Faith Ellen Antonina Kolokolova

Jrg-Rdiger Sack (Eds.)

LNCS 10389

Algorithms
and Data Structures
15th International Symposium, WADS 2017
St. John's, NL, Canada, July 31 August 2, 2017
Proceedings

123
Lecture Notes in Computer Science 10389
Commenced Publication in 1973
Founding and Former Series Editors:
Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen

Editorial Board
David Hutchison
Lancaster University, Lancaster, UK
Takeo Kanade
Carnegie Mellon University, Pittsburgh, PA, USA
Josef Kittler
University of Surrey, Guildford, UK
Jon M. Kleinberg
Cornell University, Ithaca, NY, USA
Friedemann Mattern
ETH Zrich, Zrich, Switzerland
John C. Mitchell
Stanford University, Stanford, CA, USA
Moni Naor
Weizmann Institute of Science, Rehovot, Israel
C. Pandu Rangan
Indian Institute of Technology, Madras, India
Bernhard Steffen
TU Dortmund University, Dortmund, Germany
Demetri Terzopoulos
University of California, Los Angeles, CA, USA
Doug Tygar
University of California, Berkeley, CA, USA
Gerhard Weikum
Max Planck Institute for Informatics, Saarbrcken, Germany
More information about this series at http://www.springer.com/series/7407
Faith Ellen Antonina Kolokolova

Jrg-Rdiger Sack (Eds.)

Algorithms
and Data Structures
15th International Symposium, WADS 2017
St. Johns, NL, Canada, July 31 August 2, 2017
Proceedings

123
Editors
Faith Ellen Jrg-Rdiger Sack
University of Toronto Carleton University
Toronto, ON Ottawa, ON
Canada Canada
Antonina Kolokolova
Memorial University of Newfoundland
St. Johns, NL
Canada

ISSN 0302-9743 ISSN 1611-3349 (electronic)

Lecture Notes in Computer Science
ISBN 978-3-319-62126-5 ISBN 978-3-319-62127-2 (eBook)
DOI 10.1007/978-3-319-62127-2

Library of Congress Control Number: 2017945725

LNCS Sublibrary: SL1 Theoretical Computer Science and General Issues

Springer International Publishing AG 2017

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the
material is concerned, specically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microlms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specic statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, express or implied, with respect to the material contained herein or for any errors or
omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in
published maps and institutional afliations.

Printed on acid-free paper

This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
 Preface

This volume contains the papers presented at the 15th International Algorithms and
Data Structures Symposium (WADS 2017), which was held from July 31 to August 2,
2017, in St. Johns, Newfoundland, Canada. WADS, which alternates with the Scan-
dinavian Symposium and Workshops on Algorithm Theory, SWAT, is a forum for
researchers in the area of design and analysis of algorithms and data structures.
In response to the call for papers, 109 papers were submitted. From these sub-
missions, the Program Committee selected 49 papers for presentation at WADS 2017,
using a combination of online discussion in EasyChair and a one-day video conference.
In addition, invited lectures were given by Pankaj Agarwal (Duke University), Michael
Saks (Rutgers University), and Virginia Vassilevska Williams (MIT).
Special issues of papers selected from WADS 2017 are planned for two journals,
Algorithmica and Computational Geometry: Theory and Applications.
We gratefully acknowledge the support of the WADS 2017 sponsors: Memorial
University of Newfoundland, The Fields Institute for Research in Mathematical
Sciences, Elsevier, and Springer.

July 2017 Faith Ellen

Antonina Kolokolova
Jrg-Rdiger Sack
 Organization

Conference Chair and Local Arrangements Chair

Antonina Kolokolova Memorial University of Newfoundland, Canada

Program Committee Co-chairs

Faith Ellen University of Toronto, Canada
Antonina Kolokolova Memorial University of Newfoundland, Canada
Jrg-Rdiger Sack Carleton University, Canada

Steering Committee
Frank Dehne Carleton University, Canada
Faith Ellen University of Toronto, Canada
Ian Munro University of Waterloo, Canada
Jrg-Rdiger Sack Carleton University, Canada
Roberto Tamassia Brown University, USA

Program Committee
Ittai Abraham VMware Research, Israel
Hee-Kap Ahn Pohang University of Science and Technology, Korea
Nina Amenta University of California, Davis, USA
Sayan Bhattacharya University of Warwick, UK
Therese Biedl University of Waterloo, Canada
Joan Boyar University of Southern Denmark, Denmark
Martin Dietzfelbinger Technische Universitt Ilmenau, Germany
Stephane Durocher University of Manitoba, Canada
Funda Ergun Indiana University, USA
Martn Farach-Colton Rutgers University, USA
Fedor Fomin University of Bergen, Norway
Travis Gagie Diego Portales University, Chile
Thore Husfeld IT University of Copenhagen, Denmark
and Lund University, Sweden
Michael Kerber Graz University of Technology, Austria
David Kirkpatrick University of British Columbia, Canada
Ramesh Krishnamurti Simon Fraser University, Canada
Kasper Green Larsen Aarhus University, Denmark
Vangelis Markakis Athens University of Economics and Business, Greece
Aleksandar Nikolov University of Toronto, Canada
Naomi Nishimura University of Waterloo, Canada
VIII Organization

Richard Peng Georgia Institute of Technology, USA

Seth Pettie University of Michigan, USA
Liam Roditty Bar-Ilan University, Israel
Sandeep Sen Indian Institute of Technology, Delhi, India
Monique Teillaud Inria Nancy-Grand Est, France
Takeshi Tokuyama Tohoku University, Japan
Virginia Vassilevska Massachusetts Institute of Technology, USA
Williams
Todd Wareham Memorial University of Newfoundland, Canada
Qin Zhang Indiana University Bloomington, USA

Additional Reviewers

Mohammad Ali Abam Olivier Devillers

Josh Alman Michael Dinitz
Spyros Angelopoulos Yago Diez Donoso
Esther Arkin Feodor Dragan
Martin Aumller Ran Duan
Franz Aurenhammer Charilaos Efthymiou
Mauricio Ayala-Rincon Eduard Eiben
Erfan Sadeqi Azer Matthias Englert
Sang Won Bae David Eppstein
Aritra Banik Leah Epstein
Gill Barequet William Evans
Djamal Belazzougui Rolf Fagerberg
Robert Benkoczi Lene Favrholdt
Binay Bhattacharya Andreas Emil Feldmann
Arijit Bishnu Yuval Filmus
Hans L. Bodlaender Johannes Fischer
Piotr Borowiecki Kyle Fox
Jean Cardinal Mathew Francis
Paz Carmi Jie Gao
Timothy M. Chan Daya Gaur
Steven Chaplick Ellen Gethner
Jiecao Chen Arijit Ghosh
Man Kwun Chiu Archontia Giannopoulou
Keerti Choudhary Petr Golovach
Nachshon Cohen Carsten Grimm
Vincent Cohen-Adda Allan Grnlund
Alexander Conway Martin Gro
Sren Dahlgaard Joachim Gudmundsson
Peter Damaschke Manoj Gupta
Syamantak Das Shalmoli Gupta
Mark de Berg Sariel Har-Peled
Jean-Lou De Carufel Meng He
Organization IX

Martin Hoefer Debajyoti Mondal

Chien-Chung Huang Gila Morgenstern
Thomas Hull Miguel A. Mosteiro
Ehsan Iranmanesh Wolfgang Mulzer
Mark Jerrum Ian Munro
Hossein Jowhari Nabil Mustafa
Stasys Jukna Petra Mutzel
Mark Keil Jesper Nederlof
Kamyar Khodamoradi Stefan Neumann
Elena Khramtcova Huy L. Nguyen
Philipp Kindermann Patrick Nicholson
Peter Kling Jesper Sindahl Nielsen
Chip Klostermeyer Arnur Nigmetov
Christian Knauer Nicolas Nisse
Mathias Bk Tejs Knudsen Jakob Nordstrm
Janne H. Korhonen Eunjin Oh
Matias Korman Yota Otachi
Daniel Kral John Owens
Stefan Kratsch Jeff M. Phillips
Ravishankar Krishnaswamy Rameshwar Pratap
Janardhan Kulkarni Sharath Raghvendra
Amit Kumar Benjamin Raichel
Woden Kusner Rajiv Raman
Jakub acki Mathias Rav
Kim Skak Larsen Saurabh Ray
Samuli Leppnen Jean-Florent Raymond
Jian Li Andr van Renssen
Pak Ching Li Marcel Roeloffzen
Yi Li Luis M.S. Russo
Bingkai Lin Yogish Sabharwal
Daniel Lokshtanov Sushant Sachdeva
Aaron Lowe Vera Sacristan
Anna Lubiw Toshiki Saitoh
Bin Ma Pawe Schmidt
Pasin Manurangsi Hannah Schreiber
Euripides Markou Oded Schwartz
Tyler Mayer Roy Schwartz
Saeed Mehrabi Akiyoshi Shioura
Moti Medina Rakesh Sinha
Victor Milenkovic Michiel Smid
Tillmann Miltzow Shay Solomon
Gopinath Misra Bettina Speckmann
Joseph Mitchell Abhinav Srivastav
Dieter Mitsche Frank Staals
Valia Mitsou Thomas Steinke
Reza Modarres Torstein Strmme
X Organization

Jan Arne Telle Josh Wang

Sharma V. Thankachan Sebastian Wild
Csaba Toth Ryan Williams
Meng-Tsung Tsai Steve Wismath
Charalampos Tsourakakis Prudence Wong
Ryuhei Uehara Christian Wulff-Nilsen
Seeun William Umboh Jie Xue
Birgit Vogtenhuber Sang Duk Yoon
Erik Waingarten Hamid Zarrabi-Zadeh
Di Wang Haoyu Zhang
Haitao Wang Yuan Zhou
Junxing Wang
Abstracts of Invited Lectures
 Algorithms for Geometric Similarity:
Recent Developments

Pankaj K. Agarwal

Department of Computer Science, Duke University, Durham, USA

[email protected]

Abstract. A basic problem in classifying, or searching for similar objects, in a

large set of geometric objects is computing similarity between two objects.
There has been extensive work on computing geometric similarity between two
objects. In many applications, it is not sufcient to return a single similarity
score. Instead, a map between two objects that identies shared structures is
needed.
This talk discusses some recent work on computing maps between two or
more objects. The talk consists of three parts. The rst part focuses on com-
puting maps between two weighted point sets, say, distributions. The second
part is devoted to computing maps between a pair of trajectories. The third part
will briefly discuss computing Gromov-Hausdorff distance between two metric
spaces.

This work is supported in part by NSF under grants CCF-15-13816, CCF-15-46392, and IIS-14-08846,
by ARO grant W911NF-15-1-0408, and by grant 2012/229 from the U.S.-Israel Binational Science
Foundation.
 How efciently can easy dynamic programs be
approximated?

Michael Saks

Department of Mathematics, Rutgers University, New Brunswick, USA

[email protected]

Abstract. In many of the simplest examples of dynamic programming, inputs of

size n are processed by constructing an n
n matrix, where each entry is
obtained by a simple function of a few entries above and to the left. This yields a
simple On2 algorithm for such problems. These algorithms naturally arise, for
example, in evaluating various distance measures between two strings, such as
LCS (longest common subsequence) distance, Edit Distance, Frechet Distance,
and Dynamic Time Warping Distance, and the i; j entry of the matrix gives the
desired measure between the length i prex of the rst string, and the length
j prex of the second. With few exceptions (such as the Longest Increasing
Subsequence (LIS) problem where the quadratic time algorithm has been
improved to On logn), these quadratic time dynamic programming algo-
rithms remain essentially the fastest exact algorithms (except for no1 factor
improvements). This phenomenon has been the focus of much recent research in
ne grain complexity, and it has been shown that for many such problems,
reducing the running time to On2e would contradict the Strong Exponential
Time Hypothesis (e.g., Bringmann [7], Abboud, Backurs and Williams [2],
Backurs and Indyk [6] and Bringmann and Kunnermann [8].)
If we are willing to accept a good approximation (rather than the exact
answer), then there is much less evidence that quadratic complexity is needed.
Bringmann [7] proved that the Strong Exponential Hypothesis implies that truly
subquadratic algorithms cannot achieve approximation factors arbitrarily close
to 1. Abboud and Backurs [1] provided complexity theoretic evidence that truly
subquadratic deterministic algorithms cannot achieve approximation factors
arbitrarily close to 1 for edit distance and LCS-distance.
If we allow randomized algorithms, it is quite possible that problems such as
edit distance and LCS distance have constant factor approximation algorithms
that are signicantly faster than quadratic. Andoni, Krauthgamer and Onak [4]
gave a nearly linear time algorithm that achieves a polylogarithmic approxi-
mation to edit distance. For certain special cases of LCS-distance, arbitrarily
good additive en error approximation algorithms are known that are substan-
tially faster than the best exact algorithms. For the LIS Problem, Saks and
Seshadhri [11] (following the work of Ailon, Chazelle, Comandur and Liu [3]
and Parnas, Ron and Rubinfeld [10]) developed such an additive approximation

Supported by Simons Foundation Award 332622.

How efciently can easy dynamic programs be approximated? XV

whose running time is only polylogarithmic in the length of the input. For the
special case of LCS-distance between two permutations of f1; . . .; ng (Ulam
Distance), Naumovitz, Saks and Seshadhri [9] (following earlier work of
Andoni and Nguyen [5]) obtained such an additive approximation running in
~ pn.
time O

References
1. Abboud, A., Backurs, A.: Towards hardness of approximation for polynomial time prob-
lems. In: Conference on Innovations in Theoretical Computer Science (ITCS) (2017)
2. Abboud, A., Backurs, A., Williams, V.V.: Quadratic-Time Hardness of LCS and other
Sequence Similarity Measures. CoRR, abs/1501.07053 (2015)
3. Ailon, N., Chazelle, B., Comandur, S., Liu, D.: Estimating the distance to a monotone
function. Random Struct. Algorithms 31, 371383 (2017)
4. Andoni, A., Krauthgamer, R., Onak, K.: Polylogarithmic approximation for edit distance and
the asymmetric query complexity. In: IEEE Symposium on Foundations of Computer
Science (FOCS), pp. 377386 (2010)
5. Andoni, A., Nguyen, H.L.: Near-optimal sublinear time algorithms for Ulam distance. In:
Proceedings of the 21st Symposium on Discrete Algorithms (SODA), pp. 7686 (2010)
6. Backurs, A., Indyk, P.: Edit distance cannot be computed in strongly subquadratic time
(unless SETH is false). In: Proceedings of the Forty-Seventh Annual ACM on Symposium
on Theory of Computing (STOC), pp. 5158 (2015)
7. Bringmann, K.: Why walking the dog takes time: Frechet distance has no strongly sub-
quadratic algorithms unless SETH fails. CoRR, abs/1404.1448 (2014)
8. Bringmann, K., Kunnemann, M.: Quadratic conditional lower bounds for string problems
and dynamic time warping. In: Proceedings of the 56th Annual IEEE Symposium on
Foundations of Computer Science (FOCS), pp. 7997 (2015)
9. Naumovitz, T., Saks, M., Seshadhri, C.: Accurate and nearly optimal sublinear approxi-
mations to Ulam distance. In: SIAM-ACM Symposium on Discrete Algorithms, pp. 2012
2031 (2017)
10. Parnas, M., Ron, D., Rubinfeld, R.: Tolerant property testing and distance approximation.
J. Comput. Syst. Sci. 6, 10121042 (2006)
11. Saks, M., Seshadhri, C.: Estimating the longest increasing sequence in polylogarithmic time.
SIAM J. Comput. 46(2), 774823 (2017)
 Fine-Grained Complexity of Problems in P

Virginia Vassilevska Williams

Computer Science and Articial Intelligence Laboratory,

Massachusetts Institute of Technology, Cambridge, USA
[email protected]

Abstract. A central goal of algorithmic research is to determine how fast

computational problems can be solved in the worst case. Theorems from
complexity theory state that there are problems that, on inputs of size n, can be
solved in tn time but not in tn1e time for e [ 0. The main challenge is to
determine where in this hierarchy various natural and important problems lie.
Throughout the years, many ingenious algorithmic techniques have been
developed and applied to obtain blazingly fast algorithms for many problems.
Nevertheless, for many other central problems, the best known running times are
essentially those of their classical algorithms from the 1950s and 1960s.
Unconditional lower bounds seem very difcult to obtain, and so practically
all known time lower bounds are conditional. For years, the main tool for
proving hardness of computational problems have been NP-hardness reductions,
basing hardness on P 6 NP. However, when we care about the exact running
time (as opposed to merely polynomial vs non-polynomial), NP-hardness is not
applicable, especially if the problem is already solvable in polynomial time. In
recent years, a new theory has been developed, based on ne-grained reduc-
tions that focus on exact running times. Mimicking NP-hardness, the approach
is to (1) select a key problem X that is conjectured to require essentially tn
time for some t, and (2) reduce X in a ne-grained way to many important
problems. This approach has led to the discovery of many meaningful rela-
tionships between problems, and even sometimes to equivalence classes.
The main key problems used to base hardness on have been: the 3SUM
problem, the CNF-SAT problem (based on the Strong Exponential Time
Hypothesis (SETH)) and the All Pairs Shortest Paths Problem. Research on
SETH-based lower bounds has flourished in particular in recent years showing
that the classical algorithms are optimal for problems such as Approximate
Diameter, Edit Distance, Frechet Distance, Longest Common Subsequence,
many dynamic graph problems, etc.
In this talk I will give an overview of the current progress in this area of
study, and will highlight some exciting new developments.

Supported by an NSF CAREER Award, NSF Grants CCF-1417238, CCF-1528078 and CCF-1514339,
and BSF Grant BSF:2012338.
 Contents

Covering Segments with Unit Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Ankush Acharyya, Subhas C. Nandy, Supantha Pandit, and Sasanka Roy

Replica Placement on Bounded Treewidth Graphs. . . . . . . . . . . . . . . . . . . . 13

Anshul Aggarwal, Venkatesan T. Chakaravarthy, Neelima Gupta,
Yogish Sabharwal, Sachin Sharma, and Sonika Thakral

Fast Exact Algorithms for Survivable Network Design

with Uniform Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Akanksha Agrawal, Pranabendu Misra, Fahad Panolan,
and Saket Saurabh

The Complexity of Tree Partitioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

Zhao An, Qilong Feng, Iyad Kanj, and Ge Xia

Effectiveness of Local Search for Art Gallery Problems . . . . . . . . . . . . . . . . 49

Sayan Bandyapadhyay and Aniket Basu Roy

Parameterized Complexity of Geometric Covering Problems

Having Conflicts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
Aritra Banik, Fahad Panolan, Venkatesh Raman, Vibha Sahlot,
and Saket Saurabh

Obedient Plane Drawings for Disk Intersection Graphs . . . . . . . . . . . . . . . . 73

Bahareh Banyassady, Michael Hoffmann, Boris Klemz,
Maarten Lffler, and Tillmann Miltzow

d-Greedy t-spanner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
Gali Bar-On and Paz Carmi

Dynamic Graph Coloring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

Luis Barba, Jean Cardinal, Matias Korman, Stefan Langerman,
Andr van Renssen, Marcel Roeloffzen, and Sander Verdonschot

Universal Hinge Patterns for Folding Strips Efficiently

into Any Grid Polyhedron . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Nadia M. Benbernou, Erik D. Demaine, Martin L. Demaine,
and Anna Lubiw

An Optimal XP Algorithm for Hamiltonian Cycle on Graphs

of Bounded Clique-width . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
Benjamin Bergougnoux, Mamadou Moustapha Kant,
and O-Joung Kwon
XVIII Contents

Improved Algorithms for Computing k-Sink on Dynamic

Flow Path Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Binay Bhattacharya, Mordecai J. Golin, Yuya Higashikawa,
Tsunehiko Kameda, and Naoki Katoh

A 2-Approximation for the Height of Maximal Outerplanar

Graph Drawings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
Therese Biedl and Philippe Demontigny

Splitting B2 -VPG Graphs into Outer-string and Co-comparability Graphs . . . . 157

Therese Biedl and Martin Derka

A Deterministic Algorithm for Online Steiner Tree Leasing . . . . . . . . . . . . . 169

Marcin Bienkowski, Artur Kraska, and Pawe Schmidt

The I/O Complexity of Strassens Matrix Multiplication

with Recomputation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
Gianfranco Bilardi and Lorenzo De Stefani

Maximum Plane Trees in Multipartite Geometric Graphs . . . . . . . . . . . . . . . 193

Ahmad Biniaz, Prosenjit Bose, Kimberly Crosbie, Jean-Lou De Carufel,
David Eppstein, Anil Maheshwari, and Michiel Smid

Local Routing in Spanners Based on WSPDs . . . . . . . . . . . . . . . . . . . . . . . 205

Prosenjit Bose, Jean-Lou De Carufel, Vida Dujmovi,
and Frdrik Paradis

Relaxing the Irrevocability Requirement for Online Graph Algorithms . . . . . . 217

Joan Boyar, Lene M. Favrholdt, Michal Kotrbk, and Kim S. Larsen

Approximating Small Balanced Vertex Separators in Almost Linear Time . . . 229

Sebastian Brandt and Roger Wattenhofer

Balanced Line Separators of Unit Disk Graphs . . . . . . . . . . . . . . . . . . . . . . 241

Paz Carmi, Man Kwun Chiu, Matthew J. Katz, Matias Korman,
Yoshio Okamoto, Andr van Renssen, Marcel Roeloffzen,
Taichi Shiitada, and Shakhar Smorodinsky

All-Pairs Shortest Paths in Geometric Intersection Graphs . . . . . . . . . . . . . . 253

Timothy M. Chan and Dimitrios Skrepetos

The Complexity of Drawing Graphs on Few Lines and Few Planes. . . . . . . . 265
Steven Chaplick, Krzysztof Fleszar, Fabian Lipp, Alexander Ravsky,
Oleg Verbitsky, and Alexander Wolff

Modular Circulation and Applications to Traffic Management. . . . . . . . . . . . 277

Philip Dasler and David M. Mount
Contents XIX

The Homogeneous Broadcast Problem in Narrow and Wide Strips . . . . . . . . 289

Mark de Berg, Hans L. Bodlaender, and Sndor Kisfaludi-Bak

Minimizing the Continuous Diameter when Augmenting

a Tree with a Shortcut . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
Jean-Lou De Carufel, Carsten Grimm, Stefan Schirra, and Michiel Smid

Inapproximability of the Standard Pebble Game

and Hard to Pebble Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
Erik D. Demaine and Quanquan C. Liu

Capacitated Center Problems with Two-Sided Bounds and Outliers . . . . . . . . 325

Hu Ding, Lunjia Hu, Lingxiao Huang, and Jian Li

Faster Randomized Worst-Case Update Time for Dynamic Subgraph

Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
Ran Duan and Le Zhang

Improved Distance Sensitivity Oracles via Tree Partitioning . . . . . . . . . . . . . 349

Ran Duan and Tianyi Zhang

Delta-Fast Tries: Local Searches in Bounded Universes

with Linear Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
Marcel Ehrhardt and Wolfgang Mulzer

Split Packing: Packing Circles into Triangles with Optimal Worst-Case

Density . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
Sndor P. Fekete, Sebastian Morr, and Christian Scheffer

Fast and Compact Planar Embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385

Leo Ferres, Jos Fuentes, Travis Gagie, Meng He,
and Gonzalo Navarro

When Can Graph Hyperbolicity be Computed in Linear Time?. . . . . . . . . . . 397

Till Fluschnik, Christian Komusiewicz, George B. Mertzios,
Andr Nichterlein, Rolf Niedermeier, and Nimrod Talmon

Optimal Query Time for Encoding Range Majority . . . . . . . . . . . . . . . . . . . 409

Pawe Gawrychowski and Patrick K. Nicholson

Conditional Lower Bounds for Space/Time Tradeoffs . . . . . . . . . . . . . . . . . 421

Isaac Goldstein, Tsvi Kopelowitz, Moshe Lewenstein, and Ely Porat

Posimodular Function Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437

Magns M. Halldrsson, Toshimasa Ishii, Kazuhisa Makino,
and Kenjiro Takazawa
XX Contents

How to Play Hot and Cold on a Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449

Herman Haverkort, David Kbel, Elmar Langetepe,
and Barbara Schwarzwald

Faster Algorithm for Truth Discovery via Range Cover . . . . . . . . . . . . . . . . 461

Ziyun Huang, Hu Ding, and Jinhui Xu

Searching Edges in the Overlap of Two Plane Graphs . . . . . . . . . . . . . . . . . 473

John Iacono, Elena Khramtcova, and Stefan Langerman

Improved Average Complexity for Comparison-Based Sorting . . . . . . . . . . . 485

Kazuo Iwama and Junichi Teruyama

An EPTAS for Scheduling on Unrelated Machines

of Few Different Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 497
Klaus Jansen and Marten Maack

A Polynomial Kernel for Distance-Hereditary Vertex Deletion . . . . . . . . . . . 509

Eun Jung Kim and O-joung Kwon

Busy Time Scheduling on a Bounded Number of Machines

(Extended Abstract). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 521
Frederic Koehler and Samir Khuller

Algorithms for Covering Multiple Barriers . . . . . . . . . . . . . . . . . . . . . . . . . 533

Shimin Li and Haitao Wang

An Improved Algorithm for Diameter-Optimally Augmenting

Paths in a Metric Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 545
Haitao Wang

Covering Uncertain Points in a Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557

Haitao Wang and Jingru Zhang

Stochastic Closest-pair Problem and Most-likely Nearest-neighbor

Search in Tree Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 569
Jie Xue and Yuan Li

On the Expected Diameter, Width, and Complexity of a Stochastic

Convex-hull . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 581
Jie Xue, Yuan Li, and Ravi Janardan

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 593

 Covering Segments with Unit Squares

Ankush Acharyya, Subhas C. Nandy, Supantha Pandit, and Sasanka Roy

Indian Statistical Institute, Kolkata, India

Abstract. We study several variations of line segment covering problem

with axis-parallel unit squares in the plane. Given a set S of n line seg-
ments, the objective is to nd the minimum number of axis-parallel unit
squares which cover at least one end-point of each segment. The varia-
tions depend on the orientation and length of the input segments. We
prove some of these problems to be NP-complete, and give constant factor
approximation algorithms for those problems. For the general version of
the problem, where the segments are of arbitrary length and orientation,
and the squares are given as input, we propose a factor 16 approxima-
tion result based on multilevel linear programming relaxation technique.
This technique may be of independent interest for solving some other
problems. We also show that our problems have connections with the
problems studied by Arkin et al. [2] on conict-free covering problem.
Our NP-completeness results hold for more simplied types of objects
than those of Arkin et al. [2].

Keywords: Segment cover, unit square, NP-hardness, linear program-

ming, approximation algorithms, PTAS.

1 Introduction

In this paper, we study dierent interesting variations of line segment covering

problem. Here, a set S = {s1 , s2 , . . . , sn } of line segments in IR2 is given. An
axis-parallel square t is said to cover a line segment s S if t contains at least
one end-point of s. We deal with two classes of covering problem: i) continuous,
and ii) discrete.
Continuous Covering Segments by Unit Squares (CCSUS ):
Given a set S of segments in the plane, the goal is to nd a smallest set T of
unit squares which covers all the segments in S.
Discrete Covering Segments by Unit Squares (DCSUS ): Given
a set S of segments and a set T of unit squares in the plane, the goal is to nd
a subset T  T of minimum cardinality which can cover all the segments in S.
The motivation of studying this problem comes from an applications to net-
work security [9]. Here, a set of physical devices is deployed over a geographical
area. These devices communicate with each other through physical links. The
objective is to check the security of the network by placing minimum number
of devices which can sense at least one end point of every link. This problem

Springer International Publishing AG 2017 1

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 112, 2017.
DOI: 10.1007/978-3-319-62127-2_1
2 A. Acharyya et al.

can be modelled as line segment covering problem, where the links can be
interpreted as straight-line segments and the objects can be interpreted as unit
squares. In [9], several other applications are also stated.
We study the following variations of covering problem for line segments which
are classied depending upon their lengths and orientations.
Continuous Covering

 CCSUS-H1-US: Horizontal unit segments inside a unit height strip.

 CCSUS-H1: Horizontal unit segments.
 CCSUS-HV1: Horizontal and vertical unit segments.
 CCSUS-ARB: Segments with arbitrary length and orientation.

Discrete Covering

 DCSUS-ARB: Segments with arbitrary length and orientation.

We dene some terminologies and denitions used in this paper. We use seg-
ment to denote a line segment, and unit square to denote an axis-parallel unit
square. For a given non-vertical segment s, we dene l(s) and r(s) to be its left
and right end-points. For a vertical segment s, l(s) and r(s) are dened to be the
end-points of s with highest and lowest y-coordinates respectively. The center
of a square t is the point of intersection of its two diagonals. We use t(a, b) to
denote a square whose center is at the point a and whose side length is b.

Denition 1. Two segments in S are said to be independent if no unit square

can cover both the segments. A subset S  S is an independent set if every
pair of segments in S  is independent. A subset S  S of segments is a maximal
independent set if for any s S \ S  , S  {s} is not an independent set.

Known results: Arkin et al. [2] studied a related problem the conict-free
covering . Given a set P of n color classes, where each color class contains exactly
two points, the goal is to nd a set of conict-free objects of minimum cardinality
which covers at least one point from each color class. An object is said to be
conict-free if it contains at most one point from each color class. Arkin et al.
[1, 2] showed that, both discrete and continuous versions of conict-free covering
problem are NP-complete where the points are on a real line and objects are
intervals of arbitrary length. These results are also valid for covering arbitrary
length segments on a line with unit intervals. They provided 2- and 4-factor
approximation algorithms for the continuous and discrete versions of conict-
free covering problem with arbitrary length intervals respectively. If the points
of the same color class are either vertically or horizontally unit separated, then
they proved that the continuous version of the conict-free covering problem with
axis-parallel unit squares is NP-complete, and proposed a factor 6 approximation
algorithm. Finally, they remarked the existence of a polynomial time dynamic
programming based algorithm for the continuous version of conict-free covering
problem with unit intervals where the points are on a real line and each pair of
Covering Segments with Unit Squares 3

same color points is unit separated. Recently, Kobylkin [9] studied the problem
of covering the edges of a given straight line embedding of a planar graph by
minimum number of unit disks, where an edge is said to be covered by a disk if
any point on that edge lies inside that disk. He proved NP-completeness results
for some special graphs. A similar study is made in [10], where a set of line
segments is given, the objective is to cover these segments with minimum number
of unit disks, where the covering of a segment by a disk is dened as in [9]. For
continuous version of the problem, they proposed a PTAS where the segments
are non-intersecting. For the discrete version, they showed that the problem is
APX-hard.
Our contributions: In Section 2, we show that the CCSUS-H1 problem is NP-
complete, and propose an O(n log n) time factor 3 approximation algorithm for
the CCSUS-HV1 problem. This improves the factor 6 approximation result of
Arkin et al. [2] while keeping the time complexity unchanged. We also provide
a PTAS for CCSUS-HV1 problem. For the CCSUS-ARB problem, we give an
O(n log n) time factor 6 approximation algorithm. In Section 3, give a polynomial
time factor 16 approximation algorithm for the DCSUS-ARB problem.

2 Continuous covering

In this version, the segments are given, and the objective is to place minimum
number of unit squares for covering at least one end-point of all the segments.

2.1 CCSUS-H1-US problem

Let S be a set of n horizontal unit segments inside a unit height horizontal strip.
Start with an empty set T  . Sort the segments in S from left to right with respect
to their right end-points. Repeat the following steps until all segments of S are
processed. In each step, select the left-most segment s among the uncovered
segments in S. Add a unit square t T  which is inside the strip aligning its
left boundary at r(s). Mark all the segments that are covered by t as processed.
Finally, return T  as the output. Thus, we have an O(n log n) time algorithm for
the CCSUS-H1-US problem.

2.2 CCSUS-H1 problem

We prove that CCSUS-H1 is NP-complete by using a reduction from the recti-

linear version of planar 3 SAT (RPSAT(3)) problem [8]. We also propose an
O(n log n)-time 2-approximation algorithm for this problem.
RPSAT(3) [8]: Given a 3 SAT instance with n variables and m clauses,
where the variables are positioned on a horizontal line and each clause containing
3 literals is formed with three vertical and one horizontal line segment. Each
clause is connected with its three variables either from above or from below such
4 A. Acharyya et al.

that no two line segments corresponding to two dierent clauses intersect. The
objective is to nd a satisfying assignment of . See Figure 1(a) for an instance
of RPSAT(3) problem. Here the solid (resp. dotted) vertical segment attached to
the horizontal line of a clause represents that the corresponding variable appears
as a positive (resp. negative) literal in that clause.

(a) (b)

(c)

Fig. 1. (a) RPSAT(3) representation. (b) Connection of a cycle and a chain. (c) Cycle
gadget for a variable xi .

We now describe the construction of an instance I of CCSUS-H1 problem from

an instance of RPSAT(3) problem. Let {x1 , x2 , . . . , xn } be n variables and
{C1 , C2 , . . . , Cm } be m clauses of . We describe the construction for the clauses
connecting to the variables from above. A similar construction can be done for
the clauses connecting to the variables from below.
Let d be the maximum number of vertical segments connected to every variable
from dierent clauses either from above or from below. Let, = 4d + 3. Each
variable gadget for xi may consist of a cycle and at most 2d chains. The cycle
consists of 2 unit horizontal segments {si1 , si2 , . . . , si2 } in two sides of a (dotted)
horizontal line (see Figure 1(c)). The segments {si1 , si2 , . . . , si } are above the
horizontal line and the segments {si+1 , si+2 , . . . , si2 } are below the horizontal
line. The chains correspond to the vertical segments connecting a variable xi
with the clauses containing it. There are three types of chains: (i) , (ii)
, and (iii) (see Figure 1(a)). The gadget corresponding to three types
of chains are shown in Figures 2(a), 2(b), and 2(c) respectively. The chains are
connected to the cycle, and together it forms a chain of big-cycle (see Figure
1(b)). It needs to be mentioned that the number of segments is not xed for every
chain, even for similar chains of dierent clauses. Note that, at the joining point
(to construct a big-cycle) we slightly perturb two unit segments little upward.
Covering Segments with Unit Squares 5

(a) (b)

(c) (d)

Fig. 2. Gadget for (a) type (i) chain, (b) type (ii) chain (c) type (iii) chain, and (d)
Demonstration of clause-segment s corresponding to the clause C = (xi xj xk );
shaded portions in (a), (b), (c) represent the connection of s with the variables in C .

Let 0, 1, 2, . . . , ( d) be the left to right order of the vertical segments

corresponding to the clauses which are connected to the gadget corresponding to
the variable xi . Consider the -th clause C in this order. If xi is a positive literal,
then the segments si3+4 and si3+4+1 are perturbed (moved upward as shown
using upward arrow in Figures 2(a), 2(b), 2(c)) to connect the corresponding
chain of C with the cycle of variable xi . Otherwise, if xi is a negative literal,
then the segments si3+4+1 and si3+4+2 are perturbed.
Note that, the squares are not given as a part of the input. In Figures 1(c), 2(a),
2(b), and 2(c) a possible set of unit squares are also depicted. Each square can
cover exactly two segments. Therefore, we have the following observation:

Observation 1 Exactly half of the squares (either all empty or all shaded) can
cover all the segments in the big-cycle corresponding to the variable xi . This
solution represents the truth value (empty for true and shaded for false) of the
corresponding variable xi .

Further, for the clause C , we take a single unit horizontal segment s that
connects the chain corresponding to three variables. This is referred to as a
clause-segment. The placement of s is shown in Figure 2(d). Note that, in
order to maintain the alternating empty and shaded vertical layers in a variable
6 A. Acharyya et al.

gadget we may need to reduce the distance between two consecutive vertical
layers of squares. But, the segments are placed suciently apart so that no
unit square can cover more than two segments from a variable gadget. As the
number of segments (Q) considering all variable gadgets, is even, we need exactly
Q
2 squares to cover them. Now, if a clause C is satisable then at least one square
connected to s will be chosen, and hence s will be covered; if C is not satisable
then the square adjacent to s of each variable chain will not be chosen in the
solution, and hence we need one more square to cover s (see Figure 2(d)). Thus,
we have the following result.

Lemma 1. The given RPSAT(3) formula is satisable if the number N of

squares needed to cover all the unit segments in the construction is exactly
N0 = 12 ( i=1 Qi ), where Qi is the number of squares in the big-cycle correspond-
n

ing to the gadget of variable xi . If the formula is not satisable then N > N0 ,

Clearly, (i) CCSUS-H1 problem is in NP, and (ii) number of squares and seg-
ments used in the reduction is polynomial in the the size of RPSAT(3) problem.

Theorem 1. CCSUS-H1 is NP-complete.

Approximation algorithm Let S be a set of n unit horizontal segments on

the plane. We rst partition the entire region into a set of  disjoint unit height
horizontal strips H1 , H2 , . . . , H . Let Si S be the set of segments in the strip
Hi , for i = 1, . . . , . Clearly, (i) Si Sj = , for i = j, and (ii) any unit square
cannot cover two segments, one from Si and the other from Sj where j i 2,
for i = 1, . . . ,  2, and j = 3, . . . , .
We calculate minimum number of unit squares covering Si using the algorithm
in Section 2.1. Let Qi be the set of squares returned by our algorithm. Let
Qodd = {Q1 Q3 . . .} and Qeven = {Q2 Q4 . . .} be the optimum solutions
for the segments in odd and even numbered strips respectively. We have Qodd
Qeven = , and we report Q = Qodd Qeven . Let OP T be a minimum sized
set of unit squares covering S. Now, |OP T | max(|Qodd |, |Qeven |). Thus, |Q| =
|Qodd | + |Qeven | 2|OP T |. Since Si Sj = for i = j, and computing Qi needs
O(|Si | log |Si |) time, the overall running time of the algorithm is O(n log n).

Theorem 2. A 2-factor approximation solution for the CCSUS-H1 problem can

be computed in O(n log n) time.

2.3 CCSUS-HV1 problem

Here, we have both horizontal and vertical segments in S which are of unit length.
An easy way to get a factor 4 approximation algorithm for this problem is as fol-
lows. Let S = SH SV , where SH and SV are the sets of horizontal and vertical
unit segments respectively. By Theorem 2, we already have a factor 2 approxima-
tion algorithm for covering the members in SH (resp. SV ). If QH and QV be the
Covering Segments with Unit Squares 7

(a) (b)

Fig. 3. (a) Placement of unit squares for a horizontal and vertical unit segment s.

set of squares returned by our algorithm for covering SH and SV respectively,

and OP TH and OP TV be the optimum solution for SH and SV respectively,
then |QH | + |QV | 2|OP TH | + 2|OP TV | 4|OP T |, where OP T be the overall
optimum solution for SH SV (since |OP T | max(|OP TH |, |OP TV |)).
We now propose a factor 3 approximation algorithm for this problem using a
sweep-line technique. During the execution of the algorithm, we maintain a set
of segments LB such that no two members in LB can be covered by one unit
square. Elements in LB form an independent set. For each segment in S we
maintain a f lag variable; its value is 1 or 0 depending on whether it is covered
or not by the chosen set of squares corresponding to the members in LB. We
also maintain a range tree T with the end-points of the members in S. Each
element in T has a pointer to the corresponding element in S.
Let OU T P U T be a set containing the chosen squares to cover the members in
LB. Initially, we set OU T P U T = . We sort the unit segments in S from top
to bottom with respect to their r(.) values. Next, we consider the elements in S
in order. For each segment s S, if its f lag bit is 0, then (i) if s is horizontal
then we dene three unit squares {t1 , t2 , t3 } (see Figure 3(a)), and (ii) if s is
vertical then we dene two unit squares {t1 , t2 } (see Figure 3(b)). Store s in LB
and the generated squares are put in OU T P U T , and for each newly generated
squares t, we search in the range tree T to identify the members in S covered
by t. The f lag bit of s and those members in S covered by t are set to 1. The
process continues until the ag bit of all the members in S are set to 1.

Theorem 3. The proposed algorithm produces a 3-factor approximation result

for the CHSUS-HV1 problem in O(n log n) time using O(n log n) space.

Polynomial time approximation scheme We propose a PTAS for the CCSUS-

HV1 problem using the shifting strategy of Hochbaum and Maass [7]. We are
given a set S of n horizontal and vertical unit segments. Enclose the segments
inside a integer length square box B; partition B into vertical strips of width
1, and also partition B into horizontal strips of height 1. We choose a constant
k Z+ , and dene a k-strip consisting of at most k consecutive strips. Now,
we dene the concept of shifts. We have k dierent shifts in the vertical di-
rection. Each vertical shift consists of some disjoint k-strips. In the i-th shift
(i = 0, 1, . . . , k 1), the rst k-strip consists of i unit vertical strips at extreme
8 A. Acharyya et al.

left, and then onwards each k-strip is formed with k consecutive unit vertical
strips except the last k-strip which may contain lesser than k vertical strips.
Similarly, k shifts are dened in horizontal direction. Now consider shift(i, j)
as the i-th vertical shift and j-th horizontal shift. This splits the box B into
rectangular cells of size at most k k. The following observation is important
to analyze the complexity of our algorithm.

Observation 2 There exists an optimal solution that contains squares such that
two boundaries of each square is attached to an end-point(s) of some segments.
2
Lemma 2. Finding a feasible solution for each shif t(i, j) requires O(n2k ) time.

In our algorithm, for each shif t(i, j) we calculate optimal solution in each cell
and combine them to get a feasible solution. Finally, return the minimum among
these k 2 feasible solutions.
Let OP T be an optimum set of unit squares covering S, and Q be a feasible so-
lution returned by our algorithm described above. Following the proof technique
of Lemma 2.1 of [7], we can prove the following theorem.

Theorem 4. |Q| (1 + k1 )2 |OP T | and the running time of the above algorithm
2
is O(k 2 n2k ).

2.4 CCSUS-ARB problem

Here, we rst design a 8-approximation algorithm for CCSUS-ARB problem as

follows. Next, we improve the approximation factor to 6.

Observation 3 Let s1 and s2 be two segments in S. If neither of the squares

t(l(s1 ), 2) and t(r(s1 ), 2) covers s2 , then s1 and s2 are independent.

Here also we start with an empty set OU T P U T and LB, and each segment in S
is attached with a ag bit. We maintain a range tree T with the end-points of the
members in S. Each time, an arbitrary segment s S with f lag(s) = 0 is chosen,
and inserted in LB. Its ag bit is set to 1. Insert four unit squares {t1 , t2 , t3 , t4 }
which fully cover the square t(l(s), 2) and four unit squares {t1 , t2 , t3 , t4 } which
fully cover the square t(r(s), 2) in OU T P U T . Remove all the segments in S that
are covered by {t1 , t2 , t3 , t4 , t1 , t2 , t3 , t4 } by performing range searching in T as
stated in Section 2.3. The end-points of the deleted segments are also deleted
from T . This process is repeated until all the members in S are agged. Finally,
return the set OU T P U T . Observation 3 suggests the following result.

Lemma 3. The above algorithm for CCSUS-ARB problem runs in O(n log n)
time, and the produced solution is factor 8 approximation of the optimal solution.

We now improve the approximation factor to 6 using a sweep-line technique

introduced in Biniaz et al. [4]. We sort the segments in S with respect to their
Covering Segments with Unit Squares 9

left end-points1 , and process the elements in S in order. When an element s S

is processed, if f lag(s) = 0 then we put six squares, two of them covering the
1 2 rectangle t1 = right-half of t(l(s), 2) and four of them covering the 2 2
square t2 = t(r(s), 2). We also identify the segments in S that are covered by t1
and t2 , and their ag bit is set to 1. Thus, we have the following result:
Theorem 5. The above algorithm for the CCSUS-ARB problem produces a fac-
tor 6 approximation result in O(n log n) time.

3 Discrete covering: DCSUS-ARB problem

In this section, we give a 16 factor approximation algorithm for DCSUS-ARB
problem. We use the following problem to solve the DCSUS-ARB problem.
DCPUS Problem: Given a point set P in IR2 and a set T of unit squares, nd
a subset of T of minimum cardinality to cover all the points in P .
For an ILP Zw , the corresponding LP is denoted by Zw . We denote OP TwI and
OP TwF as the value of the optimal solution for Zw and Zw respectively.
Lemma 4. If there exists an LP-based factor approximation algorithm for the
DCPUS problem, then there exists a LP-based factor 2 approximation result for
the DCSUS-ARB problem.

Proof. Let T1 T (resp. T2 T ) be the set of all squares which cover the left
(resp. right) end-points of the segments in S. Consider binary variable xi for
each square ti T1 , and yj for each square tj T2 . Create an ILP as follows.

 
Z0 : min xi + yj
i|ti T1 j|tj T2
 
s.t. xi + yj 1 k | sk S; xi , yj {0, 1} i | ti T1 & j | tj T2
i|l(sk )ti T1 j|r(sk )tj T2

In the solution of Z0 , xi = 1 or 0 (resp. yj = 1 or 0) depending on whether the

square ti T1 (resp. tj T2 ) is in an optimal solution or not. We solve the
corresponding LP , Z0 . Create two partitions: S1 S with those segments sk
such that i|l(sk )ti T1 xi 1/2, and S2 S with those segments s for which

j|r(s )tj T2 yi 1/2. Now, consider the following two ILP s Z1 and Z2 .

 
Z1 : min xi Z2 : min yj
i|ti T1 j|tj T2
 
s.t. xi 1, k | sk S1 s.t. yj 1, k | sk S2
i|l(sk )ti T1 j|r(sk )tj T2

xi {0, 1} i | ti T1 yj {0, 1} j | tj T 2
1
The left end-point of a vertical segment is dened as its top end-point.
10 A. Acharyya et al.

By an analysis identical to Gaur et al. [6], we conclude that OP T1F + OP T2F

2OP T0F 2OP T0I . 


Observe that, both the ILP s, Z1 and Z2 are the problems of covering points by
unit squares. We consider the problem Z1 , and use a multilevel LP relaxation
(see Theorem 6 stated below) to have a factor 8 approximation algorithm for the
DCPUS problem. It consists of 3 steps of LP relaxations. In each step, we use
linear programming to partition the considered points into two disjoint subsets,
and nally we reach to the Restricted-Point-Cover problem, as follows.
Restricted-Point-Cover : Given a point set P in IR2 above the x-axis and
a set R of unit width (x-span) rectangles such that bottom boundary of each
rectangle in R coincides with x-axis, nd a subset of R of minimum size to cover
all the points in P .

Lemma 5. For the standard ILP formulation of the Restricted-Point-Cover

I F
problem ZRP C , if OP TRP C and OP TRP C are the optimum solutions of ZRP C
C 2OP TRP C .
I F
and its LP-relaxation respectively, then OP TRP

C OP TRP C
I F
In Section 3.2 of [3], Bansal and Pruhs showed that OP TRP
for some positive constant for a more generic version of this problem. In our
simplied case 2 for the proof of Lemma 5).
Chan et al. [5] proposed an O(1)-approximation algorithm for the DCPUS prob-
lem using quasi-uniform sampling. Thus using Lemma 4, we have an O(1) ap-
proximation for our DCSUS-ARB problem. However, the following theorem says
that one can get a factor 8 approximation algorithm for the DCPUS problem.

Theorem 6. For the standard ILP formulation of the DCPUS problem Z1 with
a set of points P1 and a set of unit squares T1 , if OP T1I and OP T1F are the op-
timum solutions of Z1 and its LP-relaxation respectively, then OP T1I 8OP T1F .

Proof. It consists of three steps as follows:

Step 1: Divide the plane into unit strips by drawing horizontal lines. No unit
square in T1 can intersect more than one line. Partition T1 into two sets T11 and
T12 , where T11 (resp. T12 ) consists of all squares which intersect even (resp. odd)
indexed lines. Dene binary variables xi for ti T11 and yj for tj T12 . Then,
Z1 is equivalent to the following ILP .

 
Z1 : min xi + yj
i|ti T11 j|tj T12
 
s.t. xi + yj 1 p P1 ; xi , yj {0, 1} i | ti T11 & j | tj T12
i|pti T11 j|ptj T12

create two subsets P11 , P12 P1 . P11 consists of those points

We solve Z1 . Now,
p P1 such that i|pti T11 xi 1/2, and P12 consists of those points p P1 for

which j|ptj T12 yi 1/2. Again consider two ILP s, Z11 and Z12 as follows.
Covering Segments with Unit Squares 11

 
Z11 : min xi Z12 : min yj
i|ti T11 j|tj T12
 
s.t. xi 1, p P11 s.t. yj 1, p P12
i|pti T11 j|ptj T12

xi {0, 1} i | ti T11 yj {0, 1} j | tj T12

By an analysis identical to Gaur et al. [6], we have OP T11 F
+ OP T12F
2OP T1F .
Step 2: In step 1, there are two ILP s, Z11 and Z12 corresponding to Z1 . Observe
that, Z11 (resp. Z12 ) is the problem of covering the points of P11 P1 (resp.
P12 P1 ) by those unit squares which intersect even (resp. odd) indexed lines.
Now, focus our attention on Z11 . Since the subset of points covered by the squares
intersected by two dierent even indexed lines i and j (i = j) are disjoint, we
can split Z11 into dierent ILPs corresponding to each even indexed line, which
can be independently solved.
Consider Z11 corresponding to the line  ( is even). Let P11
be the set of points

which are to be covered by the set of squares T11 , intersected by  . We split

P11 into disjoint sets P111 and P112 ; P111 (resp. P112 ) are the set of points which

are above (resp. below)  . The objective is to cover the members of P111 (resp.

P112 ) by minimum number of squares. Thus, the ILP Z11 can be written as,


Z11 : min xi

i|ti T11


s.t. xi 1 p P111 , and xi 1 p P112

i|pti T11 i|pti T11

xi {0, 1} i | ti T11

Again, since the sets P111 and P112 are disjoint, we may consider the following

two ILP s, Z111 and Z112 as follows.
 
Z111 : min xi Z112 : min xi

i|ti T11 i|ti T11


s.t. xi 1 p P111 s.t. xi 1 p P112

i|pti T11 i|pti T11

xi {0, 1} i | ti T11 xi {0, 1} i | ti T11
 be an optimal fractional solution of Z11 . Clearly, x
Let x  satises all the con-
F F
straints in both Z111 and Z112 . Also, it is observe that OP T111 OP T11 and
F F F F F
OP T112 OP T11 . Combining, we conclude that OP T111 +OP T112 2OP T11 .
A similar equation can be shown for Z12 as follows: OP T121
F F
+OP T122 F
2OP T12 .
Finally, we have the following four equations,
12 A. Acharyya et al.
 F  F  F
1. OP T111 + even OP T112 2 even OP T11 2OP T11
F
,
 even F  F  F
odd OP T122 2 odd OP T12 2OP T12 ,
F
2. odd OP T121 +

Step 3: In this step we apply Lemma 5 independently on each of Z111 , Z112 , where
is even, and Z121 , Z122 where is odd to get the following four equations.
I F I F
(i) OP T111 2OP T111 , (ii) OP T112 2OP T112 ,
I F I F
(iii) OP T121 2OP T121 , (iv) OP T122 2OP T122 .

Now combining all the inequalities of Step 3, we have

 I  I  I  I
even OP T111 +
 even OP T112 + odd OP T121 + odd OP T122 
F  F  F  F
2 even OP T 111 + even OP T 112 + odd OP T121 + odd OP T 122
8 OP T0I , by applying the inequalities in Steps 3, 2 and 1 in this order.  

Lemma 4 and Theorem 6 lead to the following theorem.

Theorem 7. There exists a factor 16 approximation algorithm for DCSUS-

ARB problem that runs in polynomial time.

References
1. E. M. Arkin, A. Banik, P. Carmi, G. Citovsky, M. J. Katz, J. S. B. Mitchell, and
M. Simakov. Choice is hard. In ISAAC, pages 318328, 2015.
2. E. M. Arkin, A. Banik, P. Carmi, G. Citovsky, M. J. Katz, J. S. B. Mitchell, and
M. Simakov. Conict-free covering. In CCCG, 2015.
3. N. Bansal and K. Pruhs. The geometry of scheduling. SIAM J. Comput.,
43(5):16841698, 2014.
4. A. Biniaz, P. Liu, A. Maheshwari, and M. Smid. Approximation algorithms for the
unit disk cover problem in 2D and 3D. Computational Geometry, 60:818, 2016.
5. T. M. Chan, E. Grant, J. Konemann and M. Sharpe. Weighted capacited, priority,
and geometric set cover via improved quasi-uniform sampling In SODA, Pages
15761585, 2012.
6. D. R. Gaur, T. Ibaraki, and R. Krishnamurti. Constant ratio approximation algo-
rithms for the rectangle stabbing problem and the rectilinear partitioning problem.
Journal of Algorithms, 43(1):138152, 2002.
7. D. S. Hochbaum and W. Maass. Approximation schemes for covering and packing
problems in image processing and VLSI. J. ACM, 32(1):130136, January 1985.
8. D. E. Knuth and A. Raghunathan. The problem of compatible representatives.
SIAM Journal on Discrete Mathematics, 5(3):422427, 1992.
9. K. Kobylkin. Computational complexity of guarding of proximity graphs. CoRR,
abs/1605.00313, 2016.
10. R. R. Madireddy and A. Mudgal. Stabbing line segments with disks and related
problems. In CCCG, 2016.
 Replica Placement on Bounded Treewidth
Graphs

Anshul Aggarwal1 , Venkatesan T. Chakaravarthy2 , Neelima Gupta1 , Yogish

Sabharwal2 , Sachin Sharma1 , and Sonika Thakral1
1
University of Delhi, India [email protected], [email protected]
2
IBM Research, India {vechakra,ysabharwal}@in.ibm.com

Abstract. We consider the replica placement problem: given a graph

and a set of clients, place replicas on a minimum set of nodes of the
graph to serve all the clients; each client is associated with a request and
maximum distance that it can travel to get served; there is a maximum
limit (capacity) on the amount of request a replica can serve. The prob-
lem falls under the general framework of capacitated set cover. It admits
an O(log n)-approximation and it is NP-hard to approximate within a
factor of o(log n). We study the problem in terms of the treewidth t of
the graph and present an O(t)-approximation algorithm.

1 Introduction
We study a form of capacitated set cover problem [5] called replica placement
(RP) that nds applications in settings such as data distribution by internet
service providers (ISPs) and video on demand service delivery (e.g., [6, 8]). In
this problem, we are given a graph representing a network of servers and a set
of clients. The clients are connected to the network by attaching each client to
a specic server. The clients need access to a database. We wish to serve the
clients by placing replicas (copies) of the database on a selected set of servers
and clients. While the selected clients get served by the dedicated replicas (i.e.,
cached copies) placed on themselves, we serve the other clients by assigning
them to the replicas on the servers. The assignments must be done taking into
account Quality of Service (QoS) and capacity constraints. The QoS constraint
stipulates a maximum distance between each client and the replica serving it.
The clients may have dierent demands (the volume of database requests they
make) and the capacity constraint species the maximum demand that a replica
can handle. The objective is to minimize the number of replicas opened. The
problem can be formally dened as follows.
Problem Denition (RP): The input consists of a graph G = (V, E), a set
of clients A and a capacity W . Each client a is attached to a node u V, denoted
att(a). For each client a A, the input species a request r(a) and a distance
dmax (a). For a client a A and a node u V, let d(a, u) denote the length
of the shortest path between u and att(a), the node to which a is attached

Corresponding author

Springer International Publishing AG 2017 13

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 1324, 2017.
DOI: 10.1007/978-3-319-62127-2_2
14 A. Aggarwal et al.

- the length is measured by the number of edges and we take d(a, u) = 0, if

u = att(a). We say that a client a A can access a node u V, if d(a, u)
is at most dmax (a). A feasible solution consists of two parts: (i) it identies a
subset of nodes S V where a replica is placed at each node in S; (ii) for each
client a A, it either opens a dedicated replica at a itself for serving the clients
request or assigns the request to the replica at some node u S accessible to
a. The solution must satisfy the constraint that for each node u S, the sum
of requests assigned to the replica at u does not exceed W . The cost of the
solution is the number of replicas opened, i.e., cardinality of S plus the number
of dedicated replicas opened at the clients. The goal is to compute a solution
of minimum cost. In order to ensure feasibility, without loss of generality, we
assume r(a) W , a A. 

The RP problem falls under the framework of the capacitated set cover prob-
lem, the generalization of the classical set cover problem wherein each set is
associated with a capacity specifying the number of elements it can cover. The
latter problem is known to have an O(log n)-approximation algorithm [5]. Using
the above result, we can derive an O(log n)-approximation algorithm for the RP
problem as well. On the other hand, we can easily reduce the classical dominat-
ing set problem to RP: given a graph representing an instance of the dominating
set problem, we create a new client for each vertex and attach it to the vertex;
then, we set dmax () = 1 for all the clients and W = . Since it is NP-hard to
approximate the dominating set problem within a factor of o(log n) [7], by the
above reduction, we get the same hardness result for the RP problem as well.
The RP problem is NP-hard even on the highly restricted special case where
the graph is simply a path, as can be seen via the following reduction from the
bin packing problem. Given K bins of capacity W and a set of items of sizes
s1 , s2 , . . . , sn , for each item i, we create a client a with demand r(a) = si . We
then construct a path of nodes of length K and attach all the clients to one end
of the path and take W to be the capacity of the nodes.
Prior Results: Prior work has studied a variant of the RP problem where
the network is a directed acyclic graph (DAG), and a client a can access a node
u only if there is a directed path from a to u of the length at most dmax (a).
Under this setting, Benoit et al. [3] considered the special case of rooted trees
and presented a greedy algorithm with an approximation ratio of O(), where
is the maximum degree of the tree. For the same problem, Arora et al. [2]
(overlapping set of authors) devised a constant factor approximation algorithm
via LP rounding.
Progress has been made on generalizing the above result to the case of
bounded treewidth DAGs. Recall that treewidth [4] is a classical parameter
used for measuring how close a given graph is to being a tree. For a DAG,
the treewidth refers to the treewidth t of the underlying undirected graph. No-
tice that the reduction from the bin-packing problem shows that the problem is
NP-hard even for trees (i.e., t = 1) and rules out the possibility of designing an
exact algorithm running in time nO(t) (say via dynamic programming) or FPT
algorithms with parameter t.
Replica Placement on Bounded Treewidth Graphs 15

Arora et al. [1] made progress towards handling DAGs of bounded treewidth
and designed an algorithm for the case of bounded-degree, bounded-treewidth
graphs. Their algorithm achieves an approximation ratio of O( + t), where
is the maximum degree and t is the treewidth of the DAG. Their result also ex-
tends for networks comprising of bounded-degree bounded-treewidth subgraphs
connected in a tree like fashion.
Our Result and Discussion: We study the RP problem on undirected
graphs of bounded treewidth. Our main result is an O(t)-approximation algo-
rithm running in polynomial time (the polynomial is independent of t and the
approximation guarantee). In contrast to prior work, the approximation ratio
depends only on the treewidth and is independent of parameters such as the
maximum degree.
Our algorithm is based on rounding solutions to a natural LP formulation, as
in the case of prior work [2, 1]. However, the prior algorithms exploit the acyclic
nature of the graphs and the bounded degree assumption to transform a given
LP solution to a solution wherein each client is assigned to at most two replicas.
In other words, they reduce the problem to a capacitated vertex cover setting,
for which constant factor rounding algorithms are known [10].
The above reduction does not extend to the case of general bounded treewidth
graphs. Our algorithm is based on an entirely dierent approach. We introduce
the notion of clustered solutions, wherein the partially open nodes are grouped
into clusters and each client gets served only within a cluster. We show how
to transform a given LP solution to a new solution in which a partially-open
node participates in at most (t + 1) clusters. This allows us to derive an overall
approximation ratio O(t). The notion of clustered solutions may be applicable
in other capacitated set cover settings as well.
Other Related Work: The RP problem falls under the framework of the
capacitated set cover problem (CSC), which admits an O(log n)-approximation
[5]. Two versions of the CSC problem have been studied: soft capacity and hard
capacity settings. Our work falls under the more challenging hard capacity set-
ting, wherein a set can be picked at most once. The capacitated versions of the
vertex cover problem (e.g., [10]) and dominating set problem (e.g., [9]) have also
been studied. Our result applies to the capacitated dominating problem with
uniform capacities and yields O(t)-approximation.
Full Version: Due to space constraints, some of the proofs and details of
analysis could not be included in this version. A full version of the paper is
available as an Arxiv preprint (https://arxiv.org/abs/1705.00145).

2 Overview of the Algorithm

Our O(t)-approximation algorithm is based on rounding solution to a natural LP

formulation. In this section, we present an outline of the algorithm highlighting
its main features, deferring a detailed description to subsequent sections. We
assume that the input includes a decomposition T of treewidth t of the input
network G = (V, E).
16 A. Aggarwal et al.

LP Formulation: For each node u V, we introduce a variable y(u) to

represent the extent to which a replica is opened at u and similarly, for each
client a A, we add a variable y(a) to represent the extent to which a dedicated
replica is opened at a itself. For each client a A and each node u V accessible
to a, we use a variable x(a, u) to represent the extent to which a is assigned to
u. For a client a A and a node u V, we use the shorthand a u to mean
that a can access u.
 
min y(a) + y(u)
aA uV

y(a) + x(a, u) 1 for all a A (1)
uV : au

x(a, u) r(a) y(u) W for all u V (2)
aA : au
x(a, u) y(u) for all a A and u V with a u (3)
0 y(u), y(a) 1 for all u V and a A (4)

Constraint (3) stipulates that a client a cannot be serviced at a node u for an

amount exceeding the extent to which u is open. For an LP solution = x, y,
let cost() denote the objective value of .
The following simple notations will be useful in our discussion. With respect
to an LP solution , we classify the nodes into three categories based on the
extent to which they are open. A node u is said to be fully-open, if y(u) = 1;
partially-open, if 0 < y(u) < 1; and fully-closed, if y(u) = 0. A client a is said to
be assigned to a node u, if x(a, u) > 0. For a set of nodes U , let y(U ) denote the
extent to which the vertices in U are open, i.e., y(U ) = uU y(u).
Outline: The major part of the rounding procedure involves transforming
a given LP solution in = xin , yin  into an integrally open solution: wherein
which each node u V is either fully open or closed. Such a solution diers
from an integral solution as a client may be assigned to multiple nodes (possibly
to its own dedicated replica as well). We address the issue easily via a cycle
cancellation procedure to get an integral solution.
The procedure for obtaining an integrally open solution works in two stages.
First it transforms the input solution into a clustered solution, which is then
transformed into an integrally open solution. The notion of clustered solution
lies at the heart of the rounding algorithm. Intuitively, in a clustered solution,
the set of partially open (and closed) nodes are partitioned into a collection of
clusters C and the clients can be partitioned into a set of corresponding groups
satisfying three useful properties, as discussed below.
Let = x, y be an LP solution. It will be convenient to express the three
properties using the notion of linkage: we say that a node u is linked to a node
v, if there exists a client a assigned to both u and v. For constants and , the
solution is said to be (, )-clustered, if the set of partially-open nodes can be
partitioned into a collection of clusters, C = {C1 , C2 , . . . , Ck } (for some k), such
that the the following properties are true:
Replica Placement on Bounded Treewidth Graphs 17

(a) (b)

Fig. 1. (a) Illustration for clustered solution. Three clusters are shown C1 , C2 and C3 ,
open to an extent of 0.4, 0.4 and 0.5; the clusters are linked to the sets of fully-open
nodes {v1 , v2 , v4 }, {v1 , v2 , v3 , v4 }, and {v2 , v4 , v5 , v6 }. The solution is (0.5, 4)-clustered.
(b) Illustration for regions. The gure shows an example tree decomposition. The bags
lled solidly represent already identied boundary bags. All checkered bags belong to
the region headed by P .

Localization: assignments from clients to the partially-open nodes is local-

ized, i.e., two partially-open nodes are linked only if they belong to the same
cluster.
Distributivity: assignments from the clients to fully-open nodes are restricted,
i.e., for any Cj , there are at most  fully-open nodes that are linked to the
nodes in Cj .
Bounded opening: clusters are tiny, i.e., the total extent to which any cluster
is open is at most , i.e., y(Cj ) < .
Figure 1 (a) provides an illustration. In the rst stage of the rounding algorithm,
we transform the input solution in into an (, t + 1)-clustered solution with the
additional guarantee that the number of clusters is at most a constant factor of
cost(in ), where [0, 1/2] is a tunable parameter. The lemma below species
the transformation performed by the rst stage.

Lemma 1. Fix any constant 1/2. Any LP solution can be transformed

into a (, t + 1)-clustered solution  such that cost(  ) is at most 2 + 6(t +
1)cost()/. Furthermore, the number of clusters is at most 3 + 8 cost()/.

At a high level, the lemma is proved by considering the tree decomposition

T of the input graph G = (V, E) and performing a bottom-up traversal that
identies a suitable set of boundary bags. We use these boundary bags to split
the tree into a set of disjoint regions and create one cluster per region. We
then fully open the nodes in the boundary bags and transfer assignments from
the nodes that stay partially-open to these fully-open nodes. The transfer of
assignments is performed in such a manner that clusters get localized and have
distributivity of (t+1). By carefully selecting the boundary bags, we shall enforce
that each cluster is open to an extent of only and that the number of clusters
is also bounded. The proof is discussed in Section 3.
The goal of the second stage is to transform a (1/4, t + 1)-clustered solution
(obtained from Lemma 1) into an integrally open solution. At a high level, the
18 A. Aggarwal et al.

localization property allows us to independently process each cluster C C

and its corresponding group of clients A. The clients in A are assigned to a
set of fully-open nodes, say F . For each node u F , we identify a suitable
node v C called the consort of u C and fully open v. Then the idea is
to transfer assignments from the non-consort nodes to the nodes in F and their
consorts in such a manner that at the end, no client is assigned to the non-consort
nodes. This allows us to fully close the non-consort nodes. The localization and
bounded opening properties facilitate the above maneuver. On the other hand,
the distributivity property ensures that F is at most (t + 1). This means that
we fully open at most (t + 1) consorts per cluster. Thus, overall increase in cost
is at most (t + 1)|C|. Since |C| is guaranteed to be linear in cost(in ), we get an
O(t) approximation factor.
Lemma 2. Let = x, y be a (1/4, t + 1)-clustered solution via a collection
of clusters C. The solution can be transformed into an integrally open solution
 = x , y   such that cost(  ) 2 cost() + 2(t + 1)|C|.
Once we obtain an integrally open solution, it can be transformed to an inte-
gral solution by applying a cycle cancellation strategy, as given by the following
lemma.
Lemma 3. Any integrally open solution = x, y can be transformed to an
integral solution  = x , y   such that cost(  ) 4 cost().
We can transform any input LP solution in into an integral solution out by
applying the above three transformations leading to the following main result of
the paper: the RP problem admits on O(t)-approximation poly-time algorithm.

3 Clustered Solutions: Proof of Lemma 1

The goal is to transform a given solution into an (, t+1)-clustered solution with
the properties claimed in the lemma. The idea is to select a set of partially-open
or closed nodes and open them fully, and then transfer assignments from the
other partially-open nodes to them in such a manner that the partially-open
nodes get partitioned into clusters satisfying the three properties of clustered
solutions. An issue in executing the above plan is that the capacity at a newly
opened node may be exceeded during the transfer. We circumvent the issue by
rst performing a pre-processing step called de-capacitation.

3.1 De-capacitation
Consider an LP solution = x, y and let u be a partially-open or closed node.
The clients that can access u might have been assigned to other partially-open
nodes under . We call the node u de-capacitated, if even when all the above
assignments are transferred to u, the capacity at u is not exceeded; meaning,
 
x(a, v) < W,
au v: av vPO
Replica Placement on Bounded Treewidth Graphs 19

For each partially-open node v (considered in an arbitrary order)

For each client a that can access both u and v (considered
 in an arbitrary order)
Compute capacity available at u: cap(u) = W bA : bu x(b, u) r(b)
If cap(u)= 0 exit the procedure

cap(u)
= min x(a, v), r(a)
Increment x(a, u) by and decrement x(a, v) by .

Fig. 2. Pulling procedure for a given partially-open or closed node u.

where PO is the set of partially-open nodes under (including u). The solution
is said to be de-capacitated, if all the partially-open and the closed nodes are
de-capacitated.
The preprocessing step transforms the input solution into a de-capacitated
solution by performing a pulling procedure on the partially-open and closed
nodes. Given a partially-open or closed node u, the procedure transfers assign-
ments from other partially-open nodes to u, as long as the capacity at u is not
violated. The procedure is shown in Figure 2, which we make use of in other
components of the algorithm as well.

Lemma 4. Any LP solution = x, y can be transformed into a de-capacitated

solution  = x , y   such that cost(  ) 2 cost().

Proof. We consider the partially-open and closed nodes, and process them in
an arbitrary order, as follows. Let u be a partially-open or closed node. Hy-
pothetically, consider applying the pulling procedure on u. The procedure may
terminate in one of two ways: (i) it exits mid-way because of reaching the ca-
pacity limit; (ii) the process executes in its entirety. In the former case, we fully
open u and perform the pulling procedure on u. In the latter case, the node u
is de-capacitated and so, we leave it as partially-open or closed, without per-
forming the pulling procedure. It is clear that the above method produces a
de-capacitated solution  . We next analyze the cost of  . Let s be the num-
ber of partially-open or closed nodes converted to be fully-open. Apart from
these conversions, the method does not alter the cost and  so, cost(  ) is at most
s + cost(). Let the total amount of requests be rtot = aA r(a). The extra
cost s is at most rtot /W , since any newly opened node is lled to its capacity.
Due to the capacity constraints, the input solution must also incur a cost of
at least rtot /W . It follows that cost(  ) is at most 2 cost(). 


3.2 Clustering

Given Lemma 4, assume that we have a de-capacitated solution = x, y.

We next discuss how to transform into an (, t + 1)-clustered solution. The
transformation would perform a bottom-up traversal of the tree decomposition
and identify a set of partially-open or closed nodes. It would then fully open
20 A. Aggarwal et al.

them and perform the pulling procedure on these nodes. The advantage is that
the above nodes are de-capacitated and so, the pulling procedure would run to its
entirety (without having to exit mid-way because of reaching capacity limits).
As a consequence, the linkage between the nodes gets restricted, leading to a
clustered solution. Below we rst describe the transformation and then present
an analysis.
Transformation: Consider the given tree decomposition T . We select an
arbitrary bag of T and make it the root. A bag P is said to be an ancestor of a
bag Q, if P lies on the path connecting Q and the root; in this case, Q is called a
descendant of P . We consider P to be both an ancestor and descendant of itself.
A node u may occur in multiple bags; among these bags the one closest to the
root is called the anchor of u and it is denoted anchor(u). A region in T refers
to any set of contiguous bags (i.e., the set of bags induce a connected sub-tree).
In transforming into a clustered solution, we shall encounter three types
of nodes and it will be convenient to color them as red, blue and brown. To
start with, all the fully-open nodes are colored red and the remaining nodes
(partially-open nodes and closed nodes) are colored blue. The idea is to carefully
select a set of blue nodes, fully-open them and perform the pulling procedure on
these nodes; these nodes are then colored brown. Thus, while the blue nodes are
partially-open or closed, the red and the brown nodes are fully-open, with the
brown and blue nodes being de-capacitated.
The transformation identies two kinds of nodes to be colored brown: helpers
and boundary nodes. We say that a red node u V is proper, if it has at least
one neighbor v V which is a blue node. For each such proper red node u, we
arbitrarily select one such blue neighbor v V and declare it to be the helper
of u. Multiple red nodes are allowed to share the same helper. Once the helpers
have been identied, we color them all brown. The boundary brown nodes are
selected via a more involved bottom-up traversal of T that works by identifying
a set B of bags, called the boundary bags. To start with, B is initialized to be
the empty set. We arrange the bags in T in any bottom-up order (i.e., a bag
gets listed only after all its children are listed) and then iteratively process each
bag P as per the above order. Consider a bag P . We dene the region headed
by P , denoted Region(P ), to be the set of bags Q such that Q is a descendant
of P , but not the descendant of any bag already in B. See Figure 1 (b) for an
illustration. A blue node u is said to be active at P , if it occurs in some bag
included in Region(P ). Let active(P ) denote the set of blue nodes active at P .
We declare P to be a boundary bag and add it to B under three scenarios: (i) P
is the root bag. (ii) P is the anchor of some red node.(iii) the extent to which
the nodes in active(P ) are open is at least , i.e., uactive(P ) y(u) . If
P is identied as a boundary bag, then we select all the blue nodes appearing
in the bag and change their color to be brown. Once the bottom-up traversal
is completed, we have a set of brown nodes (helpers and boundary nodes). We
consider these nodes in any arbitrary order, open them fully, and perform the
pulling procedure on them. We take  to be the solution obtained by the above
process. This completes the construction of  . We note that a node may change
Replica Placement on Bounded Treewidth Graphs 21

its color from blue to brown in the above process, and the new color is to be
considered while determining the active sets thereafter. Notice that during the
whole process of the above transformation, the solution continues to remain
de-capacitated.
Analysis: We now show that  is an (, t + 1)-clustered solution. To start
with, we have a set of red nodes that are fully-open and a set of blue nodes that
are either partially-open or closed under . The red nodes do not change color
during the transformation. On the other hand, each blue node u becomes active
at some boundary bag P . If u occurs in the bag P , it changes its color to brown,
otherwise it stays blue. Thus, the transformation partitions the set of originally
blue nodes into a set of brown nodes and a set of nodes that stay blue. In the
following discussion, we shall use the term blue to refer to the nodes that stay
blue. With respect to the solution  , the red and brown nodes are fully-open,
whereas the blue nodes are partially-open or closed.
Recall that with respect to  , two nodes u and v are linked, if there is a
client a assigned to both u and v. In order to prove the properties of (, t + 1)-
clustering, we need to analyze the linkage information for the blue nodes. We
rst show that the blue nodes cannot be linked to brown nodes, by proving the
following stronger observation.

Proposition 1. If a client a A is assigned to a blue node u under  , then a

cannot access any brown node v.

Proposition 1 rules out the possibility of a blue node u being linked to any
brown node. Thus, u may be linked to a red node or another blue node. The
following lemma establishes a crucial property on the connectivity in these two
settings.

Lemma 5. (a) If two blue nodes u and v are linked under  , then there must
exist a path connecting u and v consisting of only blue nodes. (b) If a blue node
u is linked to a red node v under  , then there must exist a path p connecting u
and v such that barring v, the path consists of only blue nodes.

The transformation outputs a set of boundary bags B; let B denote the set
of non-boundary bags. If we treat the bags in B as cut-vertices and delete them
from T , the tree splits into a collection R of disjoint regions. Alternatively, these
regions can be identied in the following manner. For each bag P B and each
of its non-boundary child Q B, add the region headed by Q (Region(Q)) to
the collection R. Let the collection derived be R = {R1 , R2 , . . . , Rk }. It is easy
to see that R partitions B and that the regions in R are pairwise disconnected
(not connected by edges of the tree decomposition).
In order to show that  is an (, t + 1)-clustered solution, let us suitably
partition the set of blue nodes into a collection of clusters C. For each region Rj ,
let Cj be the set of partially-open nodes that occur in some bag of Rj . We take
C to be the collection {C1 , C2 , . . . , Ck }. It can be veried that the collection C
is a partitioning of the set of partially-open nodes. Based on Lemma 5 we can
establish the following result.
22 A. Aggarwal et al.

Lemma 6. The solution  is (, t + 1)-clustered.

We next analyze the cost of the solution  = x , y  . Let Red, Blue and
Brown denote the set of red, brown and blue nodes, respectively. We can see that
in constructing  the brown nodes are the only new nodes opened fully and
hence, cost(  ) cost() + |Brown|. We create a brown helper node for each red
node and furthermore, for each boundary bag P B, we convert all the blue
nodes in P to be brown and the number of blue nodes per bag is at most (t + 1).
Thus, the number of brown nodes is at most |Red| + (t + 1)|B|. A bag P is made
into a boundary bag under one of the three scenarios: (i) P is the root bag; (ii)
P is the anchor of some red node; (iii) the total extent to which the nodes in
active(P ) are open is at least . The number of boundary bags of the rst two
types are 1 + |Red| and those of the third type can be (1/) times the extent to
which the blue nodes are open, which is in turn, at most cost(). Using the above
arguments, we can show that |B| is at most 2 + |Red| + cost()/ and cost(  ) is
at most 2 + 3(t + 1)cost()/. The preprocessing step of de-capacitation incurs a
2-factor increase in cost. Taking this into account, we get the cost bound claimed
in the statement of Lemma 1.
As mentioned earlier, an issue with the collection C is that it may have
more clusters than the bound claimed in Lemma 1. The issue can be resolved
as follows. Consider each boundary bag P . All the non-boundary children of
P have a corresponding cluster in C and let CP denote the collection of these
clusters. We merge any two clusters C, C  from CP having y(C), y(C  ) /2.
The process is stopped when we cannot nd two such clusters.
It can be shown that the process of merging does not aect distributivity
and the total number of clusters in the transformed solution is at most 3 +
4cost()/. The preprocessing step of de-capacitation incurs a 2-factor increase
in cost. Taking this into account, we get the bound on number of clusters claimed
in the statement of Lemma 1.

4 Integrally Open Solution: Proof of Lemma 2

Our goal is to transform a given (1/4, t + 1)-clustered solution = x, y into
an integrally open solution  . We classify the clients into two groups, small and
large, based on the extent to which they are served by dedicated replicas: a client
a A said to be small, if y(a) 1/2, and it is said to be large otherwise. Let
As and Al denote the set of small and large clients, respectively.
We pre-process the solution by opening a dedicated replica at each large
client a and removing its assignments to the nodes (set y(a) = 1 and set x(a, u) =
0 for all nodes u accessible to a). We see that the transformation at most doubles
the cost and the solution remains (1/4, t + 1)-clustered.
Consider the pre-processed solution . Let C denote the set of clusters (of
the partially-open nodes) under . For each cluster C C, we shall fully open a
selected set of at most 2(t + 1) nodes and fully close rest of the nodes in it.
We now describe the processing for a cluster C C. Let A As denote the set
of clients assigned to the nodes in C. By the distributivity property, these clients
Replica Placement on Bounded Treewidth Graphs 23

are assigned to at most (t + 1) fully-open nodes, denoted F = {u1 , u2 , . . . , ut+1 }.

A client a A may be assigned to multiple nodes from F . In our procedure, it
would be convenient if each client is assigned to at most one node from F and
we obtain such a structure using the following transformation.

Proposition 2. Given a solution = x, y, a set of fully-open nodes F and a

set of clients A, we can obtain a solution  = x , y   such that each client a A
is assigned to at most one node from F . Furthermore, the transformation does
not alter the other assignments, i.e., for any node u V and any client a A,
if u  F or a  A, then x (a, u) = x(a, u). Moreover, cost(  ) cost() + |F |.

The above proposition is proved via a cycle cancellation procedure that trans-
fers assignments amongst the nodes in F . The procedure can ensure that, except
for at most |F | clients, every other client a A is assigned to at most one node
from F . We open dedicated replicas at the exceptional clients and this results in
an cost increase of at most |F |.
The proposition does not alter the other assignments and so, its output solu-
tion is also (1/4, t + 1)-clustered. Given the proposition and the pre-processing,
we can assume that = x, y is (1/4, t+1)-clustered wherein each client a A is
assigned to at most one node from F and that y(a) 1/2. For each node ui F ,
let Ai A denote the set of clients assigned to the node ui . The proposition
guarantees that these sets are disjoint.
For a node v and a client a, let load(a, v) denote the amount of load imposed
by a on v towards the capacity: load(a, v) = x(a, v)r(a). It will be convenient
to dene the notion over sets of clients and nodes. For a set of clients B and
a set of nodes U , let load(B, U )  denote the load imposed by the clients in
B on the nodes U : load(B, U ) = aB,vU :av x(a, v)r(a); when the sets are

singletons, we

 shall omit the curly braces. Similarly, for a subset C C, let
load(C ) = vC  load(v).
The intuition behind the remaining transformation is as follows. We shall
identify a suitable set of nodes L = {v1 , v2 , . . . , vt+1 } from C, with vi being called
the consort of ui in C, and fully open all these nodes. Then, we consider the non-
consort nodes C  = C L and for each i t+1, we transfer the load load(Ai , C  )
to the node ui . As a result, no clients are assigned to the non-consort nodes any
more and so, they can be fully closed. In order to execute the transfer, for each
i t+1, we create space in ui by pushing a load equivalent to load(Ai , C  ) from
ui to its (fully-opened) consort vi . The amount of load load(Ai , C  ) involved in
the transfer is very small: the bounded opening property ensures that y(C) <
1/4 and thus, load(Ai , C  ) < W/4. The fully-opened consort vi has enough
additional space to receive the load: y(vi ) 1/4 and so, load(A, vi ) W/4,
which means that if we fully open the consort, we get an additional space of
(3/4)W . However, an important issue is that a consort vi may not be accessible to
all the clients in Ai . Therefore, we need to carefully choose the consorts in such a
manner that each fully open node ui has enough load accessible to the consort vi
that can be pushed to vi . Towards this purpose, we dene the notion of pushable
load. For a node ui F and a node v C, let pushable(ui , v) denote the amount
24 A. Aggarwal et al.

of load on ui that is accessible to v: pushable(ui , v) = aAi :av x(a, ui )r(a).
We next show how to identify a suitable set of consorts such that the pushable
load is more than the load that we wish to transfer.
Lemma 7. We can nd a set of nodes L = {v1 , v2 , . . . , vt+1 } such that for all
i t + 1, pushable(ui , v) load(Ai , C  ).
We have shown that each node ui has a load of at least load(Ai , C  ) which
can be pushed to its consort vi . As observed earlier load(Ai , C  ) < W/4 and
load(Ai , vi ) W/4. Hence, when we fully open the consort, we get an additional
space of (3/4)W , which is sucient to receive the load from ui .
Given the above discussion, we iteratively consider each cluster Cj C and
perform the above transformation. This results in (t + 1) consorts from Cj being
fully-opened and all the other nodes in Cj being fully closed. At the end of
processing all the clusters, we get a solution in which each node either fully open
or fully close. For each cluster Cj , we incur an extra cost of at most (t + 1)
while applying Proposition 2, and an additional cost of (t + 1) for opening the
consorts. Thus, the cost increases by at most 2(t + 1)|C|.

References
1. S. Arora, V. Chakaravarthy, K. Gupta, N. Gupta, and Y. Sabharwal. Replica place-
ment on directed acyclic graphs. In V. Raman and S. Suresh, editors, Proceedings
of the 34th International Conference on Foundation of Software Technology and
Theoretical Computer Science (FSTTCS), pages 213225, 2014.
2. S. Arora, V. Chakaravarthy, N. Gupta, K. Mukherjee, and Y. Sabharwal. Replica
placement via capacitated vertex cover. In A. Seth and N. Vishnoi, editors, Proceed-
ings of the 33rd International Conference on Foundations of Software Technology
and Theoretical Computer Science (FSTTCS), pages 263274, 2013.
3. A. Benoit, H. Larcheveque, and P. Renaud-Goud. Optimal algorithms and approxi-
mation algorithms for replica placement with distance constraints in tree networks.
In Proceedings of the 26th IEEE International Parallel and Distributed Processing
Symposium (IPDPS), pages 10221033, 2012.
4. H. Bodlaender and A. Koster. Combinatorial optimization on graphs of bounded
treewidth. Computer Journal, 51(3):255269, 2008.
5. J. Chuzhoy and J. Naor. Covering problems with hard capacities. SIAM Journal
of Computing, 36(2):498515, 2006.
6. I. Cidon, S. Kutten, and R. Soer. Optimal allocation of electronic content. Com-
puter Networks, 40:205218, 2002.
7. U. Feige. A threshold of ln n for approximating set cover. Journal of the ACM,
45(4):634652, 1998.
8. K. Kalpakis, K. Dasgupta, and O. Wolfson. Optimal placement of replicas in trees
with read, write, and storage costs. IEEE Transactions on Parallel and Distributed
Systems, 12:628637, 2001.
9. M. Kao, H. Chen, and D. Lee. Capacitated domination: Problem complexity and
approximation algorithms. Algorithmica, 72(1):143, 2015.
10. B. Saha and S. Khuller. Set cover revisited: Hypergraph cover with hard capacities.
In A. Czumaj, K. Mehlhorn, A. Pitts, and R. Wattenhofer, editors, Proceedings
of the 39th International Colloquium on Automata, Languages, and Programming
(ICALP), volume 7391 of LNCS, pages 762773. Springer, 2012.
 Fast Exact Algorithms for Survivable Network
Design with Uniform Requirements 

Akanksha Agrawal1 , Pranabendu Misra1 , Fahad Panolan1 , and Saket

Saurabh1,2
1
Department of Informatics, University of Bergen, Norway.
{Akanksha.Agrawal|Pranabendu.Misra|Fahad.Panolan}@uib.no
2
The Institute of Mathematical Sciences, HBNI, Chennai, India. [email protected]

Abstract. We design exact algorithms for the following two problems

in survivable network design: (i) designing a minimum cost network with
a desired value of edge connectivity, which is called Minimum Weight
-connected Spanning Subgraph and (ii) augmenting a given net-
work to a desired value of edge connectivity at a minimum cost which is
called Minimum Weight -connectivity Augmentation. Many well
known problems such as Minimum Spanning Tree, Hamiltonian Cy-
cle, Minimum 2-Edge Connected Spanning Subgraph and Min-
imum Equivalent Digraph reduce to these problems in polynomial
time. It is easy to see that a minimum solution to these problems con-
tains at most 2(n1) edges. Using this fact one can design a brute-force
algorithm which runs in time 2O(n(log n+log ) . However no better algo-
rithms were known. In this paper, we give the rst single exponential time
algorithm for these problems, i.e. running in time 2O(n) , for both undi-
rected and directed networks. Our results are obtained via well known
characterizations of -connected graphs, their connections to linear ma-
troids and the recently developed technique of dynamic programming
with representative sets.

1 Introduction
The survivable network design problem involves designing a cost eective com-
munication network that can survive equipment failures. The failure may be
caused by any number of things such as a hardware or software breakage, human
error or a broken link between two network components. Designing a network
which satises certain connectivity constraints, or augmenting a given network
to a certain connectivity are important and well studied problems in network
design. In terms of graph theory these problems correspond to nding a spanning
subgraph of a graph which satises given connectivity constraints and, augment-
ing the given graph with additional edges so that it satises the given constraints,
respectively. Designing a minimum cost network which connects all the nodes,
is the well-known Minimum Spanning Tree(MST) problem. However such a

Supported by Parameterized Approximation ERC Starting Grant 306992 and
Rigorous Theory of Preprocessing ERC Advanced Investigator Grant 267959.

Springer International Publishing AG 2017 25

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 2536, 2017.
DOI: 10.1007/978-3-319-62127-2_3
26 A. Agrawal et al.

network fails on the failure of a single link. This leads to the question of design-
ing a minimum cost network which can survive one or more link failures. Such a
network must be -connected, in order to survive 1 link failures (we use the
term -connected to represent -edge connected). This problem is NP-hard (for
2), and a 2-approximation algorithm is known [19]. In the special case when
the weights are 1 or , i.e. we wish to nd a minimum spanning -connected
2
subgraph, a 1 + +1 approximation may be obtained in polynomial time [6]. The
above results also hold in the case of directed graphs. The case of = 1 for
digraphs, known as Minimum Strong Spanning Subgraph(MSSS), is NP-
hard as it is a generalization of the Hamiltonian Cycle. Further, the Minimum
Equivalent Graph(MEG) problem reduces to it in polynomial time.
Adding a minimum number of edges to make the graph satisfy certain con-
nectivity constrains is known as minimum augmentation problem. Minimum
augmentation nd application in designing survivable networks [12, 16] and in
data security [14, 17]. Watanabe and Nakamura [25] gave a polynomial time algo-
rithm for solving the -edge connectivity augmentation in an undirected graph,
where we want to add minimum number of edges to the graph to make it -edge
connected. Frank gave a polynomial time algorithm for the same problem in
directed graphs [11]. However in the weighted case, or when the augmenting set
must be a subset of a given set of links, the problem becomes NP-Hard problem.
Even the restricted case of augmenting the edge connectivity of a graph from
1 to remains NP-hard [1]. A 2-approximation may be obtained for these
problems, by choosing a suitable weight function and applying the algorithm of
[19]. We refer to [1, 4, 18, 20] for more details, other related problems and further
applications. A few results are also known in the frameworks of parameterized
complexity and exact exponential time algorithms. Marx and Vegh gave an FPT
algorithm for computing a minimum cost set of at most k links, which augments
the connectivity of an undirected graph from 1 to [22]. Basavaraju et al.
[2] improved the running time of this algorithm and, also gave an algorithm for
another variant of this problem. Bang-Jensen and Gutin [1, Chapter 12] obtain
an FPT algorithm for a variant of MSSS in unweighted graphs. The rst exact
algorithms for MEG and MSSS, running in time O(2O(m) nO(1) ), where m is
the number of edges in the graph, were given in by Moyles and Thompson [23] in
1969. Only recently, Fomin et al. [10] gave the rst single-exponential algorithm
for MEG and MSSS, i.e. with a running time of 2O(n) . For the special case of
Hamiltonian Cycle, a O(2n ) time algorithm is known [15, 3] for digraphs from
1960s. It was recently improved to O(1.657n ) for undirected graphs [5], and to
O(1.888n ) for bipartite digraphs [9] (but these are randomized algorithms). For
other results and more details we refer to Chapter 12 of [1].
In this paper we consider the problem of designing an exact algorithm for
nding a minimum weight spanning subgraph of a given -connected (di)graph.
Minimum Weight -connected Spanning Subgraph
Input: A graph G (or digraph D), and a weight function w on the edges(or
the arcs).
Output: A minimum weight spanning -connected subgraph.
Fast Exact Algorithms for Survivable Network Design with Uniform Requirements 27

One can observe that such a subgraph contains at most (n1) edges (2(n1)
arcs for digraphs). Hence a solution can be obtained by enumerating all possible
subgraphs with at most these many edges and testing if it is -connected. How-
ever such an algorithm will take 2O(n(log n+log )) time. One may try a more
clever approach, by using the observation that we can enumerate all possible
minimal -connected graphs in 2O(n) time. Then we test if any of these graph
is isomorphic to a subgraph of the input graph. However, subgraph isomorphism
requires 2n(log n+log ) unless the Exponential Time Hypothesis fails [7]. In this
paper, we give the rst single exponential algorithm for this problem that runs
in time 2O(n) . As a corollary, we also obtain single exponential time algorithm
for the minimum weight connectivity augmentation problem.
Minimum Weight -connectivity Augmentation
Input: A graph G (or a digraph D), a set of links L V V (ordered pairs
in case of digraphs), and a weight function w : L N.
Output: A minimum weight subset L of L such that G L (or D L) is
-connected

Our Methods and Results. We extend the algorithm of Fomin et al. for nding a
Minimum equivalent Graph [10], to solve Minimum weight - Connected
sub-digraph, exploiting the structural properties of -connected (di)graphs. A
digraph D is -connected if and only if for some r V (D), there is a collection
I of arc disjoint in-branchings rooted at r and a collection O of arc disjoint
out-branchings rooted at r. Then computing a I and a O with the largest pos-
sible intersection yields a minimum weight -connected spanning sub-digraph.
We show that the solution can be embedded in a linear matroid of rank O(n),
and then compute the solution by a dynamic programming algorithm with rep-
resentative sets over this matroid.
Theorem 1. Let D be a -edge connected digraph on n vertices and w : A(D)
N. Then we can nd a min-weight -edge connected subgraph of D in 2O(n) time.
For the case of undirected graphs, no equivalent characterization is known. How-
ever, we obtain a characterization by converting the graph to a digraph with
labels on the arcs, corresponding to the undirected edges. Then computing a
solution that minimizes the number of labels used, gives the following theorem.
Theorem 2. Let G be a -edge connected graph on n vertices and w : E(G)
N. Then we can nd a min-weight -edge connected subgraph of G in 2O(n)
time.
For the problem of augmenting a network to a given connectivity requirement,
at a minimum cost, we obtain the following results by applying the previous
theorems with suitably chosen weight functions.
Theorem 3. Let D be a digraph (or a graph) on n vertices, L V (D) V (D)
be a collection of links with weight function w : L N. For any integer , we can
nd a min-weight L L such that D = (V (D), A(D) L ) is -edge connected,
in time 2O(n) .
28 A. Agrawal et al.

Preliminaries. Due to space constraints, standard denitions, notations related

graphs and matroids have been omitted from this extended abstract. These pre-
liminaries and other results on matriods and representative sets, may be found
in [8, 10]. We only mention the following. We say that a family S = {S1 , . . . , St }
of subsets of a universe U is a p-family if each set in S has cardinality at most
p. For two families S1 and S2 of a universe U , dene S1 S2 = {Si Sj |
Si S1 , Sj S2 and Si Sj = }. We use to denote the exponent of matrix
multiplication.

Denition 1 (Min/Max q-Representative Family [8, 10, 21]). Given a

matroid M = (E, I), a p-family B of E and a non-negative weight function
w : B N. We say that B B is a min (max) q-representative for B if for
every set Y E of size at most q, if there is a set X B, such that X Y =
and X Y I, then there is a set X B such that X
Y = , X Y I and

w(X) w(X) (w(X) w(X)). If B B is a min (max) q-representative for B
then we denote it by B qminrep B (B qmaxrep B).

Theorem 4 ([8, 10]). Let M = (E, I) be a linear matroid of rank k = p +

q, and matrix AM be a representation of M over a eld F. Also, let B =
{B1 , B2 , . . . , Bt } be a p-family of independent sets in E and w : B N be a
non-negative weight function. Then, there exists B qminrep B (B qmaxrep B) of


q q
size at most p+q p . Moreover, B minrep B (B maxrep B) can be computed in

p+q
p+q1
at most O( p tp + t p ) operations over F.

2 Directed Graphs
In this section, we give a single exponential exact algorithm, that is of running
time 2O(n) , for computing a minimum weight spanning -connected subgraph of
a connected n-vertex digraph. We rst consider the unweighted version of the
problem and it will be clear that the same algorithm works for weighted version
as well. In a digraph D, we dene OutD (v) = {(v, w) A(D)} and InD (v) =
{(u, v) A(D)} to be the set of out-edges and in-edges of v, respectively. We
begin with the following characterization of -connectivity in digraphs.

Lemma 1 (3 ). Let D be a digraph. Then D is -connected if and only if for

any r V (D), there is a collection of arc disjoint in-branchings rooted at r,
and a collection of arc disjoint out-branchings rooted at r.

Let D be the input to our algorithm, which is a -connected digraph on n

vertices. Let us x a vertex r V (D). By Lemma 1, any minimal -connected
subgraph of D is a union of a collection I of arc disjoint in-branchings and a
collection O of arc disjoint out-branchings which are all rooted at vertex r.
The following lemma relates the size of such a minimal subgraph to the number
of arcs which appear in both I and O and it follows easily from Lemma 1. Here,
3
Proof of the results marked () are omitted due to space constraints.
Fast Exact Algorithms for Survivable Network Design with Uniform Requirements 29

A(I) denotes the set of arcs which are present in some I I and A(O) denotes
the set of arcs which are present in some O O.

Lemma 2. Let D be a -connected digraph, r be a vertex in V (D) and 

[(n 2)]. Then a subdigraph D  with at most 2(n 1)  arcs, is a minimal
-connected spanning subdigraph of D if and only if D  is a union of a collection
I of arc disjoint in-branchings rooted at r, and a collection O of arc disjoint
out-branchings rooted at r such that |A(I) A(O)|  (i.e. they have at least 
common arcs).

By Lemma 2, a minimum connected subgraph of D is I O, where O =

{O1 , O2 , . . . O } is a collection of arc disjoint out-branchings rooted at r, I =
{I1 , I2 , . . . I } is a collection of arc disjoint in-branchings rooted at r, and
A(O) A(I) is maximized. To explain the concept of the algorithm let us assume
that the number of arcs in a minimum connected spanning subdigraph D is
2(n 1)  and let A(D ) = A(O) A(I), where O = {O1 , O2 , . . . O } is a
collection of arc disjoint out-branchings rooted at r and I = {I1 , I2 , . . . I }
is a collection of arc disjoint in-branchings rooted at r. Note that |A(O)
A(I)| = . The rst step of our algorithm is to construct the set A(O) A(I),
and then, given the intersection, we can construct O and I in polynomial time.
Observe that A(O) and A(I) can intersect in at most (n 2) arcs. The main
idea is to enumerate a subset of potential candidates for the intersection, via
dynamic programming. But note that there could be as many as nO(n) such
candidates, and enumerating them all will violate the claimed running time. So
we try a dierent approach. We rst observe that the arcs in a solution, O I,
can be embedded into a linear matroid of rank O(n). Then we prove that, it
is enough to keep a representative family of the partial solutions in the dynamic
programming table. Since, the size of the representative family is bounded by
2O(n) , our algorithm runs in the claimed running time.
Let us delve into the details of the algorithm. Let Dr be the digraph obtained
from D after removing the arcs in OutD (r). Similarly, let Dr+ be the digraph
obtained from D after removing the arcs in InD (r). Observe  that the arc sets of
Dr and Dr+ can be partitioned as follows. A(Dr ) = vV (Dr ) OutDr (v) and

A(Dr+ ) = vV (Dr+ ) InDr+ (v). We construct a pair of matroids corresponding
to each of the in-branching in I and each of the out-branching in O. For
each in-branching Ii I, we have a matroid MiI,1 = (EI,1 i
, II,1
i
) which is a
i
graphic matroid in D and EI,1 is a copy of the arc set of D. And similarly,
for each out-branching Oi O, we have a matroid MiO,1 = (EO,1 i
, IO,1
i
) which
i
is a graphic matroid in D and EO,1 is again a copy of the arc set of D. Note
that the rank of these graphic matroids is n 1. Next, for each Ii , we dene
matroid MiI,2 = (EI,2 i
, II,2
i
) which is a partition matroid where EI,2 i
is a copy

of the arc set of Dr and II,2 = {X | X EI,2 , |X OutDr (v)| 1, for all v
i i

V (Dr )} 4 . Since OutDr (r) = and |V (Dr )| = n, we have that the rank of
4
We slightly abuse notation for the sake of clarity, as strictly speaking X and
OutDG
r (v) are disjoint, since they are subsets of two dierent copies of the arc set.
30 A. Agrawal et al.

these partition matroids, MiI,2 , i [], is n 1. Similarly, for each Oi , we dene

MiO,2 = (EO,2i
, IO,2
i
) as the partition matroid, where EO,2 i
is a copy of the arc
set of Dr and IO,2 = {X | X EO,2 , |X InDr+ (v)| 1, for all v V (Dr+ )}
+ i i

Since InDr+ (r) = and V (Dr+ ) = n, we have that the rank of these partition
matroids, MiO,2 , i [], is n 1. We dene two uniform matroids MI and
MO of rank (n 1), corresponding to I and O, on the ground sets EI and
EO , respectively, where Ei and EO are copies of the arc set of D. We dene
matroid M = (E, I) as the direct sum of MI , MO , MiI,j , MiO,j , for i [] and
 
j {1, 2}. That is, M = i[],j{1,2} (M i
I,j M i
O,j ) MI MO Since the
rank of MiI,j , MiO,j where i [] and j {1, 2}, are n 1 each, and rank of MI
and MO is (n 1), we have that the rank of M is 6(n 1). We briey discuss
the representation of these matroids. The matroids MiI,1 , MiO,1 for i [] are
graphic matroids, which are representable over any eld of size at least 2. The
matroids MiI,2 , MiO,1 are partition matroids with partition size 1, and therefore
they are representable over any eld of at least 2 as well. Finally, the two uniform
matroids, MI and MO , are representable over any eld with at least |A(D)| + 1
elements. Hence, at the start of our algorithm, we choose a representation of all
these matroids over a eld F of size at least |A(D)| + 1. So M is representable
over any eld of size at least |A(D)| + 1 (see [8, 10]).
For an arc e A(D) not incident to r there are 4 + 2 copies of it in M.
Let eiJ,j denotes its copy in EJ,j i
, where i [], j {1, 2} and J {I, O}.
An arc incident to r has only 3 + 2 copies in M. For an arc e InD (r) we
i i i
will denote its copies in EI,1 , EO,1 , EI,2 by eiI,1 , eiO,1 , eiI,2 , and similarly for an
arc e OutD (r) we will denote its copies in EI,1 i i
, EO,1 i
, EO,2 by eiI,1 , eiO,1 , eiO,2 .
And nally, for any arc e A(D), let eI and eO denote its copies in EI and
EO , respectively. For e A(D) \ OutD (r) and i [], let SI,e i
= {eiI,1 , eiI,2 }.
Similarly for e A(D) \ InD (r), i [], let SO,e = {eO,1 , eO,2 }. Let Se =
i i i
 j 
(i=1 SI,e
i
) (j=1 SO,e ) {eI , eO }. For X I, let AX denote the set of arcs
e A(D) such that Se X = .

Observation 1 (i) Let I be an in-branching in D rooted at r. Then for any

i [], {eiI,1 | e A(I)} is a basis in MiI,1 and {eiI,2 | e A(I)} is a basis
in MiI,2 . And conversely, let X and Y be basis of MiI,1 and MiI,2 , respectively,
such that AX = AY . Then AX is an in-branching rooted at r in D.
(ii) Similarly, let O be an out-branching in D. Then for any i [], {eiO,1 | e
A(O)} is a basis in MiO,1 and {eiO,2 | e A(O)} is a basis in MiO,2 . And
conversely, let X and Y be basis of MiO,1 and MiO,2 , respectively, such that
AX = AY . Then AX is an out-branching rooted at r in D.

Observe that any arc e A(D) can belong to at most one in-branching in I
and at most one out-branching in O, because both I and O are collection of arc
disjoint subgraphs of D. Because of Observation 1, if we consider that each Ii I
is embedded into MiI,1 and MiI,2 and each Oi O is embedded into MiO,1 and
MiO,2 , then we obtain an independent set Z  of rank 4(n 1) corresponding
Fast Exact Algorithms for Survivable Network Design with Uniform Requirements 31

to I O in the matroid M. Further, since the collection I is arc disjoint, {eI |

e A(I)} is a basis of MI . And similarly, {eO | e A(O)} is a basis of MO .
Therefore, Z = Z  {eI | e A(I)} {eO | e A(O)} is a basis of M. Now
observe that, each arc in the intersection I O has six copies in the independent
set Z. The remaining arcs in I O have only three copies each, and this includes
any arc which is incident on r. Now, we dene a function : I A(D) {0, 1},
where for W I and e A(D), (W, e) = 1 if and only if exactly one of the
following holds. Either, W Se = . Or, {eI , eO } W and there exists t, t []
t
t
such that SI,e W and SO,e W . And for each i [] \ {t} and j [] \ {t },
j
i
SI,e W = and SO,e W = . Using function we dene the following collection
of independent sets of M. B 6 = {W  | W I, |W | = 6, e A(D) (W, e) = 1}
By the denitions of , I and O, eA(O)A(I) Se is an independent set of M,
which is contained in B 6 . In fact, for the optimal value of , the collection B 6
contains all possible candidates for the intersection of O and I , where O and
I are collections of arc disjoint in-branchings and arc disjoint out-branchings
which form an optimum solution. Our goal is to nd one such candidate partial
solution from B 6 . We are now ready to state the following lemma which shows
that a representative family of B 6 always contains a candidate partial solution
which can be extended to a complete solution.

Lemma 3. Let D be a -connected digraph on n vertices, r V (D) and 

[(n 2)]. There exists a -connected spanning subdigraph D  of D with at
most 2(n 1)  arcs if and only if, there exists T B 6 nrep6 B 6 , where



n = 6(n 1), such that D has arc disjoint in-branchings containing AT and
arc disjoint out-branchings containing AT , which are all rooted at r.

Proof. In the forward direction consider a -connected spanning subdigraph D

of D with at most 2(n 1)  arcs. By Lemma 2, D is union of a collection
I = {I1 , I2 , . . . , I } of arc disjoint in-branchings rooted at r, and a collection
O = {O1 , O2 , . . . , O } of arc disjoint out-branchings rooted at r such that |A(I)
A(O)| . By Observation 1, for all i [], {eiI,1 | e A(Ii )} is a basis in MiI,1
and {eiI,2 | e A(Ii )} is a basis in MiI,2 . Similarly, by Observation 1, for all
i [], {eiO,1 | e A(Oi )} is a basis in MiO,1 and {eiO,2 | e A(Oi )} is a basis in
MiO,2 . Further {eI | e A(I)} and {eO | e A(O)} are bases of MI and MO ,
respectively. Hence the set ZD = {eiI,1 , eiI,2 | e A(Ii ), i []} {eiO,1 , eiO,2 |
e A(Oi ), i []} {eI | e A(I)} {eO | e A(O)} is an independent
set in M. Since |ZD | = 6(n 1), ZD is actually a basis in M. Consider
T A(I)A(O) with exactly  arcs. Let T  = {eiI,1 , eiI,2 | e T Ii , for some i
[]} {eiO,1 , eiO,2 | e T Oi , for some i []} {eI , eO | e T }. Note that T 
is a set of six copies of the  arcs that are common to a pair of an in-branching in
I and an out-branching in O. Therefore, by the denition of B 6 , T  B 6 . Then,
by the denition of representative family, there exists T B 6 nrep6 B 6 , such


that Z = (ZD \ T  ) T is an independent set in M. Note that |Z| = 6(n 1),

and hence it is a basis in M. Also note that AT AZ .
32 A. Agrawal et al.

Claim. 1 () (i) For any i [] and e A(D), either {eiI,1 , eiI,2 } Z or

{eiI,1 , eiI,2 } Z = . And further for every e A(D) such that eiI,1 Z for
some i [], Z also contains eI . Similarly, for any i [] and e A(D), either
{eiO,1 , eiO,2 } Z or {eiO,1 , eiO,2 } Z = , and further, for every e A(D) such
that eiO,1 Z for some i [], Z also contains eO .
(ii) And for any i, j [], i = j, either {ei , ei } Z = or {ej , ej }
I,1 I,2 I,1 I,2
Z = . Similarly, for any i, j [], i = j, either {eiO,1 , eiO,2 } Z = or
{eiO,1 , eiO,2 } Z = .

Since Z is a basis in M, for any i [], j {1, 2} and k {I, O}, we have
that Z Ek,j
i i = Z E i
is a basis in Mik,j (see [8, 10]). For each i [], let X 1 I,1
and X i = Z E i . By Claim 1, A  i = A  i and hence, by Observation 1,
2 I,2 X1 X2
I i = A  i forms an in-branching rooted at r. Because of Claim 1, {I i | i } are
X1
pairwise arc disjoint as I i I j = for every i = j []. Further AT is covered
in arc disjoint in-branchings {AIi,1 | i }, as T EI,j
i
X i for j {1, 2}. By
j
similar arguments we can show that there exist a collection {O i | i []} of
out-branchings rooted at r containing AT . The reverse direction of the lemma
follows from Lemma 2. 

Lemma 4. Let D be a connected digraph on n vertices and  [(n 2)].


In time 2O(n) we can compute B 6 nrep6 B 6 such that |B 6 | n6 . Here


n = 6(n 1).
Proof. We give an algorithm via dynamic programming. Let D be an array of size
 + 1. For i {0, 1, . . . , } the entry D[i] will store the family B 6i nrep6i B 6i . We


will ll the entries in array D according to the increasing order of index i, i.e. from
0, 1, . . . , . For i = 0, we have B 0 = {}. Let W = {{eI , eO , eiI,1 , eiI,2 , ejO,1 , ejO,2 } |
i, j [], e A(D)} and note that |W| = 2 m, where m = |A(D)|. Given that
we have lled all the entries D[i ], where i < i + 1, we ll the entry D[i + 1] at
step i + 1 as follows. Let F 6(i+1) = (B 6i W) I.
n 6(i+1)
Claim. 2 () F 6(i+1) rep B 6(i+1) , for all i {0, 1, . . .  1}
Now the entry for D[i+1] is F 6(i+1) which is n 6(i+1) representative family
for F 6(i+1) , and it is computed as follows. By Theorem 4 we have that |B 6i |

n   6(i+1)


6i , Hence it follows that |F | 2 m n6i and moreover, we can compute

 n 6(i+1)
F 6(i+1) in time O(2 mn n6i ). We use Theorem 4 to compute F 6(i+1) rep

n 
n 
F 6(i+1) of size at most 6(i+1) . This step can be done in time O( 6(i+1) tp +

n 1

t 6(i+1) ), where t = |F 6(i+1) | = 2 m 6i . We know from Claim 2 that
n

n 6(i+1)
F 6(i+1) rep B 6(i+1) . Therefore by the transitive property of representative
n 6(i+1) 6(i+1)
sets [8, 10], we have B 6(i+1) = F 6(i+1) rep B . Finally, we assign the
family B 6(i+1)
to D[i + 1]. This completes the description of the algorithm and
Fast Exact Algorithms for Survivable Network Design with Uniform Requirements 33

its correctness. Now, since  n /6, we can bound the total running time of this

 
n 
n 1  2
n 
algorithm as O i=1 i 6(i+1) + 6(i+1) m 6i 2O(n) . 


We have the following algorithm for computing I and O given A(I) A(O).
This algorithm extends a given set of arcs to an minimum weight collection of
arc disjoint out-branchings. This is a simple corollary of [24, Theorem 53.10]
and it also follows from the results of Gabow [13].

Lemma 5 (). Let D be a digraph and w be a weight function on the arcs.

For any subset X of arcs of D, a vertex r and an integer , we can nd a
minimum weight collection O of arc disjoint out-branchings rooted at r, such
that X A(O), if it exists, in polynomial time.

Theorem 5. Let D be a edge connected digraph on n vertices. Then we can

nd a minimum edge connected subgraph of D in 2O(n) time.

Proof. Let n = 6(n 1). We x an arbitrary r V (D) and for each choice
of , the cardinality of |A(I) A(O)|, we attempt to construct a solution. By
Lemma 3 we know that there exists a -connected spanning subdigraph D of
D with at most 2(n 1)  arcs if and only if there exists T B 6 nrep6 B 6 ,



where n = 6(n 1), such that D has a collection I = {I1 , I2 , . . . , I } of arc
disjoint in-branchings rooted at r and a collection O = {O1 , O2 , . . . , O } of arc
disjoint out-branchings rooted at r such that AT A(I) A(O). Using Lemma 4
we compute B 6l nrep6 B 6 in time 2O(n) , and for every F B 6 we check if


AF can be extended to a collection of arc disjoint out-branchings rooted at

r and a collection of arc disjoint in-branchings rooted at r, using Lemma 5.
Since  (n 2), the running time of the algorithm is bounded by 2O(n) .  

An similar algorithm can be obtained for the weighted version of the problem
using the notion of weighted representative sets, thus proving Theorem 1.

3 Undirected Graphs

In this section, we give an algorithm for computing a minimum -connected sub-

graph of an undirected graph G. As before, we only consider the unweighted ver-
sion of the problem. While there is no equivalent characterization of -connected
graphs as there was in the case of digraphs, we show that we can obtain a charac-
terization by converting the graph to a digraph with labels on the arcs. Then, as
in the previous section, we embed the solutions in a linear matroid and compute
them by a dynamic programming algorithm with representative families. Let
DG be the digraph with V (DG ) = V (G) and for each edge e = (u, v) E(G),
we have two arcs ae = (u, v) and ae = (v, u) in A(DG ). We label the arcs ae
and ae by the edge e, which is called the type of these arcs. For X A(DG ) let
Typ(X) = {e E(G) | ae X or ae X}. The following two lemmata relate
-connected subgraphs of G with collections of out-branchings in DG .
34 A. Agrawal et al.

Lemma 6 (). Let G be an undirected graph and DG be the digraph constructed

from G as described above. Then G is -connected if and only if for any r
V (DG ), there are arc disjoint out-branchings rooted at r in DG .
By Lemma 6 we know that G is -connected if and only if for any r V (D),
there is a collection O of arc disjoint out-branchings rooted at r in DG . Given
a collection of out-branchings, we can obtain a -connected subgraph of G with
at most (n 1) edges. For an edge e E(G) which is not incident on r, the
two arcs corresponding to it in DG may appear in two distinct out-branchings of
O, but for an edge e incident on r in G, only the corresponding outgoing arc of
r may appear in O. Since there are (n 1) arcs in total that appear in O and
at least of those are incident on r, the number of edges of G such that both
the arcs corresponding to it appear in O is upper bounded by (n2) 2 . So any
(n2)
minimal -connected subgraph of G has (n 1)  edges where  [ 2 ].
Lemma 7 (). Let G be an undirected -connected graph on n vertices and
 [ (n2)
2 ]. G has a -connected subgraph G with at most (n 1) 
edges if and only if for any r V (DG ), DG has arc disjoint out-branchings
O = {O1 , O2 , . . . , O } rooted at r such that |Typ(A(O))| (n 1) .
By Lemma 7, a collection O of out-branchings rooted at some vertex r,
that minimizes |Typ(A(O))| corresponds to a minimum -connected subgraph
of G. In the rest of this section, we design an algorithm that nds a collection
of arc disjoint out-branchings O in DG such that |Typ(A(O))| is minimized.
The rst step of our algorithm is to compute the set of edges of G such that
both the arcs corresponding to it appear in the collection O, and then we can
r
extend this to a full solution in polynomial time. Fix a vertex r. Let DG denote
the digraph obtained from DG by removing the arcs in InDG (r). Observe that
r r
A(DG ) can be partitioned as follows. A(DG ) = vV (Dr ) InDGr (v) We construct
G
a pair of a graphic matroid and a partition matroid, corresponding to each of
the out-branching that we want to nd. For each i [], we dene a matroid
Mi1 = (Ai1 , I1i ) which is a graphic matroid of DG r
whose ground set Ai1 is a copy
of the arc set A(DG ). Similarly, for each i [] we dene matroid Mi2 = (Ai2 , I2i ),
r

which is a partition matroid on the ground set Ai2 , which is a copy of the arc
set A(DG r
), such that the following holds. I2i = {I | I Ai2 , |I InDG r (v)|

1, for all v V (DG r

)} Next, let MO be a uniform matroid of rank (n 1)
r
on the ground set AO where AO is also a copy of A(DG ). Finally, we dene
the matroid  M = (A M , I) as the
 direct sum of M O and M 1 , M2 , for i [],
i i

i.e. M = i[] (M1 M2 ) MO . Note that the rank of this matroid is

i i

3(n 1) and it is representable over any eld of size at least |A(DG r

)| + 1. For
an arc a A(DG ), we denote its copies in A1 , A2 and AO by a1 , ai2 and aO
r i i i

respectively. For a collection O of out-branchings in DG r

, by A(O) we denote
the set of arcs which is present in some  O O. For X I, by AX we denote the
set of arcs a A(DG r
) such that X i=1 {ai1 , ai2 } = . For e E(G) and i [],

 
we let Sei = {(ae )i1 , (ae )i2 , (ae )i1 , (ae )i2 } and Se = {(ae )O , (ae )O } i
i=1 Se . We
dene a function : I E(G) {0, 1}, where for W I, e E(G), (W, e) = 1
Fast Exact Algorithms for Survivable Network Design with Uniform Requirements 35

if and only if exactly one of the following holds. Either W Se = ; or, there exists
t, t [], t = t , such that, (i) (ae )O , (ae )O W , (ii) Set W = {(ae )t1 , (ae )t2 },
  
(iii) Set W = {(ae )t1 , (ae )t2 }, and (iv) i [] \ {t, t }, Sei W = . Now
for each  [(n 2)/2], we dene the following set. B 6 = {W | W
I, |W | = 6 and e E(G), (W, e) = 1} Observe that for every W B 6 ,
|Typ(AW )| =  and, ae AW if and only if ae AW . Therefore, any set in
this collection corresponds to a potential candidate for the subset of arcs which
appear in exactly two out-branchings in O. The following lemma, relates the
computation of out-branchings minimizing types and representative sets.

Lemma 8 (). Let G be a -connected undirected graph on n vertices, DG its

corresponding digraph and  [ (n2)
2 ]. Let n = 3(n 1). Then there exists
a set O of out-branchings rooted at r, with |Typ(A(O))| (n 1)  in DG if
of
and only if there exists T B 6 nrep6 B 6 , such that DG has a collection O


out-branchings rooted at r, AT A(O) and |Typ(O)| (n 1) . Further,


we can compute B 6 n 6 B 6 such that |B 6 | n in time 2O(n) .

rep 6

Finally, Lemmata 5, 7 and 8 give us the following theorem. As before, this

theorem can be extended to prove Theorem 2.
Theorem 6 (). Let G be a edge connected graph on n vertices. Then we can
nd a minimum edge connected subgraph of G in 2O(n) time.

4 Augmentation Problems
The algorithms for Minimum Weight -connected Spanning Subgraph
may be used to solve instances of Minimum Weight -connectivity Aug-
mentation as well. Given an instance (D, L, w, ) of the augmentation problem,
we construct an instance (D , w , ) of Minimum Weight -connected Span-
ning Subgraph, where D = D L and w is a weight function that gives a
weight 0 to arcs in A(D) and it is w for the arcs from L. It is easy to see that
the solution returned by our algorithm contains a minimum weight augmenting
set. A similar approach works for undirected graphs as well, proving Theorem 3.

References
1. Bang-Jensen, J., Gutin, G.Z.: Digraphs: theory, algorithms and applications.
Springer Science & Business Media (2008)
2. Basavaraju, M., Fomin, F.V., Golovach, P., Misra, P., Ramanujan, M., Saurabh,
S.: Parameterized algorithms to preserve connectivity. In: Automata, Languages,
and Programming, pp. 800811. Springer (2014)
3. Bellman, R.: Dynamic programming treatment of the travelling salesman problem.
Journal of the ACM (JACM) 9(1), 6163 (1962)
4. Berman, P., Dasgupta, B., Karpinski, M.: Approximating transitive reductions
for directed networks. In: Proceedings of the 11th International Symposium on
Algorithms and Data Structures. pp. 7485. Springer-Verlag (2009)
36 A. Agrawal et al.

5. Bjorklund, A.: Determinant sums for undirected hamiltonicity. SIAM Journal on

Computing 43(1), 280299 (2014)
6. Cheriyan, J., Thurimella, R.: Approximating minimum-size k-connected spanning
subgraphs via matching. SIAM Journal on Computing 30(2), 528560 (2000)
7. Cygan, M., Fomin, F.V., Golovnev, A., Kulikov, A.S., Mihajlin, I., Pachocki, J.,
Socala, A.: Tight bounds for graph homomorphism and subgraph isomorphism. In:
Proceedings of the Twenty-Seventh Annual ACM-SIAM Symposium on Discrete
Algorithms, SODA 2016, Arlington, VA, USA, January 10-12, 2016. pp. 16431649
(2016)
8. Cygan, M., Fomin, F.V., Kowalik, L., Lokshtanov, D., Marx, D., Pilipczuk, M.,
Pilipczuk, M., Saurabh, S.: Parameterized Algorithms. Springer (2015)
9. Cygan, M., Kratsch, S., Nederlof, J.: Fast hamiltonicity checking via bases of per-
fect matchings. In: Proceedings of the forty-fth annual ACM Symposium on The-
ory of Computing. pp. 301310. ACM (2013)
10. Fomin, F.V., Lokshtanov, D., Panolan, F., Saurabh, S.: Ecient computation of
representative families with applications in parameterized and exact algorithms. J.
ACM 63(4), 29:129:60 (Sep 2016), http://doi.acm.org/10.1145/2886094
11. Frank, A.: Augmenting graphs to meet edge-connectivity requirements. SIAM
Journal on Discrete Mathematics 5(1), 2553 (1992)
12. Frank, H., Chou, W.: Connectivity considerations in the design of survivable net-
works. Circuit Theory, IEEE Transactions on 17(4), 486490 (1970)
13. Gabow, H.N.: A matroid approach to nding edge connectivity and packing ar-
borescences. Journal of Computer and System Sciences 50(2), 259273 (1995)
14. Guseld, D.: A graph theoretic approach to statistical data security. SIAM Journal
on Computing 17(3), 552571 (1988)
15. Held, M., Karp, R.M.: A dynamic programming approach to sequencing problems.
Journal of the Society for Industrial and Applied Mathematics 10(1), 196210
(1962)
16. Jain, S., Gopal, K.: On network augmentation. Reliability, IEEE Transactions on
35(5), 541543 (1986)
17. Kao, M.Y.: Data security equals graph connectivity. SIAM Journal on Discrete
Mathematics 9(1), 87100 (1996)
18. Khuller, S.: Approximation algorithms for nding highly connected subgraphs.
Vertex 2, 2 (1997)
19. Khuller, S., Vishkin, U.: Biconnectivity approximations and graph carvings. Jour-
nal of the ACM (JACM) 41(2), 214235 (1994)
20. Kortsarz, G., Nutov, Z.: Approximating minimum cost connectivity problems. In:
Dagstuhl Seminar Proceedings. Schloss Dagstuhl-Leibniz-Zentrum fur Informatik
(2010)
21. Marx, D.: A parameterized view on matroid optimization problems. Theor. Com-
put. Sci. 410(44), 44714479 (2009), http://dx.doi.org/10.1016/j.tcs.2009.07.027
22. Marx, D., Vegh, L.A.: Fixed-parameter algorithms for minimum-cost edge-
connectivity augmentation. ACM Transactions on Algorithms (TALG) 11(4), 27
(2015)
23. Moyles, D.M., Thompson, G.L.: An algorithm for nding a minimum equivalent
graph of a digraph. Journal of the ACM (JACM) 16(3), 455460 (1969)
24. Schrijver, A.: Combinatorial optimization: polyhedra and eciency, vol. 24.
Springer Science & Business Media (2003)
25. Watanabe, T., Narita, T., Nakamura, A.: 3-edge-connectivity augmentation prob-
lems. In: Circuits and Systems, 1989., IEEE International Symposium on. pp.
335338. IEEE (1989)
 The Complexity of Tree Partitioning

Zhao An1 Qilong Feng1 Iyad Kanj2 Ge Xia3

1
School of Information Science and Engineering, Central South University, China.
[email protected], [email protected]
2
School of Computing, DePaul University, Chicago, IL. [email protected]
3
Dept. of Computer Science, Lafayette College, Easton, PA. [email protected]

Abstract. Given a tree T on n vertices, and k, b, s1 , . . . , sb N, the

Tree Partitioning problem asks if at most k edges can be removed
from T so that the resulting components can be grouped into b groups
such that the number of vertices in group i is si , for i = 1, . . . , b. The
case when s1 = = sb = n/b, referred to as the Balanced Tree
Partitioning problem, was shown to be N P-complete for trees of max-
imum degree at most 5, and the complexity of the problem for trees of
maximum degree 4 and 3 was posed as an open question. The parame-
terized complexity of Balanced Tree Partitioning was also posed as
an open question in another work.
In this paper, we answer both open questions negatively. We show that
Balanced Tree Partitioning (and hence, Tree Partitioning) is
N P-complete for trees of maximum degree 3, thus closing the door on
the complexity of Balanced Tree Partitioning, as the simple case
when T is a path is in P. In terms of the parameterized complexity of
the problems, we show that both Balanced Tree Partitioning and
Tree Partitioning are W [1]-complete. Finally, using a compact repre-
sentation of the solution space for an instance of the problem, we present
a dynamic programming algorithm for Tree Partitioning (and hence,
for Balanced Tree Partitioning) that runs in subexponential-time
2O( n) , adding a natural problem to the list of problems that can be
solved in subexponential time.

1 Introduction
Problem Denition and Motivation. We consider the Tree Partition-
ing problem dened as follows:

Tree Partitioning
Given: A tree T ; k, b, s1 , . . . , sb N
Parameter: k
Question: Does there exist a subset E  E(T ) of at most k edges such that
the components of T E  can be grouped into b groups, where group i contains
si vertices, for i = 1, . . . , b?
The special case of the problem when s1 = = sb = |V (T )|/b is referred to
as Balanced Tree Partitioning.

Springer International Publishing AG 2017 37

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 3748, 2017.
DOI: 10.1007/978-3-319-62127-2_4
38 Z. An et al.

The two problems are special cases of the Balanced Graph Partitioning
problem, which has applications in the areas of parallel computing [3], computer
vision [3], VLSI circuit design [4], route planning [8], and image processing [22].
The special case of Balanced Graph Partitioning, corresponding to b =
2, is the well-known N P-complete problem Bisection [16]. The Balanced
Graph Partitioning problem has received a lot of attention from the area of
approximation theory (for instance, see [2, 12, 21]). Moreover, the complexity
and the approximability of the problem restricted to special graph classes, such
as grids, trees, and bounded degree trees [1214, 20], have been studied.

Our Results. We study the complexity and the parameterized complexity of

Tree Partitioning and Balanced Tree Partitioning, and design subex-
ponential time algorithms for these problems. Our results are:

(A) We prove that Balanced Tree Partitioning, and hence Tree Parti-
tioning, is N P-complete for trees with maximum degree at most 3. This
answers an open question in [13] about the complexity of Balanced Tree
Partitioning for trees of maximum degree 4 and 3, after they had shown
the N P-completeness of the problem for trees of maximum degree at most
5. This also closes the door on the complexity of these problems on trees, as
the simple case when the tree is a path is in P.
(B) We prove that both Tree Partitioning and Balanced Tree Parti-
tioning are W [1]-hard. This answers an open question in [23]. We also prove
the membership of the problems in the class W [1], using the characterization
of W [1] given by Chen et al. [7].
(C) We present an exact subexponential-time algorithm for Tree Partition-
ing, and hence for Balanced Tree Partitioning, that runs in time
2O( n) , where n is the number of vertices in the tree.

For the lack of space, many details and proofs in this paper have been omitted,
and can found in [1].

Related Work and Our Contributions. Feldmann and Foschini [13] stud-
ied Balanced Tree Partitioning. They showed that the problem is N P-
complete for trees of maximum degree at most 5, and left the question about
the complexity of the problem for maximum degree 4 and 3 open. Whereas the
reduction used in the current paper to prove the N P-hardness of Balanced
Tree Partitioning on trees of maximum degree at most 3 starts from the
same problem (3-Partition) as in [13], and is inspired by their construction,
the reduction in this paper is much more involved in terms of the gadgets em-
ployed and the correctness proofs.
Bevern et al. [23] showed that the parameterized complexity of Balanced
Graph Partitioning is W [1]-hard when parameterized by the combined pa-
rameters (k, ), where k is (an upper bound on) the cut size, and is (an upper
bound on) the number of resulting components after the cut. It was observed
in [23], however, that the employed FPT -reduction yields graphs of unbounded
The Complexity of Tree Partitioning 39

treewidth, which motivated the authors to ask about the parameterized com-
plexity of the problem for graphs of bounded treewidth, and in particular for
trees. We answer their question by showing that the problem is W [1]-complete.
Bevern et al. [23] also showed that Balanced Graph Partitioning is
W [1]-hard on forests by a reduction from the Unary Bin Packing problem,
which was shown to be W [1]-hard in [18]. We note that the disconnectedness of
the forest is crucial to their reduction, as they represent each number x in an
instance of Bin Packing as a separate path of x vertices. For Balanced Tree
Partitioning, in contrast to Unary Bin Packing (and hence, to Balanced
Graph Partitioning on forests), the diculty is not in grouping the compo-
nents into groups (bins) because enumerating all possible distributions of k + 1
components (resulting from cutting k edges) into b k + 1 groups can be done
in FPT -time; the diculty, however, stems from not knowing which tree edges
to cut. The FPT -reduction we use to show the W [1]-hardness is substantially
dierent from both of those in [18, 23], even though we use the idea of non-
averaging sets in our constructiona well-studied notion in the literature (e.g.,
see [5]), which was used for the W [1]-hardness result of Unary Bin Packing
in [18].
Many results in the literature have shown that certain N P-hard graph prob-
lems are solvable in subexponential time. Some of these rely on topological
properties of the underlying graph that guarantee the existence of a balanced
graph-separator of sub-linear size, which can then be exploited in a divide-and-
conquer approach (e.g., see [6, 9]). There are certain problems on restricted
graph classes that resist such approaches due to the the problem specications;
designing subexponential-time algorithms for such problems usually requires ex-
ploiting certain properties of the solution itself, in addition to properties of the
graph class (see [15, 19] for such recent results). In the case of Tree Parti-
tioning and Balanced Tree Partitioning, since every tree has a balanced
separator consisting of a single vertex, yet the two problems remain N P-hard
on trees, clearly a divide-and-conquer approach based solely on balanced sep-
arators does not yield subexponential-time algorithms for these problems. To
design subexponential-time algorithms for them, we rely on the observation that
the number of possible partitions of an integer n N is subexponential in n; this
allows for
a compact representation of all solutions using a solution space of
size 2O( n) , enabling a dynamic programming approach that solves the problems
within the same time upper bound.

Terminologies. We refer the reader to [10, 11] for more information about
graph theory and parameterized complexity.
Let T be a rooted tree. For an edge e = uv in T such that u is the parent of
v, by the subtree of T below e we mean the subtree Tv of T rooted at v. For two
edges e, e in T , e is said to be below e if e in an edge of the subtree of T below
e . A nice binary tree T is dened recursively as follows. If |V (T )| 1 then T
is a nice binary tree. If V (T ) > 1, then T is nice if (1) each of the left-subtree
and right-subtree of T is nice and (2) the sizes of the left-subtree and the right-
subtree dier by at most 1. For any n N, there is a nice binary tree of order
40 Z. An et al.

n. A star S is a tree consisting of a single vertex r, referred to as the root of

the star, attached to degree-1 vertices, referred to each as a star-leaf; we refer
to an edge between r and a leaf in S as a star-edge; we refer to a subtree of S
containing r as a substar of S.
A solution P to an instance (T, k, b, s1 , . . . , sb ) of Tree Partitioning is a
pair (EP , P ), where EP is a set of k edges in T , and P is an assignment that
maps the connected components in T EP into b groups so that the total number
of vertices assigned to group i is si , for i [b]. We call a connected component
in T EP a P -component, and denote by CP the set of all P -components.
By a cut in a tree T we mean the removal of an edge from T . A solution
P = (EP , P ) to (T, k, b, s1 , . . . , sb ) cuts an edge e in T if e EP . Let T  be a
subtree of T such that P cuts at least one edge in T  . By a lowest P -component
in T  we mean a subtree T  below an edge e of T  such that T  is a P -component
(i.e., P does not cut any edge below e in T  ).
The restriction of Tree Partitioning to instances in which s1 = = sb =
|T |/b is denoted Balanced Tree Partitioning; an instance of Balanced
Tree Partitioning is a triplet (T, k, b). The restriction of Tree Partition-
ing and Balanced Tree Partitioning to trees of maximum degree at most 3
are denoted Degree-3 Tree Partitioning and Balanced Degree-3 Tree
Partitioning, respectively. For  1 N, we write [] for the set {1, . . . , }.

2 N P-completeness

In this section, we show that Balanced Degree-3 Tree Partitioning, and

hence Degree-3 Tree Partitioning, is N P-complete. Without loss of gener-
ality, we will consider the version of Balanced Degree-3 Tree Partitioning
in which we ask for a cut of size exactly k, as opposed to at most k; it is easy to
see that the two problems are polynomial-time reducible to one another.
To prove that Balanced Degree-3 Tree Partitioning is N P-hard, we
will show that the strong N P-hard problem 3-Partition [16] is polynomial-
time reducible to it. Our reduction is inspired by the construction of Feldmann
and Foschini [13]. Whereas the construction in [13] uses gadgets such that each
consists of ve chains joined at a vertex, the construction in this paper uses
gadgets consisting of nearly-complete binary trees, that are referred to as nice
binary trees. The idea behind using nice binary trees is that we can combine
them to construct a degree-3 tree in which the cuts must happen at specic
edges in order to produce components of certain sizes.
An instance of the 3-Partition problem consists of an integer s > 0 and
a collection S = a1 , . . . , a3k  of 3k positive integers, where each ai satises
s/4 < ai < s/2, for i [3k]. The problem is to decide whether S can be
partitioned into k groups S1 , . . . , Sk , each of cardinality 3, such that the sum
of the elements in each Si is s, for i [k]. By pre-processing the instance of
3k
3-Partition, we can assume that i=1 ai = k s, and that s is divisible by 4.
For the reduction, we construct a degree-3 tree T as follows. For each ai S,
we create a binary tree Ti , whose left subtree Li is a nice binary tree of size ai ,
The Complexity of Tree Partitioning 41

and whose right subtree Ri is a nice binary tree of size s 2. We denote by Ril
and Rir the left and right subtrees of Ri , respectively. Let H = (p1 , . . . , p3k ) be a
path on 3k vertices. The tree T is constructed by adding an edge between each
pi in H and the root of Ti , for i [3k]. See Figure 1 for illustration. It is clear
from the construction that T is a degree-3 tree of 4k s vertices, since each Ti has
size ai + s 1 and P has 3k vertices. We will show that (S, s) is a yes-instance of
3-Partition if and only if the instance I = (T, 6k 1, b = 4k) is a yes-instance
of Balanced Degree-3 Tree Partitioning.

3k

p1 H pi p3k

Li Ri
 
a1 ai Ril Rir a3k
  
s s s s
2
-1 2
-2 s s 2
-1 2
-2
2
-1 2
-2
T1 Ti T3k
Fig. 1. Illustration of the construction of the tree T .

Note that the size of T is 4k s, and hence, if the vertices in T can be grouped
into 4k groups of equal size, then each group must contain s vertices. From the
aforementioned statement, it follows that at least one cut is required in each tree
Ti because the size of each Ti is ai + s 1 > s.
Suppose that the instance I has a solution P that cuts 6k 1 edges in T .

Lemma 1. For i [3k], Ri is not a P -component in CP and Ti does not contain

a lowest P -component of size less than s/4.

Lemma 2. For i [3k], Li is the only P -component contained in Ti , and the

subtree of T induced by (V (Ti ) V (Li )) {pi } is a P -component of size s.

Proof. Since |Ti | > s, any Ti must contain at least one P -component. Since
CP has 6k P -components, at least one of the 3k Ti s contains at most one P -
component, because otherwise the P -components containing vertices in H are
not accounted for. Therefore, at least one Ti contains exactly one P -component
C, which must be a lowest P -component in Ti . By Lemma 1, C = Ri and
|C| s/4, and hence C cannot be any proper subtree of Li , Ril , or Ril . This
leaves Li , Ril , and Ril as the only possible choices for C.
Suppose that C = Ril . After removing C, the partial-Ti , denoted Ti , has
size s 1 + ai (s/2 1) = s/2 + ai , and contains no P -components. Let D
be the set of vertices that are not in Ti , and are in the same group as Ti .
Observe that for any j = i, j [3k], if a vertex in Lj is in D then all vertices
42 Z. An et al.

in Lj are in D. This is true because, by Lemma 1, all vertices in Lj belong

to the same P -component; otherwise Lj would have a lowest P -component of
size less than s/4. This means that the P -component containing Ti has size
|Ti | + |D| s/2 + ai + aj > s. Therefore, D does not include any vertex in Lj .
Similarly, D does not include any vertex in Rjl or Rjr . It follows that D consists
only of vertices in H, the roots of the Ti s, and the roots of the Ri s, i [3k].
However, there are only 9k such vertices, which means |D| 9k and hence the
P -component containing Ti has size |Ti | + |D| s/2 + ai + 9k < s. The last
inequality is true because ai s/2 18k, for i [k].
Therefore, C = Ril . By a similar argument, C = Rir . It follows that C = Li .
After Li is removed, the resulting partial-Ti along with pi in H induces a subtree
Ci of size exactly s, and hence must be a P -component by itself.
After both Li and Ci are removed, there are (3k 1)-many Ti s and 6k 2 P -
components remaining in T . Thus, there is at least one Tj containing exactly one
P -component. By the same argument above, the only P -component contained
in Tj is Lj . Repeating this argument 3k times in total proves the lemma. 

Lemma 2 shows that, in a solution P of (T, 6k 1, 4k), each of L1 , . . . , L3k is

a P -component, and the remaining part of each of the 3k Ti s is a P -component
whose size is s. Based on this, the theorem below easily follows:

Theorem 1. Balanced Degree-3 Tree Partitioning is N P-complete.

3 W [1]-completeness
To show that Tree Partitioning is W [1]-hard (membership in W [1] is shown
using a characterization of W [1] given in [7]), we give a xed-parameter tractable
reduction (FPT -reduction) from the k-Multi-Colored Clique problem (k-
MCC), which is W [1]-complete ([11]), and is dened as follows: Given a graph
M = (V (M ), E(M )) and a proper k-coloring of the vertices f : V (M ) C,
where C = {1, 2, ..., k} and each color class has the same cardinality, decide
whether there exists a clique Q V (M ) of size k such that, u, v Q, f (u) =
f (v). For i [k], we dene Ci = {v M | f (v) = i} to be the color class
consisting of all vertices whose color is i. Let n = |Ci |, i [k], and let N = k n.
We label the vertices in Ci arbitrarily as v1i , . . . , vni . We rst introduce some
terminologies.
For a nite set X N and  Z+ , we say that X is -non-averaging if for
any  numbers x1 , . . . , x X, and for any number x X, the following holds:
if x1 + + x =  x then x1 = = x = x.
Let X = {x1 , . . . , xn } be a (k 1)-non-averaging set. It is known that we can
construct such a set X such that each element xi X, i [n], is polynomial in n
(for instance, see [5]). Jensen et al. [18] showed that a (k1)-non-averaging set of
cardinality n, in which each number is at most k 2 n2 n4 , can be constructed   in
polynomial time in n; we will assume that X is such a set. Let k  = k + k2 , and
let z = k 2 n5 . Choose 2k numbers b1 , . . . , bk , c1 , . . . , ck N such that bj = k 2j z
for j [k], and cj = k 2(k+j) z for j [k]. Observe that each number in the
The Complexity of Tree Partitioning 43

sequence b1 , . . . , bk , c1 , . . . , ck is equal to the preceding number multiplied by

k 2 , and that the smallest number b1 in this sequence is k 2 z k 4 n5 . For
 
each j, j  [k], j < j  , we choose a number cjj = ck k 2((j1)kj(j1)/2+j j) .
That is, each number in the sequence c21 , . . . , ck1 , c32 , . . . , ck2 , . . . , ckk1 is equal to
the preceding one multiplied by k 2 , and the smallest number c21 in this sequence
is equal to k 2 ck .
We construct a tree T rooted at a vertex r as follows. For a vertex vij ,
i [n], j [k], we correspond a vertex-gadget (for vertex vij ) that is a star Svj
i
with cj (k 1)bj (k 1)xi 1 leaves, and hence with cj (k 1)bj (k 1)xi
vertices; we label the root of the star rvj , and add the edge rrvj to T . See Figure 2
i i

for illustration. For each edge e in M between two vertices vij and vpq , i, p
[n], j, q [k], j < q, we create two stars Sv j and Sv pq , with bj +xi 1 and bq +xp 1
i
leaves, respectively, and of roots rv j and rv pq , respectively. We introduce a star Se
i
with root re and cqj 1 leaves, and connect re to rv j and rv pq to form a tree Te with
i
root re that we call an edge-gadget (for edge e). We connect re to r. See Figure 3
for illustration. Note that the number of vertices in Te that are not in Sv j Sv pq is
i
exactly cqj . Finally, we create k  +1 copies of a star Sf ix consisting of ckk1 +k  +1
many vertices, and connect the root r of T to the root of each of these copies.
This completes the construction of T . Let t = |T |. We dene the reduction from
k-Multi-Colored Clique to Tree Partitioning to be the map that takes
an instance I = (M, f) of k-Multi-Colored Clique and produces the instance
I  = (T, k  , b = k + k2 , c1 , . . . , ck , c21 , . . . , ck1 , c32 , . . . , ck2 . . . , ckk1 , t ), where k  =
  k 
k + 3 k2 and t = t j=1 cj j,q[k],j<q cqj . Clearly, this reduction is an
FPT -reduction. Next, we describe the intuition behind this reduction.

r
 Se

re

 
Svj
i
Sv j
i
Sv q
p 
rv j rv q
i p cqj 1


  
 

cj (k 1)bj (k 1)xi 1 bj + xi 1 bq + xp 1

Fig. 2. Illustration of the vertex- Fig. 3. Illustration of the edge-gadget for e =

gadget for vij . vij vpq .
44 Z. An et al.

Each number cj , j [k], chosen above, will serve as a signature for class
Cj , in the sense that it will ensure that in any solution to the instance, a vertex-
gadget corresponding to a vertex in class Cj is cut and placed in the group

of size cj . Each number cjj , j, j  [k], j < j  , will serve as a signature for
the class-pair (Cj , Cj  ), in the sense it will ensure that in a solution exactly one
edge-gadget corresponding to an edge e between classes Cj and Cj is cut and

the star Se is placed in the group whose size is cjj . Each number bj , j [k], will
serve as a signature for any edge such that one of its endpoints is in Cj (i.e.,
a signature for an arbitrary vertex in Cj ), ensuring that in a solution, k 1 of
these edges are cut. Finally, the choice of the xi s, for i [n], to be elements of
a (k 1)-non-averaging set, will ensure that all the edges cut that are incident
to vertices in the same class Cj , j [k], are incident to the same vertex in Cj .
Next, we prove the correctness of the reduction. One direction is easy:
Lemma 3. If (M, f ) is a yes-instance of k-Multi-Colored Clique then I 
is a yes-instance of Tree Partitioning.
To prove the converse,
  let P = (EP , P ) be a solution to the instance
I  = (T, k  , b = k + k2 , c1 , . . . , ck , c21 , . . . , ck1 , c32 , . . . , ck2 . . . , ckk1 , t ) of Tree Par-
titioning. Let Gj , j [k], denote the group of size cj , Gqj , j, q [k], j < q,
denote the group of size cqj , and Grest denote the group of size t . We have:
 
Lemma 4. There is a solution P that cuts exactly k  = k + 3 k2 edges from T
as follows. For each j [k], P cuts exactly one edge between the root r of T and
the root of a vertex-gadget corresponding to a vertex in color class Cj ; moreover,
P assigns the resulting vertex-gadget to group Gj . For each j, q [k], j < q, P
cuts exactly 3 edges from one edge-gadget Te , corresponding to an edge e between
a vertex vij , i [n], in color classes Cj , and a vertex vpq , p [n], in color class
Cq ; those 3 edges are the edges rre , re rv j , and re rv pq , where re is the root of star
i
Se in Te , and rv j , rv pq are the roots of stars Sv j , Sv pq in Te , respectively; moreover,
i i
P assigns Se to group Gqj .
Lemma 5. If I  is a yes-instance of Tree Partitioning then (M, f ) is a yes-
instance of k-MCC.

kBy Lemma 4, we can assume that I has a solution P = (EP , P ) that cuts
Proof.
k+3 2 edges, and that satises the properties in the lemma. Let rrvi1 , . . . , rrvik ,
1 k
i1 , . . . , ik [n], be the edges between the root r of T and the roots of the
vertex-gadgets Svi1 , . . . , Svik that P cuts. We claim that the set of vertices Q =
1 1
{vi11 , . . . , vikk } induce a multi-colored clique in M . To show that, it suces to show
 
that each of the k2 edges rre cut by P , between r and the root of an edge-gadget
Te , where e = vij vpq , i, p [n], p, q [k], p < q, satises that vij , vpq Q.
Consider an arbitrary group Gj , j [k]. The size of Gj is cj , and by Lemma 4,
P assigns the star Svj of size cj (k 1)bj (k 1)xij to Gj . Each star Se is
ij

assigned to some group Gqp whose size is exactly |Se |. Therefore, each group Gj
The Complexity of Tree Partitioning 45

contains a vertex-gadget and some of the stars S  j , i [n], j  [k]. Observe

vi
that group Gj , j < k, cannot contain a star S  j such that j  > j because the
vi
size of such a star is at least bj  > k 2 bj , and hence the size of such a star plus
the size of Svj would exceed the size of Gj . Since there are exactly k 1 stars,
ij

of the form Sv k contained in edge-gadgets corresponding to edges incident to

class Ck , it follows that all these stars must be assigned by P to group Gk .
Moreover, no other star Sv j , j < k, can be assigned to Gk , as the size of such a

star would be at least b1 > (k 1)xi for any i [n]; hence, Gk would contain
vertex gadget Svik of size ck (k 1)bk (k 1)xik , plus k 1 stars Sv k of total
k
size greater than (k 1)bk , plus a star of size at least b1 > (k 1)xik , and the
size of Gk would exceed ck .
Similarly, all the k 1 stars of the form Sv k1 contained in edge-gadgets

corresponding to edges incident to class Ck1 are assigned to group Gk1 , and
following this argument, we obtain that for each j [k], the (k 1) stars of
the form Sv j must be assigned to group Gj . We claim that all these stars must

correspond to the same vertex vijj . Observe that this will prove that Q is a clique,
 
since it will imply that each vertex in Q is incident to exactly k 1 of the k2
many edges between the color classes.
Let Sv j , . . . , Sv j be the k 1 stars placed in Gj . The sizes of these stars
i i
1 k1
are bj + xi1 , . . . , bj + xik1 , respectively. The size cj of Gj is equal to the sum of
the sizes of these k 1 stars, plus that of Svj . Therefore: cj = cj (k 1)b
ij

(k 1)xij + (k 1)b + xi1 + + xik1 , and hence, (k 1) xij = xi1 + + xik1 .

Since the set X is (k 1)-non-averaging, it follows that xij = xi1 = = xik1 ,
and hence, the (k 1) stars Sv j must correspond to vertex vijj . 

Theorem 2. Tree Partitioning and Balanced Tree Partitioning are

W [1]-complete.

4 Subexponential-time Algorithms
+
 n Z . A partition of n is a collection X of positive integers such that
Let
xX x = n. Let p(n) denote the total number of (distinct) partitions of n.

It is well known that p(n) = 2O( n) [17]. It follows  that the total number
of
partitions of all integers n , where 0 < n n, is 0<n n p(n ) = 2O( n) .
Let L be a list of numbers in N that are not necessarily distinct. We denote
by L(i) the ith number in L, and by Li the sublist of L consisting of the rst i
numbers. The length of L, denoted |L|, is the number of elements in L.
Let (T, k, b, s1 , . . . , sb ) be an instance of Tree Partitioning. Let n = |T |.
Consider a partial assignment of n n vertices of T to the b groups, with the
possibility of some groups being empty. Since the groups are indistinguishable,
such an assignment corresponds to a partition of the n vertices into at most b
46 Z. An et al.

parts, and can be represented by a sorted list L of b numbers in N whose sum

is n , where L(i) n for i [b], is the number of vertices assigned to group
i; we call such a representation of the groups, under a partial assignment, a
size representation, denoted as -representation. Note that the zeroes in a -
representation appear at the beginning. Since each -representation corresponds
uniquely to a partition of a number n n prexed by less than b nzeroes, it
follows that the total number of -representations is n 2O( n) = 2O( n) .
Let X, Y, Z be three lists of the same length. We write X = Y Z if there is
a list Y  obtained via a permutation of the numbers in Y , and a list Z  obtained
via a permutation of the numbers in Z, such that X(i) = Y  (i) + Z  (i), for every
i [|X|]; that is, in the context when the lists are -representations, X = Y Z
if each group-size in X can be obtained, in a one-to-one fashion, by adding a
group-size in Y to a group-size in Z (including group-sizes zero). We have:

Proposition 1. There is a subroutine

Check-Realizability(X, Y, Z) that de-
termines if X = Y Z in time 2O( n) .

Let (T, k, b, s1 , . . . , sb ) be an instance of Tree Partitioning. The key ob-

servation that leads to a subexponential-time algorithm is that the b groups are
indistinguishable. Therefore, all assignments of the n vertices in T to the b groups
can be compactly represented by lists of numbers, where each list corresponds
to a partition of n into b parts. This simple, yet crucial, observation allows for a
compact representation of all solutions using a solution space of size 2O( n) .
Suppose that T is rooted at an arbitrary vertex r. The algorithm uses dy-
namic programming, starting from the leaves of T , and climbing T up to its root
r. At each vertex v in T , we construct a table v that contains the following
information. For each -representation X, for each k  = 0, . . . , n, and for each
s [n], v (k  , X, s) is true if and only if there is a cut C of k  edges in Tv
such that the component Pv containing v in Tv C has size s (note that this
component, so far, is still attached to the rest of the tree above v), and such
that there is an assignment to the components in Tv C Pv to the b groups
whose -representation is X; otherwise, v (k  , X, s) is false. If v (k  , X, s) is
true, we store a witness that realizes such a partial solution. To compute v ,
we consider the children of v one by one. After a child ui of v is considered, we
have computed a partial table i containing partial solutions up to child ui ; this
is done by considering the two possibilities of whether or not the edge vui is in
the cut C. Although the above may seem like we are enumerating all possibilities
for the edges between v and its children to be cut or not, the crucial ingredient
for this approach to achieve the desired running time is that the table v at
vertex vcan
be computed based on the tables corresponding to the children of
v in 2O( n) time. This analysis works similarly to iterative compression, as the
table i is a compressed table, storing 2O( n) many entries.
Suppose that the algorithm is at vertex v whose children are u1 , . . . , ud ,
and that the tables u1 , . . . , ud associated with u1 , . . . , ud , respectively, have
been constructed. To compute v , we iterate through the edges vu1 , . . . , vud .
Let Tp , for p [d], be the subtree of T rooted at v that is induced by the
The Complexity of Tree Partitioning 47
p
vertex-set ( j=1 V (Tuj )) {v}. Consider edge vui , and assume inductively, that
a table i1 has been computed (based on tables u1 , . . . , ui1 ) that contains
the following information. For each k  = 0, . . . , n, for each s [n], and for each
-representation X, i1 (k  , X, s) is true if and only if there is a cut C of k 
edges in Ti1 , with s being the size of the component Pv containing v in Ti1 C,
and an assignment to the components in Ti1 C Pv that realizes X. After
considering vui , we will compute a table i such that, for each k  = 0, . . . , n, for
each s [n], and for each -representation X, i (k  , X, s) is true if and only
if there is a cut C of k  edges in Ti , with s being the size of the component Pv
containing v in Ti C, and an assignment to the components in Ti C Pv that
realizes X. We explain how the Boolean value i (k  , X, s) is computed, and omit
how the witness is stored. After we are done computing d , we set v = d .
To compute i , we compute two tables i and i+ , and set i = i i .
Table i contains the solutions that can be obtained by cutting edge uvi , and
i+ contains those that can be obtained by not cutting edge vui .
1. To compute i , we enumerate each possible triplet (k  , X, s), where k  =
0, . . . , n, s [n], and X is a -representation. Fix such a triplet (k  , X, s). To com-
pute i (k  , X, s), we iterate through every entry in ui containing (kui , Y, sui )
and every entry of i1 containing (ki1 , Z, si1 ) such that k  = kui + ki1 + 1
(because 1 more cut is introduced, corresponding to the edge vui ), and s = si1
because the component Pui containing ui of size sui becomes a separate compo-
nent after vui is cut. Since Pui becomes a separate component, it will be placed
into one of the groups, and hence, it contributes its size to one of the numbers
in the -representation X. We enumerate each number in X as the number that
Pui contributes to. For each number j in X satisfying j |Pui |, we subtract
|Pui | from j in X to obtain a new -representation X  from X, and then call
Check-Realizability(X  , Y, Z); i (k  , X, s) is true i for some number j in
X, Check-Realizability(X  , Y, Z) returns true.
2. To compute i+ , we enumerate each triplet (k  , X, s), where k  = 0, . . . , n,
s [n], and X is a -representation. Fix such a triplet (k  , X, s). To compute
i+ (k  , X, s), we iterate through every entry in ui containing (kui , Y, sui ), and
every entry in i1 containing (ki1 , Z, si1 ), such that k  = kui + ki1 , and s =
sui +si1 (because sui is attached to v). We call Check-Realizability(X, Y, Z),
and set i+ (k  , X, s) to true i Check-Realizability(X, Y, Z) returns true.

Theorem 3. The dynamic programming algorithm described above

solves Tree
Partitioning and Balanced Tree Partitioning in time 2O( n) .

Bibliography

[1] Z. An, Q. Feng, I. Kanj, and G. Xia. The Complexity of Tree Partitioning.
Available at http://arxiv.org/abs/1704.05896.
[2] K. Andreev and H. Rcke. Balanced graph partitioning. Theory of Com-
puting Systems, 39(6):929939, 2006.
48 Z. An et al.

[3] P. Arbenz, G. van Lenthe, U. Mennel, R. Mller, and M. Sala. Multi-level

-nite element analysis for human bone structures. In PARA 2006, pages
240250, 2006.
[4] S. Bhatt and F. Leighton. A framework for solving VLSI graph layout
problems. Journal of Computer and System Sciences, 28(2):300343, 1984.
[5] . Boscznay. On the lower estimation of non-averaging sets. Acta Mathe-
matica Hungariga, 53(1-1):155157, 1989.
[6] J. Chen, I. Kanj, L. Perkovic, E. Sedgwick, and G. Xia. Genus characterizes
the complexity of certain graph problems: Some tight results. Journal of
Computer and System Sciences, 73(6):892907, 2007.
[7] Y. Chen, J. Flum, and M. Grohe. Machine-based methods in parameterized
complexity theory. Theoretical Computer Science, 339(2-3):167199, 2005.
[8] D. Delling, A. Goldberg, T. Pajor, and R. Werneck. Customizable route
planning. In SEA 2011, pages 376387, 2011.
[9] E. Demaine, F. Fomin, M. Hajiaghayi, and D. Thilikos. Subexponential pa-
rameterized algorithms on bounded-genus graphs and H-minor-free graphs.
J. ACM, 52:866893, 2005.
[10] Reinhard Diestel. Graph Theory, 4th Edition. Springer, 2012.
[11] R. Downey and M. Fellows. Fundamentals of Parameterized Complexity.
Springer, New York, 2013.
[12] A. Feldmann. Balanced partitions of grids and related graphs, 2012. Ph.D.
thesis, ETH, Zurich, Switzerland.
[13] A. Feldmann and L. Foschini. Balanced partitions of trees and applications.
Algorithmica, 71(2):354376, 2015.
[14] A. Feldmann and P. Widmayer. An O(n4 ) time algorithm to compute the
bisection width of solid grid graphs. Algorithmica, 71(1):181200, 2015.
[15] F. Fomin, S. Kolay, D. Lokshtanov, F. Panolan, and S. Saurabh. Subexpo-
nential algorithms for rectilinear steiner tree and arborescence problems. In
SoCG 2016, pages 39:139:15, 2016.
[16] M. Garey and D. Johnson. Computers and Intractability: A Guide to the
Theory of NP-Completeness. W. H. Freeman, New York, 1979.
[17] G. Hardy and S. Ramanujan. Asymptotic formulae in combinatory analysis.
Proceedings of the London Mathematical Society, 17(2):75115, 1918.
[18] K. Jansen, S. Kratsch, D. Marx, and I. Schlotter. Bin packing with
xed number of bins revisited. Journal of Computer and System Sciences,
79(1):3949, 2013.
[19] P. Klein and D. Marx. A subexponential parameterized algorithm for subset
TSP on planar graphs. In SODA 2014, pages 18121830, 2014.
[20] R. MacGregor. On partitioning a graph: a theoretical and empirical study,
1978. Ph.D. thesis, University of California at Berkeley, California, USA.
[21] H. Rcke and R. Stotz. Improved approximation algorithms for balanced
partitioning problems. In STACS 2016, pages 58:158:14, 2016.
[22] J. Shi and J. Malik. Normalized cuts and image segmentation. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence, 22(8):888905, 2000.
[23] R. van Bevern, A. Feldmann, M. Sorge, and O. Such. On the parame-
terized complexity of computing balanced partitions in graphs. Theory of
Computing Systems, 57(1):135, 2015.
 Eectiveness of Local Search for Art Gallery
Problems

Sayan Bandyapadhyay1 and Aniket Basu Roy2

1
Computer Science, University of Iowa, Iowa City, USA
2
Computer Science and Automation, Indian Institute of Science, Bangalore, India
[email protected], [email protected]

Abstract. We study the variant of the art gallery problem where we

are given an orthogonal polygon P (possibly with holes) and we want to
guard it with the minimum number of sliding cameras. A sliding camera
travels back and forth along an orthogonal line segment s in P and a
point p in P is said to be visible to the segment s if the perpendicular
from p onto s lies in P . Our objective is to compute a set containing the
minimum number of sliding cameras (orthogonal segments) such that
every point in P is visible to some sliding camera. We study the following
two variants of this problem: Minimum Sliding Cameras problem, where
the cameras can slide along either horizontal or vertical segments in P ,
and Minimum Horizontal Sliding Cameras problem, where the cameras
are restricted to slide along horizontal segments only. In this work, we
design local search PTASes for these two problems improving over the
existing constant factor approximation algorithms. We note that in the
rst problem, the polygons are not allowed to contain holes. In fact, there
is a family of polygons with holes for which the performance of our local
search algorithm is arbitrarily bad.

1 Introduction

Local search is a popular technique for designing time and cost ecient approx-
imation algorithms. It has a long history in combinatorial optimization and has
proved to be very eective for achieving near-optimum solutions. The use of this
technique in geometric approximation is relatively new, but has resulted in im-
proved approximation for metric and geometric versions of many combinatorial
problems. In fact, this technique has been used in this domain to achieve several
breakthrough results.
In this article, we restrict ourselves to the works based on local search in
geometric approximation. One of the rst results of local search for the problems
in metric space is a 3+ approximation algorithm for k-median due to Arya et al.
[1]. An arguably simplied analysis was later given by Gupta and Tangwongsan
[15]. Building on the work of Arya et al., Kanungo et al. [16] have designed a
9 + approximation algorithm for k-means. In a celebrated work Mustafa and

The author has been supported by NSF under Grant CCF-1615845.

Springer International Publishing AG 2017 49

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 4960, 2017.
DOI: 10.1007/978-3-319-62127-2_5
50 S. Bandyapadhyay and A. Basu Roy

Ray [20] designed the rst PTASes for the Hitting Set problem for a wide class
of geometric range spaces. Around the same time Chan and Har-Peled [7] gave
PTASes for Maximum Independent Set problem of geometric objects including
fat objects and pseudodisks in the plane. Cohen-Addad and Mathieu [8] have
shown the eectiveness of local search for achieving PTASes for facility location
type problems. In a recent breakthrough, Cohen-Addad et al. [9] and Friggstad et
al. [13] have independently designed the rst PTAS for k-means which was a long
standing open problem. Very recently, Govindarajan et al. [14] have obtained the
rst PTASes for the Set Cover and Dominating Set problems for non-piercing
regions. See also [2, 3, 4, 5] for related work on local search.
In this paper, we consider art gallery problems and design PTASes using local
search proving the eectiveness of this technique for a new family of problems.

1.1 Art Gallery Problems

In the classical art gallery problem, we are given a polygon P (possibly with
holes), and the goal is to determine the number of guards (points) needed that
can see all parts of P . A guard can see the portion of the polygon not obstructed
by the polygon boundary, considering the boundary to be opaque. Over the
years dierent variants of the problem have been studied based on the shape
of the polygon, the type of guards employed, and the notion of visibility. The
polygons can be of specic type like orthogonal3 , monotone etc. Guards can be
stationary or mobile. In case of stationary guards, either one can place the guards
anywhere inside the polygon (point guards) or restrict them only to the vertices
of the polygon (vertex guards). The notion of mobile guards was introduced by
Toussaint [21], where each guard can travel to and fro along a segment inside
the polygon, and every point in the polygon must be seen by at least one guard
at some point of time along its path. If the notion of visibility is altered, it gives
rise to a lot of other variants like k-transmitters [22], multi-guarding [18] etc.
One such variant is called sliding cameras, where the polygon is considered to
be orthogonal. A sliding camera is a mobile guard that can move back and forth
along an orthogonal (or axis-parallel) segment. In the following, we formally
dene this notion.
Denition 1 (Sliding Camera). Given an orthogonal polygon P , any orthog-
onal (horizontal or vertical) segment contained in P is called a sliding camera. If
the corresponding segment is horizontal (resp. vertical), then the sliding camera
is horizontal (resp. vertical). A sliding camera s can guard a point p P if there
is a point q s such that the segment pq is perpendicular to s and is contained
in P . A set S of sliding cameras can guard the polygon P , if for every point
p P , there is a sliding camera in S that can guard p.
In this article, we study the following two problems involving sliding cameras.
Denition 2 (Minimum Horizontal Sliding Cameras Problem (MHSC)).
Given an orthogonal polygon P , compute a set S containing the minimum num-
ber of horizontal sliding cameras, that can guard P .
3
A polygon is said to be orthogonal if all of its sides are axis-parallel.
Effectiveness of Local Search for Art Gallery Problems 51

Denition 3 (Minimum Sliding Cameras Problem (MSC)). Given an

orthogonal polygon P , compute a set S containing the minimum number of
sliding cameras, that can guard P .
The MSC and MHSC problems were introduced by Katz and Morgenstern
[17]. They showed that if P is simple (has no holes), MHSC can be solved in
polynomial time. For monotone P , they obtain a polynomial time 2 approxima-
tion for MSC. Later for this special case, de Berg et al. [10] gave a linear time
exact algorithm. MHSC and MSC both are known to be N P -hard for polygons
with holes [6, 11]. Recently, Biedl et al. [6] have given the rst constant factor
approximation algorithms for MHSC and MSC for polygons possibly with holes.
Currently, this is the best known approximation factor for both problems.

1.2 Related Work

The art gallery problem is related to the 1.5D Terrain Guarding problem. In the
latter problem, the input is a terrain T consisting of an x-monotone polygonal
chain, and we need to guard the terrain T by choosing the minimum number of
point guards on it. The discrete version of the problem comes with two nite sets
G, X T and one needs to compute the minimum sized subset G G such that
G guards X. A local search framework similar to the ones in [7, 20] was used by
Krohn et al. [19] to give a PTAS for the discrete version of the problem. Building
on this result, Friedrichs et al. [12] gave a PTAS for the continuous version. Also
note that the collection of problems considered in the local search framework of
[2] includes problems on guarding polygons and terrains with limited visibility
and appropriate shallowness assumption.

1.3 Our Results and Techniques

In this article, we use a standard local search framework to obtain PTASes
for the MHSC and MSC problems. In fact, we give PTASes for certain hitting
set problems involving segments that yield the PTASes for MHSC and MSC.
We note, that we get a PTAS for MHSC even in the case where the polygon
contains holes. However, we get the PTAS for MSC only if the polygon is simple
(has no holes). In fact, one can show that in case of MSC where the polygon
contains holes, the performance of local search can be arbitrarily bad. Our main
contribution is the design of two planar graph embedding schemes which are used
to prove the planarity of certain graphs. In Section 2 we dene some notations
and describe the local search framework that we use. In Section 3 and 4 we use
the framework for MHSC and MSC, respectively, and prove that the framework
yields PTASes for both problems. Due to lack of space, many proofs are omitted
which can be found in the full version of the paper.

2 Preliminaries
Consider the orthogonal polygon P in the MSC or MHSC problem. An orthog-
onal segment s P is called canonical if (i) s is an extension of a side of P ,
52 S. Bandyapadhyay and A. Basu Roy

and (ii) s is maximal in terms of length, i.e., there is no other segment in P that
strictly contains it. We note that the endpoints of a canonical segment lie on the
boundary of P . Also the number of canonical segments is at most the number
of sides of P . It is not hard to see that there exists a canonical segment s for
any segment s P , such that s can guard all the points in P guarded by s.
Henceforth, for any solution S of MHSC or MSC, without loss of generality we
assume that S consists of only canonical segments. Obviously, for MHSC, any
solution consists of only horizontal canonical segments.

2.1 The Local Search Framework

Now we describe a local search framework similar to the ones in [7, 20] for
designing and analyzing local search algorithms. In the following sections we will
apply this framework to get near optimum approximations for MHSC and MSC.
We note that this framework is applicable for discrete optimization problems,
where given a set S, the goal is to nd an optimum subset S  S that is feasible.
Notably, not every subset S  S is a feasible solution. Moreover, there is an
initial feasible solution that can be found in polynomial time, and given any
S  S, one can decide the feasibility of S  in polynomial time. We note that
for MHSC (resp. MSC), S is the nite set of all horizontal (resp. horizontal and
vertical) canonical segments, which is a feasible solution. Let n = |S|. Fix > 0.
As we deal with only minimization problems in this article, the local search
algorithm we consider is the following.

Local Search Algorithm. Choose a parameter k. Start with some feasible

solution A. At each iteration, search for A A and C S \ A such that
|A | k, |C| < |A |, and (A \ A ) C is a feasible solution. If such A and C
exist, update A to (A \ A ) C and reiterate. Otherwise, return A.
As we shoot for a (1 + ) approximation we set k := c/ 2 for some constant
c. From the above discussion, it is not hard to see that the running time of the
algorithm is polynomial. A feasible solution A is said to be local optimum if there
is no A A and C S \ A such that |A | k, |C| < |A |, and (A \ A ) C is
a feasible solution. Note that the local search algorithm always returns a local
optimum solution. Let R and B be an optimum solution and the local search
solution, respectively. For simplicity, we assume that R B = . Otherwise, one
can remove the common elements R B from both R and B, and perform a
similar analysis. As we remove the same number of elements from both R and
B the approximation ratio of the original instance is at most the approximation
ratio of the restricted one. Now we state the conditions required for the local
search algorithm to be a PTAS.

Theorem 1. [7, 20]. Consider a minimization problem . Suppose there exists

a planar bipartite graph H=(R B, E), that satises the local exchange property:
For any subset B B, (B \ B  ) N (B  ) is a feasible solution. Then the Local
Search Algorithm is a PTAS. Here N (B  ) is the set of neighbors of B  in H.
Effectiveness of Local Search for Art Gallery Problems 53

3 The MHSC Problem

Instead of directly working with MHSC we consider the following Orthogonal
Segment Covering (OSC) problem: Given a nite set H of non-intersecting hor-
izontal segments and a nite set V of vertical segments in the plane, nd a
minimum cardinality set S H such that any segment in V is intersected by S.
From the work due to Biedl et al. [6] we have the following theorem.
Theorem 2. [6] MHSC reduces to the Orthogonal Segment Covering problem.
In particular, an approximation for the Orthogonal Segment Covering problem
yields an approximation for MHSC.
We note that the non-intersecting assumption of the horizontal segments
comes from the fact that we consider only the canonical horizontal segments in
MHSC to guard P which are non-intersecting. Biedl et al. [6] gave a constant
approximation for OSC which implies a constant approximation for MHSC. We
use the local search framework to obtain a PTAS for OSC; by Theorem 2 we
obtain a PTAS for MHSC as well.
Let L be the local search solution and G be an optimum solution. We show
that there exists a planar bipartite graph H=(G L, E) that satises the local
exchange property; by Theorem 1 the local search algorithm is a PTAS for
OSC. As mentioned before, we can assume WLOG that L G = . For any
two intersecting vertical segment v and horizontal segment h, let I(h, v) be the
intersection point.
Construction of H. Initially the set of vertices of H consists of the intersection
points of the segments in L G V and the endpoints of the segments in L G.
We rst join the vertices on any horizontal segment using edges. In particular, for
any segment h LG, let {p1 , . . . , pt } be the vertices in H corresponding to h in
increasing order of their x-coordinates. For each 1 i t 1, draw a horizontal
edge between pi and pi+1 . Then we join the consecutive horizontal segments.
Formally, for any two segments h1 L and h2 G, if h1 and h2 intersect a
segment v of V, and h3 L G \ {h1 , h2 } such that I(h3 , v) is on the segment
joining I(h1 , v) and I(h2 , v), draw a vertical edge between I(h1 , v) and I(h2 , v).
For any segment h L G, contract all the vertices in H corresponding to h
into a single vertex. The modied graph is the desired graph H.
It is obvious that before the contractions, H is a planar graph, as it is a
subgraph of the graph corresponding to the arrangement of the segments in
L G V. Now note that the segments in L G are non-intersecting. Thus
there is a 1-1 correspondence between the vertices in nal H and the segments
in L G. Furthermore, any edge in H is between a vertex corresponding to a
segment in L and a vertex corresponding to a segment in G. As planar graphs
are closed under contraction of edges, H is planar.
Observation 4. H = (G L, E) is a planar bipartite graph.
Observation 5. For any segments h1 , h2 , v such that h1 L, h2 G, v V, if
there is an edge in H between I(h1 , v) and I(h2 , v) before the contractions, then
there is an edge corresponding to h1 and h2 in the nal graph H.
54 S. Bandyapadhyay and A. Basu Roy

Lemma 3. H = (G L, E) satises the local exchange property.

Proof. It is sucient to show that for any v V, there exist h1 L and h2 G

such that h1 and h2 intersect v, and there is an edge between h1 and h2 in
H. Consider any v V. Also consider the subset of segments S L G that
intersect v. Then S must contain two segments h1 , h2 such that h1 L, h2 G
and h3 such that I(h3 , v) is on the segment joining I(h1 , v) and I(h2 , v). Thus
there must be an edge in H before the contractions between I(h1 , v) and I(h2 , v).
From Observation 5 it follows that there is an edge in H between h1 and h2 which
completes the proof.

We conclude this section with the following theorem.

Theorem 4. There exists a local search PTAS for the MHSC problem, where
the input polygon may or may not contain holes.

4 The MSC Problem

We use the local search framework to get a PTAS for MSC problem assuming P
does not contain any holes. In fact, one can construct a family of polygons with
holes for which our local search scheme performs very poorly (see full paper).
We start the local search algorithm with the set of canonical segments. Now
consider the local search solution L and an optimum solution R. Then we have
the following lemma which will be useful later.

Lemma 5. Consider the local search solution L and an optimum solution R.

There is a local optimum solution B such that |B| = |L|, and for every point
p P , there exists segments b B and r R both of which guard p and either
of the following is true.

1. b and r are both horizontal or both vertical.

2. One of b and r is horizontal and the other is vertical, and r b = .

Henceforth, we consider the local optimum solution B in Lemma 5 and show

that |B| (1 + )|R|. As |L| = |B| the local search algorithm is a PTAS. Like
before, WLOG we can assume that B R = . We will show the existence of a
bipartite planar graph G(R B) that satises the local exchange property.
Denote the arrangement of a set of segments S by A(S). Consider the edges
and the vertices of A(S). One can visualize these edges and vertices as a plane
graph and dene its connected components accordingly. Now consider any such
component C and let S  be the subset of segments of S corresponding to C. We
dene the closure of C (denoted by closure(C)) as the union of the points on
the segments in S  and the points in the bounded cells of A(S  ). Dene H(S)
as the union of the closures of the components in A(S), i.e., H(S) := CA(S)
closure(C)). Also dene H(S) as the union of the boundaries of the closures in
H(S) (see Figure 1b). Later, we will prove the following theorem.
Effectiveness of Local Search for Art Gallery Problems 55

(a) (b)

Y

(c) (d)

Fig. 1: (a) A connected component of the arrangement. (b) The boundary of the
closure of the component. (c) A horizontal segment X and a vertical segment
Y (bolded). The two slabs generated by taking the Minkowski sum of the two
segments with S , a square of side length . (d) The fattened boundary of the
closure.

Theorem 6. Given a set of canonical horizontal segments X and a set of canon-

ical vertical segments Y, one can construct a planar graph G(X Y) satisfying
the following properties.
1. The vertices in G(X Y) correspond to the segments in X Y.
2. For each Xi , Xj X , such that there exists a vertical segment v contained in
P that intersects Xi and Xj , and Xk X that intersects v and lies between
Xi and Xj , (Xi , Xj ) is in G(X Y).
3. For each Yi , Yj Y, such that there exists a horizontal segment h contained
in P that intersects Yi and Yj , and Yk Y that intersects h and lies between
Yi and Yj , (Yi , Yj ) is in G(X Y).
4. For each Xi X , and Yj Y, such that Xi Yj is on H(X Y), (Xi , Yj )
is in G(X Y).
We construct the planar graph G(R B) by applying Theorem 6 on the
horizontal and vertical segments of RB. We remove all the edges from G(RB)
that joins two segments of R or two segments of B making it bipartite for the
partition (R, B). The following lemma completes our claim that the local search
algorithm is a PTAS. The proof of the lemma partly follows from Lemma 5.
Lemma 7. G(R B) satises the local exchange property.

4.1 Graph Construction Algorithm

In this section, we prove Theorem 6. Given a set of horizontal segments X and a
set of vertical segments Y, we dene a graph G(X Y) on X Y. The construction
56 S. Bandyapadhyay and A. Basu Roy

is a bit involved and hence we describe it in a sequence of steps. After dening

the graph G(X Y) we show a planar embedding of this graph. First we dene
three graphs GX = (X , EX ), GY = (Y, EY ) and GX ,Y = (X Y, EX ,Y ) with
respect to the simple orthogonal polygon P . The edge set of G(X Y) is dened
to be the union of EX , EY , and EX ,Y .
A pair of horizontal segments Xi , Xj X are called consecutive if there
exists a vertical segment v contained in P that joins Xi and Xj , and there does
not exist Xk X that intersects v and lies between Xi and Xj . We add an
edge to EX between each pair of consecutive segments in X . If Xi and Xj are
consecutive but there does not exist a vertical segment in Y intersecting them,
then we call (Xi , Xj ) a soft edge. Likewise, we dene the consecutive segments
in Y, the graph GY = (Y, EY ) and the soft edges of GY . We make the following
claim.

Lemma 8. GX and GY are forests.

We assume that the arrangement A(X Y) of the segments in X Y has only

one connected component and later show how to remove this assumption. Then
for each pair Xi , Xj X , there is a chain of vertical and horizontal segments in
X Y that connects Xi and Xj . Thus by denition GX is also connected and
hence is a tree. Similarly, GY is a tree. We refer to the closure of the component
in A(X Y) by H and its boundary by H. Dene GX ,Y = (X Y, EX ,Y ) to be
the bipartite graph on X and Y, such that (X, Y ) EX ,Y i X Y H = .
We note that, by denition G(X Y) satises the four properties mentioned
in Theorem 6. Thus it is sucient to show that G(X Y) is planar, to which
we turn next. First we use the embedding of X Y to get an intermediate plane
graph H. Then we perform some planarity preserving operations on H to obtain
G(X Y). Hence we get a planar drawing of G(X Y), thus proving G(X Y)
is planar.
To construct the plane graph H we use the given embedding of the segments
in X Y. We consider the endpoints of the segments in X Y and the intersection
points of the segments in X Y to be the vertex points of H. We rst join the
vertices on any horizontal segment using edges. For any segment h X Y, let
{p1 , . . . , pt } be the vertices in H corresponding to h in increasing order of their
x-coordinates. For each 1 i t 1, draw a horizontal edge between pi and
pi+1 . Similarly, we join the vertices on any vertical segment using vertical edges
and call them the edges corresponding to that vertcal segment. It is not hard to
see that H is a plane graph. One can visualize H as a graph where each X X
is being represented by its left endpoint, and each Y Y is being represented by
the segment Y itself. Later we will contract the edges corresponding to Y Y to
represent them using points in the plane. Note that the edges of EY are already
present in H. Thus to get a drawing of G(X Y) we need to draw the edges
of EX and EX ,Y . In the remainder of the section, we will show the planarity
preserving operations on H that leads to the drawing of G(X Y). This will be
done in two stages. In the rst stage, we add some edges to H. In the second
stage, we contract several edges.
Effectiveness of Local Search for Art Gallery Problems 57

First, we x some notations. For every X X , denote its left (resp. right)
endpoint by (X) (resp. r(X)), and the set of vertical segments in Y intersecting
it by vert(X). Also, denote any non-empty intersection point X Y by p(X, Y ),
where X X and Y Y.
One can visualize H as a self-intersecting curve. For simplicity we will dis-
entangle it to get a simple closed curve. To this end, we consider the object
H = H S , which is the Minkowski-sum of H and a square S with side length
, where is an arbitrarily small positive quantity (see Figure 1c and 1d). This is
basically fattening the boundary H such that the boundary of H is a simple
closed curve. Also note that it fattens only towards the positive horizontal and
vertical axes. Let H be the boundary of H . Now for every p(X, Y ) H,
there exists at least one point on H among the following points p(X, Y ),
p(X, Y ) + (0, ), p(X, Y ) + (, 0), and p(X, Y ) + (, ). Refer to that point as
p (X, Y ). If multiple of these points exist on H , then choose one arbitrarily.
For convenience we use p (X, Y ) as a proxy for p(X, Y ) while adding edges
to H whose one endpoint is p(X, Y ). At the end we will use the original points
p(X, Y ). As the two points are arbitrarily close the latter conversion does not af-
fect the planarity. Hereafter, for any reference to a point p (X, Y ), we will assume
that p(X, Y ) H. Now note that the only edges that are needed to be drawn
are the edges in EX and EX ,Y . The endpoints of the edges in EX will be the left
endpoints of the segments in X . Also the endpoints of the edges in EX ,Y will be
the left endpoints of the segments in X and the intersection
 points
 p(X, Y ) for
X X and Y Y such that p(X, Y ) H. Let Q = X (X) X,Y p (X, Y ),
the set of endpoints of the edges we will draw. Note that all the points of Q are
on H .
We consider GX as a tree rooted at some segment Xr X . Later we will
use this tree structure to add edges to H in an inductive manner. We denote
the parent of a segment Xi with respect to GX by parent(Xi ). Let r(Xr ) =
(x , y  ). We cut open the closed curve H at (x , y  + ) such that the open
curve is homeomorphic to an interval M , where r(Xr ) gets mapped before (Xr ).
For simplicity, we denote the mapped points by their original notations. Thus
r(Xr ) < (Xr ) in M . Note that one can consider another homeomorphism to an
interval (the reection of M with respect to its middle point) so that (Xr ) <
r(Xr ). One can visualize M in a way such that H is disconnected at the point
(x , y  + ) and straightened out to get M . The way we have dened M it has the
following property: for any child Xi of Xr , if Xi lies above Xr , r(Xr ) < (Xr ) <
(Xi ) < r(Xi ) and if Xi lies below Xr , r(Xr ) < r(Xi ) < (Xi ) < (Xr ). Observe,
that the endpoints of M are not in Q and we already have an embedding of the
points of Q on the interval M . We use this embedding of the points to draw the
edges between them so that the edges do not cross each other.
For any edge (a, b) EX EX ,Y , (a, b) is called an edge corresponding to a
segment X of GX if either a or b is X, and the other endpoint is not parent(X).
For any subtree, we call all the edges corresponding to its nodes as its corre-
sponding edges. Now we proceed towards the inductive process of addition of
edges. For that we need to dene a concept called zone for every segment Xi of
58 S. Bandyapadhyay and A. Basu Roy

4
s4 t4
6
1

2 t2 s2
3 t3 s3

Fig. 2: The bolded curves corresponding to the children (2, 3, 4) of 1 are mapped to
the corresponding zones in M . si and ti are the respective left and right endpoint
of the zone of i for i = 2, 3, 4.

GX , which is basically a sub-interval of M . We will ensure that the endpoints of

the edges corresponding to Xi lies in its zone. The structure of H or M ensures
that the zones have many useful properties, which we will exploit heavily.
For the root segment Xr , zone(Xr ) is dened to be the entire interval M .
For any other Xi X , let Y and Yr be the leftmost and the rightmost ver-
tical segment in vert(Xi ) vert(parent(Xi )), respectively. If Xi lies above its
parent, then zone(Xi ) = [p(Xi , Y ), p(Xi , Yr ) + (, 0)]. Otherwise, zone(Xi ) =
[p(Xi , Yr ) + (, ), p(Xi , Y ) + (0, )]. See Figure 2. Now we have the following
observation.
Observation 6. The following statements are true regarding zones.
1. zone(Xi ) zone(parent(Xi )) for every Xi X .
2. For any segment X X , Y Y, p (X, Y ) and (X) lies outside zone(Xj ),
where Xj is a child of X.
3. The endpoints of the edges corresponding to a segment X X lie within the
zone of X. Moreover, for all edge corresponding to the subtree rooted at X,
both of the endpoints lie on the same side of (X).
4. The zones of the segments corresponding to the children of X are pairwise
disjoint.
In the following lemma, we show the rst stage of our construction, where
the edges are drawn without crossing.

Lemma 9. Given the plane graph H as dened before, the edges of EX and
EX ,Y can be drawn in a non-crossing manner.

Proof. We note that the edges of EX EX ,Y are drawn outside of the closed
curve H . Thus they do not cross the edges of H as the latter edges lie inside
H . Consider the set Q of the endpoints of the edges to be drawn. We use the
Effectiveness of Local Search for Art Gallery Problems 59

embedding of the points of Q on M for drawing edges between them. As M is

homeomorphic to the open curve we obtain by cutting open H this lemma
follows. Now we give an explicit construction of a drawing of the edges. In each
step, we draw the edges corresponding to a subtree of GX rooted at some node
X, given the drawings corresponding to the subtrees rooted at the children of
X. In the base case, we draw the edges corresponding to the leaves of GX . Note
that each leaf is corresponding to a segment X X . The only edges we draw
corresponding to X are of the form ((X), p (X, Y )), and hence it is easy to see
that these edges can be drawn without any crossing.
Now consider any non-leaf node X. By induction we already have a crossing-
free drawing corresponding to the subtree rooted at any child of X. From state-
ments (1) and (3) of Observation 6 it follows that the endpoints of the edges
in this drawing must lie within the zone of the child. As zones of the children
are disjoint by statement (4) of the same Observation, we get a crossing-free
drawing with respect to the children of X. Thus the only additional edges we
need to draw have (X) as an endpoint. The other endpoint could be l(Xi ) for
a child Xi of X or p (X, Y ) for (X, Y ) EX ,Y . Now l(X) and p (X, Y ) do not
belong to the zone of any child of X by statement (4) of Observation 6. Also
by statement (3) of the same observation, both of the endpoints of any existing
edge corresponding to the subtree rooted at a child Xj of X lie on the same
side of l(Xj ). Hence all the additional edges can be drawn without crossing any
existing edge.
Now we begin the second stage of planarity preserving operations, i.e. the
edge contractions. We contract (p(X, Y ), r(X)) for every X X where Y is the
rightmost vertical segment among the ones intersecting X. After this for every
Y Y we contract every edge subdividing Y such that a single point remains
at the end corresponding to Y . Thus for every Y , there exists a unique vertex
point. Hence, by construction all the properties in Theorem 6 are satised.
One can extend the construction for the case when A(X Y) has more than
one components. We conclude this section with the following theorem.

Theorem 10. There exists a local search PTAS for the MSC problem, where
the input polygon does not contain holes.

Acknowledgments
We would like to thank the anonymous referees, Abhiruk Lahiri, and Kasturi
Varadarajan for their valuable comments, which have helped us improve the
presentation of the paper.

References
[1] V. Arya, N. Garg, R. Khandekar, A. Meyerson, K. Munagala, and V. Pandit.
Local search heuristics for k-median and facility location problems. SIAM
J. Comput., 33(3):544562, 2004.
60 S. Bandyapadhyay and A. Basu Roy

[2] R. Aschner, M. J. Katz, G. Morgenstern, and Y. Yuditsky. Approximation

schemes for covering and packing. In WALCOM 2013., pages 89100.
[3] P. Ashok, A. Basu Roy, and S. Govindarajan. Local search strikes again: PTAS
for variants of geometric covering and packing. In COCOON, 2017.
[4] S. Bandyapadhyay and K. R. Varadarajan. On variants of k-means cluster-
ing. In SoCG 2016,, pages 14:114:15, 2016.
[5] V. V. S. P. Bhattiprolu and S. Har-Peled. Separating a voronoi diagram via
local search. In SoCG 2016,, pages 18:118:16, 2016.
[6] T. C. Biedl, T. M. Chan, S. Lee, S. Mehrabi, F. Montecchiani, and
H. Vosoughpour. On guarding orthogonal polygons with sliding cameras.
CoRR, abs/1604.07099, 2016.
[7] T. M. Chan and S. Har-Peled. Approximation algorithms for maximum
independent set of pseudo-disks. DCG, 48(2):373392, 2012.
[8] V. Cohen-Addad and C. Mathieu. Eectiveness of local search for geometric
optimization. In SoCG 2015,, pages 329343, 2015.
[9] V. Cohen-Addad, P. N. Klein, and C. Mathieu. Local search yields approx-
imation schemes for k-means and k-median in euclidean and minor-free
metrics. In FOCS 2016, pages 353364, 2016.
[10] M. de Berg, S. Durocher, and S. Mehrabi. Guarding monotone art galleries
with sliding cameras in linear time. In COCOA 2014, pages 113125, 2014.
[11] S. Durocher and S. Mehrabi. Guarding orthogonal art galleries using sliding
cameras: Algorithmic and hardness results. In MFCS, pages 314324, 2013.
[12] S. Friedrichs, M. Hemmer, J. King, and C. Schmidt. The continuous 1.5d
terrain guarding problem: Discretization, optimal solutions, and PTAS.
JoCG, 7(1):256284, 2016.
[13] Z. Friggstad, M. Rezapour, and M. R. Salavatipour. Local search yields a
PTAS for k-means in doubling metrics. In FOCS 2016, pages 365374.
[14] S. Govindarajan, R. Raman, S. Ray, and A. Basu Roy. Packing and covering
with non-piercing regions. In ESA 2016, pages 47:147:17, 2016.
[15] A. Gupta and K. Tangwongsan. Simpler analyses of local search algorithms
for facility location. CoRR, abs/0809.2554, 2008.
[16] T. Kanungo, D. M. Mount, N. S. Netanyahu, C. D. Piatko, R. Silverman,
and A. Y. Wu. A local search approximation algorithm for k-means clus-
tering. Comput. Geom., 28(2-3):89112, 2004.
[17] M. J. Katz and G. Morgenstern. Guarding orthogonal art galleries with
sliding cameras. Int. J. Comput. Geometry Appl., 21(2):241250, 2011.
[18] D. G. Kirkpatrick. An o(lg lg opt)-approximation algorithm for multi-
guarding galleries. DCG, 53(2):327343, 2015.
[19] E. Krohn, M. Gibson, G. Kanade, and K. Varadarajan. Guarding terrains
via local search. Journal of Computational Geometry, 5(1):168178, 2014.
[20] N. H. Mustafa and S. Ray. Improved results on geometric hitting set prob-
lems. Discrete & Computational Geometry, 44(4):883895, 2010.
[21] J. ORourke. Art Gallery Theorems and Algorithms. Oxford University
Press, Inc., New York, NY, USA, 1987. ISBN 0-19-503965-3.
[22] J. ORourke. Computational geometry column 52. SIGACT News, 43(1):
8285, Mar. 2012. ISSN 0163-5700.
 Parameterized Complexity of Geometric
Covering Problems Having Conicts 

Aritra Banik1 , Fahad Panolan2 , Venkatesh Raman3 , Vibha Sahlot1 , and Saket
Saurabh3
1
Indian Institute of Technology, Jodhpur, India.
{aritrabanik|sahlotvibha}@gmail.com
2
Department of Informatics, University of Bergen, Norway.
[email protected]
3
The Institute of Mathematical Sciences, HBNI, Chennai, India.
{vraman|saket}@imsc.res.in

Abstract. The input for the Geometric Coverage problem consists

of a pair = (P, R), where P is a set of points in Rd and R is a set
of subsets of P dened by the intersection of P with some geometric
objects in Rd . These coverage problems form special instances of the
Set Cover problem which is notoriously hard in several paradigms
including approximation and parameterized complexity. Motivated by
what are called choice problems in geometry, we consider a variation of the
Geometric Coverage problem where there are conicts on the covering
objects that precludes some objects from being part of the solution if
some others are in the solution.
As our rst contribution, we propose two natural models in which the
conict relations are given: (a) by a graph on the covering objects, and
(b) by a representable matroid on the covering objects. We consider the
parameterized complexity of the problem based on the structure of the
conict relation. Our main result is that as long as the conict graph has
bounded arboricity (that includes all the families of intersection graphs
of low density objects in low dimensional Euclidean space), there is a
parameterized reduction to the problem without conicts on the covering
objects. This is achieved through a randomization-derandomization trick.
As a consequence, we have the following results when the conict graph
has bounded arboricity.
If the Geometric Coverage problem is xed parameter tractable
(FPT), then so is the conict free version.
If the Geometric Coverage problem admits a factor -approximation,
then the conict free version admits a factor -approximation algo-
rithm running in FPT time.
As a corollary to our main result we get a plethora of approximation
algorithms running in FPT time. Our other results include an FPT algo-
rithm and a W[1]-hardness proof for the conict-free version of Covering
Points by Intervals. The FPT algorithm is for the case when the con-
icts are given by a representable matroid, and the W[1]-hardness result

Supported by Parameterized Approximation, ERC Starting Grant 306992, and
Rigorous Theory of Preprocessing, ERC Advanced Investigator Grant 267959.

Springer International Publishing AG 2017 61

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 6172, 2017.
DOI: 10.1007/978-3-319-62127-2_6
62 A. Banik et al.

is for all the families of conict graphs for which the Independent Set
problem is W[1]-hard.

1 Introduction, Motivation, Model and Our Results

There are many real life geometric covering problems, for which there exist
additional constrains that need to be enforced. In this paper, we attempt to
address these problems and hope that this will initiate a new line of research
directed at bridging the gap between theory and practice.
To dene our model of covering with conicts, we start by dening the classic
covering problem. The input to a covering problem consists of a universe U of size
n, a family F of size m of subsets of U and a positive integer k. Our objective is
to check whether there exists a subfamily F  F of size at most k satisfying
some desired properties. If F  is required to contain all the elements of U , then
it corresponds to the classical Set Cover problem and F  is called a set cover.
The Set Cover problem is part of Karps 21 NP-complete problems [11].
We begin the development with a conict free problem already studied,
Conflict Free Interval Covering, introduced in [1,2,3]. Let P be a set of
points on the x-axis, and let I = {I1 , . . . , Im } be a set of intervals on the x-axis.
Furthermore, let C = {C1 , C2 , . . . , C } denote a set of color classes, where each
color class Ci consists of a pair of intervals from I. Moreover, for any pair of
integers i, j (1 i < j ), Ci Cj = . We term C a matching family. For a
set of intervals Q I, Q is conict free if Q contains at most one interval from
each color class, i.e. 1i |Q Ci | 1. Finally, for an interval I = [a, b] and a
point c on x-axis, we say I covers p if and only if a c b. Now we are ready to
dene the problem formally.
Rainbow Covering
Input: A set of points P on the x-axis, a set of intervals I = {I1 , . . . , Im } on
the x-axis and a matching family C = {C1 , C2 , . . . , C }.
Question: Does there exist a conict free subset Q of intervals which covers
all the points in P ?
Our rst goal is to dene a model in which we can express much more
generalized version of conicts beyond the matching family of conict graphs.
To dene our model we revisit Set Cover, as the model is best dened in
the most general setting. Recall that the input to a Set Cover consists of a
universe U of size n, a family F of subsets of U of size m. A natural way to
model conict is by using graphs. Formally stated, we have a graph CGF , on the
vertex set F and there is an edge between two sets Fi , Fj F if Fi and Fj are in
conict. We call CGF a conict graph. Observe that in the Rainbow Covering
problem, the family C would corresponds to CGC with degree at most one. That
is, edges of CGC form a matching. And the question of nding a conict free
subset Q of intervals covering all the points in P becomes a problem of nding a
set Q of intervals that covers all the points in P and CGC [Q] is an independent
set. The set cover F  such that CGF [F  ] is an independent set will be called
conict free set cover.
Parameterized Complexity of Geometric Covering Problems Having Conflicts 63

Our Contributions. In this paper we study the following problems in geo-

metric settings in the realm of Parameterized Complexity. For more details about
parameterized complexity we refer to monographs [4].
Graphical Conflict Free Set Cover (Graphical CF-SC)
Input: A universe U of size n, a family F of size m of subsets of U , a conict
graph CGF and a positive integer k.
Parameter: k
Question: Does there exist a set cover F  F of size at most k such that
CGF [F  ] is an independent set?
Let (A, B)-Set Cover denote a restriction of Set Cover, where every
instance (U, F, k) of Set Cover satises the property that U A and F B.
For example in this setting, Covering Points by Intervals corresponds to
(A, B)-Set Cover where A is the set of points on x-axis and B is the set of
intervals on x-axis. Given (A, B)-Set Cover, the corresponding Graphical
CF-SC corresponds to (A, B)-Graphical CF-SC.
Observe that Graphical CF-SC becomes Set Cover if CGF is an indepen-
dent set. As the general Set Cover is hard in the parameterized framework, to
design an FPT algorithm for Graphical CF-SC, it is important that the base
Set Cover problem is FPT. This restricts us to (A, B)-Set Cover which are
either FPT or polynomial time solvable. If we are seeking FPT approximation
algorithms then we can also restrict ourselves to (A, B)-Set Cover which has
either polynomial time approximation scheme (PTAS), constant factor approxi-
mation algorithm or FPT approximation algorithms, even if the problem is not
in FPT. For example (A, B)-Set Cover, where A is set of points in R2 and B is
a set of unit discs in R2 is known to be W[1] hard [14] but admits a PTAS [10].
We will call (A, B)-Set Cover tractable if it admits one of the following: a
polynomial time algorithm, an FPT algorithm, an (E)PTAS, a constant factor
approximation algorithm, an FPT approximation algorithm.
The next natural question is if we restrict ourselves to tractable (A, B)-Set
Cover, can an arbitrary conict graph CGF yield tractable algorithms for the
conict-free versions of (A, B)-Set Cover? To formalize this question, let G
denote a family of graphs. Then, the question is for which family of graphs G, does
(A, B)-Graphical CF-SC admit an FPT algorithm or an FPT approximation
algorithm when CGF belongs to G. For example, if G is the family of cliques,
then even Graphical CF-SC trivially becomes polynomial time solvable when
CGF belongs to this family of cliques.
A problem that will be central to our study is the following. Let P and I
denote a set of points and a set of intervals on the x-axis, respectively.
(P, I )-Graphical CF-SC Parameter: k
Input: A set of points P P, a set of intervals I = {I1 , . . . , Im } I , a
conict graph CGI and a positive integer k.
Question: Does there exist a conict free set cover of size at most k?
In (P, I )-Graphical CF-SC, when CGI belongs to the family of matchings
then the problem becomes Parameterized Rainbow Covering. This problem
64 A. Banik et al.

was studied in [1] and shown to be NP-complete. In fact, even if we do not care
about the size of the conict free set cover we seek, just the decision version of a
conict free set cover set is the same as Rainbow Covering, which is known to
be NP-complete. Thus, seeking a conict free set cover can transform a problem
from being tractable to intractable.
In order to restrict the family of graphs to which a conict graph belongs, we
need to dene the notion of arboricity. The arboricity of an undirected graph is the
minimum number of forests into which its edges can be partitioned. A graph G is
said to have arboricity d if the edges of G can be partitioned into at most d forests.
Let Gd denote the family of graphs of arboricity d. This family includes the family
of intersection graphs of low density objects in low dimensional Euclidean space
as explained in [8,9]. Specically, this includes planar graphs, graphs excluding a
xed graph as a minor, graphs of bounded expansion, and graphs of bounded
degeneracy. Har-Peled and Quanrud [8,9] showed that low-density geometric
objects form a subclass of the class of graphs that have polynomial expansion,
which in turn, is contained in the class of graphs of bounded arboricity. Thus,
our restriction of the family of conict graphs to a family of graphs of bounded
arboricity covers a large class of low-density geometric objects.

Theorem 1. Let (A, B)-Set Cover be tractable and let Gd be the family of
graphs of arboricity d. Then, the corresponding (A, B)-Graphical CF-SC is
also tractable if CGF belongs to Gd . In particular we obtain following results
when CGF belongs to Gd :
If (A, B)-Set Cover admits an FPT algorithm with running time (k)nO(1) ,
then (A, B)-Graphical CF-SC admits an FPT algorithm with running time
2O(dk) (k) nO(1) .
If (A, B)-Set Cover admits a factor -approximation running in time
nO(1) then (A, B)-Graphical CF-SC admits a factor -FPT-approximation
algorithm running in time 2O(dk) nO(1) .

The proof of Theorem 1 is essentially a black-box reduction to the non-conict

version of the problem. Thus, Theorem 1 covers a number of conict-free version of
many fundamental geometric coverage problems as illustrated in Table 1. In light
of Theorem 1, it is natural to ask whether or not, these problems admit polynomial
time approximation algorithms. Unfortunately, we cannot expect these problems
to admit even a factor o(n)-approximation algorithm. This is because for most
of these problems even deciding whether there exists a conict free solution, with
no restriction on the size of the solution, is NP-complete (for example Rainbow
Covering is NP-complete [1]). Thus, having an o(n)-approximation algorithm
would imply a polynomial time algorithm for the decision version of the problem,
which we do not expect unless P=NP. Hence, the best we can expect for the
(A, B)-Graphical CF-SC problems is an FPT-approximation algorithm, as for
many of them we can neither have an FPT algorithm, nor a polynomial time
approximation algorithm.
We complement our algorithmic ndings by a hardness reduction. Let G
denote a family of graphs. Let G -Independent Set be the problem where the
Parameterized Complexity of Geometric Covering Problems Having Conflicts 65

(R2 , A)-SC Complexity of Complexity of

(R2 , A)-SC (R2 , A)-Graphical CF-SC
Disks/pseudo-disks PTAS -FPT approx., > 1
Fat triangles of same size O(1) O(1)-FPT approx.
Fat objects in R2 O(log OPT) O(log OPT)-FPT approx.
O(1) density objects in R2 PTAS -FPT approx., > 1

Objects with polylog density QPTAS 2O(k) nO(log n) time approx.,
> 1
Objects with density O(1) in Rd PTAS -FPT approx., > 1
(A, B)-Set Cover where every in- O(d log(dOPT)) O(d log(dOPT))-FPT ap-
stance (U, F ) has VC dimension d prox.
Point Guard Art Gallery O(log OPT) O(log OPT)-FPT approx.
Terrain Guarding PTAS -FPT approx., > 1
(P, I )-Set Cover Polynomial Time 2O(dk) nO(1) -FPT algorithm
Table 1. Corollaries of Theorem 1. Here (R2 , A)-Set Cover ((R2 , A)-SC) is a geometric
set cover problem where R2 is a set of points in the plane and the covering objects are
specied in the rst column. The conict graph for all the problems is Gd , family of
graphs of arboricity d, for some constant d. For the denitions of density and fatness
we refer to [8]. The entries in the second column give the approximation ratio of the
(R2 , A)-SC problem based on Theorem 1.

input is a graph G G and a positive integer k, and the objective is to decide

whether there is a set S of size at least k such that G[S] is an independent set.
Theorem 2. Let G denote a family of graphs such that G -Independent Set
is W[1]-hard. If CGI belongs to G , then (P, I )-Graphical CF-SC does not
admit an FPT algorithm, unless FPT =W[1].
The proof of Theorem 2 is a Turing reduction based on (n, k)-perfect hash
families [16] that takes time 2O(k) nO(1) . In fact, for any xed A and B, one
should be able to follow this proof and show W[1]-hardness for (A, B)-Graphical
CF-SC, where CGF belongs to a graph family G for which G -Independent
Set is W[1]-hard. Due to paucity of space the proof of Theorem 2 is deferred to
the full version of the paper.
Theorem 1 captures those families of conict graphs that are everywhere
sparse. However, the (A, B)-Graphical CF-SC problem is also tractable if the
conict graphs belong to the family of cliques. When the conict graph belongs to
a dense family of graphs, we design a general theorem using matroid machinery.
Let (U, F, k) be an instance of Set Cover. In the matroidal model of
representing conicts, we are given a matroid M = (E, J ), where the ground set
E = F, and J is a family of subsets of F satisfying all the three properties of a
matroid. In this paper we assume that M = (E, J ) is a linear or representable
matroid, and the corresponding linear representation is given as part of the input.
In the Rainbow Covering problem, let Q denote the family of conict free
subsets of intervals in I. One can dene a partition matroid on F such that
J = Q. Thus, the question of nding a conict free subset of intervals covering all
the points in P becomes a problem of nding an independent set in J that covers
66 A. Banik et al.

all the points in P . The Matroidal Conflict Free Set Cover problem
(Matroidal CF-SC, in short) is dened similarly to Graphical CF-SC. In
particular, the input consists of a linear matroid M = (F, J ) over the ground
set F such that the set cover F  J .
Theorem 3. (P, I )-Matroidal CF-SC is FPT for all representable matroids
M = (I, J ) dened over I. In fact, given a linear representation, the algorithm
runs in time 2k (n + m)O(1) . Here, is the exponent in the running time of
matrix multiplication.
A graph is called a cluster graph, if all its connected components are cliques.
Since cluster graphs can be captured by partition matroids, Theorem 3 implies
that (P, I )-Matroidal CF-SC is FPT if CGF is a cluster graph.
Notations. For t N, we use [t] as a shorthand for {1, 2, . . . , t}. A family of sets
A is called a p-family, if the cardinality of all the sets in A is p. Given two families
of sets A and B, we dene A B = {X Y | X A and Y B and X Y = }.
Given a graph G, V (G) and E(G) denote its vertex-set and edge-set, respectively.
We borrow notations from the book of Diestel [5] for graph-related notations.

2 FPT Algorithms
In this section we prove Theorems 1 and Theorem 3. The Proof of Theorem 1 is
based on a randomization scheme while the proof of Theorem 3 uses the idea of
ecient computation of representative families [6].

2.1 FPT Algorithms for Graphical CF-SC

Our algorithm for Theorem 1 is essentially a randomized reduction from (A, B)-
Graphical CF-SC to (A, B)-Set Cover, when the conict graph has bounded
arboricity. Towards this, we start with a forest decomposition of graphs of
bounded arboricity and then apply a randomized process to obtain an instance
of (A, B)-Set Cover. However, to design a deterministic algorithm we use the
construction of universal sets. For this, we will exploit the following denition
and theorem.
Denition 1 ([16]). An (n, t)-universal set F is a set of functions from {1, . . . , n}
to {0, 1}, such that for every subset S {1, . . . , n}, |S| = t, the set F |S =
{f |S | f F } is equal to the set 2S of all the functions from S to {0, 1}.
Theorem 4 ([16]). There is a deterministic algorithm with O(2t tO(log t) n log n)
run time that constructs an (n, t)-universal set F such that |F | = 2t tO(log t) log n.
Now we are ready to give the proof of Theorem 14
Proof (Proof of Theorem 1). Let (U, F, CGF , k) be an instance of (A, B)-
Graphical CF-SC, where CGF belongs to Gd . Our algorithm has the following
phases.
4
The idea used in the proof of Theorem 1 is inspired by a proof used in [13].
Parameterized Complexity of Geometric Covering Problems Having Conflicts 67

Decomposing CGF into Forests. We apply the known polynomial time

algorithm [7] to decompose the graph CGF into T1 , . . . , Td where Ti is a forest
d
in CGF and i=1 E(Ti ) = E(CGF ). Let vroot be a special vertex such that vroot
does not belong to V (CGF ) = F. Now for every Ti , and for every connected
component of Ti , we pick an arbitrary vertex and connect it to vroot . Now if
we look at the tree induced on V (Ti ) {vroot } then it is connected and we will
denote this tree by Ti . Furthermore, we will treat each Ti as a tree rooted at
vroot . This automatically denes parent-child relationship among the vertices of
Ti . This completes the partitioning of the edge set of CGF into forests.
Step 1: Randomized event and probability of success. Independently
color the vertices of CGF into blue and green uniformly at random. That is, we
color the vertices of CGF blue and green with probability 12 . Furthermore, we
color {vroot } to blue. Let F  be a conict free set cover of size at most k. We
consider the following event to be good.
Every vertex in F  is colored green and every parent of every vertex in
F  in every tree Ti is colored blue.
Let Sparent denote the set of parents of every vertex in F  in every tree Ti . Since,
we have at most d trees and the size of F  is upper bounded by k we have that
|Sparent | kd. We say that F  (Sparent ) is green (blue) to mean that every vertex
in F  (Sparent ) is colored green (blue). Thus,
Pr[good event happens] = Pr[F  is green Sparent is blue]
1
= Pr[F  is green] Pr[Sparent is blue] .
2k(d+1)
The second equality follows from the following fact. The set F  is an independent
set in CGF and Sparent NCGF (F  ) {vroot }. Thus, these sets are pairwise
disjoint and hence the events F  is colored green and Sparent is colored blue are
independent.
1
Step 2: A cleaning process. Let p = 2kd . Now we apply a cleaning procedure
so that we get a set Z such that CGF [Z] is an independent set in CGF and it
contains F  . Let B denote the set of vertices that have been colored blue. We start
by deleting every vertex in B. Now for every edge (f1 , f2 ) in CGF [V (CGF ) \ B],
we do as follows. We know that (f1 , f2 ) belongs to some tree Ti and thus either
f1 is a child of f2 or vice-versa. If f1 is a child then we delete f1 , otherwise
we delete f2 . Let the resulting set of vertices be Z. By construction Z is an
independent set in CGF . Next we show that F  Z with probability p/2k .
Clearly, with probability 21k we know that no vertex of F  is colored blue and
thus with probability 21k we know that F  V (CGF ) \ B. Observe that with
probability p, we have that all the parents of F  in any tree Ti have been colored
blue. Thus, a vertex x V (CGF ) \ B, colored green, can not belong to F  , if it
is a child of some vertex in some tree Ti after deleting the vertices of B. This
is the reason when we delete a vertex from an edge (f1 , f2 ), we delete the one
which is a child in some tree Ti . Thus, by deleting a vertex that is a child in
an edge (f1 , f2 ), we do not delete any vertex from F  . This implies that with
1
probability 2k(d+1) , we have that F  Z. This completes the proof.
68 A. Banik et al.

Solving the problem. Let Q be a parameterized algorithm for (A, B)-Set

Cover running in time (k) nO(1) . Recall that (U, F, CGF , k) is an instance of
(A, B)-Graphical CF-SC. Now to test whether there exists a conict free set
cover F  of size at most k, we run Q on (U, Z, k). If the algorithm return Yes,
we return the same for (A, B)-Graphical CF-SC. Else, we repeat the process
by randomly nding another Z by following Steps 1 and 2 and then running
the algorithm Q on the instance (U, Z , k) and returning the answer accordingly.
We repeat the process 2k(d+1) time. If we fail to detect whether (U, F, k, CGF )
is a Yes instance of (A, B)-Graphical CF-SC in 2k(d+1) rounds, then we return
that the given instance is a No instance. Thus, if (U, F, k, CGF ) is No instance of
(A, B)-Graphical CF-SC, then we always return No. However, if (U, F, k, CGF )
is a Yes instance of (A, B)-Graphical CF-SC then there exists a set F  , that is
a conict free set cover of size at most k. The probability that we will notqnd a

set Z containing F in q = 2 k(d+1)
rounds is upper bounded by 1 q 1e .
1

Thus, the probability that we will nd a set Z containing F  in q rounds is at

least 1 1e 12 . Thus, if the given instance is a Yes instance then the algorithm
succeeds with probability at least 12 . The running time of the algorithm is upper
bounded by (k) 2k(d+1) nO(1) .
Derandomizing the algorithm. Now to design our deterministic algorithm
all we will need to do is to replace the randomized coloring function with
a deterministic coloring function that colors the vertices in F  green and all
the vertices in Sparent to blue. To design such a coloring function we set t =
k(d + 1), and use Theorem 4 to construct an (n, t)-universal set F such that
|F | = 2t tO(log(t)) log n. The algorithm to construct F takes O(2t tO(log(t)) n log n).
Finally, to derandomize our algorithm, rather than randomly coloring vertices
with {blue, green}, we go through each function f in the family F and view the
vertices that have assigned 0 as blue and others as green. By the properties of
(n, t)-universal set we know that there exists a function f that correctly colors the
vertices in F  with 1 and every vertex in Sparent with 0. Thus, the set Zf we will
obtain by applying Step 2 will contain the set F  . After this the correctness of the
algorithm follows from the correctness of the algorithm Q. Thus, the running time
of the algorithm is upper bounded by (k)|F |nO(1) = (k)2k(d+1)+o(kd) nO(1) .
This completes the proof of the rst part.
Let S be a factor -approximation algorithm for (A, B)-Set Cover running
in time nO(1) . To obtain the desired FPT approximation algorithm with factor ,
we do as follows. We only give the deterministic version of the algorithm based on
the uses of universal sets. As before, let (U, F, CGF , k) be an instance of (A, B)-
Graphical CF-SC, where CGF , belongs to Gd . We again set t = k(d+1), and use
Theorem 4 to construct an (n, t)-universal set F such that |F | = 2t tO(log(t)) log n.
The algorithm to construct F takes O(2t tO(log(t)) n log n). We go through each
function f in the family F and view the vertices that have been assigned 0 as
blue and others as green. If there exists a conict free set cover F  of size at
most k, then by the properties of (n, t)-universal set we know that there exists
a function f that correctly color the vertices in F  with 1 and every vertex in
Sparent with 0. Thus, the set Zf we will obtain by applying Step 2, will contain
Parameterized Complexity of Geometric Covering Problems Having Conflicts 69

the set F  . Thus, to design the approximation algorithm, for every f F , we

rst construct Zf . And for each such Zf we run S on (U, Zf , k). This could
either return that there is No solution, or returns a solution F  which is a factor
-approximation to the instance (U, Zf , k). If for some f F , S returns F 
of size at most k when run on (U, Zf , k) then the algorithm returns F  . In all
other cases the algorithm returns that the given instance is a No instance. The
correctness of the algorithm follows from the properties of universal sets and
the correctness of the algorithm S . The running time of the algorithm is upper
bounded by: |F | Running time of S = 2k(d+1)+o(kd) nO(1) . This completes
the proof. 


2.2 FPT Algorithm for (P, I )-Matroidal CF-SC

In this section we will design an FPT algorithm proving Theorem 3. Towards that
we need to dene some basic notions related to representative families and results
regarding their fast and ecient computation. For denitions related to matroids
and a broad overview of representative families we refer to [4, Chapter 12].
Denition 2 (q-Representative Family [15,4]). Given a matroid M = (E, J )
and a family S of subsets of E, we say that a subfamily S S is q-representative
for S if the following holds: for every set Y E of size at most q, if there is a
set X S disjoint from Y with X Y J , then there is a set X S disjoint
from Y with X Y J . If S S is q-representative for S we write S q S.
rep

Lemma 1 ([6]). Let M = (E, J ) be a matroid and S be a family of subsets of

E. If S  qrep S and S qrep S  , then S qrep S.
Lemma 2 ([12]). Let M = (E, J ) be a linear matroid of rank n and let S =
{S1 , . . . , St } be a p-family of independent sets. Let A be a n|E| matrix represent-
ing M over a eld F, where F = Fp or F is Q. 
Then there is a deterministic algo-


p+q  3 2
1

rithm computing S S of size np
q
in O p+q
tp n + t p+q (pn)1 +
rep p p q
O(1)
(n + |E|) operations over F.
Now we are ready to prove Theorem 3. Let (P, I, k, M = (I, J )) be an
instance of (P, I )-Matroidal CF-SC, where P is a set of points on the x-axis,
I = {I1 , . . . , Im } is a set of intervals on the x-axis and M = (I, J ) is a matroid
over the ground set I. The objective is to nd a set cover S I of size at most
k such that S J .
To design our algorithm for (P, I )-Matroidal CF-SC, we will use e-
cient computation of representative families applied on a dynamic programming
algorithm. Let P = {p1 , . . . , pn } denote the set of points sorted from left to right.
Next we introduce the notion of family of partial solutions. Let
  

P i = X  X I, X J , |X| k, X covers p1 , . . . , pi

denote the family of subsets of intervals of size at most k that covers rst i points
and are independent in the matroid M = (I, J ). Furthermore, for every j [k],
by P ij , we denote the subset of P i containing sets of size exactly j. Thus,
70 A. Banik et al.


k
Pi = P ij .
j=1

In this subsection whenever we talk about independent sets, these are independent
sets of the matroid M = (I, J ). Furthermore, we assume that we are given, AM ,
the linear representation of M . Without loss of generality we can assume that
AM is a n |I| matrix, where n |I|.
Observe that (P, I, k, M = (I, J )) is a Yes instance of (P, I )-Matroidal
CF-SC if and only if P n is non-empty. This implies that P n is non-empty if and
only if P n 0 P n is non-empty. We capture this into the following lemma.
rep

Lemma 3. Let (P, I, k, M = (I, J )) be an instance of (P, I )-Matroidal

CF-SC. Then, (P, I, k, M = (I, J )) is a Yes instance of (P, I )-Matroidal
n 0 P n is non-empty.
CF-SC if and only if P n is non-empty if and only if P rep

For an ease of presentation by P 0 , we denote the set {}. The next lemma
provides an ecient computation of the family P i 1k P i . In particular, for
rep
every 1 i n, we compute

k 
 
i =
P ij kj P ij .
P rep
j=1

Lemma 4. Let (P, I, k, M = (I, J )) be an instance of (P, I )-Matroidal

CF-SC. Then for every 1 i n, a collection of families P i 1k P i , of size
rep
O(1)
at most 2 |I| k can be found in time 2 (n + |I|)
k k
.

Proof. We describe a dynamic programming based algorithm. Let P = {p1 , . . . , pn }

denote the set of points sorted from left to right and D be a n + 1-sized array
indexed with {0, . . . , n}. The entry D[i] will store a family P i 1k P i . We ll
rep
the entries in the matrix D in the increasing order of index. For i = 0, D[i] = {}.
Let i {0, 1, . . . , n} and assume that we have lled all the entries until the row
i (i.e, D[i] will contain a family P i 1k P i ). For any interval I I, let I be
rep
the lowest index in [n] such that pI is covered by I. Let Zi+1 denote the set of
intervals I I that covers the point pi+1 . Now we compute

N i+1 = (D[I 1] {I}) J (1)
IZi+1

Notice that in the Equation 1, the union is taken over I Zi+1 . Since for any
I Zi+1 , I covers pi+1 , the value I 1 is strictly less than i + 1 and hence
Equation 1 is well dened. Let N (i+1)j denote the subset of N i+1 containing
subsets of size exactly j.

Claim. N i+1 1k
rep P
i+1
.
Parameterized Complexity of Geometric Covering Problems Having Conflicts 71

Proof. Let S P (i+1)j and Y be a set of size k j (which is essentially an

independent set of M ) such that S Y = and S Y J . We will show that
there exists a set S N (i+1) such that S Y = and S Y J . This will
imply the desired result.
Since S covers {p1 , . . . , pi+1 }, there is an interval J in S which covers pi+1 .
Since S covers {p1 , . . . , pi+1 } and J covers pi+1 , the set of intervals S  = S \ {J}
covers {p1 , . . . , pi+1 } \ {pJ , . . . , pi+1 } and J covers {pJ , . . . pi+1 }. Let Y  =
Y {J}. Notice that S  Y  = S Y J , |S  | = j 1, |Y  | = k j + 1 and S 
covers {p1 , . . . , pi+1 } \ {pJ , . . . , pi+1 }. This implies that S  P (J 1)(j1) and
by our assumption that D[J 1] contain P (J 1)(j1) kj+1 P (J 1)(j1) , we
rep
have that there exists S D[J 1] such that S Y  = and S Y  J . By

Equation 1, S {J} in N i+1 , because S {J} J . Now we set S = S {J}.

Observe that S Y = and S Y J . This completes the proof of the claim.  

We ll the entry for D[i + 1] as follows.

k  
D[i + 1] = N (i+1)j kj N (i+1)j (2)
rep
j=1

In Equation 2, for every 1 j k, N (i+1)j denote the subset of N (i+1) containing

sets of size exactly j and N (i+1)j can be computed using Lemma 2. Lemma 1
and Claim 2.2 implies that D[i + 1] 1k rep P
i+1
.
Now we analyse the running time of the algorithm. Consider the time to
compute D[i + 1]. We already have computed the family corresponding to D[r] for
all r [i]. By Lemma 2, for any r [i] and j [k], the
subset of D[r] containing
sets of size exactly j is upper bounded by |I| k kj . Hence, the cardinality


of N (i+1)j is upper bounded by |I|2 n k kj . Thus, by Lemma 2, the time


 
to compute N (i+1)j kj N (i+1)j is bounded by k 2 + k (n + |I|)O(1) =
rep j j

k  O(1)
(n + |I|) number of operation over the eld in which AM is given and
j

k 

|N (i+1)j
| |I| k . Hence the total running time to compute D[i + 1] for any
j
i + 1 [n] is
k  
 k
(n + |I|)O(1) ) = 2k (n + |I|)O(1) .
j=1
j

By Lemma 2, the cardinality of D[i + 1] is bounded by,


k 
k  
(i+1)j | k
|D[i + 1]| = |N |I| k = 2k |I| k.
j=1 j=1
j

This completes the proof. 



Theorem 3 follows from Lemmata 3 and 4. Now we explain an application of

Theorem 3. Consider the problem (P, I )-Graphical CF-SC, where CGI is a
cluster graph. Let (P, I, CGI , k) be an instance of (P, I )-Graphical CF-SC.
72 A. Banik et al.

Let C1 , . . . Ct be the connected components of CGI , where each Ci is a clique for

all i [t]. In any solution we are allowed to pick at most one vertex (an interval)
from Ci for any i [t]. This information can be encoded using a partition matroid
M = (I = V (C1 )  . . .  V (Ct ), J ) where any subset I  I is independent in M
if and only if |I  V (Ci )| 1 for any i [t]. Moreover, a linear representation of
a partition matroid can be found in polynomial time ([15, Proposition 3.5]). As
a result, by applying Theorem 3 and Proposition 3.5 of [15], we get the following
corollary.
Corollary 1. (P, I )-Graphical CF-SC, when CGI is a cluster graph, can
be solved in time 2k (n + |I|)O(1) .

References
1. Arkin, E.M., Banik, A., Carmi, P., Citovsky, G., Katz, M.J., Mitchell, J.S.B.,
Simakov, M.: Choice is hard. In: ISAAC. pp. 318328 (2015)
2. Arkin, E.M., Banik, A., Carmi, P., Citovsky, G., Katz, M.J., Mitchell, J.S.B.,
Simakov, M.: Conict-free covering. In: CCCG (2015)
3. Banik, A., Panolan, F., Raman, V., Sahlot, V.: Frechet distance between a line and
avatar point set. FSTTCS pp. 32:132:14 (2016)
4. Cygan, M., Fomin, F.V., Kowalik, L., Lokshtanov, D., Marx, D., Pilipczuk, M.,
Pilipczuk, M., Saurabh, S.: Parameterized algorithms. Springer (2015)
5. Diestel, R.: Graph Theory, 4th Edition, Graduate texts in mathematics, vol. 173.
Springer (2012)
6. Fomin, F.V., Lokshtanov, D., Panolan, F., Saurabh, S.: Ecient computation of
representative families with applications in parameterized and exact algorithms. J.
ACM 63(4), 29 (2016)
7. Gabow, H.N., Westermann, H.H.: Forests, frames, and games: Algorithms for
matroid sums and applications. Algorithmica 7(5&6), 465497 (1992)
8. Har-Peled, S., Quanrud, K.: Approximation algorithms for low-density graphs.
CoRR abs/1501.00721 (2015)
9. Har-Peled, S., Quanrud, K.: Approximation algorithms for polynomial-expansion
and low-density graphs. In: ESA. vol. 9294, pp. 717728. Springer (2015)
10. Hunt III, H.B., Marathe, M.V., Radhakrishnan, V., Ravi, S.S., Rosenkrantz, D.J.,
Stearns, R.E.: NC-approximation schemes for NP- and PSPACE-hard problems for
geometric graphs. J. Algorithms 26(2), 238274 (1998)
11. Karp, R.M.: Reducibility among combinatorial problems. In: Proceedings of a
symposium on the Complexity of Computer Computations. pp. 85103 (1972)
12. Lokshtanov, D., Misra, P., Panolan, F., Saurabh, S.: Deterministic truncation of
linear matroids. In: ICALP, Proceedings, Part I. vol. 9134, pp. 922934. Springer
(2015)
13. Lokshtanov, D., Panolan, F., Saurabh, S., Sharma, R., Zehavi, M.: Covering small
independent sets and separators with applications to parameterized algorithms.
ArXiv e-prints (May 2017)
14. Marx, D.: Parameterized complexity of independence and domination on geometric
graphs. In: IPEC, pp. 154165. Springer (2006)
15. Marx, D.: A parameterized view on matroid optimization problems. Theor. Comput.
Sci. 410(44), 44714479 (2009)
16. Naor, M., Schulman, J.L., Srinivasan, A.: Splitters and near-optimal derandomiza-
tion. In: FOCS. pp. 182191 (1995)
 Obedient Plane Drawings for Disk Intersection Graphs

Bahareh Banyassady1, , Michael Hoffmann2 ,

Boris Klemz1 , Maarten Lofer3, , Tillmann Miltzow4,  
1
Institute of Computer Science, Freie Universitat Berlin, Germany
2
Department of Computer Science, ETH Zurich, Switzerland
3
Dept. of Computing and Information Sciences, Utrecht University, the Netherlands
4
Universite libre de Bruxelles (ULB), Brussels, Belgium

Abstract. Let D be a set of disks and G be the intersection graph of D. A

drawing of G is obedient to D if every vertex is placed in its corresponding disk.
We show that deciding whether a set of unit disks D has an obedient plane straight-
line drawing is N P-hard regardless of whether a combinatorial embedding is
prescribed or an arbitrary embedding is allowed. We thereby strengthen a result by
Evans et al., who show N P-hardness for disks with arbitrary radii in the arbitrary
embedding case. Our result for the arbitrary embedding case holds true even if G
is thinnish, that is, removing all triangles from G leaves only disjoint paths. This
contrasts another result by Evans et al. stating that the decision problem can be
solved in linear time if D is a set of unit disks and G is thin , that is, (1) the (graph)
distance between any two triangles is larger than 48 and (2) removal of all disks
within (graph) distance 8 of a triangle leaves only isolated paths. A path in a disk
intersection graph is isolated if for every pair A, B of disks that are adjacent along
the path, the convex hull of A B is intersected only by disks adjacent to A or B.
Our reduction can also guarantee the triangle separation property (1). This leaves
only a small gap between tractability and N P-hardness, tied to the path isolation
property (2) in the neighborhood of triangles. It is therefore natural to study the
impact of different restrictions on the structure of triangles. As a positive result, we
show that an obedient plane straight-line drawing is always possible if all triangles
in G are light and the disks are in general position (no three centers collinear). A
triangle in a disk intersection graph is light if all its vertices have degree at most
three or the common intersection of the three corresponding disks is empty. We
also provide an efcient drawing algorithm for that scenario.

1 Introduction
Disk intersection graphs have been long studied in mathematics and computer science
due to their wide range of applications in a variety of domains. They can be used
to model molecular bonds as well as the structure of the cosmic web, interference in
communication networks and social interactions in crowds. Finding planar realizations of

supported by DFG project MU/3501-2

supported by the Netherlands Organisation for Scientic Research (NWO) under project no.
639.021.123 and 614.001.504

supported by the ERC grant PARAMTIGHT: Parameterized complexity and the search for
tight complexity results, no. 280152

Springer International Publishing AG 2017 73

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 7384, 2017.
DOI: 10.1007/978-3-319-62127-2_7
74 B. Banyassady et al.

D G

(a)

Fig. 1: A set D of disks, the induced graph G, and an obedient plane straight-line drawing of G.

disk intersection graphs is an important tool to model sensor networks so as to translate

the network connectivity information into virtual coordinates [11]. The planar model
can help with topology extraction [10, 15] and enable geometric routing schemes [5, 12].
Rapidly increasing data collection rates call for efcient algorithms to run simulations.
But computational problems involving sets of disks are notoriously difcult to grasp: they
often lie in a transition zone between tractable and intractable classes of shapes. In this
paper, we investigate the computational complexity of one such problem: determining
whether a disk intersection graph admits an obedient plane straight-line drawing.
Disk-obedient Drawings. A set D of disks in the plane induces a graph G = (V, E),
called the disk intersection graph, which has a vertex for every disk in D and an edge
between two vertices whose (closed) disks intersect. A straight-line drawing of G is
an injective map : V R2 so that for every edge uv E the open line segment
(u)(v) is disjoint from (V ). A drawing is plane (also called an embedding) if the
edges do not intersect except at common endpoints. A drawing is obedient if every
vertex is contained in its corresponding disk. Disk-obedient drawings were introduced by
Evans et al. [8], who prove that recognizing whether G admits an obedient plane straight-
line drawing is N P-hard. This decision problem is called P LANAR D ISK O BEDIENCE
R ECOGNITION. The motivation to study disk-obedience stems from dealing with data
uncertainty [7, 13]. The problem is strongly related to Anchored Planar Graph Drawing
(AGD) (shown to be N P-hard by Angelini et al. [2]): Given a planar graph G and an
associated unit disk for each circle, produce a planar embedding of G such that each
vertex is contained in its disk. Our problem is different from AGD since for us, G itself
is dened by the disks. Keszegh et al. [9] also study a related problem of placing vertices
of a disk intersection graph inside their respective disks such that the resulting drawing
is C-oriented; however, the disks in [9] can only touch (their interiors are disjoint),
ensuring the drawing is always a plane embedding.
Results and Overview. In this paper we show that several natural restrictions of
P LANAR D ISK O BEDIENCE R ECOGNITION remain N P-hard: even when the disks
are unit disks (Section 2), and / or the combinatorial embedding of the graph is given
(Section 4), the problem is still hard. In the former case, our result holds true even if
removing all triangles leaves only disjoint paths and even if all triangles are far apart. This
creates an interesting contrast to results by Evans et al. [8]. Since it is unclear whether
the problem is in N P (see Section 5), our results indicate that the problem is indeed
very hard, and it is probably difcult to attack the general problem using a combinatorial
Obedient Plane Drawings for Disk Intersection Graphs 75

approach. On the positive side, we show that the problem can be solved efciently if the
degree of vertices belonging to ply-3 triangles (that is, the three corresponding disks have
a common intersection) is bounded by three (Section 3). In this result no assumption
concerning the uniformity of the disks radii is needed. Due to space restrictions, some
of the proofs are sketched or omitted in this short paper.
Notation. Throughout this paper disks are closed disks in the plane. Let D be a
set of disks, let G = (V, E) be the intersection graph of D and let be a straight-line
drawing of G. We use capital letters to denote disks and non-capital letters to denote
the corresponding vertices, e.g., the vertex corresponding to disk D D is d V .
We use the shorthand uv to denote the edge {u, v} E. Further, uvw refers to the
triangle composed of the edges uv, vw and wu. We identify vertices and edges with their
geometric representations, e.g., uv also refers to the line segment between the points
representing u and v in . We use int(D) to refer to the interior of a disk D D. The
ply of a point p R2 with respect to D is the cardinality |{D D | p D}|. The ply
of D is the maximum ply of any point p R2 with respect to D.
Planar Montone 3-Satisability. Let = (U , C) be a 3-S ATISFIABILITY (3SAT)
formula where U denotes the set of variables and C denotes the set of clauses. We call
the formula monotone if each clause c C is either positive or negative, that is, all
literals of c are positive or all literals of c are negative. Note that this is not the standard
notion of monotone Boolean formulas. Formula = (U , C) is planar if its variable
clause graph G = (U  C, E) is planar. The graph G is bipartite and every edge in E
is incident to both a clause vertex from C and a variable vertex from U . The edge {c, u}
is contained in E if and only if a literal of variable u U occurs in c C. In the decision
problem P LANAR M ONOTONE 3-S ATISFIABILITY we are given a planar and monotone
3SAT formula together with a monotone rectilinear representation R of the variable
clause graph of and we need to decide whether is satisable. The representation R
is a contact representation on an integer grid, see Figure 2a. In R the variables are
represented by horizontal line segments arranged on a line . The clauses are represented
by E-shapes. All positive clauses are placed above  and all negative clauses are placed
below . P LANAR M ONOTONE 3-S ATISFIABILITY is N P-complete [4].

2 Thinnish Unit Disk Intersection Graphs

Evans et al. [8] showed that P LANAR D ISK O BEDIENCE R ECOGNITION is N P-hard.
Further, they provide a polynomial time algorithm to recognize disk-obedience graphs
for the case that the respective disk intersection graph is thin and unit. A disk intersection
graph G is thin if (i) the graph distance between any two triangles of G is larger than 48
and (ii) removal of all disks within graph distance 8 of a triangle decomposes the graph
into isolated paths. A path is isolated if for any pair of adjacent disks A and B of the
path, the convex hull of A B is intersected only by disks adjacent to A or B.
In this section we strengthen the N P-hardness result by Evans et al. by showing that
P LANAR D ISK O BEDIENCE R ECOGNITION is N P-hard even for unit disks. Further,
we show that the path-isolation property of thin disk intersection graphs is essential to
make the problem tractable. This is implied by the fact that our result holds even for
disk intersection graphs that are thinnish, that is, removing all disks that belong to a
76 B. Banyassady et al.

c c
c c
1 1
s s
u1 u2 u5
v v v v v v v v v v
u3 u4 u5 s s
1 1
u1 u2 u3 u4 u5 s s s s s s
1 1 1 1 1 1
u3 u 4 u 5 c c
u2 u3 u5 c c
u1 u2 u5 c c
(a) (b) (c)

Fig. 2: (a) A monotone rectilinear representation. For the purposes of illustration, the line segments
representing variables and clauses are thickened and labeled. (b) Gadget layout mimicking the
monotone rectilinear representation from (a) in the proof of Theorem 1. The v-nodes are variable
gadgets, the c-nodes are clause gadgets, the s-nodes are splitter gadgets and the (1)-nodes are
inverter gadgets. The input of the splitter gadgets are marked with an arrow. The black polygonal
paths represent wires. (c) Gadget layout for the proof of Theorem 3

triangle decomposes the graph into disjoint paths (which are not necessarily isolated).
Being thinnish does not impose any distance or other geometric constraint on the set of
disks and its intersection graph. Nevertheless, our reduction also works if the distance
between any two triangles is lower bounded by some value. In particular, this implies
that spacial separation of triangles is not sufcient for tractability.

Theorem 1. P LANAR D ISK O BEDIENCE R ECOGNITION is N P-hard even under any

combination of the following restrictions: (1) all disks have the same radius (unit disks);
(2) the intersection graph of the disks is thinnish; (3) the graph distance between any
two triangles is lower bounded by some value that is polynomial in the number of disks.

Proof. We describe a polynomial-time reduction from P LANAR M ONOTONE 3-SAT.

Let = (U, C) be a planar monotone 3SAT formula where U is the set of variables and C
is the set of clauses. On an intuitive level our reduction works as follows. We introduce
ve different types of gadgets. A variable gadget is created for each variable of U . The
gadget has two combinatorial states that are used to encode the truth state of its variable.
Wire gadgets are used to propagate these states to other gadgets. In particular, we create
a clause gadget for each clause c C and use wires to propagate the truth states of the
variables occurring in c to the clause gadget of c. The purpose of the clause gadget is to
enforces that at least one of the literals of c is satised. In order to appropriately connect
the variables with the clauses we require two more gadgets. The splitter gadget splits
a wire into two wires and the inverter gadget inverts the state transported along a wire.
The gadgets are arranged according to the monotone rectilinear representation R for ,
see Figure 2b. We proceed by describing our gadgets in detail.
Variables, Wires and Inverters. For each variable vertex we create a variable
gadget as depicted in Figure 3a. Note how on the left side of the obedient plane straight-
line drawing of the disk intersection graph the vertices of the lower path belong to the
Obedient Plane Drawings for Disk Intersection Graphs 77

(a) (b)

Fig. 3: The variable gadget has two possible states: narrow and wide. It can also be used as an
inverter as either the wire to the left or to the right has to be in narrow position.

upper path of disks and vice versa. We say these paths are in a narrow position. On
the other hand, the paths on the right side are in a wide position: the upper path in the
obedient plane straight-line drawing of the disk intersection graph belongs to the upper
path of disks. By ipping the combinatorial embedding of the subgraph induced by the
edges incident to the triangle vertices of the gadget, we can allow the left paths to be
in wide position and the right path to be in narrow position; see Figure 3b. However,
observe that (i) without ipping the embedding it is not possible to switch between
the narrow and the wide positions and (ii) exactly one side has to be in narrow and the
other in wide position in order to avoid edge-crossings. We shall use these two states of
the variable gadget to encode the truth state of the corresponding variable. The parallel
paths of disks to either side of the triangle act as wires that propagate the state of the
variable gadget. Figure 4a illustrates the information propagation and shows that it is not
necessary that the disk centers of the two parallel paths are collinear. Thus, wires are
very exible structures that allow us to transmit the truth states of the variable gadgets to
other gadgets in our construction. Observe that our variable gadget can also be used as
an inverter. If the narrow state is propagated to the triangle from one side, the other side
is forced to propagate the wide state and vice versa.
Clauses. For each clause vertex we create a clause gadget as depicted in Figure 4b.
Each of the three sets of disks {R1 , R2 , ...}, {G1 , G2 , ...} and {B1 , B2 , ...} belong to
one wire. Note that if a wire is in narrow position, the position of the vertices of the
wire is essentially unique up to an arbitrarily small wiggle room. The clause gadget is
designed such that if all three of its wires are in narrow position the disk intersection
graph can not be drawn obediently. The reason for this is that the unique positions of
the vertices of the triangle r1 g1 b1 in the middle of the gadget enforce a crossing, see
Figure 4b. However, if at least one of the incident wires is in wide position and, thus,
at least one of u, v or w can be placed freely in its disk, then the gadget can be drawn
obediently.
Splitters. The nal gadget in our construction is the splitter gadget, see Figure 5a.
It works as follows. The three sets of disks {R1 , R2 , ...}, {G1 , G2 , ...} and {B1 , B2 , ...}
belong to three wires r, g, b respectively. We also created a disk O which almost com-
pletely overlaps with G2 . Without the disk O the gadget would contain a vertex of
degree 3 that is not part of a triangle. The disk O articially creates a triangle so that the
resulting disk intersection graph is thinnish. We refer to the wire b as the input of the
gadget and to the wires r and g as the outputs. If the input is narrow then both outputs
have to be wide due to the unique positions of the vertices in the narrow wires. However,
78 B. Banyassady et al.

b4
B3 B4
b3 B2
G2 R2
b2
B1
G4 R4
g2 g1 r1
b1 r2
g3 r3 r4
g4
g5 G1 R1 r5
G3 R3
(a) (b)

Fig. 4: (a) A curved wire that propagates the narrow position between the two disks with dashed
boundary. (b) The clause gadget.

if the input is wide, then any of the outputs can have any combination of states. For
instance, both can be narrow. For our reduction we shall always connect the two outputs
to inverters so that if the input of the splitter is narrow, then the inverted outputs are also
narrow.
Layout. Figure 2b illustrates how we layout and combine our gadgets. We mimic
the monotone rectilinear representation R for . We place the variable gadgets and
clause gadgets according to R. Consider a variable u U . From the variable gadget of u
one wire wt leads to the top; another wire wb leads to the bottom. If u occurs as a literal
only once in a positive clause, the top wire wt leads directly to the corresponding clause
gadget. Otherwise, it leads to the input of a splitter. As stated earlier, we connect the
outputs of the splitter to inverters. We split the resulting wires recursively until we have
created as many wires as there are literals of u in positive clauses. We call the created
wires the children of wt and wt is their origin. Similarly, the wire wb that leads from the
variable gadget of u to the bottom is connected to the negative clauses in which u occurs
and the resulting wires are the children of wb and wb is their origin. Further, we refer
to u as the variable of wt and wb . Note that while in some of our gadgets we require
very precise coordinates, the required precision does not depend on the input size. Thus,
the construction can be carried out in polynomial time.
Correctness. It remains to argue that our reduction is correct. Recall that if the input
of a splitter is narrow then the outputs are wide. Since we place inverters at the outputs of
each splitter it follows that all children of a narrow wire w are also narrow. Conversely,
if a wire connected to a clause gadget is wide then it is a child of a wide wire.
Assume there exists an obedient plane straight-line drawing of the disk intersection
graph of the set of disks we created. We create a satisfying truth assignment for . In
an obedient plane straight-line drawing, for each clause gadget c there is at least one
wire w connected to c that is wide; otherwise there is a crossing in the subdrawing of the
clause gadget. Consequently, the origin of w is wide as well. If c is positive, we set the
variable of w to true and if c is negative we set the variable of w to false. Thus, we have
created a truth assignment in which for each clause, there is at least one satised literal.
Note that it is not possible that we set a variable to both true and false since a wire can
Obedient Plane Drawings for Disk Intersection Graphs 79

B3
G1 B2
G3
b3 r1
b2

B1 b1
o g1
g3 g2 r1
g4
R2
b1
R1
r2
O R4
G2 r3
G4

g1
r4
g2 o R3
r5

(a)

Fig. 5: The splitter gadget.

only be the origin of either positive of negative clauses due to the fact that in a monotone
rectilinear representation all negative clauses are below and all positive clauses are above
the variable line.
Finally, assume that is satisable. For each variable u we orient its variable gadget
such that the wire that leads to the positive clauses is wide if and only if u is true. We
draw the splitter gadgets such that all children of wide wires are wide. Since every clause
has a satised literal, the corresponding clause gadget is connected to a wide wire and,
thus, can be drawn without introducing crossing.
Spacial Separation. Note that the splitter, variable, clause and inverter gadgets each
contain one triangle and recall that all gadgets are connected by wires, which do not
contain triangles. Thus, it is straightforward to ensure a minimum distance between any
two triangles by simply increasing the length of our wires accordingly. 


3 Disk Intersection Graphs with Light Triangles

In Section 2 we give an N P-hardness proof for a very restricted class of instances of
P LANAR D ISK O BEDIENCE R ECOGNITION. A key ingredient of our reduction from
3SAT is a triangle in the middle of each variable gadget. The two truth states of the
gadget are encoded by the combinatorial embedding of the subgraph induced by edges
incident to the triangle vertices. It seems natural to study recognition of disk obedience
graphs where the degree of vertices in triangles and, thus, the number of combinatorial
80 B. Banyassady et al.

H C
c
h r R A pac pbc
x q p
l p a pab b
t
T B
L

(a) (b)

Fig. 6: (a) Every crossing between centered edges implies the existence of an arrow. (b) In a
centered drawing, every point in a ply-3 triangle abc belongs to one of the three disks.

embeddings for the triangle-induced subgraphs is bounded. This motivates the following
denition. Let D be a set of disks and G = (V, E) be the intersection graph of D. A
triangle abc of G is called light if deg(a), deg(b), deg(c) 3 or if A B C = . We
say that D is light if every triangle of G is light. In this section we show that for any
light set D of disks there always exists an obedient plane straight-line drawing of the
intersection graph of D. Note that we do not require the disks in D to have unit radius.
We begin by introducing some notations and by stating some helpful observations. A
set of disks is connected if the union of all disks is connected. A set D of disks is said
to be in general position if for any connected subset D D, |D | = 3 the disk centers
of D are non-collinear. Let G = (V, E) be the intersection graph of a set of disks D
and be a straight-line drawing of G. A vertex v V that is placed at its respective
disks center in is called centered. An edge e E between two centered vertices in
is called centered. The drawing is called centered if all vertices in V are centered. An
arrow (h, t, l, r) in a straight-line drawing of a graph G = (V, E) is a sequence of
vertices h, t, l, r V such that ht, hl, hr, lr E and such that ht and lr cross in , see
Figure 6a. We refer to h, t, l, r as the arrows head, tail, left and right vertex respectively.

Evans et al. [8] show that any set of disks with ply 2 in general position admits an
obedient plane straight-line drawing. We observe that with some minor adaptations, their
observation furthermore yields an explicit statement regarding the graph structure in
non-plane centered drawings. We restate their proof together with our modications
to show that every crossing between centered edges implies the existence of an arrow.
Furthermore, we strengthen this statement by showing that if the intersection point x of
the two edges is contained in the interior of one of the corresponding disks, then there
always is an arrow whose heads disk contains x in its interior.

Lemma 1. Let G = (V, E) be the disk intersection graph of a set of disks D in general
position. Let be a straight-line drawing of G. For any crossing in between centered
edges ab and cd there exists an arrow (h, t, l, r) where H L R lr = and either
(i) ht = ab and lr = cd, or (ii) ht = cd and lr = ab. Furthermore, if x = ab cd is
contained in the interior of at least one of A, B, C, D, then x int(H).
Obedient Plane Drawings for Disk Intersection Graphs 81

Next, we observe that if A B C = , then in a centered drawing any point in the

triangle abc is contained in one of the disks A, B or C.

Observation 1. Let D = {A, B, C} be a set of disks whose intersection graph G is the

triangle abc and let be a centered drawing of G. If the closed triangle abc contains a
point p AB C then any point q in the closed triangle abc is contained in AB C.

Our nal auxiliary result states that under certain conditions, the number of crossings
along one edge in a centered drawing is at most one.

Lemma 2. Let D be a light set of disks in general position and let G = (V, E) be the
intersection graph of D. Let be a straight-line drawing of G. Let (a, b, c, d) be an
arrow in where a, b, c, d are centered and A C D cd = . Then there exists no
centered edge other than ab that crosses cd in .

Proof Sketch. According to the denition of an arrow, we know that ab, cd, ac, ad E,
see Figure 7c (left). Assume that some centered edge f g = ab crosses cd. According
to Lemma 1, the intersection of f g and cd implies existence of an arrow A consisting
of c, d and two other vertices f and g, at least one of which has to be different from a
and b. Since acd is a light triangle in G and since A C D = , the degree of a, c, d
is bounded by 3. Due to deg(a) 3 the vertex a can not be connected to any vertex
other than b, c, d, which means that a cannot be an endpoint of the edge f g. Note that b
could be equal to f or g. We perform a case distinction regarding the containment of the
edges bc and bd in E.
First, assume that bc E and bd E, see Figure 7a (left). Because of the degree
restrictions for c and d, neither of them can be adjacent to a vertex different from a, b, c, d.
Thus, cd can not be the head-tail edge of A since the head of A is connected to f and
g. If cd is the left-right edge of A, then the head of A has to be b and without loss of
generality f is equal to b. Due to Lemma 1 we know that B C D = and, thus,
since bcd is light the degree of b is bounded by 3. However, this is a contradiction to the
fact that b is connected to g.
Now, assume that bc E and bd / E, see Figure 7b. Similar to the last case, the
degree restriction for c implies that c can not be adjacent to any vertex other than a, b, d.
This implies that c can not be the head of A and it can be the right or the left vertex
of A only if b is the head of A. In this case cd is the left-right edge of A and d has to
be connected to the head b of A, which contradicts bd / E. Thus, c can only be the tail
of A, and so d is the head. The head d is connected to f and g both of which have to
be different from a and b due to deg(a) 3 and due to bd / E respectively. This is a
contradiction to deg(d) 3.
We sketch the nal case. Assume bc / E and bd / E, see Figure 7c (left). Due to
degree restrictions we see that cd is the left-right edge of A and w.l.o.g. g = b is the
head. Head g can not be located inside triangle acd since this would imply an additional
crossing with ab, which contradicts degree restrictions. If g is exterior to acd and f is
interior, Observation 1 implies a contradiction to the degree bounds of a, c, d. If g and
f are exterior to acd, f g has to cross ad or ac, which implies the existence of another
arrow, again contradicting degree restrictions. 

82 B. Banyassady et al.

a d d a d

a

c b c b c b
(a) (b)
a d d d e d e
 
a b
a
a b
c b c b c f c f
(c) (d)

Fig. 7: Removing crossing by moving vertices.

Theorem 2. Let D be a light set of disks in general position whose intersection graph
is G. Then G has a plane straight-line drawing obedient to D.
Proof. We describe an iterative approach that transforms a centered drawing of G into a
crossing-free drawing obedient to D. In each step we change the position of precisely
one, two or three vertices to remove one or two crossings. During the entire procedure,
each vertex is moved at most once. We maintain the following invariant: After and before
each step, all crossing edges are centered. We proceed by describing our algorithm. After
that we show that the invariant is maintained and, thus, that the algorithm is correct.
Algorithm. Let ab and cd be two centered edges that cross in a point x. By Lemma 1
there exists an arrow consisting of a, b, c, d, w.l.o.g. (a, b, c, d), where AC Dcd = ,
see Figure 7c (left). Note that this implies that the degree of a, c, d is bounded by 3
since D is light and since A C D = . In order to remove the crossing we move
some of a, b, c, d and we use a , b , c , d to denote the new postions of these vertices.
We distinguish two cases. First assume that x A B C D and x / int(A)
int(B) int(C) int(D), i.e. x is on the boundary of all four disks. In this case we
set a = x and we move vertices c, d by a distance R+ in the direction given by


the vector ba, see Figure 7a. Value should be chosen small enough such that c C,
d D and c d does not cross an edge ef , unless cd already crosses ef .


Now we consider the case that x is in the interior of at least one of A, B, C, D.

By Lemma 1 we can assume without loss of generality that x int(A). In order to


remove the crossing we move a in the direction given by the vector ab for distance
|ax| + < |ab|, see Figure 7c. The value R+ should be chosen small enough
such that a A and such that there is no crossing between a c or a d and any other
edge ef , unless ef also intersects cd. To shorten notation, we refer to this procedure
as removing the crossing by moving a to x, although technically a is close to x
but a = x. It remains to treat the special case that ab has a crossing with some additional
edge ef = cd. In this case, in addition to moving a to x, we move b to y = ab ef , see
Figure 7d. In the following paragraphs we will see that there exists at most one such
additional crossing edge ef and that e, f are distinct from c, d as illustrated.
Correctness. Clearly our invariant holds for the initial centered drawing. In order
to show that our invariant is maintained it sufces to show that none of the moved
Obedient Plane Drawings for Disk Intersection Graphs 83

vertices is incident to an edge that has a crossing. In our algorithm we considered two
main cases. In the rst case we moved vertices a, c and d, which formed a complete
graph with vertex b. Note that all these vertices have exactly degree three due to the
degree restrictions. Therefore none of them can be adjacent to a vertex different from
a, b, c, d. In the second case we moved vertex a, which is adjacent to exactly b, c and d
due the degree restriction on a. Therefore, in both cases the moved vertices can only be
incident to edges with crossing if there exists some edge ef = ab, cd that has a crossing
with ab or cd. According to Lemma 2 there is no edge that crosses cd except for ab.
Due to Lemma 1, if ef crosses ab, there exists an arrow A composed of the edge ef
and ab. According to Lemma 2, ab can not be the left-right edge of A. The degree of a
is bounded by 3 and, thus, a has to be the tail, b has to be the head and e, f have to be
the left and right vertex of A.
We perform a case distinction regarding the equivalence of e, f and c, d. First,
asumme that {e, f } = {c, d} and without loss of generality e = d and f = c. Then A
consists of a, b, c, d. Next, assume that exactly one of e, f is equal to one of c, d and
without loss of generality e = d and f = a, b, c, d. Then d is adjacent to a, c, b, f ,
which contradicts the degree bound for d. Finally assume that both e, f are distinct
from a, b, c, d. By Lemma 2, neither cd nor ef can have a crossing with any edge other
than ab. Hence, the situation looks like the one illustrated in Figure 7d. In this case, in
addition to moving a to x, we move b to y = ab ef as described above. Now, a b
does not have any crossing and, thus, a is not incident to an edge with a crossing. For
symmetric reasons, neither is b . 


4 Embedded Unit Disk Intersection Graphs

In Section 2 we proved that P LANAR D ISK O BEDIENCE R ECOGNITION is N P-hard
even for disk intersection graphs that are unit and thinnish. In the reduction from 3SAT
used for the proof, the truth state of a variable gadget corresponds to the combinatorial
embedding of the respective subgraph. The N P-hardness proof by Evans et al. [8]
also establishes a correspondence between truth states and combinatorial embeddings.
This raises the question, whether N P-hardness holds if a combinatorial embedding
is prescribed. The following theorem answers this question in the afrmative. On a
high level, the proof idea is a reduction from P LANAR M ONOTONE 3-S ATISFIABILITY
similar to the one in the proof of Theorem 1. However, for the reduction in Theorem 3
we have to heavily rely on geometric arguments rather than combinatorial embeddings
to encode the truth states of variable and wire gadgets.

Theorem 3. P LANAR D ISK O BEDIENCE R ECOGNITION is N P-hard even for embed-

ded unit disk intersection graphs.

5 Remarks and Open Problems

Other Shapes. The notion of obedient drawings naturally extends to other shapes. The
reduction strategies used in the hardness proofs in this paper and the paper by Evans et
al. [8] seem to apply for several other shapes as well, e.g., for unit squares. This raises
84 B. Banyassady et al.

the interesting question whether a more general statement can be made that captures all
these hardness results at once.
NP-Membership. For many combinatorial problems, showing N P-membership is
an easy exercise. For disk-obedience the question turns out to be much more intricate.
A naive idea to show NP-membership would be to guess the coordinates of all vertices.
However, it is not obvious that there always exists a rational representation of bounded
precision. Indeed, there are several geometric problems where this approach is known to
fail. In some cases an explicit rational representation may require an exponential number
of bits, in others optimal solutions may require irrational coordinates, see [1,3,14]. Many
problems initially not known to lie in N P turned out to be R-complete. The complexity
class R captures all computational problems that are equivalent under polynomial time
reductions to the satisability of arbitrary polynomial equations and inequalities over
the reals, see [6, 14]. We leave it as an open problem to determine the relation of disk
obedient plane straight-line drawings with respect to N P and R.
Acknowledgments. This work was initiated during the Fixed-Parameter Computa-
tional Geometry Workshop at Lorentz Center, April 48, 2016. We thank the organizers
and all participants for the productive and positive atmosphere.

References
1. Abrahamsen, M., Adamaszek, A., Miltzow, T.: Irrational guards are sometimes needed. arXiv
preprint arXiv:1701.05475 (2017)
2. Angelini, P., Lozzo, G.D., Bartolomeo, M.D., Battista, G.D., Hong, S., Patrignani, M., Roselli,
V.: Anchored drawings of planar graphs. In: Graph Drawing. Lecture Notes in Computer
Science, vol. 8871, pp. 404415. Springer (2014)
3. Bajaj, C.: The algebraic degree of geometric optimization problems. Discrete & Computational
Geometry 3(2), 177191 (1988)
4. de Berg, M., Khosravi, A.: Optimal binary space partitions for segments in the plane. Int. J.
Comput. Geometry Appl. 22(3), 187206 (2012)
5. Bose, P., Morin, P., Stojmenovic, I., Urrutia, J.: Routing with guaranteed delivery in ad hoc
wireless networks. Wireless networks 7(6), 609616 (2001)
6. Cardinal, J.: Computational geometry column 62. SIGACT News 46(4), 6978 (2015)
7. Evans, W., Kirkpatrick, D., Lofer, M., Staals, F.: Minimizing co-location potential of moving
entities. SIAM J. Comput. 45(5), 18701893 (2016)
8. Evans, W.S., van Garderen, M., Lofer, M., Polishchuk, V.: Recognizing a DOG is hard, but
not when it is thin and unit. In: FUN 2016. pp. 16:116:12 (2016)
9. Keszegh, B., Pach, J., Palvolgyi, D.: Drawing planar graphs of bounded degree with few
slopes. SIAM Journal on Discrete Mathematics 27(2), 11711183 (2013)
10. Kroller, A., Fekete, S.P., Psterer, D., Fischer, S.: Deterministic boundary recognition and
topology extraction for large sensor networks. In: SoDA 2006. pp. 10001009 (2006)
11. Kuhn, F., Moscibroda, T., Wattenhofer, R.: Unit disk graph approximation. In: Proceedings of
the 2004 joint workshop on Foundations of mobile computing. pp. 1723. ACM (2004)
12. Kuhn, F., Wattenhofer, R., Zhang, Y., Zollinger, A.: Geometric ad-hoc routing: of theory and
practice. In: PoDC. pp. 6372 (2003)
13. Lofer, M.: Data Imprecision in Computational Geometry. Ph.D. thesis, Utrecht University
(2009)
14. Matousek, J.: Intersection graphs of segments and R. arXiv preprint arXiv:1406.2636 (2014)
15. Wang, Y., Gao, J., Mitchell, J.S.B.: Boundary recognition in sensor networks by topological
methods. In: MobiCom 2006. pp. 122133 (2006)
 
-Greedy t-spanner

Gali Bar-On and Paz Carmi

Department of Computer Science,

Ben-Gurion University of the Negev, Israel

Abstract. We introduce a new geometric spanner, -Greedy, whose con-

struction is based on a generalization of the known Path-Greedy and
Gap-Greedy spanners. The -Greedy spanner combines the most desir-
able properties of geometric spanners both in theory and in practice.
More specically, it has the same theoretical and practical properties as
the Path-Greedy spanner: a natural denition, small degree, linear num-
ber of edges, low weight, and strong (1 + )-spanner for every > 0. The
-Greedy algorithm is an improvement over the Path-Greedy algorithm
with respect to the number of shortest path queries and hence with re-
spect to its construction time. We show how to construct such a spanner
for a set of n points in the plane in O(n2 log n) time.
The -Greedy spanner has an additional parameter, , which indicates
how close it is to the Path-Greedy spanner on the account of the number
of shortest path queries. For = t the output spanner is identical to the
Path-Greedy spanner, while the number of shortest path queries is, in
practice, linear.
Finally, we show that for a set of n points placed independently at ran-
dom in a unit square the expected construction time of the -Greedy al-
gorithm is O(n log n). Our analysis indicates that the -Greedy spanner
gives the best results among the known spanners of expected O(n log n)
time for random point sets. Moreover, analysis implies that by setting
= t, the -Greedy algorithm provides a spanner identical to the Path-
Greedy spanner in expected O(n log n) time.

1 Introduction

Given a set P of points in the plane, a Euclidean t-spanner for P is an undirected

graph G, where there is a t-spanning path in G between any two points in P . A
path between points p and q is a t-spanning path if its length is at most t times
the Euclidean distance between p and q (i.e., t|pq|).
The most known algorithm for computing t-spanner is probably the Path-
Greedy spanner. Given a set P of n points in the plane, the Path-Greedy spanner
algorithm creates a t-spanner for P as follows.

 It starts with a graph G having
a vertex set P , an empty edge set E and n2 pairs of distinct points sorted in
a non-decreasing order of their distances. Then, it adds an edge between p and

The research is partially supported by the Lynn and William Frankel Center for
Computer Science.

Springer International Publishing AG 2017 85

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 8596, 2017.
DOI: 10.1007/978-3-319-62127-2_8
86 G. Bar-On and P. Carmi

q to the set E if the length of the shortest path between p and q in G is more
than t|pq|, see Algorithm 1 for more details. It has been shown in [9, 8, 12, 11,
17, 21] that for every set of points, the Path-Greedy spanner has O(n) edges, a
bounded degree and total weight O(wt(M ST (P ))), where wt(M ST (P )) is the
weight of a minimum spanning tree of P . The main weakness of the Path-Greedy
algorithm is its time complexity the naive implementation

 of the Path-Greedy
algorithm runs in near-cubic time. By performing n2 shortest path queries,
where each query uses Dijkstras shortest path algorithm, the time complexity
of the entire algorithm reaches O(n3 log n), where n is the number of points in
P . Therefore, researchers in this eld have been trying to improve the Path-
Greedy algorithm time complexity. For example, the Approximate-Greedy algo-
rithm generates a graph with the same theoretical properties as the Path-Greedy
spanner in O(n log n) time [13, 19]. However, in practice there is no correlation
between the expected and the unsatisfactory resulting spanner as shown in [15,
16]. Moreover, the algorithm is complicated and dicult to implement.
Another attempt to build a t-spanner more eciently is introduced in [14,
15]. This algorithm uses a matrix to store the length of the shortest path between
every two points. For each pair of points, it rst checks the matrix to see if there
is a t-spanning path between these points. In case the entry in the matrix for this
pair indicates that there is no t-spanning path, it performs a shortest path query
and updates the matrix. The authors in [15] have conjectured that the number
of performed shortest path queries is linear. This has been shown to be wrong
in [5], as the number of shortest path queries may be quadratic. In addition, Bose
et al. [5] have shown how to compute the Path-Greedy spanner in O(n2 log n)
time. The main idea of their algorithm is to compute a partial shortest path
and then extend it when needed. However, the drawback of this algorithm is
that it is complex and dicult to implement. In [1], Alewijnse et al. compute
the Path-Greedy spanner using linear space in O(n2 log2 n) time by utilizing the
Path-Greedy properties with respect to the Well Separated Pair Decomposition
(WSPD). In [2], Alewijnse et al. compute a t-spanner in O(n log2 n log2 log n)
expected time by using bucketing for short edges and by using WSPD for long
edges. Their algorithm is based on the assumption that the Path-Greedy spanner
consists of mostly short edges.
Additional eort has been put in developing algorithms for computing t-
spanner graphs, such as -Graph algorithm [10, 18], Sink spanner, Skip-List
spanner [3], and WSPD-based spanners [6, 7]. However, none of these algorithms
produces a t-spanner as good as the Path-Greedy spanner in all aspects: size,
weight and maximum degree, see [15, 16].
Therefore, our goal is to develop a simple and ecient algorithm that achieves
both the theoretical and practical properties of the Path-Greedy spanner. In this
paper we introduce the -Greedy algorithm that constructs such a spanner for
a set of n points in the plane in O(n2 log n) time. Moreover, we show that for
a set of n points placed independently at random in a unit square the expected
running time of the -Greedy algorithm is O(n log n).
-Greedy t-spanner 87

Algorithm 1 Path-Greedy(P, t)
Input: A set P of points in the plane and a constant t > 1
Output: A
t-spanner
 G(V, E) for P
1: sort the n2 pairs of distinct points in non-decreasing order of their distances
and store them in list L
2: E
3: for (p, q) L consider pairs in increasing order do
4: length of the shortest path in G between p and q
5: if > t|pq| then
6: E := E |pq|
7: return G = (P, E)

2 -Greedy
In this section we describe the -Greedy algorithm (Section 2.1) for a given set
P of points in the plane, and two real numbers t and , such that 1 < t.
Then, in Section 2.2 we prove that the resulting graph is indeed a t-spanner with
bounded degree. Throughout this section we assume that < t (for example,
4
= t 5 or = 1+4t
5 ), except in Lemma 4, where we consider the case that = t.

2.1 Algorithm description

For each point p P we maintain a collection of cones Cp with the property
that for each point q P that lies in Cp there is a t-spanning path between p
and q in the current graph. The main idea of the -Greedy algorithm is to ensure
that two cones of a constant angle with apexes at p and q are added to Cp and
to Cq , respectively, each time the algorithm runs a shortest path query between
points p and q.
The algorithm starts with a graph G having a vertex set P , an empty edge
set, and an initially empty collection of cones Cp for each point p P . The
algorithm considers all pairs of distinct points of P in a non-decreasing order of
their distances. If p Cq or q Cp , then there is already a t-spanning path that
connects p and q in G, and there is no need to check this pair. Otherwise, let d
be the length of the shortest path that connects p and q in G divided by |pq|.

Let cp (, q) denote the cone with apex at p of angle , such that the ray pq is
its bisector. The decision whether to add the edge (p, q) to the edge set of G is
made according to the value of d. If d > , then we add the edge (p, q) to G, a
cone cp (2, q) to Cp , and a cone cq (2, p) to Cq , where = 4 arcsin( 12t ). If
d , then we do not add this edge to G, however, we add a cone cp (2, q) to
Cp and a cone cq (2, p) to Cq , where = 4 arcsin( d2t ).
In Algorithm 2, we give the pseudo-code description of the -Greedy algo-
rithm. In Figure 1, we illustrate a cone collection Cp of a point p and how it
is modied during the three scenarios of the algorithm. The gure contains the
point p, its collection Cp colored in gray, and three points v, u, and w, such
88 G. Bar-On and P. Carmi

that |pv| < |pu| < |pw|. Point v lies in Cp representing the rst case, where the
algorithm does not change the spanner and proceeds to the next pair without
performing a shortest path query. The algorithm runs a shortest path query be-
tween p and u, since u / Cp (for the purpose of illustration assume p / Cu ).
Figure 1(b) describes the second case of the algorithm, where the length of the
shortest path between p and u is at most |pu|. In this case the algorithm adds
a cone to Cp without updating the spanner. Figure 1(c) describes the third case
of the algorithm, where the length of the shortest path between p and w is more
than |pw|. In this case the algorithm adds a cone to Cp and the edge (p, w) to
the spanner.

Algorithm 2 -Greedy
Input: A set P of points in the plane and two real numbers t and s.t. 1 < t
Output: A
t-spanner
 for P
1: sort the n2 pairs of distinct points in non-decreasing order of their distances
(breaking ties arbitrarily) and store them in list L
2: E /* E is the edge set */
3: Cp p P /* Cp is set of cones with apex at p */
4: G (P, E) /* G is the resulting t-spanner */
5: for (p, q) L consider pairs in increasing order do
6: if (p / Cq ) and (q / Cp ) then
7: d length of the shortest path in G between p and q divided |pq|
8: if d > then
9: E E {(p, q)}
10: d 1
1
11: 4 arcsin( d2t ) /* cos sin t
= d */

12: cp (2, q) cone of angle 2 with apex at p and bisector pq

13: cq (2, p) cone of angle 2 with apex at q and bisector qp
14: Cp Cp cp (2, q)
15: Cq Cq cq (2, p)
16: return G = (P, E)

u u u
v v v
w w w

p p p

(a) (b) (c)

Fig. 1. Illustration of the three scenarios of the -Greedy algorithm. (a) v Cp ;

/ Cp and d ; (c) w
(b) u / Cp and d > .
-Greedy t-spanner 89

2.2 Algorithm analysis

In this section we analyze several properties of the -Greedy algorithm, including
the spanning ratio and the degree of the resulting graph.
The following lemma is a generalization of Lemma 6.4.1. in [20].
Lemma 1. Let t, and be real numbers, such that 1 t and 0
4.
Let p, q, and r be points in the plane, such that
1. p = r,
2. |pr| |pq|,
1
, where is the angle rpq (i.e., rpq = 4 arcsin( 2t )).
t
3. cos sin

Then |pr| + t|rq| t|pq|.

Proof. Let r be the orthogonal projection of r onto segment pq. Then, |rr | =
|pr| sin , |pr | = |pr| cos , and |r q| = |pq| |pr |. Thus, |r q| = |pq| |pr| cos .
By the triangle inequality

|rq| |rr | + |r q|

|pr| sin + |pq| |pr| cos
= |pq| |pr|(cos sin ).

We have, |pr| + t|rq| |pr| + t(|pq| |pr|(cos sin ))

= t|pq| t|pr|(cos sin ) + |pr|
t|pq| t|pr|(cos sin ) + t(cos sin )|pr|
t|pq|.

Lemma 2. The number of shortest path queries performed by -Greedy algo-

1
rithm for each point is O( t/1 ).

Proof. Clearly, the number of shortest path queries performed for each point is
at most n 1. Thus, we may assume that t/ > 1 + 1/n. Consider a point p P
and let (p, q) and (p, r) be two pairs of points that -Greedy algorithm has run
shortest path queries for. Assume w.l.o.g. that the pair (p, r) has been considered
before the pair (p, q), i.e., |rp| |pq|. Let d be the length of the path computed
by the shortest path query for (p, r) divide by |pr|. If d , then the cone added
to the collection Cp has an angle of at least 4 arcsin( 2t ). Otherwise, the
algorithm adds the edge (p, r) to G and a new cone to the collection of cones Cp ,
where the angle of this cone is 4 arcsin( 12t ). Thus, after the shortest path
query performed for the pair (p, r), the collection Cp contains a cone cp (, r),
where is at least 2 2 arcsin( 2t ). The -Greedy algorithm performs a shortest
path query for (p, q) only if p / Cq and q / Cp . Thus, the angle rpq is at least

arcsin( ), and we have at most k = 2 shortest path queries for a point.

4 2t
90 G. Bar-On and P. Carmi

Let us consider the case where t > 1 and t 1. The equation =

4
1
arcsin( 2t ) implies that cos sin t
= . Then, we have

t t
0, 1 + , and 1.

2 1
Thus, we have k = O( t/1 ). 

1
t

x1
Observation 1 For = t x , where x > 1 is a xed integer, the number of
x
shortest path queries performed by -Greedy algorithm for each point is O( t1 ).

Proof. As in Lemma 2, let us consider the case where t > 1 and t

1. Then,
we have
t t 1
0, 1 + , x1 1 + , t( x ) 1 + ,
(
t x )

t1
t (1 + )x , t 1 + x , and .
x
2x
Thus, we have k t1
x
= O( t1 ).


2
n log n
Observation 2 The running time of -Greedy algorithm is O( (t/1) 2 ).


Proof. The algorithm sorts the n2 pairs of distinct points in non-decreasing
order of their distances, this takes O(n2 log n) time. A shortest path query is
n
done by Dijkstras shortest path algorithm on a graph with O( t/1 ) edges and
n 1
takes O( t/1 +n log n) time. By Lemma 2 each point performs O( t/1 ) shortest
path queries. Therefore, we have that the running time of -Greedy algorithm is
n
O(( t/1 )2 log n). 


Observation 3 The number of cones that each point has in its collection along
1
the algorithm is constant depending on t and (O( t/1 )).

Proof. As shown in Lemma 2, the number of shortest path queries for each point
1
is O( t/1 ). The subsequent step of a shortest path query is the addition of two
cones, meaning that for each point p the number of cones in the collection of
1
cones Cp is O( t/1 ). 


Corollary 1. The additional space for each point p for the collection Cp is con-
stant.

Lemma 3. The output graph G = (P, E) of -Greedy algorithm (Algorithm 2)

is a t-spanner for P (for 1 < < t).

Proof. Let G = (P, E) be the output graph of the -Greedy algorithm. To prove
that G is a t-spanner for P we show that for every pair (p, q) P , there exists a
t-spanning path between them in G. We prove the above statement by induction
-Greedy t-spanner 91

on the rank of the distance |pq|, i.e., the place of (p, q) in a non-decreasing
distances order of all pairs of points in P .
Base case: Let (p, q) be the rst pair in the ordered list (i.e., the closest pair).
The edge (p, q) is added to E during the rst iteration of the loop in step 9 of
Algorithm 2, and thus there is a t-spanning path between p and q in G.
Induction hypothesis: For every pair (r, s) that appears before the pair (p, q)
in the ordered list, there is a t-spanning path between r and s in G.
The inductive step: Consider the pair (p, q). We prove that there is a t-
spanning path between p and q in G. If p / Cq and q / Cp , we check whether
there is a -spanning path in G between p and q. If there is a path which length
is at most |pq|, then |pq| t|pq|, meaning there is a t-spanning path between
p and q in G. If there is no path of length of at most |pq|, we add the edge (p, q)
to G, which forms a t-spanning path.
Consider that p Cq or q Cp , and assume w.l.o.g. that q Cp . Let (p, r)
be the edge handled in Step 5 in Algorithm 2 when the cone containing q has
been added to Cp (Step 12 in Algorithm 2). Notice that |pr| |pq|. Step 7 of
Algorithm 2 has computed the value d for the pair (p, r). In the algorithm there
are two scenarios depending on the value of d.
The rst scenario is when d > , then the algorithm has added the edge (p, r)
to G and a cone cp (, r) to Cp , where = 2( 4 arcsin( 12t )). Thus, the angle
between (p, q) and (p, r) is less than /2. Hence, |rq| < |pq| and by the induction
hypothesis there is a t-spanning path between r and q. Consider the shortest
path between p and q that goes through the edge (p, r). The length of this path
is at most |pr| + t|rq|. By Lemma 1, we have |pr| + t|rq| |pr| + t|rq| t|pq|
for = 1. Therefore, we have a t-spanning path between p and q.
The second scenario is when d , then the algorithm has added a cone
cp (, r) to Cp , where = 2( 4 arcsin( d2t )). Thus, the angle between (p, q)
and (p, r) is less than /2. Hence, |rq| < |pq| and by the induction hypothesis
there is a t-spanning path between r and q. Consider the shortest path between
p and q that goes through r. The length of this path is at most d|pr| + t|rq|. By
Lemma 1, we have d|pr| + t|rq| t|pq|. Therefore, we have a t-spanning path
between p and q. 

Theorem 4. The -Greedy algorithm computes a t-spanner for a set of points
P with the same properties as the Path-Greedy t-spanner, such as degree and
n
weight, in O(( t/1 )2 log n) time.
Proof. Clearly, the degree of the -Greedy is at most the degree of the Path-
Greedy -spanner. The edges of the -Greedy spanner satisfy the -leap frog
property, thus, the weight of the -Greedy is as Path-Greedy t-spanner. Hence,
we can pick close to t, such that we will have the required bounds. 

Lemma 4. If t = , the result of the -Greedy algorithm is identical to the result
of the Path-Greedy algorithm.
Proof. Assume towards contradiction that for t = the resulting graph of the
-Greedy algorithm, denoted as G = (P, E), diers from the result of the Path-
Greedy algorithm, denoted as G = (P, E  ). Assuming the same order of the
92 G. Bar-On and P. Carmi

sorted edges, let (p, q) be the rst edge that is dierent in G and G . Notice
that -Greedy algorithm decides to add the edge (p, q) to G when there is no
t-spanning path between p and q in G. Since until handling the edge (p, q) the
graphs G and G are identical, the Path-Greedy algorithm also decides to add
the edge (p, q) to G . Therefore, the only case we need to consider is (p, q) E 
and (p, q) / E. The -Greedy algorithm does not add an edge (p, q) to G in
two scenarios: (i) there is a t-spanning path between p and q in the current
graph G which contradicts that the Path-Greedy algorithm adds the edge
(p, q) to G ; (ii) p Cq or q Cp the -Greedy algorithm does not perform
a shortest path query between p and q. Assume w.l.o.g., q Cp , and let (p, r)
be the edge considered in Step 5 in Algorithm 2 when the cone containing q
has been added to Cp . The angle of the added cone is = 2 2 arcsin( d2t ),
where d is the length of the shortest path between p and r divided |pr|. Thus,
1
we have |pr| |pq| and cos sin d , where is the angle rpq. Then, by
t

Lemma 1, |pr| + t|rq| t|pq|, and since there is a path from p to r of length
at most |pr|, we have that there is t-spanning path between p and q in the
current graph. This is in contradiction to the assumption that the Path-Greedy
algorithm adds the edge (p, q) to E  . 


3 -Greedy in Expected O(n log n) Time for Random Set

In this section we show how a small modication in the implementation im-

proves the running time of the -Greedy algorithm. This improvement yields
an expected O(n log n) time for random point sets. The rst modication is to
run the shortest path query between points p to q up to |pq|. That is, running
Dijkstras shortest path algorithm with source p and terminating as soon as the
minimum key in the priority queue is larger than |pq|.
Let P be a set of n points in the plane uniformly distributed in a unit
square. To prove that -Greedy algorithm computes a spanner for P in expected
O(n log n) time, we need to show that:

each point runs a constant number of shortest path queries follows from
Lemma 2;
the expected number of points visited in each query is constant The fact
that the points are randomly chosen uniformly in the unit square implies
that the expected number of points at distance of at most r from point p
is (r2 n). A shortest path query from a point p to a point q terminates
as soon as the minimum key in the priority queue exceeds |pq|, thus, it is
expected to visit O(n (|pq|)2 ) points.
By Lemma 2 the number of shortest path queries performed by the algorithm
1
for a point p is O( t/1 ). Each such query denes a cone with apex at p of
angle (t/ 1), such that no other shortest path query from p will be
1
performed to a point in this cone. By picking k = t/1 and r = kn , we
have that the expected number of points around each point in a distance of
1
r is (k 2 ) = ( (t/1) 2 ).
-Greedy t-spanner 93

Assume we partition the plane into k equal angle cones with apex at point p.
The probability that there exists a cone that does not contain a point from
2
the set of points of distance kn is at most k (1 k1 )k . Let Q be the set
of points that p computed a shortest path query to, and let q Q be the
farthest point in Q from p. Then, the expected Euclidean distance between
p and q is less than kn . Thus, the expected number of points visited by the
2 2 2
k
entire set of shortest path queries from a point is O( t/1 ) = O( (t) 3 );

the next pair to be processed can be obtained in expected O(log n) time

without sorting all pairs of distinct points Even-though this is quite
straight forward, for completeness we give a short description how this can
be done. Divide the unit square to n n grid cells of side length 1/n. A hash
table of size 3n is initialized, and for each non-empty grid cell (at most n
such cells) we map the points in it to the hash table. In addition, we maintain
a minimum heap Hp for each point p P (initially empty), and one main
minimum heap H that contains the top element of each Hp . Each heap Hp
contains a subset of the pairs that include p.
For each point p P , all the cells of distance at most kn from p are scanned
(using the hash table) to nd all the points in these cells, where k is a
parameter that we x later. All the points found in these cells are added to
Hp according to their Euclidean distance from p.
The heap H holds the relevant pairs in an increasing order, therefore the
pairs are extracted from the main heap H. After extracting the minimum
pair in H that belongs to a point p, we add to H the next minimum in
Hp . To insure the correctness of the heaps, when needed we increase the
distance to the scanned cells. Observe that there may be a pair (p, q) such
that |pq| < |rw|, where the pair (r, w) is the top pair in H. This can occur
only when the pair (p, q) has not been added to Hp nor Hq , and this happens
when p Cq or q Cp . However, in this case we do not need to consider the
pair (p, q).
Notice that the only cells that are not contained in Cp are scanned to add
more pairs to Hp . Thus, points that are in Cp are ignored.
2

Therefore, the total expected running time of the algorithm is O( (t) 3 n log n).

Since both t and t/ are constants bigger than one, the expected running time
of the -Greedy algorithm is O(n log n).
A very nice outcome of -Greedy algorithm and its analysis can be seen
when is equal to t. Assume that -Greedy algorithm (for = t) has computed
a shortest path query for two points p and q and the length of the received path
is d|pq|. If the probability that t/d > 1 + is low (e.g, less than 1/2), for some
constant > 0, then -Greedy algorithm computes the Path-Greedy spanner
with linear number of shortest path queries. Thus -Greedy algorithm computes
the Path-Greedy spanner for a point set uniformly distributed in a square in
expected O(n log n) time.
Not surprisingly our experiments have shown that this probability is indeed
low (less than 1/100), since most of the shortest path queries are performed on
94 G. Bar-On and P. Carmi

pairs of points placed close to each other (with respect to Euclidean distance),
and thus with a high probability their shortest path contains a constant number
of points. Moreover, it seems that for a real-life input this probably is low.
Thus, there is a very simple algorithm to compute the Path-Greedy spanner in
expected O(n2 log n) time for real-life inputs, based on the -Greedy algorithm
For real-life input we mean that our analysis suggests that in the current
computers precision (Memory) one cannot create an instance of points set with
more than 1000 points, where the Path-Greedy spanner based on the -Greedy
algorithm has more than O(n2 log n) constructing time.

4 Experimental Results
In this section we discuss the experimental results by considering the properties
of the graphs generated by dierent algorithms and the number of shortest path
queries performed during these algorithms. We have implemented the Path-
Greedy, -Greedy, Gap-Greedy, -Graph, Path-Greedy on -Graph algorithms.
The Path-Greedy on -graph t-spanner, rst computes a -graph t -spanner,
where t < t, and then runs the Path-Greedy t/t -spanner on this t -spanner. The
shortest path queries criteria is used for an absolute running time comparison
that is independent of the actual implementation. The known theoretical bounds
for the algorithms can be found in Table 1.
The experiments were performed on a set of 8000 points, with dierent values
of the parameter (between 1 and t). We have chosen to present the parameter
for the values t, t0.9 and t. This values do not have special properties, they
where chosen arbitrary to present the behavior of the spanner.
To avoid the eect of specic instances, we have run the algorithms several
times and taken the average of the results. However, in all the cases the dierence
between the values is negligible. Table 24 show the results of our experiments
for dierent values of t and . The columns of the weight (divided by wt(M ST ))
and the degree are rounded to integers, and the columns of the edges are rounded
to one digit after the decimal point (in k).
The implementation details and the results analysis appear in [4].

Acknowledgments We would like to thank Rachel Saban for implementing

the algorithms.

Weight
Algorithm Edges wt(M ST ) Degree Time
1
n
Path-Greedy O( t1 ) O(1) O( t1 ) O(n3 log n)
n
Gap-Greedy O( t1 ) O(log n) 1
O( t1 ) O(n log2 n)
n
-Graph O( ) O(n) O(n) O( n log n)
1 1
-Greedy n
O( t/1 ) O(1) O( t/1 ) O( t/1 n2 log n)

Table 1. Theoretical bounds of dierent t-spanner algorithms

-Greedy t-spanner 95

Algorithm Edges (in K) Weight Degree Shortest path

wt(M ST ) queries (in K)
Path-Greedy - 35.6 10 17 31996
-Greedy 1.1 35.6 10 17 254
-Greedy 1.0896 37.8 12 18 242
-Greedy 1.048 51.6 19 23 204
-Graph - 376.6 454 149 -
Greedy on -Graph 1.0896 37.8 12 18 3005
Greedy on -Graph 1.048 52 19 23 693
Gap-Greedy - 51.6 19 23 326

Table 2. Comparison between several t-spanner algorithms for t = 1.1

Algorithm Edges (in K) Weight Degree Shortest path

wt(M ST ) queries (in K)
Path-Greedy - 15.1 3 7 31996
-Greedy 1.5 15.1 3 7 82
-Greedy 1.44 16 3 8 77
-Greedy 1.224 22.5 5 11 63
-Graph - 118.6 76 53 -
Greedy on -Graph 1.44 16 3 8 817
Greedy on -Graph 1.224 22.5 6 11 198
Gap-Greedy - 22.6 5 11 95

Table 3. Comparison between several t-spanner algorithms for t = 1.5

Algorithm Edges (in K) Weight Degree Shortest path

wt(M ST ) queries (in K)
Path-Greedy - 11.4 2 5 31996
-Greedy 2 11.4 2 5 55
-Greedy 1.866 11.9 2 5 52
-Greedy 1.414 16.3 3 8 44
-Graph - 85.3 48 42 -
Greedy on -Graph 1.866 11.9 3 6 493
Greedy on -Graph 1.414 16.5 3 8 129
Gap-Greedy - 16 3 8 63

Table 4. Comparison between several t-spanner algorithms for t = 2

96 G. Bar-On and P. Carmi

References
1. S. P. A. Alewijnse, Q. W. Bouts, A. P. ten Brink, and K. Buchin. Computing the
greedy spanner in linear space. Algorithmica, 73(3):589606, 2015.
2. S. P. A. Alewijnse, Q. W. Bouts, A. P. ten Brink, and K. Buchin. Distribution-
sensitive construction of the greedy spanner. Algorithmica, pages 123, 2016.
3. S. Arya, D. M. Mount, and M. H. M. Smid. Randomized and deterministic algo-
rithms for geometric spanners of small diameter. In FOCS, pages 703712, 1994.
4. G. Bar-On and P. Carmi. -greedy t-spanner. CoRR, abs/1702.05900, 2017.
5. P. Bose, P. Carmi, M. Farshi, A. Maheshwari, and M. H. M. Smid. Computing the
greedy spanner in near-quadratic time. In SWAT, pages 390401, 2008.
6. P. B. Callahan. Optimal parallel all-nearest-neighbors using the well-separated
pair decomposition. In FOCS, pages 332340, 1993.
7. P. B. Callahan and S. R. Kosaraju. A decomposition of multi-dimensional point-
sets with applications to k-nearest-neighbors and n-body potential elds. In STOC,
pages 546556, 1992.
8. B. Chandra. Constructing sparse spanners for most graphs in higher dimensions.
Inf. Process. Lett., 51(6):289294, 1994.
9. B. Chandra, G. Das, G. Narasimhan, and J. Soares. New sparseness results on
graph spanners. Int. J. Comp. Geom. and Applic., 5:125144, 1995.
10. K. L. Clarkson. Approximation algorithms for shortest path motion planning. In
STOC, pages 5665, 1987.
11. G. Das, P. J. Heernan, and G. Narasimhan. Optimally sparse spanners in 3-
dimensional Euclidean space. In SoCG, pages 5362, 1993.
12. G. Das and G. Narasimhan. A fast algorithm for constructing sparse Euclidean
spanners. Int. J. Comp. Geom. and Applic., 7(4):297315, 1997.
13. G. Das and G. Narasimhan. A fast algorithm for constructing sparse Euclidean
spanners. Int. J. Comp. Geom. and Applic., 7(4):297315, 1997.
14. M. Farshi and J. Gudmundsson. Experimental study of geometric t-spanners. In
ESA, pages 556567, 2005.
15. M. Farshi and J. Gudmundsson. Experimental study of geometric t-spanners: A
running time comparison. In WEA, pages 270284, 2007.
16. M. Farshi and J. Gudmundsson. Experimental study of geometric t-spanners. ACM
Journal of Experimental Algorithmics, 14, 2009.
17. J. Gudmundsson, C. Levcopoulos, and G. Narasimhan. Fast greedy algorithms for
constructing sparse geometric spanners. SIAM J. Comput., 31(5):14791500, 2002.
18. J. M. Keil. Approximating the complete Euclidean graph. In SWAT, pages 208
213, 1988.
19. C. Levcopoulos, G. Narasimhan, and M. H. M. Smid. Improved algorithms for
constructing fault-tolerant spanners. Algorithmica, 32(1):144156, 2002.
20. G. Narasimhan and M. Smid. Geometric Spanner Networks. Cambridge University
Press, New York, NY, USA, 2007.
21. J. Soares. Approximating Euclidean distances by small degree graphs. Discrete &
Computational Geometry, 11(2):213233, 1994.
 Dynamic Graph Coloring

Luis Barba1 , Jean Cardinal2 , Matias Korman3 , Stefan Langerman2 ,

Andre van Renssen4,5 , Marcel Roelozen4,5 , and Sander Verdonschot6
1
Dept. of Computer Science, ETH Zurich, Switzerland. [email protected]
2
Department dInformatique, Universite Libre de Bruxelles, Brussels, Belgium.
{jcardin,stefan.langerman}@ulb.ac.be
3
Tohoku University, Sendai, Japan. [email protected]
4
National Institute of Informatics, Tokyo, Japan. {andre,marcel}@nii.ac.jp
5
JST, ERATO, Kawarabayashi Large Graph Project.
6
School of Computer Science, Carleton University, Ottawa, Canada.
[email protected]

Abstract. In this paper we study the number of vertex recolorings that

an algorithm needs to perform in order to maintain a proper coloring of
a graph under insertion and deletion of vertices and edges. We present
two algorithms that achieve dierent trade-os between the number
of recolorings and the number of colors used. For any d > 0, the rst
algorithm maintains a proper O(CdN 1/d )-coloring while recoloring at most
O(d) vertices per update, where C and N are the maximum chromatic
number and maximum number of vertices, respectively. The second
algorithm reverses the trade-o, maintaining an O(Cd)-coloring with
O(dN 1/d ) recolorings per update. We also present a lower bound,
showing that any algorithm that maintains a c-coloring of a 2-colorable
2
graph on N vertices must recolor at least (N c(c1) ) vertices per update,
for any constant c 2.

1 Introduction

It is hard to underestimate the importance of the graph coloring problem in

computer science and combinatorics. The problem is certainly among the most
studied questions in those elds, and countless applications and variants have
been tackled since it was rst posed for the special case of maps in the mid-
nineteenth century. Similarly, the maintenance of some structures in dynamic
graphs has been the subject of study of several volumes in the past couple
of decades [1,2,11,18,19,20]. In this setting, an algorithmic graph problem is
modelled in the dynamic environment as follows. There is an online sequence of
insertion and deletion of edges or vertices, and our goal is to maintain the solution
of the graph problem after each update. A trivial way to maintain this solution is
to run the best static algorithm for this problem after each update; however, this

M. K. was partially supported by MEXT KAKENHI grant Nos. 12H00855, and
17K12635. A. v. R. and M. R. were supported by JST ERATO Grant Number
JPMJER1305, Japan. L. B. was supported by the ETH Postdoctoral Fellowship.
S. V. was partially supported by NSERC and the Carleton-Fields postdoctoral award.

Springer International Publishing AG 2017 97

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 97108, 2017.
DOI: 10.1007/978-3-319-62127-2_9
98 L. Barba et al.

is clearly not optimal. A dynamic graph algorithm seeks to maintain some clever
data structure for the underlying problem such that the time taken to update
the solution is much smaller than that of the best static algorithm.
In this paper, we study the problem of maintaining a coloring in a dynamic
graph undergoing insertions and deletions of both vertices and edges. At rst sight,
this may seem to be a hopeless task, since there exist near-linear lower bounds
on the competitive factor of online graph coloring algorithms [9], a restricted
case of the dynamic setting. In order to break through this barrier, we allow a
fair number of vertex recolorings per update. We focus on the combinatorial
aspect of the problem the trade-o between the number of colors used versus
the number of recolorings per update. We present a strong general lower bound
and two simple algorithms that provide complementary trade-os.
Denitions and Results. Let C be a positive integer. A C-coloring of a graph
G is a function that assigns a color in {1, . . . , C} to each vertex of G. A C-coloring
is proper if no two adjacent vertices are assigned the same color. We say that G
is C-colorable if it admits a proper C-coloring, and we call the smallest such C
the chromatic number of G.
A recoloring algorithm is an algorithm that maintains a proper coloring of a
simple graph while that graph undergoes a sequence of updates. Each update
adds or removes either an edge or a vertex with a set of incident edges. We say
that a recoloring algorithm is c-competitive if it uses at most c Cmax colors,
where Cmax is the maximum chromatic number of the graph during the updates.
For example, an algorithm that computes the optimal coloring after every
update is 1-competitive, but may recolor every vertex for every update. At the
other extreme, we can give each vertex a unique color, resulting in a linear
competitive factor for an algorithm that recolors at most 1 vertex per update. In
this paper, we investigate intermediate solutions that use more than C colors but
recolor a sublinear number of vertices per update. Note that we do not assume
that the value C is known in advance, or at any point during the algorithm.
In Section 2, we present two complementary recoloring algorithms: an O(dN 1/d )-
competitive algorithm with an amortized O(d) recolorings per update, and an
O(d)-competitive algorithm with an amortized O(dN 1/d ) recolorings per update,
where d is a positive integer parameter and N is the maximum number of vertices
in the graph during a sequence of updates. Interestingly, for d = (log N ), both
are O(log N )-competitive with an amortized O(log N ) vertex recolorings per
update. Using standard techniques, the algorithms can be made sensitive to the
current (instead of the maximum) number of vertices in the graph.
We provide lower bounds in Section 3. In particular, we show that for any
recoloring algorithm A using c colors, there exists a specic 2-colorable graph on N
vertices and a sequence of m edge insertions and deletions that forces A to perform
2
at least (m N c(c1) ) vertex recolorings. Thus, any x-competitive recoloring
1
algorithm performs in average at least (N x(2x1) ) recolorings per update.
To allow us to focus on the combinatorial aspects, we assume that we have
access to an algorithm that, at any time, can color the current graph (or an
induced subgraph) using few colors. Of course, nding an optimal coloring of an
Dynamic Graph Coloring 99

n-vertex graph is NP-complete in general [13] and even NP-hard to approximate

to within n1 for any  > 0 [22]. Still, this assumption is not as strong as it
sounds. Most practical instances can be colored eciently [4], and for several
important classes of graphs the problem is solvable or approximable in polynomial
time, including bipartite graphs, planar graphs, k-degenerate graphs, and unit
disk graphs [15].
Related results. Dynamic graph coloring. The problem of maintaining a col-
oring of a graph that evolves over time has been tackled before, but to our
knowledge, only from the points of view of heuristics and experimental results.
This includes for instance results from Preuveneers and Berbers [17], Ouerfelli
and Bouziri [16], and Dutot et al. [7]. A related problem of maintaining a graph-
coloring in an online fashion was studied by Borowiecki and Sidorowicz [3]. In
that problem, vertices lose their color, and the algorithm is asked to recolor them.
Online graph coloring. The online version of the problem is closely related
to our setting, except that most variants of the online problem only allow the
coloring of new vertices, which then cannot be recolored later. Near-linear lower
bounds on the best achievable competitive factor have been proven by Halldorsson
and Szegedy more than two decades ago [9]. They show their bound holds even
when the model is relaxed to allow a constant fraction of the vertices to change
color over the whole sequence. This, however, does not contradict our results. We
allow our algorithms to recolor all vertices at some point, but we bound only the
number of recolorings per update. Algorithms for online coloring with competitive
factor coming close, or equal to this lower bound have been proposed by Lovasz
et al. [14], Vishwanathan [21], and Halldorsson [8].
Dynamic graphs. Several techniques have been used for the maintenance of
other structures in dynamic graphs, such as spanning trees, transitive closure, and
shortest paths. Surveys by Demetrescu et al. [5,6] give a good overview of those.
Recent progress on dynamic connectivity [12] and approximate single-source
shortest paths [10] are witnesses of the current activity in this eld.

2 Upper bound: Recoloring-algorithms

For the description of our algorithms we consider only inserting a vertex with its
incident edges. Deletions cannot invalidate the coloring and edge insertions can
be done by removing and adding one of the vertices with the appropriate edges.
Our algorithms partition the vertices into a set of buckets, each of which has
its own distinct set of colors. All our algorithms guarantee that the subgraph
induced by the vertices inside each bucket is properly colored and this implies
that the entire graph is properly colored at all times.
The algorithms dier in the number of buckets they use and the size (maximum
number of vertices) of each bucket. Typically, there is a sequence of buckets of
increasing size, and one reset bucket that can contain arbitrarily many vertices
and that holds vertices whose color has not changed for a while. Initially, the size
of each bucket depends on the number of vertices in the input graph. As vertices
are inserted and deleted, the current number of vertices changes. When certain
100 L. Barba et al.

si si si si
1 1 1 1

s s 1
Fig. 1: The small-buckets algorithm uses d levels, each with s buckets of capacity si ,
1/d
where i is the level, s = NR , and NR is the number of vertices during the last reset.

buckets are full, we reset everything, to ensure that we can accommodate the
new number of vertices. This involves emptying all buckets into the reset bucket,
computing a proper coloring of the entire graph, and recomputing the sizes of
the buckets in terms of the current number of vertices.
We refer to the number of vertices during the most recent reset as NR , and
1/d
we express the size of the buckets in s = NR , where d > 0 is an integer
parameter that allows us to achieve dierent trade-os between the number
of colors and number of recolorings used. Since s = O(N 1/d ), where N is the
maximum number of vertices thus far, we state our bounds in terms of N . Note
that it is also possible to keep NR within a constant factor of the current number
of vertices by triggering a reset whenever the current number of vertices becomes
too small or too large. We omit these details for the sake of simplicity.

2.1 Small-buckets algorithm

Our rst algorithm, called the small-buckets algorithm, uses a lot of colors, but
needs very few recolorings. In addition to the reset bucket, the algorithm uses
ds buckets, grouped into d levels of s buckets each. All buckets on level i, for
0 i < d, have capacity si (see Fig. 1). Initially, the reset bucket contains
all vertices, and all other buckets are empty. Throughout the execution of the
algorithm, we ensure that every level always has at least one empty bucket. We
call this the space invariant.
When a new vertex is inserted, we place it in any empty bucket on level 0.
The space invariant guarantees the existence of this bucket. Since this bucket
has a unique set of colors, assigning one of them to the new vertex establishes a
proper coloring. Of course, if this was the last empty bucket on level 0, lling it
violates the space invariant. In that case, we gather up all s vertices on this level,
place them in the rst empty bucket on level 1 (which has capacity s and must
exist by the space invariant), and compute a new coloring of their induced graph
using the set of colors of the new bucket. If this was the last free bucket on level
1, we move all its vertices to the next level and repeat this procedure. In general,
if we lled the last free bucket on level i, we gather up all at most s si = si+1
vertices on this level, place them in an empty bucket on level i + 1 (which exists
by the space invariant), and recolor their induced graph with the new colors. If
we ll up the last level (d 1), we reset the structure, emptying each bucket into
Dynamic Graph Coloring 101

the reset bucket and recoloring the whole graph.

Theorem 2.1. For any integer d > 0, the small-buckets algorithm is an O(dN 1/d )-
competitive recoloring algorithm that uses at most O(d) amortized vertex recolor-
ings per update.
Proof. (sketch) The total number of colors is bounded by the maximum number
of non-empty buckets (1 + d(s 1)), multiplied by the maximum number of
colors used by any bucket. Let C be the maximum chromatic number of the
graph. Since any induced subgraph of a C-colorable graph is also C-colorable,
each bucket requires at most C colors. Thus, the total number of colors is at most
(1 + d(s 1))C, and the algorithm is O(dN 1/d )-competitive.
To analyze the number of recolorings, we use a simple charging scheme that
places coins in the buckets and pays one coin for each recoloring. Whenever we
place a vertex in a bucket on level 0, we give d + 2 coins to that bucket. One of
these coins is immediately used to pay for the vertexs new color, leaving d + 1
coins. In general, we can maintain the invariant that each non-empty bucket on
level i has si (d i + 1) coins from which the result follows. 


2.2 Big-buckets algorithm

Our second algorithm, called sd
the big-buckets algorithm, is s si
similar to the small-buckets
algorithm, except it merges
all buckets on the same level
Fig. 2: Besides the reset bucket, the big-buckets algo-
into a single larger bucket. rithm uses d buckets, each with capacity si+1 , where i
Specically, the algorithm is the bucket number.
uses d buckets in addition
to the reset bucket. These buckets are numbered sequentially from 0 to d 1,
with bucket i having capacity si+1 , see Fig. 2. Since we use far fewer buckets, an
upper bound on the total number of colors drops signicantly, to (d + 1)C. Of
course, as we will see later, we pay for this in the number recolorings. Similar
to the space invariant in the small-buckets algorithm, the big-buckets algorithm
maintains the high point invariant: bucket i always contains at most si+1 si
vertices (its high point).
When a new vertex is inserted, we place it in the rst bucket. Since this
bucket may already contain other vertices, we recolor all its vertices, so that the
subgraph induced by these vertices remains properly colored. This revalidates the
coloring, but may violate the high point invariant. If we lled bucket i beyond its
high point, we move all its vertices to bucket i + 1 and compute a new coloring
for this bucket. We repeat this until the high point invariant is satised, or we
ll bucket d 1 past its high point. In the latter case we reset, adding all vertices
to the reset bucket and computing a new coloring for the entire graph.
Theorem 2.2. For any integer d > 0, the big-buckets algorithm is an O(d)-
competitive recoloring algorithm that uses at most O(dN 1/d ) amortized vertex
recolorings per update.
102 L. Barba et al.

n1/3 1-trees
...

... ...
... ... ... ... ... ...

1/3

n2/3 1 leaves n2/3 1 leaves

n
9 1-trees

Fig. 3: (left) A 1-conguration is any forest that has many 1-trees as induced subgraphs.
(right) A 2-tree is constructed by connecting the roots of many 1-trees.

3 Lower bound
In this section we prove a lower bound on the amortized number of recolorings for
any algorithm that maintains a c-coloring of a 2-colorable graph, for any constant
c 2. We say that a vertex is c-colored if it has a color in [c] = {1, . . . , c}. For
simplicity of description, we assume that a recoloring algorithm only recolors
vertices when an edge is inserted and not when an edge is deleted, as edge deletions
do not invalidate the coloring. This assumption causes no loss of generality, as
we can delay the recolorings an algorithm would perform in response to an edge
deletion until the next edge insertion.
The proof for the lower bound consists of several parts. We begin with a specic
initial conguration and present a strategy for an adversary that constructs a
large conguration with a specic colouring and then repeatedly performs costly
operations in this conguration. In light of this strategy, a recoloring algorithm
has a few choices: it can allow the conguration to be built and perform the
recolorings required, it can destroy the conguration by recoloring parts of it
instead of performing the operations, or it can prevent the conguration from
being built in the rst place by recoloring parts of the building blocks. We show
that all these options require a large number of amortized recolorings.

3.1 Maintaining a 3-coloring

To make the general lower bound easier to understand, we rst show that to
maintain a 3-coloring, we need at least (n1/3 ) recolorings on average per update.
Lemma 3.1. For any suciently large n and any m 2n1/3 , there exists a
forest with n vertices, such that for any recoloring algorithm A, there exists a
sequence of m updates that forces A to perform (m n1/3 ) vertex recolorings to
maintain a 3-coloring throughout this sequence.
Proof. Let A be any recoloring algorithm that maintains a 3-coloring of a forest
under updates. We use an adversarial strategy to choose a sequence of updates
on a specic forest with n nodes that forces A to recolor many vertices. We
start by describing the initial forest structure.
A 1-tree is a rooted (star) tree with a distinguished vertex as its root and
Dynamic Graph Coloring 103

n2/3 1 leaf nodes attached to it. Initially, our forest consists of n1/3 pairwise
disjoint 1-trees, which account for all n vertices in our forest. The sequence of
updates we construct never performs a cut operation among the edges of a 1-tree.
Thus, the forest remains a 1-conguration: a forest of rooted trees with the n1/3
independent 1-trees as induced subgraphs; see Fig. 3 (left). We require that the
induced subtrees are not upside down, that is, the root of the 1-tree should be
closer to the root of the full tree than its children. Intuitively, a 1-conguration
is simply a collection of our initial 1-trees linked together into larger trees.
Let F be a 1-conguration. We assume that A has already chosen an initial
3-coloring of F . We assign a color to each 1-tree as follows. Since each 1-tree is
properly 3-colored, the leaves cannot have the same color as the root. Thus, a
2/3
1-tree T always has at least n 2 1 leaves of some color C, and C is dierent
from the color of the root. We assign the color C to T . In this way, each 1-tree
is assigned one of the three colors. We say that a 1-tree with assigned color C
becomes invalid if it has no children of color C left. Notice that to invalidate
2/3
a 1-tree, algorithm A needs to recolor at least n 2 1 of its leaves. Since the
1/3
coloring uses only three colors, there are at least n 3 1-trees with the same
assigned color, say X. In the remainder, we focus solely on these 1-trees.
1/3
A 2-tree is a tree obtained by merging n 9 1-trees with assigned color X, as
follows. First, we cut the edge connecting the root of each 1-tree to its parent, if
it has one. Next, we pick a distinguished 1-tree with root r, and connect the root
1/3
of each of the other n 9 1 1-trees to r. In this way, we obtain a 2-tree whose
1/3
root r has n2/3 1 leaf children from the 1-tree of r, and n 9 1 new children
that are the roots of other 1-trees; see Fig. 3 (right) for an illustration. This
1/3 1/3
construction requires n 9 1 edge insertions and at most n 9 edge deletions (if
every 1-tree root had another parent in the 1-conguration).
1/3 1/3
We build 3 such 2-trees in total. This requires at most 6( n 9 ) = 2n3 updates.
If none of our 1-trees became invalid, then since our construction involves only
1-trees with the same assigned color X, no 2-tree can have a root with color X.
Further, since the algorithm maintains a 3-coloring, there must be at least two
2-trees whose roots have the same color. We can now perform a matching link,
by connecting the roots of these two trees by an edge (in general, we may need
to perform a cut rst). To maintain a 3-coloring after a matching link, A must
recolor the root of one of the 2-trees and either recolor all its non-leaf children
or invalidate a 1-tree. If no 1-tree has become invalidated, this requires at least
n1/3
9 recolorings, and we again have two 2-trees whose roots have the same color.
Thus, we can perform another matching link between them. We keep doing this
1/3
until we either performed n 6 matching links, or a 1-tree is invalidated.
1/3
Therefore, after at most n1/3 updates ( 2n3 for the construction of the 2-trees,
n1/3
and 3 for the matching links), we either have an invalid 1-tree, in which case
n2/3 1 n1/3
A recolored at least 2 nodes, or we performed 6 matching links, which
1/3 1/3 2/3
forced at least n 6 n 9 = n54 recolorings. In either case, we forced A to
2/3
perform at least (n ) vertex recolorings, using at most n1/3 updates.
104 L. Barba et al.

Since no edge of a 1-tree was cut, we still have a valid 1-conguration, where
the process can be restarted. Consequently, for any m 2n1/3 , there exists a
sequence of m updates that starts with a 1-conguration and forces A to perform
2/3
1/3 (n
 nm ) = (m n1/3 ) vertex recolorings. 


3.2 On k-trees

We are now ready to

{
describe a general lower n2/c

{
{
{
2(c2) 2(ck+1) 2(ck)
bound for any number 0-trees n c(c1)
n c(c1) n c(c1)
of colors c. The general
1-trees (k 2)-trees (k 1)-trees
approach is the same as
when using 3 colors: We Fig. 4: A k-tree is constructed by connecting the roots of
construct trees of height a large number of (k 1)-trees.
up to c + 1, each exclud-
ing a dierent color for the root of the merged trees. By now connecting two such
trees, we force the algorithm A to recolor the desired number of vertices.
A 0-tree is a single node, and for each 1 k c, a k-tree is a tree obtained
2(ck)
recursively by merging 2 n c(c1) (k 1)-trees as follows: Pick a (k 1)-tree and
2(ck)
let r be its root. Then, for each of the 2 n c(c1) 1 remaining (k 1)-trees,
connect their root to r with an edge; see Fig. 4 for an illustration.
As a result, for each 0 j k 1, a k-tree T consists of a root r with
2(cj1)
2 n c(c1) 1 j-trees, called the j-subtrees of T , whose root hangs from r. The
root of a j-subtree of T is called a j-child of T . By construction, r is also the
root of a j-tree which we call the core j-tree of T .
Whenever a k-tree is constructed, it is assigned a color that is present among
a large fraction of its (k 1)-children. Indeed, whenever
 a k-tree is assigned a
2(ck)
color ck , we guarantee that it has at least 2c n c(c1) (k 1)-children of color ck .
We describe later how to choose the color that is assigned to a k-tree.
We say that a k-tree that was assigned color ck has a color violation if its
root no longer has a (k 1)-child with color ck . We say that a k-tree T becomes
invalid if either (1) it has a color violation or (2) if a core j-tree of T has a color
violation for some 1 j < k; otherwise we say that T is valid.
Observation 1 To obtain a color violation
 in a k-tree constructed by the above
2(ck)
2
procedure, A needs to recolor at least c n c(c1) vertices.

Notice that a valid c-colored k-tree of color ck cannot have a root with color ck .
Formally, color ck is blocked for the root of a k-tree if this root has a child with
color ck . In particular, the color assigned to a k-tree and the colors assigned to
its core j-trees for 1 j k 1 are blocked as long as the tree is valid.

3.3 On k-congurations
A 0-conguration is a set F0 of c-colored nodes, where |F0 | = T0 = n, for
some suciently large constant which will be specied later. For 1 k < c, a
Dynamic Graph Coloring 105

k-conguration is a set Fk of Tk k-trees, where

k 2(ci)
Tk = k
n1 i=1 c(c1) .
(4c)

Note that the trees of a k-conguration may be part of m-trees for m > k. If at
least T2k k-trees in a k-conguration are valid, then the conguration is valid.
For our construction, we let the initial conguration F0 be an arbitrary
c-colored 0-conguration in which each vertex is c-colored. To construct a k-
conguration Fk from a valid (k 1)-conguration Fk1 , consider the at least
Tk1
2 valid (k 1)-trees from Fk1 . Recall that the trees of Fk1 may be part
of larger trees, but since we consider edge deletions as free operations we can
separate the trees. Since each of these trees has a color assigned, among them at
least Tk1
2c have the same color assigned to them. Let ck1 denote this color.
2(ck)
Because each k-tree consists of 2 n c(c1) (k 1)-trees, to obtain Fk we merge
Tk1
2c (k 1)-trees of color ck1 into Tk k-trees, where

Tk1 1 k 2(ci)
Tk = 2(ck)
= k
n1 i=1 c(c1) .
2c 2 n c(c1) (4c)

Once the k-conguration Fk is constructed, we perform a color assign-

ment to each k-tree in Fk as follows: For a k-tree of Fk whose root has
2(ck)
2 n c(c1) 1 c-colored (k 1)-children, we assign a color that is shared by
2(ck)
at least 2c n c(c1) 1 of these (k 1)-children. Therefore, has at least
 2(ck)

2
c n children of its assigned color. After these color assignments, if each
c(c1)

(k 1)-tree used is valid, then each of the Tk k-trees of Fk is also valid. Thus,
Fk is a valid conguration. Moreover, for Fk to become invalid, A would need
to invalidate at least T2k of its k-trees. Since we use (k 1)-trees with the same
assigned color to construct k-trees, we can conclude the following about the use
of colors in any k-tree.

Lemma 3.2. Let Fk be a valid k-conguration. For each 1 j < k, each core
j-tree of a valid k-tree of Fk has color cj assigned to it. Moreover, ci = cj for
each 1 i < j < k.

We also provide bounds on the number of updates needed to construct a

k-conguration.
k
Lemma 3.3. Using ( i=j Ti ) = (Tj ) edge insertions, we can construct a
k-conguration from a valid j-conguration.

Proof. To merge Tk12c (k 1)-trees to into Tk k-trees, we need (Tk1 ) edge

insertions. Thus,
k in total, to construct a k-conguration from a j-conguration,
we need ( i=j Ti ) = (Tj ) edge insertions. 

106 L. Barba et al.

3.4 Reset phase

Throughout the construction of a k-conguration, the recoloring-algorithm A

may recolor several vertices which could lead to invalid subtrees in Fj for any
1 j < k. Because A may invalidate some trees from Fj while constructing Fk
from Fk1 , one of two things can happen. If Fj is a valid j-conguration for each
1 j k, then we continue and try to construct a (k + 1)-conguration from Fk .
Otherwise a reset is triggered as follows.
Let 1 j < k be an integer such that Fi is a valid i-conguration for each
0 i j 1, but Fj is not valid. Since Fj was a valid j-conguration with
at least Tj valid j-trees when it was rst constructed, we know that in the
T
process of constructing Fk from Fj , at least 2j j-trees where invalidated by A.
We distinguish two ways in which a tree can be invalid:

(1) the tree has a color violation, but all its j 1-subtrees are valid and no
core i-tree for 1 i j 1 has a color violation; or
(2) A core i-tree has a color violation for 1 i j 1, or the tree has a
color violation and at least one of its (j 1)-subtrees is invalid.

In case (1) the algorithm A has to perform fewer recolorings, but the tree can be
made valid again with a color reassignment, whereas in case (2) the j-tree has to
be rebuild.
Let Y0 , Y1 and Y2 respectively be the set of j-trees of Fj that are either valid,
T
or are invalid by case (1) or (2) respectively. Because at least 2j j-trees were
T
invalidated, we know that |Y1 | + |Y2 | > 2j . Moreover, for each tree in Y1 , A
2(cj)
recolored at least 2c n c(c1) 1 vertices to create the color violation on this
j-tree by Observation 1. For each tree in Y2 however, A created a color violation
in some i-tree for i < j. Therefore, for each tree in Y2 , by Observation 1, the
2(ci) 2(cj+1)
number of vertices that A recolored is at least 2c n c(c1) 1 > 2c n c(c1) 1.
Case 1: |Y1 | > |Y2 |. Recall that each j-tree in Y1 has only valid (j1)-subtrees
by the denition of Y1 . Therefore, each j-tree in Y1 can be made valid again by
performing a color assignment on it while performing no update. In this way,
T
we obtain |Y0 | + |Y1 | > 2j valid j-trees, i.e., Fj becomes a valid j-conguration
contained in Fk . Notice that when a color assignment is performed on a j-tree,
vertex recolorings previously performed on its (j 1)-children cannot be counted
again towards invalidating this tree.
Since we have a valid j-conguration instead of a valid k-conguration, we
wasted some edge insertions. We say that the insertion of each edge in Fk that is
not an edge of Fj is a wasted edge insertion. By Lemma 3.3, to construct Fk from
Fj we used (Tj ) edge insertions. That is, (Tj ) edge insertions became wasted.
However, while we wasted (Tj ) edge insertions, we also forced A to perform
2(cj) 2(cj)
(|Y1 | n c(c1) ) = (Tj n c(c1) ) vertex recolorings. Since 1 j < k c 1, we
2(cj) 2 2
know that n c(c1) n c(c1) . Therefore, we can charge A with (n c(c1) ) vertex
recolorings per wasted edge insertion. Finally, we remove each edge corresponding
to a wasted edge insertion, i.e., we remove all the edges used to construct Fk
Dynamic Graph Coloring 107

from Fj . Since we assumed that A performs no recoloring on edge deletions, we

are left with a valid j-conguration Fj .
T
Case 2: |Y2 | > |Y1 |. In this case |Y2 | > 4j . Recall that Fj1 is a valid
(j 1)-conguration by our choice of j. In this case, we say that the insertion
of each edge in Fk that is not an edge of Fj1 is a wasted edge insertion. By
Lemma 3.3, we constructed Fk from Fj1 using (Tj1 ) wasted edge insertions.
However, while we wasted (Tj1 ) edge insertions, we also forced A to perform
2(cj+1) 2(cj+1)
(|Y2 | n c(c1) ) = (Tj n c(c1) ) vertex recolorings. That is, we can charge
2(cj+1)
Tj
A with ( Tj1 n c(c1) ) vertex recolorings per wasted edge insertions. Since
2(cj) 2
Tj1
Tj = 4c n , we conclude that A was charged (n c(c1) ) vertex recolorings
c(c1)

per wasted edge insertion. Finally, we remove each edge corresponding to a wasted
edge insertion, i.e., we go back to the valid (j 1)-conguration Fj1 as before.

Regardless of the case, we know that during a reset consisting of a sequence

2
of h wasted edge insertions, we charged A with the recoloring of (h n c(c1) )
vertices. Notice that each edge insertion is counted as wasted at most once as the
edge that it corresponds to is deleted during the reset phase. A vertex recoloring
may be counted more than once. However, a vertex recoloring on a vertex v can
count towards invalidating any of the trees it belongs to. Recall though that v
belongs to at most one i-tree for each 0 i c. Moreover, two things can happen
during a reset phase that count the recoloring of v towards the invalidation of a
j-tree containing it: either (1) a color assignment is performed on this j-tree or
(2) this j-tree is destroyed by removing its edges corresponding to wasted edge
insertions. In the former case, we know that v needs to be recolored again in order
to contribute to invalidating this j-tree. In the latter case, the tree is destroyed
and hence, the recoloring of v cannot be counted again towards invalidating it.
Therefore, the recoloring of a vertex can be counted towards invalidating any
j-tree at most c times throughout the entire construction. Since c is assumed to
be a constant, we obtain the following result.

Lemma 3.4. After a reset phase in which h edge insertions become wasted, we
2
can charge A with (h n c(c1) ) vertex recolorings. Moreover, A will be charged
at most O(1) times for each recoloring.

If A stops triggering resets, then at some point we reach a (c1)-conguration

with at least Tc1 2(c + 1) trees and c + 1 valid ones. By linking together two
such trees with the same color we can force algorithm A to trigger a reset.

Theorem 3.5. Let c be a constant. For any suciently large integers n and
depending only on c, and any m = (n) suciently large, there exists a forest
F with n vertices, such that for any recoloring algorithm A, there exists a
2
sequence of m updates that forces A to perform (m n c(c1) ) vertex recolorings
to maintain a c-coloring of F .
108 L. Barba et al.

References
1. S. Baswana, M. Gupta, and S. Sen. Fully dynamic maximal matching in O(log n)
update time. SIAM J. on Comp., 44(1):88113, 2015.
2. S. Baswana, S. Khurana, and S. Sarkar. Fully dynamic randomized algorithms for
graph spanners. ACM Trans. on Alg., 8(4):35, 2012.
3. P. Borowiecki and E. Sidorowicz. Dynamic coloring of graphs. Fundamenta
Informaticae, 114(2):105128, 2012.
4. O. Coudert. Exact coloring of real-life graphs is easy. In Proc. 34th Design Autom.
Conf., pages 121126. ACM, 1997.
5. C. Demetrescu, D. Eppstein, Z. Galil, and G. F. Italiano. Dynamic graph algorithms.
In M. J. Atallah and M. Blanton, editors, Algorithms and Theory of Computation
Handbook. Chapman & Hall/CRC, 2010.
6. C. Demetrescu, I. Finocchi, and P. Italiano. Dynamic graphs. In D. Mehta and
S. Sahni, editors, Handbook on Data Structures and Applications, Computer and
Information Science. CRC Press, 2005.
7. A. Dutot, F. Guinand, D. Olivier, and Y. Pigne. On the decentralized dynamic
graph-coloring problem. In Proc. Worksh. Compl. Sys. and Self-Org. Mod., 2007.
8. M. M. Halldorsson. Parallel and on-line graph coloring. J. Alg., 23(2):265280,
1997.
9. M. M. Halldrsson and M. Szegedy. Lower bounds for on-line graph coloring. Theo.
Comp. Sci., 130(1):163 174, 1994.
10. M. Henzinger, S. Krinninger, and D. Nanongkai. A subquadratic-time algorithm
for decremental single-source shortest paths. In Proc. 25th ACM-SIAM Symp. on
Discr. Alg., pages 10531072, 2014.
11. J. Holm, K. De Lichtenberg, and M. Thorup. Poly-logarithmic deterministic
fully-dynamic algorithms for connectivity, minimum spanning tree, 2-edge, and
biconnectivity. J. ACM, 48(4):723760, 2001.
12. B. M. Kapron, V. King, and B. Mountjoy. Dynamic graph connectivity in polylog-
arithmic worst case time. In Proc. 24th ACM-SIAM Symp. on Discr. Alg., pages
11311142, 2013.
13. R. M. Karp. Reducibility among combinatorial problems. In Complexity of computer
computations, pages 85103. Plenum, New York, 1972.
14. L. Lovasz, M. E. Saks, and W. T. Trotter. An on-line graph coloring algorithm
with sublinear performance ratio. Discr. Math., 75(1-3):319325, 1989.
15. M. V. Marathe, H. Breu, H. B. Hunt, III, S. S. Ravi, and D. J. Rosenkrantz. Simple
heuristics for unit disk graphs. Networks, 25(2):5968, 1995.
16. L. Ouerfelli and H. Bouziri. Greedy algorithms for dynamic graph coloring. In
Proc. Int. Conf. on Comm., Comp. and Control App., pages 15, 2011.
17. D. Preuveneers and Y. Berbers. ACODYGRA: an agent algorithm for coloring
dynamic graphs. In Symb. Num. Alg. Sci. Comp., pages 381390, 2004.
18. L. Roditty and U. Zwick. Improved dynamic reachability algorithms for directed
graphs. In Proc. 43rd IEEE Sym. Found. Comp. Sci., pages 679688, 2002.
19. L. Roditty and U. Zwick. Dynamic approximate all-pairs shortest paths in undi-
rected graphs. In Proc. 45th IEEE Sym. Found. Comp. Sci., pages 499508, 2004.
20. M. Thorup. Fully-dynamic min-cut. Combinatorica, 27(1):91127, 2007.
21. S. Vishwanathan. Randomized online graph coloring. J. Alg., 13(4):657669, 1992.
22. D. Zuckerman. Linear degree extractors and the inapproximability of max clique
and chromatic number. Theory Comp., 3:103128, 2007.
 Universal Hinge Patterns for Folding Strips
Eciently into Any Grid Polyhedron

Nadia M. Benbernou1  , Erik D. Demaine2 ,

Martin L. Demaine2 , and Anna Lubiw3
1
Google Inc., [email protected].
2
MIT Computer Science and Articial Intelligence Laboratory,
32 Vassar St., Cambridge, MA 02139, USA, {edemaine,mdemaine}@mit.edu
3
David R. Cheriton School of Computer Science, University of Waterloo,
Waterloo, Ontario N2L 3G1, Canada, [email protected]

Abstract. We present two universal hinge patterns that enable a strip

of material to fold into any connected surface made up of unit squares on
the 3D cube gridfor example, the surface of any polycube. The folding
is ecient: for target surfaces topologically equivalent to a sphere, the
strip needs to have only twice the target surface area, and the folding
stacks at most two layers of material anywhere. These geometric results
oer a new way to build programmable matter that is substantially more
ecient than what is possible with a square N N sheet of material,
which can fold into all polycubes only of surface area O(N ) and may
stack (N 2 ) layers at one point. We also show how our strip foldings
can be executed by a rigid motion without collisions (albeit assuming
zero thickness), which is not possible in general with 2D sheet folding.
To achieve these results, we develop new approximation algorithms
for milling the surface of a grid polyhedron, which simultaneously give a
2-approximation in tour length and an 8/3-approximation in the number
of turns. Both length and turns consume area when folding a strip, so
we build on past approximation algorithms for these two objectives from
2D milling.

1 Introduction
In computational origami design, the goal is generally to develop an algorithm
that, given a desired shape or property, produces a crease pattern that folds into
an origami with that shape or property. Examples include folding any shape
[9], folding approximately any shape while being watertight [10], and optimally
folding a shape whose projection is a desired metric tree [14,15]. In all of these
results, every dierent shape or tree results in a completely dierent crease pat-
tern; two shapes rarely share many (or even any) creases.
The idea of a universal hinge pattern [6] is that a nite set of hinges (possible
creases) suce to make exponentially many dierent shapes. The main result
along these lines is that an N N box-pleat grid suces to make any polycube

Work performed while at MIT.

Springer International Publishing AG 2017 109

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 109120, 2017.
DOI: 10.1007/978-3-319-62127-2_10
110 N.M. Benbernou et al.

made of O(N ) cubes [6]. The box-pleat grid is a square grid plus alternating
diagonals in the squares, also known as the tetrakis tiling. For each target
polycube, a subset of the hinges in the grid serve as the crease pattern for that
shape. Polycubes form a universal set of shapes in that they can arbitrarily
closely approximate (in the Hausdor sense) any desired volume.
The motivation for universal hinge patterns is the implementation of pro-
grammable matter material whose shape can be externally programmed. One
approach to programmable matter, developed by an MITHarvard collabora-
tion, is a self-folding sheeta sheet of material that can fold itself into several
dierent origami designs, without manipulation by a human origamist [12,1].
For practicality, the sheet must consist of a xed pattern of hinges, each with
an embedded actuator that can be programmed to fold or not. Thus for the
programmable matter to be able to form a universal set of shapes, we need a
universal hinge pattern.
The box-pleated polycube result [6], however, has some practical limitations
that prevent direct application to programmable matter. Specically, using a
sheet of area (N 2 ) to fold N cubes means that all but a (1/N ) fraction of the
surface area is wasted. Unfortunately, this reduction in surface area is necessary
for a roughly square sheet, as folding a 11N tube requires a sheet of diameter
(N ). Furthermore, a polycube made from N cubes can have surface area as
low as (N 2/3 ), resulting in further wastage of surface area in the worst case.
Given the factor-(N ) reduction in surface area, an average of (N ) layers of
material come together on the polycube surface. Indeed, the current approach
can have up to (N 2 ) layers coming together at a single point [6]. Real-world
robotic materials have signicant thickness, given the embedded actuation and
electronics, meaning that only a few overlapping layers are really practical [12].
Our results: strip folding. In this paper, we in-
troduce two new universal hinge patterns that avoid
these ineciencies, by using sheets of material that
(a)
are long only in one dimension (strips). Specically,
Fig. 1 shows the two hinge patterns: the canonical
strip is a 1 N strip with hinges at integer grid lines
and same-oriented diagonals, while the zig-zag strip
is an N -square zig-zag with hinges at just integer grid
lines. We show in Section 2 that any grid surface
any connected surface made up of unit squares on the
3D cube gridcan be folded from either strip. The (b)
strip length only needs to be a constant factor larger
than the surface area, and the number of layers is Fig. 1. Two universal
at most a constant throughout the folding. Most of hinge patterns in strips.
our analysis concerns (genus-0) grid polyhedra, that (a) A canonical strip of
is, when the surface is topologically equivalent to a length 5. (b) A zig-zag
sphere (a manifold without boundary, so that every strip of length 6. The
edge is incident to exactly two grid squares, and with- dashed lines are hinges.
out handles, unlike a torus). We show in Section 4 that a grid polyhedron of
Universal Hinge Patterns for Folding Strips Efficiently into Any Grid Polyhedron 111

surface area N can be folded from a canonical strip of length 2N with at most
two layers everywhere, or from a zig-zag strip of length 4N with at most four
layers everywhere.
The improved surface eciency and reduced layering of these strip results
seem more practical for programmable matter. In addition, the panels of either
strip (the facets delineated by hinges) are connected acyclically into a path,
making them potentially easier to control. One potential drawback is that the
reduced connectivity makes for a imsier device; this issue can be mitigated by
adding tabs to the edges of the strips to make full two-dimensional contacts
across seams and thereby increase strength.
We also show in Section 5 an important practical result for our strip foldings:
under a small assumption about feature size, we give an algorithm for actually
folding the strip into the desired shape, while keeping the panels rigid (at) and
avoiding self-intersection throughout the motion. Such a rigid folding process
is important given current fabrication materials, which put exibility only in
the creases between panels [12]. An important limitation, however, is that we
assume zero thickness of the material, which would need to be avoided before
this method becomes practical.
Our approach is also related to the 1D chain robots of [7], but based on thin
material instead of thick solid chains. Most notably, working with thin material
enables us to use a few overlapping layers to make any desired surface without
scaling, and still with high eciency. Essentially, folding long thin strips of sheet
material is like a fusion between 1D chains of [7] and the square sheet folding of
[6,12,1].
Milling tours. At the core of our ecient strip foldings are ecient approxi-
mation algorithms for milling a grid polyhedron. Motivated by rapid-fabrication
CNC milling/cutting tools, milling problems are typically stated in terms of a
2D region called a pocket and a cutting tool called a cutter, with the goal
being to nd a path or tour for the cutter that covers the entire pocket. In our
situation, the pocket is the surface of the grid polyhedron, and the cutter
is a unit square constrained to move from one grid square of the surface to an
(intrinsically) adjacent grid square.
The typical goals in milling problems are to minimize the length of the tour
[3] or to minimize the number of turns in the tour [2]. Both versions are known
to be strongly NP-hard, even when the pocket is an integral orthogonal polygon
and the cutter is a unit square. We conjecture that the problem remains strongly
NP-hard when the pocket is a grid polyhedron, but this is not obvious.
In our situation, both length and number of turns are important, as both
inuence the required length of a strip to cover the surface. Thus we develop
one algorithm that simultaneously approximates both measures. Such results
have also been achieved for 2D pockets [2]; our results are the rst we know for
surfaces in 3D. Specically, we develop in Section 3 an approximation algorithm
for computing a milling tour of a given grid polyhedron, achieving both a 2-
approximation in length and an 8/3-approximation in number of turns.
112 N.M. Benbernou et al.

Fig. 2. Strip folding of individual letters typeface, AZ and 09: unfolded font (top)
and folded font (bottom), where the face incident to the bottom edge remains face-up.

Fonts. To illustrate the power of strip folding, we designed a typeface, rep-

resenting each letter of the alphabet by a folding of a 1 x strip for some x,
as shown in Fig. 2. The individual-letters typeface consists of two fonts: the un-
folded font is a puzzle to gure out each letter, while the folded font is easy to
read. These crease patterns adhere to an integer grid with orthogonal and/or
diagonal creases, but are not necessarily subpatterns of the canonical hinge pat-
tern. This extra exibility gives us control to produce folded half-squares as
desired, increasing the fonts delity.
We have developed a web app that visualizes the font.4 Currently in devel-
opment is the ability to chain letters together into one long strip folding; Fig. 9
at the end of the paper shows one example.
The full version of this paper [5] contains details omitted from this extended
abstract.

2 Universality

In this section, we prove that both the canonical strip and zig-zag strip of Fig. 1,
of sucient length, can fold into any grid surface. We begin with milling tours
4
http://erikdemaine.org/fonts/strip/
Universal Hinge Patterns for Folding Strips Efficiently into Any Grid Polyhedron 113

(a) (b)
Fig. 3. (a) Left and (b) right turn with a canonical strip.

which provide an abstract plan for routing the strip, and then turn to the details
of how to manipulate each type of strip.
Dual graph. Recall that a grid surface consists of one or more grid squares
that is, squares of the 3D cube gridglued edge-to-edge to form a connected
surface (ignoring vertex connections). Dene the dual graph to have a dual vertex
for each such grid square, and any two grid squares sharing an edge dene a dual
edge between the two corresponding dual vertices. Our assumption of the grid
surface being connected is equivalent to the dual graph being connected.
Milling tours. A milling tour is a (not necessarily simple) spanning cycle
in the dual graph, that is, a cycle that visits every dual vertex at least once (but
possibly more than once). Equivalently, we can think of a milling tour as the
path traced by the center of a moving square that must cover the entire surface
while remaining on the surface, and return to its starting point. Milling tours
always exist: for example, we can double a spanning tree of the dual graph to
obtain a milling tour of length less than double the given surface area.
At each grid square, we can characterize a milling tour as going straight,
turning, or U-turningintrinsically on the surfaceaccording to which two sides
of the grid square the tour enters and exits. If the sides are opposite, the tour is
straight; if the sides are incident, the tour turns; and if the sides are the same, the
tour U-turns. Intuitively, we can imagine unfolding the surface and developing
the tour into the plane, and measuring the resulting planar turn angle at the
center of the grid square.
Strip folding. To prove universality, it suces to show that a canonical strip
or zig-zag strip can follow any milling tour and thus make any grid polyhedron. In
particular, it suces to show how the strip can go straight, turn left, turn right,
and U-turn. Then, in 3D, the strip would be further folded at each traversed
edge of the grid surface, to stay on the surface. Indeed, U-turns can be viewed
as folding onto the opposite side of the same surface, and thus are intrinsically
equivalent to going straight; hence we can focus on going straight and making
left/right turns.
Canonical strip. Fig. 3 shows how a canonical strip can turn left or right;
it goes straight without any folding. Each turn adds 1 to the length of the strip,
and adds 2 layers to part of the grid square where the turn is made. Therefore
a milling tour of length L with t turns of a grid surface can be followed by a
canonical strip of length L + t. Furthermore, if the milling tour visits each grid
square at most c times, then the strip folding has at most 3c layers covering any
point of the surface.
114 N.M. Benbernou et al.

Fig. 4. Going straight with a zig-zag strip requires at most two unit squares per grid
square. Left and right crease patterns show two dierent parities along the strip.

(a) (b)
Fig. 5. Turning with a zig-zag strip has two cases because of parity. (a) Turning left
at an odd position requires three grid squares, whereas turning right requires one grid
square. (b) Turning left at an even position requires one grid square, whereas turning
right requires three grid squares.

Zig-zag strip. Fig. 4 shows how to fold a zig-zag strip in order to go straight.
In this straight portion, each square of the surface is covered by two squares of
the strip. Fig. 5 shows left and right turns. Observe that turns require either one
or three squares of the strip. Therefore a milling tour of length L with t turns
can be followed by a zig-zag strip of length at most 2L + t. Furthermore, if the
milling tour visits each grid square at most c times, then the strip folding has
at most 3c layers covering any point of the surface.

Proposition 1. Every grid surface of area N can be folded from a canonical

strip of length 4N , with at most eight layers stacked anywhere, and from a zig-
zag strip of length 6N , with at most twelve layers stacked anywhere.

The goal in the rest of this paper is to achieve better bounds for grid poly-
hedra, using more carefully chosen milling tours.
Universal Hinge Patterns for Folding Strips Efficiently into Any Grid Polyhedron 115

3 Milling Tour Approximation

This section presents a constant-factor approximation algorithm for milling a
(genus-0) grid polyhedron P with respect to both length and turns. Specically,
our algorithm is a 2-approximation in length and an 8/3-approximation in turns.
Our milling tours also have special properties that make them more amenable
to strip folding.
Our approach is to reduce the milling problem to vertex cover in a tripartite
graph. Then it follows that our algorithm is a 2-approximation in turns, where
is the best approximation factor for vertex cover in tripartite graphs. The
best known bounds on are 34/33 4/3. Clementi et al. [8] proved
that minimum vertex cover in tripartite graphs is not approximable within a
factor smaller than 34/33 = 1.03 unless P = NP. Theorem 1 of [13] implies
a 4/3-approximation for minimum weighted vertex cover for tripartite graphs
(assuming we are given the 3-partition of the vertex set, which we know in our
case). Thus we use = 4/3 below. An improved approximation ratio would
improve our approximation ratios, but may also aect the stated running times,
which currently assume use of [13].

3.1 Bands
The basis for our approximation algorithms is the notion of bands for a grid
polyhedron P . Let xmin and xmax respectively be the minimum and maximum
x coordinates of P ; dene ymin , ymax , zmin , zmax analogously. These minima and
maxima have integer values because the vertices of P lie on the integer grid.
Dene the ith x-slab Sx (i) to be the slab bounded by parallel planes x = xmin +i
and x = xmin + i + 1, for each i {0, 1, . . . , xmax xmin 1}. The intersection of
P with the ith x-slab Sx (i) (assuming i is in the specied range) is either a single
band (i.e., a simple cycle of grid squares in that slab), or a collection of such
bands, which we refer to as x-bands. Dene y-bands and z-bands analogously.
Two bands overlap if there is a grid square contained in both bands. Each grid
square of P is contained in precisely two bands (e.g., if a grid squares outward
normal were in the +z-direction, then it would be contained in one x-band and
one y-band). Two bands B1 and B2 are adjacent if they do not overlap, and a
grid square of B1 shares an edge with a grid square of B2 . A band cover for P is
a collection of x-, y-, and z-bands that collectively cover the entire surface of P .
The size of a band cover is the number of its bands.

3.2 Cover Bands

The starting point for the milling approximation algorithm is to nd an approx-
imately minimum band cover, as the minimum band cover is a lower bound on
the number of turns in any milling tour:

Proposition 2. [2, Lemma 4.9] The size of a minimum band cover of a grid
polyhedron P is a lower bound on the number of turns in any milling tour of P .
116 N.M. Benbernou et al.

Next we describe how to nd a near-optimal band cover. Consider the graph

GP with one vertex per band of a grid polyhedron P , connecting two vertices by
an edge if their corresponding bands overlap. It turns out that an (approximately
minimum) vertex cover in GP will give us an (approximately minimum) band
cover in P :

Proposition 3. A vertex cover for GP induces a band cover of the same size
and vice versa.

Because the bands fall into three classes (x-, y-, and z-), with no over-
lapping bands within a single class, GP is tripartite. Hence we can use an
-approximation algorithm for vertex cover in tripartite graphs to nd an -
approximate vertex cover in GP and thus an -approximate band cover of P .

3.3 Connected Bands

Our next goal will be to eciently tour the bands in the cover. Given a band
cover S for a grid polyhedron P , dene the band graph GS to be the subgraph
of GP induced by the subset of vertices corresponding to S. We will construct
a tour of the bands S based on a spanning tree of GS . Our rst step is thus to
show that GS is connected (Lemma 5 below). We do so by showing that adjacent
bands (as dened in Section 3.1) are in the same connected component of GS ,
using the following lemma of Genc [11]:

Lemma 4. [11]5 For any band B in a grid polyhedron P , let Nb be the bands of
P overlapping B. (Equivalently, Nb is the set of neighbors of B in GP ). Then
the subgraph of GP induced by NB is connected.

Lemma 5. If S is a band cover for a grid polyhedron P , then the graph GS is

connected.

3.4 Band Tour

Now we can present our algorithm for transforming a band cover into an ecient
milling tour.

Theorem 6. Let P be a grid polyhedron with N grid squares. In O(N 2 log N )

time, we can nd a milling tour of P that is a 2-approximation in length and an
8/3-approximation (or more generally, a 2-approximation) in turns.

We now state some additional properties of any milling tour produced by the
approximation algorithm of Theorem 6, which will be useful for later applications
to strip folding in Section 4.
5
Genc [11] uses somewhat dierent terminology to state this lemma: straight cycles
in the dual graph are our bands, and crossing is our overlapping. The induced
subgraph is also dened directly, instead of as an induced subgraph of GP .
Universal Hinge Patterns for Folding Strips Efficiently into Any Grid Polyhedron 117

Proposition 7. Let P be a grid polyhedron, and consider a milling tour of P

obtained from the approximation algorithm of Theorem 6. Then the following
properties hold:

1. A grid square of P is either visited once, in which case it is visited by a

straight part of the tour; or it is visited twice, by two straight junctions or
by two turn junctions.
2. In the case of a turn junction, the length of the milling tour between the
two visits to the grid square (counting only one of the two visits to the grid
square in the length measurement) is even.
3. The tour can be modied to alternate between left and right turns (without
changing its length or the number of turns).

3.5 Polynomial Time

The algorithm described above is polynomial in the surface area N of the grid
polyhedron, or equivalently, polynomial in the number of unit cubes making
up the polyomino solid. For our application to strip folding, this is polynomial
in the length of the strip, and thus sucient for most purposes. On the other
hand, polyhedra are often encoded as a collection of n vertices and faces, with
faces possibly much larger than a unit square. In this case, the algorithm runs
in pseudopolynomial time.
Although we do not detail the approach here, our algorithm can be modied
to run in polynomial time. To achieve this result, we can no longer aord to deal
with unit bands directly, because their number is polynomially related to the
number N of grid squares, not the number n of vertices. To achieve polynomial
time in n, we concisely encode the output milling tour using fat bands rather
than unit bands, which can then be easily decoded into a tour of unit bands. By
making each band as wide as possible, their number is polynomially related to
n instead of N . Details of an O(n3 log n)-time milling approximation algorithm
(with the same approximation bounds as above) can be found in [4].

4 Strip Foldings of Grid Polyhedra

In this section, we show how we can use the milling tours from Section 3 to fold a
canonical strip or zig-zag strip eciently into a given (genus-0) grid polyhedron.
For both strip types, dene the length of a strip to be the number of grid squares
it contains; refer to Fig. 1. For a strip folding of a polyhedron P , dene the
number of layers covering a point q on P to be the number of interior points of
the strip that map to q in the folding, excluding crease points, that is, points lying
on a hinge that gets folded by a nonzero angle. (This denition may undercount
the number of layers along one-dimensional creases, but counts correctly at the
remaining two-dimensional subset of P .) We will give bounds on the length of
the strip and also on the maximum number of layers of the strip covering any
point of the polyhedron.
118 N.M. Benbernou et al.

The main idea for canonical strips is that Properties (1)

and (3) of Proposition 7 allow us to make turns as shown in
Fig. 6, so that we do not waste an extra square of the strip
per turn.
Fig. 6. A turn
Theorem 8. Let P be a grid polyhedron, and let N be the junction for a
number of grid squares of P . Then P can be covered by a canonical strip.
folding of a canonical strip of length 2N , and with at most two layers covering
any point of P .

For zig-zag strips, we instead use Properties (1) and (2) of Proposition 7:

Theorem 9. Let P be a grid polyhedron, and let N be the number of grid squares
of P . Then P can be covered by a folding of a zig-zag strip of length 4N , and
with at most four layers covering any point of P .

By coloring the two sides of the zig-zag strip dierently, we can also bicolor
the surface of P in any pattern we wish, as long as each grid square is assigned
a uniform color. We do not prove this result formally here, but mention that
the bounds in length would become somewhat worse, and the rigid motions
presented in Section 5 do not work in this setting.

5 Rigid Motion Avoiding Collision

So far we have focused on just specifying a nal folded state for the strip, and
ignored the issue of how to continuously move the strip into that folded state.
In this section, we show how to achieve this using a rigid folding motion, that is,
a continuous motion that keeps all polygonal faces between hinges rigid/planar,
bending only at the hinges, and avoids collisions throughout the motion. Rigid
folding motions are important for applications to real-world programmable mat-
ter made out of sti material except at the hinges. Our approach may still suer
from practical issues, as it requires a large (temporary) accumulation of many
layers in an accordion form.
We prove that, if the grid polyhedron P has fea-
ture size at least 2, then we can construct a rigid mo-
tion of the strip folding without collision. (By feature (a) (b)
size at least 2, we mean that every exterior voxel of P Fig. 7. Accordion for (a)
is contained in some empty 2 2 2 box.) If the grid canonical strip and (b)
polyhedron we wish to fold has feature size 1, then zig-zag strip, with hinges
one solution is to scale the polyhedron by a factor drawn thick for increased
of 2, and then the results here apply. visibility.
Our approach is to keep the yet-unused portion of the strip folded up into an
accordion and then to unfold only what is needed for the current move: straight,
left, or right. Fig. 7 shows the accordion state for the canonical strip and for
the zig-zag strip. We will perform the strip folding in such a way that the strip
never penetrates the interior of the polyhedron P , and it never weaves under
Universal Hinge Patterns for Folding Strips Efficiently into Any Grid Polyhedron 119

(a)

(b)

Fig. 8. Canonical strip, face-up cases. (a) Straight. (b) Turn where e23 is at.

previous portions of the strip. Thus, we could wrap the strip around P s surface
even if P s interior were already a lled solid. This restriction helps us think
about folding the strip locally, because some of P s surface may have already
been folded (and it thus should not be penetrated) by earlier parts of the strip.
It suces to show, regardless of the local geometry of the polyhedron at the
grid square where the milling tour either goes straight or turns, and regardless of
whether the accordion faces up or down relative to the grid square it is covering,
that we can maneuver the accordion in a way that allows us to unroll as many
squares as necessary to perform the milling-tour move. Fig. 8 shows two key
cases for unrolling part of the accordion of a canonical strip. See [5] for details.
Acknowledgments. We thank ByoungKwon An and Daniela Rus for several
helpful discussions about programmable matter that motivated this work.

References

1. B. An, N. Benbernou, E. D. Demaine, and D. Rus. Planning to fold multiple

objects from a single self-folding sheet. Robotica, 29(1):87102, 2011. Special issue
on Robotic Self-X Systems.
120 N.M. Benbernou et al.

2. E. M. Arkin, M. A. Bender, E. D. Demaine, S. P. Fekete, J. S. B. Mitchell, and

S. Sethia. Optimal covering tours with turn costs. SIAM Journal on Computing,
35(3):531566, 2005.
3. E. M. Arkin, S. P. Fekete, and J. S. B. Mitchell. Approximation algorithms for
lawn mowing and milling. Computational Geometry: Theory and Applications,
17(12):2550, 2000.
4. N. M. Benbernou. Geometric Algorithms for Recongurable Structures. PhD thesis,
Massachusetts Institute of Technology, September 2011.
5. N. M. Benbernou, E. D. Demaine, M. L. Demaine, and A. Lubiw. Universal hinge
patterns for folding strips eciently into any grid polyhedron. arXiv:1611.03187,
2016. https://arXiv.org/abs/1611.03187.
6. N. M. Benbernou, E. D. Demaine, M. L. Demaine, and A. Ovadya. Universal
hinge patterns to fold orthogonal shapes. In Origami5 : Proceedings of the 5th In-
ternational Conference on Origami in Science, Mathematics and Education, pages
405420. A K Peters, Singapore, 2010.
7. K. C. Cheung, E. D. Demaine, J. Bachrach, and S. Grith. Programmable assem-
bly with universally foldable strings (moteins). IEEE Transactions on Robotics,
27(4):718729, 2011.
8. A. E. F. Clementi, P. Crescenzi, and G. Rossi. On the complexity of approximating
colored-graph problems. In Proceedings of the 5th Annual International Conference
on Computing and Combinatorics, volume 1627 of Lecture Notes in Computer
Science, pages 281290, 1999.
9. E. D. Demaine, M. L. Demaine, and J. S. B. Mitchell. Folding at silhouettes and
wrapping polyhedral packages: New results in computational origami. Computa-
tional Geometry: Theory and Applications, 16(1):321, 2000.
10. E. D. Demaine and T. Tachi. Origamizer: A practical algorithm for folding any
polyhedron. Manuscript, 2017.
11. B. Genc. Reconstruction of Orthogonal Polyhedra. PhD thesis, University of Wa-
terloo, 2008.
12. E. Hawkes, B. An, N. M. Benbernou, H. Tanaka, S. Kim, E. D. Demaine, D. Rus,
and R. J. Wood. Programmable matter by folding. Proceedings of the National
Academy of Sciences of the United States of America, 107(28):1244112445, 2010.
13. D. S. Hochbaum. Ecient bounds for the stable set, vertex cover and set packing
problems. Discrete Applied Mathematics, 6(3):243254, 1983.
14. R. J. Lang. A computational algorithm for origami design. In Proceedings of
the 12th Annual ACM Symposium on Computational Geometry, pages 98105,
Philadelphia, PA, May 1996.
15. R. J. Lang and E. D. Demaine. Facet ordering and crease assignment in uniaxial
bases. In Origami4 : Proceedings of the 4th International Conference on Origami
in Science, Mathematics, and Education, pages 189205, Pasadena, California,
September 2006. A K Peters.

Fig. 9. An example of joining together a few letters from our typeface in Fig. 2. Un-
folding (bottom) not to scale with folding (top).
An optimal XP algorithm for Hamiltonian Cycle
on graphs of bounded clique-width

Benjamin Bergougnoux1 , Mamadou Moustapha Kante1 , and O-joung Kwon2 

1
Universite Clermont Auvergne, LIMOS, CNRS, Aubiere, France
2
Logic and Semantics, TU Berlin, Berlin, Germany
[email protected], [email protected], [email protected]

Abstract. For MSO2 -expressible problems like Edge Dominating Set

or Hamiltonian Cycle, it was open for a long time whether there is
an algorithm which given a clique-width k-expression of an n-vertex
graph runs in time f (k) nO(1) for some function f . Recently, Fomin
et al. (SIAM. J. Computing, 2014) presented several lower bounds; for
instance, there are no f (k)no(k) -time algorithms for Edge Dominating
Set and for Hamiltonian Cycle unless the Exponential Time Hypoth-
esis (ETH) fails. They also provided an algorithm running in time nO(k)
for Edge Dominating Set, but left open whether Hamiltonian Cycle
can be solved in time nO(k) .
In this paper, we prove that Hamiltonian Cycle can be solved in time
2
nO(k) . This improves the naive algorithm that runs in time nO(k ) by Es-
pelage et al. (WG 2001). We present a general technique of representative
sets using two-edge colored multigraphs on k vertices. The essential idea
behind is that for a two-edge colored multigraph, the existence of an
Eulerian trail that uses edges with dierent colors alternatively can be
determined by two information: the number of colored edges incident
with each vertex, and the connectedness of the multigraph. With this
idea, we avoid the bottleneck of the naive algorithm, which stores all
the possible multigraphs on k vertices with at most n edges. We can
apply this technique to other problems such as q-Cycle Covering or
Directed Hamiltonian Cycle as well.

1 Introduction

Tree-width is one of the graph width parameters that plays an important role
in graph algorithms. Various problems which are NP-hard on general graphs,
have been shown to be solvable in polynomial time on graphs of bounded tree-
width [1,2]. A celebrated algorithmic meta-theorem by Courcelle [3] states that
every graph property expressible in monadic second-order logic which allows

B. Bergougnoux and M.M. Kante are supported by French Agency for Research
under the GraphEN project (ANR-15-CE-0009). O. Kwon is supported by the Eu-
ropean Research Council (ERC) under the European Unions Horizon 2020 research
and innovation programme (ERC consolidator grant DISTRUCT, agreement No.
648527).

Springer International Publishing AG 2017 121

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 121132, 2017.
DOI: 10.1007/978-3-319-62127-2_11
122 B. Bergougnoux et al.

quantications over edge and vertex sets (MSO2 ) can be decided in linear time
on graphs of bounded tree-width. Minimum Dominating Set, q-Coloring,
and Hamiltonian Cycle problems are such graph problems.
Courcelle and Olariu [5] dened the notion of clique-width of graphs, whose
modeling power is strictly stronger than tree-width. The motivation of clique-
width came from the observation that many algorithmic problems are tractable
on classes of graphs that can be recursively decomposed along vertex partitions
(A, B) where the number of neighbourhood types between A and B is small. We
formally dene clique-width in Section 2. Courcelle, Makowsky, and Rotics [4]
extended the meta-theorem on graphs of bounded tree-width [3] to graphs of
bounded clique-width, at a cost of a smaller set of problems, namely, the class of
problems expressible in MSO1 , which allows quantications on vertex sets only.
Some of the known examples of graph problems that are MSO2 -denable, but not
MSO1 -denable are Max-Cut, Edge Dominating Set, and Hamiltonian
Cycle problems.
A natural question is whether such problems allow an algorithm with running
time f (k) nO(1) for some function f , when a clique-width k-expression of an
n-vertex input graph is given. This question has been carefully answered by
Fomin et al. [8,9]. In particular, they showed that for Max-Cut and Edge
Dominating Set, there is no f (k) no(k) -time algorithm unless the Exponential
Time Hypothesis (ETH) fails, and proposed for both problems algorithms with
running time nO(k) . They proved that Hamiltonian Cycle also cannot be
solved in time f (k) no(k) , unless ETH fails, but left open the question of nding
an algorithm running in time nO(k) . Until now, the best algorithm is the one by
2
Espelage, Gurski, and Wanke [7] which runs in time nO(k ) .
Our Contribution and Approach. In this paper, we prove that Hamiltonian
Cycle can be solved in time nO(k) , thereby resolving the open problem in [9].
A Hamiltonian cycle in a graph is a cycle containing all vertices of the graph.
The Hamiltonian Cycle problem asks whether given a graph G, G contains
a Hamiltonian cycle or not.

Theorem 1. Given an n-vertex graph and its clique-width k-expression, one

can solve Hamiltonian Cycle in time nO(k) .
A k-labeled graph is a graph whose vertices are labeled by integers in {1, . . . , k}.
Clique-width k-expressions are expressions which allow to recursively construct
a graph with the following graph operations: (1) creating a graph with a single
vertex labeled i, (2) taking the disjoint union of two k-labeled graphs, (3)
adding all edges between vertices labeled i and vertices labeled j, for some i = j,
(4) renaming all vertices labeled i into j. The clique-width of a graph is the min-
imum k such that it can be constructed using labels in {1, . . . , k}. One observes
that if a graph contains a Hamiltonian cycle C, then each k-labeled graph H
introduced in the clique-width k-expression admits a partition of its vertex set
into pairwise vertex-disjoint paths, path-partitions, which is the intersection of
H and C. A natural approach is to enumerate all path-partitions. Furthermore,
since the adjacency relations between this k-labeled graph and the remaining
An optimal XP algorithm for Hamiltonian Cycle on graphs of bounded clique-width 123

part only depends on the labels, it is sucient to store for each pair of labels
(i, j), the number of paths whose end vertices are labeled by i and j. As the
2
number of paths of a path-partition is bounded by n, there are at most nO(k )
possibilities. This is the basic idea of the XP algorithm developed by Espelage,
Gurski, and Wanke [7].
The essential idea of our algorithm is to introduce an equivalence relation
between two path-partitions. Given a path-partition P that is arestriction of
a Hamiltonian cycle C, we consider the maximal paths in C P P E(P ) as
another path-partition Q. As depicted in Figure 1, we can construct a multigraph
associated with P and Q on the vertex set {v1 , . . . , vk }, by adding a red edge
vi vj (an undashed edge) if there is a path in P with end vertices labeled by i and
j, and by adding a blue edge vi vj (a dashed edge) if there is a path in Q with end
vertices labeled by i and j. A crucial observation is that this multigraph admits
an Eulerian trail where red edges and blue edges are alternatively used. This is
indeed a characterisation of the fact that two such path-partitions can be joined
into a Hamiltonian cycle. To determine the existence of such an Eulerian trail,
it is sucient to know the degree of each vertex and the connected components
of the corresponding multigraphs of the two path-partitions. This motivates
an equivalence relation between path-partitions. As a byproduct, we can keep
in each equivalence class a representative and since the number of equivalence
classes is bounded by 2k log2 k nk , we can turn the naive algorithm into an nO(k) -
time algorithm (there are at most 2k log2 k partitions of k-elements set). A more
detailed explanation of our algorithm is provided in Section 3.

1
v1
2
v2
3
v3
4
v4

Fig. 1. The restriction of a Hamiltonian cycle to a k-labeled graph. The complement

part can be considered as another set of paths.

The paper is organized as follows. Section 2 contains the necessary prelimi-

naries and required notions. Section 3 is devoted to the overview of the algorithm
and the proof of the existence of Eulerian trails in two-edge colored multigraphs.
We introduce in Section 4 the equivalence relation between multigraphs on the
vertex set {v1 , . . . , vk }, and introduce operations related to the update of path-
partitions in clique-width k-expressions, and prove that they preserve the equiv-
alence relation. We dene the notion of representatives and give the algorithm in
Section 5. We conclude with more applications of our representatives in Section
6. Some proofs are omitted (statements with ) because of space constraints.
124 B. Bergougnoux et al.

2 Preliminaries
The size of a set V is denoted by |V |, and we write [V ]2 to denote the set of
all subsets of V of size 2. We denote by N the set of non-negative integers. We
essentially follow [6] for our graph terminology, but we deal only with nite
graphs. The vertex set of a graph G is denoted by V (G) and its edge set by
E(G) [V (G)]2 . We write xy to denote an edge {x, y}. Let G be a graph.
For X V (G), we denote by G[X] the subgraph of G induced by X, and for
F E(G), we write G F for the subgraph (V (G), E(G) \ F ). The degree of a
vertex x, denoted by degG (x), is the number of edges incident with x. A cut-edge
in a connected graph is an edge e such that G {e} is disconnected. For two
sets A, B V (G), A is complete to B if for every v A and w B, vw E(G).
A graph is non-trivial if it contains an edge, otherwise it is said trivial. A
walk of a graph is an alternating sequence of vertices and edges, starting and
ending at some vertices, where for every consecutive pair of a vertex x and an
edge e, x is incident with e. A trail of a graph is a walk where each edge is used
at most once. A trail is closed if its rst and last vertices are the same.
A multigraph is essentially a graph, but we allow two edges to be incident
with the same set of vertices. Formally, a multigraph G is a pair (V (G), E(G))
of disjoint sets, also called sets of vertices and edges, respectively, together with
a map multG : E(G) V (G) [V (G)]2 , which maps every edge to one or two
vertices, still called its end vertices. Note that we admit loops in multigraphs,
while we do not in our denition of graphs. If there is e E(G) such that
multG (e) = {x, y} (or multG (e) = {x}), we use the term multiedge to refer to
{x, y} (or {x}). The degree of a vertex x in a multigraph G, is dened analo-
gously as in graphs, except that each loop is counted twice, and similarly for
other notions. If there are exactly k edges e such that multG (e) = {x, y} (or
multG (e) = {x}), then we denote these distinct edges by {x, y}1 , . . . , {x, y}k (or
{x}1 , . . . , {x}k ); if k = 1, then for the sake of clarity, we write {x, y} (or {x})
instead of {x, y}1 (or {x}1 ).
An Eulerian trail in a multigraph is a closed trail containing all edges.
Clique-width. A graph is k-labeled if there is a labeling function f : V (G)
{1, . . . , k}, and we call f (v) the label of v. For a k-labeled graph G, we simply
call the set of all vertices with label i as the label class i of G.
The clique-width of a graph G is the minimum number of labels needed to
construct G using the following four operations:
1. Creation of a new vertex v with label i (denoted by i(v)).
2. Disjoint union of two labeled graphs G and H (denoted by G H).
3. Joining by an edge each vertex with label i to each vertex with label j (i = j,
denoted by i,j ).
4. Renaming label i to j (denoted by ij ).
Such an expression is called a clique-width k-expression or simply a k-expression
if it uses at most k distinct labels. We can naturally represent this expression
as a tree-structure. Such trees are known as syntactic trees associated with k-
expressions.
An optimal XP algorithm for Hamiltonian Cycle on graphs of bounded clique-width 125

A clique-width k-expression is called irredundant if whenever i,j is applied,

the constructed graph contains no prior edges between vertices with label i and
vertices with label j. Courcelle and Olariu [5] proved that given a clique-width k-
expression, it can be transformed into an irredundant k-expression in linear time.
Therefore, we can assume that a given clique-width expression is irredundant.

Path-partition. For a graph G, a set P= {P1 , . . . , Pm } of vertex-disjoint paths

in G is called a path-partition of G if 1im V (Pi ) = V (G). A path-partition
P is k-labeled if the end vertices of each path in P are labeled by some integer
in {1, . . . , k}. For a path-partition P of a k-labeled graph G, the labeling of
P induced by the labeling of G consists in assigning to each end vertex of a
path in P its label in G. Lastly, for a k-labeled path-partition P of a graph, we
dene the  auxiliary multigraph Aux(P) with vertex set {v1 , . . . , vk } and edge
set {{vi , vj }1 , . . . , {vi , vj }ij } where ij is the number of paths in P
i,j{1,...,k}
with end vertices labeled i and j respectively.

3 Overview of the algorithm

In an irredundant clique-width k-expression dening a given graph G, G is

recursively constructed using k-labeled graphs. Such k-labeled graphs H arising
in the k-expression dening G are subgraphs3 of G and satisfy the following
properties: (1) for two vertices v, w V (H) with same labels in H, NG (v)
(V (G) \ V (H)) = NG (w) (V (G) \ V (H)), and (2) for some label class i, say
Li , and label class j, say Lj , in H with i = j, if there exist v Li , w Lj with
vw E(G) \ E(H), then every vertex of Li is adjacent to every vertex of Lj
in G, and there are no edges between Li and Lj in H. The former statement is
because when we add vw, all vertices in each set Li or Lj have the same label, and
the latter statement is because of the irredundancy of . Given such a k-labeled
subgraph H of G, for any Hamiltonian cycle C, the restriction of C to H induces
a k-labeled path-partition of H. Because of the two properties (1) and (2), it is
sucient to store the end vertices of each path. This is naturally represented as a
multigraph on k vertices, motivating the denition of the auxiliary multigraphs
associated with k-labeled path-partitions. Our algorithm will be based on the
following characterization of equivalent path-partitions.
Proposition 1 (). Let P1 , P2 be k-labeled path-partitions of H whose labelings
are induced by the labeling of H such that (degAux(P1 ) (v1 ), . . . , degAux(P1 ) (vk )) =
(degAux(P2 ) (v1 ), . . . , degAux(P2 ) (vk )), and {V (C) | C is a component of Aux(P1 )}
= {V (C) | C is a component of Aux(P2 )}. Then G has a Hamiltonian cy-
cle C1containing each path in P1 as a subpath and such that every edge in
C1 ( P P1 E(P )) is contained in E(G) \ E(H) if and only if G has a Hamilto-
nian cycleC2 containing each path in P2 as a subpath and such that every edge
in C2 ( P P2 E(P )) is contained in E(G) \ E(H).
3
Disregarding the labels.
126 B. Bergougnoux et al.

Suppose there are two such path-partitions P1 and P2 where there is a cycle
C1 satisfying the conditions of Proposition 1. Let Q be the subpaths of C1 con-
necting two consecutive paths in P1 . See Figure 1 for an illustration. Note that
if a path in Q is an edge, then it is an edge between two label classes and contained
in E(G) \ E(H). By the property (2), these two label classes are complete in G.
We label each end vertex of a path in Q as the label of H, We consider the
auxiliary multigraph Aux(P1 ) and the auxiliary multigraph  Aux(Q) by consid-
ering Q as a path-partition of the underlying graph on QQ V (Q). We obtain
a multigraph F from the disjoint union of Aux(P1 ) and Aux(Q) by identifying
each vi . Following the Hamiltonian cycle C, one easily checks that there is an
Eulerian trail which alternates between edges in Aux(P1 ) and edges in Aux(Q).
We will prove that if we replace Aux(P1 ) with Aux(P2 ) in F , then the
new graph also admits an Eulerian trail, because of the given conditions in
Proposition 1. To see this, we observe the following, which is a strengthening
of Eulers theorem on Eulerian trails. It is well known that a connected graph
contains an Eulerian trail if and only if every vertex has even degree. Moreover,
when edges are colored by two colors, say red and blue, and each vertex is incident
with the same number of edges for both colors, then we can nd an Eulerian
trail where the two colors appear alternatively. We call such an Eulerian trail a
red-blue alternating Eulerian trail. For a multigraph G colored by red and blue
and v V (G), let rdegG (v) denote the number of red edges incident with v, and
let bdegG (v) denote the number of blue edges incident with v.
Lemma 1 (). Let G be a connected multigraph whose edges are colored by red
and blue. Then G has a red-blue alternating Eulerian trail if and only if for every
vertex v, bdegG (v) = rdegG (v).
Indeed, when we replace Aux(P1 ) with Aux(P2 ) in F , the set of components
does not change (thus consists of one non-trivial component), and each vertex is
incident with same number of red and blue edges, and by Lemma 1, the resulting
graph has an Eulerian trail. We will show that one can construct a Hamiltonian
cycle of G from paths of P2 using the properties (1) and (2).
Motivated by Proposition 1, we dene in Section 4 an equivalence relation
between two sets of multigraphs on the same vertex set {v1 , . . . , vk }. We further
dene operations on those multigraphs, corresponding to procedures of updating
path-partitions, and prove that the equivalence between two sets is preserved
under such operations. These results will form the skeleton of the main algorithm.

4 An equivalence relation between families of k-vertex

multigraphs
For two multigraphs G and H on the same vertex set {v1 , . . . , vk } and with dis-
joint edge sets, we denote by G//H the multigraph with vertex set {v1 , . . . , vk }
and edge set E(G) E(H).
For families F, F1 , F2 of multigraphs on the vertex set {v1 , . . . , vk } and two
distinct integers i, j {1, . . . , k}, we dene the following operations:
An optimal XP algorithm for Hamiltonian Cycle on graphs of bounded clique-width 127

1. F + (i, j) is the set of all multigraphs F  where F  can be obtained from

a multigraph F in F as follows4 : choose two distinct edges {vi , vi }t and
{vj , vj }s , and let F  be the multigraph on vertex set {v1 , . . . , vk } and edge
set (E(F ) \ {{vi , vi }t , {vj , vj }s }) {e} with e
/ E(F ) mapped to {vi , vj }.
2. F + t(i, j) is the set constructed from F by doing the operation +(i, j) t
times.
3. F|ij is the set of all multigraphs F where F can be obtained from a multi-
graph in F by replacing every edge with an end vertex vi by an edge with
an end vertex vj .
4. F1
F2 := {F1 //F2 : F1 F1 , F2 F2 }.
Let F1 and F2 be two families of multigraphs on the vertex set {v1 , . . . , vk }.
We write F1  F2 if for every multigraph H on the vertex set {v1 , . . . , vk },
- whenever there exists G2 F2 such that (degG2 (v1 ), . . . , degG2 (vk )) =
(degH (v1 ), . . . , degH (vk )) and G2 //H has at most one non-trivial component, there
exists G1 F1 such that (degG1 (v1 ), . . . , degG1 (vk )) = (degH (v1 ), . . . , degH (vk ))
and G1 //H has at most one non-trivial component.

We say that F1 is equivalent to F2 , written F1 F2 , if F1  F2 and F2  F1 .

We prove that the equivalence between two families is preserved by the op-
eration +(i, j).
Proposition 2. Let F1 and F2 be two families of multigraphs on the vertex set
{v1 , . . . , vk }. If F1 F2 , then F1 + (i, j) F2 + (i, j).

Proof. Suppose F1 F2 . It is sucient to prove that F1 +(i, j)  F2 +(i, j). For

this, suppose there exist a graph H on {v1 , . . . , vk } and G2 F2 +(i, j) such that
(degG2 (v1 ), . . . , degG2 (vk )) = (degH (v1 ), . . . , degH (vk )) and G2 //H has at most
one non-trivial component. Since G2 F2 + (i, j), there exist F2 F2 , edges
{vi , vi }t , {vj , vj }s in F2 such that G2 = (V (F2 ), E(F2 ) \ {{vi , vi }t , {vj , vj }s }
{e}) with e / E(F2 ) mapped to {vi , vj } in G2 . Let H  := (V (H), E(H) {e })

with e / E(H) mapped to {vi , vj } in H. We claim that

(degF2 (v1 ), . . . , degF2 (vk )) = (degH  (v1 ), . . . , degH  (vk )) and

F2 //H  has at most one non-trivial component.

By the construction of G2 from F2 , for every v V (F2 ) \ {vi , vj }, v has

the same degree in F2 and G2 , and the degrees of vi and vj in G2 are one
less than the degrees in F2 , respectively. Since the degrees of vi and vj in
H  are one more than the degrees in H, we have (degF2 (v1 ), . . . , degF2 (vk )) =
(degH  (v1 , . . . , degH  (vk )). Assume now that F2 //H  has at least two non-trivial
components. First observe that the four vertices vi , vj , vi , vj are in the same non-
trivial component C of F2 //H  , and {vi , vj } are in a same non-trivial component
of G2 //H. If C  is another non-trivial component of F2 //H  , then it does not
intersect {vi , vi , vj , vj }, that is, C  is non-trivial component in G2 //H that does
not intersect the one containing {vi , vj }, yielding a contradiction.
4
We allow vi (or vj ) to be equal to vi (or vj ).
128 B. Bergougnoux et al.

Since F1 F2 , there exists F1 F1 such that (degF1 (v1 ), . . . , degF1 (vk )) =

(degH  (v1 ), . . . , degH  (vk )) and F1 //H  has at most one non-trivial component.
By Lemma 1, F1 //H  contains an Eulerian trail where edges in F1 and edges
in H  are alternatively used. Let {vi , vi }t and {vj , vj }s be the edges where
{vi , vi }t , e , {vj , vj }s appear in the Eulerian trail in this order (recall that e
is mapped to {vi , vj }). Clearly, if we remove the edges {vi , vi }t , e , {vj , vj }s
and add an edge f mapped to {vi , vj } in F1 //H  , then the obtained multi-
graph K still admits an alternating Eulerian trail. Let G1 = (V (F1 ), E(F1 ) \
{{vi , vi }t , {vj , vj }s }{f }) with f / E(F1 ) mapped to {vi , vj } in G1 . One eas-
ily checks that K = G1 //H, and since K has an Eulerian trail where edges in G1
and edges in H are alternatively used, by Lemma 1, (degG1 (v1 ), . . . , degG1 (vk )) =
(degH (v1 ), . . . , degH (vk )) and G1 //H has at most one non-trivial component.
Because G1 F1 + (i, j), we can thus conclude that F1 + (i, j)  F2 + (i, j). 
We prove a similar property for the other operations.
Proposition 3 (). Let F1 and F2 be families of multigraphs on the vertex set
{v1 , . . . , vk } and let i, j {1, . . . , k} be two distinct integers. If F1 F2 , then
F1 |ij F2 |ij .
Proposition 4 (). Let F1 , F2 , F3 be families of multigraphs on the vertex set
{v1 , . . . , vk }. If F1 F2 , then F1
F3 F2
F3 .

5 Hamiltonian Cycle problem

We prove the main result of this paper. We recall the statement.
Theorem 2. Given a graph G and its clique-width k-expression, one can solve
Hamiltonian Cycle in time nO(k) .
We now dene formally our notion of representatives based only on the degree
sequence and connected components of auxiliary multigraphs associated with k-
labeled path-partitions.
Denition 1 (Representatives by auxiliary multigraphs). Let G and H
be multigraphs on vertex set {v1 , . . . , vk }. We write G  H whenever
(degG (v1 ), . . . , degG (vk )) = (degH (v1 ), . . . , degH (vk )) and {V (C) | C is a com-
ponent of G} is equal to {V (C) | C is a component of H}. One easily checks
that  is an equivalence relation on any set F of multigraphs on vertex set
{v1 , . . . , vk }.
For a family F of multigraphs on vertex-set {v1 , . . . , vk }, let reduce(F) be
the operation which takes in each equivalence class of F/  a representative.
The following is a rephrasing of Proposition 1.
Proposition 5 (). Let G be a graph with its irredundant clique-width k-expression
, and let t be a node in the syntactic tree. Let Gt be the k-labeled graph con-
structed at t, and let P1 and P2 be k-labeled path-partitions of Gt whose labelings
are induced by the labeling of Gt . If Aux(P1 )  Aux(P2 ), then the following are
equivalent.
An optimal XP algorithm for Hamiltonian Cycle on graphs of bounded clique-width 129

1. G has a Hamiltonian cycle C1containing each path in P1 as a subpath and

such that every edge in C1 ( P P1 E(P )) is contained in E(G) \ E(Gt ).
2. G has a Hamiltonian cycle C2containing each path in P2 as a subpath and
such that every edge in C2 ( P P2 E(P )) is contained in E(G) \ E(Gt ).

Proof (Sketch of Proof ). Suppose G has a Hamiltonian cycle C1 satisfying 1.

LetQ be the set of all maximal paths in C1 ( P P1 E(P )), and let H :=
G[ QQ V (Q)]. We consider Q as the path-partition of H where the end vertices
of paths in Q are labeled by their labels in Gt . Since C1 is a Hamiltonian cycle,
Aux(P1 )//Aux(Q) has an Eulerian trail where edges in Aux(P1 ) and edges
in Aux(Q) are alternatively used. Since Aux(P2 )  Aux(P1 ), by Lemma 1,
Aux(P2 )//Aux(Q) admits an Eulerian trail where the edges in Aux(P2 ) and
the edges in Aux(Q) are alternatively used.
To construct a Hamiltonian cycle C2 satisfying 2, let e1 , e2 , . . . , e2m be the
sequence of the edges in an Eulerian trail of Aux(P2 )//Aux(Q) where edges in
Aux(P2 ) and edges in Aux(Q) are alternatively used such that e1 E(Aux(P2 )).
If e2i corresponds to a path of length 1 in Q between two label classes, then we
add a direct edge between end vertices of paths in P corresponding to e2i1 and
e2i+1 . If e2i corresponds to a path of length 2 in Q, then we add this path to
the subgraph and connect to the paths in P corresponding to e2i1 and e2i+1 . In
particular, the former procedure is possible because is irredundant, and thus
two label classes are complete in G. In this way, we can construct a Hamiltonian
cycle C2 satisfying 2.


Proposition 6 tells us that if F is the set of possible k-labeled path-partitions

at a node t of the syntactic tree, it is enough to store reduce({Aux(P) | P F }).

Proposition 6 (). Let F be a family of graphs on the vertex set {v1 , . . . , vk }.

Then F reduce(F).

Proof (Proof of Theorem 1). We assume that G has at least 3 vertices, otherwise
we can automatically say it is a No-instance. Since every k-expression can be
transformed into an irredundant k-expression in linear time, we may assume that
G is given with an irredundant k-expression. Let be the given irredundant k-
expression dening G, and T be the syntactic tree of . For every node t of T ,
let Gt be the subgraph of G dened at node t, and for each i {1, . . . , k}, let
Gt [i] be the subgraph of Gt induced by the vertices with label i.
For each node t and each vector (a1 , . . . , ak ) {0, 1, . . . , n}k, let c[t, (a1 , . . . , ak )]
be the set of all multigraphs F on the vertex set {v1 , . . . , vk } where
F = Aux(P) for some k-labeled path-partition P of Gt ,
for each i {1, . . . , k}, ai is the degree of vi in F .
Instead of computing the whole set c[t, (a1 , . . . , ak )], we will compute a subset
r[t, (a1 , . . . , ak )] of c[t, (a1 , . . . , ak )] of size 2O(k log k) such that r[t, (a1 , . . . , ak )]
c[t, (a1 , . . . , ak )].
We explain how to decide whether G has a Hamiltonian cycle. Let troot be
the root node of T , and let tlastjoin be the node taking the disjoint union of two
130 B. Bergougnoux et al.

graphs and closest to the root node. We can observe that G has a Hamiltonian
cycle if and only if there are some node t between troot and tlastjoin with child
t and a path-partition P of Gt such that t is a join node labeled by i,j ,
and degAux(P) (vi ) = degAux(P) (vj ) > 0 and degAux(P) (vi ) = 0 for all i
{1, . . . , k} \ {i, j}. This is equivalent to that c[t , (a1 , . . . , ak )] = for some vector
(a1 , . . . , ak ) where ai = aj > 0 and ai = 0 for all i {1, . . . , k} \ {i, j}.
Therefore, if there is a Hamiltonian cycle, then r[t , (a1 , . . . , ak )] = for such a
tuple of t, t , and (a1 , . . . , ak ), and we can correctly say that G has a Hamiltonian
cycle, and otherwise, there are no such tuples, and we can correctly say that G
has no Hamiltonian cycles.
Now, we explain how to recursively generate r[t, (a1 , . . . , ak )].

1. (Creation of a vertex v with label i)

If ai = 2 and aj = 0 for all j = i, then c[t, (a1 , . . . , ak )] consists of one graph
on the vertex set {v1 , . . . , vk } with a loop incident with vi , and otherwise, it is
an empty set. So, we add the graph ({v1 , . . . , vk }, {vi vi }) to r[t, (a1 , . . . , ak )]
when ai = 2 and aj = 0 for all j = i, and set r[t, (a1 , . . . , ak )] := otherwise.
2. (Disjoint union node with two children t1 and t2 )
Since every path-partition of Gt is obtained by taking the disjoint union of
a path-partition of Gt1 and a path-partition of Gt2 , we have

c[t,(a1 , . . . , ak )]

:= c[t1 , (a11 , . . . , a1k )]
c[t2 , (a21 , . . . , a2k )].
(a11 ,...,a1k )+(a21 ,...,a2k )=(a1 ,...,ak )

We assign

r[t, (a1 , . . . , ak )]


:= reduce r[t1 , (a11 , . . . , a1k )]
r[t2 , (a21 , . . . , a2k )] .
(a11 ,...,a1k )+(a21 ,...,a2k )=(a1 ,...,ak )

3. (Join node with the child t such that each vertex with label i is joined to
each vertex with label j)
Note that every path-partition of Gt is obtained from a path-partition of
Gt by adding some edges between end vertices of label i and end vertices of
label j. We can observe that when we add an edge between an end vertex
v of a path P1 with label i, and an end vertex w of a path P2 with label j,
these two paths P1 and P2 will be unied into a path whose end vertices are
end vertices of P1 and P2 other than v and w. Thus, it corresponds to the
operation +(i, j) on auxiliary multigraphs. We observe that

c[t,(a1 , . . . , ak )] := (c[t , (a1 , . . . , ak )] + (i, j)).
ai ai =aj aj =0
ax =ax for x = i, j
An optimal XP algorithm for Hamiltonian Cycle on graphs of bounded clique-width 131

We take all possible vectors (a1 , . . . , ak ) where ai ai = aj aj 0, and
for all t {1, . . . , k} \ {i, j}, at = at . Assume  = ai ai . For each 
{0, 1, . . . , n}, we assign

r := reduce( reduce(reduce(r[t , (a1 , . . . , ak )] + (i, j)) + (i, j)) + (i, j)),

where we repeat  times, and assign

r[t, (a1 , . . . , ak )] := reduce(r0 r1 rn ).

4. (Renaming node with a child t such that the label of each vertex with label
i is changed to j)
Every path-partition of Gt is also a path-partition of Gt , and vice versa.
Since just labelings of vertices are changed, we can observe that if ai = 0,
then c[t, (a1 , . . . , ak )] is the empty set, and otherwise, we have

c[t,(a1 , . . . , ak )] := c[t , (a1 , . . . , ak )]|ij .
ax =ax for all x = i, j
ai +aj =aj

If ai = 0, then we assign the empty set to r[t, (a1 , . . . , ak )], and otherwise,
we assign



r[t,(a1 , . . . , ak )] := reduce r[t , (a1 , . . . , ak )]|ij .

ax =ax for all x = i, j
ai +aj =aj

One can prove by induction that r[t, (a1 , . . . , ak )] c[t, (a1 , . . . , ak )] for each t
and (a1 , . . . , ak ). Therefore, we can correctly decide whether G has a Hamiltonian
cycle or not using sets r[t, (a1 , . . . , ak )].
Running time. Each constructed set r[t, (a1 , . . . , ak )] consists of at most 2O(k log k)
graphs, as we keep at most one graph for each partition of {v1 , . . . , vk } after the
reduce operation. For the node taking the disjoint union of two graphs, for a xed
vector (a1 , . . . , ak ), there are nO(k) ways to take two vectors A1 and A2 such that
A1 + A2 = (a1 , . . . , ak ). So, we can update r[, ] in time nO(k) 2O(k log k) . For the
node joining edges between two classes, the value  can be taken from 0 to n.
Since each operation +(i, j) take k 2 2O(k log k) time, we can update r[, ] in time
n2 2O(k log k) . Clearly, we can update r[, ] in time n 2O(k log k) for the relabeling
nodes. Therefore, we can solve Hamiltonian Cycle for G in time nO(k) .


6 More applications
Let q be a positive integer. The q-Cycle Covering problem asks for a given
graph G whether there is a set of at most q pairwise vertex-disjoint cycles in
132 B. Bergougnoux et al.

G whose union contains all vertices of G. Denitely, 1-Cycle Covering is the

Hamiltonian Cycle problem. In the q-Cycle Covering problem, we relax
the denition of path-partitions so that it may contain at most q cycles, and
we keep the number of cycles in the path-partition. Also, we dene its auxiliary
multigraph Aux(P) using those remaining paths. One can easily check that two
such modied path-partitions P1 and P2 are equivalent for q-Cycle Covering
if they contain the same number of cycles and Aux(P1 )  Aux(P2 ).
The second application is for Directed Hamiltonian Cycle. Clique-width
was also considered for directed graphs by Courcelle and Olariu [5]. The clique-
width operations for directed graphs are the same as for the undirected graphs,
except the one that add edges between two label classes, dened as follows: (3 )
Adding an arc (u, v) for each vertex u with label i to each vertex v with label j
(i = j, denoted by i,j ). The clique-width of a directed graph G is the minimum
number of labels needed to construct G using these operations. In this case, we
use directed auxiliary multigraphs. Similar to Lemma 1, we can show that
Lemma 2. Let G be a connected directed multigraph whose arcs are colored by
red and blue. Then the following are equivalent.
1. For every vertex v, the number of blue edges leaving v is the same as the
number of red edges entering v, and the number of red edges leaving v is the
same as the number of blue edges entering v.
2. G has a red-blue alternating Eulerian directed trail.
Using Lemma 2, we can proceed same as Theorem 1.

References
1. Arnborg, S., Proskurowski, A.: Linear time algorithms for NP-hard problems re-
stricted to partial k-trees. Discrete Appl. Math. 23, 1124 (1989)
2. Bodlaender, H.L.: A partial k-arboretum of graphs with bounded treewidth. Theor.
Comput. Sci. 209, 145 (1998)
3. Courcelle, B.: The monadic second-order logic of graphs I: Recognizable sets of nite
graphs. Information and Computation 85, 1275 (1990)
4. Courcelle, B., Makowsky, J.A., Rotics, U.: Linear time solvable optimization prob-
lems on graphs of bounded clique width. Theor. Comput. Sci. 33, 125150 (2000)
5. Courcelle, B., Olariu, S.: Upper bounds to the clique width of graphs. Discrete Appl.
Math. 101(1-3), 77114 (2000)
6. Diestel, R.: Graph Theory. No. 173 in Graduate Texts in Mathematics, Springer,
third edn. (2005)
7. Espelage, W., Gurski, F., Wanke, E.: How to solve NP-hard graph problems on
clique-width bounded graphs in polynomial time. In: Graph-theoretic concepts in
computer science (Boltenhagen, 2001), Lecture Notes in Comput. Sci., vol. 2204,
pp. 117128. Springer, Berlin (2001)
8. Fomin, F.V., Golovach, P.A., Lokshtanov, D., Saurabh, S.: Intractability of clique-
width parameterizations. SIAM J. Comput. 39(5), 19411956 (2010)
9. Fomin, F.V., Golovach, P.A., Lokshtanov, D., Saurabh, S.: Almost optimal lower
bounds for problems parameterized by clique-width. SIAM J. Comput. 43(5), 1541
1563 (2014)
 Improved Algorithms for Computing
k-Sink on Dynamic Flow Path Networks

Binay Bhattacharya1 , Mordecai J. Golin2 , Yuya Higashikawa3,  , ,

Tsunehiko Kameda1 , and Naoki Katoh4,  
1
School of Computing Science, Simon Fraser University, Burnaby, Canada
2
Dept. of Computer Science, Hong Kong Univ. of Science and Technology, China
3
Dept. of Information and System Engineering, Chuo University, Tokyo, Japan
4
School of Science and Technology, Kwansei Gakuin University, Hyogo, Japan

Abstract. We address the problem of locating k sinks on dynamic

ow path networks with n vertices in such a way that the evacua-
tion completion time to them is minimized. Our two algorithms run
in O(n log n + k2 log4 n) and O(n log3 n) time, respectively. When all
edges have the same capacity, we also present two algorithms which run
in O(n + k2 log2 n) time and O(n log n) time, respectively. These algo-
rithms together improve upon the previously most ecient algorithms,
which have time complexities O(kn log2 n) [1] and O(kn) [11], in the
general and uniform edge capacity cases, respectively. The above re-
sults are achieved by organizing relevant data for subpaths in a strategic
way during preprocessing, and the nal results are obtained by extract-
ing/merging them in an ecient manner.

1 Introduction
Ford and Fulkerson [5] introduced the concept of dynamic ow which models
movement of commodities in a network. In this model, each vertex is assigned
some initial amount of supply, each edge has a capacity, which limits the rate
of commodity ow into it per unit time, and the transit time to traverse it.
One variant of the dynamic ow problem is the quickest transshipment problem,
where the source vertices have specied supplies and sink vertices have specied
demands. The problem is to send exactly the right amount of commodity out of
sources into sinks in minimum overall time. Hoppe and Tardos [12] provided a
polynomial time algorithm for this problem in the case where the transit times
are integral. However, the complexity of their algorithm is very high. Finding a
practical polynomial time solution to this problem is still open. The reader is
referred to a recent paper by Skutella [18] on dynamic ows.
This paper discusses a related problem, called the evacuation problem [8,
14], in which the supplies (i.e., evacuees) are discrete, and the sinks and their
demands are not specied. In fact, the locations of the sinks are the output of the
problem. Many disasters, such as earthquakes, nuclear plant accidents, volcanic

Partially supported by a Discovery Grant from NSERC of Canada

Partially supported by Hong Kong RGC GRF grant 16208415

Supported by JSPS KAKENHI Grant-in-Aid for Young Scientists (B) (17K12641)

Supported by JST CREST (JPMJCR1402)

Springer International Publishing AG 2017 133

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 133144, 2017.
DOI: 10.1007/978-3-319-62127-2_12
134 B. Bhattacharya et al.

eruptions, ooding, have struck in recent years in many parts of the world, and
it is recognized that orderly evacuation planning is urgently needed.

A k-sink is a set of k sinks such that the evacuation completion time to

sinks is minimized, and our objective is to nd a k-sink on a dynamic ow
path network. Congestion is said to occur when an evacuee cannot move at the
maximum speed constrained only by transit time. Thus, when the capacities of
the edges are suciently large, no congestion occurs and each evacuee can follow
the shortest path to its nearest sink at the maximum speed. This is equivalent
to the classical k-center problem in networks, which is known to be NP-hard
even on bipartite planar graphs of maximum degree 4 [17]. To the best of our
knowledge the most general polynomially solvable case for general k is where the
underlying graphs are cacti or partial t-trees with constant t. Congestion could
occur if vertex capacities are limited, in which case edges may get clogged and
congestion backs up. Our results are valid regardless of whether vertex capacities
(the number of evacuees that they can accommodate) are limited or not.

Mamada et al. [15] solved the 1-sink problem for the dynamic ow tree net-
works in O(n log2 n) time under the condition that only a vertex can be a sink,
where n is the number of vertices. When edge capacities are uniform, we have
presented O(n log n) time algorithms with a more relaxed condition that the
sink can be on an edge, as well as on a vertex [3, 10]. Dealing with congestion is
non-trivial even in path networks. On dynamic ow path networks with uniform
edge capacities, it is straightforward to compute the 1-sink in linear time, as
shown by Cheng et al. [4]. Arumugam et al. [1] showed that the k-sink problem
for dynamic ow path networks can be solved in O(kn log2 n) time, and when
the edge capacities are uniform Higashikawa et al. [11] showed that it can be
solved in O(kn) time.

In this paper we present two algorithms for the k-sink problem on dynamic
ow path networks with general edge capacities. A path network can model an
airplane aisle, a hall way in a building, a street, a highway, etc., to name a few.
Unlike the previous algorithm for the k-sink problem [1] which uses dynamic
programming, our algorithms adopt Megiddos parametric search [16] and the
sorted matrices introduced by Frederickson and Johnson [6, 7]. Together, they
outperform all other known algorithms, and they are the rst sub-quadratic
algorithms for any value of k. These improvements were made possible by our
method of merging evacuation times of subpaths stored in a hierarchical data
structure. We also present two algorithms for the dynamic ow path networks
with uniform edge capacities.

This paper is organized as follows. In the next section, we dene our model
and the terms that are used throughout the paper. Sec. 3 introduces a new
data structure, named the capacities and upper envelopes tree, which plays a
central role in the rest of the paper. In Sec. 4 we identify two important tasks
that form building blocks of our algorithms, and also discuss a feasibility test.
Sec. 5 presents several algorithms for uniform and general edge capacities. Fi-
nally, Sec. 6 concludes the paper.
Improved Algorithms for Computing k -Sink on Dynamic Flow Path Networks 135

2 Preliminaries
2.1 Denitions
Let P = (V, E) be a path network, whose vertices v1 , v2 , . . . , vn are arranged from
left to right in this order. For i = 1, 2, . . . , n, vertex vi has an integral weight
wi (> 0), representing the number of evacuees, and each edge ei = (vi , vi+1 ) has
a xed non-negative length li and an integral capacity ci , which is the upper limit
on the number of evacuees who can enter an edge per unit time. We assume that
a sink has innite capacity, so that the evacuees coming from the left and right
of a sink do not interfere with each other. An evacuation starts at the same time
from all the vertices, and all the evacuees from a vertex evacuate to the same
sink. This is called conuent ow in the parlance of the network ow theory.
This constraint is desirable in evacuation in order to avoid confusion among the
evacuees at a vertex as to which way they should move.
By x P , we mean that point x lies on either an edge or a vertex of P . For
two points a, b P , a b or b  a means that a lies to the left of b. Let d(a, b)
denote the distance (sum of the edge lengths) between a and b. If a and/or b lies
on an edge, we use the prorated distance. The transit time for a unit distance
is denoted by , so that it takes d(a, b) time to travel from a to b, and is
independent of the edge. Let c(a, b) denote the minimum capacity of the edges
on the subpath of P between a and b. The point that is arbitrarily close to vi
on its left (resp. right) side is denoted by vi (resp. vi+ ). Let P [a, b] denote the
subpath of P between a and b satisfying a b. If a, b or both are excluded, we
denote them by P (a, b], P [a, b) or P (a, b), respectively. Let V [a, b] (resp. V (a, b],
V [a, b) or V (a, b)) denotes the set of vertices on P [a, b] (resp. P (a, b], P [a, b) or
P (a, b)). We introduce a weight array W [], dened by

W [i]  wj , for i = 1, 2, . . . , n, (1)
vj V [v1 ,vi ]

and let W [vi , vj ]  W [j] W [i 1] for i j.

2.2 Completion time functions

In our model, a set of k sinks accepts evacuees from k disjoint subpaths of P .
We thus need to be able to compute the completion time for each such subpath
P [vi , vj ]. For simplicity, from now on, we assume that the optimal k sinks are
on edges, not on vertices. Small modications will be necessary if we allow some
sinks to be on vertices. We dene the completion time from left (L-time for short)
to x  vj of vertex vp on P [vi , vj ] to be the evacuation completion time to x
for the evacuees on the vertices on P [vi , vp ], assuming that they all arrive at
x continuously at a uniform rate c(vp , x). We similarly dene the completion
time from right (R-time for short) to x vi of vertex vp on P [vi , vj ] to be the
evacuation completion time to x for all the evacuees on the vertices on P [vp , vj ],
arriving at x continuously at a uniform rate c(x, vp ). For any vertex vp V [vi , vj ],
its L-time and R-time are given mathematically as
[i,j]
L (x, vp )  d(vp , x) + W [vi , vp ]/c(vp , x) for x  vj , (2)
[i,j]
R (x, vp )  d(x, vp ) + W [vp , vj ]/c(x, vp ) for x vi , (3)
136 B. Bhattacharya et al.

8 14 5 10 7 15 8 12 7 3 10
weights

lengths 3 2 4 3 2 3 2 4 2 1 4
v1 v2 v3 v4 v5 v6 v7 v8 v9 x v10 v11

capacities
5 3 4 2 6 5 3 4 5 1

Fig. 1. An example path P = (V, E) with x e9 = (v9 , v10 ).

respectively. For convenience we sometimes refer to the rst (resp. second) term
in the righthand side of (2) and (3) as the distance time (resp. weight time).
Note that the distance time is linear in the distance to x. Consider an arbitrary
subpath P [vi , vj ], where i j.
Fig. 1 shows an example, where vertices v1 , v2 , v3 , . . . (represented by black
circles) have weights 8, 14, 5, . . ., and edges e1 , e2 , e3 , . . . have lengths 3, 2, 4, . . .
and capacities 5, 3, 4, . . . Point x is located on e9 = (v9 , v10 ) so that d(v9 , x) = 2
(represented by a white circle). Assuming = 1, let us compute the L-time to
x of vertex v5 on P [v2 , v7 ]. From d(v5 , x) = 13, W [v2 , v5 ] = 36 and c(v5 , x) = 3,
[2,7]
we obtain L (x, v5 ) = 25.
To be more precise, the weight time should be W [vi , vp ]/c(vp , x) and
W [vp , vj ]/c(x, vp ) in (2) and (3), respectively, since the evacuees are discrete
entities. Although only small modications are necessary to get exact solutions
as shown in [4], we use (2) and (3) for simplicity.

Lemma 1. [9] Let s be the sink for a subpath P [vi , vj ] of a path network P . The
evacuation completion time to s (vi  vh s vh+1  vj ) for the evacuees on
P [vi , vj ] is given by
 
[i,h] [h+1,j]
[i,j] (s)  max max {L (s, v)},  max {R (s, v  )} . (4)
vV [vi ,s) v V (s,vj ]

[i,j]
Referring to (2) and (3), the vertex vp V [vi , vj ] that maximizes L (vj+ , vp )
[i,j]
(resp. R (vi , vp )) is called the L-critical vertex (resp. R-critical vertex) of
[i,j] [i,j]
P [vi , vj ], and is denoted by cL (resp. cR ). Note that (vj+ , vp ) (resp. (vi , vp ))
is used instead of (vj , vp ) (resp. (vi , vp )), and that we have d(vp , vj+ ) = d(vp , vj )
and c(vp , vj+ ) = min{c(vp , vj ), cj } (resp. d(vi , vp ) = d(vi , vp ) and c(vi , vp ) =
min{c(vi , vp ), ci1 }).
Using the example in Fig. 1 again, let us nd the L-critical vertex of P [v2 , v7 ].
[2,7] [2,7]
We rst compute L (v7+ , vp ) for p = 2, . . . , 7: L (v7+ , v2 ) = 14 + 14/2 = 21,
[2,7] + [2,7] + [2,7]
L (v7 , v3 ) = 12 + 19/2 = 21.5, L (v7 , v4 ) = 8 + 29/2 = 22.5, L (v7+ , v5 ) =
[2,7] [2,7]
5+36/3 = 17, L (v7+ , v6 ) = 3+51/3 = 20, and L (v7+ , v7 ) = 0+59/3 19.7.
[2,7]
Comparing these values, we obtain cL = v4 .
Improved Algorithms for Computing k -Sink on Dynamic Flow Path Networks 137

[i,j] [i,j] [i,j]

Proposition 1. Critical vertex vp = cL (resp. vp = cR ) maximizes L (x, vp )
[i,j]
(resp. R (x, vp )) for any point x (vj , vj+1 ] (resp. x [vi1 , vi )).

3 Data structures
A problem instance is said to be t-feasible if there are k sinks such that every
evacuee can reach a sink within time t. In our algorithms, we want to perform
t-feasibility tests for many dierent values of completion time t. Therefore, it is
worthwhile to spend some time during preprocessing to construct data structures
which facilitate these tests.

3.1 Capacities and upper envelopes (CUE) tree

[i,j] [i,j]
We want to design a data structure with which critical vertices cL and cR can
be found eciently for an arbitrary pair (i, j) with 1 i j n. To this end we
introduce the capacities and upper envelopes tree (CUE tree, for short), denoted
by T , with root , whose leaves are the vertices of P arranged from left to right.
It is a balanced tree with height O(log n). In balancing, the vertex weights are
not considered. For a node5 u of T , let T (u) denote the subtree rooted at u,
and let l(u) (resp. r(u)) denote the index of the leftmost (resp. rightmost) vertex
on P that belongs to T (u). See Fig. 2. Let ul , ur and up denote the left child
of u, the right child of u, and the parent of u, respectively. We say that node

(vi , ) (vj , )

vi vl(u) vr(u) vj

Fig. 2. Illustration of a part of CUE tree T . The small gray disks represent nodes of
N [vi , vj ] and dashed circles enclose subpaths in P[vi , vj ].

u spans subpath P [vl(u) , vr(u) ]. At node u, we store l(u), r(u) and the capacity
c(vl(u) , vr(u) ) among others. This information at every node can be computed
bottom up in O(n) time by performing heap-like operations.
For two nodes u, u of T , let (u, u ) denote the path from u to u along
edges of T . Suppose that for an index pair (i, j) with 1 i j n, node  is
5
We use the term node here to distinguish it from the vertices on the path. A
vertex, being a leaf of T , is considered a node, but an interior node of T is not a
vertex.
138 B. Bhattacharya et al.

the lowest common ancestor of vi and vj in T . Consider every node of T that

is the right child of a node on (vi ,  ) or the left child of a node on (vj ,  ),
but which itself is not on (vi ,  ) or (vj ,  ). Let N [vi , vj ] denote the set of
such nodes plus vi and vj . Then clearly N [vi , vj ] consists of O(log n) nodes. Let
P[vi , vj ] denote the set of O(log n) subpaths spanned by nodes of N [vi , vj ].

[i,j]
In order to determine cL for a given pair (i, j), we need to compute

max {d(vp , vj ) + W [vi , vp ]/c(vp , vj+1 )} . (5)

vp V [vi ,vj ]

To facilitate such a computation for an arbitrary pair (i, j), at each node u,
we precompute and store two upper envelope functions associated with sub-
path P [vl(u) , vr(u) ]. Then for u N [vi , vj ] that spans vp , we have W [vi , vp ] =
W [vi , vl(u)1 ]+W [vl(u) , vp ] and c(vp , vj+1 ) = min{c(vp , vr(u)+1 ), c(vr(u)+1 , vj+1 )}.
Since (i, j), hence W [vi , vl(u)1 ] and c(vr(u)+1 , vj+1 ), is not known during pre-
processing, we replace these values with variables W and c, respectively, and
express the two upper envelopes stored at u as functions of W = W [vi , vl(u)1 ]
and c = c(vr(u)+1 , vj+1 ), respectively. We can now break (5) down into a number
of formulea, one for each u N [vi , vj ], which is given by
!
max d(vp , vr(u) ) + (W + W [vl(u) , vp ])/ min{c(vp , vr(u)+1 ), c} . (6)
vp V [vl(u) ,vr(u) ]

Using the concrete values of W and c, we can evaluate (5) by nding the maxi-
mum of the |N [vi , vj ]| = O(log n) values, computed by (6).

Now we want to compute (6) eciently for arbitrary W and c, but of

course we have W = W [vi , vl(u)1 ] and c = c(vr(u)+1 , vj ) in mind for some i
and j. Consider two extreme cases, where for any p with vp V [vl(u) , vr(u) ] (i)
c > c(vp , vr(u)+1 ), and (ii) c c(vp , vr(u)+1 ), respectively. In Case (i), we have
from (6)
!
u
L,1 (W )  max d(vp , vr(u) ) +(W +W [vl(u) , vp ])/c(vp , vr(u)+1 ) (7)
vp V [vl(u) ,vr(u) ]
 
[l(u),r(u)] +
= max L (vr(u) , vp ) + W/c(vp , vr(u)+1 ) . (8)
vp V [vl(u) ,vr(u) ]

Note that c(vp , vr(u)+1 ) gets smaller as vp moves to the left. From (7) it is seen
u
that L,1 (W ) is the upper envelope of linear functions of W and each coecient
u
of W is positive, which means that L,1 (W ) is piecewise linear, continuous, and
increasing in W . Thus it can be encoded as a sequence of bending points. In
Case (ii), we have from (6)
u
!
L,2 (c) = max d(vp , vr(u) ) + W [vl(u) , vp ]/c . (9)
vp V [vl(u) ,vr(u) ]

Note that (9) was obtained from (6) by removing the term W/c, which does
u
not depend on vp . If we plot L,2 (c) vs. (1/c) as a graph, it is also piecewise
linear, continuous, and increasing in (1/c), and can be encoded as a sequence of
bending points.
Improved Algorithms for Computing k -Sink on Dynamic Flow Path Networks 139

u u
At node u we store both L,1 (W ) and L,2 (c) in encoded form with bending
points, which can be computed in O(r(u) l(u)) time. Similarly, in order to
[i,j]
determine cR for an arbitrary pair (i, j), we store two functions which are
u u u u
symmetric to L,1 (W ) and L,2 (c), respectively, named R,1 (W ) and R,2 (c),
in linear time. We can now prove the following lemma.
Lemma 2. Given a dynamic ow path network with n vertices, CUE tree T
with associated data can be constructed in O(n log n) time and O(n log n) space.

3.2 Using CUE tree

[i,j]
Suppose we want to nd the L-critical vertex cL for subpath P [vi , vj ]. We
work on P [vl(u) , vr(u) ] P[vi , vj ] for each node u N [vi , vj ]. Each such subpath
[i,j]
provides a candidate for cL . Clearly, among those candidates, the one that has
[i,j]
the largest L-time is cL .
Let us rst compute c = c(vr(u)+1 , vj+1 ). For this purpose, we trace the path
(vr(u)+1 , vj+1 ) in T and, at each node u N [vr(u)+1 , vj+1 ] \ {vj+1 }, retrieve
c(vl(u ) , vr(u ) ) and cr(u )+1 . Taking the minimum of the retrieved capacities, we
obtain c(vr(u)+1 , vj+1 ), which costs O(log n) time.
Using binary search, we then nd the largest index q (l(u) q r(u)), if
any, such that c(vq , vr(u)+1 ) < c = c(vr(u)+1 , vj+1 ) holds. Note that c(vq , vr(u)+1 )
is monotonically non-increasing as q decreases. To nd q we trace the path
(vr(u) , u) in T as follows. Set cmin to cr(u)+1 and u to vr(u) . If cmin < c, q is
determined as r(u). Otherwise update u to up . While u
= u and

min{cmin , c(vl(ul ) , vr(ul ) ), cr(ul )+1 } c, (10)

update cmin to the L.H.S. of (10), and u to up . If u = u and (10) holds, such
q does not exist. If (10) stops holding at some node u , then update u to ul .
While
min{cmin , c(vl(ur ) , vr(ur ) ), cr(ur )+1 } c, (11)
update cmin to the L.H.S. of (11) and u to ur . If (11) stops holding at some node
u , then update u to ur . This way we will eventually reach vq , if it exists, in
O(log n) time. If q exists, we partition P [vl(u) , vr(u) ] into two subpaths P [vl(u) , vq ]
and P [vq+1 , vr(u) ]. Letting V1 = V [vl(u) , vq ], V2 = V [vq+1 , vr(u) ], and W = W [vi ,
vl(u)1 ], we dene
 
u [l(u),r(u)] +
L,1 (W ) = max L (vr(u) , vp ) + W/c(vp , vr(u)+1 ) , (12)
vp V1
u
!
L,2 (c) = max d(vp , vr(u) ) + W [vl(u) , vp ]/c . (13)
vp V2

Note that the range of maximization vp V1 in (12) (resp. vp V2 in (13)) is

u
limited compared with (8) (resp. (9)). If q does not exist, we set L,1 (W ) = 0
and V2 = V [vl(u) , vr(u) ]. It is clear that
 
u u
max L,1 (W ), L,2 (c) + W/c (14)
140 B. Bhattacharya et al.

is equal to (6), and its maximizing vertex corresponds to a candidate from

P [vl(u) , vr(u) ] for the L-critical vertex of P [vi , vj ].
Let v1 (resp. v2 ) be a vertex in V1 V2 = V [vl(u) , vr(u) ] which maximizes the
bracketed term in (8) (resp. (9)). Once W = W [vi , vl(u)1 ] and c = c(vr(u)+1 , vj+1 )
are given, we can obtain v1 and v2 by binary search on the bending points of
u u
L,1 (W ) and L,2 (c), respectively, which can be done in O(log n) time. We can
now prove the following lemma.
Lemma 3.
(a) If v2 V1 , we have
u u
L,1 (W ) > L,2 (c) + W/c. (15)

(b) If v1 V2 , we have
u
L,1 (W ) L,2
u
(c) + W/c. (16)

(c) v1 V2 and v2 V1 cannot happen at the same time.

If v1 V1 and v2 V2 , clearly v1 and v2 also achieve the maxima in (12)
and (13), respectively. Therefore, if L,1 u
(W ) > L,2u u
(c) + W/c (resp. L,1 (W )
u
L,2 (c) + W/c), v1 (resp. v2 ) is a candidate critical vertex from P [vl(u) , vr(u) ].
Otherwise, by Lemma 3, v1 , v2 V1 or v1 , v2 V2 holds. Also by Lemma 3, if
v1 , v2 V1 (resp. v1 , v2 V2 ), v1 (resp. v2 ) is a candidate critical vertex. Based
on the above arguments, we can prove the following lemma.

Lemma 4. Suppose that CUE tree T is available. Consider subpath P [vi , vj ]

with 1 i < j n.
(a) For each node u N [vi , vj ], candidates from P [vl(u) , vr(u) ] for L-critical and
R-critical vertices of P [vi , vj ] can be computed in O(log n) time.
(b) The L-critical and R-critical vertices for P [vi , vj ] can be computed in O(log2 n)
time.

4 Building blocks
There are two useful tasks that we can call upon repeatedly. Given the starting
vertex va , the rst task is to nd the rightmost vertex vd such that all the
evacuees on V [va , vd ] can evacuate to a sink within time t. The second task is to
nd the cost of the 1-sink on a given subpath P [vi , vj ]. To perform these tasks,
we start with more basic procedures.

4.1 Basic algorithms

To implement the rst task, note that for a given index h > a, there are O(log n)
nodes in N [va , vh ]. For each such node u, we want to test where a sink s should
be placed: to the left of vl(u) , to the right of vr(u) , or between vl(u) and vr(u) .
Here is an algorithm for the rst task.
Improved Algorithms for Computing k -Sink on Dynamic Flow Path Networks 141

Algorithm 1 1-Sink(t, va )
1. Compute an integer b by binary search over h with a h n such that
[a,b] [a,b+1]
the L-time of cL to vb+ does not exceed t but the L-time of cL +
to vb+1
exceeds t.
[a,b] [a,b]
2. Solve L (vb+ , cL ) + x = t, and place a sink s (vb , vb+1 ] satisfying
d(vb , s) = x.
3. If s (vb , vb+1 ), set c to b + 1. If s = vb+1 , set c to b + 2. Compute an integer
[c,d]
d by binary search over h with c h n such that the R-time of cR to s
[c,d+1]
does not exceed t but the R-time of cR to s exceeds t.

Lemma 5. If CUE tree T is available, 1-Sink(t, va ) runs in O(log3 n) time.

[a,h] [a,h]
Proof. In Step 1, for a xed h, nding cL and computing the L-time of cL
to vh+ take O(log2 n) time by Lemma 4. Clearly, we repeat this computation
O(log n) times, thus Step 1 takes O(log3 n) time. Step 2 takes O(1) time and
Step 3 takes O(log3 n) time similarly to Step 1. Summarizing these, we complete
the proof. 

Here is an algorithm for the second task.
Algorithm 2 Local-Cost(vi , vj )
1. Let u be the node where the two paths (vi , ) and (vj , ) meet.
[i,r(u )] [l(u ),j]
2. If the L-time of cL l and the R-time of cR r have the same value at
some point x on the edge (vr(ul ) , vl(ur ) ), then output x as the 1-sink.
[i,r(u )] [l(u ),j]
3. If the L-time of cL l is higher (resp. lower) than the R-time of cR r at
every point on edge (vr(ul ) , vl(ur ) ), then let u = ul (resp. u = ur ) and repeat
Step 2, using the new ul and ur .
We have the following lemma.
Lemma 6. If CUE tree T is available, Local-Cost(vi , vj ) nds a 1-sink on
subpath P [vi , vj ] in O(log3 n) time.

4.2 t-feasibility test

We carry out 1-Sink(t, v) repeatedly, starting from the left end of P , i.e., v1 .
Clearly, the problem instance is t-feasible if and only if the rightmost vertex vn
belongs to the l-th isolated subpath, where l k.
Lemma 7. Given a dynamic ow path network, if CUE tree T is available, we
can test its t-feasibility in O(min{n log2 n, k log3 n}) time.
Proof. Starting at the leftmost vertex v1 of P , invoke 1-Sink(t, v1 ), which iso-
lates the rst subpath in O(log3 n) time by Lemma 5, and remove it from P .
We repeat this at most k 1 more times on the remaining subpath, spending
O(k log3 n) time.
On the other hand, when each 1-Sink(t, va ) is executed, suppose we compute
[a,h]
the L-time of cL to vh+ for h = a, a + 1, . . . one by one at Step 1, and similarly
142 B. Bhattacharya et al.

[c,h]
the R-time of cR to s for h = c, c + 1, . . . one by one, instead of binary search.
Then, the computations of L-time and R-time are invoked at most n times during
a t-feasibility test. Since each computation of L-time or R-time takes O(log2 n)
time by Lemma 4, the total time is O(n log2 n) in this way. 

4.3 Uniform edge capacity case

The problem is much simplied if the edges have the same capacity. In particular,
we can compute the critical vertex of a subpath resulting from concatenating
two subpaths in constant time. At each node u of T bottom up, we compute
+
and record the L- and R-critical vertices of P [vl(u) , vr(u) ] with respect to vr(u)
and their costs, based on the following lemma.
Lemma 8. [11] For a node u of CUE tree T , let vl(ul ) = vh , vr(ul ) = vi , vl(ur ) =
[h,j] [h,j] [i+1,j]
vi+1 , and vr(ur ) = vj , and assume that the critical vertices, cL , cR , cL ,
[i+1,j]
and cR have already been computed.
[h,j] [h,i] [i+1,j]
(a) The L-critical vertex cL is either cL or cL .
[h,j] [h,i] [i+1,j]
(b) The R-critical vertex cR is either cR or cR .
The following two lemmas provide counterparts to Lemmas 2 and 4, respec-
tively.
Lemma 9. Given a dynamic ow path network with n vertices and uniform
edge capacities, CUE tree T with associated data can be constructed in O(n)
time and O(n) space.
Lemma 10. Suppose that CUE tree T is available. For any i and j (1 i < j
n), we can comput the L-critical and R-critical vertices for P [vi , vj ] in O(log n)
time.
Similarly to Lemma 7, we can prove the following lemma.
Lemma 11. Given a dynamic ow path network with uniform edge capacities,
if CUE tree T is available, we can test its t-feasibility in O(min{n, k log n}) time.

5 Optimization
5.1 Parametric search approach
Lemma 12. [1] If t-feasibility test can be tested in (t) time, then the k-sink
can be found in O(k(t) log n) time, excluding the preprocessing time.
By Lemma 2 it takes O(n log n) time to construct T with weight and capacity
data, and (t) = O(k log3 n) by Lemma 7. Lemma 12 thus implies
Theorem 1. Given a dynamic ow path network with n vertices, we can nd
an optimal k-sink in O(n log n + k 2 log4 n) time.
Applying Megiddos main theorem in [16] to Lemma 11, we obtain
Theorem 2. Given a dynamic ow path network with n vertices and uniform
edge capacities, we can nd an optimal k-sink in O(n + k 2 log2 n) time.
Improved Algorithms for Computing k -Sink on Dynamic Flow Path Networks 143

5.2 Sorted matrix approach

Let OP T (l, r) denote the evacuation time for the optimal 1-sink on subpath
P [vl , vr ]. Dene an n n matrix A whose entry (i, j) entry is given by

OPT (n i + 1, j) if n i + 1 j
A[i, j] = (17)
0 otherwise.

It is clear that matrix A includes OPT (l, r) for every pair of integers (l, r)
such that 1 l r n. There exists a pair (l, r) such that OPT (l, r) is
the evacuation time for the optimal k-sink on the whole path. Then the k-sink
location problem can be formulated as: Find the smallest A[i, j] such that the
given problem instance is A[i, j]-feasible. Note that we do not actually compute
all the elements of A[ ], but element A[i, j] is computed on demand as needed.
A matrix is called a sorted matrix if each row and column of it is sorted in the
nondecreasing order. In [6, 7], Frederickson and Johnson show how to search for
an element in a sorted matrix. The following lemma is implicit in their papers.

Lemma 13. Suppose that A[i, j] can be computed in g(n) time, and feasibility
can be tested in f (n) time with h(n) preprocessing time. Then we can solve the
k-sink problem in O(h(n) + ng(n) + f (n) log n) time.

We have h(n) = O(n log n) by Lemma 2, g(n) = O(log3 n) by Lemma 6, and

f (n) = O(n log2 n) by Lemma 7. Lemma 13 thus implies

Theorem 3. Given a dynamic path network with n vertices and general edge
capacities, we can nd an optimal k-sink in O(n log3 n) time.

In the uniform edge capacity case, we have h(n) = O(n) by Lemma 9, g(n) =
O(log n) by Lemma 10, and f (n) = O(n) by Lemma 11. Lemma 13 thus implies

Theorem 4. Given a dynamic path network with n vertices and uniform edge
capacities, we can nd the k-sink in O(n log n) time.

6 Conclusion and discussion

We have presented more ecient algorithms than the existing ones to solve the
k-sink problem on dynamic ow path networks. Due to lack of space, we could
not present all the proofs. All our results are valid if the model is changed slightly,
so that the weights and edge capacities are not restricted to be integers. Then
it becomes conuent transshipment problem.
For dynamic ow tree networks with uniform edge capacities, it is known
that computing evacuation time to a vertex can be transformed to that on a
path network [13]. We believe that our method is applicable to each spine,
which is a path in the spine decomposition of a tree [2], and we think we may be
able to solve the k-sink problem on dynamic ow tree networks more eciently.
This is work in progress.
144 B. Bhattacharya et al.

References
1. G. P. Arumugam, J. Augustine, M. J. Golin, and P. Srikanthan. A polynomial time
algorithm for minimax-regret evacuation on a dynamic path. arXiv:1404,5448v1,
2014.
2. R. Benkoczi, B. Bhattacharya, M. Chrobak, L. Larmore, and W. Rytter. Faster
algorithms for k-median problems in trees. Mathematical Foundations of Computer
Science, Springer-Verlag, LNCS 2747:218227, 2003.
3. Binay Bhattacharya and Tsunehiko Kameda. Improved algorithms for computing
minmax regret sinks on path and tree networks. Theoretical Computer Science,
607:411425, Nov. 2015.
4. S. W. Cheng, Y. Higashikawa, N. Katoh, G. Ni, B. Su, and Y. Xu. Minimax
regret 1-sink location problem in dynamic path networks. In Proc. Annual Conf.
on Theory and Applications of Models of Computation (T-H.H. Chan, L.C. Lau,
and L. Trevisan, Eds.), Springer-Verlag, volume LNCS 7876, pages 121132, 2013.
5. L. R. Ford and D. R. Fulkerson. Constructing maximal dynamic ows from static
ows. Operations research, 6(3):419433, 1958.
6. G. N. Frederickson. Optimal algorithms for tree partitioning. In Proc. 2nd ACM-
SIAM Symp. Discrete Algorithms, pages 168177, 1991.
7. G. N. Frederickson and D. B. Johnson. Finding kth paths and p-centers by gener-
ating and searching good data structures. J. Algorithms, 4:6180, 1983.
8. H.W. Hamacher and S.A. Tjandra. Mathematical modeling of evacuation prob-
lems: a state of the art. in: Pedestrian and Evacuation Dynamics, Springer Verlag,,
pages 227266, 2002.
9. Y. Higashikawa. Studies on the space exploration and the sink location under
incomplete information towards applications to evacuation planning. PhD thesis,
Kyoto University, Japan, 2014.
10. Y. Higashikawa, M. J. Golin, and N. Katoh. Minimax regret sink location problem
in dynamic tree networks with uniform capacity. J. of Graph Algorithms and
Applications, 18.4:539555, 2014.
11. Y. Higashikawa, M. J. Golin, and N. Katoh. Multiple sink location problems in
dynamic path networks. Theoretical Computer Science, 607:215, 2015.
12. B. Hoppe and E. Tardos. The quickest transshipment problem. Mathematics of
Operations Research, 25(1):3662, 2000.
13. N. Kamiyama, N. Katoh, and A. Takizawa. An ecient algorithm for evacuation
problem in dynamic network ows with uniform arc capacity. IEICE Transactions,
89-D(8):23722379, 2006.
14. S. Mamada, K. Makino, and S. Fujishige. Optimal sink location problem for dy-
namic ows in a tree network. IEICE Trans. Fundamentals, E85-A:10201025,
2002.
15. S. Mamada, T. Uno, K. Makino, and S. Fujishige. An O(n log2 n) algorithm for
a sink location problem in dynamic tree networks. Discrete Applied Mathematics,
154:23872401, 2006.
16. N. Megiddo. Combinatorial optimization with rational objective functions. Math.
Oper. Res., 4:414424, 1979.
17. N. Megiddo and A. Tamir. New results on the complexity of p-center problems.
SIAM J. Comput., 12:751758, 1983.
18. M. Skutella. An introduction to network ows over time. In Research Trends in
Combinatorial Optimization, pages 451482. Springer, 2009.
 A 2-Approximation for the Height of Maximal
Outerplanar Graph Drawings 

Therese Biedl and Philippe Demontigny

David R. Cheriton School of Computer Science, University of Waterloo, Waterloo,

Ontario N2L 1A2, Canada.
[email protected], [email protected]

Abstract In this paper, we study planar drawings of maximal outer-

planar graphs with the objective of achieving small height. (We do not
necessarily preserve a given planar embedding.) A recent paper gave an
algorithm for such drawings that is within a factor of 4 of the optimum
height. In this paper, we substantially improve the approximation fac-
tor to become 2. The main ingredient is to dene a new parameter of
outerplanar graphs (the umbrella depth, obtained by recursively splitting
the graph into graphs called umbrellas). We argue that the height of any
poly-line drawing must be at least the umbrella depth, and then devise
an algorithm that achieves height at most twice the umbrella depth.

1 Introduction

Graph drawing is the art of creating a picture of a graph that is visually appeal-
ing. In this paper, we are interested in drawings of so-called outerplanar graphs,
i.e., graphs that can be drawn in the plane such that no two edges have a point
in common (except at common endpoints) and all vertices are incident to the
outerface. All drawings are required to be planar, i.e., to have no crossing. The
drawing model used is that of at visibility representations where vertices are
horizontal segments and edges are horizontal or vertical segments, but any such
drawing can be transformed into a poly-line drawing (or even a straight-line
drawing if the width is of no concern) without adding height [6].
Every planar graph with n vertices has a straight-line drawing in an n n-
grid [19,9]. Minimizing the area is NP-complete [17], even for outerplanar graphs
[7]. In this paper, we focus on minimizing just one direction of a drawing (we use
the height; minimizing the width is equivalent after rotation). It is not known
whether minimizing the height of a planar drawing is NP-hard (the closest related
result concerns minimizing the height if edges must connect adjacent rows [16]).
Given the height H, testing whether a planar drawing of height H exists is xed
parameter tractable in H [12], but the run-time is exceedingly large in H. As
such, approximation algorithms for the height of planar drawings are of interest.

TB supported by NSERC. Part of this work appeared as PDs Masters thesis [10].

Springer International Publishing AG 2017 145

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 145156, 2017.
DOI: 10.1007/978-3-319-62127-2_13
146 T. Biedl and P. Demontigny

It is known that any graph G with a planar drawing of height H has pw(G)
H [13], where pw(G) is the so-called pathwidth of G. This makes the path-
width a useful parameter for approximating the height of a planar graph draw-
ing. For a tree T , Suderman gave an algorithm to draw T with height at
most  32 pw(T ) [20], making this an asymptotic 32 -approximation algorithm. It
was discovered later that optimum-height drawings can be found eciently for
trees [18]. Approximation-algorithms for the height or width of order-preserving
and/or upward tree drawing have also been investigated [1,2,8].
For outerplanar graphs, the rst author gave two results that will be improved
upon in this paper. In particular, every maximal outerplanar graph has a drawing
of height at most 3 log n1 [3], or alternatively of height 4pw(G)3 [5]. Note that
the second result gives a 4-approximation on the height of drawing outerplanar
graphs, and improving this 4 is the main objective of this paper. A number
of results for drawing outerplanar graphs have been developed since paper [3].
In particular, any outerplanar graph with maximum degree admits a planar
straight-line drawing with area O(n1.48 ) [15], or with area O(n log n) [14].
The former bound was improved to O(n1.48 ) area[11]. Also, every so-called
balanced outerplanar graph can be drawn in an O( n) O( n)-grid [11].
In this paper, we present a 2-approximation algorithm for the height of planar
drawings of maximal outerplanar graphs. The key ingredient is to dene the
so-called umbrella depth ud(G) in Section 3. In Section 4, we show that any
outerplanar graph G has a planar drawing of height at most 2ud(G) + 1. This
algorithm is a relatively minor modication of the one in [5], albeit described
dierently. The bulk of the work for proving a better approximation factor hence
lies in proving a better lower bound, which we do in Section 5: Any maximal
outerplanar graph G with a planar drawing of height H has ud(G) H 1.

2 Preliminaries

Throughout this paper, we assume that G is a simple graph with n 3 vertices

that is maximal outerplanar. Thus, G has a standard planar embedding in which
all vertices are in the outer face (the innite connected region outside the draw-
ing) and form an n-cycle, and all interior faces are triangles. We call an edge
(u, v) of G a cutting edge if G {u, v} is disconnected, and a non-cutting edge
otherwise. In a maximal outerplanar graph, any cutting edge (u, v) has exactly
two cut-components, i.e., there are two maximal outerplanar subgraphs G1 , G2
of G such that G1 G2 = {u, v} and G1 G2 = G.
The dual tree T of G is the weak dual graph of G in the standard embedding,
i.e., T has a vertex for each interior face of G, and an edge between two vertices
i their corresponding faces in G share an edge. An outerplanar path P is a
maximal outerplanar graph whose dual tree is a path. P connects edges e and
e if e is incident to the rst face and e is incident to the last face of the path
that is the dual tree of P . An outerplanar path P with n = 3 is a triangle and
connects any pair of its edges. Since any two interior faces are connected by a
path in T , any two edges e, e of G are connected by some outerplanar path.
A 2-Approximation for the Height of Maximal Outerplanar Graph Drawings 147

r2
r2

2
r1 r1
2

1 1

(a) (b)

Figure 1: (a) A straight-line drawing in the standard embedding, including the

dual tree (dashed edges) and an outerplanar path (shaded) connecting (1 , 2 )
with (r1 , r2 ). (b) A at visibility representation. Both drawings have height 4.

Graph drawing: A drawing of a graph assigns to each vertex a point or an

axis-aligned box, and to each edge a polygonal curve connecting its endpoints.
We only consider planar drawings where none of the points, boxes, or curves
intersect unless the corresponding elements do in the graph. In this paper, a
planar drawing is not required to reect a graphs given planar embedding.
We require that all dening features (points, endpoints of segments, bends) are
placed at points with integer y-coordinates. A layer (or row) is a horizontal line
with integer y-coordinate that intersects elements of the drawing, and the height
is the number of layers.
In a at visibility representation vertices are horizontal line segments, and
edges are vertical or horizontal straight-line segments. (For ease of reading, we
draw vertices as boxes of small height in our illustrations.) In a poly-line drawing
vertices are points and edges are polygonal curves, while in a straight-line draw-
ing vertices are points and edges are line segments. In this paper, we only study
planar at visibility representations, but simply speak of a planar drawing, be-
cause it is known that any planar at visibility representation can be converted
into a planar straight-line drawing of the same height and vice versa [6].

3 Umbrellas, bonnets and systems thereof

In this section, we introduce a method of splitting maximal outerplanar graphs

into systems of special outerplanar graphs called umbrellas and bonnets.

Denition 1. Let G be a maximal outerplanar graph, let U be a subgraph of

G with n 3, and let (u, v) be a non-cutting edge of G. We say that U is an
umbrella with cap (u, v) if

1. U contains all neighbours of u and v,

2. there exists a non-empty outerplanar path P U (the handle) that connects
(u, v) to some non-cutting edge of G, and
3. any vertex of U is either in P or a neighbour of u or v.
148 T. Biedl and P. Demontigny

See also Figure 2(a). For such an umbrella U , the fan at u is the outerplanar
path that starts at an edge (u, x) of the handle P , contains all neighbours of
u, and that is minimal with respect to these constraints. If all neighbours of u
belong to P , then the fan at u is empty. Dene the fan at v similarly, using v.
Any edge (a, b) of U that is a cutting edge of G, but not of U , is called
an anchor-edge of U in G. (In the standard embedding, such edges are on the
outerface of U but not on the outerface of G.) The hanging subgraph with respect
to anchor-edge (a, b) of U in G is the cut-component Sa,b of G with respect to
cutting-edge (a, b) that does not contain the cap (u, v) of U . We often omit of
U in G when umbrella and super-graph are clear from the context.

Denition 2. Let G be a maximal outerplanar graph with n 3, and let (u, v)

be a non-cutting edge of G. An umbrella system U on G with root-edge (u, v) is
a collection U = U0 U1 Uk of subgraphs of G for some k 0 that satisfy
the following:

1. U0 contains only one subgraph U0 (the root umbrella), which is an umbrella

with cap (u, v).
2. U0 has k anchor-edges. We denote them by (ui , vi ) for i = 1, . . . , k, and let
Si be the hanging subgraph with respect to (ui , vi ).
3. For i = 1, . . . , k, Ui (the hanging umbrella system) is an umbrella system of
Si with root-edge (ui , vi ).

The depth of such an umbrella system is dened recursively to be d(U) := 1 +

max1ik d(Ui ); in particular d(U ) = 1 if k = 0.

u v u v

a b
x

(a) (b)

Figure 2: (a) An umbrella system of depth 3. The root umbrella is shaded, with
its handle darker shaded. (b) The same graph has a bonnet system of depth 2,
with the root bonnet shaded and its ribbon darker shaded.

See also Figure 2(a). A graph may have many dierent umbrella systems
with the same root-edge. Dene ud(G; u, v) (the (rooted) umbrella depth of G)
to be the minimum depth over all umbrella systems with root-edge (u, v). Note
that the umbrella depth depends on the choice of the root-edge; dene the free
umbrella depth ud(G) := udfree (G) to be the minimum umbrella depth over all
A 2-Approximation for the Height of Maximal Outerplanar Graph Drawings 149

possible root-edges. (One can show that the free umbrella depth is at most one
unit less than the rooted umbrella depth for any choice of root-edge; see [10].)
Bonnets: A bonnet is a generalization of an umbrella that allows two handles,
as long as they go to dierent sides of the interior face at (u, v). Thus, condition
(2) of the denition of an umbrella gets replaced by
2. There exists a non-empty outerplanar path P U (the ribbon) that connects
two non-cutting edges and contains u, v and their common neighbour.
Other than that, bonnets are dened exactly like umbrellas. See also Figure 2(b).
We dene bonnet system, root bonnet, etc., exactly as for an umbrella system,
except that bonnet is substituted for umbrella everywhere. Let bd(G; u, v)
(the rooted bonnet-depth of G) be the minimum possible depth of a bonnet
system with root-edge (u, v), and let bdfree (G) = bd(G) be the minimum bonnet-
depth over all choices of root-edge. Since any umbrella is a bonnet, we have
bd(G) ud(G).
By denition the root bonnet U0 must contain all edges incident to the ends
u, v of the root-edge. If follows that no edge incident to u or v can be an anchor-
edge of U0 , else the hanging subgraph at it would contain further neighbours of
u (resp. v). We note this trivial but useful fact for future reference:
Observation 1 In a bonnet system with root-edge (u, v), no edge incident to u
or v is an anchor-edge of the root bonnet.

4 From Bonnet System to Drawing

In this section, we show that any outerplanar graph G has a at visibility rep-
resentation of height at most 2ud(G) + 1. We actually show a slightly stronger
bound, namely a height of 2bd(G) + 1 2ud(G) + 1. So x a bonnet system of
G of depth bd(G) with root-edge (u, v). For merging purposes, we want to draw
(u, v) in a special way: It spans the top layer, which means that u touches the
top left corner of the drawing, and v touches the top right corner, or vice versa
(see for example Figure 3(d)). We rst explain how to draw the root bonnet U0 .

Lemma 1. Let U0 be the root bonnet of a bonnet system with root-edge (u, v).
Then there exists a at visibility representation of U0 on three layers such that
1. (u, v) spans the top layer of .
2. Any anchor-edge of U0 is drawn horizontally in the middle or bottom layer.

Proof. As a rst step, we draw the ribbon P of U0 on 2 layers in such a way

that (u, v) and all anchor-edges are drawn horizontally; see Figure 3(a) for an
illustration. (This part is identical to [5].) To do this, consider the standard
embedding of P in which the dual tree is a path, say it consists of faces f1 , . . . , fk .
We draw k + 1 vertical edges between two layers, with the goal that the region
between two consecutive ones belong to f1 , . . . , fk in this order. Place u and
v as segments in the top layer, and with an x-range such that they touch all
150 T. Biedl and P. Demontigny

the regions of faces that u and v are incident to. Similarly create segments for
all other vertices. The placement for the vertices is uniquely determined by the
standard planar embedding, except for the vertices incident to f1 and fk . We
place those vertices such that the leftmost/rightmost vertical edge is not an
anchor-edge. To see that this is possible, recall that P connects two non-cutting
edges e1 , e2 of G that are incident to f1 and fk . If e1
= (u, v), then choose the
layer for the vertices of f1 such that e1 is drawn vertically. If e1 = (u, v), then one
of its ends (say u) is the degree-2 vertex on f1 and drawn in the top-left corner.
The other edge e incident to u is not an anchor-edge of U by Observation 1, and
we draw e vertically. So the leftmost vertical edge is either a non-cutting edge
(hence not an anchor-edge) or edge e (which is not an anchor-edge). We proceed
similarly at fk so that the rightmost vertical edge is not an anchor-edge. Finally
all other vertical edges are cutting edges of U0 and hence not anchor-edges.
The drawing of P obtained in this rst step has (u, v) in the top layer. As a
second step, we now release (u, v) as in [5]. This operation adds a layer above
the drawing, moves (u, v) into it, and re-routes edges at u and v by expanding
vertical ones and turning horizontal ones into vertical ones. In the result, (u, v)
spans the top layer. See Figure 3(b) for an illustration and [5] for details.

u v

fan at v
br
u v br fan at u
rest of P
e  f1 f2 f3 f4 rest of P e2
a
a
(c) Adding the fans. The resulting drawing is not in
(a) Drawing the ribbon. the standard embedding.
u v

x y

u v Sa,b

new layers
br
rest of P Sx,y

a a b

(b) Releasing (u, v). (d) Merging hanging subgraphs.

Figure 3: From bonnet system to drawing.

As the third and nal step, we add the fans. Consider the fan at v, and let
(v, br ) be the edge that it has in common with the ribbon P . Assume rst that
(v, br ) was drawn horizontally after the rst step, see Figure 3(a). After releasing
(u, v) therefore no edge at br attaches on its left, see Figure 3(b). Into this space
A 2-Approximation for the Height of Maximal Outerplanar Graph Drawings 151

we insert, after adding columns, the remaining vertices of the fan at v, in order
in which they appear around v in the standard embedding. See Figure 3(c)).
Else, (v, br ) was drawn vertically after the rst step. (Figure 3(c) does not
illustrate this case for v, but illustrates the corresponding case for u.) Since the
drawing of the rst step is in the standard embedding, and (v, br ) is on the
outerface of the ribbon, therefore (v, br ) must be the rightmost vertical edge.
We can then simply place the vertices of the fan to the right of br and extend v.
The fan at u is placed in a symmetric fashion. It remains to show that all
anchor-edges are horizontal and in the bottom two layers. We ensured that this
is the case in the rst step. Releasing (u, v) adds more vertical edges, but all of
them are incident to u or v and not anchor-edges by Observation 1. Likewise,
all vertical edges added when inserting the fans are incident to u or v. The only
horizontal edge in the top layer is (u, v), which is not an anchor-edge.


Now we explain how to merge hanging subgraphs.

Theorem 1. Any maximal outerplanar graph G has a planar at visibility rep-

resentation of height at most 2bdfree (G) + 1.

Proof. We show by induction that any graph with a bonnet system U of depth
H has a drawing of height 2H + 1 where the root-edge (u, v) spans the top
layer. This proves the theorem when using a bonnet system U of depth bdfree (G).
Let U0 be the root bonnet of the bonnet system, and draw U0 on 3 layers
using Lemma 1. Thus (u, v) spans the top and any anchor-edge (a, b) of U0
is drawn as a horizontal edge in the bottom two layers of 0 . If H = 1 then
there are no hanging subgraphs and we are done. Else add 2H 2 layers to 0
between the middle and bottom layers. For each anchor-edge (a, b) of U0 , the
hanging subgraph Sa,b of U0 has a bonnet system of depth at most H 1 with
root-edge (a, b). By induction Sa,b has a drawing 1 on at most 2H 1 layers
with (a, b) spanning the top layer.
If (a, b) is in the bottom layer of 0 , then we can rotate (and reect, if
necessary) 1 so that (a, b) is in the bottom layer of 1 and the left-to-right
order of a and b in 1 is the same as their left-to-right order in 0 . This updated
drawing of 1 can then be inserted in the space between (a, b) in 0 . This ts
because 1 has height at most 2H 1, and in the insertion process we can re-use
the layer spanned by (a, b). If (a, b) is in the middle layer of U0 , then we can
reect 1 (if necessary) so that (a, b) has the same left-to-right order in 1 as in
0 . This updated drawing of 1 can then be inserted in the space between (a, b)
in 0 . See Figure 3(d). Since we added 2H 2 layers to a drawing of height 3,
the total height of the nal drawing is 2H + 1 as desired.


Our proof is algorithmic, and nds a drawing, given a bonnet system, in

linear time. One can also show (see [10]) that the rooted bonnet depth, and an
associated bonnet system, can be found in linear time using dynamic program-
ming in the dual tree. The free bonnet depth can be found in quadratic time by
trying all root-edges, but one can argue [10] that this will save at most one unit
of depth and hence barely seems worth the extra run-time.
152 T. Biedl and P. Demontigny

Comparison to [5]: The algorithm in [5] has only two small dierences. The
main one is that it does not do the third step when drawing the root bonnet,
thus it draws the ribbon but not the fans. Thus in the induction step our algo-
rithm always draws at least as much as the one in [5]. Secondly, [5] uses a special
construction if pw(G) = 1 to save a constant number of levels. This could easily
be done for our algorithm as well in the case where pw(G) = 1 but bd(G) = 2.
As such, our construction never has worse height (and frequently it is better).
Comparison to [3]: One can argue that bd(G) log(n + 1) (see [10]). Since
[3] uses 3 log n 1 levels while ours uses 2bd(G) + 1 2 log(n + 1) + 1 levels, the
upper bound on the height is better for n 9.

5 From Drawing to Umbrella System

The previous section argued that given an umbrella system (or even more gen-
erally, a bonnet system) of depth H, we can nd a drawing of height at most
2H 1. To show that this is within a factor of 2 of the optimum, we show in this
section that any drawing of height H gives rise to an umbrella system of depth
at most H 1. (Any umbrella system is also a bonnet system, so it also has a
bonnet system of depth at most H 1.)
We rst briey sketch the idea. We assume that we have a at visibility
representation, and further, for some non-cutting edge (u, v) we have an escape
path, i.e., a poly-line to the outerface that does not intersect the drawing. Now
nd an outerplanar path that connects the leftmost vertical edge (x, y) of the
drawing with (u, v). This becomes the handle of an umbrella U with cap (u, v).
One can now argue that any hanging subgraph of U is drawn with height at
most H 1, and furthermore, has an escape path from its anchor-edge. The
claim then holds by induction.
We rst clarify some denitions illustrated in Figure 4(a). Let be a at
visibility representation, and let B be a minimum-height bounding box of .
A vertex w G has a right escape path in if there exists a polyline inside B
from w to a point on the right side of B that is vertex-disjoint from except
at w, and for which all bends are on layers. We say that (r1 , r2 ) is a right-free
edge of if it is vertical, and any layer intersected by (r1 , r2 ) is empty, except
for vertices r1 , r2 , to the right of the edge. In particular, for both r1 and r2 the
rightward ray on its layer is an escape path. Dene left escape paths and left-free
edges symmetrically; an escape path is a left escape path or a right escape path.
Observe that in any at visibility representation any leftmost vertical edge
(v, w) is left-free. (Such vertical edges exist, presuming the graph has minimum
degree 2, since the leftmost vertex in each layer has at most one incident hori-
zontal edge.) For in any layer spanned by (v, w), no vertical edge is farther left
by choice of (v, w), and no vertex can be farther left, else the incident vertical
edge of the leftmost of them would be farther left. So (v, w) is left-free.
For the proof of the lower bound, we use as handle an outerplanar path
connecting to a left-free edge. Recall that the denition of handle requires that
it connects to a non-cutting edge, so we need a left-free edge that is not a cutting
A 2-Approximation for the Height of Maximal Outerplanar Graph Drawings 153

B = B A
1 1 1

w r1 1 w r1

2

2 r2 2 2 r2

(a) (b)

Figure 4: w has a right escape path, (1 , 2 ) is left-free and (r1 , r2 ) is right-free.
After ipping the cutting component at (1 , 2 ), the non-cutting edge (1 , 2 )
becomes left-free.

edge. This does not exist in all drawings (see e.g. Figure 4(a)), but as we show
now, we can modify the drawing without increasing the height such that such an
edge exists. To be able to apply it later, we must also show that this modication
does not destroy a given escape path.
Lemma 2. Let be a at visibility representation of a maximal outerplanar
graph G.
1. Let (r1 , r2 ) be a right-free edge of , and let w be a vertex that has a right
escape path. Then there exists a drawing  in which w has a right escape
path, (r1 , r2 ) is a right-free edge, and there exists a left-free edge that is not
a cutting edge of G.
2. Let (1 , 2 ) be a left-free edge of , and let w be a vertex that has a left
escape path. Then there exists a drawing  in which w has a left escape
path, (1 , 2 ) is a left-free edge, and there exists a right-free edge that is not
a cutting edge of G.
In either case, the y-coordinates of all vertices in are unchanged in  , and
in particular both drawings have the same height.
Proof. We prove the claim by induction on n and show only the rst claim (the
other is symmetric). Let (1 , 2 ) be the leftmost vertical edge of ; this is left-free
as argued above. If (1 , 2 ) is not a cutting edge of G, then we are done with
 = . This holds in particular if n = 3 because then G has no cutting edge.
So assume n 4 and (1 , 2 ) is a cutting edge of G. Let A and B be the cut-
components of (1 , 2 ), named such that w A. Let A [resp. B ] be the drawing
of A [B] induced by . Edge (1 , 2 ) is left-free for both A and B . Reect B
horizontally (this makes (1 , 2 ) right-free) to obtain B . By induction, we can
create a drawing B from B in which (1 , 2 ) is right-free and there is a left-free
edge (1 , 2 ) that is not a cutting edge of B. We have (1 , 2 )
= (1 , 2 ), because
the common neighbour of 1 , 2 in B forces a vertex or edge to reside to the left
of the right-free edge (1 , 2 ). So (1 , 2 ) is not a cutting edge of G either.
As in Figure 4(b), create a new drawing that places B to the left of A and
extends 1 and 2 to join the two copies; this is possible since (1 , 2 ) has the
154 T. Biedl and P. Demontigny

same y-coordinates in A , , B and B , and it is left-free in A and right-free

in B . Also delete one copy of (1 , 2 ). The drawing A is unchanged, so w will
have the same right escape path in  as in , and  will have right-free edge
(r1 , r2 ) and left-free non-cutting edge (1 , 2 ), as desired.


We are now ready to prove the lower bound if there is an escape path.

Lemma 3. Let be a at visibility representation of a maximal outerplanar

graph G with height H, and let (u, v) be a non-cutting edge of G. If there exists
an escape path from u or v in , then G has an umbrella system with root-edge
(u, v) and depth at most H 1.

Proof. We proceed by induction on H. Assume without loss of generality that

there exists a right escape path from v (all other cases are symmetric). Using
Lemma 2, we can modify without increasing the height so that v has a right
escape path, and there is a left-free edge (1 , 2 ) in that is a not a cutting edge
of G. Let P be the outerplanar path that connects edge (1 , 2 ) and (u, v). Let
U0 be the union of P , the neighbors of u, and the neighbors of v; we use U0 as
the root umbrella of an umbrella system.
We now must argue that all hanging subgraphs of U0 are drawn with height
at most H 1 and have escape paths from their anchor-edges; we can then nd
umbrella systems for them by induction and combining them with U0 gives the
umbrella system for G as desired. To prove the height-bound, dene dividing
paths as follows. The outerface of U0 in the standard embedding contains (1 , 2 )
(since it is not a cutting edge) as well as v. Let P1 and P2 be the two paths from
1 and 2 to v along this outerface in the standard embedding. Dene the dividing
path i (for i = 1, 2) to be the poly-line in that consists of the leftward ray
from i , the drawing of the path Pi (i.e., the vertical segments of its edges and
parts of the horizontal segments of its vertices), and the right escape path from
v. See Figure 5.

u v
u
P1
Sa,b
b 1
Sa,b v

a P2 1
a
b 2
1

2 2

Figure 5: Extracting dividing paths from a at visibility representation. P1 /1

is dotted while P2 /2 is dashed.

Now consider any hanging subgraph Sa,b of U0 with anchor-edge (a, b). No
edge incident to v is an anchor-edge, and neither is (1 , 2 ), since it is not a
cutting edge. So (a, b) is an edge of P1 or P2 (say P1 ) that is not incident to v.
A 2-Approximation for the Height of Maximal Outerplanar Graph Drawings 155

Therefore (a, b) (and with it Sa,b ) is vertex-disjoint from P2 . It follows that the
drawing S of Sa,b induced by is disjoint from the dividing path 2 . Since
2 connects a point on the left boundary with a point on the right boundary,
therefore S must be entirely above or entirely below 2 , say it is above. Since
2 has all bends at points with integral y-coordinate, therefore the bottom layer
of is not available for S , and S has height at most H 1 as desired.
Recall that (a, b) belongs to P1 and is not incident to v. After possible re-
naming of a and b, we may assume that b is closer to 1 along P1 than a. Then
the sub-path of P1 from b to 1 is interior-disjoint from Sa,b . The part of 1
corresponding to this path is a left escape path from b that resides within the
top H 1 layers, because it does not contain v and hence is disjoint from 2 .
We can hence apply induction to Sa,b to obtain an umbrella system of depth at
most H 2 with root-edge (a, b). Repeating this for all hanging subgraphs, and
combining the resulting umbrella systems with U0 , gives the result.

Theorem 2. Let G be a maximal outerplanar graph. If G has a at visibility
representation of height H, then udfree (G) H 1.
Proof. Using Lemma 2, we can convert into a drawing  of the same height in
which some edge (u, v) is a right-free non-cutting edge. This implies that there is
a right escape path from v, and by Lemma 3 we can nd an umbrella system of
G with root-edge (u, v) and depth H 1. So udfree (G) ud(G; u, v ) H 1. 

6 Conclusions and Future Work

We presented an algorithm for drawing maximal outerplanar graphs that is a
2-approximation for the optimal height. To this end, we introduced the umbrella
depth as a new graph parameter for maximal outerplanar graphs, and used as
key result that any drawing of height H implies an umbrella-depth of at least
H 1. Our result improves the previous best result, which was based on the
pathwidth and gave a 4-approximation. We close with some open problems:
Our result only holds for maximal outerplanar graphs. Can the algorithm be
modied so that it becomes a 2-approximation for all outerplanar graphs?
Clearly one could apply the algorithm after adding edges to make the graph
maximal, but which edges should be added to keep the umbrella depth small?
The algorithm from Section 4 creates a drawing that does not place all
vertices on the outerface. Can we create an algorithm that approximates the
optimal height in the standard planar embedding?
What is the width achieved by the algorithm from Section 4 if we enforce
integral x-coordinates? Any visibility representation can be modied without
changing the height so that the width is at most m+n, where m is the number
of edges and n is the number of vertices [6]. Thus the width is O(n), but
what is the constant?
Finally, can we determine the optimal height for maximal outerplanar graphs in
polynomial time? This question is of interest both if (as in our algorithm) the
embedding can be changed, or if the drawing must be in the standard embedding.
156 T. Biedl and P. Demontigny

References
1. Md. J. Alam, Md. A.H. Samee, M. Rabbi, , and Md. S. Rahman. Minimum-layer
upward drawings of trees. J. Graph Algorithms Appl., 14(2):245267, 2010.
2. J. Batzill and T. Biedl. Order-preserving drawings of trees with approximately
optimal height (and small width), 2016. CoRR 1606.02233 [cs.CG]. In submission.
3. T. Biedl. Drawing outer-planar graphs in O(n log n) area. In S. Kobourov and
M. Goodrich, editors, Graph Drawing (GD01), volume 2528 of LNCS, pages 54
65. Springer-Verlag, 2002. Full version included in [4].
4. T. Biedl. Small drawings of outerplanar graphs, series-parallel graphs, and other
planar graphs. Discrete and Computational Geometry, 45(1):141160, 2011.
5. T. Biedl. A 4-approximation algorithm for the height of drawing 2-connected out-
erplanar graphs. In T. Erlebach and G. Persiano, editors, Workshop on Approxi-
mation and Online Algorithms (WAOA12), volume 7846 of LNCS, pages 272285.
Springer-Verlag, 2013.
6. T. Biedl. Height-preserving transformations of planar graph drawings. In C. Dun-
can and A. Symvonis, editors, Graph Drawing (GD14), volume 8871 of LNCS,
pages 380391. Springer, 2014.
7. T. Biedl. On area-optimal planar grid-drawings. In J. Esparza, P. Fraigniaud,
T. Husfeldt, and E. Koutsoupias, editors, International Colloquium on Automata,
Languages and Programming (ICALP 14), volume 8572 of LNCS, pages 198210.
Springer-Verlag, 2014.
8. T. Biedl. Ideal tree-drawings of approximately optimal width (and small height).
Journal of Graph Algorithms and Applications, 21(4):631648, 2017.
9. H. de Fraysseix, J. Pach, , and R. Pollack. How to draw a planar graph on a grid.
Combinatorica, 10:4151, 1990.
10. P. Demontigny. A 2-approximation for the height of maximal outerplanar graphs.
Masters thesis, University of Waterloo, 2016. See also CoRR report 1702.01719.
11. G. Di Battista and F. Frati. Small area drawings of outerplanar graphs. Algorith-
mica, 54(1):2553, 2009.
12. V. Dujmovic, M. Fellows, M. Kitching, G. Liotta, C. McCartin, N. Nishimura,
P. Ragde, F. Rosamond, S. Whitesides, , and D. Wood. On the parameterized
complexity of layered graph drawing. Algorithmica, 52:267292, 2008.
13. S. Felsner, G. Liotta, , and S. Wismath. Straight-line drawings on restricted integer
grids in two and three dimensions. J. Graph Alg. Appl, 7(4):335362, 2003.
14. F. Frati. Straight-line drawings of outerplanar graphs in O(dn log n) area. Comput.
Geom., 45(9):524533, 2012.
15. A. Garg and A. Rusu. Area-ecient planar straight-line drawings of outerplanar
graphs. Discrete Applied Mathematics, 155(9):11161140, 2007.
16. L.S. Heath and A.L. Rosenberg. Laying out graphs using queues. SIAM Journal
on Computing, 21(5):927958, 1992.
17. M. Krug and D. Wagner. Minimizing the area for planar straight-line grid drawings.
In S. Hong, T. Nishizeki, and W. Quan, editors, Graph Drawing (GD07), volume
4875 of LNCS, pages 207212. Springer-Verlag, 2007.
18. D. Mondal, Md. J. Alam, , and Md. S. Rahman. Minimum-layer drawings of trees.
In N. Katoh and A. Kumar, editors, Algorithms and Computations (WALCOM
2011), volume 6552 of LNCS, pages 221232. Springer, 2011.
19. W. Schnyder. Embedding planar graphs on the grid. In ACM-SIAM Symposium
on Discrete Algorithms (SODA 90), pages 138148, 1990.
20. M. Suderman. Pathwidth and layered drawings of trees. Intl. J. Comp. Geom.
Appl, 14(3):203225, 2004.
Splitting B2 -VPG Graphs into Outer-string and
Co-comparability Graphs

Therese Biedl and Martin Derka

David R. Cheriton School of Computer Science,

University of Waterloo, Waterloo, ON, N2L 3G1, Canada
{biedl,mderka}@uwaterloo.ca

Abstract. A B2 -VPG representation of a graph is an intersection rep-

resentation that consists of orthogonal curves with at most 2 bends. In
this paper, we show that the curves of such a representation can be
partitioned into O(log n) groups that represent outer-string graphs or
O(log3 n) groups that represent permutation graphs. This leads to bet-
ter approximation algorithms for hereditary graph problems, such as
independent set, clique and clique cover, on B2 -VPG graphs.

1 Introduction

An intersection representation of a graph is a way of portraying a graph using

geometric objects. In such a representation, every object corresponds to a vertex
in the graph, and there is an edge between vertices u and v if and only if their two
objects u and v intersect. One example are the string graphs, where the objects
are (open) curves in the plane with no intersections that are overlaps or touch
points. An outer-string representation is one where all the curves are inside a
polygon P and touch the boundary of P at least once. A string representation
is called a 1-string representation if any two strings intersect at most once. It is
called a Bk -VPG-representation1 (for some k 0) if every curve is an orthogonal
curve with at most k bends. We naturally use the term outer-string graph for
graphs that have an outer-string representation, and similarly for other types of
intersecting objects.
Our contribution: This paper is concerned with partitioning string graphs (and
other classes of intersection graphs) into subgraphs that have nice properties,
such as being outer-string graphs or permutation graphs (dened formally be-
low). We can then use such a partition to obtain approximation algorithms for
some graph problems, such as weighted independent set, clique, clique cover and
colouring. More specically, partitioning in this paper usually means a vertex
partition, i.e., we split the vertices of the graph as V = V1 Vk such that
the subgraph induced by each Vi has nice properties. In one case we also do

T.B. was supported by NSERC; M.D. was supported by Vanier CGS.
1
Vertex intersection graphs of k-bend Paths in a Grid; see [8] and the references
therein.

Springer International Publishing AG 2017 157

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 157168, 2017.
DOI: 10.1007/978-3-319-62127-2_14
158 T. Biedl and M. Derka

an edge-partition where we partition E = E1 E2 and then work on the two

subgraphs Gi = (V, Ei ), for i = 1, 2.
Our paper was inspired by a paper by Lahiri et al. [8] in 2014. They gave
an algorithm to approximate the maximum (unweighted) independent set in a
B1 -VPG graph within a factor of 4 log2 n (log in this paper denotes log2 ). We
greatly expand on their approach as follows. First, rather than solving maxi-
mum independent set directly, we instead split such a graph into subgraphs.
This allows us to approximate not just independent set, but more generally any
hereditary graph problem that is solvable in such graphs.
Secondly, rather than using co-comparability graphs for splitting as Lahiri
et al. did, we use outer-string graphs. This allows us to stop the splitting ear-
lier, reducing the approximation factor from 4 log2 n to 2 log n, and to give an
algorithm for weighted independent set (wIS).
Finally, we allow much more general shapes. For splitting into outer-string
graphs, we can allow any shape that can be described as the union of one vertical
and any number of horizontal segments (we call such intersection graphs single-
vertical). Our results imply a 2 log n-approximation algorithm for wIS in such
graphs, which include B1 -VPG graphs, and a 4 log n-approximation for wIS in
B2 -VPG graphs.
In the second part of the paper, we consider splitting the graph such that the
resulting subgraphs are co-comparability graphs. This type of problem was rst
considered by Keil and Stewart [7], who showed that so-called subtree lament
graphs can be vertex-partitioned into O(log n) co-comparability graphs. The
work of Lahiri et al. [8] can be seen as proving that every B1 -VPG graph can be
vertex-partitioned into O(log2 n) co-comparability graphs. We focus here on the
bigger class of B2 -VPG-graphs, and show that they can be vertex-partitioned
into O(log3 n) co-comparability graphs. Moreover, these co-comparability graphs
have poset dimension 3, and if the B2 -VPG representation was 1-string, then
they are permutation graphs. This leads to better approximation algorithms for
clique, colouring and clique cover for B2 -VPG graphs.

2 Decomposing into outer-string graphs

We argue in this section how to split a graph into outer-string graphs if it

has an intersection representation of a special form. A single-vertical object is a
connected set S R2 of the form S = s0 s1 sk , where s0 is a vertical
segment and s1 , . . . , sk are horizontal segments, for some nite k. Given a number
of single-vertical objects S1 , . . . , Sn , we dene the intersection graph of it in
the usual way, by dening one vertex per object and adding an edge whenever
objects have at least one point in common (contacts are considered intersections).
We call such a representation a single-vertical representation and the graph a
single-vertical intersection graph. The x-coordinate of one single-vertical object
is dened to be the x-coordinate of the (unique) vertical segment. We consider
a horizontal segment to be a single-vertical object as well, by attaching a zero-
length vertical segment at one of its endpoints.
Splitting B2-VPG Graphs into Outer-string and Co-comparability Graphs 159

Theorem 1. Let G be a single-vertical intersection graph. Then the vertices of

G can be partitioned into at most max{1, 2 log n} sets such that the subgraph
induced by each is an outer-string graph.2
Our proof of Theorem 1 uses a splitting technique implicit in the recursive
approximation algorithm of Lahiri et al. [8]. Let R be a single-vertical represen-
tation on G and S be an ordered list of the x-coordinates of all the objects in R.
We dene the median m of R to be a value such that at most |S| 2 x-coordinates
|S|
in S are smaller than m and at most 2 x-coordinates in S are bigger than m.
(m may or may not be the x-coordinate of at least one object.) Now split R into
three sets: The middle set M of objects that intersect the vertical line m with
x-coordinate m; the left set L of objects whose x-coordinates are smaller than m
and that do not belong to M , and the right set R of objects whose x-coordinates
are bigger than m and that do not belong to M . Split M further into ML = {
c | the x-coordinate of c is less than m} and MR = M \ ML .

L ML MR R

Fig. 1: The split of a representation into L, M = ML MR and R.

Lemma 2. The subgraph induced by the objects in ML is outer-string.

Proof. All the objects in ML intersect curve m. Since all the x-coordinates of
those objects are smaller than m, all the intersections of the objects occur left
of m. If an object is not a curve, one can replace it by a closed curve that traces
around the part of the object that is left of m and that intersects the same
set of objects (possibly repeatedly). Doing so for every object, one obtains a
representation that induces the same graph as ML and where all curves are in
the left half-space of m and intersect m, hence an outer-string representation.


2
This bound is not tight; a more careful analysis shows that we get at most
max{1, 2log n 2} graphs.
160 T. Biedl and M. Derka

A similar proof shows that the graph induced by objects in MR is an outer-

string graph. Now we can prove our main result:

Proof (of Theorem 1). Let G be a graph with a single-vertical representation.

We proceed by induction on the number of vertices n in G. If n 2, then the
graph is outer-string and we are done, so assume n 3, which implies that
log n 32 . By Lemma 2, both ML and MR individually induce an outer-string
graph. Applying induction, we get at most

max{1, 2 log |L|} max{1, 2 log(n/2)} = max{1, 2 log n 2} = 2 log n 2

outer-string subgraphs for L, and similarly at most 2 log n 2 outer-string sub-

graphs for R. Since the objects in L and R are separated by the vertical line
m, there are no edges between the corresponding vertices. Thus any outer-string
subgraph dened by L can be combined with any outer-string subgraph dened
by R to give one outer-string graph. We hence obtain 2 log n 2 outer-string
graphs from recursing into L and R. Adding to this the two outer-string graphs
dened by ML and MR gives the result.


Our proof is constructive, and nds the partition within O(log n) recursions.
In each recursion we must nd the median m and then determine which objects
intersect the line m. If we presort three lists of the objects (once by x-coordinate
of the vertical segment, once by leftmost x-coordinate, and once by rightmost
x-coordinate), and pass these lists along as parameters, then each recursion
can be done in O(n) time, without linear-time median-nding. The presorting
takes O(N + n log n) time, where N is the total number of segments in the
representation. Hence the run-time to nd the partition is O(N + n log n).
The above results were for single-vertical graphs. However, the main focus of
this paper is Bk -VPG-graphs, for k 2. Clearly B1 -VPG graphs are single-
vertical by denition. It is not obvious whether B2 -VPG graphs are single-
vertical graphs or not. Note that a B2 -VPG representation may not be a single-
vertical representationit may consist of both curves with two horizontal seg-
ments and curves with two vertical segments (so no rotation of the representation
can give a single-vertical representation). However, we can still handle them by
doubling the number of graphs into which we split.

Lemma 3. Let G be a B2 -VPG graph. Then the vertices of G can be partitioned

into 2 sets such that the subgraph induced by each is a single-vertical B2 -VPG
graph.

Proof. Fix a B2 -VPG-representation of G. Let Vv be the vertices that have at

most one vertical segment in their curve, and Vh be the remaining vertices. Since
every curve has at most three segments, and all curves in Vh have at least two
vertical segments, each of them has at most one horizontal segment. Clearly Vv
induces a single-vertical B2 -VPG graph, and after rotating all curves by 90 Vh
also induces a single-vertical B2 -VPG graph.


Combining this with Theorem 1, we immediately obtain:

Splitting B2-VPG Graphs into Outer-string and Co-comparability Graphs 161

Corollary 4. Let G be a B2 -VPG graph. Then the vertices of G can be parti-

tioned into at most max{1, 4 log n}sets such that the subgraph induced by each is
an outer-string graph.

3 Decomposing into co-comparability graphs

We now show that by doing further splits, we can actually decompose B2 -VPG
graphs into so-called co-comparability graphs of poset dimension 3 (dened for-
mally below). While we require more subgraphs for such a split, the advantage is
that numerous problems are polynomial for such co-comparability graphs, while
for outer-string we know of no problem other than weighted independent set
that is poly-time solvable.
We rst give an outline of the approach. Given a B2 -VPG-graph, we rst
use Lemma 3 to split it into two single-vertical B2 -VPG-graphs. Given a single-
vertical B2 -VPG-graph, we next use a technique much like the one of Theorem 1
to split it into log n single-vertical B2 -VPG-graphs that are centered in some
sense. Any such graph can easily be edge-partitioned into two B1 -VPG-graphs
that are grounded in some sense. We then apply the technique of Theorem 1
again (but in the other direction) to split a grounded B1 -VPG-graph into log n
B1 -VPG-graphs that are cornered in some sense. The latter graphs can be
shown to be permutation graphs. This gives the result after arguing that the
edge-partition can be un-done at the cost of combining permutation graphs into
co-comparability graphs.
We assume for this section that the B2 -VPG representation is in general
position in the sense that no two horizontal or vertical segments overlap each
other. Since curves do not overlap or touch, this is not a restriction for B2 -VPG
representations.

3.1 Co-comparability graphs

We start by dening the graph classes that we use in this section only. A graph
G with vertices {1, . . . , n} is called a permutation graph if there exist two permu-
tations 1 , 2 of {1, . . . , n} such that (i, j) is an edge of G if and only if 1 lists
i, j in the opposite order as 2 does. Put dierently, if we place 1 (1), . . . , 1 (n)
at points along a horizontal line, and 2 (1), . . . , 2 (n) at points along a parallel
horizontal line, and use the line segment (1 (i), 2 (i)) to represent vertex i, then
the graph is the intersection graph of these segments.
A co-comparability graph G is a graph whose complement can be directed
in an acyclic transitive fashion. Rather than dening these terms, we describe
here only the restricted type of co-comparability graphs that we are interested
in. A graph G with vertices {1, . . . , n} is called a co-comparability graph of poset
dimension k if there exist k permutations 1 , . . . , k such that (i, j) is an edge if
and only if there are two permutations that list i and j in opposite order. (See
Golumbic et al. [5] for more on these characterizations.) Note that a permutation
graph is a co-comparability graph of poset dimension 2.
162 T. Biedl and M. Derka

(a) (b) (c) 3 (d)

Fig. 2: (a) A graph that has simultaneously (b) a permutation representation;

(c) a co-comparability representation of poset dimension 3; and (d) a cornered
B1 -VPG graph.

3.2 Cornered B1 -VPG graphs

A B1 -VPG-representation is called cornered if there exists a horizontal and a
vertical ray emanating from the same point such that any curve of the repre-
sentation intersects both rays outside their common end. See Fig. 2(d) for an
example.

Lemma 5. If G has a cornered B1 -VPG-representation, say with respect to rays

r1 and r2 , then G is a permutation graph. Further, the two permutations dening
G are exactly the two orders in which vertex-curves intersect r1 and r2 .

Proof. Since the curves have only one bend, the intersections with r1 and r2
determine the curve of each vertex. In particular, two curves intersect if and
only if the two orders along r1 and r2 are not the same, which is to say, if their
orders are dierent in the two permutations of the vertices dened by the orders
along the rays. Hence using these orders shows that G is a permutation graph.



3.3 From grounded to cornered

We call a B1 -VPG representation grounded if there exists a horizontal line
segment H that intersects all curves, and has all horizontal segments of all
curves above it. See also Fig. 3 and [1] for more properties of graphs that have
a grounded representation. We now show how to split a grounded B1 -VPG-
representation into cornered ones. It will be important later that not only can
we do such a split, but we know how the curves intersect H afterwards. More
precisely, the curves in the resulting representations may not be identical to
the ones we started with, but they are modied only in such a way that the
intersections points of curves along H is unchanged.

Lemma 6. Let R be a B1 -VPG-representation that is grounded with respect

to segment H . Then R can be partitioned into at most max{1, 2 log n} sets
R1 , . . . , RK such that each set Ri is cornered after upward translation and segment-
extension of some of its curves.

Proof. A single curve with one bend is always cornered, so the claim is easily
shown for n 4 where max{1, 2 log n} n. For n 5, it will be helpful to
Splitting B2-VPG Graphs into Outer-string and Co-comparability Graphs 163

split R rst into two sets, those curves of the form |h and those that form h| (no
other shapes can exist in a grounded B1 -VPG-representation). The result follows
if we show that each of them can be split into log n many cornered B1 -VPG-
representations.
So assume that R consists of only |hs. We apply essentially the same idea as
in Theorem 1. Let again m be the vertical line along the median of x-coordinates
of vertical segments of curves. Let M be all those curves that intersect m. Since
curves are |hs, any curve in M intersects H to the left of m, and intersects
m above H . Hence taking the two rays along H and m emanating from their
common point shows that M is cornered.

m
m

lh GL GR lh

Fig. 3: An illustration for the proof of Lemma 6. (left) Splitting a cornered B1 -

VPG graph. (right) Combining a graph GL (solid black) from L with a graph
GR (solid blue) from R so that the result is a cornered B1 -VPG graph.

We then recurse both in the subgraph L of vertices entirely left of m and the
subgraph R of vertices entirely right of m. Each of them is split recursively into
at most max{1, log(n/2)} = log n 1 subgraphs that are cornered. We must now
argue how to combine two such subgraphs GL and GR (of vertices from L and
R) such that they are cornered while modifying curves only in the permitted
way.
Translate curves of GL upward such that the lowest horizontal segment of
GL is above the highest horizontal segment of GR . Extend the vertical segments
of GL so that they again intersect H . Extend horizontal segments of both GL
and GR rightward until they all intersect one vertical line segment. The resulting
representation satises all conditions.
Since we obtain at most log n 1 such cornered representations from the
curves in R L, we can add M to it and the result follows.

Corollary 7. Let G be a graph with a grounded B1 -VPG representation. Then
the vertices of G can be partitioned into at most max{1, 2 log n} sets such that
the subgraph induced by each is a permutation graph.

3.4 From centered to grounded

We now switch to VPG-representations with 2 bends, but currently only allow
those with a single vertical segment per curve. So let R be a single-vertical B2 -
164 T. Biedl and M. Derka

VPG-representation. We call R centered if there exists a horizontal line segment

H that intersects the vertical segment of each curve. Given such a representa-
tion, we can cut each curve apart at the intersection point with H . Then the
parts above H form a grounded B1 -VPG-representation, and the parts below
form (after a 180 rotation) also a grounded B1 -VPG-representation. Note that
this split corresponds to splitting the edges into E = E1 E2 , depending on
whether the intersection for each edge occurs above or below H . If curves may
intersect repeatedly, then an edge may be in both sets. See Fig. 4 for an example.
With this, we can now split into co-comparability graphs.



Fig. 4: Splitting a centered single-vertical B2 -VPG-representation into two

grounded B1 -VPG-representations.

Lemma 8. Let G be a graph with a single-vertical centered B2 -VPG represen-

tation. Then the vertices of G can be partitioned into at most max{1, 4 log2 n}
sets such that the subgraph induced by each is a co-comparability graph of poset
dimension 3.

Proof. The claim clearly holds for n 4, so assume n 5. Let H be the

horizontal segment along which the representation is centered. Split the edges
into E1 and E2 as above, and let R1 and R2 be the resulting grounded B1 -
VPG-representations, which have the same order of vertical intersections along
H . Split R1 into K 2 log n sets of curves R11 , . . . , R1K , each of which forms a
cornered B1 -VPG-representation that uses the same order of intersections along

H . Similarly split R2 into K  2 log n sets R21 , . . . , R2K of cornered B1 -VPG-
representations.
Now dene Ri,j to consist of all those curves r where the part of r above H
belongs to R1i and the part below belongs to R2j . This gives K K  4 log2 n
sets of curves. Consider one such set Ri,j . The parts of curves in Ri,j that
were above H are cornered at H and some vertical upward ray, hence dene
Splitting B2-VPG Graphs into Outer-string and Co-comparability Graphs 165

a permutation 1 along the vertical ray and 2 along H . Similarly the parts
of curves below H dene two permutations, say 2 along H and 3 along
some vertical downward ray. But the split into cornered B1 -VPG-representation
ensured that the intersections along H did not changed, so 2 = 2 . The three
permutations 1 , 2 , 3 together hence dene a co-comparability graph of poset
dimension 3 as desired.

We can do slightly better if the representation is additionally 1-string.
Corollary 9. Let G be a graph with a single-vertical centered 1-string B2 -VPG
representation.Then the vertices of Gcan be partitioned into at most max{1, 4 log2 n}
sets such that the subgraph induced by each is a permutation graph.
Proof. The split is exactly the same as in Lemma 8. Consider one of the sub-
graphs Gi and the permutations 1 , 2 , 3 that came with it, where 2 is the
permutation of curves along the centering line H . We claim that Gi is a per-
mutation graph, using 1 , 3 as the two permutations. Clearly if (u, v) is not an
edge of Gi , then all of 1 , 2 , 3 list u and v in the same order. If (u, v) is an
edge of Gi , then two of 1 , 2 , 3 list u, v in opposite order. We claim that 1
and 3 list u, v in opposite order. For if not, say u comes before v in both 1 and
3 , then (to represent edge (u, v)) we must have u after v in 2 . But then the
curves of u and v intersect both above and below H , contradicting that we have
a 1-string representation. So the two permutations 1 , 3 dene graph Gi .


3.5 Making single-vertical B2 -VPG-representations centered

Lemma 10. Let G be a graph with a single-vertical B2 -VPG representation.
Then the vertices of G can be partitioned into at most max{1, log n} sets such that
the subgraph induced by each has a single-vertical centered B2 -VPG-representation.
Proof. The approach is quite similar to the one in Theorem 1, but uses a hor-
izontal split and a dierent median. The claim is easy to show for n 3, so
assume n 4. Recall that there are n vertical segments, hence 2n endpoints of
such segments. Let m be a value such that at most n of these endpoints each
are below and above m, and let m be the horizontal line with y-coordinate m.
Let M be the curves that are intersected by m; clearly they form a single-
vertical centered B2 -VPG-representation. Let B be all those curves whose ver-
tical segment (and hence the entire curve) is completely below m. Each such
curve contributes two endpoints of vertical segments, hence |B| n/2 by choice
of m. Recursively split B into at most max{1, log(n/2)} = log n 1 sets, and
likewise split the curves U above m into at most log n 1 sets.
Each chosen subset GB of B is centered, as is each chosen subset GU of
U . Since GB uses curves below m while GU uses curves above, there are no
crossings between these curves. We can hence translate the curves of GB such
they are centered with the same horizontal line as GU . Therefore GB GU has a
centered single-vertical B2 -VPG-representation. Repeating this for all of R U
gives log n 1 centered single-vertical B2 -VPG-graphs, to which we can add the
one dened by M .

166 T. Biedl and M. Derka

3.6 Putting it all together

We summarize with our main result about splits into co-comparability graphs:
Theorem 11. Let G be a B2 -VPG-graph. Then the vertices of G can be parti-
tioned into at most max{1, 8 log3 n} sets such that the subgraph induced by each
is co-comparability graph of poset dimension 3. If G is a 1-string B2 -VPG graph,
then the subgraphs are permutation graphs.
Proof. The claim is trivial for n 3 since then n 8 log3 n, so assume n 4.
Fix a B2 -VPG-representation R, and split it into two single-vertical B2 -VPG-
representations (Lemma 3). Split each of them into log n single-vertical centered
B2 -VPG-representations using Lemma 10, for a total of at most 2 log n sets of
curves. Split each of them into 4 log2 n co-comparability graphs (or permutation
graphs if the representation was 1-string) using Lemma 8 or Corollary 9.

We can do better for B1 -VPG-graphs. The subgraphs obtained in the result
below are the same ones that were used implicitly in the 4 log2 n-approximation
algorithm given by Lahiri et al.[8].
Theorem 12. Let G be a B1 -VPG-graph. Then the vertices of G can be parti-
tioned into at most max{1, 4 log2 n} sets such that the subgraph induced by each
is a permutation graph.
Proof. The claim is trivial if n = 1, so assume n > 1. Fix a B1 -VPG-representation
R, and split it into log n single-vertical centered B1 -VPG-representations using
Lemma 10. Split each of them into two centered B1 -VPG-representations, one
of those curves with the horizontal segment above the centering line, and one
with the rest. Each of the resulting 2 log n centered B1 -VPG-representations is
grounded (possibly after a 180 rotation) and can be split into 2 log n permuta-
tion graphs using Corollary 7, for a total of 4 log2 n permutation graphs.


4 Applications
We now show how Theorem 1 and 11 can be used for improved approximation
algorithms for B2 -VPG-graphs. The techniques used here are virtually the same
as the one by Keil and Stewart [7] and require two things. First, the problem
considered needs to be solvable on the special graphs class (such as outer-string
graph or co-comparability graph or permutation graph) that we use. Second,
the problem must be hereditary in the sense that a solution in a graph implies
a solution in an induced subgraph, and solutions in induced subgraphs can be
used to obtain a decent solution in the original graph.
We demonstrate this in detail using weighted independent set, which Keil
et al. showed to be polynomial-time solvable in outer-string graphs [6]. Recall
that this is the problem, given a graph with vertex-weights, of nding a subset
I of vertices
 that has no edges between them. The objective is to maximize
w(I) := vI w(v), where w(v) denotes the weight of vertex v. The run-time to
solve weighted independent set in outer-string graphs is O(N 3 ), where N is the
number of segments in the given outer-string representation.
Splitting B2-VPG Graphs into Outer-string and Co-comparability Graphs 167

Theorem 13. There exists a (2 log n)-approximation algorithm for weighted in-
dependent set on single-vertical graphs with run-time O(N 3 ), where N is the total
number of segments used among all single-vertical objects.
Proof. If n = 1, then the unique vertex is the maximum weight independent set.
Else, use Theorem 1 to partition the vertices of the given graph G into at most
2 log n sets, each of which induces an outer-string graph, and nd the largest
weighted independent set in each applying the algorithm of Keil et al. If Gi had
an outer-string representation with Ni segments in total, then this takes time
O( Ni3 ) time. Note that if a single-vertical object consisted of one vertical
and  horizontal segments, then we can trace around it with a curve with O()
segments. Hence all curves together have O(N ) segments and the total run-time
is O(N 3 ).
Let Ii be the maximum-weight independent set in Gi , and return as set I
the set in I1 , . . . , Ik that has the maximum weight. To argue the approximation-
factor, let I be the maximum-weight independent set of G, and dene Ii to
be all those elements of I that belong to Ri , for i = 1, . . . , k. Clearly Ii is an
independent set of Gi , and so w(Ii ) w(Ii ). But on the other hand maxi w(Ii )
w(I )/k since we split I into k sets. Therefore w(I) = maxi w(Ii ) w(I )/k,
and so w(I) is within a factor of k 2 log n of the optimum.

We note here that the best polynomial algorithm for independent set in
general string graphs achieves an approximation factor of O(n ), under the as-
sumption that any two strings cross each other at most a constant number of
times [3]. This algorithm only works for unweighted independent set; we are not
aware of any approximation results for weighted independent set in arbitrary
string graphs.
Because B2 -VPG-graphs can be vertex-split into two single-vertical B2 -VPG-
representations, and the total number of segments used is O(n), we also get:
Corollary 14. There exists a (4 log n)-approximation algorithm for weighted in-
dependent set on B2 -VPG-graphs with run-time O(n3 ).
Another hereditary problem is colouring: Find the minimum number k such
that we can assign numbers in {1, . . . , k} to vertices such that no two adjacent
vertices receive the same number. Fox and Pach [3] pointed out that if we have
a c-approximation algorithm for Independent Set, then we can use it to ob-
tain an O(c log n)-approximation algorithm for colouring. Therefore our result
also immediately implies an O(log2 n)-approximation algorithm for colouring in
single-vertical graphs and B2 -VPG-graphs.
Another hereditary problem is weighted clique: Find the maximum-weight
subset of vertices such that any two of them are adjacent. (This is indepen-
dent set in the complement graph.) Clique is NP-hard in outer-string graphs
even in its unweighted version [2]. For this reason, we use the split into co-
comparability graphs instead; weighted clique can be solved in quadratic time in
co-comparability graphs (because weighted independent set is linear-time solv-
able in comparability graphs [4]). Weighted clique is also linear-time solvable on
permutation graphs [4]. We therefore have:
168 T. Biedl and M. Derka

Theorem 15. There exists an (8 log3 n)-approximation algorithm for weighted

clique on B2 -VPG-graphs with run-time O(n2 ). The run-time becomes O(n) if
the graph is a 1-string B2 -VPG graph, and the approximation factor becomes
4 log2 n if the graph is a B1 -VPG-graph.
In a similar manner, we can get poly-time (8 log3 n)-approximation algo-
rithms for any hereditary problem that is solvable on co-comparability graphs.
This includes clique cover, maximum k-colourable subgraph, and maximum h-
coverable subgraph. See [7] for the denition of these problems, and the argument
that they are hereditary.

5 Conclusions
We presented a technique for decomposing single-vertical graphs into outer-string
subgraphs, B2 -VPG-graphs into co-comparability graphs, and 1-string B2 -VPG-
graphs into permutation graphs. We then used these results to obtain approxi-
mation algorithms for hereditary problems, such as weighted independent set.
As for open problems, we are very interested in approximation algorithms
for Bk -VPG graphs, where k is a constant. Also, if curves are not required to
be orthogonal, but have few bends, are there approximation algorithms better
than those for arbitrary string graphs?

References
1. Jean Cardinal, Stefan Felsner, Tillmann Miltzow, Casey Tompkins, and Birgit
Vogtenhuber. Intersection graphs of rays and grounded segments. Technical Report
1612.03638 [cs.DM], ArXiV, 2016.
2. Sergio Cabello, Jean Cardinal and Stefan Langerman. The Clique Problem in Ray
Intersection Graphs. Discrete & Computational Geometry 50(3), 771783, 2013.
3. Jacob Fox and Janos Pach. Computing the independence number of intersection
graphs. In Dana Randall, editor, Proceedings of the Twenty-Second Annual ACM-
SIAM Symposium on Discrete Algorithms, SODA 2011, San Francisco, California,
USA, January 23-25, 2011, pages 11611165. SIAM, 2011.
4. M. C. Golumbic. Algorithmic graph theory and perfect graphs. Academic Press, New
York, 1st edition, 1980.
5. Martin Charles Golumbic, Doron Rotem, and Jorge Urrutia. Comparability graphs
and intersection graphs. Discrete Mathematics, 43(1):3746, 1983.
6. J. Mark Keil, Joseph S. B. Mitchell, Dinabandhu Pradhan, and Martin Vatshelle. An
algorithm for the maximum weight independent set problem on outerstring graphs.
Comput. Geom., 60:1925, 2017.
7. J. Mark Keil and Lorna Stewart. Approximating the minimum clique cover and
other hard problems in subtree lament graphs. Discrete Applied Mathematics,
154(14):19831995, 2006.
8. Abhiruk Lahiri, Joydeep Mukherjee, and C. R. Subramanian. Maximum indepen-
dent set on B1 -VPG graphs. In Zaixin Lu, Donghyun Kim, Weili Wu, Wei Li,
and Ding-Zhu Du, editors, Combinatorial Optimization and Applications (COCOA
2015), volume 9486 of Lecture Notes in Computer Science, pages 633646. Springer,
2015.
 A Deterministic Algorithm
for Online Steiner Tree Leasing

Marcin Bienkowski1 , Artur Kraska1 , and Pawel Schmidt1

Institute of Computer Science, University of Wroclaw, Poland

Abstract. We study the Online Steiner Tree Leasing (OSTL) problem,

dened in a weighted undirected graph with a distinguished root node r.
There is a known set L of available lease types, where each type  L
is characterized by its duration D and cost factor C . As an input,
an online algorithm is given a sequence of terminals and has to connect
them to the root r using leased edges. An edge of length d can be leased
using lease type  for cost C d and remains valid for time D .
The OSTL problem contains the online Steiner tree and the single-source
rent-or-buy problems as specic subcases. We present the rst determin-
istic online algorithm for OSTL, whose competitive ratio is O(|L| log k),
where k is the number of dierent terminals in the input. The currently
best randomized algorithm attains the ratio of O(log |L| log n), where
n k is the number of nodes in the graph.

Keywords: Steiner tree Leasing Competitive analysis Online al-

gorithms

1 Introduction

The traditional network design focuses on graph optimization problems, in which

an algorithm purchases bandwidth on links to maintain certain graph proper-
ties, such as connectivity or throughput. A standard feature of most considered
models is the permanence of bandwidth allocations. For example, in the Steiner
tree problem [23], the goal is to buy a subset of edges connecting a given set
of terminals to the chosen root node r. Even the online avor of this prob-
lem [4,5,16,18,22] has this feature: the terminals arrive in online manner, and
an algorithm irrevocably buys additional links, so that the terminals seen so far
are connected to the root r. In this setting, each purchase is everlasting, i.e., the
problem should be rather termed incremental Steiner tree.

Rent-or-Buy Variants. A well-studied modication of the online Steiner tree

scenario is to relax the need of upfront commitment and additionally allow an al-
gorithm to rent edges (at a fraction of the edge purchase price). Each terminal
must be connected to the root r using either type of edges, but rented edges are

Partially supported by the Polish National Science Centre grant
2016/22/E/ST6/00499.

Springer International Publishing AG 2017 169

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 169180, 2017.
DOI: 10.1007/978-3-319-62127-2_15
170 M. Bienkowski et al.

valid only for a single terminal and cannot be reused by subsequent ones. This
variant is called online single-source rent-or-buy [3,10,11,12,15,17,21] and is also
equivalent to the le replication problem.
Note that the rent-or-buy variant is still incremental: bought edges persist in
the graph till the end and if a request to connect a specic terminal appears
suciently many times in the input any reasonable algorithm nally buys
a path connecting this terminal to the root r.

Leasing Variants. Many markets give an option of temporary leasing of re-

sources. In particular, the advent of digital services in cloud computing changed
the business model from buying physical servers to leasing virtual ones. This al-
lowed companies to adapt quickly to varying requirements of their customers [7].
Furthermore, the software-dened networking enabled similar mechanisms on the
network level, allowing companies to lease network links on the y [13]. Typ-
ically, possible leases have dierent lengths and costs, and obey economies of
scale, e.g., leasing a link for a week is more expensive than leasing it for a day,
but not more than seven times. One can view rent-or-buy variants as an extreme
case of leasing, where only two leases are available: a lifetime one (buying) and
a lease of innitesimal duration (renting).
These trends motivate the study of algorithmic leasing variants of popular
network design mechanisms [1,2,6,19,20]. Note that from the algorithmic stand-
point, leasing variants have a truly online nature: leases have nite duration and
expire after some time. An online algorithm has to adapt itself to varying access
patterns, e.g., by acquiring longer leases in response to increased demand.

1.1 The Model

In this paper, we study the Online Steiner Tree Leasing (OSTL) problem in-
troduced by Meyerson [19]. The problem is dened in a weighted undirected
graph G with a distinguished root node r V (G). For each pair of nodes u
and v, by dG (u, v) we denote the length (with respect to edge weights) of the
shortest path between u and v. There is a known set L of available lease types,
where each type  L is characterized by its duration D and cost ratio C . We
denote the number of leases by L = |L|.
An input to the problem consists of a sequence of requests, each being
a terminal (a node of G), arriving sequentially in an online manner. We treat
the root r also as a terminal. We assume that each request arrives at a dierent
time (arrival times are real non-negative numbers). In response to a requested
terminal t , which appears at time t, an online algorithm has to connect t to r
using leased edges in G, leasing additional ones if necessary. (We note that a
terminal may occur multiple times in sequence .) If an algorithm acquires a
lease type  L of an edge e = (u, v), it pays C dG (u, v). The edge leased at
time t remains available for the period [t, t + D ); afterwards the lease expires.
To recap, an input instance I is a tuple (G, dG , r, L; ), where G, dG , r, and L
are known a priori, and is presented in an online fashion to an algorithm. For
any algorithm A, A(I) is the cost of A on input I and is subject to minimization.
A Deterministic Algorithm for Online Steiner Tree Leasing 171

The goal is to minimize the competitive ratio, dened as a worst-case ratio of the
cost of an online algorithm to the cost of the optimal oine solution (denoted
Opt) on the same input.

1.2 Previous Results

The OSTL problem on a single edge is known as the parking permit problem
for which optimally competitive algorithms were given by Meyerson [19]: a de-
terministic O(L)-competitive one and a randomized O(log L)-competitive one.
(Note that the rent-or-buy variant on a single edge is equivalent to the classic
ski rental problem with a trivially achievable constant competitive ratio.)
The randomized algorithm can be extended to trees [19]. As any n-node graph
can be approximated by a random tree with expected distortion of O(log n) [14],
this approach yields a randomized O(log Llog n)-competitive solution for graphs.
Better or non-randomized algorithms were known only for specic variants of
the OSTL. In particular, for a rent-or-buy variant (recall that it corresponds to
a special 2-lease variant, where the cheaper lease suces only for serving a single
request, and the more expensive lease lasts forever) Awerbuch, Azar and Bartal
gave a randomized O(log k)-competitive algorithm and O(log2 k)-deterministic
one [9]. The latter result was improved to O(log k) only recently by Umboh [21].
In these results, k denotes the number of dierent terminals in an input.

1.3 Our Contribution

In this paper, we present the rst deterministic online algorithm for the OSTL
problem. Our algorithm is O(L log k)-competitive. It outperforms the random-
ized O(log L log n)-competitive solution by Meyerson [19] when k (the number
of dierent terminals in the input) is small.
While the result might not be optimal, neither O(L) nor O(log k) can be
beaten by a deterministic solution: (L) bound follows by the lower bound on
the parking permit problem (which is equivalent to the OSTL on a single edge)
and (log k) bound follows by the online Steiner tree problem (which is a specic
case of the OSTL with a single lease of the innite duration).
In our solution (presented in Sect. 4), a path that connects a requested termi-
nal to an already existing Steiner tree is chosen greedily. However, we still have
to decide which lease type to use for such path. To this end, we check how many
requests were recently served in a neighborhood of the currently requested
terminal; once certain thresholds are met, more expensive leases are acquired.
While such approach is natural, the main diculty stems from the dynamics
of the leased edges. Namely, while in the rent-or-buy scenario the Steiner tree
maintained by an algorithm may only grow, in the leasing variant it may also
shrink as edge leases expire. As a result, the already aggregated serving cost may
cease to be sucient to cover a more expensive lease for the new connection.
Coping with this issue is the main challenge we tackle in this paper.
We use a recent analysis technique by Umboh [21]: the online algorithm is
run on a graph G, but its cost is compared to the cost of Opt run on a tree T .
172 M. Bienkowski et al.

This tree T is a hierarchically separated tree (HST), whose leaves are requested
terminals and whose distances dominate graph distances. By showing that our
algorithm is O(L)-competitive against Opt on T , for any choice of T , we obtain
that it is O(L log k)-competitive against Opt on the original graph G. The
details of this reduction are presented in Sect. 2.
We emphasize that the competitive ratio of our algorithm is a function of the
number of dierent terminals, k, and not the number of nodes in the graph, n,
as it is the case for the randomized algorithm of [19]. In fact, our algorithm and
its analysis work without changes also in any (innite) metric space, e.g., on the
Euclidean plane; in the paper, we use the graph terminology for simplicity.

1.4 Related Work

Other network design problems were also studied in leasing context. In par-
ticular, a randomized O(log L log n)-competitive algorithm was given for the
Online Steiner Forest Leasing by Meyerson [19]. Deterministic algorithms for
this problem are known only for the rent-or-buy subcase, for which an opti-
mal competitive ratio of O(log k) was achieved by Umboh [21]. Other problems
include the facility location [1,20] and the set cover [2].
The leasing setting was also applied to oine scenarios of the problems above
by Anthony and Gupta [6], who showed an interesting reduction between leasing
and stochastic optimization variants.

2 HST Embeddings

In this section, we show how to use hierarchically separated trees (HSTs) for the
analysis of an algorithm for the OSTL problem. Unlike many online construc-
tions for network design problems (see, e.g., [8,19]), here HSTs are not used for
an algorithm construction. Moreover, in our analysis, an HST will approximate
not the whole graph, but only the subgraph spanned by terminals.

Denition 1 (Dominating HST embedding of terminals). Fix any input

instance I = (G, dG , r, L; ) of the OSTL problem. Let X V (G) be the set
of terminals requested in (including the root r). Assume that the minimum
distance between any pair of nodes from X is at least 1.1 A dominating HST
embedding of terminals of I is a rooted tree T with pairwise distances given by
metric dT , satisfying the following properties.

1. The leaves of T are exactly the nodes of X and they are on the same level.
2. The distance from any leaf of T to its parent is 1.
3. The edge lengths increase by a factor of 2 on any leaf-to-root path.
4. dT dominates dG , i.e., dT (u, v) dG (u, v) for any pair of nodes u, v X.
1
For analysis, we may always scale the instance, so that this property holds.
A Deterministic Algorithm for Online Steiner Tree Leasing 173

Fix now any instance I = (G, dG , r, L; ) of the OSTL and let (T, dT ) be any
dominating HST embedding of terminals of I. Let IT = (T, dT , r, L; ) be the
instance I, where graph G was replaced by tree T with distances given by dT .
While estimating Opt(I) directly may be quite involved, lower-bounding
Opt(IT ) is much easier. In particular, for each request t there is a unique path
in T connecting t with r. As dT dominates dG , it is also feasible to compare
the cost of an online algorithm on I to Opt(IT ). Finally, it is possible to relate
Opt(IT ) to Opt(I) as stated in the following lemma, due to Umboh [21].

Lemma 2. Let I = (G, dG , r, L; ) be an instance of the OSTL problem and

let X be the set of terminals of I. There exists a dominating HST embedding
(T , dT ) of terminals X, such that Opt(IT ) O(log |X|) Opt(I), where
IT = (T , dT , r, L; ).

Proof. Fix any dominating HST embedding (T, dT ) of terminals X. The solution
Opt(I) is a schedule that leases particular edges of G at particular times. Let
Off(IT ) be an oine solution that, for any leased edge e = (u, v) in Opt(I),
leases all edges on the unique path in T from u to v, using the same lease
type. While it is not necessary for the proof, it is worth observing that, by the
domination property (cf. Denition 1), Opt(I) Off(IT ).
By the FRT approximation [14], there exists a probability distribution D
over dominating HST embeddings (T, dT ) of X, such that ET D [dT (u, v)]
O(log |X|) dG (u, v) for all u, v X. This relation summed over all edges (u, v)
used in the solution of Opt(I) yields that

ET D [Off(IT )] O(log |X|) Opt(I) .

By the average argument, there exists a dominating HST embedding (T , dT ),

such that Off(IT ) O(log |X|) Opt(I), and the proof follows by observing
that Opt(IT ) is at most Off(IT ).


The lemma can be generalized to any network design problem whose objective
function is a linear combination of edge lengths. In Sect. 4, we will construct an
algorithm for the OSTL which is O(L)-competitive against the cost of Opt on
any HST embedding. By Lemma 2, this algorithm is O(L log k)-competitive.

3 Interval Model
In this section, we make several assumptions on the available leases. At the
expense of a constant increase of the competitive ratio, they will make the con-
struction of our algorithm easier. Similar assumptions were also made for the
parking permit problem [19].

Denition 3. In the interval model, the following conditions hold for the input
instance.
Costs factors and durations of all leases are powers of two.
174 M. Bienkowski et al.

Lease types are sorted both by their costs and durations, i.e., if  < , then
D < D and C < C .
Fix any lease type  and let Jm = [m D , (m + 1) D ) for any m N.
For any time t and any edge, there is a unique lease of type  that can be
acquired by an algorithm: it is the lease for period Jm containing t.
The last property of the interval model means that, unlike the standard
leasing model outlined in Sect. 1.1, if an algorithm leases an edge at a time t
using a lease type  L, such transaction may occur within the lease duration.
Hence, the acquired lease may expire earlier than at time t + D . We also dene
J [t] to be the period Jm containing time t.
Observation 4. In the interval model, when lease of type  expires, all leases
of smaller types expire as well.
Lemma 5. Any (online or oine) algorithm for the original leasing model can
be transformed into an algorithm for the interval model (and back) without chang-
ing its cost by more than a constant factor.
The lemma above follows by standard rounding arguments (its proof is omit-
ted; see [19] for a similar argument). Hence, if an algorithm is R-competitive for
the interval model, it is O(R)-competitive for the original leasing model. There-
fore, we will assume the interval model in the remaining part of the paper.

4 Algorithm Construction
We present our algorithm Accumulate-and-Lease-Greedily (Alg). For sim-
plicity of the description, we assume that a given graph G is complete (with the
metric given by dG ). Such assumption is without loss of generality, as leasing
the edge (u, v) can be always replaced by leasing a shortest path connecting u
and v.
We will say that an edge e is -leased at time t, if an algorithm leased e for
period J [t] using lease type . Additionally, a request t is -leased if at time t
an algorithm -leases an edge e = (t , u) for some u.
By F [t] and Fm we denote the set of all requests that arrived during J [t]
and Jm , respectively. Furthermore, T [t] denotes the set of requests that are
connected, at time t, to the root r using edges of lease types at least .

High-level idea. In the execution of Alg, at any time t, the set of all currently
leased edges will be a single (possibly empty) tree, called the Steiner tree of Alg.
Furthermore, on any path from the root r that consists of leased edges, the closer
we are to the root, the longer leases we have. In eect, T [t] always forms a tree.
Moreover, when leases expire, the set of leased edges shrinks, but it remains
connected.
When a request t arrives, we check whether we can aord a lease  for t ,
starting from the longest (and the most expensive) available lease: We compute
the distance d from t to T [t]. Then, we check if there were suciently many
requests served recently in a small (compared to d) neighborhood of t . If
so, then we connect t to T [t] using lease type .
A Deterministic Algorithm for Online Steiner Tree Leasing 175

Algorithm 1 Accumulate-and-Lease-Greedily for the OSTL problem

1: while request t arrives do
2: for  L . . . 1 do
3: if t is not connected to the root r with a path of leased edges then
4: /* Check whether we can aord lease  for t */
5: let x be the node of T [t] closest to t
6: j log dG (t , x ) 
7: Nt {f F [t] : dG (t , f ) 2j2 and class(f ) = j}
8: if |Nt | C1 C or  = 1 then
9: -lease the edge (t , x )
10: class(t ) j

Algorithm description. More precisely, x any time t when a request t is pre-

sented to Alg. Alg checks, for each lease type  starting from the most expensive
(the L-th one), what the cost of connecting t to the tree T [t] would be. That
is, among all the nodes of T [t], it nds the node x closest to t . If we -lease
the edge (t , x ) at the cost C dG (t , x ), then t becomes connected to the
root r via a path of leased edges (of lease type at least ). We round the distance
dG (t , x ) up to the smallest power of two, denoted 2j . Then, we look at the
set Nt (cf. Line 7 in Algorithm 1) of requests that

arrived at any time in J [t],

are at distance at most 2j /4 from t ,
are of class j (i.e., upon their arrival, Alg connected them to its Steiner
tree, using an edge of length from (2j1 , 2j ], i.e., roughly dG (t , x )).

Note that the terminals of Nt are not necessarily connected to the Steiner
tree of Alg at time t. If the number of requests in Nt is at least C /C1 , then Alg
-leases the edge (t , x ) and sets the class of t to j. Otherwise, Alg proceeds
to cheaper lease types. If no lease type  satises the condition |Nt | C /C1 ,
Alg eventually 1-leases the edge (t , x1 ). Note that Alg leases exactly one edge
for each terminal that is not connected to the tree at the time of its arrival.
Pseudocode of Alg is given in Algorithm 1. We recall the property of Alg
stated earlier in its informal description.

Observation 6. For any time t and lease type  L, T [t] is a single tree.

5 Analysis

Throughout this section, we x an input instance I = (G, dG , r, L; ) and a cor-

responding tree instance IT = (T, dT , r, L; ), where (T, dT ) is a dominating
HST embedding of terminals of I (cf. Sect. 2).
Without loss of generality, we assume that when a request t arrives, it
is not yet connected to the Steiner tree of Alg (otherwise t would be ignored
by Alg). If t was -leased, then we call Nt (computed in Line 7 of Algorithm 1)
176 M. Bienkowski et al.

its neighbor set. We denote the set of all requests of class j by Wj . Additionally,
let Wj consist of all requests from Wj that were -leased.
A brief idea of the proof is as follows. Suppose each request t Wj receives
a credit of C1 2j when it arrives. If t was -leased, the actual cost paid by
Alg was C 2j . While the latter amount can be much larger for an individual
request t , in Sect. 5.1, we show that, for any xed lease type , the total cost
paid for -leased edges is bounded by the sum of all requests credits. In Sect. 5.2,
we exploit properties of dominating HST embeddings to show how all credits can
be charged (up to constant factors) to the leasing costs Opt pays for particular
edges of the tree T . Altogether, this will show that Alg(I) O(L) Opt(IT ).
Along with Lemma 2, this will bound the competitive ratio of Alg (see Sect. 5.3).

5.1 Upper Bound on ALG

The core of this section is Lemma 8, which essentially states that for any lease
type , all requests credits can cover all leases of type . Before proceeding to
its proof, we rst show the following structural property.

Lemma 7. Fix a class j, a lease type , and a pair of distinct requests s , t

Wj . Their neighbor sets, Nt and Ns , are disjoint.

Proof. Without loss of generality, s < t. We will prove the lemma by contradic-
tion. Assume there exists a request u Ns Nt .
By the denition of neighbor sets, u F [s] F [t]. In the interval model,
there are only two possibilities: either periods J [s] and J [t] are equal or they
are disjoint. As in the latter case the corresponding sets F [s] and F [t] would
be disjoint as well, it holds that J [s] = J [t].
As the leases of type  that Alg bought for t and s started and expired
at the same time, s was in the tree T [t] when t arrived. Thus, the dis-
tance between t and the tree T [t] was at most dG (t , s ). From the trian-
gle inequality and diameters of sets Ns and Nt , it follows that dG (t , s )
dG (t , u ) + dG (u , s ) 2j2 + 2j2 = 2j1 . Hence, the request t would be
of class j 1 or lower, which would contradict its choice.


Lemma 8. For any class j and a lease type  L, C |Wj | C1 |Wj |.

Proof. The lemma follows trivially for  = 1, and therefore we assume that  2.
We look at any request t Wj and its neighbor set Nt . As t is of class j,
N contains requests only of class j, i.e., Nt Wj . 
t
By Lemma 7, the neighbor
sets of all requests from Wj are disjoint, and hence t W  |Nt | |Wj |.
j
t
As Alg -leases requestt , its neighbor set N contains at least C /C1
requests. Therefore, |Wj | t W  |N | t W  C /C1 = |Wj | C /C1 . 
t

j j


Lemma 9. For any input I, it holds that Alg(I) L j |Wj | C1 2j .

Proof. The cost of serving any request t Wj is at most C 2j . Using Lemma 8,
  
we obtain Alg(I) L j |Wj | C 2j L j |Wj | C1 2j .

A Deterministic Algorithm for Online Steiner Tree Leasing 177

Fig. 1. An example of an HST embedding. Square nodes (leaves of the HST) repre-
sent terminals from an input sequence (including root r). Edge e is a 2-level edge (of
length 22 ). D(e) is the set of leaves below edge e.

5.2 Lower Bound on OPT

In this part, we bound the sum of all requests credits by O(1) Opt(IT ).
We number edge levels of T starting from bottom ones and counting from 0.
That is, a j-level edge is of length 2j . Moreover, for an edge e, we denote the set
of all leaves below it by D(e), see Fig. 1. We denote the set of all j-level edges
by Ej . The next observation follows immediately by Denition 1.
Observation 10. Fix any j-level edge e. For any two leaves u, v D(e), it holds
that dG (u, v) dT (u, v) 2j+1 .
Lemma 11. For any request t of class j 3, there exists an edge e Ej3 ,
such that t D(e) and e lies on the unique path from t to the root r in T .
Proof. When Alg -leases request t and assigns class j to it, the distance
between t and the tree T [t] is larger than 2j1 , and thus dG (t , r) > 2j1 .
By Observation 10, the unique path in T between r and t must cross two
(j 2)-level edges, and hence also two (j 3)-level edges. We pick e to be the
(j 3)-level edge that is closer to t (see Fig. 1).

The lemma above implicitly creates a mapping from the set of all requests
of class j 3 to edges in tree T : a request of class j (i.e., connected by Alg
with an edge of length at most 2j ) is mapped to a tree edge of level j 3 (of
length 2j3 ). Note that a request of class j could be mapped to an edge in Ej2 .
However, the next lemma requires that all the requests mapped to an edge e are
close to each other. We extend to include also the requests of classes j 2,
by mapping any such request t to the edge e E0 adjacent to t in tree T . In
these terms, 1 (e) is a set of requests mapped to e. For an edge e of level j 1,
3
1 (e) = D(e) Wj+3 , and for e E0 , we have 1 (e) = D(e) j=0 Wj .
Let Opt(e) be the total leasing cost of e in the optimal solution for IT .
Our goal now is to show that the sum of credits of requests in 1 (e) is at
most O(Opt(e)). To do so, we rst prove a general bound on the amount of credit
that holds for all possible periods Jm . Later on, we will apply it to periods Jm
when Opt leased edge e.
Lemma 12. Fix a lease type  > 1 and a j-level edge e of T . Then, for any
m N, |1 (e) Fm | 8 C /C1 .
178 M. Bienkowski et al.

Proof. We rst assume that j 1 and we will show that |1 (e) Fm |
C /C1 + 1 2 C /C1 .
For a contradiction, assume that 1 (e) Fm contains more than b =
C /C1 + 1 requests. Let s and t be the b-th and the (b + 1)-th of them,
respectively. By Lemma 11, all requests of 1 (e) are of class j + 3 and are
contained in D(e). By Observation 10, they are all within a distance of 2j+1
from s in the graph. Therefore, Ns , the neighbor set of s considered by Alg,
contains all previous requests from 1 (e) Fm (there are b 1 = C /C1 many
of them). The condition at Line 8 of Algorithm 1 is thus fullled, and therefore
Alg buys a lease of type at least  for s .
In eect, when t arrives, s is in T [t]. Hence, the distance from t to the
tree T [t] in the graph is at most dG (t , s ) dT (t , s ) 2j+1 . Therefore,
the class of t is at most j + 1, which contradicts the choice of t .
The analysis above can be extended to any 0-level edge e. Because D(e) for
e E0 contains exactly one terminal, all requests from 1 (e) Fm are always
contained in the appropriate neighbor set. This implies that |1 (e) Fm
3
Wi | 2 C /C1 for any class i {0, 1, 2, 3}. As 1 (e) = D(e) i=0 Wi , we
obtain |1 (e) Fm | 4 2 C /C1 .

Lemma 13. Fix a j-level edge e of T. Then, |1 (e)| C1 2j 8 Opt(e).
Proof. By Lemma 11, for each request t in 1 (e), Opt has to have edge e
leased at time t, as e lies on the only path between t and the root r (see also
Fig. 1).
Let P (e) be the set
 of all pairs (, m), such that Opt -leases e for period Jm .
That is, Opt(e) = (,m)P (e) C 2 . In the optimal solution Jm periods are
j

pairwise disjoint for all pairs (, m) in P (e), and hence so are sets Fm . Thus,

|1 (e)| C1 2j = |1 (e) Fm | C1 2j
(,m)P (e)

8 C 2j = 8 Opt(e) ,
(,m)P (e)

where the inequality follows by Lemma 12.


Lemma 14. For any input  I and any dominating HST embedding (T, dT ) of
terminals of I, it holds that j |Wj | C1 2j O(1) Opt(IT ).
Proof. Fix any level j 1. Recall that all requests of class j + 3 (and only them)
are mapped by to edges from Ej . Hence, we obtain

|Wj+3 | = eEj |1 (e)| . (1)

On the other hand, all requests of class j

 {0, 1, 2, 3} (and
 only them) are
mapped by to edges from E0 . Therefore, j3 |Wj | = eE0 |1 (e)|, and
consequently  
|Wj | 2j 8 |1 (e)| . (2)
j3 eE0
A Deterministic Algorithm for Online Steiner Tree Leasing 179

We use (1) and (2) to bound j |Wj | 2j :
   
|Wj | C1 2j 8 |1 (e)| C1 + 2j+3 |1 (e)| C1
j0 eE0 j1 eEj
 
1
= 8 2 |
j
(e)| C1
j0 eEj
 
8 8 Opt(e) = O(1) Opt(IT ) .
j0 eEj

The second inequality is a consequence of Lemma 13.



5.3 The Competitive Ratio

Theorem 15. Accumulate-and-Lease-Greedily is O(Llog k)-competitive.

Proof. Fix an instance I = (G, dG , r, L; ). By Lemma 2, there exists a domi-

nating HST embedding (T, dT ), such that Opt(IT ) O(log k) Opt(I), where
IT = (T, dT , r, L; ). By Lemma
 9, the total cost of Alg is at most L times the
sum of all requests credits, j |Wj |C1 2j . By Lemma 14, the latter amount is at
most O(1)Opt(IT ), and hence Alg(I) O(L)Opt(IT ) O(Llog k)Opt(I),
which concludes the proof.


6 Conclusions
We showed that the technique of analyzing greedy algorithms using HSTs can be
also applied to the leasing variant of the online Steiner tree (the OSTL problem).
A natural research direction is to employ it for other leasing variants of graph
problems, such as Steiner forest or facility location.
Closing the gap between the current upper and lower bounds for the deter-
ministic algorithms solving the OSTL problem (O(L log k) and (L + log k),
respectively) is an intriguing open problem. In particular, it seems that improv-
ing the competitive ratio requires a very careful interplay between path-choosing
and lease-upgrade routines. We remark that analogous gaps exist also for ran-
domized algorithms for the OSTL problem and for a leasing variant of the facility
location problem [20].

References
1. Absho, S., Kling, P., Markarian, C., Meyer auf der Heide, F., Pietrzyk, P.: Towards
the price of leasing online. Journal of Combinatorial Optimization pp. 120 (2015)
2. Absho, S., Markarian, C., Meyer auf der Heide, F.: Randomized online algorithms
for set cover leasing problems. In: Proc. 8th Int. Conf. on Combinatorial Optimiza-
tion and Applications (COCOA). pp. 2534 (2014)
3. Albers, S., Koga, H.: New on-line algorithms for the page replication problem.
Journal of Algorithms 27(1), 7596 (1998)
180 M. Bienkowski et al.

4. Alon, N., Azar, Y.: On-line Steiner trees in the Euclidean plane. In: Proc. 8th ACM
Symp. on Computational Geometry (SoCG). pp. 337343 (1992)
5. Angelopoulos, S.: On the competitiveness of the online asymmetric and Euclidean
Steiner tree problems. In: Proc. 7th Workshop on Approximation and Online Al-
gorithms (WAOA). pp. 112 (2009)
6. Anthony, B.M., Gupta, A.: Infrastructure leasing problems. In: Proc. 12th Int.
Conf. on Integer Programming and Combinatorial Optimization (IPCO). pp. 424
438 (2007)
7. Armbrust, M., Fox, A., Grith, R., Joseph, A.D., Katz, R.H., Konwinski, A.,
Lee, G., Patterson, D.A., Rabkin, A., Stoica, I., Zaharia, M.: Above the clouds: A
Berkeley view of cloud computing. Tech. Rep. UCB/EECS-2009-28, EECS Depart-
ment, University of California, Berkeley (2009), https://www2.eecs.berkeley.
edu/Pubs/TechRpts/2009/EECS-2009-28.pdf
8. Awerbuch, B., Azar, Y.: Buy-at-bulk network design. In: Proc. 38th IEEE Symp.
on Foundations of Computer Science (FOCS). pp. 542547 (1997)
9. Awerbuch, B., Azar, Y., Bartal, Y.: On-line generalized Steiner problem. Theoret-
ical Computer Science 324(23), 313324 (2004)
10. Awerbuch, B., Bartal, Y., Fiat, A.: Competitive distributed le allocation. In: Proc.
25th ACM Symp. on Theory of Computing (STOC). pp. 164173 (1993)
11. Bartal, Y., Fiat, A., Rabani, Y.: Competitive algorithms for distributed data man-
agement. Journal of Computer and System Sciences 51(3), 341358 (1995)
12. Black, D.L., Sleator, D.D.: Competitive algorithms for replication and migra-
tion problems. Tech. Rep. CMU-CS-89-201, Department of Computer Science,
Carnegie-Mellon University (1989)
13. Chowdhury, N.M.K., Boutaba, R.: A survey of network virtualization. Computer
Networks 54(5), 862876 (2010)
14. Fakcharoenphol, J., Rao, S., Talwar, K.: A tight bound on approximating arbitrary
metrics by tree metrics. Journal of Computer and System Sciences 69(3), 485497
(2004)
15. Fleischer, R., Glazek, W., Seiden, S.S.: New results for online page replication.
Theoretical Computer Science 324(23), 219251 (2004)
16. Imase, M., Waxman, B.M.: Dynamic Steiner tree problem. SIAM Journal on Dis-
crete Mathematics 4(3), 369384 (1991)
17. Lund, C., Reingold, N., Westbrook, J., Yan, D.C.K.: Competitive on-line algo-
rithms for distributed data management. SIAM Journal on Computing 28(3), 1086
1111 (1999)
18. Matsubayashi, A.: Non-greedy online Steiner trees on outerplanar graphs. In: Proc.
14th Workshop on Approximation and Online Algorithms (WAOA). pp. 129141
(2016)
19. Meyerson, A.: The parking permit problem. In: Proc. 46th IEEE Symp. on Foun-
dations of Computer Science (FOCS). pp. 274284 (2005)
20. Nagarajan, C., Williamson, D.P.: Oine and online facility leasing. Discrete Op-
timization 10(4), 361370 (2013)
21. Umboh, S.: Online network design algorithms via hierarchical decompositions. In:
Proc. 26th ACM-SIAM Symp. on Discrete Algorithms (SODA). pp. 13731387
(2015)
22. Westbrook, J., Yan, D.C.K.: The performance of greedy algorithms for the on-line
Steiner tree and related problems. Mathematical Systems Theory 28(5), 451468
(1995)
23. Wu, W., Huang, Y.: Steiner trees. In: Encyclopedia of Algorithms, pp. 21022107
(2016)
 The I/O Complexity of Strassens Matrix
Multiplication with Recomputation

Gianfranco Bilardi1 and Lorenzo De Stefani2

1
Department of Information Engineering, University of Padova,
Via Gradenigo 6B/Padova, Italy
[email protected]
2
Department of Computer Science, Brown University,
115 Waterman Street/Providence, United States of America
[email protected]

Abstract. A tight ((n/ M )log2 7 M ) lower bound is derived on the
I/O complexity of Strassens algorithm to multiply two n n matrices,
in a two-level storage hierarchy with M words of fast memory. A proof
technique is introduced, which exploits the Grigorievs ow of the matrix
multiplication function as well as some combinatorial properties of the
Strassen computational directed acyclic graph (CDAG). Applications to
parallel computation are also developed. The result generalizes a similar
bound previously obtained under the constraint of no-recomputation,
that is, that intermediate results cannot be computed more than once.

1 Introduction
Data movement is increasingly playing a major role in the performance of
computing systems, in terms of both time and energy. This technological trend [1]
is destined to continue, since the very fundamental physical limitations on
minimum device size and on maximum message speed lead to inherent costs
when moving data, whether across the levels of a hierarchical memory system
or between processing elements of a parallel system [2]. The communication
requirements of algorithms have been the target of considerable research in the
last four decades; however, obtaining signicant lower bounds based on such
requirements remains an important and challenging task.
In this paper, we focus on the I/O complexity of Strassens matrix multiplica-
tion algorithm. Matrix multiplication is a pervasive primitive utilized in many
applications. Strassen [3] showed that two n n matrices can be multiplied with
O(n ) operations, where = log2 7 2.8074, hence with asymptotically fewer
than the n3 arithmetic operations required by the straightforward implementation
of the denition of matrix multiplication. This result has motivated a number of
eorts which have lead to increasingly faster algorithms, at least asymptotically,
with the current record being at < 2.3728639 [4].

This work was supported, in part, by MIUR of Italy under project AMANDA
2012C4E3KT 004 and by the University of Padova under projects CPDA121378/12,
and CPDA152255/15.

Springer International Publishing AG 2017 181

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 181192, 2017.
DOI: 10.1007/978-3-319-62127-2_16
182 G. Bilardi and L. De Stefani

Previous and Related Work: I/O complexity has been introduced in the
seminal work by Hong and Kung [5]; it is essentially the number of data transfers
between the two levels of a memory hierarchy with a fast memory of M words and
a slow memory with an unbounded number of words. Hong and Kung presented
techniques to develop lower bounds to the I/O complexity of computations
modeled by computational directed acyclic graphs (CDAGs). The resulting lower
bounds apply to all the schedules of the given CDAG, including those with
recomputation, that is, where some vertices of the CDAG
 are evaluated multiple
3

times. Among other results, they established an n / M lower bound to
the I/O complexity of the denition-based matrix multiplication algorithm,
which matched a known upper bound [6]. The techniques of [5] have also been
extended to obtain tight communication bounds for the denition-based matrix
multiplication in some parallel settings [79] and for the special case of sparse
matrix multiplication [10]. Ballard et al. generalized the results on matrix
multiplication of Hong and Kung [5] in [11, 12] by using the approach proposed
in [8] based on the Loomis-Whitney geometric theorem [13, 14]. The same papers
present tight I/O complexity bounds for various classical linear algebra algorithms,
for problems such as LU/Cholesky/LDLT/QR factorization and eigenvalues and
singular values computation.
It is natural to wonder what is the impact of Strassens reduction of the
number of arithmetic operations on the number of data transfers. In an important
contribution, Ballard et al. [15], obtained an ((n/ M )log2 7 M ) I/O lower bound
for Strassens algorithm, using the edge expansion approach. The authors extend
their technique to a class of Strassen-like fast multiplication algorithms and to
fast recursive multiplication algorithms for rectangular matrices [16]. This result
was later generalized to a broader class of Strassen-like algorithms by Scott
et. al [17] using the path routing technique. In [18] (Chap. 4.5), De Stefani
presented an alternative technique for obtaining I/O lower bounds for a large
class of Strassen-like algorithms characterized by a recursive structure. This result
combines the concept of Grigorievs ow of a function and the dichotomy width
technique [19]; it generalizes previous results and simplies the analysis.
A parallel, communication avoiding implementation of Strassens algorithm
whose performance matches the known lower bound [15, 17], was proposed by
Ballard et al. [20]. A communication ecient algorithm for the special case of
sparse matrices based on Strassens algorithm was presented in [21].
On the impact of recomputation: The edge expansion technique of [15],
the path routing technique of [17], and the closed dichotomy width technique
of [19] all yield I/O lower bounds that apply only to computational schedules for
which no intermediate result is ever computed more than once (nr-computations).
While it is of interest to know what is the I/O complexity achievable by nr-
computations, it is also important to investigate what can be achieved with
recomputation. In fact, for some CDAGs, recomputing intermediate values reduces
the space and/or the I/O complexity of an algorithm [22]. In [23], it is shown
that some algorithms admit a portable schedule (i.e., a schedule which achieves
optimal performance across memory hierarchies with dierent access costs) only
The I/O Complexity of Strassen's Matrix Multiplication with Recomputation 183

if recomputation is allowed. Recomputation can also enhance the performance

of simulations among networks (see [24] and references therein) and plays a key
role in the design of ecient area-universal VLSI architectures with constant
slowdown [25]. A number of lower bound techniques that allow for recomputation
have been presented in the literature, including the S-partition technique [5],
the S-span technique [22], and the S-covering technique [26] which merges
and extends aspects from both [5] and [22]. However, none of these have been
previously applied to fast matrix multiplication
algorithms.
Our results: We extend the ((n/ M )log2 7 M ) I/O complexity lower bound
for Strassens algorithm to schedules with recomputation. A matching upper
bound is known, and obtained without recomputation; hence, we can conclude
that, for Strassens algorithm, recomputation does not help in reducing I/O
complexity if not, possibly, by a constant factor. Our proof technique is of
independent interest, since it exploits to a signicant extent the divide and
conquer nature exhibited by many algorithms. We follow the dominator set
approach pioneered by Hong and Kung in [5]. However, we focus the dominator
analysis only on a select set of target vertices, specically the outputs of the
sub-CDAGs of Strassens CDAG that correspond to sub-problems of a suitable
size (i.e., chosen as a function of the fast memory capacity M ). Any dominator
set of a set of target vertices can be partitioned into two subsets, one internal
and one external to the sub-CDAGs. The analysis of the internal component
can be carried out based only on the fact that the sub-CDAGs compute matrix
products, irrespective of the algorithm (in our case, Strassens) by which the
products are computed. To achieve this independence of the algorithm, we resort
on the concept of Grigorievs ow of a function [27] and on a lower bound to
such ow established by Savage [28] for matrix multiplication..
In order to obtain our general lower bound for the I/O complexity, we then
build on this result combining it with the analysis of the external component
of the dominator, which requires instead rather elaborate arguments that are
specic to Strassens CDAG. The paper is organized as follows: In the rst part
of Sect. 2, we provide the details of our model and of several theoretical notions
needed in our analysis. In the second part of Sect. 2, we analyze the relation
between the Grigorievs ow of a function and the size of the dominator sets of
subsets of output vertices of a CDAG. In Sect. 3, we present the I/O complexity
lower bound for Strassens algorithm when recomputation is allowed. Extensions
of the result to a parallel model are also discussed.

2 I/O Complexity, Dominator Sets and Grigorievs Flow

We consider algorithms which compute the product C = AB of n n matrices

A, B with entries from a ring R. We focus on algorithms whose execution, for
any given n, can be modeled as a computational directed acyclic graph (CDAG)
G = (V, E), where each vertex v V represents either an input value or the
result of a unit time operation (i.e., an intermediate result or one of the output
values), while the directed edges in E represent data dependences. A directed
184 G. Bilardi and L. De Stefani

7 5 4 1 3 2 6 7 5 4 1 3 2 6 C1,1 C1,2 C2,1 C2,2

A1,1 A1,2 A2,1 A2,2 B1,1 B1,2 B2,1 B2,2 M7 M 5 M4 M 1 M3 M 2 M 6

(a) EncA (b) EncB (c) Dec

Fig. 1: Basic building blocks of Strassens CDAG. EncA and EncB are isomorphic.

C1,1 C1,2 C2,1 C2,2 C1,1 C1,2 C2,1 C2,2

Dec n2 Dec
M7 M5 M4 M1 M3 M2 M6
H7nn H5nn H4nn H1nn H3nn H2nn H6nn

EncA EncB n2 EncA n2 EncB

A1,1 A1,2 A2,1 A2,2 B1,1 B1,2 B2,1 B2,2 A1,1 A1,2 A2,1 A2,2 B1,1 B1,2 B2,1 B2,2

(a) Strassens H 22 CDAG (b) Recursive construction of H 2n2n

Fig. 2: Black vertices represent combinations of the input values from the factor
matrices A and B which are used as input values for the sub-problems Mi ;
Grey vertices represent the output of the seven sub-problems which are used to
compute the output values of the product matrix C.

path connecting vertices u, v V is an ordered sequence of vertices for which u

and v are respectively the rst and last vertex such that there is in E a (directed)
edge pointing from each vertex in the sequence to its successor. We say that
G = (V  , E  ) is a sub-CDAG of G = (V, E) if V  V and E  E (V  V  ).
Properties of Strassens CDAG: Consider Strassens algorithm [3] when
used to compute C = AB, where A and B are n n matrices with entries
from the ring R. Let H nn denote the corresponding CDAG. For n 2, H nn
can be obtained by using a recursive construction which mirrors the recursive
structure of the algorithm. The base of the construction is the H 22 CDAG which
corresponds to the multiplication of two 2 2 matrices using Strassens algorithm
(Fig. 2a). H 2n2n can then be constructed by composing seven copies of H nn ,
each corresponding to one of the seven sub-products generated by the algorithm
(see Fig. 2b): n2 disjoint copies of CDAG EncA (resp., EncB ) are used to connect
the input vertices of H 2n2n , which correspond to the values of the input matrix
A (resp., B) to the appropriate input vertices of the seven sub-CDAGs Hinn ;
the output vertices of the sub-CDAGs Hinn (which correspond to the outputs
of the seven sub-products) are connected to the appropriate output vertices of
the entire H 2n2n CDAG using n2 copies of the decoder sub-CDAG Dec.
We will exploit the following recursive structure of Strassens CDAG:
The I/O Complexity of Strassen's Matrix Multiplication with Recomputation 185

Lemma 1. Let H nn denote the CDAG of Strassens algorithm for input matri-
ces of size n n. For 0 i log n 1, there are exactly 7i disjoint sub-CDAGs
H n/2 n/2 .
i i

We will also capitalize on the existence of vertex-disjoint paths connecting the

global input vertices of H nn to the local input vertices of the sub-CDAGs
H n/2 n/2 for 0 i log n 1, with the help of the following lemma.
i i

Lemma 2. Given an encoder CDAG, for any subset Y of its output vertices,
there exists a subset X of its input vertices, with min{|Y |, 1 + (|Y | 1) /2}
|X| |Y |, such that there exist |X| vertex-disjoint paths connecting the vertices
in X to vertices in Y .
We refer the reader to the extended on-line version of this paper [29] for a detailed
presentation of Strassens algorithm and for the proofs of Lemmas 1 and 2.
Model: We assume that sequential computations are executed on a system
with a two-level memory hierarchy consisting of a fast memory or cache of size
M , measured in words, and a slow memory of unlimited size. A memory word
can store at most one value from R. An operation can be executed only if all its
operands are in cache. Data can be moved from the slow memory to the cache by
read operations, and, in the other direction, by write operations. Read and write
operations are also called I/O operations. We assume the input data to be stored
in slow memory at the beginning of the computation. The evaluation of a CDAG
in this model can be analyzed by means of the red-blue pebble game [5]. The
number of I/O operations executed when evaluating a CDAG depends on the
computational schedule, that is, on the order in which vertices are evaluated and
on which values are kept in/discarded from cache. The I/O complexity IOG (M )
of a CDAG G is dened as the minimum number of I/O operations over all
possible computational schedules.
We also consider a parallel model where P processors, each with a local
memory of size M , are connected by a network. We assume that the input
is initially distributed among the processors, thus requiring that M P 2n2 .
Processors can exchange point-to-point messages among each other. For this
model, we derive lower bounds to the number of words that must be either sent
or received by at least one processor during the CDAG evaluation.
Grigorievs ow and dominator sets: The concept of dominator set was
originally introduced in [5]. We use the following, slightly dierent, denition:
Denition 1 (Dominator set). Given a CDAG G = (V, E), let I V denote
the set of input vertices. A set D V is a dominator set for V  V with respect
to I  I if every path from a vertex in I  to a vertex in V  contains at least a
vertex of D. When I  = I, D is simply referred as a dominator set for V  V .
The ow of a function was introduced by Grigoriev [27]. We use a revised
formulation by Savage [28]. The ow is an inherent property of a function, not of
a specic algorithm by which the function may be computed.
Denition 2 (Grigorievs ow). A function f : Rp Rq has a w (u, v)
Grigorievs ow if for all subsets X1 and Y1 , of its p input and q output variables,
186 G. Bilardi and L. De Stefani

with |X1 | u and |Y1 | v, there is a sub-function h of f obtained by making

some assignment to variables of f not in X1 and discarding output variables not
in Y1 , such that h has at least |R|w(u,v) points in the image of its domain.
A lower bound on the Grigorievs ow for the square matrix multiplication
2 2
function fnn : R2n Rn over the ring R was presented in [28] (Thm. 10.5.1).
2 2 2
Lemma 3 (Grigorievs ow of fnn : R2n Rn [28]). fnn : R2n
2
Rn has a wnn (u, v) Grigorievs ow, where:
"
2 2 #
1 2n u
wnn (u, v) v , for 0 u 2n2 , 0 v n2 . (1)
2 4n2

The ow of a function measures the amount of information that suitable

subsets of outputs encode about suitable subsets of inputs. Such information
must be encoded by any dominator of those outputs, thus implying the following
lower bound on the size of dominators.
Lemma 4. Let G = (V, E) be a CDAG computing f : Rp Rq with Grigorievs
ow wf (u, v). Let I (resp., O) denote the set of input (resp., output) vertices of
G. Any dominator set D for any subset O O with respect to any subset I  I
satises |D| wf (|I  |, |O |).
Proof. Given I  I and O O, suppose the values of the input variables in
I \ I  to be xed. Let D be a dominator set for O O with respect to I  I.
The lemma follows combining statements (i) and (ii):
(i) By Denition 2, there exists an assignment of the input variables in I  , such
 
that the output variables in O can assume |R|wf (|I |,|O |) distinct values.
(ii) Since all paths I  to O intercept D, the values of the outputs in O are
determined by the inputs in I \ I  , which are xed, and by the values of the
vertices in D; hence, the outputs in O can take at most |R||D| distinct values. 
We let Gnn denote the CDAG corresponding to the execution of an unspeci-
ed algorithm for the square matrix multiplication function.
Lemma 5. Given Gnn , let O O be a subset of its output vertices O. For
any subset D of the vertices of Gnn with |O |$ 2|D|, there exists a set I  I of
the input vertices I with cardinality |I  | 2n |O | 2|D|, such that all vertices
in I  are connected to some vertex in O by directed paths with no vertex in D.
Proof. Lemma 5 follows by applying the results in Lemmas 3 and 4 to the CDAG
Gnn . Let I  I denote the set of all input vertices of Gnn , such that all paths
connecting these vertices to the output vertices in O include at least a vertex in
D (i.e., I  is the largest subset of I with respect to whom D is a dominator set
for O ). From Lemmas 3 and 4 the following must hold:
"
2 2 #
1  2n |I  |
|D| wnn |O | . (2)
2 4n2
The I/O Complexity of Strassen's Matrix Multiplication with Recomputation 187

Let I  = I \ I  . By the denition of I  , the vertices in I  are exactly those that

are connected to vertices in O by directed paths with no vertex in D. Since
|I| = 2n2 , from (2) we have |I  |2 4n2 (|O | 2|D|).


3 Lower Bounds for Schedules with Recomputation

Without recomputation, once an input value is loaded in memory or an intermedi-

ate result is computed, it must be kept in memory (either cache or slow) until the
result of each operation which uses it has been evaluated. This is exploited by the
dichotomy width technique [19], the boundary ow technique [30], and those
yielding I/O lower bounds for Strassens algorithm [1618]. With recomputation,
intermediate results can instead be deleted from all memory and recomputed
starting from the global input values. This considerably complicates the analysis
of the I/O cost (see [11] for an extensive discussion). In this section, we present
a technique which addresses these complications. First, we obtain a lower bound
for the minimum size of the dominator
set of subset of vertices corresponding
to theoutput values of the (n/(2 M ))log2 7 Strassens sub-problems with input
size 2 M 2 M . In turn, this dominator bound yields an asymptotically tight
I/O lower bound both in the sequential and the parallel model.
For 1 M n2 /4, with M a power of four, we focus on the subset Y oflogthe7
input vertices and
the
subset Z of the output vertices of the (n/(2 M )) 2
2 M 2 M
sub-CDAGs H of H nn
. Further, we let X be the set of the global
input vertices of H nn which correspond to the entries of matrices A and B.

Lemma 6. Given H nn , let Q be a set of internal (i.e., not input) vertices of

 log2 7
its n/(2 M ) sub-CDAGs H 2 M 2 M . For any Z Z with |Z| 2|Q|
$
there exist X X and Y Y with |X| = |Y | 4 M (|Z| 2|Q|) such that,
(a) there are |X| = |Y | vertex-disjoint paths from X to Y , and (b) each vertex in
Y is connected to some vertex in Z by a directed path with no vertex in Q.

Proof. For a xed M , we proceed
by induction on n = 2 M , 4 M , . . . In the
base case, H nn = H 2 M 2 M , and
the sets Y and X coincide.

The statement
is a consequence of Lemma 5 as H 2 M 2 M is a G2 M 2 M CDAG.
Assuming now inductively that the statement holds for H nn , with n 2 M ,
we shall show it also holds for H 2n2n . Let H1nn , H2nn , . . . , H7nn denote the
seven sub-CDAGs of H 2n2n , each corresponding to one of the seven sub-products
generated by the rst recursive step of Strassens algorithm.
Let Zi , Yi and Qi respectively denote the subsets of Z, Y and Q in Hinn .
Since, from Lemma 1, the seven sub-CDAGs Hinn are mutually vertex-disjoint,
7 Z, Y1 , Y2 , . . . , Y7 partition
clearly Z1 , Z2 , . . . , Z7 partition 7 Y and Q1 , Q2 , . . . , Q7
partition Q. This implies i=1 |Zi | = |Z|, and i=1 |Qi | = |Q|. Letting i =
7
max{0, |Zi | 2|Qi |}, we have = i=1 i |Z| 2|Q|.
Applying the inductive hypothesis to each Hinn , we have that there is a
subset Yi Yi with |Yi | 4 M i such that vertices of Yi are connected to
188 G. Bilardi and L. De Stefani

vertices in Zi via paths with no vertex in Qi . In the sequel, the set Y referred
to in the statement will be identied as a suitable subset of 7i=1 Yi so that
property (b) will be automatically satised. Towards property (a), we observe
by the inductive hypothesis that vertices in Yi can be connected to a subset Ki
of the input vertices of Hinn with |Ki | = |Yi | using vertex-disjoint paths. Since
the sub-CDAGs Hinn are vertex-disjoint, so are the paths $ connecting vertices in
Yi to vertices in Ki . It remains to show that at least 4 M (|Z| 2|Q|) of these
paths can be extended to X while maintaining them vertex-disjoint.
In Strassens CDAG H 2n2n (Sect. 2), vertices in X corresponding to input
matrix A (resp., B) are connected to vertices in K1 , K2 , . . . , K7 by means of
n2 encoding sub-CDAGs EncA (resp., EncB ). None of these 2n2 encoding sub-
CDAGs share any input or output vertices. No two output vertices of the same
encoder sub-CDAG belong to the same sub-CDAG Hinn . This fact ensures that
for a single sub-CDAG Hinn it is possible to connect all the vertices in Ki to a
subset of the vertices in X via vertex-disjoint paths.
For each of the 2n2 encoder sub-CDAGs, let us consider the vector yj {0, 1}7
such that yj [i] = 1 i the corresponding i-th output vertex (respectively according
to the numbering indicated in Fig. 1a or Fig. 1b) is in Ki . Therefore, |yj |
equals the number of output vertices of the j-th encoder sub-CDAG which
are in K. From Lemma 2, for each encoder sub-CDAG there exists a subset
Xj X of the input vertices of the j-th encoder sub-CDAG for which it is
possible to connect each vertex in Xj to a distinct output vertex of the j-th
encoder sub-CDAG using vertex-disjoint paths, each constituted by a singular
edge with min{|yj |, 1 + (|yj | 1) /2} |Xj | |yj |. Therefore, the number
of vertex-disjoint paths connecting vertices in X to vertices in 7i=1 Ki is at
2n2 2n2
least j=1 min{|yj |, 1 + (|yj | 1) /2} under the constraint that j=1 yj [i] =

4 M i . Let us assume, w.l.o.g., that 1 2 . . . 7 . As previously stated, it
is possible to connect all vertices in K1 to vertices in X through vertex-disjoint
paths. Consider now all possible dispositions of the vertices in 7i=2 Ki over
the outputs of the 2n2 encoder sub-CDAGs. Recall that the output vertices
of an encoder sub-CDAG belong each to a dierent H nn sub-CDAG. From
Lemma 2, we have that for each encoder, there exists a subset Xj X of
the input vertices of the j-th encoder sub-CDAG with |Xj | min |yj |, 1 +
  
7
(|yj | 1) /2 yj [1] + i=2 yj [i] /2, for which it is possible to connect all
vertices in Xj to |Xj | distinct output vertices of the j-th encoder sub-CDAG
which are in 7i=1 Ki using |Xj |, thus using vertex-disjoint paths. As all the
Enc sub-CDAGs are vertex-disjoint, we can add their contributions so that the
number of vertex-disjoint paths connecting vertices in X to vertices in 7i=1 Ki is
1
7  7
at least |K1 | + 2 i=2 |Ki | = 4 M 1 + 12 i=2 i . Squaring this quantity
leads to:
" " ##2 " 7 #2
7 7 $
$ 1 $ $  1 $
4 M 1 + i = 16M 1 + 1 i + i ,
2 i=2 i=2
2 i=2
The I/O Complexity of Strassen's Matrix Multiplication with Recomputation 189

since, by assumption, 1 . . . 7 , we have: 1 i i for i = 2, . . . , 7. Thus:
" " ##2
$ 7
1 $ 7  $ 2
4 M 1 + i 16M i 4 M (|Z| 2|Q|) .
2 i=2 i=1
$
Thus,there are at least 4 M (|Z| 2|Q|) vertex-disjoint paths connecting vertices
in X to vertices in 7i=2 Ki as desired.

Lemma 7. For 1 M n2 /4, and for any subset Z Z in H nn with
|Z| = 4M , any dominator set D of Z satises |D| |Z|/2 = 2M .
Proof. Suppose for contradiction that D is a dominator set for Z in H nn
such that |D| 2M 1. Let D D be the subset of the vertices
of D
composed by vertices which are not internal to the sub-CDAGs H 2 M 2 M .
From Lemma $ 6, with Q = D \ D , there exist X X and Y Y with |X| =
|Y | 4 M (|Z| 2 (|D| |D |)) such that vertices in X are connected to
vertices in Y by vertex-disjoint paths. Hence, each vertex in D can be on
 
at most one of these
 
$ paths. Thus, there exists X X and Y  Y  with

|X | = |Y | = 4 M (|Z| 2 (|D| |D |)) |D | paths from X to Y with
no vertex in D . From Lemma 6, we also have that all vertices in Y , and, hence,
in Y  , are connected to some vertex in Z by a path with no vertex in D \ D .
Thus, there are at least paths connecting vertices in X  X to vertices in
Z with no vertex in D. We shall now show that the contradiction assumption
|D| 2M 1 implies > 0:
 $ 2
4 M (|Z| 2 (|D| |D |)) = 16M (|Z| 2 (|D| |D |)) ,
= 16M (|Z| 2|D|) + 32M |D |.

By |D| 2M 1, we have |Z| 2|D| > 4M 2(M 1) > 0. Furthermore, from

D D, we have 32M > 2M 1 > |D| |D |. Therefore:
 $ 2
2
( + |D |) = 4 M (|Z| 2 (|D| |D |)) > |D |2 . (3)

Again, |D| 2M 1 implies M (|Z| 2 (|D| |D |)) > 0. Hence, we can take
the square root on both sides of (3) and conclude that > 0. Therefore, for
|D| 2M 1 there are at least > 0 paths connecting a global input vertex
to a vertex in Z with no vertex in D, contradicting the assumption that D is a
dominator of Z.

Lemma 7 provides us with the tools required to obtain our main result.
Theorem 1 (Lower bound I/O complexity Strassens algorithm). The
I/O-complexity of Strassens algorithm to multiply two matrices A, B Rnn ,
on a sequential machine with cache of size M n2 , satises:
 log2 7
1 n
IOH nn (M ) M. (4)
7 M
190 G. Bilardi and L. De Stefani

On P processors, each with a local memory of size M n2 , the I/O complexity

satises:
 log2 7
1 n M
IOH nn (P, M ) . (5)
7 M P

Proof. We start by proving (4). Let n = 2a and M = 2b for some a, b N.
At least 3n2 3M I/O operations are necessary in order to read the 2n2 input
values from slow memory to the cache and to write the n2 output
values to the
slow memory. The bound in (4) is therefore veried if n 2 M .
 log2 7
For n 4 M , let Z denote the set of output vertices of the n/(2 M )

sub-CDAGs H 2 M 2 M of H nn . Let C be any computation schedule for the se-
quential execution of Strassens algorithm using a cache of size M . We partition C
into segments C1 , C2 , . . . such that during each Ci exactly 4M distinct vertices in Z
 log 7
(denoted as Zi ) are evaluated for the rst time. Since |Z| = 4M n/(2 M ) ,
 log 7
there are n/(2 M ) such segments. Below we show that the number qi of
I/O operations executed during each Ci satises qi M , from which (4) follows.
To bound qi , consider the set Di of vertices of H nn corresponding to the at
most M values stored in the cache at the beginning of Ci and to the at most qi
values loaded into the cache from the slow memory during Ci by means of a read
I/O operation. Clearly, |Di | M + qi . In order for the 4M values from Zi to be
computed during Ci there cannot be any path connecting any vertex in Zi to any
input vertex of H nn which does not have at least one vertex in Di ; that is, Di
has to be a dominator set of Zi . We recall that |Zi | = 4M and, from Lemma 7,
we have that any subset of 4M elements of Z has dominator size at least 2M ,
whence M + qi |Di | 2M , which implies qi M as stated above.
The proof for the bound for the parallel model in (5), follows a similar strategy:
At least one of the P processors being used, denoted as P , must compute at
 log 7
least |Z|/P = 4M n/(2 M ) /P values corresponding to vertices in Z. The
bound follows by applying the same argument discussed for the sequential case to
the computation executed by P (details the extended on-line version [29]). 

Ballard et al. [20] presented a version of Strassens algorithm whose I/O cost
matches the lower bound of Theorem 1 to within a constant factor. Therefore, our
bound is asymptotically tight, and the algorithm in [20] is asymptotically I/O opti-
mal. Since in this algorithm no intermediate result is recomputed, recomputation
can lead at most to a constant factor reduction of the I/O complexity.

The lower bound of Theorem 1 generalizes to ((n/ M )log2 7 M B ) in the
External Memory Model introduced by Aggarwal and Vitter [31], where B 1
values can be moved between cache and consecutive slow memory locations with
a single I/O operation.
The I/O Complexity of Strassen's Matrix Multiplication with Recomputation 191

4 Conclusion
This work has contributed to the characterization of the I/O complexity of
Strassens algorithm by establishing asymptotically tight lower bounds that
hold even when recomputation is allowed. Our technique exploits the recursive
nature of the CDAG, which makes it promising for the analysis of other recursive
algorithms, e.g., for fast rectangular matrix multiplication [32].
The relationship we have exploited between dominator size and Grigorievs ow
points at connections between I/O complexity, (pebbling) space-time tradeos [28],
and VLSI area-time tradeos [33]; these connections deserve further attention.
Some CDAGs for which non-trivial I/O complexity lower bounds are known
only in the case of no recomputations are described in [19]. These CDAGs are of
interest in the limiting technology model, dened by fundamental limitations
on device size and message speed, as they allow for speedups super-linear in the
number of processors. Whether such speedups hold even when recomputation is
allowed remains an open question, which our new technique might help answer.
While we know that recomputation may reduce the I/O complexity of some
CDAGs, we are far from a characterization of those CDAGs for which recompu-
tation is eective. This broad goal remains a challenge for any attempt toward a
general theory of the communication requirements of computations.

References
1. Patterson, C.A., Snir, M., Graham, S.L.: Getting Up to Speed:: The Future of
Supercomputing. National Academies Press (2005)
2. Bilardi, G., Preparata, F.P.: Horizons of parallel computation. Journal of Parallel
and Distributed Computing 27(2) (1995) 172182
3. Strassen, V.: Gaussian elimination is not optimal. Numerische Mathematik 13(4)
(1969) 354356
4. Le Gall, F.: Powers of tensors and fast matrix multiplication. In: Proc. ACM
ISSAC, ACM (2014) 296303
5. Hong, J., Kung, H.: I/o complexity: The red-blue pebble game. In: Proc. ACM
STOC, ACM (1981) 326333
6. Cannon, L.E.: A cellular computer to implement the Kalman lter algorithm.
Technical report, DTIC Document (1969)
7. Ballard, G., Demmel, J., Holtz, O., Lipshitz, B., Schwartz, O.: Brief announce-
ment: strong scaling of matrix multiplication algorithms and memory-independent
communication lower bounds. In: Proc. ACM SPAA, ACM (2012) 7779
8. Irony, D., Toledo, S., Tiskin, A.: Communication lower bounds for distributed-
memory matrix multiplication. Journal of Parallel and Distributed Computing
64(9) (2004) 10171026
9. Scquizzato, M., Silvestri, F.: Communication lower bounds for distributed-memory
computations. arXiv preprint arXiv:1307.1805 (2013)
10. Pagh, R., Stockel, M.: The input/output complexity of sparse matrix multiplication.
In: Proc. ESA, Springer (2014) 750761
11. Ballard, G., Demmel, J., Holtz, O., Schwartz, O.: Minimizing communication in
numerical linear algebra. SIAM Journal on Matrix Analysis and Applications 32(3)
(2011) 866901
192 G. Bilardi and L. De Stefani

12. Ballard, G., Demmel, J., Holtz, O., Schwartz, O.: Communication-optimal parallel
and sequential Cholesky decomposition. SIAM Journal on Scientic Computing
32(6) (2010) 34953523
13. Loomis, L.H., Whitney, H.: An inequality related to the isoperimetric inequality.
Bull. Amer. Math. Soc. 55(10) (10 1949) 961962
14. V. A. Zalgaller, A. B. Sossinsky, Y.D.B. The American Mathematical Monthly
96(6) (1989) 544546
15. Ballard, G., Demmel, J., Holtz, O., Schwartz, O.: Graph expansion and communi-
cation costs of fast matrix multiplication. JACM 59(6) (2012) 32
16. Ballard, G., Demmel, J., Holtz, O., Lipshitz, B., Schwartz, O.: Graph expansion
analysis for communication costs of fast rectangular matrix multiplication. In:
Design and Analysis of Algorithms. Springer (2012) 1336
17. Scott, J., Holtz, O., Schwartz, O.: Matrix multiplication I/O complexity by Path
Routing. In: Proc. ACM SPAA. (2015) 3545
18. De Stefani, L.: On space constrained computations. PhD thesis, University of
Padova (2016)
19. Bilardi, G., Preparata, F.: Processor-time trade os under bounded speed message
propagation. Lower Bounds. Theory of Computing Systems 32(5) (1999) 531559
20. Ballard, G., Demmel, J., H., O., Lipshitz, B., Schwartz, O.: Communication-optimal
parallel algorithm for Strassens matrix multiplication. In: Proc. ACM SPAA. (2012)
193204
21. Jacob, R., Stockel, M.: Fast output-sensitive matrix multiplication. In: Proc. ESA.
Springer (2015) 766778
22. Savage, J.E.: Extending the Hong-Kung model to memory hierarchies. In: Com-
puting and Combinatorics. Springer (1995) 270281
23. Bilardi, G., Peserico, E.: A characterization of temporal locality and its portability
across memory hierarchies. In: Automata, Languages and Programming. Springer
(2001) 128139
24. Koch, R.R., Leighton, F.T., Maggs, B.M., Rao, S.B., Rosenberg, A.L., Schwabe,
E.J.: Work-preserving emulations of xed-connection networks. JACM 44(1) (1997)
104147
25. Bhatt, S.N., Bilardi, G., Pucci, G.: Area-time tradeos for universal VLSI circuits.
Theoret. Comput. Sci. 408(2-3) (2008) 143150
26. Bilardi, G., Pietracaprina, A., DAlberto, P.: On the space and access complexity of
computation DAGs. In: Graph-Theoretic Concepts in Computer Science, Springer
(2000) 4758
27. Grigorev, D.Y.: Application of separability and independence notions for proving
lower bounds of circuit complexity. Zapiski Nauchnykh Seminarov POMI 60 (1976)
3848
28. Savage, J.E.: Models of Computation: Exploring the Power of Computing. 1st edn.
Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA (1997)
29. Bilardi, G., Stefani, L.D.: The i/o complexity of strassens matrix multiplication
with recomputation. arXiv preprint arXiv:1605.02224 (2016)
30. Ranjan, D., Savage, J.E., Zubair, M.: Upper and lower I/O bounds for pebbling
r-pyramids. Journal of Discrete Algorithms 14 (2012) 212
31. Aggarwal, A., Vitter, Jerey, S.: The input/output complexity of sorting and
related problems. Commun. ACM 31(9) (September 1988) 11161127
32. Le Gall, F.: Faster algorithms for rectangular matrix multiplication. In: Proc. IEEE
FOCS, IEEE (2012) 514523
33. Thompson, C.: Area-time complexity for VLSI. In: Proc. ACM STOC, ACM (1979)
8188
 Maximum Plane Trees in Multipartite
Geometric Graphs

Ahmad Biniaz1 , Prosenjit Bose1 , Kimberly Crosbie1 , Jean-Lou De Carufel2 ,

David Eppstein3 , Anil Maheshwari1 , and Michiel Smid1

Abstract. A geometric graph is a graph whose vertices are points in the

plane and whose edges are straight-line segments between the points. A
plane spanning tree in a geometric graph is a spanning tree that is non-
crossing. Let R and B be two disjoint sets of points in the plane where
the points of R are colored red and the points of B are colored blue, and
let n = |R B|. A bichromatic plane spanning tree is a plane spanning
tree in the complete bipartite geometric graph with bipartition (R, B).
In this paper we consider the maximum bichromatic plane spanning tree
problem, which is the problem of computing a bichromatic plane span-
ning tree of maximum total edge length.
1. For the maximum bichromatic plane spanning tree problem, we
present an approximation algorithm with ratio 1/4 that runs in
O(n log n) time.
2. We also consider the multicolored version of this problem where the
input points are colored with k > 2 colors. We present an approxi-
mation algorithm that computes a plane spanning tree in a complete
k-partite geometric graph, and whose ratio is 1/6 if k = 3, and 1/8
if k  4.
3. We also revisit the special case of the problem where k = n, i.e., the
problem of computing a maximum plane spanning tree in a complete
geometric graph. For this problem, we present an approximation
algorithm with ratio 0.503; this is an extension of the algorithm
presented by Dumitrescu and Toth (2010) whose ratio is 0.502.

1 Introduction
Let P be a set of n points in the plane in general position, i.e., no three points
are collinear. Let K(P ) be the complete geometric graph with vertex set P . It
is well known that the standard minimum spanning tree (MinST) problem in
K(P ) can be solved in (n log n) time. Also, any minimum spanning tree in
K(P ) is plane, i.e., its edges do not cross each other. The maximum spanning
tree (MaxST) problem is the problem of computing a spanning tree in K(P )
whose total edge length is maximum. Monma et al. [5] showed that this problem
1
Carleton University, Ottawa. Supported by NSERC. [email protected], kim-
[email protected], {jit, anil, michiel}@scs.carleton.ca
2
University of Ottawa, Ottawa. Supported by NSERC. [email protected]
3
University of California, Irvine. Supported by NSF grant CCF-1228639. epp-
[email protected]

Springer International Publishing AG 2017 193

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 193204, 2017.
DOI: 10.1007/978-3-319-62127-2_17
194 A. Biniaz et al.

can be solved in (n log n) time. However, a MaxST is not necessarily plane.

Alon et al. [1] started the maximum plane spanning tree (MaxPST) problem,
which is the problem of computing a plane spanning tree in K(P ) whose total
edge length is maximum. It is not known whether or not this problem is NP-hard.
They presented an approximation algorithm with ratio 0.5 for this problem. This
approximation ratio was improved to 0.502 by Dumitrescu and Toth [4].
Let R and B be two disjoint sets of points in the plane such that R B is in
general position, and let n = |RB|. Suppose that the points of R are colored red
and the points of B are colored blue. Let K(R, B) be the complete bipartite ge-
ometric graph with bipartition (R, B). The minimum bichromatic spanning tree
(MinBST) problem is to compute a minimum spanning tree in K(R, B). The
maximum bichromatic spanning tree (MaxBST) problem is to compute a span-
ning tree in K(R, B) whose total edge length is maximum. Recently, Biniaz et
al. [2] showed that both the MinBST and the MaxBST problems can be solved
in (n log n) time. We note that none of MinBST and MaxBST is necessarily
plane; they might have crossing edges. Borgelt et al. [3] studied the problem of
computing a minimum bichromatic plane spanning tree, which we refer to as
the MinBPST problem. They showed that this problem is NP-hard, and also
presenteda polynomial-time approximation algorithm with approximation ra-
tio of O( n). In this paper we study the problem of computing a maximum
bichromatic plane spanning tree, which we refer to as the MaxBPST problem.
See Figure 1.

MaxBST MaxBPST Max-4-ST Max-4-PST

Fig. 1. Maximum spanning trees.

A natural extension of the MinBST and the MaxBST problems is to have

more than two colors. In this multicolored version, the input points are colored
by k > 2 colors, and we are looking for a minimum (resp. maximum) spanning
tree in which the two endpoints of every edge have distinct colors. In other words,
we look for a minimum (resp. maximum) spanning tree in a complete k-partite
geometric graph. We refer to these problems as the Min-k-ST and the Max-k-ST
problems, respectively. Biniaz et al. [2] showed that both these problems can be
solved in O(n log n log k) time. Notice that the MinST and the MaxST problems
are special cases of the Min-k-ST and the Max-k-ST problems in which every
input point has a unique color, i.e., k = n. In this paper we also study the
problem of computing a plane Max-k-ST, which we refer to as the Max-k-PST
problem. See Figure 1.
Maximum Plane Trees in Multipartite Geometric Graphs 195

1.1 Our contributions

In this paper we study the maximum plane spanning tree problems. In Section 3
we present an approximation algorithm with ratio 1/4 for the MaxBPST prob-
lem. We study the Max-k-PST problem in Section 4. For this problem, we present
an approximation algorithm whose ratio is 1/6 if k = 3, and 1/8 if k  4. In
Section 5 we consider the MaxPST problem, where we modify the algorithm pre-
sented by Dumitrescu and Toth [4] for this problem; this modication improves
the approximation ratio to 0.503. All the presented approximation algorithms
run in O(n log n) time, where n is the number of input points.

2 Preliminaries
For any two points p and q in the plane, we refer to the line segment between
p and q as pq, and to the Euclidean distance between p and q as |pq|. The lune
between p and q, which we denote by lune(p, q), is the intersection of the two
disks of radius |pq| that are centered at p and q.
For a point set P , the diameter of P is the maximum Euclidean distance
between any two points of P . A pair of points that realizes the diameter of P is
referred to as a diametral pair.
Let G be a geometric graph with colored vertices. We denote by L(G) the
total Euclidean length of the edges of G. A star is a tree with one internal node,
which we refer to as the center of the star. For a color c, a c-star in G is a star
whose center is colored c and the colors of its leaves are dierent from c.

3 The MaxBPST problem

In this section we consider the MaxBPST problem. Recall that in this problem we
are given two sets R and B of red and blue points in the plane, respectively, and
we are looking for a maximum plane spanning tree in K(R, B). Let n = |R B|.
We present an approximation algorithm with ratio 1/4 for this problem that
runs in O(n log n) time.
We will show that the length of the longest star in K(R, B) is at least 1/4
times the length of an optimal MaxBPST. In fact, we present an algorithm that
returns such a star. Moreover, we show that this estimate is the best possible for
the length of a longest star. The longest star can easily be augmented to form
a plane spanning tree as follows. The longest star has exactly one point of one
color as its center, and all points of other color as its leaves. The edges of this
star can be extended to partition the plane into convex cones, possibly except
one cone; we split this cone into two convex cones by adding its bisector. Then,
we connect all the remaining points in each cone to one of the leaves that is on
the boundary of that cone.
If |R| = 1 or |B| = 1, then the problem is trivial. Assume |R|  2 and |B|  2.
Our algorithm rst computes a diametral pair (a, b) in R and a diametral pair
(p, q) in B. Then, returns the longest star Sx in K(R, B) that is centered at
196 A. Biniaz et al.

a point x {a, b, p, q}. Since diametral pairs of R and B can be computed in

O(n log n) time, the running time of the algorithm follows. In the rest of this
section we will show that the longest of these stars satises the approximation
ratio.
Let T be an optimal MaxBPST and let L denote the length of T . We
make an arbitrary point the root of T and partition the edges of T into two

sets as follows. Let ER be the set of edges (u, v) in T where u is a red point

and u is the parent of v. Let EB be the set of edges (u, v) where u is a blue

point and u is the parent of v. The edges of ER (resp. EB ) form a forest in
which each component is a red-star (resp. blue-star). Let LR and LB denote the


total lengths of the edges of ER and EB , respectively. Without loss of generality

assume that LR  LB . Then,
L = LB + LR  2LB . (1)
We will show that, in this case, the longest of Sp and Sq is a desired tree.
To that end, let FB be the set of edges that is obtained by connecting every
red point to its farthest blue point. Notice that the edges of FB form a forest in
which every component is a blue-star. Moreover, observe that
LB  L(FB ). (2)
Lemma 1. L(FB )  2 (L(Sp ) + L(Sq )).
3

Proof. Let Cp and Cq be the two disks of radius |pq| that are centered at p and
q, respectively. Since (p, q) is a diameter of B, all blue points lie in lune(p, q);
see Figure 2(a).

r r
q
p q p q p q
1
f (r) p x

f (r) 1

f (r)
3
Cp Cq
s s
(a) (b) (c)

Fig. 2. Illustration of Lemmas 1 and 2: r is a red point, and p, q, f (r) are blue points.

For any red point r R, let f (r) denote its neighbor in FB . Recall that f (r)
is the farthest blue point to r, and note that f (r) is in lune(p, q). See Figure 2(a).
Maximum Plane Trees in Multipartite Geometric Graphs 197

We are going to show that |rf (r)|  23 (|rp| + |rq|). Depending on whether or
not r lune(p, q) we consider the following two cases.

r / lune(p, q). Thus, we have that r / Cp or r / Cq . Without loss of

generality assume r / Cp ; see Figure 2(a). By the triangle inequality we have
|rf (r)|  |rq|+|qf (r)|. Since pq is a diameter of B, we have |qf (r)|  |pq|. In
addition, since r / Cp , we have |pq|  |rp|. By combining these inequalities,
we get
|rf (r)|  |rq| + |qf (r)|  |rq| + |pq|  |rq| + |rp|.

r lune(p, q). Without loss of generality assume that pq has unit length, pq
is horizontal, r is above pq, and r is closer to q than to p. If f (r) is on or
above pq, then |rf (r)| is smaller than |pq|, and hence smaller than |rp| + |rq|.
Assume f (r) is below pq. Let s be the intersection point of the boundaries
of Cp and Cq that is below pq as in Figure 2(b).

Claim 1. |rf (r)|  |rs|.

Proof of Claim 1. Let f  (r) be the intersection point of the ray that is
emanating from r and passing through f (r) with the boundary of lune(p, q).
Note that |rf (r)|  |rf  (r)|. If f  (r) is on the boundary of Cq , then the
perpendicular bisector of the segment sf  (r) passes through q. In this case
r is in the same side of this perpendicular as f  (r), and thus, |rf  (r)|  |rs|;
see Figure 2(b). Similarly, if f  (r) is on the boundary of Cp , then both r and
f  (r) are on a same side of the perpendicular bisector of sf  (r) which passes
through p. This proves the claim.

Extend the line segment sp from the endpoint p. Let p be the point on the
extended line that is closest to r. Dene q  similarly. Note that |rp |  |rp|
and |rq  |  |rq|. Based on this and Claim 1, in order to show that |rf (r)| 
2 (|rp| + |rq|), it suces to show that |rs|  2 (|rp | + |rq  |). Let = rsq,
3 3
and note that 0   6 ; see Figure 2(c). Since the triangles rsp and
rsq  are right-angled and spq is equilateral, we have |rq  | = |rs| sin
and |rp | = |rs| sin(/3 ). Thus,

3
|rq  | + |rp | = |rs| (sin + sin(/3 ))  2
|rs|,

where the inequality is valid because sin + sin(/3 ) is at least 3/2
for all 0   6 . This implies that |rs|  23 (|rp | + |rq  |).

Since in both previous cases |rf (r)|  2 (|rp| + |rq|), we have

3
  2 2
L(FB ) = |rf (r)|  (|rp| + |rq|) = (L(Sp ) + L(Sq )).
3 3
rR rR


198 A. Biniaz et al.

Combining Inequalities (1), (2), and Lemma 1, we get

4
L  2LB  2L(FB )  (L(Sp ) + L(Sq )).
3

Therefore, the length of the longest of Sp and Sq is at least 83 0.215 times

L . In the following lemma we improve the bound of Lemma 1 by proving that
L(FB )  L(Sp ) + L(Sq ); this improves the approximation ratio to 1/4. However,
the proof of this lemma is algebraic.
Lemma 2. L(FB )  L(Sp ) + L(Sq ).
Proof. For any red point r R, let f (r) denote its neighbor in FB . In order to
prove this lemma, as we have seen in the proof of Lemma 1, it suces to show
that |rf (r)|  |rp| + |rq|. Dene lune(p, q) as in the proof of Lemma 1. As we
have seen there, if r
/ lune(p, q), then |rf (r)|  |rp|+|rq|. Assume r lune(p, q).
Without loss of generality assume that pq has unit length, pq is horizontal, r is
above pq, and r is closer to q than to p; see Figure 2(c). Dene the point s as in
the proof of Lemma 1. By Claim 1 in the proof of Lemma 1, in order to show
that |rf (r)|  |rp| + |rq|, it suces to showthat |rs|  |rp| + |rq|. Let x = |rs|
and = rsq; see Figure 2(c). Note that 3/2  x  3 and 0   6 . By
the cosine rule we have
$ $
|rp| = 1 + x2 2x cos(/3 ) and |rq| = 1 + x2 2x cos .

Dene
$ $
f (x, ) = 1 + x2 2x cos(/3 ) + 1 + x2 2x cos x. (3)

Then, |rp| + |rq| |rs| =

f (x, ). In
the full version of the paper we show
that f (x, )  0 for all 3/2  x  3 and 0   6 . This implies that
|rs|  |rp| + |rq|.

To this end, we have proved the following theorem:
Theorem 1. Let R and B be two disjoint sets of points in the plane such that
R B is in general position, and let n = |R B|. One can compute, in O(n log n)
time, a plane spanning tree in K(R, B) whose length is at least 1/4 times the
length of a maximum plane spanning tree.

3.1 A matching upper bound

In this section we show that the above estimate is best possible for the length
of the longest star in K(R, B). Consider a set R of n/2 red points and a set B
of n/2 blue points that are equally distributed in two circles of arbitrary very
small radius with their centers at distance 1; see Figure 3. The bichromatic plane
spanning tree/path that is shown in this gure, has n 2 edges of unit length
and one small edge. Any star in K(R, B) has n/4 edges of unit length, plus n/4
edges of very small length. Thus, the length of the longest star, in this example,
is about 1/4 times the length of an optimal MaxBPST (at the limit).
Maximum Plane Trees in Multipartite Geometric Graphs 199

Fig. 3. Illustration of the upper bound.

4 The Max-k-PST problem

In the multicolored version of the maximum plane spanning tree problem, the
input points are colored by more than two colors, and we want the two endpoints
of every edge in the tree to have distinct colors. Formally, we are given a set
P of n points in the plane in general position that is partitioned into subsets
P1 , . . . , Pk , with k  3. For each c {1, . . . , k}, assume the points of Pc are
colored c. Let K(P1 , . . . , Pk ) be the complete multipartite geometric graph on
P , which has edges between every point of each set in the partition to all points of
the other sets. The Max-k-PST problem is the problem of computing a maximum
plane spanning tree in K(P1 , . . . , Pk ). The standard MaxPST problem can be
interpreted as an instance of this multicolored version in which k = n, i.e., each
point has a unique color. In this section, we present an approximation algorithm,
for the Max-k-PST problem, whose ratio is 1/6 if k = 3, and 1/8 if k  4.
We will show that the length of the longest star in K(P1 , . . . , Pk ) is at least
1/8 (resp. 1/6) times the length of an optimal Max-k-PST if k  4 (resp. k =
3). In fact, we present an O(n log n)-time algorithm that returns such a star.
The algorithm is as follows. Compute a bichromatic diameter (p, q) of P , i.e.,
two points of dierent colors that have the maximum distance. It can easily be
veried, by contradiction, that the MaxST in K(P ), which can be computed in
O(n log n) time [5], contains a bichromatic diameter of P . Notice that the length
of any edge in K(P1 , . . . , Pk ) is at most |pq|. Without loss of generality assume
that p Pi and q Pj . Notice that all points of P \ (Pi Pj ) lie in lune(p, q),
because, otherwise (p, q) cannot be a bichromatic diameter of P . To simplify
the notation, we write R, B, and G, for Pi , Pj , and P \ (Pi Pj ), respectively.
Moreover, we assume that the points of R, B, and G are colored red, blue,
and green, respectively. Compute a diametral pair (r, r  ) in R, a diametral pair
(b, b ) in B, and a diametral pair (g, g  ) in G. Return the longest star Sx in
K(P1 , . . . , Pk ) that is centered at a point x {p, q, r, r , b, b , g, g  }. We show
that the length of Sx is at least 1/8 times the length of an optimal tree.
Let T be an optimal Max-k-PST, and let L denote the length of T . We
make an arbitrary point the root of T and partition the edges of T into four

sets as follows. Let ER be the set of edges (u, v) in T where u is a red point

and u is the parent of v. Let EB be the set of edges (u, v) where u is a blue

point and u is the parent of v. Let EG be the set of edges (u, v) where u is a
green point, v is not a green point, and u is the parent of v. Let E be the set

of edges (u, v) where both u and v are green points. The edges of each of ER ,
200 A. Biniaz et al.

EB , and EG form a forest in which each component is a red-star, a blue-star,
and a green-star, respectively. Let LR , LB , and LG denote the total lengths of

the edges of ER , EB , and EG , respectively. Let LE denote the total length of
the edges in E . Then,

L = LR + LB + LG + LE .

We consider the following two cases depending on where or not LE is larger than
max{LR , LB , LG }.

LE  max{LR , LB , LG }. In this case LE  14 L . The number of edges in E

is at most n(G)1, where n(G) is the number of points in G. Recall that the
length of every edge in E is at most |pq|. Thus LE  (n(G) 1) |pq|. Each
of the stars Sp and Sq has an edge to every point of G. Thus L(Sp )+L(Sq ) 
n(G) |pq|. Therefore,
1
4
L  LE  (n(G) 1) |pq| < n(G) |pq|  L(Sp ) + L(Sq ),

which implies that the longest of Sp and Sq has length at least 18 L .

LE < max{LR , LB , LG }. Without loss of generality assume that LB =
max{LR , LB , LG }. Thus, LB  14 L . Let FB be the set of edges that is
obtained by connecting every point of R G to its farthest blue point.
Notice that the edges of FB form a forest in which every component is
a blue-star. Observe that LB  L(FB ). Moreover, by Lemma 2 we have
L(FB )  L(Sb ) + L(Sb ). Therefore,
1
4
L  LB  L(FB )  L(Sb ) + L(Sb ),

which implies that the longest of Sb and Sb has length at least 18 L .

The point set {p, q, r, r , b, b , g, g  } can be computed in O(n log n) time, and thus,
the running time of the algorithm follows.

Note 1. If k = 3, then E is empty, and thus, the longest star Sx with x

{r, r , b, b , g, g  } has length at least 16 L .

Note 2. If the diameter pair of P is monochromatic, then we get the ratio of

1/6. Assume both points of a diametral pair (p, q) of P belong to Pi . Let R = Pi
and G = R \ Pi . Then, B is empty and L = LR + LG + LE . Moreover, we have
r = p and r = q. If LG  13 L , then the longest of Sg and Sg has length at
least 16 L . If LG < 13 L , then LR  L(Sr ) + L(Sr ) and LE  L(Sr ) + L(Sr ).
Thus, the longest of Sr and Sr has length at least 16 L .

Theorem 2. Let P be a set of n points in the plane in general position that is

partitioned into subsets P1 , . . . , Pk , with k  3. One can compute, in O(n log n)
time, a plane spanning tree in K(P1 , . . . , Pk ) whose length is at least 1/8 (resp.
1/6) times the length of a maximum plane spanning tree if k  4 (resp. k = 3).
Maximum Plane Trees in Multipartite Geometric Graphs 201

5 The MaxPST problem

In this section we study the MaxPST problem, a special case of the Max-k-PST
problem where every input point has a unique color. Formally, given a set P
of points in the plane in general position and we want to compute a maximum
plane spanning tree in K(P ). We revisit this problem which was rst studied by
Alon, Rajagopalan and Suri (1993), and then by Dumitrescu and Toth (2010).
Alon et al. [1] presented an approximation algorithm with ratio 1/2 for this
problem. In fact they proved that the length of the longest star in K(P ) is at
least 0.5 times the length of an optimal tree; this bound is tight for a longest
star. Dumitrescu and Toth [4] improved the approximation ratio to 0.502. They
proved that the length of the longest double-star in K(P ) (a double-star is a tree
with two internal nodes) is at least 0.502 times the length of an optimal solution.
They left as an open problem a more precise analysis of the approximation ratio
of their algorithm. In this section we modify the algorithm of Dumitrescu and
Toth [4], and slightly improve the approximation ratio to 0.503. We will describe
their algorithm briey, and provide detail on the parts that we will modify. The
algorithm outputs the longest of ve plane trees Sa , Sb , Sh , Ea , Eb , that are
describe below.

c y=0.558
v
u
y=0.15
a b a b
p
y= 0.15
Va Vb

Vm
Va Vb
Vm
x=0.2 x=0.8 x=0.2 x=0.8

(a) (b)

Fig. 4. (a) The plane tree Ea . The solid, dashed, and dotted edges belong to Ea1 , Ea2 ,
and Ea3 respectively. (b) The distance between p and c is less than 0.948.

Let n = |P |, and let L denote the length of an optimal MaxPST in K(P ).

Compute a diametral pair (a, b) of P . Without loss of generality assume that ab
is a horizontal, |ab| = 1, a = (0, 0), and b = (1, 0). Since (a, b) is a diametral
pair, the length of any edge in K(P ) is at most 1. Thus, L  n 1. Moreover,
all points of P are in lune(a, b). See Figure 4. Let h = (xh , yh ) be a point in P
with the largest value of |y| (absolute value of the y-coordinate). Without loss
202 A. Biniaz et al.

of generality assume that yh  0. Dene Sa , Sb , and Sh as three spanning stars

that are centered at a, b, and h respectively. We compute plane trees Ea and Eb
as follows; this computation is dierent from the one that is presented in [4].
Set w = 0.6. Let Va be the vertical strip between the lines x = 0 and x =
1w 1+w
2 = 0.2, Vm be the vertical strip between the lines x = 0.2 and x = 2 = 0.8,
and Vb be the vertical strip between the lines x = 0.8 and x = 1. Note that a Va
and b Vb . See Figure 4(a). We describe how to construct Ea ; the construction
of Eb is analogous. Connect a to each point in Vb and let Ea1 denotes the set
of edges of the resulting star (the solid edges in Figure 4(a)). The edge ab has
length 1 and every other edge has length at least 1+w 1
2 . The edges of Ea partition
Va into convex regions. Each of these regions is bounded by at least one edge of
Ea1 from above or below. Take any region A of the partition of Va . Let av be an
edge of Ea1 that bounds A either from above or blow. Note that v Vb . Connect
all points that are in A (excluding a) to v. Let Ea2 be the set of all such edges
after considering all regions (the dashed edges in Figure 4(a)). Since any edge in
Ea2 connects a point in Va to a point in Vb , each edge in Ea2 has length at least
w. The edges of Ea1 Ea2 partition Vm into convex regions. Each of these regions
is bounded by at least one edge of Ea1 Ea2 . Take any region M of the partition
of Vm . Consider the following two cases: (a) If M is bounded by an edge uv of
Ea2 , then connect all points that are in M to one of u and v that maximizes
the total length of the new edges, and (b) if M is not bounded by any edge of
Ea2 , then it is bounded by an edge av of Ea1 , connect all points that are in M
to one of a and v that maximizes the total length of the new edges. Let Ea3 be
the set of all such edges after considering all regions of Vm (the dotted edges in
Figure 4(a)). Let Ea be the graph with vertex set P and edge set Ea1 Eb2 Ea3 .
The following is a restatement of Lemma 3 in [4].
Lemma 3 (Dumitrescu and Toth [4]). Let a and b be two points in the
plane that are at distance at least from each other, for some real > 0. Let S
be a set of m points in the plane. Let Sa and Sb be two star that are centered at
a and b, respectively, and connected to all points of S. Then, the length of the
longest of Sa and Sb is at least m
2 .
Recall that the length of each edge in Ea1 is at least 1+w
2 and the length of
each edge in Eb2 is at least w. By Lemma 3 the total length of the edges in Ea3
is at least w2 times the number of points in Vm .
Lemma 4. Let na and nb denote the number of points in Va and Vb respectively.
Then L(Ea )  n2b + w2 (n + na ) + 13w
2 , L(Eb )  2 + 2 (n + nb ) +
na w 13w
2 , and
na +nb
consequently L(Ea ) + L(Eb )  2 + 2 (2n + na + nb ) + 1 3w.
w

Proof. Since Ea1 , Ea2 , and Ea3 are pairwise disjoint, we have L(Ea ) = L(Ea1 ) +
L(Ea2 ) + L(Ea3 ). Ea1 contains nb edges (including ab) of length at least 1+w
2 with
the length of ab is 1. Thus, L(Ea1 )  1+w 2 (n b 1) + 1. E 2
a has na 1 edges of
length at least w. Thus, L(Ea2 )  w(na 1). Vm contains n na nb points,
and thus, the length of Ea3 is at least w2 (n na nb ). Therefore,
1+w w nb w 13w
L(Ea )  2
(nb 1) + 1 + w(na 1) + 2 (n na nb ) = 2
+ 2
(n + na ) + 2 .
Maximum Plane Trees in Multipartite Geometric Graphs 203

The estimation of L(Eb ) is analogous.


The following lemma summarizes Lemmas 3, 4, 5, and 7 in [4]. Let P =
{p1 , . . . , pn }, where pi = (xi , yi ). Let dmax (pi ) denote the maximum distance
from pi to other points in P .
Lemma 5 (Dumitrescu and Toth [4]). Let and t be two constants where
0   t  1. Then
+ L(Sb )  n.
1. L(Sa )
2. L  i=1 dmax (pi ).
n
n %
3. if i=1 |yi |  n, then L(Sa ) + L(Sb )  2n 14 + 2 .
n
4. if i=1 |yi |  n and yh  t, then L(Sh )  (t )n.
Set = 0.055, t = 0.558, and z = 0.49; these constants are dierent from the
ones that are chosen in [4].
Lemma 6. Assume that |yh |  t. Let pi = (xi , yi ) be a point of P in Vm with
|yi |  0.15. Then dmax (pi )  0.948.
Proof. The maximum distance is attained when pi = (0.2, 0.15) or pi =
(0.8, 0.15). Because of symmetry we assume that pi = (0.2, 0.15). Let c =
(xc , yc ) be the rightmost intersection point of the line y = t and lune(a, b). See
Figure 4(b). The furthest point from pi in the allowed region is c. Note that
yc = t $ and xc = 1 t2 < 0.83. Thus, xc xi < 0.63. Therefore dmax (pi ) 
|pi c| < 0.632 + (0.15 + t)2 < 0.948.

To prove the approximation ratio, we show that the length of the longest of
Sa , Sb , Sh , Ea , Eb is at least 0.503 times L . In order to do that, we recall the
following four cases that are considered in [4].
n
1. If i=1 |yi |  n, then the output of the algorithm is not shorter than the
longest of Sa and Sb . By Lemma 5, the approximation ratio is at least
&
L(Sa ) + L(Sb ) 1
 + 2  0.503.
2L 4
n
2. If i=1 |yi | < n and yh  t, then the output of the algorithm is not shorter
than Sh . By Lemma 5, the approximation ratio is at least
L(Sh )
 t = 0.503.
L
n
3. If i=1 |yi | < n, yh < t, and na + nb  (1 z)n, then the output of the
algorithm is not shorter than the longest of Ea and Eb . As a consequence of
Lemma 4, the approximation ratio is at least
L(Ea ) + L(Eb ) na + nb w(2n + na + nb ) 1 3w

 + +
2L 2 2n 2 2n 2n
(1 z)(1 + w) w 1 3w
 + +  0.503,
4 2 2n
204 A. Biniaz et al.

where the last inequality is valid for all n  400. Based on the above calcu-
lations and the details that are provided in the proof of Lemma 4, it turns
outthat ratio of 0.502 that is claimed in [4] is valid when n  100.
n
4. If i=1 |yi | < n, yh < t, and na + nb < (1 z)n, then the output of the
algorithm is not shorter than the longest of Sa and Sb . There are at least
11
zn = .49npoints in Vm . At most 30 points have |yi |  0.15 because otherwise
n 11n
we have i=1 |yi |  0.15 30 = 0.055n = n, which is a contradiction. Thus,
at least 49n 11n 37n
100 30 = 300 points in Vm have |yi |  0.15. By Lemma 5 and
Lemma 6 we have
263n 37n
L  + 0.948 < 0.994n.
300 300
The approximation ratio is at least
L(Sa ) + L(Sb ) n
 > 0.503.
2L 2 0.994n
Theorem 3. Let P be a set of n points in the plane in general position. One
can compute, in O(n log n) time, a plane spanning tree in K(P ) whose length is
at least 0.503 times the length of a maximum plane spanning tree.

6 Concluding Remarks
In this paper we presented constant factor approximation algorithms for the
problem of computing a maximum plane tree in a multipartite geometric graph.
It is not known whether or not this problem is NP-hard. A natural open problem
is to improve any of the presented approximation ratios. Specically, when the
number of sets in the partition is more than two, we conjecture that the length
of the longest star is at least 1/3 times the length of a maximum spanning tree.

References
1. N. Alon, S. Rajagopalan, and S. Suri. Long non-crossing congurations in the plane.
Fundamenta Informaticae, 22(4):385394, 1995. Also in Proceedings of the 9th ACM
Symposium on Computational Geometry (SoCG), 257263, 1993.
2. A. Biniaz, P. Bose, D. Eppstein, A. Maheshwari, P. Morin, and M. Smid. Spanning
trees in multipartite geometric graphs. CoRR, abs/1611.01661, 2016. Also submitted
to Algorithmica.
3. M. G. Borgelt, M. J. van Kreveld, M. Loer, J. Luo, D. Merrick, R. I. Silveira,
and M. Vahedi. Planar bichromatic minimum spanning trees. Journal of Discrete
Algorithms, 7(4):469478, 2009.
4. A. Dumitrescu and C. D. Toth. Long non-crossing congurations in the plane.
Discrete & Computational Geometry, 44(4):727752, 2010. Also in Proceedings
of the 27th International Symposium on Theoretical Aspects of Computer Science
(STACS), 311322, 2010.
5. C. L. Monma, M. Paterson, S. Suri, and F. F. Yao. Computing Euclidean maximum
spanning trees. Algorithmica, 5(3):407419, 1990.
 
Local Routing in Spanners Based on WSPDs

Prosenjit Bose1 , Jean-Lou De Carufel2 , Vida Dujmovic2 , and Frederik Paradis2

1
School of Computer Science, Carleton University
Ottawa, Canada
[email protected]
2
School of Electrical Engineering and Computer Science, University of Ottawa
Ottawa, Canada
{jdecaruf,vida.dujmovic,fpara058}@uottawa.ca

Abstract. The well-separated pair decomposition (WSPD) of the com-

plete Euclidean graph dened on points in R2 (Callahan and Kosaraju
[JACM, 42 (1): 67-90, 1995]) is a technique for partitioning the edges of
the complete graph based on length into a linear number of sets. Among
the many dierent applications of WSPDs, Callahan and Kosaraju proved
that the sparse subgraph that results by selecting an arbitrary edge from
each set (called WSPD-spanner) is a 1 + 8/(s 4)-spanner, where s > 4
is the separation ratio used for partitioning the edges.
Although competitive local-routing strategies exist for various spanners
such as Yao-graphs, -graphs, and variants of Delaunay graphs, few
local-routing strategies are known for any WSPD-spanner. Our main
contribution is a local-routing algorithm with a near-optimal competi-
tive routing ratio of 1 + O(1/s) on a WSPD-spanner. Specically, we
present a 2-local and a 1-local routing algorithm on a WSPD-spanner
with competitive routing ratios of 1 + 6/(s 2) + 4/s and 1 + 6/(s 2) +
6/s + 4/(s2 2s) + 8/s2 , respectively.

1 Introduction

A fundamental problem in networking is the routing of a message from one

vertex to another in a graph. Because network resources are limited, it is often
desirable that routing algorithms use as little memory as possible. At one extreme
in this direction are local routing algorithms in which the routing algorithm must
choose the next vertex to forward a message to based solely on knowledge of the
destination vertex, the current vertex and some information about all vertices
directly connected to the current vertex. When local routing is not possible, it
is still desirable that a routing algorithm use little memory.
In many settings it is natural to model a network as a geometric graph, that is,
a graph whose vertices are points and each edge is a line-segment whose weight
is the Euclidean distance between its two endpoints. For example, geometric
routing algorithms are important in wireless sensor networks (see [11] for a survey
of the area) since routing strategies can take advantage of the fact that nodes in

Research supported in part by NSERC and OGS.

Springer International Publishing AG 2017 205

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 205216, 2017.
DOI: 10.1007/978-3-319-62127-2_18
206 P. Bose et al.

these networks have physical locations that can be used to help guide a packet
to its destination.
A geometric routing strategy is said to be competitive if the length of the path
found by the routing strategy is not more than a constant times the Euclidean
distance between its endpoints. This constant is called the routing ratio. In order
to nd a competitive path between any two vertices of a graph, such a path
must rst exist. Graphs that meet this criterion are called (geometric) spanners.
Formally, given a geometric graph G, the distance, dG (u, v), between two vertices
u and v in G is the sum of the weights of the edges in the shortest path between
u and v in G. G is a t-spanner if for all pairs of vertices u and v in G, dG (u, v)
t |uv| for t 1. Here |uv| denotes the Euclidean distance between u and v. The
smallest value t for which G is a t-spanner is the spanning ratio or stretch factor
of G. A family of graphs that are t-spanners, for some xed constant t, are often
referred to as simply spanners. Spanners have been extensively studiedfor a
detailed overview of results on geometric spanners, see [12].
Geometric spanners tend to fall into three categories: (i) Long-known ge-
ometric graphs that happen to be spanners, such as Delaunay triangulations;
(ii) cone-based constructions, such as Keils -graphs [10]; and (iii) well-separated
pair decomposition (WSPD) based constructions introduced by Callaghan and
Kosaraju [6] in the early 90s. Note that graphs in the rst category have xed
worst-case spanning ratios bounded away from 1. Constructions in the second
and third categories are designed for a given parameter. They can achieve span-
ning ratios arbitrarily close to 1 by choosing arbitrarily small values for this
parameter. Signicant work has gone into nding competitive local and low-
memory routing algorithms for graphs in the rst category, including Delaunay
graphs (classical, L1 , L , and TDDelaunay triangulations) [7,4,1,3]. In most
cases, proving tight results about graphs in this category is dicult. For exam-
ple, even the exact worst-case spanning ratio of the Delaunay triangulation is
unknown, despite over 30 years of study [8,10,2,13].
For the second categorycone-based spannerscompetitive local routing al-
gorithms are usually trivial. These spanners are designed so that greedy choices
produce paths of low stretch. Still, for certain cone-based spanners, there have
been some rened results on competitive routing algorithms that produce ex-
ceptionally low competitive ratios. For example, Bose et al [3] present a routing
algorithm for the Half-6 -graph with a competitive ratio of 2.887.
In this paper, we consider routing algorithms for the third category: WSPD-
based spanners. Intuitively, a WSPD of a pointset is a partition of the edges
of the complete geometric graph (on that pointset) such that all edges in the
same partition are approximately of equal length.3 Since its introduction by
Callahan and Kosaraju [6] in early the 90s, the WSPD and WSPD-based span-
ners have found a plethora of applications in solving distance problems [12].
The main diculty about local routing in these spanners stems from the fact
that WSPD-spanners are based on WSPDs which are built globally and capture

3
See the next section for the formal denition.
Local Routing in Spanners Based on WSPDs 207

global distance properties of the given pointset. As such WSPD-spanners pose

a challenge in designing local routing strategies.
Given a pointset and a separation ratio s, a WSPD with separation ratio s is
(typically) not unique. Callahan and Kosarajus original construction of a WSPD
is based on fair split-trees and it computes a WSPD containing a linear number
of edge partitions [6]. From this WSPD, we show how to construct a WSPD-
spanner that facilitates local routing by selecting a well-chosen edge from each
partition rather than picking an arbitrary edge (Section 2). As a side benet, our
WSPD-spanner has a slightly improved spanning ratio, 1+4/s+4/(s2), over the
original one, 1+8/(s4). This improvement stems from the additional properties
our well-chosen edges have. On this WSPD-spanner, we present a 2-local and a
1-local routing algorithm with competitive routing ratios of 1 + 4/s + 6/(s 2)
and 1 + 6/(s 2) + 6/s + 4/(s2 2s) + 8/s2 , respectively (Sections 3 and 4). A
routing algorithm on a graph G is k-local if each vertex v of G stores information
about vertices that are at hop distance at most k from v. Ideally, one would like
the routing ratio to be identical to the spanning ratio, however, this is rarely
the case when routing locally since an adversary can often force an algorithm to
stray from the actual shortest path. Finally, we prove a lower bound of 1+8/s on
the spanning ratio of our WSPD-spanner, thereby proving the near-optimality of
the spanning ratio of our WSPD-spanner and the near-optimality of the routing
ratio of both our routing algorithms.
WSPDs have been used before as an aid to routing in unit-disk graphs by
Kaplan et al. [9]. Their scheme applies to our setting when the unit distance
is the diameter of the point set. Their routing scheme requires a header of
O(log n log D) bits, where D is the diameter, and routing tables of total size
O(n5 log2 n log2 D) bits. Their routing ratio is 1 +  where  = (/s) log |pq|
with 192. Our routing scheme does not use a header and requires routing
tables of total size O(s2 nB) bits, where B is the maximum number of bits to
store a bounding box. Our routing ratio is 1 +  where  = 6/(s 2) + 4/s for
our 2-local algorithm and  = 6/(s 2) + 6/s + 4/(s2 2s) + 8/s2 for our 1-local
algorithm. The main advantage of our scheme is that we do not require a header.

2 Construction of t-Spanners Using WSPDs

In this section, we explain how to construct a WSPD-spanner on which our

results are based. We also prove some useful geometric lemmas concerning these
spanners. We begin with some denitions. Two point sets A and B are well-
separated with separation ratio s if there exist two circles with the same radius,
, one containing A and the other containing B such that the minimum distance
between the two circles is s . A well-separated pair decomposition (WSPD)
of a point set S R2 with separation ratio s is a set of pairs of subsets of S:
{{A1 , B1 }, . . . , {Am , Bm }} such that for each pair {Ai , Bi }, 1 i m, Ai and
Bi are well-separated with separation ratio s and for any distinct points p and
q in S, there is a unique pair {Ai , Bi }, 1 i m, such that p Ai and q Bi .
208 P. Bose et al.

Callahan and Kosarajus [6] classical construction of WSPDs uses the fair-
split tree. We denote the bounding box of a point set A, the smallest axis-aligned
rectangle enclosing A, by R(A). The fair split tree is dened as follows. Take
the bounding box R(S) of the point set S and make it the root u of the split
tree. Then, split R(S) on its longest axis and make the bounding boxes of two
resulting point sets the children of u. Repeat for each child until the leaves are
the points of S. Callahan and Kosarajus [6] classical construction of a spanner
given a WSPD proceeds as follows: for each pair {A, B}, select an arbitrary
point a A as a representative of the set A and an arbitrary point b B as a
representative of the set B and add the edge ab to the graph. We call any such
spanner a WSPD-spanner. Callahan and Kosaraju [5] proved that any WSPD-
spanner has a spanning ratio of at most 1 + 8/(s 4), where s is the separation
ratio of the WSPD. At the heart of their proof is the following lemma.
Lemma 1 (Callahan and Kosaraju [5]). Let {A, B} be a well-separated pair
with respect to the separation ratio s > 0. Let p, p , p A and q, q  B. Then,
|p p | (2/s)|pq| and |p q  | (1 + 4/s)|pq|.
Any WSPD-spanner that is built using the WSPD resulting from Callahan
and Kosarajus fair-split tree, we call ASW-Spanner (standing for Arbitrary
representative Split tree based WSPD-spanner). To facilitate our routing al-
gorithm, rather than selecting an arbitrary point as representative of a set in
a pair, we choose the rightmost point in each set as its representative. If there
is more than one candidate, we choose the topmost point among the rightmost
ones. We refer to an ASW-Spanner constructed in this way as an RSW-Spanner
(standing for Rightmost representative Split tree based WSPD-spanner). It is
on RSW-Spanners that we prove our routing results. Unless stated otherwise,
for the remainder of this paper, we focus on RSW-Spanners.
By exploiting properties of the fair-split tree, we prove that the spanning ratio
of RSW-Spanners is at most 1 + 4/(s 2) + 4/s which is a slight improvement
over the spanning ratio of 1 + 8/(s 4), shown for any WSPD-spanner. When
using a fair-split tree to construct a WSPD, each set A from a well-separated pair
corresponds to a unique node u in the fair-split tree. The points in A are precisely
the points in the subtree rooted at u. The following lemmas are consequences of
the denition of fair-split trees, the construction of the WSPD, and the choice
of representatives in an RSW-Spanner.
Lemma 2. For any two nodes u and v in the fair split tree, their corresponding
bounding boxes are either disjoint or one is a subset of the other.
Lemma 3. In an RSW-Spanner, consider two sets P and Q, each from a pair
of the WSPD. Let p be a representative of P . If Q P and p Q, then p is
also the representative of Q.
Lemma 4. In an RSW-Spanner, let {A, B} be a pair in the WSPD, a, x A
be two points such that a is the representative of A, and x
= a. There is a well-
separated pair {C, D} such that: (1) a C, (2) x D, (3) a is the representative
of C, (4) C is a proper subset of A and (5) D is a proper subset of A.
Local Routing in Spanners Based on WSPDs 209

Lemma 5. In an RSW-Spanner, let A be a set in a pair from the WSPD and

let a, x A be two points such that a is the representative of A and x
= a. Let
C be a set in a pair from the WSPD such that A C and a point c C is the
representative of C. If xc is an edge, then ac is an edge.
Lemma 6. In an RSW-Spanner, let {A, B} and {C, D} be two distinct pairs
from the WSPD, such that A C. Let c be the representative of C. Let x be any
point in B and let {E, F } be the unique pair from the WSPD separating c E
from x F . Then, c is the representative of E.
Algorithm 1 nds a path between p and q in any WSPD-spanner and is
inspired from the proof of Theorem 9.2.1 by Narasimhan and Smid in [12].

Algorithm 1 FindPath(p, q)
Precondition: p
= q
Let {A, B} be the unique pair in the WSPD separating p A from q B.
Let a and b be the representatives of A and B.
return FindPathRec(p, a, A), FindPathRec(q, b, B)

Algorithm 2 FindPathRec(v, w, E)
Precondition: v, w E and w is the representative of E
if v = w then return v
else
Let {C, D} be the pair in the WSPD separating v C from w D.
Let c and d be the representatives of C and D, respectively.
return FindPathRec(v, c, C), FindPathRec(w, d, D)

Lemma 7. Let p, q S. Consider a call to FindPath(p, q) in a ASW-Spanner

of S. Consider the call to FindPathRec(v, w, E) at recursion depth k 1.
For any two points e, f E, |ef | (2/s)k |pq|.
Theorem 1. The spanning ratio t of an RSW-Spanner is at most 4/(s 2) +
4/s + 1.
Proof. We nd an upper bound on the spanning ratio by analyzing the path
found by FindPath. Consider any of the calls to FindPathRec in FindPath(p,
q). Since w is the representative of E, by Lemma 4, we know that w is the
representative of D and, thus, w = d. This means that for each level k 1,
the call to FindPathRec(w, d, D) returns immediately. In other words, for all
k 1, there is exactly one edge of level k. Notice that v, w E according to the
preconditions of FindPathRec(v, w, E). Therefore, by Lemma 2, c and d are
also both in E. From Lemma 7, we get |cd| (2/s)k |pq|. The fact that there is
exactly one edge of level k allows us to get only (2/s)k as the sum of the length of
all edges at level k. Then, if wesum up the lengthof all edges [c, d] from level 1 to
m
a maximum depth m, we nd i=1 (2/s)i |pq| i=1 (2/s)i |pq| = (2/(s2))|pq|.
Let {A, B} be the pair separating p A and q B. Let a A and b B
be the representatives of A and B, respectively. From Lemma 1, we have that
210 P. Bose et al.

|ab| (1 + 4/s)|pq|. To bound the path found by FindPath(p, q), we take

the length of the path found by the call to FindPathRec(p, a, A), add the
length of the edge [a, b], and add the length of the path found by the call to
FindPathRec(q, b, B). Thus, the path found in FindPath(p, q) has a length
of at most 2 (2/(s 2))|pq| + (1 + 4/s)|pq| = (4/(s 2) + 4/s + 1)|pq|.


Theorem 2. For any s > 0, there exist an RSW-Spanner with a spanning ratio
arbitrarily close to 1 + 8/s.

Proof. Let 0 <  < be a real number. Let S = {p, p , q, q  } be a point set
such that p = (cos(/2 + ), sin(/2 + )), p = (cos(/2 + ), sin(/2 + )),
q = (cos(/2), sin(/2)+s+2) and q  = (cos(/2), sin(/2)+s+2).
Let A = {p, p } and B = {q, q  }. By construction, there is a pair {A, B} in the
WSPD. Again by construction, p is the representative of R(A) and q  is the
representative of R(B). Hence, the only path between p and q is pp q  q. We can
show that the spanning ratio of the path between p and q and, therefore, the
spanning ratio of the graph approaches lim 0 (|pp |+|p q  |+|q  q|)/|pq| = 1+8/s
as  approaches 0.


3 2-Local-Routing Algorithm
In this section, we present our competitive 2-local routing algorithm on RSW-
Spanners. We begin with some notation. For two distinct points t, u S, with
{T, U } being the unique pair in the WSPD with t T and u U , let Btu (t) :=
R(T ), and let Btu (u) := R(U ). Let u be the representative of U , then notice
that u is the representative of Btu (u ) = Btu (u) = R(U ). Therefore, u has
an edge to the representative of Btu (t).
In order to describe our 2-local routing algorithm, we need to describe the
precise information available at each vertex of the RSW-Spanner Let c be the
current vertex of the routing path. Let d be any neighbor of c. Let e be any
neighbor of d. We assume that the following information is available at c: (1)
Bcd (c) and Bcd (d); (2) Bde (d) and Bde (e). Notice that we know Bde (d) and
Bde (e) even though the current point is c. This information makes our routing
algorithm 2-local. In Section 4, we will modify our algorithm so that it does not
need to know Bde (e). This will lead to a 1-local routing algorithm with a slightly
larger routing ratio. We want to nd a path between two points p S and q S.
Let {A, B} be the unique pair in the WSPD separating p from q. Let a and b be
the representatives of A and B. The goal for our algorithm is to nd competitive
path from p to a, take the edge ab, and then, nd a competitive path b from q.
To nd a competitive path from p to a, we use the following strategy. Let
v be the current point on the path from p to a produced by our algorithm (at
the beginning v = p). Here is how our algorithm selects the next edge. For each
vertex v  adjacent to v, let Bv be the largest bounding box such that v  is the
representative of Bv . The next edge chosen by our algorithm is the edge vw
such that the size of Bw is maximized, Bw contains p, and Bw is contained in
Bpq (p) = R(A). Note that we do not know yet whether Bw is contained in Bpq (p).
Local Routing in Spanners Based on WSPDs 211

However, we will prove, in Section 3.1, that this is the case. Upon reaching a,
we take the edge ab. To nd a competitive path from b to q, we notice that
b already has an edge to the representative of a bounding box smaller than
R(B) containing q. Thus, the algorithm simply takes that edge. Then, we repeat
this procedure until we nd q. Our algorithm is summarized in Algorithm 3.
Note that sizeof(Bv ) denotes the area of Bv and N (v) denotes the set of all
neighbours of v.

Algorithm 3 FindPathTwoLocal(v, p, q)
Input: the current point v, the source p and the destination q.
Output: The next point w on the path.
1: Consider all bounding boxes of every neighbor v  of v.
2: if v  is the representative of Bvq (q) then w v  // Reducing step
3: else// Enlarging step
4: if v  is the representative of Bpq (p) then w v 
5: else
6: Let Bv is the largest bounding box that v  is the representative of.
7: Let V = {v  N (v) | p Bv , v  is not the representative of Bv q (v  )}
8: w argmaxv V sizeof(Bv )
9: return w

As said previously, Algorithm 3 is 2-local. Indeed, the information used is

the bounding boxes of: the current point v, every neighbor v  of v, and every
neighbor v  of every neighbor of v. However, it is not obvious how the bounding
boxes of the v  s are used. To know whether v  is the representative of Bpq (p)
or Bv q (v  ), we need to know if v  is the representative of a bounding box which
is separated from another bounding box containing q.

3.1 Correctness: 2-Local

In this section, we prove the correctness of Algorithm 3 (see Theorem 3). For
the rest of this paper, we denote by Pt (p, q) the path from p to q with spanning
ratio t, found by the FindPath algorithm, and, we denote by P (p, q) the path
from p to q found by Algorithm 3. The following lemma is used to prove the
correctness of Algorithm 3 and to establish an upper bound on the routing ratio
of Algorithm 3 (see Theorem 4). By using the fact that Bv is the largest bounding
box that v is the representative of and supposing that p Bv , it establishes a
relation between Bv and the bounding boxes containing p of the vertices of
Pt (p, q). This relation allows us to nd the conclusion of Lemma 8.

Lemma 8. Consider any RSW-Spanner. Let v be a point inside Bpq (p) that is
not the representative of Bpq (p) such that p is in Bv . Among all edges of Pt (p, q),
let de be the edge such that Bde (d) is the smallest bounding box containing p that
is larger than Bv . Then, there is an edge between v and d.
212 P. Bose et al.

Proof. We rst argue that the edge de is well-dened. Since v is inside Bpq (p)
but is not the representative of Bpq (p), we know that Bv is smaller than and
inside Bpq (p) by Lemma 3 and 2, respectively. This implies that the set of edges
[, ] from Pt (p, q) such that p is in B () and Bv is smaller than B ()
is non-empty. Indeed, the edge ab from Bpq (p) to Bpq (q) is in that set since
Bab (a) = Bpq (p), and Bv is smaller than and inside Bpq (p). Therefore, the edge
de is well-dened.
Let c be the point before d in Pt (p, q). Since d is in Pt (p, q), then d is repre-
sentative of Bde (d). Therefore, by Lemma 4, we know that d is the representative
of the bounding box separating d from p. Then, c is the representative of Bcd (c)
such that p is in Bcd (c).
Because Bde (d) is the smallest bounding box containing p larger than Bv ,
c Bcd (c) Bv . If v = c is the representative of Bcd (c), then v has an edge to
d. Otherwise, Bcd (c) Bv and we apply Lemma 5 in the following way. We have
that c is in Bv , Bv Bde (d), d is the representative of Bde (d) and there is an
edge between c and d. Therefore, there is an edge from v to d by Lemma 5. 
Recall that {A, B} is the pair separating p A from q B and that a A
and b B are the representatives of A and B, respectively.
Lemma 9. The Enlarging step of Algorithm 3 nds a path in an RSW-Spanner
from p A to a A.
Proof. In this proof, we use some notation introduced in Section 3. Recall that, in
Algorithm 3, we dene V = {v  N (v) | p Bv , v  is not the representative of
Bv q (v  )}. We prove that each edge vw taken in the Enlarging step of Algorithm 3
leads to a bounding box Bw that is larger than Bv but not larger than Bpq (q).
Thus, Algorithm 3 nds a path from p to the representative of Bpq (q) (i.e. the
representative a of A). Suppose that the current point v is inside but is not the
representative of Bpq (p). From Lemma 8, we get that v has an edge to a point
of Pt (p, q) that has a bounding box larger than Bv . This proves that there is
always a choice of edges in the Enlarging step such that Bw is larger than Bv .
Now, we prove that the next edge vw is chosen such that w is inside Bpq (p).
We prove this by contradiction. Suppose Algorithm 3 takes the edge vw where
w is outside of Bpq (p). Therefore, w must be the representatives of Bpq (p) or
must be in V. Since w is outside of Bpq (p), it cannot be the representative of
Bpq (p). Thus, it must be in V. Since p is in Bw and w is outside of Bpq (p), we
have that Bw is larger than Bpq (p). Since the representative of Bpq (p) has an
edge to a bounding box containing q, from Lemma 6, we also get that w has an
edge to a bounding box containing q which contradicts the denition of V.
Because v = p is inside Bpq (p) in the rst call of Algorithm 3, we then get
that each edge vw taken in the Enlarging step of Algorithm 3 leads to a bounding
box Bw that is larger than Bv but not larger than Bpq (q).

Once a is found, then Algorithm 3 follows the edge ab since a has an edge
leading to a bounding box containing q. We get the following lemma from the
fact that the Reducing step of Algorithm 3 follows exactly what the recursive
algorithm FindPath does. The following lemma and theorem follow.
Local Routing in Spanners Based on WSPDs 213

Lemma 10. The Reducing step of Algorithm 3 nds a path in an RSW-Spanner

from b B to q B.

Theorem 3. Algorithm 3 nds a path in an RSW-Spanner from p to q.

3.2 Routing Ratio: 2-Local

In this section, we nd an upper bound on the routing ratio of Algorithm 3. We

get the following lemma from the fact that the Reducing step of Algorithm 3
follows exactly what the recursive algorithm FindPath does.

Lemma 11. When the Reducing step of Algorithm 3 is executed on an RSW-

Spanner, the length of the part of the path from b B to q B is at most
(2/(s 2))|pq|.

Lemma 12. When the Enlarging step of Algorithm 3 is executed on an RSW-

Spanner, the length of the part of the path from p A to a A is at most
(4/(s 2))|pq|.

Proof. For the purpose of this proof, we consider the edges of P (p, q) as directed
edges. Thus, if uv is an edge in P (p, q), then u precedes v in P (p, q). We say
that u is the source of the edge and that v is the target of the edge.
Let cde be a subpath of Pt (p, q) such that c, d A and the edge cd is at
level i in the analysis of Theorem 1. Consider the set Ti of edges vw such that
vw is an edge of P (p, q) and the target w is in Bde (d) but not in Bcd (c). We
claim that the sum of the lengths of the edges in Ti is at most 2(2/s)i |pq|, i.e.

vwTi |vw| 2(2/s) |pq|. Then, using the analysis of Theorem 1, if we sum
i

up the lengths of all edges vw from levelm 1 to a maximumdepth m, we get

m
that the length of the path is at most i=1 vwTi |vw| i=1 2(2/s)i |pq|

i=1 2(2/s) |pq| = (4/(s 2))|pq|.
i

We now prove our claim. If Ti is empty, then the sum is zero. Otherwise, let
an edge wj1 wj Ti , where wj is the i-th edge in P (p, q). From Lemma 7, we get
|wj1 wj | (2/s)i |pq|. We consider two cases: either (1) wj is the representative
of Bde (d) or (2) it is not.

1. Consider the edge wj2 wj1 . We consider two subcases: either (a) wj1 is
in Bcd (c) or (b) it is not.
(a) In this case, only wj1 wj has its target in Bde (d) and |wj1 wj |
(2/s)i |pq| 2(2/s)i |pq|. Notice that, in this case, wj1 is the repre-
sentative of Bcd (c) (thus wj1 = c) because wj1 = c can only belong
to one pair separating it from wj = d.
(b) In this case, the edge wj2 wj1 falls in case 2a, i.e. wj2 is in Bcd (c)
but not the representative of Bcd (c), and wj1 is in Bde (d) but not in
Bcd (c). Therefore, the sum of the lengths of all edges having their target
in Bde (d) is |wj2 wj1 | + |wj1 wj | 2(2/s)i |pq|. See Figure 1a.
214 P. Bose et al.

2. From Lemma 8, we get that wj1 is in Bcd (c) but not the representative of
Bcd (c). Otherwise, there would be an edge from wj1 to d. Since wj1 is in
Bcd (c), there is no other edge wk1 wk of P (p, q) preceding wj1 wj , where
wk is in Bde (d) but not in Bcd (c).
Now, consider the edge wj wj+1 . We consider two subcases: either (a) wj+1
is the representative of Bde (d) or (b) it is not.
(a) From Lemma 7, we get |wj wj+1 | (2/s)i |pq|. Therefore, the sum of the
lengths of all edges having their target in Bde (d) is |wj1 wj |+|wj wj+1 |
2(2/s)i |pq|. See Figure 1a.
(b) From Lemma 8, we get that wj has an edge to d. Because wj+1 is not
the representative of Bde (d), wj+1 must be outside of Bde (d). Therefore,
only wj1 wj has its target in Bde (d) and not in Bcd (c) and |wj1 wj |
(2/s)i |pq| 2(2/s)i |pq|. See Figure 1b.

These cases cover all possibilities of edges in Ti .



B wj p wj1
Bwj1 p wj2 Bwj1 wj
Bwj2 wj1 wj+1
Bc,d (c) c
c d P (p, q)
Bc,d (c) Bd,e (d)
d = wj Pt (p, q)
Pt (p, q)
Bd,e (d)
P (p, q)
(a) (b)

Fig. 1. (a) An illustration of the case 1b (and 2a.)(b) Illustration of the case 2b) of Lemma 12.

Theorem 4 combines Lemmas 11 and 12.

Theorem 4. For a given RSW-Spanner, the routing ratio of Algorithm 3 is at
most 6/(s 2) + 4/s + 1.

4 Improvement - 1-Local Routing Algorithm

An important aspect of routing algorithms is how much information each vertex

needs to store. In this section, we present an algorithm that is slightly dierent
from Algorithm 3. The main dierence is that it is 1-local instead of 2-local. Let
c be the current point, d be any neighbor of c and e be any neighbor of d. For
Algorithm 3, we assumed that the following information was available at c: (1)
Bcd (c) and Bcd (d); (2) Bde (d) and Bde (e). In this section, we explain how to
design a routing algorithm that does not need to know Bde (e). As a result, this
increases the upper bound on the routing ratio by 8/s2 + 2/s + 4/(s(s 2)).
In our modied algorithm, only the Enlarging step diers from Algorithm 3
since the Reducing step in Algorithm 3 is already 1-local. To nd a competitive
path from p to a, we use the following modied strategy. Let v be the current
vertex on the path from p to a produced by our algorithm (at the beginning
v = p). Here is how our new algorithm selects the next edge. For each vertex v 
adjacent to v, we modify the denition of Bv . In this section, Bv is the largest
Local Routing in Spanners Based on WSPDs 215

bounding box such that v  is the representative of Bv and the distance between
the enclosing circle Cv of Bv and q is at least s times the radius v of Cv .
The next edge chosen by our algorithm is the edge vw such that the size of Bw
is maximized and Bw contains p. The strategy to a competitive path from b
to q stays the same. Algorithm 4 below outlines the modied algorithm. Notice

Algorithm 4 FindPathOneLocal(v, p, q)
Input: the current point v, the source p and the destination q.
Output: The next point w on the path.
1: Consider all bounding boxes of every neighbor v  of v.
2: if v  is the representative of Bvq (q) then w v  // Reducing step
3: else// Enlarging step
4: v  N (v), let Bv be the largest bounding box that v  is the representative of
such that |Cv q| sv .
5: Let V = {v  N (v) | p Bv }
6: w argmaxv V sizeof(Bv )
7: return w

that this new algorithm does not guarantee that the path stays inside Bpq (p).
However, as shown in the proof of Lemma 9, the rst edge that goes outside of
Bpq (p) has an edge to the representative of a bounding box containing q. Thus,
Algorithm 4 is entering the Reducing step right after this edge is taken. Before
we prove an upper bound on the routing ratio of Algorithm 4 (see Theorem 5),
we need the following lemma.

Lemma 13. Consider any RSW-Spanner. In Algorithm 4, the diameter of the

last enclosing circle in the Enlarging step is at most (2/s)|pq|.

Theorem 5. For a given RSW-Spanner, the routing ratio of Algorithm 4 is at

most 8/s2 + 6/s + 4/(s(s 2)) + 6/(s 2) + 1.

Proof. Consider the analysis of Lemma 12. Let cde be a subpath of Pt (p, q) such
that c, d A and the edge cd is at level i in the analysis of Theorem 1. Recall
that Ti is the set of edges vw such that vw is an edge of P (p, q) and the target
w is in Bde (d) but not in Bcd (c).
From Lemma 9, Algorithm 3 guarantees that the edge ab is taken while
Algorithm 4 does not. However, as said above, the rst edge that goes outside
of Bpq (p) has an edge to a bounding box containing q. This edge is the edge
wj wj+1 in Case 2b, where the level i is 1 and Bde (d) = Bpq (p). From Lemma 13,
we know that |wj wj+1 | (2/s)|pq|. Since T1 in Case 2b only contains the edge
wj1 wj , adding the edge wj wj+1 to T1 gives that the sum of the lengths of the
edges of T1 is at most 2(2/s)|pq|. Thus, if we sum up the lengths of all edges vw
from level 1 to a maximum depth m, we mgetthat the lengthofmthe path found
in the Enlarging step is still at most i=1 vwTi |vw| i=1 2(2/s)i |pq|

i=1 2(2/s) |pq| = 4/(s 2)|pq|.
i
216 P. Bose et al.

Suppose we want to nd a path between the point wj+1 and q with Algo-
rithm 4. Let us rename wj+1 as u. Since u already has an edge to a bounding
box containing q, Algorithm 4 enters immediately the Reducing step at u. This
part of the algorithm does not dier from Algorithm 3. Therefore, the rout-
ing ratio is the same for the path from u to q with both algorithms. Because
the Enlarging step does not occur, we get that the length of P (u, q) is at most
(1 + 4/s + 2/(s 2))|uq|.
Since the diameter of the enclosing circle of Bu is at most (2/s)|pq| from
Lemma 13 and p is in Bu , we have |up| (2/s)|pq|. By the triangle inequality,
we get that |uq| |up| + |pq| (2/s)|pq| + |pq| = (1 + 2/s)|pq|.
Let Ppu be the subpath from p to u of P (p, q) and Puq be the subpath
from u to q of P (p, q). We then get that the length of the path is at most
|Ppu | + |Puq |, which in turn is at most (4/(s 2))|pq| + (1 + 4/s + 2/(s 2))|uq|
(4/(s 2))|pq| + (1 + 4/s + 2/(s 2))(1 + 2/s)|pq|, which simplies to (8/s2 +
6/s + 4/(s(s 2)) + 6/(s 2) + 1)|pq|.


References
1. Bonichon, N., Bose, P., De Carufel, J.-L., Perkovic, L., van Renssen, A.: Upper and
lower bounds for online routing on Delaunay triangulations. In: ESA. pp. 203214
(2015)
2. Bose, P., Devroye, L., Loer, M., Snoeyink, J., Verma, V.: Almost all Delaunay
triangulations have stretch factor greater than pi/2. Comput. Geom. 44(2), 121
127 (2011)
3. Bose, P., Fagerberg, R., van Renssen, A., Verdonschot, S.: Optimal local routing on
Delaunay triangulations dened by empty equilateral triangles. SIAM J. Comput.
44(6), 16261649 (2015)
4. Bose, P., Morin, P.: Online routing in triangulations. SIAM J. Comput. 33(4),
937951 (2004)
5. Callahan, P.B., Kosaraju, S.R.: Faster algorithms for some geometric graph prob-
lems in higher dimensions. In: SODA. pp. 291300 (1993)
6. Callahan, P.B., Kosaraju, S.R.: A decomposition of multidimensional point sets
with applications to k-nearest-neighbors and n-body potential elds. J. ACM 42(1),
6790 (1995)
7. Chew, P.: There is a planar graph almost as good as the complete graph. In: SOCG.
pp. 169177 (1986)
8. Dobkin, D.P., Friedman, S.J., Supowit, K.J.: Delaunay graphs are almost as good
as complete graphs. Discrete Comput. Geom. 5, 399407 (1990)
9. Kaplan, H., Mulzer, W., Roditty, L., Seiferth, P.: Routing in unit disk graphs. In:
LATIN. pp. 536548 (2016)
10. Keil, J.M., Gutwin, C.A.: Classes of graphs which approximate the complete eu-
clidean graph. Discrete Comput. Geom. 7, 1328 (1992)
11. Misra, S., Woungang, I., Misra, S.C.: Guide to Wireless Sensor Networks. Springer
Publishing Company, Incorporated, 1st edn. (2009)
12. Narasimhan, G., Smid, M.H.M.: Geometric spanner networks. Cambridge Univer-
sity Press (2007)
13. Xia, G.: The stretch factor of the Delaunay triangulation is less than 1.998. SIAM
J. Comput. 42(4), 16201659 (2013)
 Relaxing the Irrevocability Requirement for
Online Graph Algorithms

Joan Boyar1 , Lene M. Favrholdt1 , Michal Kotrbck2 , and Kim S. Larsen1

1
University of Southern Denmark, Odense, Denmark,
{joan,lenem,kslarsen}@imada.sdu.dk
2
The University of Queensland, Brisbane, Australia, [email protected]

Abstract. Online graph problems are considered in models where the

irrevocability requirement is relaxed. Motivated by practical examples
where, for example, there is a cost associated with building a facility
and no extra cost associated with doing it later, we consider the Late
Accept model, where a request can be accepted at a later point, but
any acceptance is irrevocable. Similarly, we also consider a Late Reject
model, where an accepted request can later be rejected, but any rejection
is irrevocable (this is sometimes called preemption). Finally, we consider
the Late Accept/Reject model, where late accepts and rejects are both
allowed, but any late reject is irrevocable. For Independent Set, the Late
Accept/Reject model is necessary to obtain a constant competitive ratio,
but for Vertex Cover the Late Accept model is sucient and for Mini-
mum Spanning Forest the Late Reject model is sucient. The Matching
problem has a competitive ratio of 2, but in the Late Accept/Reject
model, its competitive ratio is 32 .

1 Introduction
For an online problem, the input is a sequence of requests. For each request,
the algorithm has to make some decision without any knowledge about possible
future requests. Often (part of) the decision is whether to accept or reject the
request and the decision is usually assumed to be irrevocable. However, many
online problems have applications for which total irrevocability is not inher-
ent or realistic. Furthermore, when analyzing the quality of online algorithms,
relaxations of the irrevocability constraint often result in dramatically dierent
results, especially for graph problems. This has already been realized and several
papers study various relaxations of the irrevocability requirement. In this paper
we initiate a systematic study of the nature of irrevocability and of the implica-
tions for the performance of the algorithms. Our aim is to understand whether
it is the absence of knowledge of the future or the irrevocability restrictions on
the manipulation of the solution set that makes an online problem dicult.
We consider graph problems and focus on four classical problems, Indepen-
dent Set, Matching, Vertex Cover, and Minimum Spanning Forest. Independent

Supported in part by the Danish Council for Independent Research, Natural Sci-
ences, grant DFF-1323-00247, and the Villum Foundation, grant VKR023219.

Springer International Publishing AG 2017 217

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 217228, 2017.
DOI: 10.1007/978-3-319-62127-2_19
218 J. Boyar et al.

Set and Vertex Cover are studied in the vertex arrival model. In this model,
vertices arrive one by one together with all the edges between the newly arrived
vertex and previous vertices. Matching and Minimum Spanning Forest are stud-
ied in the edge arrival model, but the results hold in the vertex arrival model as
well. In the edge arrival model, edges arrive one by one, and if a vertex incident
with the newly-arrived edge was not seen previously, it is also revealed.

Relaxed irrevocability
For the four problems considered in this paper, the online decision is whether to
accept or reject the current request. In the standard model of online problems,
this decision is irrevocable and has to be made without any knowledge about
possible future requests. We relax the irrevocability requirement by allowing
the algorithm to perform two additional operations, namely late accept and late
reject. Late accept allows the algorithm to accept not only the current request but
also requests that arrived earlier. Thus, late accept relaxes irrevocability by not
forcing the algorithm to discard the items that are not used immediately. Late
reject allows the algorithm to remove items from the solution being constructed,
relaxing the irrevocability of the decision to accept an item. When the algorithm
is allowed to perform late accept or late reject, but not both, we speak of a
Late Accept model and Late Reject model, respectively. Note that, in these two
models, the late operations are irrevocable. We also consider the situation where
the algorithm is allowed to perform both late accepts and late rejects, focusing
on the Late Accept/Reject model, where any item can be late-accepted and late-
rejected, but once it is late-rejected, this decision is irrevocable. In other words,
if the algorithm performs both late accept and late reject on a single item, the
late accept has to precede the late reject.
We believe that the Late Accept, Late Reject, and Late Accept/Reject models
are appropriate modeling tools corresponding to many natural settings. Match-
ing, for example, in the context of online gaming or chats, functions in the Late
Accept model. Indeed, the users are in the pool until assigned, allowing the late
accept, but once the users are paired, the connection should not be broken by
the operator. Note that the matching problem is a maximization problem. For
minimization problems, accepting a request may correspond to establishing a re-
source at some cost. Often there is no natural reason to require the establishment
to happen at a specic time. Late acceptance was considered for the dominating
set problem in [2], which also contains further feasible practical applications and
additional rationale behind the model.
When the knapsack problem is studied in the Late Reject model, items are
usually called removable; see for example [16, 13, 12, 4, 14]. For most other prob-
lems, late rejection is usually called preemption and has been studied in variants
of many online problems, for example call control [1, 9], maximum coverage [23,
22], and weighted matching problems [6, 7]. Preemption was also previously con-
sidered for one of the problems we consider here, independent set, in [19], but in
a model where advice is used, presenting lower bounds on the amount of advice
necessary to achieve given competitive ratios in a stated range.
Relaxing the Irrevocability Requirement for Online Graph Algorithms 219

Online Vertex Cover was studied in [5], where they considered the possibility
of swapping some of the accepted vertices for other vertices at the very end, at
some cost depending on the number of vertices involved. A similar concept is
studied in, for example, [15, 21, 10, 11] for online Steiner tree problems, MST,
and TSP. Here, replacing an accepted edge with another is allowed, and the
objective is to minimize the number of times this occurs while obtaining a good
competitive ratio. The problem is said to allow rearrangements or recourse. TSP
has also been studied [17] in a model where the actual acceptances and rejections
(rejections carry a cost) are made at any time.

Competitive analysis

For each graph problem, we study online algorithms in the standard, Late Ac-
cept, Late Reject, and Late Accept/Reject models using the standard tool of
competitive analysis [24, 18], where the performance of an online algorithm is
compared to the optimum algorithm Opt via the competitive ratio. For any
algorithm (online or oine), A, we let A() denote the value of the objective
function when A is applied to the input sequence .
For minimization problems, we say that an algorithm, Alg, is c-competitive,
if there exists a constant such that, for all inputs , Alg() c Opt() + .
Similarly, for maximization problems, Alg is c-competitive, if there exists a
constant such that, for all inputs , Opt() c Alg() + . In both cases, if
the inequality holds for = 0, the algorithm is strictly c-competitive. The (strict)
competitive ratio of Alg is the inmum over all c such that Alg is (strictly)
c-competitive. The competitive ratio of a problem P is the inmum over the
competitive ratio of all online algorithms for the problem. For all combinations
of the problem and the model, we obtain matching lower and upper bounds on
the competitive ratio.
For ease of notation for our results, we adopt the following conventions to
express that a problem essentially has competitive ratio n, i.e., it is true up to
an additive constant. We say that a problem has competitive ratio n (1) if
(i) for any algorithm, there is a constant b > 0 such that the strict competitive
ratio is at least n b, and (ii) for any constant b, there is a strictly (n b)-
competitive algorithm for graphs with at least b + 1 vertices. Similarly, we say
that a problem has competitive ratio n/(1) if (i) for any algorithm, there is a
constant b > 0 such that the strict competitive ratio is at least n/b, and (ii) for
any constant b, there is an n/b-competitive algorithm for graphs with at least
b vertices. This notation is used in Theorems 3 and 12. For all other results,
the upper bounds hold for the strict competitive ratio. For convenience, when
stating results containing both an upper bound on the strict competitive ratio
and a lower bound on the competitive ratio, we use the term competitive ratio
even though the result holds for the strict competitive ratio as well.
220 J. Boyar et al.

Our results

The paper shows that for some problems the Late Accept model allows for algo-
rithms with signicantly better competitive ratios, while for others it is the Late
Reject model which does. For other problems, the Late Accept/Reject model is
necessary to get these improvements. See Table 1. Note that only deterministic
algorithms are considered, not randomized algorithms.
Our results on Minimum Spanning Forest follow from previous results. Thus,
they are mainly included to give an example where late rejects bring down
the competitive ratio dramatically. The technical highlights of the paper are
the results for Independent Set in the Late Accept/Reject model, where, in
Theorems 4 and 5, we prove matching lower and upper bounds of 3 3/2 on the
competitive ratio.

Table 1. Competitive ratios of the four problems in each of the four models. W is the
ratio of the largest weight to the smallest.

Problem Standard Late Accept Late Reject Late Accept/Reject

n
3 3
Independent Set n1 n
(1) 2 2
2.598
3
Matching 2 2 2 2
Vertex Cover n1 2 n (1) 2
Min. Spanning Forest W W 1 1

We consider only undirected graphs G = (V, E). Throughout the paper, G

will denote the graph under consideration, and V and E will denote its vertex
and edge set, respectively. Moreover, n = |V | will always denote the number of
vertices in G. We use uv for the undirected edge connecting vertices u and v, so
vu denotes the same edge.
The missing proofs all appear in the full paper [3].

2 Independent Set

An independent set for a graph G = (V, E) is a subset I V such that no two

vertices in I are connected by an edge. For the problem called Independent Set,
the objective is to nd an independent set of maximum cardinality. We consider
online Independent Set in the vertex arrival model.
In the standard model, no online algorithm can be better than (n 1)-
competitive, since the adversary can give independent vertices, until the algo-
rithm accepts a vertex, v, and then only give vertices adjacent to v. On the other
hand, the greedy algorithm is (n 1)-competitive; if it ever rejects a vertex, the
graph contains at least one edge and hence, Opt accepts at most n 1 vertices.
Relaxing the Irrevocability Requirement for Online Graph Algorithms 221

Theorem 1. For Independent Set in the standard model, the strict competitive
ratio is n 1.

In the Late Reject model, the greedy algorithm becomes n/2-competitive

with the following modication. If a new vertex is adjacent to exactly one ac-
cepted vertex, v, then v is rejected and the new vertex is accepted. If the algo-
rithm accepts only one vertex, the graph has a path containing all n vertices and
Opt can accept at most  n2  vertices. For the lower bound, the adversary can
give a bipartite graph by always connecting the new vertex to the only vertex
(if any) accepted by the algorithm.

Theorem 2. For Independent Set in the Late Reject model, the strict competi-
tive ratio is n/2.

Allowing late accepts helps further, but not enough to obtain a nite (con-
stant) competitive ratio. For a given positive constant c, an algorithm which does
not accept any vertex until the presented graph has an independent set of size at
least c, and then accepts any such set, is n/c-competitive. For the lower bound,
if the adversary starts the input sequence with isolated vertices, any algorithm
with a bounded competitive ratio will have to accept a vertex, v, at some point.
From this point on, the adversary can give vertices with only v as a neighbor. If
v was the cth vertex in the input, the algorithm can accept at most the rst c
vertices.

Theorem 3. For Independent Set in the Late Accept model, the competitive
ratio is n/(1).

The following two theorems show that, in the Late Accept/Reject model,
the optimal competitive ratio for Independent set is 3 3/2. The upper bound
comes from a variant of the greedy algorithm, Algorithm 1, rejecting a set of
vertices if it can be replaced by a set at least 3 as large. The algorithmic
idea
is natural and has been used before (with other parameters than 3) in [22, 23],
for example. Thus, the challenge lies in deciding the parameter and proving the
resulting competitive ratio. Pseudocode for Algorithm 1 is given below.
For Algorithm 1, we introduce the following notation. Let S be the current
set of vertices that have been accepted and not late-rejected. Let R be the set
of vertices that have been late-rejected, and let P denote the set V (R S) of
vertices that have not been accepted (and, hence, not late-rejected).
For a set U of vertices, let N (U ) = vU N (v), where N (v) is the neighbor-
hood of a vertex v (not including v). We call a set, T , of vertices admissible if
all the following conditions are satised:

1) T is an independent set;
2) T P;
3) |T | 3 |N (T ) S|.
222 J. Boyar et al.

Algorithm 1: Algorithm for Independent Set in the Late Accept/Reject

model.
Result: Independent set S
1 S=
2 while a vertex v is presented do
3 if S {v} is independent then
4 S = S {v}
5 else
6 while there exists an admissible set do
7 Let T be an admissible set minimizing |S N (T )|
8 S = (S N (T )) T
For the analysis of Algorithm 1, we partition S into the set, A, of vertices
accepted in line 4 and the set, B, of vertices accepted in line 8. We let O be
the independent set constructed by Opt. For any set, U , of vertices, we let
U + = U O and U = U O. Thus, O = P + S + R+ = P + A+ B + R+ .
The following lemma follows fromthe fact that each time a set, X, of vertices
is moved from S to R, a set at least 3 times as large as X is added to B.

Lemma 1. When Algorithm 1 terminates, |B| ( 3 1)|R|.

The next lemma follows from the facts that when the algorithm terminates,
P + is not admissible and P + S + is independent, since P + S + O.

Lemma 2. When Algorithm 1 terminates, |P + | < 3 |S |.

Lemma 3. When Algorithm 1 terminates, |B | + |R | 3 |R+ |.

Proof. Consider a set, T , added to B in line 8. Let Q = N (T ) S. We prove

that

|T | 3|Q+ | (1)

If |Q+ | = 0, this is trivially true. Thus, we can assume that Q is a proper

subset of Q. Since T is admissible, it follows that

|T | 3|Q| (2)

Note that (S Q ) T + is independent, since (S Q) T is independent and

there are no edges between Q+ and T + . Since the algorithm chooses T such that
|Q| is minimized, this means that

|T + | < 3|Q | (3)

Subtracting Ineq. (3) from Ineq. (2), we obtain Ineq. (1).

Let T1 , T2 , . . . , Tk be all the admissible sets that are chosen in line 8 during
the run of the algorithm, and let Q1 , Q2 , . . . , Qk be the corresponding sets that
Relaxing the Irrevocability Requirement for Online Graph Algorithms 223

are removed from S. Then, ki=1 Ti B R, and thus, ki=1 Ti B R .

Furthermore, R = ki=1 Qi . Hence,


k 
k

|B | + |R | |Ti | 3|Q+
i |= 3|R+ |,
i=1 i=1

where the second inequality follows from Ineq. (1).



Using ( 3 + 1)(|B + | + |R+ |) = 3(|B + | + |R+ |) + |B| + |R| (|B | + |R |),
we obtain the following lemma via simple calculations using Lemmas 1 and 3.

Lemma 4. When Algorithm 1 terminates, |B + | + |R+ | 3 |B + | 3
3+1
+ 2 |B|.

The upper bound now follows from simple calculations using Lemmas 2 and 4:
Theorem 4. For Independent Set in the Late Accept/Reject model, Algorithm 1
is strictly 3 3/2-competitive.
We prove a matching lower bound:

Theorem 5. For Independent

Set in the Late Accept/Reject model, the compet-
itive ratio is at least 3 3/2.

Proof (Sketch of a proof.). Assume that Alg is strictly c-competitive for some
c > 1. We rst show that c is at least 3 3/2 and then lift the strictness restric-
tion. Assume for the sake of contradiction that c < 3 3/2.
Incrementally, we construct an input consisting of a collection of bags, where
each bag is an independent set. Whenever a new vertex v belonging to some
bag B is given, we make it adjacent to every vertex not in B, except vertices
that have been late-rejected by Alg. Thus, if Alg accepts v, it cannot hold any
vertex in any other bag. This implies that the currently accepted vertices of Alg
always form a subset of a single bag, which we refer to as Algs bag, and this is
the crucial invariant in the proof. We say that Alg switches when it rejects the
vertices of its current bag and accepts vertices of a dierent bag.
For the incremental construction, the rst bag is special in the sense that
Alg cannot switch to another bag. We discuss later when we decide to create
the second bag, but all we will need is that the rst bag is large enough. From
the point where we have created a second bag, Alg has the option of switching.
Whenever Alg switches to a bag, B  , we start the next bag, B  . All that this
means is that the vertices we give from this point on and until Alg switches
bag again belong to B  , and Alg never holds vertices in the newest bag.
Now we argue that as long as we keep giving vertices, Alg will repeatedly
have to switch bag in order to be c-competitive. Choose some > 0, let B
be Algs bag, B  be the new bag, and s be the number of vertices which are
not adjacent to any vertices in B  . If Alg has accepted a vertices of B after
(c + )a s vertices of the new bag B  have been given, Alg has to accept at
least one additional vertex to be c-competitive, since at this point Opt could
accept all of the vertices in B  and s additional vertices. Since B  is the new bag,
224 J. Boyar et al.

B has reached its nal size, so eventually Alg will have to switch to a dierent
bag.
For the proof, we keep track of relevant parts of the behavior of Alg using
a tree structure. The rst bag is the root of the tree. Recall that whenever Alg
switches to a bag, say X, we start a new bag Y . In our tree structure we make
Y a child of X.
Since Alg is c-competitive and always holds vertices only from a single bag
B, the number a of vertices held in B satises a |B|/c. Since, by assumption,
c < 3, it follows that Alg can accept and then reject disjoint sets of vertices
of B at most twice, or equivalently, that each bag in the tree has at most two
children. As we proved above, Alg will have to keep switching bags, so if we
keep giving vertices, this will eventually lead to leaves arbitrarily far from the
root.
Consider a bag Bm that Alg holds after a long enough sequence has been
presented. Label the bags from the root to Algs bag by B1 , . . . , Bm , where
Bi+1 is a child of Bi for each i = 1, . . . , m 1. Let aj , 1 j < m, be the number
of vertices of Bj held by Alg immediately before it rejected already accepted
vertices from Bj for the rst time and let am be the number of vertices currently
accepted in Bm . Let nj = |Bj |, 1 j m.
Furthermore, for each j, if j is even, let sj = a2 + a4 + + aj , and if j is
odd, let sj = a1 + a3 + + aj . Note that our choice of adjacencies between bags
implies that Opt can hold at least sj vertices in bags B1 , B2 , . . . , Bj .
Thus, just before Alg rejects the vertices in Bj1 (just before the nj th vertex
of Bj is given), we must have caj1 nj 1 + sj2 , by the assumption that
Alg is c-competitive. We want to introduce the arbitrarily small chosen above
and eliminate the 1 in this inequality: Since Opt can always hold the a1
vertices from the root bag, caj a1 must hold for all j. Since a1 (n1 1)/c,
we get that aj (n1 1)/c2 . Thus, at the beginning of the input sequence, we
can keep giving vertices for the rst bag, making n1 large enough such that aj
becomes large enough that aj1 1. This establishes (c + )aj1 nj + sj2 .
Trivially, nj aj , so
(c + )aj1 sj2 aj . (4)

Next, we want to show that for any 1 c < 3 3/2, there exists an m such that
sm > cam , (5)
contradicting the assumption that Alg was c-competitive. To accomplish this,
we repeatedly strengthen Ineq. (5) by replacing aj with the bound from Ineq. (4),
eventually arriving at an inequality which can be proven to hold, and then this
will imply all the strengthened inequalities and, nally, Ineq. (5).


3 Matching
A matching in a graph G = (V, E) is a subset of E consisting of pairwise non-
incident edges. For the problem called Matching, the objective is to nd a match-
ing of maximum cardinality. We study online Matching in the edge arrival model,
Relaxing the Irrevocability Requirement for Online Graph Algorithms 225

but note that the results hold in the vertex arrival model as well: For the upper
bounds, an algorithm in the vertex arrival model can process the edges incident
to the current vertex in any order. For the lower bounds, all adversarial sequences
used in this section consist of paths, and hence, exactly the same input can be
given in the vertex arrival model.
It is well known and easy to prove that the greedy algorithm which adds an
edge to the matching whenever possible is 2-competitive and this is optimal in
the standard model. The rst published proof of this is perhaps in the classical
paper of Korte and Hausmann [20].
For late accept, we can use the same adversarial sequence as for the standard
model: First a number of isolated edges are presented. Then, for each edge, uv,
accepted by the algorithm, two edges, xu and vy, are presented.

Theorem 6. For Matching in the Late Accept model, the competitive ratio is 2.

Late rejects do not improve the competitive ratio either. This can be seen
from a sequence starting with a number of isolated edges. For each accepted edge,
uv, the adversary presents an edge vx. If the algorithm late-rejects uv (and thus
accepts vx), an edge xy is presented. Otherwise, an edge zu is presented.

Theorem 7. For Matching in the Late Reject model, the competitive ratio is 2.

In the Late Accept/Reject model, the competitive ratio is 3/2. Again, the
adversarial sequence starts with a number of isolated edges. If an edge uv is
accepted at any point, the adversary presents edges xu and vy. If uv is then
late-rejected, edges x x and yy  are presented.

Theorem 8. For Matching in the Late Accept/Reject model, the competitive

ratio is at least 3/2.

To prove a matching upper bound, we give an algorithm, Algorithm 2, which

is strictly 32 -competitive in the Late Accept/Reject model.
Recall that for a matching M , a path P = e1 , . . . , ek is alternating with
respect to M , if for all i {1, . . . , k}, ei belongs to M if and only if i is even.
Moreover, an alternating path P is called augmenting if neither endpoint of P
is incident to a matched edge. Note that the symmetric dierence of a matching
M and an augmenting path with respect to M is a matching of size larger than
M . We focus on local changes, called short augmentations in [26].
Algorithm 2: Algorithm for maximal matching in the Late Accept/Reject
model.
Result: Matching M
1 M =
2 while an edge e is presented do
3 if M {e} is a matching then
4 M = M {e}
5 if there is an augmenting path xuvy of length 3 then
6 M = (M {ux, vy}) {uv}
226 J. Boyar et al.

The fact that Algorithm 2 is a Late Accept/Reject algorithm follows from

the observation that no matched vertex ever becomes unmatched again. For the
upper bound, we use that if a maximal matching M does not admit augmenting
paths of length 3, then 3|M | 2|OP T |. This fact is easy to prove and can be
found as Lemma 2 of [8], for example.

Theorem 9. For Matching in the Late Accept/Reject model, Algorithm 2 is

strictly 3/2-competitive.

4 Vertex Cover

A vertex cover for a graph G = (V, E) is a subset C V such that for any edge,
uv E, {u, v} C
= . For the problem called Vertex Cover, the objective is
to nd a vertex cover of minimum cardinality. We study online Vertex Cover in
the vertex arrival model.
In the standard model, no online algorithm can be better than (n 1)-
competitive: The adversary can present isolated vertices until some vertex, v,
is rejected, and then vertices that are adjacent only to v. On the other hand,
this competitive ratio is obtained by the algorithm that accepts only the second
endpoint of each uncovered edge.

Theorem 10. For Vertex Cover in the standard model, the strict competitive
ratio is n 1.

Late accept changes the situation dramatically, since then the 2-approx-
imation algorithm adding both endpoints of each uncovered edge can be used.
Adding late rejects does not change the situation further; if the algorithm ever
late-rejects a vertex, v, the adversary can add arbitrarily many neighbors of v.

Theorem 11. For Vertex Cover in the Late Accept model and the Late Ac-
cept/Reject model, the competitive ratio is 2.

In the Late Reject model, the competitive ratio is n (1). For the lower
bound, the adversary can give isolated vertices until some vertex, v, is rejected,
and then arbitrarily many neighbors of v. The upper bound is obtained by a
family Algb of algorithms accepting the rst b + 1 vertices and then rejecting
the vertices not part of an optimal vertex cover of the graph seen so far.

Theorem 12. For Vertex Cover in the Late Reject model, the competitive ratio
is n (1).

5 Minimum Spanning Forest

A spanning forest for a graph G = (V, E) is a subset T E which forms

a spanning tree on each of the connected components of G. Given a weight
function w : E R+ , the objective of the Minimum Spanning Forest problem
Relaxing the Irrevocability Requirement for Online Graph Algorithms 227

is to nd a spanning forest of minimum total weight. We let W denote the ratio

between the largest and the smallest weight of any edge in the graph.
We study online Minimum Spanning Forest in the edge arrival model, but the
results also hold in the vertex arrival model: For the upper bounds, an algorithm
in the vertex arrival model can process the edges incident to the current vertex
in any order. In the lower bound sequences presented here, all edges from a new
vertex to all previous vertices are presented together in an arbitrary order.
Even in the standard model, the competitive ratio cannot be higher than W ,
since all spanning forests contain the same number of edges. A matching lower
bound follows from the sequence consisting of a tree of edges of weight W and
then a vertex with edges of weight 1 to all previous vertices. Since an online
algorithm, even in the Late Accept model model, does not know when the input
ends, it must always have a forest spanning all the vertices seen so far:
Theorem 13. For Minimum Spanning Forest in the standard model or the Late
Accept model, the competitive ratio is W .
On the other hand, in the Late Reject model, the greedy online algorithm
mentioned by Tarjan in [25] can be used: Each new edge is accepted, and if
this results in a cycle, the heaviest edge in the cycle is (late-)rejected. Since the
Late Reject model leads to an optimal spanning tree, any model allowing that
possibility inherits the result.
Theorem 14. For Minimum Spanning Forest in the Late Reject model or the
Late Accept/Reject model, the competitive ratio is 1.

Future Work
Since we prove tight results for all combinations of problems and models con-
sidered, we leave no immediate open problems. However, one could reasonably
consider late operations a resource to be used sparingly, as for the rearrange-
ments in [15, 21, 10, 11], for example. Thus, an interesting continuation of our
work would be a study of trade-os between the number of late operations em-
ployed and the quality of the solution (in terms of competitiveness). Obviously,
one could also investigate other online problems and further model variations.

References
1. Y. Bartal, A. Fiat, and S. Leonardi. Lower bounds for on-line graph problems with
application to on-line circuit and optical routing. In 28th STOC, pages 531540.
ACM, 1996.
2. J. Boyar, S.J. Eidenbenz, L.M. Favrholdt, M. Kotrbck, and K.S. Larsen. Online
dominating set. In 15th SWAT, volume 53 of LIPIcs, pages 21:121:15. Schloss
Dagstuhl Leibniz-Zentrum fur Informatik GmbH, 2016.
3. J. Boyar, L.M. Favrholdt, M. Kotrbck, and K.S. Larsen. Relaxing the irrevocabil-
ity requirements for online graph algorithms. Technical Report arXiv:1704.08835
[cs.DS], arXiv, 2017.
228 J. Boyar et al.

4. M. Cygan, L  . Jez, and J. Sgall. Online knapsack revisited. Theor. Comput. Syst.,
58(1):153190, 2016.
5. M. Demange and V.Th. Paschos. On-line vertex-covering. Theor. Comput. Sci.,
332:83108, 2005.
6. L. Epstein, A. Levin, J. Mestre, and D. Segev. Improved approximation guarantees
for weighted matching in the semi-streaming model. SIAM J. Discrete Math.,
25(3):12511265, 2011.
7. L. Epstein, A. Levin, D. Segev, and O. Weimann. Improved bounds for online
preemptive matching. In 30th STACS, volume 20 of LIPIcs, pages 389399. Schloss
Dagstuhl Leibniz-Zentrum fur Informatik GmbH, 2013.
8. J. Feigenbaum, S. Kannan, A. McGregor, S. Suri, and J. Zhang. On graph problems
in a semi-streaming model. Theor. Comput. Sci., 348(23):207216, 2005.
9. J.A. Garay, I.S. Gopal, S. Kutten, Y. Mansour, and M. Yung. Ecient on-line call
control algorithms. J. Algorithm., 23(1):180194, 1997.
10. A. Gu, A. Gupta, and A. Kumar. The power of deferral: Maintaining a constant-
competitive steiner tree online. SIAM J. Comput., 45(1):128, 2016.
11. A. Gupta and A. Kumar. Online steiner tree with deletions. In 25th SODA, pages
455467, 2014.
12. X. Han, Y. Kawase, and K. Makino. Randomized algorithms for online knapsack
problems. Theor. Comput. Sci., 562:395405, 2015.
13. X. Han, Y. Kawase, K. Makino, and H. Guo. Online removable knapsack problem
under convex function. Theor. Comput. Sci., 540:6269, 2014.
14. X. Han and K. Makino. Online minimization knapsack problem. Theor. Comput.
Sci., 609:185196, 2016.
15. M. Imase and B.M. Waxman. Dynamic steiner tree problem. SIAM J. Discrete
Math., 4(3):369384, 1991.
16. K. Iwama and S. Taketomi. Removable online knapsack problems. In 29th ICALP,
volume 2380 of LNCS, pages 293305. Springer, 2002.
17. P. Jaillet and X. Lu. Online traveling salesman problems with rejection options.
Networks, 64:8495, 2014.
18. A.R. Karlin, M.S. Manasse, L. Rudolph, and D.D. Sleator. Competitive snoopy
caching. Algorithmica, 3:79119, 1988.
19. D. Komm, R. Kralovic, R. Kralovic, and C. Kudahl. Advice complexity of the
online induced subgraph problem. In 41st MFCS, volume 58 of LIPIcs, pages
59:159:13. Schloss Dagstuhl - Leibniz-Zentrum fur Informatik, 2016.
20. B. Korte and D. Hausmann. An analysis of the greedy heuristic for independence
systems. Ann. Discrete Math., 2:6574, 1978.
21. N. Megow, M. Skutella, J. Verschae, and A. Wiese. The power of recourse for
online MST and TSP. SIAM J. Comput., 45(3):859880, 2016.
22. D. Rawitz and A. Rosen. Online Budgeted Maximum Coverage. In 24th ESA,
volume 57 of LIIPCcs, pages 73:173:17. Schloss Dagstuhl Leibniz-Zentrum fur
Informatik GmbH, 2016.
23. B. Saha and L. Getoor. On maximum coverage in the streaming model & appli-
cation to multi-topic blog-watch. In 9th SDM, pages 697708. SIAM, 2009.
24. D.D. Sleator and R.E. Tarjan. Amortized eciency of list update and paging rules.
Communications of the ACM, 28(2):202208, 1985.
25. R.E. Tarjan. Data Structures and Network Algorithms, volume 44 of CBMS-NSF
regional conference series in applied mathematics. SIAM, 1983.
26. D.E.D. Vinkemeier and S. Hougardy. A linear-time approximation algorithm for
weighted matchings in graphs. ACM T. Algorithms, 1(1):107122, 2005.
 Approximating Small Balanced Vertex
Separators in Almost Linear Time

Sebastian Brandt and Roger Wattenhofer

ETH Zurich, Zurich, Switzerland,

[email protected],
[email protected]

Abstract. For a graph G with n vertices and m edges, we give a ran-

domized Las Vegas algorithm that approximates a small balanced vertex
separator of G in almost linear time. More precisely, we show the fol-
lowing, for any 23 < 1 and any 0 < < 1 : If G contains an
-separator of size K, then our algorithm nds an ( + )-separator of
size O(1 K 2 log1+o(1) n) in time O(1 K 3 m log2+o(1) n) w.h.p. In par-
ticular, if K O(polylog n), then we obtain an ( + )-separator of
size O(1 polylog n) in time O(1 m polylog n) w.h.p. The presented
algorithm does not require knowledge of K.
Due to space restrictions, no proofs are included in this version of the
paper; the full version with a lot of additional material can be found at
http://disco.ethz.ch/publications/wads2017-vertexsep.pdf.

1 Introduction
Motivation In order to solve a large computational problem, the problem is
typically divided into smaller parts, and each part is solved on a single processor,
in parallel. Some problems can be chopped into pieces in a straightforward way,
e.g., using MapReduce or Spark. Other computational problems cannot be par-
titioned easily. Such dicult problems can frequently be represented as graphs:
Each vertex represents some piece of work whereas an edge between two vertices
denotes a relation between the two pieces, i.e., change at one vertex will directly
aect the other (and possibly vice versa). There are dozens of software packages
for distributed graph processing, e.g., Googles Pregel or PowerGraph.
In order to use multiple processors, the input graph has to be partitioned
into multiple components, ideally of similar size. Then the vertices of a com-
ponent are simulated on a single processor whereas edges between two vertices
in dierent components are handled by the two processors responsible for the
two components by exchanging messages. A natural objective of designing such
a partition is to reduce the inter-processor communication as it is the expensive
part in terms of runtime.
We argue that an input graph should be partitioned by means of a balanced
vertex separator (and not a balanced edge cut), since vertex separators are often
more ecient. For a simple example, consider a star graph, a tree where one
single root is connected to all leaves. We want to partition the star for two

Springer International Publishing AG 2017 229

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 229240, 2017.
DOI: 10.1007/978-3-319-62127-2_20
230 S. Brandt and R. Wattenhofer

processors. A star graph does not feature a small balanced edge cut, whereas
the root is a perfectly good vertex separator. The root is simply replicated on
both processors, and communication is reduced to the exchange between the
two copies of the root vertex. In general, the computation and communication
overhead of a vertex separator is asymptotically never worse than that of a
balanced edge cut, whereas in some cases (such as the star graph) it can be a
factor of n better, where n is the number of vertices in the graph.
In the last decades, algorithms research has made a lot of progress regarding
balanced vertex separators, cf. [2, 69, 14]. To a large extent, these works focus
on the fundamental case of dividing the input graph into two parts. For the
remainder of this paper, we will also exclusively consider vertex separators that
cut the input graph into two similar-sized pieces. Even though the algorithms
given in the works cited above only need polynomial time, this is often too slow
for practical purposes, as partitioning the input graph is the only non-parallel
part of the whole process. What is needed is a quick and dirty way to compute
a balanced vertex separator, i.e., an algorithm that (apart from a polylogarithmic
factor) only reads the input once. So far, to the best of our knowledge, it is not
known how to compute a balanced vertex separator for general graphs quickly.
Our goal is to nd a reasonably small balanced vertex separator if there
exists a small balanced vertex separator, e.g., of polylogarithmic size, and we
want to achieve this in almost linear time. For a graph with n vertices and
m edges, we show the following, for any 23 < 1 and any 0 < < 1
: If the graph contains an -balanced vertex separator of size K, then our
randomized Las Vegas algorithm nds an ( + )-balanced vertex separator of
size O(1 K 2 log1+o(1) n) in time O(1 K 3 m log2+o(1) n) w.h.p. Of course, this
result can also be used for other practical applications related to balanced vertex
separators, e.g., for determining quickly if a network has serious bottlenecks and
locating them in the armative case. If no xed K is considered, by successive
doubling we can quickly reach a size K for which an -separator exists, yielding
only an additional small constant factor for the time complexity. In particular,
using this technique, our algorithm does not require knowledge of K. If, on the
other hand, the input graph does not contain a small separator, our algorithm
will report the lack thereof. Note that graphs without small vertex separators
may not be amenable to distributed graph processing in the rst place, and one
may wonder whether parallelism can speed up processing such graphs at all.

Related Work As discussed above, nding a balanced edge separator does

not yield a balanced vertex separator with a similar approximation guarantee
in general. Since there is an abundance of results regarding edge separators, we
will only mention them if they are also related to vertex separators.
Let G = (V, E) be a graph with n vertices and m edges. An -separator of
G is a triple (A, S, B) of disjoint subsets of V s.t. V = A S B, there are no
edges between A and B, and max{|A|, |B|} |V |. Its size is |S|.
The problem of nding an -separator of minimum size is NP-hard, as shown
by Bui and Jones [4]. Hence, one main focus of research in the context of balanced
Approximating Small Balanced Vertex Separators in Almost Linear Time 231

vertex separators has been to nd approximation algorithms, cf., e.g., [2,6,7]. In

their seminal paper [14], Leighton and Rao gave an O(log n)-approximation for
balanced edge separator, incurring only an arbitrarily small loss in the balance.
As they showed, their result extends to the case of directed edge separators
and thereby to vertex separators. Feige, Hajiaghayi and Lee [8] proved that a
1
3
4 -separator of size K log K can be found in polynomial time if the input graph
2

2
contains a 3 -separator of size K. Subsequently, Feige and Mahdian [9] showed,
for any 23 < 1, how to nd an -separator of size K if such a separator
exists, except when there is an ( + )-separator of smaller size in which case
they nd the latter. Their runtime is polynomial if K O(log n), for xed .
As shown by Marx [17], the problem of nding an -separator of minimum
size is even W [1]-hard. In their work [9] mentioned above, Feige and Mahdian
solve this issue by showing that the problem becomes xed parameter tractable
if the balance requirement is relaxed, obtaining a runtime of nO(1) 2O(K) which
is polynomial for K O(log n). We show that if we relax the requirements on
balance and size of the separator, then we can achieve an almost linear runtime.
The techniques used in the works above, e.g., linear or semidenite pro-
gramming, focus on achieving as good approximation ratios as possible while
having polynomial time complexity. By applying their primal-dual approach for
semidenite programs [3] to the problem of approximating minimum balanced
3 3
separators, Arora and Kale achieved a runtime of O(m 2 + n2+ ) (resp. O(m 2 )),
1
for obtaining an approximation ratio of O(log 2 n) (resp. O(log n)). Although
achieved in the context of directed edge separators, the given runtimes and ap-
proximation ratios apply directly to our problem of undirected vertex separators.
A dierent line of research consists in searching for primarily fast algorithms
that yield separators of not necessarily near-optimal size. For graph classes with
certain restrictions, there are a number of results obtaining good runtimes, of-
ten at the expense of the separator size depending polynomially on n. Gilbert,
Hutchinson and Tarjan [10] gave a linear-time algorithm for nding a 23 -separator
of size O((gn)1/2 ) where g is the genus of the given graph, thereby extending the
famous planar separator theorem by Lipton and Tarjan [15]. The same linear
runtime was achieved independently by Djidjev [5].
A further extension to graphs excluding certain minors was given by Alon,
Seymour and Thomas [1]. They showed how to nd, for a graph containing no
1 3
minor Kj for some xed integer j, a 23 -separator of size O(n 2 ) in time O(n 2 ).
Reed and Wood [19] gave an algorithm which solves the same problem in linear
2
time except that the separators are of somewhat larger size O(n 3 ). Further-
more, they showed how to trade runtime for separator size in a parametrized
way, bounded by those two results. Kawarabayashi and Reed [13] improved the
1
runtime for nding a separator of size O(n 2 ) to O(n1+ ), for any > 0, ad-
ditionally improving the dependency of the separator size on the number j of
vertices of the excluded minor. Unfortunately, the runtime depends heavily on
j, making the algorithm infeasible in practice. Wul-Nilsen [20] gave an algo-
rithm which depends only polynomially on j, at a slight expense of runtime and
separator size. Moreover, he showed how to nd, for constant c < 1 and xed
232 S. Brandt and R. Wattenhofer

j, a separator of size O(nc ) in linear time. We are not aware of any results for
general graphs (regarding balanced vertex separators) that focus on achieving a
near-linear time complexity.
As mentioned earlier, recently various software packages to handle large
graphs have been introduced, e.g., Pregel [16] or PowerGraph [11]. Some of
them include simple heuristics to partition the input graph into pieces. Power-
Graph, for instance, merely removes vertices with large degrees until the graph
falls into small enough pieces. In practice, this seems to work well on power-law
graphs, which include many interesting application areas such as, e.g., social
networks. We believe that our work will help to nd a theoretical foundation for
this practical problem while also providing an implementable solution.

Our Approach In the following, we give an overview of our approach with-

out providing formal accuracy. Exact denitions will follow in the next section.
Our approach is based on maximum s-t-ows. By the very nature of ows, it
is likely that such an approach can only nd a near-optimally sized balanced
vertex separator quickly if the considered graph actually contains a reasonably
small balanced vertex separator. As explained before, this restricted problem is
still very important in practice, thus we deem the presented approach to be a
worthwile endeavour while having the advantage of (conceptual) simplicity.
Assume we are given a graph G containing a small vertex separator and we
have vertices s and t on dierent sides of the separator. Then, by Mengers
Theorem (cf. [18]), the maximum number of pairwise vertex-disjoint s-t-paths
is also small. We start by computing a set of maximum cardinality of pairwise
vertex-disjoint s-t-paths. By using the Ford-Fulkerson algorithm (cf. [12]), this
can be done in almost linear time as such a path collection corresponds to a
maximum s-t-ow in an unweighted directed graph obtained from G by a simple
transformation. From this collection of k paths we extract s-t-vertex cuts of the
same cardinality k by taking one vertex from each path. These vertices have to
be chosen carefully in order to actually separate s and t, but the existence of the
s-t-vertex cuts is ensured, again, by Mengers Theorem. Using binary search,
we determine two of the best-balanced of all these s-t-cuts, one closer to s
and one closer to t. If one of these two cuts is suciently balanced, then we
have found the desired small balanced vertex separator. Otherwise, consider the
connected components cut o by the two s-t-cuts. We contract the two connected
components containing s and t into new vertices s and t , respectively.
All s -t -vertex cuts in the newly obtained graph are also s-t-cuts in G and
additionally better-balanced than the above two s-t-cuts. We will prove that the
maximum number of pairwise vertex-disjoint s -t -paths is larger than k (and
therefore the same is true for the cardinality of any s -t -cut corresponding to
such a path collection). We iterate the above process of nding vertex-disjoint
paths, extracting some of the best-balanced s-t-cuts and contracting vertex sets,
until we obtain s-t-cuts whose cardinality is equal to some predetermined value
K (or observe that no such cut of cardinality K exists).
Approximating Small Balanced Vertex Separators in Almost Linear Time 233

Consider an -separator of size at most K separating s and t, where 23 <

1. If the iterative process described above does not yield cuts whose (combined)
balance is at least as good as , then, as we will prove, at least one vertex of
the -separator must have been involved in one of the performed contractions.
Thus, by iterating the whole process at most K times (with newly chosen s, t in
each iteration), we obtain a balanced vertex separator (by collecting the relevant
cuts obtained in the process). We will show that if G contains a small balanced
vertex separator, then the obtained balanced vertex separator is also small.
Up to now, we assumed that we can nd vertices s and t on dierent sides
of a balanced separator. But because of the balance of the separator, this is
actually the case with a large enough probability. By choosing s and t uniformly
at random, applying the iterative process described above and then iterating the
whole procedure on the largest obtained connected component, we obtain an
almost linear runtime for nding a reasonably small balanced vertex separator,
provided the given graph contains a small balanced vertex separator.
Due to space restrictions, all proofs (which show many of the intricacies
of the presented work) are deferred to the full version, which can be found at
http://disco.ethz.ch/publications/wads2017-vertexsep.pdf.

2 Conventions and Basic Denitions

In this work, we consider simple, undirected, connected graphs G = (V, E) with-
out self-loops, with n nodes and m edges. We call a triple (A, S, B) of pairwise
disjoint subsets of V a vertex separator of G if V = A S B and there is no
{u, v} E s.t. u A, v B. We call |S| the size of (A, S, B). Let 0 < < 1.
If max{|A|, |B|} |V |, then we call (A, S, B) -balanced or, equivalently, an
-separator. Let s, t V s.t. s A, t B. Then we call (A, S, B) an s-t-vertex
separator. Let s, t V, s
= t and let {f1 , ..., fk }, k N>0 be a set of s-t-paths in
G. Then we say that f1 , ..., fk are pairwise vertex-disjoint if there are no vertices
except s and t that appear in more than one of these paths. For all subsets
X V , we denote the induced subgraph of G whose vertex set is X by G[X].
We will often consider two special vertices s, t. For the remainder of this work,
we will assume that s and t are dierent, non-adjacent vertices if not specied
otherwise. Moreover, for convenience, we will be not too technical regarding the
distinction between sets and tuples.
In 1927, Karl Menger [18] stated the following famous theorem:
Theorem 1. The maximum number of pairwise vertex-disjoint s-t-paths in a
graph G is equal to the minimum number of vertices v, s
= v
= t, which have to
be removed from G in order that there is no s-t-path in the resulting graph.
Consider a set of maximum cardinality of pairwise vertex-disjoint s-t-paths. By
Mengers Theorem, we can disconnect s from t by removing one vertex from
each of those paths. Of course, if we choose these vertices arbitrarily (but still
one per path), then it is not ensured that there is no s-t-path left. We call such
a set of arbitrarily chosen vertices a slice whereas we call it a cut if its removal
results in a disconnection of s from t. In more formal terms:
234 S. Brandt and R. Wattenhofer

Denition 2. Let G be a graph and s, t V . Let {f1 , ..., fk } be a set of pairwise

vertex-disjoint s-t-paths in G. Then we call a tuple (w1 , ..., wk ) a slice (w.r.t.
(f1 , ..., fk )) if wi fi , s
= wi
= t for all 1 i k. Let X be an arbitrary
subset of V s.t. s, t / X. If there is no s-t-path in G[V \X], then we say that X
separates s and t. We call a slice that separates s and t a cut.

Following Marx [17], the set of slices (w.r.t some xed set of s-t-paths) can
be partially ordered by their relative closeness to s. The following denition
adapts the denition of the dominance relation given in [17] to our setting.

Denition 3. Let {f1 , ..., fk } be a set of pairwise vertex-disjoint s-t-paths in G.

Let U = (u1 , ...uk ) and W = (w1 , ..., wk ) be slices w.r.t. (f1 , ..., fk ) s.t., for all
1 i k, ui is a predecessor of wi in fi or ui = wi . Then we say that U is closer
to s than W and write U  W . If additionally ui
= wi for some 1 i k, then
we say that U is strictly closer to s than W and write U W . Analogously, we
say that W is (strictly) closer to t than U . For convenience, we dene the above
analogously for the tuples (s, s, ..., s) and (t, t, ..., t). Thus we can, e.g., say that
(s, s, ..., s) is closer to s than any slice.

The removal of a cut decomposes G into at least two connected components

as s and t are not connected anymore. The component containing s and the
component containing t are of special interest to us since we will develop a
method to make them larger (by choosing better cuts) which, in turn, aids in
nding cuts (or, more precisely, s-t-vertex separators) of better balance.

Denition 4. Let U be an arbitrary cut. We dene Vs (U ) as the vertex set

of the connected component of G[V \U ] containing s, Vt (U ) as the vertex set
of the connected component of G[V \U ] containing t and Vr (U ) as the union of
the vertex sets of the remaining connected components of G[V \U ], i.e., those
containing neither s nor t (so Vr (U ) may be empty).

3 Closest Cuts
Consider a slice U . Among all cuts that are closer to t (resp. s) than U , we
would like to single the closest one out. Our partial order  provides a very
intuitive way to do so, resulting in Denition 5. A proof for the uniqueness of
such a cut is given in the full version. Lemma 6 shows that closest cuts can be
computed in linear time. For the remainder of this paper, let {f1 , ..., fk } be a
set of maximum cardinality of pairwise vertex-disjoint s-t-paths.

Denition 5. Let U be a slice w.r.t. (f1 , ..., fk ). Let X be a cut s.t. U  X and
there is no cut X  satisfying U  X  X. Then we dene U + := X. If there
exists no X as described above, then set U + := (t, t, ..., t). Analogously, let Y be
a cut s.t. Y  U and there is no cut Y  satisfying Y Y   U . Then we dene
U := Y . If there exists no Y as described above, then set U := (s, s, ..., s).

Lemma 6. We can compute U + and U in time O(m).

Approximating Small Balanced Vertex Separators in Almost Linear Time 235

Algorithm 1 Find Innermost s-sided Cut of Minimum Size

Initialization: Given weights g(s), g(t) N>0 and the maximum number of pairwise
vertex-disjoint s-t-paths f1 , ..., fk , let vij , 0 j i , be the jth vertex of the path
fi where i is the length of fi and s is considered to be the 0th vertex of every fi .
Set wi := vi1 for all 1 i k and valid := false.

1: for i = 1 to k do
2: c := 0 // indexes the start of path fi
3: d := i // indexes the end of path fi
4: while d
= c + 1 do
5: e :=  c+d
2
 // binary search on fi
6: W := (w1 , ..., wi1 , vie , wi+1 , ..., wk ) // get new slice by moving vertex on fi
7: if W +
= (t, t, ..., t) and |Vs (W + )| + g(s) |Vr (W + ) Vt (W + )| + g(t) then
8: c := e // W is suitable, continue binary search in direction towards t
9: valid := true // suitable cut found
10: else
11: d := e // W is not suitable, continue binary search in direction towards s
12: end if
13: end while
14: wi := vic // x best vertex found on fi , continue with next path
15: end for
16: if valid then
17: return (w1 , ..., wk )
18: else
19: return (s, s, ..., s)
20: end if

4 An Algorithm for Finding Good Cuts of Bounded Size

Consider a graph G which contains an s-t-vertex separator of size at most K, for

some xed s, t. The goal of this section is to nd a good cut of size at most K.
Intuitively, a cut U is good if the connected components obtained by removing
U from G can be divided into two groups in a balanced way. Unfortunately, it is
not easy to nd such a cut quickly. Thus, we relax our notion of good slightly
and say that a cut U , that does not admit a balanced partition of the connected
components, is still good if it satises the following property: For any s-t-vertex
separator (A , S , B ) of size at most K with a better balance than U , S is
not contained in any of the connected components obtained by removing U . The
idea is to iterate the process of nding a good cut on the largest component
obtained by removing the (previous) good cut and to benet from the fact that
the size of (A , S , B ) restricted to this component decreases by at least 1 in
each iteration. The details of this idea will be discussed in Section 5.
The rst step in order to design an algorithm that nds a good cut is to
develop a method for nding a cut U (if it exists) s.t. |Vs (U )| |Vr (U ) Vt (U )|
and all cuts which are strictly closer to t than U violate that property. This
can be done eciently using binary search, as given by Algorithm 1. Essentially,
236 S. Brandt and R. Wattenhofer

this algorithm moves a vertex of some initial slice closer to t along an s-t-path,
thereby obtaining a new slice, and checks if the cut closest to this slice in direction
towards t still satises the above inequality. In the armative case, it iterates
starting from this new slice, otherwise it goes back and tries another vertex.
Since, later on, we will have to deal with graphs which are the result of a series
of contractions, we design Algorithm 1 in a rather general way where weights
are assigned to s and t. Regarding output and runtime, we prove the following,
using Lemma 6 in the process:

Lemma 7. Let g(s), g(t) be positive integers. If there is a cut X s.t. |Vs (X)| +
g(s) |Vr (X) Vt (X)| + g(t), then Algorithm 1 returns such a cut X with the
additional property that |Vs (X  )| + g(s) > |Vr (X  ) Vt (X  )| + g(t) for all cuts X 
satisfying X X  . If there is no such cut, then the algorithm returns the tuple
(s, s, ..., s). In both cases Algorithm 1 terminates in time O(km log n).

For reasons of symmetry, Algorithm 1 and Lemma 7 also work if s and t are
reversed. The respective versions are given in the full paper. We will denote the
reversed algorithm and lemma by Algorithm 1 , resp. Lemma 7 .
With the tools gathered above, we are now able to design and analyze an
algorithm (Algorithm 2) which nds a pair of tuples that contains a good cut. As
we will perform contractions on a given graph G in the process, we give a short
overview of the technical details. The contraction of a subset U of V transforms
G into a graph H where V (H) := (V \U ) {u} and E(H) contains an edge
{u, w} for each edge {v, w} E satisfying v U , w V \U while all edges in G
between vertices in V \U remain edges in H. We call u the contraction of U .
Essentially, Algorithm 2 uses Algorithm 1 and Algorithm 1 as subroutines
in order to nd two cuts that cut a preferably large part containing s, resp. t,
o. Then it contracts these two parts into new nodes s and t and iterates on the
obtained graph. We show in the following that the number of pairwise vertex-
disjoint s-t-paths grows in each iteration and that the performed contractions
ensure that s and t remain non-adjacent. This enables us to bound the run-
time of Algorithm 2 (in the process using Ford and Fulkersons maximum ow
algorithm). For the proofs of both lemmas, we need Lemmas 7 and 7 .

Lemma 8. Consider an iteration of the while loop in Algorithm 2 where (not

necessarily non-trivial) contractions are performed. Then, at the end of the it-
eration, there is no edge {s, t} E(H) and there are (at least) k + 1 pairwise
vertex-disjoint s-t-paths in H.

Lemma 9. Algorithm 2 terminates in time O(K 2 m log n).

What is left to show is that the returned pair contains indeed a good cut (good
as described in the introduction of Section 4). Theorem 10 takes care of that,
using again Lemmas 7 and 7 in the proof. For ease of presentation, dene, for any
subset U of V , LU := max{|V (C)| | C is a connected component of G[V \U ]}.
Approximating Small Balanced Vertex Separators in Almost Linear Time 237

Algorithm 2 Find Innermost Cut of Bounded Size

Initialization: Given a positive integer K and two vertices s, t V , set g(s) := g(t) :=
1, k := 0, H := G and S := T := {}.

1: while the max. number of pairwise vertex-disjoint s-t-paths in H is at most K do

2: nd max. number of pairwise vertex-disjoint s-t-paths f1 , ..., fk in H, update k
3: execute Algorithm 1, let U be output // U cuts large part containing s o
4: execute Algorithm 1 , let W be output // W cuts large part containing t o
5: if U
= (s, s, ..., s) then
6: Ms := Vs (U ) U // collect vertex set which is to be contracted
7: S := U // U is best s-sided cut we found so far
8: else
9: Ms := {s} // no s-sided cut found, so nothing to contract
10: end if
11: if W
= (t, t, ..., t) then
12: Mt := Vt (W ) W // collect vertex set which is to be contracted
13: T := W // W is best t-sided cut we found so far
14: else
15: Mt := {t} // no t-sided cut found, so nothing to contract
16: end if
17: if Ms Mt
= or there is an edge from Ms to Mt in H then
18: break // contracting Ms and Mt impossible/problematic for later iterations
19: else
20: contract Ms and denote the contraction by s (and update H accordingly)
21: contract Mt and denote the contraction by t (and update H accordingly)
22: replace parallel edges of H by a single edge
23: g(s) := g(s) + |Ms | 1 // update total number of vertices contracted into s
24: g(t) := g(t) + |Mt | 1 // update total number of vertices contracted into t
25: end if
26: end while
27: return the pair (S, T )

Theorem 10. Let K be some positive integer and let s, t be two vertices of G.
Let (S, T ) be the pair returned by Algorithm 2 and suppose that there is an s-t-
vertex separator (A , S , B ) of G of size K  K. Now if there is a connected
component C of G[V \(S T )] s.t. S V (C), then one of the following holds:

(i) LS LS (ii) LT LS (iii) LS 12 |V | (iv) LT 12 |V |

5 Approximating Small Balanced Vertex Separators

Quickly

In this section, we nally design an algorithm that uses Algorithm 2 as a subrou-

tine in order to nd a small and balanced vertex separator. The idea of Algorithm
3 is based on Algorithm 2 in conjunction with Theorem 10. Let 23 < 1, and
238 S. Brandt and R. Wattenhofer

assume that there is a small -separator (A , S , B ) of G. By removing vertices

of the given graph G, we obtain dierent connected components of which we
then choose the largest one and iterate on this component. The goal is to reduce
thereby the size of the largest component to approximately |V | at most while
removing only a small number of vertices in the process.
In more detail, Algorithm 3 chooses two vertices s, t in each iteration and then
uses Algorithm 2 to nd a small number of vertices which it then removes. There
is a large enough probability that s and t are on dierent sides of the separator
whose existence we assumed above since this separator is -balanced. If s and
t are on dierent sides, then Theorem 10 ensures that the size of the largest
resulting connected component is at most |V | or that the above separator
contains strictly fewer vertices in the separating set S when restricted to this
component. By iterating, we can reduce the number of vertices in this separating
set to 0 (if the latter case occurs repeatedly) and then the balance of the above
separator ensures that the largest component is of size at most |V |.
Since the balance of the separator may decrease radically by restricting it
to the largest obtained component, so does the probability of s and t being on
dierent sides of the separator. Thus, in order to obtain a good runtime, we stop
when we achieve a balance close to . Lemma 12 formalizes some of the above
considerations. It enables us to prove Corollary 13 which gives an upper bound
for the number of so-called successful iterations, an intuitive concept dened in
the following. Using Corollary 13, we prove Lemma 14 which provides an upper
bound for the runtime of Algorithm 3 in terms of the number of iterations. Note
that the proof of Lemma 12 relies heavily on Theorem 10.
Denition 11. Let (A , S , B ) be a vertex separator. Then we call an iteration
of the while loop of Algorithm 3 unsuccessful (w.r.t. (A , S , B )) if, for the
vertices s, t chosen in that iteration, we have s, t A or s, t B . If this is
not the case, then we call the iteration successful.
Lemma 12. Let 23 < 1 and let (A , S , B ) be an -separator of G of size
at most K. Consider a successful iteration of the while loop of Algorithm 3 w.r.t.
(A , S , B ). Let H0 and H1 denote the graph H at the beginning, resp. the end,
of this iteration. Then |V (H1 ) S | |V (H0 ) S | 1 or |V (H1 )| |V |.
Corollary 13. Let 23 < 1 and let (A , S , B ) be an -separator of G of
size at most K. Then the number of successful iterations w.r.t. (A , S , B ) is
at most K in any execution of Algorithm 3.
Lemma 14. Let 23 < 1. If there exists an -separator of size at most K,
then Algorithm 3 terminates after O(1 K log1+o(1) n) iterations w.h.p.
Using Lemma 9 and Lemma 14, we are able to prove our main result. The given
bound on the separator size follows from the fact that the cardinality of the
vertex sets S and T returned by Algorithm 2 never exceeds K.
Theorem 15. Let 23 < 1 and 0 < < 1 . If G contains an -
separator of size at most K, then Algorithm 3 nds an ( + )-separator of
size O(1 K 2 log1+o(1) n) in time O(1 K 3 m log2+o(1) n) w.h.p.
Approximating Small Balanced Vertex Separators in Almost Linear Time 239

Algorithm 3 Find Small Balanced Vertex Separator

2
Initialization: Given a graph G, a positive integer K, some 3
< 1 and some
0 < < 1 , set H := G and S  := {}.

1: while |V (H)| > ( + )|V | do

2: choose two vertices s, t in H uniformly at random // can be identical/adjacent
3: if s = t or s and t are adjacent then
4: nd a largest connected component C of H[V (H)\({s} {t})] // remove s, t
5: H := C // continue on the resulting largest connected component
6: S  := S  {s} {t} // remember the removed vertices
7: else
8: execute Algorithm 2 (with input H, K, s, t) and denote the output by (S, T )
9: compute a largest connected component C of H[V (H)\(S T {s} {t})]
10: H := C // continue on the resulting largest connected component
11: S  := S  S T {s} {t} // remember the removed vertices
12: end if
13: end while
14: order the connected components C1 , C2 , ... of G[V \S  ] s.t. |V (C1 )| |V (C2 )| ...
15: A := V (C1 ) // start collecting (vertex sets of) components
16: add the vertex sets V (C2 ), V (C3 ), ... successively to A as long as the resulting A
satises |A| |V | // note that V (C1 ) could already contain ( + )|V | vertices
17: B := V (G)\(A S  ) // collect the other side of the vertex separator
18: return (A, S  , B)

By Theorem 15, we can nd a reasonably small (+)-separator in almost linear

time, provided that G contains a small -separator. In particular, we obtain:
If K O(polylog n), then Algorithm 3 nds an ( + )-separator of size
O(1 polylog n) in time O(1 m polylog n) w.h.p.
If K O(log n), then Algorithm 3 nds an ( + )-separator that has size
O(1 log3+o(1) n) in time O(1 m log5+o(1) n) w.h.p.
Throughout this work, we supposed that we have a xed K which gives us an
upper bound for the size of the -separator whose existence we assume. If we
do not want to consider some specic K, but rather nd some (ideally small) K
for which Algorithm 3 returns a separator as specied in Theorem 15, we can do
this by successively doubling K, each time executing Algorithm 3. The obtained
total runtime is asymptotically the same as the runtime of Algorithm 3.
We consider our approach to be a rst step in a new direction. We are con-
dent that future work building on this approach can improve the presented
theoretical bounds signicantly. One reason (of many) for this is the following:
One factor of K in the runtime is due to the possibility that, in each of K suc-
cessful iterations, the number of vertices in S contained in the largest connected
component potentially decreases by only 1 (see the proof of Corollary 26, full
version). We expect that graphs exhibiting such an incremental decrease must
have very specic structures that can be exploited. Moreover, from a practical
(and entirely informal) standpoint, we note that the hidden constants are fairly
240 S. Brandt and R. Wattenhofer

small and that all three factors of K in the runtime should be signicantly lower
than K on average.

References
1. Noga Alon, Paul D. Seymour, and Robin Thomas. A separator theorem for graphs
with an excluded minor and its applications. STOC 1990.
2. Eyal Amir, Robert Krauthgamer, and Satish Rao. Constant factor approximation
of vertex-cuts in planar graphs. STOC 2003.
3. Sanjeev Arora and Satyen Kale. A combinatorial, primal-dual approach to semidef-
inite programs. STOC 2007.
4. Thang Nguyen Bui and Curt Jones. Finding good approximate vertex and edge
partitions is NP-hard. Information Processing Letters, 42(3):153159, 1992.
5. H. N. Djidjev. A linear algorithm for partitioning graphs of xed genus. Serdica,
11(4):369387, 1985.
6. Guy Even, Joseph Naor, Satish Rao, and Baruch Schieber. Divide-and-conquer
approximation algorithms via spreading metrics. FOCS 1995.
7. Guy Even, Joseph Naor, Satish Rao, and Baruch Schieber. Fast approximate graph
partitioning algorithms. SODA 1997.
8. Uriel Feige, Mohammad Taghi Hajiaghayi, and James R. Lee. Improved approxi-
mation algorithms for minimum-weight vertex separators. STOC 2005.
9. Uriel Feige and Mohammad Mahdian. Finding small balanced separators. STOC
2006.
10. John R. Gilbert, Joan P. Hutchinson, and Robert Endre Tarjan. A separator
theorem for graphs of bounded genus. Journal of Algorithms, 5(3):391407, 1984.
11. Joseph E. Gonzalez, Yucheng Low, Haijie Gu, Danny Bickson, and Carlos Guestrin.
Powergraph: Distributed graph-parallel computation on natural graphs. OSDI
2012.
12. L. R. Ford Jr. and D. R. Fulkerson. Maximal ow through a network. Canadian
Journal of Mathematics, 8:399404, 1956.
13. Ken-ichi Kawarabayashi and Bruce A. Reed. A separator theorem in minor-closed
classes. FOCS 2010.
14. Frank Thomson Leighton and Satish Rao. Multicommodity max-ow min-cut
theorems and their use in designing approximation algorithms. Journal of the
ACM, 46(6):787832, 1999.
15. Richard J. Lipton and Robert E. Tarjan. A separator theorem for planar graphs.
SIAM Journal on Applied Mathematics, 36(2):177189, 1979.
16. Grzegorz Malewicz, Matthew H. Austern, Aart J. C. Bik, James C. Dehnert, Ilan
Horn, Naty Leiser, and Grzegorz Czajkowski. Pregel: a system for large-scale graph
processing. SIGMOD 2010.
17. Daniel Marx. Parameterized graph separation problems. Theoretical Computer
Science, 351(3):394406, 2006.
18. Karl Menger. Zur allgemeinen Kurventheorie. Fundamenta Mathematicae,
10(1):96115, 1927.
19. Bruce A. Reed and David R. Wood. A linear-time algorithm to nd a separator
in a graph excluding a minor. ACM Transactions on Algorithms, 5(4), 2009.
20. Christian Wul-Nilsen. Separator theorems for minor-free and shallow minor-free
graphs with applications. FOCS 2011.
 Balanced Line Separators of Unit Disk Graphs

Paz Carmi1 , Man Kwun Chiu2,3 , Matthew J. Katz1 , Matias Korman4 ,

Yoshio Okamoto5 , Andre van Renssen2,3 , Marcel Roelozen2,3 ,
Taichi Shiitada5 , and Shakhar Smorodinsky1
1
Ben-Gurion University of the Negev, Beer-Sheva, Israel
2
National Institute of Informatics, Tokyo, Japan
3
JST, ERATO, Kawarabayashi Large Graph Project, Japan
4
Tohoku University, Sendai, Japan
5
The University of Electro-Communications, Tokyo, Japan

Abstract. We prove a geometric version of the graph separator theorem

for the unit disk intersection graph: for any set of n unit disks
 in the plane
there exists a line  such that  intersects at most O( (m + n) log n)
disks and each of the halfplanes determined by  contains at most 2n/3
unit disks from the set, where m is the number of intersectingpairs of
disks. We also show that an axis-parallel line intersecting O( m + n)
disks exists, but each halfplane may contain up to 4n/5 disks. We give
an almost tight lower bound (up to sublogarithmic factors) for our ap-
proach, and also show that no line-separator of sublinear size in n exists
when we look at disks of arbitrary radii, even when m = 0. Proofs are
constructive and suggest simple algorithms that run in linear time. Ex-
perimental evaluation has also been conducted, which shows that for
random instances our method outperforms the method by Fox and Pach

(whose separator has size O( m)).

1 Introduction

Balanced separators in graphs are a fundamental tool and used in many divide-
and-conquer-type algorithms as well as for proving theorems by induction. Given
an undirected graph G = (V, E) with V as its vertex set and E as its edge set,
and a non-negative real number [1/2, 1], we say that a subset S V is an
-separator if the vertex set of G \ S can be partitioned into two sets A and
B, each of size at most |V | such that there is no edge between A and B. The
parameter determines how balanced the two sets A and B are in terms of size.

Chiu, van Renssen and Roelozen were supported by JST ERATO Grant Num-
ber JPMJER1305, Japan. Korman was supported in part by KAKENHI Nos.
12H00855 and 17K12635. Katz was partially supported by grant 1884/16 from the
Israel Science Foundation. Okamoto was partially supported by KAKENHI Grant
Numbers JP24106005, JP24220003 and JP15K00009, JST CREST Grant Number
JPMJCR1402, and Kayamori Foundation for Informational Science Advancement.
Smorodinskys research was partially supported by Grant 635/16 from the Israel
Science Foundation.

Springer International Publishing AG 2017 241

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 241252, 2017.
DOI: 10.1007/978-3-319-62127-2_21
242 P. Carmi et al.

For a balanced separator to be useful we want both the size |S| of the separator
and 1/2 to be small.
Much work has been done to prove the existence of separators with certain
properties in general sparse graphs. For example, the well-known LiptonTarjan
planar separator theorem [12] states that for any n-vertex planar graph, there
exists a 2/3-separator of size O( n). Similar theorems have been proven for
bounded-genus graphs [7], minor-free graphs [2], low-density graphs, and graphs
with polynomial expansion [17, 8]. Note that graphs in each of these graph classes
contain only O(n) edges, where n is the number of vertices.
These separator results are for abstract planar graphs and are very general.
Our focus of interest is geometric graphs, which often encode additional infor-
mation other than an adjacency matrix. Even though one can use the separator
tools in geometric graphs, often the additional information is lost in the pro-
cess. As such, a portion of the literature has focused on the search of balanced
separators that also preserve the geometric properties of the geometric graph.
Among several others, we highlight the work of Miller et al. [16], and Smith
and Wormald [18]. They considered intersection graphs of n balls in Rd and
proved that if every point in d-dimensional space is covered by at most k of the
given balls, then there exists a (d + 1)/(d + 2)-separator of size O(k 1/d n11/d )
(and such a separator can be found in deterministic linear time [4]). More inter-
estingly, the separator itself and the two sets it creates have very nice properties;
they show that there exists a (d1)-dimensional sphere that intersects at most
O(k 1/d n11/d ) balls and contains at most (d + 1)n/(d + 2) balls in its interior
and at most (d + 1)n/(d + 2) balls in its exterior. In this case, the sphere acts as
the separator (properly speaking, the balls that intersect the sphere), whereas
the two sets A and B are the balls that are inside and outside the separator
sphere, respectively. Note that the graph induced by the set A consists of the
intersection graph of the balls inside the separator (similarly, B for the balls
outside the separator and S for the balls intersecting the sphere).
We emphasize that, even though the size of the separator is larger than the
one from LiptonTarjan (specially for high values of d), the main advantage
is that the three subgraphs it creates are geometric graphs of the same family
(intersection graphs of balls in Rd ). The bound onthe separator size does not
hold up well when k is large, even for d = 2: if n disks overlap at a single
point and the other disks form a path we have k = n and m = (n), where m
is the
number of edges in the intersection graph. Hence, the separator has size
O( kn) = O(m3/4 ).
Fox and Pach [5] gave another separator result that follows the same spirit:
the intersection
graph of a set of Jordan curves in the plane has a 2/3-separator
of size O( m) if every pair of curves intersects at a constant number of points.6
A set of disks in R2 satises this condition, and thus the theorem applies to disk
graphs. Their proof can be turned into a polynomial-time algorithm. However,
we need to construct the arrangement of disks, which takes O(n2 2(n) ) time,
6
Without restriction on the number of intersection points for every pair of curves,

the bound of O( m log m) can be achieved [14].
Balanced Line Separators of Unit Disk Graphs 243

(a) (b)

Fig. 1. An example of a line separator of a unit disk graph. (a) A family of unit disks
(blue) and a line (red). (b) Removing the disks intersected by the red line leaves a
disconnected graph.

where (n) is the inverse Ackermann function [3], and in practice an ecient
implementation is non-trivial.
From a geometric perspective these two results show that, given a set of unit
disks in the plane, we can always nd a closed curve in the plane (a circle [16, 18]
and a Jordan curve [5], respectively) to partition the set. The disks intersected
by the curve are those in the separator, and the two disjoint sets are the disks
inside and outside the curve, respectively.

Results and Paper Organization. In this paper we continue the idea of geo-
metric separators and show that a balanced separator always exists, even if we
constrain the separator to be a line (see Fig. 1). Given a set of n unit disks
with
$ m pairwise distinct intersections, we show that a line 2/3-separator of size
O( (m + n) log n) can be found
in expected O(n) time, and that an axis-parallel
line 4/5-separator of size O( m + n) can be found in deterministic O(n) time.
Comparing our results with the previous work, our algorithm matches or
improves in four ways, see also Table 1. (i) simplicity of the shape: circle [16,
18] vs. Jordan curve [5] vs. our line, (ii) balance of the sets A and B: 3/4 [16,
3/4
and7 us, (iii) size of the separator: O(m ) [16, 18] vs.
18]vs. 2/3 for both [5]
O( m) [5] vs. our O( m). Finally, (iv) our algorithms are easy to implement
and asymptotically faster: O(n) [16, 18] vs. O(n2 ) [5] vs. our O(n).

Table 1. Comparison of our results with other geometric separator results.

result sep. shape balance sep. size run-time observations

[16], [18] circle 3/4 O(m3/4 ) O(n) arbitrary disks

[5] Jordan curve 2/3 O( m) O(n2 ) pseudodisks

Thm. 1 line 2/3 O( m) O(n) unit disks

Thm. 2 axis-parallel line 4/5 O( m) O(n) unit disks

[1], [13], [9] line 1/2 O(
 n log n) O(n) disjoint unit disks
[9] line 1 O( n(1 2)) O(n) disjoint unit disks

[13] axis-parallel line 9/10 O( n) O(n) disjoint unit disks

7
The O() notation suppresses sublogarithmic factors. In particular, we note that our
separator is slightly larger than the Fox-Pach separator.
244 P. Carmi et al.

We emphasize that our results focus on unit disk graphs, while the other
results hold for disk graphs of arbitrary radii, too. Indeed, if we want to separate
disks of arbitrary radii with a line, we show that the separators size may be as
large as (n). We also prove that for unit disks
$ our algorithm may fail to nd

a line 2/3-separator of size better than O( m log(n/ m)) in the worst case;
the exact statement can be found in Section 3. In this sense, the size of our
separators is asymptotically almost tight. In Section 4, experimental results are
presented. We evaluate the performance of our algorithm, compare it with the
method by Fox and Pach [5] in terms of the size of the produced separators for
random instances, and conclude that our algorithm outperforms theirs.

Other Related Work. In a dierent context, line separators of pairwise disjoint

unit disks have also been studied. Since the disks are pairwise disjoint, the inter-
section graph is trivially empty and can be easily separated. Instead, the focus is
now to nd a closed curve that intersects few disks, such that the two connected
components it denes contain roughly the same number of disks.
8
Alon et al. [1] proved that for a given set D of n pairwise disjoint
unit disks,
there exists a slope a such that every line with slope a intersects O( n log n) unit
disks of D. In particular, the halving line ofthat slope will be a nice separator
(each halfplane will have at most n/2 O( n log n) disks fully contained in).
Their proof is probabilistic, which can be turned into an expected O(n)-time
randomized algorithm [13]. A deterministic O(n)-time algorithm was afterwards
given by Homann
$ et al. [9], who also showed how to nd a line  that intersects
at most O( n/(1 2)) unit disks and each halfplane contains at most (1)n
disks (for any 0 < < 1/2). Loer and Mulzer [13] proved that there exists
an axis-parallel line  such that  intersects O( n) disks, and each halfplane
contains at most 9n/10 unit disks. For comparison purposes, these three results
are also shown in Table 1.

Preliminaries. In this paper, all disks are assumed to be closed (i.e., the bound-
aries are part of the disks), and a unit disk has radius one (thus diameter two).
For a set S of n points in R2 , there always exists a point p R2 such that every
halfplane containing p contains at least n/3 points from S. Such a point p is
called a centerpoint of S, and can be found in O(n) time [11]. Let  be a line
through a centerpoint of S. Then, each of the two closed halfplanes bounded
by  contains at least n/3 points of S, which in turn means that each of the two
open halfplanes bounded by  contains at most 2n/3 points of S each. Here, a
halfplane H (closed or open) contains a point p if p H. We also say a half-
plane H contains a disk D if D H. Due to lack of space the proofs of some
claims are deferred to the full version.

8
The result extends to pairwise disjoint fat objects that are convex and of similar area
(see Theorem 4.1 of [1]). For the sake of conciseness we only talk about unit disks.
Balanced Line Separators of Unit Disk Graphs 245

4 2 3
1 4
3
4 1 p 2 3
2
3 1 2 4
2 2 4
3 3 (a) (b)

Fig. 2. Proof of Lemma 1. Fig. 3. Proof of Lemma 2. (a) A grid of 2 2 squares
The number i in each disk is laid over the family of unit disks. (b) Disks associated
means that it intersects the to the same cell intersect, but disks associated to dier-
circle of radius 2i centered ent cells of the same color do not intersect. If j is the
at p. index for color yellow, then nj = 7, lj0 = 4, and lj1 = 1
in this example.

2 Upper Bounds

Let D be a set of n unit disks in the plane. We rst consider the case where the
disks in D are pairwise disjoint. The results from this case will also be used for
the more general case where the disks in D are not necessarily disjoint.

Lemma 1. Let D be a set of n pairwise disjoint unit disks in the plane and
let p be any point in the plane.Then the expected number of disks intersected by
a random line through p is O( n).

We note that the lemma has a avor similar to a theorem by Alon et al. [1].
Proof. Let Ci be the circle of radius 2i centered at p, for i = 0, 1, . . .. Then each
disk in D is intersected by at most two of these circlesif a disk is intersected by
two circles the intersection takes place on the boundary of the disk. Let Di D
be the set of disks that
 have non-empty intersection with Ci , for i = 0, 1, . . ..
See Fig. 2. Note that i |Di | 2n.
Given a random line  through p, the number ki of disks of Di that are inter-
sected by  is at most four, due to disjointness, and its expectation is O(|Di |/i).
Therefore, by the linearity of expectation, the expected number of disks of D
intersected by  is bounded by

  
E ki = E ki + E ki

i0 i : i n i : i> n

4 n+ O(|Di |/ n) = O( n) + O(n/ n) = O( n).


i : i> n

Corollary 1. Let D be a set of n pairwise

disjoint unit disks in the plane. Then,
there exists a line  that intersects O( n) disks of D such that each of the two
open halfplanes bounded by  contains at most 2n/3 disks of D. Moreover, such
a line can be found in O(n) time with probability at least 3/4.
246 P. Carmi et al.

Proof. Let p be a centerpoint of the set of centers of disks in D. By Lemma 1,

some line through p must intersect at most O( n) disks from D. Since p is a
centerpoint, each of the two open halfplanes bounded by  contains at most 2n/3
centers of disks in D, which means that the halfplane contains at most 2n/3 disks
from D.
The argument above suggests the following algorithm: rst we compute a cen-
terpoint p of the centers of a given set of disks, and then choose a line through p
uniformly at random. By Lemma 1 and Markovs
inequality, the probability that
the random line intersects more than c n disks is at most 1/4 for some con-
stant c. Thus, a desired line can be found with probability at least 3/4. The
running time is linear in the number of disks since a centerpoint can be found
in linear time [11].


In the statement, the exact value of 3/4 for the lower bound to the success
probability is not important (any positive probability will suce).
We now consider the general case where the disks are not necessarily disjoint.

Lemma 2. Let D be a set of n unit disks in the plane with m intersecting pairs,
and let p be any point in the plane.$Then the expected number of disks intersected
by a random line through p is O( (m + n) log n).

Proof. Consider a grid of 2 2 squares. Eachgrid cell is treated
as right-open
and top-open so that it is of the form of [x, x + 2) [y, y + 2). Associate each
disk in D with the grid cell containing its center, see Fig. 3(a).
Observe that one can color the grid cells with nine colors for every 3 3
block of grid cells, so that no two disks that are associated with dierent grid
cells of the same color intersect, see Fig. 3(b). Consider one of the colors j, with
1 j 9, and let Cj be the collection of subsets of D associated with the grid
cells of this color:

Cj = {C D | the center of disks in C lie in the same grid cell of color j}.

Then, Cj has the following two properties: (i) each subset C Cj in the same grid
cell is a clique, i.e., any two disks in the subset intersect each other; (ii) any two
disks from two dierent subsets in Cj are pairwise disjoint. Let nj = CCj |C|
denote the number of disks in D associated to grid cells of color j.
We divide the cliques in Cj into O(log nj ) buckets Bj0 , Bj1 , . . ., where Bji
consists of all cliques of Cj whose size is in the range [2i , 2i+1 ). Set lji = |Bji |,
sizes of the cliques in Bjiis in the range
for i = 0, 1, . . .. Then, the sum xji of the 
lji 2i xji < lji 2i+1 . We also know that i xji = nj . Let mji = CBji |C|(|C|
  
1)/2. Then, lji (22i1 2i1 ) mji < lji (22i+1 2i ) and i mji j i mji
m.
We rst compute the expected number of disks in cliques of Bji intersected by
a random line through p. Since the union of the disks in each clique is contained
in a disk$ of radius 2 such that they are disjoint, a random line through$ p intersects
only O( lji ) cliques of Bji by Lemma 1, and therefore only O( lji 2i+1 ) disks
Balanced Line Separators of Unit Disk Graphs 247

of D. By the denition of Bji , we have lji 22i1 mji + lji 2i1 mji + xji /2.

Thus, a random line through p intersects O( mji + xji ) disks in expectation.
Then, by the linearity of expectation, we sum the numbers for all j = 1, . . . , 9
and i = 0, . . . , log nj :

 $ +  + 
O( mji + xji ) O (mji + xji ) 1
j i j i j i
$
O( (m + n) log n),

where the rst inequality follows from the Cauchy-Schwarz inequality.



In the same way as Corollary 1 follows from Lemma 1, the following theorem
follows from Lemma 2.

Theorem 1. Let D be a set of n unit disks in the$plane with m intersecting

pairs. Then, there exists a line  that intersects O( (m + n) log n) disks of D
such that each of the two open halfplanes bounded by  contains at most 2n/3
disks of D. Moreover, such a line can be found in O(n) time with probability at
least 3/4.

As before, the exact value of 3/4 for the success probability is not important.

2.1 Axis-Parallel Separators

In this section we show an alternative, more restricted separator.
Specically,
we show that a line separator that intersects fewer disks (O( m + n)) exists,
even if we restrict ourselves to axis-parallel lines. However, this comes at the
cost that the balancing parameter is worsened: on each side of the line we can
certify only that there are at most 4n/5 disks. The theorem below and its proof
have a avor similar to Loer and Mulzer [13].

Theorem 2. Let D be a set of n unit disks in the plane with m intersecting

pairs. Then, there exists an axis-parallel line  that intersects O( m + n) disks
of D such that each of the two open halfplanes bounded by  contains at most
4n/5 disks of D. Moreover, such a line can be found in O(n) time.

Proof (sketch). Let P be the set of disk centers and assume that there are no
two points (centers) in P with the same x-coordinate or the same y-coordinate.
Let d (resp., u ) be a horizontal line such that there are exactly n/5 points of
P below it (resp., above it). Let H be the distance between d and u . Similarly,
let l (resp., r ) be a vertical line such that there are exactly n/5 points of P
to its left (resp., to its right). Let V be the distance between l and r . From
the above denitions, it follows that any horizontal line  between d and u ,
and any vertical line  between l and r have at most 4n/5 disks of D on each
of the two open halfplanes bounded by . We will show that one of these lines
intersects O( m + n) disks of D.
248 P. Carmi et al.

Consider the rectangle R dened by d , u , l and r . Now let hi be the

horizontal line above d whose distance from d is exactly i, for i = 1, . . . , H 1,
and similarly, let vi be the vertical line to the right of l whose distance from l
is exactly i, for i = 1, . . . , V 1. We may assume that at least one of H or V is
at least 2; otherwise it is easy. Then, we prove that one$ (or more) of the lines hi
or one (or more) of the lines vi intersects at most m + n/10 disks, for some
constant .


3 Almost Tightness of Our Approach

We now show that the approach used in Theorem 1 cannot be drastically im-
proved. Specically, we present a family D of n unit disks with a centerpoint p
of the centers of those unit disks such that any line that passes through p will
intersect many disks. Although this example can be constructed for any number
of disks n > 0 and any desired number of intersecting pairs m > n, the details
are a bit tedious. Instead, given the desired values n and m, we nd n n
and m m that satisfy the properties. This greatly simplies the proof and,
asymptotically speaking, the bounds are unaected.
Theorem 3. For any n, m N such that 9n m n2 /6, there exist n , m
N, where n n 2n and m/9 m 6m, and a set D of n unit disks in the
plane with m intersecting pairs that have the following property. There exists a
$ of unit disks in D such that any line  that passes
centerpoint p of the centers
through p intersects ( m log(n/ m)) disks of D.

% n and m, we choose k N as the smallest natural number

Proof (sketch). Given
  6m
k such that k 1+ln(n/k ) . First observe that such k exists and satises
k n. Indeed, this follows from the assumption m n2 /6.
1
Let  = n/k, and consider a suciently small positive real number 2 .
Consider now the  concentric circles Ci centered at the origin with radius
2i(1 + ) for i = 1, . . . , . On each such circle, we place k unit disks uniformly
(i.e., the arc spacing between the centers of two consecutive disks is 4i(1+)/k).
Let D be the collection of these disks, see Fig. 4 (left).
Let n = k. By construction, D has n disks, and n n n + k 2n
as claimed. It turns out that the number m of intersecting pairs in D satises
m/9 m 6m.
By symmetry, we can see that the origin is a centerpoint of the centers of
disks in D. Thus, it remains to show that any line that passes through the origin
must intersect many disks of D. In the following we show $ something stronger:
any ray emanating from a point p inside C1 will cross ( m log(n/ m)) disks.
Partition the disks of D into  layers D1 , . . . , D depending on which con-
centric circle their center lies on. Since we placed unit disks on circles that are
2(1+) units apart, only disks that belong to the same layer may have nonempty
intersection.
Let i be the maximum arc length of Ci such that two unit disks centered
at two endpoints of the arc touch each other, see Fig. 4 (middle). Note that
Balanced Line Separators of Unit Disk Graphs 249

Ci
i 1
2i(1 + )
p

p
i
2i(1 + )

Fig. 4. Left: Almost tightness construction for  = 3 (and k = 16). Middle: upper and
lower bounds for i . Right: construction for disks of arbitrary radii.

two disks on the same layer overlap if and only if the arc distance between the
centers is i or less.
We count the number of intersections on each layer independently. Each
unit disk in Di covers 2i(1+) i
radians of Ci . Since there are k evenly spread
1
disks in Di and 4 , each point of the circle Ci is contained in at least
1
 2i(1+)
i k
2  > i(4+1)
i k
1 disks.
%
6m 
By substituting k 1+ln  , i > 2, ln( + 1) (1 + ln )/2,  = n /k and

2m/9 n , we obtain that the number of disks in D intersected by the ray is at
least
   "& #"  #
 i k 2 6m 1
1 > 
i=1
i(4 + 1) 4 + 1 1 + ln  i=1
i
%
= ( m log(n/ m)).


Theorem 4. For any n N there exists a set D of O(n) pairwise disjoint

disks such that any line  such that both halfplanes contain (n) disks of D also
intersects (n) disks of D.
Proof (sketch). See Fig. 4 (right).

Note that in this construction the radii of the disks grow at an exponential
rate. By using a bucketing technique we can show that our upper bounds work
for disks of arbitrary radii at an expense of an additional O(log ) multiplicative
factor, where is the ratio between the smallest and largest radius of the disks
in D. This shows that our upper bounds are optimal also in this sense.

4 Experiments
In our experiments, we evaluate the quality of separator algorithms by the sep-
arator size. Theorem 1 suggests a simple algorithm: nd a centerpoint (which
250 P. Carmi et al.

can be done in linear time [11]) and try random lines passing through that point
until a good separator is found. Since implementing the centerpoint algorithm
is not trivial, we give an alternative method that is asymptotically slower but
much easier to implement: for a slope a selected uniformly at random nd a 2/3-
separator with slope a that intersects the minimum number of disks (this step
can be done in O(n log n) time by sorting the disks in orthogonal direction of a
and making a plane sweep). Repeat this
$ process for dierent slopes until we nd
a separating line that intersects O( (m + n) log n) disks. Clearly, a separator
found by the modied algorithm intersects at most as many disks as the line of
the same slope passing through a centerpoint. Thus, as in Theorem 1, a random
direction will be good with positive probability.
We compare our algorithm with the method by Fox and Pach [5] which
guarantees the separator size of O( m). For the implementation of our algorithm
we use the simpler variation described above.

The Method by Fox and Pach. Fox and Pach [5] proved that the intersection
graph of a set of Jordan curves in the plane has a 2/3-separator of size O( m)
if every pair of curves intersects in a constant number of points. Their proof is
constructive, as outlined below.
First, we build the arrangement of curves, and obtain a plane graph whose
vertex set are the vertices of the arrangements and and consecutive vertices on
a curve are joined by an edge.9 We triangulate the obtained plane graph to
make it maximal planar. Then, we nd a simple cycle 2/3-separator
C (i.e., a
2/3-separator that forms a cycle in the graph) of size O( m + n), which always
exists [15]. We output all curves containing a vertex in C.
In our implementation, we construct the circle arrangement in a brute-force
manner, and we use a simple cycle separator algorithm by Holzer et al. [10], called
the fundamental cycle separator (FCS) algorithm. Although the FCS algorithm
has no theoretical guarantee for the size of the obtained separator, the recent
experimental study by Fox-Epstein et al. [6] showed that it has a comparable
performance to the state-of-the-art cycle separator algorithm with theoretical
guarantee for most of the cases.

Instance Generation and Experiment Setup. All instances for our experiments
are randomly generated. We x a d d square S and generate n unit disks in S
independently and uniformly at random. If the graph is disconnected, we discard
it and generate again. All experiments have been performed on Intel  CoreTM
i7-5600U CPU @2.60GHz 4, with 7.7GB memory and 976.0GB hard disk,
running Ubuntu 14.04.3 LTS 64bit.

Experiment 1: Quality of the Proposed Method. In the rst experiment we em-

pirically examine the size of a separator obtained by our proposed algorithm
with the modication proposed at the beginning of this section. For instance
9
The method of Fox and Pach needs to add a constant number of additional vertices,
but the main feature is that the overall complexity of the graph is O(m + n).
Balanced Line Separators of Unit Disk Graphs 251

Average of the separator size

1000
Average of the separator size

Fox Pach k=5

50
optimal line separator k = 50
k=1 k = 100
40

30

20

10
k=1 k = 10 k = 20
k=2 k = 15 0

26

59

94

193

271

436

526

643

825

1077

1306

1551

1927

2250
100
100000 1x106
Number of edges in a unit disk graph Number of edges in a unit disk graph

Fig. 5. Left: The average separator size of our algorithm for random instances. The
horizontal axis shows the number of edges in logscale, and the vertical axis shows the av-
erage separator size in logscale when the algorithm is run k times, k {1, 2, 10, 15, 20}.
The solid line shows the square root of the number of edges. Right: Comparison of our
method with the Fox-Pach method. The horizontal axis shows the number of edges,
and the vertical axis shows the separator size; the red bar is obtained by the Fox-Pach
method, the blue bar is an optimal line separator, and the remaining four bars are
obtained by our method after trying k random directions (k {1, 5, 50, 100}) and re-
turning the minimum size separators. For all methods we repeat this process 20 times
and display the average minimum size.

generation, we x d = 100, and vary n from 10, 000 to 30, 000 with an increment
of 50. Since our algorithm is randomized, we run the algorithm k times, where
k {1, 2, 5, 10, 15, 20}, and compute the average separator size.
Fig. 5 (left) shows the result. We can observe that as the number of edges
increases, the size of the obtained line separators also increases, but the tendency
is close to ( m) as the solid black line shows. We thus empirically conclude
that the algorithm from Theorem 1 tends to output line separators of size O( m)
for random instances.

Experiment 2: Comparison with the Method by Fox and Pach. In the second
experiment we compare our algorithm with the method by Fox and Pach in its
separator size. For instance generation, we x d = 16, and generated 14 random
instances with various numbers of edges (the number of disks is not xed). For
each instance, we run our implementation of the Fox-Pach method as outlined
above and our simplied algorithm described in Section 4. Our algorithm is
iterated k times, where k {1, 5, 100, 500}, and we take the minimum separator
size among k runs. In this way, we can observe how the minimum separator
size converges to the real minimum. This process is repeated 20 times, and we
compute the average of the minimum separator sizes over those 20 repetitions.
We also examine the size of an optimal line separator to investigate the
limitation of the line-separator approach. Here, an optimal line separator means
a line 2/3-separator of minimum size, which can be found in O(n2 ) time by
looking at all possible bitangents to pairs of unit disks (but we implemented a
simpler O(n3 )-time algorithm), where the O() notation suppresses logarithmic
factors.
252 P. Carmi et al.

Fig. 5 (right) shows the result. On average, our method is strictly better than
the Fox-Pach method even when k = 1. As k increases, the average separator
size approaches to the size of optimal line separators. Empirically, fty iterations
are enough to obtain a reasonably good line separator.

Acknowledgments. The authors thank Michael Homann and Eli Fox-Epstein

for motivating discussion on the topic.

References
1. Alon, N., Katchalski, M., Pulleyblank, W.R.: Cutting disjoint disks by straight
lines. Discrete & Computational Geometry 4, 239243 (1989)
2. Alon, N., Seymour, P., Thomas, R.: A separator theorem for nonplanar graphs. J.
Amer. Math. Soc. 3, 801808 (1990)
3. Edelsbrunner, H., Guibas, L.J., Pach, J., Pollack, R., Seidel, R., Sharir, M.: Ar-
rangements of curves in the planetopology, combinatorics and algorithms. Theor.
Comput. Sci. 92(2), 319336 (1992)
4. Eppstein, D., Miller, G.L., Teng, S.: A deterministic linear time algorithm for
geometric separators and its applications. Fundam. Inform. 22(4), 309329 (1995)
5. Fox, J., Pach, J.: Separator theorems and Turan-type results for planar intersection
graphs. Advances in Mathematics 219(3), 10701080 (2008)
6. Fox-Epstein, E., Mozes, S., Phothilimthana, P.M., Sommer, C.: Short and simple
cycle separators in planar graphs. In: Proc. of the 15th Meeting on Algorithm
Engineering and Experiments. pp. 2640. SIAM (2013)
7. Gilbert, J.R., Hutchinson, J.P., Tarjan, R.E.: A separator theorem for graphs of
bounded genus. J. Algorithms 5(3), 391407 (1984)
8. Har-Peled, S., Quanrud, K.: Approximation algorithms for polynomial-expansion
and low-density graphs. In: Proc. of the 23rd Annual European Symposium. pp.
717728. Springer (2015)
9. Homann, M., Kusters, V., Miltzow, T.: Halving balls in deterministic linear time.
In: Proc. of the 22th Annual European Symposium. pp. 566578. Springer (2014)
10. Holzer, M., Schulz, F., Wagner, D., Prasinos, G., Zaroliagis, C.D.: Engineering
planar separator algorithms. ACM Journal of Experimental Algorithmics 14 (2009)
11. Jadhav, S., Mukhopadhyay, A.: Computing a centerpoint of a nite planar set of
points in linear time. Discrete & Computational Geometry 12, 291312 (1994)
12. Lipton, R.J., Tarjan, R.E.: A separator theorem for planar graphs. SIAM Journal
on Applied Mathematics 36(2), 177189 (1979)
13. Loer, M., Mulzer, W.: Unions of onions: Preprocessing imprecise points for fast
onion decomposition. JoCG 5(1), 113 (2014)
14. Matousek, J.: Near-optimal separators in string graphs. Combinatorics, Probability
& Computing 23(1), 135139 (2014)
15. Miller, G.L.: Finding small simple cycle separators for 2-connected planar graphs.
J. Comput. Syst. Sci. 32(3), 265279 (1986)
16. Miller, G.L., Teng, S., Thurston, W.P., Vavasis, S.A.: Separators for sphere-
packings and nearest neighbor graphs. J. ACM 44(1), 129 (1997)
17. Nesetril, J., Ossona de Mendez, P.: Grad and classes with bounded expansion II.
Algorithmic aspects. Eur. J. Comb. 29(3), 777791 (2008)
18. Smith, W.D., Wormald, N.C.: Geometric separator theorems & applications. In:
Proc. of the 39th Annual Symposium on Foundations of Computer Science. pp.
232243 (1998)
 All-Pairs Shortest Paths in Geometric
Intersection Graphs

Timothy M. Chan1 and Dimitrios Skrepetos2

1
Department of Computer Science, University of Illinois at Urbana-Champaign,
[email protected]
2
Cheriton School of Computer Science, University of Waterloo,
[email protected]

Abstract. We address the All-Pairs Shortest Paths (APSP) problem

for a number of unweighted, undirected geometric intersection graphs.
We present a general reduction of the problem to static, oine in-
tersection searching (specically detection). As a consequence,

we can
solve APSP for intersection graphs of n arbitrary disks in O n2 log n


time, axis-aligned line segments in O n2 log log n time, arbitrary line

7/3 1/3 
segments in O n log n time, d-dimensional axis-aligned boxes in


O n2 logd1.5 n time for d 2, and d-dimensional axis-aligned unit hy-



percubes in O n2 log log n time for d = 3 and O n2 logd3 n time for
d 4.
In addition, we show how to solve the Single-Source Shortest Paths
(SSSP) problem in unweighted intersection graphs of axis-aligned line
segments in O (n log n) time, by a reduction to dynamic orthogonal point
location.

Keywords: shortest paths, geometric intersection graphs, intersection

searching data structures, disk graphs

1 Introduction
As a motivating example, consider the following toy problem: given a set S of
n axis-aligned line segments in the plane representing a road network, and two
points p1 and p2 lying on two segments of S, compute a path from p1 to p2 that
stays on S while minimizing the number of turns. (See Figure 1.)
To solve the problem, we can create a vertex for each segment of S and an
(unweighted, undirected) edge between two vertices if their corresponding seg-
ments intersect. This denes the intersection graph G(S). Then given two points
p1 and p2 , lying on the segments s and t of S, a minimum-turn path from p1 to p2
corresponds precisely to an unweighted shortest path from s to t in G(S). Naively
constructing G(S) and running breadth-rst search (BFS) would require O(n2 )
worst-case time. In Section 4, however, we observe an O(n log n)-time algorithm,
which is new to the best of the authorss knowledge. In fact, the algorithm solves
the more general, Single-Source Shortest Paths (SSSP) problem in G(S), by an
application of data structures for dynamic orthogonal point location [21,6].

Springer International Publishing AG 2017 253

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 253264, 2017.
DOI: 10.1007/978-3-319-62127-2_22
254 T.M. Chan and D. Skrepetos
p2
p1 t
s

Fig. 1. A set S of axis-aligned line segments is shown. The path from p1 , lying on
segment s, to p2 , lying on segment t, staying on S and using the minimum number of
turns is marked in bold.
Our main focus in this paper will be a similar problem, namely the All-Pairs
Shortest Path (APSP) problem in geometric intersection graphs. More generally,
given a set S of n geometric objects, its intersection graph G(S) is dened
by creating one vertex for every object of S and an (undirected, unweighted)
edge between two vertices if their corresponding objects intersect. We want to
compute a representation of an unweighted shortest path between s and t for
every pair of objects s, t S. For general unweighted, undirected graphs the
problem can be solved in O (n ) time (e.g., see [5,26]), where < 2.373 is the
matrix multiplication exponent [28], but better results are possible for geometric
intersection graphs.
Our main results are as follows:


For arbitrary disks, we solve APSP in O n2 log n time. The disk case is
naturally motivated by applications in ad hoc communication networks. Fol-
lowing work by Cabello and Jeji [8] on SSSP for unit-disk graphs, a pre-
% by the authors [15] studied APSP for unit-disk graphs and gave
vious paper
log log n
an O n2 log n -time algorithm, but the approach cannot be extended
to arbitrary disks. A paper by Kaplan et al. [23] contains an algorithm for
SSSP for disks (which can be used for APSP), but this is for a directed vari-
ant of intersection graphs (called transmission graphs), and the running
time has multiple logarithmic factors unless we assume that the maximum-
to-minimum radius ratio is bounded.

For axis-aligned line segments, we solve APSP in O n2 log log n time, which
is better than running n times the O(n log n)-time algorithm for SSSP that
we have mentioned for the toy problem at the beginning. No previous results
have been reported, to the best of the authorss knowledge.
When the line segments are not  axis-alignedbut have arbitrary orientations
instead, we solve APSP in O n7/3 log1/3 n time, which is a little better
than the general O(n ) result, at least with the current upper bound on .
(Regardless, our algorithm has the advantage of being combinatorial.)
See Table 1 for further results on axis-aligned boxes, unit hypercubes, and
fat triangles of roughly equal size.

All these results stem from one single, general technique, which reduces APSP
to the design of data structures for static, oine intersection detection, i.e., given
a query object, decide whether there is an input object intersecting it (and report
All-Pairs Shortest Paths in Geometric Intersection Graphs 255

Geometric Objects Running Time

arbitrary disks O(n2 log n)


axis-aligned line segments O n2 log log n


arbitrary line segments O n7/3 log1/3 n


d-dimensional axis-aligned boxes O n2 logd1.5 n for d 2


d-dimensional axis-aligned unit O n2 log log n for d = 3 and


hypercubes O n2 logd3 n for d 4


fat triangles of roughly equal size O n2 log4 n

Table 1. The results for APSP

one if the answer is yes). Our technique, described in Section 2, works by visiting
vertices in an order prescribed by a spanning tree; given the BFS tree from a
source vertex s as a guide, we can generate the BFS tree from an adjacent source
vertex s quickly, by exploiting the fact that distances to s are approximately
known up to 1, and by using the right geometric data structures. Some form of
this simple idea has appeared before for general graphs (e.g., see [4,10]), but it is
somehow overlooked by previous researchers in the context of geometric APSP.
To appreciate the advantages of the new technique, we should compare it
with other known general approaches:
First, a naive approach is to solve SSSP n times from every source indepen-
dently, i.e., generate the BFS trees from each source from scratch. Geometric
SSSP problems can often be reduced to dynamic data structuring problems,
for example, as observed in Chan and Efrats paper [12] (the reduction is
much simplied in the unweighted, undirected setting). In fact, our solution
to the toy problem at the beginning is done via this approach. However, dy-
namic data structures for geometric intersection or range searching usually
are more complicated and have slower query times than their static counter-
parts, sometimes by multiple logarithmic factors. For example, the arbitrary
disk case requires dynamic data structures for additively

weighted
 nearest
neighbor search, and a BFS therein takes nearly O n log10 n time [24]. Our
reduction to static data structuring problems yields better results.
Another general approach is to employ biclique covers [19,2] to sparsify
the intersection graph rst and then solve the problem on the sparsied
graph. Biclique covers are related to static, oine intersection searching
data structures (e.g., as noted in [9]). However, the complexity of biclique
covers also tends to generate extra logarithmic factors. For example, for d-
dimensional boxes, the sparsied graph has O n logd n edges, leading to
   
an O n2 logd n -time algorithm, but our solution requires O n2 logd1.5 n
time.

For arbitrary
 disks, the complexity

of the biclique covers is even worse
(O n3/2+ [3]), leading to an O n5/2+ -time algorithm, which is much
256 T.M. Chan and D. Skrepetos


slower than our O n2 log n result. The underlying issue is that intersection
searching (as implicitly needed in biclique covers) may in general be harder
than intersection detection.

In deriving
 our result for axis-aligned boxes, we also obtain a new
O n log n -time algorithm for oine rectangle stabbing in two dimensions (pre-
process n axis-aligned rectangles so that we can nd a rectangle stabbing each
query point). This result (see the full paper) may be of independent interest.
For the rest of the paper, for s, t S, where S is a set of geometric objects, let
dist[s, t] denote the distance of the shortest path from s to t in the intersection
graph of S and pred [s, t] denote the predecessor of t in that path. In SSSP we
want to compute dist[s, t] and pred [s, t] for a given s S and t S, while
in APSP we want to compute dist[s, t] and pred [s, t] s, t S. All algorithms
assume the standard unit-cost RAM model of computation where the word size
is at least log n in bits.

2 Reducing APSP to static, oine intersection detection

In this section, we reduce the problem of solving APSP in unweighted, undirected

geometric intersection graphs of objects of constant-description complexity to
static, oine intersection detection. We assume that the graph is connected; if
not, then we can simply work with every connected component independently.
We rst compute an arbitrary spanning tree T0 of G(S), root it at an arbitrary
object s0 S, and then compute the shortest path tree of s0 . Then, we visit
each object s of T0 in a pre-order manner, and compute the shortest path tree
of s by using the shortest path tree of s as a guide, where s is the parent of s
in T0 . The pseudocode of the algorithm is given in Algorithm 1. The initial call
is APSP(S, s0 ).

Algorithm 1: APSP(S, s0 )
1 build G(S)
2 compute any spanning tree T0 of G(s) and root it at any s0 S
3 compute the shortest path tree of s0
4 for each s S {s0 } following a pre-order traversal of T0 do
5 compute the shortest path tree T (s) of s, using the shortest path tree
T (s ) of its parent s in T0 , by calling SSSP (S, s, T (s ))

It remains to describe how to compute the shortest path tree of a vertex

s S, given the shortest path tree of a vertex t at unit distance from it, i.e.,
how to implement Line 5 in Algorithm 1. From the triangle inequality and from
dist[s, s ] = 1, we know that if dist[s , z] =  for an object z S, then 
1 dist[s, z]  + 1. Thus we already have an 1-additive approximation of the
All-Pairs Shortest Paths in Geometric Intersection Graphs 257

distances dist[s , z] for any z S. To compute the exact distances from s to any
object z S, we follow the procedure of the next paragraph.
As in classical BFS, we proceed in n 1 steps, where in step  we assume
that we have found all the objects at distance at most  1 from s and want to
produce the objects at distance exactly . The objects at distance exactly  1
from s are called the frontier objects, while the ones whose distance has not yet
been found are called the undiscovered objects. Then we need to procure quickly
all the undiscovered objects that intersect the frontier objects. Because of the
1-additive approximation, an object z can be at distance  from s only if it is at
distance  1, , or  + 1 from s . These points are called the candidate objects.
Hence we need to determine, for each candidate object, whether it intersects any
frontier object. This is an instance of intersection searching, or more specically,
intersection detection:
Preprocess a set of input objects into a data structure so that we can
quickly decide if a given query object intersects any input object, and
report one such input object if it exists.
In our application, the input objects are static, and the query objects are
oine, i.e., are all given in advance.
To summarize, the pseudocode is presented in Algorithm 2. Thus we obtain
the following theorems:

Algorithm 2: SSSP(S, s, T (s ))

1 dist[s, s] = 0
2 dist[s, z] = z S {s}
3 pred [s, z] = NULL z S
4 for  = 0 to n 1 do
5 A = {z | dist[s , z] = } // objects at distance  from s
6 for  = 1 to n 1 do
7 F = {z S | dist[s, z] =  1} // frontier objects
8 C = A1 A A+1 // candidate objects
9 build a static, oine intersection detection data structure for F and C
10 for z C do
11 if dist[s, z] = then
12 query the data structure for z
13 let w be the answer
14 if w not NULL then
15 dist[s, z] = 
16 pred [s, z] = w

Theorem 1. Given a set S of n objects of constant-description complexity and

the shortest path tree of an object s S in the unweighted, undirected intersec-
tion graph of S, we can compute the shortest path tree of an object s S, where
258 T.M. Chan and D. Skrepetos

dist[s, s ] = 1, in the same graph in O(SI (n, n)) time, where SI (n, m) is the time
to construct a static, oine intersection detection data structure for n objects
and query it m times, assuming the property that SI (n1 , m1 ) + SI (n2 , m2 )
SI (n1 + n2 , m1 + m2 ).

Proof. Let n (resp. m ) be the number of frontier (resp. candidate) objects

in step  of the BFS. During the algorithm, an object is in thefrontier ex-
n1
actly once and in the candidate at most thrice; in other words, =1 n n
n1
 =1 m 3n.Then the time to compute the shortest path tree of s is
and
n1
O =1 SI (n , m ) = O(SI(n, n)).


Theorem 2. We can solve APSP in an unweighted

geometric intersection graph
of n objects of constant-description complexity in O n2 + nSI (n, n) time, where
SI (, ) is dened as in Theorem 1.


Proof. In Lines 13 of Algorithm 1, we can build G(S) in O n2 time,
 nd
a spanning tree T0 , and compute the shortest-path tree of s0 , in O n2 time
naively. In each of the n 1 iterations, Line 3 of Algorithm 1 takes O (SI (n, n))
time by Theorem 1.


3 Applications

In this section we apply Theorem 2 and known data structures for static, oine
intersection detection to obtain ecient APSP algorithms in specic families of
geometric intersection graphs. Some of the data structures we employ are in fact
online.

Arbitrary disks in the plane. We rst consider intersection graphs of disks of

arbitrary radii, also known as disk graphs. The static intersection detection data
structure for disks will be based on an additively weighted Voronoi diagram,
where the distance between a site w corresponding to a disk of radius rw and a
point x is dened as d(w, x) = ||w x|| rw . This Voronoi diagram allows us to
determine the disk whose boundary is closest to a query point. We construct the
Voronoi diagram for the centers of the frontier disks and a point location data
structure for the diagrams cells. Then we query the Voronoi diagram with the
center of each query disk. We can check if the query disk and the disk returned
by the query intersect in constant time.
The time for building the additively weighted Voronoi diagram of n disks is
O(n log n) [20]. We build a point location data structure in O(n log n) time, so
that (online) queries take O(log n) time [27]. Therefore, SI (n, n) = O(n log n).

Theorem

3. We can solve APSP in an unweighted intersection graph of n disks
in O n2 log n time.
All-Pairs Shortest Paths in Geometric Intersection Graphs 259
v4
v2
h1 v3
h5
v1

h2
v5

h3

h4

Fig. 2. This gure depicts a set of horizontal input segments, its vertical decomposition
(shown by the dashed lines), and a set of vertical query segments.
Axis-aligned line segments in the plane. We now turn our attention to inter-
section graphs of axis-aligned line segments. We describe a static intersection
detection data structure for horizontal input segments and vertical query seg-
ments. (Vertical input segments and horizontal query segments can be handled
by a symmetric structure.) The data structure is composed of the vertical de-
composition of the horizontal input segments, stored in a point location data
structure. Given a vertical query segment, we perform a point location query
for its bottom endpoint. If the top endpoint lies in the same cell, there is no
intersection; otherwise, we can report the segment bounding the top side of the
cell. (See Figure 2 for an example.)
We can apply the static orthogonal point location data structure of Chan [11]
(Theorem 2.1), with O(n log log U ) preprocessing time and O(log log U ) query
time, under the assumption that all coordinates are integers bounded by U .
Thus,
SI (n, n) = O(n log log U ). This implies an APSP algorithm running
in O n2 log log U time. At the beginning, we can presort all coordinates in
O(n log n) time and replace each coordinate value with its rank; this ensures
that U = n. Thus, we obtain:

Theorem 4. We can solve

APSP in an unweighted intersection graph of n axis-
aligned line segments in O n2 log log n time.

The result can be easily be extended to any set of line segments with a
constant number of dierent orientations.

Arbitrary line segments. Next we consider the case of arbitrary

 line segments.
Chazelle [16] (Theorem 4.4) has given an O n4/3 log1/3 n -time algorithm to
count the number of intersections among n line segments. The algorithm can
be modied to count the number of intersections between n red (input) line
segments and n blue (oine query) line segments. In fact, it is straightforward
to adapt the algorithm to decide, for each blue segment, whether it intersects
any
 red segment  and, if yes, report one such red segment. Thus, SI (n, n) =
O n4/3 log1/3 n .
260 T.M. Chan and D. Skrepetos

Theorem 5. We can solve APSP in anunweighted intersection graph of n ar-

bitrary line segments in O n7/3 log1/3 n time.

Axis-aligned boxes in d dimensions. For the case of axis-aligned rectangles in d =

2 dimensions, oine rectangle intersection counting is known to be reducible [18]
to oine orthogonal

 range counting, for which Chan and Ptracu [14] have
given an O n log n -time algorithm, under the assumption that all coordinates
have been presorted. Consequently, we can decide for each query box whether
it intersects any input box. With more eort, we can adapt their technique to
report a witness input box for each query box
with
a yes answer, and thus solve
oine intersection detection in SI (n, n) = O n log n time; see full paper for
details. At the beginning, we can presort all coordinates in O(n log n) time.
For axis-aligned boxes in d 3 dimensions, we can use standard range

trees [17] with the above d = 2 base case to obtain SI (n, n) = O n logd1.5 n .

Theorem 6. We can solve APSP in  an unweighted

 intersection graph of n d-
dimensional axis-aligned boxes in O n2 logd1.5 n time for d 3.

Axis-aligned unit hypercubes in d dimensions. When the axis-aligned boxes are

unit hypercubes, the time bound for oine intersection detection can be im-
proved. We build a uniform grid with unit side length and solve the problem
inside each grid cell separately. Each input or query unit hypercube participates
in at most a constant (2d ) number of grid cells. Inside a grid cell, each unit
hypercube is eectively unbounded along d sides. Without loss of generality, we
may assume that each input box is of the form (, a1 ] (, ad ] and
each query box is of the form [b1 , ) [bd , ). Thus, the problem reduces
to oine dominance detection: decide for each query point (b1 , . . . , bd ) whether
it is dominated by some input point (a1 , . . . , ad ) and, if yes, report one such
input point.
For d = 3, Gupta et al. [22] gave an algorithm to answer n oine dominance
reporting queries in O((n + K) log log U ) time where K is the total output size,
under the assumption that all coordinates are integers bounded by U . Their al-
gorithm can be easily adapted to answer n oine dominance detection
queries in
O(n log log U ) time. This implies an APSP algorithm running in O n2 log log U
time. At the beginning, we can presort all coordinates in O(n log n) time and
replace each coordinate value with its rank; this ensures that U = n.
For d 4, Afshani et al. [1] (following Chan et al. [13]) gave a deterministic

algorithm to answer n oine dominance reporting queries in O n logd3 n + K
time where K is the total output size. It can be checked
 that their
 algorithm can
answer n oine dominance detection queries in O n logd3 n time. (One step
in their algorithm which involves reversing the role between input and query
points becomes unnecessary for the detection problem.)
All-Pairs Shortest Paths in Geometric Intersection Graphs 261

Theorem 7. We can solve APSP in an unweighted

2 intersection
 graph of n d-
dimensional
 axis-aligned
 unit hypercubes in O n log log n time for d = 3 and
in O n2 log d3
n time for d 4.

Fat triangles in the plane. Finally we consider the intersection graph of fat
triangles (i.e., triangles that have bounded inradius-to-circumradius ratios) with
roughly equal size. Katz [25] (Theorem
4.1 (i)and (iii)) has given an (online)
data structure achieving SI (n, n) = O n log4 n . Thus:
Theorem 8. We can solve APSP in
an unweighted  intersection graph of n fat
triangles with roughly equal size in O n2 log4 n time.

4 Reducing SSSP to decremental intersection detection

We give in this section a reduction of SSSP in intersection graphs to dynamic
intersection detection.
We will emulate the classic BFS algorithm in the following way. Let s S be
the given source vertex. We proceed iteratively in n1 steps and follow the same
process in each one. In step  we assume that we have found all the distances
and predecessors for all the objects that are at distance no more than  1 from
s. We employ the denitions of the frontier and undiscovered objects as given
in Section 2. The goal is to compute the distances and predecessors for all the
undiscovered objects that are at distance  from s. Those objects are the ones
that have at least one intersection with a frontier object. To nd those intersec-
tions we maintain an intersection detection data structure for the undiscovered
objects that supports deletiona deletion-only dynamic data structure is often
referred to as a decremental data structure. We query the structure with the
frontier objects; each time we detect an intersection of a frontier object with an
undiscovered one, we properly update the latters distance and predecessor and
delete it from the data structure. The pseudocode of the algorithm is given in
Algorithm 3.
We conclude this section with the following theorem.
Theorem 9. We can solve SSSP in an unweighted, undirected geometric in-
tersection graph in O(DI (n, n)) time, where DI (n, m) is the time to construct
a decremental intersection detection data structure of n objects and perform n
deletions and m queries.
Proof. The correctness of the algorithm can be easily proved by induction.
For the running time, we notice that an object can be in the frontier in only
one step of the algorithm. In the beginning all the objects except the source
are undiscovered (thus in the decremental intersection detection data structure
as well), and once an object is deleted from that set, it is never inserted again.
When querying the intersection detection data structure with a frontier object
t there are two possible outcomes. If the query returns an undiscovered object
z that intersects t, then z is deleted from the data structure, and since it is
262 T.M. Chan and D. Skrepetos

Algorithm 3: SSSP(S, s)
1 dist[s, s] = 0
2 dist[s, t] = t S {s}
3 pred [s, t] = NULL t S
4 build a decremental intersection detection data structure for S {s}
// i.e., for undiscovered objects
5 for  = 1 to n 1 do
6 F = {t S | dist[s, t] =  1} // frontier
7 for each t F do
8 while true do
9 query the data structure with t
10 let z be the answer
11 if z not NULL then
12 dist[s, z] = 
13 pred [s, z] = t
14 delete z from the data structure
15 else
16 break

never reinserted, this type of query happens only once z S. If the query
returns nothing, then this is the last query that t performs in that step, and
since t can be in the frontier at most once, this type of query happens only
once t S. Consequently the total number of queries in the data structure is
O(n). Furthermore, the number of deletions in the decremental data structure
is obviously O(n). Thus the total running time is O(DI (n, n)).


Application to axis-aligned line segments. We need a decremental intersection de-

tection data structure for horizontal input segments and vertical query segments.
(Vertical input segments and horizontal query segments can be handled by a
symmetric structure.) Giyora and Kaplan [21] (Theorem 5.3) and Blelloch [6]
(Theorem 6.1) provided a data structure for supporting vertical ray shooting
queries in O(log n) time and insertions and deletions of horizontal segments in
O(log n) timethe problem is sometimes referred to as dynamic orthogonal point
location. This immediately implies DI (n, n) = O(n log n). Thus:

Theorem 10. We can solve SSSP in an unweighted intersection graph of n

axis-aligned line segments in O(n log n) time.

The result can be easily be extended to any set of line segments with a
constant number of dierent orientations.

5 Conclusion

Interesting open problems in unweighted, undirected geometric intersection

graphs include constructing ecient

 distance oracles and computing the diam-
eter in truly subquadratic O n2 time for some > 0, in view of Cabellos
All-Pairs Shortest Paths in Geometric Intersection Graphs 263

recent breakthrough for the diameter problem in planar graphs [7]. For certain
geometric objects such as arbitrary line segments, even a quadratic-time APSP
algorithm is already open. Finally, solving APSP in the weighted case seems to
be more dicult, as we can no longer exploit the general reduction from APSP
to static, oine intersection detection.

References

1. Peyman Afshani, Timothy M. Chan, and Konstantinos Tsakalidis. Deterministic

rectangle enclosure and oine dominance reporting on the RAM. In Proceedings
of the Forty-First Annual International Colloquium on Automata, Languages, and
Programming, pages 7788, 2014.
2. Pankaj K. Agarwal, Noga Alon, Boris Aronov, and Subhash Suri. Can visibility
graphs be represented compactly? Discrete & Computational Geometry, 12(3):347
365, 1994.
3. Pankaj K. Agarwal, Marco Pellegrini, and Micha Sharir. Counting circular arc
intersections. SIAM Journal on Computing, 22(4):778793, 1993.
4. Donald Aingworth, Chandra Chekuri, Piotr Indyk, and Rajeev Motwani. Fast
estimation of diameter and shortest paths (without matrix multiplication). SIAM
Journal on Computing, 28(4):11671181, 1999.
5. Noga Alon, Zvi Galil, Oded Margalit, and Moni Naor. Witnesses for boolean matrix
multiplication and for shortest paths. In Proceedings of the Thirty-Third Annual
IEEE Symposium on Foundations of Computer Science, pages 417426, 1992.
6. Guy E. Blelloch. Space-ecient dynamic orthogonal point location, segment inter-
section, and range reporting. In Proceedings of the Nineteenth Annual ACM-SIAM
Symposium on Discrete Algorithms, pages 894903, 2008.
7. Sergio Cabello. Subquadratic algorithms for the diameter and the sum of pairwise
distances in planar graphs. In Proceedings of the Twenty-Eighth Annual ACM-
SIAM Symposium on Discrete Algorithms, pages 21432152. SIAM, 2017.
8. Sergio Cabello and Miha Jeji. Shortest paths in intersection graphs of unit disks.
Computational Geometry, 48(4):360367, 2015.
9. Timothy M. Chan. Dynamic subgraph connectivity with geometric applications.
SIAM Journal on Computing, 36(3):681694, 2006.
10. Timothy M. Chan. All-pairs shortest paths for unweighted undirected graphs in
o(mn) time. ACM Transactions on Algorithms, 8:117, 2012.
11. Timothy M. Chan. Persistent predecessor search and orthogonal point location on
the word RAM. ACM Transactions on Algorithms, 9(3):22, 2013.
12. Timothy M. Chan and Alon Efrat. Fly cheaply: On the minimum fuel consumption
problem. Journal of Algorithms, 41(2):330337, 2001.
13. Timothy M. Chan, Kasper Green Larsen, and Mihai Ptracu. Orthogonal range
searching on the RAM, revisited. In Proceedings of the Twenty-Seventh Annual
ACM Symposium on Computational Geometry, pages 110, 2011.
14. Timothy M. Chan and Mihai Ptracu. Counting inversions, oine orthogonal
range counting, and related problems. In Proceedings of the Twenty-First Annual
ACM-SIAM Aymposium on Discrete Algorithms, pages 161173, 2010.
15. Timothy M. Chan and Dimitrios Skrepetos. All-pairs shortest paths in unit-disk
graphs in slightly subquadratic time. In Proccedings of the Twenty-Seventh Annual
International Symposium on Algorithms and Computation, pages 24:124:13, 2016.
264 T.M. Chan and D. Skrepetos

16. Bernard Chazelle. Cutting hyperplanes for divide-and-conquer. Discrete & Com-
putational Geometry, 9(2):145158, 1993.
17. Herbert Edelsbrunner and Hermann A. Maurer. On the intersection of orthogonal
objects. Information Processing Letters, 13(4/5):177181, 1981.
18. Herbert Edelsbrunner and Mark H. Overmars. On the equivalence of some rect-
angle problems. Information Processing Letters, 14(3):124127, 1982.
19. Toms Feder and Rajeev Motwani. Clique partitions, graph compression and
speeding-up algorithms. Journal of Computer and System Sciences, 51(2):261
272, 1995.
20. Steven Fortune. A sweepline algorithm for Voronoi diagrams. Algorithmica, 2:153
174, 1987.
21. Yoav Giora and Haim Kaplan. Optimal dynamic vertical ray shooting in rectilinear
planar subdivisions. ACM Transactions on Algorithms, 5(3):28, 2009.
22. Prosenjit Gupta, Ravi Janardan, Michiel H. M. Smid, and Bhaskar DasGupta. The
rectangle enclosure and point-dominance problems revisited. International Journal
of Computational Geometry and Applications, 7(5):437455, 1997.
23. Haim Kaplan, Wolfgang Mulzer, Liam Roditty, and Paul Seiferth. Spanners and
reachability oracles for directed transmission graphs. In LIPIcs-Leibniz Interna-
tional Proceedings in Informatics, volume 34, 2015.
24. Haim Kaplan, Wolfgang Mulzer, Liam Roditty, Paul Seiferth, and Micha Sharir.
Dynamic planar Voronoi diagrams for general distance functions and their algo-
rithmic applications. In Proceedings of the Twenty-Eighth Annual ACM-SIAM
Symposium on Discrete Algorithms, pages 24952504, 2017.
25. Matthew J. Katz. 3-D vertical ray shooting and 2-D point enclosure, range
searching, and arc shooting amidst convex fat objects. Computational Geometry,
8(6):299316, 1997.
26. Raimund Seidel. On the all-pairs-shortest-path problem in unweighted undirected
graphs. Journal of Computer and System Sciences, 51(3):400403, 1995.
27. Jack Snoeyink. Point location. In Handbook of Discrete and Computational Ge-
ometry, pages 767785. CRC Press, second edition, 2004.
28. Virginia Vassilevska Williams. Multiplying matrices faster than Coppersmith-
Winograd. In Proceedings of the Forty-Fourth Annual ACM Symposium on Theory
of Computing, pages 887898, 2012.
 The Complexity of Drawing Graphs
on Few Lines and Few Planes

Steven Chaplick1 , Krzysztof Fleszar2 , Fabian Lipp1 , Alexander Ravsky3 ,

Oleg Verbitsky4   , and Alexander Wol1
1
Lehrstuhl fur Informatik I, Universitat Wurzburg, Germany,
www1.informatik.uni-wuerzburg.de/en/staff, [email protected]
2
Department of Mathematical Engineering, Universidad de Chile, Chile,
[email protected].
3
Pidstryhach Institute for Applied Problems of Mechanics and Mathematics,
National Academy of Sciences of Ukraine, Lviv, Ukraine, [email protected]
4
Institut fur Informatik, Humboldt-Universitat zu Berlin, Germany,
[email protected]

Abstract. It is well known that any graph admits a crossing-free

straight-line drawing in R3 and that any planar graph admits the
same even in R2 . For a graph G and d {2, 3}, let 1d (G) denote the
minimum number of lines in Rd that together can cover all edges
of a drawing of G. For d = 2, G must be planar. We investigate the
complexity of computing these parameters and obtain the following
hardness and algorithmic results.
For d {2, 3}, we prove that deciding whether 1d (G) k for a
given graph G and integer k is R-complete.
Since NP R, deciding 1d (G) k is NP-hard for d {2, 3}.
On the positive side, we show that the problem is xed-parameter
tractable with respect to k.
Since R PSPACE, both 12 (G) and 13 (G) are computable in
polynomial space. On the negative side, we show that drawings
that are optimal with respect to 12 or 13 sometimes require
irrational coordinates.
Let 23 (G) be the minimum number of planes in R3 needed to
cover a straight-line drawing of a graph G. We prove that decid-
ing whether 23 (G) k is NP-hard for any xed k 2. Hence,
the problem is not xed-parameter tractable with respect to k
unless P = NP.


K. Fleszar was supported by Conicyt PCI PII 20150140 and Millennium Nucleus
Information and Coordination in Networks RC130003.

F. Lipp was supported by Cusanuswerk. ORCID: orcid.org/0000-0001-7833-0454

O. Verbitsky was supported by DFG grant VE 652/1-2.

ORCID: orcid.org/0000-0001-5872-718X

Springer International Publishing AG 2017 265

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 265276, 2017.
DOI: 10.1007/978-3-319-62127-2_23
266 S. Chaplick et al.

1 Introduction
As is well known, any graph can be drawn in R3 without crossings so that all
edges are segments of straight lines. Suppose that we have a supply L of lines in
R3 , and the edges are allowed to be drawn only on lines in L. How large does
L need to be for a given graph G? For planar graphs, a similar question makes
sense also in R2 , since planar graphs admit straight-line drawings in R2 by the
WagnerFaryStein theorem. Let 13 (G) denote the minimum size of L which is
sucient to cover a drawing of G in R3 . For a planar graph G, we denote the cor-
responding parameter in R2 by 12 (G). The study of these parameters was posed
as an open problem by Durocher et al. [10]. The two parameters are related to
several challenging graph-drawing problems such as small-area or small-volume
drawings [9], layered or track drawings [8], and drawing graphs with low visual
complexity. Recently, we studied the extremal values of 13 (G) and 12 (G) for
various classes of graphs and examined their relations to other characteristics
of graphs [6]. In particular, we showed that there are planar graphs where the
parameter 13 (G) is much smaller than 12 (G). Determining the exact values of
13 (G) and 12 (G) for particular graphs seems to be tricky even for trees.
In fact, the setting that we suggested is more general [6]. Let 1 l < d. We
dene the ane cover number ld (G) as the minimum number of l-dimensional
planes in Rd such that G has a straight-line drawing that is contained in the
union of these planes. We suppose that l 2 as otherwise ld (G) = 1.
Moreover, we can focus on d 3 as every graph can be drawn in 3-space
as eciently as in higher dimensions, that is, ld (G) = l3 (G) if d 3 [6]. This
implies that, besides the line cover numbers in 2D and 3D, 12 (G) and 13 (G),
the only interesting ane cover number is the plane cover number 23 (G). Note
that 23 (G) = 1 if and only if G is planar. Let Kn denote the complete graph
on n vertices. For the smallest non-planar graph K5 , we have 23 (K5 ) = 3. The
parameters 23 (Kn ) are not so easy to determine even for small values of n. We
have shown that 23 (K6 ) = 4, 23 (K7 ) = 6, and 6 23 (K8 ) 7 [6]. It is not hard
to show that 23 (Kn ) = (n2 ), and we determined the asymptotics of 23 (Kn ) up
to a factor of 2 using the relations of these numbers to Steiner systems.
The present paper is focused on the computational complexity of the ane
cover numbers. A good starting point is to observe that, for given G and k, the
statement ld (G) k can be expressed by a rst-order formula about the reals
of the form x1 . . . xm (x1 , . . . , xm ), where the quantier-free subformula is
written using the constants 0 and 1, the basic arithmetic operations, and the
order and equality relations. If, for example, l = 1, then we just have to write
that there are k pairs of points, determining a set L of k lines, and there are n
points representing the vertices of G such that the segments corresponding to
the edges of G lie on the lines in L and do not cross each other. This observation
shows that deciding whether or not ld (G) k reduces in polynomial time to
the decision problem (Hilberts Entscheidungsproblem) for the existential theory
of the reals. The problems admitting such a reduction form the complexity class
R introduced by Schaefer [23], whose importance in computational geometry
has been recognized recently [4,16,24]. In the complexity-theoretic hierarchy,
The Complexity of Drawing Graphs on Few Lines and Few Planes 267

this class occupies a position between NP and PSPACE. It possesses natural

complete problems like the decision version of the rectilinear crossing number [1],
the recognition of segment intersection graphs [15] or unit disk graphs [13].
Below, we summarize our results on the computational complexity of the
ane cover numbers.

The complexity of the line cover numbers in 2D and 3D. We begin by showing
that it is R-hard to compute, for a given graph G, its line cover numbers 12 (G)
and 13 (G); see Section 2.
Our proof uses some ingredients from a paper of Durocher et al. [10] who
showed that it is NP-hard to compute the segment number segm(G) of a graph G.
This parameter was introduced by Dujmovic et al. [7] as a measure of the vi-
sual complexity of a planar graph. A segment in a straight-line drawing of a
graph G is an inclusion-maximal connected path of edges of G lying on a line,
and the segment number segm(G) of a planar graph G is the minimum num-
ber of segments in a straight-line drawing of G in the plane. Note that while
12 (G) segm(G), the parameters can be far apart, e.g., as shown by a graph
with m isolated edges. For connected graphs, we have shown earlier [6] that
segm(G) O(12 (G)2 ) and that
this bound is optimal as there exist planar tri-
angulations with 12 (G) O( n) and segm(G) (n). Still, we follow Durocher
et al. [10] to some extent in that we also reduce from Arrangement Graph
Recognition (see Theorem 1).

Parameterized complexity of computing the line cover numbers in 2D and 3D. It

follows from the inclusion NP R that the decision problems 12 (G) k and
13 (G) k are NP-hard if k is given as a part of the input. On the positive side,
in Section 3, we show that both problems are xed-parameter tractable. To this
end, we rst describe a linear-time kernelization procedure that reduces the given
2
graph to one of size O(k 4 ). Then, in k O(k ) time, we carefully solve the problem
on this reduced instance by using the exponential-time decision procedure for
the existential theory of the reals by Renegar [20,21,22] as a subroutine. To the
best of our knowledge, this is the rst application of Renegars algorithm for
obtaining an FPT result, in particular, in the area of graph drawing where FPT
algorithms are widely known.

The space complexity of 1d -optimal drawings. Since R belongs to PSPACE (as

shown by Canny [3]), the parameters 1d (G) for both d = 2 and 3 are com-
putable in polynomial space. On the negative side, we construct a graph G with
a 12 -optimal drawing requiring irrational coordinates; we provide a more com-
plex argument to show that any 12 -optimal drawing of G requires irrational
coordinates; for details see the full version [5].

The complexity of the plane cover number. Though the decision problem 23 (G)
k also belongs to R, its complexity status is dierent from that of the line
cover numbers. In Section 4, we establish the NP-hardness of deciding whether
23 (G) k for any xed k 2, which excludes an FPT algorithm for this problem
268 S. Chaplick et al.

unless P = NP. To show this, we rst prove NP-hardness of Positive Planar

Cycle 1-in-3-Sat (a new problem of planar 3-SAT type), which we think is of
independent interest.

Weak ane cover numbers. We previously dened the weak ane cover number
dl (G) of a graph G similarly to ld (G) but under the weaker requirement that
the l-dimensional planes in Rd whose number has to be minimized contain the
vertices (and not necessarily the edges) of G [6]. Based on our combinatorial
characterization of 31 and 32 [6], we show in Section 5 that the decision problem
3l (G) 2 is NP-complete, and that it is NP-hard to approximate 3l (G) within
a factor of O(n1 ), for any  > 0. Asymmetrically to the ane cover numbers
12 , 13 , and 23 , here it is the parameter 21 (for planar graphs) whose complexity
remains open. For more open problems, see Section 6.

2 Computational Hardness of the Line Cover Numbers

In this section, we show that deciding, for a given graph G and integer k, whether
12 (G) k or 13 (G) k is an R-complete problem. The R-hardness results
are often established by a reduction from the Pseudoline Stretchability
problem: Given an arrangement of pseudolines in the projective plane, decide
whether it is stretchable, that is, equivalent to an arrangement of lines [17,18].
Our reduction is based on an argument of Durocher et al. [10] who designed a
reduction of the Arrangement Graph Recognition problem, dened below,
to the problem of computing the segment number of a graph.
A simple line arrangement is a set L of k lines in R2 such that each pair
of lines has one intersection point and no three lines share a common point. In
the following, we assume that every line arrangement is simple. We dene the
arrangement graph for a set of lines as follows [2]: The vertices correspond to the
intersection points of lines and two vertices are adjacent in the graph if and only
if they are adjacent along some line. The Arrangement Graph Recognition
problem is to decide whether a given graph is the arrangement graph of some
set of lines.
Bose et al. [2] showed that this problem is NP-hard by reduction from a
version of Pseudoline Stretchability for the Euclidean plane, whose NP-
hardness was proved by Shor [25]. It turns out that Arrangement Graph
Recognition is actually an R-complete problem [11, page 212]. This stronger
statement follows from the fact that the Euclidean Pseudoline Stretchabil-
ity is R-hard as well as the original projective version [16,23].

Theorem 1. Given a planar graph G and an integer k, it is R-hard to decide

whether 12 (G) k and whether 13 (G) k.

Proof. We rst treat the 2D case. We show hardness by a reduction from Ar-
rangement Graph Recognition. Let G be an instance of this problem. If
G is an arrangement graph, there must be an integer  such that G consists of
( 1)/2 vertices and ( 2) edges, and each of its vertices has degree d where
The Complexity of Drawing Graphs on Few Lines and Few Planes 269

d [2, 4]. So, we rst check these easy conditions to determine  and reject G if
one of them fails. Let G be the graph obtained from G by adding one tail (i. e.,
a degree-1 vertex) to each degree-3 vertex and two tails to each degree-2 vertex.
So every vertex of G has degree 1 or 4. Note that, if G is an arrangement graph,
then there are exactly 2 tails in G (2 for each line) if this is not true we can
already safely reject G. We now pick k = , and show that G is an arrangement
graph if and only if 12 (G ) k.
For the rst direction, let G be an arrangement graph. By our choice of k,
it is clear that G corresponds to a line arrangement of k lines. Clearly, all edges
of G lie on these k lines and the tails of G can be added without increasing the
number of lines. Hence, 12 (G ) k.
For the other direction, assume 12 (G ) k and let
 be a straight-line
drawing of G on 12 (G ) lines. The graph G contains k2 degree-4 vertices.
As each of these vertices lies on the intersection of two lines in  , we need k
lines to get enough intersections, that is, 12 (G ) = k. Additionally, there are
no intersections of more than two lines. The most extreme points on any line
have degree 1, that is, they are tails, because degree 4 would imply a more
extreme vertex. We can assume that there are exactly 2k tails, otherwise G
would have been rejected before as it could not be an arrangement graph. Each
line contains exactly two of them. Let n2 (resp. n3 ) be the number of degree-2
(resp. degree-3) vertices. As we added 2 (resp. 1) tails to each of these vertices,
we have 2k = 2n2 + n3 . By contradiction, we show that the edges on each line
form a single segment. Otherwise, there would be a line with two segments.
Note that the vertices at the ends of each segment have degree less than 4 (that
is, degree 1). This would imply more than two degree-1 vertices on one line, a
contradiction. So  is indeed a drawing of G using k segments. By removing
the tails, we obtain a straight-line drawing of G using k = n2 + n3 /2 segments.
The result by Durocher et al. [10, Lemma 2] implies that G is an arrangement
graph.
Now we turn to 3D. Let G be a graph and let G be the augmented graph as
above. We show that 13 (G ) = 12 (G ), which yields that deciding 13 (G ) is also
NP-hard. Clearly, 13 (G ) 12 (G ).
Conversely,
 assume that G can be drawn
 k
on k lines in 3-space. Since G has 2 vertices of degree 4, each of them must
be a crossing point of two lines. It follows that each of the k lines crosses all
the others. Fix any two of the lines and consider the plane that they determine.
Then all k lines must lie in this plane, which shows that 12 (G ) 13 (G ).


It remains to notice that the decision problems under consideration lie in

the complexity class R. To this end, we transform the inequalities ld (G) k
into rst-order existential expressions about the reals. For details, see the full
version [5].

Lemma 2. Each of the following decision problems belongs to the complexity

class R
(a) deciding, for a planar graph G and an integer k, whether 12 (G) k;
(b) deciding, for a graph G and an integer k, whether 13 (G) k;
270 S. Chaplick et al.

(c) deciding, for a graph G and an integer k, whether 23 (G) k.

3 Fixed-Parameter Tractability of the Line Cover

Numbers

In this section we show that, for an input graph G and integer k, both testing
whether 12 (G) k, and testing whether 13 (G) k are decidable in FPT time
(in k). Moreover, for both the 2D and 3D cases, for positive instances (G, k), we
can compute the combinatorial description of a solution also in time FPT in k.
One subtle point here is that there are graphs where each 12 -optimal drawing
requires irrational coordinates; see the full version for details [5]. Thus, in some
sense, a combinatorial description of a solution can be seen as a best possible
output from an algorithm for these problems. Note that, by a k-line cover in Rd
of a graph G, we mean a drawing D of G together with a set L of k lines such
that (D, L) certies 1d (G) k.
Our FPT algorithm follows from a simple kernelization/pre-processing pro-
cedure in which we reduce a given instance (G, k) to a reduced instance (H, k)
where H has O(k 4 ) vertices and edges, and G has a k-line cover if and only if
H does as well. After this reduction, we can then apply any decision procedure
for the existential theory of the reals since we have shown in Lemma 2 that
both k-line cover problems are indeed both members of this complexity class.
Our kernelization approach is given as Theorem 3 and our FPT result follows
as described in Corollary 4. We denote the number of vertices and the number
of edges in the input graph by n and m respectively.

Theorem 3. For each d {2, 3}, graph G, and integer k, the problem of de-
ciding whether 1d (G) k admits a kernel of size O(k 4 ), i.e., we can produce a
graph H such that H has O(k 4 ) vertices and edges and 1d (G) k if and only if
1d (H) k. Moreover, H can be computed in O(n + m) time.

Proof. For a graph G, if G is going to have a k-line cover (D, L), then there are
several necessary conditions about G which we can exploit to shrink G. First,
notice that any connected components of G which are paths can easily be placed
on any line in L without interfering with the other components, i.e., these can

be disregarded. This provides a new instance G . Second, there are at most k2


intersection points among the lines in L. Thus, G has at most k2 vertices with
degree larger than two. Moreover, each line  L will contain at most k 1
of these vertices. Thus, the total number of edges which are incident to vertices
with degree larger than two, is at most 2 (k 1) per line, or 2 (k 2 k) in
total. Thus, G contains at most 2 (k 2 k) vertices of degree one (since each
one occurs at the end of a path originating from a vertex of degree larger than
two where all the internal vertices have degree 2). Similarly, G contains at most
2 (k 2 k) paths where every internal vertex has degree two and the end vertices
either have
 degree one or degree larger than two. Finally, for each such path,
at most k2 vertices are mapped to intersection points in L. Thus, any path
The Complexity of Drawing Graphs on Few Lines and Few Planes 271


with more than k2 vertices can be safely contracted to a path with at most

k  
2 vertices. This results in our nal graph G which can easily be seen to have
O(k ) vertices and O(k ) edges (when G has a k-line cover). Now, if G does
4 4

not satisfy one of the necessary conditions described above, we use the graph
K1,2k+1 as H, i.e., this way H has no k-line cover.
We conclude by remarking that this transformation of G to G can be per-
formed in O(n + m) time. The transformation from G to G is trivial. The
transformation from G to G can be performed by two traversals of the graph
(e.g., breadth rst searches) where we rst measure the lengths of the paths of
degree-2 vertices, then we shrink them as needed.

In the notation of the above proof, note that the statement 1d (G ) k can
be expressed as a prenex formula in the existential rst-order theory of the
reals. The proof of Lemma 2 shows that such a formula can be written using
O(k 4 ) rst-order variables and involving O(k 4 ) polynomial inequalities, each of
total degree at most 4 and with coecients 1. We could now directly apply
the decision procedure of Renegar [20,21,22] to and obtain an FPT algorithm
for deciding whether 1d (G) k, but that would only provide a running time of
4
(k O(k ) + O(n + m)). We can be a little more clever and reduce the exponent
from O(k 4 ) to O(k 2 ). This is described in the proof of the following corollary.
Corollary 4. For each d {2, 3}, graph G, and integer k, we can decide whether
2
1d (G) k in k O(k ) + O(n + m) time, i.e., FPT time in k.
Proof. First, we apply to the given graph G the kernelization procedure from the
 
proof of Theorem 3 to obtain a reduced graph
k G . Now, notice that G has at
4
most O(k ) vertices of degree two, but only 2 of these can be bend points and


are actually important in a solution, i.e., at most k2 of these vertices are mapped
to intersection points of the lines. Thus, we can simply enumerate all possible

4 
O (kk) subsets which will occur as intersection points, and, for each of these,
2
test whether this further reduced instance has a k-line cover using Renegars
2
decision algorithm. This leads to a total running time of k O(k ) + O(n + m) as
needed.

We have now seen how to decide if a given graph G has a k-line cover in both
2D and 3D. Moreover, when G is a positive instance, our approach provides a
reduced graph G where G also has a k-line cover, G has O(k 2 ) vertices and
edges, and any k-line cover of G naturally induces a k-line cover of G. In the
following theorem, whose proof can be found in the full version [5], we show that
we can further determine the combinatorial structure of some k-line cover of G
2
in k O(k ) time and use this to recover a corresponding combinatorial structure
for G. Here, the combinatorial structure is a set of k linear forests since each
line in a k-line cover naturally induces a linear forest in G. Recall that a linear
forest is a forest whose connected components are paths.
3
Theorem 5. For each d {2, 3}, graph G, and integer k, in 2O(k ) + O(n + m)
time we can not only decide whether 1d (G) k but, if so, also partition the edge
set of G into linear forests accordingly to a k-line cover of G in Rd .
272 S. Chaplick et al.

4 Computational Complexity of the Plane Cover Number

While graphs with 23 -value 1 are exactly the planar graphs, recognizing graphs
with 23 -value k, for any k > 1, immediately becomes NP-hard. This requires a
detour via the NP-hardness of a new problem of planar 3-SAT type, which we
think is of independent interest. Full proofs for this section are given in the full
version [5].

Denition 6 ([19]). Let be a Boolean formula in 3-CNF. The associated

graph of , G(), has a vertex vx for each variable x in and a vertex vc for
each clause c in . There is an edge between a variable-vertex vx and a clause-
vertex vc if and only if x or x appears in c. The Boolean formula is called
planar if G() is planar.

Kratochvl et al. [14] proved NP-hardness of Planar Cycle 3-Sat, which is

a variant of Planar 3-Sat where the clauses are connected by a simple cycle in
the associated graph without introducing crossings. Their reduction even shows
hardness of a special case, where all clauses consist of at least two variables.
We consider only this special case. Mulzer and Rote [19] proved NP-hardness
of Positive Planar 1-in-3-Sat, another variant of Planar 3-Sat where all
literals are positive and the assignment must be such that, in each clause, exactly
one of the three variables is true. We combine proof ideas from the two to show
NP-hardness of the following new problem.

Denition 7. In the Positive Planar Cycle 1-in-3-Sat problem, we are

given a collection of clauses each of which contains exactly three variables,
together with a planar embedding of G() + C where C is a cycle through all
clause-vertices. Again, all literals are positive. The problem is to decide whether
there exists an assignment of truth values to the variables of such that exactly
one variable in each clause is true.

Lemma 8. Positive Planar Cycle 1-in-3-Sat is NP-complete.

Proof (sketch). We reduce from Planar Cycle 3-Sat. We iteratively replace

the clauses by positive 1-in-3-Sat clauses while maintaining the cycle through
these clauses. Our reduction uses some of the gadgets from the proof of Mulzer
and Rote [19]. We show how to maintain the cycle when inserting these gadgets.
We consider the interaction between the cycle and the clauses. Every clause
consists of two or three literals and thus there are two or three faces around a
clause in the drawing. There are two options for the cycle: (O1) it can touch
the clause, that is, the incoming and the outgoing edge are drawn in the same
face; (O2) it can pass through the clause, that is, incoming and outgoing edge
are drawn in dierent faces. As an example, Fig. 1a shows how we weave the
cycle through the inequality gadget by Mulzer and Rote. As a replacement for
the clauses with 2 variables we cannot use the gadget described by Mulzer and
Rote as it does not allow us to add a cycle through the clauses. Therefore, we
use a new gadget that is depicted in Fig. 1b.

The Complexity of Drawing Graphs on Few Lines and Few Planes 273

O2
b O1
a
=
x y y
x
c
=
O1 a
d b c O2

(a) Mulzer and Rotes gadget for x

= y. (b) Our gadget for the clause (x y).

x O1 u1
O2
r u2
a u
= e u1
v1 ,

= v2 ,
q v3
b d
c
= u4 u3
y z
(d) The intersection line gadget
(c) Mulzer and Rotes gadget for the clause: xyz. and how it is depicted in Fig. 2.

Fig. 1: Gadgets for our NP-hardness proof. Variables are drawn in circles, clauses are
represented by squares. The boxes with the inequality sign represent the inequality
gadget. The dashed line shows how we weave the cycle through the clauses. There are
two variants of the cycle, which dier only in one edge: (O1) The cycle touches the
gadget; (O2) the cycle passes through the gadget.

We now introduce what we call the intersection line gadget; see Fig. 1d. It
consists of a K3,4 in which the vertices in the smaller set of the bipartition
denoted by v1 , v2 , and v3 are connected by a path. We denote the vertices in
the other set by u1 , u2 , u3 , and u4 .
Lemma 9. If a graph containing the intersection line gadget can be embedded
on two non-parallel planes, the vertices v1 , v2 , and v3 must be drawn on the
intersection line of the two planes while the vertices u1 , u2 , u3 , and u4 cannot
lie on the intersection line.

Theorem 10. Let G be a graph. Deciding whether 23 (G) = 2 is NP-hard.

Proof (sketch). We show NP-hardness by reduction from Positive Planar

Cycle 1-in-3-Sat. We build the graph G () = (V, E) for formula that
consists of n clauses as follows: Each clause c is represented by a clause gadget
that consists of three vertices vc1 , vc2 , and vc3 that are connected by a path. Let
x be a variable that occurs in the clauses ci1 , ci2 , . . . , cil with i1 < i2 < < il .
Each variable x is represented by a tree with the vertices w1x , w2x , . . . , wlx that
are connected to the relevant clauses, and the vertices v1x , v2x , . . . , vlx that lie on a
path and are connected to these vertices. To each of the vertices v1x , v2x , . . . , vlx one
274 S. Chaplick et al.

v1x v2x x v3x

w1x w2x w3x

v1b v2b v7b

vc11 vc31vc12 vc32 vc17 vc37

Fig. 2: Example for the graph G () constructed from a Positive Planar Cycle
1-in-3-Sat instance . The clauses are depicted by the black boxes with three vertices
inside and denoted by c1 , . . . , c7 from left to right. The variables are drawn in pale
red (true) and blue (false). The variable x is highlighted by a shaded background. The
ellipses attached to variable-vertices stand for the intersection line gadget (see Fig. 1d).
The depicted vertices incident to the gadget correspond to u1 in Fig. 1d; u2 to u4 are
not shown. If is true, one plane covers the blue variable gadgets and one plane covers
the blocking caterpillar (bold black) and the pale red variable gadgets.

instance of the intersection line gadget is connected. Finally, we add a blocking

caterpillar consisting of the vertices v1b , . . . , vnb that are connected to the clauses
in their cyclic order, which exists for the Positive Planar Cycle 1-in-3-Sat
instance by denition. See Fig. 2 for an example of this construction.
We show that the formula has a truth assignment with exactly one true
variable in each clause if and only if the graph G () can be drawn onto two
planes. The idea of our construction is that only two variables can be connected
to a clause gadget on each of the planes. One plane contains the blocking cater-
pillar and one variable per clause (corresponding to true variables). The other
plane contains two variables per clause (corresponding to false variables). Our
construction ensures that the vertices of a variable cannot be partitioned onto
both planes in any drawing.

Corollary 11. Deciding whether 23 (G) = k is NP-hard for any k 2.
Proof (sketch). We add a blocking gadget for each additional plane.


5 Complexity of the Weak Ane Cover Numbers 31 / 32

Recall that a linear forest is a forest whose connected components are paths.
The linear vertex arboricity lva(G) of a graph G equals the smallest size r of a
partition V (G) = V1 Vr such that every Vi induces a linear forest. The
vertex thickness vt(G) of a graph G is the smallest size r of a partition V (G) =
V1 Vr such that G[V1 ], . . . , G[Vr ] are all planar. Obviously, vt(G) lva(G).
We recently used these notions to characterize the 3D weak ane cover numbers
in purely combinatorial terms [6]: 31 (G) = lva(G) and 32 (G) = vt(G).
The Complexity of Drawing Graphs on Few Lines and Few Planes 275

Theorem 12. For l {1, 2},

(a) deciding whether or not 3l (G) 2 is NP-complete, and
(b) approximating 3l (G) within a factor of O(n1 ), for any  > 0, is NP-hard.

Proof. (a) The membership in NP follows directly from the above combinatorial
characterization [6], which also allows us to deduce NP-hardness from a much
more general hardness result by Farrugia [12]: For any two graph classes P and
Q that are closed under vertex-disjoint unions and taking induced subgraphs,
deciding whether the vertex set of a given graph G can be partitioned into two
parts X and Y such that G[X] P and G[Y ] Q is NP-hard unless both P
and Q consist of all graphs or all empty graphs. To see the hardness of our two
problems, we set P = Q to the class of linear forests (for l = 1) and to the class
of planar graphs (for l = 2).
(b) The combinatorial characterization [6] given above implies that (G)
4 vt(G) = 432 (G) (by the four-color theorem). Note that each color class can be
placed on its own line, so 31 (G) (G). As 32 (G) 31 (G), both parameters are
linearly related to the chromatic number of G. Now, the approximation hardness
of our problems follows from that of the chromatic number [26].


6 Conclusion and Open Problems

1. We have determined the computational complexity of the ane cover numbers
12 and 13 . The corresponding decision problems 12 (G) k and 13 (G) k
turn out to be R-complete. On the positive side, these problems admit an
FPT algorithm (Corollary 4). This is impossible for the plane cover number
23 , unless P = NP, because the decision problem 23 (G) k is NP-hard even
for k = 2 (Theorem 10 in Section 4). If k is arbitrary and given as a part of
the input, then this problem is in R (Lemma 2)but is it R-hard?
2. Is the segment number segm(G) introduced in [7] xed-parameter tractable?
3. Our proof of Theorem 1 implies that computing 12 (G) and 13 (G) is hard even
for planar graphs of maximum degree 4. Can 12 (G) and 13 (G) be computed
eciently for trees? This is true for the segment number segm(G) [7].
4. How hard is it to approximate 12 , 13 , and 23 ?

References
1. Bienstock, D.: Some provably hard crossing number problems. Discrete Comput.
Geom. 6, 443459 (1991)
2. Bose, P., Everett, H., Wismath, S.K.: Properties of arrangement graphs. Int. J.
Comput. Geom. Appl. 13(6), 447462 (2003)
3. Canny, J.F.: Some algebraic and geometric computations in PSPACE. In: Proc.
20th Annu. ACM Symp. Theory Comput. (STOC88). pp. 460467 (1988)
4. Cardinal, J.: Computational geometry column 62. SIGACT News 46(4), 6978
(2015)
5. Chaplick, S., Fleszar, K., Lipp, F., Ravsky, A., Verbitsky, O., Wol, A.: The com-
plexity of drawing graphs on few lines and few planes (2016), arxiv.org/1607.06444
276 S. Chaplick et al.

6. Chaplick, S., Fleszar, K., Lipp, F., Ravsky, A., Verbitsky, O., Wol, A.: Drawing
graphs on few lines and few planes. In: Hu, Y., Nollenburg, M. (eds.) GD 2016.
LNCS, vol. 9801, pp. 166180. Springer-Verlag (2016), arxiv.org/1607.01196
7. Dujmovic, V., Eppstein, D., Suderman, M., Wood, D.R.: Drawings of planar graphs
with few slopes and segments. Comput. Geom. Theory Appl. 38(3), 194212 (2007)
8. Dujmovic, V., Por, A., Wood, D.R.: Track layouts of graphs. Discrete Math. &
Theor. Comput. Sci. 6(2), 497522 (2004)
9. Dujmovic, V., Whitesides, S.: Three-dimensional drawings. In: Tamassia, R. (ed.)
Handbook of Graph Drawing and Visualization, chap. 14, pp. 455488. CRC Press
(2013)
10. Durocher, S., Mondal, D., Nishat, R.I., Whitesides, S.: A note on minimum-segment
drawings of planar graphs. J. Graph Alg. Appl. 17(3), 301328 (2013)
11. Eppstein, D.: Drawing arrangement graphs in small grids, or how to play planarity.
J. Graph Alg. Appl. 18(2), 211231 (2014)
12. Farrugia, A.: Vertex-partitioning into xed additive induced-hereditary properties
is NP-hard. Electr. J. Comb. 11(1) (2004)
13. Kang, R.J., Muller, T.: Sphere and dot product representations of graphs. Discrete
Comput. Geom. 47(3), 548568 (2012)
14. Kratochvl, J., Lubiw, A., Nesetril, J.: Noncrossing subgraphs in topological lay-
outs. SIAM J. Discrete Math. 4(2), 223244 (1991)
15. Kratochvl, J., Matouek, J.: Intersection graphs of segments. J. Comb. Theory,
Ser. B 62(2), 289315 (1994)
16. Matouek, J.: Intersection graphs of segments and R. arxiv.org/1406.2636 (2014)
17. Mnev, N.E.: On manifolds of combinatorial types of projective congurations and
convex polyhedra. Sov. Math., Dokl. 32, 335337 (1985)
18. Mnev, N.E.: The universality theorems on the classication problem of congu-
ration varieties and convex polytopes varieties. In: Viro, O. (ed.) Topology and
Geometry, Rohlin Seminar 19841986, Lect. Notes Math., vol. 1346, pp. 527543.
Springer-Verlag, Berlin (1988)
19. Mulzer, W., Rote, G.: Minimum-weight triangulation is NP-hard. J. ACM 55(2),
11 (2008)
20. Renegar, J.: On the computational complexity and geometry of the rst-order the-
ory of the reals, part I: Introduction. preliminaries. the geometry of semi-algebraic
sets. the decision problem for the existential theory of the reals. J. Symb. Comput.
13(3), 255300 (1992)
21. Renegar, J.: On the computational complexity and geometry of the rst-order the-
ory of the reals, part II: The general decision problem. preliminaries for quantier
elimination. J. Symb. Comput. 13(3), 301328 (1992)
22. Renegar, J.: On the computational complexity and geometry of the rst-order
theory of the reals, part III: Quantier elimination. J. Symb. Comput. 13(3), 329
352 (1992)
23. Schaefer, M.: Complexity of some geometric and topological problems. In: Epp-
stein, D., Gansner, E.R. (eds.) GD 2009. LNCS, vol. 5849, pp. 334344. Springer-
Verlag (2010)
24. Schaefer, M., Stefankovic, D.: Fixed points, Nash equilibria, and the existential
theory of the reals. Theory Comput. Syst. pp. 122 (2015)
25. Shor, P.W.: Stretchability of pseudolines is NP-hard. In: Gritzmann, P., Sturm-
fels, B. (eds.) Applied Geometry and Discrete MathematicsThe Victor Klee
Festschrift, DIMACS Series, vol. 4, pp. 531554. Amer. Math. Soc. (1991)
26. Zuckerman, D.: Linear degree extractors and the inapproximability of max clique
and chromatic number. Theory Comput. 3(1), 103128 (2007)
Modular Circulation and Applications to Trac
Management

Philip Dasler and David M. Mount

Department of Computer Science and

Institute for Advanced Computer Studies
University of Maryland, College Park, Maryland 20742
{daslerpc,mount}@cs.umd.edu

Abstract. We introduce a variant of the well-known minimum-cost cir-

culation problem in directed networks, where vertex demand values are
taken from the integers modulo , for some integer 2. More formally,
given a directed network G = (V, E), each of whose edges is associated
with a weight and each of whose vertices is associated with a demand
taken over the integers modulo , the objective is to compute a ow of
minimum weight that satises all the vertex demands modulo . This
problem is motivated by a problem of computing a periodic schedule for
trac lights in an urban transportation network that minimizes the total
delay time of vehicles. We show that this modular circulation problem is
solvable in polynomial time when = 2 and that the problem is NP-hard
when = 3. We also present a polynomial time algorithm that achieves
a 4( 1)-approximation.

Keywords: Network ows and circulations, Trac management, Ap-

proximation algorithms, NP-hard problems

1 Introduction

Minimum (and maximum) cost network ows and the related concept of circula-
tions are fundamental computational problems in discrete optimization. In this
paper, we introduce a variant of the circulation problem, where vertex demand
values are taken from the integers modulo , for some integer 2. For ex-
ample, if = 10 a vertex with demand 6 can be satised by any net incoming
ow of 6, 16, 26 and so on or a net outgoing ow of 4, 14, 24, and so on. Our
motivation in studying this problem stems from an application in synchronizing
the trac lights of an urban transportation system.
Throughout, let G = (V, E) denote a directed graph, and let 2 be an
integer. Each edge (u, v) E is associated with a nonnegative integer weight,
wt(u, v), and each vertex u V is associated with a demand, d(u), which is an
integer drawn from Z , the integers modulo . Let f be an assignment of values

Research supported by NSF grant CCF-1618866.

Springer International Publishing AG 2017 277

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 277288, 2017.
DOI: 10.1007/978-3-319-62127-2_24
278 P. Dasler and D.M. Mount

from Z to the edges of G. For each vertex v V , dene

 
fin (v) = f (u, v) and fout (v) = f (v, w),
(u,v)E (v,w)E

and dene the net ow into a vertex v to be fin (v) fout (v). We say that f
is a circulation with -modular demands, or -CMD for short, if it satises the
modular ow-balance constraints, which state that for each v V ,

fin (v) fout (v) d(v) (mod ).

Observe that a demand of d(v) is equivalent to the modular supply requirement

that the net ow out of this vertex modulo is d(v).
Dene the cost of a circulation f to be the weighted sum of the ow values
on all the edges, that is,

cost(f ) = wt(u, v) f (u, v).
(u,v)E

Given a directed graph G and the vertex demands d, the -CMD problem is
that of computing a -CMD of minimum cost. (Observe that there is no loss in
generality in restricting the ow value on each edge to Z , since the cost could be
reduced by subtracting from this value without aecting the ows validity.)
The standard minimum-cost circulation problem (without the modular as-
pect) is well studied. We refer the reader to any of a number of standard sources
on this topic, for example, [1, 2, 6, 8]. In contrast, -CMD is complicated by the
wrap-around eect due to the modular nature of the demand constraints. A
vertexs demand of d(u) units can be satised in the traditional manner by hav-
ing a net incoming ow of d(u), but it could also be met by generating a net
outgoing ow of d(u) (not to mention all variants thereof that involve adding
multiples of ). Our main results are:

2-CMD can be solved exactly in polynomial time (see Section 4).

3-CMD is NP-hard (see Section 5).
There is a polynomial time 4(1)-approximation to -CMD (see Section 6).

In Section 2 we discuss the relevance of the -CMD problem to a trac-

management problem. In Section 3 we present some preliminary observations
regarding this problem. In Sections 46 we present each of our three main results.

2 Application to Trac Management

Our motivation in studying the -CMD problem arises from an application in

trac management. In urban settings, intersections are the shared common re-
source between vehicles traveling in dierent directions, and their control is
essential to maximizing the utilization of a transportation network [9]. There
Modular Circulation and Applications to Traffic Management 279

are numerous approaches to modeling trac ow and diverse computational ap-

proaches to solve and analyze the associated trac management problems [4,11].
Despite the popular interest in automated trac systems, there has been rela-
tively little work on this problem from the perspective of algorithm design.
In an earlier paper [3], we considered the problem of scheduling the move-
ments of a collections of vehicles through a system of unregulated crossing. Our
approach was based on the idealized assumption that the motion of individual
vehicles in the system is controlled by a central server. A more practical ap-
proach is based on aggregating vehicles into groups, or platoons, and planning
motion at the motion of these groups [7, 10].
We consider the problem in this aggregated form, but from a periodic perspec-
tive. Consider an urban transportation network consisting of a grid of horizontal
and vertical roads as laid out on a map. Each pair of horizontal and vertical
roads meets at a unique intersection controlled by a trac light that alternates
between horizontal and vertical trac, such that the pattern repeats over a time
interval . We assume throughout that has been discretized to a reasonably
small integer value, say in terms of seconds or tens of seconds.
More formally, we say that a trac-light schedule is -periodic if repeats
every time units. We consider a trac management system of the foreseeable
future where the trac light schedule is transmitted to the vehicles, which in
turn may adjust their speeds to avoid excessive waiting at intersections. While
vehicles may turn at intersections, the schedule is designed to minimize the delay
of straight-moving trac.
To motivate the connection with modular circulations, consider a four-sided
city block (see Fig. 1). Let a, b, c, and d denote the intersections, and let tab , tbc ,
tcd , tad denote the travel times between successive intersections along each road
segment. If the road segment is oriented counterclockwise around the block (as
shown in our example), these travel times are positive, and otherwise they are
negative. Suppose that the trac-light schedule is -periodic, and that at time
t = ta the light at intersection a transitions so that the eastbound trac can
move horizontally through the intersection (see Fig. 1(a)). In order for these ve-
hicles to proceed without delay through intersection b, this light must transition
from vertical to horizontal at time tb = ta + tab (see Fig. 1(b)).

t=0 t = tab t = tab + tbc t = tab + tbc + tcd + tda

tcd
d c d c d c d c

tbc tbc tda

t=0 a b a b a b a b
tab

(a) (b) (c) (d)

Fig. 1. Delay-free trac-light schedule.

280 P. Dasler and D.M. Mount

Reasoning analogously, for the other intersections, it follows that the vertical-
to-horizontal transition times for intersections c and d are tc = tb + tbc and
td = tc + tcd , respectively (see Fig. 1(c)). On returning to a (see Fig. 1(d)), we
nd that
ta ta + (tab + tbc + tcd + tda ) (mod ).
Thus, in order to achieve delay-free ow around the intersection in a -periodic
context, we must satisfy the constraint

tab + tbc + tcd + tda 0 (mod ).

Since the transportation times along the road segments are not under our control,
in order to satisfy this constraint, we introduce an (ideally small) delay ij 0
along each road segment ij. This yields the new constraint

(tab + ab ) + (tbc + bc ) + (tcd + cd ) + (tda + da ) 0 (mod ),

or equivalently, if we dene T = tab + tbc + tcd + tda to be the sum of (signed)

travel times of the road segments around this block, we have

ab + bc + cd + da T (mod ). (1)

The upshot is that if vehicles travel at a reduced speed so that the transit time
along each of the road segments includes the associated delay, then the straight-
line vehicular trac along each road need never pause or wait at any trac
signal. The objective is to minimize the sum of delay values over all the road
segments in the network, which we refer to as the total delay.
More formally, the transportation network is modeled as a set of horizontal
and vertical roads. This denes a directed grid graph whose vertices are the
intersections, whose edges are the road segments, and whose (bounded) faces are
the blocks of the city. For each pair of adjacent intersections i and j, let tij denote
the delay-free travel time along this road segment. For each block u, dene the
total signed travel time about u to be the sum of the travel times for each of
the road segments bounding u, where the travel time is counted positively if
the segment is oriented counterclockwise about u and negatively otherwise. Let
T (u) denote this value modulo . A -periodic trac-light schedule assigns a
delay to each road segment so that for each block, these delays satisfy Eq. (1).
The objective is to minimize the total delay, which is dened to be the sum of
delays over all the segments in the network.
To express this in the form of an instance of -CMD, let G = (V, E) denote
the directed dual of the graph, by which we mean that the vertex set V consists
of the city blocks, and there is a directed edge (u, v) E if the two blocks are
incident to a common road segment, and the direction of the road segment is
counterclockwise about u (and hence, clockwise about v). The demand of each
vertex u, denoted d(u), is set to T (u), and the weight of each edge is set to unity.
There remains one impediment to linking the -periodic trac-light schedule
and the -CMD problems. The issue is that the delay associated with any road
segment (which may be as large as 1) can be signicantly larger than the
Modular Circulation and Applications to Traffic Management 281

time to traverse the road segment. If so, the capacity of the road segment to hold
the vehicles that are waiting for the next signal may spill backwards and block
the preceding intersection. In order to deal with this issue without complicating
our model, we introduce the assumption that is smaller than the time to
traverse any road segment. The link between the two problems is presented in
the following lemma. Due to space limitations, the proofs of this and many other
lemmas have been omitted and will appear in the full version of the paper.

Lemma 1. Given a transportation network and integer 2, let G be the as-

sociated directed graph with vertex demands and edge weights as described above.

(i) If there exists a -periodic trac-light schedule with total delay , then there
exists a -CMD for G of cost .
(ii) If there exists a -CMD for G of cost and for all road segments ij, tij ,
then there exists a -periodic trac-light schedule with total delay .

3 Preliminaries

In this section we present a few denitions and observations that will be used
throughout the paper. Given an instance G = (V, E) of the -CMD problem,
consider any subset V  V . Let G = (V  , E  ) be the associated induced sub-
graph of G, and let d(V  ) denote the sum of demands of all the nodes in V  . We
refer to E  as the internal edges of this subgraph, and we refer to the edges of G
that cross the cut (V  , V \ V  ) as the interface. Given such a subgraph and any
ow f on G, dene its internal ow to be only the ow on the internal edges,
and dene the internal cost to be the cost of the ow restricted to these edges.
Dene the interface ow and interface cost analogously for the interface edges.
Dene fin (V  ) to be the sum of ow values on the interface edges that are di-
rected into V  , and dene fout (V  ) analogously for outward directed edges. The
following lemma provides necessary and sucient conditions for the existence of
a -CMD.

Lemma 2. Given an instance G = (V, E) of the -CMD problem:

(i) For any induced subgraph G = (V  , E  ) and any -CMD f , we have

fin (V  ) fout (V  ) d(V  ) (mod ).

(ii) If G is weakly connected, then a -CMD exists for G if and only if d(V ) 0
(mod ).

It follows from this lemma the -CMD instance associated with any trac-
light scheduling problem has a solution. The reason is that each edge (u, v)
contributes its travel time tuv positively to d(u) and negatively to d(v), and
therefore the sum of demands over all the vertices of the network is zero, irre-
spective of the travel times.
282 P. Dasler and D.M. Mount

4 Polynomial Time Solution to 2-CMD

In this section we show that 2-CMD, which we also call binary CMD, can be
solved in polynomial time by a reduction to minimum-cost matching in general
graphs. Intuitively, the binary case is simpler because the edge directions are
not signicant. If a vertex is incident to an even number of ow-carrying edges
(whether directed into or out of this vertex), then the net ow into this vertex
modulo is zero, and otherwise it is one. Thus, solving the problem reduces
to computing a minimum-cost set of paths that connect each pair of vertices of
nonzero demand, which is essentially a minimum-cost perfect matching in a com-
plete graph whose vertex set consists of the subset vertices of nonzero demand
and whose edge weights are distances between vertices ignoring edge directions.
The remainder of this section is devoted to providing a formal justication of
this intuition.
Recall G = (V, E) is a directed graph, and d(v) denotes the demand of vertex
v. Since = 2, for each v V , we have d(v) {0, 1}. Let G = (V, E  ) denote
the graph on the same vertices as G but with directions removed from all the
edges. We may assume that G is connected, for otherwise it suces to solve the
problem separately on each connected component of G . We set the weight of
each edge of G to the weight of the corresponding edge of G. If there are two
oppositely directed edges joining the same pair of vertices, the weight is set to
the minimum of the two.
Let U = U (G) denote the subset of vertices of V whose demand values are
equal to 1. By Lemma 2(ii), we may assume that d(V ) 0 (mod ). Therefore,
d(V ) is even, which implies that |U | is also even. For each u, v U , let (u, v)
denote the shortest weight path between them in G , and let wt((u, v)) denote
this weight. Dene G = (U, E)
to be a complete, undirected graph on the vertex
set U , where for each u, v U , the weight of this edge wt((u, v)). (This is well
dened by our assumption that G is connected.) Since G is complete and has an
even number of vertices, it has a perfect matching. The reduction of 2-CMD to
the minimum-cost perfect matching problem is implied by the following lemma.
Lemma 3. lem:2-cmd Given an instance G = (V, E) of the 2-CMD problem,
the minimum cost of any 2-CMD for G is equal to the minimum cost of a perfect

matching in G.
Since G is dense, a minimum-cost perfect matching can be constructed in
O(|U |3 ) time. The graph can be computed in O(n3 ) time, where n = |V |, by
applying the Floyd-Warshall algorithm for computing shortest paths [2]. Thus,
the overall running time is O(n3 ).
Theorem 1. It is possible to solve the 2-CMD problem in O(n3 ) time on any
instance G = (V, E), where n = |V |.

5 Hardness of 3-CMD
In this section, we present the following hardness result for -CMD.
Modular Circulation and Applications to Traffic Management 283

Theorem 2. For 3, the -CMD problem is NP-hard.

The reduction is from positive 1-in-3-SAT [5]. For the sake of brevity and
simplicity, we show the proof for the case = 3 here, but the method easily
generalizes (as will be explained in the full version of the paper). Let F denote a
boolean formula in 3-CNF, where each literal is in positive form. Throughout, for
{0, 1, . . . , 1}, we use the term -vertex to denote a vertex whose demand
is . The reduction involves two principal components, a variable gadget which
associates truth values with the variables of F and a clause gadget which enforces
the condition that exactly one variable in each clause is assigned the value True.

5.1 Variable Gadget

Before discussing the general gadget, we describe a fundamental building block
from which all variables will be constructed. The block consists of six vertices,
three 1-vertices and three ( 1)-vertices (i.e., 2-vertices), connected together
with edges as shown in Figure 2(a). Edges connecting 1-vertices have weight
wt(u, v) = 1.5, while all other edges are of weight wt(u, v) = 1.
If a ow of 1 is sent from each 2-vertex to its connected 1-vertex, the 2-vertices
overow and all demands are satised with cost(f ) = 3 (see Figure 2(b)). This
ow, in which there is no ow across the interface edges, represents a logical
value of False.
If instead a ow of 1 is sent across the interface edges, then the demands
of the 2-vertices are satised. A ow of 1 across each edge originating at the
central 1-vertex will cause it to overow and will satisfy each of the connected
1-vertices. This ow, in which each interface edge carries a ow of 1, represents
a logical value of True and again has cost(f ) = 3 (see Figure 2(c)).
False True
1 1
1 1.5 1 1.5 1 1 1.5 1 1.5 1 1 1.5 1 1.5 1
1 1 1

2 2 2 2 2 2 2 2 2
1 1 1

(a) (b) (c)

Fig. 2. (a) The fundamental building block used to build variable gadgets. Interface
edges are dashed gray segments. (b,c) CMDs representing the assignment of False and
True values, respectively, with the ow values in boxes.

If every interface edge of a variable gadget carries the same ow and that ow
is either 0 or 1, that variable is said to be interface-consistent. If all variables
are consistent for a given ow, then that ow is said to be variable-consistent.
Notice that both logical values above are realized via interface-consistent ows.
Lemma 4. Given a fundamental block, a satisfying ow has cost 3 if and
only if that ow is interface-consistent.
284 P. Dasler and D.M. Mount

Proof. Each 2-vertex can only be satised by: (1) sending a ow of 1 across one
of its edges or (2) sending a ow of 2 across both of its edges (in both cases the
vertexs demand overows).
In the second case, the 2-vertex sends a ow of 2 to its neighboring 1-vertex.
As per Lemma 2(i), that vertex now requires a ow of 2 across its other edge in
order to have its demand satised. Together these ows come at a cost of 5 (one
of these edges has a weight of 1.5), therefor no 2-vertex may be satised by a
ow greater than 1 without a cost greater than 3.
There exists a satisfying ow for a fundamental block if and only if the total
ow across its interface edges is equivalent to 0 (mod ) (see Lemma 2). Given
this and the fact that no single interface ow may equal 2, the ows across the
interface edges must either all be 0 or all be 1, i.e., the over all ow must be
interface consistent for it to be a satisfying ow.
As variables may appear in multiple clauses, we need a mechanism by which
existing variables can be expanded. For this, we create an expansion module,
any number of which can be added to a variable so that there are three interface
edges for each clause in which that variable appears.
To understand how this module functions, let us rst look at the case when
two fundamental blocks are connected together. This connection occurs through
a shared 2-vertex, so that what was an interface edge for one block becomes the
connection between a 2-vertex and 1-vertex in the other block (see Figure 3).
Recall that the value assignment of a variable is determined by the direction of
ow from the 2-vertices, with ow along the interface representing True and
internal ow (i.e., ow to the connected 1-vertices) representing False. Because
the outgoing edges of the shared 2-vertex are simultaneously an interface edge
of one block and an internal edge of the other, pushing a ow across either edge
will assign opposing values to the blocks.

1 1
1 1 1 1 1 1

2 2 2 1 1 1 2 2 2 1 1 1
1 1
1 1

2 2 1 2 2

(a) (b)

Fig. 3. Two fundamental blocks connected via a shared 2-vertex, with gure (b) show-
ing a ow that is satisfying but not interface-consistent.

Knowing this, the module is constructed as a double-negative, ensuring that

it is assigned the same value as the variable it extends. A fundamental block
is used as a hub and to this hub we attach two more fundamental blocks (see
Figure 4). When attached to a variable, this module creates four new interface
edges and consumes one, thus extending the variable by three interface edges.
Modular Circulation and Applications to Traffic Management 285
1.5 1.5 1.5 1.5
1 1 1 1 1 1 2 1 1 1 2

2 2 2 1 1 1 2 2 2 1 2 1
1.5 1.5
1.5 1.5
1 1 1 1 1 1 1
2 2 2
2 2 2 2 2 2
1 1 1

(a) (b)

Fig. 4. (a) A fundamental block with a single extension module attached. (b) The same
structure rearranged to emphasize the two clause outputs, each with three interface
edges.

While the structure of the fundamental blocks is as described above, the

weighting must be adjusted to maintain equal costs between the True and False
states. Rather than each fundamental block having two edges of weight 1.5, only
the rightmost block in the expansion module has such edges; all others are of
weight 1. In this way, the minimal cost of a consistent satisfying ow across the
module is 8, regardless of the value assigned to the variable. This is easily veried
by assigning a truth value to the gadget (xing an interface-consistent ow on
the interface edges of 0 or 1) and then traversing the structure, satisfying the
demand in each vertex by assigning ow to its unused edge.
Given this, Lemma 4, and the fact that the expansion module is constructed
from fundamental blocks, we have the following:
Lemma 5. Given an expansion module, there exists an interface-consistent ow
of internal cost 8 (in either the True or False cases), and any other satisfying
ow has a strictly larger internal cost.
To construct a gadget for a variable vi , appearing in c(vi ) clauses, begin with
a fundamental block and connect c(vi ) 1 expansion modules to it. Doing so
provides c(vi ) interface edges and yields the following result:
Lemma 6. Given a variable gadget, there exists an interface-consistent ow of
internal cost 3 + 8[c(vi ) 1] (in either the True or False cases), and any other
satisfying ow has a strictly larger internal cost.

5.2 Clause Gadget

The basis for the clause gadget is a single 1-vertex with three incoming edges,
one for each literal. These edges have a weight wt(u, v) = and are connected
to the appropriate variables as their outgoing interface edges. If a single literal
is True, one of these edges will carry a ow of 1, satisfying the demand of the
clause vertex. If more literals are True, the demand underows and the vertex
is left unsatised. It is possible to satisfy the vertex by creating ows on these
edges greater than 1, but such ows can be made cost-prohibitive by setting the
edge weights suciently high.
286 P. Dasler and D.M. Mount

Recall that each variable gadget produces copies of its respective variable
(three interface edges in this example) per clause in which it appears. Because of
this, the clause gadget must also be created in triplicate. Every clause consists of
three 1-vertices, each with an incoming edge from its three literals (see Figure 5).
Their weighting and behavior are as described above. Since there are no internal
edges in the clause gadgets, they do not contribute to the cost of the ow (but
their interface edges will).

Fig. 5. A full clause gadget, with three inputs from each of three literals.

5.3 Final Construction

Each variable in F is represented by a fundamental block connected to c(vi ) 1

expansion modules, creating c(vi ) outputs. Thus, outputs are linked to each
of the appropriate clause gadgets. The size of the variable gadget is a linear
function of the number of clauses in which that variable appears and can thus
be constructed in polynomial time.
If F is satisable, then a 3-CMD exists that is variable-consistent. In this case,
each fundamental block incurs a cost of 3, and each expansion module incurs
an additional cost of 8 for a ow representing a consistent truth value across
its interface and the interfaces of the modules/fundamental block to which it is
attached, as per Lemma 5.
For each clause, create 1-vertices, each connected to the clauses three
literals by incoming edges. As there are no edges between these vertices, there is
no ow possible within the clause gadget, resulting in an internal cost of 0. The
size of the clause gadget is constant.
Finally, the ow on the edges between the variable gadgets and clause gadgets
has yet to be counted as they are interface edges for both gadgets. Each clause
gadget contains 1-vertices, with each receiving a ow of 1 across edges of weight
. Thus, these add a cost of 3|C|, where |C| is the number of clauses in F .
If F is not satisable, then some set of variables must have inconsistent out-
puts in order to create a valid CMD. As shown in Lemma 4, these inconsistencies
will always lead to a strictly greater cost. Thus:

Lemma 7. Given a positive boolean formula F in 3-CNF, in polynomial time

it is possible
 to construct an instance of 3-CMD that has a satisfying ow with
cost(f ) vi V (3 + 8[c(vi ) 1] + 3|C|) if and only if F is 1-in-3 satisable.
Modular Circulation and Applications to Traffic Management 287

6 Approximation Algorithm

In this section, we present an 4( 1)-factor approximation to the -CMD prob-

lem for 2. Before presenting the algorithm, we introduce some terminology.
Consider an instance G = (V, E) of the -CMD problem, with vertex demands
d. Let G = (V, E  ) be (as dened in Section 4) the undirected version of G. Let
us assume that G is connected. Let U = U (G) denote the subset of vertices of V
whose demand values are nonzero. Dene SMT (U ) to be a Steiner minimal tree
in G whose terminal set is U (that is, a connected subgraph of G of minimum
weight that contains all the vertices of U ).
As in Section 4, dene G = (U, E)
to be the complete, undirected graph over
the vertex set U , where for each u, v U , the weight of this edge is the weight of
a minimum weight path between u and v in G . Given any U V , let MST (U )
denote any minimum spanning tree on the subgraph of G induced on U . From
standard results on Steiner and minimum spanning trees we have the following.

Lemma 8. For any U V , wt(MST (U )) 2 wt(SMT (U )).

Dene a balanced partition to be a partition {U1 , . . . , Uk } of U such that

for 1 i k, the total demand within Ui (that is, d(Ui )) is equivalent to
zero modulo . By Lemma 2(ii), we may assume that d(V ) 0 (mod ), and
so there is always a trivial partition, namely {V } itself. Dene cost(Ui ) to be
cost(SMT (Ui )), and dene the cost of a balanced partition to be the sum of costs
over its components. A minimum balanced partition for G is a balanced parti-
tion of minimum cost. The following lemma establishes the connection between
balanced partitions and minimum modular circulations.

Lemma 9. Consider an instance G = (V, E) of -CMD. Let = (U1 , . . . , Uk )

denote a minimum balanced partition of G, as dened above, and let f denote
any minimum cost -CMD for G. Then cost( ) |f | ( 1) cost( ).

By the above lemma, it suces to compute a balanced partition for G of low

cost. We will present a simple approximation algorithm that outputs a balanced
partition whose cost is within a factor of 4 of the optimum.
The construction begins with the metric closure G dened above. In a manner
similar to Kruskals algorithm, we sort the edges of G in increasing order, and

start with each vertex of G in a separate component. All these components are
labeled as active. We process the edges one by one. Letting (u, v) denote the
next edge being processed, if u and v are in distinct components, and both
components are active, we merge these components into a single component. If
the sum of the demands of the vertices within this component is equivalent to
zero modulo , we label the resulting component as nished, and output its set of
vertices. Because the total sum of demands of all the nodes is equivalent to zero
modulo , it follows that every vertex is placed within a nished component, and
therefore the algorithm produces a balanced partition of G (and by extension, a
balanced partition of G).
288 P. Dasler and D.M. Mount

This algorithm has the same running time as Kruskals algorithm. (Observe
that we can associate each component with its sum of demands, thus enabling
us to determine the sum of merged components in constant time.) The following
lemma establishes the approximation factor for this construction.
Lemma 10. Let  denote the balanced partition generated by the above al-
gorithm, and let denote the optimum balanced partition. Then cost(  )
4 cost( ).
Combining Lemmas 9 and 10(ii), it follows that our algorithm achieves an
approximation factor of 4( 1). While obtaining the best running time has not
been a focus of this work, it is easy to see that this procedure runs in polynomial
time. Let n = |V |. The graph G can be computed in O(n3 ) time by the Floyd-
Warshall algorithm [2]. The Kruskal-like algorithm for computing the balanced
partition can be performed in O(n2 log n) time, as can the algorithm of Lemma 9.
Thus, the overall running time is O(n3 ).
Theorem 3. Given an instance G = (V, E) of the -CMD problem for 2,
it is possible to compute a 4( 1)-approximation in time O(n3 ), where n = |V |.

References
1. R. K. Ahuja, T. L. Magnanti, and J. B. Orlin. Network Flows: Theory, Algorithms,
and Applications. Prentice-Hall, 1993.
2. T. H. Cormen, C. Stein, R. L. Rivest, and C. E. Leiserson. Introduction to Algo-
rithms. McGraw-Hill Higher Education, 2nd edition, 2001.
3. P. Dasler and D. M. Mount. On the complexity of an unregulated trac crossing.
In Proc. 14th Internat. Sympos. Algorithms Data Struct., volume 9214 of Lecture
Notes Comput. Sci., pages 224235. Springer-Verlag, 2015.
4. K. M. Dresner and P. Stone. A multiagent approach to autonomous intersection
management. J. Artif. Int. Res., 31:591656, 2008.
5. M. R. Garey and D. S. Johnson. Computers and Intractability: A Guide to the
Theory of NP-Completeness. W. H. Freeman & Co., 1979.
6. A. V. Goldberg and R. E. Tarjan. Finding minimum-cost circulations by canceling
negative cycles. Journal of the ACM, 36:873886, 1989.
7. S. I. Guler, M. Menendez, and L. Meier. Using connected vehicle technology to
improve the eciency of intersections. Transportation Research Part C: Emerging
Technologies, 46:121131, 2014.
8. J. B. Orlin. A polynomial time primal network simplex algorithm for minimum
cost ows. Mathematical Programming, 78:109129, 1997.
9. R. Tachet, P. Santi, S. Sobolevsky, L. I. Reyes-Castro, E. Frazzoli, D. Helbing, and
C. Ratti. Revisiting street intersections using slot-based systems. PLOS ONE,
11(3):e0149607, 2016.
10. J. J. B. Vial, W. E. Devanny, D. Eppstein, and M. T. Goodrich. Scheduling
autonomous vehicle platoons through an unregulated intersection. In M. Goerigk
and R. Werneck, editors, 16th Wkshp. Alg. Approaches Transport. Model., Opt.,
and Syst. (ATMOS 2016), volume 54, pages 114, 2016.
11. J. Yu and S. M. LaValle. Multi-agent path planning and network ow. In E. Fraz-
zoli, T. Lozano-Perez, N. Roy, and D. Rus, editors, Algorithmic Foundations of
Robotics X, pages 157173, Berlin, Heidelberg, 2013. Springer Berlin Heidelberg.
 The Homogeneous Broadcast Problem in
Narrow and Wide Strips

Mark de Berg1 , Hans L. Bodlaender12 , and Sandor Kisfaludi-Bak1

1
TU Eindhoven, The Netherlands
[email protected]; [email protected]; [email protected]
2
Utrecht University, The Netherlands

Abstract. Let P be a set of nodes in a wireless network, where each

node is modeled as a point in the plane, and let s P be a given source
node. Each node p can transmit information to all other nodes within unit
distance, provided p is activated. The (homogeneous) broadcast problem
is to activate a minimum number of nodes such that in the resulting
directed communication graph, the source s can reach any other node.
We study the complexity of the regular and the hop-bounded version of
the problem (in the latter, s must be able to reach every node within a
specied number of hops), with the restriction that all points lie inside
a strip of width w. We almost completely characterize the complexity of
both the regular and the hop-bounded versions as a function of the strip
width w.

1 Introduction

Wireless networks give rise to a host of interesting algorithmic problems. In the

traditional model of a wireless network each node is modeled as a point p R2 ,
which is the center of a disk (p) whose radius equals the transmission range
of p. Thus p can send a message to another node q if and only if q (p). Using
a larger transmission radius may allow a node to transmit to more nodes, but
it requires more power and is more expensive. This leads to so-called range-
assignment problems, where the goal is to assign a transmission range to each
node such that the resulting communication graph has desirable properties, while
minimizing the cost of the assignment. We are interested in broadcast problems,
where the desired property is that a given source node can reach any other node
in the communication graph. Next, we dene the problem more formally.
Let P be a set of n points in Rd and let s P be a source node. A range
assignment is a function : P R0 that assigns a transmission range (p)
to each point p P . Let G = (P, E ) be the directed graph where (p, q) E
i |pq|  (p). The function is a broadcast assignment if every point p P is
reachable from s in G . If every p P is reachable within h hops, for a given
parameter h, then is an h-hop broadcast assignment. The (h-hop) broadcast

This research was supported by the Netherlands Organization for Scientic Research
(NWO) under project no. 024.002.003.

Springer International Publishing AG 2017 289

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 289300, 2017.
DOI: 10.1007/978-3-319-62127-2_25
290 M. de Berg et al.

problem is to nd an (h-hop) broadcast assignment whose cost pP cost((p))
is minimized. Often the cost of assigning transmission radius x is dened as
cost(x) = x for some constant . In R1 , both the basic broadcast problem and
the h-hop version are solvable in O(n2 ) time [7]. In R2 the problem is np-hard
for any > 1 [6,10], and in R3 it is even apx-hard [10]. There are also several
approximation algorithms [1,6]. For the 2-hop broadcast problem in R2 an O(n7 )
algorithm is known [2] and for any constant h there is a PTAS [2]. Interestingly,
the complexity of the 3-hop broadcast problem is unknown.
An important special case of the broadcast problem is where we allow only
two possible transmission ranges for the points, (p) = 1 or (p) = 0. In this
case the exact cost function is irrelevant and the problem becomes to minimize
the number of active points. This is called the homogeneous broadcast problem
and it is the version we focus on. From now on, all mentions of broadcast and
h-hop broadcast refer to the homogeneous setting. Observe that if (p) = 1 then
(p, q) is an edge in G if and only if the disks of radius 1/2 centered at p and q
intersect. Hence, if all points are active then G in the intersection graph of a
set of congruent disks or, in other words, a unit-disk graph (UDG). Because of
their relation to wireless networks, UDGs have been studied extensively.
Let D be a set of congruent disks in the plane, and let GD be the UDG
induced by D. A broadcast tree on GD is a rooted spanning tree of GD . To send
a message from the root to all other nodes, each internal node of the tree has to
send the message to its children. Hence, the cost of broadcasting is related to the
internal nodes in the broadcast tree. A cheapest broadcast tree corresponds to a
minimum-size connected dominating set on GD , that is, a minimum-size subset
D such that the subgraph induced by is connected and each node in
GD is either in or a neighbor of a node in . The broadcast problem is thus
equivalent to the following: given a UDG GD with a designated source node s,
compute a minimum-size connected dominated set D such that s .
In the following we denote the dominating set problem by ds, the connected
dominating set problem by cds, and we denote these problems on UDGs by ds-
udg and cds-udg, respectively. Given an algorithm for the broadcast problem,
one can solve cds-udg by running the algorithm n times, once for each possible
source point. Consequently, hardness results for cds-udg can be transferred
to the broadcast problem, and algorithms for the broadcast problem can be
transferred to cds-udg at the cost of an extra linear factor in the running time.
It is well known that ds and cds are np-hard, even for planar graphs [11]. ds-
udg and cds-udg are also np-hard [13,15]. The parameterized complexity of
ds-udg has also been investigated: Marx [14] proved that ds-udg is W[1]-hard
when parameterized by the size of the dominating set. (The denition of W[1]
and other parameterized complexity classes can be found in the book by Flum
and Grohe [9].)

Our contributions. Knowing the existing hardness results for the broadcast prob-
lem, we set out to investigate the following questions. Is there a natural special
case or parameterization admitting an ecient algorithm? Since the broadcast
problem is polynomially solvable in R1 , we study how the complexity of the
The Homogeneous Broadcast Problem in Narrow and Wide Strips 291

problem changes as we go from the 1-dimensional problem to the 2-dimensional

problem. To do this, we assume the points (that is, the disk centers) lie in a
strip of width w, and we study how the problem complexity changes as we in-
crease w. We give an almost complete characterization of the complexity, both
for the general and for the hop-bounded version of the problem. More precisely,
our results are as follows.
We rst study strips of width at most 3/2. Unit disk graphs restricted
to such narrow strips are a subclass of co-comparability graphs [16], for which
an O(nm) time cds algorithm is known [12,3]. (Here m denotes the number
of edges in the graph.) The broadcast problem is slightly dierent because it
requires s to be in the dominating set; still, one would expect better running
times in this restricted graph class. Indeed, we show that for narrow strips the
broadcast problem can be solved in O(n log n) time. The hop condition in the
h-hop broadcast problem has not been studied yet for co-comparability graphs
to our knowledge. This condition complicates the problem considerably. Never-
theless, we show that the h-hop broadcast problem in narrow strips is solvable in
polynomial time. Our algorithm runs in O(n6 ) and uses a subroutine for 2-hop
broadcast, which may be of independent interest: we show that the 2-hop broad-
cast problem is solvable in O(n4 ) time. Our subroutine is based on an algorithm
by Ambuhl et al. [2] for the non-homogeneous case, which runs in O(n7 ) time.
This result is can be found in the full version of this paper.
Second, we investigate what happens for wider strips. We show that the
broadcast problem has an nO(w) dynamic-programming algorithm for strips of
width w. We prove a matching lower bound of n(w) , conditional on the Expo-
nential Time Hypothesis (ETH). Interestingly, the h-hop broadcast problem has
no such algorithm (unless p = np): we show this problem is already np-hard on
a strip of width 40. One of the gadgets in this intricate construction can also be
used to prove that a cds-udg and the broadcast problem are W[1]-hard param-
eterized by the solution size k. The W[1]-hardness proof is discussed only in the
full version. It is a reduction from Grid
Tiling based on ideas by Marx [14],
o( k)
and it implies that there is no f (k)n algorithm for cds-udg unless ETH
fails.

2 Algorithms for broadcasting inside a narrow strip

In this section we present polynomial algorithms (both for broadcast and for
h-hop broadcast) for inputs that lie inside a strip S := R [0, w], where 0 < w 

3/2 is the width of the strip. Without loss of generality, we assume that the
source lies on the y-axis. Dene S0 := [0, )[0, w] and S0 := (, 0][0, w].
Let P be the set of input points. We dene x(p) and y(p) to be the x- and
y-coordinate of a point p P , respectively, and (p) to be the unit-radius disk
centered at p. Let G = (P, E) be the graph with (p, q) E i q (p), and let
P  := P \ (s) be the set of input points outside the source disk. We say that
a point p P is left-covering if pp E for all p P  with x(p ) < x(p); p is
right-covering if p p E for all p P  with x(p ) > x(p). We denote the set of
 292 M. de Berg et al.

left-covering and right-covering points by Q and Q+ respectively. Finally, the

core area of a point p, denoted by core(p), is [x(p) 12 , x(p) + 12 ] [0, w]. Note

that core(p) (p) because w  3/2, i.e., the disk of p covers a part of the
strip that has horizontal
length at least one. This is a key property of strips of
width at most 3/2, and will be used repeatedly.
We partition P into levels L0 , L1 , . . . Lt , based on hop distance from s in G.
Thus Li := {p P : dG (s, p) = i}, where dG (s, p) denotes the hop-distance. Let
L +
i and Li denote the points of Li with negative and nonnegative coordinates,
respectively. We will use the following observation multiple times.

Observation 1. Let G = (P, E) be a unit disk graph on a narrow strip S.

(i) Let be a path in G from a point p P to a point q P . Then the region
[x(p) 12 , x(q) + 12 ] [0, w] is fully covered by the disks of the points in .
(ii) The overlap of neighboring levels is at most 12 in x-coordinates: max{x(p)|p
L+ + 1 +
i1 }  min{x(q)|q Li } + 2 for any i > 0 with Li
= ; similarly,
min{x(p)|p Li1 }  max{x(q)|q Li } 2 for any i > 0 with L
1
i
= .
(iii) Let p be an arbitrary point in L+ i for some i > 0. Then the disks of any
path (s, p) cover all points in all levels L0 L1 L+ 2 L +
i1 . A similar
statement holds for points in L i .

2.1 Minimum broadcast set in a narrow strip

A broadcast set is a point set D P that gives a feasible broadcast, i.e., a

connected dominating set of G that contains s. Our task is to nd a minimum
broadcast set inside a narrow strip. Let p, p P be points with maximum and
minimum x-coordinate, respectively. Obviously there must be paths from s to p
and p in G such that all points on these paths are active, except possibly p and p .
If p and p are also active, then these paths alone give us a feasible broadcast set:
by Observation 1(i), these paths cover all our input points. Instead of activating
p and p , it is also enough to activate the points of a path that reaches Q and
a path that reaches Q+ . In most cases it is sucient to look for broadcast sets
with this structure.
Lemma 1. If there is a minimum broadcast set with an active point on L2 , then
there is a minimum broadcast set consisting of the disks of a shortest path
from s to Q and a shortest path + from s to Q+ . These two paths share s and
they may or may not share their rst point after s.

s b b s b b

a a
a a a

Fig. 1. A swap operation. The edges of the broadcast tree are solid lines.
 The Homogeneous Broadcast Problem in Narrow and Wide Strips 293

Proof sketch. If a minimum broadcast set does not contain a point from Q+ , then
we can nd two active points a and b whose disks uniquely cover two non-active
points a and b, respectively; see Fig. 1. By deactivating a and activating b we
get a new feasible solution, since (b) covers all points previously only covered
by (a). By using such operations repeatedly, we can nd a solution containing a
point from Q+ . Using similar arguments, we can nd a solution also containing
a point from Q . Finally, using Observation 1(iii), we can show that a shortest
path + from s to Q+ and a shortest path from s to Q together form a
feasible and minimum-size solution. 

Lemma 2 below fully characterizes optimal broadcast sets. To deal with the
case where Lemma 1 does not apply, we need some more terminology. We say
that the disk (q) of an active point q in a feasible broadcast set is bidirectional
if there are two input points p L + +
2 and p L2 that are covered only by

(q). See points p and p in Fig. 2 for an example. Note that q core(s), because
core(s) = [ 12 , 12 ] [0, w] is covered by (s), and our bidirectional disk has to
cover points both in (, 12 ] [0, w] and [ 12 , ) [0, w]. Active disks that are
not the source disk and not bidirectional are called monodirectional.
Lemma 2. For any input P that has a feasible broadcast set, there is a minimum
broadcast set D that has one of the following structures.
(i) Small: |D|  2.
(ii) Path-like: |D|  3, and D consists of a shortest path from s to Q and
a shortest path + from s to Q+ ; + and share s and may or may not
share their rst point after s.
(iii) Bidirectional: |D| = 3, and D contains two bidirectional disk centers and s.
As it turns out, the bidirectional case is the most dicult one to compute
eciently. (It is similar to cds-udg in co-comparability graphs, where the case
of a connected dominating set of size at most 3 dominates the running time.)

Lemma 3. In O(n log n) time we can nd a bidirectional broadcast if it exists.

p
P P+
p
s

Fig. 2. A bidirectional broadcast.

294 M. de Berg et al.

Proof. Let P := {u1 , u2 , . . . , uk } be the set of points to the left of the source
disk (s), where the points are sorted in increasing y-order with ties broken
arbitrarily. Similarly, let P + := {v1 , v2 , . . . , vl } be the set of points to the

right of (s), again sorted in order of increasing y-coordinate. Dene Pi :=
+ +
{u1 , . . . , ui }, and dene P>i , and Pi and P>i analogously. Our algorithm is
based on the following observation: There is a bidirectional solution if and only
+
if there are indices i, j and points p, p core(s) such that (p) covers Pi Pj
 +
and (p ) covers P>i P>j ; see Fig. 2.

Now for a point p core(s), dene Z (p) := max{i : Pi (p)} and
+ + +
Z> (p) := min{i : P>i (p)}, and Z (p) := max{i : Pi (p)}, and Z> (p) :=
+
min{i : P>i (p)}. Then the observation above can be restated as:
There is a bidirectional solution if and only if there are points p, p
 + + 
core(s) such that Z (p)  Z> (p ) and Z (p)  Z> (p ).
It is easy to nd such a pairif it existsin O(n log n) time once we have
+ +
computed the values Z (p), Z> (p), Z (p), and Z> (p) for all points p (s). It
remains to show that these values can be computed in O(n log n) time.

Consider the computation of Z (p); the other values can be computed sim-
ilarly. Let T be a balanced binary tree whose leaves store the points from
P in order of their y-coordinate. For a node in T , let F () := {(ui ) :
ui is stored in the subtree rooted at }. We start by computing at each node
the intersection of the disks
, in F (). More precisely, for each we compute the
region I() := core(s) F (). Notice that I() is y-monotone and convex,
and each disk (ui ) contributes at most one arc to I(). (Here I() refers to
the boundary of I() that falls inside S.) Moreover, I() = I(left-child())
I(right-child()). Hence, we can compute the regions I() of all nodes in T in
O(n log n) time in total, in a bottom-up manner. Using the tree T we can now

compute Z (p) for any given p core(s) by searching in T , as follows. Suppose
we arrive at a node . If p I(left-child()), then descend to right-child(),
otherwise descend to left-child(). The search stops when we reach a leaf, stor-

ing a point ui . One easily veries that if p (ui ) then Z (p) = i, otherwise

Z (p) = i 1.
Since I() is a convex region, we can check if p I() in O(1) time if we can
locate the position of py in the sorted list of y-coordinates of the vertices of I().
We can locate py in this list in O(log n) time, leading to an overall query time of
O(log2 n). This can be improved to O(log n) using fractional cascading [5]. Note
that the application of fractional cascading does not increase the preprocessing

time of the data structure. We conclude that we can compute all values Z (p)
in O(n log n) time in total.


In order to compute a minimum broadcast, we can rst check for small and
bidirectional solutions. To nd path-like solutions, we rst compute the sets Q
and Q+ , and compute shortest paths starting from these sets back to the source
disk. The path computation is very similar to the shortest path algorithm in
UDGs by Cabello and Jejcic [4].
The Homogeneous Broadcast Problem in Narrow and Wide Strips 295

Theorem 2. The broadcast problem inside a strip of width at most 3/2 can
be solved in O(n log n) time.

Remark 1. If we apply this algorithm to every disk as source, we get an O(n2 log n)
algorithm for cds in narrow strip UDGs. We can compare this to O(mn), the run-
ning time that we get by applying the algorithm for co-comparability graphs [3].
Note that in the most dicult case, when the size of the minimum connected
dominating set is at most 3, the unit disk graph has constant diameter, which
implies that the graph is dense, i.e., the number of edges is m = (n2 ). Hence,
we get an (almost) linear speedup for the worst-case running time.

2.2 Minimum-size h-hop broadcast in a narrow strip

In the hop-bounded version of the problem we are given P and a parameter h,

and we want to compute a broadcast set D such that every point p P can be
reached in at most h hops from s. In other words, for any p P , there must be a
path in G from s to p of length at most h, all of whose vertices, except possibly p
itself, are in D. We start by investigating the structure of optimal solutions in
this setting, which can be very dierent from the non-hop-bounded setting.
As before, we partition P into levels Li according to the hop distance from
s in the graph G, and we dene L+
i and Li to be the subsets of points at level i
with positive and nonnegative x-coordinates, respectively. Let Lt be the highest
non-empty level. If t > h then clearly there is no feasible solution.
If t < h then we can safely use our solution for the non-hop-bounded case,
because the non-hop-bounded algorithm gives a solution which contains a path
with at most t + 1 hops to any point in P . This follows from the structure of the
solution; see Lemma 2. (Note that it is possible that the solution given by this
algorithm requires t + 1 hops to some point, namely, if Q+ Q Lt .) With the
t < h case handled by the non-hop-bounded algorithm, we are only concerned
with the case t = h.
We deal with one-sided inputs rst, where the source is the leftmost input
point. Let G be the directed graph obtained by deleting edges connecting points
inside the same level of G, and orienting all remaining edges from lower to higher
levels. A Steiner arborescence of G for the terminal set Lh is a directed tree
rooted at s that contains a (directed) path p from s to p for each p Lh . From
now on, whenever we speak of arborescence we refer to a Steiner arborescence
in G for terminal set Lh . We dene the size of an arborescence to be the
number of internal nodes of the arborescence. Note that the leaves in a minimum-
size arborescence are exactly the points in Lh : these points must be in the
arborescence by denition, they must be leaves since they have out-degree zero
in G , and leaves that are not in Lh can be removed.

Remark 2. In the minimum Steiner Set problem, we are given a graph G and
a vertex subset T of terminals, and the goal is to nd a minimum-size vertex
subset S such that T S induces a connected subgraph. This problem has a
polynomial algorithm in co-comparability graphs [3], and therefore in narrow
296 M. de Berg et al.

1 2
s

1 2 3

Fig. 3. Two dierent arborescences, with vertices labeled with their level. The arbores-
cence made of the bottom path does not dene a feasible broadcast for h = 3, since it
would take four hops to reach the top right node.

strip unit disk graphs. However, the broadcast set given by a solution does not
t our hop bound requirements. Hence, we have to work with a dierent graph
(e.g. the edges within each level Li have been removed), and this modied graph
is not necessarily a co-comparability graph.
Lemma 4 below states that either we have a path-like solutionfor the one-
sided case a path-like solution is a shortest s Q+ path or any minimum-
size arborescence denes a minimum-size broadcast set. The latter solution is
obtained by activating all non-leaf nodes of the arborescence. We denote the
broadcast set obtained from an arborescence A by DA .
Lemma 4. Any minimum-size Steiner arborescence for the terminal set Lh de-
nes a minimum broadcast set, or there is a path-like minimum broadcast set.
Notice that a path-like solution also corresponds to an arborescence. However,
it can happen that there are minimum-size arborescences that do not dene a
feasible broadcast; see Fig. 3. Lemma 4 implies that if this happens, then there
must be an optimal path-like solution. The lemma also implies that for non-path-
like solutions we can use the Dreyfus-Wagner dynamic-programming algorithm
to compute a minimum Steiner tree [8], and obtain an optimal solution from this
tree.3 Unfortunately the running time is exponential in the number of terminals,
which is |Lh | in our case. However, our setup has some special properties that
we can use to get a polynomial algorithm.
We dene an arborescence A to be nice if the following holds. For any two
arcs uu and vv  of A that go between the same two levels, with u
= v, we have:
y(u ) < y(v  ) y(u) < y(v). Intuitively, a nice arborescence is one consisting of
paths that can be ordered vertically in a consistent manner, see the left of Fig. 4.
We dene an arborescence A to be compatible with a broadcast set D if D = DA .
Note that there can be multiple arborescencesthat is, arborescences with the
same node set but dierent edge setscompatible with a given broadcast set D.
Lemma 5. Every optimal broadcast set D has a nice compatible arborescence.
Proof sketch. To nd a nice compatible arborescence we will associate a unique
arborescence with D. To this end, we dene for each p (D Lh ) \ {s} a unique
3
The Dreyfus-Wagner algorithm minimizes the number of edges in the arborescence.
In our setting the number of edges equals the number of internal nodes plus |Lh | 1,
so this also minimizes the number of internal nodes.
The Homogeneous Broadcast Problem in Narrow and Wide Strips 297

s
pred (p) p ray(p)
z

Fig. 4. Left: A nice Steiner arborescence. Note that arc crossings are possible. Right:
Dening the pred function.


predecessor pred (p), as follows. Let i be the boundary of {(p)|p Li D}.
The two lines bounding the strip S cut i into four parts: a top and a bot-
tom part that lie outside the strip, and a left and a right part that lie inside the
strip. Let i be the part on the right inside the strip. We then dene the function
pred : (D Lh ) \ {s} D the following way. Consider a point p (D Lh ) \ {s}
and let i be its level. Let ray(q) be the horizontal ray emanating from q to the
right; see the right of Fig. 4. It follows from Observation 1(iii) that ray(q) can-
not enter any disk from level i 1. We can prove that any point p D Lh is
contained in a disk from ps previous level, so pred (p) is well dened for these
points. The edges pred (p)p for p D Lh thus dene an arborescence. We can
prove that it is nice by showing that the y-order of the points in a level Li cor-
responds
 to the vertical order in which the boundaries of their disks appear on
{(p) : p Li D}. 
Let q1 , q2 , . . . , qm be the points of Lh in increasing y-order. The crucial property
of a nice arborescence is that the descendant leaves of a point p in the arbores-
cence form an interval of q1 , q2 , . . . , qm . Using the above lemmas, we can adapt
the Dreyfus-Wagner algorithm and get the following theorem.
Theorem
3. The one-sided h-hop broadcast problem inside a strip of width at
most 3/2 can be solved in O(n4 ) time.
In the general (two-sided) case, we can have path-like solutions and arborescence-
based solutions on both sides, and the two side solutions may or may not share
points in L1 . We also need to handle small solutionsnow these are 2-hop
solutionsseparately.

Theorem 4. The h-hop broadcast problem inside a strip of width at most 3/2
can be solved in O(n6 ) time.

3 Broadcasting in a wide strip

Theorem 5. The broadcast problem and cds-udg can be solved in nO(w) time
on a strip of width w. Moreover, there is no algorithm for cds-udg or the
broadcast problem with runtime f (w)no(w) unless ETH fails.

Surprisingly, the h-hop version has no nO(w) algorithm (unless p =np).

298 M. de Berg et al.

h hops
x4
s x3
x2
x1

Fig. 5. The gadget representing the variables. The dotted paths form the x2 -string.

Theorem 6. The h-hop broadcast problem is np-complete in strips of width 40.

(The theorem of course refers to the decision version of the problem: given a
point set P , a hop bound h, and a value K, does P admit an h-hop broadcast
set of size at most K?) Our reduction is from 3-SAT. Let x1 , x2 , . . . xn be the
variables and C1 , . . . , Cm be the clauses of a 3-CNF.
Fig. 5 shows the structural idea for representing the variables, which we call
the base bundle. It consists of (2h1)n+1 points arranged as shown in the gure,
where h is an appropriate value. The distances between the points are chosen
such that the graph G, which connects two points if they are within distance 1,
consists of the edges in the gure plus all edges between points in the same level.
Thus (except for the intra-level edges, which we can ignore) G consists of n pairs
of paths, one path pair for each variable xi . The i-th pair of paths represents
the variable xi , and we call it the xi -string. By setting the target size, K, of
the problem appropriately, we can ensure the following for each xi : any feasible
solution must use either the top path of the xi -string or the bottom path, but it
cannot use points from both paths. Thus we can use the top path of the xi -path
to represent a true setting of the variable xi , and the bottom path to represent
a false setting. A group of consecutive strings is called a bundle. We denote the
bundle containing all xt -strings with t = i, i + 1, . . . , j by bundle(i, j).
The clause gadgets all start and end in the base bundle, as shown in Fig. 6.
The gadget to check a clause involving variables xi , xj , xk , with i < j < k,
roughly works as follows; see also the lower part of Fig. 6, where the strings for
xi , xj , and xk are drawn with dotted lines.
First we split o bundle(1, i1) from the base bundle, by letting the top i1
strings of the base bundle turn left. (In Fig. 6 this bundle consists of two strings.)
We then separate the xi -string from the base bundle, and route the xi -string into
a branching gadget. The branching gadget creates a branch consisting of two
tapesthis branch will eventually be routed to the clause-checking gadgetand
a branch that returns to the base bundle. Before the tapes can be routed to the
clause-checking gadget, they have to cross each of the strings in bundle(1, i 1).
For each string that must be crossed we introduce a crossing gadget. A crossing
gadget lets the tapes continue to the right, while the string being crossed can
return to the base bundle. The nal crossing gadget turns the tapes into a side
string that can now be routed to the clause-checking gadget. The construction
guarantees that the side string for xi still carries the truth value that was selected
for the xi -string in the base bundle. Moreover, if the true path (resp. false
The Homogeneous Broadcast Problem in Narrow and Wide Strips 299

Clause 1 Clause 2 Clause m

Lh
s

string
string
tape
side string
branching
crossing
clause check

Fig. 6. The overall construction, and the way a single clause is checked. Note that in
this gure each string (which actually consists of two paths) is shown as a single curve.

path) of the xi -string was selected to be part of the broadcast set initially, then
the true path (resp. false path) of the rest of the xi -string that return to the
base bundle must be in the minimum broadcast set as well.
After we have created a side string for xi , we create side strings for xj and
xk in a similar way. The three side strings are then fed into the clause-checking
gadget. The clause-checking gadget is a simple construction of four points. In-
tuitively, if at least one side string carries the correct truth valuetrue if the
clause contains the positive variable, false if it contains the negated variable,
then we activate a single disk in the clause check gadget that corresponds to a
true literal. Otherwise we need to change truth value in at least one of the side
strings, which requires an extra disk.
The nal construction contains (n4 m) points that all t into a strip of
width 40. The details are given in the full version.

4 Conclusion
We studied the complexity of the broadcast problem in narrow and wider strips.
For narrow strips we obtained ecient polynomial algorithms, both for the non-
hop-bounded and for the h-hop version, thanks to the special structure of the
problem inside such strips. On wider strips, the broadcast problem has an nO(w)
algorithm, while the h-hop broadcast becomes np-complete on strips of width 40.
With the exception of a constant width range (between 3/2 and 40) we charac-
terized the complexity when parameterized by strip width. We have also proved
that the planar problem (and, similarly, cds-udg) is W[1]-hard when parame-
terized by the solution size. The problem of nding a planar h-hop broadcast set
seems even harder: we can solve it in polynomial time for h = 2 (see full version)
but already for h = 3 we know no better algorithm than brute force. Interesting
open problems include:
300 M. de Berg et al.

- What is the complexity of planar 3-hop broadcast? In particular, is there a

constant value t such that t-hop broadcast is np-complete?
- What is the complexity of h-hop broadcast in planar graphs?

References
1. Ambuhl, C.: An optimal bound for the MST algorithm to compute energy ecient
broadcast trees in wireless networks. In: ICALP, Proceedings. pp. 11391150 (2005)
2. Ambuhl, C., Clementi, A.E.F., Ianni, M.D., Lev-Tov, N., Monti, A., Peleg, D.,
Rossi, G., Silvestri, R.: Ecient algorithms for low-energy bounded-hop broadcast
in ad-hoc wireless networks. In: STACS, Proceedings. pp. 418427 (2004)
3. Breu, H.: Algorithmic aspects of constrained unit disk graphs. Ph.D. thesis, Uni-
versity of British Columbia (1996)
4. Cabello, S., Jejcic, M.: Shortest paths in intersection graphs of unit disks. Comput.
Geom. 48(4), 360367 (2015)
5. Chazelle, B., Guibas, L.J.: Fractional cascading: I. A data structuring technique.
Algorithmica 1(1), 133162 (1986)
6. Clementi, A.E., Crescenzi, P., Penna, P., Rossi, G., Vocca, P.: A worst-case analysis
of an mst-based heuristic to construct energy-ecient broadcast trees in wireless
networks. In: STACS, Proceedings. pp. 121131 (2001)
7. Das, G.K., Das, S., Nandy, S.C.: Range assignment for energy ecient broadcasting
in linear radio networks. Theor. Comput. Sci. 352(1-3), 332341 (2006)
8. Dreyfus, S.E., Wagner, R.A.: The Steiner problem in graphs. Networks 1(3), 195
207 (1971)
9. Flum, J., Grohe, M.: Parameterized Complexity Theory. Texts in Theoretical Com-
puter Science. An EATCS Series, Springer (2006)
10. Fuchs, B.: On the hardness of range assignment problems. Networks 52(4), 183195
(2008)
11. Garey, M.R., Johnson, D.S.: Computers and Intractability: A Guide to the Theory
of NP-Completeness. W. H. Freeman (1979)
12. Kratsch, D., Stewart, L.: Domination on cocomparability graphs. SIAM J. Discrete
Math. 6(3), 400417 (1993)
13. Lichtenstein, D.: Planar formulae and their uses. SIAM J. Comput. 11(2), 329343
(1982)
14. Marx, D.: Parameterized complexity of independence and domination on geometric
graphs. In: Parameterized and Exact Computation, Second International Work-
shop, IWPEC, Proceedings. pp. 154165 (2006)
15. Masuyama, S., Ibaraki, T., Hasegawa, T.: The computational complexity of the
m-center problems on the plane. IEICE Transactions 64(2), 5764 (1981)
16. Matsui, T.: Approximation algorithms for maximum independent set problems
and fractional coloring problems on unit disk graphs. In: JCDCG, Proceedings.
pp. 194200. Springer (1998)
 Minimizing the Continuous Diameter when
Augmenting a Tree with a Shortcut

Jean-Lou De Carufel1 , Carsten Grimm2,3 , Stefan Schirra3 , and Michiel Smid2

1
School of Electrical Engineering and Computer Science, University of Ottawa
800 King Edward Avenue, Ottawa, Ontario, K1N 6N5, Canada
2
School of Computer Science, Carleton University
1125 Colonel By Drive, Ottawa, Ontario, K1S 5B6, Canada
3
Institut fur Simulation und Graphik, Otto-von-Guericke-Universitat Magdeburg
Universitatsplatz 2, D-39106 Magdeburg, Germany

Abstract. We augment a tree T with a shortcut pq to minimize the

largest distance between any two points along the resulting augmented
tree T + pq. We study this problem in a continuous and geometric set-
ting where T is a geometric tree in the Euclidean plane, a shortcut is
a line segment connecting any two points along the edges of T , and we
consider all points on T + pq (i.e., vertices and points along edges) when
determining the largest distance along T + pq. The continuous diameter
is the largest distance between any two points along edges. We establish
that a single shortcut is sucient to reduce the continuous diameter of a
geometric tree T if and only if the intersection of all diametral paths of T
is neither a line segment nor a point. We determine an optimal shortcut
for a geometric tree with n straight-line edges in O(n log n) time.

Keywords: Network Augmentation Continuous Diameter Minimization

1 Introduction
A network is a connected, undirected graph with positive edge weights. A curve
is rectiable if it has a nite length. A geometric network is a network that is
embedded in the Euclidean plane whose edges are rectiable curves weighted
with their length. We describe our algorithmic results for straight-line edges,
even though they extend to more general edges. We say that a point p lies on a
geometric network G and write p G when there is an edge e of G such that p is
a point along the embedding of e. A point p on an edge e of length l subdivides
e into two sub-edges of lengths (1 ) l and l for some value [0, 1]. We
represent the points on G in terms of their relative positions (expressed by )
along their containing edges, thereby avoiding ambiguity in case of crossings.
The network distance between any two points p and q on a geometric net-
work G is the length of a shortest weighted path from p to q in G and it
is denoted by dG (p, q). The continuous diameter of G is the largest network

This work was partially supported by NSERC and FQRNT.

Springer International Publishing AG 2017 301

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 301312, 2017.
DOI: 10.1007/978-3-319-62127-2_26
302 J.-L. De Carufel et al.

distance between any two points on G, and it is denoted by diam(G), i.e.,

diam(G) = maxp,qG dG (p, q). In contrast, for a network with vertex set V ,
the discrete diameter is the largest distance between any two vertices, i.e.,
maxu,vV dG (u, v). A pair p, q G is diametral when their distance is the con-
tinuous diameter, i.e., diam(G) = dG (p, q). A diametral path in G is a shortest
weighted path in G that connects a diametral pair of G. We denote the Euclidean
distance between two points p and q in the plane by |pq|. A line segment pq with
endpoints p, q G is a shortcut for G. We augment a geometric network G with
a shortcut pq as follows. If they do not exist already, we introduce vertices at p
and at q, thereby subdividing the edges containing p and q. We add the line seg-
ment pq as an edge of length |pq| to G without introducing vertices at crossings
between pq and other edges. We denote the resulting network by G + pq.
Our goal is to locate a shortcut pq
for a geometric tree T that minimizes
the continuous diameter of the augmented 1 1
tree T + pq, as illustrated in Fig. 1. This 2 p 4
3
means we seek two points p, q T with
diam(T +pq) = minr,sT diam(T +rs). We q
call a shortcut that minimizes the contin- 2
3
uous diameter an optimal shortcut. 4
The backbone of a tree T is the inter-
Fig. 1. An optimal shortcut pq for a ge-
section of all diametral paths of T ; we de-
ometric tree T with the diametral pairs
note the backbone of T by B. The absolute of T +pq marked by matching numbers.
center of a geometric tree T is the unique
point c T that minimizes the largest
network distance from c, i.e., maxqT dT (c, q) = minpT maxqT dT (p, q). Note
that we always have c B. It takes O(n) time to determine the absolute center
and, thereby, the backboneof a geometric tree with n vertices [9].

1.1 Related Work

We summarize related work on minimum-diameter network augmentation.
In the abstract and discrete setting, the goal is to minimize the discrete di-
ameter of an abstract graph with positive weights for the edges and non-edges
by inserting non-edges as shortcuts. If the edges and non-edges have unit weight,
then it is NP-hard to decide whether the diameter can be reduced below D 2
by adding at most k shortcuts [2, 10, 13]. This problem remains NP-hard, even
for restricted graph classes such as trees [2]. The weighted version of this problem
falls into the parameterized complexity class W[2]-hard [5, 6]. Minimum-diameter
augmentation has also been studied as a bicriteria optimization in which both
the diameter and the number (or cost) of the additional edges are minimized.
For an overview on bicriteria approximation algorithms refer to Frati et al. [5].
Groe et al. [7] introduce the geometric and discrete setting in which the
problem is to minimize the discrete diameter of a geometric network by connect-
ing vertices with line segments. Groe et al. [7] determine an optimal shortcut
for a polygonal path with n vertices in O(n log3 n) time. The stretch factor,
Minimizing the Continuous Diameter when Augmenting a Tree with a Shortcut 303

i.e., the largest ratio of the network distance between any two vertices and their
Euclidean distance, has also been considered as a target function [4, 11].
In the geometric and continuous setting [3], the task is to minimize the con-
tinuous diameter of a geometric network by inserting line segments that may
connect any two points along the edges. For a polygonal path of length n, one
can determine an optimal shortcut in O(n) time. For a cycle, one shortcut can
never decrease the continuous diameter while two always suce. For convex
cycles with n vertices, one can determine an optimal pair of shortcuts in O(n).
In the model studied in this work, a crossing of a shortcut with an edge or
another shortcut is not a vertex: a path may only enter edges at their endpoints.
In the planar model [1, 14], every crossing is a vertex of the resulting network,
which leads to a dierent graph structure and, thus, continuous diameter. In the
planar model, Yang [14] characterizes optimal shortcuts for a polygonal path.
Caceres et al. [1] determine in polynomial time whether the continuous diameter
of a plane geometric network can be reduced with a single shortcut.
Recently, Oh and Ahn [12] develop an O(n2 log3 n)-time algorithm to deter-
mine an optimal shortcut for a tree with n vertices in the discrete setting.

1.2 Structure and Results

We present the following structural results about optimal shortcuts for geometric
trees. A shortcut pq is useful for a geometric tree T when augmenting T with pq
decreases the continuous diameter. A geometric tree T admits a useful shortcut
if and only if its backbone B is not a line segment. We consider a point to be
a degenerate line segment. Every geometric tree T has an optimal shortcut pq
with p, q B such that the absolute center c lies on the path from p to q in T .
Based on these insights, we determine an optimal shortcut for a geometric
tree T with n vertices in O(n log n) time. Conceptually, we slide a candidate
shortcut pq along the backbone of T while balancing the diametral paths in T +
pq. The diametral pairs of the augmented tree guide our search: each diametral
pair in T + pq rules out a better shortcut in one of four possible directions of
our search. To implement this approach, we discretize this movement such that
the shortcut jumps from one change in the diametral paths to the next.

2 Structural Results

We say a shortcut pq is useful for T when diam(T + pq) < diam(T ), we say
pq is indierent for T when diam(T + pq) = diam(T ), and we say pq is useless
for T when diam(T + pq) > diam(T ). In the discrete setting, every shortcut
is useful or indierent, as the discrete diameter only considers vertices of T .
In the continuous setting, a shortcut may be useless for T , since the points on
the shortcut pq matter as well, as exemplied in Fig. 2. For some trees, we
cannot reduce the continuous diameter with a single shortcut, as demonstrated
in Fig. 3. In this example, the diametral paths of T intersect in a single vertex
the absolute centerand, thus, the backbone is a degenerate line segment.
304 J.-L. De Carufel et al.

x p s q y
x
p
z

t c q
y z
Fig. 2. A geometric tree T with a short-
cut pq that is useless for T . Each edge has Fig. 3. A geometric tree T where
weight one and |pq| = 8. The continuous di- no shortcut decreases the continu-
ameter increases when we augment T with ous diameter. Every shortcut pq is
pq, since diam(T ) = dT (x, y) = 16 and useless for at least one pair among
diam(T + pq) = dT +pq (s, t) = 17. the three leaves x, y, and z.

Theorem 1. A geometric tree has a useful shortcut if and only if its backbone
is not a line segment.

Proof (Sketch). If the backbone B of a geometric tree T consists of a single point

(i.e., B is a degenerate line segment), then T does not possess a useful shortcut.
Thus, if T does possess a useful shortcut pq, then B is a path with distinct
endpoints a and b such that pq is useful for {a, b}, i.e., dT +pq (a, b) < dT (a, b).
This means that the path from a to b in T , i.e., B, cannot be a line segment.
Conversely, if B is not a line segment, then B is a path from a to b with
|ab| < dT (a, b). If ab is not a useful shortcut for T , then T + ab has a diametral
pair with one partner on ab and one partner in a sub-tree attached to B. We
nd p, q B where the path from p to q is not a line segment and the cycle in
T + pq is suciently small to rule out diametral pairs in T + pq with a partner
on pq. Therefore, if B is not a line segment, then T has a useful shortcut.


Let T be a geometric tree where the

backbone B is a path from a to b with
a
= b. This path contains the absolute x2 y1
Sb
center c of T . We prove that there is an Sa y2
optimal shortcut pq for T such that p lies x1
a b y3
on the path from a to c along B and q lies rk
r1
on the path from c to b along B. Suppose Sk
r2 rk1
we pick the absolute center c B as the S1 r3
root of T . Let Sa and Sb be the sub-trees
... Sk1
of T attached to the backbone at a and S2 S3
at b, respectively. We refer to Sa and Sb
as the primary B-sub-trees of T ; we refer Fig. 4. A geometric tree T with its
to any other sub-trees of T attached to backbone B and the primary B-sub-
B as secondary B-sub-trees. The root of trees Sa and Sb , as well as the sec-
a B-sub-tree S is the vertex r connecting ondary B-sub-trees S1 , S2 , . . . , Sk .
S and the backbone B. As illustrated in
Fig. 4, every diametral pair in T consists
of a leaf x in Sa and a leaf y in Sb .
Minimizing the Continuous Diameter when Augmenting a Tree with a Shortcut 305

Theorem 2. For every geometric tree T , there exists an optimal shortcut pq

where both endpoints lie on the backbone B of T , i.e., p, q B, and where the
path from p to q along B contains the absolute center c of T .
Proof (Sketch). If T has no useful shortcuts, then the degenerate shortcut cc
satises the claim, since it is indierent and therefore optimal for T with c B.
Suppose T does possess a useful shortcut, i.e., the backbone B of T is a
path from a to b with a
= b and |ab| < dT (a, b). Let S be a B-sub-tree with
root r. First, we show that if pq is useful for T , then p and q cannot lie in
the same B-sub-tree of T , i.e., we have p / S or q / S. Then, we show that
if p S then the diameter never increases if we move p to the root r of S,
i.e., diam(T + rq) diam(T + pq). Since our argument does not depend on the
position of q, we conclude that there is an optimal shortcut for T with both
endpoints on the backbone B.
To prove diam(T + rq) diam(T + pq) we consider a diametral pair s, t of
T + rq. We distinguish the three cases: (1) s, t T with dT (s, t) = dT +pq (s, t),
(2) s, t T with dT (s, t) < dT +pq (s, t), and (3) s / T or t
/ T . Groe et al. [8]
show the claim for Cases (1) and (2). The argument for Case (3) is as follows.
(3) Suppose s / T or t / T . We assume, without loss of generality, that
s
T ; otherwise, we swap s and t. Then s rq with r
= s
= q, as in Fig. 5.
Let C(r, q) be the simple cycle in T + rq. The
path from s to t leaves C(r, q) at the point s that S
p s
is the farthest point from s on C(r, q) . Note that q
s
r
= s
= q, because r
= s
= q and |rq| dT (r, q).
Since p S, we have q / S. This means the path r
s t
connecting r and q in T lies outside of S and,
therefore, s / S and t / S. The cycle C(p, q)
is formed by pq and the path from p to q in T . Fig. 5. A diametral pair s, t of
Since r lies on the path from p to q, we know that T + rq with s rq.
s C(p, q). Let s be the farthest point from s on C(p, q). We obtain
diam(T + rq) = dT +rq (s, t) (s and t are diametral in T + rq)
= dT +rq (s, s) + dT (s, t) (s is on any path from t to s in T + rq)
dT (r, q) + |rq|
= + dT (s, t) (s and s are antipodals on C(r, q))
2
dT (r, q) + dT (r, p) + |pq|
+ dT (s, t) (triangle inequality)
2
dT (p, q) + |pq|
= + dT (s, t) (r is on the path from p to q in T )
2
= dT +pq (s , s) + dT (s, t) (s and s are antipodals on C(p, q))
= dT +pq (s , t) (s is on any path from t to s in T + pq)
diam(T + pq) .
Suppose pq is an optimal shortcut for T with p, q B where c does not lie on
the path from p to q. We argue that if q lies on the path from p to c, then the
diameter never increases as q moves to c, i.e., diam(T + pc) diam(T + pq). 
306 J.-L. De Carufel et al.

3 Preparations for the Algorithm

Initially, we place the endpoints of the shortcut, p and q, at the absolute center
c of T . Then, we move p and q along the backbone B balancing the diametral
paths in T + pq. During this movement p remains along the path from a to c
and q remains on the path from c to b. The diametral pairs in T + pq guide our
search: each diametral pair in T + pq rules out some direction in which we could
search for a better shortcut. We describe our algorithm along the following steps.
1. We simplify the geometric tree T by compressing the B-sub-trees, thereby
simplifying the discussion about diametral pairs and paths in T + pq.
2. We dene algorithm states in terms of the diametral paths and diametral
pairs that are present in the augmented tree, and we distinguish four types
of movements for the shortcut as the operations of our algorithm.
3. Each type of diametral pair rules out a better shortcut in some direction;
some combinations of pair types imply that the current shortcut is optimal.
4. We describe the continuous, conceptual movement of the shortcut that is
guided by the set of types of diametral pairs that are present in T + pq. We
identify the invariants that guarantee that we nd an optimal shortcut.
5. We determine the speeds at which the endpoints of the shortcut would move
in the continuous algorithm, depending on the diametral paths in T + pq.
6. We bound the number of events of the discrete algorithm by O(n) by ruling
out some transitions between the algorithm states and by identifying when
we can safely ignore events without compromising optimality.
7. Finally, we process each of the O(n) events in O(log n) amortized time and,
thus, bound the running time of our algorithm by O(n log n).

Simplifying the Tree Let Sa and Sb be

the primary B-sub-trees of a geomet- y
ric tree T and let S1 , S2 , . . . , Sk be the
secondary B-sub-trees of T that are x
a b
attached to the backbone B at their rk
r1
roots r1 , r2 , . . . , rk , respectively, as il- s1 rk1 sk
lustrated in Fig. 6. Let x be a far- r2 r
3
thest leaf from a in Sa , let y be a far-
...
thest leaf from b in Sb and, for every s2 s3 sk1
i = 1, 2, . . . , k, let si be a farthest leaf
from ri in Si . We replace each B-sub- Fig. 6. The perspective from the backbone
tree Si with an edge from ri to a ver- for the geometric tree from Fig. 4.
tex representing si of length dT (ri , si ).
Likewise, we replace Sa and Sb with edges of appropriate lengths.
There is no need to distinguish diametral pairs from two dierent B-sub-
trees: for any i, j = 1, 2, . . . , k with i
= j, the pair si , sj is diametral in T + pq
if and only if every farthest leaf from ri in Si forms a diametral pair with every
farthest leaf from rj in Sj . There is no need to consider diametral pairs from the
same B-sub-tree S: if the augmented tree T + pq has a diametral pair u, v such
that u and v lie in the same B-sub-tree S of T , then the shortcut pq is optimal.
Minimizing the Continuous Diameter when Augmenting a Tree with a Shortcut 307

Every diametral pair of T consists of x Sa and y Sb . In the following,

the symbols x and y indicate diametral partners in Sa and in Sb , respectively.

Pair Types Every point v of a diametral pair u, v in T + pq is either: (x, y) a

point in the tree that occurs in a diametral pair, () some other point on the
tree, or () a point on the shortcut. More specically, v is either (x, y) a leaf of a
primary B-sub-tree, () a leaf of a secondary B-sub-tree, or v is a point on the
simple cycle C(p, q) in the augmented tree that () lies in the original tree or ()
on the shortcut. This leads to the following distinction of the diametral pairs.
1. (x-y): Diametral pairs x, y of T + pq with x Sa and y Sb .
2. (x-): Diametral pairs x, v of T + pq with x Sa and v T \ (Sa Sb ).
(a) (x-): Pairs x, sj with x Sa and sj Sj for some j = 1, 2, . . . , k.
(b) (x-): Pairs x, x with x Sa and where x is farthest from x on C(p, q).
3. (-y): Diametral pairs u, y of T + pq with u T \ (Sa Sb ) and y Sb .
(a) (-y): Pairs si , y with si Si and y Sb for some i = 1, 2, . . . , k.
(b) (-y): Pairs y, y with y Sb and where y is farthest from y on C(p, q).
4. (-): Diametral pairs u, v of T + pq with u, v T \ (Sa Sb ).
(a) (-): Pairs si , sj with si Si and sj Sj for i, j = 1, 2, . . . , k.
(b) (-): Pairs si , si with si , si T where si Si for some i = 1, 2, . . . , k
and si T is farthest from si on C(p, q).
5. (-): Diametral pairs u, v of T + pq with v / T , i.e., v pq with p
= v
= q.
(a) (-): Pairs si , si with si Si , si pq, and p
= si
= q, for i = 1, 2, . . . , k,
and where si is farthest from si on C(p, q).
There is no need to consider diametral pairs of type - or -: the distance
from x or y to their respective farthest points on the cycle C(p, q) is always larger
than the distance between any two points on C(p, q)unless T + pq = C(p, q).
There are no diametral pairs of type x- or -y, since x, y T : If x, v is of type
x- in T + pq, then v T , and if u, y is of type -y in T + pq, then u T .
The pair state is the set of types of diametral pairs in T +pq. If T +pq has the
diametral pairs x, y; x, s3 ; x, s5 ; and x, x, then T + pq is in pair state {x-y, x-}.

Path Types Every diametral path that connects u, v T contains the shortcut
or not. Every diametral path that connects u T with v / T contains p or q.
There are no diametral paths connecting u, v
/ T . We distinguish the following.
1. (-pq-): A diametral path that does contain pq and connects u, v T .
2. (-T -): A diametral path that does not contain pq and connects u, v T .
3. (-p-): A diametral path that contains p and connects u T with v / T.
4. (-q-): A diametral path that contains q and connects u T with v / T.
Any type of endpoint (e.g., x, y, , , , ) may appear in place of . For
instance, we denote a diametral path from x to x via the shortcut by x-pq-.
The path state is the set of types of diametral paths that are present in T +pq.
For instance, if T + pq has the diametral pairs x, y; x, s3 ; x, s5 ; and x, x such
that pq is useful for x, y and x, s3 but useless for x, s5 then T + pq is in path
state {x-pq-y, x-pq-, x-T -, x-pq-, x-T -}. Figure 7 illustrates the distinction
between diametral paths of type -pq- and -T - when x-pq-y is diametral.
308 J.-L. De Carufel et al.

x a y x a y
b b

p q p q
u ru
ru rv v u rv v
rv ru rv
ru
q p
(a) The diametral path of type (b) The diametral path of type
-pq- connects u with v via pq. -T - connects u with v via T .

Fig. 7. An illustration of an augmented tree T + pq with (a) diametral paths of type

x-pq-y and -pq- and (b) diametral paths of type x-pq-y and -T -. The points ru
and rv mark the farthest points from ru and rv , respectively along the simple cycle in
T + pq. Their position indicates wether pq is useful or useless for the pair u, v.

Operations We distinguish the four types of movements for the shortcut. Suppose
we move a shortcut pq for T with p, q B such that d(a, p) d(a, q) to p q  with
p , q  B such that d(a, p ) d(a, q  ). The movement from pq to p q  is
an out-shift if p moves to a and q moves to b, i.e., if d(a, p ) d(a, p) and
d(b, q  ) d(b, q) and, thus d(a, p ) d(a, p) d(a, q) d(a, q  ),
an in-shift if p moves away from a and q moves away from b, i.e., if d(a, p)
d(a, p ) and d(b, q) d(b, q  ) and, thus, d(a, p) d(a, p ) d(a, q  ) d(a, q),
an x-shift if p moves to a and q moves away from b, i.e., if d(a, p ) d(a, p)
and d(b, q) d(b, q  ) and, thus, d(a, p ) d(a, p) d(a, q  ) d(a, q), or
a y-shift if p moves away from a and q moves to b, i.e., if d(a, p) d(a, p )
and d(b, q  ) d(b, q) and, thus, d(a, p) d(a, p ) d(a, q) d(a, q  ).
The types of movements intentionally overlap when one of the endpoints remains
stationary, e.g., when p = p every shift towards y is also an outwards shift.
Lemma 3 (Blocking Lemma). Let pq be a shortcut for a tree T with p, q B.
1. If T +pq has a diametral pair of type x-y, then diam(T +pq) diam(T +p q  )
for every shortcut p q  reached with an in-shift from pq.
2. If T +pq has a diametral pair of type x-, then diam(T +pq) diam(T +p q  )
for every shortcut p q  reached with a y-shift from pq.
3. If T +pq has a diametral pair of type -y, then diam(T +pq) diam(T +p q  )
for every shortcut p q  reached with an x-shift from pq.
4. If T + pq has a diametral pair of type - or -, then diam(T + pq)
diam(T + p q  ) for every shortcut p q  reached with an out-shift from pq.
As an immediate consequence of Lemma 3, the shortcut pq must be optimal for
T when each of the four types of movements is blocked by some diametral pair
T + pq. If the augmented tree T + pq has a diametral pair of type x-y, then there
is no need to consider diametral pairs of type -, due to the following.
Lemma 4. If T + pq has diametral pairs x-y and - then T + pq also has
diametral pairs of type x- and -y and, thus, the shortcut pq is optimal for T .
Minimizing the Continuous Diameter when Augmenting a Tree with a Shortcut 309

4 Continuous Algorithm
Inspired by the plane-sweep paradigm, weconceptuallymove the shortcut
continuously while changing its speed and direction at certain events, i.e., when
the pair state or path state changes. Figure 8 describes the continuous algorithm
in terms of the pair states. Initially, we place the shortcut with both endpoints
on the absolute center c and, thus, we start in pair state {x-y}. The algorithm
consists of at most three phases: an outwards shift, a shift towards x or a shift
towards y, and another outwards shift. For the sake of simplicity, we omit pair
states containing x-y and - in Fig. 8, due to Lemma 4, we omit transitions
implied by transitivity, and we omit pair states that are supersets of nal states.

Phase I
x-y Phase II
out-shift Phase III

x-y x-y x-y x-y x-y x-y x-y

- x- x- x- -y -y -

x-shift - x-shift -y y-shift - y-shift

x-shift out-shift y-shift

x-y x-y
x- x- x-
- -
-y -y -y
-y x-
- out-shift -

Fig. 8. The pair states during our search for an optimal shortcut for a tree. There
are three types of states: First, regular states (solid) indicate the pair state and the
operation applied (out-shift, x-shift, or y-shift). Second, transition states (dotted) are
visited only momemtarily while transitioning from one regular state to another. Third,
nal states (double) where we terminate our search and report the best shortcut en-
countered. Under certain conditions, the search may terminate in non-nal states. We
start in state {x-y} with an outward shift. When we reach the pair state {x-y, -}
from {x-y} then we perform both a shift towards x and separately a shift towards y.

Phase I: Shifting Outwards In Phase I, we shorten all diametral paths of type

x-y with an outwards shift, i.e., we move p with unit speed from c towards a and
we move q with unit speed from c towards b. If p reaches a before q reaches b,
then p stays at a while q continues towards b. Likewise, p continues towards a
if q reaches b. In Phase I, the current shortcut is the best shortcut encountered
so far. Phase I ends when the shortcut reaches the end of the backbone, i.e.,
pq = ab, or when a second type of diametral pair appears and Phase II begins.

Phase II: Shifting Sideways Phase II begins when we leave pair state {x-y}.
If we transit from {x-y} to {x-y, x-}, then we shift towards x. If we transit
310 J.-L. De Carufel et al.

from {x-y} to {x-y, -y}, then we shift towards y. If we transit from {x-y} to
{x-y, -}, then we branch into a shift towards x and a shift towards y. Suppose
we reach the pair state {x-y, x-} from {x-y}. All diametral paths in T + pq
contain the path from a to p. We move p closer to a and we move q with a speed
towards a that keeps all diametral paths in balance. This ensures that we remain
in the current pair state until another diametral pair appears. In this state, the
current shortcut is the best shortcut encountered so far. When we reach the pair
state {x-y, -}, we move p towards a and adjust the position of q to balance
the diametral paths of type x-pq-y with those of type -p- and -q-. In this
state, the diameter shrinks and grows with |pq| and the shortest shortcut since
we entered this state is the best shortcut so far. Phase II ends when p reaches
a, when we transit to a pair state containing x-y and -, when we transit to
the nal state {x-y, -, -y} or to {x-y, x-, -y} where Phase III begins.

Phase III: Shifting Outwards Phase III begins when we reach {x-y, x-, -y}.
As x-y, x-, and -y block all other movements, we shift outwards balancing
x- and -y. We immediately transit to {x-, -y}, as the path from x to y via
the shortcut shrinks faster than the paths connecting x- and -y. If we reach
Phase III, then the shortest shortcut encountered during Phase III is optimal.
Phase III ends when p hits a, when q hits b, or when - or - appears.

Optimality Balancing the diametral paths when moving pq ensures that we re-
main in the current pair state until another diametral pair appears. This allows
us to derive invariants from the blocking lemma (Lemma 3) that guarantee that
we encounter an optimal shortcut and it restricts our search considerably: we
are essentially conducting a linear search, since the speed of q is determined by
the speed of p, the path state, and the change in the length of the shortcut.
We elucidate our invariants and their use through an example: In pair state
{x-y, -} of Phase II, the blocking lemma implies that there is an optimal
shortcut p q that we reach from the current shortcut pq with an x-shift, with a
y-shift, or without moving (pq = p q ). Suppose we miss p q during an x-shift
from pq to the next change in the pair state at p q  . This means we reach p q
with an x-shift from pq and with an y-shift from p q  . Let p q  be the last position
during the shift from pq to p q  where the shift from p q  to p q leads towards
x. Then T + p q  is in pair state {x-y, -} and we have p = p or q  = q .
If p = p , then the movement from p q  to p q is an inward shift towards x.
Since x-y blocks any inward shift, we have diam(T + p q  ) diam(T + p q ).
If q  = q , then the movement from p q  to p q is an outward shift towards x.
Since - blocks any outward shift, we have diam(T + p q  ) diam(T + p q ).
Since p q is optimal, p q  is an optimal shortcut that we encounter.
If we could implement and run it, the continuous algorithm would produce an
optimal shortcut pq for T with p on the path from a to c and q on the path from
c to b: In Phase I and III p and q move away from c. In Phase II a diametral pair
of type -y appears before q reaches c during a shift towards x and a diametral
pair of type x- appears before p reaches c during a shift towards y.
Minimizing the Continuous Diameter when Augmenting a Tree with a Shortcut 311

5 Discretization
To discretize the continuous algorithm, we subdivide the continuous motion of
the shortcut with events such that we can calculate the next event and the
change in the continuous diameter of T + pq between subsequent events. We
introduce events, for instance, when the shortcut hits a vertex, when the path
state changes, and when the shortcut changes between shrinking and growing.
During Phase I, the continuous di-
ameter decreases or remains constant.
Therefore, it is sucient for the dis- x-pq-y
crete algorithm to determine where x-pq-
Phase I of the continuous algorithm x-T -
3
ends. At the end of Phase I, we have
x-pq-y
either reached the end of the back- 1 x-pq-y x-pq-y
-p-
bone, i.e., pq = ab, or a diametral pair 2 x-T - 3 x-pq-
-q-
of type x-, -y, or - has appeared
alongside the diametral pairs of type 3
x-y. We may ignore diametral pairs
of type - as they appear together Fig. 9. The transitions between the path
with x- and -y when x-y is diame- states during Phase II. We omit transitory
tral, as shown in Lemma 4. We detect states like {x-pq-y, x-pq-, x-pq-, x-T -}.
Transition 1 is possible when the shortcut
changes in the path state by monitor-
grows. Transition 2 is possible when the
ing the candidates for each type of di- shortcut shrinks. If we take any of the tran-
ametral pathexcept for those con- sitions with label 3 from {x-pq-y, x-pq-sj }
necting diametral pairs of type -. for j = 1, 2, . . . , k, then the pair x, sj ceases
We argue that there are O(n) events to be diametral and cannot become diame-
where we spend O(log n) time to de- tral in Phase II again.
tect whether Phase I is about to end.
To discretize Phase II, we calculate the speeds at which p and q move to
balance the diametral paths and how this impacts the diameter. For instance,
if we balance the paths x-pq-y, x-pq-, and x-T - during an x-shift from pq to
p q  , then dT (q, q  ) = 13 (dT (p, p ) + |pq| |p q  |) and the diameter changes by
2   
3 (dT (p, p ) + |pq| |p q |). With the assumption that p moves with unit speed,
this allows us to predict the position of q. The respective speeds for the path
states and their dependence on the length of the shortcut allow us to bound the
number of path state changes in Phase II by O(n), as hinted in Fig. 9.
During Phase III, the path state has two components: the diametral path
of type x- and the diametral path of type -y. We encounter two challenges
when discretizing Phase III: First, the path state might change (n2 ) times. This
occurs, for instance, when the shortcut alternates (n) times between growing
and shrinking such that there are (n) candidates for diametral pairs, for each
of which the shortcut becomes useful whenever it shrinks and useless whenever it
grows. We circumvent this issue by ignoring certain superuous path state events.
Second, since x-y is no longer diametral, we need to detect diametral pairs of type
- of which there are (k 2 ) candidates, where k is the number of B-sub-trees
of T . The diculty lies in detecting diametral paths of type -pq-. It suces to
312 J.-L. De Carufel et al.

nd the rst position pq where a path si -pq-sj , for i, j {1, 2, . . . , k}, becomes
a diametral path of type -pq-: If we shift outwards from pq, then si -pq-sj
will remain diametral while increasing in length. We proceed as follows. First,
we simulate the modied Phase III without attempting to detect if a diametral
path of type -pq- appears. We record the sequence of edge pairs that we
visit during this simulation. This sequence contains O(n) edge pairs. After the
simulation, we perform a binary search for pq in the sequence of visited edge
pairs. The binary search for pq takes O(n log n) time, since we can determine
the largest path of type -pq- in O(n) time for a xed position of the shortcut.
Therefore, we discretize all phases of the continuous algorithmwith modi-
cations that do not impact optimalitywith O(n) events that we can process
in O(n log n) total time, followed by an O(n log n)-time post-processing step.
Theorem 5. For a geometric tree T with n vertices, it takes O(n log n) time to
determine a shortcut pq that minimizes the continuous diameter of T + pq. 

References
1. Caceres, J., Garijo, D., Gonzalez, A., Marquez, A., Puertas, M.L., Ribeiro, P.:
Shortcut Sets for Plane Euclidean Networks. Electron Notes Discrete Math 54,
163168 (2016)
2. Chepoi, V., Vaxes, Y.: Augmenting Trees to Meet Biconnectivity and Diameter
Constraints. Algorithmica 33(2), 243262 (2002)
3. De Carufel, J.L., Grimm, C., Maheshwari, A., Smid, M.: Minimizing the Continu-
ous Diameter when Augmenting Paths and Cycles with Shortcuts. In: SWAT 2016.
pp. 27:127:14 (2016)
4. Farshi, M., Giannopoulos, P., Gudmundsson, J.: Improving the Stretch Factor of a
Geometric Network by Edge Augmentation. SIAM J. Comp. 38(1), 226240 (2008)
5. Frati, F., Gaspers, S., Gudmundsson, J., Mathieson, L.: Augmenting Graphs to
Minimize the Diameter. Algorithmica 72(4), 9951010 (2015)
6. Gao, Y., Hare, D.R., Nastos, J.: The Parametric Complexity of Graph Diameter
Augmentation. Discrete Appl Math 161(10-11), 16261631 (2013)
7. Groe, U., Gudmundsson, J., Knauer, C., Smid, M., Stehn, F.: Fast Algorithms
for Diameter-Optimally Augmenting Paths. In: ICALP 2015. pp. 678688 (2015)
8. Groe, U., Gudmundsson, J., Knauer, C., Smid, M., Stehn, F.: Fast Algorithms
for Diameter-Optimally Augmenting Paths and Trees (2016), arXiv:1607.05547
9. Hakimi, S.L.: Optimum locations of switching centers and the absolute centers and
medians of a graph. Operations Research 12(3), 450459 (1964)
10. Li, C.L., McCormick, S., Simchi-Levi, D.: On the Minimum-Cardinality-Bounded-
Diameter and the Bounded-Cardinality-Minimum-Diameter Edge Addition Prob-
lems. Operations Research Letters 11(5), 303308 (1992)
11. Luo, J., Wul-Nilsen, C.: Computing Best and Worst Shortcuts of Graphs Embed-
ded in Metric Spaces. In: ISAAC 2008. pp. 764775 (2008)
12. Oh, E., Ahn, H.K.: A Near-Optimal Algorithm for Finding an Optimal Shortcut
of a Tree. In: ISAAC 2016. pp. 59:159:12 (2016)
13. Schoone, A.A., Bodlaender, H.L., van Leeuwen, J.: Diameter Increase Caused by
Edge Deletion. Journal of Graph Theory 11(3), 409427 (1987)
14. Yang, B.: Euclidean Chains and their Shortcuts. Theor Comput Sci 497, 5567
(2013)
Inapproximability of the Standard Pebble Game
and Hard to Pebble Graphs

Erik D. Demaine, Quanquan C. Liu

MIT CSAIL, Cambridge, Massachusetts

Abstract. Pebble games are single-player games on DAGs involving

placing and moving pebbles on nodes of the graph according to a certain
set of rules. The goal is to pebble a set of target nodes using a minimum
number of pebbles. In this paper, we present a possibly simpler proof of
the result in [4] and strengthen the result to show that it is PSPACE-hard
to determine the minimum number of pebbles to an additive n1/3 term
for all > 0, which improves upon the currently known additive constant
hardness of approximation [4] in the standard pebble game. We also
introduce a family of explicit, constant indegree graphs with n nodes

where there exists a graph in the family such that using 0 < k < n
k
pebbles requires ((n/k) ) moves to pebble in both the standard and
black-white pebble games. This independently answers an open question
summarized in [14] of whether a family of DAGs exists that meets the
upper bound of O(nk ) moves using constant k pebbles with a dierent
construction than that presented in [1].

1 Introduction

Pebble games were originally introduced to study compiler operations and pro-
gramming languages. For such applications, a DAG represents the computational
dependency of each operation on a set of previous operations and pebbles repre-
sent register allocation. Minimizing the amount of resources allocated to perform
a computation is accomplished by minimizing the number of pebbles placed on
the graph [16]. The standard pebble game (also known as the black pebble game)
is traditionally used to model such behavior. In the standard pebble game, one
is given a DAG, G = (V, E), with n nodes and constant indegree and told to
perform a set of pebbling moves that places, removes, or slides pebbles around
the nodes of G.
The premise of such games is given some input modeled by source nodes
S V one should compute some set of outputs modeled as target nodes T V .
In terms of G, S is typically the set of nodes without incoming edges and T
is typically the set of nodes without outgoing edges. The rules of the standard
pebble game are as follows:

Springer International Publishing AG 2017 313

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 313324, 2017.
DOI: 10.1007/978-3-319-62127-2_27
314 E.D. Demaine and Q.C. Liu

Standard Pebble Game

Input: Given a DAG, Gn, = (V, E). Let pred(v) = {u V : (u, v) E}. Let
S V be the set of sources of G and T V be the set of targets of G. Let
P = {P0 , . . . , P } be a valid pebbling strategy that obeys the following rules
where Pi is a set of nodes containing pebbles at timestep i and P0 = and
P = {T }. Let Peb(G, P) = maxi[ ] {|Pi |}.

Rules:
1. At most one pebble can be placed or removed from a node at a time.
2. A pebble can be placed on any source, s S.
3. A pebble can be removed from any vertex.
4. A pebble can be placed on a non-source vertex, v, at time i if and only if its
direct predecessors are pebbled, pred(v) Pi1 .
5. A pebble can slide from vertex v to vertex w at time i if and only if (v, w) E
and pred(w) Pi1 .
Goal: Determine minP {Peb(G, P)} using a valid strategy P.

In addition to the standard pebble game, other pebble games are useful for
studying computation. The red-blue pebble game is used to study I/O complex-
ity [12], the reversible pebble game is used to model reversible computation [3],
and the black-white pebble game is used to model non-deterministic straight-line
programs [5]. Although we will be proving a result about the black-white pebble
game in Section 4, we will defer introducing the rules of the game to our full
paper [6] since the black-white pebble game is not central to the main results of
this paper.
Much previous research has focused on proving lower and upper bounds on the
pebbling space cost (i.e. the maximum number of pebbles over time) of pebbling a
given DAG under the rules of each of these games. For all of the aforementioned
pebble games (except the red-blue pebble game since it relies on a dierent set
of parameters), any DAG can be pebbled using O(n/ log n) pebbles [9,11,15].
Furthermore, there exist DAGs for each of the games that require (n/ log n)
pebbles [9,11,15].
It turns out that nding a strategy to optimally pebble a graph in the standard
pebble game is computationally dicult even when each vertex is allowed to be
pebbled only once. Specically, nding the minimum number of black pebbles
needed to pebble a DAG in the standard pebble game is PSPACE-complete [8]
and nding the minimum number of black pebbles needed in the one-shot case is
NP-complete [16]. In addition, nding the minimum number of pebbles in both
the black-white and reversible pebble games have been recently shown to be
both PSPACE-complete [4,10]. But the result for the black-white pebble game is
proven for unbounded indegree [10]. A key open question in the eld is whether
hardness results can be obtained for constant indegree graphs for the black-white
pebble game. However, whether it is possible to nd good approximate solutions
to the minimization problem has barely been studied. In fact, it was not known
Inapproximability of the Standard Pebble Game and Hard to Pebble Graphs 315

until this paper whether it is hard to nd the minimum number of pebbles

within even a non-constant additive factor [4]. The best known multiplicative
approximation factor is the very loose (n/ log n) which is the pebbling space
upper bound [11], leaving much room for improvement.
Our results deal primarily with the standard pebble game, but we believe
that the techniques could be extended to show hardness of approximation for
other pebble games. We prove the following:

Theorem 1. The minimum number of pebbles needed in the standard pebble

game on DAGs with maximum indegree 2 is PSPACE-hard to approximate to
within an additive n1/3 for any > 0.

In addition to determining the pebbling space cost, we sometimes also care

about pebbling time which refers to the number of operations (placements,
removals, or slides) that a strategy uses. For example, such a situation arises if
we care not only about the memory used in computation but also the time of
computation. It is previously known that there exists a family n
of graphs such
that, given ( logn n ) pebbles, one is required to use (2( log n ) ) moves to pebble
any graphs with n nodes in the family [13].
Less is known about the trade-os when a small number (e.g. constant k) of
pebbles is used until the very recent, independent result presented in [1]. It can
be easily shown through a combinatorial argument that the maximum number of
moves necessary using k = O(1) pebbles to pebble n nodes is O(nk ) [14]. It is an
open question whether it is possible to prove a time-space trade-o such that
using k = O(1) pebbles requires (nk ) time. In this paper, we resolve this open
question for both the standard pebble and the black-white pebble games using
an independent construction from that presented in [1].

Theorem 2. There exists a family of graphs with n vertices and maximum

2
indegree 2 such that (( nk k
k ) )moves are necessary to pebble any graph with n
vertices in the family using k < n pebbles in both the standard and black-white
pebble games.

In particular, when k = O(1), the number of moves necessary to pebble a

graph in the family is (nk ).
The organization of the paper is as follows. First, in Section 2, we provide the
denitions and terminology we use in the remaining parts of the paper. Then, in
Section 3, we provide a proof for the inapproximability of the standard pebble
game to an n1/3 additive factor.

In Section 4, we present our hard to pebble graph families using k < n
pebbles and prove that the family takes (nk ) moves to pebble in both the
standard and black-white pebble games when k = O(1).
Finally, in Section 5, we discuss some open problems resulting from this
paper.
316 E.D. Demaine and Q.C. Liu

2 Denitions and Terminology

In this section, we dene the terminology we use throughout the rest of the paper.
All of the pebble games we consider in this paper are played on directed acyclic
graphs (DAGs). In this paper, we only consider DAGs with maximum indegree 2.
We dene such a DAG as G = (V, E) where |V | = n and |E| = m.
The purpose of any pebble game is to pebble a set of targets T V using
minimum number of pebbles. In all pebble games we consider, a player can always
place a pebble on any source node, S V . Usually, S consists of all nodes with
indegree 0 and T consists of all nodes with outdegree 0.
A sequential pebbling strategy, P = [P0 , . . . , P ] is a series of congurations
of pebbles on G where each Pi is a set of pebbled vertices Pi V . Pi follows
from Pi1 by the rules of the game and P0 = and P = T . Then, by denition,
|Pi | is the number of pebbles used in conguration Pi . For a sequential strategy,
|Pi1 | 1 |Pi | |Pi1 | + 1 for all i [ ] = [1, . . . , ] (i.e. at most one pebble
can be placed, removed, or slid on the graph at any time). In this paper, we only
consider sequential strategies.
Given any strategy P for pebbling G, the pebbling space cost, Peb(G, P), of P
is dened as the maximum number of pebbles used by the strategy at any time:
Peb(G, P) = maxi[ ] {|Pi |}.
The minimum pebbling space cost of G, Peb(G), is dened as the smallest
space cost over the set of all valid strategies, P, for G:

Denition 1 (Minimum Pebbling Space Cost).

Peb(G) = min{Peb(G, P)}.

PP

The pebbling time cost, Time(G, P, s) = |P| 1, of a strategy P using s

pebbles is the number of moves used by the strategy. The minimum pebbling
time cost of any strategy that has pebbling space cost s is the minimum number
of moves used by any such strategy.

Denition 2 (Minimum Pebbling Time Cost).

Time(G, s) = min {Time(G, P  , s)} n.

P  {PP:|P|s}

3 Inapproximability of the Standard Pebble Game

In this section, we provide an alternative proof of the result presented in [4] that
the standard pebble game is inapproximable to any constant additive term. Then,
we show that our proof technique can be used to show our main result stated in
Theorem 1. We make modications to the proof presented by [8] to obtain our
main result. A quick explanation of the relevant results presented in [8] can be
found in our full paper [6].
Inapproximability of the Standard Pebble Game and Hard to Pebble Graphs 317

3.1 Inapproximability to n1/3 additive term for any > 0

We now prove our main result. For our reduction we modify the variable, clause,
and quantier gadgets in [8] to produce a gap reduction from the PSPACE-hard
problem, QBF.

Important Graph Components Before we dive into the details of our con-
struction, we rst mention two subgraphs and the properties they exhibit.
The rst graph is the pyramid graph (please refer to [6] for a gure), h with
height h, which requires a number of pebbles that is equal to the height, h, of the
pyramid to pebble [8]. Therefore, in order to pebble the apex of such a graph, at
least h pebbles must be available. As in [8], we depict such pyramid graphs by a
triangle with a number indicating the height (hence number of pebbles) needed
to pebble the pyramid (see Figure 1 for an example of the triangle symbolism).
We make use of the following denition and lemma (restated and adapted)
from [8] in our proofs:

Denition 3 (Frugal Strategy [8]). A pebbling strategy, P, is frugal if the

following are true:

1. Suppose vertex v G is pebbled for the rst time at time t . Then, for all
times, t > t , some path from v to q1 (the only target node) contains a pebble.
2. At all times after v is pebbled for the last time, all paths from v to q1 contain
a pebble.
3. The number of pebble placements on any vertex v G where v =  q1 , is
bounded by the number of pebble placements on pred(v).

Lemma 1 (Normal Pebbling Strategy [8]). If the target vertex is not inside
a pyramid, h , and each of the vertices in the bottom level of the pyramid has at
most one predecessor each, then any pebbling strategy can be transformed into a
normal pebbling strategy without increasing the number of pebbles used. A normal
pebbling strategy is one that is frugal and after the rst pebble is placed on any
pyramid, h , no placements of pebbles occurs outside h until the apex of h is
pebbled and all other pebbles are removed from h .

The other important subgraph is the road graph (see [6] for gure), Rw with
width w, which requires a number of pebbles that is the width of the graph to
pebble any of the outputs [7,14]. Therefore, we state as an immediately corollary
of their proof:

Corollary 1 (Road Graph Pebbling). To pebble O {o1 , . . . , ow } of the

outputs of Rw , with a valid strategy, P = [P0 , . . . , P ] where P = O, requires
w + |O| 1 pebbles.

We dene a regular pebbling strategy for road graphs similarly to the normal
pebbling strategy for pyramids.
318 E.D. Demaine and Q.C. Liu

Lemma 2 (Regular Pebbling Strategy). If each input, ij {i1 , . . . , iw }, to

the road graph has at most 1 predecessor, any pebbling strategy can be transformed
into a regular pebbling strategy without increasing the number of pebbles used. A
regular pebbling strategy is one that is frugal and after the rst pebble is placed
on any road graph, Rw , no placements of pebbles occurs outside Rw until a set of
desired outputs, O {o1 , . . . , ow }, of Rw all contain pebbles and all other pebbles
are removed from Rw .

We immediately obtain the following corollary from Lemmas 1 and 2:

Corollary 2. Any pebbling strategy, P, can be transformed into a pebbling strat-

egy, P  , that is normal and regular if no target vertices lie inside a pyramid or
road graph and each input node to either the pyramid or road graph has at most
one predecessor.

Modied Graph Constructions We rst describe the changes we made to

each of the gadgets used in the PSPACE-completeness proof presented by [8] and
then prove our inapproximability result using these gadgets in Section 3.1. Given
a QBF instance, B = Q1 x1 Qu xu F , with c clauses, we create the following
gadgets:
Variable Nodes: We replace all variable nodes in the proof provided in [8]
with road graphs each of width K. The modied variable nodes are shown
in Figure 1. Each variable node as in the original proof by [8] has 3 possible
congurations which are also shown in Figure 1.

Fig. 1: Modied variable gadget with 3 possible congurations using the road
graph subgraph previously described. Here K = 3. a) xi is True. b) xi is True. c)
Double false.

Quantier Blocks: Each universal and existential quantier blocks are also
modied to account for the new variable nodes. See Figure 2 and Figure 3 which
depict the new quantier gadgets that use the new variable nodes. Note that
instead of each quantier gadget requiring 3 total pebbles, each gadget requires
3K pebbles to remain on each block before the clauses are pebbled. The basic
idea is to expand all nodes ai , bi , ci ...etc. into a path of length K to account for
each of the K copies of xi and each of the K copies of xi . Each si = si1 3K
and s1 = 3Ku + 3K + 1.
Inapproximability of the Standard Pebble Game and Hard to Pebble Graphs 319

Fig. 2: Modied universal quantier block. Here K = 3.

Fig. 3: Modied existential quantier block. Here K = 3.

320 E.D. Demaine and Q.C. Liu

Clause Gadgets: Each clause gadget is modied to be a pyramid of height

3K + 1 where the bottom layer is connected to 2 nodes from two dierent literals.
Therefore, for a given clause (li , lj , lk ), K nodes are connected to li and lj , K
nodes to lj and lk , and K nodes to li and lk . See Figure 4 for an example of the
modied clause gadget.

Fig. 4: Modied clause gadget. The clause here is (xi , xl , xr ) where xi = T rue,
xl = F alse, and xr = F alse. Here K = 3.

Proofs of the Construction We construct a graph G using the gadgets

described above in Section 3.1 for any given QBF instance, B = Q1 x1 Qk xu F .
In short, the proof relies on the fact that each quantier gadget requires 2K
pebbles to set the corresponding variable to true or false (i.e. the corresponding
literals to true or false). An additional K pebbles need to remain on each quantier
in order to be able to repebble quantiers when checking for universal variables
satisfaction. Furthermore, a clause would consist of modied pyramids of height
3K + 1 connected to pairs of nodes from dierent literals. Following the proof
in [8], the quantier gadgets are pebbled rst with 3Ku pebbles remaining on
the quantier gadgets. Then, the clauses are pebbled with 3K + 1 pebbles.
If B is satisable, then clauses can be pebbled with 3K +1 pebbles. Otherwise,
4K pebbles are needed to pebble one or more unsatised clauses in G, resulting
in a gap of K 1 pebbles between when B is satisable and unsatisable. Thus,
if given an approximation algorithm that estimates the number of pebbles needed
within additive K 1, we can distinguish between the case when B is satisable
(at most 3Km + 3K + 1 pebbles are needed) and the case when B is unsatisable
(when at least 3Km + 4K pebbles are needed).
In this construction, K can be any polynomial function of u where u is the
number of variables in B and c is the number of clauses (in other words, K = ua cb
Inapproximability of the Standard Pebble Game and Hard to Pebble Graphs 321

for any constants a and b). The total minimum number of pebbles necessary is
O(Ku) and the total number of nodes in the graph is O(K 3 (u + c)).
We rst prove that the number of pebbles needed to pebble each quantier
gadget is 3K and 3K pebbles remain on the quantier blocks throughout the
pebbling of the clauses.

Lemma 3. Every regular and normal strategy, P  , must be one where each
quantier gadget must be pebbled with 3K pebbles before the clauses are pebbled.
Furthermore, each quantier gadget must be pebbled wth 3K pebbles when qu is
pebbled.

Next we prove that provided 3Ku pebbles stay on the quantier blocks, each
unsatised clause requires 4K pebbles.

Lemma 4. Given a clause gadget, Ci , its corresponding variable, ci is true if

and only if Ci can be pebbled with 3K + 1 pebbles. Furthermore, if ci is false and
all literals in Ci are set in the false conguration, then at least 4K pebbles are
necessary to pebble the clause.

Given the previous proofs, we now prove the following key lemmas:

Lemma 5. Given G which is constructed from the provided QBF instance, B =

Q1 x1 Qu xu F , using our modied reduction in Section 3.1, B is satisable if
and only if Peb(G) 3Ku + 3K + 1.

Before, we prove the next crucial lemma (Lemma 7), we rst prove the
following lemma which will help us prove Lemma 7:

Lemma 6. Let Ni be the conguration such that some number of pebbles are on
the rst i 1 quantier blocks and the i-th is being pebbled. Therefore, Nu+1 is
the conguration when some number of pebbles are on all u quantier blocks and
the rst clause gadget is being pebbled. There does not exist a frugal strategy, P,
that can pebble our reduction construction, G, such that Ni contains less than
s si pebbles on the rst i 1 quantier blocks when the i-th quantier block or
when the rst clause is being pebbled.

Lemma 7. Given G which is constructed from the provided QBF instance, B =

Q1 x1 Qu xu F , using our modied reduction in Section 3.1, B is unsatisable
if and only if Peb(G) 3Ku + 4K.

Proof of Inapproximability Using Lemmas 5 and 7, we prove that it is

PSPACE-hard to approximate the minimum number of black pebbles needed
given a DAG, G, to an additive n1 factor.

Theorem 3 (Restatement of Theorem 1). The minimum number of peb-

bles needed in the standard pebble game on DAGs with maximum indegree 2 is
PSPACE-hard to approximate to additive n1/3 for > 0.
322 E.D. Demaine and Q.C. Liu

Proof. From Lemmas 5 and 7, the cost of pebbling a graph constructed from
a satisable B is at most 3Ku + 3K + 1 whereas the cost of pebbling a graph
constructed from an unsatisable B is at least 3Ku + 4K. As we can see, the
aforementioned reduction is a gap-producing reduction with a gap of K 1
pebbles. Then, all that remains to be shown is that for any > 0, it is the case
that K (K 3 (u + c))(1/3) . (Note that for > 1/3, setting K to any positive
integer achieves this bound.) Suppose we set K = max(u, c)a where a > 0. Given
an 0 < 1/3, a = 1/3
1+3 0 precisely when is in the stated range.
For values of a 0, we can duplicate the clauses and variables gadgets so that
u and c are large enough such that K = max(u, c)a 2. Let d = max(u, c). Then,
we need d to be large enough so that da 2 (i.e. we want da to be some integer).
Then, we can set d 21/a . Thus, we can duplicate the number of variables and
1+3
clauses so that d 2 1/3 .
Therefore, for every > 0, we can construct a graph with a specic K
calculated from such that it is PSPACE-hard to nd an approximation within
an additive n1/3 factor where n is the number of nodes in the graph.

4 Hard to Pebble Graphs for Constant k Pebbles

It is long known that the maximum number of moves necessary to pebble any
graph with constant k pebbles is O(nk ). (Note that the maximum number of
moves necessary to pebble any graph is either O(nk1 ) or O(nk ) depending on
whether or not sliding is allowed. Here, we allow sliding in all of our games. The
bound of O(nk1 ) proven in [14] is one for the case when sliding is not allowed.)
The upper bound of O(nk ) for any constant k number of pebbles submits to a
simple combinatorial proof adapted from [14]) to account for sliding. However,
to the best of the authors knowledge, examples of such families of graphs that
require O(nk ) moves to pebble using k pebbles did not exist until very recently
in an independent work [1]. In this section, we present an independent, simple
to construct family of graphs that require (nk ) time for constant k number of
pebbles in both the standard and black-white pebble games. We further reduce
the indegree of nodes in this family of graphs to 2 and show that our results still
hold. Furthermore, we show this family of graphs to exhibit a steep time-space
trade-o (from exponential in k to linear) even when k is not constant. Such
families of graphs could potentially have useful applications in cryptography in
the domain of proofs of space and memory-hard functions [2].
We construct the following family of graphs, Hn, , below with n nodes and
indegree and show that for constant k pebbles, the number of steps it takes to
pebble the graph Hn,k Hn, with k pebbles is (nk ). We also show a family of
graphs, Hn,2 with indegree 2 that shows the same asymptotic tradeo.
We construct the family of graphs in the following way.
Denition
4. Given a set of n nodes and maximum number of pebbles k where
k < n, we lexicographically order the nodes (from 1 to n) and create the
following set of edges between the nodes where directed edges are directed from vi
to vj where i < j:
Inapproximability of the Standard Pebble Game and Hard to Pebble Graphs 323

1. vi and vi+1 for all i [k, n]

2. vi and vj for all i [l 1] for all 2 l k and j {f (l) + 2r 2} for all
r [ nk
2k + 1] where f (l) = k + (l 1)( k ) + 1.
nk

3. vi and vj for all i = f (l) 1 and j {f (l) + 2r 1} for all i [ nk

2k + 1]
where l [1, k 1].
The target node (the only sink) is vn . Note that the sources in our construction
are vj for all j [1, k].
Due to the space constraints, we leave all proofs of the properties of the graph
family as well as an example gure of a member of the family in our full paper [6].

5 Open Problems
There are a number of open questions that naturally follow the content of this
paper.
The rst obvious open question is whether the techniques introduced in this
paper can be tweaked to allow for a PSPACE-hardness of approximation to an
n1 additive factor for any > 0. We note that the trivial method of attempting
to reduce the size of the subgraph gadgets used in the variables (i.e. use a dierent
construction than the road graph such that less than K 3 nodes are used) is not
sucient since the number of nodes in the graph is still (K 3 (u + c)). This is not
to say that such an approach is not possible; simply that more changes need to
be made to all of the other gadgets. The next logical step is to determine whether
Peb(G) can be approximated to a constant 2 factor multiplicative approximation.
Another open question is whether the techniques introduced in this paper
can be applied to show hardness of approximation results for other pebble games
such as the black-white or reversible pebble games. The main open question in
the topic of hardness of approximation of pebble games is whether the standard
pebble game can be approximated to any factor smaller than n/ log n or whether
the games are PSPACE-hard to approximate to any constant factor, perhaps
even logarithmic factors.
With regard to hard to pebble graphs, we wonder if our graph family could
be improved to show (nk ) for any 0 < k n/ log n. This would be interesting
because to the best of the authors knowledge we do not yet know of any graph
families that exhibit sharp (asymptotically tight) time-space trade-os for this
entire range of pebble number.

References
1. Joel Alwen, Susanna F. de Rezende, Jakob Nordstrom, and Marc Vinyals. Cumula-
tive space in black-white pebbling and resolution. In Innovations in Theoretical
Computer Science, ITCS 2017, Berkeley, CA, USA, 9-11 January, 2017, 2017.
2. Joel Alwen and Vladimir Serbinenko. High parallel complexity graphs and memory-
hard functions. In Proceedings of the Forty-Seventh Annual ACM on Symposium on
Theory of Computing, STOC 2015, Portland, OR, USA, June 14-17, 2015, pages
595603, 2015. Available from: http://doi.acm.org/10.1145/2746539.2746622.
324 E.D. Demaine and Q.C. Liu

3. Charles H. Bennett. Time/space trade-os for reversible computation. SIAM J.

Comput., 18(4):766776, August 1989. Available from: http://dx.doi.org/10.
1137/0218053.
4. Siu Man Chan, Massimo Lauria, Jakob Nordstrom, and Marc Vinyals. Hardness of
approximation in PSPACE and separation results for pebble games. In IEEE 56th
Annual Symposium on Foundations of Computer Science, FOCS 2015, Berkeley,
CA, USA, 17-20 October, 2015, pages 466485, 2015. Available from: http://dx.
doi.org/10.1109/FOCS.2015.36.
5. Stephen Cook and Ravi Sethi. Storage requirements for deterministic / polynomial
time recognizable languages. In Proceedings of the Sixth Annual ACM Symposium
on Theory of Computing, STOC 74, pages 3339, New York, NY, USA, 1974. ACM.
Available from: http://doi.acm.org/10.1145/800119.803882.
6. Erik D. Demaine and Quanquan C. Liu. Inapproximability of the standard pebble
game and hard to pebble graphs. CoRR, 2017.
7. Peter Emde Boas and Jan Leeuwen. Theoretical Computer Science 4th GI Con-
ference: Aachen, March 2628, 1979, chapter Move rules and trade-os in the
pebble game, pages 101112. Springer Berlin Heidelberg, Berlin, Heidelberg, 1979.
Available from: http://dx.doi.org/10.1007/3-540-09118-1_12.
8. John R. Gilbert, Thomas Lengauer, and Robert Endre Tarjan. The pebbling
problem is complete in polynomial space. In Proceedings of the Eleventh Annual
ACM Symposium on Theory of Computing, STOC 79, pages 237248, New York, NY,
USA, 1979. ACM. Available from: http://doi.acm.org/10.1145/800135.804418.
9. John R. Gilbert and Robert E Tarjan. Variations of a pebble game on graphs.
Technical report, Stanford, CA, USA, 1978.
10. Philipp Hertel and Toniann Pitassi. The PSPACE-completeness of black-white
pebbling. SIAM J. Comput., 39(6):26222682, April 2010. Available from: http:
//dx.doi.org/10.1137/080713513.
11. John Hopcroft, Wolfgang Paul, and Leslie Valiant. On time versus space. J. ACM,
24(2):332337, April 1977. Available from: http://doi.acm.org/10.1145/322003.
322015.
12. Hong Jia-Wei and H. T. Kung. I/O complexity: The red-blue pebble game. In
Proceedings of the Thirteenth Annual ACM Symposium on Theory of Computing,
STOC 81, pages 326333, New York, NY, USA, 1981. ACM. Available from:
http://doi.acm.org/10.1145/800076.802486.
13. Thomas Lengauer and Robert Endre Tarjan. Upper and lower bounds on time-space
tradeos. In Proceedings of the Eleventh Annual ACM Symposium on Theory of
Computing, STOC 79, pages 262277, New York, NY, USA, 1979. ACM. Available
from: http://doi.acm.org/10.1145/800135.804420.
14. Jakob Nordstrom. New wine into old wineskins: A survey of some pebbling classics
with supplemental results. 2015.
15. Wolfgang J. Paul, Robert Endre Tarjan, and James R. Celoni. Space bounds
for a game on graphs. In Proceedings of the Eighth Annual ACM Symposium on
Theory of Computing, STOC 76, pages 149160, New York, NY, USA, 1976. ACM.
Available from: http://doi.acm.org/10.1145/800113.803643.
16. Ravi Sethi. Complete register allocation problems. SIAM J. Comput., 4(3):226248,
1975. Available from: http://dx.doi.org/10.1137/0204020.
 Capacitated Center Problems with Two-Sided
Bounds and Outliers

Hu Ding1 , Lunjia Hu2 , Lingxiao Huang2 and Jian Li2

1
Computer Science and Engineering, Michigan State University, East Lansing, MI,
USA 
2
Institute for Interdisciplinary Information Sciences, Tsinghua University, Beijing,
China

Abstract. In recent years, the capacitated center problems have at-

tracted a lot of research interest. Given a set of vertices V , we want
to nd a subset of vertices S, called centers, such that the maximum
cluster radius is minimized. Moreover, each center in S should satisfy
some capacity constraint, which could be an upper or lower bound on
the number of vertices it can serve. Capacitated k-center problems with
one-sided bounds (upper or lower) have been well studied in previous
work, and a constant factor approximation was obtained.
We are the rst to study the capacitated center problem with both ca-
pacity lower and upper bounds (with or without outliers). We assume
each vertex has a uniform lower bound and a non-uniform upper bound.
For the case of opening exactly k centers, we note that a generaliza-
tion of a recent LP approach can achieve constant factor approximation
algorithms for our problems. Our main contribution is a simple combi-
natorial algorithm for the case where there is no cardinality constraint
on the number of open centers. Our combinatorial algorithm is simpler
and achieves better constant approximation factor compared to the LP
approach.

1 Introduction
The k-center clustering is a fundamental problem in theoretical computer science
and has numerous applications in a variety of elds. Roughly speaking, given a
metric space containing a set of vertices, the k-center problem asks for a subset
of k vertices, called centers, such that the maximum radius of the induced k
clusters is minimized. Actually k-center clustering falls in the umbrella of the
general facility location problems which have been extensively studied in the
past decades. Many operation and management problems can be modeled as
facility location problems, and usually the input vertices and selected centers
are also called clients and facilities respectively. In this paper, we consider a

HD is supported by the start-up fund from Michigan State University.

LJH, LXH and JL are supported in part by the National Basic Research Program
of China grants 2015CB358700, 2011CBA00300, 2011CBA00301, and the National
NSFC grants 61033001, 61632016, 61361136003.

Springer International Publishing AG 2017 325

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 325336, 2017.
DOI: 10.1007/978-3-319-62127-2_28
326 H. Ding et al.

signicant generalization of the k-center problem, where each vertex is associated

with a capacity interval; that is, the cardinality of the resulting cluster centered
at the vertex should satisfy the given lower and upper capacity bounds (the
formal denition is shown in Section 1.2). In addition, we also consider the case
where a given number of vertices may be excluded as outliers.
Besides being a natural combinatorial problem on its own, the k-center prob-
lem with both capacity upper and lower bounds is also strongly motivated by
several realistic issues raised in a variety of application contexts.

1. In the context of facility location, each open facility may be constrained by

the maximum number of clients it can serve. The capacity lower bounds
also come naturally, since an open facility needs to serve at least a certain
number of clients in order to generate prot.
2. Several variants of the k-center clustering have been used in the context of
preserving privacy in publication of sensitive data (see e.g., [1, 23, 26]). In
such applications, it is important to have an appropriate lower bound for
the cluster sizes, in order to protect the privacy to certain extent (roughly
speaking, it would be relatively easier for an adversary to identify the clients
inside a too small cluster).
3. Consider the scenario where the data is distributed over the nodes in a large
network. We would like to choose k nodes as central servers, and aggregate
the information of the entire network. We need to minimize the delay (i.e.,
minimize the cluster radius), and at the same time consider the balancedness,
for the obvious reason that the machines receiving too much data could be
the bottleneck of the system and the ones receiving too little data is not
suciently energy-ecient [11].

Our problem generalizes the classic k-center problem as well as many impor-
tant variants studied by previous authors. The optimal approximation results for
the classic k-center problem appeared in the 80s: Gonzalez [15] and Hochbaum
and Shmoys [17] provided a 2-approximation in a metric graph; moreover, they
proved that any approximation ratio c < 2 would imply P = N P . The rst study
on capacitated (with only upper bounds) k-center clustering is due to Bar-Ilan et
al. [5] who provided a 10-approximation algorithm for uniform capacities (i.e., all
the upper bounds are identical). Further, Khuller and Sussmann [20] improved
the approximation ratio to be 6 and 5 for hard and soft uniform capacities, re-
spectively. 3 The recent breakthrough for non-uniform (upper) capacities is due
to Cygan et al. [10]. They developed the rst constant approximation algorith-
m based on LP rounding, though their approximation ratio is about hundreds.
Following this work, An et al. [3] provided an approximation algorithm with the
much lower approximation ratio 9. On the imapproximability side, it is impos-
sible to achieve an approximation ratio lower than 3 for non-uniform capacities
unless P = N P [10].
3
We can open more than one copies of a facility in the same node in the soft capacity
version. But in the hard capacity version, we can only open at most one copy.
Capacitated Center Problems with Two-Sided Bounds and Outliers 327

For the ordinary k-center with outliers, a 3-approximation algorithm was

obtained by Charikar et al. [8]. Kociumaka and Cygan [21] studied k-center with
non-uniform upper capacities and outliers, and provided a 25-approximation
algorithm.
k-center clustering with lower bounds on cluster sizes was rst studied in
the context of privacy-preserving data management [26]. Aggarwal et al. [1]
provided a 2-approximation and a 4-approximation for the cases without and
with outliers, respectively. Further, Ene et al. [13] presented a near linear time
(4 + )-approximation algorithm in constant dimensional Euclidean space. Note
that both [1, 13] are only for uniform lower bounds. Recently, Ahmadian and
Swamy [2] provided a 3-approximation and a 5-approximation for the non-
uniform lower bound case without and with outliers.
Our main results. To the best of our knowledge, we are the rst to study
the capacitated center with both capacity lower and upper bounds (with or
without outliers). Recently, Ding [12] also studies k-center clustering with two-
sided bounds in high dimension or any metric space when k is a constant, and
provides a nearly linear time 4-approximation. Given a set V of n vertices, we
focus on the case where the capacity of each vertex u V has a uniform lower
bound Lu = L and a non-uniform upper bound Uu . Sometimes, we consider a
generalized supplier version where we are only allowed to open centers among
a facility set F, see Denition 1 for details. We mainly provide rst constant
factor approximation algorithms for the following variants, see Table 1 for other
results. Due to the lack of space, we defer many details and proofs to a full
version.

1. (L,U ,soft-,p)-Center (Section 2.2): In this problem, both the lower bounds
and the upper bounds are uniform, i.e., Lu = L, Uu = U for all u V . The
number of open centers can be arbitrary, i.e., there is no requirement to
choose exactly k open centers. Moreover, we allow multiple open centers at
a single vertex u V (i.e., soft capacity). We may exclude n p outliers. We
provide the rst polynomial time combinatorial algorithm which can achieve
an approximate factor of 5.
2. (L,{Uu },,p)-Center(Section 2.3): In this problem, the lower bounds are
uniform, i.e., Lu = L for all u V , but the upper bound can be nonuniform.
The number of open centers can be arbitrary. We may exclude n p outliers.
We provide the rst polynomial time combinatorial 11-approximation for this
problem.
3. (L,{Uu },k)-Center (Section 3.3): In this problem, we would like to open
exactly k centers, such that the maximum cluster radius is minimized. All
vertices have the same capacity lower bounds, i.e., Lu = L for all u V .
But the capacity upper bounds may be nonuniform, i.e., each vertex u has
an individual capacity upper bound Uu . Moreover, we do not exclude any
outlier. We provide the rst polynomial time 9-approximation algorithm for
this problem, based on LP rounding.
4. (L,{Uu },k,p)-Center (Section 3.3): This problem is the outlier version of
the (L,{Uu },k)-Center problem. The problem setting is exactly the same
328 H. Ding et al.

except that we can exclude np vertices as outliers. We provide a polynomial

time 25-approximation algorithm for this problem.

Approximation Ratio
Problem Setting
Center Version Supplier Version
(L,U ,soft-,p) 5 5
(L,U ,,p) 10 23
Without k Constraint
(L,{Uu },soft-,p) 11 11
(L,{Uu },,p) 11 25
(L,U ,k) 6 9
(L,{Uu },k) 9 13
(L,U ,soft-k,p) 13 13
With k Constraint
(L,U ,k,p) 23 23
(L,{Uu },soft-k,p) 25 25
(L,{Uu },k,p) 25 25
Table 1. A summarization table for our results in this paper.

Our main techniques. In Section 2, we consider the rst two variants which
allow to open arbitrarily many centers. We design simple and faster combinatori-
al algorithms which can achieve better constant approximation ratios compared
to the LP approach. For the simpler case (L,U ,soft-,p)-Center, we construct a
data structure for all possible open centers. We call it a core-center tree (CCT).
Our greedy algorithm mainly contains two procedures. The rst procedure pass-
up greedily assigns vertices to open centers from the leaves of CCT to the root.
After this procedure, there may exist some unassigned vertices around the root.
We then introduce the second procedure called pass-down, which assigns these
vertices in order by nding an exchange route each time. For the more general
case (L,{Uu },,p)-Center, our greedy algorithm is similar but somewhat more
subtle. We still construct a CCT and run the pass-up procedure. Then we obtain
an open center set F , which may contain redundant centers. However, since we
deal with hard capacities and outliers, we need to nd a non-redundant open
center set which is not too far from F (see Section 2.3 for details) and have
enough total capacities. Then by a pass-down procedure, we can assign enough
vertices to their nearby open centers.
In Section 3 and 3.3, we consider the last two variants which require to open
exactly k centers. We generalized the LP approach developed for k-center with
only capacity upper bounds [3, 21] and obtain constant approximation schemes
for two-sided capacitated bounds. The omitted proofs can be found in the full
version in this paper.
1.1 Other Related Work
The classic k-center problem is quite fundamental and has been generalized in
many ways, to incorporate various constraints motivated by dierent application
scenarios. Recently, Fernandes et al. [14] also provided constant approximations
for the fault-tolerant capacitated k-center clustering. Chen et al. [9] studied the
Capacitated Center Problems with Two-Sided Bounds and Outliers 329

matroid center problem where the selected centers must form an independent
set of a given matroid, and provided constant factor approximation algorithms
(with or without outliers).
There is a large body of work on approximation algorithms for the facility lo-
cation and k-median problems (see e.g., [4, 6, 7, 16, 18, 19, 22, 24, 25]). Moreover,
Dick et al. [11] studied multiple balanced clustering problems with uniform ca-
pacity intervals, that is, all the lower (upper) bounds are identical; they also
consider the problems under the stability assumption.
1.2 Preliminaries
In this paper, we usually work with the following more general problem, called
the capacitated k-supplier problem. It is easy to see it generalizes the capaci-
tated k-center problem since we can not open centers at any vertex. The formal
denition is as follows.

Denition 1. (Capacitated k-supplier with two-sided bounds and outliers) Sup-

pose that we have
1. Two integers k, p ZZ 0 ;
2. A nite set C of clients, and a nite set F of facilities;
3. A symmetric distance function d : (C F) (C F) IR0 satisfying the
triangle inequality;
4. A capacity interval [Lu , Uu ] for each facility u F, where Lu , Uu ZZ 0
and Lu Uu .
Our goal is to nd a client set C C of size at least p, an open facility set F F
of size exactly k, and a function : C F satisfying that Lu |1 (u)| Uu
for each u F , which minimize the maximum cluster radius maxvC d(v, (v)).
If the maximum cluster radius is at most r, we call the tuple (C, F, ) a distance-r
solution.

By the similar approach of Cygan et al. [21], we can reduce the ({Lu },{Uu },k,p)-
Supplier problem to a simpler case. We rst introduce some denitions.
Denition 2. (Induced distance function) We say the distance function dG :
(C F) (C F) IR0 is induced by an undirected unweighted connected
graph G = (C F, E) if
1. (u, v) E, we have u F and v C.
2. a1 , a2 C F, the distance dG (a1 , a2 ) between a1 and a2 equals to the
length of the shortest path from a1 to a2 .

Denition 3. (Induced ({Lu },{Uu },k,p)-Supplier instance)An({Lu },{Uu },k,p)-

Supplier instance is called an induced ({Lu },{Uu },k,p)-Supplier instance if
the following properties are satised:
1. The distance function dG is induced by an undirected connected graph G =
(C F, E).
2. The optimal capacitated k-supplier value is at most 1.
330 H. Ding et al.

Moreover, we say this instance is induced by G.

When the graph of interest G is clear from the context, we will use d in-
stead of dG for convenience. We then show a reduction from solving the gener-
alized ({Lu },{Uu },k,p)-Supplier problem to solving induced ({Lu },{Uu },k,p)-
Supplier instances by Lemma 1.
Lemma 1. Suppose we have a polynomial time algorithm A that takes as input
any induced ({Lu },{Uu },k,p)-Supplier instance, and outputs a distance- so-
lution. Then, there exists a -approximation algorithm for the ({Lu },{Uu },k,p)-
Supplier problem with polynomial running time.
2 Capacitated Center with Two-Sided Bounds and
Outliers
In this section, we consider the version that the number of open centers can be
arbitrary. By the LP approach in Section 3.3 and enumerating the number of
open centers, we can achieve approximation algorithms for dierent variants in
this case. However, the approximation factor is not small enough. In this section,
we introduce a new greedy approach in order to achieve better approximation
factors. Since our algorithm is combinatorial, it is easier to be implemented and
saves the running time compared to the LP approach.
2.1 Core-center tree (CCT)
Consider the (L,{Uu },,p)-Supplier problem. By Lemma 1, we only need to
consider induced (L,{Uu },,p)-Supplier instances induced by an undirected un-
weighted connected graph G = (C F, E). We rst propose a new data structure
called core-center tree ( CCT) as follows.
Denition 4.(Core-center tree( CCT)) Given an induced (L,{Uu },,p)-Supplier
instance induced by an undirected unweighted connected graph G = (C F, E),
we call a tree T = (F, ET ) a core-center tree( CCT) if the following properties
hold.
1. For each edge (u, u ) ET , we have dG (u, u ) 2;
2. Suppose the root of T is at layer 0. Denote I to be the set of vertices in the
even layers of T . We call I the core-center set of T . For any two distinct
vertices u, u I, we have dG (u, u ) 3.
Lemma 2. Given an induced (L,{Uu },,p)-Supplier instance induced by an
undirected unweighted connected graph G = (C F, E), we can construct a CCT
in polynomial time.
For any u F, denote NG [u] = {v C : (u, v) E} to be the collec-
tion of all neighbors of u F. 4 W.l.o.g., we assume that Uu |NG (u)| for
every facility u F in this section. In fact, we can directly delete all u F
satisfying that |NG [u]| < L from the facility set F, since u can not be open
in any optimal feasible solution. 5 Otherwise if L |NG [u]| < Uu , we set
4
If u C is also a client, then u NG [u].
5
If this deletion causes the induced graph unconnected, similar to Lemma 6 in [21],
we divide the graph into dierent connected components, and consider each smaller
induced instance based on dierent connected components.
Capacitated Center Problems with Two-Sided Bounds and Outliers 331

Uu min{Uu , |NG [u]|}, which has no inuence on any optimal feasible solu-
tion of the induced (L,{Uu },,p)-Supplier instance. The following lemma gives
a useful property of CCT.

Lemma 3. Given an induced (L,{Uu },,p)-Supplier instance induced by an

undirected unweighted connected graph G = (C F, E), and a core-center tree
T = (F, ET ), suppose I is the core-center set of T . Then, we can construct a
function : C F satisfying the following properties in polynomial time.
1. For all v C, we have ((v), v) E;
2. For all u I, we have | 1 (u)| L.

2.2 A Simple Case: (L,U ,soft-,p)-Supplier

We rst consider a simple case where the capacity bounds (upper and lower)
are uniform and soft. In this setting, we want to nd an open facility set F =
{ui | ui F}i . Note that we allow multiple open centers in F . We also need to
nd an assignment function : C F , representing that we assign every client
v C to facility (v). The main theorem is as follows.

Theorem 1. (main theorem) There exists a 5-approximation polynomial time

algorithm for the (L,U ,soft-,p)-Supplier problem.

By Lemma 1, we only consider induced (L,U ,soft-,p)-Supplier instances.

Given an induced (L,U ,soft-,p)-Supplier instance induced by an undirected
unweighted connected graph G = (C F, E), recall that we can assume |NG [u]|
Uu L for each u F. We rst construct a CCT T = (F, ET ) rooted at node
u , and a function : C F satisfying Lemma 3. For a facility set P F, we
denote 1 (P ) = uP 1 (u) to be the collection of clients assigned to some
facility in P by .
Our algorithm mainly includes two procedures. The rst procedure is called
pass-up, which is a greedy algorithm to map clients to facilities from the leaves
of T to the root. After the pass-up procedure, we still leave some unassigned
clients nearby the root. Then we use a procedure called pass-down to allocate
those unassigned clients by iteratively nding an exchange route. In the following,
we give the details of both procedures.
Procedure Pass-Up. Assume that |C| = aL + b for some a IN and 0 b
L 1. In this procedure, we will nd an open facility set F of size a. We also nd
an assignment function which assigns aL clients to some nearby facility in F
except a client set S C. Here, S is a collection of b clients in 1 (u ) nearby
the root u . Our main idea is to open facility centers from the leaves of CCT T
to the root iteratively. During opening centers, we assign exactly L close clients
to each center. Thus, there are b unassigned clients after the whole procedure.
We then describe an iteration of pass-up. Assume that I is the core-center set
of T . At the beginning, we nd a non-leaf vertex u I satisfying that all of its
grandchildren (if exists) are leaves. 6 We denote P F to be the collection of all
6
If multiple non-leaf nodes satisfy this property, we choose an arbitrary one.
332 H. Ding et al.

children and all grandchildren of u. In the next step, we consider all unscanned
clients in 1 (P ), 7 and assign them to the facility u by . Note that we may
open multiple centers at u. We want that each center at u serves exactly L
centers. However, there may exist one center at u serving less than L unscanned
clients in 1 (P ). We assign some clients in 1 (u) to this center such that it
also serves exactly L clients. After this iteration, we delete the subtree rooted
at u from T except u itself.
Finally, the root u will become the only remaining node in T . We open
multiple centers at u , each serving exactly L clients in 1 (u ), until there
are less than L unassigned clients. We denote S to be the collection of those
unassigned clients. At the end of pass-up, we output an open facility set F , an
unassigned client set S and an assigned function : C \ S F . We have the
following lemma by the algorithm.

Lemma 4. Given an induced (L,U ,soft-,p)-Supplier instance induced by an

undirected unweighted connected graph G = (C F, E), assume that |C| = aL + b
for some a IN and 0 b L 1. The output of pass-up satises the following
properties:
1. Each open facility uj F satises that uj I, and |F | = a;
2. The unassigned client set S 1 (u ), and |S| = b;
3. For each facility ui F , we have |1 (ui )| = L.
4. For each client v C \ S, (v) is either (v), or the parent of (v) in T , or
the grandparent of (v) in T . Moreover, we have dG (v, (v)) 5.

Procedure Pass-Down. After the procedure pass-up, we still leave an unas-

signed client set S of size b. However, our goal is to serve at least p clients.
Therefore, we need to modify the assignment function and serve more clients.
The procedure pass-down handles the remaining b clients in S one by one.
At the beginning of pass-down, we initialize an unscanned client set B C \ S,
i.e., B is the collection of those clients allowing to be reassigned by pass-down.
In each iteration, we arbitrarily pick a client v S and assign it to the root
node u . However, if each open facility at u has already served Uu clients by ,
assigning v to u will violate the capacity upper bound. In this case, we actually
nd an open center uj F such that |1 (uj )| < Uj , i.e., there are less than Uj
clients assigned to uj by . We then construct an exchange route consisting of
open facilities in F . We rst nd a sequence of nodes w0 = u , w1 , , wm = uj
in T satisfying that wi is the grandparent of wi+1 in the core-center tree T for
all 0 i m 1. Then for each node wi (1 i m 1), we pick a client
vi 1 (wi ) which has not been reassigned so far. We call such a sequence
of clients v, v1 , . . . , vm1 an exchange route. Our algorithm is as follows: 1) we
assign v to (v1 ); 2) we iteratively reassign vi to (vi+1 ) in order (1 i m2);
3) nally we reassign vm1 to uj . We then mark all clients vi (1 i m 1)
in the exchange route by removing them from the unscanned client set B, and
7
Here, unscanned clients are those clients that have not been assigned by before
this iteration.
Capacitated Center Problems with Two-Sided Bounds and Outliers 333

remove the client {v} from the unassigned client set S. Note that our exchange
route only increases the number of clients assigned to uj by one. In fact, such
an exchange route always exists in each iteration. Thus in each iteration, the
procedure pass-down assigns one more client v S to some open facility in F .
At the end of pass-down, we output a client set C C \ S of size at least p, an
open facility set F and an assigned function : C F .
Now we prove the following lemma. Note that Theorem 1 can be directly
obtained by Lemma 1 and Lemma 5.

Lemma 5. The procedure pass-down outputs a distance-5 solution (C, F, ) of

the given induced (L,U ,soft-,p)-Supplier instance induced by G = (C F, E)
in polynomial time.

2.3 (L,{Uu },,p)-Center

In this subsection, we consider a more complicated case where the capacity upper
bounds are non-uniform, and each vertex has a hard capacity.

Theorem 2. (main theorem) There exists an 11-approximation polynomial time

algorithm for the (L,{Uu },,p)-Center problem.

By Lemma 1, we only need to consider induced (L,{Uu },,p)-Supplier in-

stances. For an induced (L,{Uu },,p)-Supplier instance induced by an undi-
rected unweighted connected graph G = (V = C F , E), recall that we can
assume Uu |NG (u)| for every vertex u F. 8 Since we consider the center
version, every vertex v C has an individual capacity interval [L, Uv ] and can
be opened as a center as well.
Similar to (L,U ,soft-,p)-Center, our algorithm rst computes a core-center
tree T = (F, E) rooted at u , a core-center set I and a function described as
in Lemma 3. Assume that |C| = aL + b for some a IN and 0 b L 1. We
still use the procedure pass-up to compute an open set F = {u1 , u2 , , ua }, an
unassigned set S 1 (u ) of size b < L, and a function : (C \ S) F .
However, we can not apply pass-down directly since we consider non-uniform
hard capacity upper bounds. Thus, we need the following lemma to modify the
open center set F . We prove this lemma by Halls theorem in the full version.

Lemma 6. Given an induced (L,{Uu },,p)-Center instance induced by G =

(V = C F, E) where |NG (u)| Uu for each u F and an open set F =
{u1 , u2 , , ua } computed by pass-up, there exists a polynomial time algorithm
that nds another open set F  = {u1 , u2 , . . . , ua } such that:

1. F  is a single set.

a all 1 i a, we have dG (ui , ui ) 6.
2. For
3. i=1 Uui p.


8
Recall that we may remove some facilities from F such that this assumption is
satised. Thus, the set F may be a subset of V .
334 H. Ding et al.

Proof of Theorem 2. By Lemma 6, we obtain another open set F  = {u1 , u2 , . . . , ua }.
We rst modify Uui to be Uui for all 1 i a. Then we apply the proce-
dure pass-down according to the modied a capacities. By Lemma 5, we obtain a
distance-5 solution (C, F, ). Since i=1 Uui p, at least p vertices are served
by . Finally, for each vertex v C and ui F such that (v) = ui , we reassign
v to ui F  , i.e., let (v) = ui . By Lemma 6, we obtain a feasible solution for the
given induced (L,{Uu },,p)-Center instance. Since d(ui , ui ) 6 (1 i a),
the capacitated center value of our solution is at most 5 + 6 = 11. Combining
with Lemma 1, we nish the proof.
3 Capacitated k-Center with Two-Sided Bounds and
Outliers
Now we study the capacitated k-center problems with two-sided bounds. We
consider that all vertices have a uniform capacity lower bound Lv = L, while
the capacity upper bounds can be either uniform or non-uniform. Similar to [3,
21], we use the LP relaxation and the rounding procedure distance-r transfer.
3.1 LP Formulation
We rst give a natural LP relaxation for ({Lu },{Uu },k,p)-Supplier.
Denition 5. (LPr (G)) Given an ({Lu },{Uu },k,p)-Supplier instance, the fol-
lowing feasibility LPr (G) that fractionally veries whether there exists a solution
that assigns at least p clients to an open center of distance at most r:

0 xuv , yu 1, u F, v C;
xuv = 0, if d(u, v) > r;
x
uv y u , u F, v C;
uF u y = k;
x p;
uF ,vC uv
uF xuv 1, v C;
Lu yu vC xuv Uu yu , u F.

Here we call xuv an assignment variable representing the fractional amount of

assignment from client v to center u, and yu the opening variable of u F. For
convenience, we use x, y to represent {xuv }uF ,vC and {yu }uF , respectively.
By Denition 3, LP1 (G) must have a feasible solution for any induced ({Lu },{Uu },k,p)-
Supplier instance. We recall a rounding procedure called distance-r transfer.

3.2 Distance-r Transfer

We rst extend the denition of distance-r transfer proposed in [3, 21] by adding
the third condition. For a vertex a C F and a set B C F, we dene
d(a, B) = minbB d(a, b).

Denition 6. Given an ({Lu },{Uu },k,p)-Supplier instance and y IRF

0 , a
 F
vector y IR0 is a distance-r transfer of y if
 

1. uF yu = uF yu ;
Capacitated Center Problems with Two-Sided Bounds and Outliers 335
 
2. U y U y for all W F;
wF :d(w,W )r w w uW u u
3. wF :d(w,W )r Lw yw uW Lu yu for all W F.

If y  is a characteristic vector of F F, we say that F is an integral distance-r

transfer of y.

In this paper, we add the third condition to satisfy the capacity lower bounds.
Like in [3, 21], we still have the following lemma.

Lemma 7. Given an ({Lu },{Uu },k,p)-Supplier problem, assume (x, y) is a

feasible solution of LP1 (G) and F F is an integral distance-r transfer of y.
Then one can nd a distance-(r + 1) solution (C, F, ) in polynomial time.

3.3 Capacitated k-Center with Two-Sided Bounds and Outliers

Now we are ready to solve the (L,{Uu },k,p)-Supplier problem. By Lemma 7,
we only need to nd an integral distance-r transfer satisfying Denition 6 given
a feasible fractional solution (x, y) of LP1 (G). Fortunately, the rounding schemes
in [3, 21] have this property. Thus, we have the following theorem by [3].

Theorem 3. There is a polynomial time 9-approximation algorithm for the

(L,{Uu },k)-Center problem. For the uniform capacity upper bound version,
the (L,U ,k)-Center problem admits a 6-approximation.

Theorem 4. There is a polynomial time 13-approximation algorithm for the

(L,{Uu },k)-Supplier problem. For the uniform capacity upper bound version,
the (L,U ,k)-Supplier problem admits a 9-approximation.
By [21], we have the following theorem.

Theorem 5. There is a polynomial time 25-approximation algorithm for the

(L,{Uu },k,p)-Supplier problem and the (L,{Uu },soft-k,p)-Supplier problem.
For the uniform capacity upper bound version, the (L,U ,k,p)-Supplier problem
admits a 23-approximation, and the (L,U ,soft-k,p)-Supplier problem admits a
13-approximation.

References
1. Aggarwal, G., Panigrahy, R., Feder, T., Thomas, D., Kenthapadi, K., Khuller, S.,
Zhu, A.: Achieving anonymity via clustering. ACM Trans. Algorithms 6(3), 49:1
49:19 (2010)
2. Ahmadian, S., Swamy, C.: Approximation algorithms for clustering problems with
lower bounds and outliers. arXiv preprint arXiv:1608.01700 (2016)
3. An, H.C., Bhaskara, A., Chekuri, C., Gupta, S., Madan, V., Svensson, O.: Central-
ity of trees for capacitated k-center. Mathematical Programming 154(1-2), 2953
(2015)
4. Arya, V., Garg, N., Khandekar, R., Meyerson, A., Munagala, K., Pandit, V.: Local
search heuristics for k-median and facility location problems. SIAM Journal on
computing 33(3), 544562 (2004)
336 H. Ding et al.

5. Barilan, J., Kortsarz, G., Peleg, D.: How to allocate network centers. Journal of
Algorithms 15(3), 385415 (1993)
6. Charikar, M., Guha, S.: Improved combinatorial algorithms for facility location
problems. SIAM Journal on Computing 34(4), 803824 (2005)
7. Charikar, M., Guha, S., Tardos, E., Shmoys, D.B.: A constant-factor approximation
algorithm for the k-median problem. In: Proceedings of the thirty-rst annual ACM
symposium on Theory of computing. pp. 110. ACM (1999)
8. Charikar, M., Khuller, S., Mount, D.M., Narasimhan, G.: Algorithms for facility
location problems with outliers. In: Proceedings of the twelfth annual ACM-SIAM
symposium on Discrete algorithms. pp. 642651. Society for Industrial and Applied
Mathematics (2001)
9. Chen, D.Z., Li, J., Liang, H., Wang, H.: Matroid and knapsack center problems.
Algorithmica 75(1), 2752 (2016)
10. Cygan, M., Hajiaghayi, M., Khuller, S.: Lp rounding for k-centers with non-uniform
hard capacities. In: Proceedings of the 2012 IEEE 53rd Annual Symposium on
Foundations of Computer Science. pp. 273282. IEEE Computer Society (2012)
11. Dick, T., Li, M., Pillutla, V.K., White, C., Balcan, M.F., Smola, A.: Data driven
resource allocation for distributed learning. arXiv preprint arXiv:1512.04848 (2015)
12. Ding, H.: Balanced k-center clustering when k is a constant. arXiv preprint arX-
iv:1704.02515 (2017)
13. Ene, A., Har-Peled, S., Raichel, B.: Fast clustering with lower bounds: No customer
too far, no shop too small. arXiv preprint arXiv:1304.7318 (2013)
14. Fernandes, C.G., de Paula, S.P., Pedrosa, L.L.: Improved approximation algorithms
for capacitated fault-tolerant k-center. In: Latin American Symposium on Theo-
retical Informatics. pp. 441453. Springer (2016)
15. Gonzalez, T.F.: Clustering to minimize the maximum intercluster distance. Theo-
retical Computer Science 38, 293306 (1985)
16. Guha, S., Khuller, S.: Greedy strikes back: Improved facility location algorithms.
Journal of algorithms 31(1), 228248 (1999)
17. Hochbaum, D.S., Shmoys, D.B.: A best possible heuristic for the k-center problem.
Mathematics of operations research 10(2), 180184 (1985)
18. Jain, K., Mahdian, M., Saberi, A.: A new greedy approach for facility location
problems. In: Proceedings of the thiry-fourth annual ACM symposium on Theory
of computing. pp. 731740. ACM (2002)
19. Jain, K., Vazirani, V.V.: Approximation algorithms for metric facility location
and k-median problems using the primal-dual schema and lagrangian relaxation.
Journal of the ACM (JACM) 48(2), 274296 (2001)
20. Khuller, S., Sussmann, Y.J.: The capacitated k-center problem. SIAM Journal on
Discrete Mathematics 13(3), 403418 (2000)
21. Kociumaka, T., Cygan, M.: Constant factor approximation for capacitated k-center
with outliers. arXiv preprint arXiv:1401.2874 (2014)
22. Korupolu, M.R., Plaxton, C.G., Rajaraman, R.: Analysis of a local search heuristic
for facility location problems. Journal of algorithms 37(1), 146188 (2000)
23. Li, J., Yi, K., Zhang, Q.: Clustering with diversity. In: International Colloquium
on Automata, Languages, and Programming. pp. 188200. Springer (2010)
24. Li, S.: A 1.488 approximation algorithm for the uncapacitated facility location
problem. Information and Computation 222, 4558 (2013)
25. Li, S., Svensson, O.: Approximating k-median via pseudo-approximation. SIAM
Journal on Computing 45(2), 530547 (2016)
26. Sweeney, L.: k-anonymity: A model for protecting privacy. International Journal
of Uncertainty, Fuzziness and Knowledge-Based Systems 10(05), 557570 (2002)
Faster Randomized Worst-Case Update Time for
Dynamic Subgraph Connectivity

Ran Duan1,2 and Le Zhang1,3

1
Institute for Interdisciplinary Information Sciences, Tsinghua University, China
2
[email protected]
3
[email protected]

Abstract. Real-world networks are prone to breakdowns. Typically in

the underlying graph G, besides the insertion or deletion of edges, the
set of active vertices changes overtime. A vertex might work actively,
or it might fail, and gets isolated temporarily. The active vertices are
grouped as a set S. The set S is subjected to updates, i.e., a failed ver-
tex restarts, or an active vertex fails, and gets deleted from S. Dynamic
subgraph connectivity answers the queries on connectivity between any
two active vertices in the subgraph of G induced by S. The problem is
solved by a dynamic data structure, which supports the updates and
answers the connectivity queries. In the general undirected graph, we
propose a randomized data structure, which has O(m  3/4 ) worst-case up-
date time. The former best results for it include O(m 2/3 ) deterministic
amortized update time by Chan, Patrascu and Roditty [4], O(m  4/5 ) by

Duan [8] and O( mn) by Baswana, Chaudhury, Choudhary and Khan
[2] deterministic worst-case update time.

1 Introduction
Dynamic subgraph connectivity is dened as follows: Given an undirected graph
G = (V, E) having m edges, n vertices with m = (n), there is a subset
S V . The set E is subjected to edge updates of the forms insert(e, E)
or delete(e, E), where e is an edge. There are vertex updates of the forms
insert(v, S) or remove(v, S). Through vertex updates, S changes overtime. The
query is on whether any two vertices s and t are connected in the subgraph of
G induced by S.
The problem was rst proposed by Frigioni and Italiano [12], and poly-
logarithmic algorithms on connectivity were described for the special case of
planar graphs. As to the general graphs, Chan [3] rst described an algorithm
- 4/(3+3) )4 , where is the matrix
of deterministic amortized update time O(m
multiplication exponent. Adopting FMM (Fast Matrix Multiplication) algorithm
of [6], the update time is O(m0.94 ). Its query time and space complexity are
- 1/3 ) and linear, respectively. Later Chan, Patrascu, and Roditty [4] pro-
O(m
posed a simpler algorithm with the improved update time of O(m - 2/3 ). The
4  hides poly-logarithmic factors.
O()

Springer International Publishing AG 2017 337

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 337348, 2017.
DOI: 10.1007/978-3-319-62127-2_29
338 R. Duan and L. Zhang

space complexity of the new algorithm increases to O(m - 4/3 ). The new algorithm
is of compact description, getting rid of the use of FMM. With the same update
time, Duan [8] presented new data structures occupying linear space. Also a
worst-case deterministic O(m- 4/5 ) algorithm was proposed by Duan [8]. Via an
application of dynamic DFS tree [2], Baswana et al. discussed a new algorithm
with O(- mn) deterministic worst-case update time. Its query time is O(1). An
improvement of it is discussed in [5]. These results are summarized in Table 1.
A close related problem is dynamic graph connectivity, which cares only
about the edge updates. Poly-logarithmic amortized update time was rst achieved
by Henzinger and King [15]. The algorithm is randomized Las Vegas. Inspired
by it, Holm et al. [16] proposed a deterministic algorithm with O(lg2 n)5 amor-
tized update time, which is now one of the classic results in the eld. A cell-
probe lower bound of (lg n) was proved by Patrascu and Demaine [21]. The
lower bound is amortized randomized. Near-optimal results were considered
by Thorup [22], where a randomized Las Vegas algorithm was described with
O(lg n(lg lg n)3 ) amortized update time. The upper bound is recently improved
to O(lg n(lg lg n)2 ) by Huang et al. [17]. Besides the classic deterministic O(lg2 n)
result, a faster deterministic algorithm was proposed by Wul-Nilsen [24], of
which the update time is O(lg2 n/ lg lg n). Turning to the worst-case dynamic
connectivity,
a deterministic O( n) update-time algorithm is Fredericksons
O( m) worst-case algorithm [11] sped up via sparsication technique proposed
by Eppstein et al. [10]. The result holds for online updating of minimum span-
ning trees. With roughly the same structure, but dierent and simpler tech-
niques, Kejlberg-Rasmussen $ et al. [19] provided the so far best deterministic
worst-case bound of O( n(lg lg n)2 / lg n) for dynamic connectivity. After the
discovery of O( n) update-time algorithm, people were wondering whether any
poly-logarithmic worst-case update time algorithm is possible, even random-
ized. The open problem stands rmly for many years. A breakthrough should
be attributed to Kapron et al. [18]. Their algorithm is Monte-Carlo, with poly-
logarithmic worst-case update time. It has several improvements until now, as
done in [13, 23]. For subgraph connectivity, the trivial update time of O(n) - fol-
lows from Kapron et al.s algorithm. The query time of it for subgraph connectiv-
ity can also be improved to O(1), as the explicit maintenance of connected com-
ponents can be done without blowing up the O(n) - update time. Very recently,
Wul-Nilsen [25] gave a Las Vegas data structure maintaining a minimum span-
ning forest in expected worst-case time polynomially faster than (n1/2 ) w.h.p.
per edge update. An independent work of Nanongkai and Saranurak [20] showed
an algorithm with O(n0.49306 ) worst-case update time w.h.p..

1.1 Our Results

- 4/5 ) deterministic worst-case subgraph connectivity structure
The former O(m
adopted as a sub-routine the O( n) deterministic worst-case algorithm for dy-
namic graph connectivity. Now the randomized poly-logarithmic worst-case con-
5
We use lg x to denote log2 x.
Faster Randomized Worst-Case Update Time for Dynamic Subgraph Connectivity 339

Table 1. Results on Dynamic Subgraph Connectivity

Update time Query time Notes

 1/3 )
 4/(3+3) ) O(m Amortized,
O(m
deterministic, linear space [3]
 mn)
O( O(1)
Worst case,

deterministic, space O(m) [2, 5]
 2/3 )  1/3 ) Amortized,
O(m O(m  4/3 ) [4]
deterministic, space O(m
 2/3 )  1/3 ) Amortized,
O(m O(m
deterministic, linear space [8]
 4/5 )  1/5 ) Worst case,
O(m O(m 
deterministic, space O(m) [8]
 3/4 )  1/4 ) Worst case,
O(m O(m
randomized, linear space, this paper

nectivity structures for dynamic graph connectivity are discovered. We consider

the question of whether it brings progress in subgraph connectivity. The an-
swer is armative. But it does not come by simple replacement. More precisely,
we tried in vain to get an improvement by carefully tuning the former setting
- 4/5 ) algorithm. Intuitively, the amortized O(m
of the O(m - 2/3 ) update time was
achieved partially because it uses the connectivity structure of poly-logarithmic
amortized update time. Now poly-logarithmic worst-case algorithms are discov-
- 2/3 ) worst-case update time is in sight. Nonetheless,
ered, it seems that the O(m
- 2/3 ) update time. Until now we ob-
we found that it is still hard to get the O(m
tain the update time of O(m- 3/4 ). The main contribution is a new organization
of the auxiliary data structures.

The O(- mn) result comes from dynamic DFS tree [2, 5], which is a periodic
rebuilding technique with fault tolerant DFS trees. Our result is always no worse
than O(- mn) as n = (m1/2 ). Faster query time can be traded with slower
- 3/4+ )
update time for the bottom four results in Table 1. As to our result, O(m
- 1/4 ) query time can be implemented. Note that the trade-
update time and O(m
os are in one direction, i.e. better query time with worse update time, but not
- 4/5 ) algorithm never gives update time
vice-versa. Consequently, the former O(m
- 3/4
of O(m ). The trade-o phenomenon is denitely hard to break, as indicated by
the OMv (Online Boolean Matrix-Vector Multiplication) conjecture proposed by
Henzinger et al. [14], which rules out polynomial pre-processing time algorithms
with the product of amortized update and query time being o(m). Based on the
conjecture of no truly subcubic combinatorial boolean matrix multiplication,
Abboud and Williams [1] showed that any combinatorial dynamic algorithm
with truly sublinear in m query time and truly subcubic in n preprocessing time
must have (m1/2 ) update time for all > 0 unless the conjecture is false.
Our result is grouped as the following theorem.
340 R. Duan and L. Zhang

Theorem 1.1 (Main Theorem) Given a graph G = (V, E), there is a data
structure for the dynamic subgraph connectivity, which has the worst-case vertex
- 3/4 ), query time O(m
(edge) update time O(m - 1/4 ), where m is the number of
edges in G, rather than in the subgraph of G induced by S. The answer to each
query is correct if the answer is yes, and is correct w.h.p. if the answer is
- 5/4 ), and the space usage is linear.
no. The pre-processing time is O(m

2 Preliminaries
Theorem 2.1 ([19]) A spanning forest F of G $ can be maintained by a de-
terministic data structure of linear space, with O( m(lg lg n)2 /lg n) worst-case
update time for an edge update in G, and constant query time to determine
whether two vertices are connected in G.
Theorem 2.2 ([18, 23]) There is a randomized data structure on dynamic
graph connectivity, which supports the worst-cast time O(lg4 n) per edge inser-
tion, O(lg5 n) per edge deletion, and O(lg n/ lg lg n) per query. For any constant
c the answer to each query is correct if the answer is yes and is correct with
probability 1 1/nc if the answer is no. The pre-processing time of it is
O(m lg3 n + n lg4 n).
Moving to subgraph connectivity, here we consider only the case of vertex
updates, with the extension to edge updates deferred to the full paper [9]. Hence
temporarily G is assumed to be static, as E does not change if there are no
edge updates. The vertex updates change S. Initially, G is slightly modied to
keep m = (n) during its lifetime, i.e., for every v V , insert a new vertex
v  and a new edge (v, v  ). The variant graph has m = (n), which facilitates
the presentation of time and space complexity as functions of m in the case of
degenerate graphs.

3 The Data Structure

We give some high-level ideas, which originate from [4]. Main diculties are the
update of S (recall that S is the set of active vertices) incurred by the high-
degree vertices, as their degrees are too high to explicitly delete their incident
edges one by one. Nonetheless, if the low-degree vertices had been removed,
the graph became smaller, and consequently former high-degree vertices were
not high-degree anymore. Hence our aim is to remove the low-degree vertices.
After that, some articial edges are added to restore the loss of connectivity
due to the removal of the low-degree vertices. Next a dynamic connectivity data
structure is maintained on the modied graph, i.e., the graph with the low-
degree vertices removed, and the articial edges added. Besides, as S evolves
dynamically, we need to update the articial edges accordingly. Hence the point
is how to maintain these articial edges consistently and eciently. We now move
to the details. We partition V according to their degrees in G. Use degG (v) to
denote the degree of v in G.
Faster Randomized Worst-Case Update Time for Dynamic Subgraph Connectivity 341

C: Vertices with degG (v) > m1/2

B: Vertices with m1/4 < degG (v) m1/2
A: Vertices with degG (v) m1/4

Denote C S, B S, and A S as VC , VB , and VA respectively. Consider

the subgraph GA of G induced by VA . Dene the degree of a component as
the sum of degG (v)s for vs in it. According to the degrees of the components,
partition the components of GA into two types: high component, with its degree
> m1/4 ; low component, with its degree m1/4 . A spanning forest FA of GA is
maintained by the deterministic connectivity structure of Theorem 2.1.

3.1 Path Graph

A path graph inserts some articial edges to reect the are connected relation
of the vertices within VB via directly linking with a component of GA . We give
a more elaborate analysis based on [8]. W.l.o.g. assume V = {0, . . . , n 1}.
Consider a spanning tree T of FA .

subpath tree: For v T , identify the set of vertices in VB that are adjacent to
v. Store the set of vertices in a balanced search tree, which has the worst-case
O(lg n) update time for the well-known search-tree operations [7]. Name the
search tree as the subpath tree of v. Given the subpath tree of v, a sequence
of articial edges is added to link the vertices stored in the subpath tree of
v. The sequence of articial edges constitutes a subpath.
path tree: Given T FA , group all v T with the non-empty subpath tree
as a balanced search tree, ordered by the Euler-tour order of T . Name it
as the path tree of T . As each vertex stored in the path tree of T has an
associated subpath, these subgraphs are also concatenated one by one via
the articial edges, generating a path. To emphasize its dierence from an
ordinary path, it is referred to as the path graph of VB w.r.t. T . An example
is shown in Fig. 1.

Lemma 3.1 The path graphs can be updated in O(m - 1/2 ) time for a vertex update
-
in VB , and in O(1) time for a link or cut on FA .

Proof. We categorize the analysis into two cases.

Reect a vertex update in VB : Suppose v VB is removed from S. The case
of insertion is similar. v has m1/2 edges adjacent to FA . Consider (v, w)
with w T . We locate w in the path tree of T . Now the subpath associated
with w is known. Update the subpath of w by removing v from the subpath.
If v happens to be the rst or the last vertex on the subpath, the path graph
of T is also updated. As the subpaths and the path graph are concerned
with the nodes stored in the subpath trees and the path tree respectively,
-
which are all balanced search trees, the removal of (v, w) needs O(1) time.
- 1/2
The removal of all such (v, w)s requires O(m ) time.
342 R. Duan and L. Zhang

15
4 11
Part of VB 3 10
2 7 14
6
ET-order
T in GA
1, 5, 8, 13, 12, 9
9
12
1 5 1, 12, 9
8
13 2, 15 3, 15
4, 10, 14

Fig. 1. The path graph of VB w.r.t. a spanning tree T in FA . The dashed edges represent
edges between VB and VA . The path tree is on sequence 1, 12, 9, and three subpath
trees are on sequences 2, 15, 4, 10, 14, and 3, 15 respectively. The resulted path
graph is a path 2, 15, 4, 10, 14, 3, 15.

Reect a link or cut on FA : We only discuss the edge cut on FA . The edge
link is similar. Assume the edge cut is (v, w) T , and the Euler tour of T
is L1 , (v, w), L2 , (w, v), L3  (The details can be found in the full paper [9].).
After the cut of (v, w), the Euler tours for the two resulted trees are L1 , L3 
and L2 . We can determine the rst vertex a and the last vertex b of L2 .
With the order tree of T (discussed in the full paper [9]), the predecessor of
a and the successor of b in the path tree of T can be found in O(lg2 n) time.
With the predecessor and the successor, the path tree of T is split. After the
split, O(1) edges in the path graph are removed to reect the split of the
path tree of T . As a conclusion, the path graph can be updated in O(1) - time
to reect a link or cut on FA .



3.2 Adjacency Structure

Given T FA and v C, we want a data structure that provides the fast query
of whether T and v are adjacent, i.e., whether an edge (u, v) exists with u T .
We give a more elaborate analysis based on [8]. Assuming v C, the adjacency
structure of v contains the following search trees.

sub-adjacency tree: Given T FA , identify the set of vertices in T that are

adjacent to v. Store the set of vertices as a balanced search tree, ordered
by the Euler-tour order of T . Name the balanced search tree as the sub-
adjacency tree of v w.r.t. T .
adjacency tree: Identify T FA by the smallest vertex in T . Group all
T FA , w.r.t. which v has non-empty sub-adjacency trees, as a balanced
search tree. Name the balanced search tree as the adjacency tree of v.

The sub-adjacency trees and the adjacency tree of v constitute the adjacency
structure of v w.r.t. FA . The query aforementioned is answered by checking
Faster Randomized Worst-Case Update Time for Dynamic Subgraph Connectivity 343

whether T is in the adjacency tree of v. Note v C, rather than VC . The

adjacency structure of v C w.r.t. FA is maintained even if v
/ S.

- 1/2 )
Lemma 3.2 The adjacency structures of C w.r.t. FA can be renewed in O(m
time for a link or cut on FA . Given a query of whether v C is adjacent to
-
T FA , it can be answered in O(1) time.

Proof. We only discuss the edge cut on FA . The edge link is similar. The adja-
cency structures of the vertices in C are renewed one by one. Consider v C.
Suppose the edge cut occurs on T , splitting T into T1 and T2 . We check whether
T is in the adjacency tree of v. If no, the update is done; if yes, remove T
from it, and update the sub-adjacent tree of v w.r.t. T to reect the edge cut on
T . For Tj (j = 1, 2), add Tj into the adjacent tree of v if it is adjacent to v (de-
termined by whether a sub-adjacent tree of v exists w.r.t. Tj ). For every vertex
in C, we need to check and update when necessary. Hence the total update time
- 1/2 ), since |C| is O(m1/2 ). The query is answered by checking whether T
is O(m
is in the adjacency tree of v. 


3.3 The Whole Structure

Now we turn to the discussion of the whole structure of our result. First, VA is
removed. After that some articial vertices and edges are added to the subgraph
of G induced by VB C, resulting in a graph H. (Note that we include the
vertices in C \ S, rather than just VC , which is C S.) The articial vertices
and edges are used to restore the loss of connectivity due to the removal of VA .
Recall that the components of GA are either low or high. We describe how the
articial edges or vertices are added as follows.

Added by the path graphs: For T FA , construct the path graph of VB

w.r.t. T .
Added by the high components: For a high component P GA , add a meta-
vertex. For v C adjacent to P , add an articial edge between v and the
meta-vertex. Identify the rst vertex of the path graph of VB w.r.t. T , where
T is the spanning tree of P . Add an articial edge between the rst vertex
and the meta-vertex.
Added by the low components: For a low component Q GA , construct a
complete graph within the vertices in C that are adjacent to Q. Similarly as
above, identify the rst vertex of the path graph of VB w.r.t. T , where T is
the spanning tree of Q. Add the articial edges between the rst vertex and
the vertices in C that are adjacent to Q.

After these, H can be dened as follows.

The vertex set V (H) of H: VB C M , where M is the set of meta-vertices.

Since the degree of a high component is > m1/4 , and the vertices in VB C
are of degree > m1/4 , H has O(m3/4 ) vertices.
344 R. Duan and L. Zhang

VB

VA

Fig. 2. An example of the whole structure. The irrelevant edges within VA , VB , and
C are omitted for clarity. The solid edges are the edges in G, while the dotted edges
denote the articial edges. The grey vertex in the VB layer indicates a meta-vertex. The
left component of VA is low; whereas the right one is high. We construct a complete
graph within the vertices in C w.r.t. the low component.

The edge set E(H) of H: The original edges of G within VB C, and the
articial edges.
Figure 2 gives an example for the construction. H is a multigraph. Use
D[u, v] > 0 of edge multiplicity to represent the edge (u, v) E(H). The main-
tenance of D[u, v]s is discussed later. Now we construct a graph G , based on
H.
The vertex set V (G ) of G : VB VC M .
The edge set E(G ) of G : The edges (u, v)s with D[u, v] > 0, where u, v
V (G ), u = v.
G is a variant of the subgraph of H induced by VB VC M . It excludes
the vertices in C \ S, i.e., only the vertices in VC of C are contained. Besides,
the multiple edges are substituted by the single ones. G is a simple graph. The
randomized connectivity structure of Theorem 2.2 is maintained on G .
About the D[u, v]s aforementioned, a balanced search tree is used to store
them, with D[u, v] indexed by u + nv (assuming u v). Only D[u, v] > 0 is
stored in the search tree. Along the process of the updates, we might increment
or decrement D[u, v]s. When D[u, v] decrements to 0, we remove it from the
search tree. If both u and v are the vertices in G and u = v, the edge (u, v)
is deleted from G . Similar updates works for incrementing. G captures the
property of connectivity, which is stated in the following lemma.
Lemma 3.3 For any two vertices u, v VB VC , they are connected in the
subgraph of G induced by S if and only if they are connected in G .
Proof. G is a variant of the subgraph of G induced by S. G removes VA from the
subgraph. Connectivity within VB via VA is restored by the path graphs. Connec-
tivity within VC via VA is restored either by linking with the same meta-vertex,
Faster Randomized Worst-Case Update Time for Dynamic Subgraph Connectivity 345

or by the complete graph constructed. Lastly, for the connectivity between VC

and VB via VA , it is restored by the rst vertex of the path graph linking with
the meta-vertex, or with all the relevant vertices in VC . Consider a path between
u and v in the subgraph induced by S, the segments of the path consisting only
of the vertices in VA can be eliminated, as the via VA connectivity is restored
as discussed. The lemma follows. 


3.4 Update and Query

The diculty of the vertex updates is to keep D[u, v]s being consistent with S.
As E(G ) is a subset of the (u, v)s with D[u, v] > 0, it might also need to be
updated.

Lemma 3.4 The whole structure constructed has the worst-case vertex update
- 3/4 ).
time O(m

Proof. We discuss the various cases of vertex updates, categorized according to

whether v A, or B, or C.

v A: Consider the case of inserting v into S. v is rst inserted as a singleton

component containing only v in GA . Next the edges incident on v are restored
in the following order: First, the edges between v and C; second, the edges
between v and VB ; third, the edges between v and VA .
Restore the edges between v and C: For every u adjacent to v where u C,
construct a sub-adjacency tree (containing only v) of u, and insert v into
the adjacency tree of u. Next the complete graph within these us in C is
constructed. Because degG (v) m1/4 , i.e. a low component, the update time
- 1/2 ), dominated by constructing the complete graph.
is O(m
Restore the edges between v and VB : Construct the subpath tree and the
path tree of v. Add the path-graph edges associated with v (Add means
incrementing the corresponding entry D[u, v]), and the edges between the
rst vertex of the path graph and the vertices in C that are adjacent to v.
- 1/4 ).
The update time is O(m
Restore the edges between v and VA : O( - m) deterministic data structure
maintaining FA is updated in O(m- 3/4 ) time. As degG (v) m1/4 , the link or
1/4
cut on FA happens O(m ) times. Consequently, according to Lemma 3.1,
the path graphs are updated in O(m- 1/4 ) time. According to Lemma 3.2, the
adjacency structures are updated in O(m- 3/4 ) time.
1/4
O(m ) components of GA are aected. For every high component, using the
adjacency structures, the edges between C and the meta-vertex (correspond-
ing to the high component) can be determined in O(m - 1/2 ) time according to
Lemma 3.2, since |C| = O(m1/2 ); for every low component, as the degree of
- 1/2 ) time suces to construct the complete
a low component is m1/4 , O(m
graph within the vertices in C that are adjacent to the low component, and
- 1/4 ) time suces to construct the edges between the rst vertex of the
O(m
346 R. Duan and L. Zhang

path graph w.r.t. the low component and the vertices in C that are adja-
cent to the low component. Hence no matter whether the component is low
or high, the update time is O(m- 1/2 ). The time needed to update all these
- 3/4
components is O(m ). Deleting of v S from S is a reverse process. In
summary, a vertex update of v A requires O(m - 3/4 ) time.
v B: Consider the case when v S is removed. The case of insertion is the
reverse. First destroy the edges between v and VA . According to Lemma 3.1,
the path graphs can be updated in O(m - 1/2 ) time. Besides, v might be the
rst vertex of some path graphs. We see how it is updated. v can be adjacent
to m1/2 components of GA , as degG (v) m1/2 . For a high component, as
only one edge linking v with the meta-vertex, the update is easy; for a low
component, since only m1/4 edges can be outward for a low component,
- 1/4 ) time suces for updating the edges between v and the vertices in C
O(m
that are adjacent to the low component. Hence the update time for v being
- 3/4 ). Until now the articial edges
the rst vertex of some path graphs is O(m
concerning v are removed. Other edges concerning v are the original edges
in G. Hence we can remove these original edges one by one in O(m - 1/2 ) time
1/2
as degG (v) m . In summary, the total update time of v B is O(m - 3/4 ).
- 3/4 ).
v C: As there are only O(m3/4 ) vertices in G , the update time is O(m
The relevant D[u, v]s are left intact, and the adjacency structure of v is not
destroyed (if v is removed from S). The total update time is O(m- 3/4 ).



The query algorithm is as follows: Given u, v S, the goal is to substitute

them with the equivalent vertices in G , where an equivalent vertex of u (or v)
is a vertex in G that is connected with u (or v). As V (G ) = VB VC M , if
u, v VB VC , the search for the equivalent vertices is done. Otherwise, if u (or
v) is in a high component, replace u (or v) with the meta-vertex corresponding
to the high component; if u (or v) is in a low component, exhaustively search
the outward edges of the low component for a vertex of G . When the equivalent
vertex of u (or v) cannot be found, it indicates that u (or v) is in a low component
of GA , and the low component is not connected with any vertex in VB VC .
Intuitively u (or v) is on an island of GA .

Lemma 3.5 The time complexity of the query algorithm is O(m - 1/4 ). The an-
swer to every query is correct if the answer is yes, and is correct w.h.p. if the
answer is no.

Proof. Connectivity within G is answered by the randomized connectivity struc-

ture on G ; whereas for the other cases, u and v are connected if and only if they
are in the same component of GA , of which the queries can be answered by the
deterministic connectivity structure on GA . The time complexity is dominated
- 1/4 ).
by the exhaustive search if u (or v) is in a low component, and thus is O(m
The correctness can be analyzed as follows. If u, v VB VC , it follows from
Lemma 3.3; otherwise, for any one not in, we only replace it with an equivalent
vertex of G . If such an equivalent vertex cannot be found, the queried vertex is
Faster Randomized Worst-Case Update Time for Dynamic Subgraph Connectivity 347

on an island aforementioned of GA . Then u and v are connected if and only if

they are on the same island. We analyze the error probability. A deterministic
connectivity structure is adopted for GA . FA is always a spanning forest of GA .
The queries are answered either by the deterministic connectivity structure if
at least one queried vertex is on an island aforementioned of GA , or by the
randomized connectivity structure if both queried vertices are (replaced with)
the vertices in G . The deterministic connectivity structure always gives the right
answer; whereas the randomized one might answer erroneously. The randomized
algorithm of [18] maintains a private witness of a spanning forest of G . The
algorithm has the property that after every update, the witness is a spanning
forest of G with probability 1 1/nc . It is the property which ensures the
answers are correct w.h.p.. Here, after every vertex update (which is transformed
into a sequence of edge updates in G ), the witness for G is also a spanning
forest of G w.h.p. after the vertex update. We can just focus on the correctness
of the witness at the point after the last transformed edge update. Consequently,
the error probability is negligible, i.e., 1/nc for any constant c. 


The proofs of the pre-processing time being O(m - 5/4 ), and the space usage
being linear can be found in the full paper [9]. Hence Theorem 1.1 follows.

Acknowledgments. This work was supported in part by the National Basic

Research Program of China Grant 2011CBA00300, 2011CBA00301, the National
Natural Science Foundation of China Grant 61033001, 61361136003. R. Duan is
supported by a China Youth 1000-Talent grant.

References
[1] Amir Abboud and Virginia Vassilevska Williams. Popular conjectures imply
strong lower bounds for dynamic problems. In IEEE 55th Annual Symposium
on Foundations of Computer Science, pages 434443. IEEE, 2014.
[2] Surender Baswana, Shreejit Ray Chaudhury, Keerti Choudhary, and Shahbaz
Khan. Dynamic DFS in undirected graphs: breaking the O(m) barrier. In Pro-
ceedings of the twenty-seventh Annual ACM-SIAM Symposium on Discrete Algo-
rithms, pages 730739. SIAM, 2016.
[3] Timothy M. Chan. Dynamic subgraph connectivity with geometric applications.
In Proceedings of the thiry-fourth annual ACM Symposium on Theory of Comput-
ing, pages 713. ACM, 2002.
[4] Timothy M. Chan, Mihai Patrascu, and Liam Roditty. Dynamic connectivity:
Connecting to networks and geometry. SIAM Journal on Computing, 40(2):333
349, 2011.
[5] Lijie Chen, Ran Duan, Ruosong Wang, and Hanrui Zhang. Improved algorithms
for maintaining DFS tree in undirected graphs. CoRR, abs/1607.04913, 2016.
[6] Don Coppersmith and Shmuel Winograd. Matrix multiplication via arithmetic
progressions. Journal of Symbolic Computation, 9(3):251 280, 1990. Computa-
tional algebraic complexity editorial.
[7] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and Cliord Stein.
Introduction to Algorithms, Third Edition. The MIT Press, 3rd edition, 2009.
348 R. Duan and L. Zhang

[8] Ran Duan. New data structures for subgraph connectivity. In Automata, Lan-
guages and Programming, pages 201212. Springer, 2010.
[9] Ran Duan and Le Zhang. Faster worst-case update time for dynamic subgraph
connectivity. CoRR, abs/1611.09072, 2016.
[10] David Eppstein, Zvi Galil, Giuseppe F. Italiano, and Amnon Nissenzweig.
Sparsicationa technique for speeding up dynamic graph algorithms. Journal
of the ACM, 44(5):669696, 1997.
[11] Greg N. Frederickson. Data structures for on-line updating of minimum spanning
trees, with applications. SIAM Journal on Computing, 14(4):781798, 1985.
[12] Daniele Frigioni and Giuseppe F. Italiano. Dynamically switching vertices in
planar graphs. Algorithmica, 28(1):76103, 2000.
[13] David Gibb, Bruce M. Kapron, Valerie King, and Nolan Thorn. Dynamic graph
connectivity with improved worst case update time and sublinear space. CoRR,
abs/1509.06464, 2015.
[14] Monika Henzinger, Sebastian Krinninger, Danupon Nanongkai, and Thatchaphol
Saranurak. Unifying and strengthening hardness for dynamic problems via the
online matrix-vector multiplication conjecture. In Proceedings of the forty-seventh
Annual ACM on Symposium on Theory of Computing, pages 2130. ACM, 2015.
[15] Monika R. Henzinger and Valerie King. Randomized fully dynamic graph algo-
rithms with polylogarithmic time per operation. Journal of the ACM, 46(4):502
516, 1999.
[16] Jacob Holm, Kristian De Lichtenberg, and Mikkel Thorup. Poly-logarithmic de-
terministic fully-dynamic algorithms for connectivity, minimum spanning tree,
2-edge, and biconnectivity. Journal of the ACM, 48(4):723760, 2001.
[17] Shang-En Huang, Dawei Huang, Tsvi Kopelowitz, and Seth Pettie. Fully dynamic
connectivity in O(log n(log log n)2 ) amortized expected time. In Proceedings of
the twenty-eighth Annual ACM-SIAM Symposium on Discrete Algorithms. SIAM,
2017.
[18] Bruce M. Kapron, Valerie King, and Ben Mountjoy. Dynamic graph connectivity
in polylogarithmic worst case time. In Proceedings of the twenty-fourth Annual
ACM-SIAM Symposium on Discrete Algorithms, pages 11311142. SIAM, 2013.
[19] Casper Kejlberg-Rasmussen, Tsvi Kopelowitz, Seth Pettie, and Mikkel Thorup.
Faster worst case deterministic dynamic connectivity. In Proceedings of the twenty-
fourth Annual European Symposium on Algorithms, 2016.
[20] Danupon Nanongkai and Thatchaphol Saranurak. Dynamic spanning forest with
worst-case update time: Adaptive, Las Vegas, and O(n1/2 )-time. In Proceedings
of the forty-ninth Annual ACM on Symposium on Theory of Computing, 2017.
[21] Mihai Patrascu and Erik D. Demaine. Logarithmic lower bounds in the cell-probe
model. SIAM Journal on Computing, 35(4):932963, 2006.
[22] Mikkel Thorup. Near-optimal fully-dynamic graph connectivity. In Proceedings of
the thirty-second annual ACM Symposium on Theory of Computing, pages 343
350, 2000.
[23] Zhengyu Wang. An improved randomized data structure for dynamic graph con-
nectivity. CoRR, abs/1510.04590, 2015.
[24] Christian Wul-Nilsen. Faster deterministic fully-dynamic graph connectivity.
In Proceedings of the twenty-fourth Annual ACM-SIAM Symposium on Discrete
Algorithms, pages 17571769, 2013.
[25] Christian Wul-Nilsen. Fully-dynamic minimum spanning forest with improved
worst-case update time. In Proceedings of the forty-ninth Annual ACM on Sym-
posium on Theory of Computing, 2017.
 Springer International Publishing AG 2017 349
F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 349360, 2017.
DOI: 10.1007/978-3-319-62127-2_30
350 R. Duan and T. Zhang
Improved distance sensitivity oracles via tree partitioning 351
352 R. Duan and T. Zhang
Improved distance sensitivity oracles via tree partitioning 353
354 R. Duan and T. Zhang
Improved distance sensitivity oracles via tree partitioning 355
356 R. Duan and T. Zhang
Improved distance sensitivity oracles via tree partitioning 357
358 R. Duan and T. Zhang
Improved distance sensitivity oracles via tree partitioning 359
360 R. Duan and T. Zhang
 Delta-Fast Tries: Local Searches in Bounded
Universes with Linear Space

Marcel Ehrhardt and Wolfgang Mulzer

Institut fur Informatik, Freie Universitat Berlin, Germany

[marehr,mulzer]@inf.fu-berlin.de

Abstract. Let w N and U = {0, 1, . . . , 2w 1} be a bounded universe

of w-bit integers. We present a dynamic data structure for predecessor
searching in U . Our structure needs O(log log ) time for queries and
O(log log ) expected time for updates, where is the dierence between
the query element and its nearest neighbor in the structure. Our data
structure requires linear space. This improves a result by Bose et al.
[CGTA, 46(2), pp. 181189].
The structure can be applied for answering approximate nearest neighbor
queries in low dimensions and for dominance queries on a grid.

1 Introduction
Predecessor searching is one of the oldest problems in theoretical computer sci-
ence [5, 12]. Let U be a totally ordered universe. The task is to maintain a set
S U , while supporting predecessor and successor queries: given q U , nd
the largest element in S smaller than q (qs predecessor) or the smallest element
in S larger than q (qs successor). In the dynamic version of the problem, we
also want to be able to modify S by inserting and/or deleting elements.
In the word-RAM model of computation, all input elements are w-bit words,
where w N is a parameter. Without loss of generality, we may assume that w
is a power of 2. We are allowed to manipulate the input elements at the bit level,
in constant time per operation. In this case, we may assume that the universe
is U = {0, . . . , 2w 1}. A classic solution for predecessor searching on the word-
RAM is due to van Emde Boas, who described a dynamic data structure that
requires space O(n) and supports insertions, deletions, and predecessor queries
in O(log log |U |) time [9, 10].
In 2013, Bose et al. [3] described a word-RAM data structure for the pre-
decessor problem that is local in the following sense. Suppose our data struc-
ture currently contains the set S U , and let q U be a query element. Let
q + := min{s S | s q} and q := max{s S | s q} be the succes-
sor and the predecessor of q in S, and let = min{|q q |, |q q + |} be the
distance between q and its nearest neighbor in S. Then, the structure by Bose
et al. can answer predecessor and successor queries in O(log log ) time. Their
solution requires O(n log log log |U |) words of space, where n = |S| is the size of

Supported in part by DFG project MU/3501-1.

Springer International Publishing AG 2017 361

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 361372, 2017.
DOI: 10.1007/978-3-319-62127-2_31
362 M. Ehrhardt and W. Mulzer

the current set. Bose et al. apply their structure to obtain a fast data structure
for approximate nearest neighbor queries in low dimensions and for answering
dominance and range searching queries on a grid.
Here, we show how to obtain a data structure with similar guarantees for the
query and update times that reduces the space requirement to O(n). This solves
an open problem from [3]. Furthermore, this also improves the space requirement
for data structures for nearest neighbor searching and dominance reporting. Full
details and pseudocode for all the algorithms and data structures described here
can be found in the Masters thesis of the rst author [8]. Belazzougui et al. give
a linear space bound for distance-sensitive queries in the static setting, using
almost the same techniques as in the present paper [2]. Our result was obtained
independently from the work of Belazzougui et al.

2 Preliminaries
We begin by listing some known structures and background information required
for our data structure.

Compressed Tries. Our data structure is based on compressed tries [5]. These
are dened as follows: we interpret the elements from S as bitstrings of length
w (the most signicant bit being in the leftmost position). The trie T  for S is a
binary tree of height w. Each node v T  corresponds to a bitstring pv {0, 1} .
The root r has pr = . For each inner node v, the left child u of v has pu = pv 0,
and the right child w of v has pw = pv 1 (one of the two children may not exist).
The bitstrings of the leaves correspond to the elements of S, and the bitstrings
of the inner nodes are prexes for the elements in S, see Figure 1.

0 1 1

00 10 11 11

001 100 110 111 001 100 110 111

Fig. 1. A trie (left) and a compressed trie (right) for the set 000, 100, 110, 111. The
longest common prex of 101 is 10. The lca of 101 in the compressed trie is the node
labeled 1.

The compressed trie T for S is obtained from T  by contracting each maximal

path of nodes with only one child into a single edge. Each inner node in T has
exactly two children, and consequently T has O(n) nodes. Maybe somewhat
Delta-Fast Tries: Local Searches in Bounded Universes with Linear Space 363

unusually, in the following, the height and depth of a node v in T will refer to the
corresponding height and depth in the (uncompressed) trie T  . This convention
will make the description of the operations more convenient.
Let q {0, 1} be a bitstring of length at most w. The longest common prex
of q with S, lcpS (q), is the longest prex that q shares with an element in S.
We say that q lies on an edge e = (u, v) of T if pu is a prex of q and q is a
proper prex of pv . If lcpS (q) lies on the edge (u, v), we call u the lowest common
ancestor of q in T , denoted by lcaT (q). One can show that lcaT (q) is uniquely
dened.

Associated Keys. Our algorithm uses the notion of associated keys. This notion
was introduced in the context of z-fast tries [1, 16], and it is also useful in our
data structure.
Associated keys provide a quick way to compute lcaT (q), for any element q
U . A natural way to nd lcaT (q) is to do binary search on the depth of lcaT (q):
we initialize (l, r) = (0, w) and let m = (l + r)/2. We denote by q  = q0 . . . qm1
the leftmost m bits of q, and we check whether T has an edge e = (u, v) such
that q  lies on e. If not, we set r = m, and we continue. Otherwise, we determine
if u is lcaT (q), by testing whether pv is not a prex of q. If u is not lcaT (q), we
set l = m and continue. In order to perform this search quickly, we need to nd
the edge e that contains a given prex q  , if it exists. For this, we precompute for
each edge e of T the rst time that the binary search encounters a prex that
lies on e. This prex is uniquely determined and depends only on e, not on the
specic string q that we are looking for. We let e be this prex, and we call e
the associated key for e = (u, v), see Figure 2.

0 1
1
1 1
2
0 0 1
3
1 1 0
w 0 0 1 0
4 = w/2
5
1 1 0 1
6
0 0 1 0 1
7
1 1 1 0 1

Fig. 2. The associated key e of an edge e: we perform a binary search on the height
of lcpS (q) in T . The associated key of an edge e is the prex of lcpS (q) in which the
search rst encounters the edge e.

The binary search needs log w steps, and since we assumed that w is a power
of two, each step determines the next bit in the binary expansion of the length
of lcpS (q). Thus, the associated key of an edge e can be computed in O(1) time
364 M. Ehrhardt and W. Mulzer

on a word RAM as follows: consider the log w-bit binary expansions u = |pu |2
and v = |pv |2 of the lengths of the prexes pu and pv , and let  be the longest
common prex of u and v . We need to determine the rst step when the binary
search can distinguish between u and v . Since u < v , and since the two binary
expansions dier in the rst bit after  , it follows that u begins with  0 and v
begins with  1. Thus, let  be obtained by taking  , followed by 1 and enough
0s to make a log w-bit word. Let l be the number encoded by . Then, the
associated key e consists of the rst l bits of pv ; see [1, 8, 16] for more details.

Hash Maps. Our data structure also makes extensive use of hashing. In particu-
lar, we will maintain several succinct hashtables that store additional information
for supporting fast queries. For this, we will use a hashtable described by De-
maine et al. [7]. The following theorem summarizes the properties of their data
structure.

Theorem 2.1. For any r 1, there exists a dynamic dictionary that stores
entries with keys from U and with associated values of r bits each. The dictionary
supports updates and queries in O(1) time, using O(n log log(|U |/n) + nr) bits
of space. The bounds for the space and the queries are worst-case, the bounds for
the updates hold with high probability. 


3 Static -fast Tries

We are now ready to describe our data structure for the static case. In the next
section, we will discuss how to add support for insertions and deletions.

3.1 The Data Structure

Our data structure is organized as follows: let S U , |S| = n, be given. We store

S in a compressed trie T . The leaves of T are linked in sorted order. Furthermore,
for each node v of T , let Tv be the subtree rooted at v. Then, v stores pointers
to the smallest and the largest leaf in Tv . To support the queries, we store three
additional hash maps: H , Hz , and Hb .
First, we describe the hash map H . Set m = log log w. For i = 0, . . . , m, we
let hi = 22 and di = w hi . The hash map H stores the following information:
i

for each s S and each di , i = 1, . . . , m, let si = s0 . . . sdi 1 be the leftmost

di -bits of s and let e = (u, v) be the edge of T such that si lies on e. Then, H
stores the entry si  u.
Next, we describe the hash map Hz . It is dened similarly as the hash map
used for z-fast tries [1, 16]. For each edge e of T , let e be the associated key of
e, as explained in Section 2. Then, Hz stores the entry e  e.
Finally, the hash map Hb is used to implement a second layer of indirection
that lets us achieve linear space. It will be described below.
Delta-Fast Tries: Local Searches in Bounded Universes with Linear Space 365

3.2 The Predecessor Query

Let q U be the query, and let q and q + be the predecessor and the successor
of q in S, as described above. We rst show how to get a running time of
O(log log ) for the queries, with = |q q + |. In Theorem 3.2, we will see
that this can easily be improved to = min{|q q |, |q q + |}.
The predecessor search works in several iterations. In iteration i, we consider
the prex qi that consists of the rst di bits of q.
First, we check whether H contains an entry for qi . If so, we know that T
contains an edge e such that qi lies on e. Hence, qi must be a prex of lcpS (q).
If one of the endpoints of e happens to be lcaT (q), we are done. Otherwise, we
consider the two edges emanating from the lower endpoint of e, nding the edge
e that lies on the path to q. We take the associated key e of e , and we use it to
continue the binary search for lcaT (q), as described in Section 2. Since |qi | = di ,
this binary search takes O(log(w di )) = O(log hi ) steps to complete. Once the
lowest common ancestor v = lcaT (q) is at hand, we can nd the predecessor of q
in O(1) additional time: it is either the rightmost element in Tv , the predecessor
of the leftmost element in Tv , or the rightmost element in the left subtree of
v. Given the pointers stored with v and the leaves of T , all these nodes can be
found in O(1) time.
If H contains no entry for qi and if qi does not consist of all 1s, we check if
H contains an entry for qi + 1. Notice that qi + 1 is the successor of qi . If such
an entry exists, we rst obtain u = H [qi + 1], and the child v of u such that
qi + 1 lies on the edge e = (u, v). Then, we follow the pointer to the leftmost
element of Tv . This is the successor q + of q. The predecessor q can then be
found by following the leaf pointers. This takes O(1) time overall.
Finally, if there is neither an entry for qi nor for qi + 1, we continue with
iteration i + 1, see Figure 3.

Fig. 3. The query algorithm: rst we perform an exponential search from the lowest
level, to nd a prex of qk or qk + 1 (left). If a prex qk is found, we perform a
binary search for lcaT (q) (middle), which can then be used to nd the predecessor and
successor of q (right). If a prex qk + 1 is found, the successor and predecessor can be
found immediately (not shown).

From the above discussion, it follows that the total time for the predecessor
query is O(k+log hk ), where k is the number of iterations and log hk is the worst-
case time for the predecessor search once one of the lookups in an iteration
366 M. Ehrhardt and W. Mulzer

succeeds. By our predecessor algorithm, we know that S contains no element

with prex qk1 or qk1 + 1, but an element with prex qk or qk + 1. Thus,
there must be at least 2wdk = 2hk consecutive elements in U \ S following
2k1
q. By our denition of hk , it follows that 2hk1 = 22 , so k 1 +
 k1 2
log log log . Furthermore, since hk = 22 = 22 = (hk1 )2 , it follows that
k

hk = O(log2 ).

3.3 Obtaining Linear Space

We now analyze the space requirement for our data structure. Clearly, the trie
T and the hash map Hz require O(n) words of space. Furthermore, as described
so far, the number of words needed for H is O(n log log w), since we store at
most n entries for each height hi , i = 0, . . . , m = log log w.
Using a trick due to Patrascu [15], we can introduce another level of indirec-
tion to reduce the space requirement to O(n). The idea is to store in H the
depth du of each branch node u in T , instead of storing u itself (here, we mean
the depth in the original trie, i.e., the length of the prex pu ). We then use an
additional hash map Hb to obtain u. This is done as follows: when trying to nd
the branch node u for a given prex qi , we rst get the depth du = |pu | of u
from H . After that, we look up the branch node u = Hb [q0 . . . qdu 1 ] from the
hash map Hb . Finally, we check whether u is actually the lowest branch node of
qi . If any of those steps fails, we return .
Let us analyze the needed space: clearly, Hb needs O(n) words, since it stores
O(n) entries. Furthermore, we have to store O(n log log w) entries in H , each
mapping a prex qi to the depth of its lowest branch node. This depth requires
log w bits. By Theorem 2.1, a retrieval only hash map for n items and r bits
of data needs O(n log log |U | 
n + n r) bits. Therefore, the space in bits for H is
proportional to

|U |
n log log w log log + n log log w log w
n log log w
= O(n log log w log w)
= o(n w),

using n = n log log w, r = log w and w = log |U |. Thus, we can store H in

O(n) words of w bits each. The following lemma summarizes the discussion

Lemma 3.1. The -fast trie needs O(n) words space.

3.4 Putting it Together

We can now obtain our result for the static predecessor problem.

Theorem 3.2. Let U = {0, . . . , 2w 1} and let S U , |S| = n. The static -

fast trie for S requires O(n) words of space, and it can answer a static predecessor
Delta-Fast Tries: Local Searches in Bounded Universes with Linear Space 367

query for an element q U on S in time O(log log min{|q q |, |q q + |}), where

q and q + denote the predecessor and successor of q in S. The preprocessing time
is O(n log log log |U |), assuming that S is sorted.
Proof. The regular search for q S can be done in O(1) time by a lookup in Hz .
We have seen that the predecessor of q can be found in O(log log |q q + |) time.
A symmetric result also holds for successor queries. In particular, we can achieve
query time O(log log |q q |) by checking for H [qi 1] instead of H [qi + 1]
in the query algorithm.
By interleaving the two searches, we obtain the desired running time of
O(log log min{|q q |, |q q + |}). Of course, in a practical implementation, it
would be more ecient to check directly for H [qi 1] and H [qi + 1] in the
query algorithm.
The trie T and the hash maps Hz and Hb can be computed in O(n) time,
given that S is sorted. Thus, the preprocessing time is dominated by the time to
ll the hash map H . Hence, the preprocessing needs O(n log log log |U |) steps,
because O(n log log w) nodes have to be inserted into H . By Lemma 3.1, the
space requirement is linear. 


4 Dynamic -fast tries

We will now explain how to extend our data structure to the dynamic case. The
basic data structure remains the same, but we need to update the hashtables
and the trie T after each insertion and deletion. In particular, our data structure
requires that for each v in Tv , we can access the leftmost and the rightmost node
in the subtree Tv . In the static case, this could be done simply by maintaining
explicit pointers from each node v T to these nodes in Tv , letting us nd the
nodes in O(1) time. In the dynamic case, we will maintain a data structure which
allows nding and updating these nodes in in O(log log ) time.

4.1 Computing Lowest Common Ancestor

To perform the update operation, we need a procedure to compute the lowest
common ancestor lcaT (q) for any given element q U . For this, we proceed as
in the query algorithm from Section 3.2, but skipping the lookups for H [qi 1]
and H [qi + 1]. By the analysis in Section 3.2, this will nd lcaT (q) in time
O(log log l), where l is height of lcaT (q) in T .
Unfortunately, it may happen that this height l is as large as w, even if q is
close to an element in the current set S. To get around this, we use a trick of
Bose et al. [3]. Namely, their idea is to perform a random shift of the universe.
More precisely, we pick a random number r U , and we add r to all query and
update elements that appear in the data structure (modulo |U |).
Lemma 4.1 (Lemma 4 in [3]). Let x, y U be two xed elements in U .
Let r U be picked uniformly at random. After a random shift of U by r, the
expected height of the lowest common ancestor of x and y in a compressed trie
is O(log |x y|). 

368 M. Ehrhardt and W. Mulzer

Corollary 4.1. Let S U and let T be a randomly shifted -fast trie stor-
ing S. Let q U . We can nd lcaT (q) in expected time O(log log ), where
= min{|q q + |, |q q |}, the elements q + and q being the predecessor and
successor of q in S. The expectation is over the random choice of the shift r.

Proof. Suppose without loss of generality that = |q q + |. By Lemma 4.1,

the expected height hk of the lowest common ancestor of q and q + is O(log ).
We perform the doubly exponential search on the prexes of q, as in Section 3.2
(without checking qi + 1) to nd the height hk . After that, we resume the search
for lcaT (q) on the remaining hk bits. Since hk = O(log ) in expectation, it
follows by Jensens inequality that the number k of loop iterations to nd hk is
O(log log log ) in expectation. Thus, the expected running time is proportional
to k + log hk = O(log log ). 


4.2 Managing the Left- and Rightmost Elements of the Subtrees

We also need to maintain for each node v T the leftmost and the rightmost
element in the subtree Tv . In the static case, it suces to have direct pointers
from v to the respective leaves, but in the dynamic case, we need an additional
data structure.

Fig. 4. For each leaf v  of T , the nodes v T for which v is the leftmost leaf in Tv if a
subpath of a root-to-leaf path in T . Considering these subpaths for all leaves in T , we
obtain a path decomposition of T (shown in bold).

To do this, we observe the following: let v  T be a leaf in T . Then, v  is

the leftmost (or rightmost) leaf in the subtrees of at most w ancestors v of v  .
Furthermore, all these nodes form a subpath (more precisely, a prex) of the
path from v to the root, see Figure 4. Hence, if we maintain the nodes of this
subpath in a concatenable queue data structure (realized by, e.g., a balanced
binary tree) [14], we can obtain O(log w) update and query time to nd the
leftmost (or rightmost) element in Tv for each v T . However, we need that
the update and query time for this data structure depend on the height hi (i.e,
the remaining bits) of the query node v. Thus, we partition the possible heights
{0, 1, . . . , w} of the nodes on a subpath into the sets T1 = {0}, Ti = [2i , 2i+1 ),
for i = 0, . . . , log w 1, and Tlog w = {w}. Each set is managed by a balanced
Delta-Fast Tries: Local Searches in Bounded Universes with Linear Space 369

binary tree, and the roots of the trees are linked together. The height of the i-th
binary search tree is log |Ti | = O(i). Furthermore, if a query node of height h is
given, the set Tlog h is responsible for it, see Figure 5.

0 [1, 2) [2, 4) [w/2, w)

T0 1
T1 2
Tlog w1 log w 1

Fig. 5. The data structure for a subpath. We group the nodes in the subpath according
to their heights, where the groups grow exponentially in size. Each group is represented
by a balanced tree. The roots are joined in a linked list. With this data structure, a
node v of height h can nd the leftmost leaf in the subtree Tv in time O(log h).

Moreover, T1 is a leaf (the depth of that node is w) in the trie and therefore
the minimum of the whole subpath. Thus, the minimum of a subpath can be
found from a given node v Ti in O(i) time by following the pointers to the
root of Ti and the pointers down to T1 .
If a node v has hk = O(log ) height (remaining bits), the node is within the
tree Tlog hk  . Thus, it takes O(log hk ) = O(log log ) time to nd the leftmost
or rightmost leaf in Tv .
Furthermore, we can support the following update operations: (i) split: given
a subpath and a node v on , split the representation of into two represen-
tations, one for the lower subpath from the leaf up to the child of v, and one
for the upper subpath starting from v; and (ii) join: given a representation of
an upper subpath starting at a node v obtained from an operation of type (i),
and a representation for a lower subpath up to a child of v, join the two repre-
sentations into the representation for a joint subpath. Given the data structure,
we can support both split and join in O(log h) time, where h is the height of
the node v where the operation occurs. This decomposition of T into dynami-
cally changing suppaths is similar to the preferred paths decomposition of Tango
trees [6].

4.3 Performing an Update

We know from the Lemma 4.1, that the lowest common ancestor of a query
element q has expected height hk = O(log ).
370 M. Ehrhardt and W. Mulzer

Lemma 4.2. Let S U , and let T be a randomly shifted -fast tree for S. Let
q U be xed. We can insert or delete q into T in O(log log ) expected time,
where the expectation is over the random choice of the shift r.

Proof. To insert q into T , we need to split an edge (u, v) of T into two edges
(u, b) and (b, v). This creates exactly two new nodes in T , an inner node and
a leaf node. The branch node is exactly lcaT (q), and it has expected height
hk = O(log ), by Lemma 4.1. Thus, it will take O(log log ) expected time to
nd the edge (u, v), by Corollary 4.1.
Once the edge (u, v) is found, the hash maps Hz and Hu can then be updated
in constant time. Now let us consider the update time of the hash map H . Recall
that H stores the lowest branch nodes for all prexes of the elements in S that
have certain lengths. This means that all prexes on the edge (b, v) which are
stored in the hash map T need to be updated. Furthermore, prexes at certain
depths which are on the new edge (b, q) need to be added. For the edge (b, v),
we will enumerate all prexes at certain depths, but we will select only those
that lie on the edge (b, v). This needs O(log log log ) insertions and updates in
total: we have to insert the prexes q0 . . . qdi for all i 1 with di < |b|. Since
we dened di = w hi = w 22 , and since |b| = w O(log ), we have that
i

di |b| as soon as c log < 22 . This holds for i > log log(c log ), and hence
i

i = (log log log ).

After that, the leftmost and rightmost elements for the subtrees of T have
to be updated. For this, we need to add one subpath for the new leaf q, and
we may need to split a subpath at a node of height hk = O(log ) and join
the resulting upper path with the newly created subpath. As we have seen, this
takes O(log hk ) = O(log log ) time.
The operations for deleting an element q from S are symmetric. 


The following theorem summarizes our result.

Theorem 4.3. Let r U be picked uniformly at random. After performing

a shift of U by r, the -fast trie provides a data structure for the dynamic
predecessor problem such that the query operations take O(log log ) worst-case
time and the update operations need O(log log ) expected time, for = min{|q
q + |, |q q |}, where q is the requested element and q + and q are the predecessor
and successor of q in the current set S. At any point in time, the data structure
needs O(n) words of space, where n = |S|.

5 Applications
Bose et al. [3] describe how to combine their structure with a technique of
Chan [4] and random shifting [11, Chapter 11] for obtaining a data structure
for distance-sensitive approximate nearest neighbor queries on a grid. More pre-
cisely, let d N be the xed dimension, U = {0, . . . , 2w 1} be the universe,
and let > 0 be given. The goal is to maintain a dynamic set S U d under
insertions, deletions, and -approximate nearest neighbor queries: given a query
Delta-Fast Tries: Local Searches in Bounded Universes with Linear Space 371

point q U d , nd a p S with d2 (p, q) (1 + )d2 (p, S). Plugging our -fast

tries into the structure of Bose et al. [3, Theorem 9], we can immediately improve
the space requirement of their structure to linear:
Theorem 5.1. Let U = {0, . . . , 2w 1} and let d be a constant. Furthermore,
let > 0 be given. There exists a data structure that supports (1+)-approximate
nearest neighbor queries over a subset S U d in (1/d ) log log ) expected time
and insertions and deletions of elements of U d in O(log log ) expected time.
Here, denotes the Euclidean distance between the query element and S. At any
point in time, the data structure requires O(n) words of space, where n = |S|.
As a second application, Bose et al. [3] present a data structure for dom-
inance queries on a grid, based on a technique of Overmars [13]. Again, let
U = {0, . . . , 2w 1}, and let S U 2 , |S| = n be given. The goal is to con-
struct a data structure for dominance queries in S. That is, given a query point
q U 2 , nd all points p in S that dominate q, i.e., for which we have px qx
and py qy , there px , py and qx , qy are the x- and y-coordinates of p and q.
Again, using -fast tries, we can immediately improve the space requirement
for the result of Bose et al. [3, Theorem 10, Corollary 13].
Theorem 5.2. Let U = {0, . . . , 2w 1}, and let S U 2 , |S| = n be given. There
exists a data structure that reports the points in S that dominate a given query
point q = (a, b) U 2 in expected time O(log log(h + v) + k), where h = 2w a,
v = 22 b, and k is the number of points in S dominated by q. The data structure
uses O(n log n) space.

6 Conclusion
We present a new data structure for local searches in bounded universes. This
structure now interpolates seamlessly between hashtables and van-Emde-Boas
trees, while requiring only a linear number of words. This provides an improved,
and in our opinion also slightly simpler, version of a data structure by Bose
et al. [3]. All the operations of our structure can be presented explicitly in pseu-
docode. This can be found in the Masters thesis of the rst author [8].

Acknowledgments. We thank the anonymous reviewers for numerous insight-

ful comments that improved the quality of the paper. In particular, we would
like to thank the anonymous reviewers for pointing us to [2].

References
1. Belazzougui, D., Boldi, P., Vigna, S.: Dynamic z-fast tries. In: Proc. 17th Int.
Symp. String Processing and Information Retrieval (SPIRE). pp. 159172 (2010)
2. Belazzougui, D., Boldi, P., Vigna, S.: Predecessor search with distance-sensitive
query time (2012), arXiv:1209.5441
3. Bose, P., Doueb, K., Dujmovic, V., Howat, J., Morin, P.: Fast local searches and
updates in bounded universes. Comput. Geom. Theory Appl. 46(2), 181189 (2013)
372 M. Ehrhardt and W. Mulzer

4. Chan, T.M.: Closest-point problems simplied on the RAM. In: Proc. 13th Annu.
ACM-SIAM Sympos. Discrete Algorithms (SODA). pp. 472473 (2002)
5. Cormen, T.H., Leiserson, C.E., Rivest, R.L., Stein, C.: Introduction to algorithms.
MIT Press, third edn. (2009)
6. Demaine, E.D., Harmon, D., Iacono, J., Patrascu, M.: Dynamic optimality al-
most. SIAM J. Comput. 37(1), 240251 (2007)
7. Demaine, E.D., Meyer auf der Heide, F., Pagh, R., Patrascu, M.: De dictionariis
dynamicis pauco spatio utentibus. In: Proc. 7th Latin American Symp. Theoretical
Inf. (LATIN). pp. 349361 (2006)
8. Ehrhardt, M.: An In-Depth Analysis of Data Structures Derived from van-
Emde-Boas-Trees. Masters thesis, Freie Universitat Berlin (2015), http://
www.mi.fu-berlin.de/inf/groups/ag-ti/theses/download/Ehrhardt15.pdf
9. van Emde Boas, P., Kaas, R., Zijlstra, E.: Design and implementation of an ecient
priority queue. Math. Systems Theory 10(2), 99127 (1976)
10. van Emde Boas, P.: Preserving order in a forest in less than logarithmic time and
linear space. Inform. Process. Lett. 6(3), 8082 (1977)
11. Har-Peled, S.: Geometric approximation algorithms, Mathematical Surveys and
Monographs, vol. 173. American Mathematical Society (2011)
12. Knuth, D.E.: The art of computer programming. Vol. 3. Sorting and searching.
Addison-Wesley, second edn. (1998)
13. Overmars, M.H.: Ecient data structures for range searching on a grid. J. Algo-
rithms 9(2), 254275 (1988)
14. Preparata, F.P., Shamos, M.I.: Computational geometry. An introduction. Springer
Verlag (1985)
15. Patrascu, M.: vEB space: Method 4 (2010), http://infoweekly.blogspot.de/
2010/09/veb-space-method-4.html
16. Ruzic, M.: Making deterministic signatures quickly. ACM Transactions on Algo-
rithms 5(3), 26:126:26 (2009)
 Split Packing: Packing Circles into Triangles
with Optimal Worst-Case Density

Sandor P. Fekete, Sebastian Morr, and Christian Scheer

Department of Computer Science, TU Braunschweig, Germany

[email protected], [email protected], [email protected]

Abstract. In the circle packing problem for triangular containers, one

asks whether a given set of circles can be packed into a given triangle.
Packing problems like this have been shown to be NP-hard. In this paper,
we present a new sucient condition for packing circles into any right
or obtuse triangle using only the circles combined area: It is possible
to pack any circle instance whose combined area does not exceed the
triangles incircle. This area condition is tight, in the sense that for any
larger area, there are instances which cannot be packed.
A similar result for square containers has been established earlier this year,
using the versatile, divide-and-conquer-based Split Packing algorithm. In
this paper, we present a generalized, weighted version of this approach,
allowing us to construct packings of circles into asymmetric triangles.
It seems crucial to the success of these results that Split Packing does
not depend on an orthogonal subdivision structure. Beside realizing all
packings below the critical density bound, our algorithm can also be
used as a constant-factor approximation algorithm when looking for the
smallest non-acute triangle of a given side ratio in which a given set of
circles can be packed.
An interactive visualization of the Split Packing approach and other
related material can be found at https://morr.cc/split-packing/.

1 Introduction

Given a set of circles, can you decide whether it is possible to pack these circles into
a given container without overlapping one another or the containers boundary?
This naturally occurring circle packing problem has numerous applications in
engineering, science, operational research and everyday life. Examples include
packaging cylinders [2], bundling tubes or cables [16, 18], the cutting industry
[17], the layout of control panels [2], the design of digital modulation schemes
[14], or radio tower placement [17]. Further applications stem from chemistry
[19], foresting [17], and origami design [9].
Despite its simple formulation, packing problems like these were shown to
be NP-hard in 2010 by Demaine, Fekete, and Lang [3], using a reduction from
3-Partition. Additionally, due to the irrational coordinates which arise when
packing circular objects, it is also surprisingly hard to solve circle packing
problems in practice. Even when the input consists of equally-sized circles, exact

Springer International Publishing AG 2017 373

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 373384, 2017.
DOI: 10.1007/978-3-319-62127-2_32
374 S.P. Fekete et al.

boundaries for the smallest square container are currently only known for up to
35 circles, see [10]. For right isosceles triangular containers, optimal results have
been published for up to 7 equal circles, see [20].
The related problem of packing square objects has long been studied. Already
in 1967, Moon and Moser [12] found a sucient condition: They proved that it
is possible to pack a set of squares into the unit square in a shelf-like manner if
their combined area does not exceed 1/2, see Figure 2. At the same time, 1/2 is
the largest upper area bound you could hope for, because two squares larger than
the quarter-squares depicted in Figure 1 cannot be packed anymore. We call the
ratio between the largest combined object area that can always be packed and
the area of the container the problems critical density, or worst-case density.

Fig. 1. Worst-case Fig. 2. Example of Fig. 3. Worst-case Fig. 4. Example

instance for pack- Moon and Mosers instance for pack- packing produced
ing squares into a shelf-packing. ing circles into a by Split Packing.
square. square.

Fig. 5. Suspected worst-case instance Fig. 6. Example packing produced by

for packing circles into a non-acute tri- Split Packing.
angle.

We recently showed a similar result for circular objects: Each circle instance
not exceeding the area of the instance shown in Figure 3 can be packed, and
this area condition is tight [13]. Proving this required a fundamentally dierent
approach than Moon and Mosers orthogonal shelf-packing, compare Figure 4.
In this paper, we consider the problem of packing circles into non-acute
triangular containers. It is obvious that circles larger than a triangles incircle
cannot be packed (compare Figure 5), but is it also possible to pack all circle
instances of up to that combined area? We will answer this question armatively
and introduce a weighted modication of the Split Packing algorithm, allowing
Split Packing: Packing Circles into Triangles with Optimal Worst-Case Density 375

us to pack circles into asymmetric non-acute triangles with critical density. See
Figure 6 for an example packing.
Many authors have considered heuristics for circle packing problems, see [7,
17] for overviews of numerous heuristics and optimization methods. The best
known solutions for packing equal circles into squares, triangles and other shapes
are continuously published on Spechts website http://packomania.com [15].
That being said, the literature on exact approximation algorithms which
actually give performance guarantees is small. Miyazawa et al. [11] devised
asymptotic polynomial-time approximation schemes for packing circles into the
smallest number of unit square bins. And recently, Hokama, Miyazawa, and
Schouery [8] developed an asymptotic approximation algorithm for the online
version of that problem. To the best of our knowledge, this paper presents the
rst approximation algorithm for packing circles into triangular containers.

1.1 Results

We show that, for any right or obtuse triangle, any circle instance with a combined
area of up to the triangles incircle can be packed into that triangle. At the same
time, for any larger area, there are instances which cannot be packed, making the
ratio between the incircles and the triangles area the triangles critical density.
For a right isosceles triangle, this density is approximately 53.91%. In the general
case, the critical density of a non-acute triangle with side lengths a, b, and c is
+
(a b c)(a + b c)(a b + c)
.
(a + b + c)3
Our proof is constructive: The Split Packing algorithm can be used to construct
the packings in polynomial time. Split Packing can also be used as a constant-
factor approximation algorithm of the smallest-area non-acute triangle of a given
side ratio which can pack a given set of circles. The approximation factor is the
reciprocal of the critical density.
While we focus on triangular containers in this paper, we see more opportu-
nities to generalize the Split Packing approach for other container and object
types. We discuss some of these extensions in the conclusion on page 11.

1.2 Key ideas

The Split Packing approach, which we successfully used for packing circles into
square containers earlier this year [13], is built on two basic ideas:
First, it applies a recursive subdivision strategy, which cuts the container
into smaller triangles, while keeping the combined area of the triangles incircles
constant. And second, it performs the splitting of the circle instance into subgroups
using an algorithm which resembles greedy scheduling. This makes sure the
resulting subgroups are close to equal in terms of their combined area. If the
groups areas deviate from the targeted 1:1 ratio, we can gain information
376 S.P. Fekete et al.

about the minimum circle size in the larger group, allowing us to round o the
subcontainer triangles.
In this paper, we introduce a weighted generalization of the Split Packing
approach: When packing into asymmetric triangles, we do not want the resulting
groups to have equal area, as it is not possible to cut the container into two
subtriangles of equal size. Instead, we target a dierent area ratio, dened by the
incircles of the two triangles created by cutting the container orthogonally its
the base through its tip, see Figure 8 on page 8. We call this desired area ratio
the split key.
The rest of the paper will detail this process.

2 Greedy splitting
The following denitions makes it easier to talk about the properties of circle
instances:
Denition 1. A circle instance is a multiset of nonnegative real numbers, which
dene the circles areas. For any circle instance C, sum(C) is the combined area
of the instances circles and min(C) is the area of the smallest circle contained
in the instance.
Denition 2. C is the set of all circle instances. C(a) consists of exactly those
circle instances C with sum(C) a. Finally, C(a, b) consists of exactly those
circle instances C C(a) with min(C) b.
Algorithm 1 takes a circle instance C, and splits it into two groups according
to the split key F , which determines the targeted ratio of the resulting groups
combined areas. Because the method resembles a greedy scheduling algorithm, we
call the process greedy splitting. The algorithm rst creates two empty buckets,
and in each step adds the largest remaining circle of the input instance to the
relatively more empty bucket:

Algorithm 1 Split(C, F )
Input: A circle instance C, sorted by size in descending order, and a split key
F = (f1 , f2 )
Output: Circle instances C1 , C2
C1
C2
for all c C do
j = arg mini sum(C
fi
i)
 Find the index of the more empty bucket.
Cj Cj {c}
end for

If the resulting groups area ratio deviates from the area ratio targeted by
the split key, we gain additional information about the relatively larger group:
Split Packing: Packing Circles into Triangles with Optimal Worst-Case Density 377

The more this group exceeds its targeted ratio, the larger the minimum size of its
elements, allowing a more rounded subcontainer in the packing. See Figure 9
on page 9 for an illustration.
Lemma 1. For any C1 and C2 produced by Split(C, (f1 , f2 )):

sum(Cj )
min(Ci ) sum(Ci ) fi
fj
sum(C )
Proof. If sum(C
fi
i)
< fj
j
, then the lemma says that min(Ci ) is larger than a
negative number, which is certainly true.
sum(Cj )
Otherwise, set r := fj . This value describes the smaller relative lling
level by the time the algorithm ends. Now assume for contradiction Ci con-
tained an element smaller than sum(Ci ) fi r. As the elements were inserted by
descending size, all elements which were put into Ci after that element would
have to be at least as small. So the nal element put into Ci (let us call it c)
would be smaller than sum(Ci ) fi r, as well.
But this means that
sum(Ci ) c sum(Ci ) (sum(Ci ) fi r)
> = r,
fi fi
meaning that at the moment before c was inserted, the relative lling level of Ci
would already have been larger than r. Recall that r is the smallest lling level
of any group by the time the algorithm ends, meaning that at the time when c
is inserted, Ci s lling level is already larger than the lling level of the other
group. This is a contradiction, as the greedy algorithm would choose to put c
not into Ci , but into the other group with the smaller lling level in this case. 
We are now going to dene a term which encapsulates all properties of the
circle instances output by Split. These properties depend on the used split key
F , and also on the combined area a and the minimum circle size b of the circle
instance, which is why it the term has three parameters.
Denition 3. For any 0 b a and any split key F = (f1 , f2 ), we say that the
tuples (a1 , b1 ), (a2 , b2 ) are (a, b, F )-conjugated if

a1 + a2 = a,
bi b, and
a
bi ai fi fjj .

Two circle instances C1 and C2 are (a, b, F )-conjugated if there are any
(a, b, F )-conjugated tuples (a1 , b1 ) and (a2 , b2 ) so that C1 C(a1 , b1 ) and C2
C(a2 , b2 ).
We can now associate this property with Split in the following theorem:
Theorem 1. For any C C(a, b) and any split key F = (f1 , f2 ), Split(C, F )
always produces two (a, b, F )-conjugated subinstances.
378 S.P. Fekete et al.

Proof. That the subinstances combined areas add up to a follows directly from
the algorithm. As the minimum size of all circles in C is b, this must also be true
for the subinstances, so min(Ci ) b. The other minimum-size property follows
from Lemma 1. 

3 Split Packing

The Split algorithm presented in the previous section, in addition to the proper-
ties of the instances it produces, are the foundations on which we now build the
central theorem of this paper. Split Packing by itself is a general framework to
pack circles and other shapes into containers. We will apply the Split Packing
theorem to triangular containers in the next section.
We will often want to state that a shape can pack all circle instances which
belong to a certain class. For this, we dene the term C-shape:

Denition 4. For any C C, a C-shape is a shape in which each C C can be

packed.

For example, if a shape is a C(a)-shape, it means that it can pack all circle
instances with a combined area of a. And a C(a, b)-shape can pack all circle
instances with a combined area of a, whose circles each have an area of at most b.
We can now state our central theorem: If it is possible to nd two subcontainers
which t in a given shape, and which can pack all possible subinstances produced
by Split, it is possible to pack the original class of circle instances into that
shape.

Theorem 2 (Split Packing). A shape s is a C(a, b)-shape if there is a split

key F , so that for all (a, b, F )-conjugated tuples (a1 , b1 ) and (a2 , b2 ) one can nd
a C(a1 , b1 )-shape and a C(a2 , b2 )-shape which can be packed into s.

Proof. Consider an arbitrary C C(a, b). We use Split(C, F ) to produce two

subinstances C1 and C2 . We know from Theorem 1 that those subinstances will
always be (a, b, F )-conjugated. So if we can indeed nd two shapes which can
pack these subinstances, and if we can pack these two shapes into s, then we also
can pack the original circle instance C into s.
Note that in the special case that C consists of a single circle, Split(C, F )
will yield two circle instances C1 = {C} and C2 = . For this case, Theorem 1
guarantees a minimum size of a for the rst group, and the associated C(a1 , b1 )-
shape is just an a-circle. This means that we can simply place the input circle in
the container, and stop the recursion at this point. 

Written as an algorithm, Split Packing looks like this:

Split Packing: Packing Circles into Triangles with Optimal Worst-Case Density 379

Algorithm 2 Splitpack(s, C)
Input: A C(a, b)-shape s and a circle instance C C(a, b), sorted by size in descending
order
Output: A packing of C into s
Determine split key F for shape s
(C1 , C2 ) Split(C, F )  See Algorithm 1.
for all i {1, 2} do
ai sum(Ci )
bi minimum guarantee for Ci  See Lemma 1.
Determine a C(ai , bi )-shape si
Splitpack(si , Ci )
end for
Pack s1 , s2 , and their contents into s

Note that the Split Packing algorithm can easily be extended to allow splitting
into more than two subgroups. For simplicity, we only describe the case of two
subgroups here, as this suces for the shapes we discuss in this paper.

3.1 Analysis

The analysis of the Split Packing approach follows exactly the same lines as in
our previous paper [13]. We will repeat the result here without proof.

Theorem 3. Split Packing requires O(n) basic geometric constructions and

O(n2 ) numerical operations.

Theorem 4. Split Packing, when used to pack circles into a C(a, b)-shape of area
A, is an approximation algorithm with an approximation factor of A a , compared
to the container of minimum area.

4 Packing into hats

After this general description of Split Packing, we will now apply it to concrete
containers. We start with an observation:
If all circles which we want to pack have a certain minimum size, sharp corners
of the container cannot be utilized anyway. This observation motivates a family
of shapes which resemble rounded triangles. We call these shapes hats:

Denition 5. For each 0 b a, an (a, b)-hat is a non-acute triangle with

an incircle of area a, whose corners are rounded to the radius of a b-circle, see
Figure 7. Call the two smaller angles of the original triangle left-angle and
right-angle. If we say right hat, the hat is based on a right triangle.
380 S.P. Fekete et al.

left-angle b b right-angle

Fig. 7. An (a, b)-hat.

We will now proceed to show that all circle instances with a combined area
of up to a with a minimum circle size of b can be packed into an (a, b)-hat.
First, it is important to choose the correct split key when packing into
asymmetric hats. We are aiming for a group ratio which will lead to a cut through
the hats tip if it is reached exactly:
Denition 6. To get a hats associated split key, split the underlying triangle
orthogonally to its base through its tip, and inscribe two circles in the two sides,
see Figure 8. The areas of these circles are the two components of the hats split
key.

f1 f2

Fig. 8. A hats associated split key equals (f1 , f2 )

Lemma 2. Consider an (a, 0)-hat with the associated split key F = (f1 , f2 ), and
call its left- and right-angles and . For all (a, 0, F )-conjugated tuples (a1 , b1 )
and (a2 , b2 ), the following two shapes can be packed into the hat:
a right (a1 , b1 )-hat with a right-angle of and
a right (a2 , b2 )-hat with a left-angle of .
The proof of this theorem is rather technical in nature. We omit it here due
to space constraints, refer to the full version [6]. See Figure 9 for an intuition
of what the resulting hats look like. Note that, as the hats incircles are getting
larger than the targeted area ratio, their corners become more rounded so that
they dont overlap the containers boundary.
Split Packing: Packing Circles into Triangles with Optimal Worst-Case Density 381

Fig. 9. Hat-in-hat packings for dierent ratios of a1 and a2 .

In the previous lemma, the container is always an (a, 0)-hat, which is essentially
a non-rounded triangle with an incircle of a. The next lemma extends this idea
to hats which are actually rounded. It is identical to Lemma 2, except that the
rounding of the container hat is no longer 0, but b.

Lemma 3. Consider an (a, b)-hat with the associated split key F = (f1 , f2 ), and
call its left- and right-angles and . For all (a, b, F )-conjugated tuples (a1 , b1 )
and (a2 , b2 ) with a1 + a2 a, the following two shapes can be packed into the hat:
a right (a1 , b1 )-hat with a right-angle of and
a right (a2 , b2 )-hat with a left-angle of .

Proof. Lemma 2 tells us that this theorem is true for b = 0. Now, the containers
corners can be rounded to the radius of a b-circle, and we need to show that the
two hats from the previous construction still t inside. But all of the two hats
corners are also rounded to (at least) the same radius (see Theorem 1), so they
will never overlap the container, see Figure 10. 

0.7a 0.7a
0.3a 0.3a

Fig. 10. Rounding all hats corners by the same radius does not aect the packing.

With these preparations, we can apply Split Packing to hats:

Theorem 5. Given an (a, b)-hat, all circle instances with a combined area of at
most a and a minimum circle size of at least b can be packed into that hat.

Proof. We proof by induction that we can pack each C C(a, b) into the hat:
If C only consists of a single circle, it can be packed into the hat, as it is at
most as big as the hats incircle.
Now assume that for any 0 b a, any (a, b)-hat could pack all circle
instances into C(a, b) with at most n circles. Consider a circle instance C C(a, b)
382 S.P. Fekete et al.

containing n + 1 circles. Denition 6 tells us how to compute the split key F .

Then we know from Theorem 1 that Split will partition C into two subinstances
C1 C(a1 , b1 ) and C2 C(a2 , b2 ), whose parameters are (a, b, F )-conjugated. As
Split can never return an empty instance (except for |C| = 1, a case which we
handled above), each subinstance will contain at most n circles. We know from
Lemma 3 that, for all pairs of (a, b, F )-conjugated tuples, we can nd two hats
with matching parameters which t into the container hat. By assumption, these
hats can now pack all instances from C(a1 , b1 ) and C(a2 , b2 ), respectively, which
means that they can especially also pack C1 and C2 . If we then pack the two hats
into the container, we have constructed a packing of C into the container hat.
By induction, we can pack each C C(a, b) into the (a, b)-hat. 

Finally, we can state this papers central result:

Theorem 6. Given a non-acute triangle with an incircle of area a, all circle

instances with a combined area of up to a can be packed into the triangle, and
this bound is tight. Expressed algebraically, for a triangle with side lengths a, b,
and c, the critical density is
+
(a b c)(a + b c)(a b + c)
.
(a + b + c)3

Proof. The triangle is an (a, 0)-hat, which by Theorem 5 is a C(a)-shape.

On the other hand, a single circle of area a + cannot be packed, as the
incircle is by denition the largest circle which ts into the triangle.
As for the algebraic formulation of the critical density, the area of the triangle
can be calculated using Herons formula:
$ a+b+c
(a, b, c) := s(s a)(s b)(s c) with s =
2
It is also known that the radius of the incircle of this triangle is

(a, b, c) a+b+c
R(a, b, c) := with s = ,
s 2
so the incircle has an area of

(a + b c)(c + a b)(b + c a)
I(a, b, c) = R(a, b, c)2 = .
4(a + b + c)
Finally, the ratio between the areas of the circle and the triangle can be
calculated to be
+
I(a, b, c) (a b c)(a + b c)(a b + c)
= ,
(a, b, c) (a + b + c)3

For a right isosceles triangle, this density is approximately 53.91%. 

Split Packing: Packing Circles into Triangles with Optimal Worst-Case Density 383

5 Conclusion
In this paper, we presented a constructive proof of the critical densities when
packing circles into right or obtuse triangles, using a weighted Split Packing
technique. We see more opportunities to apply this approach in the context of
other packing and covering problems.
It is possible to use Split Packing to pack into more container types. At this
point, we can establish the critical densities for packing circles into equilateral
triangles and rectangles exceeding a certain aspect ratio. The case of acute
triangles is still open, we discuss why the approach presented in this paper does
not work there in the full version [6].
Split Packing can also be extended to pack objects other than circles. We
can establish the critical densities for packing octagons into squares, and think
we can describe the maximum shape which can be packed into squares using
Split Packing.
Another natural extension is the online version of the problem. The current
best algorithm that packs squares into a square in an online fashion by Brubach [1],
based on the work by Fekete and Homann [4, 5], gives a density guarantee of
2
5 . It is possible to directly use this algorithm to pack circles into a square in
an online situation with a density of 10
0.3142. It would be interesting to see
whether some form of online Split Packing would give better results.
A related problem asks for the smallest area so that we can always cover
the container with circles of that combined area. For example, we conjecture
that for an isosceles right triangle, any circle instance with a total area of at least
its excircles area is sucient to cover it.

Acknowledgements
We thank the three anonymous reviewers for their helpful comments.

References
[1] Brian Brubach. Improved Bound for Online Square-into-Square Packing.
In: Approximation and Online Algorithms. Springer, 2014, pp. 4758.
[2] Ignacio Castillo, Frank J. Kampas, and Janos D. Pinter. Solving circle
packing problems by global optimization: numerical results and industrial
applications. In: European Journal of Operational Research 191(3) (2008),
pp. 786802.
[3] Erik D. Demaine, Sandor P. Fekete, and Robert J. Lang. Circle Packing for
Origami Design is Hard. In: 5th International Conference on Origami in
Science, Mathematics and Education. AK Peters/CRC Press. 2011, pp. 609
626.
[4] Sandor P. Fekete and Hella-Franziska Homann. Online Square-into-
Square Packing. In: APPROX-RANDOM. 2013, pp. 126141.
[5] Sandor P. Fekete and Hella-Franziska Homann. Online Square-into-
Square Packing. In: Algorithmica 77(3) (2017), pp. 867901.
384 S.P. Fekete et al.

[6] Sandor P. Fekete, Sebastian Morr, and Christian Scheer. Split Packing:
Algorithms for Packing Circles with Optimal Worst-Case Density. In:
CoRR abs/1705.00924 (2017). http://arxiv.org/abs/1705.00924.
[7] Mhand Hi and Rym Mhallah. A literature review on circle and sphere
packing problems: models and methodologies. In: Advances in Operations
Research Article ID 150624 (2009).
[8] Pedro Hokama, Flavio K. Miyazawa, and Rafael C. S. Schouery. A bounded
space algorithm for online circle packing. In: Information Processing Letters
116(5) (May 2016), pp. 337342. issn: 0020-0190.
[9] Robert J. Lang. A computational algorithm for origami design. In: Pro-
ceedings of the Twelfth Annual Symposium on Computational Geometry
(SoCG) (1996), pp. 98105.
[10] Marco Locatelli and Ulrich Raber. Packing equal circles in a square: a
deterministic global optimization approach. In: Discrete Applied Mathe-
matics 122(1) (2002), pp. 139166.
[11] Flavio K. Miyazawa, Lehilton L.C. Pedrosa, Rafael C.S. Schouery, Maxim
Sviridenko, and Yoshiko Wakabayashi. Polynomial-time approximation
schemes for circle packing problems. In: European Symposium on Algo-
rithms (ESA). Springer. 2014, pp. 713724.
[12] John W. Moon and Leo Moser. Some packing and covering theorems.
In: Colloquium Mathematicae. Vol. 17. 1. Institute of Mathematics, Polish
Academy of Sciences. 1967, pp. 103110.
[13] Sebastian Morr. Split Packing: An Algorithm for Packing Circles with
Optimal Worst-Case Density. In: Proceedings of the Twenty-Eighth Annual
ACM-SIAM Symposium on Discrete Algorithms (SODA). 2017, pp. 99109.
[14] R. Peikert, D. Wurtz, M. Monagan, and C. de Groot. Packing circles
in a square: A review and new results. In: Proceedings of the 15th IFIP
Conference. 1992, pp. 4554.
[15] Eckard Specht. Packomania. 2015. url: http://www.packomania.com/.
[16] Kokichi Sugihara, Masayoshi Sawai, Hiroaki Sano, Deok-Soo Kim, and
Donguk Kim. Disk packing for the estimation of the size of a wire bundle.
In: Japan Journal of Industrial and Applied Mathematics 21(3) (2004),
pp. 259278.
[17] Peter Gabor Szabo, Mihaly Csaba Markot, Tibor Csendes, Eckard Specht,
Leocadio G. Casado, and Inmaculada Garca. New Approaches to Circle
Packing in a Square. Springer US, 2007.
[18] Huaiqing Wang, Wenqi Huang, Quan Zhang, and Dongming Xu. An
improved algorithm for the packing of unequal circles within a larger
containing circle. In: European Journal of Operational Research 141(2)
(Sept. 2002), pp. 440453. issn: 0377-2217.
[19] D. Wurtz, M. Monagan, and R. Peikert. The history of packing circles in
a square. In: Maple Technical Newsletter (1994), pp. 3542.
[20] Yinfeng Xu. On the minimum distance determined by n( 7) points in
an isoscele right triangle. In: Acta Mathematicae Applicatae Sinica 12(2)
(1996), pp. 169175.
 Fast and Compact Planar Embeddings

Leo Ferres1 , Jose Fuentes2 , Travis Gagie3 , Meng He4 , and Gonzalo Navarro2
1
Faculty of Engineering, Universidad del Desarrollo
2
CeBiB; Department of Computer Science, University of Chile
3
CeBiB; EIT, Diego Portales University
4
Faculty of Computer Science, Dalhousie University

Abstract. There are many representations of planar graphs but few are
as elegant as Turans (1984): it is simple and practical, uses only four
bits per edge, can handle multi-edges and can store any specied em-
bedding. Its main disadvantage has been that it does not allow ecient
searching (Jacobson, 1989). In this paper we show how to add a sublin-
ear number of bits to Turans representation such that it supports fast
navigation, thus overcoming this disadvantage. Other data structures for
planar embeddings may be asymptotically faster or smaller but ours is
simpler, and that can be a theoretical as well as a practical advantage:
e.g., we show how our structure can be built eciently in parallel.

1 Introduction

The rate at which we store data is increasing even faster than the speed and
capacity of computing hardware. Thus, if we want to use what we store eciently,
we need to represent it in better ways. The surge in the number and complexity of
the maps we want to have available on mobile devices is particularly pronounced
and has resulted in a bewildering number of ways to store planar graphs. Each of
these representations has its disadvantages, however: e.g., some do not support
fast navigation, some are large, some cannot represent multi-edges or certain
embeddings, and some are complicated to build in practice, especially in parallel,
which is a concern when dealing with massive datasets.
Tutte [26] showed that representing a specied embedding of a connected
planar multi-graph with n vertices and m edges takes m lg 12 3.58m bits in
the worst case. Turan [25] gave a very simple representation that uses 4m bits,
but Jacobson [15] noted that it does not allow fast searching and proposed

The second and fth authors received travel funding from EU grant H2020-MSCA-
RISE-2015 BIRDS GA No. 690941. The second, third and fth authors received
funding from Basal Funds FB0001, Conicyt, Chile. The third author received funding
from Academy of Finland grant 268324. Early parts of this work were done while the
third author was at the University of Helsinki and while the third and fth authors
were visiting the University of A Coruna. Many thanks to Jeremy Barbay, Luca
Castelli Aleardi, Arash Farzan, Ian Munro, Pat Nicholson and Julian Shun. The
third author is grateful to the late David Gregory for his course on graph theory.

Springer International Publishing AG 2017 385

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 385396, 2017.
DOI: 10.1007/978-3-319-62127-2_33
386 L. Ferres et al.

one that instead uses O(m) bits and supports fast navigation. Keeler and West-
brook [17] noted in turn that the constant factor in [Jacobsons] space bound
is relatively large and gave a representation that uses m lg 12 + O(1) bits when
the graph contains either no self-loops or no vertices with degree 1, but gave
up fast navigation again. Chiang, Lin and Lu [8] gave a representation that
uses 2m + 3n + o(m) bits with fast navigation, but it is based on orderly span-
ning trees; although all planar graphs can be represented with orderly spanning
trees, some planar embeddings cannot. Blelloch and Farzan [6] extended work by
Blandford et al. [5] and gave a representation that uses m lg 12 + o(m) bits with
fast navigation of any specied embedding, but it is complicated and has not
been implemented. Barbay et al. [3] gave a data structure that uses O(n) bits
to represent a simple planar graph on n nodes with fast navigation, but the hid-
den coecient is about 18. Other authors (see, e.g., [7, 13, 14]) have considered
special kinds of planar graphs, notably as tri-connected planar graphs and trian-
gulations. We refer the reader to Munro and Nicholsons [20] and Navarros [21,
Chapter 9] recent surveys for further discussion of compact data structures for
graphs.
In this paper we show how to add o(m) bits to Turans representation such
that it supports fast navigation: we can list the edges incident to any vertex in
counter-clockwise order using constant time per edge, and determine whether
two vertices are neighbours or nd a vertexs degree in O(f (m))-time for any
given function f (m) (1). Our data structure is faster, smaller or more ex-
pressive than any of the structures listed above except Blelloch and Farzans,
and it is much simpler than theirs. Our structures simplicity is a theoretical
as well as a practical advantage, in that we can build it in parallel with linear
work and logarithmic span (albeit without support for fast neighbour and de-
gree queries). We summarize our construction algorithm in this paper and will
provide details in a subsequent paper. In contrast, we do not have such ecient
parallel algorithms for nding the book embeddings [27], orderly spanning trees
and triangulations of planar subdivisions required by, respectively, Jacobsons,
Chiang et al.s and Barbay et al.s constructions. Blandford et al.s and Blelloch
and Farzans constructions are based on nding small vertex separators [19] and,
although Kao et al. [16] designed a linear-work and logarithmic-span algorithm
for computing a cycle separator of a planar graph, both Blandford et al.s and
Blelloch and Farzans constructions decompose the input graph by repeatedly
computing separators until each piece is suciently small, which increases the
total work to O(n log n) even when this optimal parallel algorithm is used.
Turan chooses an arbitrary spanning tree of the graph, roots it at a vertex
on the outer face and traverses it, writing its balanced-parentheses representa-
tion as he goes and interleaving that with a sequence over a dierent binary
alphabet, consisting of an occurrence of one character for the rst time he sees
each edge not in the tree and an occurrence of the other character for the second
time he sees that edge. These two sequences can be written as three sequences
over {0, 1}: one of length 2n 2 encoding the balanced-parentheses representa-
tion of the tree; one of length 2m 2n + 2 encoding the interleaved sequence;
Fast and Compact Planar Embeddings 387

and one of length 2m indicating how they are interleaved. Our extension of his
representation is based on the observation that the interleaved sequence encodes
the balanced-parentheses representation of the complementary spanning tree of
the dual of the graph. By adding a sublinear number of bits to each balanced-
parentheses representation, we can support fast navigation in the trees, and by
storing the sequence indicating the interleaving as a bitvector, we can support
fast navigation in the graph.
In Section 2 we briey describe bitvectors and the balanced-parentheses rep-
resentation of trees, which are the building blocks of our extension of Turans
representation. For further discussion of these data structures, we again direct
the reader to Navarros text [21]. In Section 3 we prove the observation men-
tioned above. In Section 4 we describe our data structure and how we implement
queries. We summarize our parallel construction algorithm in Section 5 and re-
port the results of our preliminary experiments.

2 Preliminaries
A bitvector is a binary string that supports the queries rank and select in addition
to random access, where rankb (i) returns the number of bits set to b in the
prex of length  of the string and selectb (j) returns the position of the jth bit
set to b. For convenience, we dene selectb (0) = 0. There are many dierent
implementations that represent a bitvector of length n in n + o(n) bits and
support random access, rank and select in constant time.
With bitvectors we can represent an ordered tree or forest on n vertices using
2n + o(n) bits and support natural navigation queries quickly. One of the most
popular such representations is as a string of balanced parentheses: we traverse
each tree from left to right, writing an opening parenthesis when we rst visit
a vertex (starting at the root) and a closing parenthesis when we leave it for
the last time (or, in the case of the root, when we nish the traversal). We
can encode the string of parentheses as a bitvector, with 0s encoding opening
parentheses and 1s encoding closing parentheses, and achieve the space bound
stated above while supporting each of the follow queries used by our solution in
constant time:
match(i), locates the parenthesis matching the ith parenthesis,
parent(v), returns the parent of v, given as its pre-order rank in the traversal,
or 0 if v is the root of its tree.

3 Spanning trees of planar graphs

It is well known that for any spanning tree T of a connected planar graph G,
the edges dual to T are a spanning tree T of the dual of G, with T and T
interdigitating; see Figure 1 for an illustration (including multi-edges and a self-
loop) and, e.g., [4, 11, 22] for discussions. If we choose T as the spanning tree of
G for Turans representation, then we store a 0 and a 1, in that order, for each
edge in T . We now show that these bits encode a traversal of T .
388 L. Ferres et al.

T GT T

G
4
1 (1, 3) (A, B)
8
C 2 (1, 2)
7 E 6
3 (2, 3)
4 (1, 3) (A, B)
A
5 (2, 3)
F
5
D
6 (2, 4)
7 (4, 8) (A, C)
H
1
B
2
8 (2, 4)
3
9 (2, 6) (C, D)
10 (1, 2)
11 (1, 5)
12 (5, 6)
13 (2, 6) (C, D)
1 14 (6, 8) (C, E)
15 (5, 6)
2 5 7
16 (5, 7) (E, F)
3 4 6 8 17 (1, 5)
18 (1, 7)
19 (5, 7) (E, F)
A
20 (7, 8)
H G C B 21 (6, 8) (C, E)
22 (4, 8) (A, C)
E D 23 (7, 8) (A, G)
24 (7, 8)
F
25 (7, 8) (A, G)
26 (1, 7)
27 (1, 1) (A, H)
28 (1, 1) (A, H)
Fig. 1. Top left: A planar embedding of a planar graph G, with a spanning tree T
of G shown in red and the complementary spanning tree T of the dual of G shown
in blue with dashed lines. Bottom left: The two spanning trees, with T rooted at
the vertex 1 on the outer face and T rooted at the vertex A corresponding to the
outer face. Right: The list of edges we process while traversing T starting at 1 and
processing edges in counter-clockwise order, with the edges in T shown in red and the
ones in G T shown in black; the edges of T corresponding to the edges in G T are
shown in blue.
Fast and Compact Planar Embeddings 389

Lemma 1. Consider any planar embedding of a planar graph G, any spanning

tree T of G and the complementary spanning tree T of the dual of G. If we
perform a depth-rst traversal of T starting from any vertex on the outer face
of G and always process the edges (of the graph) incident to the vertex v we are
visiting in counter-clockwise order (starting from the edge immediately after the
one to vs parent or, if v is the root of T , from immediately after any incidence
of the outer face), then each edge not in T corresponds to the next edge we cross
in a depth-rst traversal of T .
Proof. Suppose the traversal of T starts at the vertex of the dual of G corre-
sponding to the outer face of G. We now prove by induction that the vertex we
are visiting in T always corresponds to the face of G incident to the vertex we
are visiting in T and to the previous and next edges in counter-clockwise order.
Our claim is true before we process any edges, since we order the edges
starting from an incidence of the outer face to the root of T . Assume it is still
true after we have processed i < m edges, and that at this time we are visiting
v in T and v in T . First suppose that the (i + 1)st edge (v, w) we process is in
T . We note that w = v, since otherwise (v, w) could not be in T . We cross from
v to w in T , which is also incident to the face corresponding to v . Now (v, w)
is the previous edge considering their counter-clockwise order at w, starting
from (v, w) and the next edge (which is (v, w) again if w has degree 1) is also
incident to v . This is illustrated on the left side of Figure 2. In fact, the next
edge is the one after (v, w) in a clockwise traversal of the edges incident to the
face corresponding to v .
Now suppose (v, w) is not in T and let w be the vertex in T corresponding
to the face on the opposite side of (v, w), which is also incident to v. We note
that w = v , since otherwise (v, w) would have to be in T . We cross from v
to w in T . Now (v, w) is the previous edge this time still considering their
counter-clockwise order at v and the next edge (which may be (v, w) again
if it is a self-loop) is also incident to w . This is illustrated on the right side
of Figure 2. In fact, the next edge is the one that follows (v, w) in a clockwise
traversal of the edges incident to the face corresponding to w .
Since our claim remains true in both cases after we have processed i + 1
edges, by induction it is always true. In other words, whenever we should process
next an edge e in G that is not in T , we are visiting in T one of the vertices
corresponding to the faces incident to e (i.e., one of the endpoints of the edge
in the dual of G that corresponds to e). Since we process each edge in G twice,
once at each of its endpoints or twice at its unique endpoint if it is a self-loop,
it follows that the list of edges we process that are not in T , corresponds to the
list of edges we cross in a traversal of T . 

We process the edges in counter-clockwise order so that the traversals of T
and T are from left to right and from right to left, respectively; processing
them in clockwise order would reverse those directions. For example, for the
embedding in Figure 1, if we start the traversal of the red tree T at vertex 1 and
start processing the edges at (1, 3), then we process them in the order shown at
the right of the gure.
390 L. Ferres et al.

y
w
v w v w

v v
u x u

Fig. 2. Left: If we process an edge (v, w) in T , then we move to w in our traversal of

T and the next edge, (w, x) in this case, is also incident to the vertex v we are visiting
in our traversal of T . Right: If (v, w) is not in T , then in T we move from v to the
vertex w corresponding to the face on the opposite side of (v, w) in G. The next edge,
(v, y) in this case, is also incident to w .

4 Data Structure
Our extension of Turans representation of a planar embedding of a connected
planar graph G with n vertices and m edges consists of the following components,
which take 4m + o(m) bits:
a bitvector A[1..2m] in which A[i] indicates whether the ith edge we process
in the traversal of T described in Lemma 1, is in T ;
a bitvector B[1..2(n 1)] in which B[i] indicates whether the ith time we
process an edge in T during the traversal, is the second time we process that
edge;
a bitvector B [1..2(m n + 1)] in which B [i] indicates whether the ith time
we process an edge not in T during the traversal, is the second time we
process that edge.
Notice B encodes the balanced-parentheses representation of T except that it
lacks the leading 0 and trailing 1 encoding the parentheses for the root. By
Lemma 1, B encodes the balanced-parentheses representation of a traversal of
the spanning tree T of the dual of G complementary to T (the right-to-left
traversal of T , in fact) except that it also lacks the leading 0 and trailing 1
encoding the parentheses for the root. Therefore, since B and B encode forests,
we can support match and parent with them.
To build A, B and B given the embedding of G and T , we traverse T as
in Lemma 1. Whenever we process an edge, if it is in T then we append a 1 to
A and append the edge to a list L; otherwise, we append a 0 to A and append
the edge to another list L . When we have nished the traversal, we replace
each edge in L or L by a 0 if it is the rst occurrence of that edge in that list,
and by a 1 if it is the second occurrence; this turns L and L into B and B ,
respectively. For the example shown in Figure 1, L and L eventually contain
the edges shown in the columns labelled T and G T , respectively, in the table
on the on the right side of the gure, and

A[1..28] = 0110110101110010110100010100
B[1..14] = 00101100110011
B [1..14] = 01001001110101 .
Fast and Compact Planar Embeddings 391

We identify each vertex v in G by its pre-order rank in our traversal of T .

Consider the following queries:
rst(v): return i such that the rst edge we process while visiting v is the ith
we process during our traversal;
next(i): return j such that if we are visiting v when we process the ith edge
during our traversal, then the next edge incident to v in counter-clockwise
order is the one we process jth;
mate(i): return j such that we process the same edge ith and jth during our
traversal;
vertex(i): return the vertex v such that we are visiting v when we process the
ith edge during our traversal.
With these it is straightforward to reenact our traversal of T and recover the
embedding of G. For example, with the following queries we can list the edges
incident to the root of T in Figure 1 and determine whether they are in T :

rst(1) = 1 mate(1) = 4 vertex(4) = 3 A[1] = 0

next(1) = 2 mate(2) = 10 vertex(10) = 2 A[2] = 1
next(2) = 11 mate(11) = 17 vertex(17) = 5 A[11] = 1
next(11) = 18 mate(18) = 26 vertex(26) = 7 A[18] = 1 .

To see why we can recover the embedding from the traversal, consider that if
we have already correctly embedded the rst i edges processed in the traversal,
then we can embed the (i + 1)st correctly given its endpoints and its rank in the
counter-clockwise order at those vertices.
We now explain our constant-time implementations of rst, next, mate and
vertex. If m = 0 then rst(v) is undened, which we indicate by returning 0.
Otherwise, we rst process an edge at v immediately after rst arriving at v.
Since we identify v with its pre-order rank in our traversal of T and B lacks the
opening parenthesis for the root, while rst arriving at any vertex v other than
the root we write the (v 1)st 0 in B and, thus, the B.select0 (v 1)st 1 in A.
If v is the root then rst(v) = 1 and so, since selectx (0) = 0, this case is also
handled by the formula below:

A.select1 (B.select0 (v 1)) + 1 if m 1
rst(v) =
0 otherwise.

In our example,

rst(5) = A.select1 (B.select0 (4)) + 1 = A.select1 (7) + 1 = 12

and indeed the twelfth edge we process, (5, 6), is the rst one we process at
vertex 5.
If the ith edge we process is the last edge we process at a vertex v then
next(i) is undened, which we again indicate by returning 0. This is the case
when i = 2m, or A[i] = 1 and B[A.rank1 (i)] = 1. Otherwise, if the ith edge we
process is not in T , then A[i] = 0, and we process the next edge at v one time
392 L. Ferres et al.

step later. Finally, if the ith edge e we process is in T and not the last one we
process at v, then we next process an edge at v immediately after returning to
v by processing e again at time mate(i). This is the case when A[i] = 1 and
B[A.rank1 (i)] = 0. In other words,

i + 1 if A[i] = 0 and i < 2m
next(i) = mate(i) + 1 if A[i] = 1 and B[A.rank1 (i)] = 0

0 otherwise.
In our example, since A[12] = 1, B[A.rank1 (12)] = B[8] = 0, the twelfth edge we
process is (5, 6) and it is also the fteenth edge we process,
next(12) = mate(12) + 1 = 16 ,
and indeed the second edge we process at vertex 5 is (5, 7).
To implement mate(i), we check A[i] and use rank to determine whether we
wrote a bit in B or in B while processing the ith edge, and to nd that bit. We
use match to nd the bit encoding the matching parenthesis, and then use select
on A to nd the bit we wrote in A when we wrote that matching bit. Therefore,

A.select0 (B .match(A.rank0 (i))) if A[i] = 0
mate(i) =
A.select1 (B.match(A.rank1 (i))) otherwise.
To compute mate(12) for our example, since A[12] = 1,
mate(12)
= A.select1 (B.match(A.rank1 (12)))
= A.select1 (B.match(8))
= A.select1 (9)
= 15 .
Suppose the ith edge e we process is not in T and we process it at vertex v.
If the preceding time we processed an edge in T was the rst time we processed
that edge, we then wrote a 0 in B, encoding the opening parenthesis for v;
otherwise, we then wrote a 1 in B, encoding the closing parenthesis for one of
vs children. Now suppose e is in T . If that is the rst time we process e, we
move to the other endpoint w of e which is a child of v and write a 0 in
B, encoding the opening parenthesis for w. If it is the second time we process e,
then we write a 1 in B, encoding the closing parenthesis for v itself. Therefore,


B.rank0 (A.rank1 (i)) + 1



if A[i] = 0 and B[A.rank1 (i)] = 0



B.parent(B.rank0 (B.match(A.rank1 (i)))) + 1


if A[i] = 0 and B[A.rank1 (i)] = 1
vertex(i) =

B.parent(B.rank0 (A.rank1 (i))) + 1


if A[i] = 1 and B[A.rank1 (i)] = 0






B.rank0 (B.match(A.rank1 (i))) + 1

otherwise.
Fast and Compact Planar Embeddings 393

In our example, since A[16] = 0 and B[A.rank1 (16)] = B[9] = 1,

vertex(16)
= B.parent(B.rank0 (B.match(A.rank1 (16)))) + 1
= B.parent(B.rank0 (B.match(9))) + 1
= B.parent(B.rank0 (8)) + 1
= B.parent(5) + 1
= 5,

and indeed we process the sixteenth edge (5, 7) while visiting 5.

We remind the reader that since B lacks parentheses for the root of T ,
B.parent(5) refers to the parent of the fth vertex in an in-order traversal of
T not including the root, i.e., the parent vertex 5 of vertex 6. Adding 1 includes
the root in the traversal, so the nal answer correctly refers to vertex 5. The
lack of parentheses for the root also means that, e.g., B.parent(4) refers to the
parent of vertex 5 and returns 0 because vertex 5 is the root of its own tree in
the forest encoded by B, without vertex 1. Adding 1 to that 0 also correctly
turns the nal value into 1, the in-order rank of the root. Of course, we have the
option of prepending and appending bits to A, B and B to represent the roots
of T and T , but that slightly confuses the relationship between the positions of
the bits and the time steps at which we process edges.
Clearly we can determine whether two vertices u and v are neighbours by
listing the neighbours of each in parallel in O(min(degree(u), degree(v))) time,
and we can nd degree(v) in O(degree(v)) time. Moreover, given a function
f (m) (1), we can make both kinds of queries take O(f (m)) time. To do
this, we store a bitvector marking the O(m/f (m)) = o(m) vertices with degree
at least f (m), which takes o(m) bits. To be able to answer neighbour queries
quickly, we consider the graph induced by those high-degree vertices and elimi-
nate multi-edges and self-loops. The resulting simple graph G is still planar
so it has average degree less than 6 and thus o(m) edges and preserves the
neighbour relation between those vertices. We can store G using Blelloch and
Farzans representation or, since the neighbour relation does not depend on the
embedding, using one of the other compact representations of planar graphs that
supports constant-time neighbour queries, which also takes o(m) bits. To answer
neighbour(u, v) now, we check whether either u or v is low-degree and, if so, list
its neighbours in O(f (m)) time; if not, we query our auxiliary representation
in O(1) time. To be able to answer degree queries quickly, we simply store the
degrees of the high-degree vertices in unary using a bitvector, which takes o(m)
bits. To nd degree(v) now, we check whether v is low-degree and, if so, list and
count its incident edges; if not, we look up degree(v).
Summarizing our results so far, we have the following theorem:

Theorem 1. We can store a given planar embedding of a connected planar

graph G with m edges in 4m + o(m) bits such that later, given a vertex v, we can
list the edges incident to v in counter-clockwise order (optionally, starting at a
394 L. Ferres et al.

given edge e incident to v) using constant time per edge, and determine whether
two vertices are neighbours or nd a vertexs degree in O(f (m))-time for any
given function f (m) (1).

5 Parallel Construction and Experiments

Due to space constraints, in this section we can only summarize our parallel
algorithm and then briey report the result of our experiments on construc-
tion and query times. We will provide the full details of the algorithm in the
subsequent paper mentioned in Section 1, of which a preprint is available at
http://arxiv.org/abs/1705.00415 .
We construct our extension of Turans representation in parallel as follows:
given a planar graph with a planar embedding G, we rst compute a spanning
tree T of G in parallel. In our experiments we used Bader and Congs algo-
rithm [2] because it works well in practice, but its theoretical bounds are for
random graphs. To obtain good worst-case bounds, we could use Shiloach and
Vishkins [23] or Awerbuch and Shiloachs [1] algorithms, which use linear work
with logarithmic span in the CRCW PRAM model. We recently learned that
Shun, Dhulipala and Blellochs [24] practical connectivity algorithm can be made
to return a spanning tree with linear work and polylogarithmic span. As a by-
product of the computation of T , we obtain an array C of length 2n 2 that
stores the number of edges of G \ T between two consecutive edges in T , in
counter-clockwise order. Notice the starting vertex for the spanning tree must
be in the outer face of G.
We construct bitvectors A, B and B by performing a parallel Euler Tour
over T [9]. During the tour, we obtain B by writing a 0 for each forward (parent-
to-child) edge and a 1 for each backward (child-to-parent) edge. We obtain A
by counting the number of edges of G \ T between two consecutive edges of T
(stored in C). We represent the former with 0s and the edges of T with 1s.
The visiting order of edges of G \ T encoded in B is implicit in the previous
Euler Tour. Therefore, with the Euler Tour and the array C, we have enough
information to compute the position of each bit in the bitvector B . We can
decide if an edge of G \ T is a forward or backward edge by checking its relative
position and the position of its complement edge on the Euler tour. Finally, in
order to support operations on A, B and B , we used Labeit et al.s algorithm
for succinct bitvectors [18], and Ferres et al.s algorithm for succinct trees [12].
We analyze our algorithm (after the computation of the spanning tree) in the
Dynamic Multithreading (DyM) model of parallel computation [10]. The DyM
model relies on two parameters: the work T1 , i.e., the running time on a single
core; and the span T , i.e., the complexity of the intrinsically sequential part of
the parallel computation. The time Tp needed to execute the computation on p
cores is bounded by Tp = (T1 /p + T ). The Euler tour and the array C can be
computed in parallel in T1 = O(n) and T = O(lg n) time. Assigning values to A
and B can be done independently for each entry of the bitvectors, which gives us
T1 = O(n) and T = O(1) time, while B takes T1 = O(m n) and T = O(1)
Fast and Compact Planar Embeddings 395

time. Rank/select structures can be constructed in parallel with T1 = O(m)

and T = O(lg m) time. Succinct trees can be constructed in T1 = O(m) and
T = O(lg m) time.

Theorem 2. Given a spanning tree of a planar embedding, the compact repre-

sentation from Theorem 1 (without the auxiliary data structures for fast neighbour
and degree queries) can be constructed in parallel with linear work and logarith-
mic span.

To provide some grounds for comparison, we also implemented a sequential

algorithm based on our parallel one, but instead of using Bader and Congs
algorithm, we used depth-rst search to compute T . This sequential implemen-
tation of Turans representation is simpler than our parallel algorithm running
on a single core, avoiding the additional steps needed for the parallel computa-
tion. To test both the sequential and parallel implementations, we synthetically
generated a planar graph (represented as an adjacency list) by computing the
Delaunay Triangulation of 25,000,000 random coordinates, yielding 25,000,000
vertices and 74,999,979 edges, with a minimum degree of 3 and a maximum
degree of 15. The experiments were carried out on a 28-core machine (two pro-
cessors with 14 physical cores each) with hyperthreading turned on (for a total
of 56 cores), per-core L1 and L2 caches of sizes 64KB and 256KB, respectively
and a per-processor shared L3 cache of 35MB and a total of 768GB DDR3 RAM
memory.
Results show that the sequential algorithm took 71.8 seconds to construct
the representation of Theorem 1, while the parallel implementation took 5.4
seconds with 28 threads and 3.3 seconds with 56 threads. The space used by the
adjacency list representation was 1.02 GB, 117.3 bits per edge. The space used
by our compact representation was 44.7 MB, 5 bits per edge, which matches
Theorem 1. Memory consumption of our parallel algorithm peaked at 1.4 GB.
With respect to queries, we tested counting (number of neighbors) and
listing (list of neighbors in counter-clockwise order) queries. For the former, the
adjacency-list representation took 0.047 microseconds per node and the compact
representation took 4.6 microseconds per node. For listing, the adjacency-list
representation took 0.046 microseconds per node listed and the compact repre-
sentation took 3.69 microseconds per node listed.
In summary, our parallel algorithm achieves a reasonable speed up in terms
of the number of threads and is an already order of magnitude faster than the se-
quential algorithm when using 28 threads; our compact representation is between
one and two orders of magnitude smaller than the adjacency-list representation,
but takes two orders of magnitude more time to answer queries. Our code and
datasets are available at https://users.dcc.uchile.cl/~jfuentess/pemb/ .

References
1. B. Awerbuch and Y. Shiloach. New connectivity and MSF algorithms for shue-
exchange network and PRAM. IEEE Trans. Computers, 36(10):12581263, 1987.
396 L. Ferres et al.

2. D. A. Bader and G. Cong. A fast, parallel spanning tree algorithm for symmetric
multiprocessors (SMPs). J. Parallel and Distributed Computing, 65:9941006, 2005.
3. J. Barbay, L. C. Aleardi, M. He, and J. I. Munro. Succinct representation of labeled
graphs. Algorithmica, 62:224257, 2012.
4. N. Biggs. Spanning trees of dual graphs. J. Comb. Theory, Series B, 11:127131,
1971.
5. D. K. Blandford, G. E. Blelloch, and I. A. Kash. Compact representations of
separable graphs. In SODA, pages 679688, 2003.
6. G. E. Blelloch and A. Farzan. Succinct representations of separable graphs. In
CPM, pages 138150, 2010.
7. Castelli Aleardi, L, O. Devillers, and G. Schaeer. Succinct representations of
planar maps. TCS, 408:174187, 2008.
8. Y.-T. Chiang, C.-C. Lin, and H.-I. Lu. Orderly spanning trees with applications.
SIAM J. Comp., 34:924945, 2005.
9. G. Cong and D. A. Bader. The Euler tour technique and parallel rooted spanning
tree. In ICPP, pages 448457, 2004.
10. T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein. Multithreaded algo-
rithms. In Introduction to Algorithms, pages 772812. MIT Press, 2009.
11. D. Eppstein. Dynamic generators of topologically embedded graphs. In SODA,
pages 599608, 2003.
12. L. Ferres, J. Fuentes-Sepulveda, M. He, and N. Zeh. Parallel construction of suc-
cinct trees. In SEA, pages 314, 2015.
13. E. Fusy, G. Schaeer, and D. Poulalhon. Dissections, orientations, and trees with
applications to optimal mesh encoding and random sampling. TALG, 4:19, 2008.
14. X. He, M. Kao, and H. Lu. Linear-time succinct encodings of planar graphs via
canonical orderings. SIAM J. Discrete Math., 12:317325, 1999.
15. G. Jacobson. Space-ecient static trees and graphs. In FOCS, pages 549554,
1989.
16. M. Kao, S. Teng, and K. Toyama. An optimal parallel algorithm for planar cycle
separators. Algorithmica, 14:398408, 1995.
17. K. Keeler and J. Westbrook. Short encodings of planar graphs and maps. DAM,
58:239252, 1995.
18. J. Labeit, J. Shun, and G. E. Blelloch. Parallel lightweight wavelet tree, sux
array and FM-index construction. In DCC, pages 3342, 2016.
19. R. J. Lipton and R. E. Tarjan. A separator theorem for planar graphs. SIAM J.
Applied Math., 36:177189, 1979.
20. J. I. Munro and P. K. Nicholson. Compressed representations of graphs. In Ency-
clopedia of Algorithms, pages 382386. Springer, 2016.
21. G. Navarro. Compact Data Structures: A Practical Approach. Cambridge Univer-
sity Press, 2016.
22. T. R. Riley and W. P. Thurston. The absence of ecient dual pairs of spanning
trees in planar graphs. Electronic J. Comb., 13, 2006.
23. Y. Shiloach and U. Vishkin. An o(log n) parallel connectivity algorithm. J. Algo-
rithms, 3(1):5767, 1982.
24. J. Shun, L. Dhulipala, and G. E. Blelloch. A simple and practical linear-work
parallel algorithm for connectivity. In SPAA, pages 143153, 2014.
25. G. Turan. On the succinct representation of graphs. DAM, 8:289294, 1984.
26. W. T. Tutte. A census of planar maps. Canadian J. Math., 15:249271, 1963.
27. M. Yannakakis. Embedding planar graphs in four pages. JCSS, 38:3667, 1989.
 When Can Graph Hyperbolicity be Computed
in Linear Time?

Till Fluschnik1, , Christian Komusiewicz2,   , George B. Mertzios3 ,

Andre Nichterlein1,3, , Rolf Niedermeier1 , and Nimrod Talmon4,
1
Institut fur Softwaretechnik und Theoretische Informatik, TU Berlin, Germany,
{till.fluschnik, andre.nichterlein, rolf.niedermeier}@tu-berlin.de
2
Institut fur Informatik, Friedrich-Schiller-Universitat Jena, Germany,
[email protected]
3
School of Engineering and Computing Sciences, Durham University, UK,
[email protected]
4
Weizmann Institute of Science, Rehovot, Israel,
[email protected]

Abstract. Hyperbolicity measures, in terms of (distance) metrics, how

close a given graph is to being a tree. Due to its relevance in model-
ing real-world networks, hyperbolicity has seen intensive research over
the last years. Unfortunately, the best known practical algorithms for
computing the hyperbolicity number of a n-vertex graph have running
time O(n4 ). Exploiting the framework of parameterized complexity anal-
ysis, we explore possibilities for linear-time FPT algorithms to com-
pute hyperbolicity. For instance, we show that hyperbolicity can be com-
puted in time 2O(k) + O(n + m) (m being the number of graph edges,
k being the size of a vertex cover) while at the same time, unless the
SETH fails, there is no 2o(k) n2 -time algorithm.

1 Introduction

(Gromov) hyperbolicity [16] of a graph is a popular attempt to capture and

measure how metrically close a graph is to being a tree. The study of hyperbol-
icity is motivated by the fact that many real-world graphs are tree-like from a
distance metric point of view [2, 3]. This is due to the fact that many of these
graphs (including Internet application networks or social networks) possess cer-
tain geometric and topological characteristics. Hence, for many applications (cf.,
e.g. [3]), including the design of (more) ecient algorithms, it is useful to know

This work was initiated at the yearly research retreat of the Algorithmics and Com-
putational Complexity (AKT) group of TU Berlin, held in in Krolpa, Thuringia,
Germany, from April 3rd till April 9th, 2016.

Supported by the DFG, project DAMM (NI 369/13-2).

Supported by the DFG, project MAGZ (KO 3669/4-1).

Supported by a postdoctoral fellowship of the DAAD while at Durham University.

Supported by a postdoctoral fellowship from I-CORE ALGO.

Springer International Publishing AG 2017 397

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 397408, 2017.
DOI: 10.1007/978-3-319-62127-2_34
398 T. Fluschnik et al.

the hyperbolicity of a graph. The hyperbolicity of a graph is a nonnegative num-

ber ; the smaller is, the more tree-like the graph is; in particular, = 0 means
that the graph metric indeed is a tree metric. Typical hyperbolicity values for
real-world graphs are below 5 [2].
Hyperbolicity can be dened via a four-point condition: Considering a size-
four subset {a, b, c, d} of the vertex set of a graph, one takes the (nonnegative)
dierence between the two largest of the three sums ab + cd, ac + bd, and ad + bc,
where, e.g., ab denotes the length of the shortest path between vertices a and b
in the given graph. The hyperbolicity is the maximum of these dierences over
all size-four subsets of the vertex set of the graph. For an n-vertex graph, this
characterization of hyperbolicity directly implies a simple (brute-force) O(n4 )-
time algorithm to compute its hyperbolicity. It has been observed that this
running time is too slow for computing the hyperbolicity of large graphs as
occurring in applications [2, 3, 4, 13]. On the theoretical side, it was shown
that relying on some (rather impractical) matrix multiplication results, one can
improve the upper bound to O(n3.69 ) [13]. Moreover, roughly quadratic lower
bounds are known [4, 13]. In practice, however, the best known algorithm still has
an O(n4 )-time worst-case bound but uses several clever tricks when compared to
the straightforward brute-force algorithm [3]. Indeed, based on empirical studies
an O(mn) running time is claimed, where m is the number of edges in the
graph. Furthermore, there are heuristics for computing the hyperbolicity of a
given graph [7].
To explore the possibility of faster algorithms for hyperbolicity in relevant
special cases is the guiding principle of this work. More specically, introducing
some graph parameters, we investigate whether one can compute hyperbolicity
in linear time when these parameters take small values. In other words, we em-
ploy the framework of parameterized complexity analysis (so far mainly used for
studying NP-hard problems) applied to the polynomial-time solvable hyperbolic-
ity problem. In this sense, we follow the recent trend of studying FPT in P [15].
Indeed, other than for NP-hard problems (where parameterized complexity is
typically applied), for some parameters we achieve not only exponential depen-
dence on the parameter but also polynomial ones. Note that such algorithms are
unlikely for metric parameters like diameter or hyperbolicity.

Our contributions. Table 1 summarizes our main results. On the positive side,
for a number of natural graph parameters we can attain linear FPT running
times. Our positive graph parameters here are the following:
the covering path number, that is, the minimum number of paths where only
the endpoints have degree greater than two and which cover all vertices;
the feedback edge number, that is, the minimum number of edges to delete
to obtain a forest;
the number of graph vertices of degree at least three;
the vertex cover number, that is the minimum number of vertices needed to
cover all edges in the graph;
When Can Graph Hyperbolicity be Computed in Linear Time? 399

Table 1. Summary of our algorithmic results. Herein, k denotes the parameter and n
and m denote the number of vertices and edges, respectively.

Parameter Running time

covering path number O(k4 (n + m)) [Theorem 5]
feedback edge number O(k4 (n + m)) [Theorem 6]
number of 3-degree vertices O(k8 (n + m)) [Theorem 8]
O(k)
vertex cover number 2 + O(n + m) [Theorem 10]
distance to cographs O(44k k7 (n + m)) [Theorem 15]

the distance to cographs, that is, the minimum number of vertices to delete
to obtain a cograph.5
On the negative side we prove that, with respect to the parameter vertex cover
number k, we cannot hope for any 2o(k) n2 algorithm unless the Strong Ex-
ponential Time Hypothesis (SETH) fails. We also obtain a quadratic-time
FPT lower bound with respect to the parameter maximum vertex degree, again
assuming SETH. Finally, we show that computing the hyperbolicity is at least
as hard as computing a size-four independent set in a graph. It is conjectured
that computing size-four independent sets needs (n3 ) time [20]. Due to lack of
space, many details and proofs (marked with ()) had to be deferred.6

2 Preliminaries and Basic Observations

We write [n] := {1, . . . , n} for every n N. For a function f : X Y and X 
X we set f (X  ) := {y Y | x X  : f (x) = y}.

Graph theory. Let G = (V, E) be a graph. We dene |G| = |V |+|E|. For W V ,

we denote by G[W ] the graph induced by W . We use G W := G[V \ W ] to
denote the graph obtained from G by deleting the vertices of W V . A path
P = (v1 , . . . , vk ) in G is a tuple of distinct vertices in V such that {vi , vi+1 } E
for all i [k 1]; we say that such a path P has endpoints v1 and vk , we call
the other vertices of P inner vertices, and we say that P is a v1 -vk path. We
denote by ab the length of a shortest a-b path if such a path exists; otherwise,
that is, if a and b are in dierent connected components, we dene ab := .
Let P = (v1 , . . . , vk ) be a path and vi , vj two vertices on P . We denote by vi vj |P
the distance of vi to vj on the path P , that is, vi vj |P = |j i|. For a graph G
we denote by VG3 the set of vertices of G that have degree at least three.

Hyperbolicity. Let G = (V, E) be graph and a, b, c, d V . We dene D1 := ab +

cd, D2 := ac+bd, and D3 := ad+bc (referred to as distance sums). Moreover, we
5
Cographs are the graphs without induced P4 s. Distance to cographs is upper-
bounded by the parameter distance to cluster graph [10] and thus also by the pa-
rameter vertex cover number.
6
A full version is available at https://arxiv.org/abs/1702.06503.
400 T. Fluschnik et al.

dene (a, b, c, d) := |Di Dj | if Dk min{Di , Dj }, for pairwise distinct i, j, k

{1, 2, 3}. If any two vertices of the quadruple {a, b, c, d} are not connected, we
set (a, b, c, d) = 0.7 The hyperbolicity of G = (V, E) is dened as (G) :=
maxa,b,c,dV {(a, b, c, d)}. Note that by our denition, if G is not connected,
(G) computes the maximal hyperbolicity over all connected components of G.
We say that the graph is -hyperbolic for some N if it has hyperbolicity at
most . That is, a graph is -hyperbolic8 . if for each 4-tuple a, b, c, d V we
have
ab + cd max{ac + bd, ad + bc} + .
Formally, the Hyperbolicity problem is dened as follows.
Hyperbolicity
Input: An undirected graph G = (V, E) and a positive integer .
Question: Is G -hyperbolic?

The following lemma will be useful later. For any quadruple {a, b, c, d},
Lemma 1 upper bounds (a, b, c, d) by twice the distance between any pair of
vertices of the quadruple.
Lemma 1 ([7, Lemma 3.1]). (a, b, c, d) 2 minu
=v{a,b,c,d} {uv}

Reduction Rule 1. As long as there are more than four vertices, remove ver-
tices of degree one.

Lemma 2 (). Reduction Rule 1 is correct and can be exhaustively applied in

linear time.

3 Polynomial Linear-Time Parameterized Algorithms

In this section, we provide polynomial linear-time parameterized algorithms with
respect to the parameters feedback edge number and number of vertices with
degree at least three; that is, we present algorithms with running time having
a linear-time dependence on the input size times a polynomial-time dependence
on the parameter value (to which we refer to as PL-FPT running time).
To this end, we rst introduce an auxiliary parameter, the minimum max-
imal path cover number, which we formally dene below and also describe a
polynomial linear-time parameterized algorithm for it.
Building upon this result, for the parameter feedback edge number we then
show that, after applying Reduction Rule 1, the number of maximal paths can
be upper-bounded by a polynomial of the feedback edge number. This implies
a polynomial linear-time parameterized algorithm for the feedback edge number
as well. For the parameter number of vertices with degree at least three, we
7
This case is often left undened in the literature. Our denition however enables to
consider also disconnected graphs.
8
Note that there is also a slightly dierent denition where graphs we call -hyperbolic
are called 2-hyperbolic [7, 17]; we follow the denition of Brinkmann et al. [6].
When Can Graph Hyperbolicity be Computed in Linear Time? 401

introduce an additional reduction rule to achieve that the number of maximal

paths is upper-bounded in a polynomial of this parameter. Again, this implies
an algorithm with PL-FPT running time.

Minimum maximal path cover number.

Denition 3 (Maximal path). Let G be a graph and P be a path in G. Then,
P is a maximal path if the following holds: (1) P contains at least two vertices;
(2) all its inner vertices have degree two in G; (3) either both its endpoints have
degree at least three in G, or one of its endpoints has degree at least three in G
while the other endpoint is of degree two in G; and (4) P is size-wise maximal
with respect to these properties.
We will be interested in the minimum number of maximal paths needed to
cover the vertices of a given graph; we call this number the minimum maximal
path cover number. While not all graphs can be covered by maximal paths (e.g.,
edgeless graphs), graphs which have minimum degree two and contain no isolated
cycles, i.e. components that form induced cycles, can be covered by maximal
paths (this follows by, e.g., a greedy algorithm which iteratively starts a path
with an arbitrary uncovered vertex and exhaustively extends it arbitrarily; since
there are no isolated cycles and the minimum degree is two, we are bound to
eventually hit at least one vertex of degree three). Based on the approximation
algorithm given in the next lemma, we assume in the following that we are given
a maximal path cover.

Lemma 4 (). There is a linear-time 2-approximation algorithm for the min-

imum maximal path cover number for graphs which have minimum degree two
and contain no isolated cycles.

Now we are ready to design a polynomial linear-time parameterized algorithm

for Hyperbolicity with respect to the minimum maximal path cover number.

Theorem 5 (). Let G = (V, E) be a graph and k be its minimum maximal

path cover number. Then, Hyperbolicity can be solved in O(k 4 (n + m)) time.

Feedback edge number. We next present a polynomial linear-time parameterized

algorithm with respect to the parameter feedback edge number k. The idea is
to show that a graph that is reduced with respect to Reduction Rule 1 contains
O(k) maximal paths.

Theorem 6 (). Hyperbolicity can be computed in O(k 4 (n + m)) time,

where k is the feedback edge number.

Number of vertices with degree at least three. We nally show a polynomial

linear-time parameterized algorithm with respect to the number k of vertices
with degree three or more. To this end, we use the following data reduction rule
additionally to Reduction Rule 1 to bound the number of maximal paths in the
graph by O(k 2 ) (in order to make use of Theorem 5).
402 T. Fluschnik et al.

Reduction Rule 2. Let G = (V, E) be a graph, u, v VG3 be two vertices of

degree at least three, and Puv be the set of maximal paths in G with endpoints u
9
and v. Let Puv Puv be the set containing the shortest path, the four longest
9
even-length paths, and the four longest odd-length paths in Puv . If Puv \Puv = ,
9
then delete in G all inner vertices of the paths in Puv \ Puv .
Lemma 7 (). Reduction Rule 2 is correct and can be exhaustively applied in
linear time.
Observe that if the graph G is reduced with respect to Reduction Rule 2
after Reduction Rule 1 was applied, then for each pair u, v VG3 there ex-
ist at most nine maximal paths with endpoints u and v. Thus, G contains at
most O(k 2 ) maximal paths and using Theorem 5 we arrive at the following.
Theorem 8. Hyperbolicity can be solved in O(k 8 (n + m)) time, where k is
the number of vertices with degree at least three.

4 Parameter Vertex Cover

A vertex cover of a graph G = (V, E) is a subset W V of vertices of G such
that each edge in G is incident to at least one vertex in W . Deciding whether
a graph G has a vertex cover of size at most k is NP-complete in general [14].
There is, however, a simple linear-time factor-2 approximation (see, e.g., [18]). In
this section, we consider the size k of a vertex cover as the parameter. We show
that we can solve Hyperbolicity in time linear in |G|, but exponential in k;
further, we show that, unless SETH fails, we cannot do asymptotically better.

A Linear-Time Algorithm Parameterized by the Vertex Cover Number. We prove

that Hyperbolicity can be solved in time linear in the size of the graph and
exponential in the size k of a vertex cover. This result is based on a linear-time
computable problem kernel of size O(2k ) that can be obtained by exhaustively
applying the following reduction rule.
Reduction Rule 3. If there are at least ve vertices v1 , v2 , . . . , v V ,  >
4, with the same (open) neighborhood N (v1 ) = N (v2 ) = . . . = N (v ), then
delete v5 , . . . , v .
We next show that the above rule is correct, can be applied in linear time, and
leads to a problem kernel for the parameter vertex cover number.
Lemma 9 (). Reduction Rule 3 is correct and can be applied exhaustively in
linear time. Furthermore, if Reduction Rule 3 is not applicable, then the graph
contains at most k + 4 2k vertices and O(k 2k ) edges, where k is the vertex
cover number.
With Reduction Rule 1 we can compute in linear time an equivalent instance
having a bounded number of vertices. Applying to this instance the trivial O(n4 )-
time algorithm yields the following.
Theorem 10. Hyperbolicity can be computed in O(24k +n+m) time, where k
denotes the size of a vertex cover of the input graph.
When Can Graph Hyperbolicity be Computed in Linear Time? 403

SETH-based Lower Bounds. We show that, unless SETH breaks, the 2O(k) +
O(n + m)-time algorithm obtained in the previous subsection cannot be im-
proved to an algorithm even with running time 2o(k) n2 . This also implies,
that, assuming SETH, there is no problem kernel with 2o(k) vertices computable
in O(n2 ) time, i. e., the kernel obtained by applying Reduction Rule 3 cannot
be improved signicantly. The proof follows by a many-one reduction from the



problem Orthogonal Vectors: herein, given two sets A and B each con-
taining n binary vectors of length  = O(log n), the question is whether there




are two vectors
a A and b B such that a and b are orthogonal, that is,


such that there is no position i for which
a [i] = b [i] = 1.
Williams and Yu [19] proved that, if Orthogonal Vectors can be
solved in O(n2 ) time, then SETH breaks. We provide a linear-time reduc-
tion from Orthogonal Vectors to Hyperbolicity where the graph G con-
structed in the reduction contains O(n) vertices and admits a vertex cover of
size O(log n) (and thus contains O(n log n) edges). The reduction then implies
that, unless SETH breaks, there is no algorithm solving Hyperbolicity in time
polynomial in the size of the vertex cover and linear in the size of the graph. We
mention that Borassi et al. [4] showed that under the SETH Hyperbolicity
cannot be solved in O(n2 ). However, the instances constructed in their reduc-
tion have a minimum vertex cover of size (n). Note that our reduction is based
on ideas from the reduction of Abboud et al. [1] for the Diameter problem.

Theorem 11. Assuming SETH, Hyperbolicity cannot be solved in 2o(k)

(n2 ) time, even on graphs with O(n log n) edges, diameter four, and domina-
tion number three. Here, k denotes the vertex cover number of the input graph.


Proof. We reduce any instance ( A , B ) of Orthogonal Vectors to an in-
stance (G, ) of Hyperbolicity, where we construct the graph G as follows
(we refer to Figure 1 for a sketch of the construction).



Make each a A a vertex a and each b B a vertex b of G, and denote
these vertex sets by A and B, respectively. Add two vertices for each of the
 dimensions, that is, add the vertex set C := {c1 , . . . , c } and the vertex set D =
{d1 , . . . , d } to G and make each of C and D a clique. Next, connect each a A
to the vertices of C in the natural way, that is, add an edge between a and ci if
and only if a [i] = 1. Similarly, add an edge between b B and di D if and


only if b [i] = 1. Moreover, add the edge set {{ci , di } | i []}. This part will
constitute the central gadget of our construction.
Our aim is to ensure that the maximum hyperbolicity is reached for 4-


tuples (a, b, c, d) such that a A, b B, and a and b are orthogonal
vectors. The construction of G is completed by adding two paths (uA , u, uB )
and (vA , v, vB ), and making uA and vA adjacent to all vertices in A C and uB
and vB adjacent to all vertices in B D.
Observe that G contains O(n) vertices, O(nlog n) edges, and that the set V \
(A B) forms a vertex cover in G of size O(log n). Moreover, observe that G has
diameter four. Note that each vertex in A B C D is at distance two to each
404 T. Fluschnik et al.

uA u uB

A C D B
a1 c1 d1 b1
.. i ai [1] = 1 ..
. i bj [2] = 1 .
ai .. .. bj
. .. .. .
.. .. .. ..
. .
an c d bn

vA v vB

Fig. 1. Sketch of the construction described in the proof of Theorem 11. Ellipses indi-
cate cliques, rectangles indicate independent sets. Multiple edges to an object indicate
that the corresponding vertex is incident to each vertex enclosed within that object.

of u and v. Moreover, vA and vB are at distance three to u. Analogously, uA ,

uB are at distance three to v. Furthermore u and v are at distance four. Finally,
observe that {uA , uB , v} forms a dominating set in G.


We complete the proof by showing that ( A , B ) is a yes-instance of Orthog-
onal Vectors if and only if G has hyperbolicity at least = 4.




() Let ( A , B ) be a yes-instance, and let a A and b B be a pair


of orthogonal vectors. We claim that (a, b, u, v) = 4. Since
a and b are or-


thogonal, there is no i [] with
a [i] = b [i] = 1 and, hence, there is no
path connecting a and b only containing two vertices in C D, and it holds
that ab = 4. Moreover, we know that uv = 4 as that au = bu = av = av = 2.
Thus, (a, b, u, v) = 8 4 = 4, and G is 4-hyperbolic.
() Let S = {a, b, c, d} be a set of vertices such that (a, b, c, d) 4. By
Lemma 1, it follows that no two vertices of S are adjacent. Hence, we assume
without loss of generality that ab = cd = 4. Observe that all vertices of C
and D have distance at most three to all other vertices. Similarly, each vertex
of {uA , vA , uB , vB } has distance at most three to all other vertices. (Consider
for example uA . By construction, uA is a neighbor of all vertices in A C {u}
and, hence, uA has distance at most two to vA and to all vertices in D. Thus,
uA has distance at most three to v, B, uB and vB and therefore to all vertices
of G. The arguments for vA , uB , and vB are symmetric.)
It follows that S A B {u, v}, and therefore at least two vertices in S are
from A B. Thus, assume without loss of generality that a is contained in A. By
the previous assumption, we have that ab = 4. This implies that b B and
a


and b are orthogonal vectors, as every other vertex in V \ B is at distance three
When Can Graph Hyperbolicity be Computed in Linear Time? 405

to a and each b B with b being non-orthogonal to

a is at distance three


to a. Hence, ( A , B ) is a yes-instance. 

We remark that, with the above reduction, the hardness also holds for the
variants in which we x one vertex (u) or two vertices (u and v). The reduc-
tion also shows that approximating the hyperbolicity of a graph within a factor
of 4/3  cannot be done in strongly subquadratic time or with a PL-FPT
running time with respect to the vertex cover number.
Next, we adapt the above reduction to obtain the following hardness result
on graphs of bounded maximum degree.
Theorem 12 (). Assuming SETH, Hyperbolicity cannot be solved in f ()
(n2 ) time, where denotes the maximum degree of the input graph.

5 Parameter Distance to Cographs

In this section we describe a linear-time parameterized algorithm for Hyper-
bolicity parameterized by the vertex deletion distance k to cographs; that is,
we present an algorithm with linear dependence on the input size but arbitrary
dependence on the parameter (to which we refer to as L-FPT). A graph is a
cograph if and only if it is P4 -free. Given a graph G we can determine in linear
time whether it is a cograph and return an induced P4 if this is not the case.
This implies that in O(k (m + n)) time we can compute a set X V of size at
most 4k such that G X is a cograph.
A further characterization is that a cograph can be obtained from graphs
consisting of one single vertex via unions and joins [5].
A union of two graphs G1 = (V1 , E1 ) and G2 = (V2 , E2 ) is the graph (V1
V2 , E1 E2 ).
A join of two graphs G1 = (V1 , E1 ) and G2 = (V2 , E2 ) is the graph (V1
V2 , E1 E2 {{v1 , v2 }|v1 V1 , v2 V2 }).
The union of t graphs and the join of t graphs are dened by taking succes-
sive unions or joins, respectively, of the t graphs in an arbitrary order. Each
cograph G can be associated with a rooted cotree TG . The leaves of TG are
the vertices of V . Each internal node of TG is labeled either as a union or join
node. For node v in TG , let L(v) denote the leaves of the subtree rooted at v.
For a union node v with children u1 , . . . , ut , the graph G[L(v)] is the union of
the graphs G[L(ui )], 1 i t. For a join node v with children u1 , . . . , ut , the
graph G[L(v)] is the join of the graphs G[L(ui )], 1 i t.
The cotree of a cograph can be computed in linear time [8]. In a subroutine
in our algorithm for Hyperbolicity we need to solve the following variant of
Subgraph Isomorphism.
Colored Induced Subgraph Isomorphism
Input: An undirected graph G = (V, E) with a vertex-coloring : V N
and an undirected graph H = (W, F ), where |W | = k, with a
vertex-coloring : W N.
Question: Is there a vertex set S V such that there is an isomorphism f
from G[S] to H such that (v) = (f (v)) for all v S?
406 T. Fluschnik et al.

Informally, the condition that (v) = (f (v)) means that every vertex is mapped
to a vertex of the same color. We say that such an isomorphism respects the
colorings. As shown by Damaschke [9], Induced Subgraph Isomorphism on
cographs is NP-complete. Since this is the special case of Colored Induced
Subgraph Isomorphism where all vertices in G and H have the same color,
Colored Induced Subgraph Isomorphism is also NP-complete (contain-
ment in NP is obvious). In the following, we show that on cographs Colored
Induced Subgraph Isomorphism can be solved by an L-FPT algorithm when
the parameter k is the order of H.

Lemma 13 (). Colored Induced Subgraph Isomorphism can be solved

in O(3k (n + m)) time in cographs.

We now turn to the algorithm for Hyperbolicity on graphs that can

be made into cographs by at most k vertex deletions. The nal step is to
reduce Hyperbolicity to the problem Distance-Constrained 4-Tuple:
herein, given an undirected graph G = (V, E) and six integers d{a,b} , d{a,c} ,
d{a,d} , d{b,c} , d{b,d} , and d{c,d} , the question is whether there is a set S V of
four vertices and a bijection f : S {a, b, c, d} such that for each x, y S we
have xy = d{f (x),f (y)} .

Lemma 14 (). Distance-Constrained 4-Tuple can be solved in O(44k k

(n + m)) time if G X is a cograph for some X V of size k.

We solve Hyperbolicity by creating O(k 6 ) instances of Distance-

Constrained 4-Tuple as shown below.
Theorem 15. Hyperbolicity can be solved in O(44k k 7 (n+m)) time, where
k is the vertex deletion distance of G to cographs.

Proof. Let G = (V, E) be the input graph and X V , |X| k, such that G X
is a cograph and observe that X can be computed in O(4k (n + m)) time. Since
every connected component of G X has diameter at most two, the maximum
distance between any pair of vertices in the same component of G is at most 4k +
2: any shortest path between two vertices u and v visits at most k vertices in X,
at most three vertices between every pair of vertices x and x from X and at
most three vertices before encountering the rst vertex of X and at most three
vertices before encountering the last vertex of X.
Consequently, for the 4-tuple (a, b, c, d) that maximizes (a, b, c, d), there
are O(k 6 ) possibilities for the pairwise distances between the four vertices. Thus,
we may compute whether there is a 4-tuple such that (a, b, c, d) = by checking
for each of the O(k 6 ) many 6-tuples of possible pairwise distances of four ver-
tices in G whether there are 4 vertices in G with these six pairwise distances and
whether this implies (a, b, c, d) . The latter check can be performed in O(1)
time, and the rst is equivalent to solving Distance-Constrained 4-Tuple
which can be done in O(44k k (n + m)) time by Lemma 14. The overall running
time follows. 

When Can Graph Hyperbolicity be Computed in Linear Time? 407

6 Reduction from 4-Independent Set

In this section, we provide a further relative lower bound for Hyperbolic-
ity. Specically, we prove that, if the running time is measured in terms of n,
then Hyperbolicity is at least as hard as the problem of nding an indepen-
dent set of size four in a graph. The currently best running time for this problem
is O(n3.257 ) [11, 20]. Hence, any improvement on the running time of Hyperbol-
icity which breaks this bound (e.g., an algorithm running in o(n3 ) time), would
also yield a substantial improvement for the 4-Independent Set problem.
To this end, we reduce from a 4-partite (or 4-colored) variant of the Inde-
pendent Set problem. The standard reduction [12] from Independent Set
to Multicolored Independent Set shows that this 4-colored variant has the
same asymptotic running time lower bound as 4-Independent Set.
Theorem 16 (). Any algorithm solving Hyperbolicity in O(nc ) time for
some constant c yields an O(nc )-time algorithm solving 4-Independent Set.

7 Conclusion
To eciently compute the hyperbolicity number, parameterization sometimes
may help. In this respect, perhaps our practically most promising results relate
to the O(k 4 (n + m)) running times (for the parameters covering path number
and feedback edge number, see Table 1). Note that they clearly improve on the
standard algorithm when k = O(n1/4 ). Moreover, the linear-time data reduc-
tion rules we presented may be of independent practical interest. On the lower
bound side, together with the work of Abboud et al. [1] our SETH-based lower
bound with respect to the parameter vertex cover number is among few known
exponential lower bounds for a polynomial-time solvable problem.
As to future work, we particularly point to the following open questions. First,
we left open whether there is an L-FPT algorithm exploiting the parameter
feedback vertex number for computing the hyperbolicity number. Second, for
parameter vertex cover number we have an SETH-based exponential lower bound
for the parameter function in any L-FPT algorithm. This does not imply that
it is impossible to achieve a polynomial parameter dependence when asking for
algorithms with running time factors such as O(n2 ) or O(n3 ).

Bibliography
[1] A. Abboud, V. Vassilevska Williams, and J. R. Wang. Approximation and
xed parameter subquadratic algorithms for radius and diameter in sparse
graphs. In Proc. 27th SODA, pages 377391. SIAM, 2016.
[2] M. Abu-Ata and F. F. Dragan. Metric tree-like structures in real-world
networks: an empirical study. Networks, 67(1):4968, 2016.
[3] M. Borassi, D. Coudert, P. Crescenzi, and A. Marino. On computing the
hyperbolicity of real-world graphs. In Proc. 23rd ESA, volume 9294 of
LNCS, pages 215226, 2015.
408 T. Fluschnik et al.

[4] M. Borassi, P. Crescenzi, and M. Habib. Into the square: On the complexity
of some quadratic-time solvable problems. Electronic Notes in Theoretical
Computer Science, 322:5167, 2016.
[5] A. Brandstadt, V. B. Le, and J. P. Spinrad. Graph Classes: a Survey,
volume 3 of SIAM Monographs on Discrete Mathematics and Applications.
SIAM, 1999.
[6] G. Brinkmann, J. H. Koolen, and V. Moulton. On the hyperbolicity of
chordal graphs. Annals of Combinatorics, 5(1):6169, 2001.
[7] N. Cohen, D. Coudert, and A. Lancin. On computing the Gromov hyper-
bolicity. ACM Journal of Experimental Algorithmics, 20:1.6:11.6:18, 2015.
[8] D. G. Corneil, Y. Perl, and L. K. Stewart. A linear recognition algorithm
for cographs. SIAM Journal on Computing, 14(4):926934, 1985.
[9] P. Damaschke. Induced subgraph isomorphism for cographs is NP-complete.
In Proc. 16th WG, volume 484 of LNCS, pages 7278. Springer, 1991.
[10] M. Doucha and J. Kratochvl. Cluster vertex deletion: A parameterization
between vertex cover and clique-width. In Proc. 37th MFCS, volume 7464
of LNCS, pages 348359. Springer, 2012.
[11] F. Eisenbrand and F. Grandoni. On the complexity of xed parameter
clique and dominating set. Theoretical Computer Science, 326(1-3):5767,
2004.
[12] M. Fellows, D. Hermelin, F. Rosamond, and S. Vialette. On the parameter-
ized complexity of multiple-interval graph problems. Theoretical Computer
Science, 410(1):5361, 2009.
[13] H. Fournier, A. Ismail, and A. Vigneron. Computing the Gromov hyperbol-
icity of a discrete metric space. Information Processing Letters, 115(6-8):
576579, 2015.
[14] M. R. Garey and D. S. Johnson. Computers and Intractability: A Guide to
the Theory of NP-Completeness. Freeman, 1979.
[15] A. C. Giannopoulou, G. B. Mertzios, and R. Niedermeier. Polynomial xed-
parameter algorithms: A case study for longest path on interval graphs. In
Proc. 10th IPEC, volume 43 of LIPIcs, pages 102113. Schloss Dagstuhl -
Leibniz-Zentrum fuer Informatik, 2015.
[16] M. Gromov. Hyperbolic groups. In Essays in Group Theory, MSRI Publ.,
vol. 8, pages 75263. Springer New York, 1987.
[17] D. Mitsche and P. Pralat. On the hyperbolicity of random graphs. The
Electronic Journal of Combinatorics, 21(2):P2.39, 2014.
[18] C. H. Papadimitriou and K. Steiglitz. Combinatorial Optimization: Algo-
rithms and Complexity. Prentice-Hall, 1982.
[19] R. Williams and H. Yu. Finding orthogonal vectors in discrete structures.
In Proc. 25th SODA, pages 18671877. SIAM, 2014.
[20] V. V. Williams, J. R. Wang, R. Williams, and H. Yu. Finding four-node
subgraphs in triangle time. In Proc. 26th SODA, pages 16711680. SIAM,
2015.
 Optimal Query Time for Encoding Range
Majority

Pawel Gawrychowski and Patrick K. Nicholson

University of Haifa, Israel and Nokia Bell Labs, Ireland

Abstract. We revisit the range -majority problem, which asks us to

preprocess an array A[1..n] for a xed value of (0, 12 ], such that for any
query range [i, j] we can return a position in A of each distinct -majority
element. A -majority element is one that has relative frequency at least
in the range [i, j]: i.e., frequency at least (j i + 1). Belazzougui et al.
[WADS 2013] presented a data structure that can answer such queries in
O(1/ ) time, which is optimal, but the space can be as much as (n lg n)
bits. Recently, Navarro and Thankachan [Algorithmica 2016] showed that
this problem could be solved using an O(n lg(1/ )) bit encoding, which is
optimal in terms of space, but has suboptimal query time. In this paper,
we close this gap and present a data structure that occupies O(n lg(1/ ))
bits of space, and has O(1/ ) query time. We also show that this space
bound is optimal, even for the much weaker query in which we must
decide whether the query range contains at least one -majority element.

1 Introduction

Misra and Gries [14] generalized a classic 2-pass algorithm by Boyer and Moore [3]
for nding majorities in lists of elements. Formally, a -majority of a list of length
n (or -heavy-hitter) is an element that appears with frequency at least n.
More recent variants and improvements [5,12] to the Misra-Gries algorithm have
become standard tools in a wide variety of applications involving streaming an-
alytics, such as IP trac monitoring, data mining, etc.
In this paper we consider the data structure variant of the problem. Suppose
we are given an array A of n elements. The goal is to preprocess the array
into a data structure that supports range -majority queries: given an arbitrary
subarray A[i..j], return all distinct elements that are -majorities in A[i..j]. As
an example, we may wish to construct such a structure on network trac logs,
to perform an analysis of how the set of frequent users changes over time.
In the last few years, this problem has received a lot of attention [2, 6, 8, 13],
nally leading to a recent result of Belazzougui et al. [1, 2]: these queries can
be supported in O(1/ ) time, using (1 + )nH0 + o(n) bits of space, where
H0 is the zero-th order empirical entropy of the array A, and is an arbitrary
positive constant.1 Since, for an arbitrary -majority query, there can be 1/ 
1
Note that, for this and all forthcoming results discussed, we assume the word-RAM
model of computation with word-size w = (lg n) bits; we use lg x to denote log2 x.

Springer International Publishing AG 2017 409

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 409420, 2017.
DOI: 10.1007/978-3-319-62127-2_35
410 P. Gawrychowski and P.K. Nicholson

answers, there is not much hope for signicantly improving the query time of
O(1/ ), except perhaps to make the time bound output-sensitive on the number
of results returned [1, Sec.7].
On the other hand, much more can be said about the space bound. Note that,
in general, if A contains elements drawn from the alphabet [1, ], then we can
 nlg  bits. If fi is the frequency of element i [1, ], then we
represent it using
have nH0 = n i ((fi /n) lg(n/fi )) nlg .2 Since the bound of Belazzougui
et al. [1] depends on the entropy of the elements in A, it can therefore can be
(n lg n) bits, if = (nc ) for any constant c 1, and the distribution is close
to uniform. However, quite recently, Navarro and Thankachan [15] showed that
this space bound can be improved signicantly in the encoding model.
In the encoding model, given array A as input, we are allowed to construct
an encoding that supports a specic query operation on A. After constructing
the encoding, the array A is deleted, and queries must be supported by accessing
only the encoding. For many query operations, we can achieve space bounds that
are much smaller than the space required to store A. One issue is that for range
-majority queries, if we return the actual element which is a -majority, then
we must store at least as many bits as are required to represent A. This follows
since an encoding supporting such queries can be used to return the contents of
the array A by querying the range A[i..i] for each 1 [1, n].
Navarro and Thankachan [15] therefore considered a dierent query, in which,
for each -majority a in the query range A[i..j], we instead return an arbitrary
position  in A such that A[] = a and i  j. In the remainder of the paper,
we use range -majority position query to refer to this positional variant of the
query operation. Navarro and Thankachan [15] showed two main results:
Theorem 1 ([15], Theorems 1 and 2).
1. For any (0, 1), there is an encoding that occupies O(nlg(1/ )) bits of
space that supports range -majority position queries in:
(a) O((1/ ) lg n) time if 1/ = o(polylog(n)).
(b) O(1/ ) time if 1/ = (polylog(n)).
(c) O(1/ lg lgw (1/ )) time if 1/ = (polylog(n)).
2. Any encoding that can support range -majority counting queries (i.e., re-
turn the total the number of -majorities)


1 in an arbitrary
 query range A[i..j]
occupies space (in bits) at least n4 lg 2 1 lg e = (n lg(1/ )).
Thus, their lower bound implies that their space bound, which depends only
on n and rather than elements in the input array A, is optimal. However,
there is gap between the query time of their encoding and the data structure of
Belazzougui et al. [1] for the case where 1/ is not (polylog(n)). Crucially, this
does not yield optimal time in the important case where 1/ is a constant. In
this paper, we close this time gap, and prove the following theorem:
We also note that Belazzougui et al. [1] also considered a slightly more dicult
problem in which can be specied at query time, rather than xed once-and-for-all
before constructing the data structure.
2
We follow the convention that (fi /n) lg(n/fi ) = 0 if fi = 0.
Optimal Query Time for Encoding Range Majority 411

Theorem 2. For any (0, 1/2], there is an encoding that occupies O(n lg(1/ ))
bits of space that can answer range -majority position queries in O(1/ ) time.3

Of course one could ask if O(1/ ) is the right bound for the query time
at all. In the output-sensitive variant of the problem the query time should
depend on the number of results returned, which might be up to O(1/ ) but
possibly smaller. However, we note that a straightforward reduction from the set
intersection conjecture indicates that a signicantly smaller query time cannot
be guaranteed even if the size of the output is 0 or 1 [9].
In terms of techniques, our approach uses the level-based decomposition of
Durocher et al. [6], but with three signicant improvements. We dene two new
methods for pruning their data structure to reduce space, and one method to
speed up queries. The rst pruning method is a top-down approach that avoids
replicating data structures at more than one level and is analysed using a charg-
ing argument. The second pruning method is bottom-up, operating on small
ranges of the input array, that we call micro-arrays, and applies one of two
strategies, depending on the parameter . One of these strategies involves boot-
strapping an optimal space (but suboptimal query time) encoding by combining
it with pre-computed lookup tables in order to speed up queries on the micro-
arrays. The other strategy stores (a rank reduced) copy of the micro-array and
solves queries in a brute-force manner. Finally, the last improvement uses wavelet
trees [11] in a non-trivial way in order to build fast ranking data structures to
improve query time for the case when 1/ = (polylog(n)).

Implications. Since the encoding yields the positions of each distinct -majority
element in the query range, we can use our optimal encoding as an alternative
to the non-encoding data structure of Belazzougui et al. This is done by rst
compressing the original array A using any compressor that supports access in
O(1) to the underlying elements.

Theorem 3. Let S(n) be the space required to store the input array in a com-
pressed form such that each position can be accessed in O(1) time. Then there is
a data structure that occupies S(n) + O(n lg(1/ )) bits of space, and can return
the range -majorities for an arbitrary range [i, j] in O(1/ ) time.

For example, using results for higher order entropy compression with O(1)
access time [7, 10] yields the following:

Corollary 1. Let A[1..n] be an array with elements drawn from [1, ]. There
is a data structure that occupies nHk + o(n lg ) + O(n lg(1/ )) bits of space4 ,
and can support arbitrary range -majority queries in time O(1/ ), for any
k = o(log n).
3
For (1/2, 1) the structure for 1/2-majorities can answer queries in O(1) time.
4
Hk denotes the k-th order empirical entropy of the sequence of elements in A: a
lower bound on the space achievable by any compressor that considers the previous
k elements in the sequence. For all k 1 we have nHk nHk1 .
412 P. Gawrychowski and P.K. Nicholson

Lower Bound. Recall the lower bound of (n lg(1/ )) bits holds for any encoding
supporting range -majority counting queries. We consider an easier problem
that we call range -majority decision queries. The query asks Is there at least
one element in the query range A[i..j] which is a -majority?. Since the previous
lower bound does not rule out a better encoding for these decision queries, it is
natural to ask whether a better encoding exists. We prove the following:

Theorem 4. Any data structure that can be used to represent an array A[1..n]
and support 1/k-majority decision queries, for any integer k 2, on any arbi-
trary query range [i, j], requires n lg ke (k 4 lg k) bits of space.

Thus, we answer this question in the negative by showing a lower bound of

(n lg(1/ )) bits for any encoding that supports these queries, which proves our
structure is space-optimal for even these restricted types of queries. Moreover,
we note that our lower bound has an improved constant factor compared to the
previous lower bound.

Related Work. Finally, we remark that the area of range queries on arrays is
quite vast, and there are many interesting related types of queries that have
been studied in the both the non-encoding and encoding models; we refer the
reader to surveys on the topics [17, 19]. The most closely related problem to the
range -majority problem is the range mode problem [4]: given a query range
[i, j] return the most frequently appearing element in the range. In contrast
with range -majority, this type of query is $
signicantly less ecient, with the
best (n lg n) bit data structures having O( n/ lg n) query time.

2 Preliminaries

Lemma 1 ([16]). Let V be a bit vector of length n bits in which m of the bits are
set to one. There is a data structure for representing V that uses m lg(n/m) +
O(n/ lgc (n)) bits for any constant c 1 such that the following queries can be
answered in O(1) time:

access(V, i): returns bit V [i].

rank(V, i): returns the number of ones in the prex V [1..i].
select(V, j): returns the index of the j-th one in V , if it exists, and 1 oth-
erwise. In other words, the inverse of the rank operation: if select(V, j) = i,
then rank(V, i) = j.

Since our proof makes heavy use of this lemma, we distinguish the m lg(n/m)
term in the space bound by calling it the leading term, and the other term
the redundancy. If we do not need the full power of rank, then we can use the
following lemma to reduce the redundancy:
Lemma 2 ([18]). If only the constant time select and access operations are
required, then we can represent V using m lg(n/m) + o(m) + O(lg lg n) bits.
Optimal Query Time for Encoding Range Majority 413

A useful fact about applying these previous Lemmas to bit vectors is that
concatenation is often helpful: if we apply either Lemma to two bit vectors
separately, both of length n containing at least m bits, then the sum of the
leading terms is no more than 2m lg(n/m). If we concatenate the bit vectors
before applying the lemma, the upper bound on the leading term is the same.

3 Upper Bound
3.1 Quadruple Decomposition
The upper bound makes use of the quadruple decomposition of Durocher et
al. [6]. For ease of description, we assume that n is a power of 2, but note that
decomposition works in general. First, at a conceptual level we build a balanced
binary tree over the array A[1..n]. Each leaf represents an element A[i]. On
the k-th level of the tree T (k), counting from the leaves at level 0, the nodes
represent a partition of A[1..n] into nk = n/2k contiguous blocks of length 2k .
Second, consider all levels containing at least four blocks. At each such level,
consider the blocks B1 , . . . , Bnk . We create a list of quadruples (i.e., groups of
four consecutive blocks) at each such level:

D(k) = [(B1 , B2 , B3 , B4 ),(B3 , B4 , B5 , B6 ), . . . , (Bnk 1 , Bnk , B1 , B2 )].

Thus, each index in A is contained in exactly two quadruples at each level, and
there is one quadruple that wraps-around to handle corner cases. The quadruples
are staggered at an oset of two blocks from each other. Moreover, given a
quadruple D = (B2+1 , B2+2 , B2+3 , B2+4 ), the two middle blocks B2+2 and
B2+3 are not siblings in the binary tree T . We call the range spanned by these
two middle blocks the middle part of D.
As observed by Durocher et al. [6], for every query range [i, j] there exists
a unique level k in the tree such that [i, j] contains at least one and at most
two consecutive blocks in T (k), and, if [i, j] contains two blocks, then the nodes
representing these blocks are not siblings in the tree T . Thus, based on our
quadruple decomposition, for every query range [i, j] we can associate it with
exactly one quadruple D = (B2+1 , B2+2 , B2+3 , B2+4 ) such that

((B2+2 [i, j]) (B2+3 [i, j])) ([i, j] B2+1 B2+2 B2+3 B2+4 .)

Moreover, Durocher et al. [6] proved the following lemma:

Lemma 3 ([6]). For each query range [i, j], in O(1) time we can compute the
level k, as well as the oset of the quadruple associated with [i, j] in the list D(k),
using o(n) bits of space.

Furthermore, if we consider any arbitrary query range [i, j] that is associated

with a quadruple D, there are at most 4/ elements in the range represented by
D that could be -majorities for the query range. Following Durocher et al., we
refer to these elements as candidates for the quadruple D.
414 P. Gawrychowski and P.K. Nicholson

For each quadruple, we compute and store all of its candidates, so that, by
Lemma 3, in O(1) time we can obtain O(1/ ) candidates. It remains to show
how to verify that a candidate is in fact a -majority in A[i..j]. At this point,
our approach deviates from Durocher et al. [6], who make use of a wavelet tree
for verication, and end up with a space bound of O(n lg n lg(1/ )) bits.
Consider such a candidate y for quadruple D = (B2+1 , . . . , B2+4 ). Our
goal is to count the number of occurrences of y in the query range [i, j]. To do
this we store a bit vector V (D, y), that represents the (slightly extended) range
B2 . . . B2+5 and marks all occurrences of y in this range with a one bit. By
counting the number of ones in the range corresponding to [i, j] in V (D, y), we
can determine if the number of occurrences exceeds the threshold (j i + 1).
If the threshold is exceeded, then we can return the rst one bit in the range, as
that position in A contains element y. Note that we have extended the range of
the bit vector beyond the range covered by D by one extra block to the left and
right. We call this extended range the extent of D, and observe the following:

Observation 5. Let E(D) be the extent of quadruple D at level k. Then for

all quadruples D at level k  < k such that the range of D has non-empty
intersection with the range of D, we have that D E(D).

We now briey analyze the total space of this method, under the assumption
that we can store a bit vector of length n with m one bits using O(m lg(n/m))
bits. This crude analysis is merely to illustrate that additional tricks are needed
to achieve optimal space. The quadruple decomposition consists of lg n levels.
On each level, we store a number of bit vectors. For each quadruple we have up
to O(1/ ) candidates Y. Thus, if fy represents the frequency of candidate y in
extent
 of quadruple D, then the space bound, for each quadruple at level k, is
yY O(f k
c lg(2 /fc )), which, by the concave version of Jensens inequality, is
bounded by O((2k ) lg(1/ )). So each level uses O(n lg(1/ )) bits, for a total of
O(n lg n lg(1/ )) bits over all levels.

3.2 Optimal Space with Suboptimal Query Time

To achieve space O(n lg(1/ )) bits, the intuition is that we should avoid dupli-
cating the same bit vectors between levels. It is easy to imagine a case where ele-
ment y is a candidate at every level and in every quadruple of the decomposition,
which results in many duplicated bit vectors. To avoid this duplication problem,
we propose a top-down algorithm for pruning the bit vectors. Initially, all indices
in A are active at the beginning. Our goal is to charge at most O(lg(1/ )) bits
to each active index in A, which achieves the desired space bound.
Let k be the current level of the quadruple decomposition, as we proceed top-
down. We maintain the invariant that for any element y in a block Bi , either
all indices storing occurrences of y are active in Bi (in which case we say y is
active in Bi ), or none are (in which case we say y is inactive in Bi ). Consider a
candidate y associated with quadruple D = (B2+1 , B2+2 , B2+3 , B2+4 ). Then:
Optimal Query Time for Encoding Range Majority 415

1. If y is active in blocks B2+1 , . . . , B2+4 , then we store the bit vector V (D, y),
and (conceptually) mark all occurrences of y inactive in these blocks after
we nish processing level k. This makes y inactive in all blocks contained in
D at lower levels. Since a block Bi is contained in two quadruples at level k,
a position storing y in Bi may be made inactive for two reasons: this is why
we mark positions inactive after processing all quadruples at level k.
2. If y is inactive in some block Bi D, then it is the case that we have
computed and stored the bit vector V (D , y) for some quadruple D at level
k  > k, such that D D = 0. Therefore, Observation 5 implies that D is
contained in the extent of D , and thus the bit vector associated with D
can be used to answer queries for D. For D we need not to store V (D, y),
though for now we do not address how to eciently answer these queries.

Next we analyse the total cost of the bit vectors that we stored during the
top-down construction. The high level idea is that we can charge the cost of bit
vector V (D, y) to the indices in D that store occurrences of y. Call these the
indices the sponsors of V (D, y). Since y is a -majority, it occurs at least O( 2k )
times in D, which has length O(2k ). Thus, we can expect to charge O(lg(1/ ))
bits to each sponsor: the expected gap between one bits is O(1/ ) and therefore
can be recorded using O(lg(1/ )) bits. There are some minor technicalities that
must be addressed, but this basic idea leads to the following intermediate result,
in which we dont concern ourselves with the query time:
Lemma 4. There is an encoding of size O(n lg(1/ )) bits such that the answer
to all range -majority position queries can be recovered.
Proof. Consider candidate y and its occurrences in extent E(D) of quadruple
D at level k, for which we stored the bit vector V (D, y). Suppose there are
fy occurrences of y in E(D). If at least one third of the occurrences of y are
contained in D, then we charge the cost of the bit vector to the (at least) fy /3
sponsor indices in D. Otherwise, this implies one of the two blocks, call it Bi such
that Bi E(D) but Bi  D contains at least fy /3 occurrences of y. Therefore,
y must also be an active candidate for the unique quadruple D that has non-
empty intersection with both Bi and D: this follows since y occurs more times
in D than in D, and y is a candidate for D. In this case we charge the cost of
the bit vector to the sponsor indices in neighbouring quadruple D .
Suppose we store the bit vectors using Lemma 2: for now ignore the O(lg lg n)
term in the space bound as we deal with it in the next paragraph. Using Lemma 2,
the cost of the bit vector V (D, y) associated with D is at most O(fy lg(1/ )),
since y is a ( /4)-majority in D. Thus, O(fy ) sponsors in D pay for at most three
bit vectors: V (D, y) and possibly the two other bit vectors that cost O(fy lg(1/ ))
bits, charged by neighbouring quadruples. Since this charge can only occur at
one level in the decomposition (the index becomes inactive at lower levels after
the rst charge occurs), each sponsor is charged O(lg(1/ )), making the total
amount charged O(n lg(1/ )) bits overall.
To make answering queries actually possible, we make use of the same tech-
nique used by Durocher et al. [6], which is to concatenate the bit vectors at level
416 P. Gawrychowski and P.K. Nicholson

k. The candidates have some implicit ordering in each quadruple, [1, ..., O(1/ )]:
the ordering can in fact be arbitrary. For each level k, we concatenate the bit
vectors associated with quadruple according to this implicit ordering of the can-
didates. Thus, since there are O(lg n) bit vectors (one per level), the O(lg lg n)
term for Lemma 2 contributes O(lg n lg lg n) to the overall space bound.
Given a query [i, j], Lemma 3 allows us to compute the level, k, and o-
set, , of the quadruple associated with [i, j]. Our goal is to remap [i, j] to the
relevant query range in the concatenated bit vector at level k. Since all bit
vectors V (D, y) at level k have the same length, we only need to know how
many bit vectors are stored for quadruples 1, ...,  1: call this quantity X.
Thus, at level k we construct and store a bit vector Lk of length O(nk / ) in
which we store the number of bit vectors associated with the quadruples in
unary. So, if the rst three quadruples have 2, 6, 4 candidates (respectively), we
store Lk = 1001000000100001 . . .. Overall, the space for Lk is O(nk lg(1/ )), or
O(n lg(1/ )) overall, if we represent each Lk using Lemma 2.
Given an oset , we can perform select(Lk , )  to get X. Once we have X,
we can use the fact that all extents have xed length at a level in order to remap
the query [i, j] to the appropriate range [i , j  ] in the concatenated bit vector
for each candidate. We can then use binary search and the select operation to
count the number of 1 bits corresponding to each candidate in the remapped
range [i , j  ] in O(lg n) time per candidate. Since some of the candidates for
the D associated with [i, j] may have been inactive, we also must compute the
frequency of each candidate in quadruples at higher levels that contain D. Since
there are O(lg n) levels, O(1) quadruples that overlap D per level, and O(1/ )
candidates per quadruple, we can answer range -majority position queries in
O(lg2 n/ ) time. Note that we have to be careful to remove possible duplicate
candidates (at each level the quadruples that overlap D may share candidates).

3.3 Optimal Space with Optimal Query Time

In Lemma 4 there are two issues that make querying inecient: 1) we have to
search for inactive candidates in O(lg n) levels; and 2) we used Lemma 2 which
does not support O(1) time rank queries. The solutions to both of these issues
are straightforward. For the rst issue, we store pointers to the appropriate bit
vector at higher levels, allowing us to access them in O(1) time. For the second
issue we can use Lemma 1 to support rank in O(1) time. However, both of these
solutions raise their own technical issues that we must resolve in this section.

Pointers to higher levels. Consider a quadruple D at level k for which candidate

y is inactive in some block contained in D. Recall that this implies the existence
of some bit vector V (D , y) for some D at level k  > k that can be used to
count occurrences of y in D. In order to access this bit vector in O(1) time,
the only information that we need to store is the number k  and also the oset
of y in the list of candidates for D : D might have a dierent ordering on its
candidates than D. Thus, in this case we store O(lg lg n + lg(1/ )) bits per
quadruple as we have O(lg n) levels and O(1/ ) candidates per quadruple. This
Optimal Query Time for Encoding Range Majority 417

is a problem, because there are O(n) quadruples, which means these pointers
can occupy O(n/ (lg lg n + lg(1/ ))) bits overall.
To deal with this problem, we simply reduce the number of quadruples using
a bottom-up pruning technique: all data associated with quadruples spanning
a range of size Z or smaller is deleted. This is good as it limits the space for
the pointers to at most O(n(lg lg n + lg(1/ ))/( Z)) bits, as there are O(n/Z)
quadruples of length greater than Z. However, we need to come up with an
alternative approach for queries associated with these small quadruples.
The value we select for Z, as well as the strategy to handle queries associated
with quadruples of size Z or smaller, depends on the value of 1/ :

1. If 1/ lg n: then we set Z = 1/ . Thus, the pointers occupy O(n(lg lg n+
lg(1/ )) = O(n lg(1/ )) bits (since lg(1/ ) = (lg lg n)). Consider the max-
imum level k such that the quadruples are of size Z or smaller. For each
quadruple D in level k, we construct a new micro-array of length 2k by
copying the range spanned by D from A. Thus, any query [i, j] associated
with a quadruple at levels k or lower can be reduced to a query on one of
these micro-arrays. Since the micro-arrays have length 1/ , we preprocess
the elements in the array by replacing them by their ranks (i.e., we reduce
the elements to rank space). Storing the micro-array therefore requires only
O(nk 2k lg(1/ )) = O(n lg(1/ )) bits. Moreover, since we have access to the
ranks of the elements directly, we can answer any query on the micro-array
directly by scanning it in O(1/ ) time. Thus, in this case, the space for the
and pointers is O(n lg(1/ )).
micro-arrays
2. If 1/ < lg n: in this branch we use the encoding of Lemma 4 that occupies
c n lg(1/ ) bits of space for an array of length n, for some constant c 1.
We set Z = lg n/(2c lg(1/ )), so that the space for the pointers becomes:
$
O(n(lg lg n + lg(1/ )) lg(1/ )/( lg n)) = O(n(lg lg n)2 / lg n) = o(n).

As in the previous case, we construct the micro-arrays for the appropriate

quadruples based on the size Z. However, this time we encode each micro-
array using Lemma 4. This gives us a set of nk encodings, taking a total
O(nk 2k lg(1/ )) = O(n lg(1/ )) bits. Moreover, the answer to a query is fully
determined by the encoding and the endpoints i, j. Since i and j are fully
contained in the micro-array, their description takes lg Z bits. Thus, using
2
an auxiliary lookup table of size O(2cZ lg(1/ ) 2lg Z ) we can preprocess the
answer for every possible encoding
and positions i, j so that a query takes
O(1) time. Because 1/ < lg n the space for this lookup table is:
2 2
O(2c(lg n/(2c lg(1/ ))) lg(1/ )+lg (lg n/(2c lg(1/ )))
) = O(2lg n/2+(lg lg n) ) = o(n).

In summary, we can apply level-based pruning to reduce the space required

by the pointers to at most O(n lg(1/ )). Note that we must be able to quickly
access the pointers associated with each quadruple D. To do this, we concatenate
the pointers at level k, and construct yet another bit vector Lk having a similar
format as Lk . The bit vector Lk allows us to easily determine how many pointers
418 P. Gawrychowski and P.K. Nicholson

are stored for the quadruples to the left of D at the current level, as well as how
many are stored for D. Thus, these additional bit vectors occupy O(n lg(1/ ))
bits of space, and allow accessing an arbitrary pointer in O(1) time.

Using the faster ranking structure. When we use the faster rank structure of
Lemma 1, we immediately get that we can verify the frequency of each can-
didate in O(1) time, rather than O(lg n) time. Recall that the bit vectors are
concatenated at each level. In the structure of Lemma 2, the redundancy at each
level was merely O(lg lg n) bits. However, with Lemma 1 we end up with a redun-
dancy of O(n/( lgc (n))) bits per level, for a total of O(n lg n/( lgc (n))) bits.
So, if 1/ = O(polylog(n)), then we can choose the constant c to be suciently
large so that this term is sublinear. Immediately, this yields:
Lemma 5. If 1/ = O(polylog(n)), there is an encoding that supports range
-majority position queries in O(1/ ) time, and occupies O(n lg(1/ )) bits.
When 1/ is (polylog(n)), we require a more sophisticated data structure to
achieve O(n lg(1/ )) bits of space. Basically, we have to replace the data struc-
ture of Lemma 1 representing the bit vectors with a more space-ecient batch
structure that groups all candidates together. The details required to nish the
proof of Theorem 2 can be found in the full version of this paper [9].

4 Lower Bound
In this section we prove Theorem 4. The idea is to show that a sequence of
permutations, each of length roughly 1/ , can be recovered using queries.
Formally, we will describe a bad string, dened using concatenation, in which
array A[i] will store the i-th symbol in the string. Conceptually, this bad string
is constructed by concatenating some padding, denoted L, before a sequence
of m permutations over the alphabet [1 , . . . , k ], denoted R = 1 . . . m .
Notationally, we use ic to denote a concatenation of the symbol i c times,
and a b to denote the concatenation of the strings a and b. In the construction
we make use of dummy symbols, , which are dened to be symbols that occur
exactly one time in the bad string. A sequence of  dummy symbols, written  ,
should be taken to mean: a sequence of  characters, each of which are distinct
from any other symbol in A.

Padding denition. Key to dening L is a gadget G(k, i), that is dened for
 
k
any integer k 2 using concatenation as follows: G(k, i) = 1k 2k . . . i1
 
k
i+1 . . . kk (i k2 )k1 i k , where k  = k 2 k + 2. Suppose we dene
A such that A[x..y] contains gadget G(k, i). Let f (x, y, ) denote the number of
occurrences of symbol in range [x, y]. We dene the density of symbol in
the query range [x, y] to be (x, y, ) = f (x, y, )/(y x + 1). We observe the
following:
1. The length of the gadget G(k, i) is k(k2 k + 2) for all i [1, k]. This fact
will be useful later when we bound the total size of the padding L.
Optimal Query Time for Encoding Range Majority 419

2. (x, y, j ) = 1/k for all j = i. This follows from the previous observation and
that, for all j = i, the number of occurrences of j in G(k, i) is k 2 k + 2.
Next, we nish dening our array A by dening L to be the concatenation
G(k, k) G(k, k 1) . . . G(k, 1). Thus, our array is obtained by embedding
the string L R into an array A. Note that the total length of the array is
k 2 (k 2 k + 2) + mk. Thus, the padding is of length (k 4 ).

Query Procedure. The following procedure can recover the position of symbol i
in j , for any i [1, k] and j [1, m]. This procedure uses (k) (1/k)-majority
decision queries: overall, recovering the contents of R uses (k 2 m) queries.
Let rj,1 , ..., rj,k denote the indices of A containing the symbols in j from
left-to-right. Moreover, consider the indices of the k occurrences of symbol i
in G(k, i), from left-to-right, and denote these as i,k , . . . , i,1 , respectively (note
that the rightmost occurrence is marked with subscript 1). Formally, the query
procedure will perform a sequence of queries, stopping if the answer is YES, and
continuing if the answer is NO. The ordered sequence of queries we execute is
[i,1 , rj,1 ], [i,2 , rj,2 ], . . . , [i,k , rj,k ].
We now claim that if the answer to a query [i,x , rj,x ] is NO, then A[rj,x ] = i .
This follows since the density of symbol i in the query range is:
x + (i 1)(k 2 k + 2) + (j 1) 1
2
<
k(x + (i 1)(k k + 2) + (j 1)) + 2 k
On the other hand, if the answer is YES, we have that the symbol i must be a
(1/k)-majority for the following reasons:
1. No other symbol j where j = i can be a (1/k)-majority. To see this, divide
the query range into a middle-part, consisting of G(k, i 1) . . . G(k, 1) 1
. . . k1 , as well as a prex (which is a sux of G(k, i)), and a sux (which
is a prex of j ). The prex of the query range contains no occurrence of j
and is at least of length k + 1. The sux contains at most one occurrence of
j . Thus, the density of j is strictly less than 1/k in the union of the prex
and sux, exactly 1/k in the middle part, and strictly less than 1/k overall.
2. No dummy symbol can be an (1/k)-majority, since these symbols appear
one time only, and all query ranges have length strictly larger than k.
3. Finally, if A[rj,x ] = i , then the density (i,x , rj,x , i ) is:
x + (i 1)(k 2 k + 2) + (j 1) + 1
1/k,
k(x + (i 1)(k 2 k + 2) + (j 1)) + 2
since k 2. Since we stop immediately after the rst YES, the procedure
therefore is guaranteed to identify the correct position of i .
As we stated, the length of the array is k 2 (k 2 k + 2) + mk = n, and for n
4
)
large enough the queries allow us to recover n(k k lg(k!) bits of information
using (1/k)-majority queries for any integer k 2, which is at least (n/k
(k 3 ))k lg(k/e) = n lg(k/e) (k 4 lg k) bits. Since there exists a unit fraction
 = 1/1/  (if (0, 1/2]), there also exists a bad input of length n in which
k = 1/  . Therefore, we have proved Theorem 4.
420 P. Gawrychowski and P.K. Nicholson

References
1. Belazzougui, D., Gagie, T., Munro, J.I., Navarro, G., Nekrich, Y.: Range majorities
and minorities in arrays. CoRR abs/1606.04495 (2016)
2. Belazzougui, D., Gagie, T., Navarro, G.: Better space bounds for parameterized
range majority and minority. In: Proc. WADS 2013. LNCS, vol. 8037, pp. 121132.
Springer (2013)
3. Boyer, R.S., Moore, J.S.: MJRTY: A fast majority vote algorithm. In: Automated
Reasoning: Essays in Honor of Woody Bledsoe. pp. 105118. Automated Reasoning
Series, Kluwer Academic Publishers (1991)
4. Chan, T.M., Durocher, S., Larsen, K.G., Morrison, J., Wilkinson, B.T.: Linear-
space data structures for range mode query in arrays. Theory Comput. Syst. 55(4),
719741 (2014)
5. Demaine, E.D., Lopez-Ortiz, A., Munro, J.I.: Frequency estimation of internet
packet streams with limited space. In: Proc. ESA 2002. LNCS, vol. 2461, pp. 348
360. Springer (2002)
6. Durocher, S., He, M., Munro, J.I., Nicholson, P.K., Skala, M.: Range majority in
constant time and linear space. Inf. Comput. 222, 169179 (2013)
7. Ferragina, P., Venturini, R.: A simple storage scheme for strings achieving entropy
bounds. Theor. Comput. Sci. 372(1), 115121 (2007)
8. Gagie, T., He, M., Munro, J.I., Nicholson, P.K.: Finding frequent elements in com-
pressed 2d arrays and strings. In: Proc. SPIRE 2011. LNCS, vol. 7024, pp. 295300.
Springer (2011)
9. Gawrychowski, P., Nicholson, P.K.: Optimal query time for encoding range major-
ity. CoRR arXiv:1704.06149 (2017), http://arxiv.org/abs/1704.06149
10. Gonzalez, R., Navarro, G.: Statistical encoding of succinct data structures. In:
Proc. CPM 2006. LNCS, vol. 4009, pp. 294305. Springer (2006)
11. Grossi, R., Gupta, A., Vitter, J.S.: High-order entropy-compressed text indexes.
In: Proc. SODA 2003. pp. 841850. ACM/SIAM (2003)
12. Karp, R.M., Shenker, S., Papadimitriou, C.H.: A simple algorithm for nding fre-
quent elements in streams and bags. ACM Trans. Database Syst. 28, 5155 (2003)
13. Karpinski, M., Nekrich, Y.: Searching for frequent colors in rectangles. In: Proc.
CCCG 2008 (2008)
14. Misra, J., Gries, D.: Finding repeated elements. Sci. Comput. Program. 2(2), 143
152 (1982)
15. Navarro, G., Thankachan, S.V.: Optimal encodings for range majority queries.
Algorithmica 74(3), 10821098 (2016)
16. Patrascu, M.: Succincter. In: Proc. FOCS 2008. pp. 305313. IEEE (2008)
17. Raman, R.: Encoding data structures. In: Proc. WALCOM 2015. LNCS, vol. 8973,
pp. 17. Springer (2015)
18. Raman, R., Raman, V., Satti, S.R.: Succinct indexable dictionaries with applica-
tions to encoding k -ary trees, prex sums and multisets. ACM Trans. Algorithms
3(4), 43 (2007)
19. Skala, M.: Array range queries. In: Space-Ecient Data Structures, Streams, and
Algorithms - Papers in Honor of J. Ian Munro on the Occasion of His 66th Birthday.
LNCS, vol. 8066, pp. 333350. Springer (2013)
 Conditional Lower Bounds for Space/Time
Tradeos
Isaac Goldstein1 , Tsvi Kopelowitz2 , Moshe Lewenstein  1 , and
Ely Porat   1
1
Bar-Ilan University , {goldshi,moshe,porately}@cs.biu.ac.il
2
University of Waterloo , [email protected]

Abstract. In recent years much eort has been concentrated towards

achieving polynomial time lower bounds on algorithms for solving various
well-known problems. A useful technique for showing such lower bounds
is to prove them conditionally based on well-studied hardness assump-
tions such as 3SUM, APSP, SETH, etc. This line of research helps to
obtain a better understanding of the complexity inside P.
A related question asks to prove conditional space lower bounds on data
structures that are constructed to solve certain algorithmic tasks after
an initial preprocessing stage. This question received little attention in
previous research even though it has potential strong impact.
In this paper we address this question and show that surprisingly many
of the well-studied hard problems that are known to have conditional
polynomial time lower bounds are also hard when concerning space. This
hardness is shown as a tradeo between the space consumed by the data
structure and the time needed to answer queries. The tradeo may be
either smooth or admit one or more singularity points.
We reveal interesting connections between dierent space hardness con-
jectures and present matching upper bounds. We also apply these hard-
ness conjectures to both static and dynamic problems and prove their
conditional space hardness.
We believe that this novel framework of polynomial space conjectures
can play an important role in expressing polynomial space lower bounds
of many important algorithmic problems. Moreover, it seems that it can
also help in achieving a better understanding of the hardness of their
corresponding problems in terms of time.

This research is supported by the Adams Foundation of the Israel Academy of
Sciences and Humanities

Part of this work took place while the second author was at University of Michigan.
This work is supported in part by the Canada Research Chair for Algorithm Design,
NSF grants CCF-1217338, CNS-1318294, and CCF-1514383

This work was partially supported by an ISF grant #1278/16

Springer International Publishing AG 2017 421

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 421436, 2017.
DOI: 10.1007/978-3-319-62127-2_36
422 I. Goldstein et al.

1 Introduction

1.1 Background

Lately there has been a concentrated eort to understand the time complexity
within P, the class of decision problems solvable by polynomial time algorithms.
The main goal is to explain why certain problems have time complexity that
seems to be non-optimal. For example, all known ecient algorithmic solutions
for the 3SUM problem, where we seek to determine whether there are three
elements x, y, z in input set S of size n such that x + y + z = 0, take O(n2 )
time1 . However, the only real lower bound that we know is the trivial (n).
Likewise, we know how to solve the all pairs shortest path, APSP, problem in
O(n3 ) time but we cannot even determine whether it is impossible to obtain
an O(n2 ) time algorithm. One may note that it follows from the time-hierarchy
theorem that there exist problems in P with complexity (nk ) for every xed k.
Nevertheless, such a separation for natural practical problems seems to be hard
to achieve.
The collaborated eort to understand the internals of P has been concen-
trated on identifying some basic problems that are conjectured to be hard to
solve more eciently (by polynomial factors) than their current known com-
plexity. These problems serve as a basis to prove conditional hardness of other
problems by using reductions. The reductions are reminiscent of NP-complete
reductions but dier in that they are restricted to be of time complexity strictly
smaller (by a polynomial factor) than the problem that we are reducing to. Ex-
amples of such hard problems include the well-known 3SUM problem, the fun-
damental APSP problem, (combinatorial) Boolean matrix multiplication, etc.
Recently, conditional time lower bounds have been proven based on the conjec-
tured hardness of these problems for graph algorithms [4, 30], edit distance [12],
longest common subsequence (LCS) [3, 14], dynamic algorithms [5, 25], jumbled
indexing [11, 19], and many other problems [1, 2, 6, 7, 13, 20, 23, 24, 29].

1.2 Motivation

In stark contrast to polynomial time lower bounds, little eort has been devoted
to nding polynomial space conditional lower bounds. An example of a space
lower bound appears in the work of Cohen and Porat [17] and Patrascu and
Roditty [27] where lower bounds are shown on the size of a distance oracle for
sparse graphs based on a conjecture about the best possible data structure for a
set intersection problem (which we call set disjointness in order to dier it from
its reporting variant).
A more general question is, for algorithmic problems, what conditional lower
bounds of a space/time tradeo can be shown based on the set disjointness (inter-
section) conjecture? Even more general is to discover what space/time tradeos
can be achieved based on the other algorithmic problems that we assumed are
1
The O and notations suppress polylogarithmic factors
Conditional Lower Bounds for Space/Time Tradeoffs 423

hard (in the time sense)? Also, what are the relations between these identied
hard problems in the space/time tradeo sense? These are the questions which
form the basis and framework of this paper.
Throughout this paper we show connections between dierent hardness as-
sumptions, show some matching upper bounds and propose several conjectures
based on this accumulated knowledge. Moreover, we conjecture that there is
a strong correlation between polynomial hardness in time and space. We note
that in order to discuss space it is often more natural to consider data structure
variants of problems and this is the approach we follow in this paper.

1.3 Our Results

Set Disjointness. In the SetDisjointness problem mentioned before, it is required

to preprocess a collection of m sets S1 , , Sm U , where U is the universe of
elements and the total number of elements in all sets is N . For a query, a pair of
integers (i, j) (1 i, j m) is given and we are asked whether Si Sj is empty
or not. A folklore conjecture, which appears in [16, 27], suggests that to achieve
a constant query time the space of the data structure constructed in the pre-
processing stage needs to be (N 2 ). We call this conjecture the SetDisjointness
conjecture. This conjecture does not say anything about the case where we allow
higher query time. Therefore, we suggest a stronger conjecture which admits a
full tradeo between the space consumed by the data structure (denoted by S)
and the query time (denoted by T ). This is what we call the Strong SetDis-
jointness conjecture. This conjecture states that for solving SetDisjointness with
a query time T our data structure needs (N 2 /T 2 ) space. A matching upper
bound exists for this problem by generalizing ideas from [16] (see also [22]). Our
new SetDisjointness conjecture can be used to admit more expressive space lower
bounds for a full tradeo between space and query time.

3SUM Indexing. One of the basic and frequently used hardness conjectures is
the celebrated 3SUM conjecture. This conjecture was used for about 20 years
to show many conditional time lower bounds on various problems. However, we
focus on what can be said about its space behavior. To do this, it is natural to
consider a data structure version of 3SUM which allows one to preprocess the
input set S. Then, the query is an external number z for which we need to answer
whether there are x, y S such that x + y = z. It was pointed out by Chan
and Lewenstein [15] that all known algorithms for 3SUM actually work within
this model as well. We call this problem 3SUM Indexing. On one hand, this
problem can easily be solved using O(n2 ) space by sorting x + y for all x, y S
and then searching for z in O(1) time. On the other hand, by just sorting S
we can answer queries by a well-known linear time algorithm. The big question
is whether we can obtain better than (n2 ) space while using just O(1) time
query? Can it be done even if we allow O(n1(1) ) query time? This leads us
to our two new hardness conjectures. The 3SUM-Indexing conjecture states that
when using O(1) query time we need (n2 ) space to solve 3SUM-Indexing. In
424 I. Goldstein et al.

the Strong 3SUM-Indexing conjecture we say that even when using O(n1(1) )
query time we need (n2 ) space to solve 3SUM-Indexing.

3SUM Indexing and Set Disjointness. We prove connections between the SetDis-
jointness conjectures and the 3SUM-Indexing conjectures. Specically, we show
that the Strong 3SUM-Indexing conjecture implies the Strong SetDisjointness
conjecture, while the SetDisjointness conjecture implies the 3SUM-Indexing con-
jecture. This gives some evidence towards establishing the diculty within the
3SUM-Indexing conjectures. The usefulness of these conjectures should not be
underestimated. As many problems are known to be 3SUM-hard these new con-
jectures can play an important role in achieving space lower bounds on their
corresponding data structure variants. Moreover, it is interesting to point on the
dierence between SetDisjointness which admits smooth tradeo between space
and query time and 3SUM-Indexing which admits a big gap between the two
trivial extremes. This may explain why we are unable to show full equivalence
between the hardness conjectures of the two problems. Moreover, it can suggest a
separation between problems with smooth space-time behavior and others which
have no such tradeo but rather two far extremes.

Generalizations. Following the discussion on the SetDisjointness and the 3SUM-

Indexing conjectures we investigate their generalizations.

I. k-Set Disjointness and (k+1)-SUM Indexing. The rst generalization is a

natural parametrization of both problems. In the SetDisjointness problem we
query about the emptiness of the intersection between two sets, while in the
3SUM-Indexing problem we ask, given a query number z, whether two numbers
of the input S sum up to z. In the parameterized versions of these problems we
are interested in the emptiness of the intersection between k sets and ask if k
numbers sum up to a number given as a query. These generalized variants are
called k-SetDisjointness and (k+1)-SUM-Indexing respectively. For each problem
we give corresponding space lower bounds conjectures which generalize those of
SetDisjointness and 3SUM-Indexing. These conjectures also have corresponding
strong variants which are accompanied by matching upper bounds. We prove
that the k-SetDisjointness conjecture implies (k+1)-SUM-Indexing conjecture via
a novel method using linear equations.

II. k-Reachability. A second generalization is the problem we call k-Reachability.

In this problem we are given as an input a directed sparse graph G = (V, E) for
preprocessing. Afterwards, for a query, given as a pair of vertices u, v, we wish
to return if there is a path from u to v consisting of at most k edges. We provide
an upper bound on this problem for every xed k 1. The upper bound admits
a tradeo between the space of the data structure (denoted by S) and the query
time (denoted by T ), which is ST 2/(k1) = O(n2 ). We argue that this upper
bound is tight. That is, we conjecture that if query takes T time, the space must
n2
be ( T 2/(k1) ). We call this conjecture the k-Reachability conjecture.
Conditional Lower Bounds for Space/Time Tradeoffs 425

We give three indications towards the correctness of this conjecture. First,

we prove that the base case, where k = 2, is equivalent to the SetDisjointness
problem. This is why this problem can be thought of as a generalization of
SetDisjointness.
Second, if we consider non-constant k then the smooth tradeo surprisingly
n2
disappears and we get extreme behavior as ( T 2/(k1) ) eventually becomes
2
(n ). This means that to answer reachability queries for non-constant path
length, we can either store all answers in advance using n2 space or simply
answer queries from scratch using a standard graph traversal algorithm. The
general problem where the length of the path from u to v is unlimited in length
is sometimes referred to as the problem of constructing ecient reachability
oracles. Patrascu in [26] leaves it as an open question if a data structure with less
than (n2 ) space can answer reachability queries eciently. Moreover, Patrascu
proved that for constant time query, truly superlinear space is needed. Our k-
Reachability conjecture points to this direction, while admitting full space-time
tradeo for constant k.
The third indication for the correctness of the k-Reachability conjecture comes
from a connection to distance oracles. A distance oracle is a data structure that
can be used to quickly answer queries about the shortest path between two
given nodes in a preprocessed undirected graph. As mentioned above, the Set-
Disjointness conjecture was used to exclude some possible tradeos for sparse
graphs. Specically, Cohen and Porat [17] showed that obtaining an approxima-
tion ratio smaller than 2 with constant query time requires (n2 ) space. Using
a somewhat stronger conjecture Patrascu and Roditty [27] showed that a (2,1)-
distance oracle for unweighted graphs with m = O(n) edges requires (n1.5 )
space. Later, this result was strengthened by Patrascu et al. [28]. However, these
results do not exclude the possibility of compact distance oracles if we allow
higher query time. For stretch-2 and stretch-3 in sparse graphs, Agarwal et.
al. [9, 10] achieved a space-time tradeo of S T = O(n2 ) and S T 2 = O(n2 ),
respectively. Agarwal [8] also showed many other results for stretch-2 and below.
We use our k-Reachability conjecture to prove that for stretch-less-than-(1+2/k)
distance oracles S T 2/(k1) is bounded by (n2 ). This result is interesting in
light of Agarwal [8] where a stretch-(5/3) oracle was presented which achieves
a space-time tradeo of S T = O(n2 ). This matches our lower bound, where
k = 3, if our lower bound would hold not only for stretch-less-than-(5/3) but
also for stretch-(5/3) oracles. Consequently, we see that there is strong evidence
for the correctness of the k-Reachability conjecture.
Moreover, these observations show that on one hand k-Reachability is a gen-
eralization of SetDisjointness which is closely related to 3SUM-Indexing. On the
other hand, k-Reachability is related to distance oracles which solve the famous
APSP problem using smaller space by sacricing the accuracy of the distance
between the vertices. Therefore, the k-Reachability conjecture seems as a conjec-
ture corresponding to the APSP hardness conjecture, while also admitting some
connection with the celebrated 3SUM hardness conjecture.
426 I. Goldstein et al.

SETH and Orthogonal Vectors. After considering space variants of the 3SUM
and APSP conjectures it is natural to consider space variants for the Strong
Exponential Time Hypothesis (SETH) and the closely related conjecture of or-
thogonal vectors. SETH asserts that for any  > 0 there is an integer k > 3
such that k-SAT cannot be solved in 2(1)n time. The orthogonal vectors time
conjecture states that there is no algorithm that for every c 1, nds if there are
at least two orthogonal vectors in a set of n Boolean vectors of length c log n in
O(n2(1) ) time. A discussion about the space variants of these conjectures will
appear in the full version of this paper. However, we note that we are unable to
connect these conjectures and the previous ones. This is perhaps not surprising
as the connection between SETH and the other conjectures even in the time
perspective is very loose (see, for example, discussions in [5, 20]).

Boolean Matrix Multiplication. Another problem which receives a lot of atten-

tion in the context of conditional time lower bounds is calculating Boolean Ma-
trix Multiplication (BMM). We give a data structure variant of this well-known
problem. We then demonstrate the connection between this problem and the
problems of SetDisjointness and k-Reachability. The discussion about BMM and
its data structure variant will appear in the full version of this paper.

Applications. Finally, armed with the space variants of many well-known con-
ditional time lower bounds, we apply this conditional space lower bounds to
some static and dynamic problems. This gives interesting space lower bound re-
sults on these important problems which sometimes also admits clear space-time
tradeo. The list of problems that we prove their conditional space-time hard-
ness includes: edge triangles, histogram indexing, distance oracles for colors,
two patterns document retrieval, forbidden pattern document retrieval, (s,t)-
reachability, bipartite perfect matching and strong connectivity. All the results
regarding the applications of our framework will appear in the full version of this
paper. We believe that this is just a glimpse of space lower bounds that can be
achieved based on our new framework and that many other interesting results
are expected to follow this promising route.

2 Set Intersection Hardness Conjectures

We rst give formal denitions of the SetDisjointness problem and its enumera-
tion variant:

Problem 1 (SetDisjointness Problem).

 Preprocess a family F of m sets, all from
universe U , with total size N = SF |S| so that given two query sets S, S  F
one can determine if S S  = .

Problem 2 (SetIntersection Problem).

 Preprocess a family F of m sets, all from
universe U , with total size N = SF |S| so that given two query sets S, S  F
one can enumerate the set S S  .
Conditional Lower Bounds for Space/Time Tradeoffs 427

Conjectures. The SetDisjointness problem was regarded as a problem that ad-

mits space hardness. The hardness conjecture of the SetDisjointness problem has
received several closely related formulations. One such formulation, given by
Patrascu and Roditty [27], is as follows:

Conjecture 1. SetDisjointness Conjecture [Formulation 1]. Any data struc-

ture for the SetDisjointness problem where |U | = logc m for a large enough con-
stant c and with a constant query time must use (m2 ) space.

Another formulation is implicitly suggested in Cohen and Porat [16]:

Conjecture 2. SetDisjointness Conjecture [Formulation 2]. Any data struc-

ture for the SetDisjointness problem with constant query time must use (N 2 )
space.

There is an important distinction between the two formulations, which is

related to the sparsity of SetDisjointness instances. This distinction follows from
the following upper bound: store an m m matrix of the answers to all possible
queries, and then queries will cost constant time. The rst formulation of the
SetDisjointness conjecture states that if we want constant (or poly-logaritmic)
query time, then this is the best we can do. At a rst glance this makes the
second formulation, whose bounds are in terms of N and not m, look rather
weak. In particular, why would we ever be interested in a data structure that
uses O(N 2 ) space when we can use one with O(m2 ) space? The answer is that
the two conjectures are the same if the sets are very sparse, and so at least in
terms of N , if one were to require a constant query time then by the second
formulation the space must be at least (N 2 ) (which happens in the very sparse
case).
Nevertheless, we present a more general conjecture, which in particular cap-
tures a tradeo curve between the space usage and query time. This formulation
captures the diculty that is commonly believed to arise from the SetDisjointness
problem, and matches the upper bounds of Cohen and Porat [16] (see also [22]).

Conjecture 3. Strong SetDisjointness Conjecture. Any data structure for the

2
SetDisjointness problem that answers queries in T time must use S = ( N
T2 )
space.

For example, a natural question to ask is what is the smallest query time
possible with linear space?. This question is addressed, at least from a lower
bound perspective, by the Strong SetDisjointness conjecture.

Conjecture 4. Strong SetIntersection Conjecture. Any data structure for the

SetIntersection problem that answers queries in O(T + op) time, where op is the
2
size of the output of the query, must use S = ( NT ) space.
428 I. Goldstein et al.

3 3SUM-Indexing Hardness Conjectures

In the classic 3SUM problem we are given an integer array A of size n and
we wish to decide whether there are 3 distinct integers in A which sum up to
zero. Gajentaan and Overmars [18] showed that an equivalent formulation of
this problem receives 3 integer arrays A1 , A2 , and A3 , each of size n, and the
goal is to decide if there is a triplet x1 A1 , x2 A2 , and x3 A3 that sum up
to zero.
We consider the data structure variant of this problem which is formally
dened as follows:

Problem 3 (3SUM-Indexing Problem). Preprocess two integer arrays A1 and A2 ,

each of length n, so that given a query integer z we can decide whether there
are x A1 and y A2 such that z = x + y.

It is straightforward to maintain all possible O(n2 ) sums of pairs in quadratic

space, and then answer a query in O(1) time. On the other extreme, if one does
not wish to utilize more than linear space then one can sort the arrays separately
during preprocssing time, and then a query can be answered in O(n) time by
scanning both of the sorted arrays in parallel and in opposite directions.
We introduce two conjectures with regards to the 3SUM-Indexing problem,
which serve as natural candidates for proving polynomial space lower bounds.

Conjecture 5. 3SUM-Indexing Conjecture: There is no solution for the 3SUM-

Indexing problem with truly subquadratic space and O(1) query time.

Conjecture 6. Strong 3SUM-Indexing Conjecture: There is no solution for

the 3SUM-Indexing problem with truly subquadratic space and truly sublinear
query time.

Notice that one can solve the classic 3SUM problem using a data structure
for 3SUM-Indexing by preprocessing A1 and A2 , and answering n 3SUM-Indexing
queries on all of the values in A3 .
Next, we prove theorems that show tight connections between the 3SUM-
Indexing conjectures and the SetDisjointness conjectures. We note that the proofs
of the rst two theorems are similar to the proofs of [23], but with space inter-
pretation. These proofs will appear in the full version of this paper.

Theorem 1. The Strong 3SUM-Indexing Conjecture implies the Strong SetDis-

jointness Conjecture.

Theorem 2. The Strong 3SUM-Indexing Conjecture implies the Strong SetInter-

section Conjecture.

Theorem 3. The SetDisjointness Conjecture implies the 3SUM-Indexing Con-

jecture.
Conditional Lower Bounds for Space/Time Tradeoffs 429

Proof. Given an instance of SetDisjointness, we construct an instance of 3SUM-

Indexing as follows. Denote with M the value of the largest element in the Set-
Disjointness instance. Notice that we may assume that M N (otherwise we can
use a straightforward renaming). For every element x U that is contained in
at least one of the sets we create two integers xA and xB , which are represented
by 2log m + log N + 3 bits each (recall that m is the number of sets).
The log N least signicant bits in xA represent the value of x. The following
bit is a zero. The following log m bits in xA represent the index of the set
containing x, and the rest of the 2 + log m are all set to zero. The log N least
signicant bits in xB represent the value of M x. The following 2 + log m
are all set to zero. The following log m bits in xB represent the index of the
set containing x, and the last bit is set to zero. Finally, the integer xA is added
to A1 of the 3SUM-Indexing instance, while the integer xB is added to A2 .
We have created two sets of n M integers. We then preprocess them to
answer 3SUM-Indexing queries. Now, to answer a SetDisjointness query on sets Si
and Sj , we query the 3SUM-Indexing data structure with an integer z which is
determined as follows. The log N least signicant bits in z represent the value
of M . The following bit is a zero. The following log m bits represent the index
i and are followed by a zero. The next log m bits represent the index j and
the last bit is set to zero.
It is straightforward to verify that there exists a solution to the 3SUM-
Indexing problem on z if and only if the sets Si and Sj are not disjoint. Therefore,
if there is a solution to the 3SUM-Indexing problem with less than (n2 ) space
and constant query time then there is a solution for the SetDisjointness problem
which refutes the SetDisjointness Conjecture.

4 Parameterized Generalization:
k-Set Intersection and (k+1)-SUM
Two parameterized generalizations of the SetDisjointness and 3SUM-Indexing
problems are formally dened as follows:

Problem 4 (k-SetDisjointness Problem). Preprocess

 a family F of m sets, all
from universe U , with total size N = SF |S| so that given k query sets
S1 , S2 , . . . , Sk F one can quickly determine if ki=1 Si = .

Problem 5 ((k+1)-SUM-Indexing Problem). Preprocess k integer arrays A1 , A2 , . . . , Ak ,

each of length n, so that given a query integer k z we can decide if there is
x1 A1 , x2 A2 , . . . , xk Ak such that z = i=1 xi .

It turn out that a natural generalization of the data structure of Cohen and
Porat [16] leads to a data structure for k-SetDisjointness as shown in the following
lemma.
Lemma 1. There exists a data structure for the k-SetDisjointness problem where
the query time is T and the space usage is S = O((N/T )k ).
430 I. Goldstein et al.

Proof. We call the f largest sets in F large sets. The rest of the sets are called
small sets. In the preprocessing stage we explicitly maintain a k-dimensional
table with the answers for all k-SetDisjointness queries where all k sets are large
sets. The space needed for such a table is S = f k . Moreover, for each set (large
or small) we maintain a look-up table that supports disjointness queries (with
this set) in constant time. Since there are f large sets and the total number of
elements is N , the size of each of the small sets is at most N/f .
Given a k-SetDisjointness query, if all of the query sets are large then we look
up the answer in the k-dimensional table. If at least one of the sets is small then
using a brute-force search we look-up each of the at most O(N/f ) elements in
each of the other k 1 sets. Thus, the total query time is bounded by O(kN/f ),
and the space usage is S = O(f k ). The rest follows.

Notice that for the case of k = 2 in Lemma 1 we obtain the same tradeo of
Cohen and Porat [16] for SetDisjointness. The following conjecture suggests that
the upper bound of Lemma 1 is the best possible.

Conjecture 7. Strong k-SetDisjointness Conjecture. Any data structure for

the k-SetDisjointness problem that answers queries in T time must use S =
k
( N
Tk
) space.

Similarly, a natural generalization of the Strong 3SUM-Indexing conjecture is

the following.
Conjecture 8. Strong (k+1)-SUM-Indexing Conjecture. There is no solution
for the (k+1)-SUM-Indexing problem with O(nk(1) ) space and truly sublinear
query time.
We also consider some weaker conjectures, similar to the SetDisjointness and
3SUM-Indexing conjectures.

Conjecture 9. k-SetDisjointness Conjecture. Any data structure for the k-

SetDisjointness problem that answers queries in constant time must use (N k )
space.

Conjecture 10. (k+1)-SUM-Indexing Conjecture. There is no solution for the

(k+1)-SUM-Indexing problem with O(nk(1) ) space and constant query time.

Similar to Theorem 3, we prove the following relationship between the k-

SetDisjointness conjecture and the (k+1)-SUM-Indexing conjecture.

Theorem 4. The k-SetDisjointness conjecture implies the (k+1)-SUM-Indexing

conjecture

Proof. Given an instance of k-SetDisjointness, we construct k instances of (k+1)-

SUM-Indexing as follows. Denote by M the value of the largest element in the
SetDisjointness instance. Notice that we may assume that M N (otherwise we
use a straightforward renaming). For every element x U that is contained in
Conditional Lower Bounds for Space/Time Tradeoffs 431

at least one of the sets we create k 2 integers in a matrix X = {xi,j } of size k k,

where each integer is represented by (k 1)log m + log N + k bits.
For integer xi,j , the log N + 1 least signicant bits represent the value of
(k 1)x if i = j, and the value of M x otherwise. The (k 1)log m + k 1
following bits are all set to zero, except for the bits in indices (j 1)(log m +
1) + 1, ..., j(log m + 1) which represent the index of the set containing x.
We now create k instances of (k+1)-SUM-Indexing where the jth input array
Aj for the ith instance is the set of integers xi,j for all x U that are contained in
at least one set of our family. Thus, the size of each array is at most N . Now, given
a k-SetDisjointness query (i1 , i2 , ..., ik ) we must decide if Si1 Si2 ...Sik = . To
answer this query we will query each of the k instances of (k+1)-SUM-Indexing
with an integer z whose binary representation is as follows: the log N + 1
least signicant bits represent the value of (k 1)M , and the bits at locations
(j 1)(log m + 1) + 1, ..., j(log m + 1) representing ij (for 1 j k). The
rest of the bits are padding zero bit (in between representations of various ij ).
If Si1 Si2 ... Sik = then by our construction it is straightforward to
verify that all of the k (k+1)-SUM-Indexing queries on z will return that there
is a solution. If Si1 Si2 ... Sik = then at least one (k+1)-SUM-Indexing
query will not be able to nd a solution. This is because we can view each
instance and query as solving a linear equation. As we construct k instances
which represent k independent linear equations, we are guaranteed that only
one solution exists. This solution is exactly the one that corresponds to nding
a specic x which is contained in all of the k sets. Therefore, we get a correct
answer to a k-SetDisjointness query by answering k (k+1)-SUM-Indexing queries.
Consequently, if for some specic constant k there is a solution to the (k+1)-
SUM-Indexing problem with less than (nk ) space and constant query time, then
with this reduction we refute the k-SetDisjointness conjecture.

5 Directed Reachability Oracles as a Generalization of

Set Disjointness Conjecture

An open question which was stated by Patrascu in [26] asks if it is possible to

preprocess a sparse directed graph in less than (n2 ) space so that Reachability
queries (given two query vertices u and v decide whether there is a path from u to
v or not) can be answered eciently. A partial answer, given in [26], states that
for constant query time truly superlinear space is necessary. In the undirected
case the question is trivial and one can answer queries in constant time using
linear space. This is also possible for planar directed graphs (see Holm et al. [21]).
We now show that Reachability oracles for sparse graphs can serve as a gener-
alization of the SetDisjointness conjecture. We dene the following parameterized
version of Reachability. In the k-Reachability problem the goal is to preprocess a
directed sparse graph G = (V, E) so that given a pair of distinct vertices u, v V
one can quickly answer whether there is a path from u to v consisting of at most
k edges. We prove that 2-Reachability and SetDisjointness are tightly connected.
432 I. Goldstein et al.

Lemma 2. There is a linear time reduction from SetDisjointness to 2-Reachability

and vice versa which preserves the size of the instance.

Proof. Given a graph G = (V, E) as an instance for 2-Reachability, we construct a

corresponding instance of SetDisjointness as follows. For each vertex v we create
the sets Vin = {u|(u, v) E} and Vout = {u|(v, u) E} {v}. We have 2n sets
and 2m + n elements in all of them (|V | = n and |E| = m). Now, a query u, v is
reduced to determining if the sets Uout and Vin are disjoint or not. Notice, that
the construction is done in linear time and preserves the size of the instance. In
the opposite direction, we are given m sets S1 , S2 , ..., Sm having N elements in
total e1 , e2 , ..., eN . We can create an instance of 2-Reachability in the following
way. For each set Si we create a vertex vi . Moreover, for each element ej we create
a vertex uj . Then, for each element ej in a set si we create two directed edges
(vi , uj ) and (uj , vi ). These vertices and edges dene a directed graph, which is
preprocessed for 2-Reachability queries. It is straightforward to verify that the
disjointness of Si and Sj is equivalent to determining if there is a path of length
at most 2 edges from vi to vj . Moreover, the construction is done in linear time
and preserves the size of the instance.

Furthermore, we consider k-Reachability for k 3. First we show an upper

bound on the tradeo between space and query time for solving k-Reachability.

Lemma 3. There exists a data structure for k-Reachability with S space and T
query time such that ST 2/(k1) = O(n2 ).

Proof. Let > 0 be an integer parameter to be set later. Given a directed graph
G = (V, E), we call vertex v V a heavy vertex if deg(v) > and a vertex
u V a light vertex if deg(u) . Notice that the number of heavy vertices
is at most n/. For all heavy vertices in V we maintain a matrix containing
the answers to any k-Reachability query between two heavy vertices. This uses
O(n2 /2 ) space.
Next, we recursively construct a data structure for (k-1)-Reachability. Given a
query u, v, if both vertices are heavy then the answer is obtained from the matrix.
Otherwise, either u or v is light vertex. Without loss of generality, say u is a light
vertex. We consider each vertex w Nout (u) (Nout (u) = {v|(u, v) E}) and
query the (k-1)-Reachability data structure with the pair w, v. Since u is a light
node, there are no more than queries. One of the queries returns a positive
answer if and only if there exists a path of length at most k from u to v.
Denote by S(k, n) the space used by our k-Reachability oracle on a graph
with n vertices and denote by Q(k, n) the corresponding query time. In our
construction we have S(k, n) = n2 /2 + S(k 1, n) and Q(k, n) = Q(k
1, n) + O(1). For k = 1 it is easy to construct a linear space data structure using
hashing so that queries can be answered in constant time. Thus, S = S(k, n) =
O((k 1)n2 /2 ) and T = Q(k, n) = O(k1 ).

Notice that for the case of k = 2 the upper bounds from Lemma 3 exactly
match the tradeo of the Strong SetDisjointness Conjecture (ST 2 = O(n2 )). We
Conditional Lower Bounds for Space/Time Tradeoffs 433

expand this conjecture by considering the tightness of our upper bound for k-
Reachability, which then leads to some interesting consequences with regard to
distance oracles.

Conjecture 11. Directed k-Reachability Conjecture. Any data structure for

n2
the k-Reachability problem with query time T must use S = ( T 2/(k1) ) space.

Notice that when k is non-constant then by our upper bound (n2 ) space is
necessary independent of the query time. This ts nicely with what is currently
known about the general question of Reachability oracles: either we spend n2
space and answer queries in constant time or we do no preprocessing and then
answer queries in linear time. This leads to the following conjecture.

Conjecture 12. Directed Reachability Hypothesis. Any data structure for

the Reachability problem must either use (n2 ) space, or linear query time.

The conjecture states that in the general case of Reachability there is no full
tradeo between space and query time. We believe the conjecture is true even if
the path is limited to lengths of some non-constant number of edges.

6 Distance Oracles and Directed Reachability

There are known lower bounds for constant query time distance oracles based on
the SetDisjointness hypothesis. Specically, Cohen and Porat [16] showed that
stretch-less-than-2 oracles need (n2 ) space for constant queries. Patrascu et
al. [28] showed a conditional space lower bound of (m5/3 ) for constant-time
stretch-2 oracles. Applying the Strong SetDisjointness conjecture to the same
argument as in [16] we can prove that for stretch-less-than-2 oracles the tradeo
between S (the space for the oracle) and T (the query time) is by ST 2 = (n2 ).
Recent eort was taken toward constructing compact distance oracles where
we allow non-constant query time. For stretch-2 and stretch-3 Agarwal et al. [10] [9]
achieves a space-time tradeo of ST = O(n2 ) and ST 2 = O(n2 ), respectively,
for sparse graphs. Agarwal [8] also showed many other results for stretch-2 and
below. Specically, Agarwal showed that for any integer k a stretch-(1+1/k)
oracle exhibits the following space-time tradeo: S T 1/k = O(n2 ). Agarwal
also showed a stretch-(1+1/(k+0.5)) oracle that exhibits the following tradeo:
S T 1/(k+1) = O(n2 ). Finally, Agarwal gave a stretch-(5/3) oracle that achieves
a space-time tradeo of S T = O(n2 ). Unfortunately, no lower bounds are
known for non-constant query time.
Conditioned on the directed k-Reachability conjecture we prove the following
lower bound.

Lemma 4. Assume the directed k-Reachability conjecture holds. Then stretch-

less-than-(1 + 2/k) distance oracles with query time T must use S T 2/(k1) =
(n2 ) space.
434 I. Goldstein et al.

Proof. Given a graph G = (V, E) for which we want to preprocess for k-Reachability,
we create a layered graph with k layers where each layer consists of a copy of
all vertices of V . Each pair of neighboring layers is connected by a copy of all
edges in E. We omit all directions from the edges. For every xed integer k, the
layered graph has O(|V |) vertices and O(|E|) edges. Next, notice that if we con-
struct a distance oracle that can distinguish between pairs of vertices of distance
at most k and pairs of vertices of distance at least k + 2, then we can answer
k-Reachability queries. Consequently, assuming the k-Reachability conjecture we
have that S T 2/(k1) = (n2 ) for stretch-less-than-(1 + 2/k) distance oracles
(For k = 2 this is exactly the result we get by the SetDisjointness hypothesis).

Notice, that the stretch-(5/3) oracle shown by Agarwal [8] achieves a space-
time tradeo of S T = O(n2 ). Our lower bound is very close to this upper
bound since it applies for any distance oracle with stretch-less-than-(5/3), by
setting k = 3.

References
1. Amir Abboud, Arturs Backurs, Thomas Deuholm Hansen, Virginia Vassilevska
Williams, and Or Zamir. Subtree isomorphism revisited. In Proc. of 27th ACM-
SIAM Symposium on Discrete Algorithms, SODA, pages 12561271, 2016.
2. Amir Abboud, Arturs Backurs, and Virginia Vassilevska Williams. If the current
clique algorithms are optimal, so is Valiants parser. 55th IEEE Annual Symposium
on Foundations of Computer Science, FOCS, pages 98117, 2015.
3. Amir Abboud, Arturs Backurs, and Virginia Vassilevska Williams. Quadratic-
time hardness of LCS and other sequence similarity measures. 55th IEEE Annual
Symposium on Foundations of Computer Science, FOCS, pages 5978, 2015.
4. Amir Abboud, Fabrizio Grandoni, and Virginia Vassilevska Williams. Subcubic
equivalences between graph centrality problems, APSP and diameter. In Proceed-
ings of the Twenty-Sixth Annual ACM-SIAM Symposium on Discrete Algorithms,
SODA 2015, San Diego, CA, USA, January 4-6, 2015, pages 16811697, 2015.
5. Amir Abboud and Virginia Vassilevska Williams. Popular conjectures imply strong
lower bounds for dynamic problems. In 55th IEEE Annual Symposium on Foun-
dations of Computer Science, FOCS 2014, Philadelphia, PA, USA, October 18-21,
2014, pages 434443, 2014.
6. Amir Abboud, Virginia Vassilevska Williams, and Oren Weimann. Consequences
of faster alignment of sequences. In Automata, Languages, and Programming - 41st
International Colloquium, ICALP 2014, Copenhagen, Denmark, July 8-11, 2014,
Proceedings, Part I, pages 3951, 2014.
7. Amir Abboud, Virginia Vassilevska Williams, and Huacheng Yu. Matching trian-
gles and basing hardness on an extremely popular conjecture. In Proceedings of
the Forty-Seventh Annual ACM on Symposium on Theory of Computing, STOC
2015, Portland, OR, USA, June 14-17, 2015, pages 4150, 2015.
8. Rachit Agarwal. The space-stretch-time tradeo in distance oracles. In Algorithms
- ESA 2014 - 22th Annual European Symposium on Algorithms, Wroclaw, Poland,
September 8-10, 2014. Proceedings, pages 4960, 2014.
9. Rachit Agarwal, Brighten Godfrey, and Sariel Har-Peled. Faster approximate dis-
tance queries and compact routing in sparse graphs. CoRR, abs/1201.2703, 2012.
Conditional Lower Bounds for Space/Time Tradeoffs 435

10. Rachit Agarwal, Philip Brighten Godfrey, and Sariel Har-Peled. Approximate
distance queries and compact routing in sparse graphs. In INFOCOM 2011. 30th
IEEE International Conference on Computer Communications, pages 17541762,
2011.
11. Amihood Amir, Timothy M. Chan, Moshe Lewenstein, and Noa Lewenstein. On
hardness of jumbled indexing. In Automata, Languages, and Programming - 41st
International Colloquium, ICALP 2014, Copenhagen, Denmark, July 8-11, 2014,
Proceedings, Part I, pages 114125, 2014.
12. Arturs Backurs and Piotr Indyk. Edit distance cannot be computed in strongly
subquadratic time (unless SETH is false). In Proceedings of the Forty-Seventh
Annual ACM on Symposium on Theory of Computing, STOC 2015, Portland, OR,
USA, June 14-17, 2015, pages 5158, 2015.
13. Karl Bringmann. Why walking the dog takes time: Frechet distance has no strongly
subquadratic algorithms unless SETH fails. In 55th IEEE Annual Symposium on
Foundations of Computer Science, FOCS 2014, Philadelphia, PA, USA, October
18-21, 2014, pages 661670, 2014.
14. Karl Bringmann and Marvin Kunnemann. Quadratic conditional lower bounds
for string problems and dynamic time warping. 55th IEEE Annual Symposium on
Foundations of Computer Science, FOCS, 2015.
15. Timothy M. Chan and Moshe Lewenstein. Clustered integer 3SUM via additive
combinatorics. In Proceedings of the Forty-Seventh Annual ACM on Symposium on
Theory of Computing, STOC 2015, Portland, OR, USA, June 14-17, 2015, pages
3140, 2015.
16. Hagai Cohen and Ely Porat. Fast set intersection and two-patterns matching.
Theor. Comput. Sci., 411(40-42):37953800, 2010.
17. Hagai Cohen and Ely Porat. On the hardness of distance oracle for sparse graph.
CoRR, abs/1006.1117, 2010.
18. A. Gajentaan and M. H. Overmars. On a class of O(n2 ) problems in computational
geometry. Comput. Geom., 5:165185, 1995.
19. Isaac Goldstein, Tsvi Kopelowitz, Moshe Lewenstein, and Ely Porat. How hard is
it to nd (honest) witnesses? In European Symposium on Algorithms, ESA 2016,
pages 45:145:16, 2016.
20. Monika Henzinger, Sebastian Krinninger, Danupon Nanongkai, and Thatchaphol
Saranurak. Unifying and strengthening hardness for dynamic problems via the
online matrix-vector multiplication conjecture. In Proceedings of the Forty-Seventh
Annual ACM on Symposium on Theory of Computing, STOC 2015, Portland, OR,
USA, June 14-17, 2015, pages 2130, 2015.
21. Jacob Holm, Eva Rotenberg, and Mikkel Thorup. Planar reachability in linear
space and constant time. In 56th Annual Symposium on Foundations of Computer
Science, FOCS 2015, Berkeley, CA, USA, 17-20 October, 2015, pages 370389,
2015.
22. Tsvi Kopelowitz, Seth Pettie, and Ely Porat. Dynamic set intersection. In Pro-
ceedings 14th Intl Symposium on Algorithms and Data Structures (WADS), pages
470481, 2015.
23. Tsvi Kopelowitz, Seth Pettie, and Ely Porat. Higher lower bounds from the 3SUM
conjecture. In Proceedings of the Twenty-Seventh Annual ACM-SIAM Symposium
on Discrete Algorithms, SODA 2016, Arlington, VA, USA, January 10-12, 2016,
pages 12721287, 2016.
24. Kasper Green Larsen, J. Ian Munro, Jesper Sindahl Nielsen, and Sharma V.
Thankachan. On hardness of several string indexing problems. Theor. Comput.
Sci., 582:7482, 2015.
436 I. Goldstein et al.

25. Mihai Patrascu. Towards polynomial lower bounds for dynamic problems. In
Proceedings of the 42nd ACM Symposium on Theory of Computing, STOC 2010,
Cambridge, Massachusetts, USA, 5-8 June 2010, pages 603610, 2010.
26. Mihai Patrascu. Unifying the landscape of cell-probe lower bounds. SIAM J.
Comput., 40(3):827847, 2011.
27. Mihai Patrascu and Liam Roditty. Distance oracles beyond the Thorup-Zwick
bound. SIAM J. Comput., 43(1):300311, 2014.
28. Mihai Patrascu, Liam Roditty, and Mikkel Thorup. A new innity of distance
oracles for sparse graphs. In 53rd Annual IEEE Symposium on Foundations of
Computer Science, FOCS 2012, New Brunswick, NJ, USA, October 20-23, 2012,
pages 738747, 2012.
29. Mihai Patrascu and Ryan Williams. On the possibility of faster SAT algorithms.
In Proceedings of the Twenty-First Annual ACM-SIAM Symposium on Discrete
Algorithms, SODA 2010, Austin, Texas, USA, January 17-19, 2010, pages 1065
1075, 2010.
30. Virginia Vassilevska Williams and Ryan Williams. Subcubic equivalences between
path, matrix and triangle problems. In 51th Annual IEEE Symposium on Founda-
tions of Computer Science, FOCS 2010, October 23-26, 2010, Las Vegas, Nevada,
USA, pages 645654, 2010.
 
Posimodular Function Optimization

Magnus M. Halldorsson1 , Toshimasa Ishii2 , Kazuhisa Makino3 , and

Kenjiro Takazawa4
1
ICE-TCS, School of Computer Science, Reykjavik University, Iceland
[email protected]
2
Graduate School of Economics, Hokkaido University, Sapporo, Japan
[email protected]
3
Research Institute for Mathematical Sciences, Kyoto University, Japan
[email protected]
4
Faculty of Science and Engineering, Hosei University, Japan
[email protected]

Abstract. A function f : 2V IR on a nite set V is posimodular if

f (X) + f (Y ) f (X \ Y ) + f (Y \ X), for all X, Y V . Posimodular
functions often arise in combinatorial optimization such as undirected
cut functions. We consider the problem of nding a nonempty subset X
minimizing f (X), when the posimodular function f is given by oracle
access.
We show that posimodular function minimization requires exponential
time, contrasting with the polynomial solvability of submodular function
minimization that forms another generalization of cut functions. On the
other hand, the problem is xed-parameter tractable in terms of the size
of the image (or range) of f .
In more detail, we show that (20.3219n Tf ) time is necessary and O(20.92n Tf )
sucient, where Tf denotes the time for one function evaluation. When
the image of f is D = {0, 1, . . . , d}, O(21.271d nTf ) time is sucient and
(20.1609d Tf ) necessary. We can also generate all sets minimizing f in
time 2O(d) n2 Tf .
Finally, we also consider the problem of maximizing a given posimodular
function, showing that it requires at least 2n1 Tf time in general, while
it has time complexity (nd1 Tf ) when D = {0, 1, . . . , d} is the image
of f , for integer d.

1 Introduction

Let V denote a nite set with n = |V |. A set function f : 2V IR is called

posimodular if
f (X) + f (Y ) f (X \ Y ) + f (Y \ X) (1.1)

This research was partially supported by Icelandic Research Fund grants 152679-05
and 174484-05, MEXT KAKENHI Grant Numbers JP24106002, JSPS KAKENHI
Grant Numbers JP25280004, JP26280001 and JP16K00001, and JST CREST Grant
Number JPMJCR1402, Japan.

Springer International Publishing AG 2017 437

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 437448, 2017.
DOI: 10.1007/978-3-319-62127-2_37
438 M.M. Halldrsson et al.

for all X, Y V , where IR denotes the set of all reals. Posimodularity is a

fundamental property in combinatorial optimization [6, 7, 12, 14, 15, 17] and
is typically the key for eciently solving undirected network optimization and
related problems, since cut functions for undirected networks are posimodular.
In comparison, cut functions for directed networks are not posimodular.
There are numerous network optimization problems where posimodularity
leads to gaps between the complexity for undirected and directed variants. One
example is the local edge-connectivity augmentation problem, which is polyno-
mially solvable in undirected networks but NP-hard in directed networks [5].
Similarly, undirected versions of the source location problem with uniform de-
mands or with uniform costs can be solved in polynomial time, [1, 19], while the
directed versions are NP-hard [8]. More generally, the current fastest algorithm
for minimizing a submodular and posimodular function runs in O(n3 Tf ) time
[13], while the one for minimizing a submodular function requires O(n5 Tf + n6 )
time [16], where a set function f : 2V IR is called submodular if

f (X) + f (Y ) f (X Y ) + f (X Y ) (1.2)

for all X, Y V , and Tf denotes the time needed to evaluate the function value
f (X) for a given X V . The submodular multiway partition problem, which
is a generalization of the graph multiway cut problems, is 2-approximable in
polynomial time, while the symmetric submodular multiway partition problem
is 1.5-approximable [2], where a set function f : 2V IR is called symmetric if
f (X) = f (V \ X) holds for any X V . We note that a function is symmetric
posimodular if and only if it is symmetric submodular, since the symmetricity
of f implies that f (X) + f (Y ) = f (V \ X) + f (Y ) and f (X \ Y ) + f (Y \ X) =
f ((V \ X) Y ) + f ((V \ X) Y ).
These phenomena can be partially explained by the following three struc-
tural properties on posimodular functions. The rst structural property is used
under the name of uncrossing techniques. There are many variants  of partition
k
problems that ask for a partition {V1 , V2 , . . . , Vk } of V minimizing i=1 f (Vi ),
for a given set function f . This includes the graph multiway cut problem, the
graph k-way cut problem, and the submodular multiway partition problem.
If f is posimodular, then after obtaining a family {V1 , V2 , . . . , Vk } of subsets
that covers V but may not be disjoint, we can apply uncrossing techniques
to obtain a partition {V1 , V2 , . . . , Vk } of V without increasing the cost (i.e.,
k k 
i=1 f (Vi ) i=1 f (Vi )). This is because the posimodularity of f implies
that f (X) + f (Y ) min{f (X) + f (Y \ X), f (Y ) + f (X \ Y )} for any two sets
X and Y . Indeed, this uncrossing technique results in a better approximation
ratio for the symmetric submodular multiway partition problem than the (non-
symmetric) submodular multiway partition problem [2] (recall that a symmetric
submodular function is posimodular). Similar uncrossing techniques have been
utilized in other partition problems [10, 18].
The second structural property holds for extreme sets. A subset X of V is
called extreme if every nonempty proper subset Y of X satises f (Y ) > f (X).
It is known that when f is posimodular, the family X (f ) of extreme sets is
Posimodular Function Optimization 439

laminar (i.e., every two members X and Y of X (f ) satisfy X Y = , X Y ,

or X Y ). Note that if X, Y X (f ) would satisfy X Y, X \ Y, Y \ X = , then
we have f (X) + f (Y ) f (X \ Y ) + f (Y \ X) > f (X) + f (Y ), a contradiction.
The family X (f ) of extreme sets for an undirected cut function f represents
the connectivity structure of a given network and helps to design many ecient
network algorithms [9, 21]. For example, the undirected source location problem
with uniform demands can be solved in O(n) time, if the family X (f ) is known
in advance, where n corresponds to the number of vertices in the network [11].
In fact, X (f ) can be computed in O(n(m + n log n)) time for any undirected cut
function [11], where m denotes the number of edges in the network. We note
that X (f ) can be found in O(n3 Tf ) time if f is posimodular and submodular
[12].
The third structural property holds for solid sets, where a subset X of V is
said to be v-solid for an element v V , if v X and every nonempty proper
subset Y of X that contains v satises  f (Y ) > f (X). Let S(f ) denote the
family of all solid sets, i.e., S(f ) = vV {X : X is v-solid}. It is known [17]
that the family S(f ) forms a tree hypergraph if f is posimodular. Similarly to
the previous case for X (f ), if a host tree T of S(f ) is known in advance, this
structure enables us to construct a polynomial time algorithm for the minimum
transversal problem for posimodular functions f , which is an extension of the
undirected source location problem with uniform costs [19] and the undirected
external network problem [20]. If f is in addition submodular, a host tree T can
be computed in polynomial time.
We remark that the above structural properties of X (f ) and S(f ) follow
from the posimodularity of f , while the submodularity is used to derive such
structures eciently.
On the other hand, to our best knowledge, all previous results for posimodular
optimization also make use of submodularity or symmetry, since undirected cut
functions, the most representative posimodular functions, are also submodular
and symmetric.
In this paper, we focus on the posimodular function minimization dened as
follows.

Posimodular Function Minimization

Input: A posimodular function f : 2V IR, (1.3)
Output: A nonempty subset X of V such that f (X ) = min f (X).
XV :X=

Here an input function f is given by an oracle that answers f (X) for a given
subset X of V , and we assume that the optimal value f (X ) is also output. The
problem was posed as an open problem on the Egres open problem list [3] in 2010,
as negamodular function maximization, where a set function f is negamodular, if
f is posimodular. We also consider the posimodular function maximization, as
submodular function maximization has been intensively studied in recent years.
440 M.M. Halldrsson et al.

Our Contributions

The main results obtained in this paper can be summarized as follows.

1. Intractability: We show that any algorithm for posimodular function min-

imization requires (20.3219n Tf ) time. On the other hand, we show that it
is possible to beat the trivial 2n Tf upper bound, giving an O(20.92n Tf )-time
algorithm.
2. Tractability on small images: We consider functions f : 2V D with
a restricted image. Our main positive result is that the problem is xed-
parameter tractable in terms of the image size |D|. Specically, we give
an algorithm with complexity O(23|D| nTf ). For the case of the image D =
{0, 1, . . . , d}, we obtain an improved bound of O(21.271d nTf ). This is matched
with an exponential lower bound of (20.1609d Tf ) time.
The most technical part of the paper is the extension of the parameterized al-
gorithm to generate all minimizers with linear delay, after initial 2O(|D|) n2 Tf -
time.
3. Hardness of Maximization: We show that posimodular function maxi-
mization requires at least 2n1 Tf time, and thus only trivial solutions are
possible. For image restricted to (a subset of) D = {0, 1, . . . , d}, for a con-
stant d, we obtain a tight bound of (nd1 Tf ) on the time complexity.

We also obtain implications for related problems. For instance, we can com-
pute all extreme sets in O(|D|23|D| nTf ) time, which implies that the source
location problem for posimodular functions can be solved in O(|D|23|D| nTf )
time.
We note that no complexity-theoretic assumptions are needed for the lower
2
bounds. For related results, Feige et al. [4] showed that at least e n/8 oracle
calls are necessary for obtaining a solution of at least (1/2 + ) times optimal for
symmetric submodular function maximization, which is equivalent to symmetric
posimodular function maximization.
The rest of this paper is organized as follows. In Section 2, we give the
hardness results and a o(2n Tf )-time algorithm for posimodular function mini-
mization. In Section 3, we consider the case where the image of f is bounded
or given by D = {0, 1, . . . , d} and show hardness results and a xed parameter
algorithm in terms of the image size. Section 4 treats the posimodular function
maximization. Due to space limitations, some proofs are omitted.

2 General Posimodular Function Minimization

2.1 Hardness Results

Let V be a nite set with n = |V | and f : 2V IR be a posimodular function.

Notice that f satises

f (X) f () for all X V, (2.1)

Posimodular Function Optimization 441

since f (X) + f (X) f () + f (). Throughout the paper, we assume that f () =

0, since otherwise, we can replace f (X) by f (X) f () for all X V .
In this section, we analyze the number of oracle calls necessary for posimod-
ular function minimization. An optimal solution to the posimodular function
minimization (1.3) is referred to as a minimizer of f (among nonempty sub-
sets).
Let g : 2V IR+ be the cardinality function dened by g(X) = |X|. Clearly,
g is posimodular since g is monotone, i.e., g(X) g(Y ) holds for all two subsets
X and Y of V with X Y .
For a given positive integer k, we construct the family Gk = {g} {gS | S
V, |S| = 2k} of functions, where gS : 2V IR+ is dened by

2k |X| if X S and |X| k + 1,
gS (X) =
g(X) = |X| otherwise.

We can see that each gS is a posimodular function close to g. We show below

that exponential number of oracles queries are necessary to distinguish between
the posimodular functions in Gk .
Let Sk = {S V | |S| = 2k} and Tk = {T V | k + 1 |T | 2k}.
Consider the following integer program that formulates the hitting set problem,
which asks for a minimum cardinality subset of Tk that hits each set in Sk .

minimize zT
T Tk
subject to T Tk :T S zT 1 for each S Sk , (2.2)
zT {0, 1} for each T Tk .

Note that every posimodular function f in Gk satises f (X) = g(X) if |X| k

or |X| 2k + 1. Oracle calls for such sets X do not help to distinguish among
posimodular functions in Gk . Therefore, we can restrict our attention to subsets
T in Tk for oracle calls.
Then, we have the following lemma, whose proof is omitted.
Lemma 2. Let qk denote the optimal value for (2.2).
(i) At least qk oracle calls are necessary to distinguish among posimodular func-
tions in Gk .
nn nn
(ii) For 2 k n/2, we have c1 (4k)k (nk) nk qk c2 2
o(n)
(4k)k (nk)nk
for
some positive constants c1 and c2 .
nn
Observe that when n = 5k, we have (4k)k (nk)nk
= 1.25n = 20.3219n . By
Lemma 2 and this, we have qk [c1 20.3219n , c2 20.3219n+o(n) ] for some positive
constants c1 and c2 . Thus, we have the following theorem.
Theorem 1. Every algorithm for posimodular function minimization makes
(20.3219n ) oracle calls in the worst case.
Finally, it follows that there is no xed parameter algorithm in terms of the
solution size |S|. Indeed, Lemma 2 implies that for the above instance, any
|S|/2
|S|/2 n oracle calls (note that |S| = 2k).
c1
algorithm requires at least (2|S|)
442 M.M. Halldrsson et al.

2.2 o(2n Tf )-time algorithm

Posimodular function minimization can trivially be solved in 2n Tf time, since the
number of subsets of V is 2n . In this subsection, we give a cn Tf time algorithm
for the problem, for c < 2.

n 
Theorem 2. Posimodular function minimization can be solved in O( n/3 n log n
0.92n
Tf ) = O(2 Tf ) time.

We say that a set X S is a splitter with respect to S if f (X {v}) >

f (X {u}) for all v V \ S and all u S \ X. Let v1 , v2 , . . . , v|V \X| be an
ordering of V \ X such that f (X {v1 }) f (X {v2 }) f (X {v|V \X| }).
Note that for each i with f (X {vi }) < f (X {vi+1 }), X is a splitter with
respect to X {v1 , v2 , . . . , vi } and that X is not a splitter with respect to any
other subset of V .
It turns out that a small splitter exists, as long as there is no singleton
minimizer. The search then reduces to nding either a small splitter or a very
large minimizer.

Lemma 5. Suppose no singleton is a minimizer of f . Then, there exists a split-

ter of cardinality at most (|S| 1)/2 with respect to a minimizer S.

Based on this lemma, we can prove Theorem 2.

Proof of Theorem 2. Lemma 5 shows that there exists either: (i) a singleton
minimizer, (ii) a minimizer of size at least 2n/3, or (iii)
n a splitter
of size at
n
most n/3. The number of sets to be examined are n, O( 2n/3 ) = O( n/3 ), and

n 
O( n/3 ), respectively. Recall that in the case a splitter is found, the minimizer
can be found in at most n oracle calls. 


3 Minimization on Small Images

In light of the hardness of minimizing general posimodular functions, we turn
our attention to parameterized algorithms. The intractability results still hold
in terms of the solution size (cardinality of the minimizer) or the value of the
minimum solution. Instead, we treat in this section the parameter |D|, the car-
dinality of the image D of the function f , with a particular focus on the case
D = {0, 1, . . . , d}, for a number d.

3.1 Fixed Parameter Algorithm

We propose a bounded-depth tree search algorithm. Each tree node corresponds
to an invocation of a recursive procedure given three set parameters, A, B, and
C. In each invocation, the algorithm either produces a solution and terminates,
or it selects an element v with which it makes two recursive call: adding v to A,
and adding v to either B or C. The crucial property maintained is that whenever
Posimodular Function Optimization 443

V \S S
A B
C

an element is added to a set, the value of the set increases. It follows immediately
that the recursion depth is at most 3|D|, and since the tree is binary, the number
of recursive calls, which dominates the time complexity, is at most 23|D| .
The challenge is in showing that a solution can be obtained once no options
for recursing remain. The parameter triplet (A, B, C) forms a valid tuple with
respect to a minimizer S, if A is disjoint from S while B and C are contained in
S (A (V \ S) B C S). One can show by induction that exactly one leaf
in the search tree corresponds to a valid tuple (for a given minimizer S). The
crucial characterization that we obtain is that once we reach a leaf with a valid
tuple, we can easily identify the rest of S: namely, S \ (B C) is given by those
elements in V \ (A B C) whose addition to neither B nor C increases the
values of those sets.
The algorithm MinPosimodular-d(f ) operates in more detail as follows.
It uses a global variable SS, initially set as the singleton set of smallest func-
tion value, which progressively improves to a minimizer. The algorithm calls a
recursive subroutine Solve with parameters A, B, and C, initially empty sets.
Note that the algorithm need not know anything about the image D.

Algorithm 1 MinPosimodular-D(f )
1: SS {v }, where v = arg minvV f (v) is a singleton of smallest value
2: Call Solve(, , )
3: Output SS

We say that an element v increases a set X if f (X {v}) > f (X), and

increases by two if f (X {v}) > f (X) + 1. Let incA(v) denote the predicate
that element v increases A, i.e., f (A {v}) > f (A). Similarly, let incBC(v)
denote the predicate that v increases B or v increases C (or both).
Let Try(v) denote the following inline macro: Recursively call Solve(A
{v}, B, C); then, call Solve(A, B {v}, C) if v increases B, and otherwise call
Solve(A, B, C {v}).
The core tools for our results are the following two lemmas. They show that
if given a set that is either properly inside or outside a minimizer, then adding
elements from the other side must increase the value of the set. This helps us
nd sets with successively larger values; once we obtain sets of maximum value,
444 M.M. Halldrsson et al.

Algorithm 2 Procedure Solve(A, B, C)

Require: Disjoint subsets A, B, C of V ; SS is a global variable representing the current
champion
1: if v V \ (A B C), incA(v) incBC(v) then
2: Try(v);
3: else
4: S B C {v V \ A : incBC(v)}
5: if f (S) < f (SS) then
6: SS S

the remaining elements can be quickly assigned, based on the characterization

of the lemmas.
A subset X is called locally minimal if f (X) < f (X \ {v}) holds for every
v X.

Lemma 7. Let S be a locally minimal set (for f ) with |S| 2. If X V \ S

and v is an element in S, then v increases X.

Lemma 8. Let S be a set with smaller value than any singleton set. Let B and
C be disjoint sets within S and let v be an element in V \ S. Then, v increases
B or it increases C by two.

We rst argue correctness.

Lemma 9. The algorithm MinPosimodular-d(f ) nds a (minimal) mini-

mizer of f .

Proof. If there is a singleton minimizer, then it is found in the rst step of

MinPosimodular-d(f ). Suppose then that there is a minimal minimizer S
with |S| 2. Recall that there is leaf in the recursion tree with a valid tuple
(A, B, C) for S. Since the recurrence ended at this leaf, each element v outside
of A B C fails either incA(v) or incBC(v). Each element v failing incA(v)
must be in V \ S, by Lemma 7, while an element w failing incBC(w) must be
in S, by Lemma 8. Then, S is given by S = B C {v V \ A : incBC(v)},
which is found in line 4 of Solve. 


The bound O(23|D| nTf ) on the time complexity is immediate. We argue a

stronger bound for when the image is a small range.

Theorem 3. Minimization of a posimodular function f : 2V D. can be

solved in time O(23|D| nTf ). For D = {0, 1, . . . , d}, the complexity improves to
O(21.271d nTf ).

Finding all minimal minimizers. We can generalize the algorithm to nd all

minimal minimizers, i.e., all minimizers S such that no proper subset of S is
also a minimizer.

Corollary 1. All minimal minimizers can be found in time O(2min(3|D|,1.271d) nTf ).

Posimodular Function Optimization 445

3.2 Related Problems: Extreme Sets, Transversals, Approximation

Extreme sets. We rst show that the family X (f ) of all extreme sets can be
obtained as an application of Theorem 3. Recall that a subset X of V is called
extreme if every nonempty proper subset Y of X satises f (Y ) > f (X). By
denition, X (f ) contains all singletons {v}, v V , and any extreme set X with
|X| 2 is locally minimal.
Consider the subfamily Xp (f ) = {X : X is extreme with f (X) = p} of ex-
treme sets with value p, for p D. The singleton sets in Xp (f ) are given precisely
by Vp = {v V : f (v) = p}, while the non-singleton sets in Xp (f ) can only con-
tain elements v with f (v) > p. Thus, to nd non-singleton sets in Xp (f ), we can
restrict attention to the universe V>p = {v V : f (v) > p}.
We observe that MinPosimodular-d(f ) restricted to V>p identies all non-
singleton sets in Xp (f ), since Lemmas 7 and 8 hold for all non-singleton extreme
sets. Thus, by iterating over the |D| possible values of p, we can produce all
extreme sets in O(|D|23|D| nTf ) time.
In summary, we have the following result.

Corollary 2. For a posimodular function f : 2V D, we can compute the

family X (f ) of all extreme sets of f in O(|D|23|D| nTf ) time.

Minimum traversal. Consider the following problem:

Minimum Transversal(f, c, r)
Input: A posimodular function f : 2V D, a cost function
c : V IR, and a demand function r : 2V IR, (3.1)

Output: A nonempty subset S of V minimizing vS c(v) such that
f (X) r(X) for every nonempty subset X V \ S.

We note that undirected source location problem with uniform demands is

a special case of this problem where f is a cut function in an undirected graph
and r is uniform [1, 11]. As is the case with the source location [11], we can nd
an optimal solution in linear time if the family X (f ) of extreme sets is known
in advance. Thus, we have the following result.

Corollary 3. The minimum transversal problem (3.1) for a posimodular func-

tion f : 2V D can be solved in O(|D|23|D| nTf ) time if r is uniform.

We note that the minimum transversal problem (3.1) for a uniform cost func-
tion c and a modulotone demand function r is studied in [17] as a generalization
of source location problem with uniform costs [19] or external network problem
[20], where a set function r : 2V IR is called modulotone if for every nonempty
subset X of V , there exists an element v X such that all proper subsets Y
of X with v Y satisfy r(Y ) > r(X). As observed in [17], this problem can be
solved if solid sets can be computed eciently. More precisely, we need to com-
pute solid sets including u and v but w for all three distinct elements u, v, and
w in V . Therefore, Corollary 2 does not imply the tractability of this problem.
446 M.M. Halldrsson et al.

Approximation. We can speed up the search signicantly when seeking (ad-

ditive) approximations.
Theorem 4. There is an algorithm for nding an additive -approximate solu-
tion, that uses O(d/) n oracle queries, for any given 1.

3.3 Generating All Minimizers

We can extend the parameterized algorithm to generate all minimizers of a
posimodular function.
Theorem 5. For a posimodular function f : 2V D, we can generate all
minimizers of f with O(nTf ) delay, after O(53|D| nTf ) time to compute the rst
minimizer.
This part is the most technical; we briey summarize the approach.
We rst observe that it suces to nd locally minimal minimizers, as other
minimizers can be quickly generated from those. The main challenge is dealing
with the set Z of singleton minimizers without those, the previous algorithm
suces. We can treat the elements of V \ Z as before, and also those elements
of Z that dont increase A. We partition the remaining elements of Z into a
collection Z of maximal minimizers, and observe that these sets dont cross (or
overlap with) other minimizers. Those sets in Z that increase both A and one
of B, C can be treated the same way as before (with Try). The rest is split into
two: Z0 , those that increase only B or C, and Z1 , those that increase only A.
We nd that a minimizer contains at most two sets from Z0 . The key idea is to
examine pairs of sets from Z1 : if there are three sets in Z1 such that any pair
increases B or C, then some pair has to be on the same side of a minimizer S,
which allows us to make progress in one of the recursive calls. Otherwise, we can
show that Z1 contains at most two sets that are not inside S, and can try all
such possibilities.

3.4 Hardness Results

We complement the parameterized algorithm with the following lower bound.
It shows that the time complexity must both be exponential in d and involve a
factor linear in n.
Theorem 6. (i) Posimodular function minimization requires (20.1609d ) oracle
calls.
(ii) Posimodular function minimization requires (n) oracle calls, even when
restricted to functions with image D = {0, 1}.
First, we show an exponential lower bound in a similar way to the proof of
Theorem 1. Let T be a subset of V with |T | = d/2. Dene g : 2V D by

0 if X = ,
g(X) = |X| if = X T,

|T | + |T X| otherwise.
Posimodular Function Optimization 447

For a positive integer k with 2k |T |, let S be a subset of T with |S| = 2k.

Dene a function gS : 2V D by

2k |X| if X S and |X| k + 1,
gS (X) =
g(X) otherwise.

We can see that g and gS are both posimodular.

If k 2 and |T | = 5k (and hence d = 10k), then by applying an argument
similar to Lemma 2, we can observe that (20.1609d ) oracle calls are necessary
for solving the problem, which proves Theorem 6 (i).
Finally, we argue Theorem 6 (ii). Let g : 2V {0, 1} be a function dened by
g(X) = 1 if X = , and g() = 0. For each element v V , dene gv : 2V {0, 1}
by 
0 if X = or X = {v}
gv (X) =
1 otherwise.
Note that both g and gv are monotone and thus posimodular. Also note that the
minimum g-value is 1, and each function gv has exactly one minimizer {v} with
gv (v) = 0 for v V . Observe that n oracle calls are necessary to distinguish
functions in {g} {gv | v V }.
Remark. Modifying the construction by multiplying the value of function by
a factor of + 1 shows that obtaining an additive -approximation also requires
exp(d/( + 1)) oracle calls, matching Thm. 4.

4 Posimodular Function Maximization

In this section, we consider posimodular function maximization dened as fol-

lows.
Posimodular Function Maximization
Input: A posimodular function f : 2V IR+ , (4.1)
Output: A nonempty subset X of V maximizing f.
Here we assume that the optimal value f (X ) is also output. Similar to posi-
modular function minimization, the problem (4.1) is in general intractable.

Theorem 7. Every algorithm for posimodular function maximization requires

at least 2n1 oracle calls in the worst case.

For the case f : 2V {0, 1, . . . , d} we have the following tight bound.

Theorem 8. Posimodular function maximization for f : 2V {0, 1, . . . , d} can

be solved in (nd1 Tf ) time.

Acknowledgments: We would like to express our thanks to S. Fujishige, M.

Grotschel, and S. Tanigawa for their helpful comments.
448 M.M. Halldrsson et al.

References
[1] K. Arata, S. Iwata, K. Makino, and S. Fujishige. Locating sources to meet ow
demands in undirected networks. Journal of Algorithms, 42:5468, 2002.
[2] C. Chekuri and A. Ene. Approximation algorithms for submodular multiway
partition. In IEEE 52nd Annual Symposium on Foundations of Computer Science,
pages 807816, 2011.
[3] Egres open problem list. http://lemon.cs.elte.hu/egres/open/Maximizing a skew-
supermodular function.
[4] U. Feige, V. S. Mirrokni, and J. Vondrak. Maximizing non-monotone submodular
functions. SIAM J. Comput., 40:11331153, 2011.
[5] A. Frank. Augmenting graphs to meet edge-connectivity requirements. SIAM
Journal on Discrete Mathematics, 5(1):2553, 1992.
[6] S. Fujishige. A laminarity property of the polyhedron described by a weakly posi-
modular set function. Discrete Applied Mathematics, 100(1-2):123126, 2000.
[7] T. Ishii and K. Makino. Posi-modular systems with modulotone requirements un-
der permutation constraints. Discrete Mathematics, Algorithms and Applications,
2(1):6176, 2010.
[8] H. Ito, K. Makino, K. Arata, S. Honami, Y. Itatsu, and S. Fujishige. Source loca-
tion problem with ow requirements in directed networks. Optimization Methods
and Software, 18:427435, 2003.
[9] E. L. Lawler. Cutsets and partitions of hypergraphs. Networks, 3(3):275285,
1973.
[10] D. Lokshtanov and D. Marx. Clustering with local restrictions. Information and
Computation, 222:278292, 2013.
[11] H. Nagamochi. Graph algorithms for network connectivity problems. Journal of
the Operations Research Society of Japan, 47(4):199223, Dec 2004.
[12] H. Nagamochi. Minimum degree orderings. Algorithmica, 56(1):1734, 2010.
[13] H. Nagamochi and T. Ibaraki. A note on minimizing submodular functions. Inf.
Process. Lett., 67(5):239244, 1998.
[14] H. Nagamochi and T. Ibaraki. Polyhedral structure of submodular and posi-
modular systems. Discrete Applied Mathematics, 107(1-3):165189, 2000.
[15] H. Nagamochi, T. Shiraki, and T. Ibaraki. Augmenting a submodular and posi-
modular set function by a multigraph. Journal of Combinatorial Optimization,
5(2):175212, 2001.
[16] J. Orlin. A faster strongly polynomial time algorithm for submodular function
minimization. Mathematical Programming, 118(2):237251, 2009.
[17] M. Sakashita, K. Makino, H. Nagamochi, and S. Fujishige. Minimum transversals
in posi-modular systems. SIAM Journal on Discrete Mathematics, 23:858871,
2009.
[18] Z. Svitkina and E. Tardos. Min-max multiway cut. In Approximation, Random-
ization, and Combinatorial Optimization, pages 207218, 2004.
[19] H. Tamura, H. Sugawara, M. Sengoku, and S. Shinoda. Plural cover problem
on undirected ow networks. IEICE Transactions, J81-A:863869, 1998. (in
Japanese).
[20] J. van den Heuvel and M. Johnson. The external network problem with edge-
or arc-connectivity requirements. In Combinatorial and Algorithmic Aspects of
Networking, volume 3405 of Lecture Notes in Computer Science, pages 114126.
Springer, 2004.
[21] T. Watanabe and A. Nakamura. Edge-connectivity augmentation problems. Jour-
nal of Computer System Sciences, 35:96144, 1987.
 How to play hot and cold on a line

Herman Haverkort1 , David Kubel2 , Elmar Langetepe2 , and Barbara

Schwarzwald2
1
TU Eindhoven, Department of Mathematics and Computer Science
2
University of Bonn, Department of Computer Science

Abstract. Suppose we are searching for a target point t in the unit

interval. To pinpoint the location of t, we issue query points q1 , . . . , qn
[0, 1]. As a response, we obtain an ordering of the query points by distance
to t. This restricts possible locations of t to a subinterval. We dene the
accuracy of a query strategy as the reciprocal of the size of the subinterval
to which we can pinpoint t in the worst case. We describe a strategy with
accuracy (n2 ), which is at most a factor two from optimal if all query
points are generated at once. With query points generated one by one
depending on the response received on previous query points, we achieve
accuracy (2.29n ), and prove that no strategy can achieve (3.66n ).

Keywords: search games, combinatorial optimization, target localiza-

tion, online/oine strategies

1 Introduction

Imagine we want to set up receivers to locate an animal that carries a tracking

device that sends signals. The strength of the signals may vary, but we know one
thing: the further the distance from the animal, the weaker the signal. Or imagine
a researcher conducting a survey, who wants to summarize respondents political
preferences by scoring them on several scales (for example, from conservative
to progressive, or from favouring a small state to a large state). Respondents
may not be able to score themselves but, given a number of hypothetical party
programmes, they can rank them and say which one they like best.
In such settings, we are essentially searching for a target that is a point in a
one- or higher-dimensional space. To pinpoint the location of the target, we issue
queries (receivers, party programmes) that are points in the same conguration
space. As a response, we obtain an ordering of the query points according to
their distance to the target. From this we try to derive the location of the target
with the highest possible accuracyor conversely, we try to reach high accuracy
with as few (expensive) queries as possible.
Searching for a stationary target is a common problem in computer science
and applied mathematics. Strategies such as evenly distributed query points or
binary search spring to mind immediately, but, as we will see in this paper, at
least in certain abstract settings of the problem we can do much better. The
problem of eciently obtaining an order on a set of objects has been studied

Springer International Publishing AG 2017 449

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 449460, 2017.
DOI: 10.1007/978-3-319-62127-2_38
450 H. Haverkort et al.

before in very general settings [3], but note that in our case, the cost measure
is the number of query points that are used, not the number of comparisons
that are made between them. The reconstruction of geometric objects based on
a sequence of geometric probes (points, lines, hyperplanes, wedges, etc.) has also
been investigated: the problem was introduced by Cole and Yap [2] and the main
focus is also on the number of queries [5].
Specically, we focus on the following setting. The target point t is a point
located at an unknown position in the unit interval [0, 1]. To pinpoint the location
of t, we may query the interval at points q1 , . . . , qn [0, 1]. As a response, we
obtain an ordering of the points by ascending distance to t. This restricts possible
positions of t to a subinterval bounded by bisectors of query points or an endpoint
of the initial interval. We measure the eciency or quality of a query strategy
in terms of the reciprocal of the size of the subinterval in which the target t
is found to lie. The worst-case of this reciprocal, that is, the minimum over all
possible locations of t, is called the accuracy of the query strategy.
With respect to the frequency of the responses, we distinguish two variants. In
the one-shot variant (Sec. 2), the response is only given after all points q1 , . . . qn
have been placed. In this case one needs to maximize the number of dierent,
well-spaced bisectors. Such combinatorial questions are classical problems in
discrete geometry; see for example [4]. In the incremental variant (Sec. 3), a
response is given after each point placement, and may aect the choice of the
next point. This enables a binary search strategy, but we can do much better
than that. The problem can be interpreted as a game where an adversary tries
to hide the target in the largest possible area. Geometric games about area
optimization have a tradition in Computational Geometry; see for example [1].
For both variants we present an ecient strategy and an upper bound on the
accuracy that can be achieved. In the second part of Section 2, we also address
lower and upper bounds for the one-shot variant in two dimensions. In Section 4,
we briey discuss room for improvement and how the strategies can be extended
to higher-dimensional settings.

2 One-shot strategies

First we consider the one-shot variant of the problem: only after generating n
query points, we get to hear their ordering by distance to the target t. This
pinpoints t to a subinterval bounded by bisectors of query points or an end
point of the interval. As the target may lie in any subinterval, our problem is
equivalent to minimizing the maximum size of such an interval.

2.1 The one-shot strategies on the unit interval

As n query points can produce at most n(n 1)/2 distinct bisectors, there are at
most n(n 1)/2 + 1 intervals. Thus, we get the following (trivial) upper bound
for the accuracy of one-shot strategies:
How to play hot and cold on a line 451

Theorem 1. The accuracy

2 of any one-shot strategy with n points is at most
1 2
2 (n n + 2) O n .
We will now develop a strategy to get close to this upper bound. As a starting
point, consider the following simple strategy, which we call EquiDist(n): place
n evenly spaced query points (q1 , q2 , . . . , qi , . . . , qn ) := (0, 1/(n 1), . . . , (i
1)/(n 1), . . . , 1). The accuracy of EquiDist(n) is only 2(n 1) because many
bisectors conincide. However, for n 7 it is possible to forgo some of these
query points, while the number of distinct bisectors, and so the accuracy, stays
the same; see Figure 1a. Hence, for some function (n) with (n) > n, we can
achieve the same accuracy as EquiDist((n)) with n query points. Specically,
we now introduce a strategy Gapx (n) that will form a subset of the query
points dened by EquiDist(x (n)) for x (n) := n(x + 1) 2x2 x 2. It uses
only the following n points from EquiDist(x (n)): q1 , q2 , . . . , qx+1 , as well as
q2x+1+k(x+1) for k {0, . . . , n (2x + 3)}, and qx (n)x , . . . , qx (n)1 , qx (n) .
This results in widely spaced query points throughout the search range, and
tightly spaced points near both ends, omitting at most x consecutive query
points from EquiDist(x (n)). Most bisectors will then be formed by one of the
tightly spaced points and one of the widely spaced points.

Lemma 1. For x, n N with n (2x + 3), the one-shot query strategy Gapx
has accuracy 2(x (n) 1).

Proof. It suces to show that every distinct bisector from EquiDist(x (n))
is also created by Gapx (n), as the points chosen by Gapx (n) are a subset of
those chosen by EquiDist(x (n)). Hence, for each 1 i < x (n), Gapx (n)
must choose two points that form a bisector in the middle between qi and qi+1 .
For each 1 < j < x (n), Gapx (n) must choose two points that form a bisector
directly on qj .
For i x and j x this is given by q1 , q2 , . . . , qx+1 . For x < i < x (n)/2,
EquiDist(x (n)) forms a bisector between qi and qi+1 with all of the pairs
(q1 , q2i ), (q2 , q2i1 ), . . . , (qx+1 , q2ix ). One of these pairs has to be in the chosen
subset of Gapx (n), since the choosing process omits at most x consecutive points.
Likewise, one of the pairs (q1 , q2j1 ), (q2 , q2j2 ), . . . , (qx+1 , q2j(x+1) ) must have
been chosen by Gapx (n). Those pairs form a bisector on qj for x < j x (n)/2.
The existence of the remaining bisectors, that is, those in the right half of the
unit interval, follows by the symmetry of the strategy.
Hence Gapx (n) produces the same set of 2(x (n) 2) distinct equidistant
bisectors as EquiDist(x (n)), resulting in an accuracy of 2(x (n) 1).

It remains to choose the optimal x, given n, maximizing x (n). As the slope

of the linear function x (n) increases with x, Gapx is eventually surpassed by
Gapx+1 . The break-even point between Gapx and Gapx+1 is at n = 4x + 3.
Hence, the optimal x = (n 3)/4 and we get:

Theorem 2. For any number of points n 3 and x = (n 3)/4 , the one-shot

strategy Gapx (n) has an accuracy of 2(x (n) 1) (n2 + 6n 27)/4 (n2 ).
452 H. Haverkort et al.

This leaves only a gap of a constant factor two between the lower and upper
bound. Note that it is (already) impossible to place four query points in such a
way that their bisectors are all distinct and divide the unit interval into equal
parts. This shows that the upper bound for the one-shot strategy is not tight.

q1 q2 q3 - q5 - - q8 - q10 q11 q12

(a) Removing q4 , q6 , q7 and q9 from EquiDist(12) (b) Gap(4) 2d-version above, an

on the unit interval does not aect the accuracy. optimal placement below?

Fig. 1: Examples for the one-shot variant of the problem: 1a considers the strate-
gies EquiDist(12) and Gap2 (8) on the unit interval. Possible placements of 4
query points on the unit disk are given in 1b.

2.2 One-shot strategies in two dimensions

Now let us turn our attention to the two dimensional case, where we use the
diameter of the resulting region as the measure for accuracy3 . So assume the
target hides somewhere in the unit square or the unit disk. Similar to the one-
dimensional case, we obtain quadratic lower and upper bounds for the accuracy
of one-shot strategies for both types.
Theorem 3. The accuracy of any two-dimensional one-shot strategy on the unit
square (or unit disk) that places n query points is O(n2 ).
Proof. Using n query points, one can construct at most n(n 1)/2 bisectors.
Consequently, the number of cells in the arrangement of bisectors is at most
(n4 2n3 + 3n2 2n + 8)/8. A cell of diameter covers an area of at most 2 /4.
This gives us an upper bound on the area that all cells can cover together if
their maximum diameter is . In turn, this gives us a lower bound 4 2

1/2
(n4 2n3 + 3n2 2n + 8) (n2 ) on the smallest maximum diameter
that can be achieved. Hence the accuracy is O(n2 ).
The same argument can be applied for the unit circle.
3
One can argue to use the area of the region instead. This however can result in very
thin and long regions contradicting the aim to precisely locate the target, as the
diameter limits the area but not vice versa.
How to play hot and cold on a line 453

Theorem 4. Any target on the unit square (or unit disk) can be located with
accuracy (n2 ) in the one-shot variant with n query points.
Proof. Dene k := n/2. As we want to prove a lower bound for the accuracy, we
may choose to use 2k query points and ignore the remaining ones. We may choose
our coordinate system such that the center of the unit square (or unit disk) lies
at (0, 0). Then, we apply Gapx (k) along each axis of the coordinate system. Now
consider only the bisectors perpendicular to the coordinate axis. These bisectors
subdivide the unit square (unit disk) into small rectangles (or parts of small
1
rectangles), each of height and width at most 1/2 (x (k) 1) , by Lemma

2 1
1. Consequently, the diameter of each square is at most 4 2 k + 6k 27 ,
which gives the quadratic lower bound for the accuracy in terms of n.
It is not clear whether Gapx is optimal in the one-dimensional case. It is certain
however, that the Gap-strategy extended to the unit disk is not, not even for
even n. Instead of placing the queries along the coordinate axis, it might be
better to place one of the query points in the center of the disk; see Figure 1b.

3 Incremental and online strategies

In this section we consider incremental strategies, that is, before placing any
query point qi we may learn the ordering of q1 , ..., qi1 by distance to the target,
and we may choose the location of qi depending on that information. We will
show upper and lower bounds on the accuracy that can be achieved.

3.1 A strategy with high accuracy

g(n) g(n) f (n) f (n) = h(n 3)

(i)
q1 q2 q3
f (n) g(n) 1
2 h(n) g(n) f (n)

h(n 3) h(n 3) h(n 4) h(n 4)

(ii)
q2 q3 q4
h(n 4) h(n 3) 1
g(n)
2 h(n)
2h(n 3) h(n 4)

Fig. 2: Location of the rst query points for our incremental strategy to locate
a target with accuracy h(n).

A simple incremental strategy could choose the query points such that the
interval containing the target is halved in each step, except the rst. Thus, with
 454 H. Haverkort et al.

n query points, we achieve accuracy 2n1 . But we can do much better, with a
recursive strategy that takes advantage of more than one new bisector in many
steps. Let h(n) be the accuracy we aim for, as a function of the number n of
query points we get to place. For ease of description, we use the interval [0, h(n)]
instead of the unit interval and pinpoint the target to an interval of length 1.
The recursion in our strategy depends on a number of conditions, labelled A to
E and presented below.
Our overall strategy to locate a target in an interval [0, h(n)] with n query
points starts with placing the rst two query points symmetrically, otherwise
the overall strategy can easily be improved.
Let g(n) be the distance of the rst two query points to their common bi-
sector, so we place q1 at 12 h(n) g(n) and q2 at 12 h(n) + g(n). Now suppose the
target lies to the right of the bisector (the other case is symmetric). We now
place q3 at some distance 2f (n) to the right of q2 , and nd that the target lies
in one of three intervals (see Figure 2(i)):

(i) [bs(q1 , q2 ), bs(q1 , q3 )] = [ 12 h(n), 12 h(n) + f (n)], of size f (n).

(ii) [bs(q1 , q3 ), bs(q2 , q3 )] = [ 12 h(n) + f (n), 12 h(n) + g(n) + f (n)], of size g(n), or
(iii) [bs(q2 , q3 ), h(n)] = [ 12 h(n) + g(n) + f (n), h(n)], of size 12 h(n) g(n) f (n).

To be able to apply our strategy recursively in the rst interval, using the re-
maining n 3 query points, we choose f (n) = h(n 3). For the third interval,
of width 12 h(n) g(n) h(n 3), we use an adapted strategy, explained be-
low, that places the remaining n 3 query points in a way that exploits the
previously placed query points q2 and q3 at distance f (n) = h(n 3) left and
right of its left boundary. The same strategy can be applied symmetrically to
the second interval, provided the second interval is not larger than the third,
that is, g(n) 14 h(n) 12 h(n 3) for n 3 (condition A). Note that condition
A also ensures that q1 and q2 lie within the interval [0, h(n)]. To be able to place
q3 , we also require g(n) + 2h(n 3) 12 h(n) for n 3 (condition B).
We now describe our adapted strategy to locate a target in an interval
[0, 12 h(n)g(n)h(n3)] (modulo translation) with n3 query points q4 , . . . , qn
that can be chosen freely and two predetermined query points q2 = h(n 3)
and q3 = h(n 3) (see Figure 2(ii)). We place q4 at h(n 3) + 2h(n 4);
this is possible if h(n 3) + 2h(n 4) 12 h(n) g(n) h(n 3) for n 4
(condition C). Again, the target lies in one of three intervals: an interval of
width h(n 4) on the left, to which we apply the overall strategy recursively,
an interval of width 12 h(n) g(n) 2h(n 3) h(n 4) on the right, with a
predetermined query point at distance h(n 4) from its left boundary, and an
interval of width h(n 3) in the middle, with a predetermined query point at
distance h(n 4) from its right boundary. We can apply the adapted strategy
recursively to the rightmost interval, using the remaining n 4 points, provided
1 1
2 h(n) g(n) 2h(n 3) h(n 4) = 2 h(n 1) g(n 1) h(n 4) for n 4
(condition D), and we can apply the adapted strategy recursively to the middle
interval if h(n 3) 12 h(n 1) g(n 1) h(n 4) for n 4 (condition E).
It remains to choose h and g as functions of n:
How to play hot and cold on a line 455

such that the conditions for recursive application are satised, so for n 4
we must have:
(A) g(n 1) 1/4 h(n 1) 1/2h(n 4);
(B) g(n 1) 1/2 h(n 1) 2h(n 4);
(C) g(n) 1/2 h(n) 2h(n 3) 2h(n 4);
(D) 1/2 h(n) g(n) 2h(n 3) = 1/2 h(n 1) g(n 1);
(E) g(n 1) 1/2 h(n 1) h(n 3) h(n 4);
such that, when no further recursion is possible because we have run out of
query points, the remaining interval is small enough: h(0) 1, h(1) 1,
h(2) 2, and 12 h(3) g(3) h(0) 1.

Theorem 5. There is an incremental strategy to locate a target in a unit inter-

val with accuracy (bn ), where b > 2.2993 is the largest root of b4 b3 6b2. The
accuracy h(n) is given by the recursion formula h(0) = 1, h(1) = 1, h(2) = 2,
h(3) = 6 and h(n) = h(n 1) + 6h(n 3) + 2h(n 4) for n 4.

Proof. We will prove that for the above recursion formula we can nd suitable
values g(3), g(4), . . . such that condition A to E and 12 h(3) g(3) h(0) 1
hold. Thus by construction, we obtain a strategy of accuracy h(n).
Since h is a non-decreasing function of n, condition B is redundant, as it is
already implied by condition E. Rewriting condition D to g(n) = g(n 1) +
1
2 (h(n) h(n 1)) 2h(n 3) and substituting it into condition C, learns that
condition C is equivalent to condition B, and equally redundant. We now satisfy
condition A by simply choosing equality g(n 1) = 14 h(n 1) 12 h(n 4)
for n 4. Substituting this in the remaining conditions we get the following.
Condition D becomes h(n) = h(n 1) + 6h(n 3) + 2h(n 4) for n 4 (which
is among the conditions of the recursion formula in the theorem). Condition E
becomes h(n 1) 4h(n 3) + 2h(n 4), by using g(n 1) this is equivalent
to h(n 3) g(n 1).
Altogether, the remaining task is that for h(0) = 1, h(1) = 1, h(2) = 2, and
g(n 1) = 14 h(n 1) 12 h(n 4) and h(n) = h(n 1) + 6h(n 3) + 2h(n 4) for
all n 4 we have to prove that 12 h(3) g(3) h(0) 1 and h(n 3) g(n 1)
for n 4 or equivalently g(n) h(n 2) for n 3.
First, we require suitable values for g(3) and h(3) which are derived from
g(3) = 14 h(3) 12 h(0) and 12 h(3) g(3) h(0) 1. The statement holds for
g(3) = 1 and h(3) = 6. The nal task is to prove condition E or h(n3) g(n1)
for n 4. From the above formulas we have h(2) = 2, h(3) = 6, h(4) = 14,
g(4) = 3 and g(5) = 6. This means g(3) h(1), g(4) h(2) and g(5) h(3).
(Due to space limitations, the proof of this condition is omitted.)
Thus, the strategy attains accuracy h(n) = h(n 1) + 6h(n 3) + 2h(n 4).
A closer analysis of the recursion formula reveals that the growth is dominated
by b = 2.29936 . . ., the greatest root of the polynomial b4 b3 6b 2.

Note that for sake of simplicity the above strategy is not quite optimal. If we
recurse with only few points left to place, we can sometimes do slightly better by
taking some already placed points into account. For example, h(2) = 2 because,
if we get to place only two points, we can only create one bisector and therefore
456 H. Haverkort et al.

we cannot achieve accuracy better than 2. However, if we start with more than
3 points and eventually we get to a nal interval in which we get to place the
last two points, then each of these two points can form a bisector together with
a previously placed query point outside the interval. This allows us to assume
h(2) = 3 in the recursive denition of h(n) for n 4, which slightly improves
the constant factor in the asymptotic bound.
In the on-line setting, the strategy is to take h(n) = 2bn2 and g(n) =
1/4 h(n) 1/2 h(n 3) = (1/2 b3 )bn2 . Thus, the locations of q1 , . . . , q4 ,
relative to the size of the interval, h(n), are independent of n, and can therefore
be decided without advance knowledge of n.
Corollary 1. Theorem 5 also holds in the on-line setting, where n is unknown
until the last query point has been placed.

3.2 Upper bound

By placing a query point qi , we obtain i 1 new bisectors of qi with previously
placed query points. We learn the rank of qi among q1 , . . . , qi with respect to the
distance to the target, which tells us where the target is with respect to the new
bisectors. This reduces the interval R where the target must be to one of at most
i subintervals of R, at least one of which must have size at least 1/i times the size
of R. Therefore, for any incremental strategy, there must be targets for which
we can increase the accuracy with a factor at most i with every query point
qi , leading to an accuracy of at most n! after placing n query points. However,
this bound is far from tight. We can show a much stronger upper bound on the
accuracy that can be achieved:

Theorem 6. No deterministic, incremental strategy can locate an arbitrary tar-

get with an accuracy (3.66n ), where n is the number of query points used.

The proof of this theorem is based on the following insight. Suppose, for example,
that we have placed n 1 query points, and placing another query point qn
generates two new bisectors, which divide the target interval into three equal
parts. Thus, the current accuracy improves by a factor three.
However, this large increase in accuracy requires two previously placed query
points qi and qj that satisfy the following conditions. First, the points qi and qj
are consecutive in the left-to-right order of all previously placed points. Second,
the distance between the bisectors bs(qn , qi ) and bs(qn , qj ) is 1/3 of the current
size of the target interval. Therefore, the distance between qi and qj must be 2/3
of the current size of the target interval.
The previous example illustrates that, for a large increase in accuracy, one
needs to have close pairs: pairs of previously placed query points with a (com-
paratively) small distance in between. The example also illustrates that, as the
accuracy increases, close pairs can stop being close as their distance relative
to the current accuracy increases. Moreover, with the placement of each query
point, one can create at most two new close pairs: one with the left neighbour
and one with the right neighbour among all query points placed so far.
How to play hot and cold on a line 457

Denition 1 (accuracy gain). Let Rn denote the interval in which the target
is known to lie after qn has been placed. Further, let rn be the size of Rn and an
the corresponding accuracy 1/rn . With gn := an /an1 = rn1 /rn , we denote
the increase of accuracy achieved by placing qn .

To create a close pair, one must place a query point qn close to a previously
placed query point, but then the newly created bisectors are also close to the
previously created bisectors. In particular, if a close pair is created, most new
bisectors inside Rn1 will be relatively close to the boundary of Rn1 and away
from the centre of Rn1 . Thus, creating close pairs cannot go together with
substantial accuracy gains if the target is near the centre of Rn1 . In eect, close
pairs are a resource whose scarcity limits the accuracy that can be obtained. We
will now make this insight more precise.

Denition 2 (d-close pair). A d-close pair is a pair of previously placed query

points that are consecutive in their left-to-right order and have a distance strictly
less than d divided by the current accuracy. Further, let pn (d) denote the number
of d-close pairs after qn has been placed.

The following three lemmas capture the trade-o between on the one hand, creat-
ing close pairs, and on the other hand, using and losing them while pin-pointing
the location of the target. Each lemma assumes that query points q1 , ..., qn have
been placed in response to learning R0 , ..., Rn1 , which we now consider to be
xed. Within this context, Rn is still a variable: the bisectors of qn with the
previously placed query points subdivide Rn1 into a set R of subintervals, and
Rn R is the subinterval among them that contains t. In turn, gn and pn (d)
are functions of Rn . Each lemma assumes a certain accuracy gain gn and de-
rives a bound for the net change in the number of d-close pairs pn (d), for all or
for some possible values of Rn . The lemmas address dierent levels of accuracy
gains: small, medium and large. (Some of the proofs are omitted due to space
limitations.)

Lemma 2 (small accuracy gain). If gn 1 + d/2 for all possible Rn R,

then pn (d) pn1 (d) + 1 for all possible Rn R.

Now, assume that the accuracy gain is somewhat higher than what we assumed
in the previous lemma. In this case no additional d-close pairs can be created
at all, as the following lemma shows. Note that it might even be the case that
several d-close pairs are lost.

Lemma 3 (medium accuracy gain). If gn 2 + d/2 for all possible Rn R,

then pn (d) pn1 (d) for all possible Rn R.

Lemma 4 (large accuracy gain). For any c > 1 and h, d such that h 2+d/2
and 2/h d 2, the following implication holds:
If gn > h for all possible Rn R, then there is a possible interval Rn R
and a k > 0, such that gn < ck /(c 1)(1 2/h) and pn (d) pn1 (d) k.
458 H. Haverkort et al.

Proof. Suppose that the placement of qn subdivides Rn1 into m + 2 regions.

Note that m 1, since otherwise it would not be possible to achieve an increase
in accuracy that is strictly greater than h > 2. Among these m + 2 regions,
each of size less than rn1 /h, there is a leftmost region and a rightmost region.
We call the other m regions interior regions and label them I1 , . . . , Im , in order
of decreasing size. Subtracting the size of the leftmost and rightmost region
from rn1 , we conclude that the total size of the interior regions is more than
(1 2/h)rn1 . Since c > 1 we may also conclude
m      
2 2 1
|Ik | > 1 rn1 > 1 rn1 1 m
h h c
k=1
   (m+1) 
2 c 1
= 1 rn1 (c 1) 1
h c1 1
m
(c 1)(1 2/h)rn1
= ,
ck
k=1

by using the geometric series. Therefore, there must be a k such that the k-th
largest interior region has size more than ck (c 1)(1 2/h)rn1 . If Rn is that
k-th largest region, the accuracy gain is upper-bounded by ck (c1)1 (12/h)1 .
It remains to show that in this case, the number of d-close pairs decreases
by at least k. Since h 2 + 2/d, by Lemma 3, no d-close pairs are created.
Each interior region Ii with i {1, . . . , k} is bounded by the bisectors of qn with
two previously placed query points q and q+ , which were neighbours in the
left-to-right order. The distance between q and q+ is
less than 2rn1 /h d rn1 , since, by the conditions of the lemma, Ii has
size less than rn1 /h;
at least 2rn d rn , since Ii is at least as big as Rn = Ik .
Thus, q and q+ constitute a d-close pair before, but not after placing qn .
We can now derive the following formula that expresses the cumulative eects
of Lemmas 2, 3 and 4 for a complete deterministic query strategy:
Lemma 5. For any h, d, c, and w such that h 2 + d/2 and 2/h d 2 and
1 < c w, for each non-negative integer n, it holds that

an wpn (d) bn

for certain locations of the target, where

    
d d c
b = max 1+ w2 , 2+ w, h, .
2 2 w(c 1)(1 2/h)
Proof (by induction). If n = 0, we have a0 = 1, p0 (d) = 0 and the claim holds.
Now suppose we add another query point qn . Distinguishing by the minimum
accuracy gain that is achieved for any target in Rn1 , one of the following cases
must hold:
How to play hot and cold on a line 459

Lemma 4 applies. Then, there is a possible Rn R and a k such that

gn < ck (c 1)1 (1 2/h)1 and pn (d) pn1 (d) k. Using the induction
hypothesis, we conclude

ck wk c
an wpn (d) < an1 wpn1 (d) bn1 bn .
(c 1)(1 2/h) w(c 1)(1 2/h)

Note that here we use the condition c w; without that condition we would
not be able to bound ck wk to c/w, or to any other constant.
Lemma 3 applies. So pn (d) pn1 (d) and, for some Rn R, we have
gn h, which implies an h an1 . Using the induction hypothesis, we may
conclude an wpn (d) han1 wpn1 (d) b bn1 = bn .
Lemma 2 applies. So pn (d) pn1 (d) + 1 and, for some Rn R, we have
gn < (2 + d/2), which implies an < (2 + d/2)an1 . With the induction
hypothesis, we get an wpn (d) < (2 + d/2)wan1 wpn1 (d) b bn1 = bn .
None of the Lemmas 2, 3, 4 apply. We have pn (d) pn1 (d) + 2, because
each new query point creates at most two d-close pairs, and gn < (1 + d/2),
so an wpn (d) < (1 + d/2)w2 an1 wpn1 (d) b bn1 = bn .

This establishes the induction step, and proves the lemma.

It remains to pick h, d, c and w such that we get the most out of Lemma 5. Our
goal is to prove a strong upper bound, so we want to minimize b. As can be seen
directly in the denition of b, it cannot hurt to choose d as small as possible
and c as large as possible, so we choose c = w and d = 2/h (this also makes the
condition h 2 + d/2 as permissive as possible). It remains to minimize:
    
1 2 1 1
max 1+ w , 2+ w, h, ,
h h (w 1)(1 2/h)
subject to h 2 + 1/h and w > 1. The maximum of the last three terms is
minimized when all are equalif either one or two of them would be higher
than the other, it would be possible to adjust h and w slightly such that the
higher terms are lowered. Solving for w and h gives us that the last three terms
are equal when h is the largest root of h3 4h2 + h + 1, so h is slightly smaller
than 3.6511 and w = h/(2 + 1/h) is approximately 1.6057. The rst term is of
no concern since for these values of h and w, which minimize the other three
terms, the rst term is even smaller. Thus we get b = h 3.6511.
Since pn (d) cannot be negative, this implies that there are targets for which
an an wpn (d) bn for all n, and thus, such targets are not located with
accuracy better than (bn ). This concludes the proof of Theorem 6.

4 Conclusions and outlook

For the one-shot variant of our location problem, we presented an upper bound
and a constructive lower bound on the accuracy that can be achieved. The
460 H. Haverkort et al.

remaining constant-factor gap between these bounds may be due to the fact
that most bisectors formed by pairs of both tightly or both widely spaced points
do not contribute anything to the accuracy. However, the upper bound might
also be improved.
For the incremental variant, we obtained non-trivial upper and lower bounds
but a gap that is exponential in n still remains. Our search strategy does not
seem to leave any space for substantial improvements, so further progress must
come from an entirely new search strategy, or from tightening the upper bound.
Observe that in each step, our search strategy only uses the bisectors that are
formed with the two previously placed query points that are closest to the target.
If we could prove that there is an optimal search strategy with this property,
then this would immediately improve the upper bound to O(3n ).
To search a d-dimensional unit cube, we could place roughly n/d query points
on each coordinate axis in a round-robin fashion; on each axis, we place the
points according to the one-shot, incremental, or on-line incremental strategy.
Thus, where we obtain accuracy h(n) in one dimension, we can pinpoint a target
to a cube of width 1/h(n/d) in d dimensions. This approach, however, fails to
take advantage of bisectors between query points placed on dierent axes, not
to mention query points placed more freely. Indeed, for d = 2, there are better
solutions for n {3, 4}.

Acknowledgements
We thank all participants of the 2016 Lorentz Center Workshop on Search Games,
and, in particular, Bengt Nilsson and Endre Csoka, for their invaluable contributions
in dening the problem and initial solutions.

References
1. H.-K. Ahn, S.-W. Cheng, O. Cheong, M. Golin, and R. van Oostrum. Competitive
facility location: the Voronoi game. Th. Comp. Sc., 310(1):457467, 2004.
2. R. Cole R. and Ch. K. Yap. Shape from probing. J. Alg., 8(1):19 38,1987.
3. J. Kahn and M. Saks Balancing poset extensions Order, 1: 113 126, 1984.
4. J. Matousek. Lectures on discrete geometry, Springer, New York, 2002.
5. S. Skiena. Interactive reconstruction via geometric probing. Proc. of the IEEE,
80(9):13641383, 1992.
Faster Algorithm for Truth Discovery via Range
Cover

Ziyun Huang1 Hu Ding2 Jinhui Xu1

1
Department of Computer Science and Engineering
State University of New York at Bualo
{ziyunhua, jinhui}@buffalo.edu
2
Department of Computer Science and Engineering
Michigan State University
[email protected]

Abstract. Truth discovery is a key problem in data analytics which

has received a great deal of attention in recent years. In this problem,
we seek to obtain trustworthy information from data aggregated from
multiple (possibly) unreliable sources. Most of the existing approaches
for this problem are of heuristic nature and do not provide any quality
guarantee. Very recently, the rst quality-guaranteed algorithm has been
discovered. However, the running time of the algorithm depends on the
spread ratio of the input points and is fully polynomial only when the
spread ratio is relatively small. This could severely restrict the applica-
bility of the algorithm. To resolve this issue, we propose in this paper
a new algorithm which yields a (1 + )-approximation in near quadratic
time for any dataset with constant probability. Our algorithm relies on
a data structure called range cover, which is interesting in its own right.
The data structure provides a general approach for solving some high
dimensional optimization problems by breaking them down into a small
number of parametrized cases.

1 Introduction
Truth discovery is an important problem arising in data analytics, and has re-
ceived a great deal of attentions in recent years in the elds of data mining,
database, and big data [3, 68, 4, 911]. Truth discovery seeks to nd trustwor-
thy information from a dataset acquired from a number of sources which may
contain false or inaccurate information. There are numerous applications for this
problem. For example, the latest search engines are able to answer user queries
directly, instead of simply listing webpages that might be relevant to the query.
This process involves retrieving answers from potentially a large number of re-
lated webpages. It is quite common that these webpages may provide inaccurate
or inconsistent information. Thus a direct answer to the query needs the search

The research of the rst and third authors was supported in part by NSF through
grants CCF-1422324, IIS-1422591, and CNS-1547167. The research of the second
author was supported by a start-up fund from Michigan State University.

Springer International Publishing AG 2017 461

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 461472, 2017.
DOI: 10.1007/978-3-319-62127-2_39
462 Z. Huang et al.

engine to be able to extract the most trustworthy information from all these
webpages, which is exactly the problem of truth discovery.
Truth discovery is an unsupervised learning problem. Besides the input data,
no prior knowledge about the reliability of each data source is provided. In such
settings, an intuitive approach is to view all data sources equally reliable and
obtain the solution by averaging or majority rule. A major issue of this approach
is that the yielded answer may be quite far away from the truth. This is because
a small number of unreliable data sources could signicantly deviate the nal
solution. To deal with this issue, truth discovery treats data sources dierently
by estimating the reliability for each of them. This greatly increases the level
of challenge for the problem. Moreover, since the truth discovery problem often
occurs in big data scenarios, the number of data sources could be quite large
and the dimensionality of the data could be rather high, which brings another
dimension of challenge to the problem.
A widely accepted geometric modeling of the truth discovery problem is the
follows. Data from each source is formulated as a set of real number attributes,
and thus can be viewed as a vector in Rd , where d is the number of attributes.
Each data source is associated with a positive variable (or weight) representing
its reliability. Formally, the truth discovery problem can be dened as follows.

Denition 1. (Truth Discovery [4, 8]). Let P = {p1 , p2 , . . . pn } be a set of points

in Rd space, where each pi represents the data acquired from the i-th source
among a set of n sources. The truth discovery problem is to nd the truth vector
p and wi (i.e., reliability) for each i-th source such that the following objective
function is minimized.
n
min i=1 wi pi p 2 , s.t. i=1
n
ewi = 1. (1)

The meaning of the above truth discovery formulation was discussed in [1]
from an information theorys point of view. It is shown that the constraint on wi
in Denition 1 ensures that the entropy is minimized when p approaches the
truth vector. For this reason, the problem is also called Entropy based Geometric
Variance problem [1].
Despite extensive studies on this problem, most of the existing techniques are
of heuristic nature, and do not provide any guarantee on the quality of solution.
It is not until very recently that the truth discovery problem has a theoretically
guaranteed solution [1]. This result ensures that a (1 + )-approximation of the
problem can be achieved in O(dn2 + (n) nd) time, where n is the number of
input points (i.e., data sources), d is the dimensionality of the space, is the
spread ratio of the input points (i.e. the ratio of the largest distance between
any two input points to the smallest distance), and is any xed small positive
number. The result is based on an elegant sampling technique called Simplex
Lemma [2] which is capable of handling high dimensional data. A main issue of
this method is that its running time depends on the spread ratio of the input
points, and is polynomial only when the spread ratio is relatively small (i.e.,
= O( n)). This could severely restrict its applicability.
Faster Algorithm for Truth Discovery via Range Cover 463

To overcome this main issue, we present in this paper a faster algorithm for
the truth discovery problem. With constant probability, our algorithm achieves
a (1 + )-approximation in O(dn2 (log n + log d)) time, and is completely inde-
pendent of the spread ratio. Our algorithm is also space ecient, using only
near linear space, while the space complexity of [1] also depends on the spread
ratio. Our algorithm relies on a new data structure called range cover, which is
interesting in its own right. Roughly speaking, range cover is a data structure
designed for a class of optimization problems (in high dimensional space) which
are decomposable into a number of easier cases, where each case can be char-
acterized by a parameterized assumption. For example, truth discovery can be
formulated as a problem of nding a truth vector p Rd from a given set P of
points in Rd so that a certain objective function (the exact formulation will be
discussed later) is minimized. We are able to show that although directly opti-
mizing the objective function is challenging, the problem is much easier to solve
if some additional information (e.g., the distance r between p and P ) is known.
Thus, by viewing the additional information as a parameterized assumption, we
can solve the truth discovery problem by searching for the best assumption. The
range cover data structure shows that even though the number of parameterized
assumptions could be very large (or even innite), it is sucient to sample only a
small number of assumptions to ensure an approximate solution. This leads to a
small-size data structure (i.e., O(n log n) space) and a faster algorithm for truth
discovery. Since the idea of decomposing problem into cases is not restricted only
to the truth discovery problem, we expect that this data structure will provide
new approaches to other problems.

2 Range Cover Data Structure

In this section, we present the aforementioned range cover data structure.

Range cover is motivated by several high dimensional optimization problems
(such as truth discovery). In these problems, an input point set P is given in Rd
space, and the objective is to nd a point q in Rd so that a certain objective func-
tion is optimized. A commonly used approach for such problems is to examine a
number of candidate points selected by some algorithms. But directly applying
such an approach could require too many (e.g., exponential in d) points to be
examined in high dimensional space. A possible way to overcome this diculty
is to characterize all possibilities of q into a small number of cases so that in each
case q is associated with a certain parametrized assumption which could help
solve the problem more eciently. For instance, in some optimization problem,
q could be much easier to obtain if we know in advance the nearest neighbor
(say p) of q in P and its distance r to q (i.e., p q = r) for some parameter r.
We expect that these parameterized assumptions form a space with much lower
dimensionality than d, and thus the overall time complexity can be signicantly
reduced.
From the above discussion we know that for the range cover data struc-
ture to be ecient, the problem needs to be decomposable into a small number
464 Z. Huang et al.

of easier cases. For this purpose, we will take advantage of the distribution
of the points in P , such as their locality and point aggregation properties. To
understand how point aggregation can be useful, consider the following param-
eterized assumption on q: Assume that p is the nearest neighbor of q in P and
r is their distance. Denote this assumption by N N q (p, r). If a subset of points,
v = {p1 , p2 , . . . , pm }, are close to each other compared to r, i.e. their diameter
D(v) is no larger than r for some predened small constant > 0, then points in
v can be viewed as a single heavy point (simply denoted by v for convenience),
and assumptions N N q (p1 , r), N N q (p2 , r), . . . , N N q (pm , r) can be covered (or
replaced) by a single assumption N N q (v, r) without losing much quality. We
formally dene N N q (v, r) for aggregated subset v as follows.

Assumption 1 N N q (v, r): For a subset v of P , N N q (v, r) is an assumption

made about q which says: D(v) r for some small constant > 0, where D(v)
is the diameter of v, and r p q (1 + )r holds for p which denotes the
nearest neighbor of q in v.

Another property of P which can be made use of is the domination relation.

If q is very close to an aggregated subset of points v P compared to points
in P \ v, it is often a degenerated case for the problem and relatively easy to
solve. To cover such cases, we dene the following assumption DOMq (v) for
predened constants > 0 and > 0.

Assumption 2 DOMq (v): For a subset v of P , DOMq (v) is an assumption

made about q, which says: there exists a point pv v such that D(v) q pv 
and pv q pv q for any point pv P \ v, where D(v) is the diameter
of v.

With the above denitions of assumption, we know that the goal of the range
cover data structure is to generate a small number of assumptions DOMq (v1 ),
DOMq (v2 ), . . ., DOMq (vh ) and N N q (v1 , r1 ), N N q (v2 , r2 ), . . . , N N q (vg , rg ), so
that for any q Rd , at least one of these assumptions holds. We call such a
collection of assumptions an assumption coverage.
The main idea of range cover is to build a series of views of P formed by
aggregated subsets from dierent scales of r, which is a controlling factor and can
be interpreted as the distance of observation. Range cover identies, for each r, a
collection of disjoint aggregated subsets v of P with diameter no larger than r
for some predened small constant > 0. The collection could be used as a sketch
of P observed from distance r, which takes much less space than P . These views
(from dierent distances r) jointly provide an easy way to access the skeleton
information of P , which allow us to produce a smaller size assumption coverage.
Particularly, for a given r, instead of generating assumptions N N q ({p}, r) for
each point p P , we produce coarse-grained assumptions N N q (v, r) for every
v in this view. Furthermore, by utilizing domination relation, we do not need to
consider small values of r, and thus can further reduce the size of the assumption
coverage. This is because the aggregation-based views of P from small enough rs
correspond to situations where q is very close to some point and the domination
Faster Algorithm for Truth Discovery via Range Cover 465

relation holds. Note that when determining point aggregation, we need not to
consider too large r as well, since for large enough r the whole point set P is an
aggregated set.
To generate the assumption coverage, an obvious challenge is how to reduce
the number of possible values for r for which we need to build a view of P . Even
though there is no need to consider too large and too small values for r, the gap
between the maximum and minimum values often depends on the spread ratio
of P , which could lead to pseudo-polynomial running time for some algorithms
using the range cover data structure. Below we will show how to overcome this
challenge and obtain a small size range cover.

2.1 Range Cover and Assumption Coverage

The range cover data structure uses the aggregation tree as an ingredient. The
aggregation tree is a version of Hierarchical Well-Separated Tree (HST)[5] which
is dened conveniently for point aggregation in a well-behaved manner. The
denition is as follows.
1. Every node v (called aggregation node) represents a subset P (v) of P , and
the root represents P .
2. Every aggregation node v is associated with a representative point l(v)
P (v) and a size s(v) which is an upper bound on the diameter of P (v).
3. Every leaf node corresponds to one point in P with size s(v) = 0, and each
point appears in exactly one leaf node.
4. The two children v1 and v2 of any internal node v form a partition of v with
max{s(v1 ), s(v2 )} < s(v).
s(vp )
5. For every aggregation node v with parent vp , rout is bounded by a polyno-
mial function P(n, d) 1 (called distortion polynomial), where rout is the
minimum distance between any point in P (v) and any point in P \ P (v).
The following theorem shows that an HST with polynomial distortion (there-
fore, the aggregation tree also) can be built within near linear time.

Theorem 1. [5] An HST with distortion polynomial O( dn5 ) can be built in
O(dn log n) time with success probability 1 1/n.

Below we will show how to build a range cover data structure from a given
aggregation tree Tp which ensures to form an assumption coverage.
Consider an aggregation node v from distance r. If the diameter of v is not
larger than r for a predened constant > 0, all points in v can be viewed as
an aggregated subset and thus is part of the view from r. If r is so large that
even the parent v  of v in Tp is an aggregated subset, v can be replaced by v  in
the view. This means that an aggregation node v should not appear in the view
from a far enough distance r. Also if r is small, either v has a too large diameter
and thus cannot be an aggregated subset or v dominates q (i.e. the solution
point). In the former case, v should be replaced by one of its descendant in the
view. In the latter case, we do not include v in the view from distance r, with
466 Z. Huang et al.

the belief (which will be proved later) that the absence of v can be compensated
by including the DOMq (v) assumption in the assumption coverage.
The above observation implies that for any aggregation node v, there exists
a range (rL , rH ) of the value of r, such that v is only visible when r lies in
the range. This immediately suggests the following scheme. Divide the set of all
positive real numbers into intervals ((1 + )t , (1 + )t+1 ], t = . . . , 2, 1, 0, 1, . . .,
and associate each of them with a bucket. If an interval (a, b] lies within the
interval (rL , rH ) of a aggregation node v, then insert v into the bucket of (a, b].
The collection of these buckets is then the desired range cover data structure.

Algorithm 1 RangeCover(Tp , , )
Input: A aggregation tree Tp built over a set P of points in Rd ; an approximation
factor 0 < < 14 , a controlling factor 0 < < 1.
Output: A number of sets of aggregation nodes, each of which is associated with an
interval ((1 + )t , (1 + )t+1 ] for some integer t.
1: For every interval ((1 + )t , (1 + )t+1 ], create an empty bucket Bt . (Note that Bt
will not be actually created until some aggregation node v is inserted into it.)
2: For every non-root node v of Tp , let vp be its parent in Tp , rH be s(vp )/, and rL
be max{s(v)/, s(vp )/(16P(n, d))}. Do
For every integer t satisfying the condition of rL (1 + )t < rH , insert v into
bucket Bt .

Given input P , for any constant factors 0 < < 1/4 and > 0 in Assump-
tion 1 and Assumption 2, we build the aggregation tree Tp and the corre-
sponding range cover data structure R by calling RangeCover(Tp , , ), and let
the assumption coverage A, (or simply A for convenience) contain the following
assumptions:
1. DOMq (v), for every aggregation node v of Tp
2. N N q (v, r), for every aggregation node v of Tp and r such that interval
(r, (1 + )r] is one of the nonempty bucket in R and v is a aggregation node
in this bucket.
Clearly obtaining A from R is quite straightforward, and |A| has a size no
larger than that of R.
The following theorem shows that A is indeed an assumption coverage.
Theorem 2. For any q in Rd , at least one of the assumptions in A holds.
Proof. Let p be the nearest neighbor of q in P . If q p  = 0, DOMq ({p })
holds. In the following we assume that q p  > 0. Let t be the integer such
 
that (1 + )t < q p  (1 + )t +1 . Let v  be a aggregation node of Tp which

is the highest ancestor of {p } in Tp such that s(v  ) (1 + )t . Since {p } is a

leaf of Tp and s({p }) = 0 (1 + )t , such a v  always exists.
Based on the relationship between v  , t and the range cover data struc-

ture, we have 4 cases to consider. (a) v  is the root of Tp , (b) (1 + )t <
Faster Algorithm for Truth Discovery via Range Cover 467

max{s(v  )/, s(vp )/(16P(n, d))}, where vp is the parent of v  in Tp , (c) (1 +
 
)t s(vp )/, and (d) max{s(v  )/, s(vp )/(16P(n, d))} (1 + )t < s(vp )/.
Below we analyze each of them.

Case (a): Since s(v  ) (1 + )t q p  and v  represents the whole
point set P (as it is the root of Tp ), we have P \ v  is empty. This means that
the assumption DOMq (v  ) holds for q.

Case (b): Note that by the denition of t , we know that (1 + )t s(v  )/.
t  
Therefore if case (b) occurs, we have (1+) s(vp )/(16P(n, d)). By (1+)t <

q p  (1 + )t +1 and < 1, it follows that q p  s(vp )/(8P(n, d)).
Let po be any point in P \ v  . Then po p  s(vp )/P(n, d) by the property of
aggregation tree. Therefore, po p  8q p . Thus, po q po p 
q p  (8/ 1)q p . By the fact < 1, we have q p  po q.
  
Also since (1 + )t s(v  )/ and (1 + )t < q p  (1 + )t +1 , we have
q p  s(v  )/. This indicates that DOMq (v  ) holds for case (b).
Case (c): This case actually never occurs. This is because, by the denition

of v  , s(vp ) > (1 + )t , since otherwise v  cannot be the highest ancestor of

{p } satisfying the inequality s(v  ) (1 + )t .


Case (d): Note that this case means that v is placed in bucket ((1+)t , (1+
t +1  t  t
) ]. Thus N N q (v , (1 + ) ) is in A. We show that N N q (v , (1 + ) ) holds
for q. Indeed, this follows immediately from previous discussion on v  : s(v  )

(1 + )t and (1 + )(1 + )t p q > (1 + )t .
Since in all cases at least one assumption in A holds for q, the theorem
follows. 

The following theorem indicates that the size of the assumption coverage is
small.
Theorem 3. Given a aggregation tree Tp and factors 0 < < 1/4 and 0 <
< 1, the range cover data structure can be built in O(1/ log(1/)n(log n +
log d)) time and takes O(1/ log(1/)n(log n+log d)) space. Consequently, |A| =
O(1/ log(1/)n(log n + log d)).
Proof. From Algorithm 1, we know that every aggregation node v is inserted
into O(log1+ rH /rL ) buckets (see Step 2 of the algorithm). Note that log1+ rH /rL
is no larger than log1+ ((s(vp )/)/ (s(vp )/16P(n, d))) = O(1/ log(1/)(log n+
log d)). Since the total number of aggregation node is O(n), the theorem follows.



3 Solving Truth Discovery with Assumption Coverage

In this section, we show how to use the assumption coverage to solve the truth
discovery problem. Given any point set P in Rd and a small constant 0 <  < 1,
we rst build an assumption coverage A with factors and whose values
depend on  only and will be determined later. We then show how to obtain a
(1 + )-approximation of the problem in polynomial time. Let p be the truth
vector (i.e., optimal solution) of the problem.
468 Z. Huang et al.

We rst borrow a useful lemma from [6]. It shows that once p is determined,
the weights wi can also be determined. Thus we only need to nd an approximate
truth vector p .
Lemma 1. [6] If the truth vector p is xed, the following value for each weight
wl minimizes the the objective function (1) (in Denition 1),
n
p pi 2
wl = log( i=1 ). (2)
p pl 2
There are two types of assumptions about p in A which covers all possibilities
of p : N N p (v, r) and DOMp (v). Below we discuss each of them.
The following lemma shows that DOMp (v) is easy to solve.
Lemma 2. By setting 1/4 and /4, if DOMp (v) holds for the truth
vector p , there exists a point p v P such that p is a (1 + )-approximation
of the truth discovery problem (using the objective function (1) in Denition
1).
From the above lemma, we know that if DOMp (v) holds for some v, then
one of the input point in P will be a (1 + )-approximation. This means that we
can handle all such cases by trying every input point as p by computing the
objective function (1) in equation (2), and choosing the one with the minimum
objective value as the solution. This takes O(dn2 ) time.
The following lemma shows that N N p (v, r) can also be handled eciently.
We leave the proof to the next subsection.
Lemma 3. If N N p (v, r) holds for any factor 0 < < 1/4, then a (1 + )-
approximation can be computed in time O(dn) with constant probability, where
 is a small constant in (0, 1).
The above lemmas suggest that we can compute an approximate p by the
following algorithm.
1. Compute an aggregation tree from P .
2. Set = /4, = 1/5, compute a range cover from the aggregation tree.
3. Compute A from the range cover.
4. Try every p P as a candidate for the truth vector. Choose the one, say p1 ,
that minimizes the objective function.
5. For every N N p (v, r) in A, compute a candidate for p . Choose the one, say
p2 , that minimizes the objective function.
6. Choose from p1 and p2 the one that minimizes the objective function
In the above algorithm, Step 1 takes O(dn log n) time. Step 2 needs O(n(log n+
log d)) time (where  is hidden in the O() notion). Step 3 costs O(n(log n+log d))
time. Step 4 can be done in O(dn2 ) time. Step 5 takes O(dn2 (log n+log d)) time,
since we test at most O(n(log n + log d)) assumptions in A. Step 6 requires only
O(1) time. For the space usage, it can be computed O(dn log n) + O(n(log n +
log d)) + O(n(log n + log d)) + O(dn) + O(dn) + O(1) = O(dn(log n + log d)). Thus
we have the following main theorem.
Faster Algorithm for Truth Discovery via Range Cover 469

Theorem 4. Given any set P of n points in Rd , with constant probability, it is

possible to compute a (1+)-approximate solution for the truth discovery problem
in O(dn2 (log n + log d)) time. The space usage can be made to O(dn(log n +
log d)).

3.1 Solving N N p (v, r)

In this section we prove Lemma 3. We assume that N N p (v, r) holds for p ,
where v P and r > 0.
Lemma 1 reveals how the weight wi of every pi P is related to p . It
is clear from the objective function (1) and Lemma 1 that p is the weighted
mean of P . Since we do not know p in advance, wi is also unknown for every
pi P . The truth discovery problem can be viewed as a problem of nding the
weighted mean of a point set with unknown weights. Our strategy for solving
this problem consists of two main steps: (1) we partition P into a number of
subsets (or sub-clusters), with each having some nice property. The weights of
the points in some clusters are approximately known, while the weights of the
points in other clusters are unknown, but have an upper and lower bound; (2)
we apply a technique in [1] to nd the approximate weighted mean point of each
subset, and combine them to estimate p .

Partitioning P for Estimating Weights We rst show how to estimate the

weights of some points by N N p (v, r) without knowing p . This is crucial for
our algorithm to be ecient for any point set P .
Let p1 v denotes the representative point l(v) of v. We label the rest of
points in P as p2 , p3 , . . ., pn . For each point pi P , dene ri = max(p1 pi , r)
and ri = p pi . For N N p (v, r), let pis v be the nearest neighbor of p in
P . Below we derive the relationship between ri and ri .
First, we consider the case that max(p1 pi , r) = r. In this case, we have
ri pis p  r = ri by assumption N N p (v, r) and the fact that pis is the
nearest neighbor of p . Also we have ri p1 p  + p1 pi  p1 p  + r,
and
p1 p  p1 pis  + p pis  D(v) + (1 + )r (1 + 2)r.
Thus, ri (2 + 2)r = (2 + 2)ri . Putting all together, we have ri ri
(2 + 2)ri .
Then, we consider the case that max(p1 pi , r) = p1 pi . In this case,
ri = p1 pi  r. Again, we have p1 p  p1 pis  + p pis 
D(v) + (1 + )r (1 + 2)r. Therefore, (1 + 2)ri p1 p . Thus,
ri = pi p  p1 pi  + p1 p  p1 pi  + (1 + 2)ri = (2 + 2)ri .
Next, we consider 2 subcases, ri 2r and ri < 2r. If ri < 2r, since ri r, we have
ri > ri /2. If ri 2r, since p1 p  (1+2)r, we have p1 p  (1+2)ri /2.
This means that
ri = pi p  p1 pi  p1 p  ri (1 + 2)ri /2 = (1 2)ri /2.
470 Z. Huang et al.

To conclude, we have (1 2)ri /2 ri (2 + 2)ri .

From the above analysis and the fact that < 1/4, we can obtain the fol-
lowing.
ri /4 ri 4ri . (3)

For each pi P , let wi = log(( pj P rj2 )/(ri2 )), i.e., wi is the optimal weight

determined by Lemma 1. Let wi = log(( pj P rj2 )/(ri2 )). From inequality (3),
we obtain the following:

wi log 256 wi wi + log 256. (4)

This means that wi can be used as an approximation of wi if wi is large

enough.
For any pi P , if wi 8/ log 256/ for any 0 < < 1, we have the
following (by (4))
(1 )wi wi (1 + )wi .
This means that wi can be well approximated by wi in this case. Let P denote
the set {pi P |wi 8/}.
Next, we further show that there is at most one point pi in P with weight
wi < log 36/25 which, if exists, can be identied by a simple procedure. By the
denition of wi , we know that wi < log 36/25 can happen only when p pi  >
5p pj  for any i = j. This means that for any j, l = i,

pj pl  p pj  + p pl  2max(p pj , p pl ). Thus, we have

pj pi  p pi  p pj 
> 5max(p pj , p pl ) max(p pj , p pl )
= 4max(p pj , p pl ) 2pj pl .
Hence, for any j, l = i, the inequality pj pl  < pi pj /2 holds. In other
words, pi is isolated from the rest of the points in P . It is easy to see that such
a pi is unique, if exists. The following procedure searches for such a pi .

1. Choose an arbitrary point p from P .

2. Find a point p in P farthest away from p.
3. Find a farthest point p from p in P .
4. Compare the pairwise distances among the three points in {p, p , p }. Throw
away the pair of points with the smallest pairwise distance. Output the
remaining point as p.

From the above discussion, it is easy to see that if there is a point pi with
weight wi < log 36/25, it must be p. Clearly, this procedure takes only O(dn)
time.
For a constant 0 < < 1/2 (whose value will be determined later), let
Pu = P \ (P {p}) and P< = {p} \ P . Then, Pu , P< , P form a partition of
P . P contains all points pi in P whose weights wi have already been roughly
determined (i.e., approximated by wi ); P< has at most one point, which will be
Faster Algorithm for Truth Discovery via Range Cover 471

Algorithm 2 (1 + O(1))-approximate Truth Discovery from N N p (v, r)

Input: A set P of n points in Rd space. Assumption N N p (v, r). = 2 . Constants
16k2
, k solved from 2 k 2 and k = log1+ 2 log1636/25  + 1. c1 = 4k
2 log 2 .

c2 = 4k2 . = 3 /48k.
Output: An approximate truth vector.

1: Identify P< , P , Pu by computing wi for each pi P .

2: Compute the weighted mean o3 of P using weights wi .
3: Randomly sample c1 points from P . Enumerate all subsets of c2 points from the
sample. Compute means of these subsets, and put all the means into a set M .
4: For every k-subset {o1 , . . . ok } of M , apply SIMPLEX(2 , k, o1 , . . . , ok ) to produce
a grid. Put all grid points in into a point set G.
5: For every o2 in G, if P< contains a point o1 , then build a grid by applying
SIMPLEX(, 3, o1 , o2 , o3 ); otherwise, build a grid using SIMPLEX(, 2, o2 , o3 ).
6: Try all the grid points produced above. Output the one that minimized the ob-
jective function (1).

the one with weight smaller than log 36/25, if exists; Pu contains all the remain-
ing points whose weights are not known yet. Pu , P< , P together with wi can be
obtained in O(dn) time since it takes a total of O(n) distance computations.
Following a similar idea in [1], we further decompose Pu by using the log-
partition technique, where > 0 is a constant to be determined later. (Note
that the log-partition cannot be explicitly obtained since we do not know the
weights wi . We assume that such a partition exists and will be used in our later
analysis.)
Denition 2. The log-partition of Pu divides points in Pu into k groups G1 , . . . Gk
as follows, where k = log1+ log16/
36/25 + 1: Gi = {pj Pu |(1 + )
i1
log 36/25
wj (1 + ) log 36/25}.
i

Note that the above partition indeed involves all points in Pu . This is because
by the denition of P< and P , and the fact that (1 )wi wi (1 + )wi for
all point pi P , we know that log 36/25 wi 16 for each point pi Pu .
This implies that G1 , . . . Gk , P< , P form a partition of P . Also, we apply log-
partition to Pu instead of P as in [1]. In this way the value of k is bounded,
making our algorithm ecient for any data.

Applying the Simplex Lemma Roughly speaking, Simplex Lemma in [1]

provides a procedure SIMPLEX(, k, o 1 , . . . , ok ) to approximate the weighted
mean of a partitioned point set Q = Qi , where  is an approximate factor,
k is an integer and every oi is a point in Rd . The procedure outputs a grid of
size ((8k/)k ) within O((8k/)k ) time which ensures that at least one of the grid
points is close to the weighted mean of Q, if oi is a good approximation of the
weighted mean of Qi .
Algorithm 2 shows how to use SIMPLEX to produce an approximate truth
vector, given P partitioned into P< , P , Pu as above. The running time and
472 Z. Huang et al.

space usage match those appear in Lemma 3. To obtain a (1 + )-approximation,

we only need to do a scaling on the constants without aecting the asymptotic
running time.
Below we briey explain the main steps of Algorithm 2. In Step 1 we partition
P into P< , P , Pu as mentioned before. In Step 2 an approximate weighted mean
of P is computed. In Steps 3 and 4, we try to guess k weighted means {o1 , . . . ok }
for the clusters G1 , . . . Gk resulted from the log-partition of Pu by using random
sampling. We apply SIMPLEX to these approximate means {o1 , . . . ok } to pro-
duce a small grid. The set G of grid points contains at least one point which is
a good approximate weighted mean of Pu . In Steps 5 and 6, we already have
approximate weighted means o1 and o3 of P< and P , respectively, and a set G
which contains an approximate weighted mean o2 of Pu . We then try all possible
o2 from G and use SIMPLEX on o1 , o2 , o3 to produce grids and one of such grids
contains the desired approximation of the truth vector.

References
1. Ding, H., Gao, J., and Xu, J.: Finding Global Optimum for Truth Discovery: En-
tropy Based Geometric Variance. Leibniz International Proceedings in Informatics
(LIPIcs), 32nd International Symposium on Computational Geometry (SoCG 2016),
Vol. 51, 34:1-34:16(2016).
2. Ding, H. and Xu, J.: A Unied Framework for Clustering Constrained Data without
Locality Property. Proceedings of ACM-SIAM Symposium on Discrete Algorithms
(SODA 2015), pp. 1471-1490, January 4-6, 2015, San Diego, California, USA.
3. Dong, X.L., Berti-Equille, L., Srivastava, D.: Integrating conicting data: The role
of source dependence. PVLDB, 2(1): 550-561(2009).
4. Li, Y., Gao, J., Meng, C., Li, Q., Su, L., Zhao, B., Fan, W., Han, J.: A Survey on
Truth Discovery, CoRR abs/1505.02463(2015).
5. Har-Peled, S.: Geometric approximation algorithms. Vol. 173. Boston: American
mathematical society(2011).
6. Li, H., Zhao, B., Fuxman, A.: The Wisdom of Minority: Discovering And Targeting
The Right Group of Workers for Crowdsourcing. Proc. of the International Confer-
ence on World Wide Web (WWW14), pp. 165-176(2014).
7. Li, Q., Li, Y., Gao, J., Su, L., Zhao, B., Demirbas, M., Fan, W., Han, J.: A
Condence- Aware Approach for Truth Discovery on Long-Tail Data. PVLDB 8(4):
425-436(2014).
8. Li, Q., Li, Y., Gao, J., Zhao, B., Fan, W., Han, J.: Resolving Conicts in Heteroge-
neous Data by Truth Discovery and Source Reliability Estimation. Proc. the 2014
ACM SIGMOD International Conference on Management of Data (SIGMOD14),
pp. 1187-1198(2014).
9. Pasternack, J., Roth, D.: Knowing what to believe (when you already know some-
thing). Proc. of the International Conference on Computational Linguistics (COL-
ING10), pp. 877-885(2010).
10. Whitehill, J., Ruvolo, P., Wu, T., Bergsma, J., Movellan, J.: Whose Vote Should
Count More: Optimal Integration of Labelers of Unknown Expertise. Advances in
Neural Information Processing Systems (NIPS09), pp. 2035-2043(2009).
11. Yin, X., Han, J., and Yu, P.S.: Truth discovery with multiple conicting informa-
tion providers on the web: Proc. of the ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining (KDD07), pp. 1048-1052(2007).
 Searching edges in the overlap of two plane graphs

John Iacono1 , Elena Khramtcova2 , and Stefan Langerman2

1
Department of Computer Science and Engineering, New York University
New York, USA, [email protected]
2
Computer Science Department, Universite libre de Bruxelles (ULB)
Brussels, Belgium, {elena.khramtcova,stefan.langerman}@ulb.ac.be

Abstract. Consider a pair of plane straight-line graphs whose edges are colored
red and blue, respectively, and let n be the total complexity of both graphs. We
present a O(n log n)-time O(n)-space technique to preprocess such a pair of
graphs, that enables efcient searches among the red-blue intersections along
edges of one of the graphs. Our technique has a number of applications to ge-
ometric problems. This includes: (1) a solution to the batched red-blue search
problem [Dehne et al. 2006] in O(n log n) queries to the oracle; (2) an algo-
rithm to compute the maximum vertical distance between a pair of 3D polyhe-
dral terrains, one of which is convex, in O(n log n) time, where n is the total
complexity of both terrains; (3) an algorithm to construct the Hausdorff Voronoi
diagram of a family of point clusters in the plane in O((n + m) log3 n) time and
O(n + m) space, where n is the total number of points in all clusters and m is
the number of crossings between all clusters; (4) an algorithm to construct the
farthest-color Voronoi diagram of the corners of n disjoint axis-aligned rectan-
gles in O(n log2 n) time; (5) an algorithm to solve the stabbing circle problem
for n parallel line segments in the plane in optimal O(n log n) time. All these
results are new or improve on the best known algorithms.

1 Introduction
Many geometric algorithms have subroutines that involve investigating intersections
between two plane graphs, often assumed being colored red and blue respectively. Such
subroutines differ in the questions that are asked about the red-blue intersections. The
most well-studied questions are to report all red-blue intersections or to count them. It
is shown how to report all the intersections in optimal O(n log n + k) time and O(n)
space [3, 4, 15, 16, 18], where n is the total complexity of both graphs, and k is the size
of the output. Note that k may be (n2 ). Counting the red-blue intersections can be
carried out in O(n log n) time and O(n) space [4, 16].
In this paper, we consider the situation where one wants to search the red-blue in-
tersections, though avoiding to compute all of them. Problems of this type appear as
building blocks in diverse geometric algorithms. The latter include: distance measure-
ment between polyhedral terrains [4], motion planning [12], construction of various
generalized Voronoi diagrams (divide-and-conquer [9, 6] or randomized incremental [5]
construction). Therefore solving such problems efciently is of high importance.
Often it is guaranteed that each red edge contains at most one sought red-blue inter-
section, and an oracle is provided, that, given a red-blue intersection, is able to quickly

Springer International Publishing AG 2017 473

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 473484, 2017.
DOI: 10.1007/978-3-319-62127-2_40
474 J. Iacono et al.

determine to which side of that intersection the sought intersection lies along the same
red edge (see Section 3.1 for more details on this setting). A particular case, when the
red graph consists of a unique edge, appeared under the name of segment query in the
randomized incremental construction algorithm for the Hausdorff Voronoi diagram [5],
and under the name of nd-change query in an algorithm to solve the stabbing circle
problem for a set of line segments [8]. If the blue graph is a tree, it can be preprocessed
in O(n log n) time using the centroid decomposition [2, 5]. Centroid decomposition
supports segment (or nd-change) queries for arbitrary line segments, requiring only
O(log n) queries to the oracle [5, 7]. If the blue graph is not a tree, then in O(n log n)
time it can be preprocessed for point location, and a nested point location along the
red edge is performed, which requires O(log2 n) queries to the oracle [6, 8]. For two
general plane straight-line graphs (where the red graph is not necessarily one edge) the
problem is called batched red-blue intersection problem (see Problem 2). It was for-
mulated in Dehne et al. [9], and solved in O(n log3 n) time and O(n log2 n) space [9]
using hereditary segment trees [4]. However, this is optimal in neither time nor space.
We present a data structure that provides a clear interface for efcient searches for
red-blue intersections along a red edge. Our data structure can be used to improve the
above result [9] (see Section 3.1), which includes an improvement on segment (or nd-
change) queries in plane straight-line graphs. Our data structure can also handle more
general search problems, e.g., a setting when a red edge may have more than one sought
red-blue intersection on it. Below we state our result and its applications.

1.1 Our result

Let R, B be a pair of plane straight-line3 graphs. We address the following problem.

Problem 1 (RB-Preprocessing problem). Given graphs R, B, construct a data structure

that for each edge e of R stores implicitly the intersections between e and the edges of
B sorted according to the order, in which these intersections appear along e. Let Te be a
perfectly balanced binary search tree built on the sorted sequence of intersections along
e. The data structure should answer efciently the following navigation queries in Te :

Return the root of Te ;

Given a non-root node of Te , return the parent of this node;
Given a non-leaf node of Te , return the left (or the right) child of this node.

We provide a solution to the RB-Preprocessing problem, where each of the naviga-

tion queries can be answered in O(1) time, and constructing the data structure requires
O(n log n) time and O(n) space, where n is the total number of vertices and edges in
both R and B (see Section 2).
The resulting data structure allows for fast searches for interesting intersections
between edges of R and the ones of B. We note that the notion of interesting is external
to the data structure: It is not known at the time of preprocessing, but rather guides the
searches on the data structure after it is built. In particular, for the input graphs R and
3
Our technique can be trivially generalized to apply to x-monotone pseudoline arcs in place of
straight-line edges of the graphs.
Searching edges in the overlap of two plane graphs 475

B, the data structure is always the same, while interesting intersections can be dened
in several ways, which of course implies that the searches may have different outputs.
Our preprocessing technique can be applied to a number of geometric problems. We
provide a list of applications, which is not exhaustive. For each application, we show
how to reduce the initial problem to searching for interesting red-blue intersections,
and how to navigate the searches, that is, how to decide, which subtree(s) of the cur-
rent node of the (implicit) tree to search. Using our technique we are able to make the
contributions listed below, and we expect it to be applicable to many more problems.

1. The batched red-blue search problem [9] for a pair of segment sets can be solved
in O(n) space and O(n log n) queries to the oracle, where n is the total number of
segments in both sets (see Section 3.1). The problem is as follows. Given are: (1)
two sets of line segments in the plane (colored red and blue, respectively), where
no two segments in the same set intersect; and (2) an oracle that, given a point
p of intersection between a red segment r and a blue segment b, determines to
which side of segment r with respect to point p the interesting red-blue intersection
lies. It is assumed that each segment contains at most one interesting intersection.
The batched red-blue search problem is to nd all interesting red-blue intersec-
tions. Our solution is an improvement on the one of Dehne et al. [9] which requires
O(n log2 n) space and O(n log3 n) queries to the oracle.
2. The maximum vertical distance between a pair of polyhedral terrains, one of which
is convex, can be computed in O(n log n) time and O(n) space (see Section 3.2).
Previously, a related notion of the minimum vertical distance between a pair of
non-intersecting polyhedral terrains was considered, and it was shown how to nd
it in O(n4/3+ ) time and space for a pair of general polyhedral terrains [4], in
O(n log n) time for one convex and one general terrain [22], and in O(n) time
for two convex terrains [22]. Our technique yields an alternative solution for the
second case within the same time bound as in [22]. The maximum distance for
non-intersecting polyhedra can be found by the above methods [4, 22], however it
is different from the minimum distance for intersecting polyhedra: asking about the
former is still interesting, while the latter is trivially zero.
3. The Hausdorff Voronoi diagram of a family of point clusters in the plane can be
constructed in O((n + m) log3 n) time, where m is the total number of pairwise
crossings of the clusters (see Section 3.3). Parameter m can be (n2 ), but is small
in practice [20, 19]. There is a deterministic algorithm to compute the diagram in
O(n2 ) time [11]. All other known deterministic algorithms [19, 20] have a running
time that depends on parameters of the input, that cannot be bounded by a function
of m.4 Each of them may take (n2 ) time even if m = 0. There is a recent random-
ized algorithm with expected time complexity O((m + n log n) log n)) [14]. For
a simpler case of non-crossing clusters (m = 0), the diagram can be computed in
deterministic O(n log5 n) time5 [9], or in expected O(n log2 n) time [5, 14]. Thus
our algorithm is the best deterministic algorithm for the case of small number of
4
The algorithms have time complexity respectively O(M + n log2 n + (m + K) log n) and
O(M  + (n + m + K  ) log n), where parameters M, M  , K, K  reect the number of pairs
of clusters such that one is enclosed in a certain type of enclosing circle of the other.
5
The time complexity claimed in [9] is O(n log4 n). See the discussion in Section 3.3.
476 J. Iacono et al.

crossings. The time complexity of our algorithm is subquadratic in n and m and

depends only on them, unlike any previous deterministic algorithm.
4. The farthest-color Voronoi diagram for a family of n point clusters, where each
cluster is all the corners of an axis-aligned rectangle,6 and these rectangles are
pairwise disjoint, can be computed in O(n log2 n) time and O(n) space. Previous
results on the topic are as follows. For arbitrary point clusters, the diagram may
have complexity (n2 ) and can be computed in O(n2 ) time and space [1, 11],
where n is the total number of points in all clusters. When clusters are pairs of
endpoints of n parallel line segments, the diagram has O(n) complexity and can be
constructed in O(n log n) time and O(n) space [8]. In this paper, we broaden the
class of inputs, for which the diagram can be constructed in subquadratic time. We
also show that the complexity of the diagram for such inputs is O(n).
5. The stabbing circle problem for line segments in the plane can be solved in time
O(THVD(S) + TFCVD(S) + (|HVD(S)| + |FCVD(S)| + m) log n), where |HVD(S)|
and |FCVD(S)| denote respectively the complexity of the Hausdorff and the farthest-
color Voronoi diagram of the pairs of endpoints of segments in S, THVD(S) and
TFCVD(S) denote the time to compute these diagrams, and m is a parameter reect-
ing the number of bad pairs of segments in S. If all segments in S are parallel to
each other, the stabbing circle problem can be solved in optimal O(n log n) time
and O(n) space. This is an improvement over the recent O(THVD(S) + TFCVD(S) +
(|HVD(S)| + |FCVD(S)| + m) log2 n) time technique for general segments, which
yielded an O(n log2 n) time algorithm for parallel segments [8].

2 The technique to preprocess a pair of graphs

Suppose we are given two plane straight-line graphs R and B, and let n be the total
number of vertices and edges in both R, B. We assume that no two vertices of the
graphs have the same x coordinate. In this section, it is more convenient to treat R
and B as two sets of line segments in the plane, where the segments in R are colored
red, and the ones in B are colored blue. No two segments of the same color intersect,
although they may share an endpoint.
Our preprocessing technique consists of three phases. In the rst phase, we invoke
an algorithm that nds the intersections between the edges of the two graphs (see Sec-
tion 2.1). After that, in the second phase, we build a linearized life table for the red
segments (see Section 2.2). Finally we sweep the life table with a line, which provides
us the resulting data structure (see Section 2.3).

2.1 Finding red-blue intersections

For the sets R (red) and B (blue), we need to nd all the intersections between segments
of different color, i.e., all the red-blue intersections.
It is known how to count the red-blue intersections in optimal O(n log n) time [4,
16, 18], or report them in optimal O(n log n + k) time [3, 4, 15, 16, 18], where k is the
6
A cluster is either the four corners of a non-degenerate axis-aligned rectangle, or the two
endpoints of a horizontal/vertical segment, or a single point.
Searching edges in the overlap of two plane graphs 477

total number of the intersections. The space requirement of each of these algorithms is
O(n). The algorithm by Mantler and Snoeyink [16] processes the red-blue intersections
in batches (called bundle-bundle intersections). In O(n log n) time and O(n) space it
can implicitly discover all the red-blue intersections, without reporting every one of
them individually. The latter feature is useful for our technique, therefore we invoke the
Mantler-Snoeyink algorithm in its rst phase. We summarize the algorithm below.
To describe the algorithm, we need to dene the following key notions: the witness
of a (bichromatic) segment intersection, a pseudoline at time i, and a (monochromatic)
bundle of segments at time i.

r
p

Fig. 1. Two line segments, r and b, the closed right wedge formed by them, and the witness p of
their intersection

Given a red segment r that intersects a blue segment b, the witness of their intersec-
tion is the leftmost of the endpoints of segments in S that are contained in the closed
right wedge formed by r and b. The closed right wedge formed by r and b is the in-
tersection of two closed right halfplanes: the one bounded by the line through r, and
the one bounded by the line through b, see the shaded area in Figure 1. Note that the
witness always exists: it may be an endpoint of a segment different from r or b (as in
Figure 1), or it may be an endpoint of either r or b.
Let n be the total number of distinct endpoints of the segments in R and B. Let
p1 , p2 , . . . , pn denote the sequence of these endpoints in the order of increasing x co-
ordinate. The basis of the Mantler-Snoeyink algorithm can be formulated as follows.

Lemma 1. For each i, 1 i n there is a y-monotone curve i that passes through

point pi , and subdivides the plane into two open regions (the left and the right one), such
that all the points pj , j < i, and all the red-blue intersections witnessed by the points
pj , j i are contained in the left region, and all the points pk , k > i together with the
intersections witnessed by them are contained in the right region, and i intersects each
segment in R or in B at most once.

We call such curve i a pseudoline at time i. Figure 2 shows a pseudoline at time 7,

i.e., 7 , in dashed black lines. Note that 7 cannot be replaced by a vertical straight line,
because it must pass through the point 7, and to the left of the intersection point between
the segments r4 and b4 , and the latter point lies to the left of the former one.
A blue bundle at time i is a maximal contiguous sequence of blue segments that
intersect the pseudoline i .7 See Figure 2, right. A red bundle is dened analogously.
7
This denition can be seen as a generalization of the one of single-edge bundles in Mount [17].
478 J. Iacono et al.

The algorithm can be seen as a topological sweep with a pseudoline, where the
only events are the endpoints p1 , . . . , pn of the segments in R and B. The sweepline
at each moment i is a pseudoline i such as dened in Lemma 1. The sweepline sta-
tus structure maintains all the red and blue bundles that intersect the current sweepline.
The sweepline status consists of (1) a balanced binary tree for each bundle, supporting
insertion, deletion of segments, and a query for the topmost and the bottommost seg-
ment in the bundle, (2) a doubly-linked list for all the bundles intersecting the sweepline
(bundles alternate colors), supporting insertion, deletion of the bundles, and sequential
search, and (3) two balanced binary trees (one per color) storing all the red and blue
bundles in order, and supporting splitting and merging of bundles.
At the event point pi the algorithm processes the intersections witnessed by pi ,
updates the sweepline from i1 to i , and makes the necessary changes to bundles (i.e.,
splits or merges them). By proceeding this way, the algorithm maintains the invariant
that all the red-blue intersections whose witness is to the left of the current event point
pi are already encountered. We summarize the result in the following.

Theorem 1 ([16]). The Mantler-Snoeyink algorithm runs in O(n log n) time, requires
O(n) space, and encounters O(n) bundle-bundle intersections in total.

2.2 Building the life table

In this section we describe our algorithm to build the life table for the sets R and B.
Figure 2 illustrates the execution of the algorithm for a simple example.
Before we start our description, recall [10] that every pointer-based data structure
with constant in-degree can be transformed into a partially persistent one. Such a per-
sistent data structure allows accessing in constant time the data structure at any moment
in the past, and performing pointer operations on it (but not modifying it); the total time
and space required to make a data structure partially persistent is linear in the number
of structural changes it underwent.
To build the life table for R and B, we rst perform the Mantler-Snoeyink plane
sweep algorithm (see Section 2.1), making the sweepline status structure partially per-
sistent. This ensures that each blue bundle that has appeared during the algorithm, can
afterwards be retrieved from the version of the sweepline status at the corresponding
moment in the past. In particular, we are interested in the blue bundles that intersect
red bundles. We assign each such blue bundle Bi a timestamp ti reecting the mo-
ment when the rst bundle-bundle intersection involving Bi was witnessed. In order to
distinguish between two different bundle-bundle intersections discovered at the same
moment tk (i.e., witnessed by the same point), we assign the moment tk +  to the in-
tersection that has smaller y coordinate. Figure 2, right, lists all such blue bundles for
the given example.
Observe that the plane sweep algorithm induces a partial order among the red seg-
ments: At any moment, the red segments crossed by the sweepline can be ordered from
bottom to top. Since the red segments are pairwise non-intersecting, no two segments
may swap their relative position. Let r1 , . . . , rn be a total order consistent with the par-
tial order along the sweepline at each moment. In Figure 2, the red segments are named
according to such an order.
Searching edges in the overlap of two plane graphs 479

We now build the life table of red segments and blue bundles, see Figure 2, bottom.
The life table is a graph dened as follows. On its y axis it has integers from 0 to nR ,
where nR is the number of red segments; the x axis of the life table coincides with the
x axis of the original setting, i.e., of the plane R2 . Each red segment ri is represented
by a horizontal line segment whose y coordinate equals i and whose endpoints x co-
ordinates coincide with the x coordinates of the endpoints of ri . Each blue bundle Bj ,
that has participated in at least one bundle-bundle intersection, is retrieved from the ver-
sion of the sweepline status at the moment tj when the rst such intersection has been
witnessed; tj is the timestamp of Bj . In the table, Bj is represented by a vertical line
segment (that could possibly be a point), whose x coordinate is tj . This vertical segment
intersects exactly the segments representing all red segments intersected by bundle Bj
(i.e., the segments of the red bundle(s) participating in the bundle-bundle intersection(s)
with Bj ). In particular, the bottom and the top endpoints of this segment lie respectively
on the two red segments that represent the rst and the last segment in R intersected by
bundle Bj , according to the topological ordering of the red segments. If Bj intersects
only one red segment, then in the life table Bj is represented by a point. In Figure 2 all
the blue bundles except B5 are represented by a point, but in a more complicated exam-
ple many bundles might be represented by line segments. Note that instead of storing
the segment list of each blue bundle explicitly, we just maintain a pointer to that bundle
as it appears in (the corresponding version of) the sweepline status structure.

5 13
3 7 r5 16
b3
r4 b4 12 18
b5 r2 15
bi a segment in B
b2 6 b8
10
b6 a segment in R
11 14 ri
4
b9
8 r1
b1 b7 19 B1 = ({b1 }, t4 )
r3 B2 = ({b1 }, t5 )
9
1 B3 = ({b3 }, t6 )
17 B4 = ({b4 , b5 , b6 }, t10 )
2 7
B5 = ({b7 }, t10 + )
B6 = ({b4 }, t13 )
5 B7 = ({b5 , b6 }, t15 )
4 B8 = ({b5 , b6 }, t16 )
3
2 B9 = ({b8 }, t17 )
1 B10 = ({b8 , b9 }, t17 + )
t4 t6 t10 t13 t16 t17 + 
t5 t10 +  t17

Fig. 2. Execution of the algorithm from Section 2.2 for R = {r1 , . . . , r5 } and B =
{b1 , . . . , b10 }. Above: the segments in R and B (solid lines, see the legend on the right); the
events of the plane sweep in order (gray numbers), and the pseudoline
7 at time 7 (dashed line).
Right: the blue bundles {B1 , . . . , B10 } encountered by the algorithm. Below: the life table for R
and {B1 , . . . , B10 }.
480 J. Iacono et al.

2.3 The resulting data structure

After the life table is built, we sweep it with a horizontal straight line from bottom to
top, again making the sweepline status partially persistent. The events now correspond
to red segments, and the version of the sweepline status at a time moment i contains all
blue bundles crossing the horizontal line y = i, sorted by x coordinate and stored in a
balanced binary tree.
Our ultimate data structure is the persistent sweepline status of the above (second)
plane sweep. We are able, in O(1) time, to retrieve the version of the sweepline status
structure at any moment i. The sweepline status at the moment i is a tree storing the
blue bundles whose beginning was witnessed before the moment i and whose end was
witnessed after that moment (in other words, those blue bundles that intersect the hor-
izontal line y = i in the life table). See Figure 2. Since each single bundle is stored in
a balanced binary tree, the tree of bundles is also a balanced binary tree. Therefore it
has height O(log n). Moreover, it can be accessed in the same way as a standard binary
tree: any navigation query in it (see Problem 1) can be performed in O(1) time.
Now suppose we wish to perform a search on (a portion of) a red segment ri among
the blue segments that cross it. We are required to be able to quickly determine where
the interesting intersection(s) lie with respect to p, for any point p of intersection be-
tween ri and a blue segment. The search in our data structure is then done as follows.
We retrieve the version of the sweepline status (of the second plane sweep) at the mo-
ment i. This sweepline status is an implicit balanced binary tree, as explained above. We
locate the endpoints of ri in that tree. Then we search in the portion of the tree between
ri s endpoints. The decisions during the search are made based on our knowledge about
the interesting intersections.
We conclude with the following.
Theorem 2. Given a pair R, B of plane straight-line graphs with n edges and vertices
in total in both graphs, the RB-Preprocessing problem for R and B can be solved in
O(n log n) time and O(n) space, such that the resulting data structure answers each of
the navigation queries in O(1) time.
Proof. The rst phase of our procedure to build the data structure is an execution of
the Mantler-Snoeyink algorithm, with the only difference that the bundle trees from
the sweepline status are made partially persistent. The latter can be performed with
amortized O(1) time and space overhead per update step and a worst-case time cost of
O(1) per access step [10]. The total number of updates made to the sweepline status
during the course of the Mantler-Snoeyink algorithm is O(n) [16]. Thus, after the rst
phase is completed, we have the order of red segments and the persistent sweepline
status. With this information, the life table can be built in O(n) time and space: we ll
the table with the horizontal red segments, and we access sequentially all the versions
of the sweepline, retrieving the blue bundles and the information on their intersections
with the red bundles, and drawing the vertical segments of the life table. Sweeping the
life table with a horizontal line, and making the sweepline status partially persistent
again costs O(n log n) time and O(n) space.
For an edge e = ri of R, the version of the sweepline status structure at time i
provides a balanced binary search tree Te , required by the RB-Preprocessing problem,
Searching edges in the overlap of two plane graphs 481

such that Te can be navigated (but not modied) in the same way and with the same time
complexity as the standard balanced binary search tree. Hence the navigation queries
of the RB-Preprocessing problem can be answered in O(1) each.

3 Applications

We proceed with more detail on the applications of our technique, which are listed in
Section 1. Due to space constraints, we give here only the following results: red-blue
batched search, the vertical distance problem, and a short version of the construction of
the Hausdorff Voronoi diagram. The omitted parts can be found in the full version of
this paper [13].

3.1 The red-blue batched search problem

Consider two sets, R (red) and B (blue), of line segments in R2 , such that the segments
in each set are pairwise interior-disjoint, and suppose that some of the red-blue intersec-
tions are interesting, and there is at most one interesting red-blue intersection per each
segment. Let O be an oracle that, given an intersection point p of a red segment r and a
blue segment b, determines to which side of p the interesting intersection on r lies.

Problem 2 (Red-blue batched search problem [9]). Given sets R, B and oracle O, nd
all interesting intersections between the segments in R and the ones in B.

Dehne et al. [9] showed how to solve the red-blue batched search problem by using
an augmentation of the hereditary segment tree data structure of Chazelle et al. [4].
Their solution requires O(n log2 n) space and O(n log3 n) queries to the oracle.
Our technique presented in Section 2.2 can be directly applied to solve the red-
blue batched search problem with better time and space: We preprocess the sets R and
B; after that for each red segment r we perform a binary search in the (implicit) tree
storing the red-blue intersections along r. The search is guided by the oracle O, and
thus it requires O(log n) queries to the oracle. Since the number of red segments is
O(n), the total number of queries to the oracle required for searching all red edges is
O(n log n). Theorem 2 implies the following.

Theorem 3. The red-blue batched search problem for the sets R, B and the oracle O
can be solved using O(n) space and O(n log n) queries to the oracle, where n is the
total number of segments in R and B.

3.2 Vertical distance for a pair of 3D polyhedral terrains, one of which is convex

Let R and B be two polyhedral terrains of complexity nR and nB respectively, where

terrain B is convex (that is, B is the upper envelope of a set of nB planes in R3 ). Let
n = nR + nB . We wish to determine the maximum vertical distance between R and B,
i.e., the length of the longest vertical line segment connecting a point in R and a point
482 J. Iacono et al.

in B.8 As an illustration, the reader may imagine a (convex) approximate model of a

mountain, and a need to compare it with the real mountain (of course, not necessarily
convex) in order to estimate the quality of the approximation.
Since both surfaces R and B are composed of planar patches, the vertical distance
between R and B is the vertical distance between these patches. The maximum vertical
distance between R and B is thus attained either between a vertex of R and a facet of B
(or vice versa), at innity along an unbounded edge of one of the surfaces, or between
an edge of R and an edge of B. In the second case the distance between R and B is +.
Both the rst and the second case can be easily processed by point location queries of
a (possibly innite) vertex of one surface into the other one, which requires O(n log n)
time in total. To deal with the last case one can preprocess the vertical projections of
R and B following our technique, and perform the binary searches along each edge e
of R for the intersection with an edge of B maximizing the vertical distance. Consider
the cross-section of B by the vertical plane containing e. This is a convex monotone
polygonal line. The sequence hB of heights of its breakpoints is unimodal, and since
all points of e lie on the same line, if we subtract from each member of hB the height
of the point in e lying on the same vertical line, the resulting sequence will still have
one maximum. It then follows that given a point p e vertically above/below an edge
of B, in constant time we can nd out in which direction this maximum lies, and this is
exactly what the oracle for the binary search along e should do. Using Theorem 2, we
conclude.
Theorem 4. Given a pair of polyhedral terrains in 3D, where one of the terrains is
convex, the maximum vertical distance between the terrains can be found in O(n log n)
time and O(n) space, where n is the total complexity of both terrains.
Notice that by slightly changing the algorithm, we could be answering the minimum
vertical distance, instead of the maximum one. In particular this gives an alternative
O(n log n) algorithm to solve the shortest watchtower problem [21, 22].

3.3 Construction of the Hausdorff Voronoi diagram

Given a set of n distinct points in the plane, we partition this set, resulting in a family
S of point clusters, where no two clusters share a point. Let the distance from a point
t R2 to a cluster P S, denoted as df (t, P ), be the maximum Euclidean distance
from t to any point in P . The Hausdorff Voronoi diagram of S, denoted as HVD(S), is
a subdivision of R2 into maximal regions such that every point within one region has
the same nearest cluster according to distance df (, ).
The diagram has worst-case combinatorial complexity (n2 ), and it can be con-
structed in optimal O(n2 ) time [11]. However, these bounds can be rened according to
certain parameters of the family S. Two clusters are called non-crossing if their convex
hulls intersect at most twice (i.e., their convex hulls are pseudocircles), and crossing
otherwise. Below we consider separately the (simpler) case of non-crossing clusters,
and the one of crossing clusters. The latter case subsumes the former one, and is given
in detail in the full version [13].
8
It may happen, that the distance keeps increasing as we move along some direction towards
innity. Then we say that the maximum vertical distance between R and B is +.
Searching edges in the overlap of two plane graphs 483

Non-crossing clusters. If all clusters in S are pairwise non-crossing, the complexity of

HVD(S) is O(n). In this case the diagram can be constructed in expected O(n log2 n)
time and expected O(n) space [5, 14]. The best deterministic algorithm to date requires
O(n log5 n) time and O(n log2 n) space [9].9 The latter algorithm follows the divide-
and-conquer strategy. To merge two recursively computed diagrams, a bottleneck pro-
cedure is formulated as a red-blue batched segment search problem (see Section 3.1),
where the two segment sets are the sets of edges of the two diagrams. The latter prob-
lem is then solved in O(n log2 n) space and O(n log3 n) queries to the oracle. The
authors dene an oracle to perform this search, which they assume can be implemented
in O(1) time. We were unable to reconstruct the claimed constant-time oracle, however
we know how to implement it in O(log n) time per query. Theorem 3 implies an algo-
rithm to construct the Hausdorff Voronoi diagram of a family of non-crossing clusters
in O(n log3 n) time and O(n) space.10 This result is subsumed by the one for arbitrary
clusters, which is strictly more general (for a family of non-crossing clusters m = 0).

Arbitrary clusters. Consider the Hausdorff Voronoi diagram of a family S of arbitrary

clusters, that could possibly cross. The essential parameter used to rene the quadratic
bounds related to the diagram in that case, is the number of crossings,11 denoted by m.
The parameter m is bounded from above by half the number of intersections between
the convex hulls of all pairs of crossing clusters. In the worst case m = (n2 ), however
it is small in known practical applications, e.g., in VLSI CAD [19, 20]. The combina-
torial complexity of the Hausdorff Voronoi diagram is shown to be O(n + m) [19].
Apart from the O(n2 ) time algorithm mentioned above, there is a plane sweep [19] and
a divide-and-conquer [20] algorithm to construct the diagram. Both of them are sensi-
tive to the parameter m, however their time complexity depends as well on some other
parameters, which are unrelated to m. In particular, these algorithms may have (n2 )
time complexity even when clusters are non-crossing.
Our technique can be applied to reduce the time complexity of the divide-and-
conquer construction of the Hausdorff Voronoi diagram of arbitrary clusters [20]. The
resulting algorithm is the fastest to date deterministic algorithm for certain input fam-
ilies, where clusters may cross, but the number of crossings is small. See the detailed
description of our algorithm in [13].

Acknowledgments. Research partially completed while J. I. was on sabbatical at the Algorithms

Research Group of the Computer Science Dept. at ULB with support from a Fulbright Research
Fellowship, F.R.S.-FNRS., and NSF grants CNS-1229185, CCF-1319648, and CCF-1533564. E.
K. was supported by F.R.S.-FNRS, and by the SNF grant P2TIP2-168563 of the Early PostDoc
Mobility program. S. L. is directeur de recherches du F.R.S.-FNRS.

9
The time complexity claimed in Dehne et al. is O(n log4 n), however we believe that in reality
the described algorithm requires O(n log5 n) time.
10
Note that if it was possible to implement the oracle in O(1) time, our algorithm would instantly
be improved by a O(log n) factor. Thus in all cases our algorithm is faster than the previous
one by a factor of O(log2 n).
11
See [19, 20] for the formal denition of the number of crossings.
484 J. Iacono et al.

References
1. Abellanas, M., Hurtado, F., Icking, C., Klein, R., Langetepe, E., Ma, L., Palop, B., Sacristan,
V.: The farthest color Voronoi diagram and related problems. In: 17th Eur. Workshop on
Comput. Geom. pp. 113116 (2001), Tech. Rep. 002 2006, Univ. Bonn
2. Aronov, B., Bose, P., Demaine, E.D., Gudmundsson, J., Iacono, J., Langerman, S., Smid,
M.: Data structures for halfplane proximity queries and incremental voronoi diagrams. In:
7th Latin American Symp. on Theoretical Informatics (LATIN). pp. 8092. Springer (2006)
3. Chan, T.M.: A simple trapezoid sweep algorithm for reporting red/blue segment intersec-
tions. In: CCCG. pp. 263268 (1994)
4. Chazelle, B., Edelsbrunner, H., Guibas, L.J., Sharir, M.: Algorithms for bichromatic line-
segment problems and polyhedral terrains. Algorithmica 11(2), 116132 (1994)
5. Cheilaris, P., Khramtcova, E., Langerman, S., Papadopoulou, E.: A randomized incremental
algorithm for the Hausdorff Voronoi diagram of non-crossing clusters. Algorithmica 76(4),
935960 (2016)
6. Cheong, O., Everett, H., Glisse, M., Gudmundsson, J., Hornus, S., Lazard, S., Lee, M., Na,
H.S.: Farthest-polygon voronoi diagrams. Comput. Geom. 44(4), 234247 (2011)
7. Claverol, M., Khramtcova, E., Papadopoulou, E., Saumell, M., Seara, C.: Stabbing circles
for some sets of Delaunay segments. In: 32th Eur. Workshop on Comput. Geom. (EuroCG).
pp. 139143 (2016)
8. Claverol, M., Khramtcova, E., Papadopoulou, E., Saumell, M., Seara, C.: Stabbing circles
for sets of segments in the plane. Algorithmica (2017), DOI 10.1007/s00453-017-0299-z
9. Dehne, F., Maheshwari, A., Taylor, R.: A coarse grained parallel algorithm for Hausdorff
Voronoi diagrams. In: Int. Conf. on Parallel Processing (ICPP). pp. 497504. IEEE (2006)
10. Driscoll, J.R., Sarnak, N., Sleator, D.D., Tarjan, R.E.: Making data structures persistent. In:
18th annual ACM Symp. on Theory of Computing. pp. 109121. ACM (1986)
11. Edelsbrunner, H., Guibas, L.J., Sharir, M.: The upper envelope of piecewise linear functions:
algorithms and applications. Discr. & Comput. Geom. 4(4), 311336 (1989)
12. Guibas, L.J., Sharir, M., Sifrony, S.: On the general motion-planning problem with two de-
grees of freedom. Discr. & Comput. Geom. 4(5), 491521 (1989)
13. Iacono, J., Khramtcova, E., Langerman, S.: Searching edges in the overlap of two plane
graphs. ArXiv e-prints (2017), arXiv:1701.02229
14. Khramtcova, E., Papadopoulou, E.: Randomized incremental construction for the Hausdorff
Voronoi diagram of point clusters. ArXiv e-prints (2016), arXiv:1612.01335
15. Mairson, H.G., Stol, J.: Reporting and counting intersections between two sets of line seg-
ments. In: Theoretical Foundations of Computer Graphics and CAD, pp. 307325 (1988)
16. Mantler, A., Snoeyink, J.: Intersecting red and blue line segments in optimal time and pre-
cision. In: Discr. and Comput. Geom.: Japanese Conference, JCDCG 2000, Revised Papers.
pp. 244251 (2001)
17. Mount, D.M.: Storing the subdivision of a polyhedral surface. Discr. & Comput. Geom. 2(2),
153174 (1987)
18. Palazzi, L., Snoeyink, J.: Counting and reporting red/blue segment intersections. CVGIP:
Graphical Models and Image Processing 56(4), 304310 (1994)
19. Papadopoulou, E.: The Hausdorff Voronoi diagram of point clusters in the plane. Algorith-
mica 40(2), 6382 (2004)
20. Papadopoulou, E., Lee, D.T.: The Hausdorff Voronoi diagram of polygonal objects: A divide
and conquer approach. Int. J. of Comput. Geom. & Appl. 14(06), 421452 (2004)
21. Sharir, M.: The shortest watchtower and related problems for polyhedral terrains. Inf. Pro-
cess. Lett. 29(5), 265270 (1988)
22. Zhu, B.: Computing the shortest watchtower of a polyhedral terrain in O(n log n) time.
Comput. Geom. 8(4), 181193 (1997)
 Improved Average Complexity for
Comparison-Based Sorting

Kazuo Iwama1 and Junichi Teruyama2,3

1
Kyoto University, [email protected]
2
National Institute of Informatics,
3
JST, ERATO, Kawarabayashi Large Graph Project, [email protected]

Abstract. This paper studies the average complexity on the number

of comparisons for sorting algorithms. Its information-theoretic lower
bound is n lg n 1.4427n + O(log n). For many ecient algorithms, the
rst n lg n term is easy to achieve and our focus is on the (negative)
constant factor of the linear term. The current best value is 1.3999 for
the MergeInsertion sort. Our new value is 1.4106, narrowing the gap
by some 25%. An important building block of our algorithm is two-
element insertion, which inserts two numbers A and B, A < B, into a
sorted sequence T . This insertion algorithm is still suciently simple for
rigorous mathematical analysis and works well for a certain range of the
length of T for which the simple binary insertion does not, thus allowing
us to take a complementary approach with the binary insertion.

1 Introduction

A majority of existing sorting algorithms, including Bubble sort, Quick sort,

Heap sort, Merge sort and Insertion sort, are so-called comparison-based sorts,
in which our basic operation is a comparison of two input numbers. The com-
plexity in terms of this measure, the number of comparisons needed to obtain
a sorted sequence, is an obvious lower bound of the running time of the algo-
rithm. Thus it has been a popular research topic in TCS to investigate its upper
and lower bounds for several sorting algorithms. Note that any sorting algo-
rithm for n elements can be described as a binary decision tree having n! leaves
corresponding to all dierent permutations of the n elements. The number of
comparisons to obtain one of them is the number of nodes on the path from the
root to the leaf corresponding to the sequence. Therefore we have an obvious
lower bound, called an information-theoretic lower bound. Namely, any sorting
algorithm needs
lg n! n lg n 1.4427n + O(log n).
comparisons in the worst case.
Usually it is not very hard to obtain an upper bound of n lg n. For instance,
consider the BinaryInsertion sort that increases the length of the sorted sequence
one by one using binary insertion. Obviously we have n 1 steps and each of
them consists of at most lg n comparisons (and much less for most of the steps).

Springer International Publishing AG 2017 485

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 485496, 2017.
DOI: 10.1007/978-3-319-62127-2_41
486 K. Iwama and J. Teruyama

Thus our interest naturally comes to the constant factor for the linear term in
n. Unfortunately, however, its analysis does not seem so easy and our knowledge
is quite limited. For instance, it is at most 0.91 for Merge (and similar other)
sort [6, 13] and the current best one is 1.32 for MergeInsertion sort obtained
by Ford and Johnson more than ve decades ago [3].
Our interest in this paper is the average-case complexity on the number of
comparisons, which should be easier to obtain than the worst-case complexity.
In fact we do have a number of better results; 1.26 for Merge sort [6], 1.38
for BinaryInsertion sort, and most recently 1.3999 for MergeInsertion sort [2].
Notice that 1.3999 is some 96.98% of 1.4427, but there still exists a gap and
seeking the exact bound for this fundamental problem should be an important
research goal.
Our Contribution We achieve 1.4034 by a new algorithm (1,2)Insertion.
Furthermore it is shown that the constant is improved to 1.4106 by combining
the new algorithm with the MergeInsertion sort. Thus we have narrowed the
previous gap between 1.3999 and 1.4427 by some 25%. Our new algorithm is
based on binary insertion. Recall that the BinaryInsertion sort repeats a binary
insertion of a new item into a sorted sequence of length i 1 for i = 2 to n.
Here the performance of binary insertion itself is optimal because it constitutes
an optimal decision tree of height lg i . However, if i is not a power of two,
this tree is not completely balanced, i.e., there is a dierence of one comparison
due to the position of the inserted element. This small dierence in each step
accumulates during the repeated steps and nally creates a large imbalance. This
is the reason for its relatively poor performance.
Our idea is to use a binary insertion if i is close to a power of two and to use
what we call a two-element merge, or 2Merge otherwise. 2Merge merges
a two-element sequence (A, B), A < B, with a sorted sequence T of length
i 2 to obtain a sorted sequence of length i. We rst insert A using a kind of
binary search, meaning A is compared with an element in T whose position is
approximately 1/3 from the smallest. If A falls into the rst third of T , then we
use a standard (with a bit of care) binary search, called right-heavy binary search
or RHBS. The key thing here is that the original bad i changes to a good
i in this binary insertion. If A falls into the right part of T , we simply recurse.
Then, we insert B into T by using a standard binary search. Thus we can reduce
the imbalance of each step of insertion, which contributes to the better bound
for the whole sorting.
Due to [2], the performance of MergeInsertion diers a lot for dierent n
and it hits a best peak when n is about one third from the previous power-
of-two number, which achieves around 1.415. This is much better than our
(1,2)Insertion (but, unfortunately, it quickly gets worse as n leaves the best
position and ends up with 1.3999 for a roughly power-of-two n). Thus here is a
natural idea: For a given sequence X of length n that is bad for MergeInsertion,
select the largest value n that is less than n and is good for MergeInsertion.
Then we use MergeInsertion sort for a length n subsequence of X and insert
the remaining elements using (1,2)Insertion, which in fact gives us 1.4106.
Improved Average Complexity for Comparison-Based Sorting 487

Related Work The idea of inserting two elements into a sorted sequence is
not new. [4] and [14] claimed two exactly optimal algorithms for such a merge op-
eration in terms of the worst-case complexity [4] and in terms of the average-case
complexity [14]. Unfortunately, both algorithms are a bit involved and their per-
formance analysis did not give closed formulas for the complexity. Our 2Merge
is probably not exactly optimal, but is suciently simple for rigorous mathe-
matical analysis.
The analysis of the BinaryInsertion sort by Edelkamp and Wei [2] gives
many hints to our new analysis. They show that the average number of compar-
isons is

2lg i
lg i + B(i), where B(i) = 1 (1)
i

for a single insertion and is

 n  
1 + ln(4pn )
(lg i + B(i)) = n lg n + 1 lg pn (2)
i=1
pn
< n lg n 1.386n
for the entire BinaryInsertion sort, where pn = 2lgn n is a parameter indicating
the deviation from a power of two. Edelkamp and Wei [2] also includes a nice
survey on this topic.
Although we have few results on the worst-case complexities for asymptot-
ically large n, we do have a rather rich literature for small ns. For instance,
the information-theoretic bound (actually its ceiling) cannot be achieved by any
comparison-based sorting for 12 n 15. The MergeInsertion sort achieves a
matching upper bound for 1 n 15, but n = 16 is still open, namely there
is a gap of one between the lower and upper bounds (45 and 46, resp.) for the
exact number of necessary comparisons. It is also known that MergeInsertion is
not optimal for some ns, for instance, for n = 47. See [1, 3, 712, 15] for these
results.
Notations and Assumptions. Our sorting algorithm takes a sequence of
all dierent n elements as input. An average complexity (or simply complexity)
of a sorting algorithm Alg is the expected number of comparisons Alg executes
to sort each of n! dierent sequences of length n. Note that the complexity of
all sorting algorithms in this paper is written as n lg n + cn + O(log n) for some
negative constant c. It is important to mention that the value of c, that is our
main issue, periodically changes depending on n usually and we are interested
in its worst (largest) value for asymptotically large n, unless otherwise stated.
We exploit the O(log n) term to make analysis simpler. In particular we assume,
without loss of generality, that n is always even throughout this paper. Also,
when summing up a cost function f (i) for i = 1 to n, an O(1/i) term in f is
not important. For notation, we write x = y z if |x y| < z, where z may be
a big-O notation like x = y O(z). We may denote a sequence of one element
(s1 ) by simply s1 .
488 K. Iwama and J. Teruyama

2 Our Algorithm and Its Analysis

See Algorithms 1, 2 and 3. The main algorithm is Algorithm 1. Note that

Algorithm 2 is improved in the next section and Algorithm 1 is combined with
MergeInsertion in Section 4. For a given sequence S = (s1 , s2 , . . . , sn1 , sn ) with
an even n, (1,2)Insertion works in Round 0, Round 2, ... up to Round n 2. In
Round 0, s1 and s2 are sorted by a single comparison to make a sorted sequence
T0 of length two. In Round i, si+1 and si+2 are inserted into Ti2 obtained in
Round i 2 by using (i) a single call of 2Merge or (ii) two calls of RHBS,
depending on the value i. Recall that we wish to obtain the average complexity
for all dierent n! sequences, in other words, we wish to obtain the expected
number of comparisons assuming that each S appears uniformly at random. It
then turns out that we can also assume that the position of si+1 (and that of
si+2 also) in each round is uniformly at random in the dierent i + 1 positions
of Ti2 that includes i elements. Thus the overall average complexity is a simple
sum of the average complexity of each round.
We rst make an analysis of 2Merge. Note that 2Merge uses RHBS which
stands for Right-Heavy Binary Search. Note that the number, say q, of compar-
isons to insert A into a sequence T = (t1 , . . . ti ) is q0 = lg(i + 1) 1 or q0 + 1
if we use the standard binary search. The feature of RHBS is that if q = q0 + 1
Algorithm 1 (1,2)Insertion(S)
Input: A (unsorted) sequence S = (s1 , s2 , ..., sn ), where n is even.
Output: Sorted sequence
Step 1: If n = 2, then sort (s1 , s2 ) with a single comparison.
Step 2: Sort S  = (s1 , ..., sn2 ) by (1,2)Insertion to obtain T  .
Step 3: If pn [0.5511, 0.888] then insert sn1 and sn into T  by calling
2Merge(sn1 , sn , T  ). Otherwise insert sn1 into T  by RHBS and then sn by RHBS.
Algorithm 2 2Merge(A, B, T )
Input: A and B are numbers and T = (t1 , t2 , ..., ti2 ) is a sorted sequence such that
i is even and i 4.
Output: Sorted sequence of length i.
Step 1. Compare A and B and swap them if A > B.
Step 2. Let (r) = 1 2r/2 . For r = 1, 2, . . . , up to 2 lg i, compare A with t(r)i and
go to Step 3 if A < t(r)i .
Step 3. Insert A to (t(r1)i + 1, . . . , t(r)i 1) using RHBS. Suppose that A falls
between t and t+1 .
Step 4. Insert B to (t+1 , . . . , ti2 ) using RHBS.
Algorithm 3 RHBS(A, T )
Input: A is a number and T = (t1 , t2 , ..., ti ) is a sorted sequence.
Output: Sorted sequence of length i + 1.
Step 1. If i 3 2lg(i+1)2 1, then let set d := 2lg(i+1)2 . Otherwise, let set
d := i 2lg(i+1)1 + 1.
Step 2. Let T1 = (t1 , . . . , td1 ) and T2 = (td+1 , . . . , ti ).
Step 3. Compare A with td . If A < td , return RHBS(A, T1 ) td T2 . Otherwise, return
T1 td RHBS(A, T2 ).
Improved Average Complexity for Comparison-Based Sorting 489

for some A, then q = q0 + 1 for any A such that A > A, in other words, the
number of comparisons is monotone. This is easily realized by selecting td (to
be compared with A) in each recursion phase such that either the number of
T s elements that is smaller than td or the number of those that is larger than
td be (a power of two)1. Suppose for instance 8 i 15. Then if i is 11 or
less, then the rst comparison is with t4 and if i is 12 or more, then the rst
comparison is with ti7 . There would be no merit of this structure if the position
of A is uniformly distributed. However, if small As are more likely than large
As, there is an obvious advantage and that provides a real merit in 2Merge.
Notice that even if our improvement in each step is a small constant, something
like 0.1, that constant signicantly aect the value of our constant factor of the
linear term.
In Step 2, we determine the range of the smaller element A. If the condition
there (A < t(r)i ) is met for r = 1, then the range is (t1 , . . . , t(11/2)i 1 ),

where (1 1/ 2) 0.2929. In general, the range is (t(r1)i +1 , . . . , t(r)i 1 )
for an integer r 1, and we wish to compute the average complexity of Step
3, i.e., the average number of comparisons to insert A into this range. Here
we
have two technical issues: (i) We introduce a parameter wr and let wr :=
( 2 1)2r/2 i. Note that wr is somehow related to the size of the above range
but it may not be integral. The idea is that the complexity does not dier
signicantly if the size of the range diers by a small constant and approximating
the size by wr makes our job much easier. (ii) Although the positions of A and
B are uniformly at random, we now know that A < B. Therefore the probability

that A falls between t1 and t under the condition that A < B is (i )/ 2i .
We also extend the denition of px = 2lgx x for a real value x.
Lemma 1. Suppose that A is to be inserted to (t(r1)i +1 , . . . , t(r)i 1 ) for
an even i. Then 2Merge requires
 r/2 
10 6 2 3 2 2 2
A(r) = lg wr + 7 4 2 + O (3)
pr p2r i
comparisons on average at Step 3. Furthermore, the expected value of A(r) is
Pr[r = 1]A(1) + Pr[r = 2]A(2) + = lg i + T(i),
where

1 1
23 pi (1/2, 1+4 2 ],
1 1 6p

i
16p2i
2
T(i) = 5 4 2 + 2 + 3pi 1
pi ( 1+4 2 , 2+4 2 ], (4)
pi 6pi

3
4 + 1
+ 13 pi ( 2+4 2 , 1].
3pi 4p2i

See Section 2.1 for the proof. Now we are going to Step 4 to insert B and
here is our analysis (see the full version of this paper [5] for its proof).
Lemma 2. For an even i, 2Merge requires
2 1
lg(i 1) + 1 + 2 + O(1/i)
pi 3pi
comparisons on average at Step 4.
490 K. Iwama and J. Teruyama

The entire complexity of 2Merge is the sum of these two quantities in Lem-
mas 1 and 2 and another two values; (+1) for comparing A and B at Step 1 and
the one for the expected number of comparisons in Step 2 that is 2 O(1/i) (the
precise analysis is shown in [5]). Thus the complexity of 2Merge is
lg i + lg(i 1) + U(i) + O(1/i)
where (T(i) is equation (4))
2 1
U(i) = 1 + T(i) + 2 . (5)
pi1 3pi1
Since this is the complexity for inserting two elements, the complexity for a
single insertion can be regarded as a half of it, or
lg i + U(i)/2 + O(1/i). (6)
It then turns out that by comparing this value with (1) of the BinaryInsertion,
2Merge is better than BinaryInsertion for 0.5511 < pi < 0.888. (Note that this
range is obtained by a numerical calculation.) Thus we use 2Merge for this
range of pi and RHBS for the other range. In summary our one step complexity
is
B(i) pi (1/2, 0.5511]
lg i + U(i)/2 + O(1/i) pi (0.5511, 0.888]

B(i) pi (0.888, 1]
By simple calculation, this is rewritten by
lg i + D(pi ) (7)
where


1 p1i pi (1/2, 0.5511] ,
 2

25
6 2 2
19 7
12pi + 32p2i pi 0.5511, 1+4 2 ,
 2
9+ 2
D(pi ) = 133 2 2 1
6pi + 4p2i pi 1+4 2 , 2+4 2 , (8)

14  2


3 2 2
13
6pi + 3
8p2i
pi 2+4 2 , 0.888 ,


1 p1i pi (0.888, 1] .
Now by using the trapezoidal rule, we have
34 4 5
n 1 pn
lg n
D(pi ) = 2 D(x)dx + D(x)dx + O(log n)
i=1 1/2 1/2

and the following theorem. We omit details of analyses, see [5].

Theorem 1. The complexity of (1,2)Insertion is at most n lg n 1.40118n.

2.1 Proof of Lemma 1

We rst prove6 formula (3). From7the assumption,

6 we call RHBS(A,
7 (t1 +1 , . . . , t2 1 )),
where 1 = (1 2(r1)/2 )i and 2 = (1 2r/2 )i . For an integer , let E
denote the event that A falls between t1 and t . Also F denotes the event that
Improved Average Complexity for Comparison-Based Sorting 491
2
A is inserted between t1 and t2 , namely F = = 1 +1
E . Let w = 2 1 and
z = 2i 1 2 1. Since Pr[E ] = i , we have
i
(2)
2
i wz Pr[E ] 2(i )
Pr[F ] =
i  =
i  and Pr[E | F ] = = .
2 2 2 Pr[F ] wz
= +1
1

lg w
Let k = 2 w. By its monotonicity, RHBS requires lg w 1 compar-
isons if t1 < A < t1 +k , and requires lg w comparisons otherwise. Therefore,
1 +k
the average number of comparisons is lg w = Pr[E | F ], we need
1 +k 1+1
1 +k 2(i)
to calculate the summation =1 +1 Pr[E | F ] = =1 +1 wz . Observing
that k/w = 1/pw 1, we have
1 +k

2(i ) k 2i 21 k 1
=
wz w z
=1 +1
k z+wk
= ( 2i 21 1 = w + z)
w  z 
k wk w
= 1+
w w z
 
1 3 1 w
= 1 + 2 + 2 (9)
pw pw pw z
6 7 6 7
Since 1 = (1 2(r1)/2 )i and 2 = (1 2r/2 )i . we have w = 2r/2 ( 2

1)i 1 and z = 2r/2 ( 2 + 1)i 1. Observe the value wz is close to 3 2 2, in
fact the dierence is bounded as

 w  4 2 2 2r/2
3 2 2  < < ( r 2 lg i) .
z z i
Therefore, because k/w = 1/pw 1 and 2 + p3w p12 14 , (9) continues as
w
   
1 3 1 2r/2
() = 1 + 2 + 2 32 2
pw pw pw i

10 6 2 3 2 2 2 r/2
= 7 + 4 2 +
pr p2r 4i
Thus, the average number of comparisons is

10 6 2 3 2 2 2r/2
lg w + 7 4 2 + . (10)
pw p2w 4i
As mentioned before the statement of the lemma, we wish to replace lg w
by lg wr , since there is no obvious way of treating the ceiling of the former
that includes another ceilings for w. Now, recall that wr = 2r/2 ( 2 1)i and
pr = 2lgwwr r  . We show that it is possible to simply replace lg w by lg wr
 r/2 
almost as it is: (i) If lg w = lg wr holds, then because | p1w p1r | = O 2 i ,
 r/2 
it is enough to replace the last (error) term with O 2 i . (ii) Otherwise suppose
that lg w = lg wr . Since |w wr | < 1 and w is an integer, w must be a power
492 K. Iwama and J. Teruyama

of two and lg wr must be lg w + 1. It then follows that pw = 1 and we can

write that 1/pr = 2 , where || = 2|wwr w|
r
< (2 2 + 2) 2r/2 /i. Substituting
pw = 1, (10) becomes
2r/2
lg w .
4i
Substituting lg wr = lg w + 1 and 1/pr = 2 , (3) becomes
 r/2   r/2 
2 2
lg w + (2 2 2) + (3 2 2)2 O = lg w O .
i i
Therefore (10) i.e., the value we want to obtain can be replaced by (3) with the
error term. Thus the former part of lemma is proved.
For formula (4) we need to give the average values of lg wr , 1/pr and 1/p2r .
Since lg i = lg i + lg pi and x = x for any value x, we have
   
lg wr = lg i + lg( 2 1) r/2 = lg i r/2 lg(pi ( 2 1)) .
Also, for any value x and integer m, we have

r/2 + m x [m, m + 1/2),
r/2 + x =
r/2 + m x [m + 1/2, m + 1).
Let  
cr (pi ) = r/2 lg(pi ( 2 1)) .

Then since lg(pi ( 2 1)) (2.5, 1), we have


r/2 + 2 pi (1/2,
1+ 2
4 ],

cr (pi ) := r/2 + 1 pi ( 1+4 2 , 2+4 2 ],


r/2 + 1 pi ( 2+4 2 , 1].
We have the following lemma about the expected values of r/2 and r/2 .

Lemma 3. E[r/2] = 2/3 O(1/i) and E[r/2 ] = 4/3 O(1/i).

This lemma implies

8/3 pi (1/2, 1+ 2
4 ],

] lg 1+ 2 2+ 2
E [lg w r = i O(1/i) 7/3 pi ( 4 , 4 ],


5/3 pi ( 2+4 2 , 1].
Similarly, we can obtain the expected value of 1/pr and 1/p2r as follows.

Lemma 4.
32+5 5(3+22)
1+ 2
p (1/2, 1+ 2
] p (1/2, 4 ]
12pi i 4 8 9 3+248p
i
2 i

3+2 2 1+ 2 2+ 2 2 2 1+ 2 2+ 2
E[1/pr ] = pi ( 4 , 4 ] , E 1/pr = 6p2i
pi ( 4 , 4 ] .
6pi
32+5

2+ 2
pi ( 4 , 1] 5(3+2 2)
pi ( 2+4 2 , 1]
6pi 12p2 i

Adding all those values, we can obtain (4) and the lemma is proved. The proofs
of Lemmas 3 and 4 are shown in [5].
Improved Average Complexity for Comparison-Based Sorting 493

3 Improvement of 2Merge

As mentioned before, the value of (r) is selected based on the observation

that (1) the probability that A falls in the left part of T should be close to 1/2
and (2) the length of the left part for r = 1 (which seems more important than
other less happening cases for r 2) should be close to a power of two. The
previous selection is perfect in terms of (1) but is not in terms of (2) since (r)
does not depend on the length i of T . In this section, we put a priority to (2) by
setting
1 1
1 2k1 + pi 2k+1 (r = 2k 1 and pi (3/4, 1]),
(r, pi ) = 1 21k pi 21k+2 (r = 2k 1 and pi (1/2, 3/4]), (11)

1 21k (r = 2k).
Note that it now depends on i and it turns out that if A falls into the left
part of T for r = 1, then the length of the left part is exactly a power of two
for any i when pi (3/4, 1]. Note that for even r, (r, pi ) is the same as the
previous (r).
We denote the modied 2Merge as 2Merge and the whole sorting algo-
rithm as (1,2)Insertion . (See Algorithm 4 and 5.) Our analysis, having two
cases for pi 3/4 and pi < 3/4, is more involved but we can obtain the average
number of comparisons for a single step is
31 3 25
2 4pi + 96p2i pi (1/2, 3/4),
lg i + B(i) O(1/i) +
1 2p3 i + 24p
13
2 pi [3/4, 1].
i

As with the previous section, comparing

: this value 2with (1), 2Merge is

better than the binary insertion for pi 34 126 , 34 + 123 . (Note that we did not
use numerical analysis this time.) Then, one step complexity of (1,2)Insertion
Algorithm 4 (1,2)Insertion*(S)
Input: A (unsorted) sequence S = (s1 , s2 , ..., sn ), where n is even.
Output: Sorted sequence
Step 1: If n = 2, then sort (s1 , s2 ) with a single comparison.
Step 2: Sort S  = (s1 , ..., sn2 ) by (1,2)Insertion*
 to obtain T  .

Step 3: If pn 3
4
6 3
,
12 4
+ 3
12
then insert sn1 and sn into T  by calling
2Merge(sn1 , sn , T  ). Otherwise insert sn1 into T  by RHBS and then sn by RHBS.
Algorithm 5 2Merge*(A, B, T )
Input: A and B are numbers and T = (t1 , t2 , ..., ti2 ) is a sorted sequence such that
i is even and i 4.
Output: Sorted sequence of length i.
Step 1. Compare A and B and swap them if A > B.
Step 2. Dene (r, pi ) as Equation (11). For r = 1, 2, . . . , up to 2 lg i, compare A with
t(r)i and go to Step 3 if A < t(r)i .
Step 3. Insert A to (t(r1)i + 1, . . . , t(r)i 1) using RHBS. Suppose that A falls
between t and t+1 .
Step 4. Insert B to (t+1 , . . . , ti2 ) using RHBS.
494 K. Iwama and J. Teruyama

22.5
Experiment
22 Analysis

Average Number of Comparisons

21.5

21

20.5

20

19.5

19

18.5

18
600 800 1000 1200 1400 1600 1800 2000
n
Fig. 1. The average number of comparisons of 2Merge*: Experiment and analysis
is
lg i + D (pi )

where
6


1 p1i (1/2, 34 12 ],


3 7 + 25
pi ( 34 126 , 3/4],
2 4pi 96p2i
D (pi ) =

2 5
+ 13
pi (3/4, 34 + 123 ],

2pi 24p2i

1 1
pi pi ( 34 + 123 , 1].

Thus, we have
34 4 5
n 1 pn
D (pi ) = 2lg n D (x)dx + D (x)dx + O(log n)
i=1 1/2 1/2

and can obtain the following theorem. The complete analysis for (1,2)Insertion
is given in the full version.

Theorem 2. The complexity of (1,2)Insertion is at most n lg n 1.4034n.

We conducted an experiment for 2Merge . We prepare sequences N =

(1, 2, . . . , n) for n up to 212 = 2046. Then two elements I1 and I2 are selected
from N and they are inserted into N {I1 , I2 } using 2Merge . We take the
average for the number of comparisons for all possible pairs of I1 and I2 . As one
can see the result matches the analysis very well. We also did a similar exper-
iment for 2Merge. The result is very close and the dierence is not visible in
such a graph.
See Fig. 1, which illustrates our analysis and results of simulations. The
symbol + means the average number of comparisons of simulation for each n.
The line represents the value of analysis:
31 3 25
2 4pi + 96p2i pi (1/2, 3/4),
lg i + B(i) +
1 2p3 i + 24p
13
2 pi [3/4, 1].
i
Improved Average Complexity for Comparison-Based Sorting 495

4 Combination with MergeInsertion

See Fig. 2, which illustrates the performance of (1,2)Insertion, (1,2)Insertion ,

and MergeInsertion [2] for the value of pn . As one can see, MergeInsertion is way
better than our algorithms in acertain range of pn . In fact, due to [3, page 389],
2k
its best case happens for n = 3 for an integer k, achieving a complexity of
n lg n (3 lg 3)n + O(lg n) n lg n 1.415n + O(lg n). This best case can be
easily included into our (1,2)Insertion , as follows (see Algorithm 6):
 our input satises pn 2/3. Then we select the largest k such
Suppose that
2k
that n := 3 n. Then we sort the rst n elements by MergeInsertion. After


2n
that the remaining elements are inserted by (1,2)Insertion . Since n = 3p n
as mentioned above, the complexity of MergeInsertion for that size is at most
4
n lg n (3 lg 3)n = n lg n n
3pn
and the additional comparisons in (1,2)Insertion cost is
n 4 pn
{lg i + D (i)} = (n n ) lg n + 2lg n D (x)dx.
i=n +1 2/3

Summing up these two quantities, the complexity of Combination is at most

3 4 pn 5
4 1
n lg n + lg pn + D (x)dx n.
3pn pn 2/3
-1.398
(1,2)Insertion
-1.4 (1,2)Insertion*
MergeInsertion
-1.402 Combined Algorithm

-1.404
Coefficient of n

-1.406

-1.408

-1.41

-1.412

-1.414

-1.416
0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
pn

Fig. 2. Performance of the algorithms

Algorithm 6 Combination(S)
Input: A (unsorted) sequence S = (s1 , s2 , ..., sn ), where n is even.
Output: Sorted sequence
Step 1: If pn 2/3, then let n := 3p 2n
n
. Otherwise, let n := 3pnn .

Step 2: Sort S = (s1 , ..., sn ) by the MergeInsertion
 sort to obtain
 T .

Step 3: For i = n + 2, n + 4, . . . , n, if pi 3
4
6 3
,
12 4
+ 3
12
, then insert si1 and si
into T by calling 2Merge*(si1 , si , T ), otherwise insert si1 into T  by RHBS and
 

then si by RHBS.
496 K. Iwama and J. Teruyama

We can use exactly the same approach for the case that pn 2/3. It turns
out however that the combined approach is worse than MergeInsertion itself for
0.638 pn 2/3. So it is better to use only MergeInsertion for this range. See
Fig. 2 for the overall performance of the combined algorithm.
Theorem 3. The complexity of Combination is at most n lg n 1.41064n.

5 Final Remarks
There is the wide agreement in the community that the information-theoretic
lower bound (= 1.4427) cannot be achieved by a specic sorting algorithm; to
prove or disprove it is a big open question. Anyway, our upper bound for the
average case seems quite close to the lower bound. So attacking the worst case
using the ideas in this paper may be more promising.

References
1. Ayala-Rincon, M., De Abreu, B. T. and De Siqueira, J.: A variant of the Ford
Johnson algorithm that is more space ecient, Information Processing Letters,
Vol. 102, No. 5, pp. 201207 (2007).
2. S. Edelkamp and A. Wei. QuickXsort: Ecient Sorting with n logn - 1.399n +
o(n) Comparisons on Average. CSR 2014: pp. 139152.
3. Ford, L. R. and Johnson, S. M.: A tournament problem, The American Mathemat-
ical Monthly, Vol. 66, No. 5, pp. 387389 (1959).
4. F. K. Hwang and S. Lin, Optimal merging of 2 elements with n elements, Acta
Informatica 1, pp.145158, 1971.
5. K. Iwama and J. Teruyama, Improved Average Complexity for Comparison-Based
Sorting, arXiv:1705.00849, 2017.
6. Knuth, D. E.: The Art of Computer Programming, Volume 3: (2nd Ed.) Sorting
and Searching, Addison Wesley Longman Publishing Co., Inc., Redwood City, CA,
USA (1998).
7. G.K. Manacher, The FordJohnson algorithm is not optimal, Journal of the Asso-
ciation for Computing Machinery 26 (1979) 441456.
8. G.K. Manacher, T.D. Bui, T. Mai, Optimum combinations of sorting and merging,
Journal of the Association for Computing Machinery 36 (1989) 290334.
9. J. Schulte Monting, Merging of 4 or 5 elements with n elements, Theoretical Com-
puter Science 14 (1981) 1937.
10. Peczarski, M.: Sorting 13 Elements Requires 34 Comparisons. In Proc. 10th An-
nual European Symposium on Algorithms. LNCS, vol. 2461, pp. 785794. Springer
(2002).
11. Peczarski, M.: New results in minimum-comparison sorting, Algorithmica, Vol. 40,
No. 2, pp. 133145 (2004).
12. Peczarski, M.: The FordJohnson algorithm still unbeaten for less than 47 elements,
Information processing letters, Vol. 101, No. 3, pp. 126128 (2007).
13. H. Steinhaus, Mathematical Snapshots, New Nork, 1950, pp. 3740.
14. M. Thanh, V.S. Alagar, T. D. Bui, Optimal Expected-Time Algorithms for Merg-
ing. J. Algorithms 7(3): pp. 341357, 1986.
15. Wells, M.: Applications of a Language for Computing in Combinatorics. In Proc.
1965 IFIP Congress, North-Holland, Amsterdam, pp. 497498 (1966).
 An EPTAS for Scheduling on Unrelated
Machines of Few Dierent Types

Klaus Jansen and Marten Maack

Christian-Albrechts-Universitat zu Kiel, 24118 Kiel, Germany,

{kj,mmaa}@informatik.uni-kiel.de

Abstract. In the classical problem of scheduling on unrelated parallel

machines, a set of jobs has to be assigned to a set of machines. The
jobs have a processing time depending on the machine and the goal is to
minimize the makespan, that is, the maximum machine load. It is well
known that this problem is NP-hard and does not allow polynomial time
approximation algorithms with approximation guarantees smaller than
1.5, unless P=NP. We consider the case that there is only a constant
number K of machine types. Two machines have the same type, if all
jobs have the same processing time for them. We present an ecient
polynomial time approximation scheme (EPTAS) for this problem, that is,
for any > 0 an assignment with makespan of length at most (1+) times
the optimum can be found in polynomial time in the input length and the
exponent is independent of 1/. In particular we achieve a running time
41
of 2O(K log(K) / log /) + poly(|I|), where |I| denotes the input length.
1

Furthermore, we study the case where the minimum machine load has to
be maximized and achieve a similar result.

1 Introduction

We consider the problem of scheduling jobs on unrelated parallel machinesor

unrelated scheduling for shortin which a set J of n jobs has to be assigned to
a set M of m machines. Each job j has a processing time pij for each machine
i and the goal is tond a schedule : J M minimizing the makespan
Cmax () = maxiM j1 (i) pij , i.e., the maximum machine load. The problem
is one of the classical scheduling problems studied in approximation. In 1990
Lenstra, Shmoys and Tardos [20] showed that there is no approximation algorithm
with an approximation guarantee smaller than 1.5, unless P=NP. Moreover, they
presented a 2-approximation and closing this gap is a rather famous open problem
in scheduling theory and approximation (see e.g. [23]).
In particular we study the special case where there is only a constant number
K of machine types. Two machines i and i have the same type, if pij = pi j holds
for each job j. In many application scenarios this scenario is plausible, e.g. when
considering computers which typically only have a very limited number of dierent

This work was partially supported by the German Research Foundation (DFG)
project JA 612/16-1.

Springer International Publishing AG 2017 497

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 497508, 2017.
DOI: 10.1007/978-3-319-62127-2_42
498 K. Jansen and M. Maack

types of processing units. We denote the processing time of a job j on a machine

of type t [K] by ptj and assume that the input consist of the corresponding
K n processing time matrix together with machine multiplicities mt for each
type t, yielding m = t[K] mt . Note that the case K = 1 is equivalent to the
classical scheduling on identical machines.
We will also consider the reverseobjective of maximizing the minimum
machine load, i.e., Cmin () = miniM j1 (i) pij . This problem is also known
as max-min fair allocation or the Santa Claus problem. The intuition behind
these names is that the jobs are interpreted as goods (e.g. presents), the machines
as players (e.g. children), and the processing times as the values of the goods from
the perspective of the dierent players. Finding an assignment that maximizes
the minimum machine load, means therefore nding an allocation of the goods
that is in some sense fair (making the least happy kid as happy as possible). We
will refer to the problem as Santa Claus problem in the following, but otherwise
will stick to the scheduling terminology.
We study approximation algorithms: Given an instance I of an optimization
problem, an -approximation A produces a solution in time poly(|I|), where |I|
denotes the input length. For the objective function value A(I) it is guaranteed
that A(I) OPT(I), in the case of an minimization problem, or A(I)
(1/)OPT(I), in the case of an maximization problem, where OPT(I) is the
value of an optimal solution. We call the approximation guarantee or rate of the
algorithm. In some cases a polynomial time approximation scheme (PTAS) can
be achieved, that is, for each > 0 an (1 + )-approximation. If for such a family
of algorithms the running time can be bounded by f (1/)poly(|I|) for some
computable function f , the PTAS is called ecient (EPTAS), and if the running
time is polynomial in both 1/ and |I| it is called fully polynomial (FPTAS).

Related work. It is well known that the unrelated scheduling problem admits
an FPTAS in the case that the number of machines is considered constant [13]
and we already mentioned the seminal work by Lenstra et al. [20]. Furthermore,
the problem of unrelated scheduling with a constant number of machine types
is strongly NP-hard, because it is a generalization of the strongly NP-hard
problem of scheduling on identical parallel machines. Therefore an FPTAS can
not be hoped for in this case. However, Bonifaci and Wiese [5] showed that
there is a PTAS even for the more general vector scheduling case. However,
in the case considered here, their algorithm has to solve m to the power of
O(K(1/)1/ log 1/ ) linear programs. Gehrke et al. [10] presented a PTAS with
2 2
an improved running time of O(Kn) + mO(K/ ) (log(m)/)O(K ) for unrelated
scheduling with a constant number of machine types. On the other hand, Chen
et al. [7] showed that there is no PTAS for scheduling on identical machines with
1
running time 2(1/) for any > 0, unless the exponential time hypothesis fails.
Furthermore, the case K = 2 has been studied: Imreh [14] designed heuristic
algorithms with rates 2+(m1 1)/m2 and 42/m1 , and Bleuse et al. [4] presented
an algorithm with rate 4/3 + 3/m2 and moreover a (faster) 3/2-approximation,
for the case that for each job the processing time on the second machine type is
An EPTAS for Scheduling on Unrelated Machines of Few Different Types 499

at most the one on the rst. Moreover, Raravi and Nelis [22] designed a PTAS
for the case with two machine types.
Interestingly, Goemans and Rothvoss [11] were able to show that unrelated
scheduling is in P, if both the number of machine types and the number of job
types is bounded by a constant. Job types are dened analogously to machine
types, i.e., two jobs j, j  have the same type, if pij = pij  for each machine i. In this
case the matrix (pij ) has only a constant number of distinct rows and columns.
Note that already in the case we study, the rank of this matrix is constant.
However the case of unrelated scheduling where the matrix (pij ) has constant
rank turns out to be much harder: Already for the case with rank 4 there is no
approximation algorithm with rate smaller than 3/2 unless P=NP [8]. In a rather
recent work, Knop and Kouteck [19] considered the number of machine types as
a parameter from the perspective of xed parameter tractability. They showed
that unrelated scheduling is xed parameter tractable for the parameters K and
max pi,j , that is, there is an algorithm with running time f (K, max pi,j )poly(|I|)
for some computable function f that solves the problem to optimality.
For the case that the number of machines is constant, the Santa Claus problem
behaves similar to the unrelated scheduling problem: there is an FPTAS that is
implied by a result due to Woeginger [24]. In the general case however, so far
no approximation algorithm with a constant approximation guarantee has been
found. The results by Lenstra et al. [20] can be adapted to show that that there is
no approximation algorithm with a rate smaller than 2, unless P=NP, and to get
an algorithm that nds a solution with value at least OPT(I) max pi,j , as was
done by Bezkov and Dani [3]. Since max pi,j could be bigger than OPT(I), this
does not provide an (multiplicative) approximation guarantee. Bezkov and Dani
also presented a simple
(nm+1)-approximation and an improved approximation
guarantee of O( n log3 n) was achieved by Asadpour and Saberi [1]. The best
rate so far is O(n ) due to Bateni et al. [2] and Chakrabarty et al. [6], with a
running time of O(n1/ ) for any > 0.

Results and Methodology. In this paper we show:

Theorem 1. There is an EPTAS for both scheduling on unrelated parallel ma-
chines and the Santa Claus problem with a constant number of dierent machine
4
types with running time 2O(K log(K)1/ log 1/) + poly(|I|).
First we present a basic version of the EPTAS for unrelated scheduling with
a running time doubly exponential in 1/. For this EPTAS we use the dual
approximation approach by Hochbaum and Shmoys [12] to get a guess T of the
optimal makespan OPT. Then we further simplify the problem via geometric
rounding of the processing times. Next we formulate a mixed integer linear
program (MILP) with a constant number of integral variables that encodes a
relaxed version of the problem. We solve it with the algorithm by Lenstra and
Kannan [21, 18]. The fractional variables of the MILP have to be rounded and
we achieve this with a cleverly designed ow network utilizing ow integrality
and causing only a small error. With an additional error the obtained solution
500 K. Jansen and M. Maack

can be used to construct a schedule with makespan (1 + O())T . This procedure

is described in detail in Section 2. Building upon the basic EPTAS we achieve
the improved running time using techniques by Jansen [15] and by Jansen, Klein
and Verschae [16]. The basic idea of these techniques is to make use of existential
results about simple structured solutions of integer linear programs (ILPs). In
particular these results can be used to guess the non-zero variables of the MILP,
because they suciently limit the search space. We show how these techniques
can be applied in our case in Section 3. Interestingly, our techniques can be
adapted for the Santa Claus Problem, which typically has a worse approximation
behaviour. We discuss the ideas needed for this in the last section of the paper.
More details and omitted proofs can be found in the long version of the paper
[17].

2 Basic EPTAS
In this chapter we describe a basic EPTAS for unrelated scheduling with a
constant number of machine types, with a running time doubly exponential in
1/. Wlog. we assume < 1. Furthermore log() denotes the logarithm with basis
2 and for k Z0 we write [k] for {1, . . . , k}.
First, we simplify the problem via the classical dual approximation concept
by Hochbaum and Shmoys [12]. In the simplied version of the problem a
target makespan T is given and the goal is to either output a schedule with
makespan at most (1 + )T for some constant Z>0 , or correctly report
that there is no schedule with makespan T . We can use a polynomial time
algorithm for this problem in the design of a PTAS in the following way. First we
obtain an upper bound B for the optimal makespan OPT of the instance with
B 2OPT. This can be done using the 2-approximation by Lenstra et al. [20].
With binary search on the interval [B/2, B] we can nd in O(log 1/) iterations
a value T for which the mentioned algorithm is successful, while T B/2 is
rejected. We have T B/2 OPT and therefore T (1 + )OPT. Hence
the schedule we obtained for the target makespan T has makespan at most
(1+)T (1+)(1+)OPT = (1+O())OPT. In the following we will always
assume that a target makespan T is given. Next we present a brief overview of
the algorithm for the simplied problem followed by a more detailed description
and analysis.

Algorithm 2.
1. Simplify the input via geometric rounding with an error of T .
2. Build the mixed integer linear program MILP(T ) and solve it with the
algorithm by Lenstra and Kannan (T = (1 + )T ).
3. If there is no solution, report that there is no solution with makespan T .
4. Generate an integral solution for MILP(T + T + 2 T ) via a ow network
utilizing ow integrality.
5. The integral solution is turned into a schedule with an additional error of
2 T due to the small jobs.
An EPTAS for Scheduling on Unrelated Machines of Few Different Types 501

Simplication of the Input. We construct a simplied instance I with mod-

ied processing times ptj . If a job j has a processing time bigger than T for a
machine type t [K] we set ptj = . We call a job big (for machine type t), if
ptj > 2 T , and small otherwise. We perform a geometric rounding step for each
job j with ptj < , that is, we set ptj = (1+)x 2 T with x = log1+ (ptj /(2 T )) .
Lemma 1. If there is a schedule with makespan at most T for I, the same
schedule has makespan at most (1 + )T for instance I and any schedule for
instance I can be turned into a schedule for I without increase in the makespan.
We will search for a schedule with makespan T = (1 + )T for I.
We establish some notation for the rounded instance. For any rounded pro-
cessing time p we denote the set of jobs j with ptj = p by Jt (p). Moreover, for
each machine type t let St and Bt be the set of small and big rounded processing
times. Obviously we have |St | + |Bt | n. Furthermore |Bt | is bounded by a
constant: Let N be such that (1 + )N 2 T is the biggest rounded processing
time for all machine type. Then we have (1 + )N 1 2 T T and therefore
|Bt | N log(1/2 )/ log(1 + ) + 1 1/ log(1/2 ) + 1 (using 1).

MILP. For any set of processing times P we call the P -indexed vectors of non-
negative integers ZP0 congurations (for P ). The size size(C) of conguration C

is given by pP Cp p. For each t [K] we consider the set Ct (T ) of congurations
C for the big processing times Bt and with size(C) T . Given a schedule , we
say that a machine i of type t obeys a conguration C, if the number of big jobs
with processing time p that assigns to i is exactly Cp for  each p Bt . Since
the processing times in Bt are bigger than 2 T we have pBt Cp 1/2 for
each C Ct (T ). Therefore the number of distinct congurations in Ct (T ) can be
2 2 2
bounded by (1/2 + 1)N < 2log(1/ +1)1/ log(1/ )+1 2O(1/ log 1/) .
We dene a mixed integer linear program MILP(T ) in which congurations
are assigned integrally and jobs are assigned fractionally to machine types. To
this amount we introduce variables zC,t Z0 for each machine type t [K]
and conguration C Ct (T ), and xj,t 0 for each machine type t [K] and job
j J . For ptj = we set xj,t = 0. Besides this, the MILP has the following
constraints:

zC,t = mt t [K] (1)
CCt (T )

xj,t = 1 j J (2)
t[K]
 
xj,t Cp zC,t t [K], p Bt (3)
jJt (p) CCt (T )
  
size(C)zC,t + p xj,t mt T t [K] (4)
CCt (T ) pSt jJt (p)

With constraint (1) the number of chosen congurations for each machine type
equals the number of machines of this type. Due to constraint (2) the variables
502 K. Jansen and M. Maack

xj,t encode the fractional assignment of jobs to machine types. Moreover, for
each machine type it is ensured with constraint (3) that the summed up number
of big jobs of each size is at most the number of big jobs that are used in the
chosen congurations for the respective machine type. Lastly, (4) guarantees that
the overall processing time of the congurations and small jobs assigned to a
machine type does not exceed the area mt T . It is easy to see that the MILP
models a relaxed version of the problem:
Lemma 2. If there is schedule with makespan T there is a feasible (integral)
solution of MILP(T ), and if there is a feasible integral solution for MILP(T )
there is a schedule with makespan at most T + 2 T . 

2
We have K2O(1/ log 1/) integral variables, i.e., a constant number. Therefore
MILP(T ) can be solved in polynomial time, with the following classical result
due to Lenstra [21] and Kannan [18]:
Theorem 3. A mixed integer linear program with d integral variables and en-
coding size s can be solved in time dO(d) poly(s).

Rounding. In this paragraph we describe how a feasible solution (zC,t , xj,t )

for MILP(T ) can be transformed into an integral feasible solution (zC,t , xj,t ) of
MILP(T + T + 2 T ). This is achieved via a ow network utilizing
 ow integrality.
For any (small or big) processing time p let t,p =  jJt (p) xj,t be the
rounded up (fractional) number of jobs with processing time p thatare assigned to
machine type t. Note that for big job sizes p Bt we have t,p CCt (T ) Cp zC,t
because of (3) and because the right hand side is an integer.

v j1 1 u1,p

1 1,p
1

1
1 m ,p
1
v jn um ,p

Fig. 1. A sketch of the ow network.

Now we describe the ow network G = (V, E) with source and sink .

For each job j J there is a job node vj and an edge (, vj ) with capacity
An EPTAS for Scheduling on Unrelated Machines of Few Different Types 503

1 connecting the source and the job node. Moreover, for each machine type t
we have processing time nodes ut,p for each processing time p Bt St . The
processing time nodes are connected to the sink via edges (ut,p , ) with capacity
t,p . Lastly, for each job j and machine type t with pt,j < , we have an
edge (vj , ut,pt,j ) with capacity 1 connecting the job node with the corresponding
processing time nodes. We outline the construction in Figure 1. Obviously we
have |V | (K + 1)n + 2 and |E| (2K + 1)n.
Lemma 3. G has a maximum ow with value n.
Proof. Obviously n is an upper bound for the maximum ow, and the solution
(zC,t , xj,t ) for MILP(T ) can be used to design a ow f  with value n, by setting
f ((, vj )) = 1, f ((vj , ut,pt,j )) = xj,t and f ((ut,y , )) = jJt (y) xj,t . 

Using the Ford-Fulkerson algorithm, an integral maximum ow f can be found
in time O(|E|f ) = O(Kn2 ). Due to ow conservation, for each job j there is
exactly one machine type t such that f ((vj , ut ,y )) = 1, and we set xj,t = 1
and xj,t = 0 for t = t . Moreover we set zC,t = zC,t . Obviously (zC,t , xj,t ) fulls
(1) and  (2). Furthermore (3) is fullled, because of the capacities and because
t,p CCt (T ) Cp zC,t for big job sizes p. Utilizing the geometric rounding

and the convergence of the geometric series, as well as jJt (p) xj,t t,p <

jJt (p) xj,t + 1, we get:

       1+
p xj,t < p xj,t + p< p xj,t + 2 T

pSt jJt (p) pSt jJt (p) pSt pSt jJt (p)
  
Hence, we have CCt (T ) size(C)zC,t + sSt jJt,s pj,t xj,t < mt (T +T +2 T )
and therefore (4) is fullled as well.

Analysis. The solution found for MILP(T ) can be turned into an integral
solution for MILP(T + T + 2 T ). This can easily be turned into a schedule with
makespan T + T + 2 T + 2 T (1 + 4)T . It is easy to see that the running
time of the algorithm by Lenstra and Kannan dominates the overall running
time. Since MILP(T ) has O(K/ log 1/ + n) many constraints, Kn fractional
2
and K2O(1/ log 1/) integral variables, the running time of the algorithm can be
O(1/ log2 1/)
bounded by 2K2 poly(|I|).

3 Better running time

We improve the running time of the algorithm using techniques that utilize
results concerning the existence of solutions for integer linear programs (ILPs)
with a certain simple structure. In a rst step we can reduce the running time
to be only singly exponential in 1/ with a technique by Jansen [15]. Then we
further improve the running time to the one claimed in Theorem 1 with a very
recent result by Jansen, Klein and Verschae [16]. Both techniques rely upon the
following result about integer cones by Eisenbrandt and Shmonin [9].
504 K. Jansen and M. Maack

Theorem
 4. Let X Zd be a nite set of integer vectors and b int-cone(X) =
{ xX x x | x Z0 }. Then there is a subset X X, such that b int-cone(X)
and |X| 2d log(4dM ), with M = maxxX x .
For the rst improvement of the running time this Theorem is used to show:
Corollary 1. MILP(T ) has a feasible solution where for each machine type
O(1/ log2 1/) of the corresponding integer variables are non-zero.
We get the better running time by guessing the non-zero variables and removing all
the others from the MILP. The number of possibilities of choosing O(1/ log2 1/)
2 2 4
elements out of a set of 2O(1/ log 1/) elements can be bounded by 2O(1/ log 1/) .
Considering all the machine types we can bound the number of guesses by
2 4
2O(K/ log 1/) . The running time of the algorithm by Lenstra and Kannan with
2
O(K/ log2 1/) integer variables can be bounded by 2O(K/ log K/ log 1/) poly(|I|).
2 41
This yields a running time of 2O(K log(K) / log /) poly(|I|).
1

In the following we rst proof Corollary 1 and then introduce the technique
from [16] to further reduce the running time.

Proof of Corollary 1. We consider the so called conguration ILP for schedul-

ing on identical machines. Let m be a given number of machines, P be a set of
processing times with multiplicities kp Z>0 for each p P and let C ZP0 be
some nite set of congurations for P . The conguration ILP for m , P , k, and
C is given by:

C p y C = kp p P (5)
CC

yC = m (6)
CC
yC Z0 C C (7)
The default case that we will consider most of the time is that C is given by a
target makespan T that upper bounds the size of the congurations.
Lets assume we had a feasible solution (zC,t , xj,t ) for MILP(T ). For t [K]

and p Bt we set kt,p = CCt (T ) Cp zC,t . We x a machine type t. By setting
yC = zC,t we get a feasible solution for the conguration ILP given by mt , Bt ,
kt and Ct (T ). Theorem 4 can be used to show the existence of a solution for
the ILP with only a few non-zero variables: Let X be the set of column vectors
corresponding to the left hand side of the ILP and b be the vector corresponding
to the right hand side. Then b int-cone(X) holds and Theorem 4 yields that
there is a subset X of X with cardinality at most 2(|Bt |+1) log(4(|Bt |+1)1/2 )
O(1/ log2 1/) and b int-cone(X). Therefore there is a solution (yC ) for the
ILP with O(1/ log2 1/) many non-zero variables. If we set zC,t = yC and
xj,t = xj,t and perform corresponding steps for each machine type, we get a
solution (zC,t , xj,t ) that obviously satises constraints (1),(2) and (3) of MILP(T ).
The last constraint is also satised, because the number of covered big jobs of
each size does not change and therefore the overall size of the congurations does
not change either for each machine type. This completes the proof of Corollary 1.
An EPTAS for Scheduling on Unrelated Machines of Few Different Types 505

Further Improvement of the Running Time. The main ingredient of the

technique by Jansen et al. [16] is a result about the conguration ILP, for the
case that there is a target makespan T  upper bounding the conguration sizes.
Let C(T  ) be the set of congurations with size at most T  . We need some further
notation. The support of any vector of numbers v is the set of indices with
non-zero entries, i.e., supp(v) = {i | vi = 0}. A conguration is called simple, if
the size of its support is at most log(T  + 1), and complex otherwise. The set of
complex congurations from C(T  ) is denoted by C c (T  ).

Theorem 5. Let the conguration ILP for m , P , k, and C(T  ) have a feasible
solution and let both the makespan T  and the processing times from P be integral.
Then there is a solution (yC ) for the ILP that satises the following conditions:

1. |supp(y|C c (T  ) )| 2(|P | + 1) log(4(|P | + 1)T  ) and yC 1 for C C c (T  ).

2. |supp(y)| 4(|P | + 1) log(4(|P | + 1)T  ).

We will call such a solution thin. Furthermore they argue:

2
(T  )+log2 (|P |))
Remark 1. The total number of simple congurations is in 2O(log .

The better running time can be achieved by determining congurations that are
equivalent to the complex congurations (via guessing and dynamic program-
ming), guessing the support of the simple congurations, and solving the MILP
with few integral variables. The approach is a direct adaptation of the one in [16]
for our case. We now explain the additional steps of the modied algorithm in
more detail and analyze its running time.
We have to ensure that the makespan and the processing times are integral
and that the makespan is small. After the geometric rounding step we scale the
makespan and the processing times, such that T = 1/3 and T = (1 + )/3
holds and the processing times have the form (1 + )x 2 T = (1 + )x /. Next we
apply a second rounding step for the big processing times, setting pt,j = pt,j
for pt,j Bt and denote the set of these processing times by Bt . Obviously we
have |Bt | |Bt | 1/ log(1/2 ) + 1. We denote the corresponding instance by I.

2
Since for a schedule with makespan T for instance I there are at most 1/ big
jobs on any machine, we get:

Lemma 4. If there is a schedule with makespan at most T for I, the same

schedule has makespan at most (1 + 2)T for instance I and any schedule for
instance I can be turned into a schedule for I without increase in the makespan.

We set T = (1 + 2)T and for each machine type t we consider the set of
congurations Ct (T ) for Bt with size at most T . Rounding down T ensures
integrality and causes no problems, because all big processing times are integral.
Furthermore let Ctc (T ) and Cts (T ) be the subsets of complex and simple
congurations. Due to Remark 1 we have:
2
T +log2 |Bt |) 21
|Cts (T )| 2O(log = 2O(log /))
(8)
506 K. Jansen and M. Maack

Due to Theorem 5 (using the same considerations concerning conguration

ILPs like in the last paragraph) we get that there is a solution (zC , xj,t ) for
MILP(T ) (adjusted to this case) that uses for each machine type t at most
4(|Bt |+1) log(4(|Bt |+1)T ) O(1/ log2 1/) many congurations from Ct (T ).
Moreover at most 2(|Bt | + 1) log(4(|Bt | + 1)T ) O(1/ log2 1/) complex con-
gurations are used and each of them is used only once. Since each conguration
corresponds to at most 1/2 jobs, there are at most O(1/3 log2 1/) many jobs
for each type corresponding to complex congurations. Hence, we can determine
the number of complex congurations mct for machine type t along with the
c
number of jobs kt,p with processing time p that are covered by a complex con-
2
guration for each p Bt in 2O(K/ log 1/) steps via guessing. Now we can use
a dynamic program to determine congurations (with multiplicities) that are
equivalent to the complex congurations in the sense that their size is bounded
by T , their summed up number is mct and they cover exactly kt,p c
jobs with
c
processing time p. The dynamic program iterates through [mt ] determining Bt -
indexed vectors y of non-negative integers with yp kt,p c
. A vector y computed
at step i encodes that yp jobs of size p can be covered by i congurations from
Ct (T ). We denote the set of congurations the program computes with Ct
and the multiplicities with zC for C Ct . It is easy;to see that the running
time of such a program can be bounded by O(mct ( pBt (kt,p c
+ 1))2 ). Using
mct O(1/ log2 1/) and kt,p c
O(1/3 log2 1/) this yields a running time of
21
K2O( / log /) , when considering all the machine types.
1

Having determined congurations that are equivalent to the complex congu-

rations, we may just guess the simple congurations. For each machine type, there
2
are at most 2O(log 1/) simple congurations and the number of congurations
we need is bounded by O(1/ log2 1/). Therefore the number of needed guesses
4
is bounded by 2O(K/ log 1/) . Now we can solve a modied version of MILP(T )
in which zC is xed to zC for C Ct and only variables zC  corresponding to
the guessed simple congurations are used. The running time for the algorithm
3
by Lenstra and Kannan can again be bounded by 2O(K log K 1/ log 1/) poly(|I|).
4
Thus we get an overall running time of 2O(K log K 1/ log 1/) poly(|I|). Considering
41 4
the two cases 2O(K log K / log /) < poly(|I|) and 2O(K log K 1/ log 1/) poly(|I|)
1

4
yields the claimed running time of of 2O(K log(K)1/ log 1/) + poly(|I|) completing
the proof of the part of Theorem 1 concerning unrelated scheduling.

4 The Santa Claus Problem

Adapting the result for unrelated scheduling we achieve an EPTAS for the Santa
Claus problem. It is based on the basic EPTAS together with the second running
time improvement. In the following we briey discuss the needed adjustments.
The dual approximation method can be applied in this case as well. However,
since we have no approximation algorithm with a constant rate, the binary search
is more expensive. For the simplication of the input it has to be taken into
An EPTAS for Scheduling on Unrelated Machines of Few Different Types 507

account that their may be huge jobs that are bigger than the optimal makespan,
but otherwise it can be done similarly.
Moreover, in the Santa Claus problem it makes sense to use congurations
of size bigger than T . Let P = T  + max{pt,j | t [K], j Bt }. It suces to
consider congurations with size at most P and for each machine type t we denote
the corresponding set of congurations by Ct (P ). Again we can bound Ct (P ) by
2
2O(1/ log 1/) . The MILP has integral variables zC,t for each such conguration
and fractional ones like before. The constraints (1) and (2) are adapted changing
only the set of congurations and for constraint (3) additionally in this case the
left-hand side has to be at least as big as the right hand side. The last constraint
(4) has to be changed more. For this we partition Ct (P ) into the set Ct (P ) of big
congurations with size bigger than T  and the set Ct (P ) of small congurations
with size at most T . The changed constraint has the following form:
   
size(C)zC,t + p xj,t (mt zC,t )T t [K] (9)
CCt (P ) pSt jJt (p) CCt (P )

To solve the MILP with the claimed running time, some additional non-trivial
considerations are needed that are omitted in this version of the paper.
Lastly, for the rounding of the MILP the ow network has to be changed as
well, using lower and upper bounds for the ow.

Acknowledgements. We thank Florian Mai and Jannis Mell for helpful discus-
sions on the problem.

References
1. Asadpour, A., Saberi, A.: An approximation algorithm for max-min fair allocation
of indivisible goods. SIAM Journal on Computing 39(7), 29702989 (2010)
2. Bateni, M., Charikar, M., Guruswami, V.: Maxmin allocation via degree lower-
bounded arborescences. In: Proceedings of the forty-rst annual ACM symposium
on Theory of computing. pp. 543552. ACM (2009)
3. Bezakova, I., Dani, V.: Allocating indivisible goods. ACM SIGecom Exchanges 5(3),
1118 (2005)
4. Bleuse, R., Kedad-Sidhoum, S., Monna, F., Mounie, G., Trystram, D.: Schedul-
ing independent tasks on multi-cores with gpu accelerators. Concurrency and
Computation: Practice and Experience 27(6), 16251638 (2015)
5. Bonifaci, V., Wiese, A.: Scheduling unrelated machines of few dierent types. arXiv
preprint arXiv:1205.0974 (2012)
6. Chakrabarty, D., Chuzhoy, J., Khanna, S.: On allocating goods to maximize fair-
ness. In: Foundations of Computer Science, 2009. FOCS09. 50th Annual IEEE
Symposium on. pp. 107116. IEEE (2009)
7. Chen, L., Jansen, K., Zhang, G.: On the optimality of approximation schemes
for the classical scheduling problem. In: Proceedings of the Twenty-Fifth Annual
ACM-SIAM Symposium on Discrete Algorithms. pp. 657668. SIAM (2014)
8. Chen, L., Ye, D., Zhang, G.: An improved lower bound for rank four scheduling.
Operations Research Letters 42(5), 348350 (2014)
508 K. Jansen and M. Maack

9. Eisenbrand, F., Shmonin, G.: Caratheodory bounds for integer cones. Operations
Research Letters 34(5), 564568 (2006)
10. Gehrke, J.C., Jansen, K., Kraft, S.E., Schikowski, J.: A ptas for scheduling unrelated
machines of few dierent types. In: International Conference on Current Trends in
Theory and Practice of Informatics. pp. 290301. Springer (2016)
11. Goemans, M.X., Rothvo, T.: Polynomiality for bin packing with a constant
number of item types. In: Proceedings of the Twenty-Fifth Annual ACM-SIAM
Symposium on Discrete Algorithms. pp. 830839. Society for Industrial and Applied
Mathematics (2014)
12. Hochbaum, D.S., Shmoys, D.B.: Using dual approximation algorithms for scheduling
problems theoretical and practical results. Journal of the ACM (JACM) 34(1),
144162 (1987)
13. Horowitz, E., Sahni, S.: Exact and approximate algorithms for scheduling noniden-
tical processors. Journal of the ACM (JACM) 23(2), 317327 (1976)
14. Imreh, C.: Scheduling problems on two sets of identical machines. Computing 70(4),
277294 (2003)
15. Jansen, K.: An EPTAS for scheduling jobs on uniform processors: using an milp
relaxation with a constant number of integral variables. SIAM Journal on Discrete
Mathematics 24(2), 457485 (2010)
16. Jansen, K., Klein, K., Verschae, J.: Closing the gap for makespan scheduling
via sparsication techniques. In: 43rd International Colloquium on Automata,
Languages, and Programming, ICALP 2016, July 11-15, 2016, Rome, Italy. pp.
72:172:13 (2016)
17. Jansen, K., Maack, M.: An EPTAS for scheduling on unrelated ma-
chines of few dierent types. arXiv preprint arXiv:1701.03263v1 (2017),
https://arxiv.org/abs/1701.03263
18. Kannan, R.: Minkowskis convex body theorem and integer programming. Mathe-
matics of operations research 12(3), 415440 (1987)
19. Knop, D., Koutecky, M.: Scheduling meets n-fold integer programming. arXiv
preprint arXiv:1603.02611 (2016)
20. Lenstra, J.K., Shmoys, D.B., Tardos, E.: Approximation algorithms for scheduling
unrelated parallel machines. Mathematical programming 46(1-3), 259271 (1990)
21. Lenstra Jr, H.W.: Integer programming with a xed number of variables. Mathe-
matics of operations research 8(4), 538548 (1983)
22. Raravi, G., Nelis, V.: A ptas for assigning sporadic tasks on two-type heterogeneous
multiprocessors. In: Real-Time Systems Symposium (RTSS), 2012 IEEE 33rd. pp.
117126. IEEE (2012)
23. Williamson, D.P., Shmoys, D.B.: The design of approximation algorithms. Cam-
bridge university press (2011)
24. Woeginger, G.J.: When does a dynamic programming formulation guarantee the
existence of a fully polynomial time approximation scheme (FPTAS)? INFORMS
Journal on Computing 12(1), 5774 (2000)
 A polynomial kernel for
Distance-Hereditary Vertex Deletion

Eun Jung Kim1 and O-joung Kwon2

1
CNRS-Universite Paris-Dauphine, Place du Marechal de Lattre de Tassigny, 75775
Paris cedex 16, France
2
Logic and Semantics, TU Berlin, Berlin, Germany
[email protected], [email protected]

Abstract. A graph is distance-hereditary if for any pair of vertices,

their distance in every connected induced subgraph containing both ver-
tices is the same as their distance in the original graph. The Distance-
Hereditary Vertex Deletion problem asks, given a graph G on n
vertices and an integer k, whether there is a set S of at most k vertices
in G such that G S is distance-hereditary. This problem is important
due to its connection to the graph parameter rank-width [19]; distance-
hereditary graphs are exactly the graphs of rank-width at most 1. Eiben,
Ganian, and Kwon (MFCS 16) proved that Distance-Hereditary Ver-
tex Deletion can be solved in time 2O(k) nO(1) , and asked whether it
admits a polynomial kernelization. We show that this problem admits a
polynomial kernel, answering this question positively. For this, we use a
similar idea for obtaining an approximate solution for Chordal Ver-
tex Deletion due to Jansen and Pilipczuk (SODA 17) to obtain an
approximate solution with O(k3 log n) vertices when the problem is a
Yes-instance, and we exploit the structure of split decompositions of
distance-hereditary graphs to reduce the total size.

1 Introduction
A graph is distance-hereditary if for every connected induced subgraph H and
two vertices u and v in H, the distance between u and v in H is the same as
their distance in G. A vertex subset X of a graph G is a distance-hereditary mod-
ulator, or a DH-modulator in short, if G X is a distance-hereditary graph. We
study the problem Distance-Hereditary Vertex Deletion (DH Vertex
Deletion) which asks, given a graph G and an integer k, whether G contains
a DH-modulator of size at most k.
The graph modication problems, in which we want to transform a graph
to satisfy a certain property with as few graph modications as possible, have
been extensively studied. For instance, the Vertex Cover and Feedback
Vertex Set problems are graph modication problems where the target graphs

O. Kwon is supported by the European Research Council (ERC) under the European
Unions Horizon 2020 research and innovation programme (ERC consolidator grant
DISTRUCT, agreement No. 648527).

Springer International Publishing AG 2017 509

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 509520, 2017.
DOI: 10.1007/978-3-319-62127-2_43
510 E.J. Kim and O.-j. Kwon

are edgeless graphs and forests, respectively. By the classic result of Lewis and
Yannakakis [18], it is known that for all non-trivial hereditary properties that
can be tested in polynomial time, the corresponding vertex deletion problems
are NP-complete. Hence, the research eort has been directed toward designing
algorithms such as approximation and parameterized algorithms.
When the target graph class C admits ecient recognition algorithms for
some NP-hard problems, the graph modication problem related to such a class
attracts more attention. Vertex deletion problems to classes of graphs of con-
stant tree-width or constant tree-depth have attracted much attention in this
context. Tree-width w Vertex Deletion is proved to admit an FPT algo-
rithm running in time 2O(k) nO(1) and a kernel with O(k g(w) ) vertices for some
function g [11,17]. Also, it was shown that Tree-depth w Vertex Deletion
admits uniformly polynomial kernels with O(k 6 ) vertices, for every xed w [12].
All these problems are categorized as vertex deletion problems for F-minor free
graphs in a general setting, when the set F contains at least one planar graph.
However, F-minor free graphs capture only sparse graphs in a sense that the
number of edges of such a graph is bounded by a linear function on the number
of its vertices. Thus these problems are not useful when dealing with very dense
graphs.
Rank-width [19] and clique-width [5] are graph width parameters introduced
for extending graph classes of bounded tree-width. Graphs of bounded rank-
width represent graphs that can be recursively decomposed along vertex par-
titions (X, Y ) where the number of neighborhood types between X and Y are
small. Thus, graphs of constant rank-width may contain dense graphs; for in-
stance, all complete graphs have rank-width at most 1. Courcelle, Makowski,
and Rotics [4] proved that every M SO1 -expressible problem can be solved in
polynomial time on graphs of bounded rank-width.
Motivated from Tree-width w Vertex Deletion, Eiben, Ganian, and
the second author [9] initiated study on vertex deletion problems to graphs of
constant rank-width. The class of graphs of rank-width at most 1 is exactly the
class of distance-hereditary graphs [19]. It was known that the vertex deletion
problem for graphs of rank-width w can be solved in FPT time [16] using a
meta-theorem [4]. Eiben et al. [9] devised the rst elementary algorithm for
this problem when w = 1, or equivalently DH Vertex Deletion, that runs
in time 2O(k) nO(1) . Furthermore, they discussed that a DH-modulator of the
size k can be used to obtain a 2O(k) nO(1) -time algorithm for problems such as
Independent Set, Vertex Cover, and 3-Coloring.
However, until now, it was not known whether DH Vertex Deletion ad-
mits a polynomial kernel or not. A kernelization of a parameterized graph prob-
lem is a polynomial-time algorithm which, given an instance (G, k) of ,
outputs an equivalent instance (G , k  ) of with |V (G )| + k  h(k) for some
computable function h. The resulting instance (G , k  ) of a kernelization is called
a kernel, and in particular, when h is a polynomial function, is said to admit
a polynomial kernel.

Our Contribution and Approach. Our main result is the following.

A polynomial kernel for Distance-Hereditary Vertex Deletion 511

Theorem 1. DH Vertex Deletion admits a polynomial kernel.

We introduce in Section 3 an approximate DH-modulator with O(k 3 log n)

vertices if the given instance is a Yes-instance. An important observation here
is that a distance-hereditary graph contains a complete bipartite subgraph (not
necessarily induced) which is a balanced separator. Thus, if G admits a small DH-
modulator, then there is balanced vertex separator XK where X is small and K
induces a complete bipartite subgraph. By recursively extracting such separators
using an approximation algorithm for nding a balanced vertex separator [10],
we will decompose the given graph into D  K1   K  X, where  =
O(k log n), Dis distance-hereditary, each Ki is a complete bipartite subgraph,
|X| = O(k 2 log k log n). In the next step, we argue that if a graph H is a
disjoint union of a distance-hereditary graph and a complete bipartite graph and
(H, k) is a Yes-instance and satises a certain property, then in polynomial time,
one can construct a DH-modulator of size O(k 2 ) for H (Proposition 2). Using
this sub-algorithm  times, we construct an approximate DH-modulator with
O(k 3 log n) vertices. This part follows a vein similar to the approach of Jansen
and Pilipczuk [15] for Chordal Vertex Deletion. Given a DH-modulator S
of size O(k 3 log n), we can obtain a new DH-modulator S  of size O(k 5 log n)
such that for every v S  , G[(V (G) \ S  ) {v}] is also distance-hereditary by
adding O(k 2 ) vertices per each vertex in S. Such a DH-modulator is called a
good DH-modulator and the details will be explained in Section 4.
The remaining part of the paper is contributed to reduce the number of
vertices in GS  . Two vertices v and w are twins if they have the same neighbors
outside {v, w}. In Section 5, we present a reduction rule that bounds the size
of each set of pairwise twins in G S  . We give, in Section 6, a reduction rule
that bounds the number of components of G S  . Lastly in Section 7, we reduce
the size of each component of G S  having at least 2 vertices. For the last
part, we use split decompositions of distance-hereditary graphs. Briey, a split
decomposition displays a tree-like structure of a distance-hereditary graph in
the form of a decomposition tree with bags for each nodes, such that each bag
consists of a maximal set of pairwise twins (possibly with an extra vertex) in G
S  . We will provide a rule that bounds the number of bags in the decomposition
tree, which results in bounding the size of each component.

2 Preliminaries
We follow [8] for basic graph terminology. A graph is trivial if it consists of a
single vertex, and non-trivial otherwise. For two sets A, B V (G), we say A is
complete to B if for every v A and w B, v is adjacent to w. Two vertices v
and w of a graph G are twins if they have the same neighbors in G {v, w}. A
vertex partition (A, B) of G is split if NG (A) is complete to NG (B).
A graph H is a biclique if there is a bipartition of V (H) into non-empty sets
A  B such that any two vertices a A and b B is adjacent. Notice that there
may be edges among the vertices of A or B. For K V (G), we say that K is a
biclique of G if G[K] is a biclique. For a connected graph G, a vertex subset S
512 E.J. Kim and O.-j. Kwon

of G is called a balanced vertex separator of G if every component of G S has

at most 23 |V (G)| vertices. We allow V (G) to be a balanced vertex separator of
G. For a vertex subset S of G, a path is called an S-path if its end vertices are
in S and all other internal vertices are in V (G) \ S.

house gem domino cycle 5

Fig. 1. The induced subgraph obstructions for distance-hereditary graphs.

A graph is called a DH obstruction if it is isomorphic to a gem, a house,

a domino or an induced cycle of length at least 5, that are depicted in Fig-
ure 1. A DH obstruction is small if it has at most 6 vertices. Bandelt and
Mulder [2] proved that a graph is distance-hereditary if and only if it has no
induced subgraph isomorphic to one of DH obstructions. A DH-modulator S is
good if G[(V (G) \ S) {v}] is distance-hereditary for every v S.

3 Approximation algorithm
We present a polynomial-time algorithm which constructs an approximate DH-
modulator of G whenever (G, k) is a Yes-instance.
Theorem 2. There is a polynomial-time algorithm which, given a graph G and
a positive integer k, either correctly reports that (G, k) is a No-instance to DH
Vertex Deletion, or returns a DH-modulator of size O(k 3 log n).
If G contains k+1 vertex-disjoint copies of small DH obstructions, then (G, k)
is a No-instance. We may assume a maximal packing of small DH obstructions
has cardinality at most k. Since a maximal packing consists of at most 6k vertices,
it is sucient to prove Theorem 2 when G has no small DH obstruction.
We prove the following two propositions, implying Theorem 2 together.
Proposition 1. There is a polynomial-time algorithm which, given an instance
(G, k), either computes a decomposition V (G) = D  K  
K1  X such that
G[D] is distance-hereditary, each Ki is a biclique, |X| = O(k 2 log k log n) and
 = O(k log n), or correctly reports that (G, k) is a No-instance.
When G does not contain any small DH obstructions, a linear program of
DH Vertex Deletion for G can be formulated as follows, where xv 0 for
each v V (G):
 
min xv s.t xv 1 induced cycle H of length at least 7
vV (G) vV (H)

A mapping x = (xv )vV (G) from V (G) to R is a feasible fractional solution to

DH Vertex Deletion for G if it is feasible
 to the above linear program for G.
For a subgraph H of G, we write x(H) := vV (H) xv and |x| := x(G).
A polynomial kernel for Distance-Hereditary Vertex Deletion 513

Proposition 2. Let (G, k) be an instance to DH Vertex Deletion such that

G has no small DH obstructions, V (G) = D  K where G[D] is distance-
hereditary, and K is a biclique. Let x be a feasible fractional solution to DH
1
Vertex Deletion for G such that xv < 20 , v V (G). Given such G and x ,
one can in polynomial time nd a DH-modulator X with O(|x |2 ) vertices.

We rst explain Proposition 1. First, we obtain an O(n3 ) bound on the num-

ber of maximal bicliques in a graph having no small DH obstructions. Secondly,
we prove that every connected distance-hereditary graph on at least two vertices
contains a balanced vertex separator that is a biclique. Combining these results,
we can show the following.
Lemma 1. Whenever (G, k) is a Yes-instance and G has no small DH obstruc-
vertex separator K  X where
tions, one can in polynomial time nd a balanced
K is a biclique or an empty set and |X| = O(k log k).

Proof
(Sketch of proof ). Over all maximal bicliques K of G, we apply the
O( log OP T )-approximation algorithm for nding a balanced vertex separa-
tor in G K, due to Feige et al [10]. One can observe that since (G, k) is a
Yes-instance, there is some set X of size at most k and a balanced vertex sepa-
rator of K  of G X that is a biclique. Thus, a maximal biclique of G containing
this K  is detected in
the algorithm, and the approximation algorithm provides
a set X  of size O(k log k) where K   X  is a balanced vertex separator. 


We set G1 := G, K0 = X0 = , and ati-th recursive step, we apply Lemma 1

to a connected component Gi of G j<i (Kj  Xj ) which is not distance-
hereditary and obtain a balanced vertex separator Ki  Xi of Gi . In the end,
we obtain a decomposition V (G) = D  K   K1  X   X1 , where
G[D]is distance-hereditary, each Ki is a biclique or an empty set, and |Xi | =
O(k log k). Since we only apply Lemma 1 to a component that is not distance-
hereditary, if (G, k) is a Yes-instance, then the size-k-modulator of G intersects
every such component. By representing the recursive procedure as a collection
of branching trees T , we can show that  = O(k log n), as the maximum length
of a root-to-leaf path in T is O(log n).
Now, we explain Proposition 2. Suppose G has no small DH obstructions
and V (G) = D  K where G[D] is distance-hereditary, and K is a biclique with
a bipartition (A, B), and x is a feasible fractional solution to DH Vertex
1
Deletion such that xv < 20 for every v V (G). We rst observe that a
new vector x where xv = 0 if v K and xv = 2xv if v D, is again a
 

feasible fractional solution. For this, we show that every induced cycle H of
length at least 7 in G satises that if G[V (H) K] has one component, then
|V (H) K| 3, and otherwise, H contains a K-path whose length is at least 3.
In the former case, we have x (H) = x (G[V (H) D]) = 2x (G[V (H) D]) 1
1
as x (G[V (H) K]) < 3 20 < 12 . In the latter case, the end vertices of the
K-path P are contained in the same part of A or B, and it forms another DH
obstruction H  with a vertex in the other part, where G[V (H  ) K] has one
component. Thus, we have x (H) x (P ) x (H  ) 1, as xv = 0 for v K.
514 E.J. Kim and O.-j. Kwon

We construct an instance (G[D], T ) of Vertex Multicut with terminal

pairs T := {(s, t) D D : distG[D],x (s, t) 1}, where distG[D],x (s, t) is the
minimum x (P ) over all (s, t)-paths P . Notice that for every terminal pair (s, t)
T , and for every (s, t)-path P in G[D], we have x (P ) distG[D],x (s, t) 1,
meaning that x is a feasible fractional solution to Vertex Multicut for the
instance (G[D], T ). Using an approximation algorithm for Vertex Multicut
by Gupta [14], we can obtain a vertex set X D of size O(|x |2 ) such that
G[D \ X] contains no (s, t)-path for every terminal pair (s, t) T in polynomial
time. We prove, in the appendix, that the obtained set X is a DH-modulator.
Proof (of Theorem 2). It is sucient to prove when G has no small DH obstruc-
tions. Let x be an optimal fractional solution to DH Vertex Deletion for G.
We may assume |x | k, otherwise we report that (G, k) is a No-instance. Let
1
X be the set of all vertices v such that xv 20 . Observe that |X| 20k since
1
otherwise, |x | 20 |X| > k, a contradiction. Also x restricted to V (G) \ X is

1
a fractional feasible solution for G X such that xv < 20 for every v.

We compute a decomposition V (GX) = D i=1 Ki X as in Proposition 1,
or correctly report (G, k) as a No-instance. Recall that  = O(k log n) and

|X| = O(k2 log k log n). Note that V (G (X X)) = D  i=1 Ki . From i = 1
up to , we want to obtain a DH-modulator Si of Gi , where G1 := G[D K1 ] and
for i = 2, . . . , , Gi is the subgraph of G induced by (V (Gi1 ) \ Si1 ) Ki . Note
that Gi Si is distance-hereditary and Ki is a blique. Hence, we can inductively
apply the algorithm of Proposition 2 and obtain a DH-modulator Si of size at
most O(|x |2 ) of Gi . Especially, G S is distance-hereditary, implying that the

set dened as S := X X i=1 Si is a DH-modulator of G. From |x | k, we
have |Si | = O(k 2 ) for each i. It follows that |S| = O(k 3 log n). 


4 Good modulator
Theorem 3. There is a polynomial-time algorithm which, given a graph G and a
positive integer k, either correctly reports that (G, k) is a No-instance, or returns
an equivalent instance (G , k  ) with a good DH-modulator of size O(k 5 log n).
Proof (Sketch of Proof ). If the algorithm of Theorem 2 reports that the instance
is a No-instance, then we are done. Let S be a DH-modulator of size O(k 3 log n)
given by Theorem 2. Let U := , and for v S, let Hv := G[(V (G) \ S) {v}].
One can in polynomial time nd either k + 1 small DH obstructions in Hv whose
pairwise intersection is v, or a vertex set Tv of V (G) \ S such that |Tv | 5k and
Hv Tv has no small DH obstructions. In the former case, we add v to U .
Assume we obtain a vertex set Tv . Since Hv Tv has no small DH obstruc-
tions, every DH obstruction in Hv Tv is an induced cycle of length at least
7. We assert that either Hv Tv contains a vertex set Xv of size O(k 2 ) such
that Hv (Tv Xv ) has no DH obstructions, or correctly reports that every
DH-modulator of size at most k contains v.
We consider an instance (Hv (Tv {v}), T ) of Vertex Multicut where
T := {(s, t) : s, t NHv Tv (v), distHv (Tv {v}) (s, t) 3}. We can show that
A polynomial kernel for Distance-Hereditary Vertex Deletion 515

X V (Hv ) \ (Tv {v}) hits all induced cycles of Hv Tv of length at least

7 if and only if X is a vertex multicut for (Hv (Tv {v}), T ), because the
restriction of an induced cycle of Hv Tv of length at least 7 is an induced path
of length at least 3 between two neighbors of v in Hv Tv , and the shortest
path between those vertices and the induced path have the same length, as
Hv (Tv {v}) is distance-hereditary. Let x be an optimal fractional solution
to Vertex Multicut, which can be eciently found using the ellipsoid method
and an algorithm for the (weighted) shortest path problem as a separation oracle.
If |x | k, then we can construct a multicut Xv V (Hv ) \ (Tv {v}) of size
O(|x |2 ) = O(k 2 ) using the approximation algorithm of Gupta [14]. If |x | > k,
then any integral solution for (Hv (Tv {v}), T ) is larger than k, and any
DH-modulator of size at most k must contain v. In this case, we add v to U .
We can conrm that (G U, k |U |) is an instance equivalent to (G, k) and
S ( vS\U (Tv Xv )) is a good DH-modulator for G U . 


5 Twin Reduction Rule

In a distance-hereditary graph, there may be a large set of pairwise twins. We
introduce a reduction rule that bounds the size of a set of pairwise twins in GS
by O(k 2 |S|3 ), where S is a DH-modulator (not necessarily good). The underlying
observation is that it suces to keep up to k + 1 vertices that are pairwise twins
with respect to each subset of S of small size. For a subset S  S, two vertices
u and v in V (G) \ S are S  -twins if u and v have the same neighbors in S  . It is
not dicult to get an upper bound O(k|S|5 ), by considering all subsets S  of S
of size min{|S|, 5} and marking up to k + 1 S  -twins. To get a better bound, we
proceed as follows.

Reduction Rule 1 Let W be a set of pairwise twins in G S, and let m :=

min{|S|, 3}. (1) Over all subsets S  S of size m, we mark up to k + 1 pairwise
S  -twins in W that are unmarked yet. (2) When |S| 4, over all subsets S  S
of size 4, if there is an unmarked vertex v of W such that G[S  {v}] is isomorphic
to the house or the gem, then we mark up to k + 1 previously unmarked vertices
in W including v that are pairwise S  -twins. (3) If there is an unmarked vertex
v of W after nishing the marking procedure, we remove v from G.

If (G, k) is irreducible with respect to Reduction Rule 1 and |S| 4, then

each set W of pairwise twins in G S contains O(k 2 |S|3 ) vertices, or (G, k) is a
No-instance. This is because, if (G, k) is a Yes-instance, then all chosen subsets
S  in (2) can be covered by at most 4k vertices. If |S| 3, then W has at most
8(k + 1) vertices. To see the safeness, suppose there is an unmarked vertex v
of W after nishing the marking procedure. It is clear that if (G, k) is a Yes-
instance, then (G v, k) is a Yes-instance. Suppose G v has a DH-modulator
T of size at most k, and G T contains a DH obstruction F containing v. In
case when F v is an induced path, let w, z be the end vertices of the path,
and choose a set S  S of size 3 containing {w, z} S. Since v is unmarked
in Reduction Rule 1, there are v1 , . . . , vk+1 W \ {v} where v1 , . . . , vk+1 , v are
516 E.J. Kim and O.-j. Kwon

pairwise S  -twins. Note that V (F ) {v1 , . . . , vk+1 } = since no other vertex in

F is adjacent to both w and z. Thus, there exists v  {v1 , . . . , vk+1 } \ T such
that G[V (F ) \ {v} {v  }] is a DH obstruction in (G v) T , contradiction.
If |S V (F )| 3, then we can proceed in the same way. We may assume
F v is not an induced path and |S V (F )| 4. If F is the house or the gem,
then we marked necessary vertices in (2), and thus we can proceed similarly. If F
is the domino, then v should be a vertex of degree 2 in F . We can prove that the
3 vertices S  in F v, two neighbors of v and the vertex farthest from v, satises
that the existence of S  -twins with v is enough to get another DH obstruction.

6 The number of non-trivial components of G S

We provide a reduction rule that bounds the number of non-trivial components

of G S, when S is a good DH-modulator. For v S and a component C of
GS, let N (v, C) := NG (v)V (C). We say that a pair (v, w) of vertices in S is a
witnessing pair (for being non-split) for a component C of G S if N (v, C) = ,
N (w, C) = and N (v, C) = N (w, C). The following lemma is essential.
Lemma 2. If C1 , C2 , . . . , Cm are distinct connected components of G S with
m 2 and v1 , v2 , . . . , vm are distinct vertices of S (vm+1 = v1 ) such that for each
i {1, . . . , m}, (vi , vi+1 ) is a witnessing pair for Ci , then G[{v1 , v2 , . . . , vm }
i{1,...,m} V (Ci )] contains a DH obstruction.

Lemma 2 for m = 2 observes that if a pair of vertices in S witnesses at least k +2

non-trivial components in GS, at least one of the pair must be contained in any
size-k DH-modulator. Furthermore, keeping exactly k+2 non-trivial components
would suce to impose this restriction. This suggests the following rule.

Reduction Rule 2 For each pair of vertices v and w in S, we mark up to k + 2

non-trivial (previously unmarked) connected components C of G S such that
(v, w) is a witnessing pair for C. If there is an unmarked non-trivial connected
component C after the marking procedure, then we remove all edges in C.

For the safeness of Reduction Rule 2, suppose there was an unmarked non-
trivial connected component C after the marking procedure, and G is the re-
sulting graph. We mainly observe that if G has a DH-modulator T of size at
most k, then (V (C) \ T, V (G) \ V (C) \ T ) is a split in G T . Otherwise, there
are v, w S and components C1 , . . . , Ck+2 where (v, w) is a witnessing pair
for C, C1 , . . . , Ck+2 in G. Then there are 2 components among C1 , . . . , Ck+2
that does not intersect T , and by Lemma 2, G T contains a DH obstruc-
tion, contradiction. Thus, if there is a DH obstruction H in G T , then since
(V (C) \ T, V (G) \ V (C) \ T ) is a split in G T and S is a good-modulator, we
have |V (H)V (C)| 1. This implies that G T also contains H, contradiction.
We prove for the other direction in the similar way.

Proposition 3. If (G, k) is irreducible with respect to Reduction Rule 2, then

either the number of non-trivial components is O(k 2 |S|) or it is a No-instance.
A polynomial kernel for Distance-Hereditary Vertex Deletion 517

Proof (of Proposition 3). Suppose (G, k) is a Yes-instance. We dene an auxil-

iary multigraph F on S such that for v, w S, the multiplicity of the edge vw
equals the number of non-trivial components that are marked by the witness of
(v, w) in Reduction Rule 2. It suces to obtain a bound on the number of edges
in F with the edge multiplicity taken into account.
Construct a maximal packing of 2-cycles in F and let S1 S be the vertices
contained in the packing. By Lemma 2, a packing of size k + 1 implies the
existence of k + 1 vertex-disjoint DH obstructions. Therefore, |S1 | 2k. Again,
due to the assumption that (G, k) is a Yes-instance, the subgraph F S1 does
not have k + 1 vertex-disjoint cycles: otherwise, G contains k + 1 vertex-disjoint
DH obstructions by Lemma 2. By the Erdos-Posa property of cycles, there exists
S2 V (F ) \ S1 hitting all cycles of F S1 with |S2 | rk log k for some constant
r. Now, the number of edges in F is at most |S1 | |S|(k + 2) + |S2 | |S \ S1 | +
(|S \ S1 \ S2 |) = 2k(k + 2)|S| + rk log k|S| + |S| (7 + r)k 2 |S|. 


7 The size of non-trivial components of G S

It remains to bound the size of each non-trivial connected component of G S.
For this, we need to use split decompositions that present tree-like structure of
distance-hereditary graphs. For the length constraint, we shortly dene here with
an example, and put the full description in the appendix (preliminary section).

B5

B4

B1 B2 B3 G
Fig. 2. An example of a split decomposition of a distance-hereditary graph. Dashed
edges denote marked edges and each Bi denotes a bag.

A connected graph G is prime if |V (G)| 5 and it has no split. A connected

graph D with a distinguished set of cut edges M (D) of D is called a marked
graph if M (D) forms a matching. An edge in M (D) is a marked edge, and every
other edge is an unmarked edge. A vertex incident with a marked edge is a
marked vertex, and every other vertex is an unmarked vertex. Each component
of D M (D) is a bag of D. See Figure 2 for an example. When G admits a split
(A, B), we construct a marked graph D on the vertex set V (G) {a , b } such
that (1) a b is a new marked edge, (2) there are no edges between A and B,
(3) {a } is complete to NG (B), {b } is complete to NG (A), and (4) G[A] = D[A]
and G[B] = D[B]. The marked graph D is a simple decomposition of G. A split
decomposition of a connected graph G is a marked graph D dened inductively
to be either G or a marked graph dened from a split decomposition D of G by
replacing a bag B with its simple decomposition.
518 E.J. Kim and O.-j. Kwon

Cunningham and Edmonds [6] developed a canonical way to decompose a

graph into a split decomposition. A split decomposition D of G is canonical if
each bag of D is either a prime graph, a star, or a complete graph, and recom-
posing any marked edge of D violates this property. It is unique up to isomor-
phism [6] and can be computed in time O(|V (G)| + |E(G)|) [7]. In particular,
Bouchet [3] proved that a graph is distance-hereditary if and only if every bag
in its canonical split decomposition is a star or a complete graph.
Let D be the canonical split decomposition of a non-trivial component H of
GS. It is known that unmarked vertices in each bag of D consist of at most two
twin sets in G S. Thus, by Reduction Rule 1, it suces to bound the number
of bags of D. Since S is a good DH-modulator, for each v S, G[V (H) {v}]
is distance-hereditary. Gioan and Paul [13, Theorem 3.4] described the way of
extending D to the canonical split decomposition of G[V (H){v}]. In particular,
there exists a bag or a marked edge that is modied when pushing v, and we
can nd this place in time O(|V (G)|). Such a bag or a marked edge is called
S-aected. A bag B is a branch bag if D V (B) contains at least 3 connected
components having at least two bags. For two adjacent bags B1 and B2 , we
denote by e(B1 , B2 ) the marked edge linking B1 and B2 .
We rst apply three reduction rules dealing with leaf bags. Firstly, we remove
a vertex of degree 1 in G. Since any DH obstruction does not contain a vertex of
degree 1, this is safe. Secondly, if there are a leaf bag B and its neighbor bag B 
such that B, B  , e(B, B  ) are S-unaected, and B  is a star bag whose center is
adjacent to B, and D V (B  ) has exactly two components, then we remove all
unmarked vertices of B  . In this case, all neighbors of a vertex in B  are contained
in B whose unmarked vertices are pairwise twins in G. Thus, any DH obstruction
does not contain a vertex of B  , and this rule is safe. Lastly, if A V (H) is a
maximal set of pairwise twins in G and ipping the adjacency between every
two vertices of A reduces the number of bags, then we ip the adjacency. For
instance, if B is a leaf bag that is a complete graph and its neighbor bag is a
star bag whose leaf is adjacent to B, and B is S-unaected, then by ipping the
adjacency between two vertices in B, we can transform B into a star bag and
merge with B  . This rule is also used in the FPT algorithm [9].
After applying those rules exhaustively, to nd a reducible part, we color the
bags of D with red and blue in the following way. (1) If a bag B is S-aected or
incident with an S-aected edge, we color B with red. (2) If a bag B is adjacent
to an S-aected leaf bag, we color B with red. (3) If B is a branch bag, then
we color B with red. (4) All other bags are colored with blue. We can show that
the number of red bags is at most 3|S|, and for each bag B, there is at most
one blue
 bag adjacent to B. We further show that the number of components of
D BRQ V (B) is at most 3|S|, where R is the set of red bags, and Q is the
set of blue leaf bags whose neighbor bags are red.

It remains to bound each connected
 component of D BRQ V (B). Let D
be a connected component of D BRQ V (B). As R contains all branch bags,
D contains no branch bags. Therefore, there is a sequence B1 B2 Bm
of bags of D that are not leaf bags, and all other bags are leaf bags adjacent
A polynomial kernel for Distance-Hereditary Vertex Deletion 519

to one of B1 , . . . , Bm . For 1 i1 < i2 < i3 m, Bi2 is a (Bi1 , Bi3 )-separator

bag, if it is a star bag whose center is adjacent to neither Bi2 1 nor Bi2 +1 . We
rst bound the number of (B1 , Bm )-separator bags, since if there are many, then
we can merge two closest separator bags into one bag. This is similar to usual
bypassing rule in Feedback Vertex Set. One interesting part is to bound
the number of the sequence of consecutive bags that are not (B1 , Bm )-separator
bags. We show that if there is such a sequence of length more than 5k + 11, then
we can always nd a vertex that can be safely  removed. In the end, we reduce
the number of bags in each component of D BRQ V (B) to 20k + 52, and
we conclude that the number of all bags is bounded by 3|S|(20k + 54) bags.
Proof (of Theorem 1). We rst prove that given an instance (G, k) and a good
DH-modulator S, one can output an equivalent instance of size O(k 5 |S|5 ). We
rst apply Reduction Rule 2 to (G, k) with S. After that, G S has O(k 2 |S|)
non-trivial components or we can correctly report that (G, k) is a No-instance
by Proposition 3. Next, we apply reduction rules for reducing the size of non-
trivial components of GS. We prove the safeness, polynomial-time applicability,
and preserving the goodness of S in the appendix. In the end, the canonical
split decomposition D of each non-trivial component of G S has at most
3|S|(20k +54) bags. Last, we apply Reduction Rule 1 exhaustively in polynomial
time. This bounds the size of a twin set in G S by O(k 2 |S|3 ). We note that the
unmarked vertices of a bag form at most two twin sets. Therefore, the number
of unmarked vertices in a bag is bounded by O(k 2 |S|3 ). Especially, the same
bounds apply to the number of trivial components in G S since they form
an independent set in G S. Combining the previous bounds altogether, we
conclude that V (G ) = O(k 5 |S|5 ), when (G , k  ) is the resulting instance.
We may assume n 2ck for some constant c. Recall that there is an algorithm
for DH Vertex Deletion running in time 2ck nO(1) by Eiben, Ganian, and
Kwon [9]. If n > 2ck , then the algorithm of [9] solves the instance (G, k) correctly
in polynomial time, in which case we can output a trivial equivalent instance. By
Theorem 3, we can obtain a good DH-modulator S of size O(k 5 log n) = O(k 6 )
in polynomial time or correctly report (G, k) as a No-instance. The previous
argument yields that in polynomial time, an equivalent instance (G , k  ) of size
O(k 35 ) can be constructed. Now, applying Theorem 3 again3 to (G , k  ), we
can either correctly conclude that (G , k  ), and thus (G, k), is a No-instance or
output a good DH-modulator S  of size O(k 5 log k). Finally we obtain a kernel
of size O(k 30 log5 k).

References
1. Agrawal, A., Lokshtanov, D., Misra, P., Saurabh, S., Zehavi, M.: Feedback vertex
set inspired kernel for chordal vertex deletion. In: Proceedings of the Twenty-Eighth
Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2017, Barcelona,
Spain, Hotel Porta Fira, January 16-19. pp. 13831398 (2017)
3
That applying kernelization twice can yield an improved bound was adequately
observed in [1].
520 E.J. Kim and O.-j. Kwon

2. Bandelt, H.J., Mulder, H.M.: Distance-hereditary graphs. J. Combin. Theory Ser.

B 41(2), 182208 (1986)
3. Bouchet, A.: Transforming trees by successive local complementations. J. Graph
Theory 12(2), 195207 (1988)
4. Courcelle, B., Makowsky, J.A., Rotics, U.: Linear time solvable optimization prob-
lems on graphs of bounded clique-width. Theory Comput. Syst. 33(2), 125150
(2000)
5. Courcelle, B., Olariu, S.: Upper bounds to the clique width of graphs. Discrete
Appl. Math. 101(1-3), 77114 (2000)
6. Cunningham, W.H., Edmonds, J.: A combinatorial decomposition theory. Canad.
J. Math. 32(3), 734765 (1980)
7. Dahlhaus, E.: Parallel algorithms for hierarchical clustering, and applications to
split decomposition and parity graph recognition. Journal of Algorithms 36(2),
205240 (2000)
8. Diestel, R.: Graph Theory. No. 173 in Graduate Texts in Mathematics, Springer,
third edn. (2005)
9. Eiben, E., Ganian, R., Kwon, O.: A Single-Exponential Fixed-Parameter Algo-
rithm for Distance-Hereditary Vertex Deletion. In: Faliszewski, P., Muscholl, A.,
Niedermeier, R. (eds.) 41st International Symposium on Mathematical Founda-
tions of Computer Science, MFCS 2016. Leibniz International Proceedings in In-
formatics (LIPIcs), vol. 58, pp. 34:134:14. Schloss DagstuhlLeibniz-Zentrum fuer
Informatik, Dagstuhl, Germany (2016)
10. Feige, U., Hajiaghayi, M., Lee, J.R.: Improved approximation algorithms for min-
imum weight vertex separators. SIAM J. Comput. 38(2), 629657 (2008)
11. Fomin, F.V., Lokshtanov, D., Misra, N., Saurabh, S.: Planar F-Deletion: approx-
imation, kernelization and optimal FPT algorithms. In: 2012 IEEE 53rd Annual
Symposium on Foundations of Computer ScienceFOCS 2012, pp. 470479. IEEE
Computer Soc., Los Alamitos, CA (2012)
12. Giannopoulou, A.C., Jansen, B.M.P., Lokshtanov, D., Saurabh, S.: Uniform kernel-
ization complexity of hitting forbidden minors. In: Automata, languages, and pro-
gramming. Part I, Lecture Notes in Comput. Sci., vol. 9134, pp. 629641. Springer,
Heidelberg (2015)
13. Gioan, E., Paul, C.: Split decomposition and graph-labelled trees: characteriza-
tions and fully dynamic algorithms for totally decomposable graphs. Discrete Appl.
Math. 160(6), 708733 (2012)
14. Gupta, A.: Improved results for directed multicut. In: Proceedings of the Four-
teenth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2003, Jan-
uary 12-14, 2003, Baltimore, Maryland, USA. pp. 454455 (2003)
15. Jansen, B.M.P., Pilipczuk, M.: Approximation and kernelization for chordal vertex
deletion. In: 28th Annual ACM-SIAM Symposium on Discrete Algorithms, SODA
2017, January 16-19, 2017, Barcelona, Spain. (2017), to appear
16. Kante, M.M., Kim, E.J., Kwon, O., Paul, C.: An fpt algorithm and a polynomial
kernel for linear rankwidth-1 vertex deletion. Algorithmica pp. 130 (2016)
17. Kim, E.J., Langer, A., Paul, C., Reidl, F., Rossmanith, P., Sau, I., Sikdar, S.:
Linear kernels and single-exponential algorithms via protrusion decompositions.
ACM Trans. Algorithms 12(2), Art. 21, 41 (2016)
18. Lewis, J.M., Yannakakis, M.: The node-deletion problem for hereditary properties
is NP-complete. J. Comput. Syst. Sci. 20(2), 219230 (1980)
19. Oum, S.: Rank-width and vertex-minors. J. Combin. Theory Ser. B 95(1), 79100
(2005)
Busy Time Scheduling on a Bounded Number of
Machines (Extended Abstract)

Frederic Koehler1 and Samir Khuller2

1
Dept. of Mathematics, MIT, Cambridge MA 02139 USA
[email protected]
2
Dept. of Computer Science, Univ. of Maryland, College Park MD 20742 USA
[email protected]

Abstract In this paper we consider a basic scheduling problem called

the busy time scheduling problem - given n jobs, with release times rj ,
deadlines dj and processing times pj , and m machines, where each ma-
chine can run up to g jobs concurrently, our goal is to nd a schedule to
minimize the sum of the on times for the machines. We develop the
rst correct constant factor online competitive algorithm for the case
when g is unbounded, and give a O(log P ) approximation for general g,
where P is the ratio of maximum to minimum processing time. When
g is bounded, all prior busy time approximation algorithms use an un-
bounded number of machines; note it is NP-hard just to test feasibility
on xed m machines. For this problem we give both oine and online
(requiring lookahead) algorithms, which are O(1) competitive in busy
time and O(log P ) competitive in number of machines used.

1 Introduction

Scheduling jobs on multiple parallel machines has received extensive attention

in the computer science and operations research communities for decades (see
reference work [3]). For the most part, these studies have focused primarily
on job-related metrics such as minimum makespan, total completion time, ow
time, tardiness and maximum throughput. Our work is part of a line of recent
results working towards a dierent goal: energy eciency, in particular aiming
to minimize the total time that a machine must be turned on, its busy time
[4,12,8,11,15,5]. Equivalently, we seek to maximize the average load of machines
while they are powered on, assuming we are free to turn machines o when they
are idle. Note in this context we are concerned with multi-processing machines,
as for machines which process only one job at a time the load is either 0 or
1 always. This measure has been studied in an eort to understand energy-
related problems in cloud computing contexts; see e.g. [11,5,4] . The busy time
metric also has connections to several key problems in optical network design,
for example in minimizing the ber costs of Optical Add Drop Multiplexers

Full version: http://math.mit.edu/~fkoehler/busytime.pdf. Research supported
by CCF 1217890 and CNS 1262805.

Springer International Publishing AG 2017 521

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 521532, 2017.
DOI: 10.1007/978-3-319-62127-2_44
522 F. Koehler and S. Khuller

(OADMs) [8], and the application of busy time models to optical network design
has been extensively outlined in the literature [8,9,18,1].
Formally the problem is dened as follows: we are given a set of n jobs, and
job j has a release time of rj , a deadline dj and a processing time of pj (it is
assumed rj + pj dj ) and a collection of m multiprocessor machines with g
processors each. The signicance of processors sharing a machine is that they
share busy time: the machine is on if a single processor on the machine is active.
Each job j is assigned to the time window [sj , sj + pj ) on some machine mj . The
assignment must satisfy the following constraints:
1. Start times respect job release times and deadlines, i.e., [sj , sj +pj ) [rj , dj ).
2. At most g jobs are running at any time on any given machine. Formally, at
any time t and on any machine m, |{j|t [sj , sj + pj ), mj = m}| g.
The busy time of a machine is the duration for which the machine is pro-
cessing any non-zero number of jobs. The objective is to minimize the total sum
of busy times of all the machines. Formally, the objective function is

 
[sj , sj + pj )
i=0 j:mj =i

where measures the geometric length of a union of disjoint half intervals by

summing their individual lengths; e.g. ([1, 2)[3, 4)[3, 5)) = 3 i.e. is Lebesgue
measure. Note that this objective is constant if g = 1.
All previous algorithms (described below) for busy time are forced to make
the assumption that m = , because the number of machines required by the
schedules they generate can be as large as (n), i.e. worst-possible. Our primary
interest in this paper is in improving on this front. Thus our primary interest
will really be in the simultaneous optimization problem of generating schedules
whose performance is bounded in two objectives simultaneously: both the busy
time and the number of machines required by the schedule. The best known
approximation
$ algorithms for each of these objectives separately is 3 [5] and
O( log n/ log log n) [6]. We conjecture that there exist schedules which achieve
a O(1) approximation in both objectives. However, as it stands the O(1) machine
minimization problem by itself remains a major open problem in combinatorial
optimization, so such a result is out of reach for now. The main result of our paper
will show that we can at least construct such a schedule under the assumption
that log P is bounded by a constant, where P = maxi,j pj /pi .

1.1 Related Work

Winkler and Zhang [18] rst studied the interval job case of busy time schedul-
ing, i.e. when pj = dj rj , and showed that even the special case when g = 2
is NP-hard. Their work was motivated by a problem in optical communication
and assigning routing requests to wavelengths. Assuming that the number of
machines available is unbounded, Alicherry and Bhatia [1], and independently
Busy Time Scheduling on a Bounded Number of Machines (Extended Abstract) 523

Kumar and Rudra [13], developed approximation algorithms with an approx-

imation factor of 2 for the case of interval jobs. Being unaware of prior work
on this problem, subsequently, Flammini et al [8] developed a very simple 4
approximation via a greedy algorithm for the interval job case.

The rst constant factor approximation for the general problem, albeit on an
unbounded number of machines, was given by Khandekar et al [11]. They rst
design a dynamic programming based algorithm for the case when g = . This
schedule is then used to x the starting times of the jobs, and the resulting in-
stance of interval jobs is scheduled by the greedy algorithm of Flammini et al [8].
Despite the restriction mapping, the approximation factor of 4 is unchanged.
Since better approximation algorithms for the interval job case were available,
it is natural to attempt to use those instead. Sadly, the restriction mapping can
actually increase the cost of an optimal solution by a factor of 2, and so even if
we use these algorithms we do not get a better bound than 4 (see [5] for a tight
example). Chang et al [5] developed a 3-approximation algorithm by giving a
new interval packing algorithm. We conjecture that the correct answer for this
is 2, matching the interval job case.

Unfortunately, the number of machines used by all of these algorithms may

be as large as (n), even in the case when all jobs are equal length and released
at time rj = 0. This is because the g = reduction chooses start times oblivious
to the true value of g. One may hope to resolve this problem from the other dir-
ection, by adapting an algorithm for minimizing the number of machines used.
It is not dicult to get a O(log n) approximation algorithm$for this problem via
randomized LP rounding. The best known result is a O( log n/ log log n) ap-
proximation algorithm by Chuzhoy et al [6] which uses a sophisticated recursive
LP relaxation to the problem. Unfortunately, it appears to us quite dicult to
adapt these LP rounding methods to account for the cost of the nonlinear busy
time objective.

When g < , very strong lower bounds for online minimization of busy
time were given by Shalom et al [17]. They show that when g < , no online
algorithm can be better than g competitive algorithm against an online adaptive
adversary. It should be noted that their general online model is harder than the
one we consider; they have no notion of time, so in the online scenario they
envision the algorithm must be able to deal with jobs arriving in arbitrary order.
However, their proof of the lower bound does not need this additional power: it
releases jobs in left-to-right order.

Some recent work [7,10] claims a 2-competitive online algorithm when g = ,

but it is incorrect; see Fig. 1. Independently and simultaneously to us, Ren and
Tang [16] recently studied the online problem when g = as well (see next
section). They proved the same lower bound on the competitive ratio of this
problem as we do and gave a slightly worse upper bound, 4 + 2 2. They also
analyzed a version of the problem where job lengths are unknown to the scheduler
and proved a strong lower bound in this setting.
524 F. Koehler and S. Khuller

# = g, p =

# = g, p = 1

Figure 1: Counter-example to online algorithm of [7]. The optimal solution delays

all the exible unit jobs to the end and gets a busy time cost of 1 + g rather
than g. Setting = g1 gives the gap. The gure shows the schedule produced by
the online algorithm with a cost of g.

1.2 Our Contributions

We divide the results into sections depending on the exibility the algorithm has
with m, the number of machines. We begin with the classic busy time model,
where m = .
Our rst result is an online 5-competitive algorithm for the busytime prob-
lem when machine capacity is unbounded g = . In addition, we show
that against an adaptive online adversary
there is no online algorithm with
competitive ratio less than = (1 + 5)/2.
The previous result is extended to the general busy time problem with g <
, and we get a competitive ratio of O(log P ). No online algorithm for this
problem was previously known. In the online setting with lookahead of 2pmax
we can give a 12-competitive algorithm.
We then present our main results, concerned with simultaneous optimization of
busytime and number of machines used:
We present a constant-factor approximation algorithm for the busy time
problem with xed number of machines m, given the assumption of identical
length jobs pj = p.
We give the rst approximation algorithm for busy time scheduling with
a non-trivial bound on the number of machines used. More precisely, for
the simultaneous optimization problem we give a schedule which is 3 + -
log P
competitive on busy time and O( log(1+) ) competitive on machine usage for
 < 1.
We give an online algorithm with O(pmax ) lookahead in time, which remains
O(1)-competitive for busy time and O(log P ) competitive on machine usage.
We also give tradeo lower bounds which show the limits on the simultaneous
optimizability of these objectives; if we optimize solely for one objective (e.g.
machine usage), we may lose a factor of (g) in the other (e.g. busy time).

1.3 Preliminaries
We recall the following fundamental scheduling lemma. The interval graph of
a collection of half-intervals {[i , i )}ni=1 is the graph with vertices the half-
Busy Time Scheduling on a Bounded Number of Machines (Extended Abstract) 525

intervals, and an edge between two half-intervals I1 and I2 i I1 I2 = . The

interval graph is perfect, i.e.:
Proposition 1 Given a collection of half-open intervals {[i , i )}ni=1 there ex-
ists a k-coloring of the corresponding interval graph i for all t R,

|{i : [i , i )  t}| k. (1)

Proposition 2 The following are lower bounds on the optimum busy time:
1. The optimal busy timefor the same input instance with g = .
n
2. The load bound (1/g) j=1 pj .

We say a job is available at time t if rj t. It is often useful to refer to the latest

start time uj = dj p of a job. An interval job is one with no choice in start
time, i.e. j is an interval job when dj rj = pj . We dene an algorithm to be a
(r1 , r2 )-approximation if it generates a schedule using at most r1 mopt machines
and r2 busyOP T busy time, where mopt is the smallest number of machines for
which a feasible schedule exists, and busyOP T (or just OPT) is the minimum
busy time on an unbounded number of machines.

2 Online Busy Time Scheduling with an Unbounded

Capacity
2.1 The g = case, Upper Bound
We give a 5-competitive deterministic online algorithm for busy time scheduling
when g = . In this setting we may assign all jobs to a single machine so we
assume w.l.o.g. m = 1. Informally, the algorithm is quite simple: everytime we
hit a latest starting time uj of an unscheduled job j, we activate the machine
from time uj to time uj + 2pj and run all the jobs that t in this window. To
analyze this, we can pick an arbitrary optimal schedule, decompose its busy time
into connected components, and then bound the cost of our schedule by charging
the cost of running jobs to the connected components containing them.
In this section we will let T denote theactive time of our machine; all jobs
are processed during this active time, i.e. j [sj , sj + pj ) T . We also maintain
a set P of primary jobs but this is only for the purposes of the analysis. Note
that at each time t the algorithm uses only information about jobs released by
time t, so it is truly online.
Algorithm Doubler:
1. Let P = . Let T = .
2. For t = 0 to dmax :
(a) Let U be the set of unscheduled, available jobs at time t.
(b) Run every unscheduled job j s.t. [t, t + pj ) T ; remove j from U .
(c) If t = uj for some j U , then pick such a j with pj maximal and set
T = T [t, t + 2pj ) (activating the machine from time t to t + 2pj ). Let
P = P {j}.
526 F. Koehler and S. Khuller

(d) Run3 every unscheduled job j s.t. [t, t + pj ) T ; j is removed from U .

Suppose the algorithm fails to schedule a job j. Then at time uj the job was
available but was not scheduled; impossible because steps 2(c) ensures that
T [uj , uj + pj ) and so step 2(d) would necessarily schedule it. Thus the al-
gorithm schedules all jobs and, because we may trivially verify it respects rj , dj
constraints, produces a valid schedule. Henceforth sj refers to the start times
chosen by algorithm Doubler; the following proposition is immediate.

Proposition 3 Let T bethe resulting active time and P the resulting set of
primary jobs. Then T = jP [sj , sj + 2pj ) and for every j P , sj = uj .

Theorem 1. Algorithm Doubler is 5-competitive.

Proof. Fix an input instance (r j , dj , pj ) and an optimal oine schedule OP T

with start times sj . Let T = j [sj , sj + pj ) so (T ) is the busy time cost of
OP T . Let P be the set of primary jobs. Let P1 P consist of those jobs j in P
with [sj , sj + 2pj ) T and P2 = P \ P1 . By the Proposition,

  
(T ) = [sj , sj + 2pj ) [sj , sj + 2pj ) + [sj , sj + 2pj )
jP jP1 jP2


(T ) + 2pj . (2)
jP2

It remains to bound the cost incurred by jobs in P2 . Decompose T into connec-

ted components {C i }ki=1 so T = C 1 C k . The endpoints of C i are inf C i and
sup C i . Let J(C i ) be the set of jobs j with [sj , sj + pj ) C i . OP T schedules all
  k 
jobs so i J(C i ) is the set of all jobs, thus jP2 2pj = i=1 jP2 J(Ci ) 2pj .
We now claim that 
pj 2(C i ). (3)
jP2 J(Ci )

k
To show the claim, rst we index so {eij }j=1
i
= P2 J(Ci ), where ki = |P2 J(Ci )|,
k
and (s(eij ))j=1 i
is a monotonically increasing sequence.
Observation: reij s(ei1 ) for all j. Suppose for contradiction that reij > s(ei1 )
for some j. We know [s (eij ), s (eij )+peij ) C i , hence reij +peij sup C i . Because
[s(ei1 ), s(ei1 ) + 2p1 )  C i we know that s(ei1 ) + 2p1 sup Ci reij + peij . Thus
[reij , reij + peij ] [s(ei1 ), s(ei1 ) + 2p1 ) T . We see then that at time reij , step 2
(b) the algorithm must have scheduled job eij . Thus eij / P P2 J(Ci ), which
contradicts the denition of eij . By contradiction reij s(ei1 ) for all j.
3
Step 2 (b) and step 2(d) are both necessary. Consider an interval job released at
time 0 of length 2 and another at time 1 of length 4. Without step 2 (b) running at
time 1, the machine will be turned on from time 5 to 9 unecessarily.
Busy Time Scheduling on a Bounded Number of Machines (Extended Abstract) 527

Now it follows that peij > 2peij1 (for j 2): suppose otherwise, then be-
cause we know eij was available at step 2 (c) at t = s(eij1 ) s(ei1 ) reij ,
job eij must have been scheduled at t with eij1 and cannot have been added

to P . By contradiction, peij > 2peij1 hence by induction pei  > 2ki j peij . Now
k
ki ki jk  j 
i

(3) follows: j=1 peij j=1 2 = 2pei  2(C i ).

ip i
ek < peik j  =0 2 k
 k i i i

Thus jP2 2pj i=1 4(C i ) = 4(T ). Combining this with (2) proves the
theorem.
Obviously we could have dened
the above algorithm replacing 2 with any > 1,

however = 2 minimizes + i=0 i and is thus optimal.

2.2 g = , Online Lower Bounds

Proposition 4 No online algorithm (without lookahead) against an online ad-
aptive adversary has competitive ratio better than = 1+2 5 1.618.
Proof. Let 0 < < 1 be a constant to be optimized later. Fix 1 >  > 0 such
that = k where k Z. Here is the strategy for the adversary:
1. Release job A of length 1 available in [0, 3).
2. Until job A is started, at each t = n for n < k Z release a single job of
length  available in [t, t + ). (The  jobs are interval jobs.)
3. If job A was started at t = n, release a nal job of length 1 available in
[2, 3).
4. Otherwise if job A is still not started at time (k 1), release no more jobs.
In the case corresponding to step (3), the online schedule has busy time n+1+1
whereas the optimal oine schedule, which runs job A at time 2, has busy time
n+2
(n + 1) + 1. The ratio is thus (n+1)+1 +2 x+2
+1 because f (x) = x+1 is mono-
tonically decreasing for x > 0. In the case corresponding to step (4), the online
schedule has busy time at least (k1)+1 = +1 whereasthe oine schedule 
has busy time 1. Thus the competitive ratio is at least min 2+1 ,  + 1
and we may take the limit as  0. The positive solution to 2
+1 = + 1 is at

51 1+ 5
= 2 , and thus we get a lower bound of = 2 .

A similar proof also gives a weaker lower bound when the algorithm is granted
lookahead of O(pmax ). Let 0 < < 1. Release job A with a very large availability
span, and simultaneously release an interval job of length , i.e. a job with rj =
0, pj = , dj = . Without loss of generality the online algorithm either schedules
job A at time 0 or chooses not to schedule job A until after time . In the former
case, release a job of length 1 at the very end of job As availability window; in
the latter case, release no more jobs. The
lower bound on the competitive
ratio
now min{ 1+1 , 2
1+ }, optimized at = 2 1, giving a ratio of 2.
Proposition 5 An algorithm with lookahead a function of pmax has competitive
ratio at least 2 1.414.
528 F. Koehler and S. Khuller

2.3 General Case, g <

Combining with the bucketing algorithm given by Shalom et al [17] this gives
a O(log ppmax
min
)-competitive online algorithm for busy time scheduling. More pre-
cisely, because the cost of their algorithm is bounded by 4 times the weight of the
input jobs, and 1 times the g = lower bound, the approximation is 9 log ppmax
min
.
Running Algorithm Doubler oine and combining with the 3-approximation
of Chang et al [5] gives a fast 7-approximation to the optimal busy time schedule.
This is because the Greedy Tracking algorithm [5] takes as input a g = sched-
ule using busytime T and outputs a schedule with cost at most T + 2w(J)/g
T + 2OP T where w(J) denotes the total processing time of all jobs. Since
T 5OP T using our algorithm, the cost is bounded by 7OP T .
If we are willing to grant the online algorithm a lookahead of 2pmax then we
can get a constant factor online algorithm. We use our g = online algorithm to
determine the time placement of jobs; this algorithm requires no lookahead so we
now know the start time of jobs 2pmax ahead of the current time. We now run the
oine machine-assignment algorithm in windows of the form [kpmax , (k+2)pmax )
for k N. We can bound the cost of even k by 5OP T + 2w(J0 )/g where w(J0 ) is
the total processing time of jobs run in windows with even k; adding the matching
term for odd k shows that this gives a fast 2 5 + 2 = 12 approximation.

3 Oine Algorithm for Equal length jobs, Bounded

Number of Machines

Although it is impossible in the general case (see Lower Bounds, Section 6), in the
case of pj = p we are able to compute a schedule which is (1, O(1))-approximate,
i.e. with the optimal number of machines and O(1) busy time vs. the busy time
optimum. Proposition 7 shows that a (O(1), 1)-approximation is impossible to
achieve, even in this scenario. Our algorithm is two-step: it starts with a feasible
schedule, and then uses a pushing scanline to push jobs together and reduce
the busytime cost.
Algorithm Compact

1. Find the minimum number of machine required to feasibly schedule the jobs
by binary search (0 mopt n), using a feasibility algorithm for the problem
with mg identical single-processor machines. Then construct a schedule on
 jobs and2 mopt machines that minimizes the sum of completion
S on these
times, Cj . A O(n ) time algorithm for these tasks is known [14].
2. Let s0j be the start time of job j in S, and let sj := s0j , K := and P := .
3. For t from rmin to dmax : (main loop)
(a) For every unscheduled job j, let sj := max{s0j , t}. Let U be the set of
unscheduled jobs.
(b) If |{j U : sj [t, t + 2p]}| mg, run each job j in this set at time
sj . Let K := K {[t, t + 3p)}. We say these jobs were run in a cluster.
Return to the main loop at t := t + 2p.
Busy Time Scheduling on a Bounded Number of Machines (Extended Abstract) 529

(c) Otherwise if t = uj for some unscheduled job j, run each job in the set
{j U : sj [t, t + p]} at its start time sj . Return to the main loop at
t := t + p. Let P := P {j}.

In step 3 it is necessary to consider only t {uj , sj 2p}, so we can run this

step in O(n log n) time.

Theorem 2. Algorithm Compact is a 6-approximation for busy time, and gen-

erates a feasible schedule using mopt machines.

4 Oine Algorithm for Bounded Number of Machines

In this section we will use the fact that scheduling jobs on a minimum number of
machines with integer release times and deadlines and with p = pj = 1 is trivial
oine. For a xed m, it is well-known that an EDF (earliest-deadline rst)
schedule, i.e. one given by running at each time up to m of the jobs with earliest
deadlines, gives a feasible schedule i the input instance is feasible. Computing
the minimum m can be done by binary search in log n steps.
We would like to describe some intuition before launching into the formal
analysis. As before, we use something like a pushing scanline approach, mov-
ing jobs rightward from a left-shifted schedule and starting a group of jobs
whenever a large number have been pushed together. To make this approach
have bounded busy time usage, we rst need to bucket jobs by similar lengths,
but this alone cannot attain our desired performance ratio, because we may need
for busy time purposes to group some long jobs with some short jobs. There-
fore, in each bucket, when a job does not nicely group together with other jobs
of the same length, we temporarily drop it. A second clean-up pass (step 3
below) runs the remaining jobs using an algorithm which has good busy-time
performance but a priori unbounded machine usage. By arguing that we drop few
jobs with overlapping availability times from each bucket, it is then possible to
bound the machine usage. Below is our (O(log pmax /pmin ), O(1))-approximation
algorithm for the general problem. Fix a constant > 1 to be optimized later.

1. Bucket jobs by processing time increasing exponentially by , so the buckets

2
contain jobs of processing time in the
 intervals [pmin , pmin ), [pmin , pmin ), . . . ,
[q1 pmin , q pmin ] where q = log ppmax min
.
2. For each bucket Bi
(a) Let p be the supremum of the processing times of jobs in this bucket.
We round job availability constraints down to multiples of p, so rj =
prj /p, uj = puj /p, and pj = p. This is a unit job scheduling problem
after we rescale by a factor of p.
(b) We generate a left-shifted feasible schedule (referred to as the initial
schedule) for the rounded (rj , dj , pj ) instance using the minimum number
of machines m. Let s0j be the start time of job j in this schedule.
(c) Execute Algorithm RunHeavy.
530 F. Koehler and S. Khuller

(d) Let Ui denote the set of jobs unscheduled in Bi after running Algorithm
RunHeavy.

3. Now let U  be the union of the Ui for all buckets, and schedule the jobs in
U  by the 3-approximation of Chang et al [5] upon a new set of machines.

Algorithm RunHeavy
1. Let U initially be the set of all jobs in the bucket. Split machines into groups
M1 and M0 ; we will require at most m machines in each group (see analysis).

2. For t = kp from rmin to umax :
(a) Let Jt = {j U : s0j = t}. Let k1 = |Jt |/g and run k1 groups of g
jobs from this set on the group of machines Mk mod 2 with start times
sj = max(s0j , rj ). Remove these jobs from U .
(b) Let Jt = {j U : s0j t uj }. Let k2 = |Jt |/g jobs, and run k2
groups of g jobs from this set on the group of machines Mk mod 2 with
start times sj = max(s0j , rj ). Remove these jobs from U .
Note in the loop in RunHeavy, we only need to do something when t = s0j for
some job j so the loop is really over polynomially many t.
Theorem 3. The above algorithm generates a schedule feasible using (2 +
1)OP T busy time on log pmax /pmin (2 mopt + 8) machines.

5 Online Algorithm for Bounded Number of Machines

Since the formal details in this section are quite long, we give a brief summary of
the main idea. In order to get an online algorithm, we still use the approach of
the previous section, but interweave an agressive variant of Algorithm Doubler in
order to pick start times for the leftover jobs which t poorly into their buckets.
In the previous section we could use that the (rj , dj , p = pj = 1) problem was
exactly solvable oine; now, we must instead rely upon the online e-competitive
online algorithm of [2] for this task. We also must use our g = algorithm in
order to schedule the jobs in U  online with bounded performance.
Theorem 4. The online algorithm with lookahead 3pmax generates a schedule
requiring at most log pmax /pmin (16 + 4e mopt ) machines and (20 + 42)
busy time.

6 Simultaneous Optimization of Busy Time and Machine

Usage
6.1 Lower Bounds
Proposition 6 For any input g, there exist input instances (with g processors
per machine) where every machine-optimal schedule uses (g)busyopt busy time
for  arbitrary small.
Busy Time Scheduling on a Bounded Number of Machines (Extended Abstract) 531

Proof. Fix 1 > > 0. Release g jobs of length 1 at time 0 with availability
windows [0, g). For k = 0 to g 1, release g 1 jobs of length with availability
windows [k, k + ), and g 1 jobs of length with availability windows [k + 1
, k + 1). The machine-optimal schedule runs all jobs on one machine, but due
to the presence of the -jobs cannot overlap the execution of the long jobs, and
thus has busy time cost g (see Fig. 2 ). The busy time optimal schedule runs the
jobs on a separate machine and runs all of the long jobs in parallel, giving a
g
busy time cost 1 + 2g. Thus the ratio is 1+2g and taking suciently small
gives the desired result.

p =1 p =1
#=g #=g

p =
p =

#=g - 1 #=g - 1

Figure 2: Illustrations of trade-o lower bounds. On one machine, the busy time
is g. On two machines the busy time is 1 + 2g.

Proposition 7 For any g, there exist input instances where every busy time
optimal schedule uses gmopt machines, even with the restriction pj = p.

Proof. We set p = pj = 1 for all jobs. For k = 0 to g 1, we release an interval

job with availability window [k/g, k/g +1), and we release g(g 1) unconstrained
jobs with availability windows [0, 2g 2 ).
There exists a busy time optimal schedule using g machines, which runs g 1
unconstrained jobs along with a single interval job together on a machine. Here
the busy time cost equals the load bound exactly. There exists a feasible schedule
using only 1 machine: for k = 0 to g 1, on processor k of the machine it runs
rst the interval job followed by g 1 unconstrained jobs, end-to-end. Thus
mopt = 1.
Now consider any schedule using fewer than g machines. By the pigeonhole
principle, it must run two interval jobs on a single machine M . Let these jobs
start at k1 /g and k2 /g respectively with k1 < k2 ; then the processor running the
job at k2 /g must be idle in [0, k2 /g) [k1 /g, k2 /g). Since the load is positive but
below g in this interval, the busy time exceeds the busy time lower bound, and
so is greater than the cost of the busy time optimal schedule described earlier.

Acknowledgements: We are grateful to Chunxing Yin for extremely useful

discussions.
532 F. Koehler and S. Khuller

References
1. Mansoor Alicherry and Randeep Bhatia. Line system design and a generalized
coloring problem. In ESA, pages 1930, 2003.
2. Nikhil Bansal, Tracy Kimbrel, and Kirk Pruhs. Speed scaling to manage energy
and temperature. J. ACM, 54(1):3:13:39, March 2007.
3. Peter Brucker. Scheduling algorithms. Springer, 2007.
4. Jessica Chang, Harold Gabow, and Samir Khuller. A model for minimizing active
processor time. Algorithmica, 70(3):368405, November 2014.
5. Jessica Chang, Samir Khuller, and Koyel Mukherjee. Lp rounding and combinat-
orial algorithms for minimizing active and busy time. In SPAA, pages 118127.
ACM, 2014.
6. Julia Chuzhoy, Sudipto Guha, Sanjeev Khanna, and Joseph Se Naor. Machine
minimization for scheduling jobs with interval constraints. In FOCS, pages 8190.
IEEE, 2004.
7. Xiaolin Fang, Hong Gao, Jianzhong Li, and Yingshu Li. Application-aware data
collection in wireless sensor networks. In Proceedings of INFOCOM, 2013.
8. Michele Flammini, Gianpiero Monaco, Luca Moscardelli, Hadas Shachnai, Morde-
chai Shalom, Tami Tamir, and Shmuel Zaks. Minimizing total busy time in parallel
scheduling with application to optical networks. In IPDPS, pages 112, 2009.
9. Michele Flammini, Gianpiero Monaco, Luca Moscardelli, Mordechai Shalom, and
Shmuel Zaks. Approximating the trac grooming problem with respect to adms
and oadms. In Proceedings of Euro-Par, pages 920929, 2008.
10. Chi Kit Ken Fong, Minming Li, Shi Li, Sheung-Hung Poon, Weiwei Wu, and
Yingchao Zhao. Scheduling tasks to minimize active time on a processor with
unlimited capacity. In MAPSP, 2015.
11. Rohit Khandekar, Baruch Schieber, Hadas Shachnai, and Tami Tamir. Minimizing
busy time in multiple machine real-time scheduling. In FSTTCS, pages 169 180,
2010.
12. Frederic Koehler and Samir Khuller. Optimal batch schedules for parallel machines.
In WADS, pages 475486, 2013.
13. Vijay Kumar and Atri Rudra. Approximation algorithms for wavelength assign-
ment. In FSTTCS, pages 152163, 2005.
14. Alejandro Lopez-Ortiz and Claude-Guy Quimper. A fast algorithm for multi-
machine scheduling problems with jobs of equal processing times. In STACS,
pages 380391, 2011.
15. George B. Mertzios, Mordechai Shalom, Ariella Voloshin, Prudence W.H. Wong,
and Shmuel Zaks. Optimizing busy time on parallel machines. In IPDPS, pages
238248, 2012.
16. Runtian Ren and Xueyan Tang. Online exible job scheduling for minimum span.
In SPAA, 2017. To Appear.
17. Mordechai Shalom, Ariella Voloshin, Prudence W.H. Wong, Fencol C.C. Yung,
and Shmuel Zaks. Online optimization of busy time on parallel machines. TAMC,
7287:448460, 2012.
18. Peter Winkler and Lisa Zhang. Wavelength assignment and generalized interval
graph coloring. In SODA, pages 830 831, 2003.
 Algorithms for Covering Multiple Barriers

Shimin Li and Haitao Wang

Department of Computer Science

Utah State University, Logan, UT 84322, USA
[email protected],[email protected]

Abstract. In this paper, we consider the problems for covering multiple

intervals on a line. Given a set B of m line segments (called barriers)
on a horizontal line L and another set S of n horizontal line segments
of the same length in the plane, we want to move all segments of S to
L so that their union covers all barriers and the maximum movement
of all segments of S is minimized. Previously, an O(n3 log n)-time algo-
rithm was given for the problem but only for the special case m = 1.
In this paper, we propose an O(n2 log n log log n + nm log m)-time algo-
rithm for any m, which improves the previous work even for m = 1. We
then consider a line-constrained version of the problem in which the seg-
ments of S are all initially on the line L. Previously, an O(n log n)-time
algorithm was known for the case m = 1. We present an algorithm of
O((n + m) log(n + m)) time for any m. These problems may have appli-
cations in mobile sensor barrier coverage in wireless sensor networks.

1 Introduction
In this paper, we study algorithms for covering multiple barriers. These are basic
geometric problems and have applications in barrier coverage of mobile sensors
in wireless sensor networks. For convenience, in the following we introduce and
discuss the problems from the mobile sensor barrier coverage point of view.
Let L be a line, say, the x-axis. Let B be a set of m pairwise disjoint segments,
called barriers, sorted on L from left to right. Let S be a set of n sensors in the
plane, and each sensor si S is represented by a point (xi , yi ). If a sensor is
moved on L, it has a sensing/covering range of length r, i.e., if a sensor s is
located at x on L, then all points of L in the interval [x r, x + r] are covered by
s and the interval is called the covering interval of s. The problem is to move all
sensors of S onto L such that each point of every barrier is covered by at least
one sensor and the$maximum movement of all sensors of S is minimized, i.e.,
the value maxsi S (xi xi )2 + yi2 is minimized, where xi is the location of si
on L in the solution (its y-coordinate is 0 since L is the x-axis). We call it the
multiple-barrier coverage problem, denoted by MBC.
We assume that covering range of the sensors is long enough so that a cover-
age of all barriers is always possible. Note that we can check whether a coverage
is possible in O(m + n) time by an easy greedy algorithm.

This research was supported in part by NSF under Grant CCF-1317143.

Springer International Publishing AG 2017 533

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 533544, 2017.
DOI: 10.1007/978-3-319-62127-2_45
534 S. Li and H. Wang

Previously, only the special case m = 1 was studied and the problem was
solved in O(n3 log n) time [10]. In this paper, we propose an O(n2 log n log log n+
nm log m)-time algorithm for any value m, which improves the algorithm in [10]
by almost a linear factor even for the case m = 1.
We further consider a line-constrained version of the problem where all sen-
sors of S are initially on L. Previously, only the special case m = 1 was stud-
ied and the problem was solved in O(n log n) time [3]. We present an O((n +
m) log(n + m)) time algorithm for any value m, and the running time matches
that of the algorithm in [3] when m = 1.

1.1 Related Work

Sensors are basic units in wireless sensor networks. The advantage of allowing
the sensors to be mobile increases monitoring capability compared to those static
ones. One of the most important applications in mobile wireless sensor networks
is to monitor a barrier to detect intruders in an attempt to cross a specic region.
Barrier coverage [9, 10], which guarantees that every movement crossing a barrier
of sensors will be detected, is known to be an appropriate model of coverage for
such applications. Mobile sensors have limited battery power and therefore their
movements should be as small as possible.
Dobrev et al. [7] studies several problems on covering multiple barriers in the
plane. They showed that these problems are generally NP-hard when sensors
have dierent ranges. They also proposed polygonal-time algorithms for some
special cases, e.g., barriers are parallel or perpendicular to each other, and sensors
have some constrained movements. In fact, if sensors have dierent ranges, by
an easy reduction from the Partition Problem as in [7], we can show that our
problem MBC is NP-hard even for the line-constrained version and m = 2.
Other previous work has been focused on the line-constrained problem with
m = 1. Czyzowicz et al. [5] rst gave an O(n2 ) time algorithm, and later, Chen
et al. [3] solved the problem in O(n log n) time. If sensors have dierent ranges,
Chen et al. [3] presented an O(n2 log n) time algorithm. For the weighted case
where sensors have weights such that the moving cost of a sensor is its moving
distance times its weight, Wang and Zhang [15] gave an O(n2 log n log log n) time
algorithm for the case where sensors have the same range.
The min-sum version of the line-constrained problem with m = 1 has also
been studied, which is to minimize the sum of the moving distances of all sensors.
If sensors have dierent ranges, the problem is NP-hard [6]. Otherwise, Czyzow-
icz et al. [6] gave an O(n2 ) time algorithm, and Andrews and Wang [1] solved
the problem in O(n log n) time. The min-num version of the problem was also
studied, where the goal is to move the minimum number of sensors to form a
barrier coverage. Mehrandish et al. [13, 14] proved that the problem is NP-hard
if sensors have dierent ranges and gave polynomial time algorithms otherwise.
Bhattacharya et al. [2] studied a circular barrier coverage problem in which
the barrier is a circle and the sensors are initially located inside the circle. The
goal is to move sensors to the circle to form a regular n-gon (so as to cover the
Algorithms for Covering Multiple Barriers 535

circle) such that the maximum sensor movement is minimized. An O(n3.5 log n)-
time algorithm was given in [2] and later Chen et al. [4] improved the algorithm
to O(n log3 n) time. The min-sum version of the problem was also studied [2, 4].

1.2 Our Approach

To solve MBC, one major diculty is that we do not know the order of the
sensors of S on L in an optimal solution. Therefore, our main eort is to nd such
an order. To this end, we rst develop a decision algorithm that can determine
whether for any value , where is the maximum sensor movement in
an optimal solution. Our decision algorithm runs in O(m + n log n) time. Then,
we solve the problem MBC by parameterizing the decision algorithm in a way
similar in spirit to parametric search [12]. The high-level scheme of our algorithm
is very similar to those in [3, 15], but many low-level computations are dierent.
The line-constrained problem is easier due to an order preserving property:
there exists an optimal solution in which the order of the sensors is the same
as in the input. This leads to a linear-time decision algorithm using the greedy
strategy. Also based on this property, we can nd a set of O(n2 m) candidate
values such that contains . To avoid computing explicitly, we implicitly
organize the elements of into O(n) sorted arrays such that each array element
can be found in O(log m) time. Finally, by applying the matrix search technique
in [8], along with our linear-time decision algorithm, we compute in O((n +
m) log(n+m)) time. We should point out that implicitly organizing the elements
of into sorted arrays is the key and also the major diculty for solving the
problem, and our technique may be interesting in its own right.
The remaining paper is organized as follows. In Section 2, we introduce some
notation. In Section 3, we present our algorithm for the line-constrained problem.
In Section 4, we present our decision algorithm for MBC. Section 5 solves the
problem MBC. Section 6 concludes the paper, with remarks that our techniques
can be used to reduce the space complexities of the algorithms in [3, 15]. Due to
the space limit, some proofs are omitted but can be found in the full paper [11].

2 Preliminaries

We denote the barriers of B by B1 , B2 , . . . , Bm sorted on L from left to right.

For each Bi , let ai and bi denote the left and right endpoints of Bi , respectively.
For ease of exposition, we make a general position assumption that ai = bi for
each Bi . The degenerated case can also be handled by our techniques, but the
discussions would be more tedious.
For any point x on L (the x-axis), we also use x to denote its x-coordinate,
and vice versa. We assume that the left endpoint of B1 is at 0, i.e., a1 = 0. Let
denote the right endpoint of Bm , i.e., = bm .
We denote the sensors of S by s1 , s2 , . . . , sn sorted by their x-coordinates.
For each sensor si located on a point x of L, x r and x + r are the left and
536 S. Li and H. Wang

right endpoints of the covering interval of si , respectively, and we call them the
left and right extensions of si , respectively.
Again, let be the maximum sensor movement in an optimal solution.
Given any value , the decision problem is to determine whether , or
equivalently, whether we can move each sensor with distance at most such
that all barriers can be covered. If yes, we say that is a feasible value. Thus,
we also call it a feasibility test on .

3 The Line-Constrained Version of MBC

In this section, we present our algorithm for the line-constrained MBC. As in the
special case m = 1 [5], a useful observation is that the following order preserving
property holds: There exists an optimal solution in which the order of the sensors
is the same as in the input. Due to this property, we have the following lemma.

Lemma 1. Given any > 0, we can determine whether is a feasible value in

O(n + m) time.

Let OP T be an optimal solution that preserves the order of the sensors. For
each i [1, n], let xi be the position of si in OP T . We say that a set of k sensors
are in attached positions if the union of their covering intervals is a single interval
of length equal to 2rk. The following lemma is self-evident and is an extension
of a similar observation for the case m = 1 in [5].

Lemma 2. There exists a sequence of sensors si , si+1 , . . . , sj in attached posi-

tions in OP T such that one of the following three cases holds. (a) The sensor
sj is moved to the left by distance and xi = ak + r for some barrier Bk (i.e.,
the sensors from si to sj together cover the interval [ak , ak + 2r(j i + 1)]). (b)
The sensor si is moved to the right by and xj = bk r for some barrier Bk .
(c) The sensor si is moved rightwards by and sj is moved leftwards by .

Cases (a) and (b) are symmetric in the above lemma. Let 1 be the set of
all possible distance values introduced by sj in Case (a). Specically, for any
pair (i, j) with 1 i j n and any barrier Bk with 1 k m, dene
(i, j, k) = xj (ak + 2r(j i) + r). Let 1 consists of (i, j, k) for all such
triples (i, j, k). We dene 2 symmetrically be the set of all possible values
introduced by si in Case (b). We dene 3 as the set consisting of the values
[xj xi 2r(j i)]/2 for all pairs (i, j) with 1 i < j n. Clearly, |3 | = O(n2 )
and both |1 | and |2 | are O(mn2 ). Let = 1 2 3 .
By Lemma 2, is in and is actually the smallest feasible value of .
Hence, we can rst compute and then nd the smallest feasible value in by
using the decision algorithm. However, that would take (mn2 ) time. To reduce
the time, we will not compute explicitly, but implicitly organize the elements
of into certain sorted arrays and then apply the matrix search technique in [8].
Since we only need to deal with sorted arrays instead of more general matrices,
we review the technique with respect to arrays in the following lemma.
Algorithms for Covering Multiple Barriers 537

Lemma 3. [8] Given a set of N sorted arrays of size at most M each, we

can compute the smallest feasible value of these arrays with O(log N + log M )
feasibility tests and the total time of the algorithm excluding the feasibility tests
is O( N log 2M
N ), where is the time for evaluating each array element (i.e.,
the number of array elements that need to be evaluated is O(N log 2M N )).

With Lemma 3, we can compute the smallest feasible values in the sets 1 ,
2 , and 3 , respectively, and then return the smallest one as . For 3 , Chen et
al. [3] (see Lemma 14) gave an approach to order in O(n log n) time the elements
of 3 into O(n) sorted arrays of O(n) elements each such that each array element
can be obtained in O(1) time. Consequently, by applying Lemma 3, the smallest
feasible value of 3 can be computed in O((n + m) log n) time.
For the other two sets 1 and 2 , in the case m = 1, the elements of each set
can be easily ordered into O(n) sorted arrays of O(n) elements each [3]. However,
in our problem for general m, it becomes signicantly more dicult to obtain a
subquadratic-time algorithm. Indeed, this is the main challenge of our method.
In what follows, our main eort is to prove Lemma 4.

Lemma 4. For the set 1 , in O(m log m) time, we can implicitly form a set
A of O(n) sorted arrays of O(m2 n) elements each such that each array element
can be computed in O(log m) time and every element of 1 is contained in one
of the arrays. The same applies to the set 2 .

We note that our technique for Lemma 4 might be interesting in its own right
and may nd other applications as well. Before proving Lemma 4, we rst prove
the following theorem by using Lemma 4.

Theorem 1. The line-constrained version of MBC can be solved in O((n +

m) log(n + m)) time.

Proof. It is sucient to compute , after which we can apply the decision

algorithm on to obtain an optimal solution.
Let 1 denote the set of all elements in the arrays of A specied in Lemma 4.
Dene 2 similarly with respect to 2 . By Lemma 4, 1 1 and 2 2 .
Since 1 2 3 , we also have 1 2 3 . Hence, is the
smallest feasible value in 1 2 3 . Let 1 , 2 , and 3 be the smallest
feasible values in the sets 1 , 2 , and 3 , respectively. As discussed before, 3
can be computed in O((n + m) log n) time. By Lemma 4, applying the algorithm
in Lemma 3 can compute both 1 and 2 in O((n + m)(log m + log n)) time.
Note that (n + m)(log m + log n) = ((n + m) log(n + m)).


3.1 Proving Lemma 4

In this section, we prove Lemma 4. We will only prove the case for the set 1 ,
since the other case for 2 is symmetric. Recall that 1 = {(i, j, k) | 1 i
j n, 1 k m}, where (i, j, k) = xj (ak + 2r(j i) + r).
538 S. Li and H. Wang

For any j and k, let A[j, k] denote the list (i, j, k) for i = 1, 2, . . . , j, which
is sorted increasingly. Let A[j]denote the union of the elements in A[j, k] for
n
all k [1, m]. Clearly, 1 = j=1 A[j]. In the following, we will organize the
elements in each A[j] into a sorted array B[j] of size O(nm2 ) such that given any
index t, the t-th element of B[j] can be computed in O(log m) time, which will
prove Lemma 4. Our technique replies on the following property: the dierence
of every two adjacent elements in each list A[j, k] is the same, i.e., 2r.
Notice that for any k [1, m 1], the rst (resp., last) element of A[j, k] is
larger than the rst (resp., last) element of A[j, k + 1]. Hence, the rst element
of A[j, m], i.e., (1, j, m), is the smallest element of A[j] and the last element of
A[j, 1], i.e., (j, j, 1), is the largest element of A[j]. Let min [j] = (1, j, m) and
max [j] = (j, j, 1).
For each k [1, m], we extend the list A[j, k] to a new sorted list B[j, k]
with the following property: (1) A[j, k] is a sublist of B[j, k]; (2) the dier-
ence every two adjacent elements of B[j, k] is 2r; (3) the rst element of B[j, k]
is in [min [j], min [j] + 2r); (4) the last element of B[j, k] is in (max [j]
2r, max [j]]. Specically, B[j, k] is dened as follows. Note that (1, j, k) and
(j, j, k) are the rst and last elements of A[j, k], respectively. We let (1, j, k)
min [j]

(1,j,k)
2r  2r and (j, j, k) +
max [j](j,j,k)
2r  2r be the rst and last
elements of B[j, k], respectively. Then, the h-th element of B[j, k] is equal to
min [j]
(1, j, k)
(1,j,k)2r  2r + 2r (h 1) for any h [1, [j]], where [j] =
max [j]min [j]
1+ 2r . Hence, B[j, k] has [j] elements. One can verify that B[j, k]
has the above four properties. Note that we can implicitly create the lists B[j, k]
in O(1) time so that given any k [1, m] and h [1, [j]], we can obtain the
h-th element of B[j, k] in O(1) time. Let B[j] be the sorted list of all elements
of B[j, k] for all 1 k m. Hence, B[j] has [j] m elements.
Let j be the permutation of 1, 2, . . . , m following the sorted order of the
rst elements of B[j, k]. For any k [1, m], let j (k) be the k-th index in j .
Lemma 5. For any t with 1 t [j] m, the t-th smallest element of B[j] is
the ht -th element of the list B[j, j (kt )], where ht =  m
t
and kt = t mod m.
By Lemma 5, if j is known, we can obtain the t-th smallest element of B[j]
in O(1) time for any t. Computing j can be done in O(m log m) time by sorting.
If we do the sorting for every j [1, n], then we wound need O(nm log m) time.
Fortunately, Lemma 6 implies that we only need to do the sorting once.
Lemma 6. The permutation j is unique for all j [1, n].
In summary, after O(m log m) time preprocessing to compute j for any j,
we can form the arrays B[j] for all j [1, n] such that given any j [1, n]
and t [1, [j] m], we can compute t-th smallest element of B[j] in O(1)
time. However, we are not done yet, because we do not have a reasonable upper
bound for [j], which is equal to 1 +  max [j]
2r
min [j]
= 1 +  (j,j,1)(1,j,m)
2r =
am a1
j +  2r . To address the issue, in the sequel, we will partition the indices
k [1, m] into groups and then apply our above approach to each group so that
the corresponding [j] values can be bounded, e.g., by O(mn).
Algorithms for Covering Multiple Barriers 539

The Group Partition Technique We consider any index j [1, m].

We partition the indices 1, 2, . . . , m into groups each consisting of a sequence
of consecutive indices, such that each group has an intra-group overlapping prop-
erty: For any index k that is not the largest index in the group, the rst ele-
ment of A[j, k] is smaller than or equal to the last element of A[j, k + 1], i.e.,
(1, j, k) (j, j, k + 1). Further, the groups have the following inter-group non-
overlapping property: For the largest index k in a group that is not the last
group, the rst element of A[j, k] is larger than the last element of A[j, k + 1],
i.e., (1, j, k) > (j, j, k + 1).
We compute the groups in O(m) time as follows. Initially, add 1 into the rst
group G1 . Let k = 1. While the rst element of A[j, k] is smaller than or equal
to the last element of A[j, k + 1], we add k + 1 into G1 and reset k = k + 1. After
the while loop, G1 is computed. Then, starting from k + 1, we compute G2 and
so on until index m is included in the last group. Let G1 , G2 , . . . , Gl be the l
groups we compute. Note that l m.
Consider any group Gg with 1 g l. We process the lists A[j][k] for all
k Gg in the same way as discussed before. Specically, for each k Gg , we
create a new list B[j][k] from A[j][k]. Based on the new lists in the group Gg ,
we form the sorted array Bg [j] with a total of |Gg | g [j] elements, where |Gg |
is the number of indices of Gg and g [j] is corresponding [j] value as dened
before but only on the group Gg , i.e., if k1 and k2 are the smallest and largest
indices of Gg respectively, then g [j] = 1 +  (j,j,k1 )(1,j,k
2r
2)
. Let B[j] be the
sorted list of all elements in the lists Bg [j] for all groups. Due to the intra-group
overlapping property of each group, it holds that g |Gg | n. Thus, the size of
l l
B[j] is at most g=1 |Gg |2 n, which is at most m2 n since g=1 |Gg | = m.
Suppose we want to nd the t-th smallest element of B[j]. As preprocess-
ing,
g we compute a sequence of values g [j] for g = 1, 2, . . . , l, where g [j] =
g =1 g [j] |Gg |, in O(m) time. To compute the t-th smallest element of
B[j], we rst do binary search on the sequence 1 [j], 2 [j], . . . , l [j] to nd in
O(log l) time the index g such that t (g1 [j], g [j]]. Due to the inter-group
non-overlapping property of the groups, the t-th smallest element of B[j] is the
(t g1 [j])-th element in the array Bg [j], which can be found in O(1) time. As
l m, the total time for computing the t-th smallest element of B[j] is O(log m).
The above discussion is on any single index j [1, n]. With O(m log m) time
preprocessing, given any t, we can nd the t-th smallest value of B[j] in O(log m)
time. For all indices j [1, n], it appears that we have to do the group partition
for every j [1, n], which would take quadratic time. To resolve the issue, we
show that it suces to only use the group partition based on j = n for all other
j [1, n 1]. The details are given below.
Suppose from now on G1 , G2 , . . . , Gl are the groups computed as above with
respect to j = n. We know that the inter-group non-overlapping property holds
respect to the index n. The following lemma shows that the property also holds
with respect to any other index j [1, n 1].

Lemma 7. The inter-group non-overlapping property holds for any j [1, n1].
540 S. Li and H. Wang

Consider any Gg with 1 g l and any j [1, n]. For each k Gg , we

create a new list B[j][k] based on A[j][k] in the same way as before. Based on
the new lists, we form the sorted array Bg [j] of |Gg | g [j] elements. We also
dene the value g [j] in the same way as before. Lemma 8 shows that g [j] and
g [j] can be computed from g [n] and g [n].

Lemma 8. For any j [1, n 1] and g [1, l], g [j] = g [n] n + j and
g
g [j] = g [n] + g g (j n), where g = g =1 |Gg |.

For each group Gg , we compute the permutation for the lists B[n, k] for all k
in the group. Computing the permutations for all groups takes O(m log m) time.
Also as preprocessing, we rst compute g , g (n) and g (n) for all g [1, l] in
O(m) time. By Lemma 8, for any j [1, n] and any g [1, l], we can compute
g [j] and g [j] in O(1) time. Because the lists B[n, k] for all k in each group Gg
have the intra-group overlapping property, it holds that g [n] |Gg | n. Hence,
l
g [n] mn. For any j [1, n 1], by Lemma 8, g [j] < g [n], and thus
g=1
l 2
g=1 g [j] mn. Note that B[j] has at most m n elements.
l
For any j [1, n] and any t [1, g=1 |Gg | g [j]], to compute the t-
th smallest element of B[j], due to the inter-group non-overlapping property in
Lemma 7, we can still use the previous binary search approach. As we can obtain
each g [j] for any g [1, l] in O(1) time by Lemma 8, we can still compute the
t-th smallest element of B[j] in O(log m) time. This proves Lemma 4.

4 The Decision Problem of MBC

In this section, we present an O(m + n log n)-time algorithm for the decision
problem of MBC: given any value > 0, determine whether . Our al-
gorithm for MBC in Section 5 will make use of this decision algorithm. The
decision problem may have independent interest because in some applications
each sensor has a limited energy and we want to know whether their energy
is enough for them to move to cover all barriers.
Consider any value > 0. We assume max1in |yi | since otherwise
some sensor cannot reach L by$moving (and thus is $not feasible). For any
sensor si S, dene xri = xi + 2 yi2 and xli = xi 2 yi2 . Note that xri
and xli are respectively the rightmost and leftmost points of L si can reach with
respect to . We call xri the rightmost (resp., leftmost) -reachable location of si
on L. For any point x on L, we use p+ (x) to denote a point x such that x > x
and x is innitesimally close to x. The high-level scheme of our algorithm is
similar to that in [15]. Below we describe the algorithm.
We use a conguration to refer to a specication on where each sensor si S
is located. For example, in the input conguration, each si is at (xi , yi ). We rst
move each sensor si to xri on L. Let C0 denote the resulting conguration. In
$ to move rightwards but can move leftwards on
C0 , each sensor si is not allowed
L by a maximum distance 2 2 yi2 .
Algorithms for Covering Multiple Barriers 541

sg(i)

x
Ri1 x Ri1

Fig. 1. Illustrating the set Si1 . Fig. 2. Illustrating the set Si2 . The segments are
The covering intervals of sensors the covering intervals of sensors. The red thick
are shown with segments (the red segments correspond to the sensors in Si2 . The
thick segments correspond to the four black points corresponding to the values xlk
sensors in Si1 ). Every sensor in r of the four sensors xk to the right of Ri1 . The
Si1 can be sg(i) . sensor sg(i) is labeled.

If , our algorithm will compute a subset of sensors with their new

locations to cover all barriers of B and the maximum movement of each sensor
of in the subset is at most .
For each step i with i 1, let Ci1 be the conguration right before the i-th
step. Our algorithm maintains the following invariants. (1) We have a subset of
sensors Si1 = {sg(1) , sg(2) , . . . , sg(i1) }, where for each 1 j i 1, g(j) is
the index of the sensor sg(j) in S. (2) In Ci1 , each sensor sk of Si1 is at a
new location xk [xlk , xrk ], and all other sensors are still in their locations of C0 .
(3) A value Ri1 is maintained such that 0 Ri1 < , Ri1 is on a barrier,
every barrier point x < Ri1 is covered by a sensor of Si1 in Ci1 . (4) If Ri1
is not at the left endpoint of a barrier, then Ri1 is covered by a sensor of Si1
in Ci1 . (5) The point p+ (Ri1 ) is not covered by any sensor in Si1 .
Initially when i = 1, we let S0 = and R0 = 0, and thus all algorithm
invariants hold for C0 . The i-th step of the algorithm nds a sensor sg(i) S\Si1
and moves it to a new location xg(i) [xlg(i) , xrg(i) ] and thus obtains a new
conguration Ci . The details are given below.
Dene Si1 as the set of sensors that cover the point p+ (Ri1 ) in Ci1 , i.e.,
Si1 = {sk | xrk r Ri1 < xrk + r}. By the algorithm invariant (5), no sensor
in Si1 covers p+ (Ri1 ). Thus, Si1 S \ Si1 . If Si1 = , then we choose an
arbitrary sensor in Si1 as sg(i) (e.g., see Fig. 1) and let xg(i) = xrg(i) . We then set
Ri = xg(i) + r, i.e., Ri is at the right endpoint of the covering interval of sg(i) .
Note that Ci is Ci1 as sg(i) is not moved.
If Si1 = , then we dene Si2 = {sk | xlk r Ri1 < xrk r} (i.e., Si2
consists of those sensors sk that does not cover Ri1 when it is at xrk but is
possible to do so when it is at some location in [xlk , xrk ]). If Si2 = , we choose
the leftmost sensor of Si2 as sg(i) (e.g., see Fig. 2), and let xg(i) = Ri1 + r (i.e.,
we move sg(i) to xg(i) and thus obtain Ci ). If Si2 = , then we conclude that
< and terminate the algorithm.
Hence, if Si1 = Si2 = , the algorithm will stop and report < . Otherwise,
a sensor sg(i) is found from either Si1 or Si2 , and it is moved to xg(i) . In either
case, Ri = xg(i) + r and Si = Si1 {sg(i) }. If Ri , then we terminate the
algorithm and report . Otherwise, we further perform the following jump-
542 S. Li and H. Wang

over procedure: We check whether Ri is located at the interior of any barrier; if

not, then we set Ri to the left endpoint of the barrier right after Ri .
This nishes the i-th step of our algorithm. One can verify that all algorithm
invariants are maintained. As S has n sensors, the algorithm will nish in at most
n steps. This nishes the description of our algorithm. The algorithm correctness
and implementation are omitted.

Theorem 2. Given any value , we can determine whether in O(m +

n log n) time.

Our algorithm in Section 5 will perform feasibility tests multiple times, for
which we have the following lemma.

Lemma 9. Suppose the values xri for all i = 1, 2, . . . , n are already sorted, we
can determine whether in O(m + n log log n) time for any .

5 Solving the Problem MBC

To solve MBC, it suces to compute . The high-level scheme of our algorithm

is similar to that in [15], although some low-level computations are dierent.
We now use xri () to refer to xri for any , so that we consider xri () as a
function on [0, ], which actually denes a half of the upper branch (on the
right side of the y-axis) of a hyperbola. Let be the order of the values xri ( )
for all i [1, n]. To use Lemma 9, we rst run a preprocessing step in Lemma 10.

Lemma 10. With O(n log3 n + m log2 n) time preprocessing, we can compute
and an interval (1 , 2 ] containing such that is also the order of the values
xri () for any (1 , 2 ].

Proof. To compute , we apply Megiddos parametric search [12] to sort the

values xri ( ) for i [1, n],
$ using the decision algorithm in Theorem 2. Indeed,
recall that xri () = xi + 2 yi2 . Hence, as increases, xri () is a (strictly)
increasing function. For any two indices i and j, there is at most one root on
[0, ) for the equation: xri () = xrj (). Therefore, we can apply Megiddos
parametric search [12] to do the sorting. The total time is O(( + n) log2 n),
where is the running time of the decision algorithm. By Theorem 2, =
O(m + n log n). Hence, the total time for computing is O(m log2 n + n log3 n).
In addition, Megiddos parametric search [12] will return an interval (1 , 2 ]
containing and is also the order of the values xri () for any (1 , 2 ]. 

As (1 , 2 ], our subsequent feasible tests will be only on values

(1 , 2 )
because if 1 , then is not feasible and if 2 , then is feasible.
Lemmas 9 and 10 together lead to the following result.

Lemma 11. Each feasibility test can be done in O(m + n log log n) time for any
(1 , 2 ).
Algorithms for Covering Multiple Barriers 543

To compute , we parameterize our decision algorithm with as a pa-

rameter. Although we do not know , we execute the decision algorithm in such
a way that it computes the same subset of sensors sg(1) , sg(2) , . . . as would be
obtained if we ran the decision algorithm on = .
Recall that for any , step i of our decision algorithm computes the sensor
sg(i) , the set Si = {sg(1) , sg(2) , . . . , sg(i) }, and the value Ri , and obtains the
conguration Ci . In the following, we often consider as a variable rather than
a xed value. Thus, we will use Si () (resp., Ri (), sg(i) (), Ci (), xri ()) to
refer to the corresponding Si (resp., Ri , sg(i) , Ci , xri ). Our algorithm has at
most n steps. Consider a general i-th step for i 1. Right before the step, we
have an interval (1i1 , 2i1 ] and a sensor set Si1 (), such that the following
algorithm invariants hold.

1. (1i1 , 2i1 ].
1 2
2. The set Si1 () is the same (with the same order) for all (
%i1 , i1 ).
3. Ri1 () on (1i1 , 2i1 ) is either constant or equal to xj + 2 yj2 + c
for some constant c and some sensor sj with 1 j i 1, and Ri1 () is
maintained by the algorithm.
4. Ri1 () < for any (1i1 , 2i1 ).

Initially when i = 1, we let 10 = 1 and 20 = 2 . Since S0 () = and

R0 () = 0 for any , by Lemma 10, all invariants hold for i = 1. In general, the
i-th step will either compute , or obtain an interval (1i , 2i ] (1i1 , 2i1 ] and
a sensor sg(i) () with Si () = Si1 () {sg(i) ()}. The running time of the step
is O((m + n log log n)(log n + log m)). The details are omitted.
The algorithm will compute after at most n steps. The total time is
O(n (m + n log log n) (log m + log n)), which is bounded by O(nm log m +
n2 log n log log n) as shown in Theorem 3. The space of the algorithm is O(n).
Theorem 3. The problem MBC can be solved in O(nm log m+n2 log n log log n)
time and O(n) space.

6 Concluding Remarks
As mentioned before, the high-level scheme of our algorithm for MBC is similar
to those in [3, 15]. However, a new technique we propose in this paper can help
reduce the space complexities of the algorithms in [3, 15]. Specically, Chen et
al. [3] solved the line-constrained problem in O(n2 log n) time and O(n2 ) space
for the case where m = 1 and sensors have dierent ranges. Wang and Zhang [15]
solved the line-constrained problem in O(n2 log n log log n) time and O(n2 ) space
for the case where m = 1, sensors have the same range, and sensors have weights.
If we apply the similar preprocessing as in Lemma 10, then the space complexities
of both algorithms [3, 15] can be reduced to O(n) while the time complexities
do not change asymptotically.
In addition, by slightly changing our algorithm for MBC, we can also solve
the following problem variant: Find a subset S  of sensors of S to move them to
544 S. Li and H. Wang

L to cover all barriers such that the maximum movement of all sensors of S  is
minimized (and sensors of S \ S  do not move). We omit the details.

References
1. Andrews, A., Wang, H.: Minimizing the aggregate movements for interval coverage.
Algorithmica 78, 4785 (2017)
2. Bhattacharya, B., Burmester, B., Hu, Y., Kranakis, E., Shi, Q., Wiese, A.: Optimal
movement of mobile sensors for barrier coverage of a planar region. Theoretical
Computer Science 410(52), 55155528 (2009)
3. Chen, D., Gu, Y., Li, J., Wang, H.: Algorithms on minimizing the maximum sensor
movement for barrier coverage of a linear domain. Discrete and Computational
Geometry 50, 374408 (2013)
4. Chen, D., Tan, X., Wang, H., Wu, G.: Optimal point movement for covering circular
regions. Algorithmica 72, 379399 (2013)
5. Czyzowicz, J., Kranakis, E., Krizanc, D., Lambadaris, I., Narayanan, L., Opa-
trny, J., Stacho, L., Urrutia, J., Yazdani, M.: On minimizing the maximum sensor
movement for barrier coverage of a line segment. In: Proc. of the 8th International
Conference on Ad-Hoc, Mobile and Wireless Networks. pp. 194212 (2009)
6. Czyzowicz, J., Kranakis, E., Krizanc, D., Lambadaris, I., Narayanan, L., Opatrny,
J., Stacho, L., Urrutia, J., Yazdani, M.: On minimizing the sum of sensor move-
ments for barrier coverage of a line segment. In: Proc. of the 9th International
Conference on Ad-Hoc, Mobile and Wireless Networks. pp. 2942 (2010)
7. Dobrev, S., Durocher, S., Eftekhari, M., Georgiou, K., Kranakis, E., Krizanc, D.,
Narayanan, L., Opatrny, J., Shende, S., Urrutia, J.: Complexity of barrier coverage
with relocatable sensors in the plane. Theoretical Computer Science 579, 6473
(2015)
8. Frederickson, G., Johnson, D.: Generalized selection and ranking: Sorted matrices.
SIAM Journal on Computing 13(1), 1430 (1984)
9. Kumar, S., Lai, T., Arora, A.: Barrier coverage with wireless sensors. In: Proc. of
the 11th Annual International Conference on Mobile Computing and Networking
(MobiCom). pp. 284298 (2005)
10. Li, S., Shen, H.: Minimizing the maximum sensor movement for barrier coverage in
the plane. In: Proc. of the 2015 IEEE Conference on Computer Communications
(INFOCOM). pp. 244252 (2015)
11. Li, S., Wang, H.: Algorithms for covering multiple barriers. arXiv:1704.06870
(2017)
12. Megiddo, N.: Applying parallel computation algorithms in the design of serial
algorithms. Journal of the ACM 30(4), 852865 (1983)
13. Mehrandish, M.: On Routing, Backbone Formation and Barrier Coverage in Wire-
less Ad Doc and Sensor Networks. Ph.D. thesis, Concordia University, Montreal,
Quebec, Canada (2011)
14. Mehrandish, M., Narayanan, L., Opatrny, J.: Minimizing the number of sensors
moved on line barriers. In: Proc. of IEEE Wireless Communications and Network-
ing Conference (WCNC). pp. 653658 (2011)
15. Wang, H., Zhang, X.: Minimizing the maximum moving cost of interval coverage.
In: Proc. of the 26th International Symposium on Algorithms and Computation
(ISAAC). pp. 188198 (2015), full version to appear in International Journal of
Computational Geometry and Application (IJCGA)
An Improved Algorithm for Diameter-Optimally
Augmenting Paths in a Metric Space

Haitao Wang

Department of Computer Science

Utah State University, Logan, UT 84322, USA
[email protected]

Abstract. Let P be a path graph of n vertices embedded in a metric

space. We consider the problem of adding a new edge to P such that
the diameter of the resulting graph is minimized. Previously (in ICALP
2015) the problem was solved in O(n log3 n) time. In this paper, based on
new algorithmic techniques and observations, we present an O(n log n)
time algorithm.

1 Introduction
Let P be a path graph of n vertices embedded in a metric space. We consider
the problem of adding a new edge to P such that the diameter of the resulting
graph is minimized. Let G be a graph and each edge has a non-negative length.
The length of any path of G is the total length of all edges of the path. For any
two vertices u and v of G, we use dG (u, v) to denote the length of the shortest
path from u to v in G. The diameter of G is dened as maxu,vG dG (u, v).
Let P be a path graph of n vertices v1 , v2 , . . . , vn with an edge e(vi1 , vi )
connecting vi1 and vi for each 1 i n 1. Let V be the vertex set of
P . We assume (V, | |) is a metric space and |vi vj | is the distance of any two
vertices vi and vj of V . Specically, the following properties hold: (1) the triangle
inequality: |vi vk | + |vk vj | |vi vj |; (2) |vi vj | = |vj vi | 0; (3) |vi vj | = 0 if i = j.
In particular, for each edge e(vi1 , vi ) of P , its length is equal to |vi1 vi |. We
assume that given any two vertices vi and vj of P , the distance |vi vj | can be
obtained in O(1) time.
Our goal is to nd a new edge e connecting two vertices of P and add e to
P , such that the diameter of the resulting graph P {e} is minimized.
The problem has been studied before. Groe et al. [11] solved the problem
in O(n log3 n) time. In this paper, we present a new algorithm that runs in
O(n log n) time. As in [11], we refer to the problem as the diameter-optimally
augmenting path problem, or DOAP for short.

1.1 Related Work

If the path P is in the Euclidean space Rd for a constant d, then Groe et al. [11]
also gave an O(n + 1/3 ) time algorithm that can nd a (1 + )-approximation

This research was supported in part by NSF under Grant CCF-1317143.

Springer International Publishing AG 2017 545

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 545556, 2017.
DOI: 10.1007/978-3-319-62127-2_46
546 H. Wang

solution for DOAP, for any  > 0. If P is in the Euclidean plane R2 , De Carufel
et al. [4] gave a linear time algorithm for adding a new edge to P to minimize
the continuous diameter (i.e., the diameter is dened with respect to all points
of P , not only vertices). For a geometric tree T of n vertices embedded in the
Euclidean plane, De Carufel et al. [5] gave an O(n log n)-time algorithm for
adding a new edge to T to minimize the continuous diameter. Unfortunately,
both algorithms [4, 5], which are particularly for the continuous diameter in the
Euclidean plane, cannot be generalized to solve our problem for the discrete
diameter in the more general metric space.
Some more general problems were also studied before, e.g., see [1, 3, 6, 7, 10,
13, 14, 16] and the references therein. Consider a general graph G in which edges
have non-negative lengths. For an integer k, the goal of the general problem is to
compute k new edges and add them to G such that the resulting graph has the
minimum diameter. The problem is NP-hard [16] and some other variants are
even W[2]-hard [7, 10]. Approximation results were given for the general problem
and many of its variations, e.g., see [3, 7, 14]. The upper and lower bounds on
the diameters of the augmented graphs were investigated, e.g., [1, 13].
Since the diameter is an important metric of network performance, which
measures the worst-case cost between any two nodes of the network, as discussed
in [3, 6], the problem of augmenting graphs for minimizing the diameter and its
variations have many applications, such as in data networks, telephone networks,
transportation networks, scheduling, etc. As an application of DOAP, consider
the following example. Suppose there is a highway that connects several cities.
In order to reduce the transportation time, we want to build a new highway
connecting two cities such that the distance between the farthest two cities using
both highways is minimized. Clearly, this is a problem instance of DOAP.

1.2 Our Approach

To tackle the problem DOAP, Groe et al. [11] rst gave an O(n log n) time
algorithm for the decision version of the problem: Given any value , determine
whether it is possible to add a new edge e into P such that the diameter of
the resulting graph is at most . Then, by implementing the above decision
algorithm in a parallel fashion and applying Megiddos parametric search [15],
they solved DOAP in O(n log3 n) time [11]. For dierentiation, we refer to the
original problem DOAP as the optimization problem. Our improvement over the
previous work [11] is twofold.
First, we solve the decision problem in O(n) time. Our algorithm is based
on the O(n log n) time algorithm in the previous work [11]. However, by dis-
covering new observations on the problem structure and with the help of the
range-minima data structure [2, 12], we avoid certain expensive operations and
eventually achieve the O(n) time complexity.
Second, comparing with the decision problem, our algorithm for the opti-
mization problem is completely dierent from the previous work [11]. Let be
the diameter of the resulting graph in an optimal solution. Instead of using the
parametric search, we identify a set S of candidate values such that is in S and
An Improved Algorithm for Diameter-Optimally Augmenting Paths in a Metric Space 547

vi
v1
vk

vn

vj vl Ii () Ii () Ii () Ii ()

Fig. 1. Illustrating the resulting graph af- Fig. 2. Illustrating f (i, j) as j changes in
ter a new edge e(vi , vj ) is added. [i, n] and Ii (f ) for f {, , , }.

then we search in S using our algorithm for the decision problem. However,
computational diculties arise for this approach due to that the set S is too
large (|S| = (n2 )) and computing certain values of S is time-consuming (e.g.,
for certain values of S, computing each of them takes O(n) time). To circum-
vent these diculties, our algorithm has several steps. In each step, we shrink S
signicantly such that always remains in S. More importantly, each step will
obtain certain information, based on which the next step can further reduce S.
After several steps, the size of S is reduced to O(n) and all the remaining values
of S can be computed in O(n log n) time. At this point we can use our decision
algorithm to nd from S in additional O(n log n) time.
The remaining paper is organized as follows. In Section 2, we introduce some
notation and observations. In Section 3, we present our algorithm for the decision
problem. The optimization problem is solved in Section 4. Due to the space limit,
some lemma proofs are omitted but can be found in the full paper [17].

2 Preliminaries

In this section, we introduce some notation and observations, some of which are
from Groe et al. [11].
For any two vertices vi and vj of P , we use e(vi , vj ) to denote the edge
connecting vi and vj in the metric space. Hence, e(vi , vj ) is in P if |i j| is 1 or
0. The length of e(vi , vj ) is |vi vj |. For any i and j with 1 i j n, we use
G(i, j) to denote the resulting graph by adding the edge e(vi , vj ) into P . If i = j
or |i j| = 1, G(i, j) is P . Let D(i, j) denote the diameter of G(i, j).
Our goal for the optimization problem DOAP is to nd a pair of indices (i, j)
with 1 i j n such that D(i, j) is minimized. Let = min1ijn D(i, j),
i.e., is the diameter in an optimal solution.
Given a value , the decision problem is to determine whether , or in
other words, determine whether there exist a pair (i, j) with 1 i j n such
that D(i, j) . If yes, we say that is a feasible value.
Recall that for any graph G, dG (u, v) refers to the length of the shortest path
between two vertices u and v in G.
Consider any pair of indices (i, j) with 1 i j n. We dene (i, j),
(i, j), (i, j), and (i, j) as follows (refer to Fig. 1).
548 H. Wang

Denition 1. 1. Dene (i, j) to be the largest shortest path length in G(i, j)

from v1 to all vertices vk with k [i, j], i.e., (i, j) = maxikj dG(i,j) (v1 , vk ).
2. Dene (i, j) to be the largest shortest path length in G(i, j) from vn to all
vertices vk with k [i, j], i.e., (i, j) = maxikj dG(i,j) (vk , vn ).
3. Dene (i, j) to be the largest shortest path length in G(i, j) from vk to vl
for any k and l with i k l j, i.e., (i, j) = maxiklj dG(i,j) (vk , vl ).
4. Dene (i, j) to be the shortest path length in G(i, j) from v1 to vn , i.e.,
(i, j) = dG(i,j) (v1 , vn ).

It can be veried (also shown in [11]) that Observation 1 holds.

Observation 1 ([11]) D(i, j) = max{(i, j), (i, j), (i, j), (i, j)}.

Further, due to the triangle inequality of the metric space, the following
monotonicity properties hold.

Observation 2 ([11])
1. For any 1 i j n 1, (i, j) (i, j + 1), (i, j) (i, j + 1),
(i, j) (i, j + 1), and (i, j) (i, j + 1).
2. For any 1 i < j n, (i, j) (i + 1, j), (i, j) (i + 1, j), (i, j)
(i + 1, j), and (i, j) (i + 1, j).

For any pair (i, j) with 1 i j n, let P (i, j) denote the subpath of P
between
 vi and vj . Hence, dP (vi , vj ) is the length of P (i, j), i.e., dP (vi , vj ) =
ikj1 |vk vk+1 | if i < j and dP (vi , vj ) = 0 if i = j.
Our algorithms will need to compute f (i, j) for each f {, , , }. Lemma 1
was already shown by Groe et al. [11].

Lemma 1. ([11]) With O(n) time preprocessing, given any pair (i, j) with 1
i j n, we can compute dP (i, j) and (i, j) in O(1) time, and compute (i, j)
and (i, j) in O(log n) time.

For computing (i, j), although one may be able to do so in O(n) time, it
is not clear to us how to make it in O(log n) time even with O(n log n) time
preprocessing. As will be seen later, this is the major diculty for solving the
problem DOAP eciently. We refer to it as the -computation diculty. Our
main eort will be to circumvent the diculty by providing alternative and
ecient solutions.
For any pair (i, j) with 1 i j n, we use C(i, j) to denote the cy-
cle P (i, j) e(vi , vj ). Consider dG(i,j) (vk , vl ) for any k and l with i k
l j. Notice that the shortest path from vk to vl in C(i, j) is also a shortest
path in G(i, j). Hence, dG(i,j) (vk , vl ) = dC(i,j) (vk , vl ). There are two paths in
C(i, j) from vk to vl : one is P (k, l) and the other uses the edge e(vi , vj ). We
use d1C(i,j) (vk , vl ) to denote the length of the latter path, i.e., d1C(i,j) (vk , vl ) =
dP (vi , vk ) + |vi vj | + dP (vl , vj ). With this notation, we have dC(i,j) (vk , vl ) =
min{dP (vk , vl ), d1C(i,j) (vk , vl )}. According to the denition of (i, j), we summa-
rize our discussion in the following observation.
An Improved Algorithm for Diameter-Optimally Augmenting Paths in a Metric Space 549

(i, j) (i + 1, j) (i + 1, j) (i, j)
(i + 1, j) (i, j) (i, j) (i + 1, j)

Ii+1 () Ii () Ii+1 ()Ii () Ii () Ii+1 () Ii () Ii+1 ()

Fig. 3. Illustrating f (i, j) and f (i + 1, j) as j changes and Ii (f ) and Ii+1 (f ) for f

{, , , }.

Observation 3 For any pair (i, j) with 1 i j n, we have (i, j) =

maxiklj dC(i,j) (vk , vl ), with dC(i,j) (vk , vl ) = min{dP (vk , vl ), d1C(i,j) (vk , vl )}
and d1C(i,j) (vk , vl ) = dP (vi , vk ) + |vi vj | + dP (vl , vj ).
In the following, to simplify the notation, when the context is clear, we will
use index i to refer to vertex vi . For example, dP (i, j) refers to dP (vi , vj ).

3 The Decision Problem

In this section, we present our O(n) time algorithm for the decision problem. For
any value , our goal is to determine whether is feasible, i.e. whether ,
or equivalently, whether there is a pair (i, j) with 1 i j n such that
D(i, j) . If yes, our algorithm can also nd such a feasible edge e(i, j).
By Observation 1, D(i, j) if and only if f (i, j) for each f
{, , , }. To determine whether is feasible, our algorithm will determine
for each i [1, n], whether there exists j [i, n] such that f (i, j) for each
f {, , , }.
For any xed i [1, n], we consider (i, j), (i, j), (i, j), and (i, j) as func-
tions of j [i, n]. In light of Observation 2, (i, j) and (i, j) are monotonically
increasing and (i, j) and (i, j) are monotonically decreasing (e.g., see Fig. 2).
We dene four indices Ii (f ) for f {, , , } as follows. Refer to Fig. 2.
Denition 2. Dene Ii () to be the largest index j [i, n] such that (i, j) .
Dene Ii () similarly to Ii (), i.e., Ii () is the largest index j [i, n] such that
(i, j) . If (i, n) , then dene Ii () to be the smallest index j [i, n]
such that (i, j) ; otherwise, let Ii () = . Dene Ii () similarly to Ii (),
i.e., if (i, n) , then Ii () is the smallest index j [i, n] such that (i, j) ;
otherwise, Ii () = .
Clearly, is feasible if and only if [1, Ii ()][Ii (), n][1, Ii ()][Ii (), n] =
for some i [1, n]. By Observation 2, we have the following lemma.
Lemma 2. For any i [1, n 1], Ii () Ii+1 (), Ii () Ii+1 (), Ii ()
Ii+1 (), and Ii () Ii+1 () (e.g., see Fig. 3).
Proof. By Observation 2, (i, j) (i + 1, j). This implies that Ii () Ii+1 ()
by the their denitions. The other three cases for , , and are similar.

550 H. Wang

3.1 Computing Ii (), Ii (), and Ii () for all i [1, n]

In light of Lemma 2, for each f {, , }, we compute Ii (f ) for all i = 1, 2, . . . , n

in O(n) time, as follows.
We discuss the case for rst. According to Lemma 1, (i, j) can be computed
in constant time for any pair (i, j) with 1 i n. We can compute Ii () for
all i [1, n] in O(n) time by the following simple algorithm. We rst compute
I1 (), which is done by computing (1, j) from j = 1 incrementally until the
rst time (1, j) . Then, to compute I2 (), we compute (2, j) from j = I1 ()
incrementally until the rst time (2, j) . Next, we compute Ii () for i =
3, 4, . . . , n in the same way. The total time is O(n). The correctness is based on
the monotonicity property of Ii () in Lemma 2.
To compute Ii () or Ii () for i = 1, 2, . . . , n, using a similar approach as
above, we can only have an O(n log n) time algorithm since computing each
(i, j) or (i, j) takes O(log n) time by Lemma 1. Lemma 3 gives an approach
that only needs O(n) time.

Lemma 3. Ii () and Ii () for all i = 1, 2, . . . , n can be computed in O(n) time.

Due to the -computation diculty mentioned in Section 2, it is not clear

to us whether it possible to compute Ii () for all i = 1, . . . , n in O(n log n)
time. Recall that is feasible if and only if there exists an i [1, n] such that
[1, Ii ()][Ii (), n][1, Ii ()][Ii (), n] = . Now that Ii (f ) for all i [1, n] and
f {, , } have been computed but the Ii ()s are not known, below we will
use an indirect approach to determine whether the intersection of the above
four intervals is empty for every i [1, n].

3.2 Determining the Feasibility of

For each i [1, n], dene Qi = [1, Ii ()] [Ii (), n] [1, Ii ()] [Ii (), n]. Our
goal is to determine whether Qi is empty for each i = 1, 2, . . . , n.
Consider any i [1, n]. Since Ii (f ) for each f {, , } is known, we can
determine the intersection [1, Ii ()] [Ii (), n] [Ii (), n] in constant time. If
the intersection is empty, then Qi = . In the following, we assume that the
intersection is not empty, and let ai be the smallest index in the intersection.
As in [11], an easy observation is that Qi = if and only if ai [1, Ii ()]. If
ai i (note that ai i actually implies ai = i since ai Ii () i), it is obvi-
ously true that ai [1, Ii ()] since i Ii (). Otherwise (i.e., i < ai ), according
to the denition of Ii (), ai [1, Ii ()] if and only if (i, ai ) . Groe et al. [11]
gave an approach that can determine whether (i, ai ) in O(log n) time after
O(n log n) time preprocessing. Below, by new observations and with the help of
the range minima data structure [2, 12], we show that whether (i, ai ) can
be determined in constant time after O(n) time preprocessing.
For each j [1, n], dene gj as the largest index k in [j, n] such that dP (j, k)
. Observe that g1 g2 gn .
An Improved Algorithm for Diameter-Optimally Augmenting Paths in a Metric Space 551

i
j

ai g j + 1 g j

Fig. 4. Illustrating the graph G(i, ai ) with gj + 1 ai .

Consider any i and the corresponding ai with i < ai . Our goal is to determine
whether (i, ai ) . Since we are talking about (i, ai ), we are essentially con-
sidering the graph G(i, ai ). Recall that C(i, ai ) is the cycle P (i, ai ) e(i, ai ). By
Observation 3, (i, ai ) = maxiklai dC(i,ai ) (k, l), and further, dC(i,j) (k, l) =
min{dP (k, l), d1C(i,ai ) (k, l)} and d1C(i,ai ) (k, l) = dP (i, k) + |vi vai | + dP (l, ai ).
For any j [i, ai 1], if gj ai 1, then vertex gj + 1 is in the cycle C(i, ai ).
By denition, d1C(i,ai ) (j, gj + 1) = dP (i, j) + |vi vai | + dP (gj + 1, ai ). See Fig. 4.

Lemma 4. (i, ai ) if and only if for each j [i, ai 1], either gj ai or

d1C(i,ai ) (j, gj + 1) .

Recall that g1 g2 gn . For each k [1, n], dene hk to be the

smallest index j in [1, k] with gj k. Observe that h1 h2 hn .
Note that if i < hai , then for each j [i, hai 1], gj < ai and gj + 1 ai .
Due to Lemma 4, we further have the following lemma.

Lemma 5. (i, ai ) if and only if either hai i or d1C(i,ai ) (j, gj + 1)

holds for each j [i, hai 1].

Let |C(i, ai )| denote the total length of the cycle C(i, ai ), i.e., |C(i, ai )| =
dP (i, ai ) + |vi vai |. The following observation is crucial because it immediately
leads to our algorithm in Lemma 6.

Observation 4 (i, ai ) if and only if either minj[i,hai 1] {dP (j, gj + 1)}

|C(i, ai )| or hai i.

Proof. Suppose hai > i. Then, for each j [i, hai 1], gj < ai and gj + 1 ai .
Note that d1C(i,ai ) (j, gj +1) = |C(i, ai )|dP (i, gj +1). Hence, d1C(i,ai ) (j, gj +1)
is equivalent to dP (j, gj +1) |C(i, ai )|. Therefore, d1C(i,ai ) (j, gj +1) holds
for each j [i, hai 1] if and only if minj[i,hai 1] {dP (j, gj + 1)} |C(i, ai )| .
By Lemma 5, the observation follows.


Lemma 6. With O(n) time preprocessing, given any i [1, n] and the corre-
sponding ai with i < ai , whether (i, ai ) can be decided in constant time.

Proof. As preprocessing, we rst compute gj for all j = 1, 2, . . . , n, which can be

done in O(n) time due to the monotonicity property g1 g2 . . . gn . Then,
we compute hk for all k = 1, 2, . . . , n, which can also be done in O(n) time due
552 H. Wang

to the monotonicity property h1 h2 . . . hn . Next, we compute an array

B[1, . . . , n] with B[j] = dP (j, gj + 1) for each j [1, n] (let dP (j, gj + 1) =
if gj + 1 > n). We build a range-minima data structure on B [2, 12]. The range
minima data structure can be built in O(n) time such that given any pair (i, j)
with 1 i j n, the minimum value of the subarray B[i j] can be returned
in constant time [2, 12]. This nishes the preprocessing step, which takes O(n)
time in total.
Consider any i and the corresponding ai with i < ai . Our goal is to determine
whether (i, ai ) , which can be done in O(1) time as follows.
According to Observation 4, (i, ai ) if and only if either hai i or
minj[i,hai 1] {dP (j, gj + 1)} |C(i, ai )| . Since hai has been computed in
the preprocessing, we check whether hai i is true. If yes, then we are done
with the assertion that (i, ai ) . Otherwise, we need to determine whether
minj[i,hai 1] {dP (j, gj + 1)} |C(i, ai )| holds. To this end, we rst compute
minj[i,hai 1] {dP (j, gj +1)} in constant time by querying the range-minima data
structure on B with (i, hai 1). Note that |C(i, ai )| can be computed in constant
time. Therefore, we can determine whether (i, ai ) in O(1) time.

With Lemma 6, the decision problem can be solved in O(n) time. The proof
of the following theorem summarizes our algorithm.
Theorem 1. Given any , we can determine whether is feasible in O(n) time,
and if is feasible, a feasible edge can be found in O(n) time.
Proof. First, we do the preprocessing in Lemma 1 in O(n) time. Then, for each
f {, , }, we compute Ii (f ) for all i = 1, 2, . . . , n, in O(n) time. We also do
the preprocessing in Lemma 6.
Next, for each i [1, n], we do the following. Compute the intersection
[1, Ii ()] [Ii (), n] [Ii (), n] in constant time. If the intersection is empty,
then we are done for this i. Otherwise, obtain the smallest index ai in the above
intersection. If ai i, then we stop the algorithm with the assertion that is
feasible and report e(i, ai ) as a feasible edge. Otherwise, we use Lemma 6 to de-
termine whether (i, ai ) in constant time. If yes, we stop the algorithm with
the assertion that is feasible and report e(i, ai ) as a feasible edge. Otherwise,
we proceed on i + 1.
If the algorithm does not stop after we check all i [1, n], then we stop the
algorithm with the assertion that is not feasible. Clearly, we spend O(1) time
on each i, and thus, the total time of the algorithm is O(n).


4 The Optimization Problem

In this section, we solve the optimization problem in O(n log n) time, by making
use of our algorithm for the decision problem given in Section 3 (we will refer to
it as the decision algorithm). We will focus on computing .
Notice that must be equal to the diameter D(i, j) of G(i, j) for some pair
(i, j) with 1 i j n. Further, by Observation 1, is equal to f (i, j) for
some f {, , , } and some pair (i, j) with 1 i j n.
An Improved Algorithm for Diameter-Optimally Augmenting Paths in a Metric Space 553

For each f {, , , }, dene Sf = {f (i, j) | 1 i j n}. Let S =

f {,,,} Sf . According to our discussion above, is in S. Further, note that
is the smallest feasible value of S. We will not compute the entire set S since
|S| = (n2 ). For each f {, , , }, let f be the smallest feasible value in
Sf . Hence, we have = min{ , , , }.
In the following, we rst compute , , in O(n log n) time.

4.1 Computing , , and

For convenience, we begin with computing .
We dene an n n matrix M [1 n; 1 n]: For each 1 i n and 1
j n, dene M [i, j] = (i, j) if j i and M [i, j] = (i, i) otherwise. By
Observation 2, the following lemma shows that M is a sorted matrix in the
sense that each row is sorted in descending order from left to right and each
column is sorted in descending order from top to bottom.
Lemma 7. For each 1 i n, M [i, j] M [i, j + 1] for any j [1, n 1]; for
each 1 j n, M [i, j] M [i + 1, j] for any i [1, n 1].
Note that each element of S is in M and vice versa. Since is the smallest
feasible value of S , is also the smallest feasible value of M . We do not
construct M explicitly. Rather, given any i and j, we can evaluate M [i, j]
in O(log n) time since (i, j) can be computed in O(log n) time if i j by
Lemma 1. Using the sorted-matrix searching techniques [8, 9], we can nd in
M by calling our decision algorithm O(log n) times and evaluating O(n) elements
of M . The total time on calling the decision algorithm is O(n log n) and the total
time on evaluating matrix elements is also O(n log n). Hence, we can compute
in O(n log n) time.
Computing and can be done similarly in O(n log n) time, although the
corresponding sorted matrices may be dened slightly dierently. We omit the
details. However, we cannot compute in O(n log n) time in the above way,
and again this is due to the -computation diculty mentioned in Section 2.
Note that having , , and essentially reduces our search space for from
S to S { , , }.
Let 1 = min{ , , }. Thus, = min{1 , }. Hence, if 1 , then
= 1 and is computed. Otherwise (i.e., < 1 ), it must be that =

and we need to compute . To compute , again we cannot use the similar way
as the above for computing . Instead, we use the following approach, whose
success relies on the information implied by < 1 .

4.2 Computing in the Case < 1

We assume < 1 . Hence, = . Let e(i , j ) be the new edge added to P
in an optimal solution. We also call e(i , j ) an optimal edge.
Since = < 1 , we have the following observation.

Observation 5 If < 1 and e(i , j ) is an optimal edge, then = (i , j ).

554 H. Wang

For any i [1, n], for each f {, , , }, with respect to 1 , we dene Ii (f )
in a similar way to Ii (f ) dened in Section 3 with respect to except that we
change to < 1 . Specically, dene Ii () to be the largest index j [i, n]
such that (i, j) < 1 . Ii () is dened similarly to Ii (). If (i, n) < 1 , then
dene Ii () to be the smallest index j [i, n] such that (i, j) < 1 ; otherwise
Ii () = . Ii () is dened similarly to Ii (). Note that similar monotonicity
properties for Ii (f ) with f {, , , } to those in Lemma 2 also hold.
Recall that e(i , j ) is an optimal edge. An easy observation is that since
1 is strictly larger than , the intersection [1, Ii ()] [Ii (), n] [Ii (), n]
cannot be empty. Let ai be the smallest index in the above intersection. Note
that i ai since i Ii () ai . The following lemma shows that e(i , ai )
is actually an optimal edge.
Lemma 8. If < 1 and e(i , j ) is an optimal edge, then j = ai .
Lemma 8 is crucial because it suggests the following algorithm.
We rst compute the indices Ii (), Ii (), Ii () for i = 1, . . . , n. This can
be done in O(n) time using the similar algorithms as those for computing
Ii (), Ii (), Ii () in Section 3.1.
Next, for each i [1, n], if [1, Ii ()] [Ii (), n] [Ii (), n] = , then we
compute ai , i.e., the smallest index in the above intersection. Let I be the set of
indices i such that the above interval intersection for i is not empty. Lemma 8
leads to the following observation.

Observation 6 If < 1 , then is the smallest feasible value of the set

{(i, ai ) | i I}.

Proof. By Lemma 8, one of the edges of {e(i, ai ) | i I} is an optimal edge. By

Observation 5, is in {(i, ai ) | i I}. Thus, is the smallest feasible value
in {(i, ai ) | i I}.

Observation 6 essentially reduces the search space for to {(i, ai ) | i I},
which has at most O(n) values. It is tempting to rst explicitly compute the
set and then nd from the set. However, again, due to the -computation
diculty, it is not clear to us how to compute the set in O(n log n) time. Alter-
natively, we use the following approach to compute .

4.3 Finding in the Set {(i, ai ) | i I}

Recall that according to Observation 3, (i, j) = maxiklj dC(i,j) (k, l), with
dC(i,j) (k, l) = min{dP (k, l), d1C(i,j) (k, l)} and d1C(i,j) (k, l) = dP (i, k) + |vi vj | +
dP (l, j). Hence, (i, j) is equal to dP (k, l) or d1C(i,j) (k, l) for some k l. There-
fore, by Observation 6, there exists i I such that is equal to dP (k, l) or
d1C(i,ai ) (k, l) for some k and l with i k l ai .
Let Sp = {dP (k, l) | 1 k l n} and Sc = {d1C(i,ai ) (k, l) | i k l
ai , i I}. Based on our above discussion, is in Sp Sc . Further, is the
smallest feasible value in Sp Sc .
An Improved Algorithm for Diameter-Optimally Augmenting Paths in a Metric Space 555

Let p be the smallest feasible value of Sp and let c be the smallest feasible
value of Sc . Hence, = min{p , c }. By using the technique of searching sorted-
matrices [8, 9], the following lemma computes p in O(n log n) time.

Lemma 9. p can be computed in O(n log n) time.

Recall that = min{p , c }. In the case p c , = p and we are done

with computing . In the following, we assume p > c . Thus, = c . With
the help of the information implied by p > c , we will compute in O(n log n)
time. The details are given below.
For any j [1, n], let gj denote the largest index k [j, n] such that the
subpath length dP (j, k) is strictly smaller than p . Note that the denition of gj
is similar to gj dened in Section 4.3 except that we change to < p .
For each k [1, n], let hk denote the smallest index j [1, k] with gj k.
Let I  be the subset of i I  such that i hai 1. Hence, for each i I  and
each j [i, hai 1], gj < ai and thus gj + 1 ai .
For each i I  , dene d1max (i, ai ) = maxj[i,ha 1] d1C(i,ai ) (j, gj + 1). The
i
following lemma gives a way to determine .

Lemma 10. If < 1 and c < p , then = d1max (i, ai ) for some i I  .

By Lemma 10, in the case of c < p , = c is the smallest feasible value

of d1max (i, ai ) for all i I  . Note that the number of such values d1max (i, ai ) is
O(n). Hence, if we can compute d1max (i, ai ) for all i I  , then can be easily
found in additional O(n log n) time using our decision algorithm, e.g., by rst
sorting these values and then doing binary search.
Lemma 11 computes d1max (i, ai ) for all i I  in O(n) time, with the help of
the range-minima data structure [2, 12].

Lemma 11. d1max (i, ai ) for all i I  can be computed in O(n) time.

In summary, we can compute in O(n log n) time in the case < 1 and
c < p . Our overall algorithm for computing an optimal solution is summarized
in the proof of Theorem 2.

Theorem 2. An optimal solution for the optimization problem can be found in

O(n log n) time.

Proof. First, we compute , , and , in O(n log n) time by using our decision
algorithm and the sorted-matrix searching techniques. Then, we compute 1 =
min{ , , }.
Second, by using 1 , we compute the indices Ii (), Ii (), and Ii () for all
i = 1, 2, . . . , n. This can be done in O(n) time. For each i [1, n], if [1, Ii ()]
[Ii (), n] [Ii (), n] = , we compute ai (i.e., the smallest index in the above
intersection) and add i to the set I (initially I = ). Hence, all such ai s and I
can be computed in O(n) time.
If I = , then we return 1 as .
556 H. Wang

If I = , then we compute p in O(n log n) time by Lemma 9. We proceed

to compute d1max (i, ai ) for all i I  by Lemma 11, and then nd the smallest
feasible value  in the set {d1max (i, ai ) | i I  } in O(n log n) time. Finally, we
return min{1 , p ,  } as .
The above computes in O(n log n) time. Applying = on our decision
algorithm can eventually nd an optimal edge in additional O(n) time.


References
1. Alon, N., Gyarfas, A., Ruszinko, M.: Decreasing the diameter of bounded degree
graphs. Journal of Graph Theory 35, 161172 (2000)
2. Bender, M., Farach-Colton, M.: The LCA problem revisited. In: Proc. of the 4th
Latin American Symposium on Theoretical Informatics. pp. 8894 (2000)
3. Bilo, D., Guala, L., Proietti, G.: Improved approximability and non-
approximability results for graph diameter decreasing problems. Theoretical Com-
puter Science 417, 1222 (2012)
4. Carufel, J.L.D., Grimm, C., Maheshwari, A., Smid, M.: Minimizing the continuous
diameter when augmenting paths and cycles with shortcuts. In: Proc. of the 15th
Scandinavian Workshop on Algorithm Theory (SWAT). pp. 27:127:14 (2016)
5. Carufel, J.L.D., Grimm, C., Schirra, S., Smid, M.: Minimizing the continuous di-
ameter when augmenting a tree with a shortcut. arXiv:1612.01370 (2016)
6. Demaine, E., Zadimoghaddam, M.: Minimizing the diameter of a network using
shortcut edges. In: Proc. of the 12th Scandinavian conference on Algorithm Theory
(SWAT). pp. 420431 (2010)
7. Frati, F., Gaspers, S., Gudmundsson, J., Mathieson, L.: Augmenting graphs to
minimize the diameter. Algorithmica 72, 9951010 (2015)
8. Frederickson, G., Johnson, D.: Generalized selection and ranking: Sorted matrices.
SIAM Journal on Computing 13(1), 1430 (1984)
9. Frederickson, G., Johnson, D.: Finding kth paths and p-centers by generating and
searching good data structures. Journal of Algorithms 4(1), 6180 (1983)
10. Gao, Y., Hare, D., Nastos, J.: The parametric complexity of graph diameter aug-
mentation. Discrete Applied Mathematics 161, 16261631 (2013)
11. Groe, U., Gudmundsson, J., Knauer, C., Smid, M., Stehn, F.: Fast algorithms for
diameter-optimally augmenting paths. In: Proc. of the 42nd International Collo-
quium on Automata, Languages and Programming (ICALP). pp. 678688 (2015)
12. Harel, D., Tarjan, R.: Fast algorithms for nding nearest common ancestors. SIAM
Journal on Computing 13, 338355 (1984)
13. Ishii, T.: Augmenting outerplanar graphs to meet diameter requirements. Journal
of Graph Theory 74, 392416 (2013)
14. Li, C.L., McCormick, S., Simchi-Levi, D.: On the minimum-cardinality-bounded-
diameter and the bounded-cardinality-minimum-diameter edge addition problems.
Operations Research Letters 11, 303308 (1992)
15. Megiddo, N.: Applying parallel computation algorithms in the design of serial
algorithms. Journal of the ACM 30(4), 852865 (1983)
16. Schoone, A., Bodlaender, H., Leeuwen, J.V.: Diameter increase caused by edge
deletion. Journal of Graph Theory 11, 409427 (1997)
17. Wang, H.: An improved algorithm for diameter-optimally augmenting paths in a
metric space. arXiv:1608.04456 (2016)
 Covering Uncertain Points in a Tree

Haitao Wang and Jingru Zhang

Department of Computer Science

Utah State University, Logan, UT 84322, USA
[email protected],[email protected]

Abstract. We consider a coverage problem for uncertain points in a tree.

Let T be a tree containing a set P of n (weighted) demand points, and the
location of each demand point Pi P is uncertain but is known to be in
one of mi points on T each associated with a probability. Given a covering
range , the problem is to nd a minimum number of points (called
centers) on T to build facilities for serving (or covering) these demand
points in the sense that for each uncertain point Pi P, the expected
distance from Pi to at least one center is no more than . The problem
has not been studied before. We present an O(|T | + M log2 M ) time
algorithm, where |T | is the number of vertices of T and Mis the total
number of locations of all uncertain points of P, i.e., M = Pi P mi .

1 Introduction
Data uncertainty is very common in many applications, such as sensor databases,
image resolution, and it is mainly due to measurement inaccuracy, sampling dis-
crepancy, outdated data sources, resource limitation, etc. Problems on uncertain
data have attracted considerable attention, e.g., [13, 9, 10, 15]. In this paper, we
study a problem of covering uncertain points on a tree, dened as follows.
Let T be a tree. We consider each edge e of T as a line segment of a positive
length so that we can talk about points on e. The distance of any two points
p and q on T , denoted by d(p, q), is dened as the sum of the lengths of all edges
on the simple path from p to q in T . Let P = {P1 , . . . , Pn } be a set of n uncertain
(demand) points on T . Each Pi P has mi possible locations on T , denoted by
{pi1 , pi2 , , pimi }, and each location pij of Pi is associated with a probability
fij 0 for Pi appearing at pij (which is independent of other locations), with
 mi
j=1 fij = 1; e.g., see Fig. 1. In addition, each Pi P has a weight wi 0.
For any point x on T , the (weighted) expected  i distance from x to Pi , denoted by
Ed(x, Pi ), is dened as Ed(x, Pi ) = wi m j=1 fij d(x, pij ).
Given a value 0, called the covering range, we say that a point x on T
covers an uncertain point Pi if Ed(x, Pi ) . The center-coverage problem is
to compute a minimum number of points on T , called centers, such that every
uncertain point of P is covered by at least one center.
To the best of our knowledge, the problem has not been studied before. nLet M
denote the total number of locations of all uncertain points, i.e., M = i=1 mi .

This research was supported in part by NSF under Grant CCF-1317143.

Springer International Publishing AG 2017 557

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 557568, 2017.
DOI: 10.1007/978-3-319-62127-2_47
558 H. Wang and J. Zhang

p2,3
0.4 0.2
0.3 0.1 p1,1
p1,3 p2,2 p1,4
p2,1 0.2
0.5 p1,2 0.3

Fig. 1. Illustrating two uncertain points P1 and P2 , where P1 has four possible locations
and P2 has three possible locations. The numbers are the probabilities.

Let |T | be the number of vertices of T . In this paper, we present an algorithm

that solves the problem in O(|T | + M log2 M ) time, which is nearly linear as the
input size of the problem is (|T | + M ).
As an application of our algorithm, we also solve a dual problem, called
the k-center problem, which is to compute a number of k centers on T such
that the covering range is minimized. The runtime of our algorithm is O(|T | +
n2 log n log M + M log2 M log n).

1.1 Related Work

Two models on uncertain data have been commonly considered: the existential
model [3, 9, 10] and the locational model [1, 2, 15]. In the existential model an
uncertain point has a specic location but its existence is uncertain while in the
locational model an uncertain point always exists but its location is uncertain
and follows a probability distribution function. Our problems belong to the loca-
tional model. In fact, the same problems under existential model are essentially
the weighted case for deterministic points (i.e., each Pi P has a single cer-
tain location), and the center-coverage problem is solvable in linear time [11]
and the k-center problem is solvable in O(n log2 n) time [6, 14].
If T is a path, both the center-coverage problem and the k-center problem on
uncertain points have been studied [17], but under a somewhat special problem
setting where mi is the same for all 1 i n. The two problems were solved in
O(M + n log k) and O(M log M + n log k log n) time, respectively. If T is tree, an
O(|T | + M ) time algorithm was given in [16] for the one-center problem under
the above special problem setting.
Recently Li and Huang [7] considered the same k-center problem under the
same uncertain model as ours but in the Euclidean space, and they gave an
approximation algorithm. Facility location problems in other uncertain models
were also considered. Loer and van Kreveld [12] gave algorithms for computing
the smallest enclosing circle for imprecise points each of which is contained in a
planar region (e.g., a circle or a square). Jrgenson et al. [8] studied the problem
of computing the distribution of the radius of the smallest enclosing circle for
uncertain points each of which has multiple locations in the plane. de Berg et
al. [4] proposed algorithms for dynamically maintaining Euclidean 2-centers for
a set of moving points in the plane (the moving points are considered uncertain).
Covering Uncertain Points in a Tree 559

1.2 Our Approach

For each uncertain point Pi P, we nd a point pi on T that minimizes the

expected distance Ed(pi , Pi ), and pi is actually the weighted median of all loca-
tions of Pi . We observe that if we move a point x on T away from pi , Ed(x, Pi )
is monotonically increasing. We compute the medians pi for all uncertain points
in O(M log M ) time. We show that there is an optimal solution in which all
centers are in Tm , where Tm is the minimum subtree of T connecting/spanning
all medians pi . Next we nd centers on Tm . For this, we propose a simple greedy
algorithm, but the challenge is on developing ecient data structures to perform
certain operations. We briey discuss it below.
We pick an arbitrary vertex r of Tm as the root. Starting from the leaves, we
consider the vertices of Tm in a bottom-up manner and place centers whenever
we have to. For example, consider a leaf v holding a median pi and let u
be the parent of v. If Ed(u, Pi ) > , then we have to place a center c on the
edge e(u, v) in order to cover Pi . The location of c is chosen to be at a point
of e(u, v) with Ed(c, Pi ) = (i.e., on the one hand, c covers Pi , and on the
other hand, c is close to u as much as possible in the hope of covering other
uncertain points as many as possible). After c is placed, we nd and remove all
uncertain points that are covered by c. Performing this operation eciently is
a key diculty for our approach. We solve the problem in an output-sensitive
manner by proposing a dynamic data structure that also supports the remove
operations. We also develop data structures for other operations needed in the
algorithm. These data structures may be of independent interest.
For solving the k-center problem, by observations, we rst identify a set of
O(n2 ) candidate values such that the covering range in the optimal solution
must be in the set. Subsequently, we use our algorithm for the center-coverage
problem as a decision procedure to nd the optimal covering range in the set.
We introduce some notations in Section 2. In Section 3, we describe our algo-
rithmic scheme for the center-coverage problem, with details in the subsequent
sections. Due to the space limit, many proofs and details, including our algorithm
for the k-center problem, are omitted but can be found in the full paper [18].

2 Preliminaries

Note that the locations of the uncertain points of P may be in the interior of
the edges of T . A vertex-constrained case happens if all locations of P are at
vertices of T and each vertex of T holds at least one location of P (but the
centers can still be in the interior of edges). As in [16], the general case can be
reduced to the vertex-constrained case in O(|T | + M ) time (see the full paper
for details). In the following, unless otherwise stated, we focus our discussion
on the vertex-constrained case and assume that our problem on P and T is a
vertex-constrained case.
For ease of exposition, we further make a general position assumption that
every vertex of T has only one location of P (we explain in the full paper that
560 H. Wang and J. Zhang

our algorithm easily extends to the degenerate case). Under the assumption,
|T | = M n.
Let e(u, v) denote the edge of T incident to two vertices u and v. For any
two points p and q on T , denote by (p, q) the simple path from p to q on T .
Let be any simple path on T and x be any point on . For any location
pij of an uncertain point Pi , the distance d(x, pij ) is a convex (and piecewise
linear) function as x changes on [13]. As a sum of multiple convex functions,
Ed(x, Pi ) is also convex (and piecewise linear) on , i.e., in general, as x moves
on , Ed(x, Pi ) rst monotonically decreases and then increases. In particular,
for each edge e of T , Ed(x, Pi ) is a linear function for x e.
For any subtree T  of T and any Pi P, we call the sum of the probabilities
of the locations of Pi in T  the probability sum of Pi in T  .
For each uncertain point Pi , let pi be a point x T that minimizes Ed(x, Pi ).
If we consider wi fij as the weight of pij , pi is actually the weighted median
of all points pij Pi . We call pi the median of Pi . Although pi may not be
unique (e.g., when there is an edge e dividing T into two subtrees such that the
probability sum of Pi in either subtree is exactly 0.5), Pi always has a median
located at a vertex v of T , and we let pi refer to such a vertex.
Recall that is the covering range of our problem. If Ed(pi , Pi ) > for some
i [1, n], then there is no solution for the problem since no point of T can cover
Pi . Henceforth, we assume Ed(pi , Pi ) for each i [1, n].

3 The Algorithmic Scheme

In this section, we describe our algorithmic scheme for the center-coverage prob-
lem, and the implementation details will be presented later. We start with com-
puting the medians pi of all uncertain points of P. We have the following lemma.

Lemma 1. The medians pi of all Pi of P can be computed in O(M log M ) time.

3.1 The Medians-Spanning Tree Tm

Denote by P the set of all medians pi . Let Tm be the minimum connected
subtree of T that spans/connects all medians. Note that each leaf of Tm must
hold a median. We pick an arbitrary median as the root of T , denoted by r. The
subtree Tm can be easily computed in O(M ) time by a post-order traversal on
T (with respect to the root r), and we omit the details. The following lemma
is based on the fact that Ed(x, Pi ) is convex for x on any simple path of T and
Ed(x, Pi ) minimizes at x = pi .

Lemma 2. There exists an optimal solution for the center-coverage problem in

which every center is on Tm .

Due to Lemma 2, we will focus on nding centers on Tm . We also consider r

as the root of Tm , and then we can talk about ancestors and descendants of the
vertices in Tm . Note that for any two vertices u and v of Tm , (u, v) is in Tm .
Covering Uncertain Points in a Tree 561

We reindex all medians and the corresponding uncertain points so that the
new indices will facilitate our algorithm, as follows. Starting from an arbitrary
child of r in Tm , we traverse down the tree Tm by always following the leftmost
child of the current node until we encounter a leaf, denoted by v . Starting from
v (i.e., v is the rst visited leaf), we perform a post-order traversal on Tm
and reindex all medians of P such that p1 , p2 , . . . , pn is the list of points of P
visited in order in the above traversal. Recall that the root r contains a median,
which is pn after the reindexing. Accordingly, we also reindex all uncertain points
of P and their corresponding locations on T , which can be done in O(M ) time.
In the following paper, we will always use the new indices.
For each vertex v of Tm , we use Tm (v) to represent the subtree of Tm rooted at
v. The reason we do the above reindexing is that for any vertex v of Tm , the new
indices of all medians in Tm (v) must form a range [i, j] for some 1 i j n,
and we use R(v) to denote the range. It will be clear later that this property
will facilitate our algorithm.

3.2 The Algorithm

Our algorithm for the center-coverage problem works as follows. Initially, all
uncertain points are active. During the algorithm, we will place centers on
Tm , and once an uncertain point Pi is covered by a center, we will deactivate
it (it then becomes inactive). The algorithm visits all vertices of Tm following
the above post-order traversal of Tm starting from leaf v . Suppose v is currently
being visited. Unless v is the root r, let u be the parent of v. Below we describe
our algorithm for processing v. There are two cases depending on whether v is a
leaf or an internal node, although the algorithm for them is essentially the same.

The Leaf Case If v is a leaf, then it holds a median pi . If Pi is inactive, we do

nothing; otherwise, we proceed as follows.
We compute a point c (called a candidate center) on the path (v, r) closest
to r such that Ed(c, Pi ) . Note that if we move a point x from v to r
along (v, r), Ed(x, Pi ) is monotonically increasing. By the denition of c, if
Ed(r, Pi ) , then c = r; otherwise, Ed(c, Pi ) = . If c is in (u, r), then we do
nothing and nish processing v. Below we assume that c is not in (u, r) and
thus is in e(u, v) \ {u} (i.e., c e(u, v) but c  u).
In order to cover Pi , by the denition of c, we must place a center in e(u, v) \
{u}. Our strategy is to place a center at c. Indeed, this is the best location
for placing a center since it is the location that covers Pi and is closest to u
(and thus is closest to every other active uncertain point). We use a candidate-
center-query to compute c in O(log n) time, whose details will be discussed later.
Next, we report all active uncertain points that are covered by c, by a coverage-
report-query in output-sensitive O(log M log n + k log n) amortized time, where
k is the number of uncertain points covered by c. The details for the operation
will be discussed later. Further, we deactivate all these uncertain points. We will
show that deactivating each uncertain point Pj can be done in O(mj log M log n)
amortized time. This nishes processing v.
562 H. Wang and J. Zhang

The Internal Node Case If v is an internal node, since we process the ver-
tices of Tm following a post-order traversal, all descendants of v have already
been processed. Our algorithm maintains an invariant that if the subtree Tm (v)
contains any active median pi (i.e., Pi is active), then Ed(v, Pi ) . When v is
a leaf, this invariant trivially holds. Our way of processing a leaf discussed above
also maintains this invariant.
To process v, we rst check whether Tm (v) has any active medians. This is
done by a range-status-query in O(log n) time, whose details will be given later.
If Tm (v) does not have any active median, then we are done with processing
v. Otherwise, by the algorithm invariant, for each active median pi in Tm (v), it
holds that Ed(v, Pi ) . If v = r, we place a center at v and nish the algorithm.
Below, we assume v = r and thus u is the parent of v.
We compute a point c on (v, r) closest to r such that Ed(c, Pi ) for all
active medians pi Tm (v), and we call c the candidate center. By the denition
of c, if Ed(r, Pi ) for all active medians pi Tm (v), then c = r; otherwise,
Ed(c, Pi ) = for some active median pi Tm (v). As in the leaf case, nding c
is done in O(log n) time by a candidate-center-query. If c is on (u, r), then we
nish processing v. Note that this implies Ed(u, Pi ) for each active median
pi Tm (v), which maintains the algorithm invariant for u.
If c  (u, r), then c e(u, v) \ {u}. In this case, by the denition of c, we
must place a center in e(u, v) \ {u} to cover Pi . As discussed in the leaf case, the
best location for placing a center is c and thus we place a center at c. Then, by
using a coverage-report-query, we nd all active uncertain points covered by c
and deactivate them. Note that by the denition of c, c covers Pj for all medians
pj Tm (v). This nishes processing v.
Once the root r is processed, the algorithm nishes.

3.3 The Time Complexity

To analyze the running time of the algorithm, it remains to discuss the three
operations: range-status-queries, coverage-report-queries, and candidate-center-
queries. For answering range-status-queries, we have the following Lemma 3.
Lemma 3. We can build a data structure in O(M ) time that can answer each
range-status-query in O(log n) time. Further, once an uncertain point is deacti-
vated, we can remove it from the data structure in O(log n) time.
For answering the coverage-report-queries and the candidate-center-queries,
we have the following Lemmas 4 and 5. To prove them, we introduce a connector-
bounded centroid decomposition on the tree T in Section 4. The proof of Lemma 4
is thus given in Section 5, and the proof of Lemma 5 is omitted. Using these
results, we can eventually obtain Theorem 1.
Lemma 4. We can build a data structure A1 in O(M log2 M ) time that can
answer in O(log M log n + k log n) amortized time each coverage-report-query,
i.e., given any point x T , report all active uncertain points covered by x, where
k is the output size. Further, if an uncertain point Pi is deactivated, we can
remove Pi from A1 in O(mi log M log n) amortized time.
Covering Uncertain Points in a Tree 563

Lemma 5. We can build a data structure A2 in O(M log M + n log2 M ) time

that can answer in O(log n) time each candidate-center-query, i.e., given any
vertex v Tm , nd the candidate center c for the active medians of Tm (v).
Further, if an uncertain point Pi is deactivated, we can remove Pi from A2 in
O(log n) time.
Theorem 1. We can nd a minimum number of centers on T to cover all
uncertain points of P in O(M log2 M ) time.

4 A Connector-Bounded Centroid Decomposition

In this section, we propose a tree decomposition of T , called a connector-bounded
centroid decomposition, which will be repeatedly used later (e.g., for Lemmas 1, 4,
5). The decomposition is dierent from the centroid decompositions used before,
e.g., [11, 14] and has certain properties that can facilitate our algorithms.
A vertex v of T is called a centroid if T can be represented as a union of two
subtrees with v as their only common vertex and each subtree has at most 23 of
the vertices of T [11], and we say the two subtrees are decomposed by v. Such a
centroid always exists and can be found in linear time [11]. For convenience, we
consider v to be contained in only one subtree but an open vertex in the other
subtree (thus, the location of P at v only belongs to one subtree).
Our decomposition of T corresponds to a decomposition tree, denoted by
and dened recursively as follows. Each internal node of has two, three, or
four children. The root of corresponds to the entire tree T . Let v be a centroid
of T , and let T1 and T2 be the subtrees of T decomposed by v. Note that T1 and
T2 are disjoint since we consider v to be contained in only one of them. Further,
we call v a connector in both T1 and T2 . Correspondingly, in , its root has two
children corresponding to T1 and T2 , respectively.
In general, consider a node of . Let T () represent the subtree of T corre-
sponding to . We assume T () has at most two connectors (initially this is true
when is the root). We further decompose T () into subtrees that correspond to
the children of in , as follows. Let v be the centroid of T () and let T1 () and
T2 () respectively be the two subtrees of T () decomposed by v. We consider v
as a connector in both T1 () and T2 ().
If T () has at most one connector, then each of T1 () and T2 () has at most
two connectors. In this case, has two children corresponding to T1 () and
T2 (), respectively.
If T () has two connectors but each of T1 () and T2 () still has at most
two connectors (with v as a new connector), then has two children correspond-
ing to T1 () and T2 (), respectively. Otherwise, one of them, say, T2 (), has
three connectors and the other T1 () has only one connector (e.g., see Fig. 2).
In this case, has a child in corresponding to T1 (), and we further perform
a connector-reducing decomposition on T2 (), as follows (this is the main dif-
ference between our decomposition and the traditional centroid decomposition
used before [11, 14]). Depending on whether the three connectors of T2 () are in
a simple path, there are two cases.
564 H. Wang and J. Zhang

y2

T1 ()
v
v

y1
T2 ()

Fig. 2. Illustrating the decomposition of T () into four subtrees in the (red) dashed
cycles, where y1 and y2 are two connectors of T (). T () is rst decomposed into
two subtrees T1 () and T2 (). However, since T2 () has three connectors, we further
decompose it into three subtrees each of which has at most two connectors.

1. If they are in a simple path, without loss of generality, we assume v is the

one between the other two connectors in the path. We decompose T2 () into
two subtrees at v such that they contain the two connectors respectively. In
this way, each subtree contains at most two connectors. Correspondingly,
has another two children corresponding the two subtrees of T2 (), and thus
has three children in total.
2. Otherwise, there is a unique vertex v  in T2 () that decomposes T2 () into
three subtrees that contain the three connectors respectively (e.g., see Fig. 2).
Note that v  and the three subtrees can be easily found in linear time by
traversing T2 (). Correspondingly, has another three children correspond-
ing to the above three subtrees of T2 (), respectively, and thus has four
children in total. Note that we consider v  as a connector in each of the above
three subtrees. Thus, each subtree contains at most two connectors.

We continue the decomposition until each subtree T () of becomes

an edge e(v1 , v2 ) of T . According to our decomposition, both v1 and v2 are
connectors of T (), but they may be only open vertices of T (). If both v1 and
v2 are open vertices of T (), then we will not further decompose T (), so is a
leaf of . Otherwise, we further decompose T () into an open edge and a closed
vertex vi if vi is contained in T () for each i = 1, 2. Correspondingly, has
either two or three children that are leaves of . In this way, for each leaf of
, T () is either an open edge or a closed vertex of T . In the former case, T ()
has two connectors that are its incident vertices, and in the latter case, T () has
one connector that is itself.
This nishes the decomposition of T . A major dierence between our decom-
position and the traditional decomposition [11, 14] is that each subtree in our
decomposition has at most two connectors. As will be clear later, this property
is crucial to guarantee the runtime of our algorithms.

Lemma 6. The height of is O(log M ) and has O(M ) nodes. The connector-
bounded centroid decomposition of T can be computed in O(M log M ) time.
Covering Uncertain Points in a Tree 565

In the following paper, we assume that our decomposition of T and the

decomposition tree have been computed. We introduce some notation that
will be used later. For each node of , we use T () to represent the subtree
of T corresponding to . If y is a connector of T (), then we use T (y, ) to
represent the subtree of T consisting of all points q of T \ T () such that (q, p)
contains y for any point p T () (i.e., T (y, ) is the outside world connecting
to T () through y; e.g., see Fig. 3). By this denition, if y is the only connector
of T (), then T = T () T (y, ); if T () has two connectors y1 and y2 , then
T = T () T (y1 , ) T (y2 , ).

5 The Data Structure A1

In this section, we present the data structure A1 for Lemma 4. The data structure
A1 is for answering the coverage-report-queries, i.e., given any point x T , nd
all active uncertain points that are covered by x. Further, it also supports the
operation of removing an uncertain point once it is deactivated.
Consider any node . If is the root, let L() = ; otherwise, dene
L() to be the sorted list of all indices i [1, n] such that Pi does not have
any locations in the subtree T () but has at least one location in T ( ), where
 is the parent of . Let y be any connector of T (). Let L(y, ) be an index
list the same as L() and each index i L(y, ) is associated with two values:
F (i, y, ), which is the probability sum of Pi in the subtree T (y, ), and D(i, y, ),
which is the expected
 distance from y to the locations of Pi in T (y, ), i.e.,
D(i, y, ) = wi pij T (y,) fij d(pij , y). We refer to L() and L(y, ) for each
connector y T () as the information lists of .
Lemma 7. Suppose L() = and the information lists of are available. Let
t be the number of indices in L(). Then, we can build a data structure of O(t )
size in O(|T ()|+t log t ) time on T (), such that given any point x T (), we
can report all indices i of L() such that Pi is covered by x in O(log n + k log n)
amortized time, where k is the output size; further, if Pi is deactivated with
i L(), then we can remove i from the data structure and all information lists
of in O(log n) amortized time.
Proof. As L() = , is not the root. Thus, T () has one or two connectors.
We only discuss the most general case where T () has two connectors since the
other case is similar but easier. Let y1 and y2 denote the two connectors of T (),
respectively. So the lists L(y1 , ) and L(y2 , ) are available. Note that for any
two points p and q in T (), (p, q) is also in T () since T () is connected.
Consider any point x T (). Suppose we traverse on T () from x to y1 , and
let qx be the rst point on (y1 , y2 ) we encounter (e.g. see Fig 4; so qx is x if
x (y1 , y2 )). Let ax = d(x, qx ) and bx = d(qx , y1 ). Thus, d(y1 , x) = ax + bx
and d(y2 , x) = ax + d(y1 , y2 ) bx .
For any i L(), since Pi does not have any location  in T (), we have
F (i, y1 , ) + F (i, y2 , ) = 1. Note that Ed(x, Pi ) = wi pij T fij d(x, pij ) = wi
  
pij T (y1 ,) fij d(x, pij )+wi pij T (y2 ,) fij d(x, pij ). Further, pij T (y1 ,) fij
566 H. Wang and J. Zhang

y2
T ()
T (y, )
T (1 ) y
qx
v x
T (2 )

y1

Fig. 3. Illustrating the subtrees T (1 ), T (2 ), Fig. 4. Illustrating the denition of qx

and T (y, ), where y is a connector of T () = in the subtree T () with two connectors
T (1 ) T (2 ). Note that T (y, ) is also y1 and y2 . The path (y1 , y2 ) is high-
T (y, 1 ) as y T (1 ). lighted with thicker (red) segments.


d(x, pij ) = [F (i, y1 , ) (ax + bx ) + D(i, y1 , )] and pij T (y2 ,) fij d(x, pij ) =
F (i, y2 , ) (ax + d(y1 , y2 ) bx ) + D(i, y2 , ). Due to F (i, y1 , ) + F (i, y2 , ) = 1,
we obtain Ed(x, Pi ) = wi [ax + (F (i, y1 , ) F (i, y2 , )) bx + D(i, y1 , ) +
D(i, y2 , ) + F (i, y2 , ) d(y1 , y2 )].
Notice that for any x T (), all above values are constant except ax and bx .
Therefore, if we consider ax and bx as two variables of x, Ed(x, Pi ) is a linear
function of them. In other words, Ed(x, Pi ) denes a plane in R3 , where the
z-coordinates correspond to the values of Ed(x, Pi ) and the x- and y-coordinates
correspond to ax and bx respectively. In the following, we also use Ed(x, Pi ) to
refer to the plane dened by it in R3 .
Remark. This nice property for calculating Ed(x, Pi ) is due to that has at most
two connectors. This is one reason our decomposition requires every subtree T ()
to have at most two connectors.
Recall that x covers Pi if Ed(x, Pi ) . Consider the plane H : z = in
R3 . In general the two planes Ed(x, Pi ) and H intersect at a line li and we
let hi represent the closed half-plane of H bounded by li and above the plane
Ed(x, Pi ). Let x be the point (ax , bx ) in the plane H . An easy observation is
that Ed(x, Pi ) if and only if x hi . Further, we say that li is an upper
bounding line of hi if hi is below li and a lower bounding line otherwise. Observe
that if li is an upper bounding line, then Ed(x, Pi ) if and only if x is below
li ; if li is a lower bounding line, then Ed(x, Pi ) if and only if x is above li .
Given any query point x T (), our goal for answering the query is to nd
all indices i L() such that Pi is covered by x. Based on the above discussions,
we do the following preprocessing. After d(y1 , y2 ) is computed, by using the
information lists of y1 and y2 , we compute all functions Ed(x, Pi ) for all i L()
in O(t ) time. Then, we obtain a set U of all upper bounding lines and a set of
all lower bounding lines on the plane H dened by Ed(x, Pi ) for all i L().
In the following, we rst discuss the upper bounding lines. Let SU denote the
indices i L() such that Pi denes an upper bounding line in U .
Given any point x T (), we rst compute ax and bx . This can be done in
constant time after O(|T ()|) time preprocessing, as follows. In the preprocessing,
for each vertex v of T (), we compute the vertex qv (dened in the similar way
as qx with respect to x) as well as the two values av and bv (dened similarly as
Covering Uncertain Points in a Tree 567

ax and bx , respectively). This can be easily done in O(|T ()|) time by traversing
T () and we omit the details. Given the point x, which is specied by an edge e
containing x, let v be the incident vertex of e closer to y1 and let be the length
of e between v and x. Then, if e is on (y1 , y2 ), we have ax = 0 and bx = bv + .
Otherwise, ax = av + and bx = bv .
After ax and bx are computed, the point x = (ax , bx ) on the plane H is also
obtained. Then, according to our discussion, all uncertain points of SU that are
covered by x correspond to exactly those lines of U above x . Finding the lines
of U above x is actually the dual problem of half-plane range reporting query in
R2 . By using the dynamic convex hull maintenance data structure of Brodal and
Jacob [5], with O(|U | log |U |) time and O(|U |) space preprocessing, for any point
x , we can easily report all lines of U above x in O(log |U |+k log |U |) amortized
time (i.e., by repeating k deletions), where k is the output size, and deleting a
line from U can be done in O(log |U |) amortized time. Clearly, |U | t .
On the set of all lower bounding lines, we do the similar preprocessing, and the
query algorithm is symmetric. Hence, the total preprocessing time is O(|T ()| +
t log t ). Each query takes O(log t + k log t ) amortized time and each remove
operation runs in O(log t ) amortized time. As t n, the lemma follows.

The preprocessing algorithm for A1 consists of the following four steps. First,
we compute the information lists for all nodes of . Second, for each node
, we compute the data structure of Lemma 7. Third, for each i [1, n],
we compute a node list L (i) containing all nodes such that i L().
Fourth, for each leaf of , if T () is a vertex v of T holding a location pij ,
then we maintain at the value Ed(v, Pi ). Before giving the details of the above
processing algorithm, we rst discuss the algorithm for answering the coverage-
report-queries by assuming that the preprocessing work has been done.
Given any point x T , we answer the coverage-report-query as follows. Note
that x is in T (x ) for some leaf x of . For each node in the path of
from the root to x , we apply the query algorithm in Lemma 7 to report all
indices i L() such that x covers Pi . In addition, if T (x ) is a vertex of T
holding a location pij such that Pi is active, then we report i if Ed(v, Pi ), which
is maintained at v, is at most . The proof of the following lemma is omitted.
Lemma 8. Our query algorithm correctly nds all active uncertain points that
are covered by x in O(k + log M log n) amortized time, where k is the output size.
If an uncertain point Pi is deactivated, then we scan the node list L (i) and
for each node L (i), we remove i from the data structure by Lemma 7. The
following lemma implies that the total time is O(mi log M log n).
Lemma 9. For each i [1, n], the number of nodes in L (i) is O(mi log M ).

Lemma 10. t = O(M log M ), and the preprocessing time for construct-
ing the data structure A1 except the second step is O(M log M ).

The proofs of Lemmas 9 and 10 are omitted. As t = O(M log M ),
applying the preprocessing of Lemma 7 on all nodes of takes O(M log2 M )
568 H. Wang and J. Zhang

time and O(M log M ) space in total. Hence, the total preprocessing time of A1
is O(M log2 M ) and the space is O(M log M ). This proves Lemma 4.

References
1. Agarwal, P., Cheng, S.W., Tao, Y., Yi, K.: Indexing uncertain data. In: Proc. of the
28th Symposium on Principles of Database Systems (PODS). pp. 137146 (2009)
2. Agarwal, P., Efrat, A., Sankararaman, S., Zhang, W.: Nearest-neighbor searching
under uncertainty. In: Proc. of the 31st Symposium on Principles of Database
Systems (PODS). pp. 225236 (2012)
3. Agarwal, P., Har-Peled, S., Suri, S., Yldz, H., Zhang, W.: Convex hulls under
uncertainty. In: Proc. of the 22nd Annual European Symposium on Algorithms
(ESA). pp. 3748 (2014)
4. de Berg, M., Roelozen, M., Speckmann, B.: Kinetic 2-centers in the black-box
model. In: Proc. of the 29th Annual Symposium on Computational Geometry
(SoCG). pp. 145154 (2013)
5. Brodal, G., Jacob, R.: Dynamic planar convex hull. In: Proc. of the 43rd IEEE
Symposium on Foundations of Computer Science (FOCS). pp. 617626 (2002)
6. Cole, R.: Slowing down sorting networks to obtain faster sorting algorithms. Jour-
nal of the ACM 34(1), 200208 (1987)
7. Huang, L., Li, J.: Stochasitc k-center and j-at-center problems. In: Proc. of the
28th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA). pp. 110
129 (2017)
8. Jrgensen, A., Loer, M., Phillips, J.: Geometric computations on indecisive points.
In: Proc. of the 12nd Algorithms and Data Structures Symposium (WADS). pp.
536547 (2011)
9. Kamousi, P., Chan, T., Suri, S.: Closest pair and the post oce problem for stochas-
tic points. In: Proc. of the 12nd International Workshop on Algorithms and Data
Structures (WADS). pp. 548559 (2011)
10. Kamousi, P., Chan, T., Suri, S.: Stochastic minimum spanning trees in Euclidean
spaces. In: Proc. of the 27th Annual Symposium on Computational Geometry
(SoCG). pp. 6574 (2011)
11. Kariv, O., Hakimi, S.: An algorithmic approach to network location problems. I:
The p-centers. SIAM J. on Applied Mathematics 37(3), 513538 (1979)
12. Loer, M., van Kreveld, M.: Largest bounding box, smallest diameter, and related
problems on imprecise points. Computational Geometry: Theory and Applications
43(4), 419433 (2010)
13. Megiddo, N.: Linear-time algorithms for linear programming in R3 and related
problems. SIAM Journal on Computing 12(4), 759776 (1983)
14. Megiddo, N., Tamir, A.: New results on the complexity of p-centre problems. SIAM
Journal on Computing 12(4), 751758 (1983)
15. Tao, Y., Xiao, X., Cheng, R.: Range search on multidimensional uncertain data.
ACM Transactions on Database Systems 32 (2007)
16. Wang, H., Zhang, J.: Computing the center of uncertain points on tree networks.
Algorithmica 78(1), 232254 (2017)
17. Wang, H., Zhang, J.: One-dimensional k-center on uncertain data. Theoretical
Computer Science 602, 114124 (2015)
18. Wang, H., Zhang, J.: Covering uncertain points in a tree. arXiv:1704.07497 (2017)
Stochastic Closest-pair Problem and Most-likely
Nearest-neighbor Search in Tree Spaces

( )
Jie Xue and Yuan Li

University of Minnesota Twin Cities, Minneapolis MN 55455, USA

{xuexx193,lixx2100}@umn.edu

Abstract. Let T be a tree space represented by a weighted tree with

t vertices, and S be a set of n stochastic points in T , each of which
has a xed location with an independent existence probability. We in-
vestigate two fundamental problems under such a stochastic setting, the
closest-pair problem and the nearest-neighbor search. For the former, we
propose the rst algorithm of computing the -threshold probability and
the expectation of the closest-pair distance of a realization of S. For the
latter, we study the k most-likely nearest-neighbor search (k-LNN) via a
notion called the k most-likely Voronoi Diagram (k-LVD), where we show
the combinatorial complexity of k-LVD is O(nk) under two reasonable
assumptions, leading to a logarithmic query time for k-LNN.

1 Introduction
In many real-world applications, due to the existence of noise or limitations
of devices, the data obtained may be imprecise or not totally reliable. In this
situation, the dataset may fail to capture well the features of the data. Motivated
by this, the topic of uncertain data has received signicant attention in the last
few decades. Many classical problems have been investigated under uncertainty,
including convex hull, minimum spanning tree, range search, linear separability,
etc. [1,2,3,4,5,7,8,10,13,15,16]. Among these, there are two common models of
uncertainty: existential uncertainty and locational uncertainty. In the former,
each (stochastic) data point has a xed location with an uncertain existence
depicted by an independent existence probability, while in the latter the location
of each point is uncertain and described as a distribution.
The closest-pair problem and nearest-neighbor search are two interrelated
fundamental problems, which have numerous applications in various areas. The
uncertain versions of both the problems have also been studied recently in
[1,9,11,12,14]. Let S be a set of n stochastic points in some metric space X . Con-
cerning the closest pair problem, a basic question one may ask is how to compute
elementary statistics about the stochastic closest-pair of S, e.g., the probability
that the closest-pair distance of a realization of S is at least , the expected
closest-pair distance, etc. Unfortunately, most problems of this kind have been
shown to be NP-hard or #P-hard for general metrics, and some of them remain
#P-hard even when X = Rd for d 2 [9,11]. Concerning the nearest-neighbor
search, an important problem is the most-likely nearest-neighbor (LNN) search

Springer International Publishing AG 2017 569

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 569580, 2017.
DOI: 10.1007/978-3-319-62127-2_48
570 J. Xue and Y. Li

[14], which looks for the data point in S with the greatest probability of being
the nearest-neighbor of a query point q. The LNN search introduces the concept
of most-likely Voronoi diagram (LVD), which decomposes X into connected cells
such that the query points in the same cell have the same LNN. However, as
in [12,14], the bound of LVD in Rd is still high even on average. Due to the
diculties of both problems in general and Euclidean space, it is then natural
to ask whether these problems are relatively easier in other metric spaces such
as a tree space. Indeed, further exploring these problems in tree spaces will be
helpful and interesting since any nite metrics (say a road network in practice)
can be embedded on a tree space under some reasonable distortions [6].
With the above motivations, in this paper, we study the stochastic closest-
pair (SCP) problem and k most-likely nearest-neighbor (k-LNN) search in tree
spaces. A tree space T is represented by a positively-weighted tree T where the
weight of each edge depicts its length. Formally, T is the geometric realization
of T , in which each edge weighted by w is isometric to the interval [0, w]. There
is a natural metric over T which denes the distance dist(x, y) as the length
of the (unique) simple path between x and y in T . See Fig. 1 for an exam-
ple of tree space. Following [9,11,14], we study the problems under existential
uncertainty: each stochastic point has a xed location (in T ) associated with
an (independent) existence probability. Due to limited space, the proofs of all
lemmas and some theorems are omitted and can be found in the full version [17].

Our results. Let T be a tree space represented by a 2.2 1.4

t-vertex weighted tree T , and S be the given set of n y 2
stochastic points in T each of which is associated with x 4.5 3.2
an existence probability. A realization of S refers to 2.8
a random sample of S in which each point is sampled dist(x, y) = 6.9
with its existence probability. 4.1

For the SCP problem, dene (S) as a random

variable indicating the closest-pair distance of a real- Fig. 1. A tree space and
ization of S. We rst show that the -threshold proba- the unique simple path (in
bility of (S) (i.e., the probability that (S) is at least bold) between x and y
) can be computed in O(t + n log n + min{tn, n2 }) time for any given positive
threshold . Based on this, we immediately obtain an O(t+min{tn3 , n4 })-time al-
gorithm for computing the expected closest-pair distance, i.e., the expectation of
(S). We then further show that one can approximate the expected closest-pair
distance within a factor of (1+) in O(t+1 min{tn2 , n3 }) time, by arguing that
the expected closest-pair distance can be approximated via O(1 n) threshold
probability queries.
For the LNN search, we rst study the size of the the k-LVD TS of S on
T . A matching O(n2 ) upper bound for the worst-case size of TS is given. More
interestingly, we show that (1) the worst-case size of TS is O(kn), if the existence
probabilities of the points in S are constant-far from 0; (2) the average-case size of
TS is O(kn), if the existence probabilities are i.i.d. random variables drawn from
a xed distribution. These results further imply the existence of an LVD data
Stochastic Closest-pair Problem and Most-likely Nearest-neighbor Search in Tree Spaces 571

structure which answers k-LNN queries in O(log n + k) time using average-case

O(t+k 2 n) space, and worst-case O(t+k 2 n) space if the existence probabilities of
the points are constant-far from 0. Finally, we give an O(t + n2 log n + n2 k)-time
algorithm to construct such a data structure.

2 The stochastic closest-pair problem

Let T be a tree space represented by a t-vertex weighted tree T and S =

{a1 , . . . , an } T be a set of stochastic points where ai has an existence proba-
bility ai . We use (S) to denote the random variable indicating the closest-pair
distance of a realization of S (if the realization is of size less than 2, we simply
set its closest-pair distance to be 0).

2.1 Computing the threshold probability

We study the problem of computing the probability that (S) is at least  for
a given threshold . We call this quantity the -threshold probability or simply
threshold probability of (S), and denote it by C (S). We show that C (S) can
be computed in O(t+n log n+min{tn, n2 }) time. This result gives us an O(t+n2 )
upper bound for t = (n) and an O(n log n + tn) bound for t = O(n). In the
rest of this section, we rst present an O(t + n3 )-time algorithm for computing
C (S), and then show how to improve it to achieve the desired bound. For
simplicity of exposition, we assume a1 , . . . , an have distinct locations in T .
An O(t + n3 )-time algorithm. In order to conveniently and eciently handle
the stochastic points in a tree space, we begin with a preprocessing step, which
reduces the problem to a more regular setting.
Theorem 1. Given T and S, one can compute in O(t + n log n) time a new
tree space T  T represented by an O(n)-vertex weighted tree T  s.t. S T 
and every point in S is located at some vertex of T  . (See [17] for a proof.)
By the above theorem, we use O(t + n log n) time to compute such a new tree
space. Using this tree space as well as the O(n)-vertex tree representing it,
the problem becomes more regular: every stochastic point in S is located at a
vertex. We can further put the stochastic points in one-to-one correspondence
with the vertices by adding dummy points with existence probability 0 to the
empty vertices. In such a regular setting, we then consider how to compute
the -threshold probability. For convenience, we still use T to denote the repre-
sentation of the (new) tree space and S = {a1 , . . . , an } the stochastic dataset
(though the actual size of S may be larger than n due to the additional dummy
points, it is still bounded by O(n)). Since the vertices of T are now in one-to-one
correspondence with the points in S, we also use ai to denote the corresponding
vertex of T .
As we are working on a tree space, a natural idea for solving the problem
is to exploit the recursive structure of the tree and to compute C (S) in a
recursive fashion. To this end, we need to dene an important concept called
572 J. Xue and Y. Li

witness. We make T rooted by setting a1 as its root. The subtree rooted at a

vertex x is denoted by Tx . Also, we use V (Tx ) to denote the set of the stochastic
points lying in Tx , or equivalently, the set of the vertices of Tx . The notations
p(x) and ch(x) are used to denote the parent of x and the set of the children of
x, respectively (for convenience we set p(a1 ) = a1 ).
Denition 1. Let dep(ai ) be the depth of ai in T , i.e., dep(ai ) = dist(a1 , ai ).
For any ai and aj , we dene ai aj if dep(ai ) < dep(aj ), or dep(ai ) = dep(aj )
and i < j. Clearly, the relation is a strict total order over S (also, over the
vertices of T ). For any subset S  S and any vertex ai of T , we dene the
witness of ai with respect to S  , denoted by (ai , S  ), as the smallest vertex in
V (Tai )S  under the -order. If V (Tai )S  = , we say (ai , S  ) is not dened.
See Fig. 2 for an illustration of witness. We say a subset S  S is legal if the
closest-pair distance of S  is at least .
The following lemma allows us to verify
a1
the legality of a subset by using the witnesses, (a1 , S  ) = a3
2.2
which will be used later. 4.1 a2
(a6 , S  ) = a9
Lemma 1. For any S  S, we have S  is a6
4.4 2
a3
legal if and only if every point ai S\{a1 } a7
1.4
a5 2.8
satises one of the following three conditions: 2 3.2
(1) (ai , S  ) is not dened; a8 a9
a4

(2) (ai , S  ) = (p(ai ), S  ); 

S = {a3 , a5 , a8 , a9 }
(3) dist((ai , S  ), (p(ai ), S  )) .
We say that S  is locally legal at ai whenever Fig. 2. An illustration of witness
ai satises one of the above conditions.
In order to compute C (S), we dene, for all
x S and y V (Tp(x) ),


S  RPr [S  is legal and (x, S  ) = y] if y V (Tx ),
V (Tx )
Py (x) =

Pr [S  {y} is legal and (p(x), S  {y}) = y] if y V (Tp(x) )\V (Tx ).
S  R V (Tx )

Here the notation R means that the former is a realization of the latter, i.e., a
random sample obtained by sampling each point with its existence probability.
n
With the above, we immediately have that C (S) = i=1 Pai (a1 ) P0 , where
P0 is the probability that a realization of S contains exactly one point. We then
show how Py (x) can be computed in a recursive way.
Lemma 2. For x S and y V (Tx ), we have that

Py (x) = Q Py (c),
cch(x)

where Q = x if x = y and Q = 1 x if x = y.
Lemma 3. For x S and y V (Tp(x) )\V (Tx ), we have that

Py (x) = (1 ai ) + Pz (x),
ai V (Tx ) z

where = {z V (Tx ) : y z and dist(z, y) }.

Stochastic Closest-pair Problem and Most-likely Nearest-neighbor Search in Tree Spaces 573

By the above two lemmas, the values of all Py (x) can be computed as follows.
We enumerate x S from the greatest to the smallest under -order. For each
x, we rst compute all Py (x) for y V (Tx ) by applying Lemma 2. After this, we
are able to compute all Py (x) for y V (Tp(x) )\V (Tx ) by applying Lemma 3. The
entire process takes O(n3 ) time. Once we have the values of all Py (x), C (S)
can be computed straightforwardly. Including the time for preprocessing, this
gives us an O(t + n3 )-time algorithm for computing C (S).
In fact, we can further improve the runtime above to O(t+n2 ) by speeding up
the computation of Py (x) for y V (Tp(x) )\V (Tx ) as they are the bottlenecks. In
addition, if t = O(n), we can even further reduce the runtime to O(t + n log n +
min(tn, n2 )). Both optimizations are nontrivial and need new insights. However,
due to the limited space, we leave these to [17] and conclude the following.
Theorem 2. Given a weighted tree T with t vertices and a set S of n stochastic
points in its tree space T , one can compute the -threshold probability of the
closest-pair distance of S, C (S), in O(t + n log n + min{tn, n2 }) time.

2.2 Computing the expected closest-pair distance

Based on our algorithm for computing the threshold probability, we further
study the problem of computing the expected closest-pair distance of S, i.e.,
the expectation of (S). It is easy to see that our algorithm in Section 2.1
immediately gives us an O(t+min{tn3 , n4 })-time algorithm

 to compute E[(S)].
This is because the random variable (S) has at most n2 distinct possible values
and hence we can compute E[(S)] via O(n2 ) threshold probability queries
with various thresholds  (note that after preprocessing our algorithm answers
each threshold probability query in O(min{tn, n2 }) time).
If we want to compute the exact value of E[(S)] (via threshold probability
queries), (n2 ) queries are necessary in worst case. So it is natural to ask whether
we can use less queries to approximate E[(S)]. In the rest of this section, we
show that one can use O(1 n) threshold probability queries to achieve a (1+)-
approximation for E[(S)], which in turn gives us an O(t + 1 min{tn2 , n3 })-
time approximation algorithm for computing E[(S)].
For simplicity of exposition, we assume that the stochastic points in S are
now one-to-one corresponding to the vertices of T (this is what we have after
preprocessing). We begin with a simple case, in which the spread of T , i.e.,
the ratio of the length of the longest edge to the length of the shortest edge is
bounded by some polynomial of n. In this case, to approximate E[(S)] is fairly
easy, and we only need O(1 log n) threshold probability queries.
Denition 2. For > > 0 and > 1, the (, , )-jump is dened as
J = {, , 2 , . . . , k , },

where k < and k+1 .

Let dmin be the length of the shortest edge of T and dmax be the sum of the
lengths of all edges of T . Also, let J be the (dmin , dmax , 1+)-jump. Suppose J =
574 J. Xue and Y. Li

{1 , . . . , |J| }. Then we do |J| threshold probability queries using the thresholds
1 , . . . , |J| , and compute
|J|

E= Ci (S) (i i1 )
i=1

as an approximation of E[(S)] (where 0 = 0). Note that |J| = O(log1+ ddmaxmin

)=
O(log1+ n) = O(1 log n). It is easy to verify that E E[(S)] (1 + )E.
The problem becomes interesting when the spread of T is unbounded. In
this case, although the above method still correctly approximates E[(S)], the
number of the threshold probability queries is no longer well bounded. Imagine
that the O(n2 ) possible values of (S) are distributed as , (1+), (1+)2 , etc.
Then the (dmin , dmax , 1 + )-jump J is of size (n2 ). Moreover, for guaranteeing
the correctness, it seems that we cannot skip any element in J. However, as
one will realize later, such an extreme situation can never happen. Recall that
we are working on a weighted tree and the O(n2 ) possible values of (S) are
indeed the pairwise distances of the vertices of the tree. As such, these values
are not arbitrary, and our insight here is to exploit the underlying properties of
the distribution of these values.
Let e1 , . . . , en1 be the edges of T where ei has the
 length (weight) wi .
m1
Assume w1 wn1 . We dene an index set I = m : i=1 wi < wm .
Suppose I = {m1 , . . . , mk } where m1 < < mk . Note that m1 = 1. For
convenience, we set mk+1 = n. We design our threshold probability  queries
as follows. Let Ji be the (wmi , si , 1 + )-jump where si = j<mi+1 j , and
w
J = J1 Jk . Suppose J = {1 , . . . , |J| } and set 0 = 0. Similarly to
the previous case, we do |J| threshold probability queries using the thresholds
1 , . . . , |J| , and compute
|J|

E= Ci (S) (i i1 )
i=1

as an approximation of E[(S)]. We rst verify the correctness, i.e., E E[(S)]

(1 + )E. The fact E E[(S)] can be easily veried. To see the inequality
E[(S)] (1 + )E, we dene a piecewise-constant function h : R+ {0} [0, 1]
as
Ci (S) if (1 + )i <  (1 + )i+1 ,
h() = 0 if  > (1 + )l|J| ,

1 if  = 0.
< < <
Then it is clear that (1+)E = 0 h()d. We claim that 0 h()d 0 C (S)d,
hence we have E[(S)] (1 + )E. Note that the jumps J1 , . . . , Jk are disjoint
and each of them contains a consecutive portion of the sequence 1 , . . . , |J| . Fur-
thermore, if i and i+1 belong to dierent jumps, then there is no possible value
of (S) within the range (i , i+1 ), i.e., C (S) is constant when  [i , i+1 ).
With this observation, it is not dicult to verify that h() C (S) for any
 0. Consequently, we have E[(S)] (1 + )E, which implies the correctness
of our method. Now the only thing remaining is to bound the number of the
threshold probability queries, which we show in Lemma 4.
Stochastic Closest-pair Problem and Most-likely Nearest-neighbor Search in Tree Spaces 575

Lemma 4. For each jump Ji , we have |Ji | = O(1 (mi+1 mi )). As a result,
the total number of the threshold probability queries, |J|, is O(1 n).

Indeed, the above method can be extended to a much more general case, in
which the stochastic dataset S is given in any metric space X (not necessarily
a tree space). In this case, one can still dene the threshold probability C (S)
as well as the expected closest-pair distance E[(S)] in the same fashion. Our
conclusion is the following.

Theorem 3. Given a set S of n stochastic points in a metric space X , one can

(1+)-approximate the expected closest-pair distance of S, E[(S)], via O(1 n)
threshold probability queries. (See [17] for a proof.)

For the expected closest-pair distance in tree space, we can eventually conclude
the following by plugging in our algorithm in Section 2.1 for computing C (S).

Corollary 1. Given a tree space T represented by a weighted tree T with t

vertices and a set S of n stochastic points in T , one can compute a (1 + )-
approximation for the expected closest-pair distance of S, E[(S)], in O(t +
1 min{tn2 , n3 }) time.

3 The most-likely nearest-neighbor search problem

In this section, we study the k most-likely nearest-neighbor (k-LNN) search in a

tree space. Again, let T be a tree space represented by a t-vertex weighted tree
T and S = {a1 , . . . , an } T be the given stochastic dataset where the point
ai has an existence probability ai . The k-LNN search problem can be dened
as follows. Let q T be any point. For each ai S, dene NNP q (ai ) as the
probability that the nearest-neighbor of q in a realization of S is ai . Clearly, the
nearest-neighbor of q in a realization is ai i ai is in the realization and any
point closer to q is not in the realization. Therefore, we have

NNP q (ai ) = ai (1 x ),
x

where = {x S : dist(q, x) < dist(q, ai )}. Given a query point q T ,

the goal of the k-LNN search is to report the k-LNN of q, which is a k-sequence
(ai1 , . . . , aik ) of points in S such that NNP q (ai1 ) NNP q (aik ) NNP q (aj )
for all j / {i1 , . . . , ik }. For convenience, we assume NNP q (ai ) = NNP q (aj ) for
any q T and ai = aj so that the k-LNN of any query point q T is uniquely
dened.
A standard tool for nearest-neighbor search is the Voronoi diagram. In stochas-
tic setting, we seek the most-likely Voronoi diagram (LVD), the concept of which
is for the rst time introduced in [14]. The k-LVD partitions the query space
into connected cells such that points in the same cell have the same k-LNN. See
[17] for an example (in color) of 1-LVD in a tree space.
576 J. Xue and Y. Li

1.5
2.3

2.5 2.2 2.9 3.6

1.4 2
2.2 3.2
2.8
3 2

1 A center c

Fig. 3. A degree-3 center involving 5 points.

3.1 The size of the tree-space LVD

We use TS to denote the k-LVD of S on T , i.e., the collection of the cells.
Formally, TS can be dened as follows. For any k-sequence = (ai1 , . . . , aik ),
let be the set of the connected components of the subspace {q T :
is the k-LNN of q}. Then TS is the union of over all possible . Clearly,
the size of TS signicantly inuences the space eciency of the LVD-based al-
gorithm for k-LNN search. Let mij T be the midpoint of ai and aj , i.e., the
midpoint of the path between
 ai and aj in T . It is easy to see that the k-LNN
only changes nearby these n2 midpoints. However, this does not immediately
imply that the size of TS is bounded by O(n2 ). The reason is that O(n2 ) points
do not necessarily decompose T into O(n2 ) pieces (cells), unless these points are
located only in the interiors of the edges. Note that throughout this section, we
do not make any spatial assumption about the midpoints. In other words, it is
allowed that dierent midpoints occupy the same location in T , and some mid-
points are located at the vertices of T . The reason why we allow this is explained
in [17]. It is not surprising that even in such a general setting, the size of TS is
still bounded by O(n2 ). We will see this later as a direct corollary of a technical
result (Lemma 5).
Denition 3. For any two midpoints mij and mi j  , we dene mij mi j  i
mij and mi j  have the same location in T and dist(ai , mij ) = dist(aj , mij ) =
dist(ai , mi j  ) = dist(aj  , mi j  ). Clearly, is an equivalence relation over the
midpoints. We call the equivalence classes (under ) centers of S and use [mij ]
to denote the center that contains mij . A stochastic point ai S is said to be
involved by a center c if c = [mij ] for some j. The degree of a center c,
denoted by deg(c), is dened as the number of the connected components of T \c
that contain at least one point involved by c, where c denotes the point in T
corresponding to c, and each such component is called a branch of c. A center
c is said to be critical if c is not in the interior of any cell C TS and there
exists at least one point involved by c that is in the k-LNN of c. (See Fig. 3 for
an intuitive illustration of a center.)

Lemma 5. Let be the set of the critical centers and = c deg(c). Then
|TS | + 1.

The above lemma immediately gives us the O(n2 ) upper bound for the size of
TS . Indeed, a center c of S contains at least (deg(c) m) midpoints, where m is
the number of the points involved by c, so + 1 is at most O(n2 ). Unfortunately,
this upper bound is tight, following from the (n2 ) worst-case lower bound for
Stochastic Closest-pair Problem and Most-likely Nearest-neighbor Search in Tree Spaces 577

the size of the 1-dim 1-LVD given by [14] (note that the 1-dim LVD is a special
case of the tree-space LVD). Surprisingly, we show that, if we make reasonable
assumptions for the existence probabilities of the stochastic points or consider
the average case, the size of TS is signicantly smaller. Our results are:

If the existence probabilities of all points in S are constant-far from 0, i.e.,

there is a xed constant > 0 such that ai for all ai S, then the size
of the k-LVD TS is O(kn). Note that this assumption about the existence
probabilities is natural and reasonable. In applications, an extremely small
existence probability means the data point is highly unreliable. Such a point
can be considered as a noise and removed from the dataset.
The average-case size of the k-LVD TS is O(kn). For the average-case anal-
ysis we assume that the existence probabilities of the points in S are i.i.d.
random variables drawn from any xed distribution (e.g., the uniform dis-
tribution among [0, 1]). In other words, we consider the expectation of |TS |
when a1 , . . . , an are such random variables. The interesting point is that
the O(kn) upper bound is totally independent of the structure of T and the
locations of the stochastic points. The randomness is only applied to the
existence probabilities in our average-case analysis.

To prove these bounds requires new ideas. By Lemma 5, to bound the size of
TS , it suces to bound the degree-sum of the critical centers. Intuitively, if a
center c is far from the points it involves (compared with other points in S),
then c is less likely to be critical, as the c-involved points are less likely to be in
the k-LNN of c. Along with this intuition, we dene the following.

Denition 4. For any center c, the diameter of c, denoted by diam(c), is

dened as the distance from c to the c-involved points. Let A T be a nite set.
We dene the depth of c with respect to A as dep A (c) = |{x A : dist(x, c) <
diam(c)}|, i.e., the number of the points in A which are closer to c than the
c-involved points.

Our idea here is to rst bound the contribution (degree-sum) of the shallow
centers, and then further bound the degree-sum of the critical ones. Specically,
we investigate in Lemma 6 the degree-sum of the d-shallow centers of S, i.e., the
centers of depth less than d with respect to S.

Lemma 6. For 1 d n 1, the degree-sum of the d-shallow centers of S is

at most 8dn.

Now we are ready to prove the O(kn) bound for |TS | under the constant-far
from 0 assumption about the existence probabilities.

Lemma 7. If the existence probabilities of the points in S are constant-far from

0, then a center of S is critical only if it is O(k)-shallow.

Theorem 4. If the existence probabilities of the points in S are constant-far

from 0, then the size of the k-LVD TS is O(kn).
578 J. Xue and Y. Li

Proof. Suppose the existence probabilities a1 , . . . , an are constant-far from 0.

Lemma 7 shows that all the critical centers of S are O(k)-shallow. By further
applying Lemma 6, the degree-sum of the critical centers is O(kn). Finally, by
Lemma 5, the size of TS is O(kn). 
To prove the bound for the average-case size requires more eorts. Let f be
a xed probability distribution function whose support is in (0, 1] and be
< 0 of the support of f . Dene two constants 0 = /(1 + ) and
the supremum
= 1 f (x)dx. Clearly, if X is a random variable drawn from f , then
= Pr[X > 0 ]. Note that is always positive by denition. The following
lemma claries the meaning of 0 .
Lemma 8. Suppose a1 , . . . , an are i.i.d. random variables drawn from f . For
any center c of S, the event c is critical does not happen if there are k (dis-
tinct) points ai1 , . . . , aik in S closer to c than the c-involved points such that
ai1 , . . . , aik are all greater than 0 .
Theorem 5. The average-case size of TS is O(kn), given that the existence
probabilities of the points in S are i.i.d. random variables drawn from a xed
distribution.
Proof. Suppose the existence probabilities a1 , . . . , an are drawn independently
from f . Lemma 8 implies that, if c is a center of S with dep S (c) = d k, then
 

k1
d i
Pr[c is critical] ud = (1 )di .
i=0
i

Then by applying Lemma 5, we have

  
n1 
E[|TS |] Pr[c is critical] deg(c) deg(c) + (ud1 ud )deg(c),
c cHk d=k+1 cHd

where Hd is the set of the d-shallow centers of S. Observe that

 
d1 k
ud1 ud = (1 )dk .
k1

Based on this and Lemma 6, we further have

 

n1
d1 k
E[|TS |] 8kn + 8n (1 )dk d.
k1
d=k+1

Note that
   k n1  

n1
d1 k  d
(1 ) dk
d=k (1 )d .
k1 1 k
d=k+1 d=k+1

By an induction argument on k, it is not dicult to see that

   

n1
d 

d (1 )k
(1 ) <
d
(1 )d = .
k k k+1
d=k+1 d=k

8kn
Finally, by combining the inequalities, E[|TS |] 8kn + = O(kn). 
Stochastic Closest-pair Problem and Most-likely Nearest-neighbor Search in Tree Spaces 579

3.2 Constructing LVD and answering queries

In this section, we show how to construct the k-LVD TS and use it to answer
k-LNN queries. Let e1 , . . . , et1 be the edges of T . Assume each edge ei has a
specied start point si (which is one of its two endpoints) and the query point
q is specied via a pair (i, ), meaning the point on ei with distance to si .
We rst explain the data structure used for storing the k-LVD TS and an-
swering queries. The LVD data structure is simple. First, it contains |TS | arrays
(called answer arrays) each of which stores the k-LNN answer of one cell of
TS . This part takes O(k|TS |) space. In addition to that, we also need to record
the structure of TS . For each edge ei of T , we use a sorted list Li to store the
cell-decomposition of ei . Specically, the intersection of each cell C TS and
ei is an interval (may be empty). These intervals are stored in Li in the order
they appear on ei . Note that this part takes O(t + |TS |) space. Indeed, if an edge
is decomposed into p pieces (intervals) by TS , then it at least entirely contains
(p 2) cells of TS (so we can charge these (p 2) pieces to the corresponding
cells and the remaining two pieces to the edge). Therefore, the total space of the
LVD data structure is O(t + k|TS |). To answer a query q = (i, ), we rst do
a binary search in the list Li to know which cell q locates in, and then use the
answer array corresponding to the cell to output the k-LNN of q directly. The
query time is clearly O(log |TS | + k).
Next, we consider the construction of the LVD data structure. The rst
step of the construction is to compute all the centers of S and sort the cen-
ters in the interior of each edge e in the order they appear on e. We are able
to get this done in O(t + n2 log n) time (see [17]). After the centers are com-
puted and sorted, we begin to construct the LVD data structure. Choose a
vertex v of T . Starting at v, we do a walk in T along with the edges of T . The
walk visits each edge of T exactly twice and nally goes back to v; see Fig. 4.
During the walk, we maintain a (balanced) binary search
tree for NNP x (a1 ), . . . , NNP x (an ) w.r.t. the current loca-
tion x. By exploiting this BST, we can work out the cell-
decomposition of each edge ei (i.e., the sorted list Li ) at the v
rst time we visit ei in the walk. Specically, we track the
k-LNN when walking along with ei , which can be obtained
by retrieving the k largest elements from the BST. When-
ever the k-LNN changes, a new cell of TS is found, so we Fig. 4. A walk in
need to create a new answer array to store the k-LNN infor- tree visiting each
mation. Also, we need to update the sorted list Li . In this edge exactly twice.
way, after we go through ei (for the rst time), Li is cor-
rectly computed. At the second time we visit ei , we do nothing but maintain the
binary search tree. When we nish the walk and go back to v, the construction
of the LVD data structure is done. Clearly, in the process of the walk, we only
need to maintain the binary search tree and retrieve the k-LNN when we arrive
at (resp., leave from) a center of S from (resp., to) one of its branches. With a
careful implementation and analysis (see [17]), we can complete the entire walk
580 J. Xue and Y. Li

and hence the entire LVD structure in O(t + n2 log n + n2 k) time. Combined
with the bounds in Section 3.1, we then have the following results.
Theorem 6. Given a tree space T represented by a t-vertex weighted tree and
a set S of n stochastic points in T , one can construct in O(t + n2 log n + n2 k)
time an LVD data structure to answer k-LNN queries in O(log n + k) time. The
LVD data structure uses worst-case O(t+kn2 ) space and average-case O(t+k 2 n)
space. Furthermore, if the existence probabilities of the points in S are constant-
far from 0, then the LVD data structure uses worst-case O(t + k 2 n) space.

References
1. Agarwal, P., Aronov, B., Har-Peled, S., Phillips, J., Yi, K., Zhang, W.: Nearest
neighbor searching under uncertainty II. In: Proc. of the 32nd Sympos. on PODS.
pp. 115126. ACM (2013)
2. Agarwal, P., Cheng, S.W., Yi, K.: Range searching on uncertain data. ACM Trans-
actions on Algorithms 8(4), 43 (2012)
3. Agarwal, P., Har-Peled, S., Suri, S., Yldz, H., Zhang, W.: Convex hulls under
uncertainty. In: Algorithms-ESA, pp. 3748. Springer (2014)
4. Agarwal, P., Kumar, N., Sintos, S., Suri, S.: Range-max queries on uncertain data.
In: Proc. of the 35th SIGMOD/PODS. pp. 465476. ACM (2016)
5. Chen, J., Feng, L.: Ecient pruning algorithm for top-k ranking on dataset with
value uncertainty. In: Proc. of the 22nd CIKM. pp. 22312236. ACM (2013)
6. Fakcharoenphol, J., Rao, S., Talwar, K.: Approximating metrics by tree metrics.
ACM SIGACT News 35(2), 6070 (2004)
7. Fink, M., Hershberger, J., Kumar, N., Suri, S.: Hyperplane separability and con-
vexity of probabilistic point sets. In: Proc. of the 32nd SoCG. ACM (2016)
8. Ge, T., Zdonik, S., Madden, S.: Top-k queries on uncertain data: on score distri-
bution and typical answers. In: Proc. of the 2009 SIGMOD. pp. 375388. ACM
(2009)
9. Huang, L., Li, J.: Approximating the expected values for combinatorial optimiza-
tion problems over stochastic points. In: Intl. Colloquium on Automata, Languages,
and Programming. pp. 910921. Springer (2015)
10. Kamousi, P., Chan, T., Suri, S.: Stochastic minimum spanning trees in Euclidean
spaces. In: Proc. of the 27th SoCG. pp. 6574. ACM (2011)
11. Kamousi, P., Chan, T., Suri, S.: Closest pair and the post oce problem for stochas-
tic points. Computational Geometry 47(2), 214223 (2014)
12. Kumar, N., Raichel, B., Suri, S., Verbeek, K.: Most likely Voronoi Diagrams in
higher dimensions. In: LIPIcs-Leibniz International Proceedings in Informatics.
vol. 65. Schloss Dagstuhl-Leibniz-Zentrum fuer Informatik (2016)
13. Loer, M., van Kreveld, M.: Largest and smallest convex hulls for imprecise points.
Algorithmica 56(2), 235269 (2010)
14. Suri, S., Verbeek, K.: On the most likely Voronoi Diagram and nearest neighbor
searching. In: ISAAC. pp. 338350. Springer (2014)
15. Suri, S., Verbeek, K., Yldz, H.: On the most likely convex hull of uncertain points.
In: AlgorithmsESA, pp. 791802. Springer (2013)
16. Xue, J., Li, Y., Janardan, R.: On the separability of stochastic geometric objects,
with applications. In: Proc. of the 32nd SoCG. ACM (2016)
17. Xue, J., Li, Y.: Stochastic closest-pair problem and most-likely nearest-neighbor
search in tree spaces. arXiv:1612.04890 (2016)
 On the Expected Diameter, Width, and
Complexity of a Stochastic Convex-hull

Jie Xue( )
, Yuan Li, and Ravi Janardan

University of Minnesota Twin Cities, Minneapolis MN 55455, USA

{xuexx193,lixx2100,janardan}@umn.edu

Abstract. We investigate several computational problems related to

the stochastic convex hull (SCH). Given a stochastic dataset consist-
ing of n points in Rd each of which has an existence probability, a SCH
refers to the convex hull of a realization of the dataset, i.e., a random
sample including each point with its existence probability. We are in-
terested in computing certain expected statistics of a SCH, including
diameter, width, and combinatorial complexity. For diameter, we estab-
lish the rst deterministic 1.633-approximation algorithm with a time
complexity polynomial in both n and d. For width, two approximation
algorithms are provided: a deterministic O(1)-approximation running in
O(nd+1 log n) time, and a fully polynomial-time randomized approxima-
tion scheme (FPRAS). For combinatorial complexity, we propose an ex-
act O(nd )-time algorithm. Our solutions exploit many geometric insights
in Euclidean space, some of which might be of independent interest.

Keywords: uncertain data, expectation, diameter, width, combinato-

rial complexity

1 Introduction
The convex hull, which is one of the most fundamental structures in computa-
tional geometry, has a wide range of applications in various areas. Traditionally,
the convex hull is studied on datasets whose information is known exactly. How-
ever, in many real-world applications, due to noise and limitation of devices,
the data obtained may be imprecise or not totally reliable. In this situation,
uncertain datasets (or stochastic datasets), in which the data points are allowed
to have some uncertainty, can better model real data. In recent years, there
have been several papers regarding the convex hull structure under uncertainty,
known as stochastic convex hull (SCH) [3,10,11,13].
In this paper, we revisit several problems related to SCH under the well-
known existential uncertainty model: each data point in the stochastic dataset
has a certain (known) location in the space with an uncertain existence depicted
by an associated (independent) existence probability. In real-world applications,
the existence probabilities can be used to express the reliability or importance
of the data points. Given a stochastic dataset S in Rd equipped with existential
uncertainty, a SCH of S refers to the convex hull of a realization of S, which can

Springer International Publishing AG 2017 581

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 581592, 2017.
DOI: 10.1007/978-3-319-62127-2_49
582 J. Xue et al.

be regarded as a probabilistic polytope in Rd . Our main focus is to compute the

expected values of some basic statistics of a SCH, including diameter, width, and
combinatorial complexity. (Formal denitions can be found in Sec. 1.1.) We give
polynomial-time algorithms (both exact and approximate) for these problems.
Related work. Geometric computation on uncertain data has received consid-
erable attentions in recent years. Many geometric problems have been studied
under uncertainty (either existential uncertainty or locational uncertainty), e.g.,
nearest-neighbor search [1,12], minimum spanning trees [8], closest pair [5,9,16],
range search [2], linear separability [4,14], dominance relation [15], etc. There
have also been several papers concerning SCH [3,6,10,11,13]. We only summarize
those that are strongly relevant to this paper. Li et al. [10] studied the expected
computation of some basic statistics of a SCH in R2 , e.g., area, perimeter, diame-
ter (their results for diameter are summarized below), etc. The results in [10] are
presented in a slightly dierent uncertainty model, but most of the algorithms
also work under existential uncertainty. Huang et al. [6] studied -coresets of a
stochastic dataset, which can be used to eciently approximate the expected
directional width of a SCH with respect to any given direction (see Sec. 1.1 for
the denition of directional width). One should note that, although the diame-
ter (resp., width) is dened as the largest (resp., smallest) directional width, the
-coresets constructed in [6] cannot be used to approximate the expected diame-
ter/width of a SCH, simply because the direction dening the diameter/width of
a SCH varies from realization to realization. Specically, the expected diameter
of a SCH was investigated in some recent works. Huang and Li [5] provided an
FPRAS for computing the expected farthest-pair distance of a stochastic dataset
in a metric space (which works under both existential and locational uncer-
tainty), which directly implies an FPRAS for computing the expected diameter
of a SCH. However, an FPRAS can only obtain the desired approximation with
high probability, and there seems no way to verify whether an answer obtained
by the
FPRAS is truly a good approximation. Li et al. [10] gave a deterministic
(2/ 3)-approximation algorithm in R2 , which is based on (exactly) computing
the expected diameter of the stochastic smallest enclosing ball. Although [10]
2
only considered
the case in R , the algorithm can be naturally extended to com-
pute a ( 2d/ d + 1)-approximation of the expected diameter of a SCH in Rd .
Nevertheless, the runtime of this algorithm grows exponentially as d increases,
since computing the expected diameter of the stochastic smallest enclosing ball
requires n(d) time [7]. The width and combinatorial complexity of a SCH have
not yet been investigated previously, to our best knowledge.
Our results. The contributions of this paper are as follows:
Expected diameter. We study the expected-diameter problem in Rd without
assuming that d is a xed. We establish an (n, d)-polynomial time (i.e., time
polynomial in both the dataset-size n and the dimension d) 1.633-approximation
algorithm for computing the expected diameter (Theorem 1).
Expected width. We study the expected-width problem in Rd with a xed
dimension d. Two approximation algorithms are proposed for computing the
expected width: a deterministic O(1)-approximation running in O(nd+1 log n)
On the Expected Diameter, Width, and Complexity of a Stochastic Convex-hull 583

time (Theorem 2), and an FPRAS (Theorem 3).

Expected combinatorial complexity. We study the expected-complexity problem
in Rd with a xed dimension d. We provide an exact algorithm for computing
the expected combinatorial complexity of a SCH in O(nd ) time.
Due to space limitations, some proofs and details are omitted in this paper. All
missing proofs can be found in the full version [17].

1.1 Preliminaries
We give the formal denitions of some basic notions used in this paper. A stochas-
tic dataset in Rd is a pair S = (S, ) where S is a set of points in Rd and
: S (0, 1] species the existence probability of each point in S. A realization
of S is a random sample R S where each point a S is sampled with proba-
bility (a). For any subset R S, we use Pr[R] ; to denote; the probability that
R occurs as a realization of S, hence Pr[R] = aR (a) bS\R (1 (b)). A
stochastic convex hull (SCH) of S refers to the convex hull of a realization of S,
which can be regarded as a probabilistic polytope in Rd .
Let P be a convex polytope in Rd . The combinatorial complexity (or simply
complexity) of P , denoted by |P |, is dened as the total number of the faces of
P (the dimensions of the faces vary from 0 to d 1). If u is a unit vector in Rd ,
we dene the directional width of P with respect to u as
widu (P ) = sup (u, p u, q) ,
p,qP

where ,  denotes the inner product. Let U be the set of unit vectors in Rd .
Then the diameter of P is dened as diam(P ) = supuU widu (P ), and the width
of P is dened as wid(P ) = inf uU widu (P ). It is clear that the diameter of P
is also the distance between the farthest-pair of points in P .
For two points x = (x1 , . . . , xd ) and y = (y1 , . . . , yd ) in Rd , we dene x y if
the d-tuple (x1 , . . . , xd ) is smaller than the d-tuple (y1 , . . . , yd ) in lexicographic
order. Then induces a (strict) total order on Rd , called -order.
The approximation algorithms in this paper use relative performance guar-
antees. Formally, a -approximation ( 1) algorithm outputs an answer within
the range [res/, res], where res is the exact answer of the problem.

2 Approximating the expected diameter

Given a stochastic dataset S = (S, ) in Rd (d is not assumed to be xed) with
|S| = n, we consider how to (approximately) compute  the expected diameter of
a SCH of S, denoted by diamS . Formally, diamS = RS Pr[R] diam(CH(R)).
Note that computing diamS exactly is #P-hard when d is not xed [17].

2.1 The witness sequence

In order to approximate diamS , we rst introduce an important notion called
witness sequence. Let P be a convex polytope in Rd with the vertex set V . For
584 J. Xue et al.

a point x Rd , we dene P (x) as the set of all points in P farthest from

x. Formally, P (x) = {y P : dist(x, y) dist(x, y  ) for any y  P }. Note
that P (x) V , and in particular P (x) is nite. We have the following two
observation about diam(P ).
Lemma 1. Let x Rd be a point. If there exist p, q P such that dist(p, q) =
diam(P ) and pxq = > /2, then for any y P (x) and z P (y) we have
dist(y, z) diam(P )/(2 cos(/4)).
Proof. Fixing x R , and let p, q P such that dist(p, q) = diam(P ) and pxq >
d

/2. Also, let y P (x) be any point. Since dist(y, z) max{dist(y, p), dist(y, q)}
for any z P (y), it suces to show
max{dist(y, p), dist(y, q)} diam(P )/(2 cos(/4)).
Without loss of generality, we may assume x = (0, . . . , 0), p = (, , 0, . . . , 0),
q = (, , 0, . . . , 0), where 0. Furthermore, we may also assume dist(x, y) =
1, hence 2 + 2 1 and 2 + 2 1. Since pxq > /2, we must have
< 0 (say > 0 and $< 0). We rst claim that max{dist(y, p), dist(y, q)}
is minimized when y = ( 1 ( + )2 /4, ( + )/2, 0, . . . , 0). Let y be the
point with these coordinates, and r = (r1 , . . . , rd ) be another point satisfying
d
dist(x, r) = 1 (i.e., i=1 ri2 = 1). First consider the case of r2 ( + )/2.
In this case, we show dist(r, p) max{dist(y, p), dist(y, q)}. Since dist(y, p) =
dist(y, q), it suces to show dist(r, p) dist(y, p). We have dist2 (r, p) = 1 +
2 + 2 2r1 2r2 and dist2 (y, p) = 1 + 2 + 2 2y1 2y2 , where
y1 and y2 are the rst two coordinates of y dened above. Now $ we only need
to show r1 + r2 y1 + y2 . Note$that r1 + r2 1 r22 + r2 as
0. Dene vectors v = (, ), u = ( 1 r22 , r2 ), w = (y1 , y2 ). Since 0,
y1 > 0, and r2 y2 < , the angle between v and u is greater $ than that
between v and w. Furthermore, u2 = w2 = 1. Therefore, 1 r22 + r2 =
u, v w, v = y1 + y2 , which implies r1 + r2 y1 + y2 . In the case
r2 ( + )/2, symmetrically, we have dist(r, q) max{dist(y, p), dist(y, q)}.
Therefore, we know that max{dist(y, p), dist(y, q)} is minimized when y has the
above coordinates. Moreover, in this situation, we have
dist(p, q) diam(P )
dist(y, p) = dist(y, q) = = . (1)
2 sin(pyq/2) 2 sin(pyq/2)
Next, we show that pyq /2 where = pxq. Since dist(x, p) dist(x, y),
xyp xpy. Also, since dist(x, q) dist(x, y), xyq xqy. It follows that
pyq = xyp+xyq xpy +xqy. But pxq +pyq +xpy +xqy = 2 and
pxq = , which implies that 2pyq 2 , as desired. Using Equation 1, we
have that dist(y, p) diam(P )/(2 sin(/2 /4)), which completes the proof. 

Lemma 2. Let v V be a vertex of P , and u P (v), w P (u) be two points.

Suppose r is the ray with initial point u which goes through v, and x is the point
on r which has distance dist(u, w)/2 from u. Then if there exist p, q P with
dist(p, q) = diam(P ) and pxq = , we have
 
diam(P )
dist(u, w) min diam(P ), .
3 sin(/2)
On the Expected Diameter, Width, and Complexity of a Stochastic Convex-hull 585

Proof. Let Bv be the (closed) ball centered at u with radius dist(v, u), and
Bu be the (closed) ball centered at u with radius dist(u, w). Then we have
P Bu Bv , because u P (v) and w P (u). Now let r and x be the
ray and the point dened in the lemma. Dene v  as the point on r which has
distance dist(u, w) from u, so x is the midpoint of the segment connecting v 
and u. Set Bv to be the (closed) ball centered at v  with radius dist(u, w). Note
that Bv Bv , since rad(Bv ) rad(Bv ) + dist(v, v  ) where rad() denotes the
radius of a ball. Therefore, P Bu Bv . Next, we claim that Bu Bv
Bx , where Bx is the (closed) ball centered at x with radius 3 dist(u, w)/2.
Suppose y Bu Bv is a point, and assume dist(y, u) dist(y, v  ) without
loss of generality (so yxu /2). Dene = dist(u, x) and = dist(y, x).
Then = sin yux/ sin uyx. Note that we have the restrictions yxu
/2 and dist(u, y) dist(u, v  ) = 2. Under these restrictions, it is easy to
is maximized when dist(u, y) = 2 and yxu = /2. In this case,
see that
= 3 = rad(Bx ). Consequently, Bu Bv Bx , which in turn implies
P Bx . With this observation, we now show the inequality in the lemma. Let
p, q P Bx be two points satisfying dist(p, q) = diam(P ) and pxq = .
If dist(p, q) dist(u, w), we are done, so assume dist(p,
q) > dist(u, w). But
both dist(x, p) and dist(x, q) are at most rad(Bx ) = 3 dist(u, w)/2. Therefore,
is the largest angle of the triangle pxy. In this case, it is easy to see that
dist(p, q) is
maximized when dist(x, p) = dist(x, q) = rad(Bx ). It follows that
dist(p, q) 3 sin(/2) dist(u, w), which completes the proof. 

Lemma 2 tells us that for a vertex v V , the distance dist(u, w) where u

P (v) and w P (u) gives a good approximation for diam(P ) as long as there
exists a pair p, q P dening diam(P ) with a small angle pxq (see the
lemma for the denition of x). However, the approximation is not satisfactory
when pxq is large. Fortunately, we have Lemma 1, which is helpful for this
case. Indeed, in the case that pxq is large, if we further take y P (x) and
z P (y), then Lemma 1 implies that dist(y, z) gives a good approximation
for diam(P ). Therefore, intuitively, by taking max{dist(u, w), dist(y, z)}, we can
well-approximate diam(P ) no matter whether pxq is small or large. Formally,
we conclude the following.

Corollary 1. Let v, u, w, x be the points dened in Lemma

2. Also, let y
P (x) and z P (y) be any two points. Set = 2 2/ 3. Then we have

diam(P )/ max{dist(u, w), dist(y, z)} diam(P ).

Proof. It is clear that max{dist(u, w), dist(y, z)} diam(P ), as u, w, y, z P .

Let p, q P such that dist(p, q) = diam(P ), and = pxq. If /2,
then Lemma 2 implies dist(u, w) diam(P )/( 3/ 2). So assume > /2.
By Lemma 1, we have dist(y,z) diam(P )/(2 cos(/4)). Then by Lemma 2, we
have dist(u, w) diam(P )/( 3 sin(/2)). Therefore,

diam(P )
max{dist(u, w), dist(y, z)} .
min{2 cos(/4), 3 sin(/2)}
586 J. Xue et al.

Note that for (/2, ], 2 cos(/4) is monotonically decreasing and 3 sin(/2)
is monotonically increasing. Thus, the right-hand side of the above inequal-
ity is minimized
when 2 cos(/4) = 3 sin(/2). We have this equality when
sin(/4) = 1/ 3. In this case, the r.h.s. is equal to diam(P )/. 
With the ve points v, u, w, y, z (which are in fact the vertices of P ) in hand,
Corollary
1 allows us to approximate the diameter of P within a factor of =
2 2/ 3 1.633. However, the choice of v, u, w, y, z is not unique in our above
construction for a given P . For later use, we need to make it unique, which can
be done by considering -order (see Sec. 1.1). We dene v V as the largest
vertex of P under -order. Also, we require u P (v), w P (u), y P (x),
z P (y) to be the largest under -order. In this way, we obtain a uniquely
dened 5-tuple (v, u, w, y, z) for the polytope P . We call this 5-tuple the witness
sequence of P , denoted by wit(P ). For a 5-tuple = (x1 , . . . , x5 ) of points in
Rd , dene () = max{dist(x2 , x3 ), dist(x4 , x5 )}. Then Corollary 1 implies
diam(P )/ (wit(P )) diam(P ) (2)

for any convex polytope P in Rd .

2.2 An (n, d)-polynomial-time approximation algorithm

We now use the notion of witness sequence dened above to establish our approx-
imation algorithm for computing diamS . Given the stochastic dataset S = (S, ),
we rst do a preprocessing to sort all the points in S in -order and compute
the pair-wise distances of the points in S. This preprocessing can be done in
O(dn2 ) time. To approximate diamS , we dene

diamS = Pr[R] (wit(CH(R))).
RS

Inequality 2 implies diamS / diamS diamS . Thus, it now suces to com-

pute diamS . Computing diamS by directly using the above formula takes ex-
ponential time, as S has 2n subsets. However, since for any R S the witness
sequence wit(CH(R)) must be a 5-tuple of points in S, we can also write

diamS = Pr[] (), (3)
S

where S is the set of all 5-tuples of points in S and Pr[] is the probability that
the witness sequence of a SCH of S is . Note that |S | = O(n5 ). Thus, we can
eciently compute diamS as long as Pr[] and () can be computed eciently
for every S . Clearly, () can be directly computed in constant time (after
our preprocessing). To compute Pr[], suppose = (p1 , . . . , p5 ) S . It is easy
to check that if p1 = p2 , then either Pr[] = 0 or () = 0. So we may assume
p1 = p2 . From the denition of witness sequence, we directly obtain the following
criterion for checking if is the witness sequence of a SCH of S. We write a b c
for a, b, c Rd if dist(a, b) < dist(c, b), or dist(a, b) = dist(c, b) and a c.
Lemma 3. Let = (p1 , . . . , p5 ) S with p1 = p2 . Suppose r is the ray with
initial point p2 which goes through p1 , and x is the point on r which has distance
On the Expected Diameter, Width, and Complexity of a Stochastic Convex-hull 587

dist(p2 , p3 )/2 from p2 . For a realization R of S, we have = wit(CH(R)) i

(1) R contains p1 , . . . , p5 , and (2) R does not contain any point a S satisfying
p1 a or p2 p1 a or p3 p2 a or p4 x a or p5 p4 a.
By Lemma 3, it is quite easy to compute Pr[] in linear time, just by multiplying
the existence probabilities of the points in and the non-existence probabilities
of all the points which should not be included in R (according to the condition (2)
in the lemma). Using Equation 5, we obtain an (n, d)-polynomial-time algorithm
to compute diamS . This algorithm runs in O(n6 + dn2 ) time. But it is fairly easy
to improve the runtime to O(n5 log n + dn2 ) by considering the 5-tuples in S
in a proper order (see [17] for details).
Theorem 1. One can 1.633-approximate diamS in (n, d)-polynomial time. Specif-
ically, the approximation can be done in O(n5 log n + dn2 ) time.
Interestingly, our witness-sequence technique also gives an O(dn)-time 1.633-
approximation algorithm for computing the diameter of the convex hull of a (non-
stochastic) point-set S in Rd , because wit(CH(S)) can be computed in O(dn)
time. To our best knowledge, there has not been any linear-time algorithm which
can achieve such an approximation factor when d is not xed.

3 Approximating the expected width

Given a stochastic dataset S = (S, ) in Rd (d is xed) with |S| = n, we consider

how to (approximately) compute
 the expected width of a SCH of S, denoted by
widS . Formally, widS = RS Pr[R] wid(CH(R)).

3.1 The witness simplex

Recall that when solving the expected-diameter problem, we developed the no-
tion of witness sequence, which well-captures the diameter of a polytope and
satises (1) the total number of the possible witness sequences of a SCH is poly-
nomial (though there are exponentially many realizations), (2) the probability
of a sequence being the witness sequence of a SCH can be easily computed. We
apply this basic idea again to the expected-width problem. To this end, we have
to design some good witness object for width, which satises the above two
conditions. The witness object to be dened is called witness simplex.
Let P be a convex polytope in Rd with wid(P ) > 0, and V be the vertex
set of P . We choose d + 1 vertices v0 , . . . , vd V of P inductively as follows.
Dene v0 V as the largest vertex of P under -order. Suppose v0 , . . . , vi
are already dened. Let Ei be the (unique) i-dim hyperplane in Rd through
v0 , . . . , vi (or the i-dim linear subspace of Rd spanned by v0 , . . . , vi ). We then
dene vi+1 V as the vertex of P which has the maximum distance to Ei , i.e.,
vi+1 = arg maxvV dist(v, Ei ). If there exist multiple vertices having maximum
distance to Ei , we choose the largest one under -order to be vi+1 . In this way,
we obtain the vertices v0 , . . . , vd . The witness simplex P of P is dened as the
588 J. Xue et al.

d-simplex with vertices v0 , . . . , vd . The (ordered) sequence (v0 , . . . , vd ) is said to

be the vertex list of P . Note that the vertex list is determined by only P
and independent of P . In other words, if we only know P without knowing the
original polytope P , we can still recover the vertex list of P , just by ordering
the d + 1 vertices of P into a sequence (v0 , . . . , vd ) such that v0 is the largest
under -order, and each vi+1 is the one having the maximum distance to Ei
(the linear subspace spanned by v0 , . . . , vi ). A useful geometric property of the
witness simplex P is that it well-captures the width of P .
Lemma 4. Let P be a convex polytope in Rd with wid(P ) > 0, then we have
wid(P ) = (wid(P )). The constant hidden in () could be exponential in d.
Proof intuition. Due to space limitation, here we only give some intuition for
proving the lemma. A detailed proof can be found in [17]. Note that wid(P )
wid(P ) since P P . To see wid(P ) = (wid(P )), let (v0 , . . . , vd ) be the
vertex list of P . Without loss of generality, we may assume that vi is in
the linear subspace of Rd spanned by the x1 , . . . , xi axes. We only need to
show widu (P ) = (widu (P )) for any unit vector u = (u1 , . . . , ud ). If |u1 | =
(widu (P )/diam(P )), then widu (P ) |u, v0  u, v1 | |u1 | dist(v0 , v1 ).
But |u1 | dist(v0 , v1 ) |u1 | diam(P )/2 = (widu (P )). On the other hand, |u1 |
is very small, so essentially we have widu (P ) widu (P ) and widu (P )
widu (P ) where u = (0, u2 , . . . , ud ). Then it suces to have widu (P ) =
(widu (P )), which can be shown using an induction argument due to our re-
cursive construction of the witness simplex P . 

3.2 An O(1)-approximation algorithm

With the notion of witness simplex in hand, we propose a O(1)-approximation
algorithm for computing widS . The basic idea is similar to what we use for
approximating diamS . We dene

widS = Pr[R] wid(CH(R) ),
RS

Lemma 4 implies widS = (widS ). Thus, in order to approximate widS within a

constant factor, it suces to compute widS . To compute widS by directly using
the above formula takes exponential time, as S has 2n subsets. However, since
CH(R) must be a d-simplex with vertices in S, widS can also be written as

widS = Pr[] wid(), (4)
d
S

where Sd is the set of all d-simplices in Rd whose vertices are (distinct) points
in S and Pr[] is the probability that the witness simplex of a SCH of S is .
Note that |Sd | = O(nd+1 ), which is polynomial. So the above formula allows us
to compute widS in polynomial time, as long as we are able to compute Pr[]
eciently for each Sd . Fixing Sd , we now investigate how to compute
Pr[]. As argued before, we can recover the vertex list (v0 , . . . , vd ) of . By
On the Expected Diameter, Width, and Complexity of a Stochastic Convex-hull 589

the construction of , v0 , . . . , vd are points in S. For i {0, . . . , d 1}, we

denote by Ei the i-dim hyperplane in Rd through v0 , . . . , vi . From the denition
of witness simplex, we directly obtain the following criterion for checking if
is the witness simplex of a SCH of S. For a hyperplane H (of any dimension)
in Rd and two points a, b Rd , we write a H b if dist(a, H) < dist(b, H), or
dist(a, H) = dist(b, H) and a b.
Lemma 5. For a realization R of S, is the witness simplex of CH(R) (i.e.,
= CH(R) ) i (1) R contains v0 , . . . , vd , and (2) R does not contain any point
a S satisfying v0 a or vi+1 Ei a for some i {0, . . . , d 1}.
Using the above lemma, we can straightforwardly compute Pr[] in linear
time, just by multiplying the existence probabilities of v0 , . . . , vd and the non-
existence probabilities of all a S which should not be included in R (according
to the condition (2) in the lemma). Therefore, we obtain an O(nd+2 )-time algo-
rithm for computing widS . It is easy to improve the runtime to O(nd+1 log n) by
considering the simplices in Sd in a proper order (see [17] for details).
Theorem 2. One can O(1)-approximate widS in O(nd+1 log n) time. The con-
stant approximation factor could be exponential in d.

3.3 A fully polynomial-time randomized approximation scheme

In this section, we develop a fully polynomial-time randomized approximation
scheme (FPRAS) for computing widS . An FPRAS should take S and a real
number > 0 as input, and output a (1 + ) approximation of widS in time
polynomial in the size of S and 1/ with probability at least 2/3.
We rst introduce some notations. As dened in the preceding section, Sd is
the set of all d-simplices in Rd whose vertices are (distinct) points in S, and for
each Sd the notation Pr[] denotes the probability that the witness simplex
of a SCH of S is . Let R be a realization of S and Sd be a simplex. From
Lemma 5, we know that = CH(R) i R contains the vertices of but does
not contain some other points in S according to (2) in the lemma. We now use
V to denote the set of the vertices of , X to denote the set of the points in S
that R must not contain if = CH(R) . Let F = S\(V X ), which is the set
of the points in S whose presence/absence in R does not inuence whether =
CH(R) . Dene F as the sub-dataset of S with the point-set F . Our FPRAS
works as follows. First, for each Sd , we randomly generate m = log n/2
realizations of F , where is a large enough constant to be determined. Let
R1 , . . . , Rm

be the generated realizations of F , and set Ti = Ri V . Note
that the witness simplex of CH(Ti ) is by Lemma 5. We then compute
 
 m
wid(CH(Ti ))
widS = Pr[] , (5)
d i=1
m
S

and output widS as the approximation of widS .

Next, we discuss the choice of the constant and verify the correctness
of our FPRAS. By Lemma 4, we can nd positive constants k1 , k2 such that
590 J. Xue et al.

k1 wid(P ) wid(P ) k2 wid(P ) for any convex polytope P in Rd with

wid(P ) > 0. We set = d(k2 /k1 )2 . Using Hoedings inequality, it is easy to
prove the following lemma, which guarantees the correctness of our FPRAS.
Lemma 6. (1 )widS widS (1 + )widS with probability at least 2/3.

Theorem 3. There exists an FPRAS for computing widS .

4 Computing the expected combinatorial complexity

Given a stochastic dataset S = (S, ) in Rd (d is xed) with |S| = n, our goal in
 complexity of a SCH of S, denoted by
this section is to compute the expected
compS . Formally, we have compS = RS |CH(R)|.

4.1 Reduction to SCH membership probability queries

Given a stochastic dataset T in Rd and a query point q Rd , the SCH mem-
bership probability (of q with respect to T ) refers to the probability that q lies
in a SCH of T , which we denote by memT (q). It is known that memT (q) can be
computed in O(md1 ) time for d 3 [4,14] and O(m log m) time for d {1, 2}
[3], where m is the number of the stochastic points in T .
We reduce our problem of computing compS to SCH membership probability
queries. Let R be a realization of S. It is clear that the faces of CH(R) must be
simplices with vertices in S. Therefore, we can rewrite the formula for compS as

  
compS = Pr[R] (R, ) = F , (6)
RS S S

where S is the set of all simplices (of dimension less than d) with vertices in
S, is a indicating function such that (R, ) = 1 if is a face of CH(R) and
(R, ) = 0 otherwise, F is the probability that is a face of a SCH of S.
We now show that for each S , the computation of F can be reduced to a
SCH membership probability query. Suppose Y is a set of m (m d + 1) points
in Rd in general position. Let y0 , . . . , yk Y be k + 1 points where 0 k d 1,
and be the k-simplex with vertices y0 , . . . , yk . Dene vectors ui = yi y0 for
i {1, . . . , k}. By the general position assumption, u1 , . . . , uk generate a k-dim
linear subspace H of Rd . Set H to be the orthogonal complement of H in Rd ,
which is by denition the (d k)-dim linear subspace of Rd orthogonal to H.
We then orthogonally project the points in Y to H , and denote the set of the
projection images by Y . Note that y0 , . . . , yk are clearly projected to the same
point in H , say y. We then have the following geometric observation.
Lemma 7. is a face of CH(Y ) i y is a vertex of CH(Y ) in H .
The above lemma allows us to reduce the computation of F for any
S to a SCH membership query as follows. For each i {0, . . . , d 1}, let
d1
Si S be the subset consisting of all i-simplices in S (then S = i=0 Si ).
On the Expected Diameter, Width, and Complexity of a Stochastic Convex-hull 591

Suppose Sk is a k-simplex with vertices v0 , . . . , vk S. As before, we

dene vectors ui = vi v0 for i {1, . . . , k}. Then u1 , . . . , uk generate a k-dim
linear subspace H of Rd , and set H to be the orthogonal complement of H in
Rd . Let : Rd H be the orthogonal projection map. We dene a multi-set
S  = {(a) : a S\{v0 , . . . , vk }} of points in H , which gives us a stochastic
dataset S  = (S  ,  ) in H where  ((a)) = (a). Set q = (v0 ) = = (vk ).
;k
Corollary 2. F = i=0 (vi ) (1 memS  (q)).
Proof. Let R be a realization of S. If is a face of CH(R), then v0 , . . . , vk
must be contained in R. Furthermore, by Lemma 7, q must be a vertex of the
projection image of CH(R) in H . By the general position assumption, this is
equivalent to saying that q is outside the projection image of CH(R\{v0 , . . . , vk }).
Conversely, if v0 , . . . , vk are contained in R and q is outside the projection image
of CH(R\{v0 , . . . , vk }), then is a face of CH(R) by Lemma 7. The probability
;k
that R contains v0 , . . . , vk is i=0 (vi ), and the probability that q is outside
the projection image of CH(R\{v0 , . . . , vk }) is 1 memS  (q). These two events
are clearly independent. Therefore, we have the formula in the corollary. 

Since H is linearly homeomorphic to R dk
, computing memS  (q) is nothing
but answering a SCH membership probability query in Rdk . Therefore, using
the algorithms for answering SCH membership probability queries [4,14], F can
be computed in O(ndk1 ) time if k {0, . . . , d 3}. Note that |Sk | = O(nk+1 ),
d3 
so we can compute the sum i=0 i F in O(nd ) time. In order to further
S 
compute compS by Equation 6, we now only need to compute d1 F and
 S
1 2
d2 F . But answering SCH membership probability queries in R and R
S
requires O(m log m) time [3] (where m is the size of the given stochastic dataset).
Thus, if we directly plug in the algorithm in [3], the computation task cannot
be done in O(nd ) time. The next section discusses how to handle this issue.

4.2 Handling k = d 2 and k = d 1

 
Set 1 = d1 F and 2 = d2 F . Fix a point o Rd such that
S S
S {o} is in general position. For every hyperplane E with o / E, we denote
by E + the connected component of Rd \E containing o, and by E the other
; Dene the S-statistic
one. of E as a 3-tuple statS (E) = (p+ , p , A) where p+ =
;

aSE + (1(a)), p = aSE (1(a)), A = S E. Let E be the collection
of the hyperplanes in Rd which go through exactly d points in S. Since S {o}
is in general position, stat(E) is dened for every E E. We say an algorithm
computes the S-statistics for E if it outputs statS (E) for all E E.
Lemma 8. One can compute 1 and 2 in O(t(n)) time, provided an algorithm
computing the S-statistics for E in O(t(n)) time.
Proof sketch. We use the provided algorithm to compute the S-statistics for
E in O(t(n)) time. With the statistics in hand, we can easily compute 1 in
O(nd ) time, because F for each Sd1 can be computed in constant time
592 J. Xue et al.

via stat(E) (using Corollary 2), where E E is the hyperplane through the d
vertices of . To compute 2 in O(nd ) time, we observe that as long as we have
the statistics, the SCH membership query needed in Corollary 2 to compute F
for each Sd2 can be answered in linear time by using the witness-edge
method in [3]. Note that t(n) = (nd ), so the computation of 1 and 2 is done
in O(t(n)) time. See [17] for a detailed proof. 
It is implicitly known that one can compute the S-statistics for E in O(nd ) time
[4,14]. We explicitly describe the algorithm in the full version [17]. With this
algorithm in hand, Lemma 8 allows us to compute 1 and 2 in O(nd ) time.
Therefore, we can nally conclude the following.
Theorem 4. One can compute the exact value of compS in O(nd ) time.

References
1. Agarwal, P.K., Aronov, B., Har-Peled, S., Phillips, J.M., Yi, K., Zhang, W.: Nearest
neighbor searching under uncertainty II. In: Proc. of the 32nd PODS. ACM (2013)
2. Agarwal, P.K., Cheng, S.W., Yi, K.: Range searching on uncertain data. ACM
Transactions on Algorithms (TALG) 8(4), 43 (2012)
3. Agarwal, P.K., Har-Peled, S., Suri, S., Yldz, H., Zhang, W.: Convex hulls under
uncertainty. In: Algorithms-ESA 2014, pp. 3748. Springer (2014)
4. Fink, M., Hershberger, J., Kumar, N., Suri, S.: Hyperplane separability and con-
vexity of probabilistic point sets. In: Proc. of the 32nd SoCG. ACM (2016)
5. Huang, L., Li, J.: Approximating the expected values for combinatorial optimiza-
tion problems over stochastic points. In: International Colloquium on Automata,
Languages, and Programming. pp. 910921. Springer (2015)
6. Huang, L., Li, J., Phillips, J.M., Wang, H.: -kernel coresets for stochastic points.
arXiv preprint arXiv:1411.0194 (2014)
7. Jrgensen, A., Loer, M., Phillips, J.: Geometric computations on indecisive
points. In: WADS. pp. 536547. Springer (2011)
8. Kamousi, P., Chan, T., Suri, S.: Stochastic minimum spanning trees in euclidean
spaces. In: Proc. of the 27th SoCG. pp. 6574. ACM (2011)
9. Kamousi, P., Chan, T.M., Suri, S.: Closest pair and the post oce problem for
stochastic points. Computational Geometry 47(2), 214223 (2014)
10. Li, C., Fan, C., Luo, J., Zhong, F., Zhu, B.: Expected computations on color
spanning sets. Journal of Combinatorial Optimization 29(3), 589604 (2015)
11. Loer, M., van Kreveld, M.: Largest and smallest convex hulls for imprecise points.
Algorithmica 56(2), 235269 (2010)
12. Suri, S., Verbeek, K.: On the most likely Voronoi Diagram and nearest neighbor
searching. In: ISAAC. pp. 338350. Springer (2014)
13. Suri, S., Verbeek, K., Yldz, H.: On the most likely convex hull of uncertain points.
In: AlgorithmsESA 2013, pp. 791802. Springer (2013)
14. Xue, J., Li, Y., Janardan, R.: On the separability of stochastic geometric objects,
with applications. In: Proc. of the 32nd SoCG. ACM (2016)
15. Xue, J., Li, Y.: Colored stochastic dominance problems. arXiv preprint
arXiv:1612.06954 (2016)
16. Xue, J., Li, Y.: Stochastic closest-pair problem and most-likely nearest-neighbor
search in tree spaces. arXiv preprint arXiv:1612.04890 (2016)
17. Xue, J., Li, Y., Janardan, R.: On the expected diameter, width, and complexity of
a stochastic convex-hull. arXiv preprint arXiv:1704.07028 (2017)
 Author Index

Acharyya, Ankush 1 Ehrhardt, Marcel 361

Aggarwal, Anshul 13 Eppstein, David 193
Agrawal, Akanksha 25
An, Zhao 37 Favrholdt, Lene M. 217
Fekete, Sndor P. 373
Bandyapadhyay, Sayan 49 Feng, Qilong 37
Banik, Aritra 61 Ferres, Leo 385
Banyassady, Bahareh 73 Fleszar, Krzysztof 265
Bar-On, Gali 85 Fluschnik, Till 397
Barba, Luis 97 Fuentes, Jos 385
Benbernou, Nadia M. 109
Bergougnoux, Benjamin 121 Gagie, Travis 385
Bhattacharya, Binay 133 Gawrychowski, Pawe 409
Biedl, Therese 145, 157 Goldstein, Isaac 421
Bienkowski, Marcin 169 Golin, Mordecai J. 133
Bilardi, Gianfranco 181 Grimm, Carsten 301
Biniaz, Ahmad 193 Gupta, Neelima 13
Bodlaender, Hans L. 289
Bose, Prosenjit 193, 205 Halldrsson, Magns M. 437
Boyar, Joan 217 Haverkort, Herman 449
Brandt, Sebastian 229 He, Meng 385
Higashikawa, Yuya 133
Hoffmann, Michael 73
Cardinal, Jean 97 Hu, Lunjia 325
Carmi, Paz 85, 241 Huang, Lingxiao 325
Chakaravarthy, Venkatesan T. 13 Huang, Ziyun 461
Chan, Timothy M. 253
Chaplick, Steven 265
Iacono, John 473
Chiu, Man Kwun 241
Ishii, Toshimasa 437
Crosbie, Kimberly 193
Iwama, Kazuo 485

Dasler, Philip 277 Janardan, Ravi 581

de Berg, Mark 289 Jansen, Klaus 497
De Carufel, Jean-Lou 193, 205, 301
De Stefani, Lorenzo 181 Kameda, Tsunehiko 133
Demaine, Erik D. 109, 313 Kanj, Iyad 37
Demaine, Martin L. 109 Kant, Mamadou Moustapha 121
Demontigny, Philippe 145 Katoh, Naoki 133
Derka, Martin 157 Katz, Matthew J. 241
Ding, Hu 325, 461 Khramtcova, Elena 473
Duan, Ran 337, 349 Khuller, Samir 521
Dujmovi, Vida 205 Kim, Eun Jung 509

Springer International Publishing AG 2017 593

F. Ellen et al. (Eds.): WADS 2017, LNCS 10389, pp. 593594, 2017.
DOI: 10.1007/978-3-319-62127-2
594 Author Index

Kisfaludi-Bak, Sndor 289 Paradis, Frdrik 205

Klemz, Boris 73 Porat, Ely 421
Koehler, Frederic 521
Komusiewicz, Christian 397 Raman, Venkatesh 61
Kopelowitz, Tsvi 421 Ravsky, Alexander 265
Korman, Matias 97, 241 van Renssen, Andr 97, 241
Kotrbk, Michal 217 Roeloffzen, Marcel 97, 241
Kraska, Artur 169 Roy, Aniket Basu 49
Kbel, David 449 Roy, Sasanka 1
Kwon, O-Joung 121, 509
Sabharwal, Yogish 13
Langerman, Stefan 97, 473
Sahlot, Vibha 61
Langetepe, Elmar 449
Saurabh, Saket 25, 61
Larsen, Kim S. 217
Scheffer, Christian 373
Lewenstein, Moshe 421
Schirra, Stefan 301
Li, Jian 325
Schmidt, Pawe 169
Li, Shimin 533
Schwarzwald, Barbara 449
Li, Yuan 569, 581
Sharma, Sachin 13
Lipp, Fabian 265
Shiitada, Taichi 241
Liu, Quanquan C. 313
Skrepetos, Dimitrios 253
Lffler, Maarten 73
Smid, Michiel 193, 301
Lubiw, Anna 109
Smorodinsky, Shakhar 241
Maack, Marten 497
Takazawa, Kenjiro 437
Maheshwari, Anil 193
Talmon, Nimrod 397
Makino, Kazuhisa 437
Teruyama, Junichi 485
Mertzios, George B. 397
Thakral, Sonika 13
Miltzow, Tillmann 73
Misra, Pranabendu 25
Morr, Sebastian 373 Verbitsky, Oleg 265
Mount, David M. 277 Verdonschot, Sander 97
Mulzer, Wolfgang 361
Wang, Haitao 533, 545, 557
Nandy, Subhas C. 1 Wattenhofer, Roger 229
Navarro, Gonzalo 385 Wolff, Alexander 265
Nicholson, Patrick K. 409
Nichterlein, Andr 397 Xia, Ge 37
Niedermeier, Rolf 397 Xu, Jinhui 461
Xue, Jie 569, 581
Okamoto, Yoshio 241
Zhang, Jingru 557
Pandit, Supantha 1 Zhang, Le 337
Panolan, Fahad 25, 61 Zhang, Tianyi 349