27.06.2013 Views

The guessing number of undirected graphs - Index of

The guessing number of undirected graphs - Index of

The guessing number of undirected graphs - Index of

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

<strong>The</strong><strong>guessing</strong><strong>number</strong><strong>of</strong><strong>undirected</strong><strong>graphs</strong><br />

DemetresChrist<strong>of</strong>idesandKlasMarkström<br />

Abstract<br />

In[Rii07],Riisintroduceda<strong>guessing</strong>gamefor<strong>graphs</strong>whichisequivalent<br />

t<strong>of</strong>indingprotocolsfornetworkcoding.Inthispaperweproveupperand<br />

lowerboundsforthewinningprobability<strong>of</strong>the<strong>guessing</strong>gameon<strong>undirected</strong><br />

<strong>graphs</strong>.Wefindoptimalboundsforperfect<strong>graphs</strong>andminimallyimperfect<br />

<strong>graphs</strong>,andpresentaconjecturerelatingtheexactvalueforall<strong>graphs</strong>to<br />

thefractionalchromatic<strong>number</strong>.<br />

Finally, we discuss extensions <strong>of</strong> our results to directed <strong>graphs</strong> and<br />

<strong>graphs</strong>withnon-uniformprobabilities.<br />

1 Introduction<br />

Anactivearea<strong>of</strong>researchincommunicationtheoryduringthelasttenyears<br />

hasbeenthedevelopment<strong>of</strong>protocolsfornetworkcoding[ACLY00].Innetwork<br />

codingthereareseveralsendersandreceiverswhowishtopassmessagesbetween<br />

eachother,howevertheroutersinthenetworkcanonlysendonemessageat<br />

atime. Inordertoavoidbottlenecks,networkcodingallowstheroutersto<br />

computeanddistributenewmessages,aslongasthereceiverscancomputethe<br />

senders’originalmessagefromthecollection<strong>of</strong>newmessagestheyreceivedin<br />

itsstead.<br />

In[Rii07],Riisconnectednetworkcodingtothemucholderproblem<strong>of</strong>finding<br />

an optimalBooleancircuitforaBooleanfunction. In thatpaper, Riis<br />

disprovedaconjecture<strong>of</strong>Valiantincircuitcomplexityandshowedthatboth<br />

networkcodingandthisform<strong>of</strong>circuitcomplexitywereequivalenttoacertain<br />

type<strong>of</strong>multiplayer<strong>guessing</strong>gameonagraph.<br />

<strong>The</strong><strong>guessing</strong>gamecanbedescribedasfollows:Eachvertexv<strong>of</strong>agraphG<br />

isassignedaplayer,alsodenotedbyv,andauniformlyrandomintegerfrom<br />

{1...s}. Aplayervcanseethe<strong>number</strong>sassignedattheneighbours<strong>of</strong>v,but<br />

notthe<strong>number</strong>atv. Withoutcommunication,eachplayermustnowmakea<br />

guessastothevalue<strong>of</strong>itsownrandom<strong>number</strong>. <strong>The</strong>team<strong>of</strong>playerswins<br />

ifall<strong>of</strong>themguesscorrectlytheirownvalue,andlosesifanyoneisincorrect.<br />

<strong>The</strong>objectiveistonowfindastrategywhichmakesthewinningprobabilityas<br />

largeaspossible.Atfirstitmightseemlikethereisnoway<strong>of</strong>gettingahigher<br />

winningprobabilitythans −n onannvertexgraph,butRiisobservedthatthis<br />

isfarfrombeingthecase. IfGisthecompletegraph,theplayerscanusethe<br />

1


followingsimplestrategy: Eachplayerpickstheunique<strong>number</strong>suchthatthe<br />

sum<strong>of</strong>thatandallother<strong>number</strong>sis0modulos. Ifeveryplayerfollowsthis<br />

strategytheywillallbecorrectwhenthesum<strong>of</strong>all<strong>number</strong>is0modulos,which<br />

happenswithprobabilitys −1 ,avaluewhichdosnotevendependonn.<br />

In[Rii07],Riisconsideredthisgameongeneraldirected<strong>graphs</strong>,whereplayer<br />

vcanseethe<strong>number</strong>atvertexuifandonlyifthereisanedgefromvtou.<br />

Motivatedby thisgeneralquestion, inthispaperwe studytheproblemfor<br />

<strong>undirected</strong><strong>graphs</strong>. Asweshallsee,theproblem<strong>of</strong>findingoptimal<strong>guessing</strong><br />

gamestrategiescanbetranslatedintoaquestionregardingthesize<strong>of</strong>thelargest<br />

independentsetinanauxiliarygraph,andusingthisgraphwecanfindgood<br />

boundsforthewinningprobabilities.<br />

Hereisanoutline<strong>of</strong>thepaper.Inthenextsectionwegivesomeformaldefinitionsandprovesomebasicboundsforthewinningprobability<strong>of</strong>the<strong>guessing</strong><br />

games. InSection3,weexplainhowthe<strong>guessing</strong><strong>number</strong>scanbedetermined<br />

bycomputingthefixedpoints<strong>of</strong>somespecificmaps. Wealsousethemethods<br />

developedtheretoshowthatthereisanaturallydefinedlimit<strong>of</strong>the<strong>guessing</strong><br />

<strong>number</strong>swhichwecalltheasymptotic<strong>guessing</strong><strong>number</strong>. InSection4,wefind<br />

lowerboundsonthe<strong>guessing</strong><strong>number</strong>susingfractionalcliquecovers.InSection<br />

5,wedefinethecodegraph<strong>of</strong>the<strong>guessing</strong>gameandprovethatthe<strong>guessing</strong><br />

<strong>number</strong>canbecomputedbydeterminingtheindependence<strong>number</strong><strong>of</strong>thisgraph.<br />

InSection6,wefindupperboundsonthe<strong>guessing</strong><strong>number</strong>susingentropyinequalitiesandposeaconjectureregardingtheasymptotic<strong>guessing</strong><strong>number</strong><strong>of</strong><br />

eachgraph.FinallyinSection7weconsidersomegeneralisations<strong>of</strong>themethods<br />

<strong>of</strong>thepaperinsimilarcontexts.<br />

2 Definitionsandsomebasicbounds<br />

Westartoutbymakingamoreformaldefinition<strong>of</strong>the<strong>guessing</strong>game.<br />

Definition2.1.Inthe<strong>guessing</strong>gameonagraphG,eachvertexvisassigned<br />

anintegerxv∈{1,...,s}uniformlyatrandom.<br />

Givenavertexv<strong>of</strong>agraphG,astrategyforplayervforthe<strong>guessing</strong>game<br />

onGisafunctionfv :{1,...,s} N(v) →{1,...,s}. <strong>The</strong>value<strong>of</strong>fviscalled<br />

v’sguess. AstrategyFforthe<strong>guessing</strong>gameonGisasequence<strong>of</strong>functions<br />

(fv) v∈V(G)suchthateachfunctionfvisastrategyforplayervforthe<strong>guessing</strong><br />

game.<br />

WesaythattheplayerswinifallguessesarecorrectandwewriteCor(F)<br />

todenotethisevent.<br />

Wenowdefinethe<strong>guessing</strong><strong>number</strong><strong>of</strong>thegraphviathewinningprobability<br />

intheoptimalstrategyforthatgraph.<br />

Definition2.2. <strong>The</strong><strong>guessing</strong><strong>number</strong>gn(G,s)<strong>of</strong>agraphGwithrespectto<br />

thepositiveintegersisthelargestβsuchthatthereexistsastrategyFforthe<br />

<strong>guessing</strong>gameonGsuchthatwithprobability 1<br />

s n−β everyplayervguessesits<br />

ownvaluexv,i.e.Pr(Cor(F))= 1<br />

s n−β.<br />

2


Ingeneral,onecouldconsiderstrategieswhereeachplayermakesarandom<br />

choicebasedontheavailableinformation. Howeverthereisalwaysanoptimal<br />

strategywhichisdeterministic,soweonlyconsiderdeterministicstrategies.<br />

Lemma2.3. Everyrandomisedstrategyforthe<strong>guessing</strong>gameonagraphG<br />

haswinningprobabilityatmost 1<br />

s n−gn(G,s).<br />

Pro<strong>of</strong>.ArandomisedstrategycanbedescribedbyassigningaprobabilityPr(F)<br />

toeachdeterministicstrategyF. Howeverthewinningprobability<strong>of</strong>sucha<br />

strategyis<br />

<br />

F<br />

Pr(F)Pr(Cor(F))max<br />

F Pr(Cor(F))=<br />

1<br />

s n−gn(G,s).<br />

Usingthisterminologywecannowgiveaformalversion<strong>of</strong>thestrategyfor<br />

thecompletegraphwhichwedescribedintheintroduction.<br />

Example 2.4. LetGbethe completegraphKn on nvertices. Wedefine<br />

a strategyforthe <strong>guessing</strong>gameon G by definingfv tomap thesequence<br />

(xu) u∈N(v)totheuniqueintegerx ′ v∈{1,...,s}suchthatx ′ v+ <br />

u∈N(v) xuis<br />

divisiblebys.Wewillcallthisthecliquestrategy.Observethatalltheplayers<br />

guesscorrectlyifandonlyifx ′ v=xvforeveryv∈V(G)whichholdsifandonly<br />

if <br />

. Since<br />

v∈V(G) xvisdivisiblebys. <strong>The</strong>probability<strong>of</strong>thiseventisthus 1<br />

s<br />

theprobabilitythatasingleplayerguessesitsownvaluecorrectlyisalso 1<br />

s ,this<br />

cannotbeimprovedandwefindthatgn(Kn,s)=n−1.<br />

Fromthecliquestrategywecandefineanaturalstrategyforgeneral<strong>graphs</strong><br />

bypartitioningthevertexsetintocliques.<br />

Definition2.5.Acliquecover,orcliquepartition<strong>of</strong>agraphGisapartition<br />

<strong>of</strong>V(G)intovertexdisjointcliques.<strong>The</strong>cliquecover<strong>number</strong>cp(G)<strong>of</strong>Gisthe<br />

minimumcardinality<strong>of</strong>acliquecover<strong>of</strong>G.<br />

Notethatthecliquesinacliquecover<strong>of</strong>Ginduceapartition<strong>of</strong>thecomplement<strong>of</strong>Gintoindependentsets,i.e.apropervertexcolouring<strong>of</strong>G.<br />

Hence<br />

cp(G)=χ G .<br />

Lemma2.6.ForeverygraphGandeverypositiveintegers,gn(G,s)n−<br />

cp(G)<br />

Pro<strong>of</strong>.Astrategygivingthisboundcanbeconstructedbytakingaminimal<br />

cliquecoverandlettingtheplayersineachcliquefollowthecliquestrategyfor<br />

thatclique.Thisgivesprobabilityatleast 1<br />

s cp(G)thatallplayersguesscorrectly.<br />

Wecallthestrategyusedinthepro<strong>of</strong><strong>of</strong>theabovelemma,thecliquecover<br />

strategy.<br />

Intheotherdirection,wecanproveanupperboundonthewinningprobabilitybyconsideringindependentsetsinthegraph.Asusual,weletα(G)denote<br />

thesize<strong>of</strong>thelargestindependentsetinG.<br />

3


Lemma2.7.ForeverygraphGandeverypositiveintegers,gn(G,s)n−α(G)<br />

Pro<strong>of</strong>.LetIbeamaximumindependentsetinGandletF =(fv) v∈V(G)be<br />

astrategy. Wechoosetherandom<strong>number</strong>(xu) u∈V(G)intwostages. Inthe<br />

firststage,wegenerateall<strong>number</strong>sxu withu /∈I. ObservethatsinceI is<br />

independentallfunctionsfuwithu∈Icannowbedetermined. Inthesecond<br />

stage,wegenerateall<strong>number</strong>sxuwithu∈Iandobservethattheprobability<br />

thatplayervguessescorrectlyisexactly1/swiththeeventsbeingindependent.<br />

ItfollowsthatPr(Cor(F))s −α(G) asrequired.<br />

Eventhesetwosimpleboundsareenoughtodeterminethe<strong>guessing</strong><strong>number</strong>s<br />

exactlyforlargeclasses<strong>of</strong><strong>graphs</strong>. Infact,thesetwoboundsdeterminethe<br />

<strong>guessing</strong><strong>number</strong><strong>of</strong>Gpreciselywhenα(G)=cp(G). Oneparticularlynatural<br />

classwhichsatisfiesthispropertyistheclass<strong>of</strong>perfect<strong>graphs</strong>,introducedby<br />

Berge[Ber63].Thisistheclass<strong>of</strong><strong>graphs</strong>suchthatχ(H)=ω(H)forallinduced<br />

sub<strong>graphs</strong><strong>of</strong>G. Aclassicaltheorem<strong>of</strong>Lovász[Lov72]tellsusthatagraphis<br />

perfectifandonlyifitscomplementisperfect. Inourcontextthismeansthat<br />

foraperfectgraphGwehaveα(G)=ω(G)=χ G =cp(G)<br />

Corollary2.8.IfGisperfectthengn(G)=n−α(G).<br />

ThistellsusamongotherthingsthatifGisbipartitethengn(G,s)=n−<br />

α(G).Onestrikingconsequenceisthatadisjointunion<strong>of</strong>nK2’s,i.e.amatching<br />

<strong>of</strong>sizenhasthesame<strong>guessing</strong><strong>number</strong>asthecompletebipartitegraphKn,n,<br />

despitebeingasubgraph<strong>of</strong>ithavingafactor<strong>of</strong>nfeweredges.Inotherwords,<br />

allthisextrainformationthattheplayershaveinplayingthe<strong>guessing</strong><strong>number</strong><br />

forKn,ncontributesnothingtothewinningprobabilityforthe<strong>guessing</strong>game<br />

onthisgraphcomparedwiththe<strong>guessing</strong>gameonthematching<strong>of</strong>sizen.<br />

3 Fixedpointsandtheasymptotic<strong>guessing</strong><strong>number</strong><br />

Anotherway<strong>of</strong>describingthe<strong>guessing</strong>problemisinterms<strong>of</strong>fixedpoints<strong>of</strong>the<br />

mappingsgivenbydifferentstrategies.NotethatastrategyFcanbeviewedas<br />

amappingfromAs={1,...,s} V(G) intoitself.<strong>The</strong>strategyFguessescorrectly<br />

onagivenoutcome<strong>of</strong>random<strong>number</strong>sx=(xv) v∈V(G)ifF(x)=x. Thusthe<br />

problem<strong>of</strong>findingagoodstrategyforthe<strong>guessing</strong>gamecanbeinterpretedas<br />

findingamappingF:As→Aswithasmanyfixedpointsaspossible,wherefv<br />

onlydependsonxu∈N(v).Wecallamappingwiththedependencestructurea<br />

strategymappinganddenotetheset<strong>of</strong>allstrategymappingsonAsbyS(G,s).<br />

IfweletFix(F)denotethe<strong>number</strong><strong>of</strong>fixedpoints,thenwecancomputethe<br />

<strong>guessing</strong><strong>number</strong>as<br />

gn(G,s)= max<br />

F∈S(G,s) log sFix(F)<br />

LetusdefineFix(G,s)=max F∈S(G,s)Fix(F).<br />

Whilethe<strong>guessing</strong><strong>number</strong>isgivenbythemaximum<strong>number</strong><strong>of</strong>fixedpoints<br />

foramapping<strong>of</strong>thistypetherearealsomappingsattheotherextremewithno<br />

fixedpoints,foreverynon-trivialgraph.<br />

4


Example 3.1. LetG=K2 andtakes=2. Letthefirstplayerguessthe<br />

samevalueashesees(i.e.theoutcome<strong>of</strong>therandomexperiment<strong>of</strong>thesecond<br />

player),andletthesecondplayerguesstheopposite<strong>of</strong>thevalueitsees.Forthis<br />

strategy,Fix(F)=0. Thusforeverygraphwithatleastoneedgethereexists<br />

astrategywhichneverguessesallvaluescorrectly,andthiscanbeextended<br />

tolargersaswell. Thisis<strong>of</strong>coursetobeexpectedastheaverage<strong>number</strong><strong>of</strong><br />

fixedpointsoverallstrategymappingsis1. Soifweareablet<strong>of</strong>indstrategy<br />

mappingswithmorefixedpoints,thentheremustalsoexiststrategymappings<br />

withn<strong>of</strong>ixedpointsatall.<br />

S<strong>of</strong>ar,foralltheexamples<strong>of</strong><strong>graphs</strong>wehaveseen,the<strong>guessing</strong><strong>number</strong>was<br />

independent<strong>of</strong>s. Ourearlierresultsshowthatinordertoseeadependence<br />

onswemustconsidernon-perfect<strong>graphs</strong>. <strong>The</strong>strongperfectgraphtheorem<br />

[CRST06]tellsusthatagraphisperfectifandonlyifneitherGnoritscomplementcontainsaninducedoddcycle<strong>of</strong>lengthatleastfive.<br />

Soweturnour<br />

attentiontothefivecycleC5,whichisthesmallestnon-perfectgraph.<br />

Example3.2.LetG=C5,bethecycleon5vertices. Sinceα(C5)=2and<br />

χ(C5)=3,Lemmas2.6and2.7give2gn(G,s)3foreverys.<br />

Fors=2,letFbethestrategywhereavertexguesses2ifbothitsneighbours<br />

havevalue1,andguesses1otherwise.Thisstrategyguessescorrectlyonthefollowingx:{11212,12112,12121,21121,21211}.Thuswehavegn(C5,2)log<br />

25,<br />

andaquickcomputercheckshowsthatthisisindeedoptimal.<br />

Fors=3,letFbethestrategywhereavertexguesses2ifbothitsneighbours<br />

havevalue1,guesses3ifatleastoneneighbourhasvalues3andnoneighbourhas<br />

value2,andguesses1otherwise.ThisstrategyhasFix(F)=11andacomputer<br />

checkshowsthatitisthebestsymmetricstrategy.Howeverbycomputercheck,<br />

usingthemethods<strong>of</strong>thenextsection,weknowthatthereexistsamorecomplex,<br />

vertexdependent,strategywhichisoptimalandhasFix(F)=12.Thuswehave<br />

gn(C5,3)=log 312


leastonu∈N(v),thenf ′ v(x)=s+1,otherwisef ′ v(x)=fv(x ′ ),wherex ′ is<br />

obtainedfromxbydefiningx ′ utobeequaltoxuifxu∈{1,...,s}andequalto<br />

1otherwise.ObservethatF ′ isindeedastrategymapping,everyfixedpoint<strong>of</strong><br />

Fisalsoafixedpoint<strong>of</strong>F ′ andmoreoverF ′ has(s+1,...,s+1)asafixed<br />

pointaswell.<br />

Lemma3.4.IfHisasubgraph<strong>of</strong>GthenFix(H,s)Fix(G,s)<br />

Pro<strong>of</strong>.GivenastrategymappingFforthe<strong>guessing</strong>gameonH,weextendit<br />

toastrategymappingF ′ forthe<strong>guessing</strong>gameonGbydefiningf ′ v(x)tobe<br />

equalt<strong>of</strong>v({xu:u∈V(H)})ifv∈V(H)andtobeidentically1otherwise.It<br />

isimmediatethatF ′ isastrategymappinghavingatleastasmanyfixedpoints<br />

asF.<br />

<strong>The</strong>lastinequalityisveryfarfrombeingstrict,asshownbyourearlier<br />

examplewiththebalancedcompletebipartitegraphon2nverticesversusthe<br />

matching<strong>of</strong>sizen.<br />

Wecanalsoboundthe<strong>number</strong><strong>of</strong>fixedpointsforcompositevalues<strong>of</strong>s<br />

Lemma3.5.Fix(G,s1s2)Fix(G,s1)Fix(G,s2)<br />

Pro<strong>of</strong>.Thisfollowsbysimplywritingeachrandom<strong>number</strong>xvas(a−1)s1+b,<br />

withb∈{1,...,s1}anda∈{1,...,s2},andusingtheoptimalmappingsfors1<br />

ands2toguessaandbindependently.<br />

<strong>The</strong>orem3.6.<strong>The</strong>limitlims→∞log sFix(G,s)exists,andisatmost|V(G)|.<br />

Moreover,itisequaltosup s∈Nlog sFix(G,s).<br />

Pro<strong>of</strong>.SupposeGhasnverticesandletuswriteas=Fix(G,s).Bythedefinition<strong>of</strong>Fix(G,s)itfollowsthatass<br />

n foreverysandsoifthelimitexiststhen<br />

itisatmostn. ByLemma3.3wehavethatas+1asforeverys∈Nandby<br />

Lemma3.5wehavethatastasatforeverys,t∈N.Ouraimistoshowthat<br />

lims→∞log sasexistsandisequaltosup s∈Nlog sas.Givens,t∈Nweclaimthat<br />

log sas k<br />

k+1 log tat,<br />

wherek=⌈logs/logt⌉. <strong>The</strong>resultimmediatelyfollows. Indeed,givenε>0,<br />

letℓ=sup s∈Nlog sasandtaketlargeenoughsuchthatlog tat>(1−ε)ℓ.With<br />

thistfixed,wecannowpickslargeenoughsuchthatk/(k+1)>(1−ε).We<br />

deducethatlog sas>(1−ε) 2 ℓ,thusliminflog sasℓandsolimlog sasexists<br />

andisequaltoℓ.Toprovetheclaim,observethat<br />

logsas= logtas logts logtatk logttk+1log tak t<br />

k+1<br />

= k<br />

k+1 log tat.<br />

Hencemayintroducethefollowingasymptoticversion<strong>of</strong>the<strong>guessing</strong><strong>number</strong>.<br />

6


Definition3.7.<strong>The</strong>asymptotic<strong>guessing</strong><strong>number</strong>gn(G)<strong>of</strong>Gis<br />

gn(G)= lim<br />

s→∞ log sFix(G,s)<br />

Example3.8.LetusreturnbacktoC5.Wehavealreadyseenthatgn(C5,2 k )<br />

gn(C5,2)=log 25.Itfollowsthatgn(C5)log 25.Inparticulargn(C5)>2.In<br />

thenextsectionwewillimprovefurtheronthisbound.<br />

4 Lowerboundsviafractionalcliquecovers.<br />

S<strong>of</strong>arwehavedeterminedexactlythe<strong>guessing</strong><strong>number</strong>s<strong>of</strong>perfect<strong>graphs</strong>and<br />

foundanon-triviallowerboundfortheasymptotic<strong>guessing</strong><strong>number</strong><strong>of</strong>C5. In<br />

fact,amuchbetterboundcanbederivedforbothC5andothernon-perfect<br />

<strong>graphs</strong>bymakinguse<strong>of</strong>afractionalcoverings.Beforedefiningfractionalcoveringsweintroducethet-foldblowup<strong>of</strong><strong>graphs</strong>andtheblow-upstrategy.<br />

Definition4.1.GivenagraphGwedefinethet-foldblowup<strong>of</strong>G,denotedby<br />

G t ,asfollows:ForeachvertexvinGtherearetverticesv1,...,vtinG t with<br />

verticesviandujbeingneighboursinG t ifandonlyifvanduareneighbours<br />

inG.<br />

Laterwewillalsoneedthefollowingspecialcase<strong>of</strong>thestronggraphproduct.<br />

Definition4.2.<strong>The</strong>strongproduct<strong>of</strong>KtandG,denotedKt∗Gisthegraph<br />

obtainedfromG t byaddingalledges<strong>of</strong>theform(vi,vj)foreachvertexvinG.<br />

Wecanusethet-blowup<strong>of</strong>Gtoobtainbetterboundsforthe<strong>guessing</strong><br />

<strong>number</strong><strong>of</strong>Gwithrespecttovalues<strong>of</strong>swhichareperfecttpowers.<br />

Definition4.3.LetGbeagraph,lettbeapositiveintegerandlets=s t 1<br />

forsomeintegers1>1. GivenanystrategyF forG t withrespecttos t 1 ,we<br />

canobtainastrategy,whichwecalltheblow-upstrategy,forGwithrespect<br />

tosasfollows: Ifxisthe<strong>number</strong>assignedtoavertexv<strong>of</strong>G,wewritex=<br />

t i−1<br />

i=1 (xi−1)s1 ,wherexi∈{1,...,s1}foreachi.<strong>The</strong>nweassignthe<strong>number</strong>s<br />

x1,...,xttotheverticesv1,...,vt<strong>of</strong>G t .WethenfollowthestrategyFonG t .<br />

Ify1,...,ybarethe<strong>number</strong>sguessedbyFatverticesv1,...,vb,thenthe<strong>number</strong><br />

guessedbytheblow-upstrategyatvertexvwillbey= t<br />

i=1<br />

(yi−1)s i−1<br />

1 .<br />

Wenotethatasimilarblow-upstrategycanbeusedinthecasethatsisa<br />

product<strong>of</strong>t(notnecessarilyequal)integersstrictlygreaterthan1.Wediscuss<br />

howthiscanbedoneinsubsection7.1.<br />

<strong>The</strong>orem4.4.LetGbeagraphandlett,s1bepositiveintegers.<strong>The</strong>ngn(G,s)<br />

gn(G,s1).<br />

Pro<strong>of</strong>.ItisimmediatethatwecanpartitionG tintotvertexdisjointcopies<strong>of</strong>G. LetFbethestrategyonG twithrespecttos1whichfollowsthebestpossible strategyoneach<strong>of</strong>thistcopies<strong>of</strong>G. Wenowfollowtheblow-upstrategy.<br />

<strong>The</strong>winningprobabilityisatleasts −t(n−gn(G,s1))<br />

1 =s−(n−gn(G,s1)) andtheresult<br />

follows.<br />

7


WecouldhaveinfactprovedthistheoremusingLemma3.5instead. We<br />

willseeamorepowerfulapplication<strong>of</strong>theblow-upstrategyafterweintroduce<br />

fractionalcoverings<strong>of</strong><strong>graphs</strong>.<br />

Definition4.5. Afractionalcliquecover<strong>of</strong>agraphGisafamily<strong>of</strong>cliques<br />

H1,...,Ht<strong>of</strong>Gtogetherwithnon-negativeweightsw1,...,wtsuchthat <br />

{i:v∈Hi} wi<br />

1forallv ∈ V(G). <strong>The</strong> maximumvalue<strong>of</strong> t i=1wi overallcliquecovers<br />

H1,...,Ht<strong>of</strong>Gisknownasthefractionalcliquecover<strong>number</strong><strong>of</strong>Gandisdenotedbycp<br />

f(G). Itisalsoknownasthefractionalchromatic<strong>number</strong><strong>of</strong>the<br />

complementG<strong>of</strong>G,denotedbyχf(G).<br />

Itiswellknownthatcp f(G)isalwaysarational<strong>number</strong>andthatα(G)<br />

cp f(G)cp(G). Similarly,wehavethatω(G)χf(G)χ(G). One<strong>of</strong>the<br />

basicresultsinfractionalgraphtheoryletsusrelatethefractionalchromatic<br />

<strong>number</strong>tothechromatic<strong>number</strong><strong>of</strong>asuitableblowup<strong>of</strong>theoriginalgraph.<br />

Fortheseandotherresultsonfractionalgraphtheorywereferthereaderto<br />

[SU97].<br />

<strong>The</strong>orem4.6.ForeachgraphGthereexistsapositiveintegert,withsuchthat<br />

χf(G)= χ(Kt∗G)<br />

t .Equivalently,consideringthecomplement<strong>of</strong>Kt∗G,wehave<br />

cp f(G)= cp(Gt )<br />

t<br />

Observethatforanypositiveintegertandanyintegermultiples<strong>of</strong>twe<br />

havethatχ(Kt∗G)/tχ(Ks∗G)/sχf(G). Inparticular,iftisaninteger<br />

forwhichtheconsequence<strong>of</strong><strong>The</strong>orem4.6holds, thenitalsoholdsforany<br />

integermultiple<strong>of</strong>t. Notethatconsideringthecomplement<strong>of</strong>G,thereisan<br />

integert(notnecessarilythesameastheonegivenby<strong>The</strong>orem4.6)suchthat<br />

cp(G t )=tcpf(G).Sowecanuseacombination<strong>of</strong>thecliquecoverstrategyand<br />

theblow-upstrategytoobtainbetterstrategiesforthe<strong>guessing</strong>game.<br />

Definition4.7.LetGbeanygraphandlettbeanypositiveintegersuchthat<br />

cp(G t )=tcpf(G). <strong>The</strong>nthefractionalcliquestrategyisdefinedbytakingthe<br />

blow-upstrategy<strong>of</strong>Gwithrespecttothecliquecoverstrategy<strong>of</strong>G t .<br />

Example4.8. ConsiderthefractionalcliquestrategyforC5. <strong>The</strong>fractional<br />

chromatic<strong>number</strong><strong>of</strong>C5 is 5<br />

2 andafractionalcliquestrategyforC5 canbe<br />

obtainedbyusingthecliquecoverstrategyonthe2-foldblowupC 2 5 <strong>of</strong>C5.This<br />

graphisdepictedinFigure1,withacliquecoveringgivenbythethickedges.<br />

<strong>The</strong>orem4.9.LettbeanypositiveintegerasinDefinition4.7.<strong>The</strong>ngn(G,s)<br />

n−cp f(G)wheneversisaperfectt-power.Inparticular,gn(G)n−cp f(G).<br />

Pro<strong>of</strong>.Weapplythefractionalcliquestrategy.Wehavecp(G t )=tcp f(G)and<br />

so,if|V(G)|=nands=s t 1 ,thenthewinningprobabilityis<br />

s −(tn−cp(Gt ))<br />

1<br />

=s −(tn−tcp f (G))<br />

1<br />

=s −(n−cp f (G)) .<br />

<strong>The</strong>resultfollows. <strong>The</strong>resultfortheasymptotic<strong>guessing</strong><strong>number</strong>als<strong>of</strong>ollows<br />

immediatelysincegn(G,s)n−cp f(G)holdsforinfinitelymanys.<br />

8


Figure1:C 2 5 withacliquecovershownasthickedges.<br />

Forthefive-cycle,andothersymmetrical<strong>graphs</strong>,wecanapplythefollowing<br />

wellknownlemma(Proposition3.1.1in[SU97]).<br />

Lemma4.10.IfGisavertextransitivegraphonnverticesthenχf(G)= n<br />

α(G)<br />

andcp f(G)= n<br />

ω(G) .<br />

Twosimpleclasses<strong>of</strong>vertextransitive<strong>graphs</strong>aretheoddcyclesandtheir<br />

complements.<br />

Example4.11.ForC5,Lemma4.10improvesourpreviousboundstogn(C5,s)<br />

5<br />

2 foreveryswhichisaperfectsquare. Thisisexactlyhalfwaybetweenthe<br />

twosimpleboundsgivenbyα(C5)andχ(C5). Fortheoddcyclesingeneral,<br />

wefindthatgn(C2k+1,s) 2k+1<br />

2 foreveryswhichisaperfectsquare,andfor<br />

theircomplementswefindgn(C2k+1,s)(2k+1)− 2k+1<br />

k<br />

=2k−1− 1<br />

k for<br />

everyswhichisaperfectk-power.Hence,bythestrongperfectgraphtheorem<br />

[CRST06],wehavefoundimprovedboundsfortheclass<strong>of</strong>minimallyimperfect<br />

<strong>graphs</strong>.<br />

Fortheoddcycles,theimprovementgivenbyconsideringcp f(G)instead<strong>of</strong><br />

cp(G)isbounded.Howevertherearefamilies<strong>of</strong><strong>graphs</strong>wheretheimprovement<br />

canbearbitrarilylarge.<br />

Example4.12.Givenpositiveintegersn,rwithn>2r,theKnesergraphG=<br />

Kn:rhasthefamily<strong>of</strong>allr-subsets<strong>of</strong>{1,...,n}asitsverticeswithtwovertices<br />

beingneighboursifandonlyifthecorrespondingsetsaredisjoint.<strong>The</strong>se<strong>graphs</strong><br />

areclearlyvertextransitiveon n<br />

r vertices.Itisimmediatethatω(G)=⌊n/r⌋.<br />

<strong>The</strong>Erdős-Ko-Radotheoremimpliesthatα(G)= n−1 r−1 andsobyLemma4.10<br />

wegetthatχf(G)=n/r. Sothe<strong>guessing</strong><strong>number</strong><strong>of</strong>Gisgn(G,s)= n n − r ,<br />

foreveryswhichisaprefectrthpower.FortheKneser<strong>graphs</strong>takingther-fold<br />

blowup<strong>of</strong>Kn:rwillgiveaworkingcliquecoverstrategy,seeChapter3<strong>of</strong>[SU97]<br />

forfulldetails<strong>of</strong>theblowups<strong>of</strong>Kneser<strong>graphs</strong>..<br />

9<br />

r


On the otherhand, by Lovász’solutionto Kneser’sconjecture, we have<br />

χ(G)=n−2r+2andsothesimplecliquecoverstrategyonlygivesgn(G,s)<br />

<br />

−(n−2r+2).Soifn=tr,thenwegetanimprovement<strong>of</strong>(t−2)(r−1).<br />

n<br />

r<br />

<strong>The</strong>family<strong>of</strong>Kneser<strong>graphs</strong>hassomeadditionalimportanceinconnection<br />

withfractionalcoverduetothefactthatthefractionalchromatic<strong>number</strong><strong>of</strong><br />

agraphcanbecompletelydescribedinterms<strong>of</strong>whichKneser<strong>graphs</strong>Ghas<br />

homomorphismsinto,seeChapter3<strong>of</strong>[SU97].<br />

Other<strong>graphs</strong>withlargegapsbetweentheboundscanbefoundbye.g.using<br />

theMycielskiconstructionorthestronggraphproduct,underwhichfractional<br />

chromatic<strong>number</strong>saremultiplicative.See[SU97]formoredetailsontheseand<br />

otherconstructionsfor<strong>graphs</strong>withprescribedfractionalchromatic<strong>number</strong>s.<br />

5 <strong>The</strong>codegraph<br />

Intheprevioussectionwehaveseenthatthe<strong>guessing</strong><strong>number</strong>canbedescribed<br />

interms<strong>of</strong>the<strong>number</strong><strong>of</strong>fixedpoints<strong>of</strong>thestrategymaps,andsincethisisa<br />

finiteset<strong>of</strong>mappings<strong>of</strong>afinitesetitisinprinciplepossiblet<strong>of</strong>indtheoptimal<br />

strategyforagivenGandsbyanexhaustivesearch.Wealsosawhowtoprovide<br />

goodlowerboundsforthe<strong>guessing</strong><strong>number</strong>interms<strong>of</strong>χf(G),butwedidnot<br />

findmatchingupperboundsingeneral. Ournextaimistointroducethecode<br />

graphforthe<strong>guessing</strong>gamewithagivenGandsand,usingthisgraph,both<br />

provideamoreefficientcomputationalprocedureforfindingoptimalstrategies<br />

andaconjectureastowhattheasymptotic<strong>guessing</strong><strong>number</strong>gn(G)shouldbe.<br />

LetΩ(n,s)denotetheset<strong>of</strong>allstrings<strong>of</strong>lengthnoverthealphabet{1,...,s}.<br />

Given a graph G on n vertices we will usually identify Ω(n,s) with As =<br />

{1,...,s} V(G) . Ourultimateaimistodeterminewhichsubsets<strong>of</strong>Ω(n,s)are<br />

fixedpointsets<strong>of</strong>somestrategymapFonG.<br />

Definition5.1.<strong>The</strong>codegraphX(G,s)hasvertexsetΩ(n,s)withtwovertices<br />

xandy<strong>of</strong>X(G,s)beingadjacentifandonlyifthereisavertexv<strong>of</strong>Gsuch<br />

thatxv=yvbutxu=yuforallu∈N(v)<br />

Asthenexttheoremshows,thecodegraphcompletelydescribestheset<strong>of</strong><br />

optimalstrategiesforthe<strong>guessing</strong>game.<br />

<strong>The</strong>orem5.2.LetIbeaset<strong>of</strong>vertices<strong>of</strong>X(G,s). <strong>The</strong>nIistheset<strong>of</strong>fixed<br />

points<strong>of</strong>anoptimalstrategymappingFifandonlyifitisamaximalindependent<br />

set.<br />

Pro<strong>of</strong>.LetFbeastrategymapping(notnecessarilyoptimal)havingIasitsset<br />

<strong>of</strong>fixedpoints. Letx,y∈IandsupposethattheyareneighboursinX(G,s).<br />

<strong>The</strong>nthereexistatleastonevertexv<strong>of</strong>Gsuchthatxv=yvbutxu=yuforall<br />

u∈N(v).Butx,yarefixedpoints<strong>of</strong>Fandsoxvandyvcanbedeterminedby<br />

thevalues<strong>of</strong>xuandyuwithu∈N(v).Thusxv=yv,acontradiction.<br />

Conversely,givenamaximumindependentsetI,wedefineamappingF:<br />

As→Asasfollows:Foreachvertexv<strong>of</strong>Gandeachvertexx<strong>of</strong>X(G,s),ifthere<br />

10


isanx ′ ∈Isuchthatx ′ u=xuforeachu∈N(V)thenwedefinefv(x)=x ′ v.<br />

Otherwise,wedefinefv(x)=1. NotethatsinceIisanindependentsetthen<br />

eachfviswell-definedandthemappingF={fv} v∈V(G)isastrategymapping.<br />

WeclaimthatF hasIasitstheset<strong>of</strong>fixedpoints. Indeed,byconstruction<br />

everyelement<strong>of</strong>Iisafixedpoint<strong>of</strong>F.Moreover,sincetheset<strong>of</strong>fixedpoints<br />

<strong>of</strong>F isanindependentsetcontainingthemaximumindependentsetI,then<br />

itmustbeequaltoI. ItnowfollowsthatF isoptimalasnootherstrategy<br />

mappingcanhavemorefixedpoints.<br />

Itfollowsthatwecancomputethe<strong>guessing</strong><strong>number</strong><strong>of</strong>Gbycomputingthe<br />

independence<strong>number</strong><strong>of</strong>X(G,s)<br />

Corollary5.3.ForeverypositiveintegersandeverygraphGwehavethat<br />

gn(G,s)=log s(α(X(G,s))).<br />

Oursimpleboundsforthe<strong>guessing</strong><strong>number</strong>caneasilybetranslatedinto<br />

properties<strong>of</strong>thecodegraph.<br />

<strong>The</strong>orem5.4.<br />

1. Ifα(G)=t,thenα(X(G,s))s n−t .<br />

2. Ifχ(G)=k,thenα(X(G,s))s k .<br />

Pro<strong>of</strong>.<br />

1. LetIbeanindependentset<strong>of</strong>sizeα(G)=tinG. Givenx,y∈Ω(n,s)<br />

wesaythattheyareequivalentifandonlyifxu=yuforeveryu /∈A.<br />

Observethatthisisanequivalencerelationwitheachequivalenceclass<br />

beingacliqueinX(G,s)andhavingsizeexactlys t . Everyindependent<br />

set<strong>of</strong>X(G,s)cancontainatmostoneelementfromeverysuchcliqueand<br />

thereforeitcanhavesizeatmosts n−t .<br />

2. LetH1,...,Hkbeacliquecover<strong>of</strong>G<strong>of</strong>sizekandletIbetheset<strong>of</strong>all<br />

elements<strong>of</strong>x∈X(G,s)suchthat <br />

v∈Hi xv≡0modsforeach1ik.<br />

<strong>The</strong>nitiseasilyseenthatIisanindependentset<strong>of</strong>X(G,s)<strong>of</strong>sizes k .<br />

<strong>The</strong>boundfrom<strong>The</strong>orem4.9canalsobetranslatedintothissettingby<br />

consideringthecodegraph<strong>of</strong>G t ,forasuitablevalue<strong>of</strong>t.<br />

<strong>The</strong>codegraphhass n vertices. Givenanypair<strong>of</strong>strings<strong>of</strong>lengthnwe<br />

candecideintimen 2 iftheyareadjacentinthecodegraphornot. Havingconstructedthecodegraphwecannowusestandardalgorithmsforfindingmaximumindependentsetst<strong>of</strong>indboththe<strong>guessing</strong><strong>number</strong>andoptimal<strong>guessing</strong>strategies.Usinge.g.themaximumindependentsetalgorithmfrom[TT77]<br />

wegetthefollowingupperboundonthecomplexity<strong>of</strong>determiningtheoptimal<br />

strategy.<br />

11


Corollary5.5.<strong>The</strong>reexistsanalgorithmwhichconstructstheoptimal<strong>guessing</strong><br />

strategyintimeatmost2 1<br />

3sn <strong>The</strong>upperboundhereis<strong>of</strong>courseprohibitivelylarge,howeveritcanbe<br />

reducedbyusingthelargeautomorphismgroup<strong>of</strong>X(G,s),e.g.usingmethods<br />

similartothose<strong>of</strong>[RAMS04].<br />

Lemma5.6.X(G,s)isvertextransitiveandAut(G)×Z n s⊆Aut(X(G,s))<br />

Pro<strong>of</strong>.Letφbeanautomorphism<strong>of</strong>Gandlety∈Z n s.<strong>The</strong>ntheautomorphism<br />

(φ,y)<strong>of</strong>Aut(X(G,s))isdefinedby<br />

(φ,y)(x)v=(x+y) φ(v)<br />

wheretheelements<strong>of</strong>Ω(n,s)areviewedasvectorsinZ n swithadditiondefined<br />

componentwise.Itisstraightforwardtoverifythatthisisindeedanautomorphism<strong>of</strong>X(G,s)andthat(φ,y)◦(φ<br />

′ ,y ′ )=(φ◦φ ′ ,y+y ′ )asrequired.<br />

6 Upperboundsviaentropyinequalities<br />

Ournextaimistoshowthatourboundonthe<strong>guessing</strong><strong>number</strong>foroddcycles<br />

andtheircomplementsgiveninExample4.12issharp. Asatoolwewilluse<br />

entropyinequalities<strong>of</strong>discreterandomvariable. Recallthatgivenadiscrete<br />

randomvariableXwithoutcomeslabelled1,...,N,itsbinaryentropyisdefined<br />

as<br />

N<br />

H(X)= Pr(X=i)log 2Pr(X=i),<br />

i=1<br />

Itwillbemoreconvenientforwhatfollowstoworkwiththes-entropyinstead<br />

H(X)=<br />

N<br />

Pr(X=i)log sPr(X=i).<br />

i=1<br />

Fromnowon,wewilldropthesubscriptsandfollowtheconventionthatallour<br />

logarithmsarewithbases.<strong>The</strong>conditionalentropy<strong>of</strong>Xgivenanotherrandom<br />

variableY istheentropy<br />

H(X|Y)= <br />

Pr(X=i,Y=j)logPr(X=i|Y=j).<br />

i,j<br />

Wewillusethefollowingproperties<strong>of</strong>theentropyfunction.<br />

<strong>The</strong>orem6.1.LetX,Y,X1,...,Xnbediscreterandomvariables<br />

1. H(X)0withequalityifandonlyifXisdetermenistic.<br />

2. IfXtakesvaluesinasetS,thenH(X)log|S|withequalityifandonly<br />

ifXisuniformlydistributedonS.<br />

12


3. H(X,Y)=H(X|Y)+H(Y).<br />

4. H(X|Y)0withequalityifandonlyifXisdeterminedbyY.<br />

5. H(X|Y)H(X)withequalityifandonlyifXandY areindependent.<br />

6. ForA⊆{1,...,n},letXAdenotestherandomvector<strong>of</strong>Xi’swithi∈A.<br />

Withthisnotation,Hisasubmodularfunction,i.e.<br />

H(XA∪B)+H(XA∩B)H(XA)+H(XB).<br />

Fortheseandotherproperties<strong>of</strong>theentropyfunctionwereferthereader<br />

to[CT06,AS08]. Recently,entropyinequalitieshavebeenusedusedtoderive<br />

boundsforotherproblemsincombinatorics,see[Rad03]forasurvey.<br />

<strong>The</strong>orem6.2.Foreachk,s∈Nwehavethatgn(C2k+1,s) 2k+1<br />

2 .<br />

Pro<strong>of</strong>.Assumethatthevertices<strong>of</strong>C2k+1are<strong>number</strong>edas1,...,2k+1with<br />

N(i)={i−1,i+1}whereadditionandsubtractionaredonemodulo2k+1.LetI<br />

beamaximumindependentset<strong>of</strong>X(C2k+1,s)andletX=(X1,X2,...,X2k+1)<br />

bethecharacteristicvector<strong>of</strong>anelementpickeduniformlyatrandomfromI.<br />

Byproperty(2),wehavethatH(X)=log|I|. Wewillnowproceedtobound<br />

H(X)fromabove. Whenestimatingentropieswewillneedtoconsidermany<br />

quantities<strong>of</strong>theformH(XB)andinordertosimplifyournotationwewillwrite<br />

themasH(B)instead.<br />

SinceXisafixedpoint<strong>of</strong>astrategymapping,weknowthatXiisdetermined<br />

byXi−1andXi+1.Inparticular,Xisdeterminedjustbytherandomvariables<br />

X2,X3,X5,X7,...,X2k+1andsobyproperties(3)and(4)wehavethat<br />

H(X)=H(X|2,3,5,...,2k+1)+H(2,3,5,...,2k+1)<br />

=H(2,3,5,...,2k+1)<br />

Applyingnowproperties(3)and(5)wegetthat<br />

andbyproperty(2)itfollowsthat<br />

Likewise,wehave<br />

H(X)H(2,3)+H(5,7,...,2k+1),<br />

H(X)H(2,3)+log(s k−1 )=H(2,3)+k−1.<br />

H(X)H(1,2,3,4)+H(6,8,...,2k)H(1,2,3,4)+k−2.<br />

Addingtheseinequalitiesandusingproperty(6)wegetthat<br />

2H(X)H(2,3)+H(1,2,3,4)+2k−3<br />

H(1,2,3)+H(2,3,4)+2k−3.<br />

13


ButsinceX2isdeterminedbyX1andX3wehavethatH(1,2,3)=H(1,3)<br />

log(s 2 )=2andsimilarlyH(2,3,4)2.Puttingallthesetogetherweget<br />

log|I|=H(X) 2k+1<br />

2<br />

andsobyLemmas5.2and5.3wegetthatgn(C2k+1,s) 2k+1<br />

2<br />

Usingalargerfamily<strong>of</strong>indexsetswecansimilarlyfindaboundforthe<br />

<strong>guessing</strong><strong>number</strong><strong>of</strong>thecomplement<strong>of</strong>anoddcycle.<br />

<strong>The</strong>orem6.3.Foreachk,s∈Nwehavethatgn(C2k+1,s)2k−1− 1<br />

k .<br />

Pro<strong>of</strong>.Assumethatthevertices<strong>of</strong>C2k+1are<strong>number</strong>edas1,...,2k+1with<br />

N(i)=[2k+1]\{i−1,i+1}whereadditionandsubtractionaredonemodulo<br />

2k+1. LetI beamaximumindependentset<strong>of</strong>X(C2k+1,s)andletX =<br />

(X1,X2,...,X2k+1)bethecharacteristicvector<strong>of</strong>anelementpickeduniformly<br />

atrandomfromI. Byproperty(2)wehavethatH(X)=log|I|. Wewillnow<br />

proceedtoboundH(X)fromabove.<br />

Foreachi∈[2k+1],letJi=[2k+1]\{i−1,i+1}andKi=[2k+1]\{i}.<br />

ObservethatXiisuniquelydeterminedfromthevalues<strong>of</strong>allXjwithj=iand<br />

soH(X)=H(Ki).ObservealsothatforanyS⊆Jiwehavethat<br />

H(X)=H(Ki−1)H(Ji)+H(S∪{i+1})−H(S)2k−2+H(S∪{i+1})−H(S).<br />

Indeed, thefirstinequalityfollowsfromproperty(6)bytakingA = Ji and<br />

B=S∪{i+1},andthesecondinequalityfollowsfromproperty(2).Similarly,<br />

wehave<br />

H(X)2k−2+H(S∪{i−1})−H(S).<br />

LetT={3,5,...,2k+1}.ApplyingthesecondinequalitywithS=Tandi=2<br />

weget<br />

H(X)2k−2+H(T∪{1})−H(T)<br />

ApplyingthefirstinequalityrepeatedlywithS=T∪{1},T∪{1,2},T∪{1,2,4},...,T∪<br />

{1,2,4,...,2k−6}andi=1,3,5,...,2k−5respectivelyweget<br />

H(X)2k−2+H(T∪{1,2})−H(T∪{1})<br />

H(X)2k−2+H(T∪{1,2,4})−H(T∪{1,2})<br />

···<br />

H(X)2k−2+H(T∪{1,2,4,...,2k−4})−H(T∪{1,2,4,...,2k−6})<br />

andsummingallthesetogetherwiththepreviousinequalityupweget<br />

Sincealso<br />

(k−1)H(X)2(k−1) 2 +H(J2k−1)−H(T)2k(k−1)−H(T).<br />

H(X)=H(3,4,...,2k+1)H(T)+H(4,6,...,2k)H(T)+(k−1),<br />

14


weget<br />

andtheresultfollows.<br />

kH(X)(2k+1)(k−1)<br />

HerewehavefoundthatifGisaminimallyimperfectgraphthenits<strong>guessing</strong><br />

<strong>number</strong>isatleastgn(G,s)n−χf(G)foreveryvalue<strong>of</strong>s. Wealsoknow<br />

fromtheprevioussectionthatequalityholdsforinfinitelymanyvalues<strong>of</strong>s.In<br />

particularitholdsforeveryswhichisaperfectb-powerforsomebwhichcan<br />

bedeterminedfromG.WealsoknowfromSection2thatgn(G,s)n−χf(G)<br />

foreveryperfectgraphG,whereinfactweevenhaveequalityforeveryvalue<strong>of</strong><br />

s.Weconjecturethatthisinequalityistrueingeneral.<br />

Conjecture6.4.ForeverygraphGandeverypositiveintegerswehavegn(G,s)<br />

n−χf(G).Inparticulargn(G,s)=n−χf(G)foreveryvalue<strong>of</strong>sforwhichthe<br />

fractionalcliquestrategycanbeusedandsogn(G)=n−χf(G).<br />

Itwouldalsobe<strong>of</strong>interestt<strong>of</strong>indthe<strong>guessing</strong><strong>number</strong>forvaluesswhere<br />

thefractionalcliquestrategycannotbeused. EvenforC5weknowonlythe<br />

valuefors=3,andwecanobtainsomeboundsforothernon-perfectsquares<br />

viaLemmas3.3and3.5.<br />

Forvalues<strong>of</strong>sotherthanperfectsquareswedonotevenhaveafullunderstanding<strong>of</strong>the<strong>guessing</strong><strong>number</strong>sforoddcycles.<br />

Example6.5.Fors=2wehavecomputedtheindependence<strong>number</strong>s<strong>of</strong>the<br />

code<strong>graphs</strong>forsmalloddcyclesusingastandardlinearprogrammingsolver.<br />

Wefoundthatα(X(C7,2))=8,α(X(C9,2))=16andα(X(C11,2))=32,i.e.<br />

exactlythevaluesattainedbythecliquestrategyonthese<strong>graphs</strong>. ForC7,we<br />

als<strong>of</strong>oundthatα(x(C7,3))=29>3 3 ,soherethecliquestrategyisnotoptimal.<br />

Weclosethissectionwithtwoproblems.<br />

Problem6.6.Isthecliquestrategyoptimalforgn(C2k+1,2)forallk3?<br />

Problem6.7.Isthecliquestrategyoptimalforgn(C2k+1,3)forallk4?<br />

7 Generalisations<br />

7.1 Nonuniformrandom<strong>number</strong>s<br />

Intheresultss<strong>of</strong>arwehaveassumedthatallrandom<strong>number</strong>sxvwerepicked<br />

uniformlyatrandomfromthesingleset{1,...,s}. Howevertherearetwoextensionswhichcouldnaturallyappear.<br />

<strong>The</strong>firstistoassumethattherandom<br />

<strong>number</strong>sareindependentasbeforebuteachxvcomesfromsomenon-uniform<br />

distributionPv,whichmaydependonv.<strong>The</strong>secondistoallowthedifferentxv<br />

tocomefromsetsSvwhichmayvaryinsizewithv.<strong>The</strong>codegraphapproach<br />

fromtheprevioussectioncaneasilybeadaptedtoboth<strong>of</strong>thesemodifications.<br />

15


InordertodealwithdistinctPvwecangiveeachstringinΩ(n,s)aweight<br />

vPv(xv)anduseastrategygivenbyamaximumweightindependentset.With thisweightingthemaximumweightwillcorrespondtothemaximumwinning<br />

probability.<br />

LikewisethecodegraphcaneasilybeadaptedtodifferingSvbyreplacingthe<br />

uniformΩ(n,s)withtheproductset <br />

vSv,usingthesameneighbourrelation asbefore.<br />

7.2 Directed<strong>graphs</strong><br />

In[Rii07]theemphasiswasondirected<strong>graphs</strong>,asthisiscaseusedinnetwork<br />

codingandbooleancircuits. Whileourdiscussionhasmainlybeeninterms<strong>of</strong><br />

<strong>undirected</strong><strong>graphs</strong>thereareseveralmethodsandboundswhichcanbegeneralizedtodirected<strong>graphs</strong>aswell.Wewillheregiveonlyafewexamples.<br />

<strong>The</strong>codegraphX(G,s)canbedefinedforadirected<strong>graphs</strong>byreplacingthe<br />

neighbourhood<strong>of</strong>vusedininDefinition5.1withtheset<strong>of</strong>out-neighbours<strong>of</strong><br />

v,i.e.theset<strong>of</strong>verticestowhichthereisadirectededgefromv.Asbeforethe<br />

<strong>guessing</strong><strong>number</strong>isgiveninterms<strong>of</strong>themaximumindependentsets<strong>of</strong>X(D,s).<br />

Recallthatadirectedgraphisacyclic ifitisnotpossibletowalkalong<br />

acyclebyfollowingthedirections<strong>of</strong>itsedges. GivenadirectedgraphDlet<br />

α(D)denotethesize<strong>of</strong>thelargestinducedsubgraph<strong>of</strong>Dwhichisacyclic.<br />

Thisparametercanbeusedinplace<strong>of</strong>theindependence<strong>number</strong>toboundthe<br />

<strong>guessing</strong><strong>number</strong><strong>of</strong>D.<br />

<strong>The</strong>orem7.1.LetDbeadigraphonnvertices.<strong>The</strong>ngn(D,s)n−α(D)<br />

Pro<strong>of</strong>.LetAbeaninducedacyclicsubgraph<strong>of</strong>G<strong>of</strong>sizeα(D). LetA0bethe<br />

set<strong>of</strong>sinksinA,i.e.thoseverticesinAwhoseonlyout-neighboursareinD\A,<br />

andrecursivelydefineAi tobetheset<strong>of</strong>verticesfromA\Ai−1whoseonly<br />

out-neighboursareinAi−1.<br />

Asinthepro<strong>of</strong><strong>of</strong>Lemma2.7givenastrategyF,weexposetherandom<br />

<strong>number</strong>sxuinstages. Atthefirststageweexposeall<strong>number</strong>sxuwithu/∈A.<br />

<strong>The</strong>nallfunctionsfuwithu∈A0areuniquelydetermined.Wenowexposeall<br />

<strong>number</strong>sxuwithu∈A0andweobservethattheprobabilitythatallplayers<br />

inA0guesscorrectlyiss −|A0| . Nowallfunctionsfuwithu∈A1areuniquely<br />

determinedandtheprobabilitythatallplayersinA1guesscorrectlyiss −|A1| .<br />

Proceedinginductivelyweconcludethatthewinningprobabilityisatmosts −|A|<br />

andtheresultfollows.<br />

Asinthecliquestrategywemayconstructstrategiesforlargerdirected<br />

<strong>graphs</strong>bypartitioningtheirvertexsetintodisjointcopies<strong>of</strong>smaller<strong>graphs</strong>,<br />

followingtheoptimalstrategiesindependentlyoneachpart<strong>of</strong>thepartition.<br />

Howeverunlikeforordinary<strong>graphs</strong>, wherethecliquesplayedapivotalrole,<br />

theredoesnotseemtobeacanonicalfamily<strong>of</strong>directed<strong>graphs</strong>topartitionthe<br />

graphinto.<strong>The</strong>mostnaturalanalogue<strong>of</strong>thecomplete<strong>graphs</strong>aretournaments,<br />

whicharedirectedcomplete<strong>graphs</strong>,butsincetherearetournamentswhichare<br />

16


acyclic,someadditionalconditionsseemtoberequired. Givenanyfamily<strong>of</strong><br />

directed<strong>graphs</strong>withknownvalues<strong>of</strong>their<strong>guessing</strong><strong>number</strong>swemayalsouse<br />

fractionalcoversusingthese<strong>graphs</strong>inthesamewayaswedidin<strong>The</strong>orem4.9.<br />

7.3 Infinite<strong>graphs</strong><br />

Itisalsonaturaltoaskwhathappensifweplaythe<strong>guessing</strong>gameoninfinite<br />

<strong>graphs</strong>.Webeginbyconsiderthecountablyinfinitecompletegraphwiths=2,<br />

say.Asacomparison,recallthatthewinningprobabilityonanyfinitecomplete<br />

graphwiths=2is1/2.<br />

Considertheset{0,1} N <strong>of</strong>countablyinfinite0,1-sequences. Wewouldlike<br />

t<strong>of</strong>inda‘large’subsetA<strong>of</strong>thissetsuchthatanytwosequencesinAdifferinat<br />

leasttwoplaces.HavingfoundsuchanA,wecanthendefineanaturalstrategy<br />

FwhichhasAasitsset<strong>of</strong>fixedpoints. Itisnotdifficulttopartition{0,1} N<br />

intotwosetsAandBsuchthatboth<strong>of</strong>themhavethepropertythatanytwo<br />

<strong>of</strong>theirsequencesdiffernatleasttwoplaces: Onejustdefinesanequivalence<br />

relation∼on{0,1} N bysayingthattwosequencesx,yareequivalentifandonly<br />

iftheydifferinafinite<strong>number</strong><strong>of</strong>places. ForeveryequivalenceclassC<strong>of</strong>∼<br />

wethenpickasequencex∈CanddefineAxtobetheset<strong>of</strong>allsequencesin<br />

Cwhichdifferfromxinanodd<strong>number</strong><strong>of</strong>placesandBxtobetheset<strong>of</strong>all<br />

sequencesinCwhichdifferfromxinaneven<strong>number</strong><strong>of</strong>places. Itiseasyto<br />

checkthatA=∪xAxandB=∪xBxhavetherequiredproperties.<br />

Thisisgreatasitseemstoshowthatthewinningprobabilityforthisgraph<br />

isalso1/2. Howeverthereisonecatch. <strong>The</strong>setsA,Babovewereconstructed<br />

usingtheaxiom<strong>of</strong>choiceandinfactitisnottoodifficulttoseethattheyare<br />

notmeasurable. Forourpurposeshowever,wewanttocalculatethewinning<br />

probabilityandsowemustrequirethateachindividualstrategyfv:{0,1} N →<br />

{0,1}ismeasurable. Sothequestionnowtranslatest<strong>of</strong>indingameasurable<br />

subsetA<strong>of</strong>{0,1} N suchthatanytwosequences<strong>of</strong>Adifferinatleasttwoplaces<br />

andAhasaslargeameasureaspossible. <strong>The</strong>nforeachk∈N,thestrategy<br />

fk:{0,1} N →{0,1}definedbyfk(x)=0ifthesequencex ′ obtainedfromxby<br />

makingitsk-thdigit0belongstoAandfk(x)=1otherwise,isameasurable<br />

strategy.Moreovertheset<strong>of</strong>fixedpointsA ′ <strong>of</strong>thestrategymappingF=(fk)k∈N<br />

isameasurablesubset<strong>of</strong>{0,1} N containingAandsoithasthesamemeasure<br />

asA.<br />

WeproceedtoshowthatanysuchAmusthavemeasure0. Observethat<br />

thisisimmediateifAisanopensubset<strong>of</strong>{0,1} N (undertheproducttopology).<br />

IndeedthesetsUx,n={y:yi=xi∀in}formabasisforthetopologyand<br />

clearlynosuchsetcanbeasubset<strong>of</strong>Aasx,x ′ ∈Ux,nwherex ′ isobtainedfrom<br />

xbychangingits(n+1)-thdigit. IfthereissuchasetA<strong>of</strong>positivemeasure,<br />

thenwecanfindabasicopensetUx,nsuchthatm(A∩Ux,n)51m(Ux,n)/100.<br />

NowletA0betheset<strong>of</strong>allsequencesinA∩Ux,nwhose(n+1)-thdigitis0and<br />

letB0betheset<strong>of</strong>allsequencesobtainedfromA0bychangingthe(n+1)-th<br />

digitto1. WedefineA1andB1analogously. NowobservethatA0,A1,B0,B1<br />

aredisjointsubsets<strong>of</strong>Ux,nwithm(A0)+m(A1)51m(Ux,n)/100andm(A0)=<br />

17


m(B0),m(A1)=m(B1). Itfollowsthatm(A0∪A1∪B0∪B1)m(Ux,n),a<br />

contradiction.<br />

Soeveninthecompletecountablyinfinitegraphitturnsoutthatthewinning<br />

probabilityforthe<strong>guessing</strong>gameis0.<br />

Acknowledgments<br />

<strong>The</strong>materialinthispaperwasfirstpresentedattheNewtonInstituteworkshop<br />

“CombinatorialandProbabilisticInequalities”in2008andthesecondauthor<br />

wouldliketothanktheinstituteforitshospitalityduringhisstaythere.<br />

References<br />

[ACLY00] RudolfAhlswede,NingCai,Shuo-YenRobertLi,andRaymondW.<br />

Yeung. Networkinformationflow. IEEETrans.Inform.<strong>The</strong>ory,<br />

46(4):1204–1216,2000.<br />

[AS08] NogaAlonandJoelH.Spencer. <strong>The</strong>probabilisticmethod. Wiley-<br />

InterscienceSeriesinDiscreteMathematicsandOptimization.John<br />

Wiley&SonsInc.,Hoboken,NJ,thirdedition,2008. Withanappendixonthelifeandwork<strong>of</strong>PaulErdős.<br />

[Ber63] C.Berge. Perfect<strong>graphs</strong>. InSixPapersonGraph<strong>The</strong>ory.Indian<br />

StatisticalInstitute,1963.<br />

[CRST06] Maria Chudnovsky, Neil Robertson, Paul Seymour, and Robin<br />

Thomas. <strong>The</strong>strongperfectgraphtheorem. Ann.<strong>of</strong>Math.(2),<br />

164(1):51–229,2006.<br />

[CT06] ThomasM.CoverandJoyA.Thomas. Elements<strong>of</strong>information<br />

theory.Wiley-Interscience[JohnWiley&Sons],Hoboken,NJ,second<br />

edition,2006.<br />

[Lov72] L.Lovász. Acharacterization<strong>of</strong>perfect<strong>graphs</strong>. J.Combinatorial<br />

<strong>The</strong>orySer.B,13:95–98,1972.<br />

[Rad03] J.Radhakrishnan. ComputationalMathematics,ModellingandAlgorithms,chapterEntropyandcounting.<br />

NarosaPublishingHouse,<br />

2003.<br />

[RAMS04]ArathiRamani,FadiA.Aloul,IgorL.Markov,andKaremA.Sakallah.Breakinginstance-independentsymmetriesinexactgraphcoloring.InProceedings<strong>of</strong>theconferenceonDesign,automationandtest<br />

inEurope-Volume1,DATE’04,pages10324–,Washington,DC,<br />

USA,2004.IEEEComputerSociety.<br />

18


[Rii07] SørenRiis. Informationflows,<strong>graphs</strong>andtheir<strong>guessing</strong><strong>number</strong>s.<br />

Electron.J.Combin.,14(1):ResearchPaper44,17,2007.<br />

[SU97] EdwardR.ScheinermanandDanielH.Ullman. Fractionalgraph<br />

theory. Wiley-InterscienceSeriesinDiscreteMathematicsandOptimization.JohnWiley&SonsInc.,NewYork,1997.<br />

Arational<br />

approachtothetheory<strong>of</strong><strong>graphs</strong>,WithaforewordbyClaudeBerge,<br />

AWiley-IntersciencePublication.<br />

[TT77] RobertEndreTarjanandAnthonyE.Trojanowski.Findingamaximumindependentset.SIAMJ.Comput.,6(3):537–546,1977.<br />

DemetresChrist<strong>of</strong>ides,Institutefor<strong>The</strong>oreticalComputerScience,Faculty<strong>of</strong>Mathematicsand<br />

Physics,MalostranskéNáměstí25,18800Prague,CzechRepublic.demetres@kam.mff.cuni.cz<br />

KlasMarkström,Department<strong>of</strong>MathematicsandMathematicalStatistics,Ume˚aUniversity,<br />

90187Ume˚a,Sweden.klas.markstrom@math.umu.se<br />

19

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!