/
aquadraticgapbetweenLasVegasandDeterminism,andallowingerror,thegapcanb aquadraticgapbetweenLasVegasandDeterminism,andallowingerror,thegapcanb

aquadraticgapbetweenLasVegasandDeterminism,andallowingerror,thegapcanb - PDF document

luanne-stotts
luanne-stotts . @luanne-stotts
Follow
368 views
Uploaded On 2015-10-14

aquadraticgapbetweenLasVegasandDeterminism,andallowingerror,thegapcanb - PPT Presentation

PrTxAxB6gxAxBforeveryinputxAxB2XAXBDenotebyCkgthecostofthebestkrounderrorprotocolforgandsimilarlyde neCAkCBkThecase0egCk0gdenotesLasVegaserrorlessprotocolsFinal ID: 160717

Pr[T(xA;xB)6=g(xA;xB)]foreveryinput(xA;xB)2XAXB.DenotebyCk(g)thecostofthebestk-round-errorprotocolforg andsimilarlyde neCA;k;CB;k.Thecase=0(e.g.Ck0(g))denotesLasVegas(errorless)protocols.Final

Share:

Link:

Embed:

Download Presentation from below link

Download Pdf The PPT/PDF document "aquadraticgapbetweenLasVegasandDetermini..." 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.


Presentation Transcript

aquadraticgapbetweenLasVegasandDeterminism,andallowingerror,thegapcanbeexponential.Weusesimpleinformationtheoreticandprobabilisticargumentstostrengthenthelowerboundof[4]intwoways.Firstweimprovetheir(k�1)-rounddeterministiclowerboundongkto (n)(forkn=logn),thusshowingthatrandomnesscanbecheaperbyafactorofk=log2nfork-roundprotocols.Thisresultalsoprovidesthelargestgapknownfork&#x-292;logninthedeterministicmodel-thepreviousonewasobtainedin[4]viacountingarguments.ThefactthatthesimulationofMcGeoch[10]isconstructive,givesthesamegapinthearbitrarypartitionmodelforanexplicitfunction,resolvinganopenquestionof[4].Second,weprovethattheprobabilisticupperboundaboveisnotveryfarfromoptimal{wegivean (n=k2)lowerbound,establishinganexponentialgapintheprob-abilisticsettingbetweenkand(k�1)-roundprotocolsforanexplicitlygivenfunction.Theexistanceofsuchfunctions(withsomewhatlargergap)wasprovedbyHalstenbergandReischuk[6],viacomplicatedcountingarguments.Theonlypreviousexponentialgapforanexplicitfunctionwasshownfork=2byYao[20].Westressthesimplicityofourprooftechnique,incontrasttothatof[6].WehaverecentlylearnedthesimilartechniqueswereusedbySmirnov[16]toprovean (n=(k(logn)k))lowerboundongk,whichismuchweakerthanourbound,butestablishanexponentialgapaswell.Finally,weusethecommunicationcomplexitycharacterizationofcircuitdepthofKarchmerandWigderson[8]toestablishgkasa\complete"problemformonotonedepth-kBooleancircuits.(ThisresultwasindependentlydiscoveredbyYanakakis[18]).Thusasimpledeterministicreductionenablestoderivethemonotoneconstant-depthhierarchyofKalwe,Paul,PippengerandYannakakis[7]fromtheconstant-roundhi-erarchyof[4].(Thereversedirectionwasprovenin[7]).Wespeculatethatournewprobabilisticlowerboundmayservetoextendthemonotonecircuithierarchyresulttodepthabovelogn,viaprobabilisticreductions(aswasdonebyRazandWigderson[15]).1.2.TheMulti-PartyModel.Chandra,FurstandLipton[2]devisedthemulti-partycommunicationcomplexitymodel.HeretplayersP1;P2;:::;PtaretryingtocomputeaBooleanfunctiong(x1;x2;:::;xt),wherexi2f0;1gni.(Untilnowallworkinthismodelconsideredequallengthinputs,i.e.ni=nforalli).Thetwististhat2 Pr[T(xA;xB)6=g(xA;xB)]foreveryinput(xA;xB)2XAXB.DenotebyCk(g)thecostofthebestk-round-errorprotocolforg,andsimilarlyde neCA;k;CB;k.Thecase=0(e.g.Ck0(g))denotesLasVegas(errorless)protocols.Finally,ifweleaveTadeterministicprotocol,andchoosetheinputuniformlyatrandom,wecande nethe-errordistributionalcomplexityDk(g)tobethecostofthebestk-roundprotocolforwhichPr[T(xA;xB)6=g(xA;xB)],underthisdistribution.Thefollowinglemmasareuseful.Lemma2.1.[20]Foreveryg;�0Dk2(g)2Ck(g).Lemma2.2.Forallconstants1 3�0�0Ck0(g)=O(Ck(g)).2.2.Results.LetVA;VBbetwodisjointsets(ofvertices)withjVAj=jVBj=nandV=VA[VB.LetFA=ffA:VA!VBg,FB=ffB:VB!VAgandf=(fA;fB):V!Vde nedbyf(v)=(fA(v)v2VAfB(v)v2VB.Foreachk0de nef(k)(v)byf(0)(v)=v;f(k+1)(v)=f(f(k)(v)).Letv02VA.Thefunctionswewillbeinterestedincomputingisgk:FAFB!Vde nedbygk(fA;fB)=f(k)(v0).Remarks:Inthefollowingtheoremsnotethatthenumberofinputbitstoeachplayerisnlogn,andthattheyholdforeveryvalueofk.WealsonotethatonecanmakegkaBooleanfunctionbytaking(say)theparityoftheoutputvertex.AllourupperandlowerboundsapplytothisBooleanfunctionaswell.Theorem2.3.[14]CA;k(gk)=O(klogn).Theorem2.4.CB;k(gk)= (n�klogn).Theorem2.5.CB;k1=3(gk)=O((n=k)logn),CB;k0(gk)=O((n=k)log2n).Theorem2.6.CB;k1=3(gk)= (n k2�klogn).Remark:Attheendoftheproofsoftheorems2and4weexplainwhythe\ugly"term�klognintheselowerboundscanbeessentiallyignored.Intheremainderweshowthe\completeness"ofgkformonotonedepthkcircuits.Letgk=gk;ntostressthateachplayergetsnvertices.Definition:Forabooleanfunctionhde neLd(h)tobethesizeoftheminimalmonotoneformulaofdepthdandunboundedfaninthatcomputesh.De neLSd(k;n)tobethemaximumofLd(h)overallfunctionshthatcanbecomputedbymonotonecircuitsofunboundedfanindepthkandtotalsizen.De neLFd(k;n)tobethemaximumofLd(h)overallfunctionshthatcanbecomputedbyaformulaofdepthk4 ProofofTheorems2.3and2.5.CA;k(gk)klognfollowseasily,sinceinroundttherightplayerknowsf(vt�1)=vtandcansendtheselognbitstothesecondplayer.Theideainbeatingthedeterminstic (n)lowerboundwhenthewrongplayerBstartsisasfollows:FirstBchoosesarandomsubsetUVBwithjUj=10n=k,andsendstoAffB(u):u2Ug.NowitisA'sturnandtheystartsendingeachotherv1;v2;:::asabove,butlaggingoneround\behindschedule".However,withprobability2=3,oneofthevi'swillbeinU,whichallowsthemtosavetworounds,and\ nishontime".ThisgivesCB;k(gk)=O((k+n=k)logn).ThisalgorithmcanbemadeLas-VegaswithanextrafactorofO(logn)inthecomplexity.ProofofTheorems2.4and2.6Letf=(fA;fB)2FAFBbetheinput.LetT0beadeterministick-roundprotocolforgkinwhichBsendsthe rstmessage.Notethatatanyroundt1,ifitisB0sturntospeak,thenvt�1=f(t�1)(v0)2VA,andviceversa.ItwillbeconvenienttoreplaceT0byaprotocolTinwhichinanyroundt1,wereplacethemessagembythemessage(m;vt�1).Byinductionont,thisisalwayspossiblefortheplayerwhoseturnitis.Inparticular,itimpliesthatlognbitsaresentperround.ThusifT0usedCbits,TusesC+klognbits.WewillassumeTusesn 2bits,(willbechosenlater),andobtainacontradiction.EverynodezoftheprotocoltreeTcanbelabeledbytherectangleFzAFzBofinputsarrivingatz.BythestructureofT,ifzisatlevelt1(therootisatlevel0),thenv0;v1;:::;vt�1aredeterminedinFzAFzB.WeshallassumetheinputischosenuniformlyatrandomfromFAFB,soinfactweshallboundfrombelowthedistributionalcomlexity.Thustheprobabilityofarrivingatzis(FzAFzB),andgiventhattheinputarrivedatz,itisuniformlydistributedinFzAFzB.Themainlemmabelowintuitivelyshowsthatiftheinputarrivedatzandtherectangleatzhasniceproperties,thenwithhigh(enough)probabilitytheinputwillproceedtoachildwofzwhichisequallynice.NicemeansthatbothFzA;FzBarelargeenough,andthattheplayernotholdingvt�1hasverylittleinformationonvt=f(vt�1).Denotebyczthetotalnumberofbitssentbytheplayersbeforearrivingatz.AssumewithoutlossofgeneralitythatAspeaksatz.LetfzAandfzBberandomvariablesuniformlydistributedoverFzAandFzB,respectively.RecallthatTuses 2nbits,andletsatisfy=Maxf4p ;400g.De neztobeniceifitsatis es:6 ifthealgorithmgivesonebit(sayparity)oftheanswer,itiscorrectwithprobability1 2+2p .ConclusionofTheorem2.4Take=10�4.TherootofTisnice,sobythemainlemmaandinductionwehaveapositiveprobability(2�k)ofreachinganiceleaf,contradictingthefactthattheprotocolnevererrs.ThisprovesonlyCB;k(gk)= (n�klogn),sinceweaugmentedanarbitraryT0toaniceprotocolT.Gettingridofthe�klognterm,achievingthelowerboundCB;k(gk)= (n)(whichisstrongerwhenkn logn)requiresamoredelicateargumentthatwesketchbelow.Theideaistofollowthesamestepsoftheproofwiththefollowingchanges.(1)WestaywiththeoriginalprotocolT0,aswecannota ordtheplayerssendinglognbitsperroundasintheniceprotocolT.(2)Westill xthevertexvt�1bytheplayersendingthemessageatroundt,butavoidpayinglognbitsforthisinformationbyremovingthisvertexfromouruniverse.ThustheinformationIismeasuredrelativetoasmallersetofpointersateveryround.(3)Weproveaweakermainlemma,whichisclearlysucientinthedeterministiccase,namelythateverynicenodezhasatleastonenicechildw.Thedetailsarelefttotheinterestedreader.ConclusionofTheorem2.6.Pick=10�4k�2.Thustheprobabilityofnotreachinganiceleafisk1 25k=1 25,andtheprobabilitythattheprotocolanswerscorrectlyislessthan1 25+(1 2+2 5p k)0:95.ThuswegetfromLemmas2.1and2.2thatDB;k1=20(gk)= (n k2�klogn).Thisboundis (n k2)forallk(n logn)1=3.Forlargerkonecanusethetriviallowerboundkwhichappliestoeveryk-roundprotocol.Notethatwhenkn1=3thistrivialboundislargerthann=k2.ProofofTheorem2.7:Asmentionedabove,theleftinequalitywasprovenin[7],soweproveonlytherightinequality.TheproofisbasedontheKarchmer-Wigdersonchar-acterizationofcircuitdepthintermsofcommunicationcomplexity,whichcanbestatedasfollows.ForeverymonotonefunctionhonnvariableswithmintermsMin(h)andmaxtermsMax(h)de neacommunicationsearchproblemRmhMin(h)Max(h)[n]inwhichplayerAgetsamintermS2Min(h),playerBgetsamaxtermT2Max(h),andtheirtaskisto ndanelementinS\T.ThenmonotoneformulaeforhandprotocolsforRmharein1-1correspondenceviathesimplesyntacticidenti cationof_gateswithplayerA'smovesand^gateswithplayerB'smoves.Inparticular,depthcorrespondstothenumberofrounds,andlogarithmofthesizetothecommunication8 ProofofTheorem3.2We rstrecallafundamentallemmafrom[11]regardingthedistributionofhashvaluesgivenlittleinformationonthehashfunctionandtheargument.Lemma3.3.[11]LetH=fh:I!OgbeacollectionofuniversalhashfunctionsfromdomainIintorangeO.LetAI,BO,CHandp=jBj=jOj.ThenjPr[h(x)2Bjx2A;h2C]�pjq pjHj=(jAjjCj)Restrictthevalueofjtobej2[p n].Thusweconsiderh:f0;1gn!f0;1gp nwhichisstillauniversalhashfunction.AssumeM3(u)p n=5.Thismeansthatthereisanewprotocol(independentofj)tocomputez=h(y)inwhichP3sendsp n=5bits,andthenplayersP1andP2cancomputeeachbitofzseparately,usingaltogethern=5bits.Let(h;y)bechosenuniformlyatrandom.Simpleaveragingshowsthattherearemessagesm1;m2;m3ofP1;P2;P3,respectively,whichunderthisdistributionsatisfyPr[m1]2�n=4,Pr[m2]2�n=4,andPr[m3]2�p n=4.Asm1corespondstoasubsetCofallhashfunctions(inputstoP1),andm2correspondstoasubsetAofallinputstoP2,wecanuseLemma3.3withjHj=jCj2n=4,1=jAj2�3n=4,B=fzgandp=2�p ntoobtainPr[h(y)=zjm1;m2;m3]2p n=4Pr[h(y)=zjm1;m2]2p n=42�p n=2=2�p n=41:Letf:f0;1gm!f0;1gnbeanarbitraryfunction,andforanym0mde negf:f0;1gm0f0;1gm�m0f0;1glogn!f0;1gbygf(x1;x2;x3)=f(x1x2)x3,wheredenotesconcatenation.Thenexttheoremgivestherelationshipofsize-depthtrade-o sincircuitsto3-roundobliviousprotocols.Theorem3.4.Iffabovecanbecomputedbyacircuitoffan-in2,sizeO(n)anddepthO(logn),thenMs(gf)=O(n=loglogn).ProofofTheorem3.4Letf:f0;1gm!f0;1gnbecomputedbyacircuitCofsizeO(n)anddepthO(logn).ByaresultofValiant[17],thereares=O(n=loglogn)wiresinC,e1,e2;:::,eswiththefollowingproperty.Foreveryinputx2f0;1gm,andeveryj2[n];f(x)jisdeterminedbythevaluese1(x);:::;es(x)onthesewires,togetherwiththevaluesofxi;i2SjwithjSjjp n.Tocomputegf,notethatP3hasaccesstox=(x1x2)10 REFERENCES[1]L.Babai,N.Nisan,M.Szegedy:Multipartyprotocolsandlogspace-hardpseudorandomsequencesProc.ofthe21stSTOC(1989)1-11.[2]A.Chandra,M.Furst,R.Lipton:Multi-partiProtocolsProc.ofthe15thSTOC(1983)94-99.[3]L.Carter,M.Wegman:UniversalhashfunctionsJournalofComp.andSys.Sci.18(1979)143-154.[4]P.Duris,Z.Galil,G.Schnitger:LowerBoundsofCommunicationComplexityProc.ofthe16thSTOC,(1984)81{91[5]J.Hastad,M.Goldmann:OnthepowerofsmalldepththresholdcircuitsProc.ofthe31stFOCS,(1990)610-618.[6]B.Halstenberg,R.Reischuk:OnDi erentModesofCommunicationProc.ofthe20thSTOC(1988)162-172.[7]M.Klawe,W.J.Paul,N.Pippenger,M.Yannakakis:OnMonotoneFormulaewithRestrictedDepthProc.ofthe16thSTOC,(1984)480{487[8]M.Karchmer,A.Wigderson:MonotoneCircuitsforConnectivityRequireSuper-LogarithmicDepthProc.ofthe20thSTOC,(1988),539{550[9]T.Lam,L.Ruzzo:ResultsonCommunicationComplexityClassesProc.ofthe4thStructuresinComplexityTheoryconference,(1989)148-157.[10]L.A.McGeoch:AStrongSeperationBetweenkandk�1RoundCommunicationComplexityforaConstructiveLanguageCMUTechnicalReportCMU-CS-86-157(1986)[11]Y.Mansour,N.Nisan,P.Tiwary:ThecomputationalcomplexityofuniversalhashingProc.ofthe22ndSTOC(1990).[12]K.Mehlhorn,E.Schmidt:LasVegasisbetterthanDeterminisminVLSIandDistributedCom-putingProc.ofthe14thSTOC(1982),330-337.[13]N.Nisan:PseudorandomGeneratorsforSpaceBoundedComputationProc.ofthe22ndSTOC,(1990)204-212.[14]P.H.Papadimitriou,M.Sipser:CommunicationComplexityProc.ofthe14thSTOC,(1982)330{337[15]R.Raz,A.Wigderson:ProbabilisticCommunicationComplexityofBooleanRelationsProc.ofthe30thFOCS,(1989)562{567[16]D.V.Smirnov:Shannon'sinformationmethodsforlowerboundsforprobabilisticcommunicationcomplexity,Manuscript(inRussian),(1989)[17]L.Valiant:Graphtheoreticargumentsinlow-levelcomplexity,TechnicalReportCS13-77,Uni-versityofEdinburgh(1977).[18]M.Yannakakis:Privatecommunication.[19]A.C.-C.Yao:SomeComplexityQuestionsRelatedtoDistributiveComputing,Proc.ofthe11thSTOC,(1979)209{213[20]A.C.-C.Yao:LowerBoundsbyProbabilisticArguments,Proc.ofthe24thFOCS,(1983)420{42812

Related Contents


Next Show more