to report in O(kn) time. Repeating this for each of the roughly n/k strips we extract yields an O(n2 )-time implementation, which would be hopelessly slow on a typical model of 20,000 triangles. How could we speed this up? It seems wasteful to rewalk from each triangle after deleting a single strip. We could maintain the lengths of all the possible future strips in a data structure. However, whenever we peel oﬀ a strip, we must update the lengths of all aﬀected strips. These strips will be shortened because they walked through a triangle that now no longer exists. There are two aspects of such a data structure: • Priority Queue – Since we were repeatedly identifying the longest remaining strip, we needed a priority queue to store the strips ordered according to length. The next strip to peel always sat at the top of the queue. Our priority queue had to permit reducing the priority of arbitrary elements of the queue whenever we updated the strip lengths to reﬂect what triangles were peeled 87 88 3. DATA STRUCTURES Model name Diver Heads Framework Bart Simpson Enterprise Torus Jaw Triangle count 3,798 4,157 5,602 9,654 12,710 20,000 75,842 Naive cost 8,460 10,588 9,274 24,934 29,016 40,000 104,203 Greedy cost 4,650 4,749 7,210 11,676 13,738 20,200 95,020 Greedy time 6.4 sec 9.9 sec 9.7 sec 20.5 sec 26.2 sec 272.7 sec 136.2 sec Figure 3.8: A comparison of the naive versus greedy heuristics for several triangular meshes away. Because all of the strip lengths were bounded by a fairly small integer (hardware constraints prevent any strip from having more than 256 vertices), we used a bounded-height priority queue (an array of buckets shown in Figure 3.7 and described in Section 12.2 (page 373)). An ordinary heap would also have worked just ﬁne. To update the queue entry associated with each triangle, we needed to quickly ﬁnd where it was. This meant that we also needed a . . . • Dictionary – For each triangle in the mesh, we needed to ﬁnd where it was in the queue. This meant storing a pointer to each triangle in a dictionary. By integrating this dictionary with the priority queue, we built a data structure capable of a wide range of operations. Although there were various other complications, such as quickly recalculating the length of the strips aﬀected by the peeling, the key idea needed to obtain better performance was to use the priority queue. Run time improved by several orders of magnitude after employing this data structure. How much better did the greedy heuristic do than the naive heuristic? Consider the table in Figure 3.8. In all cases, the greedy heuristic led to a set of strips that cost less, as measured by the total size of the strips. The savings ranged from about 10% to 50%, which is quite remarkable since the greatest possible improvement (going from three vertices per triangle down to one) yields a savings of only 66.6%. After implementing the greedy heuristic with our priority queue data structure, the program ran in O(n · k) time, where n is the number of triangles and k is the length of the average strip. Thus the torus, which consisted of a small number of very long strips, took longer than the jaw, even though the latter contained over three times as many triangles. There are several lessons to be gleaned from this story. First, when working with a large enough data set, only linear or near linear algorithms (say O(n log n)) are likely to be fast enough. Second, choosing the right data structure is often the key to getting the time complexity down to this point. Finally, using smart heuristic 3.7 HASHING AND STRINGS like greedy is likely to signiﬁcantly improve quality over the naive approach. How much the improvement will be can only be determined by experimentation. 3.7 Hashing and Strings Hash tables are a very practical way to maintain a dictionary. They exploit the fact that looking an item up in an array takes constant time once you have its index. A hash function is a mathematical function that maps keys to integers. We will use the value of our hash function as an index into an array, and store our item at that position. The ﬁrst step of the hash function is usually to map each key to a big integer. Let α be the size of the alphabet on which a given string S is written. Let char(c) be a function that maps each symbol of the alphabet to a unique integer from 0 to α − 1. The function |S|−1 H(S) = α|S|−(i+1) × char(si ) i=0 maps each string to a unique (but large) integer by treating the characters of the string as “digits” in a base-α number system. The result is unique identiﬁer numbers, but they are so large they will quickly exceed the number of slots in our hash table (denoted by m). We must reduce this number to an integer between 0 and m−1, by taking the remainder of H(S) mod m. This works on the same principle as a roulette wheel. The ball travels a long distance around and around the circumference-m wheel H(S)/m times before settling down to a random bin. If the table size is selected with enough ﬁnesse (ideally m is a large prime not too close to 2i − 1), the resulting hash values should be fairly uniformly distributed. 3.7.1 Collision Resolution No matter how good our hash function is, we had better be prepared for collisions, because two distinct keys will occasionally hash to the same value. Chaining is the easiest approach to collision resolution. Represent the hash table as an array of m linked lists, as shown in Figure 3.9. The ith list will contain all the items that hash to the value of i. Thus search, insertion, and deletion reduce to the corresponding problem in linked lists. If the n keys are distributed uniformly in a table, each list will contain roughly n/m elements, making them a constant size when m ≈ n. Chaining is very natural, but devotes a considerable amount of memory to pointers. This is space that could be used to make the table larger, and hence the “lists” smaller. The alternative is something called open addressing. The hash table is maintained as an array of elements (not buckets), each initialized to null, as shown in Figure 3.10. On an insertion, we check to see if the desired position is empty. If so, 89 90 3. DATA STRUCTURES Figure 3.9: Collision resolution by chaining 1 2 3 X 4 5 6 X X 7 8 9 X X 10 11 Figure 3.10: Collision resolution by open addressing we insert it. If not, we must ﬁnd some other place to insert it instead. The simplest possibility (called sequential probing) inserts the item in the next open spot in the table. If the table is not too full, the contiguous runs of items should be fairly small, hence this location should be only a few slots from its intended position. Searching for a given key now involves going to the appropriate hash value and checking to see if the item there is the one we want. If so, return it. Otherwise we must keep checking through the length of the run. Deletion in an open addressing scheme can get ugly, since removing one element might break a chain of insertions, making some elements inaccessible. We have no alternative but to reinsert all the items in the run following the new hole. Chaining and open addressing both require O(m) to initialize an m-element hash table to null elements prior to the ﬁrst insertion. Traversing all the elements in the table takes O(n + m) time for chaining, because we have to scan all m buckets looking for elements, even if the actual number of inserted items is small. This reduces to O(m) time for open addressing, since n must be at most m. When using chaining with doubly-linked lists to resolve collisions in an melement hash table, the dictionary operations for n items can be implemented in the following expected and worst case times: Search(L, k) Insert(L, x) Delete(L, x) Successor(L, x) Predecessor(L, x) Minimum(L) Maximum(L) Hash table (expected) O(n/m) O(1) O(1) O(n + m) O(n + m) O(n + m) O(n + m) Hash table (worst case) O(n) O(1) O(1) O(n + m) O(n + m) O(n + m) O(n + m) 3.7 HASHING AND STRINGS Pragmatically, a hash table is often the best data structure to maintain a dictionary. The applications of hashing go far beyond dictionaries, however, as we will see below. 3.7.2 Eﬃcient String Matching via Hashing Strings are sequences of characters where the order of the characters matters, since ALGORITHM is diﬀerent than LOGARITHM. Text strings are fundamental to a host of computing applications, from programming language parsing/compilation, to web search engines, to biological sequence analysis. The primary data structure for representing strings is an array of characters. This allows us constant-time access to the ith character of the string. Some auxiliary information must be maintained to mark the end of the string—either a special end-of-string character or (perhaps more usefully) a count of the n characters in the string. The most fundamental operation on text strings is substring search, namely: Problem: Substring Pattern Matching Input: A text string t and a pattern string p. Output: Does t contain the pattern p as a substring, and if so where? The simplest algorithm to search for the presence of pattern string p in text t overlays the pattern string at every position in the text, and checks whether every pattern character matches the corresponding text character. As demonstrated in Section 2.5.3 (page 43), this runs in O(nm) time, where n = |t| and m = |p|. This quadratic bound is worst-case. More complicated, worst-case linear-time search algorithms do exist: see Section 18.3 (page 628) for a complete discussion. But here we give a linear expected-time algorithm for string matching, called the Rabin-Karp algorithm. It is based on hashing. Suppose we compute a given hash function on both the pattern string p and the m-character substring starting from the ith position of t. If these two strings are identical, clearly the resulting hash values must be the same. If the two strings are diﬀerent, the hash values will almost certainly be diﬀerent. These false positives should be so rare that we can easily spend the O(m) time it takes to explicitly check the identity of two strings whenever the hash values agree. This reduces string matching to n−m+2 hash value computations (the n−m+1 windows of t, plus one hash of p), plus what should be a very small number of O(m) time veriﬁcation steps. The catch is that it takes O(m) time to compute a hash function on an m-character string, and O(n) such computations seems to leave us with an O(mn) algorithm again. But let’s look more closely at our previously deﬁned hash function, applied to the m characters starting from the jth position of string S: H(S, j) = m−1 i=0 αm−(i+1) × char(si+j ) 91 92 3. DATA STRUCTURES What changes if we now try to compute H(S, j + 1)—the hash of the next window of m characters? Note that m−1 characters are the same in both windows, although this diﬀers by one in the number of times they are multiplied by α. A little algebra reveals that H(S, j + 1) = α(H(S, j) − αm−1 char(sj )) + char(sj+m ) This means that once we know the hash value from the j position, we can ﬁnd the hash value from the (j + 1)st position for the cost of two multiplications, one addition, and one subtraction. This can be done in constant time (the value of αm−1 can be computed once and used for all hash value computations). This math works even if we compute H(S, j) mod M , where M is a reasonably large prime number, thus keeping the size of our hash values small (at most M ) even when the pattern string is long. Rabin-Karp is a good example of a randomized algorithm (if we pick M in some random way). We get no guarantee the algorithm runs in O(n+m) time, because we may get unlucky and have the hash values regularly collide with spurious matches. Still, the odds are heavily in our favor—if the hash function returns values uniformly from 0 to M − 1, the probability of a false collision should be 1/M . This is quite reasonable: if M ≈ n, there should only be one false collision per string, and if M ≈ nk for k ≥ 2, the odds are great we will never see any false collisions. 3.7.3 Duplicate Detection Via Hashing The key idea of hashing is to represent a large object (be it a key, a string, or a substring) using a single number. The goal is a representation of the large object by an entity that can be manipulated in constant time, such that it is relatively unlikely that two diﬀerent large objects map to the same value. Hashing has a variety of clever applications beyond just speeding up search. I once heard Udi Manber—then Chief Scientist at Yahoo—talk about the algorithms employed at his company. The three most important algorithms at Yahoo, he said, were hashing, hashing, and hashing. Consider the following problems with nice hashing solutions: • Is a given document diﬀerent from all the rest in a large corpus? – A search engine with a huge database of n documents spiders yet another webpage. How can it tell whether this adds something new to add to the database, or is just a duplicate page that exists elsewhere on the Web? Explicitly comparing the new document D to all n documents is hopelessly ineﬃcient for a large corpus. But we can hash D to an integer, and compare it to the hash codes of the rest of the corpus. Only when there is a collision is D a possible duplicate. Since we expect few spurious collisions, we can explicitly compare the few documents sharing the exact hash code with little eﬀort. 3.8 SPECIALIZED DATA STRUCTURES • Is part of this document plagiarized from a document in a large corpus? – A lazy student copies a portion of a Web document into their term paper. “The Web is a big place,” he smirks. “How will anyone ever ﬁnd which one?” This is a more diﬃcult problem than the previous application. Adding, deleting, or changing even one character from a document will completely change its hash code. Thus the hash codes produced in the previous application cannot help for this more general problem. However, we could build a hash table of all overlapping windows (substrings) of length w in all the documents in the corpus. Whenever there is a match of hash codes, there is likely a common substring of length w between the two documents, which can then be further investigated. We should choose w to be long enough so such a co-occurrence is very unlikely to happen by chance. The biggest downside of this scheme is that the size of the hash table becomes as large as the documents themselves. Retaining a small but well-chosen subset of these hash codes (say those which are exact multiples of 100) for each document leaves us likely to detect suﬃciently long duplicate strings. • How can I convince you that a ﬁle isn’t changed? – In a closed-bid auction, each party submits their bid in secret before the announced deadline. If you knew what the other parties were bidding, you could arrange to bid $1 more than the highest opponent and walk oﬀ with the prize as cheaply as possible. Thus the “right” auction strategy is to hack into the computer containing the bids just prior to the deadline, read the bids, and then magically emerge the winner. How can this be prevented? What if everyone submits a hash code of their actual bid prior to the deadline, and then submits the full bid after the deadline? The auctioneer will pick the largest full bid, but checks to make sure the hash code matches that submitted prior to the deadline. Such cryptographic hashing methods provide a way to ensure that the ﬁle you give me today is the same as original, because any changes to the ﬁle will result in changing the hash code. Although the worst-case bounds on anything involving hashing are dismal, with a proper hash function we can conﬁdently expect good behavior. Hashing is a fundamental idea in randomized algorithms, yielding linear expected-time algorithms for problems otherwise Θ(n log n), or Θ(n2 ) in the worst case. 3.8 Specialized Data Structures The basic data structures described thus far all represent an unstructured set of items so as to facilitate retrieval operations. These data structures are well known to most programmers. Not as well known are data structures for representing more 93 94 3. DATA STRUCTURES structured or specialized kinds of objects, such as points in space, strings, and graphs. The design principles of these data structures are the same as for basic objects. There exists a set of basic operations we need to perform repeatedly. We seek a data structure that supports these operations very eﬃciently. These eﬃcient, specialized data structures are important for eﬃcient graph and geometric algorithms so one should be aware of their existence. Details appear throughout the catalog. • String data structures – Character strings are typically represented by arrays of characters, perhaps with a special character to mark the end of the string. Suﬃx trees/arrays are special data structures that preprocess strings to make pattern matching operations faster. See Section 12.3 (page 377) for details. • Geometric data structures – Geometric data typically consists of collections of data points and regions. Regions in the plane can be described by polygons, where the boundary of the polygon is given by a chain of line segments. Polygons can be represented using an array of points (v1 , . . . , vn , v1 ), such that (vi , vi+1 ) is a segment of the boundary. Spatial data structures such as kd-trees organize points and regions by geometric location to support fast search. For more details, see Section 12.6 (page 389). • Graph data structures – Graphs are typically represented using either adjacency matrices or adjacency lists. The choice of representation can have a substantial impact on the design of the resulting graph algorithms, as discussed in Chapter 6 and in the catalog in Section 12.4. • Set data structures – Subsets of items are typically represented using a dictionary to support fast membership queries. Alternately, bit vectors are boolean arrays such that the ith bit represents true if i is in the subset. Data structures for manipulating sets is presented in the catalog in Section 12.5. The union-ﬁnd data structure for maintaining set partitions will be covered in Section 6.1.3 (page 198). 3.9 War Story: String ’em Up The human genome encodes all the information necessary to build a person. This project has already had an enormous impact on medicine and molecular biology. Algorists have become interested in the human genome project as well, for several reasons: • DNA sequences can be accurately represented as strings of characters on the four-letter alphabet (A,C,T,G). Biologist’s needs have sparked new interest in old algorithmic problems such as string matching (see Section 18.3 (page 628)) as well as creating new problems such as shortest common superstring (see Section 18.9 (page 654)). 3.9 A WAR STORY: STRING ’EM UP T A T C T T A T C G T T A T C G T T A C G T T A T C 95 C C A Figure 3.11: The concatenation of two fragments can be in S only if all sub-fragments are • DNA sequences are very long strings. The human genome is approximately three billion base pairs (or characters) long. Such large problem size means that asymptotic (Big-Oh) complexity analysis is usually fully justiﬁed on biological problems. • Enough money is being invested in genomics for computer scientists to want to claim their piece of the action. One of my interests in computational biology revolved around a proposed technique for DNA sequencing called sequencing by hybridization (SBH). This procedure attaches a set of probes to an array, forming a sequencing chip. Each of these probes determines whether or not the probe string occurs as a substring of the DNA target. The target DNA can now be sequenced based on the constraints of which strings are (and are not) substrings of the target. We sought to identify all the strings of length 2k that are possible substrings of an unknown string S, given the set of all length k substrings of S. For example, suppose we know that AC, CA, and CC are the only length-2 substrings of S. It is possible that ACCA is a substring of S, since the center substring is one of our possibilities. However, CAAC cannot be a substring of S, since AA is not a substring of S. We needed to ﬁnd a fast algorithm to construct all the consistent length-2k strings, since S could be very long. The simplest algorithm to build the 2k strings would be to concatenate all O(n2 ) pairs of k-strings together, and then test to make sure that all (k − 1) length-k substrings spanning the boundary of the concatenation were in fact substrings, as shown in Figure 3.11. For example, the nine possible concatenations of AC, CA, and CC are ACAC, ACCA, ACCC, CAAC, CACA, CACC, CCAC, CCCA, and CCCC. Only CAAC can be eliminated because of the absence of AA. We needed a fast way of testing whether the k − 1 substrings straddling the concatenation were members of our dictionary of permissible k-strings. The time it takes to do this depends upon which dictionary data structure we use. A binary search tree could ﬁnd the correct string within O(log n) comparisons, where each 96 3. DATA STRUCTURES comparison involved testing which of two length-k strings appeared ﬁrst in alphabetical order. The total time using such a binary search tree would be O(k log n). That seemed pretty good. So my graduate student, Dimitris Margaritis, used a binary search tree data structure for our implementation. It worked great up until the moment we ran it. “I’ve tried the fastest computer in our department, but our program is too slow,” Dimitris complained. “It takes forever on string lengths of only 2,000 characters. We will never get up to 50,000.” We proﬁled our program and discovered that almost all the time was spent searching in this data structure. This was no surprise since we did this k − 1 times for each of the O(n2 ) possible concatenations. We needed a faster dictionary data structure, since search was the innermost operation in such a deep loop. “How about using a hash table?” I suggested. “It should take O(k) time to hash a k-character string and look it up in our table. That should knock oﬀ a factor of O(log n), which will mean something when n ≈ 2,000.” Dimitris went back and implemented a hash table implementation for our dictionary. Again, it worked great up until the moment we ran it. “Our program is still too slow,” Dimitris complained. “Sure, it is now about ten times faster on strings of length 2,000. So now we can get up to about 4,000 characters. Big deal. We will never get up to 50,000.” “We should have expected this,” I mused. “After all, lg2 (2, 000) ≈ 11. We need a faster data structure to search in our dictionary of strings.” “But what can be faster than a hash table?” Dimitris countered. “To look up a k-character string, you must read all k characters. Our hash table already does O(k) searching.” “Sure, it takes k comparisons to test the ﬁrst substring. But maybe we can do better on the second test. Remember where our dictionary queries are coming from. When we concatenate ABCD with EF GH, we are ﬁrst testing whether BCDE is in the dictionary, then CDEF . These strings diﬀer from each other by only one character. We should be able to exploit this so each subsequent test takes constant time to perform. . . .” “We can’t do that with a hash table,” Dimitris observed. “The second key is not going to be anywhere near the ﬁrst in the table. A binary search tree won’t help, either. Since the keys ABCD and BCDE diﬀer according to the ﬁrst character, the two strings will be in diﬀerent parts of the tree.” “But we can use a suﬃx tree to do this,” I countered. “A suﬃx tree is a trie containing all the suﬃxes of a given set of strings. For example, the suﬃxes of ACAC are {ACAC, CAC, AC, C}. Coupled with suﬃxes of string CACT , we get the suﬃx tree of Figure 3.12. By following a pointer from ACAC to its longest proper suﬃx CAC, we get to the right place to test whether CACT is in our set of strings. One character comparison is all we need to do from there.” Suﬃx trees are amazing data structures, discussed in considerably more detail in Section 12.3 (page 377). Dimitris did some reading about them, then built a nice 3.9 a a c t c c t a t WAR STORY: STRING ’EM UP c t Figure 3.12: Suﬃx tree on ACAC and CACT , with the pointer to the suﬃx of ACAC suﬃx tree implementation for our dictionary. Once again, it worked great up until the moment we ran it. “Now our program is faster, but it runs out of memory,” Dimitris complained. “The suﬃx tree builds a path of length k for each suﬃx of length k, so all told there can be Θ(n2 ) nodes in the tree. It crashes when we go beyond 2,000 characters. We will never get up to strings with 50,000 characters.” I wasn’t ready to give up yet. “There is a way around the space problem, by using compressed suﬃx trees,” I recalled. “Instead of explicitly representing long paths of character nodes, we can refer back to the original string.” Compressed suﬃx trees always take linear space, as described in Section 12.3 (page 377). Dimitris went back one last time and implemented the compressed suﬃx tree data structure. Now it worked great! As shown in Figure 3.13, we ran our simulation for strings of length n = 65,536 without incident. Our results showed that interactive SBH could be a very eﬃcient sequencing technique. Based on these simulations, we were able to arouse interest in our technique from biologists. Making the actual wet laboratory experiments feasible provided another computational challenge, which is reported in Section 7.7 (page 263). The take-home lessons for programmers from Figure 3.13 should be apparent. We isolated a single operation (dictionary string search) that was being performed repeatedly and optimized the data structure we used to support it. We started with a simple implementation (binary search trees) in the hopes that it would suﬃce, and then used proﬁling to reveal the trouble when it didn’t. When an improved dictionary structure still did not suﬃce, we looked deeper into the kind of queries we were performing, so that we could identify an even better data structure. Finally, we didn’t give up until we had achieved the level of performance we needed. In algorithms, as in life, persistence usually pays oﬀ. 97 98 3. DATA STRUCTURES String length 8 16 32 64 128 256 512 1,024 2,048 4,096 8,192 16,384 32,768 65,536 Binary tree 0.0 0.0 0.1 0.3 2.4 17.1 31.6 1,828.9 11,441.7 > 2 days Hash table 0.0 0.0 0.0 0.4 1.1 9.4 67.0 96.6 941.7 5,246.7 > 2 days Suﬃx tree 0.0 0.0 0.0 0.3 0.5 3.8 6.9 31.5 553.6 out of memory Compressed tree 0.0 0.0 0.0 0.0 0.0 0.2 1.3 2.7 39.0 45.4 642.0 1,614.0 13,657.8 39,776.9 Figure 3.13: Run times (in seconds) for the SBH simulation using various data structures Chapter Notes Optimizing hash table performance is surprisingly complicated for such a conceptually simple data structure. The importance of short runs in open addressing has to more sophisticated schemes than sequential probing for optimal hash table performance. For more details, see Knuth [Knu98]. Our triangle strip optimizing program, stripe, is described in [ESV96]. Hashing techniques for plagiarism detection are discussed in [SWA03]. Surveys of algorithmic issues in DNA sequencing by hybridization include [CK94, PL94]. Our work on interactive SBH reported in the war story is reported in [MS95a]. 3.10 Exercises Stacks, Queues, and Lists 3-1. [3] A common problem for compilers and text editors is determining whether the parentheses in a string are balanced and properly nested. For example, the string ((())())() contains properly nested pairs of parentheses, which the strings )()( and ()) do not. Give an algorithm that returns true if a string contains properly nested and balanced parentheses, and false if otherwise. For full credit, identify the position of the ﬁrst oﬀending parenthesis if the string is not properly nested and balanced. 3.10 EXERCISES 3-2. [3] Write a program to reverse the direction of a given singly-linked list. In other words, after the reversal all pointers should now point backwards. Your algorithm should take linear time. 3-3. [5] We have seen how dynamic arrays enable arrays to grow while still achieving constant-time amortized performance. This problem concerns extending dynamic arrays to let them both grow and shrink on demand. (a) Consider an underﬂow strategy that cuts the array size in half whenever the array falls below half full. Give an example sequence of insertions and deletions where this strategy gives a bad amortized cost. (b) Then, give a better underﬂow strategy than that suggested above, one that achieves constant amortized cost per deletion. Trees and Other Dictionary Structures 3-4. [3] Design a dictionary data structure in which search, insertion, and deletion can all be processed in O(1) time in the worst case. You may assume the set elements are integers drawn from a ﬁnite set 1, 2, .., n, and initialization can take O(n) time. 3-5. [3] Find the overhead fraction (the ratio of data space over total space) for each of the following binary tree implementations on n nodes: (a) All nodes store data, two child pointers, and a parent pointer. The data ﬁeld requires four bytes and each pointer requires four bytes. (b) Only leaf nodes store data; internal nodes store two child pointers. The data ﬁeld requires four bytes and each pointer requires two bytes. 3-6. [5] Describe how to modify any balanced tree data structure such that search, insert, delete, minimum, and maximum still take O(log n) time each, but successor and predecessor now take O(1) time each. Which operations have to be modiﬁed to support this? 3-7. [5] Suppose you have access to a balanced dictionary data structure, which supports each of the operations search, insert, delete, minimum, maximum, successor, and predecessor in O(log n) time. Explain how to modify the insert and delete operations so they still take O(log n) but now minimum and maximum take O(1) time. (Hint: think in terms of using the abstract dictionary operations, instead of mucking about with pointers and the like.) 3-8. [6 ] Design a data structure to support the following operations: • insert(x,T) – Insert item x into the set T . • delete(k,T) – Delete the kth smallest element from T . • member(x,T) – Return true iﬀ x ∈ T . All operations must take O(log n) time on an n-element set. 3-9. [8] A concatenate operation takes two sets S1 and S2 , where every key in S1 is smaller than any key in S2 , and merges them together. Give an algorithm to concatenate two binary search trees into one binary search tree. The worst-case running time should be O(h), where h is the maximal height of the two trees. 99 100 3. DATA STRUCTURES Applications of Tree Structures 3-10. [5] In the bin-packing problem, we are given n metal objects, each weighing between zero and one kilogram. Our goal is to ﬁnd the smallest number of bins that will hold the n objects, with each bin holding one kilogram at most. • The best-ﬁt heuristic for bin packing is as follows. Consider the objects in the order in which they are given. For each object, place it into the partially ﬁlled bin with the smallest amount of extra room after the object is inserted.. If no such bin exists, start a new bin. Design an algorithm that implements the best-ﬁt heuristic (taking as input the n weights w1 , w2 , ..., wn and outputting the number of bins used) in O(n log n) time. • Repeat the above using the worst-ﬁt heuristic, where we put the next object in the partially ﬁlled bin with the largest amount of extra room after the object is inserted. 3-11. [5] Suppose that we are given a sequence of n values x1 , x2 , ..., xn and seek to quickly answer repeated queries of the form: given i and j, ﬁnd the smallest value in xi , . . . , xj . (a) Design a data structure that uses O(n2 ) space and answers queries in O(1) time. (b) Design a data structure that uses O(n) space and answers queries in O(log n) time. For partial credit, your data structure can use O(n log n) space and have O(log n) query time. 3-12. [5] Suppose you are given an input set S of n numbers, and a black box that if given any sequence of real numbers and an integer k instantly and correctly answers whether there is a subset of input sequence whose sum is exactly k. Show how to use the black box O(n) times to ﬁnd a subset of S that adds up to k. 3-13. [5] Let A[1..n] be an array of real numbers. Design an algorithm to perform any sequence of the following operations: • Add(i,y) – Add the value y to the ith number. • Partial-sum(i) – Return the sum of the ﬁrst i numbers, i.e. i j=1 A[j]. There are no insertions or deletions; the only change is to the values of the numbers. Each operation should take O(log n) steps. You may use one additional array of size n as a work space. 3-14. [8] Extend the data structure of the previous problem to support insertions and deletions. Each element now has both a key and a value. An element is accessed by its key. The addition operation is applied to the values, but the elements are accessed by its key. The Partial sum operation is diﬀerent. • Add(k,y) – Add the value y to the item with key k. • Insert(k,y) – Insert a new item with key k and value y. • Delete(k) – Delete the item with key k. 3.10 EXERCISES • Partial-sum(k) – Return the sum of all the elements currently in the set whose x. key is less than y, i.e. x ” if a > b, or “=” if a = b. 107 108 4. SORTING AND SEARCHING By abstracting the pairwise ordering decision to such a comparison function, we can implement sorting algorithms independently of such criteria. We simply pass the comparison function in as an argument to the sort procedure. Any reasonable programming language has a built-in sort routine as a library function. You are almost always better oﬀ using this than writing your own routine. For example, the standard library for C contains the qsort function for sorting: #include void qsort(void *base, size_t nel, size_t width, int (*compare) (const void *, const void *)); The key to using qsort is realizing what its arguments do. It sorts the ﬁrst nel elements of an array (pointed to by base), where each element is width-bytes long. Thus we can sort arrays of 1-byte characters, 4-byte integers, or 100-byte records, all by changing the value of width. The ultimate desired order is determined by the compare function. It takes as arguments pointers to two width-byte elements, and returns a negative number if the ﬁrst belongs before the second in sorted order, a positive number if the second belongs before the ﬁrst, or zero if they are the same. Here is a comparison function to sort integers in increasing order: int intcompare(int *i, int *j) { if (*i > *j) return (1); if (*i < *j) return (-1); return (0); } This comparison function can be used to sort an array a, of which the ﬁrst n elements are occupied, as follows: qsort(a, n, sizeof(int), intcompare); qsort suggests that quicksort is the algorithm implemented in this library function, although this is usually irrelevant to the user. 4.3 Heapsort: Fast Sorting via Data Structures Sorting is a natural laboratory for studying algorithm design paradigms, since many useful techniques lead to interesting sorting algorithms. The next several sections will introduce algorithmic design techniques motivated by particular sorting algorithms. 4.3 HEAPSORT: FAST SORTING VIA DATA STRUCTURES The alert reader should ask why we review the standard sorting when you are better oﬀ not implementing them and using built-in library functions instead. The answer is that the design techniques are very important for other algorithmic problems you are likely to encounter. We start with data structure design, because one of the most dramatic algorithmic improvements via appropriate data structures occurs in sorting. Selection sort is a simple-to-code algorithm that repeatedly extracts the smallest remaining element from the unsorted part of the set: SelectionSort(A) For i = 1 to n do Sort[i] = Find-Minimum from A Delete-Minimum from A Return(Sort) A C language implementation of selection sort appeared back in Section 2.5.1 (page 41). There we partitioned the input array into sorted and unsorted regions. To ﬁnd the smallest item, we performed a linear sweep through the unsorted portion of the array. The smallest item is then swapped with the ith item in the array before moving on to the next iteration. Selection sort performs n iterations, where the average iteration takes n/2 steps, for a total of O(n2 ) time. But what if we improve the data structure? It takes O(1) time to remove a particular item from an unsorted array once it has been located, but O(n) time to ﬁnd the smallest item. These are exactly the operations supported by priority queues. So what happens if we replace the data structure with a better priority queue implementation, either a heap or a balanced binary tree? Operations within the loop now take O(log n) time each, instead of O(n). Using such a priority queue implementation speeds up selection sort from O(n2 ) to O(n log n). The name typically given to this algorithm, heapsort, obscures the relationship between them, but heapsort is nothing but an implementation of selection sort using the right data structure. 4.3.1 Heaps Heaps are a simple and elegant data structure for eﬃciently supporting the priority queue operations insert and extract-min. They work by maintaining a partial order on the set of elements which is weaker than the sorted order (so it can be eﬃcient to maintain) yet stronger than random order (so the minimum element can be quickly identiﬁed). Power in any hierarchically-structured organization is reﬂected by a tree, where each node in the tree represents a person, and edge (x, y) implies that x directly supervises (or dominates) y. The fellow at the root sits at the “top of the heap.” In this spirit, a heap-labeled tree is deﬁned to be a binary tree such that the key labeling of each node dominates the key labeling of each of its children. In a 109 110 4. SORTING AND SEARCHING 1492 1783 1804 1918 1776 1865 2001 1945 1941 1963 1 1492 2 1783 3 1776 4 1804 5 1865 6 1945 7 1963 8 1918 9 2001 10 1941 Figure 4.2: A heap-labeled tree of important years from American history (l), with the corresponding implicit heap representation (r) min-heap, a node dominates its children by containing a smaller key than they do, while in a max-heap parent nodes dominate by being bigger. Figure 4.2(l) presents a min-heap ordered tree of red-letter years in American history (kudos to you if you can recall what happened each year). The most natural implementation of this binary tree would store each key in a node with pointers to its two children. As with binary search trees, the memory used by the pointers can easily outweigh the size of keys, which is the data we are really interested in. The heap is a slick data structure that enables us to represent binary trees without using any pointers. We will store data as an array of keys, and use the position of the keys to implicitly satisfy the role of the pointers. We will store the root of the tree in the ﬁrst position of the array, and its left and right children in the second and third positions, respectively. In general, we will store the 2l keys of the lth level of a complete binary tree from left-to-right in positions 2l−1 to 2l − 1, as shown in Figure 4.2(r). We assume that the array starts with index 1 to simplify matters. typedef struct { item_type q[PQ_SIZE+1]; int n; } priority_queue; /* body of queue */ /* number of queue elements */ What is especially nice about this representation is that the positions of the parent and children of the key at position k are readily determined. The left child of k sits in position 2k and the right child in 2k + 1, while the parent of k holds court in position n/2 . Thus we can move around the tree without any pointers. 4.3 HEAPSORT: FAST SORTING VIA DATA STRUCTURES pq_parent(int n) { if (n == 1) return(-1); else return((int) n/2); } /* implicitly take floor(n/2) */ pq_young_child(int n) { return(2 * n); } So, we can store any binary tree in an array without pointers. What is the catch? Suppose our height h tree was sparse, meaning that the number of nodes n < 2h . All missing internal nodes still take up space in our structure, since we must represent a full binary tree to maintain the positional mapping between parents and children. Space eﬃciency thus demands that we not allow holes in our tree—i.e., that each level be packed as much as it can be. If so, only the last level may be incomplete. By packing the elements of the last level as far to the left as possible, we can represent an n-key tree using exactly n elements of the array. If we did not enforce these structural constraints, we might need an array of size 2n to store the same elements. Since all but the last level is always ﬁlled, the height h of an n element heap is logarithmic because: h 2i = 2h+1 − 1 ≥ n i=0 so h = lg n . This implicit representation of binary trees saves memory, but is less ﬂexible than using pointers. We cannot store arbitrary tree topologies without wasting large amounts of space. We cannot move subtrees around by just changing a single pointer, only by explicitly moving each of the elements in the subtree. This loss of ﬂexibility explains why we cannot use this idea to represent binary search trees, but it works just ﬁne for heaps. Stop and Think: Who’s where in the heap? Problem: How can we eﬃciently search for a particular key in a heap? Solution: We can’t. Binary search does not work because a heap is not a binary search tree. We know almost nothing about the relative order of the n/2 leaf elements in a heap—certainly nothing that lets us avoid doing linear search through them. 111 112 4. SORTING AND SEARCHING 4.3.2 Constructing Heaps Heaps can be constructed incrementally, by inserting each new element into the left-most open spot in the array, namely the (n + 1)st position of a previously n-element heap. This ensures the desired balanced shape of the heap-labeled tree, but does not necessarily maintain the dominance ordering of the keys. The new key might be less than its parent in a min-heap, or greater than its parent in a max-heap. The solution is to swap any such dissatisﬁed element with its parent. The old parent is now happy, because it is properly dominated. The other child of the old parent is still happy, because it is now dominated by an element even more extreme than its previous parent. The new element is now happier, but may still dominate its new parent. We now recur at a higher level, bubbling up the new key to its proper position in the hierarchy. Since we replace the root of a subtree by a larger one at each step, we preserve the heap order elsewhere. pq_insert(priority_queue *q, item_type x) { if (q->n >= PQ_SIZE) printf("Warning: priority queue overflow insert x=%d\n",x); else { q->n = (q->n) + 1; q->q[ q->n ] = x; bubble_up(q, q->n); } } bubble_up(priority_queue *q, int p) { if (pq_parent(p) == -1) return; /* at root of heap, no parent */ if (q->q[pq_parent(p)] > q->q[p]) { pq_swap(q,p,pq_parent(p)); bubble_up(q, pq_parent(p)); } } This swap process takes constant time at each level. Since the height of an nelement heap is lg n , each insertion takes at most O(log n) time. Thus an initial heap of n elements can be constructed in O(n log n) time through n such insertions: pq_init(priority_queue *q) { q->n = 0; } 4.3 HEAPSORT: FAST SORTING VIA DATA STRUCTURES make_heap(priority_queue *q, item_type s[], int n) { int i; /* counter */ pq_init(q); for (i=0; i n <= 0) printf("Warning: empty priority queue.\n"); else { min = q->q[1]; q->q[1] = q->q[ q->n ]; q->n = q->n - 1; bubble_down(q,1); } return(min); } 113 114 4. SORTING AND SEARCHING bubble_down(priority_queue *q, int p) { int c; /* child index */ int i; /* counter */ int min_index; /* index of lightest child */ c = pq_young_child(p); min_index = p; for (i=0; i<=1; i++) if ((c+i) <= q->n) { if (q->q[min_index] > q->q[c+i]) min_index = c+i; } if (min_index != p) { pq_swap(q,p,min_index); bubble_down(q, min_index); } } We will reach a leaf after lg n bubble down steps, each constant time. Thus root deletion is completed in O(log n) time. Exchanging the maximum element with the last element and calling heapify repeatedly gives an O(n log n) sorting algorithm, named Heapsort. heapsort(item_type s[], int n) { int i; priority_queue q; /* counters */ /* heap for heapsort */ make_heap(&q,s,n); for (i=0; i n = n; for (i=0; i q[i+1] = s[i]; for (i=q->n; i>=1; i--) bubble_down(q,i); } Multiplying the number of calls to bubble down (n) times an upper bound on the cost of each operation (O(log n)) gives us a running time analysis of O(n log n). This would make it no faster than the incremental insertion algorithm described above. But note that it is indeed an upper bound, because only the last insertion will actually take lg n steps. Recall that bubble down takes time proportional to the height of the heaps it is merging. Most of these heaps are extremely small. In a full binary tree on n nodes, there are n/2 nodes that are leaves (i.e. , height 0), n/4 115 116 4. SORTING AND SEARCHING nodes that are height 1, n/8 nodes that are height 2, and so on. In general, there are at most n/2h+1 nodes of height h, so the cost of building a heap is: lg n h=0 lg n n/2 h+1 h ≤ n h/2h ≤ 2n h=0 Since this sum is not quite a geometric series, we can’t apply the usual identity to get the sum, but rest assured that the puny contribution of the numerator (h) is crushed by the denominator (2h ). The series quickly converges to linear. Does it matter that we can construct heaps in linear time instead of O(n log n)? Usually not. The construction time did not dominate the complexity of heapsort, so improving the construction time does not improve its worst-case performance. Still, it is an impressive display of the power of careful analysis, and the free lunch that geometric series convergence can sometimes provide. Stop and Think: Where in the Heap? Problem: Given an array-based heap on n elements and a real number x, eﬃciently determine whether the kth smallest element in the heap is greater than or equal to x. Your algorithm should be O(k) in the worst-case, independent of the size of the heap. Hint: you do not have to ﬁnd the kth smallest element; you need only determine its relationship to x. Solution: There are at least two diﬀerent ideas that lead to correct but ineﬃcient algorithms for this problem: 1. Call extract-min k times, and test whether all of these are less than x. This explicitly sorts the ﬁrst k elements and so gives us more information than the desired answer, but it takes O(k log n) time to do so. 2. The kth smallest element cannot be deeper than the kth level of the heap, since the path from it to the root must go through elements of decreasing value. Thus we can look at all the elements on the ﬁrst k levels of the heap, and count how many of them are less than x, stopping when we either ﬁnd k of them or run out of elements. This is correct, but takes O(min(n, 2k )) time, since the top k elements have 2k elements. An O(k) solution can look at only k elements smaller than x, plus at most O(k) elements greater than x. Consider the following recursive procedure, called at the root with i = 1 with count = k: 4.3 HEAPSORT: FAST SORTING VIA DATA STRUCTURES int heap_compare(priority_queue *q, int i, int count, int x) { if ((count <= 0) || (i > q->n) return(count); if (q->q[i] < x) { count = heap_compare(q, pq_young_child(i), count-1, x); count = heap_compare(q, pq_young_child(i)+1, count, x); } return(count); } If the root of the min-heap is ≥ x, then no elements in the heap can be less than x, as by deﬁnition the root must be the smallest element. This procedure searches the children of all nodes of weight smaller than x until either (a) we have found k of them, when it returns 0, or (b) they are exhausted, when it returns a value greater than zero. Thus it will ﬁnd enough small elements if they exist. But how long does it take? The only nodes whose children we look at are those < x, and at most k of these in total. Each have at most visited two children, so we visit at most 3k nodes, for a total time of O(k). 4.3.5 Sorting by Incremental Insertion Now consider a diﬀerent approach to sorting via eﬃcient data structures. Select an arbitrary element from the unsorted set, and put it in the proper position in the sorted set. InsertionSort(A) A[0] = −∞ for i = 2 to n do j=i while (A[j] < A[j − 1]) do swap(A[j], A[j − 1]) j =j−1 A C language implementation of insertion sort appeared in Section 2.5.2 (page 43). Although insertion sort takes O(n2 ) in the worst case, it performs considerably better if the data is almost sorted, since few iterations of the inner loop suﬃce to sift it into the proper position. Insertion sort is perhaps the simplest example of the incremental insertion technique, where we build up a complicated structure on n items by ﬁrst building it on n−1 items and then making the necessary changes to add the last item. Incremental insertion proves a particularly useful technique in geometric algorithms. 117 118 4. SORTING AND SEARCHING Note that faster sorting algorithms based on incremental insertion follow from more eﬃcient data structures. Insertion into a balanced search tree takes O(log n) per operation, or a total of O(n log n) to construct the tree. An in-order traversal reads through the elements in sorted order to complete the job in linear time. 4.4 War Story: Give me a Ticket on an Airplane I came into this particular job seeking justice. I’d been retained by an air travel company to help design an algorithm to ﬁnd the cheapest available airfare from city x to city y. Like most of you, I suspect, I’d been baﬄed at the crazy price ﬂuctuations of ticket prices under modern “yield management.” The price of ﬂights seems to soar far more eﬃciently than the planes themselves. The problem, it seemed to me, was that airlines never wanted to show the true cheapest price. By doing my job right, I could make damned sure they would show it to me next time. “Look,” I said at the start of the ﬁrst meeting. “This can’t be so hard. Construct a graph with vertices corresponding to airports, and add an edge between each airport pair (u, v) which shows a direct ﬂight from u to v. Set the weight of this edge equal to the cost of the cheapest available ticket from u to v. Now the cheapest fair from x to y is given by the shortest x-y path in this graph. This path/fare can be found using Dijkstra’s shortest path algorithm. Problem solved!” I announced, waiving my hand with a ﬂourish. The assembled cast of the meeting nodded thoughtfully, then burst out laughing. It was I who needed to learn something about the overwhelming complexity of air travel pricing. There are literally millions of diﬀerent fares available at any time, with prices changing several times daily. Restrictions on the availability of a particular fare in a particular context is enforced by a complicated set of pricing rules. These rules are an industry-wide kludge—a complicated structure with little in the way of consistent logical principles, which is exactly what we would need to search eﬃciently for the minimum fare. My favorite rule exceptions applied only to the country of Malawi. With a population of only 12 million and per-capita income of $596 (179th in the world), they prove to be an unexpected powerhouse shaping world aviation price policy. Accurately pricing any air itinerary requires at least implicit checks to ensure the trip doesn’t take us through Malawi. Part of the real problem is that there can easily be 100 diﬀerent fares for the ﬁrst ﬂight leg, say from Los Angeles (LAX) to Chicago (ORD), and a similar number for each subsequent leg, say from Chicago to New York (JFK). The cheapest possible LAX-ORD fare (maybe an AARP children’s price) might not be combinable with the cheapest ORD-JFK fare (perhaps a pre-Ramadan special that can only be used with subsequent connections to Mecca). After being properly chastised for oversimplifying the problem, I got down to work. I started by reducing the problem to the simplest interesting case. “So, you 4.4 WAR STORY: GIVE ME A TICKET ON AN AIRPLANE 119 X+Y X Y $100 $50 $110 $75 $130 $125 $150 (1,1) $160 (2,1) $175 (1,2) $180 (3,1) $185 (2,2) $205 (2,3) $225 (1,3) $235 (2,3) $255 (3,3) Figure 4.3: Sorting the pairwise sums of lists X and Y . need to ﬁnd the cheapest two-hop fare that passes your rule tests. Is there a way to decide in advance which pairs will pass without explicitly testing them?” “No, there is no way to tell,” they assured me. “We can only consult a black box routine to decide whether a particular price is available for the given itinerary/travelers.” “So our goal is to call this black box on the fewest number of combinations. This means evaluating all possible fare combinations in order from cheapest to most expensive, and stopping as soon as we encounter the ﬁrst legal combination.” “Right.” “Why not construct the m × n possible price pairs, sort them in terms of cost, and evaluate them in sorted order? Clearly this can be done in O(nm log(nm)) time.”1 “That is basically what we do now, but it is quite expensive to construct the full set of m × n pairs, since the ﬁrst one might be all we need.” I caught a whiﬀ of an interesting problem. “So what you really want is an eﬃcient data structure to repeatedly return the next most expensive pair without constructing all the pairs in advance.” This was indeed an interesting problem. Finding the largest element in a set under insertion and deletion is exactly what priority queues are good for. The catch here is that we could not seed the priority queue with all values in advance. We had to insert new pairs into the queue after each evaluation. I constructed some examples, like the one in Figure 4.3. We could represent each fare by the list indexes of its two components. The cheapest single fare will certainly be constructed by adding up the cheapest component from both lists, 1 The question of whether all such sums can be sorted faster than nm arbitrary integers is a notorious open problem in algorithm theory. See [Fre76, Lam92] for more on X + Y sorting, as the problem is known. 120 4. SORTING AND SEARCHING described (1, 1). The second cheapest fare would be made from the head of one list and the second element of another, and hence would be either (1, 2) or (2, 1). Then it gets more complicated. The third cheapest could either be the unused pair above or (1, 3) or (3, 1). Indeed it would have been (3, 1) in the example above if the third fare of X had been $120. “Tell me,” I asked. “Do we have time to sort the two respective lists of fares in increasing order?” “Don’t have to.” the leader replied. “They come out in sorted order from the database.” Good news. That meant there was a natural order to the pair values. We never need to evaluate the pairs (i + 1, j) or (i, j + 1) before (i, j), because they clearly deﬁne more expensive fares. “Got it!,” I said. “We will keep track of index pairs in a priority queue, with the sum of the fare costs as the key for the pair. Initially we put only pair (1, 1) on the queue. If it proves it is not feasible, we put its two successors on—namely (1, 2) and (2, 1). In general, we enqueue pairs (i + 1, j) and (i, j + 1) after evaluating/rejecting pair (i, j). We will get through all the pairs in the right order if we do so.” The gang caught on quickly. “Sure. But what about duplicates? We will construct pair (x, y) two diﬀerent ways, both when expanding (x−1, y) and (x, y −1).” “You are right. We need an extra data structure to guard against duplicates. The simplest might be a hash table to tell us whether a given pair exists in the priority queue before we insert a duplicate. In fact, we will never have more than n active pairs in our data structure, since there can only be one pair for each distinct value of the ﬁrst coordinate.” And so it went. Our approach naturally generalizes to itineraries with more than two legs, (a complexity which grows with the number of legs). The bestﬁrst evaluation inherent in our priority queue enabled the system to stop as soon as it found the provably cheapest fare. This proved to be fast enough to provide interactive response to the user. That said, I haven’t noticed my travel tickets getting any cheaper. 4.5 Mergesort: Sorting by Divide-and-Conquer Recursive algorithms reduce large problems into smaller ones. A recursive approach to sorting involves partitioning the elements into two groups, sorting each of the smaller problems recursively, and then interleaving the two sorted lists to totally order the elements. This algorithm is called mergesort, recognizing the importance of the interleaving operation: Mergesort(A[1, n]) Merge( MergeSort(A[1, n/2 ]), MergeSort(A[ n/2 + 1, n]) ) The basis case of the recursion occurs when the subarray to be sorted consists of a single element, so no rearrangement is possible. A trace of the execution of 4.5 MERGESORT: SORTING BY DIVIDE-AND-CONQUER M E R G E S O R T M E R G E M E R S O R T G E S O R T M E M E R G E S O R T E M E M R E G E E G M R O S R T O R S T E E G M O R R S T Figure 4.4: Animation of mergesort in action mergesort is given in Figure 4.4. Picture the action as it happens during an inorder traversal of the top tree, with the array-state transformations reported in the bottom, reﬂected tree. The eﬃciency of mergesort depends upon how eﬃciently we combine the two sorted halves into a single sorted list. We could concatenate them into one list and call heapsort or some other sorting algorithm to do it, but that would just destroy all the work spent sorting our component lists. Instead we can merge the two lists together. Observe that the smallest overall item in two lists sorted in increasing order (as above) must sit at the top of one of the two lists. This smallest element can be removed, leaving two sorted lists behind—one slightly shorter than before. The second smallest item overall must be atop one of these lists. Repeating this operation until both lists are empty merges two sorted lists (with a total of n elements between them) into one, using at most n − 1 comparisons or O(n) total work. What is the total running time of mergesort? It helps to think about how much work is done at each level of the execution tree. If we assume for simplicity that n is a power of two, the kth level consists of all the 2k calls to mergesort processing subranges of n/2k elements. The work done on the (k = 0)th level involves merging two sorted lists, each of size n/2, for a total of at most n − 1 comparisons. The work done on the (k = 1)th level involves merging two pairs of sorted lists, each of size n/4, for a total of at most n−2 comparisons. In general, the work done on the kth level involves merging 2k pairs sorted list, each of size n/2k+1 , for a total of at most n − 2k comparisons. Linear work is done merging all the elements on each level. Each of the n elements 121 122 4. SORTING AND SEARCHING appears in exactly one subproblem on each level. The most expensive case (in terms of comparsions) is actually the top level. The number of elements in a subproblem gets halved at each level. Thus the number of times we can halve n until we get to 1 is lg2 n. Because the recursion goes lg n levels deep, and a linear amount of work is done per level, mergesort takes O(n log n) time in the worst case. Mergesort is a great algorithm for sorting linked lists, because it does not rely on random access to elements as does heapsort or quicksort. Its primary disadvantage is the need for an auxilliary buﬀer when sorting arrays. It is easy to merge two sorted linked lists without using any extra space, by just rearranging the pointers. However, to merge two sorted arrays (or portions of an array), we need use a third array to store the result of the merge to avoid stepping on the component arrays. Consider merging {4, 5, 6} with {1, 2, 3}, packed from left to right in a single array. Without a buﬀer, we would overwrite the elements of the top half during merging and lose them. Mergesort is a classic divide-and-conquer algorithm. We are ahead of the game whenever we can break one large problem into two smaller problems, because the smaller problems are easier to solve. The trick is taking advantage of the two partial solutions to construct a solution of the full problem, as we did with the merge operation. Implementation The divide-and-conquer mergesort routine follows naturally from the pseudocode: mergesort(item_type s[], int low, int high) { int i; /* counter */ int middle; /* index of middle element */ if (low < high) { middle = (low+high)/2; mergesort(s,low,middle); mergesort(s,middle+1,high); merge(s, low, middle, high); } } More challenging turns out to be the details of how the merging is done. The problem is that we must put our merged array somewhere. To avoid losing an element by overwriting it in the course of the merge, we ﬁrst copy each subarray to a separate queue and merge these elements back into the array. In particular: 4.6 QUICKSORT: SORTING BY RANDOMIZATION merge(item_type s[], int low, int middle, int high) { int i; /* counter */ queue buffer1, buffer2; /* buffers to hold elements for merging */ init_queue(&buffer1); init_queue(&buffer2); for (i=low; i<=middle; i++) enqueue(&buffer1,s[i]); for (i=middle+1; i<=high; i++) enqueue(&buffer2,s[i]); i = low; while (!(empty_queue(&buffer1) || empty_queue(&buffer2))) { if (headq(&buffer1) <= headq(&buffer2)) s[i++] = dequeue(&buffer1); else s[i++] = dequeue(&buffer2); } while (!empty_queue(&buffer1)) s[i++] = dequeue(&buffer1); while (!empty_queue(&buffer2)) s[i++] = dequeue(&buffer2); } 4.6 Quicksort: Sorting by Randomization Suppose we select a random item p from the n items we seek to sort. Quicksort (shown in action in Figure 4.5) separates the n − 1 other items into two piles: a low pile containing all the elements that appear before p in sorted order and a high pile containing all the elements that appear after p in sorted order. Low and high denote the array positions we place the respective piles, leaving a single slot between them for p. Such partitioning buys us two things. First, the pivot element p ends up in the exact array position it will reside in the the ﬁnal sorted order. Second, after partitioning no element ﬂops to the other side in the ﬁnal sorted order. Thus we can now sort the elements to the left and the right of the pivot independently! This gives us a recursive sorting algorithm, since we can use the partitioning approach to sort each subproblem. The algorithm must be correct since each element ultimately ends up in the proper position: 123 124 4. SORTING AND SEARCHING Q U I C K S O R T Q I C K S O R T U Q I C K O R S T U I C K O Q R S T U I C K O Q R S T U C I K O Q R S T U Figure 4.5: Animation of quicksort in action quicksort(item_type s[], int l, int h) { int p; /* index of partition */ if ((h-l)>0) { p = partition(s,l,h); quicksort(s,l,p-1); quicksort(s,p+1,h); } } We can partition the array in one linear scan for a particular pivot element by maintaining three sections of the array: less than the pivot (to the left of firsthigh), greater than or equal to the pivot (between firsthigh and i), and unexplored (to the right of i), as implemented below: int partition(item_type s[], int l, int h) { int i; /* counter */ int p; /* pivot element index */ int firsthigh; /* divider position for pivot element */ p = h; firsthigh = l; for (i=l; i high) return (-1); middle = (low+high)/2; 3 It is actually quite interesting serving as an expert witness. /* key not found */ 4.9 BINARY SEARCH AND RELATED ALGORITHMS if (s[middle] == key) return(middle); if (s[middle] > key) return( binary_search(s,key,low,middle-1) ); else return(binary_search(s,key,middle+1,high) ); } This much you probably know. What is important is to have a sense of just how fast binary search is. Twenty questions is a popular children’s game where one player selects a word and the other repeatedly asks true/false questions in an attempt to guess it. If the word remains unidentiﬁed after 20 questions, the ﬁrst party wins; otherwise, the second player takes the honors. In fact, the second player always has a winning strategy, based on binary search. Given a printed dictionary, the player opens it in the middle, selects a word (say “move”), and asks whether the unknown word is before “move” in alphabetical order. Since standard dictionaries contain 50,000 to 200,000 words, we can be certain that the process will terminate within twenty questions. 4.9.1 Counting Occurrences Several interesting algorithms follow from simple variants of binary search. Suppose that we want to count the number of times a given key k (say “Skiena”) occurs in a given sorted array. Because sorting groups all the copies of k into a contiguous block, the problem reduces to ﬁnding the right block and then measures its size. The binary search routine presented above enables us to ﬁnd the index of an element of the correct block (x) in O(lg n) time. The natural way to identify the boundaries of the block is to sequentially test elements to the left of x until we ﬁnd the ﬁrst one that diﬀers from the search key, and then repeat this search to the right of x. The diﬀerence between the indices of the left and right boundaries (plus one) gives the count of the number of occurrences of k. This algorithm runs in O(lg n + s), where s is the number of occurrences of the key. This can be as bad as linear if the entire array consists of identical keys. A faster algorithm results by modifying binary search to search for the boundary of the block containing k, instead of k itself. Suppose we delete the equality test if (s[middle] == key) return(middle); from the implementation above and return the index low instead of −1 on each unsuccessful search. All searches will now be unsuccessful, since there is no equality test. The search will proceed to the right half whenever the key is compared to an identical array element, eventually terminating at the right boundary. Repeating the search after reversing the direction of the binary comparison will lead us to the left boundary. Each search takes O(lg n) time, so we can count the occurrences in logarithmic time regardless of the size of the block. 133 134 4. SORTING AND SEARCHING 4.9.2 One-Sided Binary Search Now suppose we have an array A consisting of a run of 0’s, followed by an unbounded run of 1’s, and would like to identify the exact point of transition between them. Binary search on the array would provide the transition point in lg n tests, if we had a bound n on the number of elements in the array. In the absence of such a bound, we can test repeatedly at larger intervals (A[1], A[2], A[4], A[8], A[16], . . .) until we ﬁnd a ﬁrst nonzero value. Now we have a window containing the target and can proceed with binary search. This one-sided binary search ﬁnds the transition point p using at most 2lg p comparisons, regardless of how large the array actually is. One-sided binary search is most useful whenever we are looking for a key that lies close to our current position. 4.9.3 Square and Other Roots The square root of n is the number r such that r2 = n. Square root computations are performed inside every pocket calculator, but it is instructive to develop an eﬃcient algorithm to compute them. First, observe that the square root of n ≥ 1 must be at least 1 and at most n. Let l = 1 and r = n. Consider the midpoint of this interval, m = (l + r)/2. How does m2 compare to n? If n ≥ m2 , then the square root must be greater than m, so the algorithm repeats with l = m. If n < m2 , then the square root must be less than m, so the algorithm repeats with r = m. Either way, we have halved the interval using only one comparison. Therefore, after lg n rounds we will have identiﬁed the square root to within ±1. This bisection method, as it is called in numerical analysis, can also be applied to the more general problem of ﬁnding the roots of an equation. We say that x is a root of the function f if f (x) = 0. Suppose that we start with values l and r such that f (l) > 0 and f (r) < 0. If f is a continuous function, there must exist a root between l and r. Depending upon the sign of f (m), where m = (l + r)/2, we can cut this window containing the root in half with each test and stop soon as our estimate becomes suﬃciently accurate. Root-ﬁnding algorithms that converge faster than binary search are known for both of these problems. Instead of always testing the midpoint of the interval, these algorithms interpolate to ﬁnd a test point closer to the actual root. Still, binary search is simple, robust, and works as well as possible without additional information on the nature of the function to be computed. Take-Home Lesson: Binary search and its variants are the quintessential divide-and-conquer algorithms. 4.10 4.10 DIVIDE-AND-CONQUER Divide-and-Conquer One of the most powerful techniques for solving problems is to break them down into smaller, more easily solved pieces. Smaller problems are less overwhelming, and they permit us to focus on details that are lost when we are studying the entire problem. A recursive algorithm starts to become apparent when we can break the problem into smaller instances of the same type of problem. Eﬀective parallel processing requires decomposing jobs into at least as many tasks as processors, and is becoming more important with the advent of cluster computing and multicore processors. Two important algorithm design paradigms are based on breaking problems down into smaller problems. In Chapter 8, we will see dynamic programming, which typically removes one element from the problem, solves the smaller problem, and then uses the solution to this smaller problem to add back the element in the proper way. Divide-and-conquer instead splits the problem in (say) halves, solves each half, then stitches the pieces back together to form a full solution. To use divide-and-conquer as an algorithm design technique, we must divide the problem into two smaller subproblems, solve each of them recursively, and then meld the two partial solutions into one solution to the full problem. Whenever the merging takes less time than solving the two subproblems, we get an eﬃcient algorithm. Mergesort, discussed in Section 4.5 (page 120), is the classic example of a divide-and-conquer algorithm. It takes only linear time to merge two sorted lists of n/2 elements, each of which was obtained in O(n lg n) time. Divide-and-conquer is a design technique with many important algorithms to its credit, including mergesort, the fast Fourier transform, and Strassen’s matrix multiplication algorithm. Beyond binary search and its many variants, however, I ﬁnd it to be a diﬃcult design technique to apply in practice. Our ability to analyze divide-and-conquer algorithms rests on our strength to solve the asymptotics of recurrence relations governing the cost of such recursive algorithms. 4.10.1 Recurrence Relations Many divide-and-conquer algorithms have time complexities that are naturally modeled by recurrence relations. Evaluating such recurrences is important to understanding when divide-and-conquer algorithms perform well, and provide an important tool for analysis in general. The reader who balks at the very idea of analysis is free to skip this section, but there are important insights into design that come from an understanding of the behavior of recurrence relations. What is a recurrence relation? It is an equation that is deﬁned in terms of itself. The Fibonacci numbers are described by the recurrence relation Fn = Fn−1 + Fn−2 and discussed in Section 8.1.1. Many other natural functions are easily expressed as recurrences. Any polynomial can be represented by a recurrence, such as the linear function: an = an−1 + 1, a1 = 1 −→ an = n 135 136 4. SORTING AND SEARCHING Any exponential can be represented by a recurrence: an = 2an−1 , a1 = 1 −→ an = 2n−1 Finally, lots of weird functions that cannot be described easily with conventional notation can be represented by a recurrence: an = nan−1 , a1 = 1 −→ an = n! This means that recurrence relations are a very versatile way to represent functions. The self-reference property of recurrence relations is shared with recursive programs or algorithms, as the shared roots of both terms reﬂect. Essentially, recurrence relations provide a way to analyze recursive structures, such as algorithms. 4.10.2 Divide-and-Conquer Recurrences Divide-and-conquer algorithms tend to break a given problem into some number of smaller pieces (say a), each of which is of size n/b. Further, they spend f (n) time to combine these subproblem solutions into a complete result. Let T (n) denote the worst-case time the algorithm takes to solve a problem of size n. Then T (n) is given by the following recurrence relation: T (n) = aT (n/b) + f (n) Consider the following examples: • Sorting – The running time behavior of mergesort is governed by the recurrence T (n) = 2T (n/2) + O(n), since the algorithm divides the data into equal-sized halves and then spends linear time merging the halves after they are sorted. In fact, this recurrence evaluates to T (n) = O(n lg n), just as we got by our previous analysis. • Binary Search – The running time behavior of binary search is governed by the recurrence T (n) = T (n/2) + O(1), since at each step we spend constant time to reduce the problem to an instance half its size. In fact, this recurrence evaluates to T (n) = O(lg n), just as we got by our previous analysis. • Fast Heap Construction – The bubble down method of heap construction (described in Section 4.3.4) built an n-element heap by constructing two n/2 element heaps and then merging them with the root in logarithmic time. This argument reduces to the recurrence relation T (n) = 2T (n/2) + O(lg n). In fact, this recurrence evaluates to T (n) = O(n), just as we got by our previous analysis. • Matrix Multiplication – As discussed in Section 2.5.4, the standard matrix multiplication algorithm for two n × n matrices takes O(n3 ), because we compute the dot product of n terms for each of the n2 elements in the product matrix. 4.10 DIVIDE-AND-CONQUER However, Strassen [Str69] discovered a divide-and-conquer algorithm that manipulates the products of seven n/2 × n/2 matrix products to yield the product of two n × n matrices. This yields a time-complexity recurrence T (n) = 7T (n/2) + O(n2 ). In fact, this recurrence evaluates to T (n) = O(n2.81 ), which seems impossible to predict without solving the recurrence. 4.10.3 Solving Divide-and-Conquer Recurrences (*) In fact, divide-and-conquer recurrences of the form T (n) = aT (n/b) + f (n) are generally easy to solve, because the solutions typically fall into one of three distinct cases: 1. If f (n) = O(nlogb a− ) for some constant > 0, then T (n) = Θ(nlogb a ). 2. If f (n) = Θ(nlogb a ), then T (n) = Θ(nlogb a lg n). 3. If f (n) = Ω(nlogb a+ ) for some constant > 0, and if af (n/b) ≤ cf (n) for some c < 1, then T (n) = Θ(f (n)). Although this looks somewhat frightening, it really isn’t diﬃcult to apply. The issue is identifying which case of the so-called master theorem holds for your given recurrence. Case 1 holds for heap construction and matrix multiplication, while Case 2 holds mergesort and binary search. Case 3 generally arises for clumsier algorithms, where the cost of combining the subproblems dominates everything. The master theorem can be thought of as a black-box piece of machinery, invoked as needed and left with its mystery intact. However, with a little study, the reason why the master theorem works can become apparent. Figure 4.9 shows the recursion tree associated with a typical T (n) = aT (n/b) + f (n) divide-and-conquer algorithm. Each problem of size n is decomposed into a problems of size n/b. Each subproblem of size k takes O(f (k)) time to deal with internally, between partitioning and merging. The total time for the algorithm is the sum of these internal costs, plus the overhead of building the recursion tree. The height of this tree is h = logb n and the number of leaf nodes ah = alogb n , which happens to simplify to nlogb a with some algebraic manipulation. The three cases of the master theorem correspond to three diﬀerent costs which might be dominant as a function of a, b, and f (n): • Case 1: Too many leaves – If the number of leaf nodes outweighs the sum of the internal evaluation cost, the total running time is O(nlogb a ). • Case 2: Equal work per level – As we move down the tree, each problem gets smaller but there are more of them to solve. If the sum of the internal evaluation costs at each level are equal, the total running time is the cost per level (nlogb a ) times the number of levels (logb n), for a total running time of O(nlogb a lg n). 137 138 4. SORTING AND SEARCHING partition size = n vertex degree = a partition size = n/b height = log n b partition size = n/b2 partition size = b partition size = 1 log n b width = a log b b = n Figure 4.9: The recursion tree resulting from decomposing each problem of size n into a problems of size n/b • Case 3: Too expensive a root – If the internal evaluation costs grow rapidly enough with n, then the cost of the root evaluation may dominate. If so, the the total running time is O(f (n)). Chapter Notes The most interesting sorting algorithms that have not been discussed in this section include shellsort, which is a substantially more eﬃcient version of insertion sort, and radix sort, an eﬃcient algorithm for sorting strings. You can learn more about these and every other sorting algorithm by browsing through Knuth [Knu98], with hundreds of pages of interesting material on sorting. This includes external sorting, the subject of this chapter’s legal war story. As implemented here, mergesort copies the merged elements into an auxiliary buﬀer to avoid overwriting the original elements to be sorted. Through clever but complicated buﬀer manipulation, mergesort can be implemented in an array without using much extra storage. Kronrod’s algorithm for in-place merging is presented in [Knu98]. Randomized algorithms are discussed in greater detail in the books by Motwani and Raghavan [MR95] and Mitzenmacher and Upfal [MU05]. The problem of 4.11 EXERCISES nut and bolt sorting was introduced by [Raw92]. A complicated but deterministic O(n log n) algorithm is due to Komlos, Ma, and Szemeredi [KMS96]. Several other algorithms texts provide more substantive coverage of divideand-conquer algorithms, including [CLRS01, KT06, Man89]. See [CLRS01] for an excellent overview of the master theorem. 4.11 Exercises Applications of Sorting 4-1. [3] The Grinch is given the job of partitioning 2n players into two teams of n players each. Each player has a numerical rating that measures how good he/she is at the game. He seeks to divide the players as unfairly as possible, so as to create the biggest possible talent imbalance between team A and team B. Show how the Grinch can do the job in O(n log n) time. 4-2. [3] For each of the following problems, give an algorithm that ﬁnds the desired numbers within the given amount of time. To keep your answers brief, feel free to use algorithms from the book as subroutines. For the example, S = {6, 13, 19, 3, 8}, 19 − 3 maximizes the diﬀerence, while 8 − 6 minimizes the diﬀerence. (a) Let S be an unsorted array of n integers. Give an algorithm that ﬁnds the pair x, y ∈ S that maximizes |x − y|. Your algorithm must run in O(n) worst-case time. (b) Let S be a sorted array of n integers. Give an algorithm that ﬁnds the pair x, y ∈ S that maximizes |x − y|. Your algorithm must run in O(1) worst-case time. (c) Let S be an unsorted array of n integers. Give an algorithm that ﬁnds the pair x, y ∈ S that minimizes |x − y|, for x = y. Your algorithm must run in O(n log n) worst-case time. (d) Let S be a sorted array of n integers. Give an algorithm that ﬁnds the pair x, y ∈ S that minimizes |x − y|, for x = y. Your algorithm must run in O(n) worst-case time. 4-3. [3] Take a sequence of 2n real numbers as input. Design an O(n log n) algorithm that partitions the numbers into n pairs, with the property that the partition minimizes the maximum sum of a pair. For example, say we are given the numbers (1,3,5,9). The possible partitions are ((1,3),(5,9)), ((1,5),(3,9)), and ((1,9),(3,5)). The pair sums for these partitions are (4,14), (6,12), and (10,8). Thus the third partition has 10 as its maximum sum, which is the minimum over the three partitions. 4-4. [3] Assume that we are given n pairs of items as input, where the ﬁrst item is a number and the second item is one of three colors (red, blue, or yellow). Further assume that the items are sorted by number. Give an O(n) algorithm to sort the items by color (all reds before all blues before all yellows) such that the numbers for identical colors stay sorted. For example: (1,blue), (3,red), (4,blue), (6,yellow), (9,red) should become (3,red), (9,red), (1,blue), (4,blue), (6,yellow). 4-5. [3] The mode of a set of numbers is the number that occurs most frequently in the set. The set (4, 6, 2, 4, 3, 1) has a mode of 4. Give an eﬃcient and correct algorithm to compute the mode of a set of n numbers. 139 140 4. SORTING AND SEARCHING 4-6. [3] Given two sets S1 and S2 (each of size n), and a number x, describe an O(n log n) algorithm for ﬁnding whether there exists a pair of elements, one from S1 and one from S2 , that add up to x. (For partial credit, give a Θ(n2 ) algorithm for this problem.) 4-7. [3] Outline a reasonable method of solving each of the following problems. Give the order of the worst-case complexity of your methods. (a) You are given a pile of thousands of telephone bills and thousands of checks sent in to pay the bills. Find out who did not pay. (b) You are given a list containing the title, author, call number and publisher of all the books in a school library and another list of 30 publishers. Find out how many of the books in the library were published by each company. (c) You are given all the book checkout cards used in the campus library during the past year, each of which contains the name of the person who took out the book. Determine how many distinct people checked out at least one book. 4-8. [4 ] Given a set of S containing n real numbers, and a real number x. We seek an algorithm to determine whether two elements of S exist whose sum is exactly x. (a) Assume that S is unsorted. Give an O(n log n) algorithm for the problem. (b) Assume that S is sorted. Give an O(n) algorithm for the problem. 4-9. [4 ] Give an eﬃcient algorithm to compute the union of sets A and B, where n = max(|A|, |B|). The output should be an array of distinct elements that form the union of the sets, such that they appear more than once in the union. (a) Assume that A and B are unsorted. Give an O(n log n) algorithm for the problem. (b) Assume that A and B are sorted. Give an O(n) algorithm for the problem. 4-10. [5] Given a set S of n integers and an integer T , give an O(nk−1 log n) algorithm to test whether k of the integers in S add up to T . 4-11. [6 ] Design an O(n) algorithm that, given a list of n elements, ﬁnds all the elements that appear more than n/2 times in the list. Then, design an O(n) algorithm that, given a list of n elements, ﬁnds all the elements that appear more than n/4 times. Heaps 4-12. [3] Devise an algorithm for ﬁnding the k smallest elements of an unsorted set of n integers in O(n + k log n). 4-13. [5] You wish to store a set of n numbers in either a max-heap or a sorted array. For each application below, state which data structure is better, or if it does not matter. Explain your answers. (a) Want to ﬁnd the maximum element quickly. (b) Want to be able to delete an element quickly. (c) Want to be able to form the structure quickly. (d) Want to ﬁnd the minimum element quickly. 4.11 EXERCISES 4-14. [5] Give an O(n log k)-time algorithm that merges k sorted lists with a total of n elements into one sorted list. (Hint: use a heap to speed up the elementary O(kn)time algorithm). 4-15. [5] (a) Give an eﬃcient algorithm to ﬁnd the second-largest key among n keys. You can do better than 2n − 3 comparisons. (b) Then, give an eﬃcient algorithm to ﬁnd the third-largest key among n keys. How many key comparisons does your algorithm do in the worst case? Must your algorithm determine which key is largest and second-largest in the process? Quicksort 4-16. [3] Use the partitioning idea of quicksort to give an algorithm that ﬁnds the median element of an array of n integers in expected O(n) time. (Hint: must you look at both sides of the partition?) 4-17. [3] The median of a set of n values is the n/2 th smallest value. (a) Suppose quicksort always pivoted on the median of the current sub-array. How many comparisons would Quicksort make then in the worst case? (b) Suppose quicksort were always to pivot on the n/3 th smallest value of the current sub-array. How many comparisons would be made then in the worst case? 4-18. [5] Suppose an array A consists of n elements, each of which is red, white, or blue. We seek to sort the elements so that all the reds come before all the whites, which come before all the blues The only operation permitted on the keys are • Examine(A,i) – report the color of the ith element of A. • Swap(A,i,j) – swap the ith element of A with the jth element. Find a correct and eﬃcient algorithm for red-white-blue sorting. There is a lineartime solution. 4-19. [5] An inversion of a permutation is a pair of elements that are out of order. (a) Show that a permutation of n items has at most n(n − 1)/2 inversions. Which permutation(s) have exactly n(n − 1)/2 inversions? (b) Let P be a permutation and P r be the reversal of this permutation. Show that P and P r have a total of exactly n(n − 1)/2 inversions. (c) Use the previous result to argue that the expected number of inversions in a random permutation is n(n − 1)/4. 4-20. [3] Give an eﬃcient algorithm to rearrange an array of n keys so that all the negative keys precede all the nonnegative keys. Your algorithm must be in-place, meaning you cannot allocate another array to temporarily hold the items. How fast is your algorithm? Other Sorting Algorithms 4-21. [5] Stable sorting algorithms leave equal-key items in the same relative order as in the original permutation. Explain what must be done to ensure that mergesort is a stable sorting algorithm. 141 142 4. SORTING AND SEARCHING 4-22. [3] Show that n positive integers in the range 1 to k can be sorted in O(n log k) time. The interesting case is when k << n. 4-23. [5] We seek to sort a sequence S of n integers with many duplications, such that the number of distinct integers in S is O(log n). Give an O(n log log n) worst-case time algorithm to sort such sequences. √ 4-24. [5] Let A[1..n] be an array such that the ﬁrst n − n elements are already sorted (though we know nothing about the remaining elements). Give an algorithm that sorts A in substantially better than n log n steps. 4-25. [5] Assume that the array A[1..n] only has numbers from {1, . . . , n2 } but that at most log log n of these numbers ever appear. Devise an algorithm that sorts A in substantially less than O(n log n). 4-26. [5] Consider the problem of sorting a sequence of n 0’s and 1’s using comparisons. For each comparison of two values x and y, the algorithm learns which of x < y, x = y, or x > y holds. (a) Give an algorithm to sort in n − 1 comparisons in the worst case. Show that your algorithm is optimal. (b) Give an algorithm to sort in 2n/3 comparisons in the average case (assuming each of the n inputs is 0 or 1 with equal probability). Show that your algorithm is optimal. 4-27. [6] Let P be a simple, but not necessarily convex, polygon and q an arbitrary point not necessarily in P . Design an eﬃcient algorithm to ﬁnd a line segment originating from q that intersects the maximum number of edges of P . In other words, if standing at point q, in what direction should you aim a gun so the bullet will go through the largest number of walls. A bullet through a vertex of P gets credit for only one wall. An O(n log n) algorithm is possible. Lower Bounds 4-28. [5] In one of my research papers√[Ski88], I discovered a comparison-based sorting algorithm that runs in O(n log( n)). Given the existence of an Ω(n log n) lower bound for sorting, how can this be possible? 4-29. [5] Mr. B. C. Dull claims to have developed a new data structure for priority queues that supports the operations Insert, Maximum, and Extract-Max—all in O(1) worstcase time. Prove that he is mistaken. (Hint: the argument does not involve a lot of gory details—just think about what this would imply about the Ω(n log n) lower bound for sorting.) Searching 4-30. [3] A company database consists of 10,000 sorted names, 40% of whom are known as good customers and who together account for 60% of the accesses to the database. There are two data structure options to consider for representing the database: • Put all the names in a single array and use binary search. • Put the good customers in one array and the rest of them in a second array. Only if we do not ﬁnd the query name on a binary search of the ﬁrst array do we do a binary search of the second array. 4.11 EXERCISES Demonstrate which option gives better expected performance. Does this change if linear search on an unsorted array is used instead of binary search for both options? 4-31. [3] Suppose you are given an array A of n sorted numbers that has been circularly shifted k positions to the right. For example, {35, 42, 5, 15, 27, 29} is a sorted array that has been circularly shifted k = 2 positions, while {27, 29, 35, 42, 5, 15} has been shifted k = 4 positions. • Suppose you know what k is. Give an O(1) algorithm to ﬁnd the largest number in A. • Suppose you do not know what k is. Give an O(lg n) algorithm to ﬁnd the largest number in A. For partial credit, you may give an O(n) algorithm. 4-32. [3] Consider the numerical 20 Questions game. In this game, Player 1 thinks of a number in the range 1 to n. Player 2 has to ﬁgure out this number by asking the fewest number of true/false questions. Assume that nobody cheats. (a) What is an optimal strategy if n in known? (b) What is a good strategy is n is not known? 4-33. [5] Suppose that you are given a sorted sequence of distinct integers {a1 , a2 , . . . , an }. Give an O(lg n) algorithm to determine whether there exists an i index such as ai = i. For example, in {−10, −3, 3, 5, 7}, a3 = 3. In {2, 3, 4, 5, 6, 7}, there is no such i. 4-34. [5] Suppose that you are given a sorted sequence of distinct integers {a1 , a2 , . . . , an }, drawn from 1 to m where n < m. Give an O(lg n) algorithm to ﬁnd an integer ≤ m that is not present in a. For full credit, ﬁnd the smallest such integer. 4-35. [5] Let M be an n × m integer matrix in which the entries of each row are sorted in increasing order (from left to right) and the entries in each column are in increasing order (from top to bottom). Give an eﬃcient algorithm to ﬁnd the position of an integer x in M , or to determine that x is not there. How many comparisons of x with matrix entries does your algorithm use in worst case? Implementation Challenges 4-36. [5] Consider an n × n array A containing integer elements (positive, negative, and zero). Assume that the elements in each row of A are in strictly increasing order, and the elements of each column of A are in strictly decreasing order. (Hence there cannot be two zeroes in the same row or the same column.) Describe an eﬃcient algorithm that counts the number of occurrences of the element 0 in A. Analyze its running time. 4-37. [6] Implement versions of several diﬀerent sorting algorithms, such as selection sort, insertion sort, heapsort, mergesort, and quicksort. Conduct experiments to assess the relative performance of these algorithms in a simple application that reads a large text ﬁle and reports exactly one instance of each word that appears within it. This application can be eﬃciently implemented by sorting all the words that occur in the text and then passing through the sorted sequence to identify one instance of each distinct word. Write a brief report with your conclusions. 143 144 4. SORTING AND SEARCHING 4-38. [5] Implement an external sort, which uses intermediate ﬁles to sort ﬁles bigger than main memory. Mergesort is a good algorithm to base such an implementation on. Test your program both on ﬁles with small records and on ﬁles with large records. 4-39. [8] Design and implement a parallel sorting algorithm that distributes data across several processors. An appropriate variation of mergesort is a likely candidate. Measure the speedup of this algorithm as the number of processors increases. Later, compare the execution time to that of a purely sequential mergesort implementation. What are your experiences? Interview Problems 4-40. [3] If you are given a million integers to sort, what algorithm would you use to sort them? How much time and memory would that consume? 4-41. [3] Describe advantages and disadvantages of the most popular sorting algorithms. 4-42. [3] Implement an algorithm that takes an input array and returns only the unique elements in it. 4-43. [5] You have a computer with only 2Mb of main memory. How do you use it to sort a large ﬁle of 500 Mb that is on disk? 4-44. [5] Design a stack that supports push, pop, and retrieving the minimum element in constant time. Can you do this? 4-45. [5] Given a search string of three words, ﬁnd the smallest snippet of the document that contains all three of the search words—i.e. , the snippet with smallest number of words in it. You are given the index positions where these words in occur search strings, such as word1: (1, 4, 5), word2: (4, 9, 10), and word3: (5, 6, 15). Each of the lists are in sorted order, as above. 4-46. [6] You are given 12 coins. One of them is heavier or lighter than the rest. Identify this coin in just three weighings. Programming Challenges These programming challenge problems with robot judging are available at http://www.programming-challenges.com or http://online-judge.uva.es. 4-1. “Vito’s Family” – Programming Challenges 110401, UVA Judge 10041. 4-2. “Stacks of Flapjacks” – Programming Challenges 110402, UVA Judge 120. 4-3. “Bridge” – Programming Challenges 110403, UVA Judge 10037. 4-4. “ShoeMaker’s Problem” – Programming Challenges 110405, UVA Judge 10026. 4-5. “ShellSort” – Programming Challenges 110407, UVA Judge 10152. 5 Graph Traversal Graphs are one of the unifying themes of computer science—an abstract representation that describes the organization of transportation systems, human interactions, and telecommunication networks. That so many diﬀerent structures can be modeled using a single formalism is a source of great power to the educated programmer. More precisely, a graph G = (V, E) consists of a set of vertices V together with a set E of vertex pairs or edges. Graphs are important because they can be used to represent essentially any relationship. For example, graphs can model a network of roads, with cities as vertices and roads between cities as edges, as shown in Figure 5.1. Electronic circuits can also be modeled as graphs, with junctions as vertices and components as edges. Stony Brook Green Port Orient Point Riverhead Shelter Island Montauk Islip Sag Harbor Figure 5.1: Modeling road networks and electronic circuits as graphs S.S. Skiena, The Algorithm Design Manual, 2nd ed., DOI: 10.1007/978-1-84800-070-4 5, c Springer-Verlag London Limited 2008 146 5. GRAPH TRAVERSAL The key to solving many algorithmic problems is to think of them in terms of graphs. Graph theory provides a language for talking about the properties of relationships, and it is amazing how often messy applied problems have a simple description and solution in terms of classical graph properties. Designing truly novel graph algorithms is a very diﬃcult task. The key to using graph algorithms eﬀectively in applications lies in correctly modeling your problem so you can take advantage of existing algorithms. Becoming familiar with many diﬀerent algorithmic graph problems is more important than understanding the details of particular graph algorithms, particularly since Part II of this book will point you to an implementation as soon as you know the name of your problem. Here we present basic data structures and traversal operations for graphs, which will enable you to cobble together solutions for basic graph problems. Chapter 6 will present more advanced graph algorithms that ﬁnd minimum spanning trees, shortest paths, and network ﬂows, but we stress the primary importance of correctly modeling your problem. Time spent browsing through the catalog now will leave you better informed of your options when a real job arises. 5.1 Flavors of Graphs A graph G = (V, E) is deﬁned on a set of vertices V , and contains a set of edges E of ordered or unordered pairs of vertices from V . In modeling a road network, the vertices may represent the cities or junctions, certain pairs of which are connected by roads/edges. In analyzing the source code of a computer program, the vertices may represent lines of code, with an edge connecting lines x and y if y is the next statement executed after x. In analyzing human interactions, the vertices typically represent people, with edges connecting pairs of related souls. Several fundamental properties of graphs impact the choice of the data structures used to represent them and algorithms available to analyze them. The ﬁrst step in any graph problem is determining the ﬂavors of graphs you are dealing with: • Undirected vs. Directed – A graph G = (V, E) is undirected if edge (x, y) ∈ E implies that (y, x) is also in E. If not, we say that the graph is directed. Road networks between cities are typically undirected, since any large road has lanes going in both directions. Street networks within cities are almost always directed, because there are at least a few one-way streets lurking somewhere. Program-ﬂow graphs are typically directed, because the execution ﬂows from one line into the next and changes direction only at branches. Most graphs of graph-theoretic interest are undirected. • Weighted vs. Unweighted – Each edge (or vertex) in a weighted graph G is assigned a numerical value, or weight. The edges of a road network graph might be weighted with their length, drive-time, or speed limit, depending upon the 5.1 147 FLAVORS OF GRAPHS 5 7 2 3 9 3 5 4 7 12 undirected directed weighted unweighted simple non−simple sparse dense cyclic acyclic embedded topological B A D F G explicit implicit unlabeled E C labeled Figure 5.2: Important properties / ﬂavors of graphs application. In unweighted graphs, there is no cost distinction between various edges and vertices. The diﬀerence between weighted and unweighted graphs becomes particularly apparent in ﬁnding the shortest path between two vertices. For unweighted graphs, the shortest path must have the fewest number of edges, and can be found using a breadth-ﬁrst search as discussed in this chapter. Shortest paths in weighted graphs requires more sophisticated algorithms, as discussed in Chapter 6. • Simple vs. Non-simple – Certain types of edges complicate the task of working with graphs. A self-loop is an edge (x, x) involving only one vertex. An edge (x, y) is a multiedge if it occurs more than once in the graph. Both of these structures require special care in implementing graph algorithms. Hence any graph that avoids them is called simple. 148 5. GRAPH TRAVERSAL • Sparse vs. Dense: Graphs are sparse when only a small fraction of the possible vertex pairs ( n2 for a simple, undirected graph on n vertices) actually have edges deﬁned between them. Graphs where a large fraction of the vertex pairs deﬁne edges are called dense. There is no oﬃcial boundary between what is called sparse and what is called dense, but typically dense graphs have a quadratic number of edges, while sparse graphs are linear in size. Sparse graphs are usually sparse for application-speciﬁc reasons. Road networks must be sparse graphs because of road junctions. The most ghastly intersection I’ve ever heard of was the endpoint of only seven diﬀerent roads. Junctions of electrical components are similarly limited to the number of wires that can meet at a point, perhaps except for power and ground. • Cyclic vs. Acyclic – An acyclic graph does not contain any cycles. Trees are connected, acyclic undirected graphs. Trees are the simplest interesting graphs, and are inherently recursive structures because cutting any edge leaves two smaller trees. Directed acyclic graphs are called DAGs. They arise naturally in scheduling problems, where a directed edge (x, y) indicates that activity x must occur before y. An operation called topological sorting orders the vertices of a DAG to respect these precedence constraints. Topological sorting is typically the ﬁrst step of any algorithm on a DAG, as will be discussed in Section 5.10.1 (page 179). • Embedded vs. Topological – A graph is embedded if the vertices and edges are assigned geometric positions. Thus, any drawing of a graph is an embedding, which may or may not have algorithmic signiﬁcance. Occasionally, the structure of a graph is completely deﬁned by the geometry of its embedding. For example, if we are given a collection of points in the plane, and seek the minimum cost tour visiting all of them (i.e., the traveling salesman problem), the underlying topology is the complete graph connecting each pair of vertices. The weights are typically deﬁned by the Euclidean distance between each pair of points. Grids of points are another example of topology from geometry. Many problems on an n × m grid involve walking between neighboring points, so the edges are implicitly deﬁned from the geometry. • Implicit vs. Explicit – Certain graphs are not explicitly constructed and then traversed, but built as we use them. A good example is in backtrack search. The vertices of this implicit search graph are the states of the search vector, while edges link pairs of states that can be directly generated from each other. Because you do not have to store the entire graph, it is often easier to work with an implicit graph than explicitly construct it prior to analysis. 5.1 Bill Clinton FLAVORS OF GRAPHS Hillary Clinton George Bush John McCain Saddam Hussein Figure 5.3: A portion of the friendship graph • Labeled vs. Unlabeled – Each vertex is assigned a unique name or identiﬁer in a labeled graph to distinguish it from all other vertices. In unlabeled graphs, no such distinctions have been made. Graphs arising in applications are often naturally and meaningfully labeled, such as city names in a transportation network. A common problem is that of isomorphism testing—determining whether the topological structure of two graphs are identical if we ignore any labels. Such problems are typically solved using backtracking, by trying to assign each vertex in each graph a label such that the structures are identical. 5.1.1 The Friendship Graph To demonstrate the importance of proper modeling, let us consider a graph where the vertices are people, and there is an edge between two people if and only if they are friends. Such graphs are called social networks and are well deﬁned on any set of people—be they the people in your neighborhood, at your school/business, or even spanning the entire world. An entire science analyzing social networks has sprung up in recent years, because many interesting aspects of people and their behavior are best understood as properties of this friendship graph. Most of the graphs that one encounters in real life are sparse. The friendship graph is good example. Even the most gregarious person on earth knows an insigniﬁcant fraction of the world’s population. We use this opportunity to demonstrate the graph theory terminology described above. “Talking the talk” proves to be an important part of “walking the walk”: • If I am your friend, does that mean you are my friend? – This question really asks whether the graph is directed. A graph is undirected if edge (x, y) always implies (y, x). Otherwise, the graph is said to be directed. The “heardof” graph is directed, since I have heard of many famous people who have never heard of me! The “had-sex-with” graph is presumably undirected, since the critical operation always requires a partner. I’d like to think that the “friendship” graph is also an undirected graph. 149 150 5. GRAPH TRAVERSAL • How close a friend are you? – In weighted graphs, each edge has an associated numerical attribute. We could model the strength of a friendship by associating each edge with an appropriate value, perhaps from -10 (enemies) to 10 (blood brothers). The edges of a road network graph might be weighted with their length, drive-time, or speed limit, depending upon the application. A graph is said to be unweighted if all edges are assumed to be of equal weight. • Am I my own friend? – This question addresses whether the graph is simple, meaning it contains no loops and no multiple edges. An edge of the form (x, x) is said to be a loop. Sometimes people are friends in several diﬀerent ways. Perhaps x and y were college classmates and now work together at the same company. We can model such relationships using multiedges—multiple edges (x, y) perhaps distinguished by diﬀerent labels. Simple graphs really are often simpler to work with in practice. Therefore, we might be better oﬀ declaring that no one is their own friend. • Who has the most friends? – The degree of a vertex is the number of edges adjacent to it. The most popular person deﬁnes the vertex of highest degree in the friendship graph. Remote hermits are associated with degree-zero vertices. In dense graphs, most vertices have high degrees, as opposed to sparse graphs with relatively few edges. In a regular graph, each vertex has exactly the same degree. A regular friendship graph is truly the ultimate in social-ism. • Do my friends live near me? – Social networks are not divorced from geography. Many of your friends are your friends only because they happen to live near you (e.g., neighbors) or used to live near you (e.g., college roommates). Thus, a full understanding of social networks requires an embedded graph, where each vertex is associated with the point on this world where they live. This geographic information may not be explicitly encoded, but the fact that the graph is inherently embedded in the plane shapes our interpretation of any analysis. • Oh, you also know her? – Social networking services such as Myspace and LinkedIn are built on the premise of explicitly deﬁning the links between members and their member-friends. Such graphs consist of directed edges from person/vertex x professing his friendship to person/vertex y. That said, the complete friendship graph of the world is represented implicitly. Each person knows who their friends are, but cannot ﬁnd out about other people’s friendships except by asking them. The “six degrees of separation” theory argues that there is a short path linking every two people in the world (e.g., Skiena and the President) but oﬀers us no help in actually ﬁnding this path. The shortest such path I know of contains three hops (Steven Skiena → Bob McGrath → John Marberger → George W. Bush), but there could 5.2 1 1 2 3 0 1 0 4 0 5 1 2 1 0 1 1 1 3 0 1 0 1 0 4 0 1 1 0 1 5 1 1 0 1 0 1 2 3 5 4 1 2 5 2 1 5 3 2 4 4 2 5 3 5 4 1 2 3 Figure 5.4: The adjacency matrix and adjacency list of a given graph be a shorter one (say, if he went to college with my dentist). The friendship graph is stored implicitly, so I have no way of easily checking. • Are you truly an individual, or just one of the faceless crowd? – This question boils down to whether the friendship graph is labeled or unlabeled. Does each vertex have a name/label which reﬂects its identify, and is this label important for our analysis? Much of the study of social networks is unconcerned with labels on graphs. Often the index number given a vertex in the graph data structure serves as its label, perhaps for convenience or the need for anonymity. You may assert that you are a name, not a number—but try protesting to the guy who implements the algorithm. Someone studying how an infectious disease spreads through a graph may label each vertex with whether the person is healthy or sick, it being irrelevant what their name is. Take-Home Lesson: Graphs can be used to model a wide variety of structures and relationships. Graph-theoretic terminology gives us a language to talk about them. 5.2 151 DATA STRUCTURES FOR GRAPHS Data Structures for Graphs Selecting the right graph data structure can have an enormous impact on performance. Your two basic choices are adjacency matrices and adjacency lists, illustrated in Figure 5.4. We assume the graph G = (V, E) contains n vertices and m edges. • Adjacency Matrix: We can represent G using an n × n matrix M , where element M [i, j] = 1 if (i, j) is an edge of G, and 0 if it isn’t. This allows fast answers to the question “is (i, j) in G?”, and rapid updates for edge insertion and deletion. It may use excessive space for graphs with many vertices and relatively few edges, however. 4 152 5. GRAPH TRAVERSAL Comparison Faster to test if (x, y) is in graph? Faster to ﬁnd the degree of a vertex? Less memory on small graphs? Less memory on big graphs? Edge insertion or deletion? Faster to traverse the graph? Better for most problems? Winner adjacency matrices adjacency lists adjacency lists (m + n) vs. (n2 ) adjacency matrices (a small win) adjacency matrices O(1) vs. O(d) adjacency lists Θ(m + n) vs. Θ(n2 ) adjacency lists Figure 5.5: Relative advantages of adjacency lists and matrices. Consider a graph that represents the street map of Manhattan in New York City. Every junction of two streets will be a vertex of the graph. Neighboring junctions are connected by edges. How big is this graph? Manhattan is basically a grid of 15 avenues each crossing roughly 200 streets. This gives us about 3,000 vertices and 6,000 edges, since each vertex neighbors four other vertices and each edge is shared between two vertices. This modest amount of data should easily and eﬃciently be stored, yet an adjacency matrix would have 3,000 × 3,000 = 9,000,000 cells, almost all of them empty! There is some potential to save space by packing multiple bits per word or simulating a triangular matrix on undirected graphs. But these methods lose the simplicity that makes adjacency matrices so appealing and, more critically, remain inherently quadratic on sparse graphs. • Adjacency Lists: We can more eﬃciently represent sparse graphs by using linked lists to store the neighbors adjacent to each vertex. Adjacency lists require pointers but are not frightening once you have experience with linked structures. Adjacency lists make it harder to verify whether a given edge (i, j) is in G, since we must search through the appropriate list to ﬁnd the edge. However, it is surprisingly easy to design graph algorithms that avoid any need for such queries. Typically, we sweep through all the edges of the graph in one pass via a breadth-ﬁrst or depth-ﬁrst traversal, and update the implications of the current edge as we visit it. Table 5.5 summarizes the tradeoﬀs between adjacency lists and matrices. Take-Home Lesson: Adjacency lists are the right data structure for most applications of graphs. We will use adjacency lists as our primary data structure to represent graphs. We represent a graph using the following data type. For each graph, we keep a 5.2 DATA STRUCTURES FOR GRAPHS count of the number of vertices, and assign each vertex a unique identiﬁcation number from 1 to nvertices. We represent the edges using an array of linked lists: #define MAXV 1000 /* maximum number of vertices */ typedef struct { int y; int weight; struct edgenode *next; } edgenode; /* adjacency info */ /* edge weight, if any */ /* next edge in list */ typedef struct { edgenode *edges[MAXV+1]; int degree[MAXV+1]; int nvertices; int nedges; bool directed; } graph; /* /* /* /* /* adjacency info */ outdegree of each vertex */ number of vertices in graph */ number of edges in graph */ is the graph directed? */ We represent directed edge (x, y) by an edgenode y in x’s adjacency list. The degree ﬁeld of the graph counts the number of meaningful entries for the given vertex. An undirected edge (x, y) appears twice in any adjacency-based graph structure, once as y in x’s list, and once as x in y’s list. The boolean ﬂag directed identiﬁes whether the given graph is to be interpreted as directed or undirected. To demonstrate the use of this data structure, we show how to read a graph from a ﬁle. A typical graph format consists of an initial line featuring the number of vertices and edges in the graph, followed by a listing of the edges at one vertex pair per line. initialize_graph(graph *g, bool directed) { int i; /* counter */ g -> nvertices = 0; g -> nedges = 0; g -> directed = directed; for (i=1; i<=MAXV; i++) g->degree[i] = 0; for (i=1; i<=MAXV; i++) g->edges[i] = NULL; } Actually reading the graph requires inserting each edge into this structure: 153 154 5. GRAPH TRAVERSAL read_graph(graph *g, bool directed) { int i; int m; int x, y; /* counter */ /* number of edges */ /* vertices in edge (x,y) */ initialize_graph(g, directed); scanf("%d %d",&(g->nvertices),&m); for (i=1; i<=m; i++) { scanf("%d %d",&x,&y); insert_edge(g,x,y,directed); } } The critical routine is insert edge. The new edgenode is inserted at the beginning of the appropriate adjacency list, since order doesn’t matter. We parameterize our insertion with the directed Boolean ﬂag, to identify whether we need to insert two copies of each edge or only one. Note the use of recursion to solve this problem: insert_edge(graph *g, int x, int y, bool directed) { edgenode *p; /* temporary pointer */ p = malloc(sizeof(edgenode)); /* allocate edgenode storage */ p->weight = NULL; p->y = y; p->next = g->edges[x]; g->edges[x] = p; /* insert at head of list */ g->degree[x] ++; if (directed == FALSE) insert_edge(g,y,x,TRUE); else g->nedges ++; } Printing the associated graph is just a matter of two nested loops, one through vertices, the other through adjacent edges: 5.3 print_graph(graph *g) { int i; edgenode *p; WAR STORY: I WAS A VICTIM OF MOORE’S LAW /* counter */ /* temporary pointer */ for (i=1; i<=g->nvertices; i++) { printf("%d: ",i); p = g->edges[i]; while (p != NULL) { printf(" %d",p->y); p = p->next; } printf("\n"); } } It is a good idea to use a well-designed graph data type as a model for building your own, or even better as the foundation for your application. We recommend LEDA (see Section 19.1.1 (page 658)) or Boost (see Section 19.1.3 (page 659)) as the best-designed general-purpose graph data structures currently available. They may be more powerful (and hence somewhat slower/larger) than you need, but they do so many things right that you are likely to lose most of the potential do-it-yourself beneﬁts through clumsiness. 5.3 War Story: I was a Victim of Moore’s Law I am the author of a popular library of graph algorithms called Combinatorica (see www.combinatorica.com), which runs under the computer algebra system Mathematica. Eﬃciency is a great challenge in Mathematica, due to its applicative model of computation (it does not support constant-time write operations to arrays) and the overhead of interpretation (as opposed to compilation). Mathematica code is typically 1,000 to 5,000 times slower than C code. Such slow downs can be a tremendous performance hit. Even worse, Mathematica was a memory hog, needing a then-outrageous 4MB of main memory to run eﬀectively when I completed Combinatorica in 1990. Any computation on large structures was doomed to thrash in virtual memory. In such an environment, my graph package could only hope to work eﬀectively on very small graphs. One design decision I made as a result was to use adjacency matrices as the basic Combinatorica graph data structure instead of lists. This may sound peculiar. If pressed for memory, wouldn’t it pay to use adjacency lists and conserve every last byte? Yes, but the answer is not so simple for very small graphs. An adjacency list representation of a weighted n-vertex, m-edge graph should use about n+2m words to represent; the 2m comes from storing the endpoint and weight components of 155 156 5. GRAPH TRAVERSAL Figure 5.6: Representative Combinatorica graphs: edge-disjoint paths (left), Hamiltonian cycle in a hypercube (center), animated depth-ﬁrst search tree traversal (right) each edge. Thus, the space advantages of adjacency lists kick in when n + 2m is substantially smaller than n2 . The adjacency matrix is still manageable in size for n ≤ 100 and, of course, half the size of adjacency lists on dense graphs. My more immediate concern was dealing with the overhead of using a slow interpreted language. Check out the benchmarks reported in Table 5.1. Two particularly complex but polynomial-time problems on 9 and 16 vertex graphs took several minutes to complete on my desktop machine in 1990! The quadratic-sized data structure certainly could not have had much impact on these running times, since 9 × 9 equals only 81. From experience, I knew the Mathematica programming language handled better to regular structures like adjacency matrices better than irregular-sized adjacency lists. Still, Combinatorica proved to be a very good thing despite these performance problems. Thousands of people have used my package to do all kinds of interesting things with graphs. Combinatorica was never intended to be a high-performance algorithms library. Most users quickly realized that computations on large graphs were out of the question, but were eager to take advantage of Combinatorica as a mathematical research tool and prototyping environment. Everyone was happy. But over the years, my users started asking why it took so long to do a modestsized graph computation. The mystery wasn’t that my program was slow, because it had always been slow. The question was why did it take so many years for people to ﬁgure this out? 5.3 WAR STORY: I WAS A VICTIM OF MOORE’S LAW Approximate year command/machine PlanarQ[GridGraph[4,4]] Length[Partitions[30]] VertexConnectivity[GridGraph[3,3]] RandomPartition[1000] 1990 Sun-3 234.10 289.85 239.67 831.68 1991 Sun-4 69.65 73.20 47.76 267.5 1998 Sun-5 27.50 24.40 14.70 22.05 2000 Ultra 5 3.60 3.44 2.00 3.12 157 2004 SunBlade 0.40 1.58 0.91 0.87 Table 5.1: Old Combinatorica benchmarks on low-end Sun workstations, from 1990 to today, (running time in seconds) The reason is that computers keep doubling in speed every two years or so. People’s expectation of how long something should take moves in concert with these technology improvements. Partially because of Combinatorica’s dependence on a quadratic-size graph data structure, it didn’t scale as well as it should on sparse graphs. As the years rolled on, user demands become more insistent. Combinatorica needed to be updated. My collaborator, Sriram Pemmaraju, rose to the challenge. We (mostly he) completely rewrote Combinatorica to take advantage of faster graph data structures ten years after the initial version. The new Combinatorica uses a list of edges data structure for graphs, largely motivated by increased eﬃciency. Edge lists are linear in the size of the graph (edges plus vertices), just like adjacency lists. This makes a huge diﬀerence on most graph related functions—for large enough graphs. The improvement is most dramatic in “fast” graph algorithms—those that run in linear or near linear-time, such as graph traversal, topological sort, and ﬁnding connected/biconnected components. The implications of this change is felt throughout the package in running time improvements and memory savings. Combinatorica can now work with graphs that are about 50-100 times larger than graphs that the old package could deal with. Figure 5.7(l) plots the running time of the MinimumSpanningTree functions for both Combinatorica versions. The test graphs were sparse (grid graphs), designed to highlight the diﬀerence between the two data structures. Yes, the new version is much faster, but note that the diﬀerence only becomes important for graphs larger than the old Combinatorica was designed for. However, the relative diﬀerence in run time keeps growing with increasing n. Figure 5.7(r) plots the ratio of the running times as a function of graph size. The diﬀerence between linear size and quadratic size is asymptotic, so the consequences become ever more important as n gets larger. What is the weird bump in running times that occurs around n ≈ 250? This likely reﬂects a transition between levels of the memory hierarchy. Such bumps are not uncommon in today’s complex computer systems. Cache performance in data structure design should be an important but not overriding consideration. 158 5. GRAPH TRAVERSAL Running Time of Minimum Spanning Tree in Old and New Combinatorica Ratio of Running Times for Mininum Spanning Tree in Old and New Combinatorica 18 45 Run time ratio New Combinatorica Old Combinatorica 40 16 14 30 12 25 Ratio Running Time in Seconds 35 10 20 8 15 6 10 4 5 2 0 0 200 400 600 Vertices in Graph 800 1000 1200 0 200 400 600 800 1000 1200 Vertices in Graph Figure 5.7: Performance comparison between old and new Combinatorica: absolute running times (l), and the ratio of these times (r). The asymptotic gains due to adjacency lists more than trumped any impact of the cache. Two main lessons can be taken away from our experience developing Combinatorica: • To make a program run faster, just wait – Sophisticated hardware eventually slithers down to everybody. We observe a speedup of more than 200-fold for the original version of Combinatorica as a consequence of 15 years of faster hardware. In this context, the further speedups we obtained from upgrading the package become particularly dramatic. • Asymptotics eventually do matter – It was my mistake not to anticipate future developments in technology. While no one has a crystal ball, it is fairly safe to say that future computers will have more memory and run faster than today’s. This gives an edge to asymptotically more eﬃcient algorithms/data structures, even if their performance is close on today’s instances. If the implementation complexity is not substantially greater, play it safe and go with the better algorithm. 5.4 War Story: Getting the Graph “It takes ﬁve minutes just to read the data. We will never have time to make it do something interesting.” The young graduate student was bright and eager, but green to the power of data structures. She would soon come to appreciate their power. 5.4 WAR STORY: GETTING THE GRAPH Figure 5.8: The dual graph (dashed lines) of a triangulation As described in a previous war story (see Section 3.6 (page 85)), we were experimenting with algorithms to extract triangular strips for the fast rendering of triangulated surfaces. The task of ﬁnding a small number of strips that cover each triangle in a mesh could be modeled as a graph problem. The graph has a vertex for every triangle of the mesh, with an edge between every pair of vertices representing adjacent triangles. This dual graph representation (see Figure 5.8) captures all the information needed to partition the triangulation into triangle strips. The ﬁrst step in crafting a program that constructs a good set of strips was to build the dual graph of the triangulation. This I sent the student oﬀ to do. A few days later, she came back and announced that it took over ﬁve CPU minutes just to construct this dual graph of a few thousand triangles. “Nonsense!” I proclaimed. “You must be doing something very wasteful in building the graph. What format is the data in?” “Well, it starts out with a list of the 3D coordinates of the vertices used in the model and then follows with a list of triangles. Each triangle is described by a list of three indices into the vertex coordinates. Here is a small example:” VERTICES 4 0.000000 240.000000 0.000000 204.000000 240.000000 0.000000 204.000000 0.000000 0.000000 0.000000 0.000000 0.000000 TRIANGLES 2 0 1 3 1 2 3 “I see. So the ﬁrst triangle uses all but the third point, since all the indices start from zero. The two triangles must share an edge formed by points 1 and 3.” “Yeah, that’s right,” she conﬁrmed. 159 160 5. GRAPH TRAVERSAL “OK. Now tell me how you built your dual graph from this ﬁle.” “Well, I can ignore the vertex information once I know how many vertices there are. The geometric position of the points doesn’t aﬀect the structure of the graph. My dual graph is going to have as many vertices as the number of triangles. I set up an adjacency list data structure with that many vertices. As I read in each triangle, I compare it to each of the others to check whether it has two end points in common. If it does, I add an edge from the new triangle to this one.” I started to sputter. “But that’s your problem right there! You are comparing each triangle against every other triangle, so that constructing the dual graph will be quadratic in the number of triangles. Reading the input graph should take linear time!” “I’m not comparing every triangle against every other triangle. On average, it only tests against half or a third of the triangles.” “Swell. But that still leaves us with an O(n2 ) algorithm. That is much too slow.” She stood her ground. “Well, don’t just complain. Help me ﬁx it!” Fair enough. I started to think. We needed some quick method to screen away most of the triangles that would not be adjacent to the new triangle (i, j, k). What we really needed was a separate list of all the triangles that go through each of the points i, j, and k. By Euler’s formula for planar graphs, the average point is incident on less than six triangles. This would compare each new triangle against fewer than twenty others, instead of most of them. “We are going to need a data structure consisting of an array with one element for every vertex in the original data set. This element is going to be a list of all the triangles that pass through that vertex. When we read in a new triangle, we will look up the three relevant lists in the array and compare each of these against the new triangle. Actually, only two of the three lists need be tested, since any adjacent triangles will share two points in common. We will add an adjacency to our graph for every triangle-pair sharing two vertices. Finally, we will add our new triangle to each of the three aﬀected lists, so they will be updated for the next triangle read.” She thought about this for a while and smiled. “Got it, Chief. I’ll let you know what happens.” The next day she reported that the graph could be built in seconds, even for much larger models. From here, she went on to build a successful program for extracting triangle strips, as reported in Section 3.6 (page 85). The take-home lesson is that even elementary problems like initializing data structures can prove to be bottlenecks in algorithm development. Most programs working with large amounts of data have to run in linear or almost linear time. Such tight performance demands leave no room to be sloppy. Once you focus on the need for linear-time performance, an appropriate algorithm or heuristic can usually be found to do the job. 5.5 5.5 TRAVERSING A GRAPH Traversing a Graph Perhaps the most fundamental graph problem is to visit every edge and vertex in a graph in a systematic way. Indeed, all the basic bookkeeping operations on graphs (such printing or copying graphs, and converting between alternate representations) are applications of graph traversal. Mazes are naturally represented by graphs, where each graph vertex denotes a junction of the maze, and each graph edge denotes a hallway in the maze. Thus, any graph traversal algorithm must be powerful enough to get us out of an arbitrary maze. For eﬃciency, we must make sure we don’t get trapped in the maze and visit the same place repeatedly. For correctness, we must do the traversal in a systematic way to guarantee that we get out of the maze. Our search must take us through every edge and vertex in the graph. The key idea behind graph traversal is to mark each vertex when we ﬁrst visit it and keep track of what we have not yet completely explored. Although bread crumbs or unraveled threads have been used to mark visited places in fairy-tale mazes, we will rely on Boolean ﬂags or enumerated types. Each vertex will exist in one of three states: • undiscovered – the vertex is in its initial, virgin state. • discovered – the vertex has been found, but we have not yet checked out all its incident edges. • processed – the vertex after we have visited all its incident edges. Obviously, a vertex cannot be processed until after we discover it, so the state of each vertex progresses over the course of the traversal from undiscovered to discovered to processed. We must also maintain a structure containing the vertices that we have discovered but not yet completely processed. Initially, only the single start vertex is considered to be discovered. To completely explore a vertex v, we must evaluate each edge leaving v. If an edge goes to an undiscovered vertex x, we mark x discovered and add it to the list of work to do. We ignore an edge that goes to a processed vertex, because further contemplation will tell us nothing new about the graph. We can also ignore any edge going to a discovered but not processed vertex, because the destination already resides on the list of vertices to process. Each undirected edge will be considered exactly twice, once when each of its endpoints is explored. Directed edges will be considered only once, when exploring the source vertex. Every edge and vertex in the connected component must eventually be visited. Why? Suppose that there exists a vertex u that remains unvisited, whose neighbor v was visited. This neighbor v will eventually be explored, after which we will certainly visit u. Thus, we must ﬁnd everything that is there to be found. We describe the mechanics of these traversal algorithms and the signiﬁcance of the traversal order below. 161 162 5. GRAPH TRAVERSAL 2 1 3 1 4 5 2 5 3 4 6 6 Figure 5.9: An undirected graph and its breadth-ﬁrst search tree 5.6 Breadth-First Search The basic breadth-ﬁrst search algorithm is given below. At some point during the course of a traversal, every node in the graph changes state from undiscovered to discovered. In a breadth-ﬁrst search of an undirected graph, we assign a direction to each edge, from the discoverer u to the discovered v. We thus denote u to be the parent of v. Since each node has exactly one parent, except for the root, this deﬁnes a tree on the vertices of the graph. This tree, illustrated in Figure 5.9, deﬁnes a shortest path from the root to every other node in the tree. This property makes breadth-ﬁrst search very useful in shortest path problems. BFS(G, s) for each vertex u ∈ V [G] − {s} do state[u] = “undiscovered” p[u] = nil, i.e. no parent is in the BFS tree state[s] = “discovered” p[s] = nil Q = {s} while Q = ∅ do u = dequeue[Q] process vertex u as desired for each v ∈ Adj[u] do process edge (u, v) as desired if state[v] = “undiscovered” then state[v] = “discovered” p[v] = u enqueue[Q, v] state[u] = “processed” The graph edges that do not appear in the breadth-ﬁrst search tree also have special properties. For undirected graphs, nontree edges can point only to vertices on the same level as the parent vertex, or to vertices on the level directly below 5.6 BREADTH-FIRST SEARCH the parent. These properties follow easily from the fact that each path in the tree must be the shortest path in the graph. For a directed graph, a back-pointing edge (u, v) can exist whenever v lies closer to the root than u does. Implementation Our breadth-ﬁrst search implementation bfs uses two Boolean arrays to maintain our knowledge about each vertex in the graph. A vertex is discovered the ﬁrst time we visit it. A vertex is considered processed after we have traversed all outgoing edges from it. Thus, each vertex passes from undiscovered to discovered to processed over the course of the search. This information could have been maintained using one enumerated type variable, but we used two Boolean variables instead. bool processed[MAXV+1]; bool discovered[MAXV+1]; int parent[MAXV+1]; /* which vertices have been processed */ /* which vertices have been found */ /* discovery relation */ Each vertex is initialized as undiscovered: initialize_search(graph *g) { int i; /* counter */ for (i=1; i<=g->nvertices; i++) { processed[i] = discovered[i] = FALSE; parent[i] = -1; } } Once a vertex is discovered, it is placed on a queue. Since we process these vertices in ﬁrst-in, ﬁrst-out order, the oldest vertices are expanded ﬁrst, which are exactly those closest to the root: bfs(graph *g, int start) { queue q; int v; int y; edgenode *p; init_queue(&q); enqueue(&q,start); discovered[start] = TRUE; /* /* /* /* queue of vertices to visit */ current vertex */ successor vertex */ temporary pointer */ 163 164 5. GRAPH TRAVERSAL while (empty_queue(&q) == FALSE) { v = dequeue(&q); process_vertex_early(v); processed[v] = TRUE; p = g->edges[v]; while (p != NULL) { y = p->y; if ((processed[y] == FALSE) || g->directed) process_edge(v,y); if (discovered[y] == FALSE) { enqueue(&q,y); discovered[y] = TRUE; parent[y] = v; } p = p->next; } process_vertex_late(v); } } 5.6.1 Exploiting Traversal The exact behavior of bfs depends upon the functions process vertex early(), process vertex late(), and process edge(). Through these functions, we can customize what the traversal does as it makes its oﬃcial visit to each edge and each vertex. Initially, we will do all of vertex processing on entry, so process vertex late() returns without action: process_vertex_late(int v) { } By setting the active functions to process_vertex_early(int v) { printf("processed vertex %d\n",v); } process_edge(int x, int y) { printf("processed edge (%d,%d)\n",x,y); } 5.6 BREADTH-FIRST SEARCH we print each vertex and edge exactly once. If we instead set process edge to process_edge(int x, int y) { nedges = nedges + 1; } we get an accurate count of the number of edges. Diﬀerent algorithms perform diﬀerent actions on vertices or edges as they are encountered. These functions give us the freedom to easily customize our response. 5.6.2 Finding Paths The parent array set within bfs() is very useful for ﬁnding interesting paths through a graph. The vertex that discovered vertex i is deﬁned as parent[i]. Every vertex is discovered during the course of traversal, so except for the root every node has a parent. The parent relation deﬁnes a tree of discovery with the initial search node as the root of the tree. Because vertices are discovered in order of increasing distance from the root, this tree has a very important property. The unique tree path from the root to each node x ∈ V uses the smallest number of edges (or equivalently, intermediate nodes) possible on any root-to-x path in the graph. We can reconstruct this path by following the chain of ancestors from x to the root. Note that we have to work backward. We cannot ﬁnd the path from the root to x, since that does not follow the direction of the parent pointers. Instead, we must ﬁnd the path from x to the root. Since this is the reverse of how we normally want the path, we can either (1) store it and then explicitly reverse it using a stack, or (2) let recursion reverse it for us, as follows: find_path(int start, int end, int parents[]) { if ((start == end) || (end == -1)) printf("\n%d",start); else { find_path(start,parents[end],parents); printf(" %d",end); } } On our breadth-ﬁrst search graph example (Figure 5.9) our algorithm generated the following parent relation: vertex parent 1 -1 2 1 3 2 4 5 5 1 6 1 165 166 5. GRAPH TRAVERSAL For the shortest path from 1 to 4, upper-right corner, this parent relation yields the path {1, 5, 4}. There are two points to remember when using breadth-ﬁrst search to ﬁnd a shortest path from x to y: First, the shortest path tree is only useful if BFS was performed with x as the root of the search. Second, BFS gives the shortest path only if the graph is unweighted. We will present algorithms for ﬁnding shortest paths in weighted graphs in Section 6.3.1 (page 206). 5.7 Applications of Breadth-First Search Most elementary graph algorithms make one or two traversals of the graph while we update our knowledge of the graph. Properly implemented using adjacency lists, any such algorithm is destined to be linear, since BFS runs in O(n + m) time on both directed and undirected graphs. This is optimal, since it is as fast as one can hope to read any n-vertex, m-edge graph. The trick is seeing when traversal approaches are destined to work. We present several examples below. 5.7.1 Connected Components The “six degrees of separation” theory argues that there is always a short path linking every two people in the world. We say that a graph is connected if there is a path between any two vertices. If the theory is true, it means the friendship graph must be connected. A connected component of an undirected graph is a maximal set of vertices such that there is a path between every pair of vertices. The components are separate “pieces” of the graph such that there is no connection between the pieces. If we envision tribes in remote parts of the world that have yet not been encountered, each such tribe would form a separate connected component in the friendship graph. A remote hermit, or extremely unpleasant fellow, would represent a connected component of one vertex. An amazing number of seemingly complicated problems reduce to ﬁnding or counting connected components. For example, testing whether a puzzle such as Rubik’s cube or the 15-puzzle can be solved from any position is really asking whether the graph of legal conﬁgurations is connected. Connected components can be found using breadth-ﬁrst search, since the vertex order does not matter. We start from the ﬁrst vertex. Anything we discover during this search must be part of the same connected component. We then repeat the search from any undiscovered vertex (if one exists) to deﬁne the next component, and so on until all vertices have been found: 5.7 APPLICATIONS OF BREADTH-FIRST SEARCH connected_components(graph *g) { int c; int i; /* component number */ /* counter */ initialize_search(g); c = 0; for (i=1; i<=g->nvertices; i++) if (discovered[i] == FALSE) { c = c+1; printf("Component %d:",c); bfs(g,i); printf("\n"); } } process_vertex_early(int v) { printf(" %d",v); } process_edge(int x, int y) { } Observe how we increment a counter c denoting the current component number with each call to bfs. We could have explicitly bound each vertex to its component number (instead of printing the vertices in each component) by changing the action of process vertex. There are two distinct notions of connectivity for directed graphs, leading to algorithms for ﬁnding both weakly connected and strongly connected components. Both of these can be found in O(n + m) time, as discussed in Section 15.1 (page 477). 5.7.2 Two-Coloring Graphs The vertex-coloring problem seeks to assign a label (or color) to each vertex of a graph such that no edge links any two vertices of the same color. We can easily avoid all conﬂicts by assigning each vertex a unique color. However, the goal is to use as few colors as possible. Vertex coloring problems often arise in scheduling applications, such as register allocation in compilers. See Section 16.7 (page 544) for a full treatment of vertex-coloring algorithms and applications. 167 168 5. GRAPH TRAVERSAL A graph is bipartite if it can be colored without conﬂicts while using only two colors. Bipartite graphs are important because they arise naturally in many applications. Consider the “had-sex-with” graph in a heterosexual world. Men have sex only with women, and vice versa. Thus, gender deﬁnes a legal two-coloring, in this simple model. But how can we ﬁnd an appropriate two-coloring of a graph, thus separating the men from the women? Suppose we assume that the starting vertex is male. All vertices adjacent to this man must be female, assuming the graph is indeed bipartite. We can augment breadth-ﬁrst search so that whenever we discover a new vertex, we color it the opposite of its parent. We check whether any nondiscovery edge links two vertices of the same color. Such a conﬂict means that the graph cannot be twocolored. Otherwise, we will have constructed a proper two-coloring whenever we terminate without conﬂict. twocolor(graph *g) { int i; /* counter */ for (i=1; i<=(g->nvertices); i++) color[i] = UNCOLORED; bipartite = TRUE; initialize_search(&g); for (i=1; i<=(g->nvertices); i++) if (discovered[i] == FALSE) { color[i] = WHITE; bfs(g,i); } } process_edge(int x, int y) { if (color[x] == color[y]) { bipartite = FALSE; printf("Warning: not bipartite due to (%d,%d)\n",x,y); } color[y] = complement(color[x]); } 5.8 DEPTH-FIRST SEARCH complement(int color) { if (color == WHITE) return(BLACK); if (color == BLACK) return(WHITE); return(UNCOLORED); } We can assign the ﬁrst vertex in any connected component to be whatever color/sex we wish. BFS can separate the men from the women, but we can’t tell them apart just by using the graph structure. Take-Home Lesson: Breadth-ﬁrst and depth-ﬁrst searches provide mechanisms to visit each edge and vertex of the graph. They prove the basis of most simple, eﬃcient graph algorithms. 5.8 Depth-First Search There are two primary graph traversal algorithms: breadth-ﬁrst search (BFS) and depth-ﬁrst search (DFS). For certain problems, it makes absolutely no diﬀerence which you use, but in others the distinction is crucial. The diﬀerence between BFS and DFS results is in the order in which they explore vertices. This order depends completely upon the container data structure used to store the discovered but not processed vertices. • Queue – By storing the vertices in a ﬁrst-in, ﬁrst-out (FIFO) queue, we explore the oldest unexplored vertices ﬁrst. Thus our explorations radiate out slowly from the starting vertex, deﬁning a breadth-ﬁrst search. • Stack – By storing the vertices in a last-in, ﬁrst-out (LIFO) stack, we explore the vertices by lurching along a path, visiting a new neighbor if one is available, and backing up only when we are surrounded by previously discovered vertices. Thus, our explorations quickly wander away from our starting point, deﬁning a depth-ﬁrst search. Our implementation of dfs maintains a notion of traversal time for each vertex. Our time clock ticks each time we enter or exit any vertex. We keep track of the entry and exit times for each vertex. Depth-ﬁrst search has a neat recursive implementation, which eliminates the need to explicitly use a stack: DFS(G, u) state[u] = “discovered” process vertex u if desired entry[u] = time 169 170 5. GRAPH TRAVERSAL time = time + 1 for each v ∈ Adj[u] do process edge (u, v) if desired if state[v] = “undiscovered” then p[v] = u DFS(G, v) state[u] = “processed” exit[u] = time time = time + 1 The time intervals have interesting and useful properties with respect to depthﬁrst search: • Who is an ancestor? – Suppose that x is an ancestor of y in the DFS tree. This implies that we must enter x before y, since there is no way we can be born before our own father or grandfather! We also must exit y before we exit x, because the mechanics of DFS ensure we cannot exit x until after we have backed up from the search of all its descendants. Thus the time interval of y must be properly nested within ancestor x. • How many descendants? – The diﬀerence between the exit and entry times for v tells us how many descendents v has in the DFS tree. The clock gets incremented on each vertex entry and vertex exit, so half the time diﬀerence denotes the number of descendents of v. We will use these entry and exit times in several applications of depth-ﬁrst search, particularly topological sorting and biconnected/strongly-connected components. We need to be able to take separate actions on each entry and exit, thus motivating distinct process vertex early and process vertex late routines called from dfs. The other important property of a depth-ﬁrst search is that it partitions the edges of an undirected graph into exactly two classes: tree edges and back edges. The tree edges discover new vertices, and are those encoded in the parent relation. Back edges are those whose other endpoint is an ancestor of the vertex being expanded, so they point back into the tree. An amazing property of depth-ﬁrst search is that all edges fall into these two classes. Why can’t an edge go to a brother or cousin node instead of an ancestor? All nodes reachable from a given vertex v are expanded before we ﬁnish with the traversal from v, so such topologies are impossible for undirected graphs. This edge classiﬁcation proves fundamental to the correctness of DFS-based algorithms. 5.8 DEPTH-FIRST SEARCH 1 2 3 2 6 1 4 5 3 4 6 5 Figure 5.10: An undirected graph and its depth-ﬁrst search tree Implementation A depth-ﬁrst search can be thought of as a breadth-ﬁrst search with a stack instead of a queue. The beauty of implementing dfs recursively is that recursion eliminates the need to keep an explicit stack: dfs(graph *g, int v) { edgenode *p; int y; if (finished) return; /* temporary pointer */ /* successor vertex */ /* allow for search termination */ discovered[v] = TRUE; time = time + 1; entry_time[v] = time; process_vertex_early(v); p = g->edges[v]; while (p != NULL) { y = p->y; if (discovered[y] == FALSE) { parent[y] = v; process_edge(v,y); dfs(g,y); } else if ((!processed[y]) || (g->directed)) process_edge(v,y); 171 172 5. GRAPH TRAVERSAL if (finished) return; p = p->next; } process_vertex_late(v); time = time + 1; exit_time[v] = time; processed[v] = TRUE; } Depth-ﬁrst search use essentially the same idea as backtracking, which we study in Section 7.1 (page 231). Both involve exhaustively searching all possibilities by advancing if it is possible, and backing up as soon as there is no unexplored possibility for further advancement. Both are most easily understood as recursive algorithms. Take-Home Lesson: DFS organizes vertices by entry/exit times, and edges into tree and back edges. This organization is what gives DFS its real power. 5.9 Applications of Depth-First Search As algorithm design paradigms go, a depth-ﬁrst search isn’t particularly intimidating. It is surprisingly subtle, however meaning that its correctness requires getting details right. The correctness of a DFS-based algorithm depends upon speciﬁcs of exactly when we process the edges and vertices. We can process vertex v either before we have traversed any of the outgoing edges from v (process vertex early()) or after we have ﬁnished processing all of them (process vertex late()). Sometimes we will take special actions at both times, say process vertex early() to initialize a vertex-speciﬁc data structure, which will be modiﬁed on edge-processing operations and then analyzed afterwards using process vertex late(). In undirected graphs, each edge (x, y) sits in the adjacency lists of vertex x and y. Thus there are two potential times to process each edge (x, y), namely when exploring x and when exploring y. The labeling of edges as tree edges or back edges occurs during the ﬁrst time the edge is explored. This ﬁrst time we see an edge is usually a logical time to do edge-speciﬁc processing. Sometimes, we may want to take diﬀerent action the second time we see an edge. But when we encounter edge (x, y) from x, how can we tell if we have previously traversed the edge from y? The issue is easy if vertex y is undiscovered: (x, y) 5.9 APPLICATIONS OF DEPTH-FIRST SEARCH Figure 5.11: An articulation vertex is the weakest point in the graph becomes a tree edge so this must be the ﬁrst time. The issue is also easy if y has not been completely processed; since we explored the edge when we explored y this must be the second time. But what if y is an ancestor of x, and thus in a discovered state? Careful reﬂection will convince you that this must be our ﬁrst traversal unless y is the immediate ancestor of x—i.e. , (y, x) is a tree edge. This can be established by testing if y == parent[x]. I ﬁnd that the subtlety of depth-ﬁrst search-based algorithms kicks me in the head whenever I try to implement one. I encourage you to analyze these implementations carefully to see where the problematic cases arise and why. 5.9.1 Finding Cycles Back edges are the key to ﬁnding a cycle in an undirected graph. If there is no back edge, all edges are tree edges, and no cycle exists in a tree. But any back edge going from x to an ancestor y creates a cycle with the tree path from y to x. Such a cycle is easy to ﬁnd using dfs: process_edge(int x, int y) { if (parent[x] != y) { /* found back edge! */ printf("Cycle from %d to %d:",y,x); find_path(y,x,parent); printf("\n\n"); finished = TRUE; } } The correctness of this cycle detection algorithm depends upon processing each undirected edge exactly once. Otherwise, a spurious two-vertex cycle (x, y, x) could be composed from the two traversals of any single undirected edge. We use the finished ﬂag to terminate after ﬁnding the ﬁrst cycle. 5.9.2 Articulation Vertices Suppose you are a vandal seeking to disrupt the telephone network. Which station in Figure 5.11 should you choose to blow up to cause the maximum amount of 173 174 5. GRAPH TRAVERSAL 1 2 6 3 4 5 Figure 5.12: DFS tree of a graph containing two articulation vertices (namely 1 and 2). Back edge (5, 2) keeps vertices 3 and 4 from being cut-nodes. Vertices 5 and 6 escape as leaves of the DFS tree damage? Observe that there is a single point of failure—a single vertex whose deletion disconnects a connected component of the graph. Such a vertex is called an articulation vertex or cut-node. Any graph that contains an articulation vertex is inherently fragile, because deleting that single vertex causes a loss of connectivity between other nodes. We presented a breadth-ﬁrst search-based connected components algorithm in Section 5.7.1 (page 166). In general, the connectivity of a graph is the smallest number of vertices whose deletion will disconnect the graph. The connectivity is one if the graph has an articulation vertex. More robust graphs without such a vertex are said to be biconnected. Connectivity will be further discussed in the catalog in Section 15.8 (page 505). Testing for articulation vertices by brute force is easy. Temporarily delete each vertex v, and then do a BFS or DFS traversal of the remaining graph to establish whether it is still connected. The total time for n such traversals is O(n(m + n)). There is a clever linear-time algorithm, however, that tests all the vertices of a connected graph using a single depth-ﬁrst search. What might the depth-ﬁrst search tree tell us about articulation vertices? This tree connects all the vertices of the graph. If the DFS tree represented the entirety of the graph, all internal (nonleaf) nodes would be articulation vertices, since deleting any one of them would separate a leaf from the root. Blowing up a leaf (such as vertices 2 and 6 in Figure 5.12) cannot disconnect the tree, since it connects no one but itself to the main trunk. 5.9 APPLICATIONS OF DEPTH-FIRST SEARCH The root of the search tree is a special case. If it has only one child, it functions as a leaf. But if the root has two or more children, its deletion disconnects them, making the root an articulation vertex. General graphs are more complex than trees. But a depth-ﬁrst search of a general graph partitions the edges into tree edges and back edges. Think of these back edges as security cables linking a vertex back to one of its ancestors. The security cable from x back to y ensures that none of the vertices on the tree path between x and y can be articulation vertices. Delete any of these vertices, and the security cable will still hold all of them to the rest of the tree. Finding articulation vertices requires maintaining the extent to which back edges (i.e. , security cables) link chunks of the DFS tree back to ancestor nodes. Let reachable ancestor[v] denote the earliest reachable ancestor of vertex v, meaning the oldest ancestor of v that we can reach by a combination of tree edges and back edges. Initially, reachable ancestor[v] = v: int reachable_ancestor[MAXV+1]; /* earliest reachable ancestor of v */ int tree_out_degree[MAXV+1]; /* DFS tree outdegree of v */ process_vertex_early(int v) { reachable_ancestor[v] = v; } We update reachable ancestor[v] whenever we encounter a back edge that takes us to an earlier ancestor than we have previously seen. The relative age/rank of our ancestors can be determined from their entry time’s: process_edge(int x, int y) { int class; /* edge class */ class = edge_classification(x,y); if (class == TREE) tree_out_degree[x] = tree_out_degree[x] + 1; if ((class == BACK) && (parent[x] != y)) { if (entry_time[y] < entry_time[ reachable_ancestor[x] ] ) reachable_ancestor[x] = y; } } The key issue is determining how the reachability relation impacts whether vertex v is an articulation vertex. There are three cases, as shown in Figure 5.13: 175 176 5. GRAPH TRAVERSAL root cutnode bridge cutnode bridge cutnode parent cutnode (of v) v Figure 5.13: The three cases of articulation vertices: root, bridge, and parent cut-nodes • Root cut-nodes – If the root of the DFS tree has two or more children, it must be an articulation vertex. No edges from the subtree of the second child can possibly connect to the subtree of the ﬁrst child. • Bridge cut-nodes – If the earliest reachable vertex from v is v, then deleting the single edge (parent[v], v) disconnects the graph. Clearly parent[v] must be an articulation vertex, since it cuts v from the graph. Vertex v is also an articulation vertex unless it is a leaf of the DFS tree. For any leaf, nothing falls oﬀ when you cut it. • Parent cut-nodes – If the earliest reachable vertex from v is the parent of v, then deleting the parent must sever v from the tree unless the parent is the root. The routine below systematically evaluates each of the three conditions as we back up from the vertex after traversing all outgoing edges. We use entry time[v] to represent the age of vertex v. The reachability time time v calculated below denotes the oldest vertex that can be reached using back edges. Getting back to an ancestor above v rules out the possibility of v being a cut-node: 5.9 APPLICATIONS OF DEPTH-FIRST SEARCH process_vertex_late(int v) { bool root; /* is the vertex the root of the DFS tree? */ int time_v; /* earliest reachable time for v */ int time_parent; /* earliest reachable time for parent[v] */ if (parent[v] < 1) { /* test if v is the root */ if (tree_out_degree[v] > 1) printf("root articulation vertex: %d \n",v); return; } root = (parent[parent[v]] < 1); /* is parent[v] the root? */ if ((reachable_ancestor[v] == parent[v]) && (!root)) printf("parent articulation vertex: %d \n",parent[v]); if (reachable_ancestor[v] == v) { printf("bridge articulation vertex: %d \n",parent[v]); if (tree_out_degree[v] > 0) /* test if v is not a leaf */ printf("bridge articulation vertex: %d \n",v); } time_v = entry_time[reachable_ancestor[v]]; time_parent = entry_time[ reachable_ancestor[parent[v]] ]; if (time_v < time_parent) reachable_ancestor[parent[v]] = reachable_ancestor[v]; } The last lines of this routine govern when we back up a node’s highest reachable ancestor to its parent, namely whenever it is higher than the parent’s earliest ancestor to date. We can alternately talk about reliability in terms of edge failures instead of vertex failures. Perhaps our vandal would ﬁnd it easier to cut a cable instead of blowing up a switching station. A single edge whose deletion disconnects the graph is called a bridge; any graph without such an edge is said to be edge-biconnected. Identifying whether a given edge (x, y) is a bridge is easily done in linear time by deleting the edge and testing whether the resulting graph is connected. In fact all bridges can be identiﬁed in the same O(n+m) time. Edge (x, y) is a bridge if (1) it is a tree edge, and (2) no back edge connects from y or below to x or above. This can be computed with a minor modiﬁcation of the reachable ancestor function. 177 178 5. GRAPH TRAVERSAL Tree Edges Forward Edge Back Edge Cross Edges Figure 5.14: Possible edge cases for BFS/DFS traversal 5.10 Depth-First Search on Directed Graphs Depth-ﬁrst search on an undirected graph proves useful because it organizes the edges of the graph in a very precise way, as shown in Figure 5.10. When traversing undirected graphs, every edge is either in the depth-ﬁrst search tree or a back edge to an ancestor in the tree. Let us review why. Suppose we encountered a “forward edge” (x, y) directed toward a descendant vertex. In this case, we would have discovered (x, y) while exploring y, making it a back edge. Suppose we encounter a “cross edge” (x, y), linking two unrelated vertices. Again, we would have discovered this edge when we explored y, making it a tree edge. For directed graphs, depth-ﬁrst search labelings can take on a wider range of possibilities. Indeed, all four of the edge cases in Figure 5.14 can occur in traversing directed graphs. Still, this classiﬁcation proves useful in organizing algorithms on directed graphs. We typically take a diﬀerent action on edges from each diﬀerent case. The correct labeling of each edge can be readily determined from the state, discovery time, and parent of each vertex, as encoded in the following function: int { if if if if edge_classification(int x, int y) (parent[y] == x) return(TREE); (discovered[y] && !processed[y]) return(BACK); (processed[y] && (entry_time[y]>entry_time[x])) return(FORWARD); (processed[y] && (entry_time[y] nvertices; i++) if (discovered[i] == FALSE) dfs(g,i); print_stack(&sorted); /* report topological order */ } We push each vertex on a stack as soon as we have evaluated all outgoing edges. The top vertex on the stack always has no incoming edges from any vertex on the stack. Repeatedly popping them oﬀ yields a topological ordering. 5.10.2 Strongly Connected Components We are often concerned with strongly connected components—that is, partitioning a graph into chunks such that directed paths exist between all pairs of vertices within a given chunk. A directed graph is strongly connected if there is a directed path between any two vertices. Road networks should be strongly connected, or else there will be places you can drive to but not drive home from without violating one-way signs. It is straightforward to use graph traversal to test whether a graph G = (V, E) is strongly connected in linear time. First, do a traversal from some arbitrary vertex v. Every vertex in the graph had better be reachable from v (and hence discovered on the BFS or DFS starting from v), otherwise G cannot possibly be strongly connected. Now construct a graph G = (V, E ) with the same vertex and edge set as G but with all edges reversed—i.e., directed edge (x, y) ∈ E iﬀ (y, x) ∈ E . Thus, any path from v to z in G corresponds to a path from z to v in G. By doing a DFS from v in G , we ﬁnd all vertices with paths to v in G. The graph is strongly connected iﬀ all vertices in G can (1) reach v and (2) are reachable from v. Graphs that are not strongly connected can be partitioned into strongly connected components, as shown in Figure 5.16 (left). The set of such components and the weakly-connecting edges that link them together can be determined using DFS. The algorithm is based on the observation that it is easy to ﬁnd a directed cycle using a depth-ﬁrst search, since any back edge plus the down path in the DFS tree gives such a cycle. All vertices in this cycle must be in the same strongly connected component. Thus, we can shrink (contract) the vertices on this cycle down to a single vertex representing the component, and then repeat. This process terminates when no directed cycle remains, and each vertex represents a diﬀerent strongly connected component. 181 182 5. GRAPH TRAVERSAL 1 1 3 2 2 4 5 6 4 3 8 6 8 7 7 5 Figure 5.16: The strongly-connected components of a graph, with the associated DFS tree Our approach to implementing this idea is reminiscent of ﬁnding biconnected components in Section 5.9.2 (page 173). We update our notion of the oldest reachable vertex in response to (1) nontree edges and (2) backing up from a vertex. Because we are working on a directed graph, we also must contend with forward edges (from a vertex to a descendant) and cross edges (from a vertex back to an nonancestor but previously discovered vertex). Our algorithm will peal one strong component oﬀ the tree at a time, and assign each of its vertices the number of the component it is in: strong_components(graph *g) { int i; /* counter */ for (i=1; i<=(g->nvertices); i++) { low[i] = i; scc[i] = -1; } components_found = 0; init_stack(&active); initialize_search(&g); for (i=1; i<=(g->nvertices); i++) if (discovered[i] == FALSE) { dfs(g,i); } } 5.10 DEPTH-FIRST SEARCH ON DIRECTED GRAPHS Deﬁne low[v] to be the oldest vertex known to be in the same strongly connected component as v. This vertex is not necessarily an ancestor, but may also be a distant cousin of v because of cross edges. Cross edges that point vertices from previous strongly connected components of the graph cannot help us, because there can be no way back from them to v, but otherwise cross edges are fair game. Forward edges have no impact on reachability over the depth-ﬁrst tree edges, and hence can be disregarded: int low[MAXV+1]; int scc[MAXV+1]; /* oldest vertex surely in component of v */ /* strong component number for each vertex */ process_edge(int x, int y) { int class; /* edge class */ class = edge_classification(x,y); if (class == BACK) { if (entry_time[y] < entry_time[ low[x] ] ) low[x] = y; } if (class == CROSS) { if (scc[y] == -1) /* component not yet assigned */ if (entry_time[y] < entry_time[ low[x] ] ) low[x] = y; } } A new strongly connected component is found whenever the lowest reachable vertex from v is v. If so, we can clear the stack of this component. Otherwise, we give our parent the beneﬁt of the oldest ancestor we can reach and backtrack: process_vertex_early(int v) { push(&active,v); } 183 184 5. GRAPH TRAVERSAL process_vertex_late(int v) { if (low[v] == v) { /* edge (parent[v],v) cuts off scc */ pop_component(v); } if (entry_time[low[v]] < entry_time[low[parent[v]]]) low[parent[v]] = low[v]; } pop_component(int v) { int t; /* vertex placeholder */ components_found = components_found + 1; scc[ v ] = components_found; while ((t = pop(&active)) != v) { scc[ t ] = components_found; } } Chapter Notes Our treatment of graph traversal represents an expanded version of material from Chapter 9 of [SR03]. The Combinatorica graph library discussed in the war story is best described in the old [Ski90]. and new editions [PS03] of the associated book. Accessible introductions to the science of social networks include Barabasi [Bar03] and Watts [Wat04]. 5.11 Exercises Simulating Graph Algorithms 5-1. [3] For the following graphs G1 (left) and G2 (right): 5.11 1 A B 6 A C 3 2 4 2 3 D 4 8 E F 7 G 10 5 11 6 I H 3 2 13 F 4 11 4 J M 1 5 23 K L 9 12 J I H 11 9 4 22 G 6 3 2 1 D 1 4 E 6 20 C 2 1 9 18 B EXERCISES 9 N O 8 P (a) Report the order of the vertices encountered on a breadth-ﬁrst search starting from vertex A. Break all ties by picking the vertices in alphabetical order (i.e., A before Z). (b) Report the order of the vertices encountered on a depth-ﬁrst search starting from vertex A. Break all ties by picking the vertices in alphabetical order (i.e., A before Z). 5-2. [3] Do a topological sort of the following graph G: A C B F E D G H J I Traversal 5-3. [3] Prove by induction that there is a unique path between any pair of vertices in a tree. 5-4. [3] Prove that in a breadth-ﬁrst search on a undirected graph G, every edge is either a tree edge or a cross edge, where x is neither an ancestor nor descendant of y, in cross edge (x, y). 5-5. [3] Give a linear algorithm to compute the chromatic number of graphs where each vertex has degree at most 2. Must such graphs be bipartite? 5-6. [5] In breadth-ﬁrst and depth-ﬁrst search, an undiscovered node is marked discovered when it is ﬁrst encountered, and marked processed when it has been completely 185 186 5. GRAPH TRAVERSAL + + 5 * 3 4 + + / * 2 3 / * 2 4 3 5 4 Figure 5.17: Expression 2 + 3 ∗ 4 + (3 ∗ 4)/5 as a tree and a DAG searched. At any given moment, several nodes might be simultaneously in the discovered state. (a) Describe a graph on n vertices and a particular starting vertex v such that Θ(n) nodes are simultaneously in the discovered state during a breadth-ﬁrst search starting from v. (b) Describe a graph on n vertices and a particular starting vertex v such that Θ(n) nodes are simultaneously in the discovered state during a depth-ﬁrst search starting from v. (c) Describe a graph on n vertices and a particular starting vertex v such that at some point Θ(n) nodes remain undiscovered, while Θ(n) nodes have been processed during a depth-ﬁrst search starting from v. (Note, there may also be discovered nodes.) 5-7. [4] Given pre-order and in-order traversals of a binary tree, is it possible to reconstruct the tree? If so, sketch an algorithm to do it. If not, give a counterexample. Repeat the problem if you are given the pre-order and post-order traversals. 5-8. [3] Present correct and eﬃcient algorithms to convert an undirected graph G between the following graph data structures. You must give the time complexity of each algorithm, assuming n vertices and m edges. (a) Convert from an adjacency matrix to adjacency lists. (b) Convert from an adjacency list to an incidence matrix. An incidence matrix M has a row for each vertex and a column for each edge, such that M [i, j] = 1 if vertex i is part of edge j, otherwise M [i, j] = 0. (c) Convert from an incidence matrix to adjacency lists. 5-9. [3] Suppose an arithmetic expression is given as a tree. Each leaf is an integer and each internal node is one of the standard arithmetical operations (+, −, ∗, /). For example, the expression 2 + 3 ∗ 4 + (3 ∗ 4)/5 is represented by the tree in Figure 5.17(a). Give an O(n) algorithm for evaluating such an expression, where there are n nodes in the tree. 5-10. [5] Suppose an arithmetic expression is given as a DAG (directed acyclic graph) with common subexpressions removed. Each leaf is an integer and each internal 5.11 EXERCISES node is one of the standard arithmetical operations (+, −, ∗, /). For example, the expression 2 + 3 ∗ 4 + (3 ∗ 4)/5 is represented by the DAG in Figure 5.17(b). Give an O(n + m) algorithm for evaluating such a DAG, where there are n nodes and m edges in the DAG. Hint: modify an algorithm for the tree case to achieve the desired eﬃciency. 5-11. [8] The war story of Section 5.4 (page 158) describes an algorithm for constructing the dual graph of the triangulation eﬃciently, although it does not guarantee linear time. Give a worst-case linear algorithm for the problem. Algorithm Design 5-12. [5] The square of a directed graph G = (V, E) is the graph G2 = (V, E 2 ) such that (u, w) ∈ E 2 iﬀ there exists v ∈ V such that (u, v) ∈ E and (v, w) ∈ E; i.e., there is a path of exactly two edges from u to w. Give eﬃcient algorithms for both adjacency lists and matrices. 5-13. [5] A vertex cover of a graph G = (V, E) is a subset of vertices V such that each edge in E is incident on at least one vertex of V . (a) Give an eﬃcient algorithm to ﬁnd a minimum-size vertex cover if G is a tree. (b) Let G = (V, E) be a tree such that the weight of each vertex is equal to the degree of that vertex. Give an eﬃcient algorithm to ﬁnd a minimum-weight vertex cover of G. (c) Let G = (V, E) be a tree with arbitrary weights associated with the vertices. Give an eﬃcient algorithm to ﬁnd a minimum-weight vertex cover of G. 5-14. [3] A vertex cover of a graph G = (V, E) is a subset of vertices V ∈ V such that every edge in E contains at least one vertex from V . Delete all the leaves from any depth-ﬁrst search tree of G. Must the remaining vertices form a vertex cover of G? Give a proof or a counterexample. 5-15. [5] A vertex cover of a graph G = (V, E) is a subset of vertices V ∈ V such that every edge in E contains at least one vertex from V . An independent set of graph G = (V, E) is a subset of vertices V ∈ V such that no edge in E contains both vertices from V . An independent vertex cover is a subset of vertices that is both an independent set and a vertex cover of G. Give an eﬃcient algorithm for testing whether G contains an independent vertex cover. What classical graph problem does this reduce to? 5-16. [5] An independent set of an undirected graph G = (V, E) is a set of vertices U such that no edge in E is incident on two vertices of U . (a) Give an eﬃcient algorithm to ﬁnd a maximum-size independent set if G is a tree. (b) Let G = (V, E) be a tree with weights associated with the vertices such that the weight of each vertex is equal to the degree of that vertex. Give an eﬃcient algorithm to ﬁnd a maximum independent set of G. (c) Let G = (V, E) be a tree with arbitrary weights associated with the vertices. Give an eﬃcient algorithm to ﬁnd a maximum independent set of G. 187 188 5. GRAPH TRAVERSAL 5-17. [5] Consider the problem of determining whether a given undirected graph G = (V, E) contains a triangle or cycle of length 3. (a) Give an O(|V |3 ) to ﬁnd a triangle if one exists. (b) Improve your algorithm to run in time O(|V |·|E|). You may assume |V | ≤ |E|. Observe that these bounds gives you time to convert between the adjacency matrix and adjacency list representations of G. 5-18. [5] Consider a set of movies M1 , M2 , . . . , Mk . There is a set of customers, each one of which indicates the two movies they would like to see this weekend. Movies are shown on Saturday evening and Sunday evening. Multiple movies may be screened at the same time. You must decide which movies should be televised on Saturday and which on Sunday, so that every customer gets to see the two movies they desire. Is there a schedule where each movie is shown at most once? Design an eﬃcient algorithm to ﬁnd such a schedule if one exists. 5-19. [5] The diameter of a tree T = (V, E) is given by max δ(u, v) u,v∈V (where δ(u, v) is the number of edges on the path from u to v). Describe an eﬃcient algorithm to compute the diameter of a tree, and show the correctness and analyze the running time of your algorithm. 5-20. [5] Given an undirected graph G with n vertices and m edges, and an integer k, give an O(m + n) algorithm that ﬁnds the maximum induced subgraph H of G such that each vertex in H has degree ≥ k, or prove that no such graph exists. An induced subgraph F = (U, R) of a graph G = (V, E) is a subset of U of the vertices V of G, and all edges R of G such that both vertices of each edge are in U . 5-21. [6] Let v and w be two vertices in a directed graph G = (V, E). Design a lineartime algorithm to ﬁnd the number of diﬀerent shortest paths (not necessarily vertex disjoint) between v and w. Note: the edges in G are unweighted. 5-22. [6] Design a linear-time algorithm to eliminate each vertex v of degree 2 from a graph by replacing edges (u, v) and (v, w) by an edge (u, w). We also seek to eliminate multiple copies of edges by replacing them with a single edge. Note that removing multiple copies of an edge may create a new vertex of degree 2, which has to be removed, and that removing a vertex of degree 2 may create multiple edges, which also must be removed. Directed Graphs 5-23. [5] Your job is to arrange n ill-behaved children in a straight line, facing front. You are given a list of m statements of the form “i hates j”. If i hates j, then you do not want put i somewhere behind j, because then i is capable of throwing something at j. (a) Give an algorithm that orders the line, (or says that it is not possible) in O(m + n) time. 5.11 EXERCISES (b) Suppose instead you want to arrange the children in rows such that if i hates j, then i must be in a lower numbered row than j. Give an eﬃcient algorithm to ﬁnd the minimum number of rows needed, if it is possible. 5-24. [3] Adding a single directed edge to a directed graph can reduce the number of weakly connected components, but by at most how many components? What about the number of strongly connected components? 5-25. [5] An arborescence of a directed graph G is a rooted tree such that there is a directed path from the root to every other vertex in the graph. Give an eﬃcient and correct algorithm to test whether G contains an arborescence, and its time complexity. 5-26. [5] A mother vertex in a directed graph G = (V, E) is a vertex v such that all other vertices G can be reached by a directed path from v. (a) Give an O(n + m) algorithm to test whether a given vertex v is a mother of G, where n = |V | and m = |E|. (b) Give an O(n+m) algorithm to test whether graph G contains a mother vertex. 5-27. [9] A tournament is a directed graph formed by taking the complete undirected graph and assigning arbitrary directions on the edges—i.e., a graph G = (V, E) such that for all u,v ∈ V , exactly one of (u, v) or (v, u) is in E. Show that every tournament has a Hamiltonian path—that is, a path that visits every vertex exactly once. Give an algorithm to ﬁnd this path. Articulation Vertices 5-28. [5] An articulation vertex of a graph G is a vertex whose deletion disconnects G. Let G be a graph with n vertices and m edges. Give a simple O(n + m) algorithm for ﬁnding a vertex of G that is not an articulation vertex—i.e. , whose deletion does not disconnect G. 5-29. [5] Following up on the previous problem, give an O(n + m) algorithm that ﬁnds a deletion order for the n vertices such that no deletion disconnects the graph. (Hint: think DFS/BFS.) 5-30. [3] Suppose G is a connected undirected graph. An edge e whose removal disconnects the graph is called a bridge. Must every bridge e be an edge in a depth-ﬁrst search tree of G? Give a proof or a counterexample. Interview Problems 5-31. [3] Which data structures are used in depth-ﬁrst and breath-ﬁrst search? 5-32. [4] Write a function to traverse binary search tree and return the ith node in sorted order. Programming Challenges These programming challenge problems with robot judging are available at http://www.programming-challenges.com or http://online-judge.uva.es. 5-1. “Bicoloring” – Programming Challenges 110901, UVA Judge 10004. 189 190 5. GRAPH TRAVERSAL 5-2. “Playing with Wheels” – Programming Challenges 110902, UVA Judge 10067. 5-3. “The Tourist Guide” – Programming Challenges 110903, UVA Judge 10099. 5-4. “Edit Step Ladders” – Programming Challenges 110905, UVA Judge 10029. 5-5. “Tower of Cubes” – Programming Challenges 110906, UVA Judge 10051. 6 Weighted Graph Algorithms The data structures and traversal algorithms of Chapter 5 provide the basic building blocks for any computation on graphs. However, all the algorithms presented there dealt with unweighted graphs—i.e. , graphs where each edge has identical value or weight. There is an alternate universe of problems for weighted graphs. The edges of road networks are naturally bound to numerical values such as construction cost, traversal time, length, or speed limit. Identifying the shortest path in such graphs proves more complicated than breadth-ﬁrst search in unweighted graphs, but opens the door to a wide range of applications. The graph data structure from Chapter 5 quietly supported edge-weighted graphs, but here we make this explicit. Our adjacency list structure consists of an array of linked lists, such that the outgoing edges from vertex x appear in the list edges[x]: typedef struct { edgenode *edges[MAXV+1]; int degree[MAXV+1]; int nvertices; int nedges; int directed; } graph; /* /* /* /* /* adjacency info */ outdegree of each vertex */ number of vertices in graph */ number of edges in graph */ is the graph directed? */ Each edgenode is a record containing three ﬁelds, the ﬁrst describing the second endpoint of the edge (y), the second enabling us to annotate the edge with a weight (weight), and the third pointing to the next edge in the list (next): S.S. Skiena, The Algorithm Design Manual, 2nd ed., DOI: 10.1007/978-1-84800-070-4 6, c Springer-Verlag London Limited 2008 192 6. WEIGHTED GRAPH ALGORITHMS typedef struct { int y; int weight; struct edgenode *next; } edgenode; /* adjacency info */ /* edge weight, if any */ /* next edge in list */ We now describe several sophisticated algorithms using this data structure, including minimum spanning trees, shortest paths, and maximum ﬂows. That these optimization problems can be solved eﬃciently is quite worthy of our respect. Recall that no such algorithm exists for the ﬁrst weighted graph problem we encountered, namely the traveling salesman problem. 6.1 Minimum Spanning Trees A spanning tree of a graph G = (V, E) is a subset of edges from E forming a tree connecting all vertices of V . For edge-weighted graphs, we are particularly interested in the minimum spanning tree—the spanning tree whose sum of edge weights is as small as possible. Minimum spanning trees are the answer whenever we need to connect a set of points (representing cities, homes, junctions, or other locations) by the smallest amount of roadway, wire, or pipe. Any tree is the smallest possible connected graph in terms of number of edges, while the minimum spanning tree is the smallest connected graph in terms of edge weight. In geometric problems, the point set p1 , . . . , pn deﬁnes a complete graph, with edge (vi , vj ) assigned a weight equal to the distance from pi to pj . An example of a geometric minimum spanning tree is illustrated in Figure 6.1. Additional applications of minimum spanning trees are discussed in Section 15.3 (page 484). A minimum spanning tree minimizes the total length over all possible spanning trees. However, there can be more than one minimum spanning tree in a graph. Indeed, all spanning trees of an unweighted (or equally weighted) graph G are minimum spanning trees, since each contains exactly n − 1 equal-weight edges. Such a spanning tree can be found using depth-ﬁrst or breadth-ﬁrst search. Finding a minimum spanning tree is more diﬃcult for general weighted graphs, however two diﬀerent algorithms are presented below. Both demonstrate the optimality of certain greedy heuristics. 6.1.1 Prim’s Algorithm Prim’s minimum spanning tree algorithm starts from one vertex and grows the rest of the tree one edge at a time until all vertices are included. Greedy algorithms make the decision of what to do next by selecting the best local option from all available choices without regard to the global structure. Since we seek the tree of minimum weight, the natural greedy algorithm for minimum 6.1 (a) (b) MINIMUM SPANNING TREES 193 (c) Figure 6.1: (a) Two spanning trees of point set; (b) the minimum spanning tree, and (c) the shortest path from center tree spanning tree repeatedly selects the smallest weight edge that will enlarge the number of vertices in the tree. Prim-MST(G) Select an arbitrary vertex s to start the tree from. While (there are still nontree vertices) Select the edge of minimum weight between a tree and nontree vertex Add the selected edge and vertex to the tree Tprim . Prim’s algorithm clearly creates a spanning tree, because no cycle can be introduced by adding edges between tree and nontree vertices. However, why should it be of minimum weight over all spanning trees? We have seen ample evidence of other natural greedy heuristics that do not yield a global optimum. Therefore, we must be particularly careful to demonstrate any such claim. We use proof by contradiction. Suppose that there existed a graph G for which Prim’s algorithm did not return a minimum spanning tree. Since we are building the tree incrementally, this means that there must have been some particular instant where we went wrong. Before we inserted edge (x, y), Tprim consisted of a set of edges that was a subtree of some minimum spanning tree Tmin , but choosing edge (x, y) fatally took us away from a minimum spanning tree (see Figure 6.2(a)). But how could we have gone wrong? There must be a path p from x to y in Tmin , as shown in Figure 6.2(b). This path must use an edge (v1 , v2 ), where v1 is in Tprim , but v2 is not. This edge (v1 , v2 ) must have weight at least that of (x, y), or Prim’s algorithm would have selected it before (x, y) when it had the chance. Inserting (x, y) and deleting (v1 , v2 ) from Tmin leaves a spanning tree no larger than before, meaning that Prim’s algorithm did not make a fatal mistake in selecting edge (x, y). Therefore, by contradiction, Prim’s algorithm must construct a minimum spanning tree. 194 6. WEIGHTED GRAPH ALGORITHMS x y x s y s v1 v2 (b) (a) Figure 6.2: Where Prim’s algorithm goes bad? No, because d(v1 , v2 ) ≥ d(x, y) Implementation Prim’s algorithm grows the minimum spanning tree in stages, starting from a given vertex. At each iteration, we add one new vertex into the spanning tree. A greedy algorithm suﬃces for correctness: we always add the lowest-weight edge linking a vertex in the tree to a vertex on the outside. The simplest implementation of this idea would assign each vertex a Boolean variable denoting whether it is already in the tree (the array intree in the code below), and then searches all edges at each iteration to ﬁnd the minimum weight edge with exactly one intree vertex. Our implementation is somewhat smarter. It keeps track of the cheapest edge linking every nontree vertex in the tree. The cheapest such edge over all remaining non-tree vertices gets added in each iteration. We must update the costs of getting to the non-tree vertices after each insertion. However, since the most recently inserted vertex is the only change in the tree, all possible edge-weight updates must come from its outgoing edges: prim(graph *g, int start) { int i; edgenode *p; bool intree[MAXV+1]; int distance[MAXV+1]; int v; int w; int weight; int dist; /* /* /* /* /* /* /* /* counter */ temporary pointer */ is the vertex in the tree yet? */ cost of adding to tree */ current vertex to process */ candidate next vertex */ edge weight */ best current distance from start */ for (i=1; i<=g->nvertices; i++) { intree[i] = FALSE; 6.1 MINIMUM SPANNING TREES distance[i] = MAXINT; parent[i] = -1; } distance[start] = 0; v = start; while (intree[v] == FALSE) { intree[v] = TRUE; p = g->edges[v]; while (p != NULL) { w = p->y; weight = p->weight; if ((distance[w] > weight) && (intree[w] == FALSE)) { distance[w] = weight; parent[w] = v; } p = p->next; } v = 1; dist = MAXINT; for (i=1; i<=g->nvertices; i++) if ((intree[i] == FALSE) && (dist > distance[i])) { dist = distance[i]; v = i; } } } Analysis Prim’s algorithm is correct, but how eﬃcient is it? This depends on which data structures are used to implement it. In the pseudocode, Prim’s algorithm makes n iterations sweeping through all m edges on each iteration—yielding an O(mn) algorithm. But our implementation avoids the need to test all m edges on each pass. It only considers the ≤ n cheapest known edges represented in the parent array and the ≤ n edges out of new tree vertex v to update parent. By maintaining a Boolean ﬂag along with each vertex to denote whether it is in the tree or not, we test whether the current edge joins a tree with a non-tree vertex in constant time. The result is an O(n2 ) implementation of Prim’s algorithm, and a good illustration of power of data structures to speed up algorithms. In fact, more sophisticated 195 196 6. WEIGHTED GRAPH ALGORITHMS 5 7 4 9 5 A 2 2 4 4 12 G 2 3 3 7 6 5 7 1 2 6 1 3 5 4 A A Prim(G,A) Kruskal(G) Figure 6.3: A graph G (l) with minimum spanning trees produced by Prim’s (m) and Kruskal’s (r) algorithms. The numbers on the trees denote the order of insertion; ties are broken arbitrarily priority-queue data structures lead to an O(m + n lg n) implementation, by making it faster to ﬁnd the minimum cost edge to expand the tree at each iteration. The minimum spanning tree itself or its cost can be reconstructed in two different ways. The simplest method would be to augment this procedure with statements that print the edges as they are found or totals the weight of all selected edges. Alternately, the tree topology is encoded by the parent array, so it plus the original graph describe everything about the minimum spanning tree. 6.1.2 Kruskal’s Algorithm Kruskal’s algorithm is an alternate approach to ﬁnding minimum spanning trees that proves more eﬃcient on sparse graphs. Like Prim’s, Kruskal’s algorithm is greedy. Unlike Prim’s, it does not start with a particular vertex. Kruskal’s algorithm builds up connected components of vertices, culminating in the minimum spanning tree. Initially, each vertex forms its own separate component in the tree-to-be. The algorithm repeatedly considers the lightest remaining edge and tests whether its two endpoints lie within the same connected component. If so, this edge will be discarded, because adding it would create a cycle in the treeto-be. If the endpoints are in diﬀerent components, we insert the edge and merge the two components into one. Since each connected component is always a tree, we need never explicitly test for cycles. Kruskal-MST(G) Put the edges in a priority queue ordered by weight. count = 0 while (count < n − 1) do get next edge (v, w) if (component (v) = component(w)) add to Tkruskal merge component(v) and component(w) 6.1 x MINIMUM SPANNING TREES y x y s v1 v2 (a) (b) Figure 6.4: Where Kruskal’s algorithm goes bad? No, because d(v1 , v2 ) ≥ d(x, y) This algorithm adds n − 1 edges without creating a cycle, so it clearly creates a spanning tree for any connected graph. But why must this be a minimum spanning tree? Suppose it wasn’t. As with the correctness proof of Prim’s algorithm, there must be some graph on which it fails. In particular, there must a single edge (x, y) whose insertion ﬁrst prevented the tree Tkruskal from being a minimum spanning tree Tmin . Inserting this edge (x, y) into Tmin will create a cycle with the path from x to y. Since x and y were in diﬀerent components at the time of inserting (x, y), at least one edge (say (v1 , v2 )) on this path would have been evaluated by Kruskal’s algorithm later than (x, y). But this means that w(v1 , v2 ) ≥ w(x, y), so exchanging the two edges yields a tree of weight at most Tmin . Therefore, we could not have made a fatal mistake in selecting (x, y), and the correctness follows. What is the time complexity of Kruskal’s algorithm? Sorting the m edges takes O(m lg m) time. The for loop makes m iterations, each testing the connectivity of two trees plus an edge. In the most simple-minded approach, this can be implemented by breadth-ﬁrst or depth-ﬁrst search in a sparse graph with at most n edges and n vertices, thus yielding an O(mn) algorithm. However, a faster implementation results if we can implement the component test in faster than O(n) time. In fact, a clever data structure called union-ﬁnd, can support such queries in O(lg n) time. Union-ﬁnd is discussed in the next section. With this data structure, Kruskal’s algorithm runs in O(m lg m) time, which is faster than Prim’s for sparse graphs. Observe again the impact that the right data structure can have when implementing a straightforward algorithm. Implementation The implementation of the main routine follows fairly directly from the psuedocode: 197 198 6. WEIGHTED GRAPH ALGORITHMS kruskal(graph *g) { int i; set_union s; edge_pair e[MAXV+1]; bool weight_compare(); /* counter */ /* set union data structure */ /* array of edges data structure */ set_union_init(&s, g->nvertices); to_edge_array(g, e); /* sort edges by increasing cost */ qsort(&e,g->nedges,sizeof(edge_pair),weight_compare); for (i=0; i<(g->nedges); i++) { if (!same_component(s,e[i].x,e[i].y)) { printf("edge (%d,%d) in MST\n",e[i].x,e[i].y); union_sets(&s,e[i].x,e[i].y); } } } 6.1.3 The Union-Find Data Structure A set partition is a partitioning of the elements of some universal set (say the integers 1 to n) into a collection of disjointed subsets. Thus, each element must be in exactly one subset. Set partitions naturally arise in graph problems such as connected components (each vertex is in exactly one connected component) and vertex coloring (a person may be male or female, but not both or neither). Section 14.6 (page 456) presents algorithms for generating set partitions and related objects. The connected components in a graph can be represented as a set partition. For Kruskal’s algorithm to run eﬃciently, we need a data structure that eﬃciently supports the following operations: • Same component(v1 , v2 ) – Do vertices v1 and v2 occur in the same connected component of the current graph? • Merge components(C1 , C2 ) – Merge the given pair of connected components into one component in response to an edge between them. The two obvious data structures for this task each support only one of these operations eﬃciently. Explicitly labeling each element with its component number enables the same component test to be performed in constant time, but updating the component numbers after a merger would require linear time. Alternately, we can treat the merge components operation as inserting an edge in a graph, but 6.1 MINIMUM SPANNING TREES 4 3 1 2 6 1 2 3 4 5 6 7 0 4 0 0 3 4 2 5 7 (l) (r) Figure 6.5: Union-ﬁnd example: structure represented as trees (l) and array (r) then we must run a full graph traversal to identify the connected components on demand. The union-ﬁnd data structure represents each subset as a “backwards” tree, with pointers from a node to its parent. Each node of this tree contains a set element, and the name of the set is taken from the key at the root. For reasons that will become clear, we will also maintain the number of elements in the subtree rooted in each vertex v: typedef struct { int p[SET_SIZE+1]; int size[SET_SIZE+1]; int n; } set_union; /* parent element */ /* number of elements in subtree i */ /* number of elements in set */ We implement our desired component operations in terms of two simpler operations, union and ﬁnd: • Find(i) – Find the root of tree containing element i, by walking up the parent pointers until there is nowhere to go. Return the label of the root. • Union(i,j) – Link the root of one of the trees (say containing i) to the root of the tree containing the other (say j) so f ind(i) now equals f ind(j). We seek to minimize the time it takes to execute any sequence of unions and ﬁnds. Tree structures can be very unbalanced, so we must limit the height of our trees. Our most obvious means of control is the decision of which of the two component roots becomes the root of the combined component on each union. To minimize the tree height, it is better to make the smaller tree the subtree of the bigger one. Why? The height of all the nodes in the root subtree stay the same, while the height of the nodes merged into this tree all increase by one. Thus, merging in the smaller tree leaves the height unchanged on the larger set of vertices. 199 200 6. WEIGHTED GRAPH ALGORITHMS Implementation The implementation details are as follows: set_union_init(set_union *s, int n) { int i; /* counter */ for (i=1; i<=n; i++) { s->p[i] = i; s->size[i] = 1; } s->n = n; } int find(set_union *s, int x) { if (s->p[x] == x) return(x); else return( find(s,s->p[x]) ); } int union_sets(set_union *s, int s1, int s2) { int r1, r2; /* roots of sets */ r1 = find(s,s1); r2 = find(s,s2); if (r1 == r2) return; /* already in same set */ if (s->size[r1] >= s->size[r2]) { s->size[r1] = s->size[r1] + s->size[r2]; s->p[ r2 ] = r1; } else { s->size[r2] = s->size[r1] + s->size[r2]; s->p[ r1 ] = r2; } } bool same_component(set_union *s, int s1, int s2) { return ( find(s,s1) == find(s,s2) ); } 6.1 MINIMUM SPANNING TREES Analysis On each union, the tree with fewer nodes becomes the child. But how tall can such a tree get as a function of the number of nodes in it? Consider the smallest possible tree of height h. Single-node trees have height 1. The smallest tree of height-2 has two nodes; from the union of two single-node trees. When do we increase the height? Merging in single-node trees won’t do it, since they just become children of the rooted tree of height-2. Only when we merge two height-2 trees together do we get a tree of height-3, now with four nodes. See the pattern? We must double the number of nodes in the tree to get an extra unit of height. How many doublings can we do before we use up all n nodes? At most, lg2 n doublings can be performed. Thus, we can do both unions and ﬁnds in O(log n), good enough for Kruskal’s algorithm. In fact, union-ﬁnd can be done even faster, as discussed in Section 12.5 (page 385). 6.1.4 Variations on Minimum Spanning Trees This minimum spanning tree algorithm has several interesting properties that help solve several closely related problems: • Maximum Spanning Trees – Suppose an evil telephone company is contracted to connect a bunch of houses together; they will be paid a price proportional to the amount of wire they install. Naturally, they will build the most expensive spanning tree possible. The maximum spanning tree of any graph can be found by simply negating the weights of all edges and running Prim’s algorithm. The most negative tree in the negated graph is the maximum spanning tree in the original. Most graph algorithms do not adapt so easily to negative numbers. Indeed, shortest path algorithms have trouble with negative numbers, and certainly do not generate the longest possible path using this technique. • Minimum Product Spanning Trees – Suppose we seek the spanning tree that minimizes the product of edge weights, assuming all edge weights are positive. Since lg(a · b) = lg(a) + lg(b), the minimum spanning tree on a graph whose edge weights are replaced with their logarithms gives the minimum product spanning tree on the original graph. • Minimum Bottleneck Spanning Tree – Sometimes we seek a spanning tree that minimizes the maximum edge weight over all such trees. In fact, every minimum spanning tree has this property. The proof follows directly from the correctness of Kruskal’s algorithm. Such bottleneck spanning trees have interesting applications when the edge weights are interpreted as costs, capacities, or strengths. A less eﬃcient 201 202 6. WEIGHTED GRAPH ALGORITHMS but conceptually simpler way to solve such problems might be to delete all “heavy” edges from the graph and ask whether the result is still connected. These kind of tests can be done with simple BFS/DFS. The minimum spanning tree of a graph is unique if all m edge weights in the graph are distinct. Otherwise the order in which Prim’s/Kruskal’s algorithm breaks ties determines which minimum spanning tree is returned. There are two important variants of a minimum spanning tree that are not solvable with these techniques. • Steiner Tree – Suppose we want to wire a bunch of houses together, but have the freedom to add extra intermediate vertices to serve as a shared junction. This problem is known as a minimum Steiner tree, and is discussed in the catalog in Section 16.10. • Low-degree Spanning Tree – Alternately, what if we want to ﬁnd the minimum spanning tree where the highest degree node in the tree is small? The lowest max-degree tree possible would be a simple path, and have n − 2 nodes of degree 2 with two endpoints of degree 1. A path that visits each vertex once is called a Hamiltonian path, and is discussed in the catalog in Section 16.5. 6.2 War Story: Nothing but Nets I’d been tipped oﬀ about a small printed-circuit board testing company nearby in need of some algorithmic consulting. And so I found myself inside a nondescript building in a nondescript industrial park, talking with the president of Integri-Test and one of his lead technical people. “We’re leaders in robotic printed-circuit board testing devices. Our customers have very high reliability requirements for their PC-boards. They must check that each and every board has no wire breaks before ﬁlling it with components. This means testing that each and every pair of points on the board that are supposed to be connected are connected.” “How do you do the testing?” I asked. “We have a robot with two arms, each with electric probes. The arms simultaneously contact both of the points to test whether two points are properly connected. If they are properly connected, then the probes will complete a circuit. For each net, we hold one arm ﬁxed at one point and move the other to cover the rest of the points.” “Wait!” I cried. “What is a net?” 6.2 (a) (b) WAR STORY: NOTHING BUT NETS (c) 203 (d) Figure 6.6: An example net showing (a) the metal connection layer, (b) the contact points, (c) their minimum spanning tree, and (d) the points partitioned into clusters “Circuit boards are certain sets of points that are all connected together with a metal layer. This is what we mean by a net. Sometimes a net consists of two points—i.e. , an isolated wire. Sometimes a net can have 100 to 200 points, like all the connections to power or ground.” “I see. So you have a list of all the connections between pairs of points on the circuit board, and you want to trace out these wires.” He shook his head. “Not quite. The input for our testing program consists only of the net contact points, as shown in Figure 6.6(b). We don’t know where the actual wires are, but we don’t have to. All we must do is verify that all the points in a net are connected together. We do this by putting the left robot arm on the leftmost point in the net, and then have the right arm move around to all the other points in the net to test if they are all connected to the left point. So they must all be connected to each other.” I thought for a moment about what this meant. “OK. So your right arm has to visit all the other points in the net. How do you choose the order to visit them?” The technical guy spoke up. “Well, we sort the points from left to right and then go in that order. Is that a good thing to do?” “Have you ever heard of the traveling salesman problem?” I asked. He was an electrical engineer, not a computer scientist. “No, what’s that?” “Traveling salesman is the name of the problem that you are trying to solve. Given a set of points to visit, how do you order them to minimize the travel time. Algorithms for the traveling salesman problem have been extensively studied. For small nets, you will be able to ﬁnd the optimal tour by doing an exhaustive search. For big nets, there are heuristics that will get you very close to the optimal tour.” I would have pointed them to Section 16.4 (page 533) if I had had this book handy. 204 6. WEIGHTED GRAPH ALGORITHMS The president scribbled down some notes and then frowned. “Fine. Maybe you can order the points in a net better for us. But that’s not our real problem. When you watch our robot in action, the right arm sometimes has to run all the way to the right side of the board on a given net, while the left arm just sits there. It seems we would beneﬁt by breaking nets into smaller pieces to balance things out.” I sat down and thought. The left and right arms each have interlocking TSP problems to solve. The left arm would move between the leftmost points of each net, while the right arm to visits all the other points in each net as ordered by the left TSP tour. By breaking each net into smaller nets we would avoid making the right arm cross all the way across the board. Further, a lot of little nets meant there would be more points in the left TSP, so each left-arm movement was likely to be short, too. “You are right. We should win if we can break big nets into small nets. We want the nets to be small, both in the number of points and in physical area. But we must be sure that if we validate the connectivity of each small net, we will have conﬁrmed that the big net is connected. One point in common between two little nets is suﬃcient to show that the bigger net formed by the two little nets is connected, since current can ﬂow between any pair of points.” Now we had to break each net into overlapping pieces, where each piece was small. This is a clustering problem. Minimum spanning trees are often used for clustering, as discussed in Section 15.3 (page 484). In fact, that was the answer! We could ﬁnd the minimum spanning tree of the net points and break it into little clusters whenever a spanning tree edge got too long. As shown in Figure 6.6(d), each cluster would share exactly one point in common with another cluster, with connectivity ensured because we are covering the edges of a spanning tree. The shape of the clusters will reﬂect the points in the net. If the points lay along a line across the board, the minimum spanning tree would be a path, and the clusters would be pairs of points. If the points all fell in a tight region, there would be one nice fat cluster for the right arm to scoot around. So I explained the idea of constructing the minimum spanning tree of a graph. The boss listened, scribbled more notes, and frowned again. “I like your clustering idea. But minimum spanning trees are deﬁned on graphs. All you’ve got are points. Where do the weights of the edges come from?” “Oh, we can think of it as a complete graph, where every pair of points are connected. The weight of the edge is deﬁned as the distance between the two points. Or is it. . . ?” I went back to thinking. The edge cost should reﬂect the travel time between between two points. While distance is related to travel time, it wasn’t necessarily the same thing. “Hey. I have a question about your robot. Does it take the same amount of time to move the arm left-right as it does up-down?” They thought a minute. “Yeah, it does. We use the same type of motor to control horizontal and vertical movements. Since the two motors for each arm are 6.3 SHORTEST PATHS independent, we can simultaneously move each arm both horizontally and vertically.” “So the time to move both one foot left and one foot up is exactly the same as just moving one foot left? This means that the weight for each edge should not be the Euclidean distance between the two points, but the biggest diﬀerence between either the x− or y-coordinate. This is something we call the L∞ metric, but we can capture it by changing the edge weights in the graph. Anything else funny about your robots?” I asked. “Well, it takes some time for the robot to come up to speed. We should probably also factor in acceleration and deceleration of the arms.” “Darn right. The more accurately you can model the time your arm takes to move between two points, the better our solution will be. But now we have a very clean formulation. Let’s code it up and let’s see how well it works!” They were somewhat skeptical whether this approach would do any good, but agreed to think about it. A few weeks later they called me back and reported that the new algorithm reduced the distance traveled by about 30% over their previous approach, at a cost of a little more computational preprocessing. However, since their testing machine cost $200,000 a pop and a PC cost $2,000, this was an excellent tradeoﬀ. It is particularly advantageous since the preprocessing need only be done once when testing multiple instances of a particular board design. The key idea leading to the successful solution was modeling the job in terms of classical algorithmic graph problems. I smelled TSP the instant they started talking about minimizing robot motion. Once I realized that they were implicitly forming a star-shaped spanning tree to ensure connectivity, it was natural to ask whether the minimum spanning tree would perform any better. This idea led to clustering, and thus partitioning each net into smaller nets. Finally, by carefully designing our distance metric to accurately model the costs of the robot itself, we could incorporate such complicated properties (as acceleration) without changing our fundamental graph model or algorithm design. Take-Home Lesson: Most applications of graphs can be reduced to standard graph properties where well-known algorithms can be used. These include minimum spanning trees, shortest paths, and other problems presented in the catalog. 6.3 Shortest Paths A path is a sequence of edges connecting two vertices. Since movie director Mel Brooks (“The Producers”) is my father’s sister’s husband’s cousin, there is a path in the friendship graph between me and him, shown in Figure 6.7—even though the two of us have never met. But if I were trying to impress how tight I am with Cousin Mel, I would be much better oﬀ saying that my Uncle Lenny grew up with him. I have a friendship path of length 2 to Cousin Mel through Uncle Lenny, while 205 206 6. WEIGHTED GRAPH ALGORITHMS Steve Dad Aunt Eve Uncle Lenny Cousin Mel Figure 6.7: Mel Brooks is my father’s sister’s husband’s cousin s t Figure 6.8: The shortest path from s to t may pass through many intermediate vertices the path is of length 4 by blood and marriage. This multiplicity of paths hints at why ﬁnding the shortest path between two nodes is important and instructive, even in nontransportation applications. The shortest path from s to t in an unweighted graph can be constructed using a breadth-ﬁrst search from s. The minimum-link path is recorded in the breadth-ﬁrst search tree, and it provides the shortest path when all edges have equal weight. However, BFS does not suﬃce to ﬁnd shortest paths in weighted graphs. The shortest weighted path might use a large number of edges, just as the shortest route (timewise) from home to oﬃce may involve complicated shortcuts using backroads, as shown in Figure 6.8. In this section, we will present two distinct algorithms for ﬁnding the shortest paths in weighted graphs. 6.3.1 Dijkstra’s Algorithm Dijkstra’s algorithm is the method of choice for ﬁnding shortest paths in an edgeand/or vertex-weighted graph. Given a particular start vertex s, it ﬁnds the shortest path from s to every other vertex in the graph, including your desired destination t. Suppose the shortest path from s to t in graph G passes through a particular intermediate vertex x. Clearly, this path must contain the shortest path from s to x as its preﬁx, because if not, we could shorten our s-to-t path by using the shorter 6.3 SHORTEST PATHS s-to-x preﬁx. Thus, we must compute the shortest path from s to x before we ﬁnd the path from s to t. Dijkstra’s algorithm proceeds in a series of rounds, where each round establishes the shortest path from s to some new vertex. Speciﬁcally, x is the vertex that minimizes dist(s, vi ) + w(vi , x) over all unﬁnished 1 ≤ i ≤ n, where w(i, j) is the length of the edge from i to j, and dist(i, j) is the length of the shortest path between them. This suggests a dynamic programming-like strategy. The shortest path from s to itself is trivial unless there are negative weight edges, so dist(s, s) = 0. If (s, y) is the lightest edge incident to s, then this implies that dist(s, y) = w(s, y). Once we determine the shortest path to a node x, we check all the outgoing edges of x to see whether there is a better path from s to some unknown vertex through x. ShortestPath-Dijkstra(G, s, t) known = {s} for i = 1 to n, dist[i] = ∞ for each edge (s, v), dist[v] = w(s, v) last = s while (last = t) select vnext , the unknown vertex minimizing dist[v] for each edge (vnext , x), dist[x] = min[dist[x], dist[vnext ] + w(vnext , x)] last = vnext known = known ∪ {vnext } The basic idea is very similar to Prim’s algorithm. In each iteration, we add exactly one vertex to the tree of vertices for which we know the shortest path from s. As in Prim’s, we keep track of the best path seen to date for all vertices outside the tree, and insert them in order of increasing cost. The diﬀerence between Dijkstra’s and Prim’s algorithms is how they rate the desirability of each outside vertex. In the minimum spanning tree problem, all we cared about was the weight of the next potential tree edge. In shortest path, we want to include the closest outside vertex (in shortest-path distance) to s. This is a function of both the new edge weight and the distance from s to the tree vertex it is adjacent to. Implementation The pseudocode actually obscures how similar the two algorithms are. In fact, the change is very minor. Below, we give an implementation of Dijkstra’s algorithm based on changing exactly three lines from our Prim’s implementation—one of which is simply the name of the function! 207 208 6. WEIGHTED GRAPH ALGORITHMS dijkstra(graph *g, int start) { int i; edgenode *p; bool intree[MAXV+1]; int distance[MAXV+1]; int v; int w; int weight; int dist; /* WAS prim(g,start) */ /* /* /* /* /* /* /* /* counter */ temporary pointer */ is the vertex in the tree yet? */ distance vertex is from start */ current vertex to process */ candidate next vertex */ edge weight */ best current distance from start */ for (i=1; i<=g->nvertices; i++) { intree[i] = FALSE; distance[i] = MAXINT; parent[i] = -1; } distance[start] = 0; v = start; while (intree[v] == FALSE) { intree[v] = TRUE; p = g->edges[v]; while (p != NULL) { w = p->y; weight = p->weight; /* CHANGED */ if (distance[w] > (distance[v]+weight)) { /* CHANGED */ distance[w] = distance[v]+weight; /* CHANGED */ parent[w] = v; } p = p->next; } v = 1; dist = MAXINT; for (i=1; i<=g->nvertices; i++) if ((intree[i] == FALSE) && (dist > distance[i])) { dist = distance[i]; v = i; } } } 6.3 SHORTEST PATHS This algorithm ﬁnds more than just the shortest path from s to t. It ﬁnds the shortest path from s to all other vertices. This deﬁnes a shortest path spanning tree rooted in s. For undirected graphs, this would be the breadth-ﬁrst search tree, but in general it provides the shortest path from s to all other vertices. Analysis What is the running time of Dijkstra’s algorithm? As implemented here, the complexity is O(n2 ). This is the same running time as a proper version of Prim’s algorithm; except for the extension condition it is the same algorithm as Prim’s. The length of the shortest path from start to a given vertex t is exactly the value of distance[t]. How do we use dijkstra to ﬁnd the actual path? We follow the backward parent pointers from t until we hit start (or -1 if no such path exists), exactly as was done in the find path() routine of Section 5.6.2 (page 165). Dijkstra works correctly only on graphs without negative-cost edges. The reason is that midway through the execution we may encounter an edge with weight so negative that it changes the cheapest way to get from s to some other vertex already in the tree. Indeed, the most cost-eﬀective way to get from your house to your next-door neighbor would be repeatedly through the lobby of any bank oﬀering you enough money to make the detour worthwhile. Most applications do not feature negative-weight edges, making this discussion academic. Floyd’s algorithm, discussed below, works correctly unless there are negative cost cycles, which grossly distort the shortest-path structure. Unless that bank limits its reward to one per customer, you might so beneﬁt by making an inﬁnite number of trips through the lobby that you would never decide to actually reach your destination! Stop and Think: Shortest Path with Node Costs Problem: Suppose we are given a graph whose weights are on the vertices, instead of the edges. Thus, the cost of a path from x to y is the sum of the weights of all vertices on the path. Give an eﬃcient algorithm for ﬁnding shortest paths on vertex-weighted graphs. Solution: A natural idea would be to adapt the algorithm we have for edge-weighted graphs (Dijkstra’s) to the new vertex-weighted domain. It should be clear that we can do it. We replace any reference to the weight of an edge with the weight of the destination vertex. This can be looked up as needed from an array of vertex weights. However, my preferred approach would leave Dijkstra’s algorithm intact and instead concentrate on constructing an edge-weighted graph on which Dijkstra’s 209 210 6. WEIGHTED GRAPH ALGORITHMS algorithm will give the desired answer. Set the weight of each directed edge (i, j) in the input graph to the cost of vertex j. Dijkstra’s algorithm now does the job. This technique can be extended to a variety of diﬀerent domains, such as when there are costs on both vertices and edges. 6.3.2 All-Pairs Shortest Path Suppose you want to ﬁnd the “center” vertex in a graph—the one that minimizes the longest or average distance to all the other nodes. This might be the best place to start a new business. Or perhaps you need to know a graph’s diameter—the longest shortest-path distance over all pairs of vertices. This might correspond to the longest possible time it takes a letter or network packet to be delivered. These and other applications require computing the shortest path between all pairs of vertices in a given graph. We could solve all-pairs shortest path by calling Dijkstra’s algorithm from each of the n possible starting vertices. But Floyd’s all-pairs shortest-path algorithm is a slick way to construct this n × n distance matrix from the original weight matrix of the graph. Floyd’s algorithm is best employed on an adjacency matrix data structure, which is no extravagance since we must store all n2 pairwise distances anyway. Our adjacency matrix type allocates space for the largest possible matrix, and keeps track of how many vertices are in the graph: typedef struct { int weight[MAXV+1][MAXV+1]; int nvertices; } adjacency_matrix; /* adjacency/weight info */ /* number of vertices in graph */ The critical issue in an adjacency matrix implementation is how we denote the edges absent from the graph. A common convention for unweighted graphs denotes graph edges by 1 and non-edges by 0. This gives exactly the wrong interpretation if the numbers denote edge weights, for the non-edges get interpreted as a free ride between vertices. Instead, we should initialize each non-edge to MAXINT. This way we can both test whether it is present and automatically ignore it in shortest-path computations, since only real edges will be used, provided MAXINT is less than the diameter of your graph. There are several ways to characterize the shortest path between two nodes in a graph. The Floyd-Warshall algorithm starts by numbering the vertices of the graph from 1 to n. We use these numbers not to label the vertices, but to order them. Deﬁne W [i, j]k to be the length of the shortest path from i to j using only vertices numbered from 1, 2, ..., k as possible intermediate vertices. What does this mean? When k = 0, we are allowed no intermediate vertices, so the only allowed paths are the original edges in the graph. Thus the initial 6.3 SHORTEST PATHS all-pairs shortest-path matrix consists of the initial adjacency matrix. We will perform n iterations, where the kth iteration allows only the ﬁrst k vertices as possible intermediate steps on the path between each pair of vertices x and y. At each iteration, we allow a richer set of possible shortest paths by adding a new vertex as a possible intermediary. Allowing the kth vertex as a stop helps only if there is a short path that goes through k, so W [i, j]k = min(W [i, j]k−1 , W [i, k]k−1 + W [k, j]k−1 ) The correctness of this is somewhat subtle, and I encourage you to convince yourself of it. But there is nothing subtle about how simple the implementation is: floyd(adjacency_matrix *g) { int i,j; int k; int through_k; /* dimension counters */ /* intermediate vertex counter */ /* distance through vertex k */ for (k=1; k<=g->nvertices; k++) for (i=1; i<=g->nvertices; i++) for (j=1; j<=g->nvertices; j++) { through_k = g->weight[i][k]+g->weight[k][j]; if (through_k < g->weight[i][j]) g->weight[i][j] = through_k; } } The Floyd-Warshall all-pairs shortest path runs in O(n3 ) time, which is asymptotically no better than n calls to Dijkstra’s algorithm. However, the loops are so tight and the program so short that it runs better in practice. It is notable as one of the rare graph algorithms that work better on adjacency matrices than adjacency lists. The output of Floyd’s algorithm, as it is written, does not enable one to reconstruct the actual shortest path between any given pair of vertices. These paths can be recovered if we retain a parent matrix P of our choice of the last intermediate vertex used for each vertex pair (x, y). Say this value is k. The shortest path from x to y is the concatenation of the shortest path from x to k with the shortest path from k to y, which can be reconstructed recursively given the matrix P . Note, however, that most all-pairs applications need only the resulting distance matrix. These jobs are what Floyd’s algorithm was designed for. 211 212 6. WEIGHTED GRAPH ALGORITHMS 6.3.3 Transitive Closure Floyd’s algorithm has another important application, that of computing transitive closure. In analyzing a directed graph, we are often interested in which vertices are reachable from a given node. As an example, consider the blackmail graph, where there is a directed edge (i, j) if person i has sensitive-enough private information on person j so that i can get j to do whatever he wants. You wish to hire one of these n people to be your personal representative. Who has the most power in terms of blackmail potential? A simplistic answer would be the vertex of highest degree, but an even better representative would be the person who has blackmail chains leading to the most other parties. Steve might only be able to blackmail Miguel directly, but if Miguel can blackmail everyone else then Steve is the man you want to hire. The vertices reachable from any single node can be computed using breadthﬁrst or depth-ﬁrst searches. But the whole batch can be computed using an all-pairs shortest-path. If the shortest path from i to j remains MAXINT after running Floyd’s algorithm, you can be sure no directed path exists from i to j. Any vertex pair of weight less than MAXINT must be reachable, both in the graph-theoretic and blackmail senses of the word. Transitive closure is discussed in more detail in the catalog in Section 15.5. 6.4 War Story: Dialing for Documents I was part of a group visiting Periphonics, which was then an industry leader in building telephone voice-response systems. These are more advanced versions of the Press 1 for more options, Press 2 if you didn’t press 1 telephone systems that blight everyone’s lives. We were being given the standard tour when someone from our group asked, “Why don’t you guys use voice recognition for data entry. It would be a lot less annoying than typing things out on the keypad.” The tour guide reacted smoothly. “Our customers have the option of incorporating speech recognition into our products, but very few of them do. Userindependent, connected-speech recognition is not accurate enough for most applications. Our customers prefer building systems around typing text on the telephone keyboards.” “Prefer typing, my pupik!” came a voice from the rear of our group. “I hate typing on a telephone. Whenever I call my brokerage house to get stock quotes some machine tells me to type in the three letter code. To make things worse, I have to hit two buttons to type in one letter, in order to distinguish between the three letters printed on each key of the telephone. I hit the 2 key and it says Press 1 for A, Press 2 for B, Press 3 for C. Pain in the neck if you ask me.” “Maybe you don’t have to hit two keys for each letter!” I chimed in. “Maybe the system could ﬁgure out the correct letter from context!” 6.4 WAR STORY: DIALING FOR DOCUMENTS “There isn’t a whole lot of context when you type in three letters of stock market code.” “Sure, but there would be plenty of context if we typed in English sentences. I’ll bet that we could reconstruct English text correctly if they were typed in a telephone at one keystroke per letter.” The guy from Periphonics gave me a disinterested look, then continued the tour. But when I got back to the oﬃce, I decided to give it a try. Not all letters are equally likely to be typed on a telephone. In fact, not all letters can be typed, since Q and Z are not labeled on a standard American telephone. Therefore, we adopted the convention that Q, Z, and “space” all sat on the * key. We could take advantage of the uneven distribution of letter frequencies to help us decode the text. For example, if you hit the 3 key while typing English, you more likely meant to type an E than either a D or F. By taking into account the frequencies of a window of three characters (trigrams), we could predict the typed text. This is what happened when I tried it on the Gettysburg Address: enurraore ane reten yeasr ain our ectherr arotght eosti on ugis aootinent a oey oation aoncdivee in licesty ane eedicatee un uhe rrorosition uiat all oen are arectee e ual ony ye are enichde in a irect aitil yar uestini yhethes uiat oatioo or aoy oation ro aoncdivee ane ro eedicatee aan loni eneure ye are oet on a irect aattlediele oe uiat yar ye iate aone un eedicate a rostion oe uiat eiele ar a einal restini rlace eor uiore yin iere iate uhdis lives uiat uhe oation ogght live it is aluniethes eittini ane rrores uiat ye rioule en ugir att in a laries reore ye aan oou eedicate ye aan oou aoorearate ye aan oou ialloy ugis iroune the arate oen litini ane eeae yin rustgilee iere iate aoorearatee it ear aante our roor rowes un ade or eeuraat the yople yill little oote oor loni renences yiat ye ray iere att it aan oetes eosiet yiat uhfy eie iere it is eor ur uhe litini rathes un ae eedicatee iere un uhe undiniside yopl yhici uhfy yin entght iere iate uiur ear ro onaky aetancde it is rathes eor ur un ae iere eedicatee un uhe irect uarl rencinini adeore ur uiat eron uhere ioooree eeae ye uale inarearee eeuotion uo tiat aaure eor yhici uhfy iere iate uhe lart eull oearure oe eeuotioo tiat ye iere iggily rerolue uiat uhere eeae riall oou iate eide io The trigram statistics did a decent job of translating it into Greek, but a terrible job of transcribing English. One reason was clear. This algorithm knew nothing about English words. If we coupled it with a dictionary, we might be onto something. But two words in the dictionary are often represented by the exact same string of phone codes. For an extreme example, the code string “22737” collides with eleven distinct English words, including cases, cares, cards, capes, caper, and bases. For our next attempt, we reported the unambiguous characters of any words that collided in the dictionary, and used trigrams to ﬁll in the rest of the characters. We were rewarded with: eourscore and seven yearr ain our eatherr brought forth on this continent azoey nation conceivee in liberty and dedicatee uo uhe proposition that all men are createe equal ony ye are engagee in azipeat civil yar uestioi whether that nation or aoy nation ro conceivee and ro dedicatee aan long endure ye are oet on azipeat battleﬁeld oe that yar ye iate aone uo dedicate a rostion oe that ﬁeld ar a ﬁnal perthni place for those yin here iate their lives that uhe nation oight live it is altogether ﬁttinizane proper that ye should en this aut in a larges sense ye aan oou dedicate ye aan oou consecrate ye aan oou hallow this ground the arate men litioi and deae yin strugglee here iate consecratee it ear above our roor power uo ade or detract the world will little oote oor long remember what ye 213 214 6. WEIGHTED GRAPH ALGORITHMS ray here aut it aan meter forget what uhfy die here it is for ur uhe litioi rather uo ae dedicatee here uo uhe toeioisgee york which uhfy yin fought here iate thus ear ro mocky advancee it is rather for ur uo ae here dedicatee uo uhe great task renagogoi adfore ur that from there honoree deae ye uale increasee devotion uo that aause for which uhfy here iate uhe last eull measure oe devotion that ye here highky resolve that there deae shall oou iate ﬁde io vain that this nation under ioe shall iate azoey birth oe freedom and that ioternmenu oe uhe people ay uhe people for uhe people shall oou perish from uhe earth If you were a student of American history, maybe you could recognize it, but you certainly couldn’t read it. Somehow, we had to distinguish between the diﬀerent dictionary words that got hashed to the same code. We could factor in the relative popularity of each word, but this still made too many mistakes. At this point, I started working with Harald Rau on the project, who proved to be a great collaborator. First, he was a bright and persistent graduate student. Second, as a native German speaker, he believed every lie I told him about English grammar. Harald built up a phone code reconstruction program along the lines of Figure 6.9. It worked on the input one sentence at a time, identifying dictionary words that matched each code string. The key problem was how to incorporate grammatical constraints. “We can get good word-use frequencies and grammatical information from a big text database called the Brown Corpus. It contains thousands of typical English sentences, each parsed according to parts of speech. But how do we factor it all in?” Harald asked. “Let’s think about it as a graph problem,” I suggested. “Graph problem? What graph problem? Where is there even a graph?” “Think of a sentence as a series of phone tokens, each representing a word in the sentence. Each phone token has a list of words from the dictionary that match it. How can we choose which one is right? Each possible sentence interpretation can be thought of as a path in a graph. The vertices of this graph are the complete set of possible word choices. There will be an edge from each possible choice for the ith word to each possible choice for the (i + 1)st word. The cheapest path across this graph deﬁnes the best interpretation of the sentence.” “But all the paths look the same. They have the same number of edges. Wait. Now I see! We have to add weight to the edges to make the paths diﬀerent.” “Exactly! The cost of an edge will reﬂect how likely it is that we will travel through the given pair of words. Perhaps we can count how often that pair of words occurred together in previous texts. Or we can weigh them by the part of speech of each word. Maybe nouns don’t like to be next to nouns as much as they like being next to verbs.” “It will be hard to keep track of word-pair statistics, since there are so many of them. But we certainly know the frequency of each word. How can we factor that into things?” 6.4 INPUT ? Blank Recognition WAR STORY: DIALING FOR DOCUMENTS ... # 4 4 8 3 ∗ 6 3 ∗ 2 ∗ 7 4 6 4 # ... Token “4483” ? • • Token Token “63” • • Token Token “2” • • Token Token “7464” • • Candidate Construction Token “4483” • • ? “63” • • ? give of hive me “2” • • “7464” ? • • ? a ping ring ? sing Sentence Disambiguating give hive PP q P of me 1 a PP q P ping ring sing ? OUTPUT GIVE ME A RING. Figure 6.9: The phases of the telephone code reconstruction process “We can pay a cost for walking through a particular vertex that depends upon the frequency of the word. Our best sentence will be given by the shortest path across the graph.” “But how do we ﬁgure out the relative weights of these factors?” “First try what seems natural to you and then we can experiment with it.” Harald incorporated this shortest-path algorithm. With proper grammatical and statistical constraints, the system performed great. Look at the Gettysburg Address now, with all the reconstruction errors highlighted: FOURSCORE AND SEVEN YEARS AGO OUR FATHERS BROUGHT FORTH ON THIS CONTINENT A NEW NATION CONCEIVED IN LIBERTY AND DEDICATED TO THE PROPOSITION THAT ALL MEN ARE CREATED EQUAL. NOW WE ARE 215 216 6. WEIGHTED GRAPH ALGORITHMS Code C 1 Code C 2 Code C 3 P(W11/C 1) 2 P(W1/C 2) P(W1/C 3) 3 P(# P(W11/#) P(W21/C 1) P(W21/#) # P(W31/#) 2 1 P(W2/W1) P(W23/C 3) P(W22/C 2) P(W31/C 1) P(W33/C 3) P(W41/C 1) P(W43/C 3) P(W41/#) Figure 6.10: The minimum-cost path deﬁnes the best interpretation for a sentence Text Clinton Speeches Herland Moby Dick Bible Shakespeare characters 1,073,593 278,670 1,123,581 3,961,684 4,558,202 characters correct 99.04% 98.24% 96.85% 96.20% 95.20% non-blanks correct 98.86% 97.89% 96.25% 95.39% 94.21% words correct 97.67% 97.02% 94.75% 95.39% 92.86% time per character 0.97ms 0.97ms 1.14ms 1.33ms 0.99ms Figure 6.11: Telephone-code reconstruction applied to several text samples ENGAGED IN A GREAT CIVIL WAR TESTING WHETHER THAT NATION OR ANY NATION SO CONCEIVED AND SO DEDICATED CAN LONG ENDURE. WE ARE MET ON A GREAT BATTLEFIELD OF THAT WAS. WE HAVE COME TO DEDICATE A PORTION OF THAT FIELD AS A FINAL SERVING PLACE FOR THOSE WHO HERE HAVE THEIR LIVES THAT THE NATION MIGHT LIVE. IT IS ALTOGETHER FITTING AND PROPER THAT WE SHOULD DO THIS. BUT IN A LARGER SENSE WE CAN NOT DEDICATE WE CAN NOT CONSECRATE WE CAN NOT HALLOW THIS GROUND. THE BRAVE MEN LIVING AND DEAD WHO STRUGGLED HERE HAVE CONSECRATED IT FAR ABOVE OUR POOR POWER TO ADD OR DETRACT. THE WORLD WILL LITTLE NOTE NOR LONG REMEMBER WHAT WE SAY HERE BUT IT CAN NEVER FORGET WHAT THEY DID HERE. IT IS FOR US THE LIVING RATHER TO BE DEDICATED HERE TO THE UNFINISHED WORK WHICH THEY WHO FOUGHT HERE HAVE THUS FAR SO NOBLY ADVANCED. IT IS RATHER FOR US TO BE HERE DEDICATED TO THE GREAT TASK REMAINING BEFORE US THAT FROM THESE HONORED DEAD WE TAKE INCREASED DEVOTION TO THAT CAUSE FOR WHICH THEY HERE HAVE THE LAST FULL MEASURE OF DEVOTION THAT WE HERE HIGHLY RESOLVE THAT THESE DEAD SHALL NOT HAVE DIED IN VAIN THAT THIS NATION UNDER GOD SHALL HAVE A NEW BIRTH OF FREEDOM AND THAT GOVERNMENT OF THE PEOPLE BY THE PEOPLE FOR THE PEOPLE SHALL NOT PERISH FROM THE EARTH. 6.5 NETWORK FLOWS AND BIPARTITE MATCHING s 217 t Figure 6.12: Bipartite graph with a maximum matching highlighted (on left). The corresponding network ﬂow instance highlighting the maximum s − t ﬂow (on right). While we still made a few mistakes, the results are clearly good enough for many applications. Periphonics certainly thought so, for they licensed our program to incorporate into their products. Figure 6.11 shows that we were able to reconstruct correctly over 99% of the characters in a megabyte of President Clinton’s speeches, so if Bill had phoned them in, we would certainly be able to understand what he was saying. The reconstruction time is fast enough, indeed faster than you can type it in on the phone keypad. The constraints for many pattern recognition problems can be naturally formulated as shortest path problems in graphs. In fact, there is a particularly convenient dynamic programming solution for these problems (the Viterbi algorithm) that is widely used in speech and handwriting recognition systems. Despite the fancy name, the Viterbi algorithm is basically solving a shortest path problem on a DAG. Hunting for a graph formulation to solve any given problem is often a good idea. 6.5 Network Flows and Bipartite Matching Edge-weighted graphs can be interpreted as a network of pipes, where the weight of edge (i, j) determines the capacity of the pipe. Capacities can be thought of as a function of the cross-sectional area of the pipe. A wide pipe might be able to carry 10 units of ﬂow in a given time, where as a narrower pipe might only carry 5 units. The network ﬂow problem asks for the maximum amount of ﬂow which can be sent from vertices s to t in a given weighted graph G while respecting the maximum capacities of each pipe. 218 6. WEIGHTED GRAPH ALGORITHMS 6.5.1 Bipartite Matching While the network ﬂow problem is of independent interest, its primary importance is in to solving other important graph problems. A classic example is bipartite matching. A matching in a graph G = (V, E) is a subset of edges E ⊂ E such that no two edges of E share a vertex. A matching pairs oﬀ certain vertices such that every vertex is in, at most, one such pair, as shown in Figure 6.12. Graph G is bipartite or two-colorable if the vertices can be divided into two sets, L and R, such that all edges in G have one vertex in L and one vertex in R. Many naturally deﬁned graphs are bipartite. For example, certain vertices may represent jobs to be done and the remaining vertices represent people who can potentially do them. The existence of edge (j, p) means that job j can be done by person p. Or let certain vertices represent boys and certain vertices represent girls, with edges representing compatible pairs. Matchings in these graphs have natural interpretations as job assignments or as marriages, and are the focus of Section 15.6 (page 498). The largest bipartite matching can be readily found using network ﬂow. Create a source node s that is connected to every vertex in L by an edge of weight 1. Create a sink node t and connect it to every vertex in R by an edge of weight 1. Finally, assign each edge in the bipartite graph G a weight of 1. Now, the maximum possible ﬂow from s to t deﬁnes the largest matching in G. Certainly we can ﬁnd a ﬂow as large as the matching by using only the matching edges and their sourceto-sink connections. Further, there can be no greater possible ﬂow. How can we ever hope to get more than one ﬂow unit through any vertex? 6.5.2 Computing Network Flows Traditional network ﬂow algorithms are based on the idea of augmenting paths, and repeatedly ﬁnding a path of positive capacity from s to t and adding it to the ﬂow. It can be shown that the ﬂow through a network is optimal if and only if it contains no augmenting path. Since each augmentation adds to the ﬂow, we must eventually ﬁnd the global maximum. The key structure is the residual ﬂow graph, denoted as R(G, f ), where G is the input graph and f is the current ﬂow through G. This directed, edge-weighted R(G, f ) contains the same vertices as G. For each edge (i, j) in G with capacity c(i, j) and ﬂow f (i, j), R(G, f ) may contain two edges: (i) an edge (i, j) with weight c(i, j) − f (i, j), if c(i, j) − f (i, j) > 0 and (ii) an edge (j, i) with weight f (i, j), if f (i, j) > 0. The presence of edge (i, j) in the residual graph indicates that positive ﬂow can be pushed from i to j. The weight of the edge gives the exact amount that can be pushed. A path in the residual ﬂow graph from s to t implies that more ﬂow can be pushed from s to t and the minimum edge weight on this path deﬁnes the amount of extra ﬂow that can be pushed. 6.5 NETWORK FLOWS AND BIPARTITE MATCHING T 5 7 3 5 S 9 4 5 5 T 2 2 2 3 3 5 7 5 S G 3 9 10 12 219 2 4 2 R(G) Figure 6.13: Maximum s − t ﬂow in a graph G (on left) showing the associated residual graph R(G) and minimum s − t cut (dotted line near t) Figure 6.13 illustrates this idea. The maximum s − t ﬂow in graph G is 7. Such a ﬂow is revealed by the two directed t to s paths in the residual graph R(G) of capacities 2 + 5, respectively. These ﬂows completely saturate the capacity of the two edges incident to vertex t, so no augmenting path remains. Thus the ﬂow is optimal. A set of edges whose deletion separates s from t (like the two edges incident to t) is called an s-t cut. Clearly, no s to t ﬂow can exceed the weight of the minimum such cut. In fact, a ﬂow equal to the minimum cut is always possible. Take-Home Lesson: The maximum ﬂow from s to t always equals the weight of the minimum s-t cut. Thus, ﬂow algorithms can be used to solve general edge and vertex connectivity problems in graphs. Implementation We cannot do full justice to the theory of network ﬂows here. However, it is instructive to see how augmenting paths can be identiﬁed and the optimal ﬂow computed. For each edge in the residual ﬂow graph, we must keep track of both the amount of ﬂow currently going through the edge, as well as its remaining residual capacity. Thus, we must modify our edge structure to accommodate the extra ﬁelds: typedef struct { int v; int capacity; int flow; int residual; struct edgenode *next; } edgenode; /* /* /* /* /* neighboring vertex */ capacity of edge */ flow through edge */ residual capacity of edge */ next edge in list */ 220 6. WEIGHTED GRAPH ALGORITHMS We use a breadth-ﬁrst search to look for any path from source to sink that increases the total ﬂow, and use it to augment the total. We terminate with the optimal ﬂow when no such augmenting path exists. netflow(flow_graph *g, int source, int sink) { int volume; /* weight of the augmenting path */ add_residual_edges(g); initialize_search(g); bfs(g,source); volume = path_volume(g, source, sink, parent); while (volume > 0) { augment_path(g,source,sink,parent,volume); initialize_search(g); bfs(g,source); volume = path_volume(g, source, sink, parent); } } Any augmenting path from source to sink increases the ﬂow, so we can use bfs to ﬁnd such a path in the appropriate graph. We only consider network edges that have remaining capacity or, in other words, positive residual ﬂow. The predicate below helps bfs distinguish between saturated and unsaturated edges: bool valid_edge(edgenode *e) { if (e->residual > 0) return (TRUE); else return(FALSE); } Augmenting a path transfers the maximum possible volume from the residual capacity into positive ﬂow. This amount is limited by the path-edge with the smallest amount of residual capacity, just as the rate at which traﬃc can ﬂow is limited by the most congested point. 6.5 NETWORK FLOWS AND BIPARTITE MATCHING int path_volume(flow_graph *g, int start, int end, int parents[]) { edgenode *e; /* edge in question */ edgenode *find_edge(); if (parents[end] == -1) return(0); e = find_edge(g,parents[end],end); if (start == parents[end]) return(e->residual); else return( min(path_volume(g,start,parents[end],parents), e->residual) ); } edgenode *find_edge(flow_graph *g, int x, int y) { edgenode *p; /* temporary pointer */ p = g->edges[x]; while (p != NULL) { if (p->v == y) return(p); p = p->next; } return(NULL); } Sending an additional unit of ﬂow along directed edge (i, j) reduces the residual capacity of edge (i, j) but increases the residual capacity of edge (j, i). Thus, the act of augmenting a path requires modifying both forward and reverse edges for each link on the path. 221 222 6. WEIGHTED GRAPH ALGORITHMS augment_path(flow_graph *g, int start, int end, int parents[], int volume) { edgenode *e; /* edge in question */ edgenode *find_edge(); if (start == end) return; e = find_edge(g,parents[end],end); e->flow += volume; e->residual -= volume; e = find_edge(g,end,parents[end]); e->residual += volume; augment_path(g,start,parents[end],parents,volume); } Initializing the ﬂow graph requires creating directed ﬂow edges (i, j) and (j, i) for each network edge e = (i, j). Initial ﬂows are all set to 0. The initial residual ﬂow of (i, j) is set to the capacity of e, while the initial residual ﬂow of (j, i) is set to 0. Analysis The augmenting path algorithm above eventually converges on the the optimal solution. However, each augmenting path may add only a little to the total ﬂow, so, in principle, the algorithm might take an arbitrarily long time to converge. However, Edmonds and Karp [EK72] proved that always selecting a shortest unweighted augmenting path guarantees that O(n3 ) augmentations suﬃce for optimization. In fact, the Edmonds-Karp algorithm is what is implemented above, since a breadth-ﬁrst search from the source is used to ﬁnd the next augmenting path. 6.6 Design Graphs, Not Algorithms Proper modeling is the key to making eﬀective use of graph algorithms. We have deﬁned several graph properties, and developed algorithms for computing them. All told, about two dozen diﬀerent graph problems are presented in the catalog, mostly in Sections 15 and 16. These classical graph problems provide a framework for modeling most applications. The secret is learning to design graphs, not algorithms. We have already seen a few instances of this idea: 6.6 DESIGN GRAPHS, NOT ALGORITHMS • The maximum spanning tree can be found by negating the edge weights of the input graph G and using a minimum spanning tree algorithm on the result. The most negative weight spanning tree will deﬁne the maximum weight tree in G. • To solve bipartite matching, we constructed a special network ﬂow graph such that the maximum ﬂow corresponds to a maximum cardinality matching. The applications below demonstrate the power of proper modeling. Each arose in a real-world application, and each can be modeled as a graph problem. Some of the modelings are quite clever, but they illustrate the versatility of graphs in representing relationships. As you read a problem, try to devise an appropriate graph representation before peeking to see how we did it. Stop and Think: The Pink Panther’s Passport to Peril Problem: “I’m looking for an algorithm to design natural routes for video-game characters to follow through an obstacle-ﬁlled room. How should I do it?” Solution: Presumably the desired route should look like a path that an intelligent being would choose. Since intelligent beings are either lazy or eﬃcient, this should be modeled as a shortest path problem. But what is the graph? One approach might be to lay a grid of points in the room. Create a vertex for each grid point that is a valid place for the character to stand; i.e. , that does not lie within an obstacle. There will be an edge between any pair of nearby vertices, weighted proportionally to the distance between them. Although direct geometric methods are known for shortest paths (see Section 15.4 (page 489)), it is easier to model this discretely as a graph. Stop and Think: Ordering the Sequence Problem: “A DNA sequencing project generates experimental data consisting of small fragments. For each given fragment f , we know certain other fragments are forced to lie to the left of f , and certain other fragments are forced to be to the right of f . How can we ﬁnd a consistent ordering of the fragments from left to right that satisﬁes all the constraints?” Solution: Create a directed graph, where each fragment is assigned a unique vertex. −→ Insert a directed edge (l, f ) from any fragment l that is forced to be to the left 223 224 6. WEIGHTED GRAPH ALGORITHMS −→ of f , and a directed edge (f, r) to any fragment r forced to be to the right of f . We seek an ordering of the vertices such that all the edges go from left to right. This is a topological sort of the resulting directed acyclic graph. The graph must be acyclic, because cycles make ﬁnding a consistent ordering impossible. Stop and Think: Bucketing Rectangles Problem: “In my graphics work I need to solve the following problem. Given an arbitrary set of rectangles in the plane, how can I distribute them into a minimum number of buckets such that no subset of rectangles in any given bucket intersects another? In other words, there can not be any overlapping area between two rectangles in the same bucket.” Solution: We formulate a graph where each vertex is a rectangle, and there is an edge if two rectangles intersect. Each bucket corresponds to an independent set of rectangles, so there is no overlap between any two. A vertex coloring of a graph is a partition of the vertices into independent sets, so minimizing the number of colors is exactly what you want. Stop and Think: Names in Collision Problem: “In porting code from UNIX to DOS, I have to shorten several hundred ﬁle names down to at most 8 characters each. I can’t just use the ﬁrst eight characters from each name, because “ﬁlename1” and “ﬁlename2” would be assigned the exact same name. How can I meaningfully shorten the names while ensuring that they do not collide?” Solution: Construct a bipartite graph with vertices corresponding to each original ﬁle name fi for 1 ≤ i ≤ n, as well as a collection of acceptable shortenings for each name fi1 , . . . , fik . Add an edge between each original and shortened name. We now seek a set of n edges that have no vertices in common, so each ﬁle name is mapped to a distinct acceptable substitute. Bipartite matching, discussed in Section 15.6 (page 498), is exactly this problem of ﬁnding an independent set of edges in a graph. 6.7 EXERCISES Stop and Think: Separate the Text Problem: “We need a way to separate the lines of text in the optical characterrecognition system that we are building. Although there is some white space between the lines, problems like noise and the tilt of the page makes it hard to ﬁnd. How can we do line segmentation? Solution: Consider the following graph formulation. Treat each pixel in the image as a vertex in the graph, with an edge between two neighboring pixels. The weight of this edge should be proportional to how dark the pixels are. A segmentation between two lines is a path in this graph from the left to right side of the page. We seek a relatively straight path that avoids as much blackness as possible. This suggests that the shortest path in the pixel graph will likely ﬁnd a good line segmentation. Take-Home Lesson: Designing novel graph algorithms is very hard, so don’t do it. Instead, try to design graphs that enable you to use classical algorithms to model your problem. Chapter Notes Network ﬂows are an advanced algorithmic technique, and recognizing whether a particular problem can be solved by network ﬂow requires experience. We point the reader to books by Cook and Cunningham [CC97] and Ahuja, Magnanti, and Orlin [AMO93] for more detailed treatments of the subject. The augmenting path method for network ﬂows is due to Ford and Fulkerson [FF62]. Edmonds and Karp [EK72] proved that always selecting a shortest geodesic augmenting path guarantees that O(n3 ) augmentations suﬃce for optimization. The phone code reconstruction system that was the subject of the war story is described in more technical detail in [RS96]. 6.7 Exercises Simulating Graph Algorithms 6-1. [3] For the graphs in Problem 5-1: (a) Draw the spanning forest after every iteration of the main loop in Kruskal’s algorithm. (b) Draw the spanning forest after every iteration of the main loop in Prim’s algorithm. 225 226 6. WEIGHTED GRAPH ALGORITHMS (c) Find the shortest path spanning tree rooted in A. (d) Compute the maximum ﬂow from A to H. Minimum Spanning Trees 6-2. [3] Is the path between two vertices in a minimum spanning tree necessarily a shortest path between the two vertices in the full graph? Give a proof or a counterexample. 6-3. [3] Assume that all edges in the graph have distinct edge weights (i.e. , no pair of edges have the same weight). Is the path between a pair of vertices in a minimum spanning tree necessarily a shortest path between the two vertices in the full graph? Give a proof or a counterexample. 6-4. [3] Can Prim’s and Kruskal’s algorithm yield diﬀerent minimum spanning trees? Explain why or why not. 6-5. [3] Does either Prim’s and Kruskal’s algorithm work if there are negative edge weights? Explain why or why not. 6-6. [5] Suppose we are given the minimum spanning tree T of a given graph G (with n vertices and m edges) and a new edge e = (u, v) of weight w that we will add to G. Give an eﬃcient algorithm to ﬁnd the minimum spanning tree of the graph G + e. Your algorithm should run in O(n) time to receive full credit. 6-7. [5] (a) Let T be a minimum spanning tree of a weighted graph G. Construct a new graph G by adding a weight of k to every edge of G. Do the edges of T form a minimum spanning tree of G ? Prove the statement or give a counterexample. (b) Let P = {s, . . . , t} describe a shortest weighted path between vertices s and t of a weighted graph G. Construct a new graph G by adding a weight of k to every edge of G. Does P describe a shortest path from s to t in G ? Prove the statement or give a counterexample. 6-8. [5] Devise and analyze an algorithm that takes a weighted graph G and ﬁnds the smallest change in the cost to a non-MST edge that would cause a change in the minimum spanning tree of G. Your algorithm must be correct and run in polynomial time. 6-9. [4] Consider the problem of ﬁnding a minimum weight connected subset T of edges from a weighted connected graph G. The weight of T is the sum of all the edge weights in T . (a) Why is this problem not just the minimum spanning tree problem? Hint: think negative weight edges. (b) Give an eﬃcient algorithm to compute the minimum weight connected subset T. 6-10. [4] Let G = (V, E) be an undirected graph. A set F ⊆ E of edges is called a feedback-edge set if every cycle of G has at least one edge in F . (a) Suppose that G is unweighted. Design an eﬃcient algorithm to ﬁnd a minimum-size feedback-edge set. 6.7 EXERCISES (b) Suppose that G is a weighted undirected graph with positive edge weights. Design an eﬃcient algorithm to ﬁnd a minimum-weight feedback-edge set. 6-11. [5] Modify Prim’s algorithm so that it runs in time O(n log k) on a graph that has only k diﬀerent edges costs. Union-Find 6-12. [5] Devise an eﬃcient data structure to handle the following operations on a weighted directed graph: (a) Merge two given components. (b) Locate which component contains a given vertex v. (c) Retrieve a minimum edge from a given component. 6-13. [5] Design a data structure that can perform a sequence of, m union and ﬁnd operations on a universal set of n elements, consisting of a sequence of all unions followed by a sequence of all ﬁnds, in time O(m + n). Shortest Paths 6-14. [3] The single-destination shortest path problem for a directed graph seeks the shortest path from every vertex to a speciﬁed vertex v. Give an eﬃcient algorithm to solve the single-destination shortest paths problem. 6-15. [3] Let G = (V, E) be an undirected weighted graph, and let T be the shortest-path spanning tree rooted at a vertex v. Suppose now that all the edge weights in G are increased by a constant number k. Is T still the shortest-path spanning tree from v? 6-16. [3] Answer all of the following: (a) Give an example of a weighted connected graph G = (V, E) and a vertex v, such that the minimum spanning tree of G is the same as the shortest-path spanning tree rooted at v. (b) Give an example of a weighted connected directed graph G = (V, E) and a vertex v, such that the minimum-cost spanning tree of G is very diﬀerent from the shortest-path spanning tree rooted at v. (c) Can the two trees be completely disjointed? 6-17. [3] Either prove the following or give a counterexample: (a) Is the path between a pair of vertices in a minimum spanning tree of an undirected graph necessarily the shortest (minimum weight) path? (b) Suppose that the minimum spanning tree of the graph is unique. Is the path between a pair of vertices in a minimum spanning tree of an undirected graph necessarily the shortest (minimum weight) path? 6-18. [5] In certain graph problems, vertices have can have weights instead of or in addition to the weights of edges. Let Cv be the cost of vertex v, and C(x,y) the cost of the edge (x, y). This problem is concerned with ﬁnding the cheapest path between vertices a and b in a graph G. The cost of a path is the sum of the costs of the edges and vertices encountered on the path. 227 228 6. WEIGHTED GRAPH ALGORITHMS (a) Suppose that each edge in the graph has a weight of zero (while non-edges have a cost of ∞). Assume that Cv = 1 for all vertices 1 ≤ v ≤ n (i.e. , all vertices have the same cost). Give an eﬃcient algorithm to ﬁnd the cheapest path from a to b and its time complexity. (b) Now suppose that the vertex costs are not constant (but are all positive) and the edge costs remain as above. Give an eﬃcient algorithm to ﬁnd the cheapest path from a to b and its time complexity. (c) Now suppose that both the edge and vertex costs are not constant (but are all positive). Give an eﬃcient algorithm to ﬁnd the cheapest path from a to b and its time complexity. 6-19. [5] Let G be a weighted directed graph with n vertices and m edges, where all edges have positive weight. A directed cycle is a directed path that starts and ends at the same vertex and contains at least one edge. Give an O(n3 ) algorithm to ﬁnd a directed cycle in G of minimum total weight. Partial credit will be given for an O(n2 m) algorithm. 6-20. [5] Can we modify Dijkstra’s algorithm to solve the single-source longest path problem by changing minimum to maximum? If so, then prove your algorithm correct. If not, then provide a counterexample. 6-21. [5] Let G = (V, E) be a weighted acyclic directed graph with possibly negative edge weights. Design a linear-time algorithm to solve the single-source shortest-path problem from a given source v. 6-22. [5] Let G = (V, E) be a directed weighted graph such that all the weights are positive. Let v and w be two vertices in G and k ≤ |V | be an integer. Design an algorithm to ﬁnd the shortest path from v to w that contains exactly k edges. Note that the path need not be simple. 6-23. [5] Arbitrage is the use of discrepancies in currency-exchange rates to make a proﬁt. For example, there may be a small window of time during which 1 U.S. dollar buys 0.75 British pounds, 1 British pound buys 2 Australian dollars, and 1 Australian dollar buys 0.70 U.S. dollars. At such a time, a smart trader can trade one U.S. dollar and end up with 0.75 × 2 × 0.7 = 1.05 U.S. dollars—a proﬁt of 5%. Suppose that there are n currencies c1 , ..., cn and an n × n table R of exchange rates, such that one unit of currency ci buys R[i, j] units of currency cj . Devise and analyze an algorithm to determine the maximum value of R[c1 , ci1 ] · R[ci1 , ci2 ] · · · R[cik−1 , cik ] · R[cik , c1 ] Hint: think all-pairs shortest path. Network Flow and Matching 6-24. [3] A matching in a graph is a set of disjoint edges—i.e., edges that do not share any vertices in common. Give a linear-time algorithm to ﬁnd a maximum matching in a tree. 6-25. [5] An edge cover of an undirected graph G = (V, E) is a set of edges such that each vertex in the graph is incident to at least one edge from the set. Give an eﬃcient algorithm, based on matching, to ﬁnd the minimum-size edge cover for G. 6.7 EXERCISES Programming Challenges These programming challenge problems with robot judging are available at http://www.programming-challenges.com or http://online-judge.uva.es. 6-1. “Freckles” – Programming Challenges 111001, UVA Judge 10034. 6-2. “Necklace” – Programming Challenges 111002, UVA Judge 10054. 6-3. “Railroads” – Programming Challenges 111004, UVA Judge 10039. 6-4. “Tourist Guide” – Programming Challenges 111006, UVA Judge 10199. 6-5. “The Grand Dinner” – Programming Challenges 111007, UVA Judge 10249. 229 7 Combinatorial Search and Heuristic Methods We can solve many problems to optimality using exhaustive search techniques, although the time complexity can be enormous. For certain applications, it may pay to spend extra time to be certain of the optimal solution. A good example occurs in testing a circuit or a program on all possible inputs. You can prove the correctness of the device by trying all possible inputs and verifying that they give the correct answer. Verifying correctness is a property to be proud of. However, claiming that it works correctly on all the inputs you tried is worth much less. Modern computers have clock rates of a few gigahertz, meaning billions of operations per second. Since doing something interesting takes a few hundred instructions, you can hope to search millions of items per second on contemporary machines. It is important to realize how big (or how small) one million is. One million permutations means all arrangements of roughly 10 or 11 objects, but not more. One million subsets means all combinations of roughly 20 items, but not more. Solving signiﬁcantly larger problems requires carefully pruning the search space to ensure we look at only the elements that really matter. In this section, we introduce backtracking as a technique for listing all possible solutions for a combinatorial algorithm problem. We illustrate the power of clever pruning techniques to speed up real search applications. For problems that are too large to contemplate using brute-force combinatorial search, we introduce heuristic methods such as simulated annealing. Such heuristic methods are important weapons in any practical algorist’s arsenal. S.S. Skiena, The Algorithm Design Manual, 2nd ed., DOI: 10.1007/978-1-84800-070-4 7, c Springer-Verlag London Limited 2008 7.1 7.1 BACKTRACKING Backtracking Backtracking is a systematic way to iterate through all the possible conﬁgurations of a search space. These conﬁgurations may represent all possible arrangements of objects (permutations) or all possible ways of building a collection of them (subsets). Other situations may demand enumerating all spanning trees of a graph, all paths between two vertices, or all possible ways to partition vertices into color classes. What these problems have in common is that we must generate each one possible conﬁguration exactly once. Avoiding both repetitions and missing conﬁgurations means that we must deﬁne a systematic generation order. We will model our combinatorial search solution as a vector a = (a1 , a2 , ..., an ), where each element ai is selected from a ﬁnite ordered set Si . Such a vector might represent an arrangement where ai contains the ith element of the permutation. Or, the vector might represent a given subset S, where ai is true if and only if the ith element of the universe is in S. The vector can even represent a sequence of moves in a game or a path in a graph, where ai contains the ith event in the sequence. At each step in the backtracking algorithm, we try to extend a given partial solution a = (a1 , a2 , ..., ak ) by adding another element at the end. After extending it, we must test whether what we now have is a solution: if so, we should print it or count it. If not, we must check whether the partial solution is still potentially extendible to some complete solution. Backtracking constructs a tree of partial solutions, where each vertex represents a partial solution. There is an edge from x to y if node y was created by advancing from x. This tree of partial solutions provides an alternative way to think about backtracking, for the process of constructing the solutions corresponds exactly to doing a depth-ﬁrst traversal of the backtrack tree. Viewing backtracking as a depthﬁrst search on an implicit graph yields a natural recursive implementation of the basic algorithm. Backtrack-DFS(A, k) if A = (a1 , a2 , ..., ak ) is a solution, report it. else k =k+1 compute Sk while Sk = ∅ do ak = an element in Sk Sk = Sk − ak Backtrack-DFS(A, k) Although a breadth-ﬁrst search could also be used to enumerate solutions, a depth-ﬁrst search is greatly preferred because it uses much less space. The current state of a search is completely represented by the path from the root to the current search depth-ﬁrst node. This requires space proportional to the height of the tree. In breadth-ﬁrst search, the queue stores all the nodes at the current level, which 231 232 7. COMBINATORIAL SEARCH AND HEURISTIC METHODS is proportional to the width of the search tree. For most interesting problems, the width of the tree grows exponentially in its height. Implementation The honest working backtrack code is given below: bool finished = FALSE; backtrack(int a[], int k, data { int c[MAXCANDIDATES]; int ncandidates; int i; /* found all solutions yet? */ input) /* candidates for next position */ /* next position candidate count */ /* counter */ if (is_a_solution(a,k,input)) process_solution(a,k,input); else { k = k+1; construct_candidates(a,k,input,c,&ncandidates); for (i=0; i y ]) { c[*ncandidates] = p->y; *ncandidates = *ncandidates + 1; } p = p->next; } } } We report a successful path whenever ak = t. is_a_solution(int a[], int k, int t) { return (a[k] == t); } process_solution(int a[], int k) { solution_count ++; } /* count all s to t paths */ 237 238 7. COMBINATORIAL SEARCH AND HEURISTIC METHODS The solution vector A must have room for all n vertices, although most paths are likely shorter than this. Figure 7.1 shows the search tree giving all paths from a particular vertex in an example graph. 7.2 Search Pruning Backtracking ensures correctness by enumerating all possibilities. Enumerating all n! permutations of n vertices of the graph and selecting the best one yields the correct algorithm to ﬁnd the optimal traveling salesman tour. For each permutation, we could see whether all edges implied by the tour really exists in the graph G, and if so, add the weights of these edges together. However, it would be wasteful to construct all the permutations ﬁrst and then analyze them later. Suppose our search started from vertex v1 , and it happened that edge (v1 , v2 ) was not in G. The next (n−2)! permutations enumerated starting with (v1 , v2 ) would be a complete waste of eﬀort. Much better would be to prune the search after v1 , v2 and continue next with v1 , v3 . By restricting the set of next elements to reﬂect only moves that are legal from the current partial conﬁguration, we signiﬁcantly reduce the search complexity. Pruning is the technique of cutting oﬀ the search the instant we have established that a partial solution cannot be extended into a full solution. For traveling salesman, we seek the cheapest tour that visits all vertices. Suppose that in the course of our search we ﬁnd a tour t whose cost is Ct . Later, we may have a partial solution a whose edge sum CA > Ct . Is there any reason to continue exploring this node? No, because any tour with this preﬁx a1 , . . . , ak will have cost greater than tour t, and hence is doomed to be nonoptimal. Cutting away such failed partial tours as soon as possible can have an enormous impact on running time. Exploiting symmetry is another avenue for reducing combinatorial searches. Pruning away partial solutions identical to those previously considered requires recognizing underlying symmetries in the search space. For example, consider the state of our TSP search after we have tried all partial positions beginning with v1 . Does it pay to continue the search with partial solutions beginning with v2 ? No. Any tour starting and ending at v2 can be viewed as a rotation of one starting and ending at v1 , for these tours are cycles. There are thus only (n − 1)! distinct tours on n vertices, not n!. By restricting the ﬁrst element of the tour to v1 , we save a factor of n in time without missing any interesting solutions. Detecting such symmetries can be subtle, but once identiﬁed they can usually be easily exploited. Take-Home Lesson: Combinatorial searches, when augmented with tree pruning techniques, can be used to ﬁnd the optimal solution of small optimization problems. How small depends upon the speciﬁc problem, but typical size limits are somewhere between 15 ≤ n ≤ 50 items. 7.3 1 3 5 6 7 7 3 8 4 1 1 8 5 2 4 6 2 6 9 8 7 5 1 4 2 3 7 1 4 9 2 3 6 8 5 3 2 5 8 6 4 9 7 1 8 7 6 2 4 5 1 3 9 9 3 1 6 7 8 2 5 4 4 5 2 1 3 9 8 6 7 5 4 9 3 8 2 7 1 6 SUDOKU 1 8 7 5 9 6 3 4 2 2 6 3 4 1 7 5 9 8 Figure 7.2: Challenging Sudoku puzzle (l) with solution (r) 7.3 Sudoku A Sudoku craze has swept the world. Many newspapers now publish daily Sudoku puzzles, and millions of books about Sudoku have been sold. British Airways sent a formal memo forbidding its cabin crews from doing Sudoku during takeoﬀs and landings. Indeed, I have noticed plenty of Sudoku going on in the back of my algorithms classes during lecture. What is Sudoku? In its most common form, it consists of a 9 × 9 grid ﬁlled with blanks and the digits 1 to 9. The puzzle is completed when every row, column, and sector (3 × 3 subproblems corresponding to the nine sectors of a tic-tac-toe puzzle) contain the digits 1 through 9 with no deletions or repetition. Figure 7.2 presents a challenging Sudoku puzzle and its solution. Backtracking lends itself nicely to the problem of solving Sudoku puzzles. We will use the puzzle here to better illustrate the algorithmic technique. Our state space will be the sequence of open squares, each of which must ultimately be ﬁlled in with a number. The candidates for open squares (i, j) are exactly the integers from 1 to 9 that have not yet appeared in row i, column j, or the 3 × 3 sector containing (i, j). We backtrack as soon as we are out of candidates for a square. The solution vector a supported by backtrack only accepts a single integer per position. This is enough to store contents of a board square (1-9) but not the coordinates of the board square. Thus, we keep a separate array of move positions as part of our board data type provided below. The basic data structures we need to support our solution are: #define DIMENSION 9 /* 9*9 board */ #define NCELLS DIMENSION*DIMENSION /* 81 cells in a 9*9 problem */ typedef struct { int x, y; } point; /* x and y coordinates of point */ 239 240 7. COMBINATORIAL SEARCH AND HEURISTIC METHODS typedef struct { int m[DIMENSION+1][DIMENSION+1]; /* matrix of board contents */ int freecount; /* how many open squares remain? */ point move[NCELLS+1]; /* how did we fill the squares? */ } boardtype; Constructing the candidates for the next solution position involves ﬁrst picking the open square we want to ﬁll next (next square), and then identifying which numbers are candidates to ﬁll that square (possible values). These routines are basically bookkeeping, although the subtle details of how they work can have an enormous impact on performance. construct_candidates(int a[], int k, boardtype *board, int c[], int *ncandidates) { int x,y; /* position of next move */ int i; /* counter */ bool possible[DIMENSION+1]; /* what is possible for the square */ next_square(&x,&y,board); /* which square should we fill next? */ board->move[k].x = x; board->move[k].y = y; /* store our choice of next position */ *ncandidates = 0; if ((x<0) && (y<0)) return; /* error condition, no moves possible */ possible_values(x,y,board,possible); for (i=0; i<=DIMENSION; i++) if (possible[i] == TRUE) { c[*ncandidates] = i; *ncandidates = *ncandidates + 1; } } We must update our board data structure to reﬂect the eﬀect of ﬁlling a candidate value into a square, as well as remove these changes should we backtrack away from this position. These updates are handled by make move and unmake move, both of which are called directly from backtrack: 7.3 SUDOKU make_move(int a[], int k, boardtype *board) { fill_square(board->move[k].x,board->move[k].y,a[k],board); } unmake_move(int a[], int k, boardtype *board) { free_square(board->move[k].x,board->move[k].y,board); } One important job for these board update routines is maintaining how many free squares remain on the board. A solution is found when there are no more free squares remaining to be ﬁlled: is_a_solution(int a[], int k, boardtype *board) { if (board->freecount == 0) return (TRUE); else return(FALSE); } We print the conﬁguration and turn oﬀ the backtrack search by setting oﬀ the global finished ﬂag on ﬁnding a solution. This can be done without consequence because “oﬃcial” Sudoku puzzles are only allowed to have one solution. There can be an enormous number of solutions to nonoﬃcial Sudoku puzzles. Indeed, the empty puzzle (where no number is initially speciﬁed) can be ﬁlled in exactly 6,670,903,752,021,072,936,960 ways. We can ensure we don’t see all of them by turning oﬀ the search: process_solution(int a[], int k, boardtype *board) { print_board(board); finished = TRUE; } This completes the program modulo details of identifying the next open square to ﬁll (next square) and identifying the candidates to ﬁll that square (possible values). Two reasonable ways to select the next square are: • Arbitrary Square Selection – Pick the ﬁrst open square we encounter, possibly picking the ﬁrst, the last, or a random open square. All are equivalent in that there seems to be no reason to believe that one heuristic will perform any better than the other. 241 242 7. COMBINATORIAL SEARCH AND HEURISTIC METHODS • Most Constrained Square Selection – Here, we check each of the open squares (i, j) to see how many number candidates remain for each—i.e. , have not already been used in either row i, column j, or the sector containing (i, j). We pick the square with the fewest number of candidates. Although both possibilities work correctly, the second option is much, much better. Often there will be open squares with only one remaining candidate. For these, the choice is forced. We might as well make it now, especially since pinning this value down will help trim the possibilities for other open squares. Of course, we will spend more time selecting each candidate square, but if the puzzle is easy enough we may never have to backtrack at all. If the most constrained square has two possibilities, we have a 1/2 probability of guessing right the ﬁrst time, as opposed to a (1/9)th probability for a completely unconstrained square. Reducing our average number of choices from (say) 3 per square to 2 per square is an enormous win, since it multiplies for each position. If we have (say) 20 positions to ﬁll, we must enumerate only 220 = 1, 048, 576 solutions. A branching factor of 3 at each of the 20 positions will result in over 3, 000 times as much work! Our ﬁnal decision concerns the possible values we allow for each square. We have two possibilities: • Local Count – Our backtrack search works correctly if the routine generating candidates for board position (i, j) (possible values) does the obvious thing and allows all numbers 1 to 9 that have not appeared in the given row, column, or sector. • Look ahead – But what if our current partial solution has some other open square where there are no candidates remaining under the local count criteria? There is no possible way to complete this partial solution into a full Sudoku grid. Thus there really are zero possible moves to consider for (i, j) because of what is happening elsewhere on the board! We will discover this obstruction eventually, when we pick this square for expansion, discover it has no moves, and then have to backtrack. But why wait, since all our eﬀorts until then will be wasted? We are much better oﬀ backtracking immediately and moving on.1 Successful pruning requires looking ahead to see when a solution is doomed to go nowhere, and backing oﬀ as soon as possible. Table 7.1 presents the number of calls to is a solution for all four backtracking variants on three Sudoku instances of varying complexity: 1 This look-ahead condition might have naturally followed from the most-constrained square selection, had it been permitted to select squares with no moves. However, my implementation credited squares that already contained numbers as having no moves, thus limiting the next square choices to squares with at least one move. 7.3 Pruning Condition next square possible values arbitrary local count arbitrary look ahead most constrained local count most constrained look ahead SUDOKU Puzzle Complexity Easy Medium Hard 1,904,832 863,305 never ﬁnished 127 142 12,507,212 48 84 1,243,838 48 65 10,374 Table 7.1: Sudoku run times (in number of steps) under diﬀerent pruning strategies • The Easy board was intended to be easy for a human player. Indeed, my program solved it without any backtracking steps when the most constrained square was selected as the next position. • The Medium board stumped all the contestants at the ﬁnals of the World Sudoku Championship in March 2006. The decent search variants still required only a few backtrack steps to dispatch this problem. • The Hard problem is the board displayed in Figure 7.2, which contains only 17 ﬁxed numbers. This is the fewest speciﬁed known number of positions in any problem instance that has only one complete solution. What is considered to be a “hard” problem instance depends upon the given heuristic. Certainly you know people who ﬁnd math/theory harder than programming and others who think diﬀerently. Heuristic A may well think instance I1 is easier than I2 , while heuristic B ranks them in the other order. What can we learn from these experiments? Looking ahead to eliminate dead positions as soon as possible is the best way to prune a search. Without this operation, we never ﬁnished the hardest puzzle and took thousands of times longer than we should have on the easier ones. Smart square selection had a similar impact, even though it nominally just rearranges the order in which we do the work. However, doing more constrained positions ﬁrst is tantamount to reducing the outdegree of each node in the tree, and each additional position we ﬁx adds constraints that help lower the degree for future selections. It took the better part of an hour (48:44) to solve the puzzle in Figure 7.2 when I selected an arbitrary square for my next move. Sure, my program was faster in most instances, but Sudoku puzzles are designed to be solved by people using pencils in much less time than this. Making the next move in the most constricted square reduced search time by a factor of over 1,200. Each puzzle we tried can now be solved in seconds—the time it takes to reach for the pencil if you want to do it by hand. This is the power of a pruning search. Even simple pruning strategies can suﬃce to reduce running time from impossible to instantaneous. 243 244 7. COMBINATORIAL SEARCH AND HEURISTIC METHODS Figure 7.3: Conﬁgurations covering 63 but not 64 squares 7.4 War Story: Covering Chessboards Every researcher dreams of solving a classical problem—one that has remained open and unsolved for over a century. There is something romantic about communicating across the generations, being part of the evolution of science, and helping to climb another rung up the ladder of human progress. There is also a pleasant sense of smugness that comes from ﬁguring out how to do something that nobody could do before you. There are several possible reasons why a problem might stay open for such a long period of time. Perhaps it is so diﬃcult and profound that it requires a uniquely powerful intellect to solve. A second reason is technological—the ideas or techniques required to solve the problem may not have existed when it was ﬁrst posed. A ﬁnal possibility is that no one may have cared enough about the problem in the interim to seriously bother with it. Once, I helped solve a problem that had been open for over a hundred years. Decide for yourself which reason best explains why. Chess is a game that has fascinated mankind for thousands of years. In addition, it has inspired many combinatorial problems of independent interest. The combinatorial explosion was ﬁrst recognized with the legend that the inventor of chess demanded as payment one grain of rice for the ﬁrst square of the board, and twice as much for the (i + 1)st square than the ith square. The king was aston64 ished to learn he had to cough up i=1 2i = 265 − 1 = 36,893,488,147,419,103,231 grains of rice. In beheading the inventor, the wise king ﬁrst established pruning as a technique for dealing with the combinatorial explosion. In 1849, Kling posed the question of whether all 64 squares on the board can be simultaneously threatened by an arrangement of the eight main pieces on the chess board—the king, queen, two knights, two rooks, and two oppositely colored bishops. Pieces do not threaten the square they sit on. Conﬁgurations that simultaneously threaten 63 squares, such as those in Figure 7.3, have been known for a long time, but whether this was the best possible remained an open problem. This problem 7.4 WAR STORY: COVERING CHESSBOARDS Figure 7.4: The ten unique positions for the queen, with respect to symmetry seemed ripe for solution by exhaustive combinatorial searching, although whether it was solvable depended upon the size of the search space. Consider the eight main pieces in chess (king, queen, two rooks, two bishops, and two knights). How many ways can they be positioned on a chessboard? The trivial bound is 64!/(64 − 8)! = 178,462,987,637,760 ≈ 1015 positions. Anything much larger than about 109 positions would be unreasonable to search on a modest computer in a modest amount of time. Getting the job done would require signiﬁcant pruning. Our ﬁrst idea was to remove symmetries. Accounting for orthogonal and diagonal symmetries left only ten distinct positions for the queen, shown in Figure 7.4. Once the queen is placed, there are 64 · 63/2 = 2,016 distinct ways to position a pair of rooks or knights, 64 places to locate the king, and 32 spots for each of the white and black bishops. Such an exhaustive search would test 2,663,550,812,160 ≈ 1013 distinct positions—still much too large to try. We could use backtracking to construct all of the positions, but we had to ﬁnd a way to prune the search space signiﬁcantly. Pruning the search meant that we needed a quick way to prove that there was no way to complete a partially ﬁlled-in position to cover all 64 squares. Suppose we had already placed seven pieces on the board, and together they covered all but 10 squares of the board. Say the remaining piece was the king. Can there be a position to place the king so that all squares are threatened? The answer must be no, because the king can threaten at most eight squares according to the rules of chess. There can be no reason to test any king position. We might win big pruning such conﬁgurations. This pruning strategy required carefully ordering the evaluation of the pieces. Each piece can threaten a certain maximum number of squares: the queen 27, the king/knight 8, the rook 14, and the bishop 13. We would want to insert the pieces in decreasing order of mobility. We can prune when the number of unthreatened squares exceeds the sum of the maximum coverage of the unplaced pieces. This sum is minimized by using the decreasing order of mobility. 245 246 7. COMBINATORIAL SEARCH AND HEURISTIC METHODS Figure 7.5: Weakly covering 64 squares When we implemented a backtrack search using this pruning strategy, we found that it eliminated over 95% of the search space. After optimizing our move generation, our program could search over 1,000 positions per second. But this was still too slow, for 1012 /103 = 109 seconds meant 1,000 days! Although we might further tweak the program to speed it up by an order of magnitude or so, what we really needed was to ﬁnd a way to prune more nodes. Eﬀective pruning meant eliminating large numbers of positions at a single stroke. Our previous attempt was too weak. What if instead of placing up to eight pieces on the board simultaneously, we placed more than eight pieces. Obviously, the more pieces we placed simultaneously, the more likely they would threaten all 64 squares. But if they didn’t cover, all subsets of eight distinct pieces from the set couldn’t possibly threaten all squares. The potential existed to eliminate a vast number of positions by pruning a single node. So in our ﬁnal version, the nodes of our search tree corresponded to chessboards that could have any number of pieces, and more than one piece on a square. For a given board, we distinguished strong and weak attacks on a square. A strong attack corresponds to the usual notion of a threat in chess. A square is weakly attacked if the square is strongly attacked by some subset of the board—that is, a weak attack ignores any possible blocking eﬀects of intervening pieces. All 64 squares can be weakly attacked with eight pieces, as shown in Figure 7.5. Our algorithm consisted of two passes. The ﬁrst pass listed boards where every square was weakly attacked. The second pass ﬁltered the list by considering blocking pieces. A weak attack is much faster to compute (no blocking to worry about), and any strong attack set is always a subset of a weak attack set. The position could be pruned whenever there was a non-weakly threatened square. This program was eﬃcient enough to complete the search on a slow 1988-era IBM PC-RT in under one day. It did not ﬁnd a single position covering all 64 squares with the bishops on opposite colored squares. However, our program showed that it is possible to cover the board with seven pieces provided a queen and a knight can occupy the same square, as shown in Figure 7.6. 7.5 HEURISTIC SEARCH METHODS 247 Figure 7.6: Seven pieces suﬃce when superimposing queen and knight (shown as a white queen) Take-Home Lesson: Clever pruning can make short work of surprisingly hard combinatorial search problems. Proper pruning will have a greater impact on search time than any other factor. 7.5 Heuristic Search Methods Heuristic methods provide an alternate way to approach diﬃcult combinatorial optimization problems. Backtracking gave us a method to ﬁnd the best of all possible solutions, as scored by a given objective function. However, any algorithm searching all conﬁgurations is doomed to be impossible on large instances. In this section, we discuss such heuristic search methods. We devote the bulk of our attention to simulated annealing, which I ﬁnd to be the most reliable method to apply in practice. Heuristic search algorithms have an air of voodoo about them, but how they work and why one method might work better than another follows logically enough if you think them through. In particular, we will look at three diﬀerent heuristic search methods: random sampling, gradient-descent search, and simulated annealing. The traveling salesman problem will be our ongoing example for comparing heuristics. All three methods have two common components: • Solution space representation – This is a complete yet concise description of the set of possible solutions for the problem. For traveling salesman, the solution space consists of (n − 1)! elements—namely all possible circular permutations of the vertices. We need a data structure to represent each element of the solution space. For TSP, the candidate solutions can naturally be represented using an array S of n − 1 vertices, where Si deﬁnes the (i + 1)st vertex on the tour starting from v1 . • Cost function – Search methods need a cost or evaluation function to access the quality of each element of the solution space. Our search heuristic 248 7. COMBINATORIAL SEARCH AND HEURISTIC METHODS identiﬁes the element with the best possible score—either highest or lowest depending upon the nature of the problem. For TSP, the cost function for evaluating a given candidate solution S should just sum up the costs involved, namely the weight of all edges (Si , Si+1 ), where Sn+1 denotes v1 . 7.5.1 Random Sampling The simplest method to search in a solution space uses random sampling. It is also called the Monte Carlo method. We repeatedly construct random solutions and evaluate them, stopping as soon as we get a good enough solution, or (more likely) when we are tired of waiting. We report the best solution found over the course of our sampling. True random sampling requires that we are able to select elements from the solution space uniformly at random. This means that each of the elements of the solution space must have an equal probability of being the next candidate selected. Such sampling can be a subtle problem. Algorithms for generating random permutations, subsets, partitions, and graphs are discussed in Sections 14.4-14.7. random_sampling(tsp_instance *t, int nsamples, tsp_solution *bestsol) { tsp_solution s; /* current tsp solution */ double best_cost; /* best cost so far */ double cost_now; /* current cost */ int i; /* counter */ initialize_solution(t->n,&s); best_cost = solution_cost(&s,t); copy_solution(&s,bestsol); for (i=1; i<=nsamples; i++) { random_solution(&s); cost_now = solution_cost(&s,t); if (cost_now < best_cost) { best_cost = cost_now; copy_solution(&s,bestsol); } } } When might random sampling do well? • When there are a high proportion of acceptable solutions – Finding a piece of hay in a haystack is easy, since almost anything you grab is a straw. When solutions are plentiful, a random search should ﬁnd one quickly. 7.5 HEURISTIC SEARCH METHODS Performance of Random Sampling for TSP 200000 Random Sampling 180000 160000 Length of Tour 140000 120000 100000 80000 60000 40000 20000 0 0 200000 400000 600000 800000 1e+06 Number of Iterations 1.2e+06 1.4e+06 1.6e+06 Figure 7.7: Search time/quality tradeoﬀs for TSP using random sampling. Finding prime numbers is domain where a random search proves successful. Generating large random prime numbers for keys is an important aspect of cryptographic systems such as RSA. Roughly one out of every ln n integers are prime, so only a modest number of samples need to be taken to discover primes that are several hundred digits long. • When there is no coherence in the solution space – Random sampling is the right thing to do when there is no sense of when we are getting closer to a solution. Suppose you wanted to ﬁnd one of your friends who has a social security number that ends in 00. There is not much you can hope to do but tap an arbitrary fellow on their shoulder and ask. No cleverer method will be better than random sampling. Consider again the problem of hunting for a large prime number. Primes are scattered quite arbitrarily among the integers. Random sampling is as good as anything else. How does random sampling do on TSP? Pretty lousy. The best solution I found after testing 1.5 million random permutations of a classic TSP instance (the capital cities of the 48 continental United States) was 101,712.8. This is more than three times the cost of the optimal tour! The solution space consists almost entirely 249 250 7. COMBINATORIAL SEARCH AND HEURISTIC METHODS of mediocre to bad solutions, so quality grows very slowly with the amount of sampling / running time we invest. Figure 7.7 presents the arbitrary up-and-down movements of random sampling and generally poor quality solutions encountered on the journey, so you can get a sense of how the score varied over each iteration. Most problems we encounter, like TSP, have relatively few good solutions but a highly coherent solution space. More powerful heuristic search algorithms are required to deal eﬀectively with such problems. Stop and Think: Picking the Pair Problem: We need an eﬃcient and unbiased way to generate random pairs of vertices to perform random vertex swaps. Propose an eﬃcient algorithm to generate elements from the n2 unordered pairs on {1, . . . , n} uniformly at random. Solution: Uniformly generating random structures is a surprisingly subtle problem. Consider the following procedure to generate random unordered pairs: i = random int(1,n-1); j = random int(i+1,n); It is clear that this indeed generates unordered pairs, since i < j. Further, n it is clear that all 2 unordered pairs can indeed be generated, assuming that random int generates integers uniformly between its two arguments. But are they uniform? The answer is no. What is the probability that pair (1, 2) is generated? There is a 1/(n−1) chance of getting the 1, and then a 1/(n−1) chance 2 of getting the 2, which yields p(1, 2) = 1/(n − 1) . But what is the probability of getting (n − 1, n)? Again, there is a 1/n chance of getting the ﬁrst number, but now there is only one possible choice for the second candidate! This pair will occur n times more often than the ﬁrst! The problem is that fewer pairs start with big numbers than little numbers. We could solve this problem by calculating exactly how unordered pairs start with i (exactly (n − i)) and appropriately bias the probability. The second value could then be selected uniformly at random from i + 1 to n. But instead of working through the math, let’s exploit the fact that randomly generating the n2 ordered pairs uniformly is easy. Just pick two integers independently of each other. Ignoring the ordering (i.e. , permuting the ordered pair to unordered pair (x, y) so that x < y) gives us a 2/n2 probability of generating each unordered pair of distinct elements. If we happen to generate a pair (x, x), we discard it and try again. We will get unordered pairs uniformly at random in constant expected time using the following algorithm: 7.5 HEURISTIC SEARCH METHODS 1 1 6 2 3 5 4 6 2 3 5 4 Figure 7.8: Improving a TSP tour by swapping vertices 2 and 6 do { i = random int(1,n); j = random int(1,n); if (i > j) swap(&i,&j); } while (i==j); 7.5.2 Local Search Now suppose you want to hire an algorithms expert as a consultant to solve your problem. You could dial a phone number at random, ask if they are an algorithms expert, and hang up the phone if they say no. After many repetitions you will probably ﬁnd one, but it would probably be more eﬃcient to ask the fellow on the phone for someone more likely to know an algorithms expert, and call them up instead. A local search employs local neighborhood around every element in the solution space. Think of each element x in the solution space as a vertex, with a directed edge (x, y) to every candidate solution y that is a neighbor of x. Our search proceeds from x to the most promising candidate in x’s neighborhood. We certainly do not want to explicitly construct this neighborhood graph for any sizable solution space. Think about TSP, which will have (n − 1)! vertices in this graph. We are conducting a heuristic search precisely because we cannot hope to do this many operations in a reasonable amount of time. Instead, we want a general transition mechanism that takes us to the next solution by slightly modifying the current one. Typical transition mechanisms include swapping a random pair of items or changing (inserting or deleting) a single item in the solution. The most obvious transition mechanism for TSP would be to swap the current tour positions of a random pair of vertices Si and Sj , as shown in Figure 7.8. 251 252 7. COMBINATORIAL SEARCH AND HEURISTIC METHODS This changes up to eight edges on the tour, deleting the edges currently adjacent to both Si and Sj , and adding their replacements. Ideally, the eﬀect that these incremental changes have on measuring the quality of the solution can be computed incrementally, so cost function evaluation takes time proportional to the size of the change (typically constant) instead of linear to the size of the solution. A local search heuristic starts from an arbitrary element of the solution space, and then scans the neighborhood looking for a favorable transition to take. For TSP, this would be transition, which lowers the cost of the tour. In a hill-climbing procedure, we try to ﬁnd the top of a mountain (or alternately, the lowest point in a ditch) by starting at some arbitrary point and taking any step that leads in the direction we want to travel. We repeat until we have reached a point where all our neighbors lead us in the wrong direction. We are now King of the Hill (or Dean of the Ditch). We are probably not King of the Mountain, however. Suppose you wake up in a ski lodge, eager to reach the top of the neighboring peak. Your ﬁrst transition to gain altitude might be to go upstairs to the top of the building. And then you are trapped. To reach the top of the mountain, you must go downstairs and walk outside, but this violates the requirement that each step has to increase your score. Hill-climbing and closely related heuristics such as greedy search or gradient descent search are great at ﬁnding local optima quickly, but often fail to ﬁnd the globally best solution. hill_climbing(tsp_instance *t, tsp_solution *s) { double cost; /* best cost so far */ double delta; /* swap cost */ int i,j; /* counters */ bool stuck; /* did I get a better solution? */ double transition(); initialize_solution(t->n,s); random_solution(s); cost = solution_cost(s,t); do { stuck = TRUE; for (i=1; i n; i++) for (j=i+1; j<=t->n; j++) { delta = transition(s,t,i,j); if (delta < 0) { stuck = FALSE; cost = cost + delta; } 7.5 HEURISTIC SEARCH METHODS else transition(s,t,j,i); } } while (!stuck); } When does local search do well? • When there is great coherence in the solution space – Hill climbing is at its best when the solution space is convex. In other words, it consists of exactly one hill. No matter where you start on the hill, there is always a direction to walk up until you are at the absolute global maximum. Many natural problems do have this property. We can think of a binary search as starting in the middle of a search space, where exactly one of the two possible directions we can walk will get us closer to the target key. The simplex algorithm for linear programming (see Section 13.6 (page 411)) is nothing more than hill-climbing over the right solution space, yet guarantees us the optimal solution to any linear programming problem. • Whenever the cost of incremental evaluation is much cheaper than global evaluation – It costs Θ(n) to evaluate the cost of an arbitrary n-vertex candidate TSP solution, because we must total the cost of each edge in the circular permutation describing the tour. Once that is found, however, the cost of the tour after swapping a given pair of vertices can be determined in constant time. If we are given a very large value of n and a very small budget of how much time we can spend searching, we are better oﬀ using it to do several incremental evaluations than a few random samples, even if we are looking for a needle in a haystack. The primary drawback of a local search is that soon there isn’t anything left for us to do as we ﬁnd the local optimum. Sure, if we have more time we could start from diﬀerent random points, but in landscapes of many low hills we are unlikely to stumble on the optimum. How does local search do on TSP? Much better than random sampling for a similar amount of time. With over a total of 1.5 million tour evaluations in our 48-city TSP instance, our best local search tour had a length of 40,121.2—only 19.6% more than the optimal tour of 33,523.7. This is good, but not great. You would not be happy to learn you are paying 19.6% more taxes than you should. Figure 7.9 illustrates the trajectory of a local search: repeated streaks from random tours down to decent solutions of fairly similar quality. We need more powerful methods to get closer to the optimal solution. 253 7. COMBINATORIAL SEARCH AND HEURISTIC METHODS Performance of Hill Climbing for TSP 200000 Hill Climbing 180000 160000 140000 Length of Tour 254 120000 100000 80000 60000 40000 20000 0 0 200000 400000 600000 800000 1e+06 Number of Iterations 1.2e+06 1.4e+06 1.6e+06 Figure 7.9: Search time/quality tradeoﬀs for TSP using hill climbing. 7.5.3 Simulated Annealing Simulated annealing is a heuristic search procedure that allows occasional transitions leading to more expensive (and hence inferior) solutions. This may not sound like progress, but it helps keep our search from getting stuck in local optima. That poor fellow trapped on the top ﬂoor of a ski lodge would do better to break the glass and jump out the window if he really wanted to reach the top of the mountain. The inspiration for simulated annealing comes from the physical process of cooling molten materials down to the solid state. In thermodynamic theory, the energy state of a system is described by the energy state of each particle constituting it. A particle’s energy state jumps about randomly, with such transitions governed by the temperature of the system. In particular, the transition probability P (ei , ej , T ) from energy ei to ej at temperature T is given by P (ei , ej , T ) = e(ei −ej )/(kB T ) where kB is a constant—called Boltzmann’s constant. What does this formula mean? Consider the value of the exponent under different conditions. The probability of moving from a high-energy state to a lowerenergy state is very high. But, there is still a nonzero probability of accepting a 7.5 HEURISTIC SEARCH METHODS transition into a high-energy state, with such small jumps much more likely than big ones. The higher the temperature, the more likely energy jumps will occur. Simulated-Annealing() Create initial solution S Initialize temperature t repeat for i = 1 to iteration-length do Generate a random transition from S to Si If (C(S) ≥ C(Si )) then S = Si else if (e(C(S)−C(Si ))/(k·t) > random[0, 1)) then S = Si Reduce temperature t until (no change in C(S)) Return S What relevance does this have for combinatorial optimization? A physical system, as it cools, seeks to reach a minimum-energy state. Minimizing the total energy is a combinatorial optimization problem for any set of discrete particles. Through random transitions generated according to the given probability distribution, we can mimic the physics to solve arbitrary combinatorial optimization problems. Take-Home Lesson: Forget about this molten metal business. Simulated annealing is eﬀective because it spends much more of its time working on good elements of the solution space than on bad ones, and because it avoids getting trapped repeatedly in the same local optima. As with a local search, the problem representation includes both a representation of the solution space and an easily computable cost function C(s) measuring the quality of a given solution. The new component is the cooling schedule, whose parameters govern how likely we are to accept a bad transition as a function of time. At the beginning of the search, we are eager to use randomness to explore the search space widely, so the probability of accepting a negative transition should be high. As the search progresses, we seek to limit transitions to local improvements and optimizations. The cooling schedule can be regulated by the following parameters: • Initial system temperature – Typically t1 = 1. • Temperature decrement function – Typically tk = α · tk−1 , where 0.8 ≤ α ≤ 0.99. This implies an exponential decay in the temperature, as opposed to a linear decay. • Number of iterations between temperature change – Typically, 100 to 1,000 iterations might be permitted before lowering the temperature. 255 7. COMBINATORIAL SEARCH AND HEURISTIC METHODS Performance of Simulated Annealing for TSP 200000 Simulated Annealing 180000 160000 140000 Length of Tour 256 120000 100000 80000 60000 40000 20000 0 0 200000 400000 600000 800000 1e+06 Number of Iterations 1.2e+06 1.4e+06 1.6e+06 Figure 7.10: Search time/quality tradeoﬀs for TSP using simulated annealing • Acceptance criteria – A typical criterion is to accept any transition from si to si+1 when C(si+1 ) < C(si ), and also accept a negative transition whenever − e (C(si )−C(si+1 )) k·ti ≥ r, where r is a random number 0 ≤ r < 1. The constant k normalizes this cost function so that almost all transitions are accepted at the starting temperature. • Stop criteria – Typically, when the value of the current solution has not changed or improved within the last iteration or so, the search is terminated and the current solution reported. Creating the proper cooling schedule is somewhat of a trial-and-error process of mucking with constants and seeing what happens. It probably pays to start from an existing implementation of simulated annealing, so check out my full implementation (at http://www.algorist.com ) as well as others provided in Section 13.5 (page 407). Compare the search/time execution proﬁles of our three heuristics. The cloud of points corresponding to random sampling is signiﬁcantly worse than the solutions 7.5 HEURISTIC SEARCH METHODS encountered by the other heuristics. The scores of the short streaks corresponding to runs of hill-climbing solutions are clearly much better. But best of all are the proﬁles of the three simulated annealing runs in Figure 7.10. All three runs lead to much better solutions than the best hill-climbing result. Further, it takes relatively few iterations to score most of the improvement, as shown by the three rapid plunges toward optimum we see with simulated annealing. Using the same 1,500,000 iterations as the other methods, simulated annealing gave us a solution of cost 36,617.4—only 9.2% over the optimum. Even better solutions are available to those willing to wait a few minutes. Letting it run for 5,000,000 iterations got the score down to 34,254.9, or 2.2% over the optimum. There were no further improvements after I cranked it up to 10,000,000 iterations. In expert hands, the best problem-speciﬁc heuristics for TSP can slightly outperform simulated annealing. But the simulated annealing solution works admirably. It is my heuristic method of choice. Implementation The implementation follows the pseudocode quite closely: anneal(tsp_instance *t, tsp_solution *s) { int i1, i2; /* pair of items to swap */ int i,j; /* counters */ double temperature; /* the current system temp */ double current_value; /* value of current state */ double start_value; /* value at start of loop */ double delta; /* value after swap */ double merit, flip; /* hold swap accept conditions*/ double exponent; /* exponent for energy funct*/ double random_float(); double solution_cost(), transition(); temperature = INITIAL_TEMPERATURE; initialize_solution(t->n,s); current_value = solution_cost(s,t); for (i=1; i<=COOLING_STEPS; i++) { temperature *= COOLING_FRACTION; start_value = current_value; for (j=1; j<=STEPS_PER_TEMP; j++) { 257 258 7. COMBINATORIAL SEARCH AND HEURISTIC METHODS /* pick indices of elements to swap */ i1 = random_int(1,t->n); i2 = random_int(1,t->n); flip = random_float(0,1); delta = transition(s,t,i1,i2); exponent = (-delta/current_value)/(K*temperature); merit = pow(E,exponent); if (delta < 0) current_value = current_value+delta; else { if (merit > flip) current_value = current_value+delta; else transition(s,t,i1,i2); } /*ACCEPT-WIN*/ /*ACCEPT-LOSS*/ /* REJECT */ } /* restore temperature if progress has been made */ if ((current_value-start_value) < 0.0) temperature = temperature/COOLING_FRACTION; } } 7.5.4 Applications of Simulated Annealing We provide several examples to demonstrate how these components can lead to elegant simulated annealing solutions for real combinatorial search problems. Maximum Cut The “maximum cut” problem seeks to partition the vertices of a weighted graph G into sets V1 and V2 to maximize the weight (or number) of edges with one vertex in each set. For graphs that specify an electronic circuit, the maximum cut in the graph deﬁnes the largest amount of data communication that can take place in the circuit simultaneously. As discussed in Section 16.6 (page 541), maximum cut is NP-complete. How can we formulate maximum cut for simulated annealing? The solution space consists of all 2n−1 possible vertex partitions. We save a factor of two over all vertex subsets because vertex v1 can be assumed to be ﬁxed on the left side of the partition. The subset of vertices accompanying it can be represented using 7.5 HEURISTIC SEARCH METHODS a bit vector. The cost of a solution is the sum of the weights cut in the current conﬁguration. A natural transition mechanism selects one vertex at random and moves it across the partition simply by ﬂipping the corresponding bit in the bit vector. The change in the cost function will be the weight of its old neighbors minus the weight of its new neighbors. This can be computed in time proportional to the degree of the vertex. This kind of simple, natural modeling is the right type of heuristic to seek in practice. Independent Set An “independent set” of a graph G is a subset of vertices S such that there is no edge with both endpoints in S. The maximum independent set of a graph is the largest such empty induced subgraph. Finding large independent sets arises in dispersion problems associated with facility location and coding theory, as discussed in Section 16.2 (page 528). The natural state space for a simulated annealing solution would be all 2n subsets of the vertices, represented as a bit vector. As with maximum cut, a simple transition mechanism would add or delete one vertex from S. One natural cost function for subset S might be 0 if S contains an edge, and |S| if it is indeed an independent set. This function ensures that we work towards an independent set at all times. However, this condition is strict enough that we are liable to move only in a narrow portion of the possible search space. More ﬂexibility in the search space and quicker cost function computations can result from allowing nonempty graphs at the early stages of cooling. Better in practice would be a cost function like C(S) = |S| − λ · mS /T , where λ is a constant, T is the temperature, and mS is the number of edges in the subgraph induced by S. The dependence of C(S) on T ensures that the search will drive the edges out faster as the system cools. Circuit Board Placement In designing printed circuit boards, we are faced with the problem of positioning modules (typically, integrated circuits) on the board. Desired criteria in a layout may include (1) minimizing the area or aspect ratio of the board so that it properly ﬁts within the allotted space, and (2) minimizing the total or longest wire length in connecting the components. Circuit board placement is representative of the kind of messy, multicriterion optimization problems for which simulated annealing is ideally suited. Formally, we are given a collection of rectangular modules r1 , . . . , rn , each with associated dimensions hi × li . Further, for each pair of modules ri , rj , we are given the number of wires wij that must connect the two modules. We seek a placement of the rectangles that minimizes area and wire length, subject to the constraint that no two rectangles overlap each other. 259 260 7. COMBINATORIAL SEARCH AND HEURISTIC METHODS The state space for this problem must describe the positions of each rectangle. To make this discrete, the rectangles can be restricted to lie on vertices of an integer grid. Reasonable transition mechanisms including moving one rectangle to a diﬀerent location, or swapping the position of two rectangles. A natural cost function would be C(S) = λarea (Sheight · Swidth ) + n n (λwire · wij · dij + λoverlap (ri ∩ rj )) i=1 j=1 where λarea , λwire , and λoverlap are constants governing the impact of these components on the cost function. Presumably, λoverlap should be an inverse function of temperature, so after gross placement it adjusts the rectangle positions to be disjointed. Take-Home Lesson: Simulated annealing is a simple but eﬀective technique for eﬃciently obtaining good but not optimal solutions to combinatorial search problems. 7.6 War Story: Only it is Not a Radio “Think of it as a radio,” he chuckled. “Only it is not a radio.” I’d been whisked by corporate jet to the research center of a large but very secretive company located somewhere east of California. They were so paranoid that I never got to see the object we were working on, but the people who brought me in did a great job of abstracting the problem. The application concerned a manufacturing technique known as selective assembly. Eli Whitney started the Industrial Revolution through his system of interchangeable parts. He carefully speciﬁed the manufacturing tolerances on each part in his machine so that the parts were interchangeable, meaning that any legal cog-widget could be used to replace any other legal cog-widget. This greatly sped up the process of manufacturing, because the workers could just put parts together instead of having to stop to ﬁle down rough edges and the like. It made replacing broken parts a snap. This was a very good thing. Unfortunately, it also resulted in large piles of cog-widgets that were slightly outside the manufacturing tolerance and thus had to be discarded. Another clever fellow then observed that maybe one of these defective cog-widgets could be used when all the other parts in the given assembly exceeded their required manufacturing tolerances. Good plus bad could well equal good enough. This is the idea of selective assembly. “Each not-radio is made up of n diﬀerent types of not-radio parts,” he told me. For the ith part type (say the right ﬂange gasket), we have a pile of si instances of this part type. Each part (ﬂange gasket) comes with a measure of the degree to which it deviates from perfection. We need to match up the parts so as to create the greatest number of working not-radios as possible.” 7.6 WAR STORY: ONLY IT IS NOT A RADIO 17 14 15 5 33 11 11 6 32 261 Figure 7.11: Part assignments for three not-radios, such that each had at most 50 points total defect The situation is illustrated in Figure 7.11. Each not-radio consists of three parts, and the sum of the defects in any functional not-radio must total at most 50. By cleverly balancing the good and bad parts in each machine, we can use all the parts and make three working not-radios. I thought about the problem. The simplest procedure would take the best part for each part type, make a not-radio out of them, and repeat until the not-radio didn’t play (or do whatever a not-radio does). But this would create a small number of not-radios drastically varying in quality, whereas they wanted as many decent not-radios as possible. The goal was to match up good parts and bad parts so the total amount of badness wasn’t so bad. Indeed, the problem sounded related to matching in graphs (see Section 15.6 (page 498)). Suppose we built a graph where the vertices were the part instances, and add an edge for all two part instances that were within the total error tolerance. In graph matching, we seek the largest number of edges such that no vertex appears more than once in the matching. This is analogous to the largest number of two-part assemblies we can form from the given set of parts. “I can solve your problem using matching,” I announced, “Provided not-radios are each made of only two parts.” There was silence. Then they all started laughing at me. “Everyone knows not-radios have more than two parts,” they said, shaking their heads. That spelled the end of this algorithmic approach. Extending to more than two parts turned the problem into matching on hypergraphs,2 —a problem which is NP-complete. Further, it might take exponential time in the number of part types just to build the graph, since we had to explicitly construct each possible hyperedge/assembly. 2 A hypergraph is made up of edges that can contain more than two vertices each. They can be thought of as general collections of subsets of vertices/elements. 262 7. COMBINATORIAL SEARCH AND HEURISTIC METHODS I went back to the drawing board. They wanted to put parts into assemblies so that no assembly would have more total defects than allowed. Described that way, it sounded like a packing problem. In the bin packing problem (see Section 17.9 (page 595)), we are given a collection of items of diﬀerent sizes and asked to store them using the smallest possible number of bins, each of which has a ﬁxed capacity of size k. Here, the assemblies represented the bins, each of which could absorb total defect ≤ k. The items to pack represented the individual parts, whose size would reﬂect its quality of manufacture. It wasn’t pure bin packing, however, since parts came in diﬀerent types. The application imposed constraints on the allowable contents of each bin. Creating the maximum number of not-radios meant that we sought a packing that maximized the number of bins which contained exactly one part for each of the m diﬀerent parts types. Bin packing is NP-complete, but is a natural candidate for a heuristic search approach. The solution space consists of assignments of parts to bins. We initially pick a random part of each type for each bin to give us a starting conﬁguration for the search. The local neighborhood operation involves moving parts around from one bin to another. We could move one part at a time, but more eﬀective was swapping parts of a particular type between two randomly chosen bins. In such a swap, both bins remain complete not-radios, hopefully with better error tolerance than before. Thus, our swap operator required three random integers—one to select the appropriate parts type (from 1 to m) and two more to select the assembly bins involved (say from 1 to b). The key decision was the cost function to use. They supplied the hard limit k on the total defect level for each individual assembly. But what was the best way to score a set of assemblies? We could just return the number of acceptable complete assemblies as our score—an integer from 1 to b. Although this was indeed what we wanted to optimize, it would not be sensitive to detect when we were making partial progress towards a solution. Suppose one of our swaps succeeded in bringing one of the nonfunctional assemblies much closer to the not-radio limit k. That would be a better starting point for further progress than the original, and should be favored. My ﬁnal cost function was as follows. I gave one point for every working assembly, and a signiﬁcantly smaller total for each nonworking assembly based on how close it was to the threshold k. The score for a nonworking assembly decreased exponentially based on how much it was over k. Thus the optimizer would seek to maximize the number of working assemblies, and then try to drive another assembly close to the limit. I implemented this algorithm, and then ran the search on the test case they provided. It was an instance taken directly from the factory ﬂoor. Not-radios turn out to contain m = 8 important parts types. Some parts types are more expensive than others, and so they have fewer available candidates to consider. The most 7.7 preﬁx AA AG GA GG AA AAAA AGAA GAAG GGAA WAR STORY: ANNEALING ARRAYS suﬃx AG AAAG AGAG GAAG GGAG GA AAGA AGGA GAGA GGGA GG AAGG AGGG GAGG GGGG Figure 7.12: A preﬁx-suﬃx array of all purine 4-mers constrained parts type had only eight representatives, so there could be at most eight possible assemblies from this given mix. I watched as simulated annealing chugged and bubbled on the problem instance. The number of completed assemblies instantly climbed (1, 2, 3, 4) before progress started to slow a bit. Then came 5 and 6 in a hiccup, with a pause before assembly 7 came triumphantly together. But tried as it might, the program could not put together eight not-radios before I lost interest in watching. I called and tried to admit defeat, but they wouldn’t hear it. It turns out that the best the factory had managed after extensive eﬀorts was only six working notradios, so my result represented a signiﬁcant improvement! 7.7 War Story: Annealing Arrays The war story of Section 3.9 (page 94) reported how we used advanced data structures to simulate a new method for sequencing DNA. Our method, interactive sequencing by hybridization (SBH), required building arrays of speciﬁc oligonucleotides on demand. A biochemist at Oxford University got interested in our technique, and moreover he had in his laboratory the equipment we needed to test it out. The Southern Array Maker, manufactured by Beckman Instruments, prepared discrete oligonucleotide sequences in 64 parallel rows across a polypropylene substrate. The device constructs arrays by appending single characters to each cell along speciﬁc rows and columns of arrays. Figure 7.12 shows how to construct an array of all 24 = 16 purine (A or G) 4-mers by building the preﬁxes along rows and the suﬃxes along columns. This technology provided an ideal environment for testing the feasibility of interactive SBH in a laboratory, because with proper programming it gave a way to fabricate a wide variety of oligonucleotide arrays on demand. However, we had to provide the proper programming. Fabricating complicated arrays required solving a diﬃcult combinatorial problem. We were given as input a set of n strings (representing oligonucleotides) to fabricate in an m × m array (where m = 64 on the Southern apparatus). We had to produce a schedule of row and column commands to realize the set of strings S. We proved that the 263 264 7. COMBINATORIAL SEARCH AND HEURISTIC METHODS problem of designing dense arrays was NP-complete, but that didn’t really matter. My student Ricky Bradley and I had to solve it anyway. “We are going to have to use a heuristic,” I told him. “So how can we model this problem?” “Well, each string can be partitioned into preﬁx and suﬃx pairs that realize it. For example, the string ACC can be realized in four diﬀerent ways: preﬁx ‘’ and suﬃx ACC, preﬁx A and suﬃx CC, preﬁx AC and suﬃx C, or preﬁx ACC and suﬃx ‘’. We seek the smallest set of preﬁxes and suﬃxes that together realize all the given strings,” Ricky said. “Good. This gives us a natural representation for simulated annealing. The state space will consist of all possible subsets of preﬁxes and suﬃxes. The natural transitions between states might include inserting or deleting strings from our subsets, or swapping a pair in or out.” “What’s a good cost function?” he asked. “Well, we need as small an array as possible that covers all the strings. How about taking the maximum of number of rows (preﬁxes) or columns (suﬃxes) used in our array, plus the number of strings from S that are not yet covered. Try it and let’s see what happens.” Ricky went oﬀ and implemented a simulated annealing program along these lines. It printed out the state of the solution each time a transition was accepted and was fun to watch. The program quickly kicked out unnecessary preﬁxes and suﬃxes, and the array began shrinking rapidly in size. But after several hundred iterations, progress started to slow. A transition would knock out an unnecessary suﬃx, wait a while, then add a diﬀerent suﬃx back again. After a few thousand iterations, no real improvement was happening. “The program doesn’t seem to recognize when it is making progress. The evaluation function only gives credit for minimizing the larger of the two dimensions. Why not add a term to give some credit to the other dimension.” Ricky changed the evaluation function, and we tried again. This time, the program did not hesitate to improve the shorter dimension. Indeed, our arrays started to be skinny rectangles instead of squares. “OK. Let’s add another term to the evaluation function to give it points for being roughly square.” Ricky tried again. Now the arrays were the right shape, and progress was in the right direction. But the progress was still slow. “Too many of the insertion moves don’t aﬀect many strings. Maybe we should skew the random selections so that the important preﬁx/suﬃxes get picked more often.” Ricky tried again. Now it converged faster, but sometimes it still got stuck. We changed the cooling schedule. It did better, but was it doing well? Without a lower bound knowing how close we were to optimal, it couldn’t really tell how good our solution was. We tweaked and tweaked until our program stopped improving. Our ﬁnal solution reﬁned the initial array by applying the following random moves: 7.7 WAR STORY: ANNEALING ARRAYS 265 Figure 7.13: Compression of the HIV array by simulated annealing – after 0, 500, 1,000, and 5,750 iterations • swap – swap a preﬁx/suﬃx on the array with one that isn’t. • add – add a random preﬁx/suﬃx to the array. • delete – delete a random preﬁx/suﬃx from the array. • useful add – add the preﬁx/suﬃx with the highest usefulness to the array. • useful delete – delete the preﬁx/suﬃx with the lowest usefulness from the array. • string add – randomly select a string not on the array, and add the most useful preﬁx and/or suﬃx to cover this string. A standard cooling schedule was used, with an exponentially decreasing temperature (dependent upon the problem size) and a temperature-dependent Boltzmann criterion for accepting states that have higher costs. Our ﬁnal cost function was deﬁned as 2 cost = 2 × max + min + (max − min) + 4(strtotal − strin ) 4 where max is the size of the maximum chip dimension, min is the size of the minimum chip dimension, strtotal = |S|, and strin is the number of strings from S currently on the chip. How well did we do? Figure 7.13 shows the convergence of a custom array consisting of the 5,716 unique 7-mers of the HIV virus. Figure 7.13 shows snapshots of the state of the chip at four points during the annealing process (0, 500, 1,000, and the ﬁnal chip at 5,750 iterations). Black pixels represent the ﬁrst occurrence of an HIV 7-mer. The ﬁnal chip size here is 130 × 132—quite an improvement over 266 7. COMBINATORIAL SEARCH AND HEURISTIC METHODS the initial size of 192 × 192. It took about ﬁfteen minutes’ worth of computation to complete the optimization, which was perfectly acceptable for the application. But how well did we do? Since simulated annealing is only a heuristic, we really don’t know how close to optimal our solution is. I think we did pretty well, but can’t really be sure. Simulated annealing is a good way to handle complex optimization problems. However, to get the best results, expect to spend more time tweaking and reﬁning your program than you did in writing it in the ﬁrst place. This is dirty work, but sometimes you have to do it. 7.8 Other Heuristic Search Methods Several heuristic search methods have been proposed to search for good solutions for combinatorial optimization problems. Like simulated annealing, many techniques relies on analogies to real-world physical processes. Popular methods include genetic algorithms, neural networks, and ant colony optimization. The intuition behind these methods is highly appealing, but skeptics decry them as voodoo optimization techniques that rely more on nice analogies to nature than demonstrated computational results on problems that have been studied using other methods. The question isn’t whether you can get decent answers for many problems given enough eﬀort using these techniques. Clearly you can. The real question is whether they lead to better solutions with less implementation complexity than the other methods we have discussed. In general, I don’t believe that they do. But in the spirit of free inquiry, I introduce genetic algorithms, which is the most popular of these methods. See the chapter notes for more detailed readings. Genetic Algorithms Genetic algorithms draw their inspiration from evolution and natural selection. Through the process of natural selection, organisms adapt to optimize their chances for survival in a given environment. Random mutations occur in an organism’s genetic description, which then get passed on to its children. Should a mutation prove helpful, these children are more likely to survive and reproduce. Should it be harmful, these children won’t, and so the bad trait will die with them. Genetic algorithms maintain a “population” of solution candidates for the given problem. Elements are drawn at random from this population and allowed to “reproduce” by combining aspects of the two-parent solutions. The probability that an element is chosen to reproduce is based on its “ﬁtness,”—essentially the cost of the solution it represents. Unﬁt elements die from the population, to be replaced by a successful-solution oﬀspring. The idea behind genetic algorithms is extremely appealing. However, they don’t seem to work as well on practical combinatorial optimization problems as simulated 7.9 PARALLEL ALGORITHMS annealing does. There are two primary reasons for this. First, it is quite unnatural to model applications in terms of genetic operators like mutation and crossover on bit strings. The pseudobiology adds another level of complexity between you and your problem. Second, genetic algorithms take a very long time on nontrivial problems. The crossover and mutation operations typically make no use of problem-speciﬁc structure, so most transitions lead to inferior solutions, and convergence is slow. Indeed, the analogy with evolution—where signiﬁcant progress require millions of years—can be quite appropriate. We will not discuss genetic algorithms further, to discourage you from considering them for your applications. However, pointers to implementations of genetic algorithms are provided in Section 13.5 (page 407) if you really insist on playing with them. Take-Home Lesson: I have never encountered any problem where genetic algorithms seemed to me the right way to attack it. Further, I have never seen any computational results reported using genetic algorithms that have favorably impressed me. Stick to simulated annealing for your heuristic search voodoo needs. 7.9 Parallel Algorithms Two heads are better than one, and more generally, n heads are better than n − 1. Parallel processing is becoming more important with the advent of cluster computing and multicore processors. It seems like the easy way out of hard problems. Indeed, sometimes, for some problems, parallel algorithms are the most eﬀective solution. High-resolution, real-time graphics applications must render thirty frames per second for realistic animation. Assigning each frame to a distinct processor, or dividing each image into regions assigned to diﬀerent processors, might be the only way to get the job done in time. Large systems of linear equations for scientiﬁc applications are routinely solved in parallel. However, there are several pitfalls associated with parallel algorithms that you should be aware of: • There is often a small upper bound on the potential win – Suppose that you have access to twenty processors that can be devoted exclusively to your job. Potentially, these could be used to speed up the fastest sequential program by up to a factor of twenty. That is nice, but greater performance gains maybe possible by ﬁnding a better sequential algorithm. Your time spent parallelizing a code might well be better spent enhancing the sequential version. Performance-tuning tools such as proﬁlers are better developed for sequential machines than for parallel models. • Speedup means nothing – Suppose my parallel program runs 20 times faster on a 20-processor machine than it does on one processor. That’s great, isn’t 267 268 7. COMBINATORIAL SEARCH AND HEURISTIC METHODS it? If you always get linear speedup and have an arbitrary number of processors, you will eventually beat any sequential algorithm. However, a carefully designed sequential algorithm can often beat an easily-parallelized code running on a typical parallel machine. The one-processor parallel version of your code is likely to be a crummy sequential algorithm, so measuring speedup typically provides an unfair test of the beneﬁts of parallelism. The classic example of this occurs in the minimax game-tree search algorithm used in computer chess programs. A brute-force tree search is embarrassingly easy to parallelize: just put each subtree on a diﬀerent processor. However, a lot of work gets wasted because the same positions get considered on diﬀerent machines. Moving from a brute-force search to the more clever alpha-beta pruning algorithm can easily save 99.99% of the work, thus dwarﬁng any beneﬁts of a parallel brute-force search. Alpha-beta can be parallelized, but not easily, and the speedups grow surprisingly slowly as a function of the number of processors you have. • Parallel algorithms are tough to debug – Unless your problem can be decomposed into several independent jobs, the diﬀerent processors must communicate with each other to end up with the correct ﬁnal result. Unfortunately, the nondeterministic nature of this communication makes parallel programs notoriously diﬃcult to debug. Perhaps the best example is Deep Blue—the world-champion chess computer. Although it eventually beat Kasparov, over the years it lost several games in embarrassing fashion due to bugs, mostly associated with its extensive parallelism. I recommend considering parallel processing only after attempts at solving a problem sequentially prove too slow. Even then, I would restrict attention to algorithms that parallelize the problem by partitioning the input into distinct tasks where no communication is needed between the processors, except to collect the ﬁnal results. Such large-grain, naive parallelism can be simple enough to be both implementable and debuggable, because it really reduces to producing a good sequential implementation. There can be pitfalls even in this approach, however, as shown in the war story below. 7.10 War Story: Going Nowhere Fast In Section 2.8 (page 51), I related our eﬀorts to build a fast program to test Waring’s conjecture for pyramidal numbers. At that point, my code was fast enough that it could complete the job in a few weeks running in the background of a desktop workstation. This option did not appeal to my supercomputing colleague, however. “Why don’t we do it in parallel?” he suggested. “After all, you have an outer loop doing the same calculation on each integer from 1 to 1,000,000,000. I can split this range of numbers into diﬀerent intervals and run each range on a diﬀerent processor. Watch, it will be easy.” 7.10 WAR STORY: GOING NOWHERE FAST He set to work trying to do our computations on an Intel IPSC-860 hypercube using 32 nodes with 16 megabytes of memory per node—very big iron for the time. However, instead of getting answers, I was treated to a regular stream of e-mail about system reliability over the next few weeks: • “Our code is running ﬁne, except one processor died last night. I will rerun.” • “This time the machine was rebooted by accident, so our long-standing job was killed.” • “We have another problem. The policy on using our machine is that nobody can command the entire machine for more than thirteen hours, under any condition.” Still, eventually, he rose to the challenge. Waiting until the machine was stable, he locked out 16 processors (half the computer), divided the integers from 1 to 1,000,000,000 into 16 equal-sized intervals, and ran each interval on its own processor. He spent the next day fending oﬀ angry users who couldn’t get their work done because of our rogue job. The instant the ﬁrst processor completed analyzing the numbers from 1 to 62,500,000, he announced to all the people yelling at him that the rest of the processors would soon follow. But they didn’t. He failed to realize that the time to test each integer increased as the numbers got larger. After all, it would take longer to test whether 1,000,000,000 could be expressed as the sum of three pyramidal numbers than it would for 100. Thus, at slower and slower intervals each new processor would announce its completion. Because of the architecture of the hypercube, he couldn’t return any of the processors until our entire job was completed. Eventually, half the machine and most of its users were held hostage by one, ﬁnal interval. What conclusions can be drawn from this? If you are going to parallelize a problem, be sure to balance the load carefully among the processors. Proper load balancing, using either back-of-the-envelope calculations or the partition algorithm we will develop in Section 8.5 (page 294), would have signiﬁcantly reduced the time we needed the machine, and his exposure to the wrath of his colleagues. Chapter Notes The treatment of backtracking here is partially based on my book Programming Challenges [SR03]. In particular, the backtrack routine presented here is a generalization of the version in Chapter 8 of [SR03]. Look there for my solution to the famous eight queens problem, which seeks all chessboard conﬁgurations of eight mutually nonattacking queens on an 8 × 8 board. The original paper on simulated annealing [KGV83] included an application to VLSI module placement problems. The applications from Section 7.5.4 (page 258) are based on material from [AK89]. 269 270 7. COMBINATORIAL SEARCH AND HEURISTIC METHODS The heuristic TSP solutions presented here employ vertex-swap as the local neighborhood operation. In fact, edge-swap is a more powerful operation. Each edge-swap changes two edges in the tour at most, as opposed to at most four edges with a vertex-swap. This improves the possibility of a local improvement. However, more sophisticated data structures are necessary to eﬃciently maintain the order of the resulting tour [FJMO93]. The diﬀerent heuristic search techniques are ably presented in Aarts and Lenstra [AL97], which I strongly recommend for those interested in learning more about heuristic searches. Their coverage includes tabu search, a variant of simulated annealing that uses extra data structures to avoid transitions to recently visited states. Ant colony optimization is discussed in [DT04]. See [MF00] for a more favorable view of genetic algorithms and the like. More details on our combinatorial search for optimal chessboard-covering positions appear in our paper [RHS89]. Our work using simulated annealing to compress DNA arrays was reported in [BS97]. See Pugh [Pug86] and Coullard et al. [CGJ98] for more on selective assembly. Our parallel computations on pyramidal numbers were reported in [DY94]. 7.11 Exercises Backtracking 7-1. [3] A derangement is a permutation p of {1, . . . , n} such that no item is in its proper position, i.e. pi = i for all 1 ≤ i ≤ n. Write an eﬃcient backtracking program with pruning that constructs all the derangements of n items. 7-2. [4] Multisets are allowed to have repeated elements. A multiset of n items may thus have fewer than n! distinct permutations. For example, {1, 1, 2, 2} has only six diﬀerent permutations: {1, 1, 2, 2}, {1, 2, 1, 2}, {1, 2, 2, 1}, {2, 1, 1, 2}, {2, 1, 2, 1}, and {2, 2, 1, 1}. Design and implement an eﬃcient algorithm for constructing all permutations of a multiset. 7-3. [5] Design and implement an algorithm for testing whether two graphs are isomorphic to each other. The graph isomorphism problem is discussed in Section 16.9 (page 550). With proper pruning, graphs on hundreds of vertices can be tested reliably. 7-4. [5] Anagrams are rearrangements of the letters of a word or phrase into a diﬀerent word or phrase. Sometimes the results are quite striking. For example, “MANY VOTED BUSH RETIRED” is an anagram of “TUESDAY NOVEMBER THIRD,” which correctly predicted the result of the 1992 U.S. presidential election. Design and implement an algorithm for ﬁnding anagrams using combinatorial search and a dictionary. 7-5. [8] Design and implement an algorithm for solving the subgraph isomorphism problem. Given graphs G and H, does there exist a subgraph H of H such that G is isomorphic to H ? How does your program perform on such special cases of subgraph isomorphism as Hamiltonian cycle, clique, independent set, and graph isomorphism? 7.11 EXERCISES 7-6. [8] In the turnpike reconstruction problem, you are given n(n − 1)/2 distances in sorted order. The problem is to ﬁnd the positions of n points on the line that give rise to these distances. For example, the distances {1, 2, 3, 4, 5, 6} can be determined by placing the second point 1 unit from the ﬁrst, the third point 3 from the second, and the fourth point 2 from the third. Design and implement an eﬃcient algorithm to report all solutions to the turnpike reconstruction problem. Exploit additive constraints when possible to minimize the search. With proper pruning, problems with hundreds of points can be solved reliably. Combinatorial Optimization For each of the problems below, either (1) implement a combinatorial search program to solve it optimally for small instance, and/or (2) design and implement a simulated annealing heuristic to get reasonable solutions. How well does your program perform in practice? 7-7. [5] Design and implement an algorithm for solving the bandwidth minimization problem discussed in Section 13.2 (page 398). 7-8. [5] Design and implement an algorithm for solving the maximum satisﬁability problem discussed in Section 14.10 (page 472). 7-9. [5] Design and implement an algorithm for solving the maximum clique problem discussed in Section 16.1 (page 525). 7-10. [5] Design and implement an algorithm for solving the minimum vertex coloring problem discussed in Section 16.7 (page 544). 7-11. [5] Design and implement an algorithm for solving the minimum edge coloring problem discussed in Section 16.8 (page 548). 7-12. [5] Design and implement an algorithm for solving the minimum feedback vertex set problem discussed in Section 16.11 (page 559). 7-13. [5] Design and implement an algorithm for solving the set cover problem discussed in Section 18.1 (page 621). Interview Problems 7-14. [4] Write a function to ﬁnd all permutations of the letters in a particular string. 7-15. [4] Implement an eﬃcient algorithm for listing all k-element subsets of n items. 7-16. [5] An anagram is a rearrangement of the letters in a given string into a sequence of dictionary words, like Steven Skiena into Vainest Knees. Propose an algorithm to construct all the anagrams of a given string. 7-17. [5] Telephone keypads have letters on each numerical key. Write a program that generates all possible words resulting from translating a given digit sequence (e.g., 145345) into letters. 7-18. [7] You start with an empty room and a group of n people waiting outside. At each step, you may either admit one person into the room, or let one out. Can you arrange a sequence of 2n steps, so that every possible combination of people is achieved exactly once? 271 272 7. COMBINATORIAL SEARCH AND HEURISTIC METHODS 7-19. [4] Use a random number generator (rng04) that generates numbers from {0, 1, 2, 3, 4} with equal probability to write a random number generator that generates numbers from 0 to 7 (rng07) with equal probability. What are expected number of calls to rng04 per call of rng07? Programming Challenges These programming challenge problems with robot judging are available at http://www.programming-challenges.com or http://online-judge.uva.es. 7-1. “Little Bishops” – Programming Challenges 110801, UVA Judge 861. 7-2. “15-Puzzle Problem” – Programming Challenges 110802, UVA Judge 10181. 7-3. “Tug of War” – Programming Challenges 110805, UVA Judge 10032. 7-4. “Color Hash” – Programming Challenges 110807, UVA Judge 704. 8 Dynamic Programming The most challenging algorithmic problems involve optimization, where we seek to ﬁnd a solution that maximizes or minimizes some function. Traveling salesman is a classic optimization problem, where we seek the tour visiting all vertices of a graph at minimum total cost. But as shown in Chapter 1, it is easy to propose “algorithms” solving TSP that generate reasonable-looking solutions but did not always produce the minimum cost tour. Algorithms for optimization problems require proof that they always return the best possible solution. Greedy algorithms that make the best local decision at each step are typically eﬃcient but usually do not guarantee global optimality. Exhaustive search algorithms that try all possibilities and select the best always produce the optimum result, but usually at a prohibitive cost in terms of time complexity. Dynamic programming combines the best of both worlds. It gives us a way to design custom algorithms that systematically search all possibilities (thus guaranteeing correctness) while storing results to avoid recomputing (thus providing eﬃciency). By storing the consequences of all possible decisions and using this information in a systematic way, the total amount of work is minimized. Once you understand it, dynamic programming is probably the easiest algorithm design technique to apply in practice. In fact, I ﬁnd that dynamic programming algorithms are often easier to reinvent than to try to look up in a book. That said, until you understand dynamic programming, it seems like magic. You must ﬁgure out the trick before you can use it. Dynamic programming is a technique for eﬃciently implementing a recursive algorithm by storing partial results. The trick is seeing whether the naive recursive algorithm computes the same subproblems over and over and over again. If so, storing the answer for each subproblems in a table to look up instead of recompute S.S. Skiena, The Algorithm Design Manual, 2nd ed., DOI: 10.1007/978-1-84800-070-4 8, c Springer-Verlag London Limited 2008 274 8. DYNAMIC PROGRAMMING can lead to an eﬃcient algorithm. Start with a recursive algorithm or deﬁnition. Only once we have a correct recursive algorithm do we worry about speeding it up by using a results matrix. Dynamic programming is generally the right method for optimization problems on combinatorial objects that have an inherent left to right order among components. Left-to-right objects includes: character strings, rooted trees, polygons, and integer sequences. Dynamic programming is best learned by carefully studying examples until things start to click. We present three war stories where dynamic programming played the decisive role to demonstrate its utility in practice. 8.1 Caching vs. Computation Dynamic programming is essentially a tradeoﬀ of space for time. Repeatedly recomputing a given quantity is harmless unless the time spent doing so becomes a drag on performance. Then we are better oﬀ storing the results of the initial computation and looking them up instead of recomputing them again. The tradeoﬀ between space and time exploited in dynamic programming is best illustrated when evaluating recurrence relations such as the Fibonacci numbers. We look at three diﬀerent programs for computing them below. 8.1.1 Fibonacci Numbers by Recursion The Fibonacci numbers were originally deﬁned by the Italian mathematician Fibonacci in the thirteenth century to model the growth of rabbit populations. Rabbits breed, well, like rabbits. Fibonacci surmised that the number of pairs of rabbits born in a given year is equal to the number of pairs of rabbits born in each of the two previous years, starting from one pair of rabbits in the ﬁrst year. To count the number of rabbits born in the nth year, he deﬁned the recurrence relation: Fn = Fn−1 + Fn−2 with basis cases F0 = 0 and F1 = 1. Thus, F2 = 1, F3 = 2, and the series continues {3, 5, 8, 13, 21, 34, 55, 89, 144, . . .}. As it turns out, Fibonacci’s formula didn’t do a very good job of counting rabbits, but it does have a host of interesting properties. Since they are deﬁned by a recursive formula, it is easy to write a recursive program to compute the nth Fibonacci number. A recursive function algorithm written in C looks like this: long fib_r(int n) { if (n == 0) return(0); if (n == 1) return(1); return(fib_r(n-1) + fib_r(n-2)); } 8.1 CACHING VS. COMPUTATION F(6)=8 F(5) F(4) F(4) F(3) F(2) F(1) F(3) F(3) F(2) F(1) F(1) F(2) F(0) F(1) F(0) F(1) F(2) F(1) F(2) F(1) F(1) F(0) F(0) F(0) Figure 8.1: The computation tree for computing Fibonacci numbers recursively The course of execution for this recursive algorithm is illustrated by its recursion tree, as illustrated in Figure 8.1. This tree is evaluated in a depth-ﬁrst fashion, as are all recursive algorithms. I encourage you to trace this example by hand to refresh your knowledge of recursion. Note that F (4) is computed on both sides of the recursion tree, and F (2) is computed no less than ﬁve times in this small example. The weight of all this redundancy becomes clear when you run the program. It took more than 7 minutes for my program to compute the ﬁrst 45 Fibonacci numbers. You could probably do it faster by hand using the right algorithm. How much time does this algorithm take to compute F (n)? Since Fn+1 /Fn ≈ √ φ = (1 + 5)/2 ≈ 1.61803, this means that Fn > 1.6n . Since our recursion tree has only 0 and 1 as leaves, summing up to such a large number means we must have at least 1.6n leaves or procedure calls! This humble little program takes exponential time to run! 8.1.2 Fibonacci Numbers by Caching In fact, we can do much better. We can explicitly store (or cache) the results of each Fibonacci computation F (k) in a table data structure indexed by the parameter k. The key to avoiding recomputation is to explicitly check for the value before trying to compute it: #define MAXN 45 #define UNKNOWN -1 long f[MAXN+1]; /* largest interesting n */ /* contents denote an empty cell */ /* array for caching computed fib values */ 275 276 8. DYNAMIC PROGRAMMING F(6)=8 F(5) F(4) F(3) F(2) F(1) F(4) F(3) F(2) F(1) F(0) Figure 8.2: The Fibonacci computation tree when caching values long fib_c(int n) { if (f[n] == UNKNOWN) f[n] = fib_c(n-1) + fib_c(n-2); return(f[n]); } long fib_c_driver(int n) { int i; f[0] = 0; f[1] = 1; for (i=2; i<=n; i++) /* counter */ f[i] = UNKNOWN; return(fib_c(n)); } To compute F (n), we call fib c driver(n). This initializes our cache to the two values we initially know (F (0) and F (1)) as well as the UNKNOWN ﬂag for all the rest we don’t. It then calls a look-before-crossing-the-street version of the recursive algorithm. This cached version runs instantly up to the largest value that can ﬁt in a long integer. The new recursion tree (Figure 8.2) explains why. There is no meaningful branching, because only the left-side calls do computation. The right-side calls ﬁnd what they are looking for in the cache and immediately return. 8.1 CACHING VS. COMPUTATION What is the running time of this algorithm? The recursion tree provides more of a clue than the code. In fact, it computes F (n) in linear time (in other words, O(n) time) because the recursive function fib c(k) is called exactly twice for each value 0 ≤ k ≤ n. This general method of explicitly caching results from recursive calls to avoid recomputation provides a simple way to get most of the beneﬁts of full dynamic programming, so it is worth a more careful look. In principle, such caching can be employed on any recursive algorithm. However, storing partial results would have done absolutely no good for such recursive algorithms as quicksort, backtracking, and depth-ﬁrst search because all the recursive calls made in these algorithms have distinct parameter values. It doesn’t pay to store something you will never refer to again. Caching makes sense only when the space of distinct parameter values is modest enough that we can aﬀord the cost of storage. Since the argument to the recursive function fib c(k) is an integer between 0 and n, there are only O(n) values to cache. A linear amount of space for an exponential amount of time is an excellent tradeoﬀ. But as we shall see, we can do even better by eliminating the recursion completely. Take-Home Lesson: Explicit caching of the results of recursive calls provides most of the beneﬁts of dynamic programming, including usually the same running time as the more elegant full solution. If you prefer doing extra programming to more subtle thinking, you can stop here. 8.1.3 Fibonacci Numbers by Dynamic Programming We can calculate Fn in linear time more easily by explicitly specifying the order of evaluation of the recurrence relation: long fib_dp(int n) { int i; long f[MAXN+1]; f[0] = 0; f[1] = 1; for (i=2; i<=n; i++) /* counter */ /* array to cache computed fib values */ f[i] = f[i-1]+f[i-2]; return(f[n]); } We have removed all recursive calls! We evaluate the Fibonacci numbers from smallest to biggest and store all the results, so we know that we have Fi−1 and Fi−2 ready whenever we need to compute Fi . The linearity of this algorithm should 277 278 8. DYNAMIC PROGRAMMING be apparent. Each of the n values is computed as the simple sum of two integers in total O(n) time and space. More careful study shows that we do not need to store all the intermediate values for the entire period of execution. Because the recurrence depends on two arguments, we only need to retain the last two values we have seen: long fib_ultimate(int n) { int i; long back2=0, back1=1; long next; /* counter */ /* last two values of f[n] */ /* placeholder for sum */ if (n == 0) return (0); for (i=2; i 0) m[0][i].parent = INSERT; else m[0][i].parent = -1; } column_init(int i) { m[i][0].cost = i; if (i>0) m[i][0].parent = DELETE; else m[i][0].parent = -1; } • Penalty Costs – The functions match(c,d) and indel(c) present the costs for transforming character c to d and inserting/deleting character c. For standard edit distance, match should cost nothing if the characters are identical, and 1 otherwise; while indel returns 1 regardless of what the argument is. But application-speciﬁc cost functions can be employed, perhaps more forgiving of replacements located near each other on standard keyboard layouts or characters that sound or look similar. 8.2 int match(char c, char d) { if (c == d) return(0); else return(1); } APPROXIMATE STRING MATCHING int indel(char c) { return(1); } • Goal Cell Identiﬁcation – The function goal cell returns the indices of the cell marking the endpoint of the solution. For edit distance, this is deﬁned by the length of the two input strings. However, other applications we will soon encounter do not have ﬁxed goal locations. goal_cell(char *s, char *t, int *i, int *j) { *i = strlen(s) - 1; *j = strlen(t) - 1; } • Traceback Actions – The functions match out, insert out, and delete out perform the appropriate actions for each edit operation during traceback. For edit distance, this might mean printing out the name of the operation or character involved, as determined by the needs of the application. insert_out(char *t, int j) { printf("I"); } delete_out(char *s, int i) { printf("D"); } match_out(char *s, char *t, int i, int j) { if (s[i]==t[j]) printf("M"); else printf("S"); } All of these functions are quite simple for edit distance computation. However, we must confess it is diﬃcult to get the boundary conditions and index manipulations correctly. Although dynamic programming algorithms are easy to design once you understand the technique, getting the details right requires carefully thinking and thorough testing. This may seem to be a lot of infrastructure to develop for such a simple algorithm. However, several important problems can now be solved as special cases of edit distance using only minor changes to some of these stub functions: • Substring Matching – Suppose that we want to ﬁnd where a short pattern P best occurs within a long text T —say, searching for “Skiena” in all its misspellings (Skienna, Skena, Skina, . . . ) within a long ﬁle. Plugging this 287 288 8. DYNAMIC PROGRAMMING search into our original edit distance function will achieve little sensitivity, since the vast majority of any edit cost will consist of deleting that which is not “Skiena” from the body of the text. Indeed, matching any scattered · · · S · · · k · · · i · · · e · · · n · · · a and deleting the rest yields an optimal solution. We want an edit distance search where the cost of starting the match is independent of the position in the text, so that a match in the middle is not prejudiced against. Now the goal state is not necessarily at the end of both strings, but the cheapest place to match the entire pattern somewhere in the text. Modifying these two functions gives us the correct solution: row_init(int i) { m[0][i].cost = 0; m[0][i].parent = -1; } /* note change */ /* note change */ goal_cell(char *s, char *t, int *i, int *j) { int k; /* counter */ *i = strlen(s) - 1; *j = 0; for (k=1; k 0 and, M [n, 1] = n si i=1 By deﬁnition, this recurrence must return the size of the optimal partition. How long does it take to compute this when we store the partial results? A total of k · n cells exist in the table. How much time does it take to compute the result 295 296 8. DYNAMIC PROGRAMMING M [n , k ]? Calculating this quantity involves ﬁnding the minimum of n quantities, each of which is the maximum of the table lookup and a sum of at most n elements. If ﬁlling each of kn boxes takes at most n2 time per box, the total recurrence can be computed in O(kn3 ) time. The evaluation order computes the smaller values before the bigger values, so that each evaluation has what it needs waiting for it. Full details are provided in the code below: partition(int s[], int n, int k) { int m[MAXN+1][MAXK+1]; int d[MAXN+1][MAXK+1]; int p[MAXN+1]; int cost; int i,j,x; /* /* /* /* /* DP table for values */ DP table for dividers */ prefix sums array */ test split cost */ counters */ p[0] = 0; for (i=1; i<=n; i++) p[i]=p[i-1]+s[i]; /* construct prefix sums */ for (i=1; i<=n; i++) m[i][1] = p[i]; for (j=1; j<=k; j++) m[1][j] = s[1]; /* initialize boundaries */ for (i=2; i<=n; i++) /* evaluate main recurrence */ for (j=2; j<=k; j++) { m[i][j] = MAXINT; for (x=1; x<=(i-1); x++) { cost = max(m[x][j-1], p[i]-p[x]); if (m[i][j] > cost) { m[i][j] = cost; d[i][j] = x; } } } reconstruct_partition(s,d,n,k); /* print book partition */ } This implementation above, in fact, runs faster than advertised. Our original analysis assumed that it took O(n2 ) time to update each cell of the matrix. This is because we selected the best of up to n possible points to place the divider, each of which requires the sum of up to n possible terms. In fact, it is easy to the avoid i need to compute these sums by storing the set of n preﬁx sums p[i] = k=1 sk , 8.5 M n 1 1 1 1 1 1 1 1 1 1 1 2 3 4 5 6 7 8 9 k 2 1 1 2 2 3 3 4 4 5 3 1 1 1 2 2 2 3 3 3 D n 1 1 1 1 1 1 1 1 1 1 – – – – – – – – – k 2 – 1 1 2 2 3 3 4 4 3 – 1 2 2 3 4 4 5 6 M n 1 2 3 4 5 6 7 8 9 297 THE PARTITION PROBLEM 1 1 3 6 10 15 21 28 36 45 k 2 1 2 3 6 9 11 15 21 24 3 1 2 3 4 6 9 11 15 17 D n 1 2 3 4 5 6 7 8 9 1 – – – – – – – – – k 2 – 1 2 3 3 4 5 5 6 3 – 1 2 3 4 5 6 6 7 Figure 8.8: Dynamic programming matrices M and D for two input instances. Partitioning {1, 1, 1, 1, 1, 1, 1, 1, 1} into {{1, 1, 1}, {1, 1, 1}, {1, 1, 1}} (l) . Partitioning {1, 2, 3, 4, 5, 6, 7, 8, 9} into {{1, 2, 3, 4, 5}, {6, 7}, {8, 9}} (r). j since k=i sk = p[j] − p[k]. This enables us to evaluate the recurrence in linear time per cell, yielding an O(kn2 ) algorithm. By studying the recurrence relation and the dynamic programming matrices of Figure 8.8, you should be able to convince yourself that the ﬁnal value of M (n, k) will be the cost of the largest range in the optimal partition. For most applications, however, what we need is the actual partition that does the job. Without it, all we are left with is a coupon with a great price on an out-of-stock item. The second matrix, D, is used to reconstruct the optimal partition. Whenever we update the value of M [i, j], we record which divider position was required to achieve that value. To reconstruct the path used to get to the optimal solution, we work backward from D[n, k] and add a divider at each speciﬁed position. This backwards walking is best achieved by a recursive subroutine: reconstruct_partition(int s[],int d[MAXN+1][MAXK+1], int n, int k) { if (k==1) print_books(s,1,n); else { reconstruct_partition(s,d,d[n][k],k-1); print_books(s,d[n][k]+1,n); } } print_books(int s[], int start, int end) { int i; /* counter */ for (i=start; i<=end; i++) printf(" %d ",s[i]); printf("\n"); } 298 8. DYNAMIC PROGRAMMING sentence ::= noun−phrase verb−phrase noun−phrase ::= article noun verb−phrase ::= verb noun−phrase article ::= the, a noun ::= cat, milk verb ::= drank sentence noun−phrase article noun verb−phrase verb noun−phrase article the cat drank the noun milk Figure 8.9: A context-free grammar (l) with an associated parse tree (r) 8.6 Parsing Context-Free Grammars Compilers identify whether the given program is legal in the programming language, and reward you with syntax errors if not. This requires a precise description of the language syntax typically given by a context-free grammar as shown in Figure 8.9(l). Each rule or production of the grammar deﬁnes an interpretation for the named symbol on the left side of the rule as a sequence of symbols on the right side of the rule. The right side can be a combination of nonterminals (themselves deﬁned by rules) or terminal symbols deﬁned simply as strings, such as “the”, “a”, “cat”, “milk”, and “drank.” Parsing a given text string S according to a given context-free grammar G is the algorithmic problem of constructing a parse tree of rule substitutions deﬁning S as a single nonterminal symbol of G. Figure 8.9(r) gives the parse tree of a simple sentence using our sample grammar. Parsing seemed like a horribly complicated subject when I took a compilers course as a graduate student. But, a friend easily explained it to me over lunch a few years ago. The diﬀerence is that I now understand dynamic programming much better than when I was a student. We assume that the text string S has length n while the grammar G itself is of constant size. This is fair, since the grammar deﬁning a particular programming language (say C or Java) is of ﬁxed length regardless of the size of the program we are trying to compile. Further, we assume that the deﬁnitions of each rule are in Chomsky normal form. This means that the right sides of every nontrivial rule consists of (a) exactly two nonterminals, i.e. X → Y Z, or (b) exactly one terminal symbol, X → α. Any context-free grammar can be easily and mechanically transformed into Chomsky normal form by repeatedly shortening long right-hand sides at the cost of adding extra nonterminals and productions. Thus, there is no loss of generality with this assumption. 8.6 PARSING CONTEXT-FREE GRAMMARS So how can we eﬃciently parse a string S using a context-free grammar where each interesting rule consists of two nonterminals? The key observation is that the rule applied at the root of the parse tree (say X → Y Z) splits S at some position i such that the left part of the string (S[1, i]) must be generated by nonterminal Y , and the right part (S[i + 1, n]) generated by Z. This suggests a dynamic programming algorithm, where we keep track of all of the nonterminals generated by each substring of S. Deﬁne M [i, j, X] to be a boolean function that is true iﬀ substring S[i, j] is generated by nonterminal X. This is true if there exists a production X → Y Z and breaking point k between i and j such that the left part generates Y and the right part Z. In other words, M [i, j, X] = j (X→Y Z)∈G i=k M [i, k, Y ] · M [k + 1, j, Z] where ∨ denotes the logical or over all productions and split positions, and · denotes the logical and of two boolean values. The one-character terminal symbols deﬁne the boundary conditions of the recurrence. In particular, M [i, i, X] is true iﬀ there exists a production X → α such that S[i] = α. What is the complexity of this algorithm? The size of our state-space is O(n2 ), as there are n(n + 1)/2 substrings deﬁned by (i, j) pairs. Multiplying this by the number of nonterminals (say v) has no impact on the big-Oh, because the grammar was deﬁned to be of constant size. Evaluating the value M [i, j, X] requires testing all intermediate values k, so it takes O(n) in the worst case to evaluate each of the O(n2 ) cells. This yields an O(n3 ) or cubic-time algorithm for parsing. Stop and Think: Parsimonious Parserization Problem: Programs often contain trivial syntax errors that prevent them from compiling. Given a context-free grammar G and input string S, ﬁnd the smallest number of character substitutions you must make to S so that the resulting string is accepted by G. Solution: This problem seemed extremely diﬃcult when I ﬁrst encountered it. But on reﬂection, it seemed like a very general version of edit distance, which is addressed naturally by dynamic programming. Parsing initially sounded hard, too, but fell to the same technique. Indeed, we can solve the combined problem by generalizing the recurrence relation we used for simple parsing. Deﬁne M [i, j, X] to be an integer function that reports the minimum number of changes to substring S[i, j] so it can be generated by nonterminal X. This symbol will be generated by some production x → yz. Some of the changes to s may be 299 300 8. DYNAMIC PROGRAMMING Figure 8.10: Two diﬀerent triangulations of a given convex seven-gon to the left of the breaking point and some to the right, but all we care about is minimizing the sum. In other words, M [i, j, X] = j min (X→Y Z)∈G min M [i, k, Y ] + M [k + 1, j, Z] i=k The boundary conditions also change mildly. If there exists a production X → α, the cost of matching at position i depends on the contents of S[i], where S[i] = α, M [i, i, X] = 0. Otherwise, it is one substitution away, so M [i, i, X] = 1 if S[i] = α. If the grammar does not have a production of the form X → α, there is no way to substitute a single character string into something generating X, so M [i, i, X] = ∞ for all i. 8.6.1 Minimum Weight Triangulation The same basic recurrence relation encountered in the parsing algorithm above can also be used to solve an interesting computational geometry problem. A triangulation of a polygon P = {v1 , . . . , vn , v1 } is a set of nonintersecting diagonals that partitions the polygon into triangles. We say that the weight of a triangulation is the sum of the lengths of its diagonals. As shown in Figure 8.10, any given polygon may have many diﬀerent triangulations. We seek to ﬁnd its minimum weight triangulation for a given polygon p. Triangulation is a fundamental component of most geometric algorithms, as discussed in Section 17.3 (page 572). To apply dynamic programming, we need a way to carve up the polygon into smaller pieces. Observe that every edge of the input polygon must be involved in exactly one triangle. Turning this edge into a triangle means identifying the third vertex, as shown in Figure 8.11. Once we ﬁnd the correct connecting vertex, the polygon will be partitioned into two smaller pieces, both of which need to be triangulated optimally. Let T [i, j] be the cost of triangulating from vertex vi to vertex vj , ignoring the length of the chord dij from vi to vj . The latter clause avoids double counting these internal chords in the following recurrence: j−1 T [i, j] = min (T [i, k] + T [k, j] + dik + dkj ) k=i+1 The basis condition applies when i and j are immediate neighbors, as T [i, i+1] = 0. 8.7 LIMITATIONS OF DYNAMIC PROGRAMMING: TSP k i j Figure 8.11: Selecting the vertex k to pair with an edge (i, j) of the polygon Since the number of vertices in each subrange of the right side of the recurrence is smaller than that on the left side, evaluation can proceed in terms of the gap size from i to j: Minimum-Weight-Triangulation(P ) for i = 1 to n − 1 do T [i, i + 1] = 0 for gap = 2 to n − 1 for i = 1 to n − gap do j = i + gap T [i, j] = minj−1 k=i+1 (T [i, k] + T [k, j] + dPi ,Pk + dPk ,Pj ) return T [1, n] There are n2 values of T , each of which takes O(j − i) time if we evaluate the sections in order of increasing size. Since j − i = O(n), complete evaluation takes O(n3 ) time and O(n2 ) space. What if there are points in the interior of the polygon? Then dynamic programming does not apply in the same way, because triangulation edges do not necessarily cut the boundary into two distinct pieces as before. Instead of only n2 possible subregions, the number of subregions now grows exponentially. In fact, the more general version of this problem is known to be NP-complete. Take-Home Lesson: For any optimization problem on left-to-right objects, such as characters in a string, elements of a permutation, points around a polygon, or leaves in a search tree, dynamic programming likely leads to an eﬃcient algorithm to ﬁnd the optimal solution. 8.7 Limitations of Dynamic Programming: TSP Dynamic programming doesn’t always work. It is important to see why it can fail, to help avoid traps leading to incorrect or ineﬃcient algorithms. Our algorithmic poster child will once again be the traveling salesman, where we seek the shortest tour visiting all the cities in a graph. We will limit attention here to an interesting special case: 301 302 8. DYNAMIC PROGRAMMING Problem: Longest Simple Path Input: A weighted graph G, with speciﬁed start and end vertices s and t. Output: What is the most expensive path from s to t that does not visit any vertex more than once? This problem diﬀers from TSP in two quite unimportant ways. First, it asks for a path instead of a closed tour. This diﬀerence isn’t substantial: we get a closed tour by simply including the edge (t, s). Second, it asks for the most expensive path instead of the least expensive tour. Again this diﬀerence isn’t very signiﬁcant: it encourages us to visit as many vertices as possible (ideally all), just as in TSP. The big word in the problem statement is simple, meaning we are not allowed to visit any vertex more than once. For unweighted graphs (where each edge has cost 1), the longest possible simple path from s to t is n − 1. Finding such Hamiltonian paths (if they exist) is an important graph problem, discussed in Section 16.5 (page 538). 8.7.1 When are Dynamic Programming Algorithms Correct? Dynamic programming algorithms are only as correct as the recurrence relations they are based on. Suppose we deﬁne LP [i, j] as a function denoting the length of the longest simple path from i to j. Note that the longest simple path from i to j had to visit some vertex x right before reaching j. Thus, the last edge visited must be of the form (x, j). This suggests the following recurrence relation to compute the length of the longest path, where c(x, j) is the cost/weight of edge (x, j): LP [i, j] = max LP [i, x] + c(x, j) (x,j)∈E The idea seems reasonable, but can you see the problem? I see at least two of them. First, this recurrence does nothing to enforce simplicity. How do we know that vertex j has not appeared previously on the longest simple path from i to x? If it did, adding the edge (x, j) will create a cycle. To prevent such a thing, we must deﬁne a diﬀerent recursive function that explicitly remembers where we have been. Perhaps we could deﬁne LP [i, j, k] to be the function denoting the length of the longest path from i to j avoiding vertex k? This would be a step in the right direction, but still won’t lead to a viable recurrence. A second problem concerns evaluation order. What can you evaluate ﬁrst? Because there is no left-to-right or smaller-to-bigger ordering of the vertices on the graph, it is not clear what the smaller subprograms are. Without such an ordering, we get are stuck in an inﬁnite loop as soon as we try to do anything. Dynamic programming can be applied to any problem that observes the principle of optimality. Roughly stated, this means that partial solutions can be optimally extended with regard to the state after the partial solution, instead of the speciﬁcs of the partial solution itself. For example, in deciding whether to extend an approximate string matching by a substitution, insertion, or deletion, we did not need to 8.7 LIMITATIONS OF DYNAMIC PROGRAMMING: TSP know which sequence of operations had been performed to date. In fact, there may be several diﬀerent edit sequences that achieve a cost of C on the ﬁrst p characters of pattern P and t characters of string T . Future decisions are made based on the consequences of previous decisions, not the actual decisions themselves. Problems do not satisfy the principle of optimality when the speciﬁcs of the operations matter, as opposed to just the cost of the operations. Such would be the case with a form of edit distance where we are not allowed to use combinations of operations in certain particular orders. Properly formulated, however, many combinatorial problems respect the principle of optimality. 8.7.2 When are Dynamic Programming Algorithms Eﬃcient? The running time of any dynamic programming algorithm is a function of two things: (1) number of partial solutions we must keep track of, and (2) how long it take to evaluate each partial solution. The ﬁrst issue—namely the size of the state space—is usually the more pressing concern. In all of the examples we have seen, the partial solutions are completely described by specifying the stopping places in the input. This is because the combinatorial objects being worked on (strings, numerical sequences, and polygons) have an implicit order deﬁned upon their elements. This order cannot be scrambled without completely changing the problem. Once the order is ﬁxed, there are relatively few possible stopping places or states, so we get eﬃcient algorithms. When the objects are not ﬁrmly ordered, however, we get an exponential number of possible partial solutions. Suppose the state of our partial solution is entire path P taken from the start to end vertex. Thus LP [i, j, P ] denotes the longest simple path from i to j, where P is the exact sequence of intermediate vertices between i and j on this path. The following recurrence relation works to compute this, where P + x denotes appending x to the end of P : LP [i, j, P + x] = max (x,j)∈E,x,j ∈P LP [i, x, P ] + c(x, j) This formulation is correct, but how eﬃcient is it? The path P consists of an ordered sequence of up to n − 3 vertices. There can be up to (n − 3)! such paths! Indeed, this algorithm is really using combinatorial search (a la backtracking) to construct all the possible intermediate paths. In fact, the max is somewhat misleading, as there can only be one value of x and one value of P to construct the state LP [i, j, P + x]. We can do something better with this idea, however. Let LP [i, j, S] denote the longest simple path from i to j, where the intermediate vertices on this path are exactly those in the subset S. Thus, if S = {a, b, c}, there are exactly six paths consistent with S: iabcj, iacbj, ibacj, ibcaj, icabj, and icbaj. This state space is at most 2n , and thus smaller than enumerating the paths. Further, this function can be evaluated using the following recurrence relation: LP [i, j, S ∪ {x}] = max (x,j)∈E,x,j ∈S LP [i, x, S] + c(x, j) 303 304 8. DYNAMIC PROGRAMMING where S ∪ {x} denotes unioning S with x. The longest simple path from i to j can then be found by maximizing over all possible intermediate vertex subsets: LP [i, j] = max LP [i, j, S] S There are only 2n subsets of n vertices, so this is a big improvement over enumerating all n! tours. Indeed, this method could certainly be used to solve TSPs for up to thirty vertices or so, where n = 20 would be impossible using the O(n!) algorithm. Still, dynamic programming is most eﬀective on well-ordered objects. Take-Home Lesson: Without an inherent left-to-right ordering on the objects, dynamic programming is usually doomed to require exponential space and time. 8.8 War Story: What’s Past is Prolog “But our heuristic works very, very well in practice.” My colleague was simultaneously boasting and crying for help. Uniﬁcation is the basic computational mechanism in logic programming languages like Prolog. A Prolog program consists of a set of rules, where each rule has a head and an associated action whenever the rule head matches or uniﬁes with the current computation. An execution of a Prolog program starts by specifying a goal, say p(a, X, Y ), where a is a constant and X and Y are variables. The system then systematically matches the head of the goal with the head of each of the rules that can be uniﬁed with the goal. Uniﬁcation means binding the variables with the constants, if it is possible to match them. For the nonsense program below, p(X, Y, a) uniﬁes with either of the ﬁrst two rules, since X and Y can be bound to match the extra characters. The goal p(X, X, a) would only match the ﬁrst rule, since the variable bound to the ﬁrst and second positions must be the same. p(a, a, a) := h(a); p(b, a, a) := h(a) ∗ h(b); p(c, b, b) := h(b) + h(c); p(d, b, b) := h(d) + h(b); “In order to speed up uniﬁcation, we want to preprocess the set of rule heads so that we can quickly determine which rules match a given goal. We must organize the rules in a trie data structure for fast uniﬁcation.” Tries are extremely useful data structures in working with strings, as discussed in Section 12.3 (page 377). Every leaf of the trie represents one string. Each node on the path from root to leaf is labeled with exactly one character of the string, with the ith node of the path corresponding to the ith character of the string. 8.8 WAR STORY: WHAT’S PAST IS PROLOG 2 1 a 2 b 2 a 3 a 1 2 a b 3 a S 2 3 b 1 3 3 3 a b b S b a d 2 3 S c a b S 4 S 1 1 b c S 2 S 3 d S 4 Figure 8.12: Two diﬀerent tries for the same set of rule heads. “I agree. A trie is a natural way to represent your rule heads. Building a trie on a set of strings of characters is straightforward: just insert the strings starting from the root. So what is your problem?” I asked. “The eﬃciency of our uniﬁcation algorithm depends very much on minimizing the number of edges in the trie. Since we know all the rules in advance, we have the freedom to reorder the character positions in the rules. Instead of the root node always representing the ﬁrst argument in the rule, we can choose to have it represent the third argument. We would like to use this freedom to build a minimum-size trie for a set of rules.” He showed me the example in Figure 8.12. A trie constructed according to the original string position order (1, 2, 3) uses a total of 12 edges. However, by permuting the character order to (2, 3, 1), we can obtain a trie with only 8 edges. “Interesting. . . ” I started to reply before he cut me oﬀ again. “There’s one other constraint. We must keep the leaves of the trie ordered, so that the leaves of the underlying tree go left-to-right in the same order as the rules appear on the page.” “But why must you keep the leaves of the trie in the given order?” I asked. “The order of rules in Prolog programs is very, very important. If you change the order of the rules, the program returns diﬀerent results.” Then came my mission. “We have a greedy heuristic for building good, but not optimal, tries based on picking as the root the character position that minimizes the degree of the root. In other words, it picks the character position that has the smallest number of distinct characters in it. This heuristic works very, very well in practice. But we need you to prove that ﬁnding the best trie is NP-complete so our paper is, well, complete.” 305 306 8. DYNAMIC PROGRAMMING I agreed to try to prove the hardness of the problem, and chased him from my oﬃce. The problem did seem to involve some nontrivial combinatorial optimization to build the minimal tree, but I couldn’t see how to factor the left-to-right order of the rules into a hardness proof. In fact, I couldn’t think of any NP-complete problem that had such a left-right ordering constraint. After all, if a given set of rules contained a character position in common to all the rules, this character position must be probed ﬁrst in any minimum-size tree. Since the rules were ordered, each node in the subtree must represent the root of a run of consecutive rules. Thus there were only n2 possible nodes to choose from for this tree. . . . Bingo! That settled it. The next day I went back to the professor and told him. “I can’t prove that your problem is NP-complete. But how would you feel about an eﬃcient dynamic programming algorithm to ﬁnd the best trie!” It was a pleasure watching his frown change to a smile as the realization took hold. An eﬃcient algorithm to compute what he needed was inﬁnitely better than a proof saying you couldn’t do it! My recurrence looked something like this. Suppose that we are given n ordered rule heads s1 , . . . , sn , each with m arguments. Probing at the pth position, 1 ≤ p ≤ m, partitioned the rule heads into runs R1 , . . . , Rr , where each rule in a given run Rx = si , . . . , sj had the same character value of si [p]. The rules in each run must be consecutive, so there are only n2 possible runs to worry about. The cost of probing at position p is the cost of ﬁnishing the trees formed by each created run, plus one edge per tree to link it to probe p: m C[i, j] = min p=1 r (C[ik , jk ] + 1) k=1 A graduate student immediately set to work implementing this algorithm to compare with their heuristic. On many inputs, the optimal and greedy algorithms constructed the exact same trie. However, for some examples, dynamic programming gave a 20% performance improvement over greedy—i.e. , 20% better than very, very well in practice. The run time spent in doing the dynamic programming was a bit larger than with greedy, but in compiler optimization you are always happy to trade oﬀ a little extra compilation time for better execution time in the performance of your program. Is a 20% improvement worth this eﬀort? That depends upon the situation. How useful would you ﬁnd a 20% increase in your salary? The fact that the rules had to remain ordered was the crucial property that we exploited in the dynamic programming solution. Indeed, without it I was able to prove that the problem was NP-complete, something we put in the paper to make it complete. 8.9 WAR STORY: TEXT COMPRESSION FOR BAR CODES 307 Figure 8.13: A two-dimensional bar-code label of the Gettysburg Address using PDF-417. Take-Home Lesson: The global optimum (found, for example, using dynamic programming) is often noticeably better than the solution found by typical heuristics. How important this improvement is depends on your application, but it can never hurt. 8.9 War Story: Text Compression for Bar Codes Ynjiun waved his laser wand over the torn and crumpled fragments of a bar code label. The system hesitated for a few seconds, then responded with a pleasant blip sound. He smiled at me in triumph. “Virtually indestructible.” I was visiting the research laboratories of Symbol Technologies, the world’s leading manufacturer of bar code scanning equipment. Next time you are in the checkout line at a grocery store, check to see what type of scanning equipment they are using. Likely it will say Symbol on the housing. Although we take bar codes for granted, there is a surprising amount of technology behind them. Bar codes exist because conventional optical character recognition (OCR) systems are not suﬃciently reliable for inventory operations. The bar code symbology familiar to us on each box of cereal or pack of gum encodes a ten-digit number with enough error correction that it is virtually impossible to scan the wrong number, even if the can is upside-down or dented. Occasionally, the cashier won’t be able to get a label to scan at all, but once you hear that blip you know it was read correctly. The ten-digit capacity of conventional bar code labels provides room enough only to store a single ID number in a label. Thus any application of supermarket bar codes must have a database mapping 11141-47011 to a particular size and brand of soy sauce. The holy grail of the bar code world has long been the development of higher-capacity bar code symbologies that can store entire documents, yet still be read reliably. “PDF-417 is our new, two-dimensional bar code symbology,” Ynjiun explained. A sample label is shown in Figure 8.13. “How much data can you ﬁt in a typical one-inch square label?” I asked him. 308 8. DYNAMIC PROGRAMMING Alpha A−Z l l s Lower Case a−z l s Mixed l s 0−9, #$%= l,s Punctuation ;<>@[\$/"*( Figure 8.14: Mode switching in PDF-417 “It depends upon the level of error correction we use, but about 1,000 bytes. That’s enough for a small text ﬁle or image,” he said. “Interesting. You will probably want to use some data compression technique to maximize the amount of text you can store in a label.” See Section 18.5 (page 637) for a discussion of standard data compression algorithms. “We do incorporate a data compaction method,” he explained. “We think we understand the diﬀerent types of ﬁles our customers will want to make labels for. Some ﬁles will be all in uppercase letters, while others will use mixed-case letters and numbers. We provide four diﬀerent text modes in our code, each with a diﬀerent subset of alphanumeric characters available. We can describe each character using only ﬁve bits as long as we stay within a mode. To switch modes, we issue a mode switch command ﬁrst (taking an extra ﬁve bits) and then the new character code.” “I see. So you designed the mode character sets to minimize the number of mode switch operations on typical text ﬁles.” The modes are illustrated in Figure 8.14. “Right. We put all the digits in one mode and all the punctuation characters in another. We also included both mode shift and mode latch commands. In a mode shift, we switch into a new mode just for the next character, say to produce a punctuation mark. This way, we don’t pay a cost for returning back to text mode after a period. Of course, we can also latch permanently into a diﬀerent mode if we will be using a run of several characters from there.” “Wow!” I said. “With all of this mode switching going on, there must be many diﬀerent ways to encode any given text as a label. How do you ﬁnd the smallest of such encoding.” “We use a greedy algorithm. We look a few characters ahead and then decide which mode we would be best oﬀ in. It works fairly well.” 8.9 WAR STORY: TEXT COMPRESSION FOR BAR CODES I pressed him on this. “How do you know it works fairly well? There might be signiﬁcantly better encodings that you are simply not ﬁnding.” “I guess I don’t know. But it’s probably NP-complete to ﬁnd the optimal coding.” Ynjiun’s voice trailed oﬀ. “Isn’t it?” I started to think. Every encoding started in a given mode and consisted of a sequence of intermixed character codes and mode shift/latch operations. At any given position in the text, we could output the next character code (if it was available in our current mode) or decide to shift. As we moved from left to right through the text, our current state would be completely reﬂected by our current character position and current mode state. For a given position/mode pair, we would have been interested in the cheapest way of getting there, over all possible encodings getting to this point. . . . My eyes lit up so bright they cast shadows on the walls. “The optimal encoding for any given text in PDF-417 can be found using dynamic programming. For each possible mode 1 ≤ m ≤ 4, and each character position 1 ≤ i ≤ n, we will maintain the cheapest encoding found of the ﬁrst i characters ending in mode m. Our next move from each mode/position is either match, shift, or latch, so there are only a few possible operations to consider.” Basically, M [i, j] = min (M [i − 1, m] + c(Si , m, j)) 1≤m≤4 where c(Si , m, j) is the cost of encoding character Si and switching from mode m to mode j. The cheapest possible encoding results from tracing back from M [n, m], where m is the value of i that minimizes min1≤i≤4 M [n, i]. Each of the 4n cells can be ﬁlled in constant time, so it takes time linear in the length of the string to ﬁnd the optimal encoding. Ynjiun was skeptical, but he encouraged us to implement an optimal encoder. A few complications arose due to weirdnesses of PDF-417 mode switching, but my student Yaw-Ling Lin rose to the challenge. Symbol compared our encoder to theirs on 13,000 labels and concluded that dynamic programming lead to an 8% tighter encoding on average. This was signiﬁcant, because no one wants to waste 8% of their potential storage capacity, particularly in an environment where the capacity is only a few hundred bytes. For certain applications, this 8% margin permitted one bar code label to suﬃce where previously two had been required. Of course, an 8% average improvement meant that it did much better than that on certain labels. While our encoder took slightly longer to run than the greedy encoder, this was not signiﬁcant, since the bottleneck would be the time needed to print the label anyway. Our observed impact of replacing a heuristic solution with the global optimum is probably typical of most applications. Unless you really botch your heuristic, you are probably going to get a decent solution. Replacing it with an optimal result, however, usually gives a small but nontrivial improvement, which can have pleasing consequences for your application. 309 310 8. DYNAMIC PROGRAMMING Chapter Notes Bellman [Bel58] is credited with developing the technique of dynamic programming. The edit distance algorithm is originally due to Wagner and Fischer [WF74]. A faster algorithm for the book partition problem appears in [KMS97]. The computational complexity of ﬁnding the minimum weight triangulation of disconnected point sets (as opposed to polygons) was a longstanding open problem that ﬁnally fell to Mulzer and Rote [MR06]. Techniques such as dynamic programming and backtracking searches can be used to generate worst-case eﬃcient (although still non-polynomial) algorithms for many NP-complete problems. See Woeginger [Woe03] for a nice survey of such techniques. The morphing system that was the subject of the war story in Section 8.4 (page 291) is described in [HWK94]. See our paper [DRR+ 95] for more on the Prolog trie minimization problem, subject of the war story of Section 8.8 (page 304). Twodimensional bar codes, subject of the war story in Section 8.9 (page 307), were developed largely through the eﬀorts of Theo Pavlidis and Ynjiun Wang at Stony Brook [PSW92]. The dynamic programming algorithm presented for parsing is known as the CKY algorithm after its three independent inventors (Cocke, Kasami, and Younger) [You67]. The generalization of parsing to edit distance is due to Aho and Peterson [AP72]. 8.10 Exercises Edit Distance 8-1. [3] Typists often make transposition errors exchanging neighboring characters, such as typing “setve” when you mean “steve.” This requires two substitutions to ﬁx under the conventional deﬁnition of edit distance. Incorporate a swap operation into our edit distance function, so that such neighboring transposition errors can be ﬁxed at the cost of one operation. 8-2. [4] Suppose you are given three strings of characters: X, Y , and Z, where |X| = n, |Y | = m, and |Z| = n + m. Z is said to be a shuﬄe of X and Y iﬀ Z can be formed by interleaving the characters from X and Y in a way that maintains the left-to-right ordering of the characters from each string. (a) Show that cchocohilaptes is a shuﬄe of chocolate and chips, but chocochilatspe is not. (b) Give an eﬃcient dynamic-programming algorithm that determines whether Z is a shuﬄe of X and Y . Hint: the values of the dynamic programming matrix you construct should be Boolean, not numeric. 8-3. [4] The longest common substring (not subsequence) of two strings X and Y is the longest string that appears as a run of consecutive letters in both strings. For example, the longest common substring of photograph and tomography is ograph. 8.10 EXERCISES (a) Let n = |X| and m = |Y |. Give a Θ(nm) dynamic programming algorithm for longest common substring based on the longest common subsequence/edit distance algorithm. (b) Give a simpler Θ(nm) algorithm that does not rely on dynamic programming. 8-4. [6] The longest common subsequence (LCS) of two sequences T and P is the longest sequence L such that L is a subsequence of both T and P . The shortest common supersequence (SCS) of T and P is the smallest sequence L such that both T and P are a subsequence of L. (a) Give eﬃcient algorithms to ﬁnd the LCS and SCS of two given sequences. (b) Let d(T, P ) be the minimum edit distance between T and P when no substitutions are allowed (i.e., the only changes are character insertion and deletion). Prove that d(T, P ) = |SCS(T, P )| − |LCS(T, P )| where |SCS(T, P )| (|LCS(T, P )|) is the size of the shortest SCS (longest LCS) of T and P . Greedy Algorithms 8-5. [4] Let P1 , P2 , . . . , Pn be n programs to be stored on a disk with capacity D megabytes. Program Pi requires si megabytes of storage. We cannot store them all n because D < i=1 si (a) Does a greedy algorithm that selects programs in order of nondecreasing si maximize the number of programs held on the disk? Prove or give a counterexample. (b) Does a greedy algorithm that selects programs in order of nonincreasing order si use as much of the capacity of the disk as possible? Prove or give a counterexample. 8-6. [5] Coins in the United States are minted with denominations of 1, 5, 10, 25, and 50 cents. Now consider a country whose coins are minted with denominations of {d1 , . . . , dk } units. We seek an algorithm to make change of n units using the minimum number of coins for this country. (a) The greedy algorithm repeatedly selects the biggest coin no bigger than the amount to be changed and repeats until it is zero. Show that the greedy algorithm does not always use the minimum number of coins in a country whose denominations are {1, 6, 10}. (b) Give an eﬃcient algorithm that correctly determines the minimum number of coins needed to make change of n units using denominations {d1 , . . . , dk }. Analyze its running time. 8-7. [5] In the United States, coins are minted with denominations of 1, 5, 10, 25, and 50 cents. Now consider a country whose coins are minted with denominations of {d1 , . . . , dk } units. We want to count how many distinct ways C(n) there are to make change of n units. For example, in a country whose denominations are {1, 6, 10}, C(5) = 1, C(6) to C(9) = 2, C(10) = 3, and C(12) = 4. (a) How many ways are there to make change of 20 units from {1, 6, 10}? 311 312 8. DYNAMIC PROGRAMMING (b) Give an eﬃcient algorithm to compute C(n), and analyze its complexity. (Hint: think in terms of computing C(n, d), the number of ways to make change of n units with highest denomination d. Be careful to avoid overcounting.) 8-8. [6] In the single-processor scheduling problem, we are given a set of n jobs J. Each job i has a processing time ti , and a deadline di . A feasible schedule is a permutation of the jobs such that when the jobs are performed in that order, every job is ﬁnished before its deadline. The greedy algorithm for single-processor scheduling selects the job with the earliest deadline ﬁrst. Show that if a feasible schedule exists, then the schedule produced by this greedy algorithm is feasible. Number Problems 8-9. [6] The knapsack problem is as follows: given a set of integers S = {s1 , s2 , . . . , sn }, and a given target number T , ﬁnd a subset of S that adds up exactly to T . For example, within S = {1, 2, 5, 9, 10} there is a subset that adds up to T = 22 but not T = 23. Give a correct programming algorithm for knapsack that runs in O(nT ) time. 8-10. [6] The integer partition takes a set of positive integers S = s1 , . . . , sn and asks if there is a subset I ∈ S such that i∈I si = si i∈I / Let s = M . Give an O(nM ) dynamic programming algorithm to solve the i∈S i integer partition problem. 8-11. [5] Assume that there are n numbers (some possibly negative) on a circle, and we wish to ﬁnd the maximum contiguous sum along an arc of the circle. Give an eﬃcient algorithm for solving this problem. 8-12. [5] A certain string processing language allows the programmer to break a string into two pieces. It costs n units of time to break a string of n characters into two pieces, since this involves copying the old string. A programmer wants to break a string into many pieces, and the order in which the breaks are made can aﬀect the total amount of time used. For example, suppose we wish to break a 20-character string after characters 3, 8, and 10. If the breaks are made in left-right order, then the ﬁrst break costs 20 units of time, the second break costs 17 units of time, and the third break costs 12 units of time, for a total of 49 steps. If the breaks are made in right-left order, the ﬁrst break costs 20 units of time, the second break costs 10 units of time, and the third break costs 8 units of time, for a total of only 38 steps. Give a dynamic programming algorithm that takes a list of character positions after which to break and determines the cheapest break cost in O(n3 ) time. 8-13. [5] Consider the following data compression technique. We have a table of m text strings, each at most k in length. We want to encode a data string D of length n using as few text strings as possible. For example, if our table contains (a,ba,abab,b) and the data string is bababbaababa, the best way to encode it is (b,abab,ba,abab,a)— a total of ﬁve code words. Give an O(nmk) algorithm to ﬁnd the length of the best 8.10 EXERCISES encoding. You may assume that every string has at least one encoding in terms of the table. 8-14. [5] The traditional world chess championship is a match of 24 games. The current champion retains the title in case the match is a tie. Each game ends in a win, loss, or draw (tie) where wins count as 1, losses as 0, and draws as 1/2. The players take turns playing white and black. White has an advantage, because he moves ﬁrst. The champion plays white in the ﬁrst game. He has probabilities ww , wd , and wl of winning, drawing, and losing playing white, and has probabilities bw , bd , and bl of winning, drawing, and losing playing black. (a) Write a recurrence for the probability that the champion retains the title. Assume that there are g games left to play in the match and that the champion needs to win i games (which may end in a 1/2). (b) Based on your recurrence, give a dynamic programming to calculate the champion’s probability of retaining the title. (c) Analyze its running time for an n game match. 8-15. [8] Eggs break when dropped from great enough height. Speciﬁcally, there must be a ﬂoor f in any suﬃciently tall building such that an egg dropped from the f th ﬂoor breaks, but one dropped from the (f − 1)st ﬂoor will not. If the egg always breaks, then f = 1. If the egg never breaks, then f = n + 1. You seek to ﬁnd the critical ﬂoor f using an n-story building. The only operation you can perform is to drop an egg oﬀ some ﬂoor and see what happens. You start out with k eggs, and seek to drop eggs as few times as possible. Broken eggs cannot be reused. Let E(k, n) be the minimum number of egg droppings that will always suﬃce. (a) Show that E(1, n) = n. 1 (b) Show that E(k, n) = Θ(n k ). (c) Find a recurrence for E(k, n). What is the running time of the dynamic program to ﬁnd E(k, n)? Graph Problems 8-16. [4] Consider a city whose streets are deﬁned by an X × Y grid. We are interested in walking from the upper left-hand corner of the grid to the lower right-hand corner. Unfortunately, the city has bad neighborhoods, whose intersections we do not want to walk in. We are given an X × Y matrix BAD, where BAD[i,j] = “yes” if and only if the intersection between streets i and j is in a neighborhood to avoid. (a) Give an example of the contents of BAD such that there is no path across the grid avoiding bad neighborhoods. (b) Give an O(XY ) algorithm to ﬁnd a path across the grid that avoids bad neighborhoods. (c) Give an O(XY ) algorithm to ﬁnd the shortest path across the grid that avoids bad neighborhoods. You may assume that all blocks are of equal length. For partial credit, give an O(X 2 Y 2 ) algorithm. 313 314 8. DYNAMIC PROGRAMMING 8-17. [5] Consider the same situation as the previous problem. We have a city whose streets are deﬁned by an X × Y grid. We are interested in walking from the upper left-hand corner of the grid to the lower right-hand corner. We are given an X × Y matrix BAD, where BAD[i,j] = “yes” if and only if the intersection between streets i and j is somewhere we want to avoid. If there were no bad neighborhoods to contend with, the shortest path across the grid would have length (X − 1) + (Y − 1) blocks, and indeed there would be many such paths across the grid. Each path would consist of only rightward and downward moves. Give an algorithm that takes the array BAD and returns the number of safe paths of length X + Y − 2. For full credit, your algorithm must run in O(XY ). Design Problems 8-18. [4] Consider the problem of storing n books on shelves in a library. The order of the books is ﬁxed by the cataloging system and so cannot be rearranged. Therefore, we can speak of a book bi , where 1 ≤ i ≤ n, that has a thickness ti and height hi . The length of each bookshelf at this library is L. Suppose all the books have the same height h (i.e. , h = hi = hj for all i, j) and the shelves are all separated by a distance of greater than h, so any book ﬁts on any shelf. The greedy algorithm would ﬁll the ﬁrst shelf with as many books as we can until we get the smallest i such that bi does not ﬁt, and then repeat with subsequent shelves. Show that the greedy algorithm always ﬁnds the optimal shelf placement, and analyze its time complexity. 8-19. [6] This is a generalization of the previous problem. Now consider the case where the height of the books is not constant, but we have the freedom to adjust the height of each shelf to that of the tallest book on the shelf. Thus the cost of a particular layout is the sum of the heights of the largest book on each shelf. • Give an example to show that the greedy algorithm of stuﬃng each shelf as full as possible does not always give the minimum overall height. • Give an algorithm for this problem, and analyze its time complexity. Hint: use dynamic programming. 8-20. [5] We wish to compute the laziest way to dial given n-digit number on a standard push-button telephone using two ﬁngers. We assume that the two ﬁngers start out on the * and # keys, and that the eﬀort required to move a ﬁnger from one button to another is proportional to the Euclidean distance between them. Design an algorithm that computes the method of dialing that involves moving your ﬁngers the smallest amount of total distance, where k is the number of distinct keys on the keypad (k = 16 for standard telephones). Try to use O(nk3 ) time. 8-21. [6] Given an array of n real numbers, consider the problem of ﬁnding the maximum sum in any contiguous subvector of the input. For example, in the array {31, −41, 59, 26, −53, 58, 97, −93, −23, 84} the maximum is achieved by summing the third through seventh elements, where 59 + 26 + (−53) + 58 + 97 = 187. When all numbers are positive, the entire array is the answer, while when all numbers are negative, the empty array maximizes the total at 0. 8.10 EXERCISES • Give a simple, clear, and correct Θ(n2 )-time algorithm to ﬁnd the maximum contiguous subvector. • Now give a Θ(n)-time dynamic programming algorithm for this problem. To get partial credit, you may instead give a correct O(n log n) divide-and-conquer algorithm. 8-22. [7] Consider the problem of examining a string x = x1 x2 . . . xn from an alphabet of k symbols, and a multiplication table over this alphabet. Decide whether or not it is possible to parenthesize x in such a way that the value of the resulting expression is a, where a belongs to the alphabet. The multiplication table is neither commutative or associative, so the order of multiplication matters. a b c a a a c b c a c c c b c For example, consider the above multiplication table and the string bbbba. Parenthesizing it (b(bb))(ba) gives a, but ((((bb)b)b)a) gives c. Give an algorithm, with time polynomial in n and k, to decide whether such a parenthesization exists for a given string, multiplication table, and goal element. 8-23. [6] Let α and β be constants. Assume that it costs α to go left in a tree, and β to go right. Devise an algorithm that builds a tree with optimal worst case cost, given keys k1 , . . . , kn and the probabilities that each will be searched p1 , . . . , pn . Interview Problems 8-24. [5] Given a set of coin denominators, ﬁnd the minimum number of coins to make a certain amount of change. 8-25. [5] You are given an array of n numbers, each of which may be positive, negative, or zero. Give an eﬃcient algorithm to identify the index positions i and j to the maximum sum of the ith through jth numbers. 8-26. [7] Observe that when you cut a character out of a magazine, the character on the reverse side of the page is also removed. Give an algorithm to determine whether you can generate a given string by pasting cutouts from a given magazine. Assume that you are given a function that will identify the character and its position on the reverse side of the page for any given character position. Programming Challenges These programming challenge problems with robot judging are available at http://www.programming-challenges.com or http://online-judge.uva.es. 8-1. 8-2. 8-3. 8-4. 8-5. “Is Bigger Smarter?” – Programming Challenges 111101, UVA Judge 10131. “Weights and Measures” – Programming Challenges 111103, UVA Judge 10154. “Unidirectional TSP” – Programming Challenges 111104, UVA Judge 116. “Cutting Sticks” – Programming Challenges 111105, UVA Judge 10003. “Ferry Loading” – Programming Challenges 111106, UVA Judge 10261. 315 9 Intractable Problems and Approximation Algorithms We now introduce techniques for proving that no eﬃcient algorithm exists for a given problem. The practical reader is probably squirming at the notion of proving anything, and will be particularly alarmed at the idea of investing time to prove that something does not exist. Why are you better oﬀ knowing that something you don’t know how to do in fact can’t be done at all? The truth is that the theory of NP-completeness is an immensely useful tool for the algorithm designer, even though all it provides are negative results. The theory of NP-completeness enables the algorithm designer to focus her eﬀorts more productively, by revealing that the search for an eﬃcient algorithm for this particular problem is doomed to failure. When one fails to show a problem is hard, that suggests there may well be an eﬃcient algorithm to solve it. Two of the war stories in this book described happy results springing from bogus claims of hardness. The theory of NP-completeness also enables us to identify what properties make a particular problem hard. This provides direction for us to model it in diﬀerent ways or exploit more benevolent characteristics of the problem. Developing a sense for which problems are hard is an important skill for algorithm designers, and only comes from hands-on experience with proving hardness. The fundamental concept we will use is that of reductions between pairs of problems, showing that the problems are really equivalent. We illustrate this idea through a series of reductions, each of which either yields an eﬃcient algorithm or an argument that no such algorithm can exist. We also provide brief introductions to (1) the complexity-theoretic aspects of NP-completeness, one of the most fundamental notions in Computer Science, and (2) the theory of approximation algorithms, which leads to heuristics that probably return something close to the optimal solution. S.S. Skiena, The Algorithm Design Manual, 2nd ed., DOI: 10.1007/978-1-84800-070-4 9, c Springer-Verlag London Limited 2008 9.1 9.1 PROBLEMS AND REDUCTIONS Problems and Reductions We have encountered several problems in this book for which we couldn’t ﬁnd any eﬃcient algorithm. The theory of NP-completeness provides the tools needed to show that all these problems are on some level really the same problem. The key idea to demonstrating the hardness of a problem is that of a reduction, or translation, between two problems. The following allegory of NP-completeness may help explain the idea. A bunch of kids take turns ﬁghting each other in the schoolyard to prove how “tough” they are. Adam beats up Bill, who then beats up Chuck. So who if any among them is “tough?” The truth is that there is no way to know without an external standard. If I told you that Chuck was in fact Chuck Norris, certiﬁed tough guy, you have to be impressed—both Adam and Bill are at least as tough as he is. On the other hand, suppose I tell you it is a kindergarten school yard. No one would call me tough, but even I can take out Adam. This proves that none of the three of them should be called be tough. In this allegory, each ﬁght represents a reduction. Chuck Norris takes on the role of satisﬁability—a certiﬁably hard problem. The part of an ineﬃcient algorithm with a possible shot at redemption is played by me. Reductions are operations that convert one problem into another. To describe them, we must be somewhat rigorous in our deﬁnitions. An algorithmic problem is a general question, with parameters for input and conditions on what constitutes a satisfactory answer or solution. An instance is a problem with the input parameters speciﬁed. The diﬀerence can be made clear by an example: Problem: The Traveling Salesman Problem (TSP) Input: A weighted graph G. n−1 Output: Which tour {v1 , v2 , ..., vn } minimizes i=1 d[vi , vi+1 ] + d[vn , v1 ]? Any weighted graph deﬁnes an instance of TSP. Each particular problem has at least one minimum cost tour. The general traveling salesman instance asks for an algorithm to ﬁnd the optimal tour for all possible instances. 9.1.1 The Key Idea Now consider two algorithmic problems, called Bandersnatch and Bo-billy. Suppose that I gave you the following reduction/algorithm to solve the Bandersnatch problem: Bandersnatch(G) Translate the input G to an instance Y of the Bo-billy problem. Call the subroutine Bo-billy on Y to solve this instance. Return the answer of Bo-billy(Y ) as the answer to Bandersnatch(G). 317 318 9. INTRACTABLE PROBLEMS AND APPROXIMATION ALGORITHMS This algorithm will correctly solve the Bandersnatch problem provided that the translation to Bo-billy always preserves the correctness of the answer. In other words, the translation has the property that for any instance of G, Bandersnatch(G) = Bo-billy(Y ) A translation of instances from one type of problem to instances of another such that the answers are preserved is what is meant by a reduction. Now suppose this reduction translates G to Y in O(P (n)) time. There are two possible implications: • If my Bo-billy subroutine ran in O(P (n)), this means I could solve the Bandersnatch problem in O(P (n) + P (n)) by spending the time to translate the problem and then the time to execute the Bo-Billy subroutine. • If I know that Ω(P (n)) is a lower bound on computing Bandersnatch, meaning there deﬁnitely exists no faster way to solve it, then Ω(P (n) − P (n)) must be a lower bound to compute Bo-billy. Why? If I could solve Bo-billy any faster, then I could violate my lower bound by solving Bandersnatch using the above reduction. This implies that there can be no way to solve Bo-billy any faster than claimed. This ﬁrst argument is Steve demonstrating the weakness of the entire schoolyard with a quick right to Adam’s chin. The second highlights the Chuck Norris approach we will use to prove that problems are hard. Essentially, this reduction shows that Bo-billy is no easier than Bandersnatch. Therefore, if Bandersnatch is hard this means Bo-billy must also be hard. We will illustrate this point by giving several problem reductions in this chapter. Take-Home Lesson: Reductions are a way to show that two problems are essentially identical. A fast algorithm (or the lack of one) for one of the problems implies a fast algorithm (or the lack of one) for the other. 9.1.2 Decision Problems Reductions translate between problems so that their answers are identical in every problem instance. Problems diﬀer in the range or type of possible answers. The traveling salesman problem returns a permutation of vertices as the answer, while other types of problems return numbers as answers, perhaps restricted to positive numbers or integers. The simplest interesting class of problems have answers restricted to true and false. These are called decision problems. It proves convenient to reduce/translate answers between decision problems because both only allow true and false as possible answers. Fortunately, most interesting optimization problems can be phrased as decision problems that capture the essence of the computation. For example, the traveling salesman decision problem could be deﬁned as: 9.2 REDUCTIONS FOR ALGORITHMS Problem: The Traveling Salesman Decision Problem Input: A weighted graph G and integer k. Output: Does there exist a TSP tour with cost ≤ k? The decision version captures the heart of the traveling salesman problem, in that if you had a fast algorithm for the decision problem, you could use it to do a binary search with diﬀerent values of k and quickly hone in on the optimal solution. With a little more cleverness, you could reconstruct the actual tour permutation using a fast solution to the decision problem. From now on we will generally talk about decision problems, because it proves easier and still captures the power of the theory. 9.2 Reductions for Algorithms An engineer and an algorist are sitting in a kitchen. The algorist asks the engineer to boil some water, so the engineer gets up, picks up the kettle from the counter top, adds water from the sink, brings it to the burner, turns on the burner, waits for the whistling sound, and turns oﬀ the burner. Sometime later, the engineer asks the algorist to boil more water. She gets up, takes the kettle from the burner, moves it over to the counter top, and sits down. “Done.” she says, “I have reduced the task to a solved problem.” This boiling water reduction illustrates an honorable way to generate new algorithms from old. If we can translate the input for a problem we want to solve into input for a problem we know how to solve, we can compose the translation and the solution into an algorithm for our problem. In this section, we look at several reductions that lead to eﬃcient algorithms. To solve problem a, we translate/reduce the a instance to an instance of b, then solve this instance using an eﬃcient algorithm for problem b. The overall running time is the time needed to perform the reduction plus that solve the b instance. 9.2.1 Closest Pair The closest pair problem asks to ﬁnd the pair of numbers within a set that have the smallest diﬀerence between them. We can make it a decision problem by asking if this value is less than some threshold: Input: A set S of n numbers, and threshold t. Output: Is there a pair si , sj ∈ S such that |si − sj | ≤ t? The closest pair is a simple application of sorting, since the closest pair must be neighbors after sorting. This gives the following algorithm: CloseEnoughPair(S,t) Sort S. Is min1≤i 3, meaning that Ci = {z1 , z2 , ..., zn } – We create n − 3 new variables and n−2 new clauses in a chain, where for 2 ≤ j ≤ n−3, Ci,j = {vi,j−1 , zj+1 , v i,j }, Ci,1 = {z1 , z2 , v i,1 }, and Ci,n−2 = {vi,n−3 , zn−1 , zn }. The most complicated case here is that of the large clauses. If none of the original literals in Ci are true, then there are not enough new variables to be able to satisfy all of the new subclauses. You can satisfy Ci,1 by setting vi,1 = false, but this forces vi,2 = false, and so on until ﬁnally Ci,n−2 cannot be satisﬁed. However, if any single literal zi = true, then we have n − 3 free variables and n − 3 remaining 3-clauses, so we can satisfy each of them. This transform takes O(m + n) time if there were n clauses and m total literals in the SAT instance. Since any SAT solution also satisﬁes the 3-SAT instance and any 3-SAT solution describes how to set the variables giving a SAT solution, the transformed problem is equivalent to the original. Note that a slight modiﬁcation to this construction would serve to prove that 4-SAT, 5-SAT, or any (k ≥ 3)-SAT is also NP-complete. However, this construction breaks down if we try to use it for 2-SAT, since there is no way to stuﬀ anything into the chain of clauses. It turns out that a depth-ﬁrst search on an appropriate graph can be used to give a linear-time algorithm for 2-SAT, as discussed in Section 14.10 (page 472). 9.5 Creative Reductions Since both satisﬁability and 3-SAT are known to be hard, we can use either of them in reductions. Usually 3-SAT is the better choice, because it is simpler to work with. What follows are a pair of more complicated reductions, designed to serve as examples and also increase our repertoire of known hard problems. Many 9.5 CREATIVE REDUCTIONS reductions are quite intricate, because we are essentially programming one problem in the language of a signiﬁcantly diﬀerent problem. One perpetual point of confusion is getting the direction of the reduction right. Recall that we must transform every instance of a known NP-complete problem into an instance of the problem we are interested in. If we perform the reduction the other way, all we get is a slow way to solve the problem of interest, by using a subroutine that takes exponential time. This always is confusing at ﬁrst, for this direction of reduction seems backwards. Make sure you understand the direction of reduction now, and think back to this whenever you get confused. 9.5.1 Integer Programming As discussed in Section 13.6 (page 411), integer programming is a fundamental combinatorial optimization problem. It is best thought of as linear programming with the variables restricted to take only integer (instead of real) values. Problem: Integer Programming Input: A set of integer variables V , a set of inequalities over V , a maximization function f (V ), and an integer B. Output: Does there exist an assignment of integers to V such that all inequalities are true and f (V ) ≥ B? Consider the following two examples. Suppose v1 ≥ 1, v2 ≥ 0 v1 + v 2 ≤ 3 f (v) : 2v2 , B=3 A solution to this would be v1 = 1, v2 = 2. Not all problems have realizable solutions, however. For the following problem: v1 ≥ 1, v2 ≥ 0 v1 + v 2 ≤ 3 f (v) : 2v2 , B=5 The maximum possible value of f (v) given the constraints is 2 × 2 = 4, so there can be no solution to the associated decision problem. We show that integer programming is hard using a reduction from 3-SAT. For this particular reduction, general satisﬁability would work just as well, although usually 3-SAT makes reductions easier. In which direction must the reduction go? We want to prove integer programming is hard, and know that 3-SAT is hard. If I could solve 3-SAT using integer 331 332 9. INTRACTABLE PROBLEMS AND APPROXIMATION ALGORITHMS programming and integer programming were easy, this would mean that satisﬁability would be easy. Now the direction should be clear; we must translate 3-SAT into integer programming. What should the translation look like? Every satisﬁability instance contains Boolean (true/false) variables and clauses. Every integer programming instance contains integer variables (values restricted to 0,1,2,. . . ) and constraints. A reasonable idea is to make the integer variables correspond to Boolean variables and use constraints to serve the same role as the clauses do in the original problem. Our translated integer programming problem will have twice as many variables as the SAT instance—one for each variable and one for its complement. For each variable vi in the set problem, we will add the following constraints: • We restrict each integer programming variable Vi to values of either 0 or 1, but adding constraints 1 ≥ Vi ≥ 0 and 1 ≥ V i ≥ 0. Thus coupled with integrality, they correspond to values of true and false. • We ensure that exactly one of the two integer programming variables associated with a given SAT variable is true, by adding constraints so that 1 ≥ Vi + V i ≥ 1. For each 3-SAT clause Ci = {z1 , z2 , z3 }, construct a constraint: V1 +V2 +V3 ≥ 1. To satisfy this constraint, at least one of the literals per clause must be set to 1, thus corresponding to a true literal. Satisfying this constraint is therefore equivalent to satisfying the clause. The maximization function and bound prove relatively unimportant, since we have already encoded the entire 3-SAT instance. By using f (v) = V1 and B = 0, we ensure that they will not interfere with any variable assignment satisfying all the inequalities. Clearly, this reduction can be done in polynomial time. To establish that this reduction preserves the answer, we must verify two things: • Any SAT solution gives a solution to the IP problem – In any SAT solution, a true literal corresponds to a 1 in the integer program, since the clause is satisﬁed. Therefore, the sum in each clause inequality is ≥ 1. • Any IP solution gives a solution to the original SAT problem – All variables must be set to either 0 or 1 in any solution to this integer programming instance. If Vi = 1, then set literal zi = true. If Vi = 0, then set literal zi = false. This is a legal assignment which must also satisfy all the clauses. The reduction works both ways, so integer programming must be hard. Notice the following properties, which hold true in general for NP-completeness proofs: 1. This reduction preserved the structure of the problem. It did not solve the problem, just put it into a diﬀerent format. 9.5 v1 v1 v2 v2 v3 v3 v3 v1 CREATIVE REDUCTIONS v4 v4 v2 v4 v1 v4 Figure 9.7: Reducing satisﬁability instance {{v1 , v¯3 , v¯4 }, {v¯1 , v2 , v¯4 }} to vertex cover 2. The possible IP instances that can result from this transformation are only a small subset of all possible IP instances. However, since some of them are hard, the general problem must be hard. 3. The transformation captures the essence of why IP is hard. It has nothing to do with having big coeﬃcients or big ranges of variables, because restricting them to 0/1 is enough. It has nothing to do with having inequalities with large numbers of variables. Integer programming is hard because satisfying a set of constraints is hard. A careful study of the properties needed for a reduction can tell us a lot about the problem. 9.5.2 Vertex Cover Algorithmic graph theory proves to be a fertile ground for hard problems. The prototypical NP-complete graph problem is vertex cover, previously deﬁned in Section 9.3.2 (page 325) as follows: Problem: Vertex Cover Input: A graph G = (V, E) and integer k ≤ |V |. Output: Is there a subset S of at most k vertices such that every e ∈ E has at least one vertex in S? Demonstrating the hardness of vertex cover proves more diﬃcult than the previous reductions we have seen, because the structure of the two relevant problems is very diﬀerent. A reduction from 3-satisﬁability to vertex cover has to construct a graph G and bound k from the variables and clauses of the satisﬁability instance. First, we translate the variables of the 3-SAT problem. For each Boolean variable vi , we create two vertices vi and v¯i connected by an edge. At least n vertices will be needed to cover all these edges, since no two of the edges will share a vertex. Second, we translate the clauses of the 3-SAT problem. For each of the c clauses, we create three new vertices, one for each literal in each clause. The three vertices of each clause will be connected so as to form c triangles. At least two vertices per 333 334 9. INTRACTABLE PROBLEMS AND APPROXIMATION ALGORITHMS triangle must be included in any vertex cover of these triangles, for a total of 2c cover vertices. Finally, we will connect these two sets of components together. Each literal in the vertex “gadgets” is connected to vertices in the clause gadgets (triangles) that share the same literal. From a 3-SAT instance with n variables and c clauses, this constructs a graph with 2n + 3c vertices. The complete reduction for the 3-SAT problem {{v1 , v¯3 , v¯4 }, {v¯1 , v2 , v¯4 }} is shown in Figure 9.7. This graph has been designed to have a vertex cover of size n + 2c if and only if the original expression is satisﬁable. By the earlier analysis, every vertex cover must have at least n + 2c vertices, since adding the connecting edges to G cannot shrink the size of the vertex cover to less than that of the disconnected vertex and clause gadgets. To show that our reduction is correct, we must demonstrate that: • Every satisfying truth assignment gives a vertex cover – Given a satisfying truth assignment for the clauses, select the n vertices from the vertex gadgets that correspond to true literals to be members of the vertex cover. Since this deﬁnes a satisfying truth assignment, a true literal from each clause must cover at least one of the three cross edges connecting each triangle vertex to a vertex gadget. Therefore, by selecting the other two vertices of each clause triangle, we also pick up all remaining cross edges. • Every vertex cover gives a satisfying truth assignment – In any vertex cover C of size n + 2c, exactly n of the vertices must belong to the vertex gadgets. Let these ﬁrst stage vertices deﬁne the truth assignment, while the remaining 2c cover vertices must be distributed at two per clause gadget. Otherwise a clause gadget edge must go uncovered. These clause gadget vertices can cover only two of the three connecting cross edges per clause. Therefore, if C gives a vertex cover, at least one cross edge per clause must be covered, meaning that the corresponding truth assignment satisﬁes all clauses. This proof of the hardness of vertex cover, chained with the clique and independent set reductions of Section 9.3.2 (page 325), gives us a library of hard graph problems that we can use to make future hardness proofs easier. Take-Home Lesson: A small set of NP-complete problems (3-SAT, vertex cover, integer partition, and Hamiltonian cycle) suﬃce to prove the hardness of most other hard problems. 9.6 The Art of Proving Hardness Proving that problems are hard is a skill. But once you get the hang of it, reductions can be surprisingly straightforward and pleasurable to do. Indeed, the dirty little secret of NP-completeness proofs is that they are usually easier to create than explain, in much the same way that it can be easier to rewrite old code than understand and modify it. 9.6 THE ART OF PROVING HARDNESS It takes experience to judge whether a problem is likely to be hard. Perhaps the quickest way to gain this experience is through careful study of the catalog. Slightly changing the wording of a problem can make the diﬀerence between it being polynomial or NP-complete. Finding the shortest path in a graph is easy, while ﬁnding the longest path in a graph is hard. Constructing a tour that visits all the edges once in a graph is easy (Eulerian cycle), while constructing a tour that visits all the vertices once is hard (Hamiltonian cycle). The ﬁrst thing to do when you suspect a problem might be NP-complete is look in Garey and Johnson’s book Computers and Intractability [GJ79], which contains a list of several hundred problems known to be NP-complete. Likely you will ﬁnd the problem you are interested in. Otherwise I oﬀer the following advice to those seeking to prove the hardness of a given problem: • Make your source problem as simple (i.e. , restricted) as possible. Never try to use the general traveling salesman problem (TSP) as a source problem. Better, use Hamiltonian cycle (i.e. , TSP) where all the weights are 1 or ∞. Even better, use Hamiltonian path instead of cycle, so you don’t have to worry about closing up the cycle. Best of all, use Hamiltonian path on directed planar graphs where each vertex has total degree 3. All of these problems are equally hard, but the more you can restrict the problem that you are reducing, the less work your reduction has to do. As another example, never try to use full satisﬁability to prove hardness. Start with 3-satisﬁability. In fact, you don’t even have to use full 3-satisﬁability. Instead, you can use planar 3-satisﬁability, where there exists a way to draw the clauses as a graph in the plane such that you can connect all instances of the same literal together without edges crossing. This property tends to be useful in proving the hardness of geometric problems. All these problems are equally hard, and hence NP-completeness reductions using any of them are equally convincing. • Make your target problem as hard as possible. Don’t be afraid to add extra constraints or freedoms to make your target problem more general. Perhaps your undirected graph problem can be generalized into a directed graph problem and can hence be easier to prove hard. Once you have a proof of hardness for the general problem, you can then go back and try to simplify the target. • Select the right source problem for the right reason. Selecting the right source problem makes a big diﬀerence in how diﬃcult it is to prove a problem hard. This is the ﬁrst and easiest place to go wrong, although theoretically any NP-complete problem works as well as any other. When trying to prove that a problem is hard, some people ﬁsh around through 335 336 9. INTRACTABLE PROBLEMS AND APPROXIMATION ALGORITHMS lists of dozens of problems, looking for the best ﬁt. These people are amateurs; odds are they never will recognize the problem they are looking for when they see it. I use four (and only four) problems as candidates for my hard source problem. Limiting them to four means that I can know a lot about each of these problems, such as which variants of the problems are hard and which are not. My favorite source problems are: – 3-SAT: The old reliable. When none of the three problems below seem appropriate, I go back to the original source. – Integer partition: This is the one and only choice for problems whose hardness seems to require using large numbers. – Vertex cover: This is the answer for any graph problem whose hardness depends upon selection. Chromatic number, clique, and independent set all involve trying to select the correct subset of vertices or edges. – Hamiltonian path: This is my choice for any graph problem whose hardness depends upon ordering. If you are trying to route or schedule something, Hamiltonian path is likely your lever into the problem. • Amplify the penalties for making the undesired selection. Many people are too timid in their thinking about hardness proofs. You are trying to translate one problem into another, while keeping the problems as close to their original identities as as possible. The easiest way to do this is to be bold with your penalties; to punish anyone for trying to deviate from your intended solution. Your thinking should be, “if you select this element, then you must pick up this huge set that prevents you from ﬁnding an optimal solution.” The sharper the consequences for doing what is undesired, the easier it is to prove the equivalence of the problems. • Think strategically at a high level, then build gadgets to enforce tactics. You should be asking yourself the following types of questions: “How can I force that A or B is chosen but not both?” “How can I force that A is taken before B?” “How can I clean up the things I did not select?” Once you have an idea of what you want your gadgets to do, you can worry about how to actually craft them. • When you get stuck, alternate between looking for an algorithm or a reduction. Sometimes the reason you cannot prove hardness is that there exists an eﬃcient algorithm to solve your problem! Techniques such as dynamic programming or reducing problems to powerful but polynomial-time graph problems, such as matching or network ﬂow, can yield surprising algorithms. Whenever you can’t prove hardness, it pays to alter your opinion occasionally to keep yourself honest. 9.7 9.7 WAR STORY: HARD AGAINST THE CLOCK War Story: Hard Against the Clock My class’s attention span was running down like sand through an hourglass. Eyes were starting to glaze, even in the front row. Breathing had become soft and regular in the middle of the room. Heads were tilted back and eyes shut in the back. There were twenty minutes left to go in my lecture on NP-completeness, and I couldn’t really blame them. They had already seen several reductions like the ones presented here. But NP-completeness reductions are easier to create than to understand or explain. They had to watch one being created to appreciate how things worked. I reached for my trusty copy of Garey and Johnson’s book [GJ79], which contains a list of over four hundred diﬀerent known NP-complete problems in an appendix in the back. “Enough of this!” I announced loudly enough to startle those in the back row. “NP-completeness proofs are suﬃciently routine that we can construct them on demand. I need a volunteer with a ﬁnger. Can anyone help me?” A few students in the front held up their hands. A few students in the back held up their ﬁngers. I opted for one from the front row. “Select a problem at random from the back of this book. I can prove the hardness of any of these problems in the now seventeen minutes remaining in this class. Stick your ﬁnger in and read me a problem.” I had deﬁnitely gotten their attention. But I could have done that by oﬀering to juggle chain-saws. Now I had to deliver results without cutting myself into ribbons. The student picked out a problem. “OK, prove that Inequivalence of Programs with Assignments is hard,” she said. “Huh? I’ve never heard of that problem before. What is it? Read me the entire description of the problem so I can write it on the board.” The problem was as follows: Problem: Inequivalence of Programs with Assignments Input: A ﬁnite set X of variables, two programs P1 and P2 , each a sequence of assignments of the form x0 ← if (x1 = x2 ) then x3 else x4 where the xi are in X; and a value set V . Output: Is there an initial assignment of a value from V to each variable in X such that the two programs yield diﬀerent ﬁnal values for some variable in X? I looked at my watch. Fifteen minutes to go. But now everything was on the table. I was faced with a language problem. The input was two programs with variables, and I had to test to see whether they always do the same thing. “First things ﬁrst. We need to select a source problem for our reduction. Do we start with integer partition? 3-satisﬁability? Vertex cover or Hamiltonian path?” Since I had an audience, I tried thinking out loud. “Our target is not a graph problem or a numerical problem, so let’s start thinking about the old reliable: 3- 337 338 9. INTRACTABLE PROBLEMS AND APPROXIMATION ALGORITHMS satisﬁability. There seem to be some similarities. 3-SAT has variables. This thing has variables. To be more like 3-SAT, we could try limiting the variables in this problem so they only take on two values—i.e. , V = {true, false}. Yes. That seems convenient.” My watch said fourteen minutes left. “So, class, which way does the reduction go? 3-SAT to language or language to 3-SAT?” The front row correctly murmured, “3-SAT to language.” “Right. So we have to translate our set of clauses into two programs. How can we do that? We can try to split the clauses into two sets and write separate programs for each of them. But how do we split them? I don’t see any natural way do it, because eliminating any single clause from the problem might suddenly make an unsatisﬁable formula satisﬁable, thus completely changing the answer. Instead, let’s try something else. We can translate all the clauses into one program, and then make the second program be trivial. For example, the second program might ignore the input and always output either only true or only false. This sounds better. Much better.” I was still talking out loud to myself, which wasn’t that unusual. But I had people listening to me, which was. “Now, how can we turn a set of clauses into a program? We want to know whether the set of clauses can be satisﬁed, or if there is an assignment of the variables such that it is true. Suppose we constructed a program to evaluate whether c1 = (x1 , x2 , x3 ) is satisﬁed.” It took me a few minutes worth of scratching before I found the right program to simulate a clause. I assumed that I had access to constants for true and false: c1 = if (x1 = true) then true else false c1 = if (x2 = false) then true else c1 c1 = if (x3 = true) then true else c1 “Great. Now I have a way to evaluate the truth of each clause. I can do the same thing to evaluate whether all the clauses are satisﬁed.” sat = if (c1 = true) then true else false sat = if (c2 = true) then sat else false .. . sat = if (cn = true) then sat else false Now the back of the classroom was getting excited. They were starting to see a ray of hope that they would get to leave on time. There were two minutes left in class. “Great. So now we have a program that can evaluate to be true if and only if there is a way to assign the variables to satisfy the set of clauses. We need a second program to ﬁnish the job. What about sat = false? Yes, that is all we need. Our language problem asks whether the two programs always output the same 9.8 WAR STORY: AND THEN I FAILED thing, regardless of the possible variable assignments. If the clauses are satisﬁable, that means that there must be an assignment of the variables such that the long program would output true. Testing whether the programs are equivalent is exactly the same as asking if the clauses are satisﬁable.” I lifted my arms in triumph. “And so, the problem is neat, sweet, and NPcomplete.” I got the last word out just before the bell rang. 9.8 War Story: And Then I Failed This exercise of picking a random NP-complete problem from the 400+ problems in Garey and Johnson’s book and proving hardness on demand was so much fun that I have repeated it each time I have taught the algorithms course. Sure enough, I got it eight times in a row. But just as Joe DiMaggio’s 56-game hitting streak came to an end, and Google will eventually have a losing quarter ﬁnancially, the time came for me to get my comeuppance. The class had voted to see a reduction from the graph theory section of the catalog, and a randomly selected student picked number 30. Problem GT30 turned out to be: Problem: Uniconnected Subgraph Input: Directed graph G = (V, A), positive integer k ≤ |A|. Output: Is there a subset of arcs A ∈ A with |A | ≥ k such that G = (V, A ) has at most one directed path between any pair of vertices? “It is a selection problem,” I realized as soon as the problem was revealed. After all, we had to select the largest possible subset of arcs so that there were no pair of vertices with multiple paths between them. This meant that vertex cover was the problem of choice. I worked through how the two problems stacked up. Both sought subsets, although vertex cover wanted subsets of vertices and uniconnected subgraph wanted subsets of edges. Vertex cover wanted the smallest possible subset, while unidirected subgraph wanted the largest possible subset. My source problem had undirected edges while my target had directed arcs, so somehow I would have to add edge direction into the reduction. I had to do something to direct the edges of the vertex cover graph. I could try to replace each undirected edge (x, y) with a single arc, say from y to x. But quite diﬀerent directed graphs would result depending upon which direction I selected. Finding the “right” orientation of edges might be a hard problem, certainly too hard to use in the translation phase of the reduction. I realized I could direct the edges so the resulting graph was a DAG. But then, so what? DAGs certainly can have many diﬀerent directed paths between pairs of vertices. Alternately, I could try to replace each undirected edge (x, y) with two arcs, from x to y and y to x. Now there was no need to chose the right arcs for my 339 340 9. INTRACTABLE PROBLEMS AND APPROXIMATION ALGORITHMS reduction, but the graph certainly got complicated. I couldn’t see how to force things to prevent vertex pairs from having unwanted multiple paths between them. Meanwhile, the clock was running and I knew it. A sense of panic set in during the last ten minutes of the class, and I realized I wasn’t going to get it this time. There is no feeling worse for a professor than botching up a lecture. You stand up there ﬂailing away, knowing (1) that the students don’t understand what you are saying, but (2) they do understand that you also don’t understand what you are saying. The bell rang and the students left the room with faces either sympathetic or smirking. I promised them a solution for the next class, but somehow I kept getting stuck in the same place each time I thought about it. I even tried to cheat and look up the proof in a journal. But the reference that Garey and Johnson cited was a 30-year old unpublished technical report. It wasn’t on the web or in our library. I dreaded the idea of returning to give my next class, the last lecture of the semester. But the night before class the answer came to me in a dream. “Split the edges,” the voice said. I awoke with a start and looked at the clock. It was 3:00 AM. I sat up in bed and scratched out the proof. Suppose I replace each undirected edge (x, y) with a gadget consisting of a new central vertex vxy with arcs going from it to x and y, respectively. This is nice. Now, which vertices are capable of having multiple paths between them? The new vertices had only outgoing edges, so they can only serve as the source of multiple paths. The old vertices had only incoming edges. There was at most one way to get from one of the new source vertices to any of the original vertices of the vertex cover graph, so these could not result in multiple paths. But now add a sink node s with edges from all original vertices. There were exactly two paths from each new source vertex to this sink—one through each of the two original vertices it was adjacent to. One of these had to be broken to create a uniconnected subgraph. How could we break it? We could pick one of these two vertices to disconnect from the sink by deleting either arc (x, s) or (y, s) for new vertex vxy . We maximize the size of our subgraph by ﬁnding the smallest number of arcs to delete. We must delete the outgoing arc from at least one of the two vertices deﬁning each original edge. This is exactly the same as ﬁnding the vertex cover in this graph! The reduction is illustrated in Figure 9.8. Presenting this proof in class provided some personal vindication, but more to the point validates the principles I teach for proving hardness. Observe that the reduction really wasn’t all that diﬃcult after all: just split the edges and add a sink node. NP-completeness reductions are often surprisingly simple once you look at them the right way. 9.9 P VS. NP 341 2 5 2 5 1 1 3 6 3 6 4 4 s Figure 9.8: Reducing vertex cover to unidirected subgraph, by dividing edges and adding a sink node 9.9 P vs. NP The theory of NP-completeness rests on a foundation of rigorous but subtle deﬁnitions from automata and formal language theory. This terminology is typically confusing to or misused by beginners who lack a mastery of these foundations. It is not really essential to the practical aspects of designing and applying reductions. That said, the question “Is P=NP?” is the most profound open problem in computer science, so any educated algorist should have some idea what the stakes are. 9.9.1 Veriﬁcation vs. Discovery The primary question in P vs NP is whether veriﬁcation is really an easier task than initial discovery. Suppose that while taking an exam you “happen” to notice the answer of the student next to you. Are you now better oﬀ? You wouldn’t dare to turn it in without checking, since an able student such as yourself could answer the question correctly if you took enough time to solve it. The issue is whether you can really verify the answer faster than you could ﬁnd it from scratch. For the NP-complete decision problems we have studied here, the answer seems obvious: • Can you verify that a graph has a TSP tour of at most k weight given the order of vertices on the tour? Sure. Just add up the weights of the edges on the tour and show it is at most k. That is easier than ﬁnding the tour, isn’t it? • Can you verify that a given truth assignment represents a solution to a given satisﬁability problem? Sure. Just check each clause and make sure it contains 342 9. INTRACTABLE PROBLEMS AND APPROXIMATION ALGORITHMS at least one true literal from the given truth assignment. That is easier than ﬁnding the satisfying assignment, isn’t it? • Can you verify that a graph G has a vertex cover of at most k vertices if given the subset S of at most k vertices forming such a cover? Sure. Just traverse each edge (u, v) of G, and check that either u or v is in S. That is easier than ﬁnding the vertex cover, isn’t it? At ﬁrst glance, this seems obvious. The given solutions can be veriﬁed in linear time for all three of these problems, while no algorithm faster than brute force search is known to ﬁnd the solutions for any of them. The catch is that we have no rigorous lower bound proof that prevents the existence of fast algorithms to solve these problems. Perhaps there are in fact polynomial algorithms (say O(n7 )) that we have just been too blind to see yet. 9.9.2 The Classes P and NP Every well-deﬁned algorithmic problem must have an asymptotically fastestpossible algorithm solving it, as measured in the Big-Oh, worst-case sense of fastest. We can think of the class P as an exclusive club for algorithmic problems that a problem can only join after demonstrating that there exists a polynomial-time algorithm to solve it. Shortest path, minimum spanning tree, and the original movie scheduling problem are all members in good standing of this class P . The P stands for polynomial-time. A less-exclusive club welcomes all the algorithmic problems whose solutions can be veriﬁed in polynomial-time. As shown above, this club contains traveling salesman, satisﬁability, and vertex cover, none of which currently have the credentials to join P . However, all the members of P get a free pass into this less exclusive club. If you can solve the problem from scratch in polynomial time, you certainly can verify a solution that fast: just solve it from scratch and see if the solution you get is as good as the one you were given. We call this less-exclusive club N P . You can think of this as standing for notnecessarily polynomial-time.1 The $1,000,000 question is whether there are in fact problems in NP that cannot be members of P. If no such problem exists, the classes must be the same and P = N P . If even one such a problem exists, the two classes are diﬀerent and P = N P . The opinion of most algorists and complexity theorists is that the classes diﬀer, meaning P = N P , but a much stronger proof than “I can’t ﬁnd a fast enough algorithm” is needed. 1 In fact, it stands for nondeterministic polynomial-time. This is in the sense of nondeterministic automata, if you happen to know about such things. 9.9 9.9.3 P VS. NP Why is Satisﬁability the Mother of All Hard Problems? An enormous tree of NP-completeness reductions has been established that entirely rests on the hardness of satisﬁability. The portion of this tree demonstrated and/or stated in this chapter (and proven elsewhere) is shown in Figure 9.2. This may look like a delicate aﬀair. What would it mean if someone does ﬁnd a polynomial-time algorithm for satisﬁability? A fast algorithms for any given NPcomplete problem (say traveling salesman) implies fast algorithm for all the problems on the path in the reduction tree between TSP and satisﬁability (Hamiltonian cycle, vertex cover, and 3-SAT). But a fast algorithm for satisﬁability doesn’t immediately yield us anything because the reduction path from SAT to SAT is empty. Fear not. There exists an extraordinary super-reduction (called Cook’s theorem) reducing all the problems in NP to satisﬁability. Thus, if you prove that satisﬁability (or equivalently any single NP-complete problem) is in P , then all other problems in NP follow and P = N P . Since essentially every problem mentioned in this book is in NP, this would be an enormously powerful and surprising result. Cook’s theorem proves that satisﬁability is as hard as any problem in NP. Furthermore, it proves that every NP-complete problem is as hard as any other. Any domino falling (i.e. , a polynomial-time algorithm for any NP-complete problem) knocks them all down. Our inability to ﬁnd a fast algorithm for any of these problems is a strong reason for believing that they are all truly hard, and probably P = N P . 9.9.4 NP-hard vs. NP-complete? The ﬁnal technicality we will discuss is the diﬀerence between a problem being NP-hard and NP-complete. I tend to be somewhat loose with my terminology, but there is a subtle (usually irrelevant) distinction between the two concepts. We say that a problem is NP-hard if, like satisﬁability, it is at least as hard as any problem in NP. We say that a problem is NP-complete if it is NP-hard, and also in NP itself. Because NP is such a large class of problems, most NP-hard problems you encounter will actually be complete, and the issue can always be settled by giving a (usually simple) veriﬁcation strategy for the problem. All the NP-hard problems we have encountered in this book are also NP-complete. That said, there are problems that appear to be NP-hard yet not in NP. These problems might be even harder than NP-complete! Two-player games such as chess provide examples of problems that are not in NP. Imagine sitting down to play chess with some know-it-all, who is playing white. He pushes his central pawn up two squares to start the game, and announces checkmate. The only obvious way to show he is right would be to construct the full tree of all your possible moves with his irrefutable replies and demonstrate that you, in fact, cannot win from the current position. This full tree will have a number of nodes exponential in its height, which is the number of moves before you lose playing your most spirited possible defense. 343 344 9. INTRACTABLE PROBLEMS AND APPROXIMATION ALGORITHMS Clearly this tree cannot be constructed and analyzed in polynomial time, so the problem is not in NP. 9.10 Dealing with NP-complete Problems For the practical person, demonstrating that a problem is NP-complete is never the end of the line. Presumably, there was a reason why you wanted to solve it in the ﬁrst place. That application will not go away when told that there is no polynomial-time algorithm. You still seek a program that solves the problem of interest. All you know is that you won’t ﬁnd one that quickly solves the problem to optimality in the worst case. You still have three options: • Algorithms fast in the average case – Examples of such algorithms include backtracking algorithms with substantial pruning. • Heuristics – Heuristic methods like simulated annealing or greedy approaches can be used to quickly ﬁnd a solution with no guarantee that it will be the best one. • Approximation algorithms – The theory of NP-completeness only stipulates that it is hard to get close to the answer. With clever, problem-speciﬁc heuristics, we can probably get close to the optimal answer on all possible instances. Approximation algorithms return solutions with a guarantee attached, namely that the optimal solution can never be much better than this given solution. Thus you can never go too far wrong when using an approximation algorithm. No matter what your input instance is and how lucky you are, you are doomed to do all right. Furthermore, approximation algorithms realizing probably good bounds are often conceptually simple, very fast, and easy to program. One thing that is usually not clear, however, is how well the solution from an approximation algorithm compares to what you might get from a heuristic that gives you no guarantees. The answer could be worse or it could be better. Leaving your money in a bank savings account guarantees you 3% interest without risk. Still, you likely will do much better investing your money in stocks than leaving it in the bank, even though performance is not guaranteed. One way to get the best of approximation algorithms and heuristics is to run both of them on the given problem instance and pick the solution giving the better result. This way, you get a solution that comes with a guarantee and a second chance to do even better. When it comes to heuristics for hard problems, sometimes you can have it both ways. 9.10 DEALING WITH NP-COMPLETE PROBLEMS Figure 9.9: Neglecting to pick the center vertex leads to a terrible vertex cover 9.10.1 Approximating Vertex Cover As we have seen before, ﬁnding the minimum vertex cover of a graph is NPcomplete. However, a very simple procedure can eﬃciently ﬁnd a cover that is at most twice as large as the optimal cover: VertexCover(G = (V, E)) While (E = ∅) do: Select an arbitrary edge (u, v) ≤ E Add both u and v to the vertex cover Delete all edges from E that are incident to either u or v. It should be apparent that this procedure always produces a vertex cover, since each edge is only deleted after an incident vertex has been added to the cover. More interesting is the claim that any vertex cover must use at least half as many vertices as this one. Why? Consider only the ≤ n/2 edges selected by the algorithm that constitute a matching in the graph. No two of these edges can share a vertex. Therefore, any cover of just these edges must include at least one vertex per edge, which makes it at least half the size of this greedy cover. There are several interesting things to notice about this algorithm: • Although the procedure is simple, it is not stupid – Many seemingly smarter heuristics can give a far worse performance in the worst case. For example, why not modify the above procedure to select only one of the two vertices for the cover instead of both. After all, the selected edge will be equally well covered by only one vertex. However, consider the star-shaped graph of Figure 9.9. This heuristic will produce a two-vertex cover, while the singlevertex heuristic can return a cover as large as n − 1 vertices, should we get unlucky and repeatedly select the leaf instead of the center as the cover vertex we retain. • Greedy isn’t always the answer – Perhaps the most natural heuristic for vertex cover would repeatedly select and delete the vertex of highest remaining degree for the vertex cover. After all, this vertex will cover the largest number 345 346 9. INTRACTABLE PROBLEMS AND APPROXIMATION ALGORITHMS v u w Figure 9.10: The triangle inequality (d(u, w) ≤ d(u, v) + d(v, w)) typically holds in geometric and weighted graph problems of possible edges. However, in the case of ties or near ties, this heuristic can go seriously astray. In the worst case, it can yield a cover that is Θ(lg n) times optimal. • Making a heuristic more complicated does not necessarily make it better – It is easy to complicate heuristics by adding more special cases or details. For example, the procedure above does not specify which edge should be selected next. It might seem reasonable to next select the edge whose endpoints have the highest degree. However, this does not improve the worst-case bound and just makes it more diﬃcult to analyze. • A postprocessing cleanup step can’t hurt – The ﬂip side of designing simple heuristics is that they can often be modiﬁed to yield better-in-practice solutions without weakening the approximation bound. For example, a postprocessing step that deletes any unnecessary vertex from the cover can only improve things in practice, even though it won’t help the worst-case bound. The important property of approximation algorithms is relating the size of the solution produced directly to a lower bound on the optimal solution. Instead of thinking about how well we might do, we must think about the worst case—i.e. , how badly we might perform. 9.10.2 The Euclidean Traveling Salesman In most natural applications of the traveling salesman problem, direct routes are inherently shorter than indirect routes. For example, if a graph’s edge weights were the straight-line distances between pairs of cities, the shortest path from x to y will always be “as the crow ﬂies.” The edge weights induced by Euclidean geometry satisfy the triangle inequality, namely that d(u, w) ≤ d(u, v) + d(v, w) for all triples of vertices u, v, and w. The general reasonableness of this condition is demonstrated in Figure 9.10. The cost of airfare is an example of a distance function that violates the triangle inequality, 9.10 DEALING WITH NP-COMPLETE PROBLEMS 1 2 4 3 7 5 6 10 8 11 9 Figure 9.11: A depth-ﬁrst traversal of a spanning tree, with the shortcut tour since it is sometimes cheaper to ﬂy through an intermediate city than to ﬂy to the destination directly. TSP remains hard when the distances are Euclidean distances in the plane. We can approximate the optimal traveling salesman tour using minimum spanning trees or graphs that obey the triangle inequality. First, observe that the weight of a minimum spanning tree is a lower bound on the cost of the optimal tour. Why? Deleting any edge from a tour leaves a path, the total weight of which must be no greater than that of the original tour. This path has no cycles, and hence is a tree, which means its weight is at least that of the minimum spanning tree. Thus the weight of the minimum spanning tree gives a lower bound on the optimal tour. Consider now what happens in performing a depth-ﬁrst traversal of a spanning tree. We will visit each edge twice, once going down the tree when discovering the edge and once going up after exploring the entire subtree. For example, in the depth-ﬁrst search of Figure 9.11, we visit the vertices in order 1 − 2 − 1 − 3 − 5 − 8 − 5 − 9 − 5 − 3 − 6 − 3 − 1 − 4 − 7 − 10 − 7 − 11 − 7 − 4 − 1, thus using every tree edge exactly twice. This tour repeats each edge of the minimum spanning tree twice, and hence costs at most twice the optimal tour. However, vertices will be repeated on this depth-ﬁrst search tour. To remove the extra vertices, we can take a shortest path to the next unvisited vertex at each step. The shortcut tour for the tree above is 1 − 2 − 3 − 5 − 8 − 9 − 6 − 4 − 7 − 10 − 11 − 1. Because we have replaced a chain of edges by a single direct edge, the triangle inequality ensures that the tour can only get shorter. Thus, this shortcut tour is also within weight and twice that of optimal. Better, more complicated approximation algorithms for Euclidean TSP exist, as described in Section 16.4 (page 533). No approximation algorithms are known for TSPs that do not satisfy the triangle inequality. 347 348 9. INTRACTABLE PROBLEMS AND APPROXIMATION ALGORITHMS 9.10.3 Maximum Acyclic Subgraph Directed acyclic graphs (DAGs) are easier to work with than general digraphs. Sometimes it is useful to simplify a given graph by deleting a small set of edges or vertices that suﬃce to break all cycles. Such feedback set problems are discussed in Section 16.11 (page 559). Here we consider an interesting problem in this class where seek to retain as many edges as possible while breaking all directed cycles: Problem: Maximum Directed Acyclic Subgraph Input: A directed graph G = (V, E). Output: Find the largest possible subset E ∈ E such that G = (V, E ) is acyclic. In fact, there is a very simple algorithm that guarantees you a solution with at least half as many edges as optimum. I encourage you to try to ﬁnd it now before peeking. Construct any permutation of the vertices, and interpret it as a left-right ordering akin to topological sorting. Now some of the edges will point from left to right, while the remainder point from right to left. One of these two edge subsets must be at least as large as the other. This means it contains at least half the edges. Furthermore each of these two edge subsets must be acyclic for the same reason only DAGs can be topologically sorted—you cannot form a cycle by repeatedly moving in one direction. Thus, the larger edge subset must be acyclic and contain at least half the edges of the optimal solution! This approximation algorithm is simple to the point of almost being stupid. But note that heuristics can make it perform better in practice without losing this guarantee. Perhaps we can try many random permutations, and pick the best. Or we can try to exchange pairs of vertices in the permutations retaining those swaps, which throw more edges onto the bigger side. 9.10.4 Set Cover The previous sections may encourage a false belief that every problem can be approximated to within a factor of two. Indeed, several catalog problems such as maximum clique cannot be approximated to any interesting factor. Set cover occupies a middle ground between these extremes, having a factorΘ(lg n) approximation algorithm. Set cover is a more general version of the vertex cover problem. As deﬁned in Section 18.1 (page 621), Problem: Set Cover Input: A collection of subsets S = {S1 , . . . , Sm } of the universal set U = {1, . . . , n}. Output: What is the smallest subset T of S whose union equals the universal set— |T | i.e. , ∪i=1 Ti = U ? 9.10 milestone class uncovered elements selected subset size 6 64 13 5 51 11 349 DEALING WITH NP-COMPLETE PROBLEMS 40 10 30 5 25 3 4 22 3 19 3 16 3 13 3 3 10 3 7 3 2 4 2 Figure 9.12: The coverage process for greedy on a particular instance of set cover The natural heuristic is greedy. Repeatedly select the subset that covers the largest collection of thus-far uncovered elements until everything is covered. In pseudocode, SetCover(S) While (U = ∅) do: Identify the subset Si with the largest intersection with U Select Si for the set cover U = U − Si One consequence of this selection process is that the number of freshly-covered elements deﬁnes a nonincreasing sequence as the algorithm proceeds. Why? If not, greedy would have picked the more powerful subset earlier if it, in fact, existed. Thus we can view this heuristic as reducing the number of uncovered elements from n down to zero by progressively smaller amounts. A trace of such an execution is shown in Figure 9.12. An important milestone in such a trace occurs each time the number of remaining uncovered elements reduces past a power of two. Clearly there can be at most lg n such events. Let wi denote the number of subsets that were selected by the heuristic to cover elements between milestones 2i+1 − 1 and 2i . Deﬁne the width w to be the maximum wi , where 0 ≤ i ≤ lg2 n. In the example of Figure 9.12, the maximum width is given by the ﬁve subsets needed to go from 25 − 1 down to 24 . Since there are at most lg n such milestones, the solution produced by the greedy heuristic must contain at most w · lg n subsets. But I claim that the optimal solution must contain at least w subsets, so the heuristic solution is no worse than lg n times optimal. Why? Consider the average number of new elements covered as we move between milestones 2i+1 − 1 and 2i . These 2i elements require wi subsets, so the average coverage is μi = 2i /wi . More to the point, the last/smallest of these subsets covers at most μi subsets. Thus, no subset exists in S that can cover more than μi of the remaining 2i elements. So, to ﬁnish the job, we need at least 2i /μi = wi subsets. The somewhat surprising thing is that there do exist set cover instances where this heuristic takes Ω(lg n) times optimal. The logarithmic factor is a property of the problem/heuristic, not an artifact of weak analysis. 1 2 1 0 1 1 350 9. INTRACTABLE PROBLEMS AND APPROXIMATION ALGORITHMS Take-Home Lesson: Approximation algorithms guarantee answers that are always close to the optimal solution. They can provide a practical approach to dealing with NP-complete problems. Chapter Notes The notion of NP-completeness was ﬁrst developed by Cook [Coo71]. Satisﬁability really is a $1,000,000 problem, and the Clay Mathematical Institute has oﬀered such a prize to any person who resolves the P=NP question. See http://www.claymath.org/ for more on the problem and the prize. Karp [Kar72] showed the importance of Cook’s result by providing reductions from satisﬁability to more than 20 important algorithmic problems. I recommend Karp’s paper for its sheer beauty and economy—he reduces each reduction to three line descriptions showing the problem equivalence. Together, these provided the tools to resolve the complexity of literally hundreds of important problems where no eﬃcient algorithms were known. The best introduction to the theory of NP-completeness remains Garey and Johnson’s book Computers and Intractability. It introduces the general theory, including an accessible proof of Cook’s theorem [Coo71] that satisﬁability is as hard as anything in NP. They also provide an essential reference catalog of more than 300 NP-complete problems, which is a great resource for learning what is known about the most interesting hard problems. The reductions claimed, but missing from this chapter can be found in Garey and Johnson, or textbooks such as [CLRS01]. A few catalog problems exist in a limbo state where it is not known whether the problem has a fast algorithm or is NP-complete. The most prominent of these are graph isomorphism (see Section 16.9 (page 550)) and integer factorization (see Section 13.8 (page 420)). That this limbo list is so short is quite a tribute to the state of the art in algorithm design and the power of NP-completeness. For almost every important problem we either have a fast algorithm or a good solid reason for why one doesn’t exist. The war story problem on unidirected subgraph was originally proven hard in [Mah76]. 9.11 Exercises Transformations and Satisﬁability 9-1. [2] Give the 3-SAT formula that results from applying the reduction of SAT to 3-SAT for the formula: (x + y + z + w + u + v) · (x + y + z + w + u + v) · (x + y + z + w + u + v) · (x + y) 9.11 EXERCISES 9-2. [3] Draw the graph that results from the reduction of 3-SAT to vertex cover for the expression (x + y + z) · (x + y + z) · (x + y + z) · (x + y + x) 9-3. [4] Suppose we are given a subroutine which can solve the traveling salesman decision problem of page 318 in, say, linear time. Give an eﬃcient algorithm to ﬁnd the actual TSP tour by making a polynomial number of calls to this subroutine. 9-4. [7] Implement a translator that translates satisﬁability instances into equivalent 3-SAT instances. 9-5. [7] Design and implement a backtracking algorithm to test whether a set of formulae are satisﬁable. What criteria can you use to prune this search? 9-6. [8] Implement the vertex cover to satisﬁability reduction, and run the resulting clauses through a satisﬁability testing code. Does this seem like a practical way to compute things? Basic Reductions 9-7. [4] An instance of the set cover problem consists of a set X of n elements, a family F of subsets of X, and an integer k. The question is, does there exist k subsets from F whose union is X? For example, if X = {1, 2, 3, 4} and F = {{1, 2}, {2, 3}, {4}, {2, 4}}, there does not exist a solution for k = 2, but there does for k = 3 (for example, {1, 2}, {2, 3}, {4}). Prove that set cover is NP-complete with a reduction from vertex cover. 9-8. [4] The baseball card collector problem is as follows. Given packets P1 , . . . , Pm , each of which contains a subset of this year’s baseball cards, is it possible to collect all the year’s cards by buying ≤ k packets? For example, if the players are {Aaron, M ays, Ruth, Skiena} and the packets are {{Aaron, M ays}, {M ays, Ruth}, {Skiena}, {M ays, Skiena}}, there does not exist a solution for k = 2, but there does for k = 3, such as {Aaron, M ays}, {M ays, Ruth}, {Skiena} Prove that the baseball card collector problem is NP-hard using a reduction from vertex cover. 9-9. [4] The low-degree spanning tree problem is as follows. Given a graph G and an integer k, does G contain a spanning tree such that all vertices in the tree have degree at most k (obviously, only tree edges count towards the degree)? For example, in the following graph, there is no spanning tree such that all vertices have a degree less than three. (a) Prove that the low-degree spanning tree problem is NP-hard with a reduction from Hamiltonian path. (b) Now consider the high-degree spanning tree problem, which is as follows. Given a graph G and an integer k, does G contain a spanning tree whose highest degree vertex is at least k? In the previous example, there exists a spanning tree with a highest degree of 8. Give an eﬃcient algorithm to solve the highdegree spanning tree problem, and an analysis of its time complexity. 351 352 9. INTRACTABLE PROBLEMS AND APPROXIMATION ALGORITHMS 9-10. [4] Show that the following problem is NP-complete: Problem: Dense subgraph Input: A graph G, and integers k and y. Output: Does G contain a subgraph with exactly k vertices and at least y edges? 9-11. [4] Show that the following problem is NP-complete: Problem: Clique, no-clique Input: An undirected graph G = (V, E) and an integer k. Output: Does G contain both a clique of size k and an independent set of size k. 9-12. [5] An Eulerian cycle is a tour that visits every edge in a graph exactly once. An Eulerian subgraph is a subset of the edges and vertices of a graph that has an Eulerian cycle. Prove that the problem of ﬁnding the number of edges in the largest Eulerian subgraph of a graph is NP-hard. (Hint: the Hamiltonian circuit problem is NP-hard even if each vertex in the graph is incident upon exactly three edges.) Creative Reductions 9-13. [5] Prove that the following problem is NP-complete: Problem: Hitting Set Input: A collection C of subsets of a set S, positive integer k. Output: Does S contain a subset S such that |S | ≤ k and each subset in C contains at least one element from S ? 9-14. [5] Prove that the following problem is NP-complete: Problem: Knapsack Input: A set S of n items, such that the ith item has value vi and weight wi . Two positive integers: weight limit W and value requirement V . Output: Does there exist a subset S ∈ S such that i∈S wi ≤ W and i∈S vi ≥ V ? (Hint: start from integer partition.) 9-15. [5] Prove that the following problem is NP-complete: Problem: Hamiltonian Path Input: A graph G, and vertices s and t. 9.11 EXERCISES Output: Does G contain a path which starts from s, ends at t, and visits all vertices without visiting any vertex more than once? (Hint: start from Hamiltonian cycle.) 9-16. [5] Prove that the following problem is NP-complete: Problem: Longest Path Input: A graph G and positive integer k. Output: Does G contain a path that visits at least k diﬀerent vertices without visiting any vertex more than once? 9-17. [6] Prove that the following problem is NP-complete: Problem: Dominating Set Input: A graph G = (V, E) and positive integer k. Output: Is there a subset V ∈ V such that |V | ≤ k where for each vertex x ∈ V either x ∈ V or there exists an edge (x, y), where y ∈ V . 9-18. [7] Prove that the vertex cover problem (does there exist a subset S of k vertices in a graph G such that every edge in G is incident upon at least one vertex in S?) remains NP-complete even when all the vertices in the graph are restricted to have even degrees. 9-19. [7] Prove that the following problem is NP-complete: Problem: Set Packing Input: A collection C of subsets of a set S, positive integer k. Output: Does S contain at least k disjoint subsets (i.e. , such that none of these subsets have any elements in common?) 9-20. [7] Prove that the following problem is NP-complete: Problem: Feedback Vertex Set Input: A directed graph G = (V, A) and positive integer k. Output: Is there a subset V ∈ V such that |V | ≤ k, such that deleting the vertices of V from G leaves a DAG? Algorithms for Special Cases 9-21. [5] A Hamiltonian path P is a path that visits each vertex exactly once. The problem of testing whether a graph G contains a Hamiltonian path is NP-complete. There does not have to be an edge in G from the ending vertex to the starting vertex of P , unlike in the Hamiltonian cycle problem. Give an O(n + m)-time algorithm to test whether a directed acyclic graph G (a DAG) contains a Hamiltonian path. (Hint: think about topological sorting and DFS.) 9-22. [8] The 2-SAT problem is, given a Boolean formula in 2-conjunctive normal form (CNF), to decide whether the formula is satisﬁable. 2-SAT is like 3-SAT, except 353 354 9. INTRACTABLE PROBLEMS AND APPROXIMATION ALGORITHMS that each clause can have only two literals. For example, the following formula is in 2-CNF: (x1 ∨ x2 ) ∧ (x̄2 ∨ x3 ) ∧ (x1 ∨ x̄3 ) Give a polynomial-time algorithm to solve 2-SAT. P=NP? 9-23. [4] Show that the following problems are in NP: • Does graph G have a simple path (i.e. , with no vertex repeated) of length k? • Is integer n composite (i.e. , not prime)? • Does graph G have a vertex cover of size k? 9-24. [7] It was a long open question whether the decision problem “Is integer n a composite number, in other words, not prime?” can be computed in time polynomial in the size of its input. Why doesn’t the following algorithm suﬃce to prove it is in P, since it runs in O(n) time? PrimalityTesting(n) composite := false for i := 2 to n − 1 do if (n mod i) = 0 then composite := true Approximation Algorithms 9-25. [4] In the maximum-satisﬁability problem, we seek a truth assignment that satisﬁes as many clauses as possible. Give an heuristic that always satisﬁes at least half as many clauses as the optimal solution. 9-26. [5] Consider the following heuristic for vertex cover. Construct a DFS tree of the graph, and delete all the leaves from this tree. What remains must be a vertex cover of the graph. Prove that the size of this cover is at most twice as large as optimal. 9-27. [5] The maximum cut problem for a graph G = (V, E) seeks to partition the vertices V into disjoint sets A and B so as to maximize the number of edges (a, b) ∈ E such that a ∈ A and b ∈ B. Consider the following heuristic for max cut. First assign v1 to A and v2 to B. For each remaining vertex, assign it to the side that adds the most edges to the cut. Prove that this cut is at least half as large as the optimal cut. 9-28. [5] In the bin-packing problem, we are given n items with weights w1 , w2 , ..., wn , respectively. Our goal is to ﬁnd the smallest number of bins that will hold the n objects, where each bin has capacity of at most one kilogram. The ﬁrst-ﬁt heuristic considers the objects in the order in which they are given. For each object, place it into ﬁrst bin that has room for it. If no such bin exists, start a new bin. Prove that this heuristic uses at most twice as many bins as the optimal solution. 9-29. [5] For the ﬁrst-ﬁt heuristic described just above, give an example where the packing it ﬁts uses at least 5/3 times as many bins as optimal. 9.11 EXERCISES 9-30. [5] A vertex coloring of graph G = (V, E) is an assignment of colors to vertices of V such that each edge (x, y) implies that vertices x and y are assigned diﬀerent colors. Give an algorithm for vertex coloring G using at most Δ + 1 colors, where Δ is the maximum vertex degree of G. Programming Challenges These programming challenge problems with robot judging are available at http://www.programming-challenges.com or http://online-judge.uva.es. Geometry 9-1. “The Monocycle” – Programming Challenges 111202, UVA Judge 10047. 9-2. “Dog and Gopher” – Programming Challenges 10310, UVA Judge 111301. 9-3. “Chocolate Chip Cookies” – Programming Challenges 111304, UVA Judge 10136. 9-4. “Birthday Cake” – Programming Challenges 111305, UVA Judge 10167. Computational Geometry 9-5. “Closest Pair Problem” – Programming Challenges 111402, UVA Judge 10245. 9-6. “Chainsaw Massacre” – Programming Challenges 111403, UVA Judge 10043. 9-7. “Hotter Colder” – Programming Challenges 111404, UVA Judge 10084. 9-8. “Useless Tile Packers” – Programming Challenges 111405, UVA Judge 10065. Note: These are not particularly relevant to NP-completeness, but are added for completeness. 355 10 How to Design Algorithms Designing the right algorithm for a given application is a major creative act—that of taking a problem and pulling a solution out of the ether. The space of choices you can make in algorithm design is enormous, leaving you plenty of freedom to hang yourself. This book is designed to make you a better algorithm designer. The techniques presented in Part I of this book provide the basic ideas underlying all combinatorial algorithms. The problem catalog of Part II will help you with modeling your application, and tell you what is known about the relevant problems. However, being a successful algorithm designer requires more than book knowledge. It requires a certain attitude—the right problem-solving approach. It is diﬃcult to teach this mindset in a book, yet getting it is essential to becoming a successful designer. The key to algorithm design (or any other problem-solving task) is to proceed by asking yourself questions to guide your thought process. What if we do this? What if we do that? Should you get stuck on the problem, the best thing to do is move onto the next question. In any group brainstorming session, the most useful person in the room is the one who keeps asking, “Why can’t we do it this way?” not the person who later tells them why, because she will eventually stumble on an approach that can’t be shot down. Towards this end, we provide a sequence of questions to guide your search for the right algorithm for your problem. To use it eﬀectively, you must not only ask the questions, but answer them. The key is working through the answers carefully by writing them down in a log. The correct answer to “Can I do it this way?” is never “no,” but “no, because. . . .” By clearly articulating your reasoning as to why something doesn’t work, you can check whether you have glossed over a possibility that you didn’t want to think hard enough about. It is amazing how often the reason S.S. Skiena, The Algorithm Design Manual, 2nd ed., DOI: 10.1007/978-1-84800-070-4 10, c Springer-Verlag London Limited 2008 10. HOW TO DESIGN ALGORITHMS you can’t ﬁnd a convincing explanation for something is because your conclusion is wrong. The distinction between strategy and tactics is important to keep aware of during any design process. Strategy represents the quest for the big picture—the framework around which we construct our path to the goal. Tactics are used to win the minor battles we must ﬁght along the way. In problem-solving, it is important to check repeatedly whether you are thinking on the right level. If you do not have a global strategy of how you are going to attack your problem, it is pointless to worry about the tactics. An example of a strategic question is “Can I model my application as a graph algorithm problem?” A tactical question might be, “Should I use an adjacency list or adjacency matrix data structure to represent my graph?” Of course, such tactical decisions are critical to the ultimate quality of the solution, but they can be properly evaluated only in light of a successful strategy. Too many people freeze up in their thinking when faced with a design problem. After reading or hearing the problem, they sit down and realize that they don’t know what to do next. Avoid this fate. Follow the sequence of questions provided below and in most of the catalog problem sections. We’ll tell you what to do next! Obviously, the more experience you have with algorithm design techniques such as dynamic programming, graph algorithms, intractability, and data structures, the more successful you will be at working through the list of questions. Part I of this book has been designed to strengthen this technical background. However, it pays to work through these questions regardless of how strong your technical skills are. The earliest and most important questions on the list focus on obtaining a detailed understanding of your problem and do not require speciﬁc expertise. This list of questions was inspired by a passage in [Wol79]—a wonderful book about the space program entitled The Right Stuﬀ. It concerned the radio transmissions from test pilots just before their planes crashed. One might have expected that they would panic, so ground control would hear the pilot yelling Ahhhhhhhhhhh —, terminated only by the sound of smacking into a mountain. Instead, the pilots ran through a list of what their possible actions could be. I’ve tried the ﬂaps. I’ve checked the engine. Still got two wings. I’ve reset the —. They had “the Right Stuﬀ.” Because of this, they sometimes managed to miss the mountain. I hope this book and list will provide you with “the Right Stuﬀ” to be an algorithm designer. And I hope it prevents you from smacking into any mountains along the way. 1. Do I really understand the problem? (a) What exactly does the input consist of? (b) What exactly are the desired results or output? (c) Can I construct an input example small enough to solve by hand? What happens when I try to solve it? (d) How important is it to my application that I always ﬁnd the optimal answer? Can I settle for something close to the optimal answer? 357 358 10. HOW TO DESIGN ALGORITHMS (e) How large is a typical instance of my problem? Will I be working on 10 items? 1,000 items? 1,000,000 items? (f) How important is speed in my application? Must the problem be solved within one second? One minute? One hour? One day? (g) How much time and eﬀort can I invest in implementation? Will I be limited to simple algorithms that can be coded up in a day, or do I have the freedom to experiment with a couple of approaches and see which is best? (h) Am I trying to solve a numerical problem? A graph algorithm problem? A geometric problem? A string problem? A set problem? Which formulation seems easiest? 2. Can I ﬁnd a simple algorithm or heuristic for my problem? (a) Will brute force solve my problem correctly by searching through all subsets or arrangements and picking the best one? i. If so, why am I sure that this algorithm always gives the correct answer? ii. How do I measure the quality of a solution once I construct it? iii. Does this simple, slow solution run in polynomial or exponential time? Is my problem small enough that this brute-force solution will suﬃce? iv. Am I certain that my problem is suﬃciently well deﬁned to actually have a correct solution? (b) Can I solve my problem by repeatedly trying some simple rule, like picking the biggest item ﬁrst? The smallest item ﬁrst? A random item ﬁrst? i. If so, on what types of inputs does this heuristic work well? Do these correspond to the data that might arise in my application? ii. On what types of inputs does this heuristic work badly? If no such examples can be found, can I show that it always works well? iii. How fast does my heuristic come up with an answer? Does it have a simple implementation? 3. Is my problem in the catalog of algorithmic problems in the back of this book? (a) What is known about the problem? Is there an implementation available that I can use? (b) Did I look in the right place for my problem? Did I browse through all pictures? Did I look in the index under all possible keywords? 10. HOW TO DESIGN ALGORITHMS (c) Are there relevant resources available on the World Wide Web? Did I do a Google web and Google Scholar search? Did I go to the page associated with this book, http://www.cs.sunysb.edu/∼algorith? 4. Are there special cases of the problem that I know how to solve? (a) Can I solve the problem eﬃciently when I ignore some of the input parameters? (b) Does the problem become easier to solve when I set some of the input parameters to trivial values, such as 0 or 1? (c) Can I simplify the problem to the point where I can solve it eﬃciently? (d) Why can’t this special-case algorithm be generalized to a wider class of inputs? (e) Is my problem a special case of a more general problem in the catalog? 5. Which of the standard algorithm design paradigms are most relevant to my problem? (a) Is there a set of items that can be sorted by size or some key? Does this sorted order make it easier to ﬁnd the answer? (b) Is there a way to split the problem in two smaller problems, perhaps by doing a binary search? How about partitioning the elements into big and small, or left and right? Does this suggest a divide-and-conquer algorithm? (c) Do the input objects or desired solution have a natural left-to-right order, such as characters in a string, elements of a permutation, or leaves of a tree? Can I use dynamic programming to exploit this order? (d) Are there certain operations being done repeatedly, such as searching, or ﬁnding the largest/smallest element? Can I use a data structure to speed up these queries? What about a dictionary/hash table or a heap/priority queue? (e) Can I use random sampling to select which object to pick next? What about constructing many random conﬁgurations and picking the best one? Can I use some kind of directed randomness like simulated annealing to zoom in on the best solution? (f) Can I formulate my problem as a linear program? How about an integer program? (g) Does my problem seem something like satisﬁability, the traveling salesman problem, or some other NP-complete problem? Might the problem be NP-complete and thus not have an eﬃcient algorithm? Is it in the problem list in the back of Garey and Johnson [GJ79]? 359 360 10. HOW TO DESIGN ALGORITHMS 6. Am I still stumped? (a) Am I willing to spend money to hire an expert to tell me what to do? If so, check out the professional consulting services mentioned in Section 19.4 (page 664). (b) Why don’t I go back to the beginning and work through these questions again? Did any of my answers change during my latest trip through the list? Problem-solving is not a science, but part art and part skill. It is one of the skills most worth developing. My favorite book on problem-solving remains Pólya’s How to Solve It [P5́7], which features a catalog of problem-solving techniques that are fascinating to browse through, both before and after you have a problem. 11 A Catalog of Algorithmic Problems This is a catalog of algorithmic problems that arise commonly in practice. It describes what is known about them and gives suggestions about how best to proceed if the problem arises in your application. What is the best way to use this catalog? First, think about your problem. If you recall the name, look up the catalog entry in the index or table of contents and start reading. Read through the entire entry, since it contains pointers to other relevant problems. Leaf through the catalog, looking at the pictures and problem names to see if something strikes a chord. Don’t be afraid to use the index, for every problem in the book is listed there under several possible keywords and applications. If you still don’t ﬁnd something relevant, your problem is either not suitable for attack by combinatorial algorithms or else you don’t fully understand it. In either case, go back to step one. The catalog entries contain a variety of diﬀerent types of information that have never been collected in one place before. Diﬀerent ﬁelds in each entry present information of practical and historical interest. To make this catalog more easily accessible, we introduce each problem with a pair of graphics representing the problem instance or input on the left and the result of solving the problem in this instance on the right. We have invested considerable thought in creating stylized examples that illustrate desired behaviors more than just deﬁnitions. For example, the minimum spanning tree example illustrates how points can be clustered using minimum spanning trees. We hope that people will be able to ﬂip through the pictures and identify which problems might be relevant to them. We augment these pictures with more formal written input and problem descriptions to eliminate the ambiguity inherent in a purely pictorial representation. S.S. Skiena, The Algorithm Design Manual, 2nd ed., DOI: 10.1007/978-1-84800-070-4 11, c Springer-Verlag London Limited 2008 364 11. A CATALOG OF ALGORITHMIC PROBLEMS Once you have identiﬁed your problem, the discussion section tells you what you should do about it. We describe applications where the problem is likely to arise and the special issues associated with data from them. We discuss the kind of results you can hope for or expect and, most importantly, what you should do to get them. For each problem, we outline a quick-and-dirty algorithm and provide pointers to more powerful algorithms to try next if the ﬁrst attempt is not suﬃcient. For each problem, we identify available software implementations that are discussed in the implementation ﬁeld of each entry. Many of these routines are quite good, and they can perhaps be plugged directly into your application. Others maybe inadequate for production use, but they hopefully can provide a good model for your own implementation. In general, the implementations are listed in order of descending usefulness, but we will explicitly recommend the best one available for each problem if a clear winner exists. More detailed information for many of these implementations appears in Chapter 19. Essentially all of the implementations are available via the WWW site associated with this book—reachable at http://www.cs.sunysb.edu/∼algorith. Finally, in deliberately smaller print, we discuss the history of each problem and present results of primarily theoretical interest. We have attempted to report the best results known for each problem and point out empirical comparisons of algorithms or survey articles if they exist. This should be of interest to students and researchers, and also to practitioners for whom our recommended solutions prove inadequate and need to know if anything better is possible. Caveats This is a catalog of algorithmic problems. It is not a cookbook. It cannot be because there are too many recipes and too many possible variations on what people want to eat. My goal is to point you in the right direction so that you can solve your own problems. I try to identify the issues you will encounter along the way—problems that you will have to work out for yourself. In particular: • For each problem, I suggest algorithms and directions to attack it. These recommendations are based on my experiences, and are aimed toward what I see as typical applications. I felt it was more important to make concrete recommendations for the masses rather than to try to cover all possible situations. If you don’t agree with my advice, don’t follow it, but before you ignore me, try to understand the reasoning behind my recommendations and articulate a reason why your application violates my assumptions. • The implementations I recommend are not necessarily complete solutions to your problem. I point to an implementation whenever I feel it might be more useful to someone than just a textbook description of the algorithm. Some programs are useful only as models for you to write your own codes. Others are embedded in large systems and so might be too painful to extract and run 11. A CATALOG OF ALGORITHMIC PROBLEMS on their own. Assume that all of them contain bugs. Many are quite serious, so beware. • Please respect the licensing conditions for any implementations you use commercially. Many of these codes are not open source and have licence restrictions. See Section 19.1 for a further discussion of this issue. • I would be interested in hearing about your experiences with my recommendations, both positive and negative. I would be especially interested in learning about any other implementations that you know about. Feel free to drop me a line at feedback@algorist.com. 365 12 Data Structures Data structures are not so much algorithms as they are the fundamental constructs around which you build your application. Becoming ﬂuent in what the standard data structures can do for you is essential to get full value from them. This puts data structures slightly out of sync with the rest of the catalog. Perhaps the most useful aspect of it will be the pointers to various implementations and data structure libraries. Many of these data structures are nontrivial to implement well, so the programs we point to will be useful as models even if they do not do exactly what you need. Certain fundamental data structures, like kd-trees and suﬃx trees, are not as well known as they should be. Hopefully, this catalog will serve to better publicize them. There are a large number of books on elementary data structures available. My favorites include: • Sedgewick [Sed98] – This comprehensive introduction to algorithms and data structures stands out for the clever and beautiful images of algorithms in action. It comes in C, C++, and Java editions. • Weiss [Wei06] – A nice text, emphasizing data structures more than algorithms. Comes in Java, C, C++, and Ada editions. • Goodrich and Tamassia [GT05] – The Java edition makes particularly good use of the author’s Java Data Structures Library (JDSL). The Handbook of Data Structures and Applications [MS05] provides a comprehensive and up-to-date survey of research in data structures. The student who took only an elementary course in data structures is likely to be impressed and surprised by the volume of recent work on the subject. S.S. Skiena, The Algorithm Design Manual, 2nd ed., DOI: 10.1007/978-1-84800-070-4 12, c Springer-Verlag London Limited 2008 12.1 Query ? DICTIONARIES . .. . quelling quenching queried queries query quest . .. . INPUT 12.1 OUTPUT Dictionaries Input description: A set of n records, each identiﬁed by one or more key ﬁelds. Problem description: Build and maintain a data structure to eﬃciently locate, insert, and delete the record associated with any query key q. Discussion: The abstract data type “dictionary” is one of the most important structures in computer science. Dozens of data structures have been proposed for implementing dictionaries, including hash tables, skip lists, and balanced/unbalanced binary search trees. This means that choosing the best one can be tricky. It can signiﬁcantly impact performance. However, in practice, it is more important to avoid using a bad data structure than to identify the single best option available. An essential piece of advice is to carefully isolate the implementation of the dictionary data structure from its interface. Use explicit calls to methods or subroutines that initialize, search, and modify the data structure, rather than embedding them within the code. This leads to a much cleaner program, but it also makes it easy to experiment with diﬀerent implementations to see how they perform. Do not obsess about the procedure call overhead inherent in such an abstraction. If your application is so time-critical that such overhead can impact performance, then it is even more essential that you be able to identify the right dictionary implementation. 367 368 12. DATA STRUCTURES In choosing the right data structure for your dictionary, ask yourself the following questions: • How many items will you have in your data structure? – Will you know this number in advance? Are you looking at a problem small enough that a simple data structure will suﬃce, or one so large that we must worry about running out of memory or virtual memory performance? • Do you know the relative frequencies of insert, delete, and search operations? – Static data structures (like sorted arrays) suﬃce in applications when there are no modiﬁcations to the data structure after it is ﬁrst constructed. Semidynamic data structures, which support insertion but not deletion, can have signiﬁcantly simpler implementations than fully dynamic ones. • Can we assume that the access pattern for keys will be uniform and random? – Search queries exhibit a skewed access distribution in many applications, meaning certain elements are much more popular than others. Further, queries often have a sense of temporal locality, meaning elements are likely to be repeatedly accessed in clusters instead of at fairly regular intervals. Certain data structures (such as splay trees) can take advantage of a skewed and clustered universe. • Is it critical that individual operations be fast, or only that the total amount of work done over the entire program be minimized? – When response time is critical, such as in a program controlling a heart-lung machine, you can’t wait too long between steps. When you have a program that is doing a lot of queries over the database, such as identifying all criminals who happen to be politicians, it is not as critical that you pick out any particular congressman quickly as it is that you get them all with the minimum total eﬀort. An object-oriented generation has emerged as no more likely to write a container class than ﬁx the engine in their car. This is good; default containers should do just ﬁne for most applications. Still, it is good sometimes to know what you have under the hood: • Unsorted linked lists or arrays – For small data sets, an unsorted array is probably the easiest data structure to maintain. Linked structures can have terrible cache performance compared with sleek, compact arrays. However, once your dictionary becomes larger than (say) 50 to 100 items, the linear search time will kill you for either lists or arrays. Details of elementary dictionary implementations appear in Section 3.3 (page 72). A particularly interesting and useful variant is the self-organizing list. Whenever a key is accessed or inserted, we always move it to head of the list. Thus, if the key is accessed again sometime in the near future, it will be near the front and so require only a short search to ﬁnd it. Most applications exhibit 12.1 DICTIONARIES both uneven access frequencies and locality of reference, so the average time for a successful search in a self-organizing list is typically much better than in a sorted or unsorted list. Of course, self-organizing data structures can be built from arrays as well as linked lists and trees. • Sorted linked lists or arrays – Maintaining a sorted linked list is usually not worth the eﬀort unless you are trying to eliminate duplicates, since we cannot perform binary searches in such a data structure. A sorted array will be appropriate if and only if there are not many insertions or deletions. • Hash tables – For applications involving a moderate-to-large number of keys (say between 100 and 10,000,000), a hash table is probably the right way to go. We use a function that maps keys (be they strings, numbers, or whatever) to integers between 0 and m − 1. We maintain an array of m buckets, each typically implemented using an unsorted linked list. The hash function immediately identiﬁes which bucket contains a given key. If we use a hash function that spreads the keys out nicely, and a suﬃciently large hash table, each bucket should contain very few items, thus making linear searches acceptable. Insertion and deletion from a hash table reduce to insertion and deletion from the bucket/list. Section 3.7 (page 89) provides a more detailed discussion of hashing and its applications. A well-tuned hash table will outperform a sorted array in most applications. However, several design decisions go into creating a well-tuned hash table: – How do I deal with collisions?: Open addressing can lead to more concise tables with better cache performance than bucketing, but performance will be more brittle as the load factor (ratio of occupancy to capacity) of the hash table starts to get high. – How big should the table be?: With bucketing, m should about the same as the maximum number of items you expect to put in the table. With open addressing, make it (say) 30% larger or more. Selecting m to be a prime number minimizes the dangers of a bad hash function. – What hash function should I use?: For strings, something like H(S, j) = m−1 αm−(i+1) × char(si+j ) mod m i=0 should work, where α is the size of the alphabet and char(x) is the function that maps each character x to its ASCII character code. Use Horner’s rule (or precompute values of αx ) to implement this hash function computation eﬃciently, as discussed in Section 13.9 (page 423). This hash function has the nifty property that H(S, j + 1) = (H(S, j) − αm−1 char(sj ))α + sj+m 369 370 12. DATA STRUCTURES so hash codes of successive m-character windows of a string can be computed in constant time instead of O(m). Regardless of which hash function you decide to use, print statistics on the distribution of keys per bucket to see how uniform it really is. Odds are the ﬁrst hash function you try will not prove to be the best. Botching up the hash function is an excellent way to slow down any application. • Binary search trees – Binary search trees are elegant data structures that support fast insertions, deletions, and queries. They are reviewed in Section 3.4 (page 77). The big distinction between diﬀerent types of trees is whether they are explicitly rebalanced after insertion or deletion, and how this rebalancing is done. In random search trees, we simply insert a node at the leaf position where we can ﬁnd it and no rebalancing takes place. Although such trees perform well under random insertions, most applications are not really random. Indeed, unbalanced search trees constructed by inserting keys in sorted order are a disaster, performing like a linked list. Balanced search trees use local rotation operations to restructure search trees, moving more distant nodes closer to the root while maintaining the in-order search structure of the tree. Among balanced search trees, AVL and 2/3 trees are now passé, and red-black trees seem to be more popular. A particularly interesting self-organizing data structure is the splay tree, which uses rotations to move any accessed key to the root. Frequently used or recently accessed nodes thus sit near the top of the tree, allowing faster searches. Bottom line: Which tree is best for your application? Probably the one of which you have the best implementation. The ﬂavor of balanced tree is probably not as important as the skill of the programmer who coded it. • B-trees – For data sets so large that they will not ﬁt in main memory (say more than 1,000,000 items) your best bet will be some ﬂavor of a B-tree. Once a data structure has to be stored outside of main memory, the search time grows by several orders of magnitude. With modern cache architectures, similar eﬀects can also happen on a smaller scale, since cache is much faster than RAM. The idea behind a B-tree is to collapse several levels of a binary search tree into a single large node, so that we can make the equivalent of several search steps before another disk access is needed. With B-tree we can access enormous numbers of keys using only a few disk accesses. To get the full beneﬁt from using a B-tree, it is important to understand how the secondary storage device and virtual memory interact, through constants such as page size and virtual/real address space. Cache-oblivious algorithms (described below) can mitigate such concerns. 12.1 DICTIONARIES Even for modest-sized data sets, unexpectedly poor performance of a data structure may result from excessive swapping, so listen to your disk to help decide whether you should be using a B-tree. • Skip lists – These are somewhat of a cult data structure. A hierarchy of sorted linked lists is maintained, where a coin is ﬂipped for each element to decide whether it gets copied into the next highest list. This implies roughly lg n lists, each roughly half as large as the one above it. Search starts in the smallest list. The search key lies in an interval between two elements, which is then explored in the next larger list. Each searched interval contains an expected constant number of elements per list, for a total expected O(lg n) query time. The primary beneﬁts of skip lists are ease of analysis and implementation relative to balanced trees. Implementations: Modern programming languages provide libraries oﬀering complete and eﬃcient container implementations. The C++ Standard Template Library (STL) is now provided with most compliers, and available with documentation at http://www.sgi.com/tech/stl/. See Josuttis [Jos99], Meyers [Mey01], and Musser [MDS01] for more detailed guides to using STL and the C++ standard library. LEDA (see Section 19.1.1 (page 658)) provides an extremely complete collection of dictionary data structures in C++, including hashing, perfect hashing, B-trees, red-black trees, random search trees, and skip lists. Experiments reported in [MN99] identiﬁed hashing as the best dictionary choice, with skip lists and 2-4 trees (a special case of B-trees) as the most eﬃcient tree-like structures. Java Collections (JC), a small library of data structures, is included in the java.util package of Java standard edition (http://java.sun.com/javase/). The Data Structures Library in Java (JDSL) is more comprehensive, and available for non-commercial use at http://www.jdsl.org/. See [GTV05, GT05] for more detailed guides to JDSL. Notes: Knuth [Knu97a] provides the most detailed analysis and exposition on fundamental dictionary data structures, but misses certain modern data structures as red-black and splay trees. Spending some time with his books is a worthwhile rite of passage for all computer science students. The Handbook of Data Structures and Applications [MS05] provides up-to-date surveys on all aspects of dictionary data structures. Other surveys include Mehlhorn and Tsakalidis [MT90b] and Gonnet and Baeza-Yates [GBY91]. Good textbook expositions on dictionary data structures include Sedgewick [Sed98], Weiss [Wei06], and Goodrich/Tamassia [GT05]. We defer to all these sources to avoid giving original references for each of the data structures described above. The 1996 DIMACS implementation challenge focused on elementary data structures, including dictionaries [GJM02]. Data sets, and codes are accessible from http://dimacs.rutgers.edu/Challenges. The cost of transferring data back and forth between levels of the memory hierarchy (RAM-to-cache or disk-to-RAM) dominates the cost of actual computation for many 371 372 12. DATA STRUCTURES problems. Each data transfer moves one block of size b, so eﬃcient algorithms seek to minimize the number of block transfers. The complexity of fundamental algorithm and data structure problems on such an external memory model has been extensively studied [Vit01]. Cache-oblivious data structures oﬀer performance guarantees under such a model without explicit knowledge of the block-size parameter b. Hence, good performance can be obtained on any machine without architecture-speciﬁc tuning. See [ABF05] for an excellent survey on cache-oblivious data structures. Several modern data structures, such as splay trees, have been studied using amortized analysis, where we bound the total amount of time used by any sequence of operations. In an amortized analysis, a single operation can be very expensive, but only because we have already beneﬁted from enough cheap operations to pay oﬀ the higher cost. A data structure realizing an amortized complexity of O(f (n)) is less desirable than one whose worst-case complexity is O(f (n)) (since a very bad operation might still occur) but better than one with an average-case complexity O(f (n)), since the amortized bound will achieve this average on any input. Related Problems: Sorting (see page 436), searching (see page 441). 12.2 PRIORITY QUEUES 373 1/1 October 30 December 7 July 4 January 1 February 2 December 25 INPUT 12.2 7/4 12/7 2/2 10/30 OUTPUT Priority Queues Input description: A set of records with numerically or otherwise totally-ordered keys. Problem description: Build and maintain a data structure for providing quick access to the smallest or largest key in the set. Discussion: Priority queues are useful data structures in simulations, particularly for maintaining a set of future events ordered by time. They are called “priority” queues because they enable you to retrieve items not by the insertion time (as in a stack or queue), nor by a key match (as in a dictionary), but by which item has the highest priority of retrieval. If your application will perform no insertions after the initial query, there is no need for an explicit priority queue. Simply sort the records by priority and proceed from top to bottom, maintaining a pointer to the last record retrieved. This situation occurs in Kruskal’s minimum spanning tree algorithm, or when simulating a completely scripted set of events. However, if you are mixing insertions, deletions, and queries, you will need a real priority queue. The following questions will help select the right one: • What other operations do you need? – Will you be searching for arbitrary keys, or just searching for the smallest? Will you be deleting arbitrary elements from the data, or just repeatedly deleting the top or smallest element? 12/25 374 12. DATA STRUCTURES • Do you know the maximum data structure size in advance? – The issue here is whether you can preallocate space for the data structure. • Might you change the priority of elements already in the queue? – Changing the priority of elements implies that we must be able to retrieve elements from the queue based on their key, in addition to being able to retrieve the largest element. Your choices are between the following basic priority queue implementations: • Sorted array or list – A sorted array is very eﬃcient to both identify the smallest element and delete it by decrementing the top index. However, maintaining the total order makes inserting new elements slow. Sorted arrays are only suitable when there will be few insertions into the priority queue. Basic priority queue implementations are reviewed in Section 3.5 (page 83). • Binary heaps – This simple, elegant data structure supports both insertion and extract-min in O(lg n) time each. Heaps maintain an implicit binary tree structure in an array, such that the key of the root of any subtree is less than that of all its descendents. Thus, the minimum key always sits at the top of the heap. New keys can be inserted by placing them at an open leaf and percolating the element upwards until it sits at its proper place in the partial order. An implementation of binary heap construction and retrieval in C appears in Section 4.3.1 (page 109) Binary heaps are the right answer when you know an upper bound on the number of items in your priority queue, since you must specify array size at creation time. Even this constraint can be mitigated by using dynamic arrays (see Section 3.1.1 (page 66)). • Bounded height priority queue – This array-based data structure permits constant-time insertion and ﬁnd-min operations whenever the range of possible key values is limited. Suppose we know that all key values will be integers between 1 and n. We can set up an array of n linked lists, such that the ith list serves as a bucket containing all items with key i. We will maintain a top pointer to the smallest nonempty list. To insert an item with key k into the priority queue, add it to the kth bucket and set top = min(top, k). To extract the minimum, report the ﬁrst item from bucket top, delete it, and move top down if the bucket has become empty. Bounded height priority queues are very useful in maintaining the vertices of a graph sorted by degree, which is a fundamental operation in graph algorithms. Still, they are not as widely known as they should be. They are usually the right priority queue for any small, discrete range of keys. • Binary search trees – Binary search trees make eﬀective priority queues, since the smallest element is always the leftmost leaf, while the largest element is 12.2 PRIORITY QUEUES always the rightmost leaf. The min (max) is found by simply tracing down left (right) pointers until the next pointer is nil. Binary tree heaps prove most appropriate when you need other dictionary operations, or if you have an unbounded key range and do not know the maximum priority queue size in advance. • Fibonacci and pairing heaps – These complicated priority queues are designed to speed up decrease-key operations, where the priority of an item already in the priority queue is reduced. This arises, for example, in shortest path computations when we discover a shorter route to a vertex v than previously established. Properly implemented and used, they lead to better performance on very large computations. Implementations: Modern programming languages provide libraries oﬀering complete and eﬃcient priority queue implementations. Member functions push, top, and pop of the C++ Standard Template Library (STL) priority queue template mirror heap operations insert, findmax, and deletemax. STL is available with documentation at http://www.sgi.com/tech/stl/. See Meyers [Mey01] and Musser [MDS01] for more detailed guides to using STL. LEDA (see Section 19.1.1 (page 658)) provides a complete collection of priority queues in C++, including Fibonacci heaps, pairing heaps, Emde-Boas trees, and bounded height priority queues. Experiments reported in [MN99] identiﬁed simple binary heaps as quite competitive in most applications, with pairing heaps beating Fibonacci heaps in head-to-head tests. The Java Collections PriorityQueue class is included in the java.util package of Java standard edition (http://java.sun.com/javase/). The Data Structures Library in Java (JDSL) provides an alternate implementation, and is available for noncommercial use at http://www.jdsl.org/. See [GTV05, GT05] for more detailed guides to JDSL. Sanders [San00] did extensive experiments demonstrating that his sequence heap, based on k-way merging, was roughly twice as fast as a well-implemented binary heap. See http://www.mpi-inf.mpg.de/∼sanders/programs/spq/ for his implementations in C++. Notes: The Handbook of Data Structures and Applications [MS05] provides several upto-date surveys on all aspects of priority queues. Empirical comparisons between priority queue data structures include [CGS99, GBY91, Jon86, LL96, San00]. Double-ended priority queues extend the basic heap operations to simultaneously support both ﬁnd-min and ﬁnd-max. See [Sah05] for a survey of four diﬀerent implementations of double-ended priority queues. Bounded-height priority queues are useful data structures in practice, but do not promise good worst-case performance when arbitrary insertions and deletions are permitted. However, von Emde Boas priority queues [vEBKZ77] support O(lg lg n) insertion, deletion, search, max, and min operations where each key is an element from 1 to n. 375 376 12. DATA STRUCTURES Fibonacci heaps [FT87] support insert and decrease-key operations in constant amortized time, with O(lg n) amortized time extract-min and delete operations. The constanttime decrease-key operation leads to faster implementations of classical algorithms for shortest-paths, weighted bipartite-matching, and minimum spanning tree. In practice, Fibonacci heaps are diﬃcult to implement and have large constant factors associated with them. However, pairing heaps appear to realize the same bounds with less overhead. Experiments with pairing heaps are reported in [SV87]. Heaps deﬁne a partial order that can be built using a linear number of comparisons. The familiar linear-time merging algorithm for heap construction is due to Floyd [Flo64]. In the worst case, 1.625n comparisons suﬃce [GM86] and 1.5n − O(lg n) comparisons are necessary [CC92]. Related Problems: Dictionaries (see page 367), sorting (see page 436), shortest path (see page 489). 12.3 XYZXYZ$ YZXYZ$ ZXYZ$ Z$ $ INPUT 12.3 $ XYZ YZ XYZ$ YZ$ 377 SUFFIX TREES AND ARRAYS (3, 3) (2, 3) XYZ 1 (7, 7) Z $ XYZ $ XYZ $ 4 5 3 6 2 OUTPUT Suﬃx Trees and Arrays Input description: A reference string S. Problem description: Build a data structure to quickly ﬁnd all places where an arbitrary query string q occurs in S. Discussion: Suﬃx trees and arrays are phenomenally useful data structures for solving string problems elegantly and eﬃciently. Proper use of suﬃx trees often speeds up string processing algorithms from O(n2 ) to linear time—likely the answer. Indeed, suﬃx trees are the hero of the war story reported in Section 3.9 (page 94). In its simplest instantiation, a suﬃx tree is simply a trie of the n suﬃxes of an n-character string S. A trie is a tree structure, where each edge represents one character, and the root represents the null string. Thus, each path from the root represents a string, described by the characters labeling the edges traversed. Any ﬁnite set of words deﬁnes a trie, and two words with common preﬁxes branch oﬀ from each other at the ﬁrst distinguishing character. Each leaf denotes the end of a string. Figure 12.1 illustrates a simple trie. Tries are useful for testing whether a given query string q is in the set. We traverse the trie from the root along branches deﬁned by successive characters of q. If a branch does not exist in the trie, then q cannot be in the set of strings. Otherwise we ﬁnd q in |q| character comparisons regardless of how many strings are in the trie. Tries are very simple to build (repeatedly insert new strings) and very fast to search, although they can be expensive in terms of memory. 7 (7, 7) 378 12. DATA STRUCTURES t w a h s e i r r h e n e Figure 12.1: A trie on strings the, their, there, was, and when A suﬃx tree is simply a trie of all the proper suﬃxes of S. The suﬃx tree enables you to test whether q is a substring of S, because any substring of S is the preﬁx of some suﬃx (got it?). The search time is again linear in the length of q. The catch is that constructing a full suﬃx tree in this manner can require O(n2 ) time and, even worse, O(n2 ) space, since the average length of the n suﬃxes is n/2. However, linear space suﬃces to represent a full suﬃx tree, if we are clever. Observe that most of the nodes in a trie-based suﬃx tree occur on simple paths between branch nodes in the tree. Each of these simple paths corresponds to a substring of the original string. By storing the original string in an array and collapsing each such path into a single edge, we have all the information of the full suﬃx tree in only O(n) space. The label for each edge is described by the starting and ending array indices representing the substring. The output ﬁgure for this section displays a collapsed suﬃx tree in all its glory. Even better, there exist O(n) algorithms to construct this collapsed tree, by making clever use of pointers to minimize construction time. These additional pointers can also be used to speed up many applications of suﬃx trees. But what can you do with suﬃx trees? Consider the following applications. For more details see the books by Gusﬁeld [Gus97] or Crochemore and Rytter [CR03]: • Find all occurrences of q as a substring of S – Just as with a trie, we can walk from the root to the node nq associated with q. The positions of all occurrences of q in S are represented by the descendents of nq , which can be identiﬁed using a depth-ﬁrst search from nq . In collapsed suﬃx trees, it takes O(|q| + k) time to ﬁnd the k occurrences of q in S. • Longest substring common to a set of strings – Build a single collapsed suﬃx tree containing all suﬃxes of all strings, with each leaf labeled with its original string. In the course of doing a depth-ﬁrst search on this tree, we can label each node with both the length of its common preﬁx and the number of distinct strings that are children of it. From this information, the best node can be selected in linear time. 12.3 SUFFIX TREES AND ARRAYS • Find the longest palindrome in S – A palindrome is a string that reads the same if the order of characters is reversed, such as madam. To ﬁnd the longest palindrome in a string S, build a single suﬃx tree containing all suﬃxes of S and the reversal of S, with each leaf identiﬁed by its starting position. A palindrome is deﬁned by any node in this tree that has forward and reversed children from the same position. Since linear time suﬃx tree construction algorithms are nontrivial, I recommend using an existing implementation. Another good option is to use suﬃx arrays, discussed below. Suﬃx arrays do most of what suﬃx trees do, while using roughly four times less memory. They are also easier to implement. A suﬃx array is in principle just an array that contains all the n suﬃxes of S in sorted order. Thus a binary search of this array for string q suﬃces to locate the preﬁx of a suﬃx that matches q, permitting an eﬃcient substring search in O(lg n) string comparisons. With the addition of an index specifying the common preﬁx length of all bounding suﬃxes, only lg n + |q| character comparisons need be performed on any query, since we can identify the next character that must be tested in the binary search. For example, if the lower range of the search is cowabunga and the upper range is cowslip, all keys in between must share the same ﬁrst three letters, so only the fourth character of any intermediate key must be tested against q. In practice, suﬃx arrays are typically as fast or faster to search than suﬃx trees. Suﬃx arrays use less memory than suﬃx trees. Each suﬃx is represented completely by its unique starting position (from 1 to n) and read oﬀ as needed using a single reference copy of the input string. Some care must be taken to construct suﬃx arrays eﬃciently, however, since there are O(n2 ) characters in the strings being sorted. One solution is to ﬁrst build a suﬃx tree, then perform an in-order traversal of it to read the strings oﬀ in sorted order! However, recent breakthroughs have lead to space/time eﬃcient algorithms for constructing suﬃx arrays directly. Implementations: There now exist a wealth of suﬃx array implementations available. Indeed, all of the recent linear time construction algorithms have been implemented and benchmarked [PST07]. Schürmann and Stoye [SS07] provide an excellent C implementation at http://bibiserv.techfak.uni-bielefeld.de/bpr/. No less than eight diﬀerent C/C++ implementations of compressed text indexes appear at the Pizza&Chili corpus (http://pizzachili.di.unipi.it/). These data structures go to great lengths to minimize space usage, typically compressing the input string to near the empirical entropy while still achieving excellent query times! Suﬃx tree implementations are also readily available. A SuffixTree class is provided in BioJava (ttp://www.biojava.org/)—an open source project providing a Java framework for processing biological data. Libstree is a C implementation of Ukkonen’s algorithm, available at http://www.icir.org/christian/libstree/. 379 380 12. DATA STRUCTURES Nelson’s C++ code [Nel96] is available from http://marknelson.us/1996/08/01/ suﬃx-trees/. Strmat is a collection of C programs implementing exact pattern matching algorithms in association with [Gus97], including an implementation of suﬃx trees. It is available at http://www.cs.ucdavis.edu/∼gusﬁeld/strmat.html. Notes: Tries were ﬁrst proposed by Fredkin [Fre62], the name coming from the central letters of the word “retrieval.” A survey of basic trie data structures with extensive references appears in [GBY91]. Eﬃcient algorithms for suﬃx tree construction are due to Weiner [Wei73], McCreight [McC76], and Ukkonen [Ukk92]. Good expositions on these algorithms include Crochmore and Rytter [CR03] and Gusﬁeld [Gus97]. Suﬃx arrays were invented by Manber and Myers [MM93], although an equivalent idea called Pat trees due to Gonnet and Baeza-Yates appears in [GBY91]. Three teams independently emerged with linear-time suﬃx array algorithms in 2003 [KSPP03, KA03, KSB05], and progress has continued rapidly. See [PST07] for a recent survey covering all these developments. Recent work has resulted in the development of compressed full text indexes that oﬀer essentially all the power of suﬃx trees/arrays in a data structure whose size is proportional to the compressed text string. Makinen and Navarro [MN07] survey these remarkable data structures. The power of suﬃx trees can be further augmented by using a data structure for computing the least common ancestor (LCA) of any pair of nodes x, y in a tree in constant time, after linear-time preprocessing of the tree. The original data structure due to Harel and Tarjan [HT84], has been progressively simpliﬁed by Schieber and Vishkin [SV88] and later Bender and Farach [BF00]. Expositions include Gusﬁeld [Gus97]. The least common ancestor of two nodes in a suﬃx tree or trie deﬁnes the node representing the longest common preﬁx of the two associated strings. That we can answer such queries in constant time is amazing, and proves useful as a building block for many other algorithms. Related Problems: String matching (see page 628), text compression (see page 637), longest common substring (see page 650). 12.4 1 2 3 5 4 INPUT 12.4 381 GRAPH DATA STRUCTURES 1 2 5 2 1 3 3 2 4 4 2 3 5 5 1 2 4 4 OUTPUT Graph Data Structures Input description: A graph G. Problem description: Represent the graph G using a ﬂexible, eﬃcient data structure. Discussion: The two basic data structures for representing graphs are adjacency matrices and adjacency lists. Full descriptions of these data structures appear in Section 5.2 (page 151), along with an implementation of adjacency lists in C. In general, for most things, adjacency lists are the way to go. The issues in deciding which data structure to use include: • How big will your graph be? – How many vertices will it have, both typically and in the worse case? Ditto for the number of edges? Graphs with 1,000 vertices imply adjacency matrices with 1,000,000 entries. This seems too be the boundary of reality. Adjacency matrices make sense only for small or very dense graphs. • How dense will your graph be? – If your graph is very dense, meaning that a large fraction of the vertex pairs deﬁne edges, there is probably no compelling reason to use adjacency lists. You will be doomed to using Θ(n2 ) space anyway. Indeed, for complete graphs, matrices will be more concise due to the elimination of pointers. • Which algorithms will you be implementing? – Certain algorithms are more natural on adjacency matrices (such as all-pairs shortest path) and others favor adjacency lists (such as most DFS-based algorithms). Adjacency matrices win for algorithms that repeatedly ask, “Is (i, j) in G?” However, most graph algorithms can be designed to eliminate such queries. • Will you be modifying the graph over the course of your application? – Eﬃcient static graph implementations can be used when no edge insertion/deletion operations will done following initial construction. Indeed, more 5 382 12. DATA STRUCTURES common than modifying the topology of the graph is modifying the attributes of a vertex or edge of the graph, such as size, weight, label, or color. Attributes are best handled as extra ﬁelds in the vertex or edge records of adjacency lists. Building a good general purpose graph type is a substantial project. For this reason, we suggest that you check out existing implementations (particularly LEDA) before hacking up your own. Note that it costs only time linear in the size of the larger data structure to convert between adjacency matrices and adjacency lists. This conversion is unlikely to be the bottleneck in any application, so you may decide to use both data structures if you have the space to store them. This usually isn’t necessary, but might prove simplest if you are confused about the alternatives. Planar graphs are those that can be drawn in the plane so no two edges cross. Graphs arising in many applications are planar by deﬁnition, such as maps of countries. Others are planar by happenstance, like trees. Planar graphs are always sparse, since any n-vertex planar graph can have at most 3n − 6 edges. Thus they should be represented using adjacency lists. If the planar drawing (or embedding) of the graph is fundamental to what is being computed, planar graphs are best represented geometrically. See Section 15.12 (page 520) for algorithms for constructing planar embeddings from graphs. Section 17.15 (page 614) discusses algorithms for maintaining the graphs implicit in the arrangements of geometric objects like lines and polygons. Hypergraphs are generalized graphs where each edge may link subsets of more than two vertices. Suppose we want to represent who is on which Congressional committee. The vertices of our hypergraph would be the individual congressmen, while each hyperedge would represent one committee. Such arbitrary collections of subsets of a set are naturally thought of as hypergraphs. Two basic data structures for hypergraphs are: • Incidence matrices, which are analogous to adjacency matrices. They require n×m space, where m is the number of hyperedges. Each row corresponds to a vertex, and each column to an edge, with a nonzero entry in M [i, j] iﬀ vertex i is incident to edge j. On standard graphs there are two nonzero entries in each column. The degree of each vertex governs the number of nonzero entries in each row. • Bipartite incidence structures, which are analogous to adjacency lists, and hence suited for sparse hypergraphs. There is a vertex of the incidence structure associated with each edge and vertex of the hypergraphs, and an edge (i, j) in the incidence structure if vertex i of the hypergraph appears in edge j of the hypergraph. Adjacency lists are typically used to represent this incidence structure. Drawing the associated bipartite graph provides a natural way to visualize the hypergraph. Special eﬀorts must be taken to represent very large graphs eﬃciently. However, interesting problems have been solved on graphs with millions of edges and 12.4 GRAPH DATA STRUCTURES vertices. The ﬁrst step is to make your data structure as lean as possible, by packing your adjacency matrix in a bit vector (see Section 12.5 (page 385)) or removing unnecessary pointers from your adjacency list representation. For example, in a static graph (which does not support edge insertions or deletions) each edge list can be replaced by a packed array of vertex identiﬁers, thus eliminating pointers and potentially saving half the space. If your graph is extremely large, it may become necessary to switch to a hierarchical representation, where the vertices are clustered into subgraphs that are compressed into single vertices. Two approaches exist to construct such a hierarchical decomposition. The ﬁrst breaks the graph into components in a natural or application-speciﬁc way. For example, a graph of roads and cities suggests a natural decomposition—partition the map into districts, towns, counties, and states. The other approach runs a graph partition algorithm discussed as in Section 16.6 (page 541). If you have one, a natural decomposition will likely do a better job than some naive heuristic for an NP-complete problem. If your graph is really unmanageably large, you cannot aﬀord to do a very good job of algorithmically partitioning it. First verify that standard data structures fail on your problem before attempting such heroic measures. Implementations: LEDA (see Section 19.1.1 (page 658)) provides the best graph data type currently implemented in C++. It is now a commercial product. You should at least study the methods it provides for graph manipulation, so as to see how the right level of abstract graph type makes implementing algorithms very clean and easy. The C++ Boost Graph Library [SLL02] (http://www.boost.org/libs/graph/doc) is more readily available. Implementations of adjacency lists, matrices, and edge lists are included, along with a reasonable library of basic graph algorithms. Its interface and components are generic in the same sense as the C++ standard template library (STL). JUNG (http://jung.sourceforge.net/) is a Java graph library particularly popular in the social networks community. The Data Structures Library in Java (JDSL) provides a comprehensive implementation with a decent algorithm library, and is available for noncommercial use at http://www.jdsl.org/. See [GTV05, GT05] for more detailed guides to JDSL. JGraphT (http://jgrapht.sourceforge.net/) is a more recent development with similar functionality. The Stanford Graphbase (see Section 19.1.8 (page 660)) provides a simple but ﬂexible graph data structure in CWEB, a literate version of the C language. It is instructive to see what Knuth does and does not place in his basic data structure, although we recommend other implementations as a better basis for further development. My (biased) preference in C language graph types is the library from my book Programming Challenges [SR03]. See Section 19.1.10 (page 661) for details. Simple graph data structures in Mathematica are provided by Combinatorica [PS03], with a library of algorithms and display routines. See Section 19.1.9 (page 661). 383 384 12. DATA STRUCTURES Notes: The advantages of adjacency list data structures for graphs became apparent with the linear-time algorithms of Hopcroft and Tarjan [HT73b, Tar72]. The basic adjacency list and matrix data structures are presented in essentially all books on algorithms or data structures, including [CLRS01, AHU83, Tar83]. Hypergraphs are presented in Berge [Ber89] The improved eﬃciency of static graph types was revealed by Naher and Zlotowski [NZ02], who sped up certain LEDA graph algorithms by a factor of four by simply switching to a more compact graph structure. An interesting question concerns minimizing the number of bits needed to represent arbitrary graphs on n vertices, particularly if certain operations must be supported eﬃciently. Such issues are surveyed in [vL90b]. Dynamic graph algorithms are data structures that maintain quick access to an invariant (such as minimum spanning tree or connectivity) under edge insertion and deletion. Sparsiﬁcation [EGIN92] is a general approach to constructing dynamic graph algorithms. See [ACI92, Zar02] for experimental studies on the practicality of dynamic graph algorithms. Hierarchically-deﬁned graphs arise often in VLSI design problems, because designers make extensive use of cell libraries [Len90]. Algorithms speciﬁcally for hierarchicallydeﬁned graphs include planarity testing [Len89], connectivity [LW88], and minimum spanning trees [Len87a]. Related Problems: Set data structures (see page 385), graph partition (see page 541). 12.5 SET DATA STRUCTURES (X X Y Z INPUT 12.5 Y) X 385 Z Y Z OUTPUT Set Data Structures Input description: A universe of items U = {u1 , . . . , un } on which is deﬁned a collection of subsets S = {S1 , . . . , Sm }. Problem description: Represent each subset so as to eﬃciently (1) test whether ui ∈ Sj , (2) compute the union or intersection of Si and Sj , and (3) insert or delete members of S. Discussion: In mathematical terms, a set is an unordered collection of objects drawn from a ﬁxed universal set. However, it is usually useful for implementation to represent each set in a single canonical order, typically sorted, to speed up or simplify various operations. Sorted order turns the problem of ﬁnding the union or intersection of two subsets into a linear-time operation—just sweep from left to right and see what you are missing. It makes possible element searching in sublinear time. Finally, printing the elements of a set in a canonical order paradoxically reminds us that order really doesn’t matter. We distinguish sets from two other kinds of objects: dictionaries and strings. A collection of objects not drawn from a ﬁxed-size universal set is best thought of as a dictionary, discussed in Section 12.1 (page 367). Strings are structures where order matters—i.e. , if {A, B, C} is not the same as {B, C, A}. Sections 12.3 and 18 discuss data structures and algorithms for strings. 386 12. DATA STRUCTURES Multisets permit elements to have more than one occurrence. Data structures for sets can generally be extended to multisets by maintaining a count ﬁeld or linked list of equivalent entries for each element. If each subset contains exactly two elements, they can be thought of as edges in a graph whose vertices represent the universal set. A system of subsets with no restrictions on the cardinality of its members is called a hypergraph. It is worth considering whether your problem has a graph-theoretical analogy, like connected components or shortest path in a graph/hypergraph. Your primary alternatives for representing arbitrary subsets are: • Bit vectors – An n-bit vector or array can represent any subset S on a universal set U containing n items. Bit i will be 1 if i ∈ S, and 0 if not. Since only one bit is needed per element, bit vectors can be very space eﬃcient for surprisingly large values of |U |. Element insertion and deletion simply ﬂips the appropriate bit. Intersection and union are done by “and-ing” or “oring” the bits together. The only drawback of a bit vector is its performance on sparse subsets. For example, it takes O(n) time to explicitly identify all members of sparse (even empty) subset S. • Containers or dictionaries – A subset can also be represented using a linked list, array, or dictionary containing exactly the elements in the subset. No notion of a ﬁxed universal set is needed for such a data structure. For sparse subsets, dictionaries can be more space and time eﬃcient than bit vectors, and easier to work with and program. For eﬃcient union and intersection operations, it pays to keep the elements in each subset sorted, so a lineartime traversal through both subsets identiﬁes all duplicates. • Bloom ﬁlters – We can emulate a bit vector in the absence of a ﬁxed universal set by hashing each subset element to an integer from 0 to n and setting the corresponding bit. Thus, bit H(e) will be 1 if e ∈ S. Collisons leave some possibility for error under this scheme, however, because a diﬀerent key might have hashed to the same position. Bloom ﬁlters use several (say k) diﬀerent hash functions H1 , . . . Hk , and set all k bits Hi (e) upon insertion of key e. Now e is in S only if all k bits are 1. The probability of false positives can be made arbitrarily low by increasing the number of hash functions k and table size n. With the proper constants, each subset element can be represented using a constant number of bits independent of the size of the universal set. This hashing-based data structure is much more space-eﬃcient than dictionaries for static subset applications that can tolerate a small probability of error. Many can. For instance, a spelling checker that left a rare random string undetected would prove no great tragedy. Many applications involve collections of subsets that are pairwise disjoint, meaning that each element is in exactly one subset. For example, consider maintaining 12.5 SET DATA STRUCTURES the connected components of a graph or the party aﬃliations of politicians. Each vertex/hack is in exactly one component/party. Such a system of subsets is called a set partition. Algorithms for generating partitions of a given set are provided in Section 14.6 (page 456). The primary issue with set partition data structures is maintaining changes over time, perhaps as edges are added or party members defect. Typical queries include “which set is a particular item in?” and “are two items in the same set?” as we modify the set by (1) changing one item, (2) merging or unioning two sets, or (3) breaking a set apart. Your basic options are: • Collection of containers – Representing each subset in its own container/dictionary permits fast access to all the elements in the subset, which facilitates union and intersection operations. The cost comes in membership testing, as we must search each subset data structure independently until we ﬁnd our target. • Generalized bit vector – Let the ith element of an array denote the number/name of the subset that contains it. Set identiﬁcation queries and single element modiﬁcations can be performed in constant time. However, operations like performing the union of two subsets take time proportional to the size of the universe, since each element in the two subsets must be identiﬁed and (at least one subset’s worth) must have its name changed. • Dictionary with a subset attribute – Similarly, each item in a binary tree can be associated a ﬁeld that records the name of the subset it is in. Set identiﬁcation queries and single element modiﬁcations can be performed in the time it takes to search in the dictionary. However, union/intersection operations are again slow. The need to perform such union operations quickly provides the motivation for the . . . • Union-ﬁnd data structure – We represent a subset using a rooted tree where each node points to its parent instead of its children. The name of each subset will be the name of the item at the root. Finding out which subset we are in is simple, for we keep traversing up the parent pointers until we hit the root. Unioning two subsets is also easy. Just assign the root of one of two trees to point to the other, so now all elements have the same root and hence the same subset name. Implementation details have a big impact on asymptotic performance here. Always selecting the larger (or taller) tree as the root in a merger guarantees logarithmic height trees, as presented with our implementation in Section 6.1.3 (page 198). Retraversing the path traced on each ﬁnd and explicitly pointing all nodes on the path to the root (called path compression) reduces the tree to almost constant height. Union ﬁnd is a fast, simple data structure that every programmer should know about. It does not support breaking up subsets created by unions, but usually this is not an issue. 387 388 12. DATA STRUCTURES Implementations: Modern programming languages provide libraries oﬀering complete and eﬃcient set implementations. The C++ Standard Template Library (STL) provides set and multiset containers. Java Collections (JC) contains HashSet and TreeSet containers and is included in the java.util package of Java standard edition. LEDA (see Section 19.1.1 (page 658)) provides eﬃcient dictionary data structures, sparse arrays, and union-ﬁnd data structures to maintain set partitions, all in C++. Implementation of union-ﬁnd underlies any implementation of Kruskal’s minimum spanning tree algorithm. For this reason, all the graph libraries of Section 12.4 (page 381) presumably contains an implementation. Minimum spanning tree codes are described in Section 15.3 (page 484). The computer algebra system REDUCE (http://www.reduce-algebra.com/) contains SETS, a package supporting set-theoretic operations on both explicit and implicit (symbolic) sets. Other computer algebra systems may support similar functionality. Notes: Optimal algorithms for such set operations as intersection and union were presented in [Rei72]. Raman [Ram05] provides an excellent survey on data structures for set operations on a variety of diﬀerent operations. Bloom ﬁlters are ably surveyed in [BM05], with recent experimental results presented in [PSS07]. Certain balanced tree data structures support merge/meld/link/cut operations, which permit fast ways to union and intersect disjoint subsets. See Tarjan [Tar83] for a nice presentation of such structures. Jacobson [Jac89] augmented the bit-vector data structure to support select operations (where is the ith 1 bit?) eﬃciently in both time and space. Galil and Italiano [GI91] survey data structures for disjoint set union. The upper bound of O(mα(m, n)) on m union-ﬁnd operations on an n-element set is due to Tarjan [Tar75], as is a matching lower bound on a restricted model of computation [Tar79]. The inverse Ackerman function α(m, n) grows notoriously slowly, so this performance is close to linear. An interesting connection between the worst-case of union-ﬁnd and the length of Davenport-Schinzel sequences—a combinatorial structure that arises in computational geometry—is established in [SA95]. The power set of a set S is the collection of all 2|S| subsets of S. Explicit manipulation of power sets quickly becomes intractable due to their size. Implicit representations of power sets in symbolic form becomes necessary for nontrivial computations. See [BCGR92] for algorithms on and computational experience with symbolic power set representations. Related Problems: Generating subsets (see page 452), generating partitions (see page 456), set cover (see page 621), minimum spanning tree (see page 484). 12.6 INPUT 12.6 KD-TREES OUTPUT Kd-Trees Input description: A set S of n points or more complicated geometric objects in k dimensions. Problem description: Construct a tree that partitions space by half-planes such that each object is contained in its own box-shaped region. Discussion: Kd-tree and related spatial data structures hierarchically decompose space into a small number of cells, each containing a few representatives from an input set of points. This provides a fast way to access any object by position. We traverse down the hierarchy until we ﬁnd the smallest cell containing it, and then scan through the objects in this cell to identify the right one. Typical algorithms construct kd-trees by partitioning point sets. Each node in the tree is deﬁned by a plane cutting through one of the dimensions. Ideally, this plane equally partitions the subset of points into left/right (or up/down) subsets. These children are again partitioned into equal halves, using planes through a diﬀerent dimension. Partitioning stops after lg n levels, with each point in its own leaf cell. The cutting planes along any path from the root to another node deﬁnes a unique box-shaped region of space. Each subsequent plane cuts this box into two boxes. Each box-shaped region is deﬁned by 2k planes, where k is the number of dimensions. Indeed, the “kd” in kd-tree is short for “k-dimensional.” We maintain 389 390 12. DATA STRUCTURES the region of interest deﬁned by the intersection of these half-spaces as we move down the tree. Flavors of kd-trees diﬀer in exactly how the splitting plane is selected. Options include: • Cycling through the dimensions – partition ﬁrst on d1 , then d2 , . . . , dk before cycling back to d1 . • Cutting along the largest dimension – select the partition dimension to make the resulting boxes as square or cube-like as possible. Selecting a plane to partition the points in half does not mean selecting a splitter in the middle of the box-shaped regions, since all the points may lie in the left side of the box. • Quadtrees or Octtrees – Instead of partitioning with single planes, use all axis-parallel planes that pass through a given partition point. In two dimensions, this means creating four child cells; in 3D, this means eight child cells. Quadtrees seem particularly popular on image data, where leaf cells imply that all pixels in the regions have the same color. • BSP-trees – Binary space partitions use general (i.e. , not just axis-parallel) cutting planes to carve up space into cells so that each cell ends up containing only one object (say a polygon). Such partitions are not possible using only axis-parallel cuts for certain sets of objects. The downside is that such polyhedral cell boundaries are more complicated to work with than boxes. • R-trees – This is another spatial data structure useful for geometric objects that cannot be partitioned into axis-oriented boxes without cutting them into pieces. At each level, the objects are partitioned into a small number of (possibly-overlapping) boxes to construct searchable hierarchies without partitioning objects. Ideally, our partitions split both the space (ensuring fat, regular regions) and the set of points (ensuring a log height tree) evenly, but doing both simultaneously can be impossible on a given input. The advantages of fat cells become clear in many applications of kd-trees: • Point location – To identify which cell a query point q lies in, we start at the root and test which side of the partition plane contains q. By repeating this process on the appropriate child node, we travel down the tree to ﬁnd the leaf cell containing q in time proportional to its height. See Section 17.7 (page 587) for more on point location. • Nearest neighbor search – To ﬁnd the point in S closest to a query point q, we perform point location to ﬁnd the cell c containing q. Since c is bordered by some point p, we can compute the distance d(p, q) from p to q. Point p is likely 12.6 KD-TREES close to q, but it might not be the single closest neighbor. Why? Suppose q lies right at the boundary of a cell. Then q’s nearest neighbor might lie just to the left of the boundary in another cell. Thus, we must traverse all cells that lie within a distance of d(p, q) of cell c and verify that none of them contain closer points. In trees with nice, fat cells, very few cells should need to be tested. See Section 17.5 (page 580) for more on nearest neighbor search. • Range search – Which points lie within a query box or region? Starting from the root, check whether the query region intersects (or contains) the cell deﬁning the current node. If it does, check the children; if not, none of the leaf cells below this node can possibly be of interest. We quickly prune away irrelevant portions of the space. Section 17.6 (page 584) focuses on range search. • Partial key search – Suppose we want to ﬁnd a point p in S, but we do not have full information about p. Say we are looking for someone of age 35 and height 5’8” but of unknown weight in a 3D-tree with dimensions of age, weight, and height. Starting from the root, we can identify the correct descendant for all but the weight dimension. To be sure we ﬁnd the right point, we must search both children of these nodes. The more ﬁelds we know the better, but such partial key search can be substantially faster than checking all points against the key. Kd-trees are most useful for a small to moderate number of dimensions, say from 2 up to maybe 20 dimensions. They lose eﬀectiveness as the dimensionality increases, primarily because the ratio of the volume of a unit sphere in kdimensions shrinks exponentially compared to the unit cube. Thus exponentially many cells will have to be searched within a given radius of a query point, say for nearest-neighbor search. Also, the number of neighbors for any cell grows to 2k and eventually become unmanageable. The bottom line is that you should try to avoid working in high-dimensional spaces, perhaps by discarding (or projecting away) the least important dimensions. Implementations: KDTREE 2 contains C++ and Fortran 95 implementations of kd-trees for eﬃcient nearest nearest neighbor search in many dimensions. See http://arxiv.org/abs/physics/0408067. Samet’s spatial index demos (http://donar.umiacs.umd.edu/quadtree/) provide a series of Java applets illustrating many variants of kd-trees, in association with his book [Sam06]. Terralib (http://www.terralib.org/) is an open source geographic information system (GIS) software library written in C++. This includes an implementation of spatial data structures. The 1999 DIMACS implementation challenge focused on data structures for nearest neighbor search [GJM02]. Data sets and codes are accessible from http://dimacs.rutgers.edu/Challenges. 391 392 12. DATA STRUCTURES Notes: Samet [Sam06] is the best reference on kd-trees and other spatial data structures. All major (and many minor) variants are developed in substantial detail. A shorter survey [Sam05] is also available. Bentley [Ben75] is generally credited with developing kd-trees, although they have the murky history associated with most folk data structures. The performance of spatial data structures degrades with high dimensionality. Projecting high-dimensional spaces onto a random lower-dimensional hyperplane has recently emerged as a simple but powerful method for dimensionality reduction. Both theoretical [IM04] and empirical [BM01] results indicate that this method preserves distances quite nicely. Algorithms that quickly produce a point provably close to the query point are a recent development in higher-dimensional nearest neighbor search. A sparse weightedgraph structure is built from the data set, and the nearest neighbor is found by starting at a random point and walking greedily in the graph towards the query point. The closest point found during several random trials is declared the winner. Similar data structures hold promise for other problems in high-dimensional spaces. See [AM93, AMN+ 98]. Related Problems: Nearest-neighbor search (see page 580), point location (see page 587), range search (see page 584). 13 Numerical Problems If most problems you encounter are numerical in nature, there is an excellent chance that you are reading the wrong book. Numerical Recipes [PFTV07] gives a terriﬁc overview to the fundamental problems in numerical computing, including linear algebra, numerical integration, statistics, and diﬀerential equations. Diﬀerent ﬂavors of the book include source code for all the algorithms in C++, Fortran, and even Pascal. Their coverage is somewhat skimpier on the combinatorial/numerical problems we consider in this section, but you should be aware of this book. Check it out at http://www.nr.com. Numerical algorithms tend to be diﬀerent beasts than combinatorial algorithms for at least two reasons: • Issues of Precision and Error – Numerical algorithms typically perform repeated ﬂoating-point computations, which accumulate error at each operation until, eventually, the results are meaningless. My favorite example [MV99] concerns the Vancouver Stock Exchange, which over a twenty-two month period accumulated enough round-oﬀ error to reduce its index to 574.081 from the correct value of 1098.982. A simple and dependable way to test for round-oﬀ errors in numerical programs is to run them both at single and double precision, and then think hard whenever there is a disagreement. • Extensive Libraries of Codes – Large, high-quality libraries of numerical routines have existed since the 1960s, which is still not yet the case for combinatorial algorithms. There are several reasons for this, including (1) the early emergence of Fortran as a standard for numerical computation, (2) the nature of numerical computations to be recognizably independent rather than S.S. Skiena, The Algorithm Design Manual, 2nd ed., DOI: 10.1007/978-1-84800-070-4 13, c Springer-Verlag London Limited 2008 394 13. NUMERICAL PROBLEMS embedded within large applications, and (3) the existence of large scientiﬁc communities needing general numerical libraries. Regardless of why, you should exploit this software base. There is probably no reason to implement algorithms for any of the problems in this section as opposed to using existing codes. Searching Netlib (see Section 19.1.5) is an excellent place to start. Many scientist’s and engineer’s ideas about algorithms culturally derive from Fortran programming and numerical methods. Computer scientists grow up programming with pointers and recursion, and so are comfortable with the more sophisticated data structures required for combinatorial algorithms. Both sides can and should learn from each other, since problems such as pattern recognition can be modeled either numerically or combinatorially. There is a vast literature on numerical algorithms. In addition to Numerical Recipes, recommended books include: • Chapara and Canale [CC05] – The contemporary market leader in numerical analysis texts. • Mak [Mak02] – This enjoyable text introduces Java to the world of numerical computation, and visa versa. Source code is provided. • Hamming [Ham87] – This oldie but goodie provides a clear and lucid treatment of fundamental methods in numerical computation. It is available in a low-priced Dover edition. • Skeel and Keiper [SK00] – A readable and interesting treatment of basic numerical methods, avoiding overly detailed algorithm descriptions through its use of the computer algebra system Mathematica. I like it. • Cheney and Kincaid [CK07] – A traditional Fortran-based numerical analysis text, with discussions of optimization and Monte Carlo methods in addition to such standard topics as root-ﬁnding, numerical integration, linear systems, splines, and diﬀerential equations. • Buchanan and Turner [BT92] – Thorough language-independent treatment of all standard topics, including parallel algorithms. It is the most comprehensive of the texts described here. 13.1 3 SOLVING LINEAR EQUATIONS 3 y = -3x/2 + 4 2 2 (2, 1) y = x-1 1 1 0 0 -1 1 2 3 -1 1 INPUT 13.1 2 3 OUTPUT Solving Linear Equations Input description: An m×n matrix A and an m×1 vector b, together representing m linear equations on n variables. Problem description: What is the vector x such that A · x = b? Discussion: The need to solve linear systems arises in an estimated 75% of all scientiﬁc computing problems [DB74]. For example, applying Kirchhoﬀ’s laws to analyze electrical circuits generates a system of equations—the solution of which puts currents through each branch of the circuit. Analysis of the forces acting on a mechanical truss generates a similar set of equations. Even ﬁnding the point of intersection between two or more lines reduces to solving a small linear system. Not all systems of equations have solutions. Consider the equations 2x + 3y = 5 and 2x + 3y = 6. Some systems of equations have multiple solutions, such as 2x + 3y = 5 and 4x + 6y = 10. Such degenerate systems of equations are called singular, and they can be recognized by testing whether the determinant of the coeﬃcient matrix is zero. Solving linear systems is a problem of such scientiﬁc and commercial importance that excellent codes are readily available. There is no good reason to implement your own solver, even though the basic algorithm (Gaussian elimination) is one you learned in high school. This is especially true when working with large systems. Gaussian elimination is based on the observation that the solution to a system of linear equations is invariant under scaling ( if x = y, then 2x = 2y) and adding 395 396 13. NUMERICAL PROBLEMS equations (the solution to x = y and w = z is the same as the solution to x = y and x + w = y + z). Gaussian elimination scales and adds equations to eliminate each variable from all but one equation, leaving the system in such a state that the solution can be read oﬀ from the equations. The time complexity of Gaussian elimination on an n × n system of equations is O(n3 ), since to clear the ith variable we add a scaled copy of the n-term ith row to each of the n − 1 other equations. On this problem, however, constants matter. Algorithms that only partially reduce the coeﬃcient matrix and then backsubstitute to get the answer use 50% fewer ﬂoating-point operations than the naive algorithm. Issues to worry about include: • Are roundoﬀ errors and numerical stability aﬀecting my solution? – Gaussian elimination would be quite straightforward to implement except for roundoﬀ errors. These accumulate with each row operation and can quickly wreak havoc on the solution, particularly with matrices that are almost singular. To eliminate the danger of numerical errors, it pays to substitute the solution back into each of the original equations and test how close they are to the desired value. Iterative methods for solving linear systems reﬁne initial solutions to obtain more accurate answers. Good linear systems packages will include such routines. The key to minimizing roundoﬀ errors in Gaussian elimination is selecting the right equations and variables to pivot on, and to scale the equations to eliminate large coeﬃcients. This is an art as much as a science, which is why you should use a well-crafted library routine as described next. • Which routine in the library should I use? – Selecting the right code is also somewhat of an art. If you are taking your advice from this book, start with the general linear system solvers. Hopefully they will suﬃce for your needs. But search through the manual for more eﬃcient procedures solving special types of linear systems. If your matrix happens to be one of these special types, the solution time can reduce from cubic to quadratic or even linear. • Is my system sparse? – The key to recognizing that you have a special-case linear system is establishing how many matrix elements you really need to describe A. If there are only a few non-zero elements, your matrix is sparse and you are in luck. If these few non-zero elements are clustered near the diagonal, your matrix is banded and you are in even more luck. Algorithms for reducing the bandwidth of a matrix are discussed in Section 13.2. Many other regular patterns of sparse matrices can also be exploited, so consult the manual of your solver or a better book on numerical analysis for details. • Will I be solving many systems using the same coeﬃcient matrix? – In applications such as least-squares curve ﬁtting, we must solve A · x = b repeatedly with diﬀerent b vectors. We can preprocess A to make this easier. The lowerupper or LU-decomposition of A creates lower- and upper-triangular matrices 13.1 SOLVING LINEAR EQUATIONS L and U such that L·U = A. We can use this decomposition to solve A·x = b, since A · x = (L · U ) · x = L · (U · x) = b This is eﬃcient since backsubstitution solves a triangular system of equations in quadratic time. Solving L · y = b and then U · x = y gives the solution x using two O(n2 ) steps instead of one O(n3 ) step, after the LU-decomposition has been found in O(n3 ) time. The problem of solving linear systems is equivalent to that of matrix inversion, since Ax = B ↔ A−1 Ax = A−1 B, where I = A−1 A is the identity matrix. Avoid it however since matrix inversion proves to be three times slower than Gaussian elimination. LU-decomposition proves useful in inverting matrices as well as computing determinants (see Section 13.4 (page 404)). Implementations: The library of choice for solving linear systems is apparently LAPACK—a descendant of LINPACK [DMBS79]. Both of these Fortran codes, as well as many others, are available from Netlib. See Section 19.1.5 (page 659). Variants of LAPACK exist for other languages, like CLAPACK (C) and LAPACK++ (C++). The Template Numerical Toolkit is an interface to such routines in C++, and is available at http://math.nist.gov/tnt/. JScience provides an extensive linear algebra package (including determinants) as part of its comprehensive scientiﬁc computing library. JAMA is another matrix package written in Java. Links to both and many related libraries are available at http://math.nist.gov/javanumerics/. Numerical Recipes [PFTV07] (www.nr.com) provides guidance and routines for solving linear systems. Lack of conﬁdence in dealing with numerical procedures is the most compelling reason to use these ahead of the free codes. Notes: Golub and van Loan [GL96] is the standard reference on algorithms for linear systems. Good expositions on algorithms for Gaussian elimination and LU-decomposition include [CLRS01] and a host of numerical analysis texts [BT92, CK07, SK00]. Data structures for linear systems are surveyed in [PT05]. Parallel algorithms for linear systems are discussed in [Gal90, KSV97, Ort88]. Solving linear systems is one of most important applications where parallel architectures are used widely in practice. Matrix inversion and (hence) linear systems solving can be done in matrix multiplication time using Strassen’s algorithm plus a reduction. Good expositions on the equivalence of these problems include [AHU74, CLRS01]. Related Problems: Matrix multiplication (see page 401), determinant/permanent (see page 404). 397 398 1 13. NUMERICAL PROBLEMS 2 3 4 5 6 7 8 1 INPUT 13.2 8 2 7 3 6 4 5 OUTPUT Bandwidth Reduction Input description: A graph G = (V, E), representing an n × n matrix M of zero and non-zero elements. Problem description: Which permutation p of the vertices minimizes the length of the longest edge when the vertices are ordered on a line—i.e. , minimizes max(i,j)∈E |p(i) − p(j)|? Discussion: Bandwidth reduction lurks as a hidden but important problem for both graphs and matrices. Applied to matrices, bandwidth reduction permutes the rows and columns of a sparse matrix to minimize the distance b of any non-zero entry from the center diagonal. This is important in solving linear systems, because Gaussian elimination (see Section 13.1 (page 395)) can be performed in O(nb2 ) on matrices of bandwidth b. This is a big win over the general O(n3 ) algorithm if b << n. Bandwidth minimization on graphs arises in more subtle ways. Arranging n circuit components in a line to minimize the length of the longest wire (and hence time delay) is a bandwidth problem, where each vertex of our graph corresponds to a circuit component and there is an edge for every wire linking two components. Alternatively, consider a hypertext application where we must store large objects (say images) on a magnetic tape. Each image has a set of possible images to visit next (i.e. , the hyperlinks). We seek to place linked images near each other on the tape to minimize the search time. This is exactly the bandwidth problem. More general formulations, such as rectangular circuit layouts and magnetic disks, inherit the same hardness and classes of heuristics from the linear versions. The bandwidth problem seeks a linear ordering of the vertices, which minimizes the length of the longest edge, but there are several variations of the problem. In linear arrangement, we seek to minimize the sum of the lengths of the edges. This 13.2 BANDWIDTH REDUCTION has application to circuit layout, where we seek to position the chips to minimize the total wire length. In proﬁle minimization, we seek to minimize the sum of oneway distances (i.e., for each vertex v) the length of the longest edge whose other vertex is to the left of of v. Unfortunately, bandwidth minimization and all these variants is NP-complete. It stays NP-complete even if the input graph is a tree whose maximum vertex degree is 3, which is about as strong a condition as I have seen on any problem. Thus our only options are a brute-force search and heuristics. Fortunately, ad hoc heuristics have been well studied and production-quality implementations of the best heuristics are available. These are based on performing a breadth-ﬁrst search from a given vertex v, where v is placed at the leftmost point of the ordering. All of the vertices that are distance 1 from v are placed to its immediate right, followed by all the vertices at distance 2, and so forth until all vertices in G are accounted for. The popular heuristics diﬀer according to how many diﬀerent start vertices are considered and how equidistant vertices are ordered among themselves. Breaking ties with low-degree vertices over to the left however seems to be a good idea. Implementations of the most popular heuristics—the Cuthill-McKee and GibbsPoole-Stockmeyer algorithms—are discussed in the implementation section. The worst case of the Gibbs-Poole-Stockmeyer algorithm is O(n3 ), which would wash out any possible savings in solving linear systems, but its performance in practice is close to linear. Brute-force search programs can ﬁnd the exact minimum bandwidth by backtracking through the set of n! possible permutations of vertices. Considerable pruning can be achieved to reduce the search space by starting with a good heuristic bandwidth solution and alternately adding vertices to the left- and rightmost open slots in the partial permutation. Implementations: Del Corso and Manzini’s [CM99] code for exact solutions to bandwidth problems is available at http://www.mfn.unipmn.it/∼manzini/bandmin. Caprara and Salazar-González [CSG05] developed improved methods based on integer programming. Their branch-and-bound implementation in C is available at http://joc.pubs.informs.org/Supplements/Caprara-2/. Fortran language implementations of both the Cuthill-McKee algorithm [CGPS76, Gib76, CM69] and the Gibbs-Poole-Stockmeyer algorithm [Lew82, GPS76] are available from Netlib. See Section 19.1.5 (page 659). Empirical evaluations of these and other algorithms on a test suite of 30 matrices are discussed in [Eve79b], showing Gibbs-Poole-Stockmeyer to be the consistent winner. Petit [Pet03] performed an extensive experimental study on heuristics for the minimum linear arrangement problem. His codes and data are available at http://www.lsi.upc.edu/∼jpetit/MinLA/Experiments/. Notes: Diaz et al. [DPS02] provide an excellent up-to-date survey on algorithms for bandwidth and related graph layout problems. See [CCDG82] for graph-theoretic and algorithmic results on bandwidth up to 1981. 399 400 13. NUMERICAL PROBLEMS Ad hoc heuristics have been widely studied—a tribute to its importance in numerical computation. Everstine [Eve79b] cites no less than 49 diﬀerent bandwidth reduction algorithms! Del Corso and Romani [CR01] investigate a new class of spectral heuristics for bandwidth minimization. The hardness of the bandwidth problem was ﬁrst established by Papadimitriou [Pap76b], and its hardness on trees of maximum degree 3 in [GGJK78]. There are algorithms that run in polynomial time for ﬁxed bandwidth k [Sax80]. Approximation algorithms oﬀering a polylogarithmic guarantee exist for the general problem [BKRV00]. Related Problems: Solving linear equations (see page 395), topological sorting (see page 481). 13.3 2 MATRIX MULTIPLICATION 3 3 4 4 5 2 3 4 3 4 5 INPUT 13.3 13 18 23 18 25 32 23 32 41 OUTPUT Matrix Multiplication Input description: An x × y matrix A and a y × z matrix B. Problem description: Compute the x × z matrix A × B. Discussion: Matrix multiplication is a fundamental problem in linear algebra. Its main signiﬁcance for combinatorial algorithms is its equivalence to many other problems, including transitive closure/reduction, parsing, solving linear systems, and matrix inversion. Thus, a faster algorithm for matrix multiplication implies faster algorithms for all of these problems. Matrix multiplication arises in its own right in computing the results of such coordinate transformations as scaling, rotation, and translation for robotics and computer graphics. The following tight algorithm computes the product of x × y matrix A and y × z matrix B runs in O(xyz). Remember to ﬁrst initialize M [i, j] to 0 for all 1 ≤ i ≤ x and i ≤ j ≤ z: for i = 1 to x do for j = 1 to z for k = 1 to y M [i, j] = M [i, j] + A[i, k] · A[k, j] An implementation in C appears in Section 2.5.4 (page 45). This straightforward algorithm would seem to be tough to beat in practice. That said, observe that the three loops can be arbitrarily permuted without changing the resulting answer. Such a permutation will change the memory access patterns and thus how eﬀectively the cache memory is used. One can expect a 10-20% variation in run time among the six possible implementations, but could not conﬁdently predict the winner (typically ikj) without running it on your machine with your given matrices. When multiplying bandwidth-b matrices, where all non-zero elements of A and B lie within b elements of the main diagonals, a speedup to O(xbz) is possible, since zero elements cannot contribute to the product. 401 402 13. NUMERICAL PROBLEMS Asymptotically faster algorithms for matrix multiplication exist, using clever divide-and-conquer recurrences. However, these prove diﬃcult to program, require very large matrices to beat the trivial algorithm, and are less numerically stable to boot. The most famous of these is Strassen’s O(n2.81 ) algorithm. Empirical results (discussed next) disagree on the exact crossover point where Strassen’s algorithm beats the simple cubic algorithm, but it is in the ballpark of n ≈ 100. There is a better way to save computation when you are multiplying a chain of more than two matrices together. Recall that multiplying an x × y matrix by a y × z matrix creates an x × z matrix. Thus, multiplying a chain of matrices from left to right might create large intermediate matrices, each taking a lot of time to compute. Matrix multiplication is not commutative, but it is associative, so we can parenthesize the chain in whatever manner we deem best without changing the ﬁnal product. A standard dynamic programming algorithm can be used to construct the optimal parenthesization. Whether it pays to do this optimization will depend upon whether your matrix dimensions are suﬃciently irregular and your chain multiplied often enough to justify it. Note that we are optimizing over the sizes of the dimensions in the chain, not the actual matrices themselves. No such optimization is possible if all your matrices are the same dimensions. Matrix multiplication has a particularly interesting interpretation in counting the number of paths between two vertices in a graph. Let A be the adjacency matrix of a graph G, meaning A[i, j] = 1 if there is an edge between i and j. Otherwise, A[i, j] = 0. Now consider the square of this matrix, A2 = A × A. If A2 [i, j] ≥ 1. This means that there must be a vertex k such that A[i, k] = A[k, j] = 1, so i to k to j is a path of length 2 in G. More generally, Ak [i, j] counts the number of paths of length exactly k between i and j. This count includes nonsimple paths, where vertices are repeated, such as i to k to i to j. Implementations: D’Alberto and Nicolau [DN07] have engineered a very eﬃcient matrix multiplication code, which switches from Strassen’s to the cubic algorithm at the optimal point. It is available at http://www.ics.uci.edu/∼fastmm/. Earlier experiments put the crossover point where Strassen’s algorithm beats the cubic algorithm at about n = 128 [BLS91, CR76]. Thus, an O(n3 ) algorithm will likely be your best bet unless your matrices are very large. The linear algebra library of choice is LAPACK, a descendant of LINPACK [DMBS79], which includes several routines for matrix multiplication. These Fortran codes are available from Netlib as discussed in Section 19.1.5 (page 659). Algorithm 601 [McN83] of the Collected Algorithms of the ACM is a sparse matrix package written in Fortran that includes routines to multiply any combination of sparse and dense matrices. See Section 19.1.5 (page 659) for details. Notes: Winograd’s algorithm for fast matrix multiplication reduces the number of multiplications by a factor of two over the straightforward algorithm. It is implementable, although the additional bookkeeping required makes it doubtful whether it is a win. Expositions on Winograd’s algorithm [Win68] include [CLRS01, Man89, Win80]. 13.3 MATRIX MULTIPLICATION In my opinion, the history of theoretical algorithm design began when Strassen [Str69] published his O(n2.81 )-time matrix multiplication algorithm. For the ﬁrst time, improving an algorithm in the asymptotic sense became a respected goal in its own right. Progressive improvements to Strassen’s algorithm have gotten progressively less practical. The current best result for matrix multiplication is Coppersmith and Winograd’s [CW90] O(n2.376 ) algorithm, while the conjecture is that Θ(n2 ) suﬃces. See [CKSU05] for an alternate approach that recently yielded an O(n2.41 ) algorithm. Engineering eﬃcient matrix multiplication algorithms requires careful management of cache memory. See [BDN01, HUW02] for studies on these issues. The interest in the squares of graphs goes beyond counting paths. Fleischner [Fle74] proved that the square of any biconnected graph has a Hamiltonian cycle. See [LS95] for results on ﬁnding the square roots of graphs—i.e. , ﬁnding A given A2 . The problem of Boolean matrix multiplication can be reduced to that of general matrix multiplication [CLRS01]. The four-Russians algorithm for Boolean matrix multiplication [ADKF70] uses preprocessing to construct all subsets of lg n rows for fast retrieval in performing the actual multiplication, yielding a complexity of O(n3 / lg n). Additional preprocessing can improve this to O(n3 / lg2 n) [Ryt85]. An exposition on the four-Russians algorithm, including this speedup, appears in [Man89]. Good expositions of the matrix-chain algorithm include [BvG99, CLRS01], where it is given as a standard textbook example of dynamic programming. Related Problems: Solving linear equations (see page 395), shortest path (see page 489). 403 404 13. NUMERICAL PROBLEMS -2 10 -3 13 2 * det 2 1 3 + 4 10 10 5 13 13 3 * det -1 * det det 4 5 -2 -3 INPUT 13.4 + 4 -2 5 -3 =0 OUTPUT Determinants and Permanents Input description: An n × n matrix M . Problem description: What is the determinant |M | or permanent perm(M ) of the matrix m? Discussion: Determinants of matrices provide a clean and useful abstraction in linear algebra that can be used to solve a variety of problems: • Testing whether a matrix is singular, meaning that the matrix does not have an inverse. A matrix M is singular iﬀ |M | = 0. • Testing whether a set of d points lies on a plane in fewer than d dimensions. If so, the system of equations they deﬁne is singular, so |M | = 0. • Testing whether a point lies to the left or right of a line or plane. This problem reduces to testing whether the sign of a determinant is positive or negative, as discussed in Section 17.1 (page 564). • Computing the area or volume of a triangle, tetrahedron, or other simplicial complex. These quantities are a function of the magnitude of the determinant, as discussed in Section 17.1 (page 564). The determinant of a matrix M is deﬁned as a sum over all n! possible permutations πi of the n columns of M : |M | = n! i=1 n (−1)sign(πi ) M [j, πj ] j=1 13.4 DETERMINANTS AND PERMANENTS where sign(πi ) denotes the number of pairs of elements out of order (called inversions) in permutation πi . A direct implementation of this deﬁnition yields an O(n!) algorithm, as does the cofactor expansion method I learned in high school. Better algorithms to evaluate determinants are based on LU-decomposition, discussed in Section 13.1 (page 395). The determinant of M is simply the product of the diagonal elements of the LUdecomposition of M , which can be found in O(n3 ) time. A closely related function called the permanent arises in combinatorial problems. For example, the permanent of the adjacency matrix of a graph G counts the number of perfect matchings in G. The permanent of a matrix M is deﬁned by perm(M ) = n! n M [j, πj ] i=1 j=1 diﬀering from the determinant only in that all products are positive. Surprisingly, it is NP-hard to compute the permanent, even though the determinant can easily be computed in O(n3 ) time. The fundamental diﬀerence is that det(AB) = det(A) × det(B), while perm(AB) = perm(A) × perm(B). There are permanent algorithms running in O(n2 2n ) time that prove to be considerably faster than the O(n!) deﬁnition. Thus, ﬁnding the permanent of a 20 × 20 matrix is not out of the realm of possibility. Implementations: The linear algebra package LINPACK contains a variety of Fortran routines for computing determinants, optimized for diﬀerent data types and matrix structures. It can be obtained from Netlib, as discussed in Section 19.1.5 (page 659). JScience provides an extensive linear algebra package (including determinants) as part of its comprehensive scientiﬁc computing library. JAMA is another matrix package written in Java. Links to both and many related libraries are available from http://math.nist.gov/javanumerics/. Nijenhuis and Wilf [NW78] provide an eﬃcient Fortran routine to compute the permanent of a matrix. See Section 19.1.10 (page 661). Cash [Cas95] provides a C routine to compute the permanent, motivated by the Kekulé structure count of computational chemistry. Two diﬀerent codes for approximating the permanent are provided by Barvinok. The ﬁrst, based on [BS07], provides codes for approximating the permanent and a Hafnian of a matrix, as well as the number of spanning forests in a graph. See http://www.math.lsa.umich.edu/∼barvinok/manual.html. The second, based on [SB01], can provide estimates of the permanent of 200 × 200 matrices in seconds. See http://www.math.lsa.umich.edu/∼barvinok/code.html. Notes: Cramer’s rule reduces the problems of matrix inversion and solving linear systems to that of computing determinants. However, algorithms based on LU-determination are faster. See [BM53] for an exposition on Cramer’s rule. 405 406 13. NUMERICAL PROBLEMS Determinants can be computed in o(n3 ) time using fast matrix multiplication, as shown in [AHU83]. Section 13.3 (page 401) discusses such algorithms. A fast algorithm for computing the sign of the determinant—an important problem for performing robust geometric computations—is due to Clarkson [Cla92]. The problem of computing the permanent was shown to be #P-complete by Valiant [Val79], where #P is the class of problems solvable on a “counting” machine in polynomial time. A counting machine returns the number of distinct solutions to a problem. Counting the number of Hamiltonian cycles in a graph is a #P-complete problem that is trivially NP-hard (and presumably harder), since any count greater than zero proves that the graph is Hamiltonian. Counting problems can be #P-complete even if the corresponding decision problem can be solved in polynomial time, as shown by the permanent and perfect matchings. Minc [Min78] is the primary reference on permanents. A variant of an O(n2 2n )-time algorithm due to Ryser for computing the permanent is presented in [NW78]. Recently, probabilistic algorithms have been developed for estimating the permanent, culminating in a fully-polynomial randomized approximation scheme that provides an arbitrary close approximation in time that depends polynomially upon the input matrix and the desired error [JSV01]. Related Problems: Solving linear systems (see page 395), matching (see page 498), geometric primitives (see page 564). 13.5 CONSTRAINED AND UNCONSTRAINED OPTIMIZATION MAX MIN INPUT 13.5 OUTPUT Constrained and Unconstrained Optimization Input description: A function f (x1 , . . . , xn ). Problem description: What point p = (p1 , . . . , pn ) maximizes (or minimizes) the function f ? Discussion: Most of this book concerns algorithms that optimize one thing or another. This section considers the general problem of optimizing functions where, due to lack of structure or knowledge, we are unable to exploit the problem-speciﬁc algorithms seen elsewhere in this book. Optimization arises whenever there is an objective function that must be tuned for optimal performance. Suppose we are building a program to identify good stocks to invest in. We have certain ﬁnancial data available to analyze—such as the priceearnings ratio, the interest rate, and the stock price—all as a function of time t. The key question is how much weight we should give to each of these factors, where these weights correspond to coeﬃcients of a formula: stock-goodness(t) = c1 × price(t) + c2 × interest(t) + c3 × PE-ratio(t) We seek the numerical values c1 , c2 , c3 whose stock-goodness function does the best job of evaluating stocks. Similar issues arise in tuning evaluation functions for any pattern recognition task. Unconstrained optimization problems also arise in scientiﬁc computation. Physical systems from protein structures to galaxies naturally seek to minimize their “energy” or “potential function.” Programs that attempt to simulate nature thus often deﬁne potential functions assigning a score to each possible object conﬁguration, and then select the conﬁguration that minimizes this potential. 407 408 13. NUMERICAL PROBLEMS Global optimization problems tend to be hard, and there are lots of ways to go about them. Ask the following questions to steer yourself in the right direction: • Am I doing constrained or unconstrained optimization? – In unconstrained optimization, there are no limitations on the values of the parameters other than that they maximize the value of f . However, many applications demand constraints on these parameters that make certain points illegal, points that might otherwise be the global optimum. For example, companies cannot employ less than zero employees, no matter how much money they think they might save doing so. Constrained optimization problems typically require mathematical programming approaches, like linear programming, discussed in Section 13.6 (page 411). • Is the function I am trying to optimize described by a formula? – Sometimes the function that you seek to optimize is presented as an algebraic formula, such as ﬁnding the minimum of f (n) = n2 − 6n + 2n+1 . If so, the solution is to analytically take its derivative f (n) and see for which points p we have f (p ) = 0. These points are either local maxima or minima, which can be distinguished by taking a second derivative or just plugging p back into f and seeing what happens. Symbolic computation systems such as Mathematica and Maple are fairly eﬀective at computing such derivatives, although using computer algebra systems eﬀectively is somewhat of a black art. They are deﬁnitely worth a try, however, and you can always use them to plot a picture of your function to get a better idea of what you are dealing with. • Is it expensive to compute the function at a given point? – Instead of a formula, we are often given a program or subroutine that evaluates f at a given point. Since we can request the value of any given point on demand by calling this function, we can poke around and try to guess the maxima. Our freedom to search in such a situation depends upon how eﬃciently we can evaluate f . Suppose that f (x1 , . . . , xn ) is the board evaluation function in a computer chess program, such that x1 is how much a pawn is worth, x2 is how much a bishop is worth, and so forth. To evaluate a set of coeﬃcients as a board evaluator, we must play a bunch of games with it or test it on a library of known positions. Clearly, this is time-consuming, so we would have to be frugal in the number of evaluations of f we use to optimize the coeﬃcients. • How many dimensions do we have? How many do we need? – The diﬃculty in ﬁnding a global maximum increases rapidly with the number of dimensions (or parameters). For this reason, it often pays to reduce the dimension by ignoring some of the parameters. This runs counter to intuition, for the naive programmer is likely to incorporate as many variables as possible into their evaluation function. It is just too hard, however, to optimize such complicated 13.5 CONSTRAINED AND UNCONSTRAINED OPTIMIZATION functions. Much better is to start with the three to ﬁve most important variables and do a good job optimizing the coeﬃcients for these. • How smooth is my function? The main diﬃculty of global optimization is getting trapped in local optima. Consider the problem of ﬁnding the highest point in a mountain range. If there is only one mountain and it is nicely shaped, we can ﬁnd the top by just walking in whatever direction is up. However, if there are many false summits, or other mountains in the area, it is diﬃcult to convince ourselves whether we really are at the highest point. Smoothness is the property that enables us to quickly ﬁnd the local optimum from a given point. We assume smoothness in seeking the peak of the mountain by walking up. If the height at any given point was a completely random function, there would be no way we could ﬁnd the optimum height short of sampling every single point. The most eﬃcient algorithms for unconstrained global optimization use derivatives and partial derivatives to ﬁnd local optima, to point out the direction in which moving from the current point does the most to increase or decrease the function. Such derivatives can sometimes be computed analytically, or they can be estimated numerically by taking the diﬀerence between the values of nearby points. A variety of steepest descent and conjugate gradient methods to ﬁnd local optima have been developed—similar in many ways to numerical root-ﬁnding algorithms. It is a good idea to try several diﬀerent methods on any given optimization problem. For this reason, we recommend experimenting with the implementations below before attempting to implement your own method. Clear descriptions of these algorithms are provided in several numerical algorithms books, in particular Numerical Recipes [PFTV07]. For constrained optimization, ﬁnding a point that satisﬁes all the constraints is often the diﬃcult part of the problem. One approach is to use a method for unconstrained optimization, but add a penalty according to how many constraints are violated. Determining the right penalty function is problem-speciﬁc, but it often makes sense to vary the penalties as optimization proceeds. At the end, the penalties should be very high to ensure that all constraints are satisﬁed. Simulated annealing is a fairly robust and simple approach to constrained optimization, particularly when we are optimizing over combinatorial structures (permutations, graphs, subsets). Techniques for simulated annealing are described in Section 7.5.3 (page 254). Implementations: The world of constrained/unconstrained optimization is suﬃciently confusing that several guides have been created to point people to the right codes. Particularly nice is Hans Mittlemann’s Decision Tree for Optimization Software at http://plato.asu.edu/guide.html. Also check out the selection at GAMS, the NIST Guide to Available Mathematical Software, at http://gams.nist.gov. NEOS (Network-Enabled Optimization System) provides a unique service—the opportunity to solve your problem remotely on computers and software at Argonne 409 410 13. NUMERICAL PROBLEMS National Laboratory. Linear programming and unconstrained optimization are both supported. Check out http://www-neos.mcs.anl.gov/ when you need a solution instead of a program. Several of the Collected Algorithms of the ACM are Fortran codes for unconstrained optimization, most notably Algorithm 566 [MGH81] , Algorithm 702 [SF92] , and Algorithm 734 [Buc94] . Algorithm 744 [Rab95] does unconstrained optimization in Lisp. They are all available from Netlib (see Section 19.1.5 (page 659)). General purpose simulated annealing implementations are available, and probably are the best place to start experimenting with this technique for constrained optimization. Feel free to try my code from Section 7.5.3 (page 254). Particularly popular is Adaptive Simulated Annealing (ASA), written in C by Lester Ingber and available at http://asa-caltech.sourceforge.net/. Both the Java Genetic Algorithms Package (JGAP) (http:// jgap.sourceforge.net/) and the C Genetic Algorithm Utility Library (GAUL) (http://gaul.sourceforge.net/) are designed to aid in the development of applications that use genetic/evolutionary algorithms. I am highly skeptical about genetic algorithms (see Section 7.8 (page 266)), but other people seem to ﬁnd them irresistible. Notes: Steepest-descent methods for unconstrained optimization are discussed in most books on numerical methods, including [BT92, PFTV07]. Unconstrained optimization is the topic of several books, including [Bre73, Fle80]. Simulated annealing was devised by Kirkpatrick et al. [KGV83] as a modern variation of the Metropolis algorithm [MRRT53]. Both use Monte Carlo techniques to compute the minimum energy state of a system. Good expositions on all local search variations, including simulated annealing, appear in [AL97]. Genetic algorithms were developed and popularized by Holland [Hol75, Hol92]. More sympathetic expositions on genetic algorithms include [LP02, MF00]. Tabu search [Glo90] is yet another heuristic search procedure with a devoted following. Related Problems: Linear programming (see page 411), satisﬁability (see page 472). 13.6 LINEAR PROGRAMMING INPUT 13.6 OUTPUT Linear Programming Input description: A set S of n linear inequalities on m variables Si := m cij · xj ≥ bi , 1 ≤ i ≤ n j=1 and a linear optimization function f (X) = m j=1 cj · xj . Problem description: Which variable assignment X maximizes the objective function f while satisfying all inequalities S? Discussion: Linear programming is the most important problem in mathematical optimization and operations research. Applications include: • Resource allocation – We seek to invest a given amount of money to maximize our return. Often our possible options, payoﬀs, and expenses can be expressed as a system of linear inequalities such that we seek to maximize our possible proﬁt given the constraints. Very large linear programming problems are routinely solved by airlines and other corporations. • Approximating the solution of inconsistent equations – A set of m linear equations on n variables xi , 1 ≤ i ≤ n is overdetermined if m > n. Such overdetermined systems are often inconsistent, meaning that there does not exist an assignment to the variables that simultaneously solves all the equations. To ﬁnd the assignment that best ﬁts the equations, we can replace each variable xi by xi + i and solve the new system as a linear program, minimizing the sum of the error terms. 411 412 13. NUMERICAL PROBLEMS • Graph algorithms – Many of the graph problems described in this book, including shortest path, bipartite matching, and network ﬂow, can be solved as special cases of linear programming. Most of the rest, including traveling salesman, set cover, and knapsack, can be solved using integer linear programming. The simplex method is the standard algorithm for linear programming. Each constraint in a linear programming problem acts like a knife that carves away a region from the space of possible solutions. We seek the point within the remaining region that maximizes (or minimizes) f (X). By appropriately rotating the solution space, the optimal point can always be made to be the highest point in the region. The region (simplex) formed by the intersection of a set of linear constraints is convex, so unless we are at the top there is always a higher vertex neighboring any starting point. When we cannot ﬁnd a higher neighbor to walk to, we have reached the optimal solution. While the simplex algorithm is not too complex, there is considerable art to producing an eﬃcient implementation capable of solving large linear programs. Large programs tend to be sparse (meaning that most inequalities use few variables), so sophisticated data structures must be used. There are issues of numerical stability and robustness, as well as choosing which neighbor we should walk to next (socalled pivoting rules). There also exist sophisticated interior-point methods, which cut through the interior of the simplex instead of walking along the outside, and beat simplex in many applications. The bottom line on linear programming is this: you are much better oﬀ using an existing LP code than writing your own. Further, you are probably better oﬀ paying money than surﬁng the Web. Linear programming is an algorithmic problem of such economic importance that commercial implementations are far superior to free versions. Issues that arise in linear programming include: • Do any variables have integrality constraints? – It is impossible to send 6.54 airplanes from New York to Washington each business day, even if that value maximizes proﬁt according to your model. Such variables often have natural integrality constraints. A linear program is called an integer program if all its variables have integrality constraints, or a mixed integer program if some of them do. Unfortunately, it is NP-complete to solve integer or mixed programs to optimality. However, there are integer programming techniques that work reasonably well in practice. Cutting plane techniques solve the problem ﬁrst as a linear program, and then add extra constraints to enforce integrality around the optimal solution point before solving it again. After suﬃciently many iterations, the optimum point of the resulting linear program matches that of the original integer program. As with most exponential-time algorithms, 13.6 LINEAR PROGRAMMING run times for integer programming depend upon the diﬃculty of the problem instance and are unpredictable. • Do I have more variables or constraints? – Any linear program with m variables and n inequalities can be written as an equivalent dual linear program with n variables and m inequalities. This is important to know, because the running time of a solver might be quite diﬀerent on the two formulations. In general, linear programs (LPs) with much more variables than constraints should be solved directly. If there are many more constraints than variables, it is usually better to solve the dual LP or (equivalently) apply the dual simplex method to the primal LP. • What if my optimization function or constraints are not linear? – In leastsquares curve ﬁtting, we seek the line that best approximates a set of points by minimizing the sum of squares of the distance between each point and the line. In formulating this as a mathematical program, the natural objective function is no longer linear, but quadratic. Although fast algorithms exist for least squares ﬁtting, general quadratic programming is NP-complete. There are three possible courses of action when you must solve a nonlinear program. The best is if you can model it in some other way, as is the case with least-squares ﬁtting. The second is to try to track down special codes for quadratic programming. Finally, you can model your problem as a constrained or unconstrained optimization problem and try to solve it with the codes discussed in Section 13.5 (page 407). • What if my model does not match the input format of my LP solver? – Many linear programming implementations accept models only in so-called standard form, where all variables are constrained to be nonnegative, the object function must be minimized, and all constraints must be equalities (instead of inequalities). Do not fear. There exist standard transformations to map arbitrary LP models into standard form. To convert a maximization problem to a minimization one, simply multiply each coeﬃcient of the objective function by −1. The remaining problems can be solved by adding slack variables to the model. See any textbook on linear programming for details. Modeling languages such as AMPC can provide a nice interface to your solver and deal with these issues for you. Implementations: The USENET Frequently Asked Question (FAQ) list is a very useful resource on solving linear programs. In particular, it provides a list of available codes with descriptions of experiences. Check it out at http://wwwunix.mcs.anl.gov/otc/Guide/faq/. There are at least three reasonable choices in free LP-solvers. Lp solve, written in ANSI C by Michel Berkelaar, can also handle integer and mixed-integer 413 414 13. NUMERICAL PROBLEMS problems. It is available at http://lpsolve.sourceforge.net/5.5/, and a substantial user community exists. The simplex solver CLP produced under the Computational Infrastructure for Operations Research is available (with other optimization software) at http://www.coin-or.org/. Finally, the GNU Linear Programming Kit (GLPK) is intended for solving large-scale linear programming, mixed integer programming (MIP), and other related problems. It is available at http://www.gnu.org/software/glpk/. In recent benchmarks among free codes (see http://plato.asu.edu/bench.html), CLP appeared to be fastest on on linear programming problems and lp solve on mixed integer problems. NEOS (Network-Enabled Optimization System) provides an opportunity to solve your problem on computers and software at Argonne National Laboratory. Linear programming and unconstrained optimization are both supported. This is worth checking out at http://www.mcs.anl.gov/home/otc/Server/ if you need an answer instead of a program. Algorithm 551 [Abd80] and Algorithm 552 [BR80] of the Collected Algorithms of the ACM are simplex-based codes for solving overdetermined systems of linear equations in Fortran. See Section 19.1.5 (page 659) for details. Notes: The need for optimization via linear programming arose in logistics problems in World War II. The simplex algorithm was invented by George Danzig in 1947 [Dan63]. Klee and Minty [KM72] proved that the simplex algorithm is exponential in worst case, but it is very eﬃcient in practice. Smoothed analysis measures the complexity of algorithms assuming that their inputs are subject to small amounts of random noise. Carefully constructed worst-case instances for many problems break down under such perturbations. Spielman and Teng [ST04] used smoothed analysis to explain the eﬃciency of simplex in practice. Recently, Kelner and Spielman developed a randomized simplex algorithm running in polynomial time [KS05b]. Khachian’s ellipsoid algorithm [Kha79] ﬁrst proved that linear programming was polynomial in 1979. Karmarkar’s algorithm [Kar84] is an interior-point method that has proven to be both a theoretical and practical improvement of the ellipsoid algorithm, as well as a challenge for the simplex method. Good expositions on the simplex and ellipsoid algorithms for linear programming include [Chv83, Gas03, MG06]. Semideﬁnite programming deals with optimization problems over symmetric positive semideﬁnite matrix variables, with linear cost function and linear constraints. Important special cases include linear programming and convex quadratic programming with convex quadratic constraints. Semideﬁnite programming and its applications to combinatorial optimization problems are surveyed in [Goe97, VB96]. Linear programming is P-complete under log-space reductions [DLR79]. This makes it unlikely to have an NC parallel algorithm, where a problem is in NC iﬀ it can be solved on a PRAM in polylogarithmic time using a polynomial number of processors. Any problem that is P-complete under log-space reduction cannot be in NC unless P=NC. See [GHR95] for a thorough exposition of the theory of P-completeness, including an extensive list of P-complete problems. Related Problems: Constrained and unconstrained optimization (see page 407), network ﬂow (see page 509). 13.7 RANDOM NUMBER GENERATION HTHTHHTHHT ? INPUT 13.7 HHHTTHHTTT HHTTTTHTHT HTHHHTTHHT HTTHHTTHTH OUTPUT Random Number Generation Input description: Nothing, or perhaps a seed. Problem description: Generate a sequence of random integers. Discussion: Random numbers have an enormous variety of interesting and important applications. They form the foundation of simulated annealing and related heuristic optimization techniques. Discrete event simulations run on streams of random numbers, and are used to model everything from transportation systems to casino poker. Passwords and cryptographic keys are typically generated randomly. Randomized algorithms for graph and geometric problems are revolutionizing these ﬁelds and establishing randomization as one of the fundamental ideas of computer science. Unfortunately, generating random numbers looks a lot easier than it really is. Indeed, it is fundamentally impossible to produce truly random numbers on any deterministic device. Von Neumann [Neu63] said it best: “Anyone who considers arithmetical methods of producing random digits is, of course, in a state of sin.” The best we can hope for are pseudorandom numbers, a stream of numbers that appear as if they were generated randomly. There can be serious consequences to using a bad random-number generator. In one famous case, a Web browser’s encryption scheme was broken with the discovery that the seeds of its random-number generator employed too few random bits [GW96]. Simulation accuracy is regularly compromised or invalidated by poor random number generation. This is an area where people shouldn’t mess around, but they do. Issues to think about include: • Should my program use the same “random” numbers each time it runs? – A poker game that deals you the exact same hand each time you play quickly loses interest. One common solution is to use the lower-order bits of the machine clock as the seed or starting point for a stream of random numbers, so that each time the program runs it does something diﬀerent. 415 416 13. NUMERICAL PROBLEMS Such methods are adequate for games, but not for serious simulations. There are liable to be periodicities in the distribution of random numbers whenever calls are made in a loop. Also, debugging is seriously complicated when program results are not repeatable. Should your program crash, you cannot go back and discover why. A possible compromise is to use a deterministic pseudorandom-number generator, but write the current seed to a ﬁle between runs. During debugging, this ﬁle can be overwritten with a ﬁxed initial value of the seed. • How good is my compiler’s built-in random number generator? – If you need uniformly-generated random numbers, and won’t be betting the farm on the accuracy of your simulation, my recommendation is simply to use what your compiler provides. Your best opportunity to mess things up is with a bad choice of starting seed, so read the manual for its recommendations. If you are going to bet the farm on the results of your simulation, you had better test your random number generator. Be aware that it is very diﬃcult to eyeball the results and decide whether the output is really random. This is because people have very skewed ideas of how random sources should behave and often see patterns that don’t really exist. Several diﬀerent tests should be used to evaluate a random number generator, and the statistical signiﬁcance of the results established. The National Institute of Standards and Technology (NIST) has developed a test suite for evaluating random number generators, discussed below. • What if I must implement my own random-number generator? – The standard algorithm of choice is the linear congruential generator. It is fast, simple, and (if instantiated with the right constants) gives reasonable pseudorandom numbers. The nth random number Rn is a function of the (n − 1)st random number: Rn = (aRn−1 + c) mod m In theory, linear congruential generators work the same way roulette wheels do. The long path of the ball around and around the wheel (captured by aRn−1 + c) ends in one of a relatively small number of bins, the choice of which is extremely sensitive to the length of the path (captured by the mod mtruncation). A substantial theory has been developed to select the constants a, c, m, and R0 . The period length is largely a function of the modulus m, which is typically constrained by the word length of the machine. Note that the stream of numbers produced by a linear congruential generator repeats the instant the ﬁrst number repeats. Further, computers are fast enough to make 232 calls to a random-number generator in a few minutes. Thus, any 32-bit linear congruential generator is in danger of cycling, motivating generators with signiﬁcantly longer periods. 13.7 RANDOM NUMBER GENERATION • What if I don’t want such large random numbers? – The linear congruential generator produces a uniformly-distributed sequence of large integers that can be easily scaled to produce other uniform distributions. For uniformly distributed real numbers between 0 and 1, use Ri /m. Note that 1 cannot be realized this way, although 0 can. If you want uniformly distributed integers between l and h, use l + (h − l + 1)Ri /m . • What if I need nonuniformly distributed random numbers? – Generating random numbers according to a given nonuniform distribution can be a tricky business. The most reliable way to do this correctly is the acceptance-rejection method. We can bound the desired geometric region to sample from by a box and then select a random point p from the box. This point can be generated by selecting the x and y coordinates independently at random. If it lies within the region of interest, we can return p as being selected at random. Otherwise we throw it away and repeat with another random point. Essentially, we throw darts at random and report those that hit the target. This method is correct, but it can be slow. If the volume of the region of interest is small relative to that of the box, most of our darts will miss the target. Eﬃcient generators for Gaussian and other special distributions are described in the references and implementations below. Be cautious about inventing your own technique, however, since it can be tricky to obtain the right probability distribution. For example, an incorrect way to select points uniformly from a circle of radius r would be to generate polar coordinates by selecting an angle from 0 to 2π and a displacement between 0 and r—both uniformly at random. In such a scheme, half the generated points will lie within r/2 of the center, when only one-fourth of them should be! This is diﬀerent enough to seriously skew the results, while being suﬃciently subtle that it can easily escape detection. • How long should I run my Monte Carlo simulation to get the best results? – The longer you run a simulation, the more accurately the results should approximate the limiting distribution, thus increasing accuracy. However, this is true only until you exceed the period, or cycle length, of your randomnumber generator. At that point, your sequence of random numbers repeats itself, and further runs generate no additional information. Instead of jacking up the length of a simulation run to the max, it is usually more informative to do many shorter runs (say 10 to 100) with diﬀerent seeds and then consider the range of results you see. The variance provides a healthy measure of the degree to which your results are repeatable. This exercise corrects the natural tendency to see a simulation as giving “the” correct answer. 417 418 13. NUMERICAL PROBLEMS Implementations: See http://random.mat.sbg.ac.at for an excellent website on random-number generation and stochastic simulation. It includes pointers to papers and literally dozens of implementations of random-number generators. Parallel simulations make special demands on random-number generators. How can we ensure that random streams are independent on each machine? L’Ecuyer et.al. [LSCK02] provide object-oriented generators with a period length of approximately 2191 . Implementations in C, C++, and Java are available at http://www.iro.umontreal.ca/∼lecuyer/myftp/streams00/. Independent streams of random numbers are supported for parallel applications. Another possibility is the Scalable Parallel Random Number Generators Library (SPRNG) [MS00], available at http://sprng.cs.fsu.edu/. Algorithms 488 [Bre74], 599 [AKD83], and 712 [Lev92] of the Collected Algorithms of the ACM are Fortran codes for generating non-uniform random numbers according to several probability distributions, including the normal, exponential, and Poisson distributions. They are available from Netlib (see Section 19.1.5 (page 659)). The National Institute of Standards [RSN+ 01] has prepared an extensive statistical test suite to validate random number generators. Both the software and the report describing it are available at http://csrc.nist.gov/rng/. True random-number generators extract random bits by observing physical processes. The website http://www.random.org makes available random numbers derived from atmospheric noise that passes the NIST statistical tests. This is an amusing solution if you need a small quantity of random numbers (say, to run a lottery) instead a random-number generator. Notes: Knuth [Knu97b] provides a thorough treatment of random-number generation, which I heartily recommend. He presents the theory behind several methods, including the middle-square and shift-register methods we have not described here, as well as a detailed discussion of statistical tests for validating random-number generators. That said, see [Gen04] for more recent developments in random number generation. The Mersenne twister [MN98] is a fast random number generator of period 219937 − 1. Other modern methods include [Den05, PLM06]. Methods for generating nonuniform random variates are surveyed in [HLD04]. Comparisons of diﬀerent random-number generators in practice include [PM88]. Tables of random numbers appear in most mathematical handbooks as relics from the days before there was ready access to computers. Most notable is [RC55], which provides one million random digits. The deep relationship between randomness, information, and compressibility is explored within the theory of Kolmogorov complexity, which measures the complexity of a string by its compressibility. Truly random strings are incompressible. The string of seemingly random digits of π cannot be random under this deﬁnition, since the entire sequence is deﬁned by any program implementing a series expansion for π. Li and Vitáni [LV97] provide a thorough introduction to the theory of Kolmogorov complexity. 13.7 RANDOM NUMBER GENERATION Related Problems: Constrained and unconstrained optimization (see page 407), generating permutations (see page 448), generating subsets (see page 452), generating partitions (see page 456). 419 420 13. NUMERICAL PROBLEMS 179424673 8338169264555846052842102071 2038074743 * 22801763489 8338169264555846052842102071 INPUT 13.8 OUTPUT Factoring and Primality Testing Input description: An integer n. Problem description: Is n a prime number, and if not what are its factors? Discussion: The dual problems of integer factorization and primality testing have surprisingly many applications for a problem long suspected of being only of mathematical interest. The security of the RSA public-key cryptography system (see Section 18.6 (page 641)) is based on the computational intractability of factoring large integers. As a more modest application, hash table performance typically improves when the table size is a prime number. To get this beneﬁt, an initialization routine must identify a prime near the desired table size. Finally, prime numbers are just interesting to play with. It is no coincidence that programs to generate large primes often reside in the games directory of UNIX systems. Factoring and primality testing are clearly related problems, although they are quite diﬀerent algorithmically. There exist algorithms that can demonstrate that an integer is composite (i.e. , not prime) without actually giving the factors. To convince yourself of the plausibility of this, note that you can demonstrate the compositeness of any nontrivial integer whose last digit is 0, 2, 4, 5, 6, or 8 without doing the actual division. The simplest algorithm for both of these problems is brute-force trial division. √ To factor n, compute the remainder of n/i for all 1 < i ≤ n. The prime factorization of n will contain at least one instance of every i such that n/i = n/i , unless n is prime. Make sure √ you handle the multiplicities correctly, and account for any primes larger than n. Such algorithms can be sped up by using a precomputed table of small primes to avoid testing all possible i. Surprisingly large numbers of primes can be represented in surprisingly little space by using bit vectors (see Section 12.5 (page 385)). A bit vector of all odd numbers less than 1,000,000 ﬁts in under 64 kilobytes. Even tighter encodings become possible by eliminating all multiples of 3 and other small primes. 13.8 FACTORING AND PRIMALITY TESTING √ Although trial division runs in O( n) time, it is not a polynomial-time algorithm. The reason is that it only takes lg2 n bits to represent n, so trial division takes time exponential in the input size. Considerably faster (but still exponential time) factoring algorithms exist, whose correctness depends upon more substantial number theory. The fastest known algorithm, the number ﬁeld sieve, uses randomness to construct a system of congruences—the solution of which usually gives a factor of the integer. Integers with as many as 200 digits (663 bits) have been factored using this method, although such feats require enormous amounts of computation. Randomized algorithms make it much easier to test whether an integer is prime. Fermat’s little theorem states that an−1 = 1(mod n) for all a not divisible by n, provided n is prime. Suppose we pick a random value 1 ≤ a < n and compute the residue of an−1 (mod n). If this residue is not 1, we have just proven that n cannot be prime. Such randomized primality tests are very eﬃcient. PGP (see Section 18.6 (page 641)) ﬁnds 300+ digit primes using hundreds of these tests in minutes, for use as cryptographic keys. Although the primes are scattered in a seemingly random way throughout the integers, there is some regularity to their distribution. The prime number theorem states that the number of primes less than n (commonly denoted by π(n)) is approximately n/ ln n. Further, there never are large gaps between primes, so in general, one would expect to examine about ln n integers if one wanted to ﬁnd the ﬁrst prime larger than n. This distribution and the fast randomized primality test explain how PGP can ﬁnd such large primes so quickly. Implementations: Several general systems for computational number theory are available. PARI is capable of handling complex number-theoretic problems on arbitrary-precision integers (to be precise, limited to 80,807,123 digits on 32-bit machines), as well as reals, rationals, complex numbers, polynomials, and matrices. It is written mainly in C, with assembly code for inner loops on major architectures, and includes more than 200 special predeﬁned mathematical functions. PARI can be used as a library, but it also possesses a calculator mode that gives instant access to all the types and functions. PARI is available at http://pari.math.u-bordeaux.fr/ LiDIA (http://www.cdc.informatik.tu-darmstadt.de/TI/LiDIA/) is the acronym for the C++ Library for Computational Number Theory. It implements several of the modern integer factorization methods. A Library for doing Number Theory (NTL) is a high-performance, portable C++ library providing data structures and algorithms for manipulating signed, arbitrary length integers, and for vectors, matrices, and polynomials over the integers and over ﬁnite ﬁelds. It is available at http://www.shoup.net/ntl/. Finally, MIRACL (Multiprecision Integer and Rational Arithmetic C/C++ Library) implements six diﬀerent integer factorization algorithms, including the quadratic sieve. It is available at http://www.shamus.ie/. Notes: Expositions on modern algorithms for factoring and primality testing include Crandall and Pomerance [CP05] and Yan [Yan03]. More general surveys of computational number theory include Bach and Shallit [BS96] and Shoup [Sho05]. 421 422 13. NUMERICAL PROBLEMS In 2002, Agrawal, Kayal, and Saxena [AKS04] solved a long-standing open problem by giving the ﬁrst polynomial-time deterministic algorithm to test whether an integer is composite. Their algorithm, which is surprisingly elementary for such an important result, involves a careful analysis of techniques from earlier randomized algorithms. Its existence serves as somewhat of a rebuke to researchers (like me) who shy away from classical open problems due to fear. Dietzfelbinger [Die04] provides a self-contained treatment of this result. The Miller-Rabin [Mil76, Rab80] randomized primality testing algorithm eliminates problems with Carmichael numbers, which are composite integers that always satisfy Fermat’s theorem. The best algorithms for integer factorization include the quadraticsieve [Pom84] and the elliptic-curve methods [Len87b]. Mechanical sieving devices provided the fastest way to factor integers surprisingly far into the computing era. See [SWM95] for a fascinating account of one such device, built during World War I. Hand-cranked, it proved the primality of 231 − 1 in 15 minutes of sieving time. An important problem in computational complexity theory is whether P = NP ∩ co-NP. The decision problem “is n a composite number?” used to be the best candidate for a counterexample. By exhibiting the factors of n, it is trivially in NP. It can be shown to be in co-NP, since every prime has a short proof of its primality [Pra75]. The recent proof that composite numbers testing is in P shot down this line of reasoning. For more information on complexity classes, see [GJ79, Joh90]. The integer RSA-129 was factored in eight months using over 1,600 computers. This was particularly noteworthy because in the original RSA paper [RSA78] they had originally predicted such a factorization would take 40 quadrillion years using 1970s technology. Bahr, Boehm, Franke, and Kleinjung hold the current integer factorization record, with their successful attack on the 200-digit integer RSA-200 in May 2005. This required the equivalent of 55 years of computations on a single 2.2 GHz Opteron CPU. Related Problems: Cryptography (see page 641), high precision arithmetic (see page 423). 13.9 ARBITRARY-PRECISION ARITHMETIC 49578291287491495151508905425869578 2 74367436931237242727263358138804367 3 INPUT 13.9 OUTPUT Arbitrary-Precision Arithmetic Input description: Two very large integers, x and y. Problem description: What is x + y, x − y, x × y, and x/y? Discussion: Any programming language rising above basic assembler supports single- and perhaps double-precision integer/real addition, subtraction, multiplication, and division. But what if we wanted to represent the national debt of the United States in pennies? One trillion dollars worth of pennies requires 15 decimal digits, which is far more than can ﬁt into a 32-bit integer. Other applications require much larger integers. The RSA algorithm for publickey cryptography recommends integer keys of at least 1000 digits to achieve adequate security. Experimenting with number-theoretic conjectures for fun or research requires playing with large numbers. I once solved a minor open problem [GKP89] 1775 . by performing an exact computation on the integer 5906 2953 ≈ 9.93285 × 10 What should you do when you need large integers? • Am I solving a problem instance requiring large integers, or do I have an embedded application? – If you just need the answer to a speciﬁc problem with large integers, such as in the number theory application above, you should consider using a computer algebra system like Maple or Mathematica. These provide arbitrary-precision arithmetic as a default and use nice Lisplike programming languages as a front end—together often reducing your problem to a 5- to 10- line program. If you have an embedded application requiring high-precision arithmetic instead, you should use an existing arbitrary precision math library. You are likely to get additional functions beyond the four basic operations for computing things like greatest common divisor in the bargain. See the Implementations section for details. • Do I need high- or arbitrary-precision arithmetic? – Is there an upper bound on how big your integers can get, or do you really need arbitrary-precision— i.e. , unbounded. This determines whether you can use a ﬁxed-length array to represent your integers as opposed to a linked-list of digits. The array is likely to be simpler and will not prove a constraint in most applications. 423 424 13. NUMERICAL PROBLEMS • What base should I do arithmetic in? – It is perhaps simplest to implement your own high-precision arithmetic package in decimal, and thus represent each integer as a string of base-10 digits. However, it is far more eﬃcient to use a higher base, ideally equal to the square root of the largest integer supported fully by hardware arithmetic. Why? The higher the base, the fewer digits we need to represent a number. Compare 64 decimal with 1000000 binary. Since hardware addition usually takes one clock cycle independent of value of the actual numbers, best performance is achieved using the highest supported base. The factor limiting √ us to base b = maxint is the desire to avoid overﬂow when multiplying two of these “digits” together. The primary complication of using a larger base is that integers must be converted to and from base-10 for input and output. The conversion is easily performed once all four high-precision arithmetical operations are supported. • How low-level are you willing to get for fast computation? – Hardware addition is much faster than a subroutine call, so you take a signiﬁcant hit on speed using high-precision arithmetic when low-precision arithmetic suﬃces. High-precision arithmetic is one of few problems in this book where inner loops in assembly language prove the right idea to speed things up. Similarly, using bit-level masking and shift operations instead of arithmetical operations can be a win if you really understand the machine integer representation. The algorithm of choice for each of the ﬁve basic arithmetic operations is as follows: • Addition – The basic schoolhouse method of lining up the decimal points and then adding the digits from right to left with “carries” runs to time linear in the number of digits. More sophisticated carry-look-ahead parallel algorithms are available for low-level hardware implementation. They are presumably used on your microprocessor for low-precision addition. • Subtraction – By fooling with the sign bits of the numbers, subtraction can be a special considered case of addition: (A − (−B)) = (A + B). The tricky part of subtraction is performing the “borrow.” This can be simpliﬁed by always subtracting from the number with the larger absolute value and adjusting the signs afterwards, so we can be certain there will always be something to borrow from. • Multiplication – Repeated addition will take exponential time on large integers, so stay away from it. The digit-by-digit schoolhouse method is reasonable to program and will work much better, presumably well enough for your application. On very large integers, Karatsuba’s O(n1.59 ) divide-and-conquer algorithm wins. Dan Grayson, author of Mathematica’s arbitrary-precision arithmetic, found that the switch-over happened at well under 100 digits. 13.9 ARBITRARY-PRECISION ARITHMETIC Even faster for very large integers is an algorithm based on Fourier transforms. Such algorithms are discussed in Section 13.11 (page 431). • Division – Repeated subtraction will take exponential time, so the easiest reasonable algorithm to use is the long-division method you hated in school. This is fairly complicated, requiring arbitrary-precision multiplication and subtraction as subroutines, as well as trial-and-error, to determine the correct digit at each position of the quotient. In fact, integer division can be reduced to integer multiplication, although in a nontrivial way, so if you are implementing asymptotically fast multiplication, you can reuse that eﬀort in long division. See the references below for details. • Exponentiation – We can compute an using n − 1 multiplications, by computing a × a × . . . × a. However, a much better divide-and-conquer algorithm is based on the observation that n = n/2 + n/2. If n is even, then an = (an/2 )2 . If n is odd, then an = a(an/2 )2 . In either case, we have halved the size of our exponent at the cost of at most two multiplications, so O(lg n) multiplications suﬃce to compute the ﬁnal value: function power(a, n) if (n = 0) return(1) x = power(a, n/2 ) if (n is even) then return(x2 ) else return(a × x2 ) High- but not arbitrary-precision arithmetic can be conveniently performed using the Chinese remainder theorem and modular arithmetic. kThe Chinese remainder theorem states that an integer between 1 and P = i=1 pi is uniquely determined by its set of residues mod pi , where each pi , pj are relatively prime integers. Addition, subtraction, and multiplication (but not division) can be supported using such residue systems, with the advantage that large integers can be manipulated without complicated data structures. Many of these algorithms for computations on long integers can be directly applied to computations on polynomials. See the references for more details. A particularly n useful algorithm is Horner’s rule for fast polynomial evaluation. When P (x) = i=0 ci · xi is blindly evaluated term by term, O(n2 ) multiplications will be performed. Much better is observing that P (x) = c0 + x(c1 + x(c2 + x(c3 + . . .))), the evaluation of which uses only a linear number of operations. Implementations: All major commercial computer algebra systems incorporate high-precision arithmetic, including Maple, Mathematica, Axiom, and Macsyma. If you have access to one of these, this is your best option for a quick, nonembedded 425 426 13. NUMERICAL PROBLEMS application. The rest of this section focuses on source code available for embedded applications. The premier C/C++ library for fast, arbitrary-precision is the GNU Multiple Precision Arithmetic Library (GMP), which operates on signed integers, rational numbers, and ﬂoating point numbers. It is widely used and well-supported, and available at http://gmplib.org/. The java.math BigInteger class provides arbitrary-precision analogues to all of Java’s primitive integer operators. BigInteger provides additional operations for modular arithmetic, GCD calculation, primality testing, prime generation, bit manipulation, and a few other miscellaneous operations. A lower-performance, less-tested, but more personal implementation of highprecision arithmetic appears in the library from my book Programming Challenges [SR03]. See Section 19.1.10 (page 661) for details. Several general systems for computational number theory are available. Each of these supports operations of arbitrary-precision integers. Information about the PARI, LiDIA, NTL and MIRACL number-theoretic libraries can be found in Section 13.8 (page 420). ARPREC is a C++/Fortran-90 arbitrary precision package with an associated interactive calculator. MPFUN90 is a similar package written exclusively in Fortran-90. Both are available at http://crd.lbl.gov/∼dhbailey/mpdist/. Algorithm 693 [Smi91] of the Collected Algorithms of the ACM is a Fortran implementation of ﬂoating-point, multiple-precision arithmetic. See Section 19.1.5 (page 659). Notes: Knuth [Knu97b] is the primary reference on algorithms for all basic arithmetic operations, including implementations of them in the MIX assembly language. Bach and Shallit [BS96] and Shoup [Sho05] provide more recent treatments of computational number theory. Expositions on the O(n1.59 )-time divide-and-conquer algorithm for multiplication [KO63] include [AHU74, Man89]. An FFT-based algorithm multiplies two n-bit numbers in O(n lg n lg lg n) time and is due to Schönhage and Strassen [SS71]. Expositions include [AHU74, Knu97b]. The reduction between integer division and multiplication is presented in [AHU74, Knu97b]. Applications of fast multiplication to other arithmetic operations are presented by Bernstein [Ber04b] Good expositions of algorithms for modular arithmetic and the Chinese remainder theorem include [AHU74, CLRS01]. A good exposition of circuit-level algorithms for elementary arithmetic algorithms is [CLRS01]. Euclid’s algorithm for computing the greatest common divisor of two numbers is perhaps the oldest interesting algorithm. Expositions include [CLRS01, Man89]. Related Problems: Factoring integers (see page 420), cryptography (see page 641). 13.10 KNAPSACK PROBLEM 427 $40 $40 $90 $90 $50 $100 $100 $30 INPUT 13.10 $30 OUTPUT Knapsack Problem Input description: A set of items S = {1, . . . , n}, where item i has size si and value vi . A knapsack capacity is C. Problem description: Find the subset S ⊂ S that maximizes the value of i∈S vi , given that i∈S si ≤ C; i.e. , all the items ﬁt in a knapsack of size C. Discussion: The knapsack problem arises in resource allocation with ﬁnancial constraints. How do you select what things to buy given a ﬁxed budget? Everything has a cost and value, so we seek the most value for a given cost. The name knapsack problem invokes the image of the backpacker who is constrained by a ﬁxed-size knapsack, and so must ﬁll it only with the most useful and portable items. The most common formulation is the 0/1 knapsack problem, where each item must be put entirely in the knapsack or not included at all. Objects cannot be broken up arbitrarily, so it is not fair taking one can of Coke from a six-pack or opening a can to take just a sip. It is this 0/1 property that makes the knapsack problem hard, for a simple greedy algorithm ﬁnds the optimal selection when we are allowed to subdivide objects. We compute the “price per pound” for each item, and take the most expensive item or the biggest part thereof until the knapsack is full. Repeat with the next most expensive item. Unfortunately, the 0/1 constraint is usually inherent in most applications. $50 428 13. NUMERICAL PROBLEMS 7 4 5 6 7 4 5 6 Figure 13.1: Integer partition is a special case of the knapsack problem Issues that arise in selecting the best algorithm include: • Does every item have the same cost/value or the same size? – When all items are worth exactly the same amount, we maximize our value by taking the greatest number of items. Therefore, the optimal solution is to sort the items in order of increasing size and insert them into the knapsack in this order until no more ﬁt. The problem is similarly solved when each object has the same size but the costs are diﬀerent. Sort by cost, and take the cheapest elements ﬁrst. These are the easy cases of knapsack. • Does each item have the same “price per pound”? – In this case, our problem is equivalent to ignoring the price and just trying to minimize the amount of empty space left in the knapsack. Unfortunately, even this restricted version is NP-complete, so we cannot expect an eﬃcient algorithm that always solves the problem. Don’t lose hope, however, because knapsack proves to be an “easy” hard problem, and one that can usually be handled with the algorithms described below. An important special case of a constant “price-per-pound” knapsack is the integer partition problem, presented in cartoon form in Figure 13.1. Here, we seek to partition the elements of S into two sets A and B such that a = a∈A b∈B , or alternately make the diﬀerence as small as possible. Integer partition can be thought of as bin packing into two equal-sized bins or knapsack with a capacity of half the total weight, so all three problems are closely related and NP-complete. The constant “price-per-pound” knapsack problem is often called the subset sum problem, because we seek a subset of items that adds up to a speciﬁc target number C; i.e. , the capacity of our knapsack. • Are all the sizes relatively small integers? – When the sizes of the items and the knapsack capacity C are all integers, there exists an eﬃcient dynamic programming algorithm to ﬁnd the optimal solution in time O(nC) and O(C) space. Whether this works for you depends upon how big C is. It is great for C ≤ 1,000, but not so great for C ≥ 10,000,000. 13.10 KNAPSACK PROBLEM The algorithm works as follows: Let S be a set of items, and let C[i, S ] be true if and only if there is a subset of S whose size adds up exactly to i. Thus, C[i, ∅] is false for all 1 ≤ i ≤ C. One by one we add a new item sj to S and update the aﬀected values of C[i, S ]. Observe that C[i, S ∪ sj ] = true iﬀ C[i, S ] or C[i − sj , S ] is true, since we either use sj in realizing the sum or we don’t. We identify all sums that can be realized by performing n sweeps through all C elements—one for each sj , 1 ≤ j ≤ n—and so updating the array. The knapsack solution is given by the largest index of a true element of the largest realizable size. To reconstruct the winning subset, we must also store the name of the item number that turned C[i] from false to true for each 1 ≤ i ≤ C and then scan backwards through the array. This dynamic programming formulation ignores the values of the items. To generalize the algorithm, use each element of the array to store the value of the best subset to date summing up to i. We now update when the sum of the cost of C[i − sj , S ] plus the cost of sj is better than the previous cost of C[i]. • What if I have multiple knapsacks? – When there are multiple knapsacks, your problem might be better thought of as a bin-packing problem. Check out Section 17.9 (page 595) for bin-packing/cutting-stock algorithms. That said, algorithms for optimizing over multiple knapsacks are provided in the Implementations section below. Exact solutions for large capacity knapsacks can be found using integer prois used to denote whether gramming or backtracking. A 0/1 integer variable xi n item i is present in the optimal subset. We maximize i=1 xi · vi given the conn straint that i=1 xi · si ≤ C. Integer programming codes are discussed in Section 13.6 (page 411). Heuristics must be used when exact solutions prove too costly to compute. The simple greedy heuristic inserts items according to the maximum “price per pound” rule described previously. Often this heuristic solution is close to optimal, but it can be arbitrarily bad depending upon the problem instance. The “price per pound” rule can also be used to reduce the problem size in exhaustive search-based algorithms by eliminating “cheap but heavy” objects from future consideration. Another heuristic is based on scaling. Dynamic programming works well if the knapsack capacity is a reasonably small integer, say ≤ Cs . But what if we have a problem with capacity C > Cs ? We scale down the sizes of all items by a factor of C/Cs , round the size down to the nearest integer, and then use dynamic programming on the scaled items. Scaling works well in practice, especially when the range of sizes of items is not too large. Implementations: Martello and Toth’s collection of Fortran implementations of algorithms for a variety of knapsack problem variants are available at http://www.or.deis.unibo.it/kp.html. An electronic copy of the associated book [MT90a] has also been generously made available. 429 430 13. NUMERICAL PROBLEMS David Pisinger maintains a well-organized collection of C-language codes for knapsack problems and related variants like bin packing and container loading. These are available at http://www.diku.dk/∼pisinger/codes.html. The strongest code is based on the dynamic programming algorithm of [MPT99]. Algorithm 632 [MT85] of the Collected Algorithms of the ACM is a Fortran code for the 0/1 knapsack problem, with the twist that it supports multiple knapsacks. See Section 19.1.5 (page 659). Notes: Keller, Pferschy, and Pisinger [KPP04] is the most current reference on the knapsack problem and variants. Martello and Toth’s book [MT90a] and survey article [MT87] are standard references on the knapsack problem, including both theoretical and experimental results. An excellent exposition on integer programming approaches to knapsack problems appears in [SDK83]. See [MPT00] for a computational study of algorithms for 0-1 knapsack problems. A polynomial-time approximation scheme is an algorithm that approximates the optimal solution of a problem in time polynomial in both its size and the approximation factor . This very strong condition implies a smooth tradeoﬀ between running time and approximation quality. Good expositions on polynomial-time approximation schemes [IK75] for knapsack and subset sum includes [BvG99, CLRS01, GJ79, Man89]. The ﬁrst algorithm for generalized public key encryption by Merkle and Hellman [MH78] was based on the hardness of the knapsack problem. See [Sch96] for an exposition. Related Problems: Bin packing (see page 595), integer programming (see page 411). 13.11 DISCRETE FOURIER TRANSFORM 431 8 1.5 1 6 0.5 50 100 150 200 250 4 -0.5 2 -1 -1.5 50 INPUT 13.11 100 150 OUTPUT Discrete Fourier Transform Input description: A sequence of n real or complex values hi , 0 ≤ i ≤ n − 1, sampled at uniform intervals from a function h. n−1 Problem description: The discrete Fourier transform Hm = k=0 hk e2πikm/n for 0 ≤ m ≤ n − 1. Discussion: Although computer scientists tend to be fairly ignorant about Fourier transforms, electrical engineers and signal processors eat them for breakfast. Functionally, Fourier transforms provide a way to convert samples of a standard timeseries into the frequency domain. This provides a dual representation of the function in which certain operations become easier than in the time domain. Applications of Fourier transforms include: • Filtering – Taking the Fourier transform of a function is equivalent to representing it as the sum of sine functions. By eliminating undesirable highand/or low-frequency components (i.e. , dropping some of the sine functions) and taking an inverse Fourier transform to get us back into the time domain, we can ﬁlter an image to remove noise and other artifacts. For example, the sharp spike in the ﬁgure above represents the period of the single sine function that closely models the input data. The rest is noise. • Image compression – A smoothed, ﬁltered image contains less information than the original, while retaining a similar appearance. By eliminating the coeﬃcients of sine functions that contribute relatively little to the image, we can reduce the size of the image at little cost in image ﬁdelity. • Convolution and deconvolution – Fourier transforms can eﬃciently compute convolutions of two sequences. A convolution is the pairwise product of elements from two diﬀerent sequences, such as in multiplying two n-variable 200 250 432 13. NUMERICAL PROBLEMS polynomials f and g or comparing two character strings. Implementing such products directly takes O(n2 ), while the fast Fourier transform led to a O(n lg n) algorithm. Another example comes from image processing. Because a scanner measures the darkness of an image patch instead of a single point, the scanned input is always blurred. A reconstruction of the original signal can be obtained by deconvoluting the input signal with a Gaussian point-spread function. • Computing the correlation of functions – The correlation function of two functions f (t) and g(t) is deﬁned by ∞ f (τ )g(t + τ )dτ z(t) = −∞ and can be easily computed using Fourier transforms. When two functions are similar in shape but one is shifted relative to the other (such as f (t) = sin(t) and g(t) = cos(t)), the value of z(t0 ) will be large at this shift oﬀset t0 . As an application, suppose that we want to detect whether there are any funny periodicities in our random-number generator. We can generate a large series of random numbers, turn them into a time series (the ith number at time i), and take the Fourier transform of this series. Any funny spikes will correspond to potential periodicities. The discrete Fourier transform takes as input n complex numbers hk , 0 ≤ k ≤ n − 1, corresponding to equally spaced points in a time series, and outputs n complex numbers Hk , 0 ≤ k ≤ n − 1, each describing a sine function of given frequency. The discrete Fourier transform is deﬁned by Hm = n−1 hk e−2πikm/n k=0 and the inverse Fourier transform is deﬁned by hm = n−1 1 Hk e2πikm/n n k=0 which enables us move easily between h and H. Since the output of the discrete Fourier transform consists of n numbers, each of which is computed using a formula on n numbers, they can be computed in O(n2 ) time. The fast Fourier transform (FFT) is an algorithm that computes the discrete Fourier transform in O(n log n). This is arguably the most important algorithm known, for it opened the door to modern signal processing. Several diﬀerent algorithms call themselves FFTs, all of which are based on a divide-and-conquer approach. Essentially, the problem of computing the discrete Fourier transform on 13.11 DISCRETE FOURIER TRANSFORM n points is reduced to computing two transforms on n/2 points each, and is then applied recursively. The FFT usually assumes that n is a power of two. If this is not the case, you are usually better oﬀ padding your data with zeros to create n = 2k elements rather than hunting for a more general code. Many signal-processing systems have strong real-time constraints, so FFTs are often implemented in hardware, or at least in assembly language tuned to the particular machine. Be aware of this possibility if the codes prove too slow. Implementations: FFTW is a C subroutine library for computing the discrete Fourier transform in one or more dimensions, with arbitrary input size, and supporting both real and complex data. It is the clear choice among freely available FFT codes. Extensive benchmarking proves it to be the “Fastest Fourier Transform in the West.” Interfaces to Fortran and C++ are provided. FFTW received the 1999 J. H. Wilkinson Prize for Numerical Software. It is available at http://www.ﬀtw.org/. FFTPACK is a package of Fortran subprograms for the fast Fourier transform of periodic and other symmetric sequences, written by P. Swartzrauber. It includes complex, real, sine, cosine, and quarter-wave transforms. FFTPACK resides on Netlib (see Section 19.1.5 (page 659)) at http://www.netlib.org/ﬀtpack. The GNU Scientiﬁc Library for C/C++ provides a reimplementation of FFTPACK. See http://www.gnu.org/software/gsl/. Algorithm 545 [Fra79] of the Collected Algorithms of the ACM is a Fortran implementation of the fast Fourier transform optimizing virtual memory performance. See Section 19.1.5 (page 659) for further information. Notes: Bracewell [Bra99] and Brigham [Bri88] are excellent introductions to Fourier transforms and the FFT. See also the exposition in [PFTV07]. Credit for inventing the fast Fourier transform is usually given to Cooley and Tukey [CT65], but see [Bri88] for a complete history. A cache-oblivious algorithm for the fast Fourier transform is given in [FLPR99]. This paper ﬁrst introduced the notion of cache-oblivious algorithms. The FFTW is based on this algorithm. See [FJ05] for more on the design of the FFTW. An interesting divide-and-conquer algorithm for polynomial multiplication [KO63] does the job in O(n1.59 ) time and is discussed in [AHU74, Man89]. An FFT-based algorithm that multiplies two n-bit numbers in O(n lg n lg lg n) time is due to Schönhage and Strassen [SS71] and is presented in [AHU74]. It is an open question of whether complex variables are really fundamental to fast algorithms for convolution. Fortunately, fast convolution can be used as a black box in most applications. Many variants of string matching are based on fast convolution [Ind98]. In recent years, wavelets have been proposed to replace Fourier transforms in ﬁltering. See [Wal99] for an introduction to wavelets. Related Problems: Data compression (see page 637), high-precision arithmetic (see page 423). 433 14 Combinatorial Problems We now consider several algorithmic problems of a purely combinatorial nature. These include sorting and permutation generations, both of which were among the ﬁrst non-numerical problems arising on electronic computers. Sorting can be viewed as identifying or imposing a total order on the keys, while searching and selection involve identifying speciﬁc keys based on their position in this total order. The rest of this section deals with other combinatorial objects, such as permutations, partitions, subsets, calendars, and schedules. We are particularly interested in algorithms that rank and unrank combinatorial objects—i.e. , that map each distinct object to and from a unique integer. Rank and unrank operations make many other tasks simple, such as generating random objects (pick a random number and unrank) or listing all objects in order (iterate from 1 to n and unrank). We conclude with the problem of generating graphs. Graph algorithms are more fully presented in subsequent sections of the catalog. Books on general combinatorial algorithms, in this restricted sense, include: • Nijenhuis and Wilf [NW78] – This book specializes in algorithms for constructing basic combinatorial objects such as permutations, subsets, and partitions. Such algorithms are often very short but hard to locate and usually are surprisingly subtle. Fortran programs for all of the algorithms are provided, as well as a discussion of the theory behind each of them. See Section 19.1.10 for details. • Kreher and Stinson [KS99] – The most recent book on combinatorial generation algorithms, with additional particular focus on algebraic problems such as isomorphism and dealing with symmetry. S.S. Skiena, The Algorithm Design Manual, 2nd ed., DOI: 10.1007/978-1-84800-070-4 14, c Springer-Verlag London Limited 2008 14. COMBINATORIAL PROBLEMS • Knuth [Knu97a, Knu98] – The standard reference on searching and sorting, with signiﬁcant material on combinatorial objects such as permutations. New material on the generation of permutations [Knu05a], subsets and partitions [Knu05b], and trees [Knu06] has been released on “fasciles,”—short paperback chunks of what is slated to be the mythical Volume 4. • Stanton and White [SW86a] – An undergraduate combinatorics text with algorithms for generating permutations, subsets, and set partitions. It contains relevant programs in Pascal. • Pemmaraju and Skiena [PS03] – This description of Combinatorica, a library of over 400 Mathematica functions for generating combinatorial objects and graph theory (see Section 19.1.9), provides a distinctive view of how diﬀerent algorithms can ﬁt together. Its second author is considered qualiﬁed to write a manual on algorithm design. 435 436 14. COMBINATORIAL PROBLEMS INPUT 14.1 OUTPUT Sorting Input description: A set of n items. Problem description: Arrange the items in increasing (or decreasing) order. Discussion: Sorting is the most fundamental algorithmic problem in computer science. Learning the diﬀerent sorting algorithms is like learning scales for a musician. Sorting is the ﬁrst step in solving a host of other algorithm problems, as shown in Section 4.2 (page 107). Indeed, “when in doubt, sort” is one of the ﬁrst rules of algorithm design. Sorting also illustrates all the standard paradigms of algorithm design. The result is that most programmers are familiar with many diﬀerent sorting algorithms, which sows confusion as to which should be used for a given application. The following criteria can help you decide: • How many keys will you be sorting? – For small amounts of data (say n ≤ 100), it really doesn’t matter much which of the quadratic-time algorithms you use. Insertion sort is faster, simpler, and less likely to be buggy than bubblesort. Shellsort is closely related to, but much faster than, insertion sort, but it involves looking up the right insert sequences in Knuth [Knu98]. When you have more than 100 items to sort, it is important to use an O(n lg n)-time algorithm like heapsort, quicksort, or mergesort. There are various partisans who favor one of these algorithms over the others, but since it can be hard to tell which is fastest, it usually doesn’t matter. Once you get past (say) 5,000,000 items, it is important to start thinking about external-memory sorting algorithms that minimize disk access. Both types of algorithm are discussed below. 14.1 SORTING • Will there be duplicate keys in the data? – The sorted order is completely deﬁned if all items have distinct keys. However, when two items share the same key, something else must determine which one comes ﬁrst. In many applications it doesn’t matter, so any sorting algorithm suﬃces. Ties are often broken by sorting on a secondary key, like the ﬁrst name or initial when the family names collide. Occasionally, ties need to be broken by their initial position in the data set. Suppose the 5th and 27th items of the initial data set share the same key. This means the 5th item must appear before the 27th in the ﬁnal order. A stable sorting algorithm preserves the original ordering in case of ties. Most of the quadratic-time sorting algorithms are stable, while many of the O(n lg n) algorithms are not. If it is important that your sort be stable, it is probably better to explicitly use the initial position as a secondary key in your comparison function rather than trust the stability of your implementation. • What do you know about your data? – You can often exploit special knowledge about your data to get it sorted faster or more easily. Of course, general sorting is a fast O(n lg n) operation, so if the time spent sorting is really the bottleneck in your application, you are a fortunate person indeed. – Has the data already been partially sorted? If so, certain algorithms like insertion sort perform better than they otherwise would. – Do you know the distribution of the keys? If the keys are randomly or uniformly distributed, a bucket or distribution sort makes sense. Throw the keys into bins based on their ﬁrst letter, and recur until each bin is small enough to sort by brute force. This is very eﬃcient when the keys get evenly distributed into buckets. However, bucket sort would perform very badly sorting names on the membership roster of the “Smith Society.” – Are your keys very long or hard to compare? If your keys are long text strings, it might pay to use a relatively short preﬁx (say ten characters) of each key for an initial sort, and then resolve ties using the full key. This is particularly important in external sorting (see below), since you don’t want to waste your fast memory on the dead weight of irrelevant detail. Another idea might be to use radix sort. This always takes time linear in the number of characters in the ﬁle, instead of O(n lg n) times the cost of comparing two keys. – Is the range of possible keys very small? If you want to sort a subset of, say, n/2 distinct integers, each with a value from 1 to n, the fastest algorithm would be to initialize an n-element bit vector, turn on the bits corresponding to keys, then scan from left to right and report the positions with true bits. 437 438 14. COMBINATORIAL PROBLEMS • Do I have to worry about disk accesses? – In massive sorting problems, it may not be possible to keep all data in memory simultaneously. Such a problem is called external sorting, because one must use an external storage device. Traditionally, this meant tape drives, and Knuth [Knu98] describes a variety of intricate algorithms for eﬃciently merging data from diﬀerent tapes. Today, it usually means virtual memory and swapping. Any sorting algorithm will run using virtual memory, but most will spend all their time swapping. The simplest approach to external sorting loads the data into a B-tree (see Section 12.1 (page 367)) and then does an in-order traversal of the tree to read the keys oﬀ in sorted order. Real high-performance sorting algorithms are based on multiway-mergesort. Files containing portions of the data are sorted into runs using a fast internal sort, and then ﬁles with these sorted runs are merged in stages using 2- or k-way merging. Complicated merging patterns and buﬀer management based on the properties of the external storage device can be used to optimize performance. • How much time do you have to write and debug your routine? – If I had under an hour to deliver a working routine, I would probably just implement a simple selection sort. If I had an afternoon to build an eﬃcient sort routine, I would probably use heapsort, for it delivers reliable performance without tuning. The best general-purpose internal sorting algorithm is quicksort (see Section 4.2 (page 107)), although it requires tuning eﬀort to achieve maximum performance. Indeed, you are much better oﬀ using a library function instead of coding it yourself. A poorly written quicksort will likely run more slowly than a poorly written heapsort. If you are determined to implement your own quicksort, use the following heuristics, which make a big diﬀerence in practice: • Use randomization – By randomly permuting (see Section 14.4 (page 448)) the keys before sorting, you can eliminate the potential embarrassment of quadratic-time behavior on nearly-sorted data. • Median of three – For your pivot element, use the median of the ﬁrst, last, and middle elements of the array to increase the likelihood of partitioning the array into roughly equal pieces. Some experiments suggest using a larger sample on big subarrays and a smaller sample on small ones. • Leave small subarrays for insertion sort – Terminating the quicksort recursion and switching to insertion sort makes sense when the subarrays get small, say fewer than 20 elements. You should experiment to determine the best switchpoint for your implementation. • Do the smaller partition ﬁrst – Assuming that your compiler is smart enough to remove tail recursion, you can minimize run-time memory by processing 14.1 SORTING the smaller partition before the larger one. Since successive stored calls are at most half as large as the previous one, only O(lg n) stack space is needed. Before you get started, see Bentley’s article on building a faster quicksort [Ben92b]. Implementations: The best freely available sort program is presumably GNU sort, part of the GNU core utilities library. See http://www.gnu.org/software/ coreutils/. Be aware that there are also commercial vendors of high-performance external sorting programs. These include Cosort (www.cosort.com), Syncsort (www.syncsort.com) and Ordinal Technology (www.ordinal.com). Modern programming languages provide libraries oﬀering complete and eﬃcient container implementations. Thus, you should never need to implement your own sort routine. The C standard library contains qsort, a generic implementation of (presumably) quicksort. The C++ Standard Template Library (STL) provides both sort and stable sort methods. It is available with documentation at http://www.sgi.com/tech/stl/. See Josuttis [Jos99], Meyers [Mey01] and Musser [MDS01] for more detailed guides to using STL and the C++ standard library. Java Collections (JC), a small library of data structures, is included in the java.util package of Java standard edition (http://java.sun.com/javase/). In particular, SortedMap and SortedSet classes are provided. For a C++ implementation of an cache-oblivious algorithm (funnelsort), check out http://kristoﬀer.vinther.name/projects/funnelsort/. Benchmarks attest to its excellent performance. There are a variety of websites that provide applets/animations of all the basic sorting algorithms, including bubblesort, heapsort, mergesort, quicksort, radix sort, and shellsort. Many of these are quite interesting to watch. Indeed, sorting is the canonical problem for algorithm animation. Representative examples include Harrison (http://www.cs.ubc.ca/spider/harrison/Java/sorting-demo.html) and Bentley [Ben99] (http://www.cs.bell-labs.com/cm/cs/pearls/sortanim.html). Notes: Knuth [Knu98] is the best book that has been written on sorting and indeed is the best book that will ever be written on sorting. It is now over thirty years old, but remains fascinating reading. One area that has developed since Knuth is sorting under presortedness measures, surveyed in [ECW92]. A noteworthy reference on sorting is [GBY91], which includes pointers to algorithms for partially sorted data and includes implementations in C and Pascal for all of the fundamental algorithms. Expositions on the basic internal sorting algorithms appear in every algorithms text. Heapsort was ﬁrst invented by Williams [Wil64]. Quicksort was invented by Hoare [Hoa62], with careful analysis and implementation by Sedgewick [Sed78]. Von Neumann is credited with having produced the ﬁrst implementation of mergesort on the EDVAC in 1945. See Knuth for a full discussion of the history of sorting, dating back to the days of punch-card tabulating machines. The primary competitive forum for high-performance sorting is an annual competition initiated by the late Jim Gray. See http://research.microsoft.com/barc/SortBenchmark/ for current and previous results, which are are either inspiring or depressing depending 439 440 14. COMBINATORIAL PROBLEMS upon how you look at it. The magnitude of progress is inspiring (the million-record instances of the original benchmarks are now too small to bother with) but it is depressing (to me) the extent that systems/memory management issues thoroughly trump the combinatorial/algorithmic aspects of sorting. Modern attempts to engineer high-performance sort programs include work on both cache-conscious [LL99] and cache-oblivious [BFV07] sorting. Sorting has a well-known Ω(n lg n) lower bound under the algebraic decision tree model [BO83]. Determining the exact number of comparisons required for sorting n elements, for small values of n, has generated considerable study. See [Aig88, Raw92] for expositions and Peczarski [Pec04, Pec07] for the latest results. This lower-bound does not hold under diﬀerent models of computation. Fredman and √ Willard [FW93] present an O(n lg n) algorithm for sorting under a model of computation that permits arithmetic operations on keys. See Andersson [And05] for a survey of algorithms for fast sorting on such nonstandard models of computation. Related Problems: Dictionaries (see page 367), searching (see page 441), topological sorting (see page 481). 14.2 SEARCHING H H D L B A F C E D J G I N K M L B O A F C G? INPUT 14.2 441 E J G I N K G OUTPUT Searching Input description: A set of n keys S, and a query key q. Problem description: Where is q in S? Discussion: “Searching” is a word that means diﬀerent things to diﬀerent people. Searching for the global maximum or minimum of a function is the problem of unconstrained optimization and is discussed in Section 13.5 (page 407). Chessplaying programs select the best move to make via an exhaustive search of possible moves using a variation of backtracking (see Section 7.1 (page 231)). Here we consider the task of searching for a key in a list, array, or tree. Dictionary data structures maintain eﬃcient access to sets of keys under insertion and deletion and are discussed in Section 12.1 (page 367). Typical dictionaries include binary trees and hash tables. We treat searching as a problem distinct from dictionaries because simpler and more eﬃcient solutions emerge when our primary interest is static searching without insertion/deletion. These little data structures can yield large performance improvements when properly employed in an innermost loop. Also, ideas such as binary search and self-organization apply to other problems and well justify our attention. Our two basic approaches are sequential search and binary search. Both are simple, yet have interesting and subtle variations. In sequential search, we start from the front of our list/array of keys and compare each successive item against the key until we ﬁnd a match or reach the end. In binary search, we start with a sorted array of keys. To search for key q, we compare q to the middle key Sn/2 . If q is before Sn/2 , it must reside in the top half of our set; if not, it must reside in the M O 442 14. COMBINATORIAL PROBLEMS bottom half of our set. By repeating this process on the correct half, we ﬁnd the key using lg n comparisons. This is a big win over the n/2 comparisons we expect with sequential search. See Section 4.9 (page 132) for more on binary search. A sequential search is the simplest algorithm, and likely to be fastest on up to about 20 elements. Beyond (say) 100 elements, binary search will clearly be more eﬃcient than sequential search, easily justifying the cost of sorting if there will be multiple queries. Other issues come into play, however, in identifying the proper variant of the algorithm: • How much time can you spend programming? – A binary search is a notoriously tricky algorithm to program correctly. It took seventeen years after its invention until the ﬁrst correct version of a binary search was published! Don’t be afraid to start from one of the implementations described below. Test it completely by writing a driver that searches for every key in the set S as well as between the keys. • Are certain items accessed more often than other ones? – Certain English words (such as “the”) are much more likely to occur than others (such as “defenestrate”). We can reduce the number of comparisons in a sequential search by putting the most popular words on the top of the list and the least popular ones at the bottom. Nonuniform access is usually the rule, not the exception. Many real-world distributions are governed by power laws. A classic example is word use in English, which is fairly accurately modeled by Zipf ’s law. Under Zipf ’s law, the ith most frequently accessed key is selected with probability (i − 1)/i times the probability of the (i − 1)st most popular key, for all 1 ≤ i ≤ n. Knowledge of access frequencies is easy to exploit with sequential search. But the issue is more complicated with binary trees. We want popular keys close to the root (so we hit them quickly) but not at the expense of losing balance and degenerating into sequential search. The answer is to employ a dynamic programming algorithm to ﬁnd the optimal binary search tree. The key observation is that each possible root node i partitions the space of keys into those to the left of i and those to the right; each of which should be represented by an optimal binary search tree on a smaller subrange of keys. The root of the optimal tree is selected to minimize the expected search costs of the resulting partition. • Might access frequencies change over time? – Preordering a list or tree to exploit a skewed access pattern requires knowing the access pattern in advance. For many applications, it can be diﬃcult to obtain such information. Better are self-organizing lists, where the order of the keys changes in response to the queries. The best self-organizing scheme is move-to-front; that is, we move the most recently searched-for key from its current position to the front of the list. Popular keys keep getting boosted to the front, while unsearched-for 14.2 SEARCHING keys drift towards the back of the list. There is no need to keep track of the frequency of access; just move the keys on demand. Self-organizing lists also exploit locality of reference, since accesses to a given key are likely to occur in clusters. A hot key will be maintained near the top of the list during a cluster of accesses, even if other keys have proven more popular in the past. Self-organization can extend the useful size range of sequential search. However, you should switch to binary search beyond 100 elements. But consider using splay trees. which are self-organizing binary search trees that rotate each searched-for node to the root. They oﬀer excellent amortized performance guarantees. • Is the key close by? – Suppose we know that the target key is to the right of position p, and we think it is nearby. A sequential search is fast if we are correct, but we will be punished severely when we guess wrong. A better idea is to test repeatedly at larger intervals (p + 1, p + 2, p + 4, p + 8, p + 16, . . .) to the right until we ﬁnd a key to the right of our target. Now we have a window containing the target and we can proceed with binary search. Such a one-sided binary search ﬁnds the target at position p + l using at most 2lg l comparisons, so it is faster than binary search when l << n, yet it can never be much worse. One-sided binary search is particularly useful in unbounded search problems, such as in numerical root ﬁnding. • Is my data structure sitting on external memory? – Once the number of keys grows too large, a binary search loses its status as the best search technique. A binary search jumps wildly around the set of keys looking for midpoints to compare, and so each comparison requires reading a new page in from external memory. Much better are data structures such as B-trees (see Section 12.1 (page 367)) or Emde Boas trees (see notes below), which cluster the keys into pages to minimize the number of disk accesses per search. • Can I guess where the key should be? – In interpolation search, we exploit our understanding of the distribution of keys to guess where to look next. An interpolation search is probably a more accurate description of how we use a telephone book than binary search. Suppose we are searching for Washington, George in a sorted telephone book. We would be safe making our ﬁrst comparison three-fourths of the way down the list, essentially doing two comparisons for the price of one. Although an interpolation search is an appealing idea, we caution against it for three reasons: First, you have to work very hard to optimize your search algorithm before you can hope for a speedup over binary search. Second, even if you do beat a binary search, it is unlikely to be by enough to have justiﬁed the exercise. Finally, your program will be much less robust and eﬃcient when the distribution changes, such as when your application gets ported to work on French words instead of English. 443 444 14. COMBINATORIAL PROBLEMS Implementations: The basic sequential and binary search algorithms are simple enough that you may consider implementing them yourself. That said, the C standard library contains bsearch, a generic implementation of (presumably) a binary search. The C++ Standard Template Library (STL) provides find (sequential search) and binary search iterators. Java Collections (JC), provides binarySearch in the java.util package of Java standard edition (http://java.sun.com/javase/). Many data structure textbooks provide extensive and illustrative implementations. Sedgewick (http://www.cs.princeton.edu/∼rs/) [Sed98] and Weiss (http://www.cs.ﬁu.edu/∼weiss/) [Wei06] provide implementation of splay trees and other search structures in both C++ and Java. Notes: The Handbook of Data Structures and Applications [MS05] provides up-to-date surveys on all aspects of dictionary data structures. Other surveys include Mehlhorn and Tsakalidis [MT90b] and Gonnet and Baeza-Yates [GBY91]. Knuth [Knu97a] provides a detailed analysis and exposition on all fundamental search algorithms and dictionary data structures, but omits such modern data structures as red-black and splay trees. The next position probed in linear interpolation search on an array of sorted numbers is given by next = (low − 1) + q − S[low − 1] × (high − low + 1) S[high + 1] − S[low − 1] where q is the query numerical key and S the sorted numerical array. If the keys are drawn independently from a uniform distribution, the expected search time is O(lg lg n) [DJP04, PIA78]. Nonuniform access patterns can be exploited in binary search trees by structuring them so that popular keys are located near the root, thus minimizing search time. Dynamic programming can be used to construct such optimal search trees in O(n lg n) time [Knu98]. Stout and Warren [SW86b] provide a slick algorithm to eﬃciently transform a binary tree to a minimum height (optimally balanced) tree using rotations. The Van Emde Boas layout of a binary tree (or sorted array) oﬀers better external memory performance than conventional binary search, at a cost of greater implementation complexity. See the survey of Arge, et al. [ABF05] for more on this and other cacheoblivious data structures. Related Problems: Dictionaries (see page 367), sorting (see page 436). 14.3 MEDIAN AND SELECTION INPUT 14.3 OUTPUT Median and Selection Input description: A set of n numbers or keys, and an integer k. Problem description: Find the key smaller than exactly k of the n keys. Discussion: Median ﬁnding is an essential problem in statistics, where it provides a more robust notion of average than the mean. The mean wealth of people who have published research papers on sorting is signiﬁcantly aﬀected by the presence of one William Gates [GP79], although his eﬀect on the median wealth is merely to cancel out one starving graduate student. Median ﬁnding is a special case of the more general selection problem, which asks for the kth element in sorted order. Selection arises in several applications: • Filtering outlying elements – In dealing with noisy data, it is usually a good idea to throw out (say) the 10% largest and smallest values. Selection can be used to identify the items deﬁning the 10th and 90th percentiles, and the outliers then ﬁltered out by comparing each item to the two selected bounds. • Identifying the most promising candidates – In a computer chess program, we might quickly evaluate all possible next moves, and then decide to study the top 25% more carefully. Selection followed by ﬁltering is the way to go. • Deciles and related divisions – A useful way to present income distribution in a population is to chart the salary of the people ranked at regular intervals, say exactly at the 10th percentile, 20th percentile, etc. Computing these values is simply selection on the appropriate position ranks. • Order statistics – Particularly interesting special cases of selection include ﬁnding the smallest element (k = 1), the largest element (k = n), and the median element (k = n/2). 445 446 14. COMBINATORIAL PROBLEMS The mean of n numbers can be computed in linear time by summing the elements and dividing by n. However, ﬁnding the median is a more diﬃcult problem. Algorithms that compute the median can readily be generalized to arbitrary selection. Issues in median ﬁnding and selection include: • How fast does it have to be? – The most elementary median-ﬁnding algorithm sorts the items in O(n lg n) time and then returns the item occupying the (n/2)nd position. The good thing is that this gives much more information than just the median, enabling you to select the kth element (for any 1 ≤ k ≤ n) in constant time after the sort. However, there are faster algorithms if all you want is the median. In particular, there is an O(n) expected-time algorithm based on quicksort. Select a random element in the data set as a pivot, and use it to partition the data into sets of elements less than and greater than the pivot. From the sizes of these sets, we know the position of the pivot in the total order, and hence whether the median lies to the left or right of this point. Now we recur on the appropriate subset until it converges on the median. This takes (on average) O(lg n) iterations, with the cost of each iteration being roughly half that of the previous one. This deﬁnes a geometric series that converges to a linear-time algorithm, although if you are very unlucky it takes the same time as quicksort, O(n2 ). More complicated algorithms can ﬁnd the median in worst-case linear time. However, the expected-time algorithm will likely win in practice. Just make sure to select random pivots to avoid the worst case. • What if you only get to see each element once? – Selection and median ﬁnding become expensive on large datasets because they typically require several passes through external memory. In data-streaming applications, the volume of data is often too large to store, making repeated consideration (and thus exact median ﬁnding) impossible. Much better is computing a small summary of the data for future analysis, say approximate deciles orfrequency moments (where the kth moment of stream x is deﬁned as Fk = i xki ). One solution to such a problem is random sampling. Flip a coin for each value to decide whether to store it, with the probability of heads set low enough that you won’t overﬂow your buﬀer. Likely the median of your samples will be close to that of the underlying data set. Alternately, you can devote some fraction of memory to retaining (say) decile values of large blocks, and then combine these decile distributions to yield more reﬁned decile bounds. • How fast can you ﬁnd the mode? – Beyond mean and median lies a third notion of average. The mode is deﬁned to be the element that occurs the greatest number of times in the data set. The best way to compute the mode sorts the set in O(n lg n) time, which places all identical elements next to each other. By doing a linear sweep from left to right on this sorted set, we can 14.3 MEDIAN AND SELECTION count the length of the longest run of identical elements and hence compute the mode in a total of O(n lg n) time. In fact, there is no faster worst-case algorithm possible to compute the mode, since the problem of testing whether there exist two identical elements in a set (called element uniqueness) can be shown to have an Ω(n log n) lower bound. Element uniqueness is equivalent to asking whether the mode occurs more than once. Possibilities exist, at least theoretically, for improvements when the mode is large by using fast median computations. Implementations: The C++ Standard Template Library (STL) provides a general selection method (nth element) implemented using the linear expected-time algorithm. It is available with documentation at http://www.sgi.com/tech/stl/. See Josuttis [Jos99], Meyers [Mey01] and Musser [MDS01] for more detailed guides to using STL and the C++ standard library. Notes: The linear expected-time algorithm for median and selection is due to Hoare [Hoa61]. Floyd and Rivest [FR75] provide an algorithm that uses fewer comparisons on average. Good expositions on linear-time selection include [AHU74, BvG99, CLRS01, Raw92], with [Raw92] being particularly enlightening. Streaming algorithms have extensive applications to large data sets, and are well surveyed by Muthukrishnan [Mut05]. A sport of considerable theoretical interest is determining exactly how many comparisons are suﬃcient to ﬁnd the median of n items. The linear-time algorithm of Blum et al. [BFP+ 72] proves that c · n suﬃce, but we want to know what c is. Dor and Zwick [DZ99] proved that 2.95n comparisons suﬃce to ﬁnd the median. These algorithms attempt to minimize the number of element comparisons but not the total number of operations, and hence do not lead to faster algorithms in practice. They also hold the current best lower bower bound of (2 + ) comparisons for median ﬁnding [DZ01]. Tight combinatorial bounds for selection problems are presented in [Aig88]. An optimal algorithm for computing the mode is given by [DM80]. Related Problems: Priority queues (see page 373), sorting (see page 436). 447 448 14. COMBINATORIAL PROBLEMS { 1, 2, 3 } { 2, 1, 3 } { 3, 2, 1 } { 3, 1, 2 } { 2, 3, 1 } { 1, 2, 3 } { 1, 3, 2 } INPUT 14.4 OUTPUT Generating Permutations Input description: An integer n. Problem description: Generate (1) all, or (2) a random, or (3) the next permutation of length n. Discussion: A permutation describes an arrangement or ordering of items. Many algorithmic problems in this catalog seek the best way to order a set of objects, including traveling salesman (the least-cost order to visit n cities), bandwidth (order the vertices of a graph on a line so as to minimize the length of the longest edge), and graph isomorphism (order the vertices of one graph so that it is identical to another). Any algorithm for solving such problems exactly must construct a series of permutations along the way. There are n! permutations of n items. This grows so quickly that you can’t really expect to generate all permutations for n > 12, since 12! = 479,001,600. Numbers like these should cool the ardor of anyone interested in exhaustive search and help explain the importance of generating random permutations. Fundamental to any permutation-generation algorithm is a notion of order, the sequence in which the permutations are constructed from ﬁrst to last. The most natural generation order is lexicographic, the sequence they would appear if they were sorted numerically. Lexicographic order for n = 3 is {1, 2, 3}, {1, 3, 2}, {2, 1, 3}, {2, 3, 1}, {3, 1, 2}, and ﬁnally {3, 2, 1}. Although lexicographic order is aesthetically pleasing, there is often no particular advantage to using it. For example, if you are searching through a collection of ﬁles, it does not matter whether the ﬁlenames are encountered in sorted order, so long as you eventually search through all of them. Indeed, nonlexicographic orders lead to faster and simpler permutation generation algorithms. 14.4 GENERATING PERMUTATIONS There are two diﬀerent paradigms for constructing permutations: ranking/unranking and incremental change methods. The latter are more eﬃcient, but ranking and unranking can be applied to solve a much wider class of problems. The key is to deﬁne the functions rank and unrank on all permutations p and integers n, m, where |p| = n and 0 ≤ m ≤ n!. • Rank(p) – What is the position of p in the given generation order? A typical ranking function is recursive, such as basis case Rank({1}) = 0 with Rank(p) = (p1 − 1) · (|p| − 1)! + Rank(p2 , . . . , p|p| ) Getting this right means relabeling the elements of the smaller permutation to reﬂect the deleted ﬁrst element. Thus Rank({2, 1, 3}) = 1 · 2! + Rank({1, 2}) = 2 + 0 · 1! + Rank({1}) = 2 • Unrank(m,n) – Which permutation is in position m of the n! permutations of n items? A typical unranking function ﬁnds the number of times (n − 1)! goes into m and proceeds recursively. U nrank(2, 3) tells us that the ﬁrst element of the permutation must be ‘2’, since (2 − 1) · (3 − 1)! ≤ 2 but (3 − 1) · (3 − 1)! > 2. Deleting (2 − 1) · (3 − 1)! from m leaves the smaller problem U nrank(0, 2). The ranking of 0 corresponds to the total order. The total order on the two remaining elements (since 2 has been used) is {1, 3}, so U nrank(2, 3) = {2, 1, 3}. What the actual rank and unrank functions are does not matter as much as the fact that they must be inverses of each other. In other words, p = U nrank(Rank(p), n) for all permutations p. Once you deﬁne ranking and unranking functions for permutations, you can solve a host of related problems: • Sequencing permutations – To determine the next permutation that occurs in order after p, we can Rank(p), add 1, and then U nrank(p). Similarly, the permutation right before p in order is Unrank(Rank(p) − 1, |p|). Counting through the integers from 0 to n! − 1 and unranking them is equivalent to generating all permutations. • Generating random permutations – Selecting a random integer from 0 to n!−1 and then unranking it yields a truly random permutation. • Keep track of a set of permutations – Suppose we want to construct random permutations and act only when we encounter one we have not seen before. We can set up a bit vector (see Section 12.5 (page 385)) with n! bits, and set bit i to 1 if permutation Unrank(i,n) has been seen. A similar technique was employed with k-subsets in the Lotto application of Section 1.6 (page 23). 449 450 14. COMBINATORIAL PROBLEMS This rank/unrank method is best suited for small values of n, since n! quickly exceeds the capacity of machine integers unless arbitrary-precision arithmetic is available (see Section 13.9 (page 423)). The incremental change methods work by deﬁning the next and previous operations to transform one permutation into another, typically by swapping two elements. The tricky part is to schedule the swaps so that permutations do not repeat until all of them have been generated. The output picture above gives an ordering of the six permutations of {1, 2, 3} using a single swap between successive permutations. Incremental change algorithms for sequencing permutations are tricky, but they are so concise that they can be expressed in a dozen-line program. See the implementation section for pointers to code. Because the incremental change is only a single swap, these algorithms can be extremely fast—on average, constant time— which is independent of the size of the permutation! The secret is to represent the permutation using an n-element array to facilitate the swap. In certain applications, only the change between permutations is important. For example, in a brute-force program to search for the optimal TSP tour, the cost of the tour associated with the new permutation will be that of the previous permutation, with the addition and deletion of four edges. Throughout this discussion, we have assumed that the items we are permuting are all distinguishable. However, if there are duplicates (meaning our set is a multiset), you can save considerable time and eﬀort by avoiding identical permutations. For example, there are only ten distinct permutations of {1, 1, 2, 2, 2}, instead of 120. To avoid duplicates, use backtracking and generate the permutations in lexicographic order. Generating random permutations is an important little problem that people often stumble across, and often botch up. The right way is to use the following two-line, linear-time algorithm. We assume that Random[i,n] generates a random integer between i and n, inclusive. for i = 1 to n do a[i] = i; for i = 1 to n − 1 do swap[a[i], a[Random[i, n]]; That this algorithm generates all permutations uniformly at random is not obvious. If you think so, convincingly explain why the following algorithm does not generate permutations uniformly: for i = 1 to n do a[i] = i; for i = 1 to n − 1 do swap[a[i], a[Random[1, n]]; Such subtleties demonstrate why you must be very careful with random generation algorithms. Indeed, we recommend that you try some reasonably extensive experiments with any random generator before really believing it. For example, generate 10,000 random permutations of length 4 and see whether all 24 distinct permutations occur approximately the same number of times. If you know how to measure statistical signiﬁcance, you are in even better shape. 14.4 GENERATING PERMUTATIONS Implementations: The C++ Standard Template Library (STL) provides two functions (next permutation and prev permutation) for sequencing permutations in lexicographic order. Kreher and Stinson [KS99] provide implementations of minimum change and lexicographic permutation generation in C at http://www.math.mtu.edu/∼kreher/cages/Src.html. The Combinatorial Object Server (http://theory.cs.uvic.ca/) developed by Frank Ruskey of the University of Victoria is a unique resource for generating permutations, subsets, partitions, graphs, and other objects. An interactive interface enables you to specify which objects you would like returned to you. Implementations in C, Pascal, and Java are available for certain types of objects. C++ routines for generating an astonishing variety of combinatorial objects, including permutations and cyclic permutations, are available at http://www.jjj.de/ fxt/. Nijenhuis and Wilf [NW78] is a venerable but still excellent source on generating combinatorial objects. They provide eﬃcient Fortran implementations of algorithms to construct random permutations and to sequence permutations in minimum-change order. Also included are routines to extract the cycle structure of a permutation. See Section 19.1.10 (page 661) for details. Combinatorica [PS03] provides Mathematica implementations of algorithms that construct random permutations and sequence permutations in minimum change and lexicographic orders. It also provides a backtracking routine to construct all distinct permutations of a multiset, and it supports various permutation group operations. See Section 19.1.9 (page 661). Notes: The best recent reference on permutation generation is Knuth [Knu05a]. Sedgewick’s excellent survey on the topic is older [Sed77], but this is not a fast moving area. Good expositions include [KS99, NW78, Rus03]. Fast permutation generation methods make only a single swap between successive permutations. The Johnson-Trotter algorithm [Joh63, Tro62] satisﬁes an even stronger condition, namely that the two elements being swapped are always adjacent. Simple lineartime ranking and unranking functions for permutations are given by Myrvold and Ruskey [MR01]. In the days before ready access to computers, books with tables of random permutations [MO63] were used instead of algorithms. The swap-based random permutation algorithm presented above was ﬁrst described in [MO63]. Related Problems: Random-number generation (see page 415), generating subsets (see page 452), generating partitions (see page 456). 451 452 14. COMBINATORIAL PROBLEMS {} {3} {1} {1, 3} {1, 2} { 1, 2, 3 } {1, 2, 3} {2} {2, 3} INPUT 14.5 OUTPUT Generating Subsets Input description: An integer n. Problem description: Generate (1) all, or (2) a random, or (3) the next subset of the integers {1, . . . , n}. Discussion: A subset describes a selection of objects, where the order among them does not matter. Many important algorithmic problems seek the best subset of a group of things: vertex cover seeks the smallest subset of vertices to touch each edge in a graph; knapsack seeks the most proﬁtable subset of items of bounded total size; while set packing seeks the smallest subset of subsets that together cover each item exactly once. There are 2n distinct subsets of an n-element set, including the empty set as well as the set itself. This grows exponentially, but at a considerably slower rate than the n! permutations of n items. Indeed, since 220 = 1,048,576, a brute-force search through all subsets of 20 elements is easily manageable. Since 230 = 1,073,741,824, you will certainly hit limits for slightly larger values of n. By deﬁnition, the relative order among the elements does not distinguish diﬀerent subsets. Thus, {1, 2, 5} is the same as {2, 1, 5}. However, it is a very good idea to maintain your subsets in a sorted or canonical order to speed up such operations as testing whether or not two subsets are identical. As with permutations (see Section 14.4 (page 448)), the key to subset generation problems is establishing a numerical sequence among all 2n subsets. There are three primary alternatives: 14.5 GENERATING SUBSETS • Lexicographic order – Lexicographic order means sorted order, and is often the most natural way to generate combinatorial objects. The eight subsets of {1, 2, 3} in lexicographic order are {}, {1}, {1, 2}, {1, 2, 3}, {1, 3}, {2}, {2, 3}, and {3}. But it is surprisingly diﬃcult to generate subsets in lexicographic order. Unless you have a compelling reason to do so, don’t bother. • Gray Code – A particularly interesting and useful subset sequence is the minimum change order, wherein adjacent subsets diﬀer by the insertion or deletion of exactly one element. Such an ordering, called a Gray code, appears in the output picture above. Generating subsets in Gray code order can be very fast, because there is a nice recursive construction. Construct a Gray code of n − 1 elements Gn−1 . Reverse a second copy of Gn−1 and add n to each subset in this copy. Then concatenate them together to create Gn . Study the output example for clariﬁcation. Further, since only one element changes between subsets, exhaustive search algorithms built on Gray codes can be quite eﬃcient. A set cover program would only have to update the change in coverage by the addition or deletion of one subset. See the implementation section below for Gray code subsetgeneration programs. • Binary counting – The simplest approach to subset-generation problems is based on the observation that any subset S is deﬁned by the items of that S are in S . We can represent S by a binary string of n bits, where bit i is 1 iﬀ the ith element of S is in S . This deﬁnes a bijection between the 2n binary strings of length n, and the 2n subsets of n items. For n = 3, binary counting generates subsets in the following order: {}, {3}, {2}, {2,3}, {1}, {1,3}, {1,2}, {1,2,3}. This binary representation is the key to solving all subset generation problems. To generate all subsets in order, simply count from 0 to 2n − 1. For each integer, successively mask oﬀ each of the bits and compose a subset of exactly the items corresponding to 1 bits. To generate the next or previous subset, increment or decrement the integer by one. Unranking a subset is exactly the masking procedure, while ranking constructs a binary number with 1’s corresponding to items in S and then converts this binary number to an integer. To generate a random subset, you might generate a random integer from 0 to 2n − 1 and unrank, although how your random number generator rounds things oﬀ might mean that certain subsets can never occur. Much better is to ﬂip a coin n times, with the ith ﬂip deciding whether to include element i in the subset. A coin ﬂip can be robustly simulated by generating a random real or large integer and testing whether it is bigger or smaller than half its range. A Boolean array of n items can thus be used to represent subsets as a sort of premasked integer. 453 454 14. COMBINATORIAL PROBLEMS Generation problems for two closely related problems arise often in practice: • K-subsets – Instead of constructing all subsets, we may only be interested in the subsets containing exactly k elements. There are nk such subsets, which is substantially less than 2n , particularly for small values of k. The best way to construct all k-subsets is in lexicographic order. The ranking n−f function is based on the observation that there are k − 1 k-subsets whose smallest element is f . Using this, it is possible to determine the smallest element in the mth k-subset of n items. We then proceed recursively for subsequent elements of the subset. See the implementations below for details. • Strings – Generating all subsets is equivalent to generating all 2n strings of true and false. The same basic techniques apply to generate all or random strings on alphabets of size α, except there will be αn strings in total. Implementations: Kreher and Stinson [KS99] provide generators for both subsets and k-subsets, including lexicographic and Gray code orders. These implementations in C are available at http://www.math.mtu.edu/∼kreher/cages/Src.html. The Combinatorial Object Server (http://theory.cs.uvic.ca/), developed by Frank Ruskey of the University of Victoria, is a unique resource for generating permutations, subsets, partitions, graphs, and other objects. An interactive interface enables you to specify which objects you would like returned to you. Implementations in C, Pascal, and Java are available for certain types of objects. C++ routines for generating an astonishing variety of combinatorial objects, including subsets and k-subsets (combinations), are available in the combinatoric