Handbook of Approximation Algorithms

and Metaheuristics, Second Edition:

Contemporary and Emerging
Applications, Volume 2 Teofilo F
Gonzalez(Editor)
Visit to download the full and correct content document:
https://textbookfull.com/product/handbook-of-approximation-algorithms-and-metaheur
istics-second-edition-contemporary-and-emerging-applications-volume-2-teofilo-f-gon
zalezeditor/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

Handbook of Approximation Algorithms and Metaheuristics

Second Edition Methologies and Traditional Applications
Volume 1 Teofilo F Gonzalez(Editor)

https://textbookfull.com/product/handbook-of-approximation-
algorithms-and-metaheuristics-second-edition-methologies-and-
traditional-applications-volume-1-teofilo-f-gonzalezeditor/

Biota Grow 2C gather 2C cook Loucas

https://textbookfull.com/product/biota-grow-2c-gather-2c-cook-
loucas/

Handbook of natural fibres Volume 2 Processing and

applications Second Edition Koz■owski

https://textbookfull.com/product/handbook-of-natural-fibres-
volume-2-processing-and-applications-second-edition-kozlowski/

Advances in Metaheuristics Algorithms Methods and

Applications Erik Cuevas

https://textbookfull.com/product/advances-in-metaheuristics-
algorithms-methods-and-applications-erik-cuevas/
Handbook of Radioactivity Analysis: Volume 2:
Radioanalytical Applications 4th Edition Michael F.
L'Annunziata (Editor)

https://textbookfull.com/product/handbook-of-radioactivity-
analysis-volume-2-radioanalytical-applications-4th-edition-
michael-f-lannunziata-editor/

Optimization Using Evolutionary Algorithms and

Metaheuristics: Applications in Engineering 1st Edition
Kaushik Kumar

https://textbookfull.com/product/optimization-using-evolutionary-
algorithms-and-metaheuristics-applications-in-engineering-1st-
edition-kaushik-kumar/

Spanish Society 1348 1700 2nd Edition Teofilo F. Ruiz

https://textbookfull.com/product/spanish-society-1348-1700-2nd-
edition-teofilo-f-ruiz/

Handbook of Optoelectronics, Second Edition Volume 2:

Enabling Technologies John P. Dakin

https://textbookfull.com/product/handbook-of-optoelectronics-
second-edition-volume-2-enabling-technologies-john-p-dakin/

Psychostimulants (Volume 2) (Neurobiology of Addiction

Series (Volume 2)) 1st Edition George F. Koob Phd

https://textbookfull.com/product/psychostimulants-
volume-2-neurobiology-of-addiction-series-volume-2-1st-edition-
george-f-koob-phd/
Contemporary and Emerging
Applications
Volume II
Chapman & Hall/CRC
Computer and Information Science Series
Series Editor: Sartaj Sahni

Computer-Aided Graphing and Simulation Tools for AutoCAD Users

P. A. Simionescu
Integration of Services into Workflow Applications
Paweł Czarnul
Handbook of Graph Theory, Combinatorial Optimization, and Algorithms
Krishnaiyan “KT” Thulasiraman, Subramanian Arumugam, Andreas Brandstädt, and Takao Nishizeki
From Action Systems to Distributed Systems
The Refinement Approach
Luigia Petre and Emil Sekerinski
Trustworthy Cyber-Physical Systems Engineering
Alexander Romanovsky and Fuyuki Ishikawa
X-Machines for Agent-Based Modeling
FLAME Perspectives
Mariam Kiran
From Internet of Things to Smart Cities
Enabling Technologies
Hongjian Sun, Chao Wang, and Bashar I. Ahmad
Evolutionary Multi-Objective System Design
Theory and Applications
Nadia Nedjah, Luiza De Macedo Mourelle, and Heitor Silverio Lopes
Networks of the Future
Architectures, Technologies, and Implementations
Mahmoud Elkhodr, Qusay F. Hassan, and Seyed Shahrestani
Computer Simulation
A Foundational Approach Using Python
Yahya E. Osais
Internet of Things
Challenges, Advances, and Applications
Qusay F. Hassan, Atta ur Rehman Khan, and Sajjad A. Madani
Handbook of Data Structures and Applications, Second Edition
Dinesh P. Mehta and Sartaj Sahni
Handbook of Approximation Algorithms and Metaheuristics, Second Edition
Methodologies and Traditional Applications, Volume 1
Teofilo F. Gonzalez
Handbook of Approximation Algorithms and Metaheuristics, Second Edition
Contemporary and Emerging Applications, Volume 2
Teofilo F. Gonzalez

For more information about this series please visit:

https://www.crcpress.com/Chapman--HallCRC-Computer-and-Information-Science-Series/book-series/CHCOMINFSCI
 Handbook of Approximation
Algorithms and Metaheuristics,
Second Edition
Contemporary and Emerging Applications
Volume II

Edited by
Teofilo F. Gonzalez
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742

© 2018 by Taylor & Francis Group, LLC

CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Printed on acid-free paper

International Standard Book Number-13: 978-1-4987-6999-0 (Hardback)

This book contains information obtained from authentic and highly regarded sources. Reasonable efforts have been made to
publish reliable data and information, but the author and publisher cannot assume responsibility for the validity of all materials
or the consequences of their use. The authors and publishers have attempted to trace the copyright holders of all material repro-
duced in this publication and apologize to copyright holders if permission to publish in this form has not been obtained. If any
copyright material has not been acknowledged please write and let us know so we may rectify in any future reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or utilized in any
form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying, microfilming,
and recording, or in any information storage or retrieval system, without written permission from the publishers.

For permission to photocopy or use material electronically from this work, please access www.copyright.com (http://www.copyright.
com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a
not-for-profit organization that provides licenses and registration for a variety of users. For organizations that have been granted a
photocopy license by the CCC, a separate system of payment has been arranged.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identifica-
tion and explanation without intent to infringe.
Visit the Taylor & Francis Web site at
http://www.taylorandfrancis.com

and the CRC Press Web site at

http://www.crcpress.com
 To my wife Dorothy, and our children:
Jeanmarie, Alexis, Julia, Teofilo, and Paolo.
 Contents

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
Contributors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii
1 Introduction, Overview, and Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Teofilo F. Gonzalez

SECTION I Computational Geometry and Graph Applications

2 Approximation Schemes for Minimum-Cost k-Connectivity Problems
in Geometric Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Artur Czumaj and Andrzej Lingas
3 Dilation and Detours in Geometric Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
Joachim Gudmundsson and Christian Knauer
4 The Well-Separated Pair Decomposition and Its Applications . . . . . . . . . . . . . . . . . . 71
Michiel Smid
5 Covering with Unit Balls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
Hossein Ghasemalizadeh and Mohammadreza Razzazi
6 Minimum Edge-Length Rectangular Partitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Teofilo F. Gonzalez and Si Qing Zheng
7 Automatic Placement of Labels in Maps and Drawings . . . . . . . . . . . . . . . . . . . . . . . 111
Konstantinos G. Kakoulis and Ioannis G. Tollis
8 Complexity, Approximation Algorithms, and Heuristics for
the Corridor Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Teofilo F. Gonzalez and Arturo Gonzalez-Gutierrez
9 Approximate Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
Ragesh Jaiswal and Sandeep Sen
10 Maximum Planar Subgraph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
Gruia Călinescu and Cristina G. Fernandes
11 Disjoint Paths and Unsplittable Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
Stavros G. Kolliopoulos

vii
viii Contents

12 The k-Connected Subgraph Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213

Zeev Nutov
13 Node-Connectivity Survivable Network Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Zeev Nutov
14 Optimum Communication Spanning Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
Bang Ye Wu, Chuan Yi Tang, and Kun-Mao Chao
15 Activation Network Design Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
Zeev Nutov
16 Stochastic Local Search Algorithms for the Graph Coloring Problem . . . . . . . . . . 299
Marco Chiarandini, Irina Dumitrescu, and Thomas Stützle
17 On Solving the Maximum Disjoint Paths Problem with
Ant Colony Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
Maria J. Blesa and Christian Blum
18 Efficient Approximation Algorithms in Random Intersection Graphs . . . . . . . . . . 333
Sotiris E. Nikoletseas, Christoforos L. Raptopoulos, and Paul G. Spirakis
19 Approximation Algorithms for Facility Dispersion . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
S. S. Ravi, Daniel J. Rosenkrantz, and Giri K. Tayi

SECTION II Large-Scale and Emerging Applications

20 Cost-Efficient Multicast Routing in Ad Hoc and Sensor Networks . . . . . . . . . . . . 369
Pedro M. Ruiz and Ivan Stojmenovic
21 Approximation Algorithm for Clustering in Mobile Ad-Hoc Networks . . . . . . . . 383
Lan Wang, Xianping Wang, and Stephan Olariu
22 Topology Control Problems for Wireless Ad Hoc Networks . . . . . . . . . . . . . . . . . . 419
Gruia Călinescu, Errol L. Lloyd, and S. S. Ravi
23 QoS Multimedia Multicast Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 441
Ion Mandoiu, Alex Olshevsky, and Alex Zelikovsky
24 Overlay Networks for Peer-to-Peer Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459
Andréa W. Richa, Christian Scheideler, and Stefan Schmid
25 Data Broadcasts on Multiple Wireless Channels: Exact and
Time-Optimal Solutions for Uniform Data and Heuristics
for Nonuniform Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483
Alan A. Bertossi, Cristina M. Pinotti, and Romeo Rizzi
26 Strategies for Aggregating Time-Discounted Information in
Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 505
Xianping Wang and Stephan Olariu
27 Approximation and Exact Algorithms for Optimally Placing a Limited
Number of Storage Nodes in a Wireless Sensor Network . . . . . . . . . . . . . . . . . . . . . 541
Gianlorenzo D’Angelo, Alfredo Navarra, and Cristina M. Pinotti
Contents ix

28 Approximation Algorithms for the Primer Selection, Planted Motif

Search, and Related Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567
Sudha Balla, Jaime Davila, Marius Nicolae, and Sanguthevar Rajasekaran
29 Dynamic and Fractional Programming-Based Approximation Algorithms
for Sequence Alignment with Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599
Abdullah N. Arslan and Ömer Eğecioğlu
30 Approximation Algorithms for the Selection of Robust Tag SNPs . . . . . . . . . . . . . 617
Yao-Ting Huang, Kui Zhang, Ting Chen, and Kun-Mao Chao
31 Large-Scale Global Placement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 635
Jason Cong and Joseph R. Shinnerl
32 Histograms, Wavelets, Streams, and Approximation . . . . . . . . . . . . . . . . . . . . . . . . . 673
Sudipto Guha
33 Color Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 691
Zhigang Xiang
34 A GSO-Based Swarm Algorithm for Odor Source Localization in
Turbulent Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 711
Joseph Thomas and Debasish Ghose
35 Digital Reputation for Virtual Communities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 739
Roberto Battiti and Anurag Garg
36 Approximation for Influence Maximization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 757
Jing Yuan, Weili Wu, and Wen Xu
37 Approximation and Heuristics for Community Detection . . . . . . . . . . . . . . . . . . . . 761
Jing Yuan, Weili Wu, and Sarat Chandra Varanasi
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 771
 Preface

More than half a century ago the research community began analyzing formally the quality of the solu-
tions generated by heuristics. The heuristics with guaranteed performance bounds eventually became
known as approximation algorithms. The idea behind approximation algorithms was to develop proce-
dures to generate provable near-optimal solutions to optimization problems that could not be solved
efficiently by the computational techniques available at that time. With the advent of the theory of
NP-completeness in the early 1970s, approximation algorithms became more prominent as the need to
generate near optimal solutions for NP-hard optimization problems became the most important avenue
for dealing with computational intractability. As it was established in the 1970s, for some problems one
could generate near optimal solutions quickly, while for other problems it was established that generat-
ing provably good suboptimal solutions was as difficult as generating optimal ones. Other approaches
based on probabilistic analysis and randomized algorithms became popular in the 1980s. The intro-
duction of new techniques to solve linear programming problems started a new wave for developing
approximation algorithms that matured and saw tremendous growth in the 1990s. To deal with the inap-
proximable problems, in a practical sense, there were a few techniques introduced in the 1980s and 1990s.
These methodologies have been referred to as metaheuristics and may be viewed as problem indepen-
dent methodologies that can be applied to sets of problems. There has been a tremendous amount of
research in metaheuristics during the past three decades. During the last 25 years or so, approximation
algorithms have attracted considerably more attention. This was a result of a stronger inapproximability
methodology that could be applied to a wider range of problems and the development of new approxi-
mation algorithms. In the last decade there has been an explosion of new applications arising from most
disciplines.
As we have witnessed, there has been tremendous growth in areas of approximation algorithms and
metaheuristics. The second edition of this handbook includes new chapters, updated chapters and chap-
ters with traditional content that did not warrant an update. For this second edition we have partitioned
the handbook into two volumes. Volume 1 covers methodologies and traditional applications. Volume
2 covers contemporary and emerging applications. More specifically volume 1 discusses the different
methodologies to design approximation algorithms and metaheuristics, as well as the application of these
methodologies to traditional combinatorial optimization problems. Volume 2 discusses application of
these methodologies to classical problems in computational geometry and graphs theory, as well as in
large-scale and emerging application areas. Chapter 1 in both of these volumes presents an overview of
approximation algorithms and metaheuristics as well as as an overview of both volumes of this handbook.
It has been a decade since the first edition and our authors have experienced all sorts of different tran-
sitions. Several authors expanded their families while writing chapters for the first edition. The babies
born at that time are now more than ten years old! A few of the authors and the editor have now retired
from their day to day obligations, but continue to be active in research. A couple of the authors became
presidents of universities while others now hold prestigious chaired positions or high level positions at
their institutions. But sadly, Rajeev Motwani and Ivan Stojmenovic, well-known researchers and authors
of first edition chapters, passed away. Since their chapters did not changed significantly, they remain as

xi
xii Preface

co-authors of their chapters. Also, Imreh Csanád, a new author for the second edition, passed away in
2017. They are all missed greatly by their families, friends, and the entire research community.
We have collected in this volume a large amount of material with the goal of making it as complete as
possible. We apologize in advance for any omissions and would like to invite all of you to propose new
chapters for future editions of this handbook. Our research area will continue to grow and we are confi-
dent that the following words from an old song “The best is yet to come, you ain’t seen nothing yet. . .”
applies to our research area. We look forward to the next decade in which new challenges and opportuni-
ties await the new, as well as the established, researchers. We look forward to working on problems arising
in new emerging applications and editing the third edition of this handbook.
I like to acknowledge the University of California, Santa Barbara (UCSB) for providing me the time
and support needed to develop the first and second editions of this handbook for the past 12 years. I also
want to thank my wife, Dorothy, our daughters, Jeanmarie, Alexis and Julia, and our sons, Teofilo and
Paolo for their moral support, love, encouragement, understanding, and patience, throughout the project
and my tenure at UCSB.

Teofilo F Gonzalez
Professor Emeritus of Computer Science
University of California, Santa Barbara
 Contributors

Abdullah N. Arslan Gruia Călinescu

Department of Computer Science Department of Computer Science
Texas A & M University Illinois Institute of Technology
Commerce, Texas Chicago, Illinois

Kun-Mao Chao
Sudha Balla
Department of Computer Science and
Department of Computer Science and
Information Engineering
Engineering
National Taiwan University
University of Connecticut
Taipei, Taiwan, Republic of China
Storrs, Connecticut

Ting Chen
Roberto Battiti Department of Computer Science and
Department of Computer Science and Technology
Telecommunications Tsinghua University
University of Trento Beijing, China
Trento, Italy
Marco Chiarandini
Department of Mathematics and
Alan A. Bertossi
Computer Science
Department of Computer Science
University of Southern Denmark
University of Bologna
Odense, Denmark
Bologna, Italy

Jason Cong
Maria J. Blesa Department of Computer Science
ALBCOM, Computer Science Department University of California
Universitat Politècnica de Catalunya (UPC) - Los Angeles, California
BarcelonaTech
Barcelona, Spain
Artur Czumaj
Department of Computer Science
Christian Blum University of Warwick
ALBCOM, Dept. Llenguatges i Coventry, United Kingdom
Sistemes Informátics
Universitat Politècnica de Catalunya (UPC) - Gianlorenzo D’Angelo
BarcelonaTech Gran Sasso Science Institute
Barcelona, Spain L’Aquila, Italy

xiii
xiv
Jaime Davila
Department of Computer Science and
Engineering
University of Connecticut
Storrs, Connecticut
Irina Dumitrescu
School of Mathematics
University of New South Wales
Sydney, Australia
Ömer Eğecioğlu
Department of Computer Science
University of California
Santa Barbara, California
Cristina G. Fernandes
Department of Computer Science
University of Sao Paulo
Sao Paulo, Brazil
Anurag Garg
Department of Computer Science and
Telecommunications
University of Trento
Trento, Italy
Hossein Ghasemalizadeh
Department of Computer Engineering
Shahid Bahonar University
Kerman, Iran
Debasish Ghose
Department of Aerospace Engineering
Indian Institute of Science
Bangalore, India
Teofilo F. Gonzalez
Department of Computer Science
University of California
Santa Barbara, California
Arturo Gonzalez-Gutierrez
Facultad de Ingeniería
Universidad Autónoma de Querétaro
Querétaro, México
Joachim Gudmundsson
University of Sydney
Sydney, Australia
Contributors
Sudipto Guha
Department of Computer
Information Sciences
University of Pennsylvania
Philadelphia, Pennsylvania
Yao-Ting Huang
Department of Computer Science and
Information Engineering
National Chung Cheng University
Chiayi County, Taiwan, Republic of China
Ragesh Jaiswal
Department of Computer Science and
Engineering
Indian Institute of Technology Delhi
New Delhi, India
Konstantinos G. Kakoulis
Department of Industrial Design Engineering
Western Macedonia
University of Applied Sciences
Florina, Greece
Christian Knauer
Universität Bayreuth
Bayreuth, Germany
Stavros G. Kolliopoulos
Department of Informatics and
Telecommunications
National and Kapodistrian
University of Athens
Athens, Greece
Andrzej Lingas
Department of Computer Science
Lund University
Lund, Sweden
Errol L. Lloyd
Department of Computer and
Information Sciences
University of Delaware
Newark, Delaware
Ion Mandoiu
Department of Computer Science
University of Connecticut
Storrs, Connecticut
Contributors
Alfredo Navarra
Department of Computer Science and
Mathematics
University of Perugia
Perugia, Italy
Marius Nicolae
Department of Computer Science and
Engineering
University of Connecticut
Storrs, Connecticut
Sotiris E. Nikoletseas
Computer Engineering &
Informatics Department
Patras University
Patras, Greece
Zeev Nutov
Department of Mathematics and
Computer Science
The Open University of Israel
Raanana, Israel
Stephan Olariu
Department of Computer Science
Old Dominion University
Norfolk, Virginia
Alex Olshevsky
Department of Electrical and
Computer Engineering
Boston University
Boston, Massachusetts
Cristina M. Pinotti
Department of Computer Science and
Mathematics
University of Perugia
Perugia, Italy
Sanguthevar Rajasekaran
Department of Computer Science and
Engineering
University of Connecticut
Storrs, Connecticut
Christoforos L. Raptopoulos
Patras University
Patras, Greece
xv
S. S. Ravi
Biocomplexity Institute of Virginia Tech
Blacksbury, Virginia
Mohammadreza Razzazi
Department of Computer Engineering
Amirkabir University of Technology
Tehran, Iran
Andréa W. Richa
Department of Computer Science
Arizona State University
Tempe, Arizona
Romeo Rizzi
Department of Computer Science
University of Verona
Verona, Italy
Daniel J. Rosenkrantz
University at Albany – SUNY
Albany, New York
Pedro M. Ruiz
DIIC
University of Murcia
Murcia, Spain
Christian Scheideler
Department of Computer Science
University of Paderborn
Paderborn, Germany
Stefan Schmid
Department of Computer Science
Aalborg University
Aalborg, Denmark
Sandeep Sen
Department of Computer Science and
Engineering
Indian Institute of Technology Delhi
New Delhi, India
Joseph R. Shinnerl
Mentor Graphics Corporation
Freemont, California
xvi Contributors

Michiel Smid Xianping Wang

School of Computer Science Department of Computer Science
Carleton University Old Dominion University
Ottawa, Canada Norfolk, Virginia

Paul G. Spirakis Bang Ye Wu

Computer Engineering & Department of Computer Science and
Informatics Department Information Engineering
Patras University National Chung Cheng University
Patras, Greece Chiayi County, Taiwan,
Republic of China
Ivan Stojmenovic
University of Ottawa
Ottawa, Canada Weili Wu
Department of Computer Science
Thomas Stützle University of Texas at Dallas
Institut de Recherches Interdisciplinaires Richardson, Texas
et de Développements en Intelligence
Artificielle (IRIDIA) Zhigang Xiang
Université Libre de Bruxelles Queens College of the City
Brussels, Belgium University of New York
Flushing, New York
Chuan Yi Tang
Department of Computer Science Wen Xu
National Tsing Hua University Department of Mathematics and
Hsinchu, Taiwan, Republic of China Computer Science
Giri K. Tayi Texas Woman’s University
University at Albany – SUNY Denton, Texas
Albany, New York
Jing Yuan
Joseph Thomas Department of Computer Science
Department of Aerospace Engineering University of Texas at Dallas
Indian Institute of Science Richardson, Texas
Bangalore, India
Alex Zelikovsky
Ioannis G. Tollis Department of Computer Science
Department of Computer Science Georgia State University
University of Crete Atlanta, Georgia
Rethimno, Greece

Sarat Chandra Varanasi Kui Zhang

Department of Computer Science Department of Mathematical Sciences
University of Texas at Dallas Michigan Technological University
Richardson, Texas Houghton, Michigan

Lan Wang Si Qing Zheng

Department of Computer Science Department of Computer Science
Old Dominion University University of Texas at Dallas
Norfolk, Virginia Richardson, Texas
 1
Introduction,
Overview, and
Notation
1.1 Introduction............................................................................... 1
1.2 Overview .................................................................................... 2
Local Search, Artificial Neural Networks, and
Metaheuristics • Multiobjective Optimization, Reoptimization,
Sensitivity Analysis, and Stability • Traditional Applications
• Computational Geometry and Graph Applications • Large-Scale and
Emerging Applications
1.3 Definitions and Notation ........................................................... 16
Time and Space Complexity • NP-Completeness • Performance
Teofilo F. Gonzalez Evaluation of Algorithms

1.1 Introduction
Approximation algorithms were formally introduced in the 1960s to generate near-optimal solutions to
optimization problems that could not be solved efficiently by the computational techniques available at
that time. With the advent of the theory of NP-completeness in the early 1970s, the area became more
prominent as the need to generate near-optimal solutions for NP-hard optimization problems became
the most important avenue for dealing with computational intractability. As it was established in the
1970s, for some problems it is possible to generate near-optimal solutions quickly, whereas for other
problems generating provably good suboptimal solutions is as difficult as generating optimal ones. Com-
putational approaches based on probabilistic analysis and randomized algorithms became popular in the
1980s. The introduction of new techniques to solve linear programming problems started a new wave
of approximation algorithms that matured and saw tremendous growth in the 1990s. There were a few
techniques introduced in the 1980s and 1990s to deal, in the practical sense, with inapproximable prob-
lems. These methodologies have been referred to as metaheuristics and include Simulated Annealing
(SA), Ant Colony Optimization (ACO), Evolutionary Computation (EC), Tabu Search (TS), Memetic
Algorithms (MAs), and so on. Other previously established methodologies such as local search, back-
tracking, and branch-and-bound were also explored at that time. There has been a tremendous amount
of research in metaheuristics during the past three decades. These techniques have been evaluated experi-
mentally and have demonstrated their usefulness for solving problems that are arising in practice. During
the last 25 years or so, approximation algorithms have attracted considerably more attention. This was a
result of a stronger inapproximability methodology that could be applied to a wider range of problems
and the development of new approximation algorithms for problems arising in established and emerg-
ing application areas. Polynomial Time Approximation Schemes (PTASs) were introduced in the 1960s

1
2 Handbook of Approximation Algorithms and Metaheuristics

and the more powerful Fully Polynomial Time Approximation Schemes (FPTASs) were introduced in the
1970s. Asymptotic PTAS (APTAS) and Asymptotic FPTAS (AFPTAS), and Fully Polynomial Randomized
Approximation Schemes (FPRASs) were introduced later on.
Today approximation algorithms enjoy a stature comparable to that of algorithms in general and the
area of metaheuristics has established itself as an important research area. The new stature is a byproduct
of a natural expansion of research into more practical areas where solutions to real-world problems are
expected, as well as by the higher level of sophistication required to design and analyze these new proce-
dures. The goal of approximation algorithms and metaheuristics is to provide the best possible solutions
and to guarantee that such solutions satisfy certain criteria. This two-volume handbook houses these two
approaches and thus covers all the aspects of approximations. We hope it will serve you as a valuable
reference for approximation methodologies and applications.
Approximation algorithms and metaheuristics have been developed to solve a wide variety of problems.
A good portion of these algorithms have only theoretical value due to the fact that their time complexity
is a high-order polynomial or they have a huge constant associated with their time complexity bound.
However, these results are important because they establish what is possible, and it may be that in the
near future these algorithms will be transformed into practical ones. Other approximation algorithms do
not suffer from this pitfall, but some were designed for problems with limited applicability. However, the
remaining approximation algorithms have real-world applications. Given this, there is a huge number of
important application areas, including new emerging ones, where approximation algorithms and meta-
heuristics have barely penetrated and we believe there is an enormous potential for their use. Our goal is
to collect a wide portion of the approximation algorithms and metaheuristics in as many areas as possible,
as well as to introduce and explain in detail the different methodologies used to design these algorithms.

1.2 Overview
Our overview in this section is devoted mainly to the earlier years. The individual chapters in the two
volumes discuss in detail the recent research accomplishments in different subareas. This section will also
serve as an overview of both volumes of this handbook. Volume 1, Chapter 2 discusses some of the basic
methodologies and applies them to classical problems.
Even before the 1960s, researchers in applied mathematics and graph theory had established upper
and lower bounds for certain properties of graphs. For example, bounds had been established for the
chromatic number, achromatic number, chromatic index, maximum clique, maximum independent set,
and so on. Some of these results could be seen as the precursors of approximation algorithms. By the
1960s it was understood that there were problems that could be solved efficiently, whereas for other
problems all the known algorithms required exponential time in the worst case. Heuristics were being
developed to find quick solutions to problems that appeared to be computationally difficult to solve.
Researchers were experimenting with heuristics, branch-and-bound procedures, and iterative improve-
ment frameworks and were evaluating their performance when solving actual problem instances. There
were many claims being made, not all of which could be substantiated, about the performance of the
procedures being developed to generate optimal and suboptimal solutions to combinatorial optimization
problems.
Half a century ago (1966), Ronald L. Graham [1] formally introduced approximation algorithms. He
analyzed the performance of list schedules for scheduling tasks on identical machines, a fundamental
problem in scheduling theory.
Problem: Scheduling tasks on identical machines.
Instance: Set of n tasks (T1 , T2 , . . . , Tn ) with processing time requirements t1 , t2 , . . . , tn , partial order
C defined over the set of tasks to enforce task dependencies, and a set of m identical machines.
Objective: Construct a schedule with minimum makespan. A schedule is an assignment of tasks to time
intervals on the machines in such a way that (1) each task Ti is processed continuously for ti units
Introduction, Overview, and Notation 3

of time by one of the machines, (2) each machine processes at most one task at a time, and (3) the
precedence constraints are satisfied (i.e., machines cannot commence the processing of a task until
all of its predecessors have been completed). The makespan of a schedule is the time at which all the
machines have completed processing the tasks.
The list-scheduling procedure is given an ordering of the tasks specified by a list L. Then the procedure
finds the earliest time t when a machine is idle, and an unassigned task is available (i.e., all its predecessor
tasks have been completed). It assigns the leftmost available task in the list L to an idle machine at time t
and this step is repeated until all the tasks have been scheduled.
The main result in Reference 1 was proving that for every problem instance I, the schedule generated
by this policy has a makespan that is bounded above by (2 − 1/m) times the optimal makespan for the
instance. This is called the approximation ratio or approximation factor for the algorithm. We also say that
the algorithm is a (2 − 1/m)-approximation algorithm. This criterion for measuring the quality of the
solutions generated by an algorithm remains as one of the most important ones in use today. The second
contribution in Reference 1 was showing that the approximation ratio (2−1/m) is the best possible for list
schedules, that is, the analysis of the approximation ratio for this algorithm cannot be improved. This was
established by presenting problem instances (for all m and n ≥ 2m − 1) and lists for which the schedule
generated by the procedure has a makespan equal to 2−1/m times the optimal makespan for the instance.
A restricted version of the list-scheduling algorithm is analyzed in detail in Volume 1, Chapter 2.
The third important aspect of the results in Reference 1 was showing that list scheduling may have
anomalies. To explain this we need to define some terms. The makespan of the list schedule for instance
I using list L is denoted by fL (I). Suppose that instance I  is a slightly modified version of instance I. The
modification is such that we intuitively expect that fL (I  ) ≤ fL (I). But this is not always true, so there is an
anomaly. For example, suppose that I  is I, except that I  has an additional machine. Intuitively fL (I  ) ≤
fL (I) because with one additional machine, tasks should be finished earlier or at worst at the same time
as when there is one fewer machine. But this is not always the case for list schedules, there are problem
instances and lists for which fL (I  ) > fL (I). This is called an anomaly. Our expectation would be valid
if list scheduling would generate minimum makespan schedules, but we have a procedure that generates
suboptimal solutions. Such guarantees are not always possible in this type of environment. List schedules
suffer from other anomalies, for example, relaxing the precedence constraints or decreasing the execution
time of the tasks. In both cases, one would expect schedules with smaller or the same makespans. But,
that is not always the case. Volume 1, Chapter 2 presents problem instances where anomalies occur.
The main reason for discussing anomalies now is that even today, numerous papers are being published
and systems are being deployed where “common sense”-based procedures are being introduced without
any analytical justification and/or thorough experimental validation. Anomalies show that since we live
for the most part in a “suboptimal world,” the effect of our decisions is not always the intended one
(unintended consequences). One can design approximation algorithms that do not suffer from certain
types of anomalies but probably not for all possible ones.
Other classical problems with numerous applications are the traveling salesperson, Steiner tree, and
spanning tree problems, which will be formally defined later on. Even before the 1960s there were several
well-known polynomial time algorithms to construct minimum weight spanning trees for edge-weighted
graphs [2]. These simple greedy algorithms have low-order polynomial time complexity bounds. It was
well known at that time that the same type of procedures does not generate an optimal tour for the trav-
eling salesperson problem (TSP) and does not construct optimal Steiner trees. However, in 1968 E. F.
Moore (as discussed in Reference 3) showed that for any set of points P in metric space LM ≤ LT ≤ 2LS ,
where LM , LT , and LS are the total weight of a minimum weight spanning tree, a minimum weight tour
(solution) for the TSP, and minimum weight Steiner tree for P, respectively. Since every spanning tree is
a Steiner tree, the above-mentioned bounds show that when using a minimum weight spanning tree to
approximate the Steiner tree we have a solution (Steiner tree) whose weight is at most twice the weight
of an optimal Steiner tree. In other words, any algorithm that generates a minimum weight spanning tree
4 Handbook of Approximation Algorithms and Metaheuristics

is a 2-approximation algorithm for the Steiner tree problem. Furthermore, this approximation algorithm
takes no more time than an algorithm that constructs a minimum weight spanning tree for edge weighted
graphs [2], as such an algorithm can be used to construct an optimal spanning tree for a set of points
in metric space. The above-mentioned bound is established by defining a transformation from any min-
imum weight Steiner tree into a TSP tour with weight at most 2LS . Therefore, LT ≤ 2LS [3]. Then by
observing that the deletion of an edge in an optimum tour for the TSP problem results in a spanning tree,
it follows that LM < LT . Volume 1, Chapter 3 discusses this approximation algorithm and its analysis
in more detail. The Steiner ratio is defined as LS /LM . The earlier arguments show that the Steiner ratio
√
is at least 12 . Gilbert and Pollak [3] conjectured that the Steiner ratio in the Euclidean plane equals 23
(the 0.86603 . . . conjecture). A proof of this conjecture and improved approximation algorithms for the
Steiner tree problem are discussed in Volume 1, Chapter 36.
The above-mentioned constructive proof can be applied to a minimum weight spanning tree to gener-
ate a tour for the TSP problem. The construction takes polynomial time and results in a 2-approximation
algorithm for the TSP problem. This approximation algorithm for the TSP is also referred to as the
double spanning tree algorithm and is discussed in Volume 1, Chapters 3 and 27. Improved approxima-
tion algorithms for the TSP and algorithms for its generalizations are discussed in Volume 1, Chapters 3,
27, 34, 35, and Volume 2, Chapter 2. The approximation algorithm for the Steiner tree problem just dis-
cussed is explained in Volume 1, Chapter 3, and improved approximation algorithms and applications
are discussed in Volume 1, Chapters 36, 37, and Volume 2, Chapter 2. Volume 2, Chapter 14, discusses
approximation algorithms for computationally intractable variations of the spanning tree problem.
In 1969, Graham [4] studied the problem of scheduling tasks on identical machines but restricted
to independent tasks, that is, the set of precedence constraints is empty. He analyzed the Largest
Processing Time first (LPT) scheduling rule, which is list scheduling where the list of tasks L is arranged
in nonincreasing order of their processing requirements. His elegant proof established that the LPT
procedure generates a schedule with makespan at most 43 − 3m 1
times the makespan of an optimal
4 1
schedule, that is, the LPT scheduling algorithm has a 3 − 3m approximation ratio. He also showed that
the analysis is best possible for all m and n ≥ 2m + 1. For n ≤ 2m tasks the approximation ratio is smaller
and under some conditions, LPT generates an optimal makespan schedule. Graham [4], following a
suggestion by D. Kleitman and D. Knuth, considered list schedules where the first portion of the list L
consists of k tasks (without loss of generality, assume k is a multiple of m) with the longest processing
times arranged by their starting times in an optimal schedule for these k tasks (only). Then the list L
has the remaining n − k tasks in any order. The approximation ratio for this list schedule using list
L is 1 + m−1
m+k . An optimal schedule for the longest k tasks can be constructed in O(n + m ) time by
k

a straight forward branch and bound algorithm. In other words, this algorithm has an approximation
ratio 1 +  and time complexity O(n + m(m−1−m)/ ). For any fixed constants m and , the algorithm
constructs in polynomial (linear) time with respect to n, a schedule with makespan at most 1 +  times
the optimal makespan. Note that for a fixed constant m the time complexity is polynomial with respect
to n, but it is not polynomial with respect to 1/. This was the first algorithm of its kind and later on,
it was called polynomial time approximation scheme (PTAS). Volume 1, Chapter 8 discusses different
PTASs. Additional PTAS appear in Volume 1, Chapter 36 and Volume 2, Chapters 2 and 5. The proof
techniques presented in References 1, 4 are outlined in Volume 1, Chapter 2 and have been extended
to apply to other problems. There is an extensive body of literature for approximation algorithms and
metaheuristics for scheduling problems. Volume 1, Chapters 38, 39, and Volume 2, Chapter 25 discuss
interesting approximation algorithms and heuristics for scheduling problems. The scheduling handbook
[5] is an excellent source for scheduling algorithms, models, and performance analysis.
The development of NP-completeness theory in the early 1970s by Cook [6], Karp [7], and others for-
mally introduced the notion that there is a large class of decision problems (the answer to these problems
is a simple yes or no) that are computationally equivalent. This means that either every problem in this
class has a polynomial time algorithm that solves it, or none of them do. Furthermore, this question is
Introduction, Overview, and Notation 5

the same as the P = NP question, a classical open problem in computational complexity. This question is
to determine whether or not the set of languages recognized in polynomial time by deterministic Turing
machines is the same as the set of languages recognized in polynomial time by nondeterministic Turing
machines. The conjecture has been that P = NP, and thus the hardest problems in NP would not be
solvable in polynomial time. The computationally equivalent decision problems in this class are called
NP-complete problems. The scheduling on identical machines problem discussed earlier is an optimiza-
tion problem. Its corresponding decision problem has its input augmented by an integer value B, and
the yes–no question is to determine whether or not there is a schedule with makespan at most B. An
optimization problem whose corresponding decision problem is NP-complete is called an NP-hard prob-
lem. Therefore, scheduling tasks on identical machines is an NP-hard problem. The TSP and the Steiner
tree problem are also NP-hard problems. The minimum weight spanning tree problem can be solved in
polynomial time and it is not an NP-hard problem under the assumption that P = NP. The next section
discusses NP-completeness in more detail. There is a long list of practical problems arising in many dif-
ferent fields of study that are known to be NP-hard problems [8]. Because of this, the need to cope with
these computationally intractable problems was recognized earlier on. Since then approximation algo-
rithms became a central area of research activity. Approximation algorithms offered a way to circumvent
computational intractability by paying a price when it comes to the quality of the solutions generated. But
a solution can be generated quickly. In other words and another language, “no te fijes en lo bien, fijate en
lo rápido.” Words used to describe my golf playing ability when I was growing up.
In the early 1970s, Garey et al. [9] as well as Johnson [10,11] developed the first set of polynomial time
approximation algorithms for the bin packing problem. The analysis of the approximation ratio for these
algorithms is asymptotic, which is different from those for the scheduling problems discussed earlier. We
will define this notion precisely in the next section, but the idea is that the ratio holds when the value of an
optimal solution is greater than some constant. Research on the bin packing problem and its variants has
attracted very talented investigators who have generated more than 1000 papers, most of which deal with
approximations. This work has been driven by numerous applications in engineering and information
sciences (Volume 1, Chapters 28, 29, 30, and 31).
Johnson [12] developed polynomial time algorithms for the sum of subsets, max satisfiability, set cover,
graph coloring, and max clique problems. The algorithms for the first two problems have a constant ratio
approximation, but for the other problems the approximation ratio is ln n and n . Sahni [13,14] developed
a PTAS for the knapsack problem. Rosenkrantz et al. [15] developed several constant ratio approximation
algorithms for the TSP that satisfy the triangle inequality (or simply, defined over metric graphs). This
version of the problem is defined over edge weighted complete graphs, rather than for points in metric
space as in Reference 3. These algorithms have an approximation ratio of two.
Sahni and Gonzalez [16] showed that there were a few NP-hard optimization problems for which
the existence of a constant ratio polynomial time approximation algorithm implies the existence of a
polynomial time algorithm to generate an optimal solution. In other words, complexity of generating a
constant ratio approximation and an optimal solution are computationally equivalent problems. For these
problems, the approximation problem is NP-hard or simply inapproximable (under the assumption that
P = NP). Later on, this notion was extended to mean that there is no polynomial time algorithm with
approximation ratio r for a problem under some complexity theoretic hypothesis. The approximation
ratio r is called the inapproximability ratio (see Chapter 17 in the first edition of this handbook).
The k-min-cluster problem is one of these inapproximable problems. Given an edge-weighted undi-
rected graph, the k-min-cluster problem is to partition the set of vertices into k sets so as to minimize the
sum of the weight of the edges with endpoints in the same set. The k-maxcut problem is defined as the
k-min-cluster problem, except that the objective is to maximize the sum of the weight of the edges with
endpoints in different sets. Even though these two problems have exactly the same set of feasible and opti-
mal solutions, there is a linear time algorithm for the k-maxcut problem that generates k-cuts with weight
at least k−1
k times the weight of an optimal k-cut [16], whereas approximating the k-min-cluster problem
6 Handbook of Approximation Algorithms and Metaheuristics

is a computationally intractable problem. The former problem has the property that a near-optimal solu-
tion may be obtained as long as partial decisions are made optimally, whereas for the k-min-cluster an
optimal partial decision may turn out to force a terrible overall solution. For the k-min-cluster problem if
one makes a mistake at some iteration, one will end up with a solution that is far from optimal. Whereas
for the k-maxcut problem one can make many mistakes and still end up with a near-optimal solution.
A similar situation arises when you make a mistake an exam where almost everyone receives a perfect
score, versus a course where the average score is about 50% of the points.
Another interesting problem whose approximation problem is NP-hard is the TSP problem [16]. This
is not exactly the same version of the TSP problem discussed earlier, which we said has several constant
ratio polynomial time approximation algorithms. Given an edge-weighted undirected graph, the TSP is
to find a least weight tour, that is, to find a least weight (simple) path that starts at vertex 1, visits each
vertex in the graph exactly once, and ends at vertex 1. The weight of a path is the sum of the weight of its
edges. The weights of the edges are unrelated and the approximation problem is NP-hard. The version of
the TSP problem studied in Reference 15 is limited to metric graphs, that is, the graph is complete (all the
edges are present) and the set of edge weights satisfies the triangle inequality (which means that the weight
of the edge joining vertex i and j is less than or equal to the weight of any path from vertex i to vertex j).
This version of the TSP problem is equivalent to the one studied by E. F. Moore [3]. The approximation
algorithms given in References 3, 15 can be easily adapted to provide a constant ratio approximation to
the version of the TSP problem where the tour is defined as visiting each vertex in the graph at least
once. Since Moore’s approximation algorithms for the metric Steiner tree and metric TSP are based on
the same idea, one would expect that the Steiner tree problem defined over arbitrarily weighted graphs
is NP-hard to approximate. However, this is not the case. Moore’s algorithm [3] can be modified to be a
2-approximation algorithm for this more general Steiner tree problem.
As pointed out in Reference 17, Levner and Gens [18] added a couple of problems to the list of prob-
lems that are NP-hard to approximate. Garey and Johnson [19] show that the max clique problem has the
property that if for some constant r there is a polynomial time r-approximation algorithm, then there is
a polynomial time r -approximation for any constant r such that 0 < r < 1. Since that time researchers
have tried many different algorithms for the clique problem, none of which were constant ratio approx-
imation algorithms, and it was conjectured that none existed under the assumption that P = NP. This
conjecture has been proved.
A PTAS is said to be an FPTAS if its time complexity is polynomial with respect to n (the problem size)
and 1/. The first FPTAS was developed by Ibarra and Kim [20] for the knapsack problem. Sahni [21]
developed three different techniques based on rounding, interval partitioning, and separation to construct
FPTAS for sequencing and scheduling problems. These techniques have been extended to other problems
and are discussed in Volume 1, Chapter 9. Horowitz and Sahni [22] developed FPTAS for scheduling
on processors with different processing speeds. Reference 17 discusses a simple O(n3 /) FPTAS for the
knapsack problem developed by Babat [23,24]. Lawler [25] developed techniques for maximum speed-up
FPTAS for the knapsack and related problems. Volume 1, Chapter 9 presents different methodologies to
design FPTAS. Garey and Johnson [26] showed that if any problem in a class of NP-hard optimization
problems that satisfy certain properties has a FPTAS, then P = NP. The properties are that the objective
function value of every feasible solution is a positive integer, and the problem is strongly NP-hard. A prob-
lem is strongly NP-hard if the problem is NP-hard even when the magnitude of the maximum number in
the input is bounded by a polynomial on the input length. For example, the TSP problem is strongly
NP-hard, whereas the knapsack problem is not, under the assumption that P = NP (see Volume 1,
Chapter 9).
Lin and Kernighan [27] developed elaborate heuristics that established experimentally that instances
of the TSP with up to 110 cities can be solved to optimality with 95% confidence in O(n2 ) time. This
was an iterative improvement procedure applied to a set of randomly selected feasible solutions. The
process was to perform k pairs of link (edge) interchanges that improved the length of the tour. However,
Papadimitriou and Steiglitz [28] showed that for the TSP no local optimum of an efficiently searchable
Introduction, Overview, and Notation 7

neighborhood can be within a constant factor of the optimal value unless P = NP. Since then there has
been quite a bit of research activity in this area. Deterministic and stochastic local search in efficiently
searchable, as well as in very large neighborhoods, are discussed in Volume 1, Chapters 16, 17, 18, and 19.
Volume 1, Chapter 13 discusses issues relating to the empirical evaluation of approximation algorithms
and metaheuristics.
Perhaps the best known approximation algorithm for the TSP defined over metric graphs is the one
by Christofides [29]. The approximation ratio for this algorithm is 32 , which is smaller than the approxi-
mation ratio of 2 for the algorithms reported in References 3, 15. However, looking at the bigger picture
that includes the time complexity of the approximation algorithms, Christofides algorithm is not of the
same order as the ones given in References 3, 15. Therefore, neither approximation algorithm dominates
the other as one has a smaller time complexity bound, whereas the other (Christofides algorithm) has a
smaller worst case approximation ratio.
Ausiello et al. [30] introduced the differential ratio, which is another way of measuring the quality of the
solutions generated by approximation algorithms. Differential ratio destroys the artificial dis-symmetry
between “equivalent” minimization and maximization problems (e.g., the k-maxcut and the k-min clus-
ter discussed earlier) when it comes to approximation. This ratio uses the difference between the worst
possible solution minus the solution generated by the algorithm, divided by the difference between the
worst solution minus the best solution. Cornuejols et al. [31] also discussed a variation of differential ratio
approximations. They wanted the ratio to satisfy the following property: “A modification of the data that
adds a constant to the objective function value should also leave the error measure unchanged.” That is,
the “error” by the approximation algorithm should be the same as before. Differential ratio and its exten-
sions are discussed in Volume 1, Chapter 15, along with other similar notions [30]. Ausiello et al. [30]
introduced reductions that preserve approximability. Since then there have been several new types of
approximation preserving reductions. The main advantage of these reductions is that they enable us to
define large classes of optimization problems that behave in the same way with respect to approxima-
tion. Informally, the class of NP Optimization (NPO) problems, is the set of all optimization problems
 which can be “recognized” in polynomial time (see Volume 1, Chapter 14 for a formal definition). An
NPO problem  is said to be in APX, if it has a constant approximation ratio polynomial time algorithm.
The class PTAS consists of all NPO problems which have PTAS. The class FPTAS is defined similarly.
Other classes, Poly-APX, Log-APX, and Exp-APX, have also been defined (see Volume 1, Chapter 14).
One of the main accomplishments at the end of the 1970s was the development of a polynomial time
algorithm for Linear Programming (LP) problems by Khachiyan [32]. This result had a tremendous
impact on approximation algorithm research and started a new wave of approximation algorithms. Two
subsequent research accomplishments were at least as significant as Khachiyan’s [32] result. The first one
was a faster polynomial time algorithm for solving linear programming problems developed by Karmakar
[33]. The other major accomplishment was the work of Grötschel et al. [34,35]. They showed that it is
possible to solve a linear programming problem with an exponential number of constraints (with respect
to the number of variables) in time which is polynomial in the number of variables and the number of
bits used to describe the input, given a separation oracle plus a bounding ball and a lower bound on the
volume of the feasible solution space. Given a solution, the separation oracle determines in polynomial
time whether or not the solution is feasible, and if it is not it finds a constraint that is violated. Volume 1,
Chapter 10 gives an example of the use of this approach. Important developments have taken place during
the past 30 years. The books [35,36] are excellent references for linear programming theory, algorithms,
and applications.
Because of the above-mentioned results, the approach of formulating the solution to an NP-hard prob-
lem as an integer linear programming problem and so solving the corresponding linear programming
problem became very popular. This approach is discussed in Volume 1, Chapter 2. Once a fractional
solution is obtained, one uses rounding to obtain a feasible solution to the original NP-hard problem.
The rounding may be deterministic or randomized, and it may be very complex (meta-rounding). LP
rounding is discussed in Volume 1, Chapters 2, 4, 7, 8, 10, 11, and Volume 2, Chapters 8 and 11.
8 Handbook of Approximation Algorithms and Metaheuristics

Independently, Johnson [12] and Lovász [37] developed efficient algorithms for the set cover with
approximation ratio of 1 + ln d, where d is the maximum number of elements in each set. Chvátal [38]
extended this result to the weighted set cover problem. Subsequently, Hochbaum [39] developed an
algorithm with approximation ratio f , where f is the maximum number of sets containing any of the ele-
ments in the set. This result is normally inferior to the one by Chvátal [38], but it is more attractive for the
weighted vertex cover problem, which is a restricted version of the weighted set cover. For this subproblem
it is a 2-approximation algorithm. A few months after Hochbaum’s initial result,* Bar-Yehuda and Even
[40] developed a primal-dual algorithm with the same approximation ratio as the one in Reference 39. The
algorithm in Reference 40 does not require the solution of an LP problem, as in the case of the algorithm in
Reference 39, and its time complexity is linear. But it uses linear programming theory to establish this
result. This was the first primal-dual approximation algorithm, though some previous algorithms may
also be viewed as falling into this category. An application of the primal-dual approach as well as related
ones are discussed in Volume 1, Chapter 2. Volume 1, Chapters 4, 34, and Volume 2, Chapter 23 discuss
several primal-dual approximation algorithms. Volume 1, Chapter 12 discusses “distributed” primal-dual
algorithms. These algorithms make decisions by using only “local” information.
In the mid 1980s Bar-Yehuda and Even [41] developed a new framework parallel to the primal-dual
methods. They called it local ratio; it is simple and requires no prior knowledge of linear programming. In
Volume 1, Chapter 2 we explain the basics of this approach, and Volume 1, Chapter 6, and Reference 42
covers extensively this technique as well as its extensions.
Raghavan and Thompson [43] were the first to apply randomized rounding to relaxations of lin-
ear programming problems to generate solutions to the problem being approximated. This field has
grown tremendously. LP randomized rounding is discussed in Volume 1, Chapters 2, 4, 7, 10, 11, and
Volume 2, Chapter 11, and deterministic rounding is discussed in Volume 1, Chapters 2, 7, 8, 10,
and Volume 2, Chapters 8 and 11. A disadvantage of LP-rounding is that a linear programming prob-
lem needs to be solved. This takes polynomial time with respect to the input length, but in this case it
means the number of bits needed to represent the input. In contrast, algorithms based on the primal-
dual approach are for the most part faster, since they take polynomial time with respect to the number of
“objects” in the input. However, the LP-rounding approach can be applied to a much larger class of prob-
lems and it is more robust since the technique is more likely to be applicable after changing the objective
function and/or constraints for a problem.
The first Asymptomatic PTAS (APTAS) was developed by Fernandez de la Vega and Lueker [44] for the
bin packing problem. The first Asymptomatic FPTAS (AFPTAS) for the same problem was developed by
Karmakar and Karp [45]. These approaches are discussed in Volume 1, Chapter 15. FPRASs are discussed
in Volume 1, Chapter 11.
In the 1980s, new approximation algorithms were developed as well as PTAS and FPTAS based on
different approaches. These results are reported throughout the handbook. One difference was the appli-
cation of approximation algorithm to other areas of research activity (very large-scale integration (VLSI),
bioinformatics, network problems) as well as other problems in established areas.
In the late 1980s Papadimitriou and Yannakakis [46] defined MaxSNP as a subclass of NPO. These
problems can be approximated within a constant factor and have a nice logical characterization. They
showed that if MAX3SAT, vertex cover, MAXCUT, and some other problems in the class could be approx-
imated in polynomial time with an arbitrary precision, then all MaxSNP problems would. This fact
was established by using approximation preserving reductions (see Volume 1, Chapter 14). In the 1990s
Arora et al. [47], using complex arguments (see Chapter 17 in the 1st edition of this handbook), showed

*
Here we are referring to the time when these results appeared as technical reports. Note that from the journal publication
dates, the order is reversed. You will find throughout the chapters similar patterns. To add to the confusion, a large number
of papers have also been published in conference proceedings. Since it would be very complex to include the dates when the
initial technical report and conference proceedings were published, we only include the latest publication date. Please keep
this in mind when you read the chapters and, in general, the computer science literature.
Introduction, Overview, and Notation 9

that MAX3SAT is hard to approximate within a factor of 1 +  for some  > 0 unless P = NP. Thus,
all problems in MaxSNP do not admit a PTAS unless P = NP. This work led to major developments in
the area of approximation algorithms, including inapproximability results for other problems, a bloom
of approximation preserving reductions, discovery of new inapproximability classes, and construction of
approximation algorithms achieving optimal or near-optimal ratios.
Feige et al. [48] showed that the clique problem could not be approximated to within some constant
value. Applying the previous results in Reference 26 it showed that the clique problem is inapproximable
to within any constant. Feige [49] showed that set cover is inapproximable within ln n. Other inapprox-
imable results appear in References 50, 51. Chapter 17 in the first edition of this handbook discusses all of
this work in detail.
There are many other very interesting results that have been published in the past 25 years. Goemans
and Williamson [52] developed improved approximation algorithms for the maxcut and satisfiability
problems using semidefinite programming (SDP). This seminal work opened a new venue for the design
of approximation algorithms. Chapter 8 of the first edition of this handbook discusses this work as
well as developments in this area. Goemans and Williamson [53] also developed powerful techniques
for designing approximation algorithms based on the primal-dual approach. The dual-fitting and factor
revealing approach is used in Reference 54. Techniques and extensions of these approaches are discussed
in Volume 1, Chapters 4, 12, 34, and Volume 2, Chapter 23.
This concludes our overview of Section I of Volume 1 of this handbook. Section 1.2.1 presents an
overview of Section II of Volume 1 dealing with local search, artificial neural nets and metaheuristics.
Section 1.2.2 presents an overview of multiobjective optimization, reoptimization, sensitivity analysis and
stability all of which is in Section III of Volume 1. In the last couple of decades we have seen approxi-
mation algorithms being applied to traditional combinatorial optimization problems as well as problems
arising in other areas of research activity. These areas include: VLSI design automation, networks (wired,
sensor, and wireless), bioinformatics, game theory, computational geometry, and graph problems. In
Sections 1.2.3 through 1.2.5 we elaborate further on these applications. Section 1.2.3 overviews tradi-
tional application covered in Section IV of Volume 1. Sections 1.2.4 and 1.2.5 overview contemporary
and emerging application which are covered in this volume.

1.2.1 Local Search, Artificial Neural Networks, and Metaheuristics

Local search techniques have a long history; they range from simple constructive and iterative
improvement algorithms to rather complex methodologies that require significant fine-tuning, such as
evolutionary algorithms (EAs) or SA. Local search is perhaps one of the most natural ways to attempt to
find an optimal or suboptimal solution to an optimization problem. The idea of local search is simple:
Start from a solution and improve it by making local changes until no further progress is possible.
Deterministic local search algorithms are discussed in Volume 1, Chapter 16. Volume 1, Chapter 17
covers stochastic local search algorithms. These are local search algorithms that make use of randomized
decisions, for example, in the context of generating initial solutions or when determining search steps.
When the neighborhood to search for the next solution is very large, finding the best neighbor to move
to is many times an NP-hard problem. Therefore, an approximation solution is needed at this step. In
Volume 1, Chapter 18 the issues related to very large-scale neighborhood search are discussed from the
theoretical, algorithmic, and applications point of view.
Reactive Search advocates the use of simple subsymbolic machine learning to automate the parameter
tuning process and make it an integral (and fully documented) part of the algorithm. Parameters are
normally tuned through a feedback loop that many times depends on the user input. Reactive search
attempts to mechanize this process. Volume 1, Chapter 19 discusses issues arising during this process.
Artificial neural networks have been proposed as a tool for machine learning and many results have
been obtained regarding their application to practical problems in robotics control, vision, pattern recog-
nition, grammatical inferences, and other areas. Recently, neural networks have found many applications
10 Handbook of Approximation Algorithms and Metaheuristics

in the forefront of Artificial Intelligence (AI) and Machine Learning (ML). For example, Google’s open-
source deep learning neural network tools as well as the Cloud, have been a catalyst for the development of
these new applications.* Recently artificial neural networks have been used to improve energy utilization
during certain periods in Google’s massive data centers with impressive results. The whole process has
been fully automated without the use of training data. Once trained (automatically or manually), the neu-
ral network will compute an input/output mapping which, if the training data was representative enough,
will closely match the unknown rule which produced the original data. Neural networks are discussed in
Volume 1, Chapter 20 and may be viewed as heuristics to solve a large class of problems.
The work of Lin and Kernighan [27] sparked the study of modern heuristics, which have evolved and
are now called metaheuristics. The term metaheuristics was coined by Glover [55] in 1986 and in general
means “to find beyond in an upper level.” The most popular metaheuristics include: TS, SA, ACO, EC,
iterated local search (ILC), MAs, plus many others that keep up popping-up every year. One of the moti-
vations for the study of metaheuristics is that it was recognized early on that constant ratio polynomial
time approximation algorithms are not likely to exist for a large class of practical problems [16]. Meta-
heuristics do not guarantee that near-optimal solutions will be found quickly for all problem instances.
However, these complex programs do find near-optimal solutions for many problem instances that arise
in practice. These procedures have wide range of applicability, which is their most appealing aspect.
There are many ways of viewing metaheuristics. Some are single point while others are population
based. In the former case one solution is modified over and over again until the algorithm terminates.
Whereas in the latter case a set of solutions is carried throughout the execution of the algorithm. Some
metaheuristics have a fixed neighborhood where moves can be made, whereas others have variable neigh-
borhoods throughout the execution of the procedure. Metaheuristic algorithms may use memory to
influence future moves and some are memoryless. Metaheuristics may be nature-inspired or algorith-
mic based. But no matter how they work, they have all been used successfully to solve many practical
problems. In what follows we discuss several metaheuristics.
The term tabu search was coined by Glover [55]. TS is based on adaptive memory and responsive explo-
ration. The former allows for an effective and efficient search of the solution space. The latter is used
to guide the search process by imposing restraints and inducements based on the information collected.
Intensification and diversification are controlled by the information collected, rather than by a random
process. Volume 1, Chapter 21 discusses many different aspects of TS as well as problems to which it
has been applied. Most recently, applications in the field of quantum computing as well an open-source
hybrid quantum solver for D-wave systems have emerged. These developments have placed TS at the
forefront of the field of quantum computing.
In the early 1980s Kirkpatrick et al. [56] and independently Černý [57] introduced SA as a randomized
local search algorithm to solve combinatorial optimization problems. SA is a local search algorithm, which
means that it starts with an initial solution and then searches through the solution space by iteratively
generating a new solution that is “near” to it. But, sometimes the moves are to a worse solution to escape
local optimal solutions. This method is based on statistical mechanics (Metropolis algorithm). It was
heavily inspired by an analogy between the physical annealing process of solids and the problem of solving
large combinatorial optimization problems. Volume 1, Chapter 25 in the 1st edition of this handbook
discusses this approach in detail.
EC is a metaphor for building, applying, and studying algorithms based on Darwinian principles of
natural selection. Algorithms that are based on evolutionary principles are called EAs. They are inspired
by nature’s capability to evolve living beings well adapted to their environment. There have been a variety
of slightly different EAs proposed over the years. Three different strands of EAs were developed indepen-
dently of each other over time. These are evolutionary programming (EP) introduced by Fogel [58] and
Fogel et al. [59], evolutionary strategies (ESs) proposed by Rechenberg [60], and genetic algorithms (GAs)

*
https://www.tensorflow.org/ and https://aiexperiments.withgoogle.com
Introduction, Overview, and Notation 11

initiated by Holland [61]. GAs are mainly applied to solve discrete problems. Genetic programming (GP)
and scatter search (SS) are more recent members of the EA family. EAs can be understood from a unified
point of view with respect to their main components and the way they explore the search space. EC is
discussed in Volume 1, Chapter 22.
Volume 1, Chapter 23 presents an overview of ACO—a metaheuristic inspired by the behavior of real
ants. ACO was proposed by Dorigo et al. [62] in the early 1990s as a method for solving hard combi-
natorial optimization problems. ACO algorithms may be considered to be part of swarm intelligence,
the research field that studies algorithms inspired by the observation of the behavior of swarms. Swarm
intelligence algorithms are made up of simple individuals that cooperate through self-organization.
MAs were introduced by Moscato [63] in the late 1980s to denote a family of metaheuristics, which
can be characterized as the hybridization of different algorithmic approaches for a given problem. It is a
population-based approach in which a set of cooperating and competing agents are engaged in periods
of individual improvement of the solutions while they sporadically interact. An important component
is problem and instance-dependent knowledge, which is used to speed-up the search process. A complete
description is given in Volume 1, Chapter 27 of the 1st edition of this handbook.

1.2.2 Multiobjective Optimization, Reoptimization,

Sensitivity Analysis, and Stability
Volume 1, Chapter 24 discusses multiobjective combinatorial optimization. This is important in practice
since quite often a decision is rarely made with only one criterion. There are many examples of such
applications in the areas of transportation, communication, biology, finance, and also computer science.
Volume 1, Chapter 24 covers stochastic local search algorithms for multiobjective optimization problems.
Volume 1, Chapter 25 discusses reoptimization which tries to address the question: Given an optimal or
nearly optimal solution to some instance of an NP-hard optimization problem and a small local change is
applied to the instance, can we use the knowledge of the old solution to facilitate computing a reasonable
solution for the new locally modified instance? As pointed out in Volume 1, Chapter 25, we should not
expect major results for optimal solutions to NP-hard problems, but there are some interesting results for
approximations. Sensitivity analysis is the dual problem, meaning that given an optimal or suboptimal
solution, find all the set of related instances for which the solution remains optimal or near-optimal.
Volume 1, Chapter 26 covers sensitivity analysis, which has been around for more than 40 years. The
aim is to study how variations affect the optimal solution value. In particular, parametric analysis studies
problems whose structure is fixed, but where cost coefficients vary continuously as a function of one or
more parameters. This is important when selecting the model parameters in optimization problems. On
the other hand, Volume 1, Chapter 27 considers a newer area which is called stability. By this we mean
how the complexity of a problem depends on a parameter whose variation alters the space of allowable
instances.

1.2.3 Traditional Applications

We have used the label “traditional applications” to refer to more established combinatorial optimization
problems. Some of these application can be categorized differently and vice-versa. The problems studied
in this part of the handbook fall into the following categories: bin packing, packing, traveling salesperson,
Steiner tree, scheduling, planning, generalized assignment, linear ordering, and submodular functions
maximization. Let us briefly discuss these categories.
One of the fundamental problems in approximations is the bin packing problem. Volume 1, Chapter 28
discusses online and offline algorithms for one-dimensional bin packing. Volume 1, Chapters 29 and 30
discuss variants of the bin packing problem. This include variations that fall into the following type of
problems: the number of items packed is maximized while keeping the number of bins fixed; there is a
bound on the number of items that can be packed in each bin; dynamic bin packing, where each item
12 Handbook of Approximation Algorithms and Metaheuristics

has an arrival and departure time; the item sizes are not known, but the ordering of the weights is known;
items may be fragmented while packing them into fixed capacity bins, but certain items cannot be assigned
to the same bin; bin stretching; variable sized bin packing problem; the bin covering problem; black and
white bin packing; bin packing with rejection; batched bin packing; maximal resource bin packing; and
bin packing with fragile items.
Volume 1, Chapter 31 discusses several ways to generalize the bin packing problem to more dimen-
sions. Two- and three-dimensional strip packing, bin packing in dimensions two and higher, vector
packing, and several other variations are discussed. Cutting and packing problems with important appli-
cations in the wood, glass, steel and leather industries, as well as in LSI and VLSI design, newspaper
paging, and container and truck loading are discussed in Volume 1, Chapter 32. For several decades,
cutting and packing problems have attracted the attention of researchers in various areas including
operations research, computer science, and manufacturing. Volume 1, Chapter 33 survey heuristics,
metaheuristics, and exact algorithms for two-dimensional packing of general shapes. These problems have
many practical applications in various industries such as the garment, shoe, and shipbuilding industries
and many variants have been considered in the literature.
Very interesting approximation algorithms for the prize collecting traveling salesperson problem is
studied in Volume 1, Chapter 34. In this problem a salesperson has to collect a certain amount of prizes
(the quota) by visiting cities. A known prize can be collected in every city. Volume 1, Chapter 35 discusses
branch and bound algorithms for the TSP problem. These algorithms have been implemented to run in
a multicomputer environment. A general software tool for running branch and bound algorithms in a
distributed environment is discussed. This framework may be used for almost any divide-and-conquer
computation. With minor adjustments this tool can take any algorithm defined as a computation over
directed acyclic graph, where the nodes refer to computations and the edges specify a precedence relation
between computations, and run in a distributed environment.
Approximation algorithms for the Steiner tree problem are discussed in Volume 1, Chapter 36. This
problem has applications in several research areas. One of this area is VLSI physical design. In Volume 1,
Chapter 37 practical approximations for a restricted Steiner tree problem are discussed.
Volume 1, Chapter 38 surveys problems at the intersection of two scientific fields: graph theory and
scheduling. These problems can either be viewed as scheduling dependent jobs where jobs have resource
requirements, or as graph coloring minimization involving different objective functions. Applications
include: wire minimization in VLSI design, minimizing the distance traveled by a robot moving in a
warehouse, session scheduling on a path, and resource constrained scheduling.
Automated planning consists of finding a sequence of actions that transforms an initial state into one
of the goal states. Planning is widely applicable and has been used in such diverse application domains
as spacecraft control, planetary rover operations, automated nursing aides, image processing, business
process generation, computer security, and automated manufacturing. Volume 1, Chapter 39 discusses
approximation algorithms and heuristics for problems falling into this category.
Volume 1, Chapter 40 presents heuristics and metaheuristics for the generalized assignment prob-
lem. This problem is a natural generalization of combinatorial optimization problems including bipartite
matching, knapsack and bin packing problems, and has many important applications in flexible manufac-
turing systems, facility location, and vehicle routing problems. Computational evaluation of the different
procedures is discussed.
The linear ordering problem is discussed in Volume 1, Chapter 41. Versions of this problem were
initially studied back in 1938 and 1958 in Economics. Exact algorithms, constructive heuristics, local
search algorithms, and metaheuristics as well as computational results are discussed extensively in this
chapter.
Volume 1, Chapter 42 discusses approximation algorithms and metaheuristics for submodular func-
tion maximization. These problems play a major role in combinatorial optimization. A few examples
of these functions include: cut functions of graphs and hypergraphs, rank functions of matroids, and
covering functions.
Another random document with
no related content on Scribd:
artist, the decorator, the paper mills’ agent, and, last of all, the printer and the
binder. This was not the way the old-time printers had planned their books.
With all their mechanical limitations, they had followed architectural lines kept
consistent and harmonious because controlled by a single mind, while the
finished volume of the eighteen-nineties was a composite production of many
minds, with no architectural plan. No wonder that the volumes manufactured,
even in the most famous Presses, failed to compare with those produced in
Venice by Jenson and Aldus four centuries earlier!
When I succeeded John Wilson as head of the University Press in 1895, I
determined to carry out the resolution I had formed four years earlier, while
sitting in on the Eugene Field conference, of following the example of the
early master-printers so far as this could be done amidst modern conditions.
Some of my publisher friends were partially convinced by my contention that
if the printer properly fulfilled his function he must know how to express his
clients’ mental conception of the physical attributes of prospective volumes in
terms of type, paper, presswork, and binding better than they could do it
themselves. The Kelmscott publications, which appeared at this time, were of
great value in emphasizing my contention, for William Morris placed printing
back among the fine arts after it had lapsed into a trade.
I had no idea, when I presented my plan, of persuading my friends to
produce typographical monuments. No demand has ever existed for volumes
of this type adequate to the excessive cost involved by the perfection of
materials, the accuracy of editorial detail, the supreme excellence of
typography and presswork, and the glory of the binding. Sweynheim and
Pannartz, Gutenberg’s successors, were ruined by their experiments in Greek;
the Aldine Press in Venice was saved only by the intervention of Jean Grolier;
Henri Étienne was ruined by his famous Thesaurus, and Christophe Plantin
would have been bankrupted by his Polyglot Bible had he not retrieved his
fortunes by later and meaner publications. Nor was I unmindful of similar
examples that might have been cited from more modern efforts, made by
ambitious publishers and printers.
What I wanted to do was to build low-cost volumes upon the same
principles as de luxe editions, eliminating the expensive materials but retaining
the harmony and consistency that come from designing the book from an
architectural standpoint. It adds little to the expense to select a type that
properly expresses the thought which the author wishes to convey; or to have
the presses touch the letters into the paper in such a way as to become a part
of it, without that heavy impression which makes the reverse side appear like
an example of Braille; or to find a paper (even made by machine!) soft to the
feel and grateful to the eye, on which the page is placed with well-considered
margins; or to use illustrations or decorations, if warranted at all, in such a way
as to assist the imagination of the reader rather than to divert him from the
text; to plan a title page which, like the door to a house, invites the reader to
open it and proceed, its type lines carefully balanced with the blank; or to bind
(even in cloth!) with trig squares and with design or lettering in keeping with
the printing inside.
By degrees the publishers began to realize that this could be done, and
when once established, the idea of treating the making of books as a
manufacturing problem instead of as a series of contracts with different
concerns, no one of which knew what the others were doing, found favor. The
authors also preferred it, for their literary children now went forth to the
world in more becoming dress. Thus serving in the capacity of book architect
and typographical advisor, instead of merely as a contrasting printer, these
years have been lived in a veritable Kingdom of Books, in company with
interesting people,—authors and artists as well as publishers,—in a delightfully
intimate way because I have been permitted to be a part of the great
adventure.

During these years I have seen dramatic changes. Wages were somewhat
advanced between 1891 and the outbreak of the World War, but even at this
latter date the cost of manufacturing books was less than half of what it is now.
This is the great problem which publishers have to face today. When the cost
of everything doubled after the World War, the public accepted the necessity
of paying twice the price for a theater ticket as a matter of course; but when
the retail price of books was advanced in proportion to the cost of
manufacture, there was a great outcry among buyers that authors, publishers,
and booksellers were opportunists, demanding an unwarranted profit. As a
matter of fact, the novel which used to sell at $1.35 per copy should now sell
at $2.50 if the increased costs were properly apportioned. The publisher today
is forced to decline many promising first novels because the small margin of
profit demands a comparatively large first edition.
 Unless a publisher can sell 5,000 copies as a minimum it is impossible for
him to make any profit upon a novel. Taking this as a basis, and a novel as
containing 320 pages, suppose we see how the $2.00 retail price distributes
itself. The cost of manufacture, including the typesetting, electrotype plates,
cover design, jacket, brass dies, presswork, paper, and binding, amounts to 42
cents per copy (in England, about 37 cents). The publisher’s cost of running
his office, which he calls “overhead,” is 36 cents per copy. The minimum
royalty received by an author is 10 per cent. of the retail price, which would
give him 20 cents. This makes a total cost of 98 cents a copy, without
advertising. But a book must be advertised.
Every fifty dollars spent in advertising on a five thousand edition adds a
cent to the publisher’s cost. The free copies distributed for press reviews
represent no trifling item. A thousand dollars is not a large amount to be spent
for advertising, and this means 20 cents a copy on a 5000 edition, making a
total cost of $1.18 per copy and reducing the publisher’s profit to 2 cents, since
he sells a two-dollar book to the retail bookseller for $1.20. The bookseller
figures that his cost of doing business is one-third the amount of his sales, or,
on a two-dollar book, 67 cents. This then shows a net profit to the retail
bookseller of 13 cents, to the publisher of 2 cents, and to the author of 20
cents a copy.
Beyond this, there is an additional expense to both bookseller and
publisher which the buyer of books is likely to overlook. It is impossible to
know just when the demand for a book will cease, and this means that the
publisher and the bookseller are frequently left with copies on hand which
have to be disposed of at a price below cost. This is an expense that has to be
included in the book business just as much as in handling fruit, flowers, or
other perishable goods.
When a publisher is able to figure on a large demand for the first edition,
he can cut down the cost of manufacture materially; but, on the other hand,
this is at least partially offset by the fact that authors whose books warrant
large first editions demand considerably more than 10 per cent. royalty, and
the advertising item on a big seller runs into large figures.

I wish I might say that I had seen a dramatic change in the methods
employed in the retail bookstores! There still exists, with a few notable
exceptions, the same lack of realization that familiarity with the goods one has
to sell is as necessary in merchandizing books as with any other commodity.
Salesmen in many otherwise well-organized retail bookstores are still painfully
ignorant of their proper functions and indifferent to the legitimate
requirements of their prospective customers.
Some years ago, when one of my novels was having its run, I happened to
be in New York at a time when a friend was sailing for Europe. He had
announced his intention of purchasing a copy of my book to read on the
steamer, and I asked him to permit me to send it to him with the author’s
compliments. Lest any reader be astonished to learn that an author ever buys a
copy of his own book, let me record the fact that except for the twelve which
form a part of his contract with the publisher, he pays cash for every copy he
gives away. Mark Twain dedicated the first edition of The Jumping Frog to “John
Smith.” In the second edition he omitted the dedication, explaining that in
dedicating the volume as he did, he had felt sure that at least all the John
Smiths would buy books. To his consternation he found that they all expected
complimentary copies, and he was hoist by his own petard!
With the idea of carrying out my promise to my friend, I stepped into one
of the largest bookstores in New York, and approached a clerk, asking him for
the book by title. My pride was somewhat hurt to find that even the name was
entirely unfamiliar to him. He ran over various volumes upon the counter, and
then turned to me, saying, “We don’t carry that book, but we have several
others here which I am sure you would like better.”
“Undoubtedly you have,” I agreed with him; “but that is beside the point. I
am the author of the book I asked for, and I wish to secure a copy to give to a
friend. I am surprised that a store like this does not carry it.”
Leaning nonchalantly on a large, circular pile of books near him, the clerk
took upon himself the education of the author.
“It would require a store much larger than this to carry every book that is
published, wouldn’t it?” he asked cheerfully. “Of course each author naturally
thinks his book should have the place of honor on the bookstalls, but we have
to be governed by the demand.”
It was humiliating to learn the real reason why this house failed to carry
my book. I had to say something to explain my presumption even in assuming
that I might find it there, so in my confusion I stammered,
 “But I understood from the publishers that the book was selling very
well.”
“Oh, yes,” the clerk replied indulgently; “they have to say that to their
authors to keep them satisfied!”
With the matter thus definitely settled, nothing remained but to make my
escape as gracefully as circumstances would permit. As I started to leave, the
clerk resumed his standing position, and my eye happened to rest on the pile
of perhaps two hundred books upon which he had been half-reclining. The
jacket was strikingly familiar. Turning to the clerk I said severely,
“Would you mind glancing at that pile of books from which you have just
risen?”
“Oh!” he exclaimed, smiling and handing me a copy, “that is the very book
we were looking for, isn’t it?”
It seemed my opportunity to become the educator, and I seized it.
“Young man,” I said, “if you would discontinue the practice of letting my
books support you, and sell a few copies so that they might support me, it
would be a whole lot better for both of us.”
“Ha, ha!” he laughed, graciously pleased with my sally; “that’s a good line,
isn’t it? I really must read your book!”

The old-time publisher is passing, and the author is largely to blame. I have
seen the close association—in many cases the profound friendship—between
author and publisher broken by the commercialism fostered by some literary
agents and completed by competitive bids made by one publishing house to
beguile a popular author away from another. There was a time when a writer
was proud to be classified as a “Macmillan,” or a “Harper” author. He felt
himself a part of the publisher’s organization, and had no hesitation in taking
his literary problems to the editorial advisor of the house whose imprint
appeared upon the title pages of his volumes. A celebrated Boston authoress
once found herself absolutely at a standstill on a partially completed novel. She
confided her dilemma to her publisher, who immediately sent one of his
editorial staff to the rescue. They spent two weeks working together over the
manuscript, solved the problems, and the novel, when published, was the most
successful of the season.
Several publishers have acknowledged to me that in offering unusually
high royalties to authors they have no expectation of breaking even, but that to
have a popular title upon their list increases the sales of their entire line. The
publisher from whom the popular writer is filched has usually done his share
in helping him attain his popularity. The royalty he pays is a fair division of the
profits. He cannot, in justice to his other authors, pay him a further premium.
Ethics, perhaps, has no place in business, but the relation between author
and publisher seems to me to be beyond a business covenant. A publisher may
deliberately add an author to his list at a loss in order to accomplish a specific
purpose, but this practice cannot be continued indefinitely. A far-sighted
author will consider the matter seriously before he becomes an opportunist.

In England this questionable practice has been of much slower growth.

The House of Murray, in London, is one of those still conducted on the old-
time basis. John Murray IV, the present head of the business, has no interest in
any author who comes to him for any reason other than a desire to have the
Murray imprint upon his book. It is more than a business. The publishing
offices at 50a, Albemarle Street adjoin and open out of the Murray home. In
the library is still shown the fireplace where John Murray III burned Byron’s
Memoirs, after purchasing them at an enormous price, because he deemed that
their publication would do injury to the reputation of the writer and of the
House itself.
John Murray II was one of the publishers of Scott’s Marmion. In those days
it was customary for publishers to share their contracts. Constable had
purchased from Scott for £1,000 the copyright of Marmion without having
seen a single line, and the honorarium was paid the author before the poem was
completed or the manuscript delivered. Constable, however, promptly
disposed of a one-fourth interest to Mr. Miller of Albemarle Street, and
another one fourth to John Murray, then of Fleet Street.
By 1829 Scott had succeeded in getting into his own hands nearly all his
copyrights, one of the outstanding items being this one-quarter interest in
Marmion held by Mr. Murray. Longmans and Constable had tried in vain to
purchase it. When, however, Scott himself approached Murray through
Lockhart, the following letter from Mr. Murray was the result:
So highly do I estimate the honour of being even in so small a degree the publisher
of the author of the poem that no pecuniary consideration whatever can induce me to
part with it. But there is a consideration of another kind that would make it painful
to me if I were to retain it a moment longer. I mean the knowledge of its being required
by the author, into whose hands it was spontaneously resigned at the same instant that
I read the request.
There has always been a vast difference in authors in the attitude they
assume toward the transformation of their manuscripts into printed books.
Most of them leave every detail to their publishers, but a few take a deep and
intelligent personal interest. Bernard Shaw is to be included in the latter group.
A leading Boston publisher once telephoned me that an unknown English
author had submitted a manuscript for publication, but that it was too
socialistic in its nature to be acceptable. Then the publisher added that the
author had asked, in case this house did not care to publish the volume, that
arrangements be made to have the book printed in this country in order to
secure American copyright.
“We don’t care to have anything to do with it,” was the statement; “but I
thought perhaps you might like to manufacture the book.”
“Who is the author?” I inquired.
“It’s a man named Shaw.”
“What is the rest of his name?”
“Wait a minute and I’ll find out.”
Leaving the telephone for a moment, the publisher returned and said,
“His name is G. Bernard Shaw. Did you ever hear of him?”
“Yes,” I replied; “I met him last summer in London through Cobden-
Sanderson, and I should be glad to undertake the manufacture of the book for
Mr. Shaw.”
“All right,” came the answer. “Have your boy call for the manuscript.”
This manuscript was Man and Superman.
 From that day and for many years, Shaw and I carried on a desultory
correspondence, his letters proving most original and diverting. On one
occasion he took me severely to task for having used two sizes of type upon a
title page. He wrote four pages to prove what poor taste and workmanship this
represented, and then ended the letter with these words, “But, after all, any
other printer would have used sixteen instead of two, so I bless you for your
restraint!”
We had another lengthy discussion on the use of apostrophes in printing.
“I have made no attempt to deal with the apostrophes you introduce,” he
wrote; “but my own usage is carefully considered and the inconsistencies are
only apparent. For instance, Ive, youve, lets, thats, are quite unmistakable, but Ill,
hell, shell, for I’ll, he’ll, she’ll, are impossible without a phonetic alphabet to
distinguish between long and short e. In such cases I retain the apostrophe, in
all others I discard it. Now you may ask me why I discard it. Solely because it
spoils the printing. If you print a Bible you can make a handsome job of it
because there are no apostrophes or inverted commas to break up the
letterpress with holes and dots. Until people are forced to have some
consideration for a book as something to look at as well as something to read,
we shall never get rid of these senseless disfigurements that have destroyed all
the old sense of beauty in printing.”
“Ninety-nine per cent. of the secret of good printing,” Shaw continued, “is
not to have patches of white or trickling rivers of it trailing down a page, like
rain-drops on a window. Horrible! White is the enemy of the printer. Black,
rich, fat, even black, without gray patches, is, or should be, his pride. Leads and
quads and displays of different kinds of type should be reserved for insurance
prospectuses and advertisements of lost dogs.…”
His enthusiasm for William Morris’ leaf ornaments is not shared by all
booklovers. Glance at any of the Kelmscott volumes, and you will find these
glorified oak leaves scattered over the type page in absolutely unrelated
fashion,—a greater blemish, to some eyes, than occasional variation in spacing.
Shaw writes:
If you look at one of the books printed by William Morris, the greatest printer of
the XIX century, and one of the greatest printers of all the centuries, you will see that
he occasionally puts in a little leaf ornament, or something of the kind. The idiots in
America who tried to imitate Morris, not understanding this, peppered such things all
over their “art” books, and generally managed to stick in an extra large quad before
each to show how little they understood about the business. Morris doesn’t do this in
his own books. He rewrites the sentence so as to make it justify, without bringing one
gap underneath another in the line above. But in printing other people’s books, which
he had no right to alter, he sometimes found it impossible to avoid this. Then, sooner
than spoil the rich, even color of his block of letterpress by a big white hole, he filled it
up with a leaf.
Do not dismiss this as not being “business.” I assure you, I have a book which
Morris gave me, a single copy, by selling which I could cover the entire cost of printing
my books, and its value is due solely to its having been manufactured in the way I
advocate; there’s absolutely no other secret about it; and there is no reason why you
should not make yourself famous through all the ages by turning out editions of
standard works on these lines whilst other printers are exhausting themselves in dirty
felt end papers, sham Kelmscott capitals, leaf ornaments in quad sauce, and then
wondering why nobody in Europe will pay twopence for them, whilst Kelmscott books
and Doves Press books of Morris’ friends, Emery Walker and Cobden-Sanderson,
fetch fancy prices before the ink is thoroughly dry.… After this I shall have to get you
to print all my future books, so please have this treatise printed in letters of gold and
preserved for future reference
CHAPTER III

Friends through Type

 III
FRIENDS THROUGH TYPE

In 1903 I again visited Italy to continue my study of the art of printing in

the old monasteries and libraries, sailing on the S. S. Canopic from Boston to
Naples. Among the passengers on board I met Horace Fletcher, returning to
his home in Venice. At that time his volume Menticulture was having a
tremendous run. I had enjoyed reading the book, and in its author I discovered
a unique and charming personality; in fact, I have never met so perfect an
expression of practical optimism. His humor was infectious, his philosophy
appealing, his quiet persistency irresistible.
To many people the name of Horace Fletcher has become associated with
the Gladstonian doctrine of excessive chewing, but this falls far short of the
whole truth. His scheme was the broadest imaginable, and thorough
mastication was only the hub into which the other spokes of the wheel of his
philosophy of life were to be fitted. The scheme was nothing less than a
cultivation of progressive human efficiency. Believing that absolute health is
the real basis of human happiness and advancement, and that health depends
upon an intelligent treatment of food in the mouth together with knowledge
of how best to furnish the fuel that is actually required to run the human
engine, Horace Fletcher sought for and found perfect guides among the
natural human instincts and physiologic facilities, and demonstrated that his
theories were facts.

During the years that followed I served as his typographic mentor. He was
eager to try weird and ingenious experiments to bring out the various points of
his theories through unique typographical arrangement (see opp. page). It
required all my skill and diplomacy to convince him that type possessed rigid
limitations, and that to gain his emphasis he must adopt less complicated
methods. From this association we became the closest of friends, and
presuming upon this relation I used to banter him upon being so casual. His
copy was never ready when the compositors needed it; he was always late in
returning his proofs. The manufacture of a Fletcher book was a hectic
experience, yet no one ever seemed to take exceptions. This was characteristic
of the man. He moved and acted upon suddenly formed impulses, never
planning ahead yet always securing exactly what he wanted, and those
inconvenienced the most always seemed to enjoy it.

A Page of Horace Fletcher Manuscript

“I believe,” he used to say, “in hitching one’s wagon to a star, but I always
keep my bag packed and close at hand ready to change stars at a moment’s
notice. It is only by doing this that you can give things a chance to happen to
you.”
 Among the volumes Fletcher had with him on board ship was one he had
purchased in Italy, printed in a type I did not recognize but which greatly
attracted me by its beauty. The book bore the imprint: Parma: Co’tipi Bodoniani.
Some weeks later, in a small, second-hand bookstore in Florence, I happened
upon a volume printed in the same type, which I purchased and took at once
to my friend, Doctor Guido Biagi, at the Laurenziana Library.
“The work of Giambattista Bodoni is not familiar to you?” he inquired in
surprise. “It is he who revived in Italy the glory of the Aldi. He and Firmin
Didot in Paris were the fathers of modern type design at the beginning of the
nineteenth century.”
“Is this type still in use?” I inquired.
“No,” Biagi answered. “When Bodoni died there was no one worthy to
continue its use, so his matrices and punches are kept intact, exactly as he left
them. They are on exhibition in the library at Parma, just as the old Plantin
relics are preserved in the museum at Antwerp.”
 GIAMBATTISTA BODONI, 1740–1813
From Engraving at the Bibliothèque Nationale, Paris

I immediately took steps through our Ambassador at Rome to gain

permission from the Italian Government to recut this face for use in America.
After considerable difficulty and delay this permission was granted, with a
proviso that I should not allow any of the type made from my proposed
matrices to get into the hands of Italian printers, as this would detract from
the prestige of the city of Parma. It was a condition to which I was quite
willing to subscribe! Within a year I have received a prospectus from a revived
Bodoni Press at Montagnola di Lugano, Switzerland, announcing that the
exclusive use of the original types of Giambattista Bodoni has been given them
by the Italian Government. This would seem to indicate that the early
governmental objections have disappeared.
While searching around to secure the fullest set of patterns, I stumbled
upon the fact that Bodoni and Didot had based their types upon the same
model, and that Didot had made use of his font particularly in the wonderful
editions published in Paris at the very beginning of the nineteenth century. I
then hurried to Paris to see whether these matrices were in existence. There,
after a search through the foundries, I discovered the original punches, long
discarded, in the foundry of Peignot, to whom I gave an order to cast the
different sizes of type, which I had shipped to America.
This was the first type based on this model ever to come into this country.
The Bodoni face has since been recut by typefounders as well as for the
typesetting machines, and is today one of the most popular faces in common
use. Personally I prefer the Bodoni letter to that of Didot (see opp. page). The
Frenchman succumbed to the elegance of his period, and by lightening the
thin lines robbed the design of the virility that Bodoni retained. I am not in
sympathy with the excessive height of the ascending letters, which frequently
extend beyond the capitals; but when one considers how radical a departure
from precedent this type was, he must admire the skill and courage of the
designers. William Morris cared little for it,—“The sweltering hideousness of
the Bodoni letter,” he exclaimed; “the most illegible type that was ever cut,
with its preposterous thicks and thins”; while Theodore L. De Vinne, in his
Practice of Typography, writes:
The beauty of the Bodoni letters consists in their regularity, in their clearness, and
in their conformity to the taste of the race, nation, and age in which the work was first
written, and finally in the grace of the characters, independent of time or place.
When authorities differ to such a wide extent, the student of type design
must draw his own conclusions!
 The Bodoni Letter (bottom) compared with the Didot Letter (top)

Fletcher’s idea of an appointment was something to be kept if or when

convenient, yet he never seemed to offend any one. He did nothing he did not
wish to do, and his methods of extricating himself from unwelcome
responsibilities always amused rather than annoyed. “If you don’t want to do a
thing very badly,” he confided to me on one such occasion, “do it very badly.”
 HORACE FLETCHER IN 1915

On board the Canopic Fletcher was surrounded by an admiring and

interested group. General Leonard Wood was on his way to study colonial
government abroad before taking up his first administration as Governor of
the Philippines. On his staff was General Hugh Lennox Scott, who later
succeeded General Wood as Chief of Staff of the United States Army. The
conversations and discussions in the smokeroom each evening after dinner
were illuminating and fascinating. General Wood had but recently completed
his work as Governor of Cuba, and he talked freely of his experiences there,
while General Scott was full of reminiscences of his extraordinary adventures
with the Indians. He later played an important part in bringing peace to the
Philippines.
 It was at one of these four-cornered sessions in the smokeroom that we
first learned of Fletcher’s ambition to revolutionize the world in its methods of
eating. That he would actually accomplish this no one of us believed, but the
fact remains. The smokeroom steward was serving the coffee, inquiring of
each one how many lumps of sugar he required. Fletcher, to our amazement,
called for five! It was a grand-stand play in a way, but he secured his audience
as completely as do the tambourines and the singing of the Salvation Army.
“Why are you surprised?” he demanded with seeming innocence. “I am
simply taking a coffee liqueur, in which there is less sugar now than there is in
your chartreuse or benedictine. But I am mixing it with the saliva, which is
more than you are doing. The sugar, as you take it, becomes acid in the
stomach and retards digestion; by my method, it is changed into grape sugar,
which is easily assimilated.”
“To insalivate one’s liquor,” he explained to us, “gives one the most
exquisite pleasure imaginable, but it is a terrific test of quality. It brings out the
richness of flavor, which is lost when one gulps the wine down. Did you ever
notice the way a tea-taster sips his tea?”
As he talked he exposed the ignorance of the entire group on physiological
matters to an embarrassing extent, clinching his remarks by asking General
Wood the question,
“Would you engage as chauffeur for your automobile a man who knew as
little about his motor as you know about your own human engine?”
No one ever loved a practical joke better than Horace Fletcher. I was a
guest at a dinner he once gave at the Graduates’ Club in New Haven. Among
the others present were President Hadley of Yale, John Hays Hammond,
Walter Camp, and Professor Lounsbury. There was considerable curiosity and
some speculation concerning what would constitute a Fletcher dinner. At the
proper time we were shown into a private room, where the table was set with
the severest simplicity. Instead of china, white crockery was used, and the chief
table decorations were three large crockery pitchers filled with ice water. At
each plate was a crockery saucer, containing a shredded-wheat biscuit. It was
amusing to glance around and note the expressions of dismay upon the faces
of the guests. Their worst apprehensions were being confirmed! Just as we
were well seated, the headwaiter came to the door and announced that by
mistake we had been shown into the wrong room, whereupon Fletcher, with
an inimitable twinkle in his eye, led the way into another private dining room,
where we sat down to one of the most sumptuous repasts I have ever enjoyed.
Today, twenty years after his campaign, it is almost forgotten that the
American breakfast was at that time a heavy meal. Horace Fletcher
revolutionized the practice of eating, and interjected the word fletcherize into
the English language. As a disciple of Fletcher Sir Thomas Barlow, physician-
in-chief to King Edward VII, persuaded royalty to set the style by cutting
down the formal dinner from three hours to an hour and a half, with a
corresponding relief to the digestive apparatus of the guests. In Belgium,
during the World War, working with Herbert Hoover, Fletcher taught the
impoverished people how to sustain themselves upon meager rations. Among
his admirers and devoted friends were such profound thinkers as William
James who, in response to a letter from him, wrote, “Your excessive reaction
to the stimulus of my grateful approval makes you remind me of those rich
soils which, when you tickle them with a straw, smile with a harvest”; and
Henry James, who closes a letter: “Come and bring with you plenary
absolution to the thankless subject who yet dares light the lamp of gratitude to
you at each day’s end of his life.”

My acquaintance with Henry James came through my close association

with the late Sir Sidney Lee, the Shakesperian authority, and Horace Fletcher.
“Don’t be surprised if he is brusque or uncivil,” Sir Sidney whispered to
me just before I met him at dinner; “one can never tell how he is going to act.”
As a matter of fact, I found Henry James a most genial and enjoyable
dinner companion, and never, during the few later occasions when I had the
pleasure of being with him, did he display those characteristics of ill humor
and brusqueness which have been attributed to him. It may not be generally
known that all his life—until he met Horace Fletcher—he suffered torments
from chronic indigestion, or that it was in Fletcherism that he found his first
relief. In a typically involved Jamesian letter to his brother William he writes
(February, 1909):
It is impossible save in a long talk to make you understand how the blessed
Fletcherism—so extra blessed—lulled me, charmed me, beguiled me, from the first
into the convenience of not having to drag myself out into eternal walking. One must