Citation for published version (APA):
Farshi, M. (2008). A theoretical and experimental study of geometric networks. Technische Universiteit
Eindhoven. https://doi.org/10.6100/IR630219
DOI:
10.6100/IR630219
Document status and date:
Published: 01/01/2008
Document Version:
Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers)
Please check the document version of this publication:
• A submitted manuscript is the version of the article upon submission and before peer-review. There can be
important differences between the submitted version and the official published version of record. People
interested in the research are advised to contact the author for the final version of the publication, or visit the
DOI to the publisher's website.
• The final author version and the galley proof are versions of the publication after peer review.
• The final published version features the final layout of the paper including the volume, issue and page
numbers.
Link to publication
General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners
and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.
• Users may download and print one copy of any publication from the public portal for the purpose of private study or research.
• You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal.
If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please
follow below link for the End User Agreement:
www.tue.nl/taverne
Take down policy
If you believe that this document breaches copyright please contact us at:
openaccess@tue.nl
providing details and we will investigate your claim.
A Theoreti al and Experimental
Study of Geometri Networks
PROEFSCHRIFT
terverkrijgingvandegraadvando tor
aan deTe hnis heUniversiteitEindhoven,opgezagvande
Re torMagni us,prof.dr.ir. C.J.vanDuijn,vooreen
ommissieaangewezendoorhetCollegevoor
Promotiesin hetopenbaarte verdedigen
opdinsdag8april2008om16.00uur
door
Mohammad Farshi
Copromotor:
dr. J.Gudmundsson
CIP-DATALIBRARYTECHNISCHEUNIVERSITEITEINDHOVEN
Farshi,Mohammad
Atheoreti alandexperimentalstudyofgeometri networks/doorMohammadFarshi.
-Eindhoven: Te hnis heUniversiteitEindhoven,2008.
Proefs hrift. - ISBN978-90-386-1135-8
NUR993
Subje theadings: omputationalgeometry /datastru tures/algorithms
Kern ommissie:
prof.dr. M.H.Overmars(Utre htUniversity)
prof.dr. M.Smid(CarletonUniversity)
prof.dr. J.J.vanWijk (Te hnis heUniversiteitEindhoven)
Ministry of Science, Research and Technology
Islamic Republic of Iran
The work in this thesis is supported by the Ministry of S ien e, Resear h and
Te hnologyofI.R.Iran unders holarshipno.800.341.
The work in this thesishas been arried outunder the auspi es of the resear h
s hoolIPA(InstituteforProgrammingresear handAlgorithmi s).
IPAdissertationseries2008-12
Mohammad Farshi 2008. All rights are reserved. Reprodu tion in whole or
in partisprohibited withoutthewritten onsentofthe opyrightowner.
Printing: EindhovenUniversityPress
Prefa e v
1 Introdu tion 1
1.1 Geometri networks . . . 1
1.2 t-Spanners. . . 3
1.3 Whyspanners? . . . 6
1.3.1 Approximateminimumspanningtree . . . 7
1.3.2 Metri spa esear hing . . . 7
1.3.3 Broad astingin ommuni ationnetworks . . . 8
1.3.4 Proteinsvisualization . . . 9
1.4 Thesisoverview . . . 10
2 Region-Fault TolerantSpanners 13 2.1 Introdu tion. . . 13
2.2 Constru ting
C
-faulttolerantspannersusingtheWSPD . . . 162.3 Well-separatedpairde omposition . . . 17
2.3.1 Constru tinga
C
-faulttolerantspanner . . . 182.3.2 Linear-sizespannersforspe ial ases . . . 20
2.3.3
C
-faulttolerantSteinerspanners . . . 222.4 Spe ial ases . . . 24
2.4.1
C
-faulttolerantfattriangulations . . . 242.4.2 Limitedboundarydire tions . . . 25
2.5
C
-fault tolerantspannersforarbitrarypointsets . . . 262.5.1 SSPDsandfault-tolerantspanners . . . 27
2.5.2 ComputinganSSPD . . . 32
2.6 Testingfor
C
-faulttoleran e . . . 392.7 Fault-tolerantgeodesi spanners . . . 40
3.2.2 A(1+")-approximationforEu lideangraphs. . . 50
3.3 Addingabottlene kedge . . . 51
3.4 A(2+")-approximationforEu lideangraphs . . . 55
3.4.1 Linearnumberof andidateedges . . . 55
3.4.2 Speedingupalgorithm3.4.1 . . . 61
3.5 Aspe ial ase:
G
has onstantdilation . . . 623.6 Con ludingremarks . . . 65
4 Dilation-OptimalEdge Deletion 67 4.1 Introdu tion. . . 67
4.2 Dilation-minimaledgedeletion ina y le. . . 69
4.2.1 Estimatingthedilationofapolygonalpath . . . 70
4.2.2 Thede isionproblem . . . 72
4.2.3 Theoptimizationalgorithm . . . 76
4.3 Dilation-maximaledgedeletion ina y le . . . 77
4.4 (1+")-Approximationalgorithm . . . 81
4.5 Con ludingremarks . . . 84
5 ComputingSpannerDiameter 85 5.1 Introdu tion. . . 85
5.2 Dynami programmingapproa h . . . 86
5.3 Improvingthe omplexitybounds . . . 90
5.4 Anal approa h . . . 91
5.5 Experimentalresults . . . 93
5.6 Con ludingremarks . . . 94
6 ExperimentalStudyofGeometri Spanners 97 6.1 Introdu tion. . . 97
6.1.1 Spannerproperties . . . 98
6.2 Spanner onstru tionalgorithms . . . 98
6.2.1 Theoriginalgreedyalgorithm andanimprovement. . . 99
6.2.2 Theapproximategreedyalgorithm . . . 101
6.2.3 The-graphalgorithm . . . 103
6.2.4 Theordered-graphalgorithm . . . 104
6.2.5 Therandomordered-graphalgorithm . . . 106
6.2.6 Thesink-spanneralgorithm . . . 106
6.2.7 Theskip-listspanneralgorithm . . . 108
6.2.8 TheWSPDalgorithm . . . 109
6.3 Experimentalresults . . . 111
6.3.6 Maximumandaveragedilation . . . 123
6.3.7 Crossings . . . 125
6.3.8 Thehybridalgorithms . . . 125
6.3.9 Numberofshortestpathqueries . . . 128
6.3.10 Runningtime . . . 130
6.4 Con ludingremarks . . . 135
7 Con lusions 137
ฬΧΧ୍ϞฬଘάනΧϛϩ
ฬϪ̶Ϋ̳߮ΧΫϩϩψ߮ι
̵ࠖέάඵහ̵Ψϊॣࠗ
ThisthesisistheresultofalmostfouryearsofworkwhereIhavebeen
a om-panied and supported by many people. I now havethe pleasantopportunity to
expressmygratitudetoallofthem.
First of all, I would like to express my deepand sin ere gratitude to my
su-pervisors,MarkdeBergandJoa himGudmundssonforgivingmethepossibility
toworkunder theirsupervision. ThankstoMarkwhotooktheriskof a epting
me, asapersonwith noknowledgein omputationalgeometry, in hisgroupand
to Joa himwho was mydaily supervisorand who helpedme to understand the
on ept of spanners,dis uss problems and read my manus ripts, notonly when
hewasin TU/ebutalsoafter heleftEindhoven. Thisworkwouldnothavebeen
possible without their support and en ouragement, and I am grateful for their
valuablefriendship.
I would alsoliketo thankmydistinguished o-authorsduringmy PhDstudy
Mohammad Ali Abam, Hee-Kap Ahn, Mark de Berg, Joa him Gudmundsson,
PanosGiannopoulos,ChristianKnauer,Mi hiel Smid, and Yajun Wang,the
re-sults in this thesis is the produ t of our joint work. I would like to express my
thanks to the people who made the joint works possible: Alexander Wolf and
XavierGoao forinvitingmeto theKoreanWorkshop onComputational
Geom-etryand theorganizers oftheworkshop ongeometri networksandmetri spa e
embedding atS hlossDagstuhl. Ialso thankJoa himGudmundssonforinviting
metoNICTAandhishospitalityduring myvisit.
Themembersofmythesis ommitteearegratefully a knowledgedforreading
thethesis,providinguseful ommentsandbeingpresentatmydefensesession. It
wasmyprivilegetohaveMarkdeBerg,Joa himGudmundsson,MarkOvermars,
Mi hiel Smid and Ja k van Wijk in thethesis ommittee and Rolf Klein in the
ofIranianstudentsinEuropefortheirhelp.
I was verylu ky to work with absolutely fantasti people in the Algorithms
group. I thank them all for making my years at Eindhoven delightful. Spe ial
thanksgotoSonjaJoostenforherhelpatthestartofmystudyandAstridVolkers
forherhelpatthenalstagesofmywork. IalsothankmygreatoÆ ematesYuval
Nir, MarkS hrodersandPeterHa henberger.
Aheart-felt thanksgoestoMohammadAli Abam forhisvaluablefriendship,
enjoyabledis ussions,whi halways omewithni eideas,andteabreaks. Ilearned
alotfrom himand I'mlookingforwardto ontinuetowork withhim.
I must thank friends/families whose ompanymade my and my family'slife
mu hmoreenjoyableandtheweekend/holidaymeetingswereourmostwonderful
timesintheNetherlands. Iwouldliketomentionthefollowingfamilies: Cheema,
Eslami, Ghasemzadeh, Mousavi, Moosavi Nezhad, Nazarpoor, Nikoufard,
Reza-eian,Talebi,Vahedi. IwouldalsoliketoexpressmygratitudetoMohammadAli
Abam, MohammadEslami, Hamed Fatemi, Amir HosseinGhamarian,
Moham-mad Ghasemzadeh,KamyarMalakpoor,MohammadRezaMousavi,Mohammad
MoosaviNezhad,MahmoudNikoufard,RezaRezaeian,MohammadSamimi,Saeid
Talebi,MostafaVahedifortheirkindfriendship.
I annot endwithout thankingmy family, onwhose onstanten ouragement
andloveIhavereliedthroughoutmylife. Iamgratefultomyparentsfortheir
un- onditionalsupport,un in hing ourageand onvi tionduringmystudy. Iwould
liketothankmywifeHamidehand mysonAlirezafortheirunwaveringsupport,
patien e and understanding during this time. It is to them that I dedi ate this
work,withloveandgratitude.
ϞฬΔࡣϩΧϪরέ̵ΧϪরϦૡঙ
߮ଶ৳ฬϦૡঙඟࢂҤΧΔ॥ϩϞ߮ଶ৳
Ϟ
1.1 Geometri networks
A network is, informally speaking, a olle tion of \obje ts" with ertain
\ on-ne tions"between the obje ts. An obvious exampleof a network is a omputer
network. Here the obje tsare omputers and thereis a onne tion betweentwo
omputers ifthere is aphysi al able onne tingthem. Other obvious examples
are road or railway networks. Inthe latter type of network the obje tsare the
stations,andthe onne tionsarethetra ks onne tingthevariousstations.
Thereare alsomanyothertypesofnetworks, however,where the onne tions
donotne essarilyhaveaphysi alrealization. For example,in so ials ien esone
studies so ial networks, where the obje ts ould be people and two people are
onne ted if they have a ertain so ial relationship| see Figure 1.1. Another
exampleisformedbybiologi alnetworkssu hasneuralnetworks,generegulatory
networks orprotein-proteinintera tionnetworks|seeFigure1.2.
Sometimesnetworks anberathersmall|thenetworkinFigure1.1for
exam-pleisquitesmall|butsometimes they an alsobehuge,liketheInternet(whi h
hasmore than 500million hosts) and thewebgraph|agraph whose nodes
or-respond to stati pageson the web and whose ar s orrespond to links between
these pages|whi h hasbillions of pages that are onne ted by billions of links.
For example,in 2003Googlesear hengineindexed1.6 billionsofURLs andthis
in reasedto4.2billionsin 2004.
From theseexamplesitis learthat networks formafundamentalmodelin a
Myriel
Mlle. Baptistine
Mme. Magloire
Countess de Lo
Geborand
Champtercier
Cravatte
Count
Old Man
Napoleon
Valjean
Labarre
Marguerite
Mme. de R
Isabeau
Gervais
Fantine
Thenardier
Cosette
Javert
Fauchelevent
Bamatabois
Simplice
Scaufflaire
Woman 1
Judge
Champmathieu
Brevet
Chenildieu
Cochepaille
Mother Innocent
Mlle. Gillenormand
Marius
Enjolras
Bossuet
Gueulemer
Babet
Claquesous
Montparnasse
Toussaint
Tholomyes
Listolier
Fameuil
Blacheville
Favourite
Dahlia
Zephine
Perpetue
Pontmercy
Eponine
Boulatruelle
Brujon
Lt. Gillenormand
Gillenormand
Gribier
Mme. Pontmercy
Mabeuf
Jondrette
Mme. Burgon
Combeferre
Prouvaire
Feuilly
Bahorel
Joly
Grantaire
Child 1
Child 2
Mme. Hucheloup
Baroness T
Mlle. Vaubois
Mother Plutarch
Anzelma
Mme. Thenardier
Woman 2
Courfeyrac
Gavroche
Magnon
Figure1.1: Thenetworkofintera tionsbetweenmajor hara tersinthenovel
LesMiserablesbyVi torHugo, dividedinto11 ommunitiesrepresentedbydierent
olors[NG04 ℄.
of resear hondesigning,analyzing, andoptimizingnetworks. Themathemati al
on ept orresponding to networks are graphs. (In the sequel, we will use the
termsgraphandnetworkinter hangeably.) AgraphGisapair(V;E)whereV is
a(usuallynite)setofnodesandE
⊂
V×
V isthesetof onne tionsbetweenthe nodes. Basedonthenetwork,we anmakethegraph(edge/vertex)weighted ordire ted/undire ted. For examplefora graphwhi h models aroadnetwork,the
weightofanedge anrepresentthelengthoftheroad. Alsobymakingitdire ted
we an showone-wayortwo-wayroads.
In some appli ations it is relevant to assume that the set of verti es of the
graph is asubset of a metri spa e and the weightof ea h edge in thegraph is
the distan e between its endpoints. A metri spa e is dened as aset where a
distan ebetweenelementsofthesetisdened. Thedistan e fun tiond( alleda
metri ) shouldbenon-negative,symmetri ,haved(x;y)=0ifandonlyifx=y,
and satisfythetriangleinequality. Obviouslyanyset ofpointsinthe planewith
the Eu lidean distan e as adistan e fun tion makes ametri spa e. As amore
omplexexample,forea hgraphG(V;E)withpositiveedgeweights,we aneasily
Figure1.2: Ayeastproteinintera tionnetwork[MS02 ℄.
thisend,wedenethedistan ebetweenea hpair(u;v)
∈
V 2asthelengthofthe
shortestpath betweenuand v in G. We all (V;d) themetri spa eindu ed by
the graphG.
A geometri network omes from adding geometry to a network. More
pre- isely,ifthevertexsetofthenetworkisasubsetofd-dimensionalEu lideanspa e,
andthemetri istheEu lideanmetri ,thenthenetworkisageometri network.
Geometri networks model naturally(atleast approximately)manyreal-life
net-works, su h as road networks, railwaynetworks, and so on. Inthis ase we an
usegeometri propertiesto designoranalyzeanetwork. Inthis thesiswealways
onsiderundire tgeometri networks,unlessexpli itlystatedotherwise.
1.2 t-Spanners
When designing a network for a given set V of points, several riteria an be
taken into a ount. In many appli ations it is important to ensure afast
on-ne tion between every pair of points in V. For this it would be ideal to have
adire t onne tion between everypair of points|the network would then be a
omplete graph|but in most appli ations this is una eptable due to the high
osts. This leadsto the on epts of spanners, as dened below. Spanners were
introdu ed byPelegand S haer[PS89℄ in the ontext ofdistributed omputing
andbyChew[Che86℄inageometri ontext.
Let V be a subset of a metri spa e and for ea h u and v in V, let d(u;v)
denotethedistan e betweenuand v inthemetri spa e. Theaimisto designa
asub-quadrati numberofedges. Notethatthenumberof edgesinthe omplete
graph is quadrati in the numberof verti es. Toobtain a\good" graphwith a
sub-quadrati numberofedgeswehavetoallowa(hopefullyshort)detourbetween
pairs of points. For example, instead of asking for a dire t onne tion between
ea hpairofpoints,weareallowedtohaveasmalldetour. Thelengthofthedetour
is givenbyareal numbert>1. Giventheparametert,a onne tionbetweenu
and v in thegraphGis \good"ifthedistan e betweenuand v in thegraphG,
denoted by d G
(u;v), is at most t times the distan e between uand v. We all
su hapathat-path betweenuandv inthegraphG. Theratiod G
(u;v)=d(u;v)
is alledthedilation betweenuandv inG. AgraphGisat-spannerofitsvertex
set ifthedilation betweenea h pairofverti esin Gisbounded byt. Figure 1.3
isanexampleof1:5-spanneron532US ities.
Figure1.3: A1:5-spannernetworkson532US ities[NS07℄.
Here istheformaldenition oft-spanner.
Denition 1.2.1(t-spannerof apointset) AgraphG(V;E)isat-spannerofV,
for areal numbert>1, iffor ea h pair of pointsu;v
∈
V, wehave thatd G
(u;v)
≤
t·
d(u;v):Thedilation,orstret hfa tor,ofanetworkG(V;E)istheminimumtforwhi hG
isat-spannerofV. Whenageometri graphsatisesthet-spanner ondition,we
Intuitively,thet-spannerpropertyisstrongerthanthegraph onne tivity
prop-erty,i.e. weshouldnotonlyhaveapathbetweenea hpairofpointsinthegraph
butalsothelengthofthepathshouldbe losetothedistan ebetweenthepairof
points. The parametert de ideshow losethe t-spannerapproximatesthe
om-plete graph. Inother words, the loser tis to one,the loserthet-spanner is to
the ompletegraph.
We an easily extendDenition 1.2.1to at-spannerof agivengraph. Re all
thatforanytwoverti esuandvinagraphGwithpositiveedgeweights,d G
(u;v)
denotes thedistan e between uand v in the graphV, that is, thelength ofthe
shortestpathbetweenuand vin G.
Denition1.2.2(t-spanner ofa graph) LetG(V;E)beagraphwithpositiveedge
weights. A graphG
′
(V;E
′
) on the samevertex set butwith edge set E
′
⊆
E is at-spannerof Giffor ea h pair ofverti esu;v
∈
V wehavethatd G
′
(u;v)≤
t·
dG (u;v):
By omparingDenition1.2.1andDenition1.2.2one anseetwodieren es.
First,at-spanner ofagivengraphGisasubgraphofG. Therefore omputinga
t-spannerofagraphGissometimes alledapruning ofG. These onddieren e
is that, to he k the t-spanner ondition, the distan e betweenea h pair in the
t-spanneris omparedtothedistan ebetweenthemintheinputgraph. Notethat
in Denition 1.2.2itis notne essarythat theverti esoftheinput graphbelong
toametri spa e. Insteadweusethemetri spa eindu edbytheinputgraphto
measurethedistan ebetweenthem.
At-spanneronapointset V|whi h issubsetofametri spa e| anbeseen
asat-spanner ofthe omplete graphon V,denoted byG
(V),where theweight
ofea hedgein the ompletegraphisthedistan ebetweenitsendpoints.
Themainquestionindesigningt-spannersiswhetherspannersexistthathave
\good"properties. Thedesirablepropertiesarethefollowing:
Size: Thenumberofedgesinthegraph. Thisisthemostimportant
measure-mentand generating spannerswith asmall, ideally near-linear, size is desirable.
The fa t that geometri spanners with anear-linear number of edges and small
dilationexisthasmadethe onstru tionofspannersone ofthefundamentaltools
inthedevelopmentofapproximationalgorithmsforgeometri alproblems.
Degree: The maximumnumber of edges in identto avertex. This property
has been shown to be useful in the development of approximation algorithms
[GLNS02a, Yao82℄ and forthe onstru tionof ad ho networks [ALW +
03, Li03℄
wheresmalldegreeisessentialin tryingtodevelopfastlo alizedalgorithms.
Weight: The weight of a network G(V;E) is the sum of the edge weights.
(Re allthat forgeometri graphs theweightof an edge is simply theEu lidean
times the weight of the minimum spanning tree, denoted wt(MST(V)). Low
weight spanners have found appli ations in areas su h as metri spa e
sear h-ing [NP03, NPC02℄|see Se tion 1.3.2|and broad astingin ommuni ation
net-works[FPZW04℄|seeSe tion 1.3.3. Also,ithasbeenusedin the onstru tionof
severalapproximationalgorithms,see[CL00,RS98℄.
Spannerdiameter: Dened asthesmallestinteger
D
su h thatforanypairof verti esuandv in V, thereis at-path in thegraphbetweenuandv ontainingat most
D
edges. For wirelessadho networksitis oftendesirableto havesmall spanner diameter sin e it determines the maximum number of times amessagehasto betransmitted inanetwork. Ifagraphhasspannerdiameter
D
then itis saidtobeaD
-hopnetwork.Ithasbeenshownthat foranyset V ofnpointsin
R
dandforanyxedt>1
thereexistsat-spannerwith
O
(n)edges, onstantdegree,andwhosetotalweight isO
(wt(MST(V))) [DN97, NS07℄. Aryaet al. [AMS99℄ designed arandomized algorithm whi h generatesat-spanner ofexpe tedlinearsize and expe tedloga-rithmi spannerdiameter.
Note that some of the above properties are ompeting, e.g., a graph with
onstantdegree annot have onstantspannerdiameter,and agraphwith small
spannerdiameter annothavealinearnumberofedges[ADM +
95℄.
In most ases we are interested in t-spanners, where t is lose to 1. More
pre isely, weare interestedin s hemes where we an spe ify anyt >1and then
obtainaspannerofdilation t. Toemphasizethefa t thatt is loseto1,wewill
sometimes speakabout(1+")-spannersinsteadoft-spanners.
1.3 Why spanners?
Spannersfor ompletegraphsaswellasforarbitraryweightedgraphsnd
appli a-tionsin roboti s,networktopologydesign,distributed systems,designofparallel
ma hines,andmanyotherareas. Re entlyspannersfoundinterestingpra ti al
ap-pli ationsinareassu hasmetri spa esear hing[NP03,NPC02℄andbroad asting
in ommuni ationnetworks[ALW +
03,FPZW04,Li03℄.
Several well-known theoreti alresults alsouse the onstru tionof t-spanners
as abuilding blo k. For example, Rao and Smith [RS98℄ made abreakthrough
byshowinganoptimal
O
(nlogn)-timeapproximations hemeforthewell-known Eu lidean traveling salesperson problem, using t-spanners (or banyans).Simi-larly,CzumajandLingas[CL00℄showedapproximations hemesforminimum- ost
multi- onne tivity problems in geometri graphs. The problem of onstru ting
geometri spannershasre eived onsiderableattentionfromatheoreti al
perspe -tive,see[ADD +
93,ADM +
95, AMS99,BGM04, DHN93,DN97,DNS95, GLN02,
thebookbyNarasimhanandSmid[NS07℄. Notethatsigni antresear hhasalso
beendoneinthe onstru tionofspannersforgeneralgraphs,seeforexample,the
book by Peleg [Pel00℄ or the re ent work by Elkin and Peleg [EP04℄ and
Tho-rup and Zwi k [TZ01℄. In this se tion we mention some of the appli ations of
t-spanners.
1.3.1 Approximate minimum spanning tree
Theproblem ofndingaminimumweightspanning tree(MST)ofagivengraph
hasre entlyattra tedalotofattention. Therearelineartimealgorithmsfor
om-putingMSTinarandomizedexpe ted asemodel[KKT95℄orwiththeassumption
that edge weightare integers[FW94℄. For deterministi omparison-based
algo-rithms,slightlysuperlinearbounds areknown[GGST86℄.
Inthegeometri ase,foranydimensiond,one an ndaminimumspanning
tree of aset of n pointsin
O
(n 2)time by onstru ting the omplete graphand
omputing all the edge weights (=pairwise distan es) in
O
(n 2) time, and then
running any MST algorithm (su h as Prim's algorithm). Most of the faster
al-gorithmsforthegeometri aseuse asimpleidea: nd asparsesubgraph ofthe
ompletegraphwhi h ontainsanMST,andthen omputeanMSTofthissparse
subgraph. Intheplane,theDelaunaygraphistheappropriategraphtousesin e
it only has
O
(n) edges and an be onstru ted inO
(nlogn) time [SH75℄. It is nothelpfulinhigherdimensionsbe ausetheDelaunaygraph anhaveaquadratinumberofedges[Eri01℄.
Salowe [Sal91℄ showed that if G
′
is a (geometri ) t-spanner of a graph G
and wt(MST(G)) denotes the weightof the minimum spanning tree of G, then
wt(MST(G
′
))
≤
t·
wt(MST(G)). Using this result, we an use any (sparse) t-spanner ofa graphto omputean approximate minimum spanning tree ofthegraph. For moredetailsaboutapproximatingtheminimumspanningtreeseethe
surveybyEppstein[Epp00℄.
1.3.2 Metri spa e sear hing
Theproblemof\approximate"proximitysear hinginmetri spa esistondthe
elementsofasetwhi hare\ lose"toagivenqueryundersomesimilarity riterion.
Similaritysear hinghasbe omeafundamental omputationaltaskinavarietyof
appli ationareas,in ludingmultimediainformationretrieval,patternre ognition,
omputervisionandbiomedi aldatabases. Insu henvironments,anexa tmat h
haslittlemeaning,and proximity/distan e on epts(similarity/dissimilarity)are
typi allymu hmorefruitful forsear hing. Inall ofthese appli ationswehavea
metri whi h shows the similarity between obje ts. The smaller the distan e is
A typi alqueryqis:
•
ndallelementsinthedatabasewhi harewithindistan er fromq.•
ndthek losestelementstoq inthedatabase.Ifthedatabase ontainsnelements,we anansweraquerybyperforming
O
(n) distan e omputations. But evaluating distan es is expensive and thegoal is toredu ethenumberofdistan eevaluations. Ingeneral,thereareseveralmethodsto
redu e thenumberofdistan eevaluations,seethesurvey[CNBYM01℄. Awidely
used te hniqueforthisisAESA[Rui86℄. Themaindrawba kofthiste hniqueis
thatit omputesallpairwisedistan esandstorestheminamatrixwhi hrequires
alotofspa e.
Navarroet al. [NPC02℄ usedat-spanner as adata stru turefor metri spa e
sear hing to redu e the spa e needed for the AESA. The key idea is to regard
thet-spanner asanapproximationofthe omplete graphof distan esamongthe
obje ts,andtouseitasa ompa tdevi etosimulatethelargematrixofdistan es
required by su essful sear h algorithms su h as AESA. The t-spanner property
impliesthatwe anusetheshortestpathinthet-spannertoestimateanydistan e
withboundederrorfa tort.
They propose several t-spanner onstru tion, update, and sear h algorithms
and experimentallyevaluated them. The experimentsshowthat their te hnique
is ompetitive against urrent approa hes, and that it has a great potential for
further improvements.
1.3.3 Broad asting in ommuni ation networks
Wireless networks onsist of a set of wireless devi es ( alled nodes) whi h are
spread overageographi alarea. These nodesare able to performpro essing as
wellas ommuni ating with ea h other. Thenodes an ommuni ate via
multi-hop wireless hannels: anode an rea h all nodes inside its transmission region
while nodes farawayfrom ea h other ommuni ate throughintermediate nodes.
Wireless ommuni ationnetworkshaveappli ationsin various situationssu h as
emergen yrelief,environmental monitoring,andso on.
There are two ommon types of wireless networks: sink-based networks and
ad ho networks. Inasink-basednetwork,like ellularwirelessnetworks, thereis
one or multiple sink nodes whi h arein hargeof olle tingdata from all nodes
and managing thewhole network. On theother hand,in ad ho networks there
are no su h sink nodes and all nodes are equalin terms of ommuni ation and
networkmanagement.
Energy onsumption andnetwork performan eare themost riti al issuesin
wirelessadho networks,be ausewirelessdevi esareusuallypoweredbybatteries
Awirelessadho networkismodeledbyasetV ofnwirelessnodesdistributed
inatwo-dimensionalplane. Ea hnodehasthesamemaximumtransmissionrange
R whi h, by aproper s aling, we an assume that all nodes havethemaximum
transmissionrangetobeequalto1. Thesewirelessnodesdeneaunitdiskgraph,
denotedbyUDG(V),inwhi hthereisanedgebetweentwonodesiftheEu lidean
distan ebetweenthemisat most1. Themost ommonpower-attenuationmodel
in the literature laims that the powerneeded to support alink (u;v) is
k
uvk
,
where
k
uvk
is the Eu lidean distan e betweenu and v and is a real onstant between 2 and 5 depending on the wireless transmission environment. So thepower onsumptionof anetwork isafun tion oftheweightof thenetwork. The
minimum possible weight for a onne ted network is the weight of a minimum
spanningtree,however,this mighthavelowperforman e.
Severalgraphtheoreti almodelsareused todesignadho networkswithlow
energy onsumption and good performan e|see [Wat05℄. Using a low weight
t-spanner of the unit disk graph is one of the ways|see[ALW +
03℄. This gives
us more exibility to onstru ta network whi h has lowenergy onsumption as
wellas goodperforman e,likebounded degree. Notethat in adho networks, a
network withsmallnodedegree,hasabetter han e tohassmallinterferen e.
1.3.4 Proteins visualization
One of the most important open problems in bioinformati s is the problem of
protein folding. A protein is a long hain of mole ules alled amino a ids. In
naturethereexist20dierentaminoa idsandseveralexperimentsshowthatthe
3D-stru tureofaproteinis ompletelydeterminedbythesequen eofaminoa ids.
The protein-folding problem is the problem of determining the 3D-stru ture
ofaproteingivenitsaminoa idsequen e. Various omputationalmethods have
beenapplied tota kletheprotein-foldingproblem,with varyingsu ess. Among
theseareneuralnetworks,approximationalgorithms,metaheuristi s,
bran h-and-bound,distributedsystemsand omputationalgeometry.
Nowadays, using high omputing power and large s ale storage, resear hers
areableto omputationallysimulate theprotein-foldingpro ess in atomisti
de-tails. Su h simulationsoftenprodu ealargenumberoffoldingtraje tories,ea h
onsisting of a series of 3D onformations of the proteinunder study. As a
re-sult,ee tivelymanagingandanalyzingsu htraje toriesisbe omingin reasingly
important.
Re ently, Russel and Guibas [RG05℄ suggested using geometri spanners for
mappingasimulationto amoredis rete ombinatorial representation. They
ap-ply geometri spanners to dis overthe proximity between dierent segments of
aprotein a rossarange of s ales, and tra k the hanges of su h proximityover
motions easier. Using their stru ture it is possible to visualize proteins in
mo-tion whi h none of the ommonly used software pa kages su h as RasMol[Ras℄,
ProteinExplorer[Mar02℄, orSPV[GP97℄havebeenabletoa hieve.
1.4 Thesis overview
In this thesis we onsider several problems relatedto the design and analysis of
geometri networks.
In Chapter 2, weintrodu e the on ept of region-fault tolerant spanners for
planarpointsets,andprovetheexisten eofregion-faulttolerantspannersofsmall
size. Forageometri graphGonapointsetV andaregionF,wedeneG
⊖
F to bewhat remainsofGafter theverti esand edgesofGinterse ting F havebeenremoved. A
C
-fault tolerantt-spanner isageometri graphGonV su hthat for any onvexregion F, thegraphG⊖
F is at-spannerforG
(V)
⊖
F, whereG(V)
isthe ompletegeometri graphonV. Fault-tolerantspannersprovidehighlevels
of availabilityand reliabilityin network onne tions. These networks keeptheir
goodproperties,evenaftersomepartofthenetworkisdestroyede.g. byanatural
disaster.
Weprovethatanyset V ofnpointsadmitsa
C
-faulttolerant(1+")-spanner ofsizeO
(nlogn),forany onstant">0;ifaddingSteinerpointsisallowedthen thesizeofthespannerredu estoO
(n),andforseveralspe ial asesweshowhow toobtainregion-faulttolerantspannersofsizeO
(n)withoutusingSteinerpoints. Wealso onsiderfault-tolerantgeodesi t-spanners: thisisavariantwhere,foranydisk D, the distan e in G
⊖
D betweenany twopoints u;v∈
V\
D is at mostt times the geodesi distan e between u and v inR
2
\
D. We provethat for any point set V we an addO
(n) Steiner points to obtain a fault-tolerant geodesi (1+")-spannerofsizeO
(n). These resultsarebasedon[AdBFG07℄.In appli ations|think of road networks, for instan e|a spanner network is
sometimesexpandedbyaddingoneormoreextra onne tions. Themainquestion
isthenhowtodotheexpansionsu hthattheresultingnetworkisasgoodas
pos-sible. InChapter3westudyaproblemofthistype. Inparti ular,we onsider the
problemofaddinganedgetoagivennetworksu hthatthedilationoftheresulting
networkisminimized. Wepresentoneexa talgorithmandseveralapproximation
algorithms. Thebest approximationalgorithm omputesa(2+")-approximation
oftheoptimalsolutionin
O
(nm+n 2logn)timeusing
O
(n 2)spa e,wherenisthe
number ofverti esand m is thenumberof edges in the inputnetwork. For the
spe ial ase, when thedilationof theinputnetwork is onstant,we an improve
the approximationfa torto 1+" and therunningtime to
O
(n 2). These results
Chapter4studiestheproblemofdilationoptimaledgedeletion. Morepre isely
we are given a geometri network in the plane and wewant to nd an edge in
the network su h that its removal minimizes, or maximizes, the dilation in the
network. Anobviousappli ationis whenwewanttoremovesome onne tionsin
anexistingnetwork,e.g. duetobudget onsideration,andwewanttoknowwhi h
edges should be removed to minimize the ee t on the quality of the network.
Wesolvetheproblem in therestri ted asewhen the network is a simple y le.
A randomized algorithm is presented whi h, given a y le on a set of n points,
omputesin
O
(nlog 3n)expe tedtime,theedgeofthe y lewhoseremovalresults
in apolygonalpath of smallestpossibledilation. It is also shown that theedge
whoseremovalgivesapolygonalpathoflargestpossibledilation anbe omputed
in
O
(nlogn)time. Iftheinput y leis a onvexpolygon,thelatterproblem an besolvedinO
(n)time. Finally,itisshownthatgivena y leC,forea hedgee ofC,a(1−
")-approximationtothedilationofthepathC\ {
e}
anbe omputed inO
(nlogn)totaltime. Theseresultsarebasedon[AFK+ 07℄.
In Chapter 5 we present algorithms for omputing the spanner diameter of
at-spanner. This is the rst algorithm for omputing spanner diameter of a
t-spanner,tothebest ofourknowledge. Thetime omplexityofthemosteÆ ient
algorithmis
O
(D
·
mn),wherenisthenumberofverti es,misthenumberofedges andD
isthespannerdiameteroftheinputgraph,anditrequiresO
(n)spa e. We also omparetherunningtimeofthepresentedalgorithmsexperimentally. Theseresultsarebasedon[FG06℄.
The empiri alstudy ofalgorithms is arapidly growingresear h area.
Imple-mentingalgorithmsandtestingtheirperforman eshowstheireÆ ien yinpra ti e
andbring thealgorithmsto thepra ti alstage. InChapter6weexperimentally
studytheperforman eandqualityofthemost ommont-spanneralgorithmsfor
points in the Eu lidean plane. The experimentsare dis ussed and ompared to
thetheoreti alresultsand in several ases wesuggestmodi ationsthat are
im-plemented and evaluated. The quality measurements that we onsider are the
numberofedges,theweight,themaximumdegree,thespannerdiameterandthe
numberof rossings. We omparetherunningtimesofthealgorithmsandsuggest
someimprovements. Thisisthersttimeanextensive omparisonhasbeenmade
between the onstru tion algorithms of t-spanners. These results are based on
[FG05℄and[FG07℄.
Spanners
2.1 Introdu tion
As we mentioned before, geometri networks have appli ations in VLSI design,
tele ommuni ations, roboti s and distributed systems. The major issues with
designing su h anetwork are performan e and reliability. The spanner on ept
apturestheperforman e whenshort onne tionsbetweenthe pointsare
impor-tant. Themainquestionisthenwhetherspannersexistthathaveasmalldilation
andasmall, ideallynear-linear,numberofedges. Other desirableproperties ofa
spannerare forexample that the total weightof the edges is small, or that the
maximumdegreeislow. Asdis ussedintheintrodu tion,su hspannersdoindeed
exist: ithas beenshownthat for any setP ofn pointsin theplane andfor any
xed">0there existsa(1+")-spannerwith
O
(n=") edges,O
(1=")degree,and whosetotalweightisO
(wt(MST(P))="4
),wherewt(MST(P))istheweightofa
minimumspanning treeofP [DN97,NS07℄.
Reliability is on erned with the fa t that in many appli ations the nodes
and/orlinks in a network may fail. In a omputernetwork, for instan e, nodes
mayfailbe ause omputers an rash,andinaroadnetworklinksmayfailbe ause
roads an be ome ina essible due to a idents or maintenan e. A network is
reliable when it retains its good properties even after some nodes or links fail.
With respe t to spanners this meansthere should still be ashort path between
Fault-tolerantspannerwereintrodu edbyLev opolousetal.[LNS98℄. Inthis
paperand a follow-up paper[LNS02℄ they showed the existen eof k-vertex (or:
k-edge) fault-tolerantgeometri spanners with
O
(nklogn) edges. This was im-provedbyLukovszki[Luk99℄, whopresentedafault-tolerantspannerwithO
(nk) edges,whi hisoptimal. LaterCzumajandZhao[CZ03℄showedthatagreedyap-proa h produ esak-vertex (or: k-edge) fault-tolerantgeometri (1+")-spanner
withdegree
O
(k)andtotalweightO
(k 2·
wt(MST(P)));theseboundsare asymp-toti allyoptimal.Thepapersonfault-tolerantspannersmentionedaboveall onsiderfaultsthat
an destroyanarbitrary olle tionofkverti esoredges. Forgeometri spanners,
however, it is natural to onsider region faults: faults that do not destroy an
arbitrary olle tionof verti esand edges,but faultsthat destroyallverti esand
edgesinterse tingsomegeometri faultregion. Thisisrelevant,forinstan e,when
thespannermodelsaroadnetworkandanatural(orother)disastermakesallthe
roads insomeregionina essible. Thisisthetopi ofthis hapter: westudythe
existen e of small spannersin the plane 1
that aretolerantagainst region faults.
Beforewepresentourresults,letusdeneregion-fault toleran emorepre isely.
Let
F
beafamilyofregionsintheplane,whi hwe allthefault regions. For afaultregionF∈ F
andageometri graphG
onapointsetP,wedeneG⊖
F to bethepartofG
that remainsafter thepointsfromP insideF andalledgesthat interse tF havebeenremovedfromthegraph|seeFigure2.1. (Forsimpli ityweassumethataregionfaultF doesnot ontainitsboundary,i.e. onlyverti esand
edgesinterse ting theinteriorofF willbeae ted.)
F
Figure2.1: Theinputgraph
G
andafaultregionF,thegraphG⊖
F,andthe graphG
(P)
⊖
F.Denition 2.1.1 An
F
-faulttolerantt-spanner isageometri graphG
onP su h that for any region F∈ F
, the graphG⊖
F is a t-spanner forG
(P)
⊖
F, whereG
(P)isthe ompletegeometri graphonP.
1
The on eptsandmanyoftheresults arryovertod-dimensionalEu lideanspa e.However,
wefeelthe on eptismainlyinterestingintheplane,sowe onneourselvestotheplanar ase
Wearemainly interestedin the asewhere
F
is thefamilyC
of onvexsets. 2Weshallalso onsider the asewhereweareallowedtoadd Steinerpointstothe
graph. Inotherwords,insteadof onstru tingageometri networkforP,weare
allowedto onstru tanetworkforP
∪
QforsomesetQ ofSteinerpoints. Then wesaythat agraphG
onP∪
Qis aC
-fault tolerant Steiner t-spanner for P if, forany F∈ C
andanytwopointsu;v∈
P\
F,thedistan e betweenuand v inG⊖
F isatmostttimestheirdistan einG
(P)
⊖
F.We also study another variant of region-fault toleran e. In this variant we
requirethat the distan ebetweenanytwopointsu;v in
G⊖
F is atmostt times thegeodesi distan ebetweenuandv inR
2
\
F. Notethatthegeodesi distan e inR
2
\
F|thatis,thelengthofashortestpathinR
2\
F|isnevermorethanthe distan ebetweenuandv inG
(P)
⊖
F. We all aspannerwith this propertyanF
-faulttolerantgeodesi t-spanner. ItisnotdiÆ ulttoshowthatnitesizeF
-fault tolerantgeodesi spannersdonotexistunlessweareallowedtouseSteinerpoints.Evenin the ase of Steiner points, nite size
F
-fault tolerant geodesi spanners do not exist whenF
is the familyC
of all onvex sets. Hen e, we restri t our attentiontoD
-faulttolerantgeodesi spanners, whereD
isthefamilyofdisks in theplane.Weobtainthefollowingresults.
•
InSe tion2.2wepresentageneralmethodto onvertawell-separatedpair de omposition (WSPD)[CK93℄forP into aC
-faulttolerantspannerforP. We use this method to obtain linear-sizeC
-fault tolerant (1+")-spanners for points in onvex position and for points distributed uniformly atran-dominsidetheunitsquare,andtoobtainlinear-size
C
-faulttolerantSteiner (1+")-spannersforarbitrarypointsets.•
InSe tion2.4we onsidertwospe ial ases,fattriangulationsandpolygonal region faults with limited number of edge dire tions, for whi h linear-sizeC
-faulttolerantspanners anbeobtained.•
In Se tion 2.5 we study smallC
-fault tolerant (non-Steiner) spanners for arbitrary point sets. By ombining a more relaxed version of the WSPDwith ideasfrom -graphs[Kei88℄, weshowthat any point set P admits a
C
-faulttolerant(1+")-spannerofsizeO
(nlogn).•
InSe tion 2.6weaddressaslightlydierentproblem. Insteadof designing aC
-faulttolerantspanner,itisalsointerestingto he kwhether anexisting networkisC
-faulttolerantornot. InSe tion2.6wegiveanalgorithmwhi h, givenagraphG
, he kswhetheritisfaulttolerantunder onvexregionfaults.2
Itiseasytoseethattherearenosmallregion-faulttolerantt-spannerswithrespe tto
non- onvexfaults: if
HH
denotes thefamilyofregionsthataretheunionof twohalf-planes,then•
InSe tion2.7westudy thegeodesi ase. WeshowthatforanysetP ofn pointsthere exists aD
-fault tolerant geodesi Steiner(1+")-spannerwithO
(n)edgesandO
(n)Steinerpoints.2.2 Constru ting
C
-fault tolerant spanners usingthe WSPD
In this se tion we show a general method to obtain a
C
-fault tolerant spanner from awell-separated-pairde omposition of apointset P. Although in generalthespanner an have(n 2
)edges, weshowthat forsomespe ial asesasmaller
bound an be proven. We also show how to use the approa h to obtain small
Steinerspanners. Before westartweproveagenerallemmashowingthat, when
onstru ting
C
-faulttolerantspanners,we aninfa trestri tourattentionto half-planefaults. This lemmawillalso beused inlaterse tions. LetH
bethefamily ofhalf-planes intheplane.Proposition2.2.1 A geometri graph
G
on a set P of points in the plane is aC
-faulttolerantt-spanner ifandonlyifitisanH
-fault tolerantt-spanner. Proof. Obviouslya graphisH
-fault tolerant ifit isC
-fault tolerant. To prove theotherdire tionassumethatG
isanH
-faulttolerantt-spannerandthatF∈ C
is anarbitrary onvexregionfault. Weneedto provethat betweeneverypairofpointsu;v
∈
P\
F thereisapathinG⊖
F oflengthatmostttimesthelengthof theshortestpathinG
′
=G
(P)⊖
F.u
v
F
G
c
(P ) ⊖ F
Π(u, v)
(a)u
v
F
h
(b)u
v
F
( )Figure2.2: IllustrationoftheProposition2.2.1
If uand v are not onne ted in
G
′
we are done. Otherwise, let(u;v) bea
shortest pathbetweenuand v in
G
′
, see Figure 2.2(a). We laimthat forevery
edge (p;q) in (u;v) there is a path in
G⊖
F of length at most t· k
pqk
. Sin e the edge(p;q) liesoutsideF andF is onvex,theremust beahalf-planeh thatA
B
C
A
C
B
≥
s × ρ
ρ
ρ
Figure2.3: Denitionofwell-separatedpair.
ontainsF but doesnotinterse t (p;q),see Figure 2.2(b). Sin e
G
is anH
-fault tolerant t-spanner there is a path (p;q) between p and q inG⊖
h of length at mostt· k
pqk
. Furthermore,sin eF⊂
hthepath(p;q)alsoexists inG⊖
F, see Figure2.2( ). The laimand, hen e,thelemmafollows.2.3 Well-separated pair de omposition
Thewell-separatedpair de omposition (WSPD)was developed byCallahan and
Kosaraju [CK95℄. This powerful data stru ture has been used to solve a wide
varietyofgeometri problemslikeN-bodyproblems(usedinastronomy,mole ular
dynami s, uiddynami sandplasmaphysi s),surfa ere onstru tion[FR02℄and
manymore appli ations. We brie y review this de omposition here be ause we
willuseitinthenext hapters.
Denition2.3.1 [CK95℄ Lets>0bearealnumber,andletAandB betwonite
sets of pointsin
R
d. We say that A and B are well-separatedwith respe t to s,
ifthere aretwo disjoint d-dimensionalballs C A
andC B
, having the sameradius,
su hthat 1. C A ontainsA, 2. C B ontainsB, and
3. the minimum distan e between C A
and C B
is at least s times the radius
of C A
|seeFigure 2.3.
Theparameterswill bereferredtoas theseparation onstant. Thenextlemma
Lemma 2.3.2 [CK95℄ Let A and B be two nite sets of points that are
well-separatedwithrespe ttos,letx andpbepointsofA, andlety andq bepoints
ofB. Then
(i)
k
xyk ≤
(1+4=s)· k
pqk
,and(ii)
k
pxk ≤
(2=s)· k
pqk
.Intuitively, byLemma2.3.2, ifAandB arewell-separated,then thedistan e
betweenapointinAandapointinBisroughlythesameasthedistan ebetween
thetwosets Aand B. Alsothedistan ebetweenapairof pointswhi hbothlie
in oneofthesetsismu hsmallerthanthedistan e betweenthetwosets.
Denition 2.3.3 [CK95 ℄ LetP beaset ofnpointsin
R
d, andlets>0beareal
number. A well-separatedpairde omposition (WSPD) for P with respe tto sis
asequen eof pairs ofnon-emptysubsets ofP, (A 1 ;B 1 );:::;(A m ;B m ), su hthat 1. A i andB i
arewell-separated withrespe ttos,for 1
≤
i≤
m.2. for any two distin t points pand q of P, there isexa tly one pair (A i
;B i
)
inthe sequen e, su hthat (i)p
∈
A i andq∈
B i , or(ii) q∈
A i andp∈
B i ,The integerm is alledthe size ofthe WSPD.
Inotherwords,awell-separatedpairde ompositionofapointsetP onsistsofa
set ofwell-separatedpairsthat overallthepairsofdistin tpoints,i.e.,anytwo
distin tpointsbelongtothedierentsetsofsomepair.
Callahanand Kosarajushowed thatfor anypoint setin Eu lideanspa e and
for any onstant s > 0, there always exist a WSPD of size m =
O
(s dn) and it
an be omputed in
O
(s dn+nlogn) time. Inthe geometri problems, when we
need all point pairs in a set, we an easily use aWSPD of the point set as an
approximationwithlinearsize.
2.3.1 Constru ting a
C
-fault tolerant spannerCallahan and Kosaraju[CK93℄ showedthat the WSPD an beused to obtaina
small(1+")-spanner. SimilarideaswereusedearlierbySalowe[Sal91,Sal92℄and
Vaidya[Vai88,Vai89,Vai91℄. Toobtainthe(1+")-spannerone simply omputes
a WSPD
W
with respe t to s :=4+8=", and then forea h well-separatedpair (A;B)∈ W
oneaddsanarbitraryedge onne tingapointfromAtoapointinB.Unfortunatelythis onstru tionisnot
C
-faulttolerant,be auseafaultF an destroy the spanner edge that onne ts a pair (A;B), while some other edgesweneedtoaddmorethanasingleedgefor(A;B). LetCH(A)andCH(B)denote
the onvexhullsof AandB, respe tively. Atrstsightitseemsthataddingthe
twoouter tangents of CH(A) and CH (B) to our spanner may lead to a
C
-fault tolerantspanner, but thisis notthe aseeither. Instead,wewill triangulatetheregionin betweenthe two onvex hulls in an arbitrarymanner, as illustrated in
Figure2.4(a).
A
B
(a) (b)
Figure2.4: (a)Illustratingthe onstru tionoftheWSPD-graph.(b)Pointsin onvex
position.
LetE(A;B)bethesetofedgesinthetriangulationaddedbetweenCH(A)and
CH(B), and let
G
betheobtainedgraphwith edgesetE
:= P(A;B)
∈W
E(A;B).
NotethatanytriangulationbetweenCH(A)and CH (B)hasthesamenumberof
edges. Throughoutthe hapterwewillusethenotation
| · |
todenotethenumber ofelementsinaset.Lemma2.3.4 The graph
G
is aC
-fault tolerant (1+")-spanner for P of size P(A;B)
∈W
|
E(A;B)
|
.Proof. Thesize of thegraphis obviously P
(A;B)
∈W
|
E(A;B)
|
, soit remainsto show that it is aC
-fault tolerant (1+")-spanner. Now weobserve that for any half-plane h,{
(A\
h;B\
h) : (A;B)∈ W}
is a WSPD for P\
h. Hen e, by Proposition 2.2.1 and the properties of the WSPD it is suÆ ient to show thefollowing: Lethbeahalf-planefault,letu;vbepointsnotinh,andlet(A;B)be
apairwith u
∈
Aandv∈
B; thenthere isanedge e∈
E(A;B)betweenCH(A) andCH (B)that isoutsideh.Tosee this werst provethat, givena point set P and atriangulationT of
P, the graph T
⊖
h is onne ted for any half-plane h. Assume without loss of generality that his belowand bounded by ahorizontal line. Sin e anypoint ofP
\
h noton the onvexhull musthavean edge onne ting it to a point further awayfrom h,we an walkfrom pawayfrom halong edges of T until we rea hapoint onthe onvex hullof P. Moreover, anytwo onvexhull points in P
\
h an be onne ted by onvexhulledges outsideh. It followsthat T⊖
his indeed onne ted.Now onsideranytriangulationT onA
∪
B thatin ludesE(A;B). ThenT⊖
h mustbe onne ted. Sin eu;v6∈
h,andu∈
Aand v∈
B,this meanstheremust beanedgee∈
E(A;B)outsideh.2.3.2 Linear-size spanners for spe ial ases
Themethoddes ribedabove anbeusedtogetsmall
C
-faulttolerantspannersfor severalspe ial ases. For example, ifP is in onvexpositionthen|
E(A;B)| ≤
3 foranypair(A;B)inthede omposition,see Figure2.4(b),soweget:Theorem 2.3.5 ForanysetP ofnpointsin onvexpositionintheplaneandany
">0,thereexistsa
C
-faulttolerant(1+")-spannerofsizeO
(n=" 2).
Next weshowthatwe analsogeta
C
-faulttolerantspannerwhose expe ted sizeislinearifthepointsetP isgeneratedbypi kingnpointsuniformlyatrandomin theunitsquare.
Lemma 2.3.6 LetP beasetofnuniformlydistributed pointsintheunit square
andAbeasub-squareoftheunitsquare. Thentheexpe tednumberofpointson
the onvexhullofP
∩
AisO
(log (n·
area (A))).Proof. Ifnpointsareuniformly distributed intheunit squarethenitis known
thattheexpe tednumberofpointsonthe onvexhullofthepointsis
O
(logn)[HP98, RS63℄.Now let X be the number of points on the onvex hull of P
∩
A and let Y :=|
P∩
A|
. Clearly EXP[Y℄ =n·
area (A). Bythe law of totalexpe tation [Ros98,Proposition4.1℄ ifX andY aretworandomvariablesthenEXP[X℄=EXP[EXP[X
|
Y℄℄;therefore
EXP [X℄ = EXP[EXP[X
|
Y℄℄ = EXP[O
(log (Y))℄≤ O
(log(EXP [Y℄) ) (Jensen's inequality[Ros98, p.418℄) =O
(log(n·
area(A))):Nowwe ombinetheideasfromthepreviousse tionwithLemma2.3.6to onstru t
a(1+")-spanneroftheuniformlydistributedpointset P.
Theorem 2.3.7 Let P be a set of n points uniformly distributed in the unit
square U. For any " >0 there is a
C
-fault tolerant(1+")-spanner of expe ted sizeO
(n="2 )forP.
Proof. Constru taquadtreepartitioningofU intosmallerandsmallersquares,
untilea hsquare hassize (side length) roughly1=
√
roughly1=nwhi hmeanstheexpe ted numberofpointsin aleafregionis
O
(1). Thequadtree hasO
(n)leaves. Level` of the quadtree orrespondsto a regular subdivisionofU intosquaresofsize1=2`
. One anshowthatthereexistsaWSPD
W
:={
(A i ;B i )}
i ofsizeO
(n=" 2)forP su hthatforea hi,thepair(A i
;B i
)either
orrespondstotwosquaresatthesamelevel,orA i
andB i
arebothsingletonpoints
thatlieinnearby ells(orthesame ell) ofthenal subdivision. Moreover,ifwe
denote byn `
the numberof pairsof the WSPD at level` of the quadtree, then
n ` =
O
(2 2` =" 2). The existen eof aWSPD with these properties followsrather
dire tlyfrom theresultsofFis herand Har-Peled[FHP05℄. For ompletenesswe
brie ysket hanargumentforoursetting.
Foranodeofthequadtree,letP()denotethesubsetofpointsfromP inside
thesquare orrespondingto . Consideralevel `of thequadtree. For ea hpair
ofnodes ;
′
atlevel`su hthatthepointsetsP()andP(
′
)arewell-separated
whilethepointsetsof theparentsof and
′
arenotwell-separated,weputthe
pair (P();P(
′
))into the WSPD. In addition, forea h pair of leaf nodes ;
′
su hthatP()andP(′
)arenotwell-separated,weputapair(
{
p}
;{
q}
)intothe WSPDforeverypairp∈
P()andq∈
P(′
). Itiseasytoverifythatthisindeed
denesaWSPD. Thebound on thenumberofpairsaddedfor ea h levelfollows
fromastandardpa kingargument.
Now onsider asquare atlevel`. ByLemma 2.3.6,be ausetheareaof is
1=2 2`
,theexpe tedsizeofthe onvexhullofthepointsin is
O
(log (n=2 2`)).
If (A;B) is an arbitrary pair in
W
whi h appears at level ` of the quadtree thenEXP[
|
E(A;B)|
℄≤
EXP[|
CH(A)|
+|
CH(B)|
℄ = EXP[|
CH(A)|
℄+EXP[|
CH(B)|
℄=
O
(log (n=2 2` )): Therefore EXP 2 4 X (A i ;B i )∈W
|
E(A i ;B i )|
3 5 = X (A i ;B i )∈W
EXP[|
E(A i ;B i )|
℄ = 1 2 logn X `=1O
n ` log(n=2 2` ) = 1 2 logn XO
(2 2` =" 2 )log(n=2 2` ) :Toboundthis summation,wesetm:= 1 2
lognandweget:
1 2 logn X `=1 2 2` log (n=2 2` ) = m X `=1 2 2` (2m
−
2`) = 2 m X `=1 2 2` (m−
`) = 2 m−
1 X k =0 2 2(m−
k )·
k (bysettingk=m−
`) = 2 2m+1 m−
1 X k =0 k 2 2k≤
2 2m+1∞
X k =0 k 2 2k =O
(n):Hen e theexpe tedsizeofthegenerated(1+")-spanneris
O
(n=" 2).
2.3.3
C
-fault tolerant Steiner spannersAboveweshowedthattheWSPD an beusedto onstru t
C
-faulttolerant span-nersofsmallsizewhenthepointsarein onvexpositionoruniformlydistributed.For arbitrarypoint sets, however,the sizeof thespanner may be (n 2
). In this
se tion wewill showthat ifweareallowedto addSteiner points,we an always
usetheabovemethod togetalinear-sizespanner:
Theorem 2.3.8 For any set P of n points in the plane and any ">0, one an
onstru t a
C
-fault tolerant Steiner(1+")-spannerof sizeO
(n=" 2)byadding at
most4(n
−
1)Steinerpoints.Theideaisto addaset Qof SteinerpointstoP su hthat
|
E(A;B)|
=O
(1)for any pair(A;B)in theWSPD of P∪
Q. Then the theorem immediately follows from Lemma2.3.4.Our method is based on the WSPD onstru tion by Fisher and
Har-Peled [FHP05℄. Their onstru tion uses a ompressed quadtree, whi h is
dened asfollows.
Let
T
(P)bethequadtreeonP. Wedenotethesquare orrespondingtoanode∈ T
(P)by(), and the subsetof pointsfrom P inside ()by P(). When someofthepointsarevery losetogether,aquadtree anhavesuperlinearsize. Aompressedquadtree
T
∗
(P)forP thereforeremovesinternalnodes from
T
(P)for whi hallpointsfromP lieinthesamequadrantof(). A ompressedquadtreehasatmostn
−
1internalnodes. FisherandHar-Peled[FHP05℄showthatone an obtainaWSPD of sizeO
(s2
n) forP that onsistsof pairs(P( 1 );P( 2 ))where 1 and 2 arenodesin
T
∗
(P).The set Q of Steinerpointsthat we use is dened as follows. Let
T
∗
(P) be
a ompressed quadtree for P. Without loss of generality, we may assume that
no point from P lies on any of the splitting lines. For ea h internal node of
T
∗
(P),weaddthefour ornerpointsof()toQ. Toavoiddegenerate ases,we
slightlymoveea hpointintotheinteriorof(). Notethattwo(ormore)squares
( 1
)and ( 2
) mayshare, for instan e, their topright orner. In this asewe
addthe(slightlyshifted) ornerpointonlyon e. Theresultingset Qhassize at
most4(n
−
1). ThenextlemmanishestheproofofTheorem2.3.8.= point from P
= Steiner point
Figure2.5: IllustrationfortheproofofLemma2.3.9 .
Lemma2.3.9 Let
T
∗
(P) be a ompressed quadtree for P := P
∪
Q, where the initialboundingsquareU isthesameasforT
∗
(P),andlet beaninternalnode
of
T
∗
(P). ThenCH(P())hasat mostfourverti es.
Proof. If the square () ontains zeroor one point from P then at most one
Steinerpointhasbeenaddedinside (),andthelemmaistrue. If() ontains
twoormorepointsthentherearetwo ases,bothillustratedinFigure2.5.
Let bethenodeof
T
∗
(P) su h that P()=P()
∩
P. Note that thefour shifted orners of () were added as Steinerpointsto Q. If () =() thenCH(P())is asquare. Otherwise, ()
⊂
(). Inthis aseCH (P()) isformed bythreeofthefour ornersof()togetherwiththeunique ornerof ()thatgeneratedaSteinerpointatsomean estorof in
T
∗
(P),seeFigure2.5. Hen e,
2.4 Spe ial ases
In this se tion wepresentalgorithms for onstru ting fault tolerant spannersin
twospe ial ases. InSe tion2.4.1, wegiveanalgorithm that onstru ta
C
-fault tolerant spanner for any point set whi h admits a fat triangulation. Then, inSe tion 2.4.2, we onstru tspannerswhi h arefault tolerant undermorelimited
regionfaults.
2.4.1
C
-fault tolerant fat triangulationsWe allatriangulationofapointset-fat ifallitstrianglesare-fator,inother
words,ifallanglesinthetriangulationareatleast. KaravelasandGuibas[KG01℄
showedthatany-fattriangulationT ofapointsetP isa2-spannerforP. To
make the spanner
C
-fault tolerant, we add some extra edges: we add an edge betweeneverypair ofpoints u;v∈
P su h that there isapathbetween uandv in T onsistingoftwoedges.Theorem 2.4.1 Let P be a set of n points in the plane and let T be a -fat
triangulation of P. Then we an augment T with aset of
O
(n=) extra edges su hthattheresultinggeometri graphisaC
-faulttolerant2-spanner.Proof. We onne t ea h node v to all other nodeswithin twosteps from v. In
otherwordsweaddanedgebetweenea hpairofpoints onne tedbyapathoftwo
edges. LetT
′
betheresult. Obviouslyweaddatmost P
v
D
·
deg(v)edges,where D isthemaximumdegreein the triangulationT and deg(v) isthe degreeofthenodevof T. Sin eforea htriangulation P
v
deg(v)
≤
6nweaddO
(D·
n)edges to thetriangulationT. NotethatD=O
(1=)sin eT isa-fat triangulation.Now the laim is that T
′
is a
C
-fault tolerant 2-spanner. Using Proposi-tion2.2.1,it suÆ estoshowthat T′
isan
H
-faulttolerant2-spanner. Lethbe an arbitraryhalf-plane andp;q∈
P\
hbe two arbitrarypoints. Karavelas and Guibas [KG01, Theorem 2.1℄ proved that there exist a 2-path (p;q) betweenp and q in T zig-zagging above and below the line onne ting p to q|see
Fig-ure 2.6(a). Note that allthe edges in thispath interse t the segmentbetweenp
and q.
If all the nodes on (p;q) lies outside h we are done. Otherwise assume
p
′
∈
(p;q)liesinsidehandletq 1andq 2
betheotherendpointsofthetwoedges
on (p;q) in identto p
′
|see Figure 2.6(b). Sin ethe segmentpq liesoutsideh
andanyedgeon(p;q)interse tpq,thepointsq 1
andq 2
lieoutsideh. Be ausewe
addededgesbetweenpairswithin twosteps|thedashededgesinFigure2.6(b)|
we anrepla ethetwofatedges(p
′
;q 1 )and(p′
;q 2 )with(q 1 ;q 2). Thiswaywe an
p
q
Π(p, q)
(a)p
q
h
p
′
q
1
q
2
(b)Figure2.6: IllustrationfortheproofofTheorem2.4.1
2.4.2 Limited boundary dire tions
Nowweputsomelimitationontheregionfaults. Let
H
′
beafamilyofhalf-planes
withat mostk boundary dire tions. Bythefollowingpro edure,whi h usesthe
WSPD, we an makean
H
′
-fault tolerant(1+")-spanner ofea hpointset P of
npointswith
O
(kn=" 2 )size. Let{
d i}
k i=1bethe setof dire tions wherethe boundaryof ea h half-plane in
H
′
is parallel to one of d i
's. For ea h 1
≤
i≤
k assume d (1) i and d (2) i are thetwodire tions perpendi ular to d i
. To onstru t a
H
′
-fault tolerantspanner we
omputeaWSPDofthepointsetwithrespe ttos:= 4("+2)
"
. Thenforea hpair
(A;B) in the WSPD and for ea h dire tion d i
, we add two edges, one between
theextremepointofAandBindire tiond (1) i
andtheotherbetweentheextreme
pointsofAandB indire tion d (2) i
. SeeAlgorithm2.4.1formoredetails.
Algorithm2.4.1: Bounded-Boundary-dire tions
Input: P,">0andaset ofdire tions
{
d i}
k i=1 . Output:H
′
-faulttolerant(1+")-spanner
G
=(P;E
).W
:=WSPDofP w.r.t. s:= 4("+2) " ; 1 forea h (A;B)∈ W
do 2 fori:=1;2;:::;kdo 3AddanedgebetweenextremepointsofA andB indire tiond (1) i
; 4
AddanedgebetweenextremepointsofA andB indire tiond (2) i ; 5 end 6 end 7 return
G
; 8Theorem 2.4.2 Let P beaset of npointstheplane and
H
′
beafamilyof
half-planeswithatmostkboundarydire tions. Thenforea h">0,we an onstru t
an
H
′
-faulttolerant(1+")-spannerofsize
O
(kn=" 2)in
O
((nlogn+kn)=" 2Proof. Obviouslyweaddatmost2k edgesforea hpairinWSPDandtherefore
the size of the graph is
O
(kn=" 2). Also the time omplexity of the algorithm
is straight forward. Therefore to omplete the proof, we show that the graph
generated byAlgorithm2.4.1is
H
′
-faulttolerant. Toshowthis,itis suÆ ientto
showthatforea hh
∈ H
′
andany(A;B)intheWSPDwhi hissituatedpartially
outsideh,wehaveanedgeoutsidehwhi h onne tAtoB.
Sin eA andB are partiallyoutsideh,theextremepointsofthem inat least
one ofthedire tions perpendi ulartotheboundaryofhisoutsideh. Thismeans
thattheedgesbetweentheextremepoints,whi hareaddedbythealgorithm,lies
outsideh.
Remark 2.4.3 Atrstitmayseemthatwe angeneralizetheresultstoanyfamily
of onvexpolygons with bounded number of edge dire tions (for example
axis-parallel polygons). However as you an see in Figure 2.7 this is not the ase.
p
p
1
p
2
q
q
1
q
2
F
A
i
B
i
Figure2.7: Counterexampleforaxis-parallelpolygonalfaults.
2.5
C
-fault tolerant spanners for arbitrary pointsets
In this se tionwe onsider the problem of onstru ting asparse
C
-fault tolerant (1+")-spannerforanarbitrarysetPofnpointsintheplanewithoutusingSteinerpoints. Themethodthatwasdes ribedinthepreviousse tiondoesnotguarantee
asmallspannerin general. Here wewilldes ribeamethodthat isguaranteedto
resultinaspannerofsize
O
(nlogn).Throughout this se tion d(
·
;·
) denotes the (Eu lidean) shortest distan e be-tween two obje ts (points, disks, et .), and radius(D) denotes the radius of a2.5.1 SSPDs and fault-tolerant spanners
Theproblem withtheWSPD inourappli ationisthat,eventhoughthenumber
ofpairsintheWSPDis
O
(n),thetotalnumberofpointsoverallthepairs anbe (n2
). Thereforewewillintrodu earelaxedversionoftheWSPD,theSSPD.
Denition2.5.1 LetAandBbetwosetsofpointsintheplane,andlets>0be
a onstant. Wesaythat A and B are semi-separatedwithrespe t to separation
onstantsiftherearetwodisjointdisksD A andD B ,su hthat (i) D A ontainsAandD B ontainsB, (ii) d(D A ;D B )
≥
s·
min (radius(D A );radius(D B )).Thus weallowthe ballsD A
and D B
to be of dierentsizes and weonly require
thatthedistan ebetweenthedisksislargerelativetothesmallerdisk. Notethat
usingthesamenotationswe anreformulatethedenition ofwell-separatedwith
respe ttosasd(D A ;D B )
≥
s·
max(radius(D A );radius(D B )).WenowdeneourSSPD.
Denition2.5.2 Let P beaset of npointsin theplane and lets>0beareal
number. Asemi-separatedpairde omposition(SSPD)forP withrespe ttosisa
olle tion
{
(A 1 ;B 1 );:::;(A m ;B m)
}
ofpairsofnon-emptysubsetsofP su h that1. A i
andB i
aresemi-separatedwithrespe ttos, foralli=1;:::;m.
2. foranytwodistin tpointspandqofP,thereisexa tlyonepair(A i
;B i
)in
the olle tion,su hthat (i)p
∈
A i andq∈
B i or(ii)q∈
A i andp∈
B i .Theweight of a set A, denoted by
|
A|
, is dened as the numberof points in A, theweightofasemi-separatedpair(A;B)is thesumoftheweightsofA andB,andtheweightofanSSPDisthetotalweightofallthepairs. Laterwewillprove
that itis possibleto omputeanSSPD of weight
O
(nlogn). First, however,we willshowhowto usetheSSPD toobtainaC
-faulttolerantspanner. Theideais to add edges to thespanner for ea h pair in the SSPD. Be ause the pairsin anSSPDareonly semi-separated,however,addingasingleedgefor everypairdoes
notne essarilyleadtoagoodspanner. Thereforeweuseanideathatisalsoused
inthe onstru tionof-graphs[Cla87,Kei88℄.
Considerapair(A;B)in anSSPDfor P. Thenthereexist twodisjointdisks
D A
andD B
that ontainAandB respe tively,andforwhi h
d(D A ;D B )
≥
s·
min (radius(D A );radius(D B )):Assume without loss of generality that radius(D A
)
≤
radius(D B), and let o A
denote the enter of D A
|see Figure 2.8(a). Theset E(A;B) ofedges added to