171MulticutFirstconsiderthemulticutproblemGivenagraphGVEKpairsofterminalverticesfsitigandacostfunctionontheedgescERthemulticutproblemasksforaminimumcostcutofGthatseparatessiandtiforalli ID: 837398
Download Pdf The PPT/PDF document "CS880ApproximationsAlgorithmsScribeMattE..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
1 CS880:ApproximationsAlgorithmsScribe:Mat
CS880:ApproximationsAlgorithmsScribe:MattElderLecturer:ShuchiChawlaTopic:SparsestCutandBalancedCutDate:3/20/07 17.1MulticutFirst,considerthemulticutproblem.GivenagraphG=(V;E),Kpairsofterminalverticesfsi;tig,andacostfunctionontheedgesc:E!R,themulticutproblemasksforaminimum-costcutofGthatseparatessiandtiforalli.Lasttime,wegaveaO(logk)approximationforthisproblem.The(relaxed)linearprogramforthisproblemisasfollows;callit\Primal1".minimizeXe2Ecedewhered(si;ti)18idisametricWecanrewriteasfollows:minimizeXe2EcedewherePi=fAllpathsfromsitotigXe2Pde18i8P2Pide08eThedualofthisLP,whichwe'llcall\Dual1",isasfollows:maximizeXiXP2Pifi;PwherePi=fAllpathsfromsitotigXiXP2PiP3efi;Pce8efi;P0Dual1solvesthemax-summulti-commodity\rowproblem:cerepresentsthecapacityofanedge,andfi;Pistheamountof\rowdirectedfromsitotialongthepathP.TheLPtriestomaximizethetotalamountofcommodity\row.Lemma17.1.1Multicutisalwayslargerthanthecorrespondingmax-summulti-commodity\row.Lemma17.1.2MulticutisatmostO(logK)timesthecorrespondingmax-summulti-commodity\row.Theorem17.1.3Whenk=2,multicutequalsmax-summulti-commodity\row.1 17.2MaximumConcurrentMulticommodityFlowAsolutiontoDual1maystarvesomecommoditieswhileroutingothers.Incontrast,max-concurrentmulticommodity\rowroutesequalfractionsofallcommoditieswhilerespectingca-pacities.Thus,wedevisethefollowingLPformax-concurrentmulticommodity\row,whichwecallDual2:maximizetwhereXiXP2PiP3efi;Pce8eXP2Pifi;Prit8ifi;P08i;8P2PIntuitively,thedierencebetweenDual1andDual2isthatDual1seekstomaxim
2 izethetotal\rowacrossindependentcommodit
izethetotal\rowacrossindependentcommodities,whileDual2seekstomaximizetheminimumofasetofweighted\rows.Thisisthemaximumconcurrentmulticommodity\rowproblem.Primal2,thedualofthemaximumconcurrentmulticommodity\rowproblem,isasfollows:minimizeXedecewhereXe2Pdeyi8i;8P2PiXiriyi1de08eyi08iThecostsceareconstantsoftheprobleminstance,soPrimal2willseektominimizethevaluesforde.Thus,theywillbenolargerthantheyareconstrainedtobe,soyi=d(si;ti),theshortestdistancefromsitotiwhereeachedgeehaslengthde.So,wecandevisethefollowinglinearprogram,equivalenttoPrimal2:minimizeXecedewhereXirid(si;ti)1disametricUptoscalingd,thisisthesameasthefollowingprogram:minimizePecede Pirid(si;ti)wheredisametric,2 whichwilltiedirectlytothesparsestcutproblem.17.3SparsestCutLetG=(E;V)besomegraph.LetT,thesetofterminals,beasetofpairsofvertices.ForanycutS,asubsetofV,let(S)denotethesparsityofS,with(S)=E(S;S) (SS)\T:Thus,(S)isthesizeofthecutSdividedbythenumberofterminalsthatSseparates.ThesparsestcutproblemtakesGandT,andndsthecutSthatminimizes(S).Wecangeneralizethisfurther,byintroducingacostcontheedgesandaweightriforeachterminali.Then,ourmoregeneral(S)lookslikethis:(S)=c(E(S;S)) Pfijsi2S,ti=2Sgri:Considertheuniformversionofsparsestcut,whereweletthesetofterminalsbeallpossiblepairs.Thatis,weletT=VVandru;v=1forallu=v.Then,thesparsityis(S)=c(E(S;S)) jSjjSj;whichisquitesimilartotheexpansionofaset,fromthecontextofexpandergraphs.Here,wegiveaO(logKlogD)-approximationforthisproblem,whereD=PiriandK=jTj.Nexttime,we'
3 llndaO(logK)-approximation.Thebestknown
llndaO(logK)-approximation.ThebestknownecientapproximationisaO(p logKloglogK)-approximation.SupposewesolvePrimal2,yieldingthemetricd.Letyidef=d(si;ti).WeknowthatPiriyi1andthat8e;de1.Weneedtorounddintoacutmetric|ametricwithonlyonesandzeroes|withoutmuchincreasingthesparsity.Considerthespecialcasewhere8i;yi1 2ymax.Letymaxdef=maxiyi,d0def=d=ymin,andy0idef=d0(si;ti)foralli.Then,8i;y01,andd0isafeasiblesolutiontothemulticutLP.So,wecanfeedd0tothemulticutapproximationalgorithmwesawlastlecture.ThatalgorithmwillyieldacutofvalueO(logK)Peced0e.Wededuce:O(logK)Xeced0eO(logK)1 yminXecedeO(logK)1 ymaxXecede:ThedemandthisseparatesisthusPiri(1=ymax)Piriyi.So,thesparsityofthisalgorithm,inthisspecialcase,isatmostO(logK)Pcede Priyi:3 So,nowconsiderthegeneralcase,withanarbitrarilylargeratiobetweenymaxandymin.DeneIx,thesetofallyiinaconveniently-denedinterval,as:Ix=nijyi2ymax 2x+1;ymax 2xio:Whenxisconstrainedtotheintegers,it'sclearthateveryyiiscontainedinexactlyoneIx.ForeachIx,ouralgorithmwillconstructamulticutinstanceasinthespecialcase,butitwillscaledby2x+1=ymaxinsteadof1=ymax.Thesparsityforeachoftheseinstancesisnottoolarge:(Ix)O(logK)Pecede Pi2Ixriyi:IfthereexistsaconstantandanxsuchthatPi2Ixriyi 1Piriyi,thenthesparsityofthisinstanceisatleastO(logK)Pcede=(Priyi).Weclaimthatwecanignoreallisuchthatyiymax=D2.Again,DisthetotaldemandPri.DenethesetWcontainingweevaluesofyi,W=ijyiymax=D2 .IgnoringWcanresultinthelossofatmostPi2WriyiPi2Wriymax=D2ymax=D1=Dfromthedenominatorofouralgorithm'sspa
4 rsity.AssumingD2,ignoringWhasonlyasmall
rsity.AssumingD2,ignoringWhasonlyasmallconstantapproximationcost.(IfD2,thisisaneasyboundarycase,whichwecaneecientlyhandleinanad-hocway.)Thus,weneedtoconsideronlythoseIxwhere2xD2.Thereareatmost2logDsuchsets,so1=(2logD).ThisyieldsaO(logKlogD)-approximation.17.4BalancedCutGivenagraphG=(V;E),thebalancedcutproblemdemandsthemin-costcutsuchthateachsidehasatleastnnodes,forsomeconstantvalueof1 2.Itisknownthatthisproblemisinapproximableton2 =OPTifP=NP.Thisisanabsurdlypoorapproximation.So,weconsiderinsteadapseudo-approximationalgorithm,inwhichweapproximateboththeobjectivefunctionoftheoptimalsolutionandtheparametersofitsinstance.So,inthiscase,whenaskedforacutwithabalanceof,weinsteadoutputacutwithabalance0,suchthat0and01=3.IftheoptimalcutofbalancehascostC,ourcutwillhavecostnogreaterthanO(logn)C=( 0).Noticethat,thoughwehaveareasonableboundontheratiobetweenthesizeofourcutandC,theratiobetweenthesizeofourcutandC0maybeunbounded.Thealgorithmemploysadirectreductiontothesparsestcutproblem,lettingT=VVandri=1.Then,thesparsityofacutSis,asbefore,c(E(S;S))=(jSjjSj).Wediscussfurtherdetailsnexttime.Eventhoughthisalgorithmisfarfromoptimal,itisactuallyuseful.Thispseudo-approximationhasapplicationsindivide-and-conqueralgorithmsongraphs.Itensuresthatwecanalwaysdivideagraphintotwopieces,eachwithsizeroughlylinearinthesizeoftheoriginalgraph,suchthatthecutbetweenthepiecesisn'ttoolarge.Thisyieldslog-depthrecursion,whichdivide-and-conqueralgorithmsdemand,whileboundingthecostofrecombiningpieces.