Tight Inefficiency Bounds for Perception-Parameterized Affine Congestion Games PieterKleer1andGuidoScha¨fer1,2 1 CentrumWiskunde&Informatica(CWI),NetworksandOptimizationGroup 2 VrijeUniversiteitAmsterdam,DepartmentofEconometricsandOperationsResearch Amsterdam,TheNetherlands [email protected], [email protected] Abstract. Congestion games constitute an important class of non-cooperative games which wasintroducedbyRosenthalin1973.Inrecentyears,severalextensionsofthesegameswere proposedtoincorporateaspectsthatarenotcapturedbythestandardmodel.Examplesofsuch 7 extensionsincludetheincorporationofrisksensitiveplayers,themodelingofaltruisticplayer 1 behaviorandtheimpositionoftaxesontheresources.Theseextensionswerestudiedintensively 0 withthegoaltoobtainapreciseunderstandingoftheinefficiencyofequilibriaofthesegames. 2 In this paper, we introduce a new model of congestion games that captures these extensions n (andadditionalones)inaunifyingway.Thekeyideahereistoparameterizeboththeperceived a costofeachplayerandthesocialcostfunctionofthesystemdesigner.Intuitively,eachplayer J perceivestheloadinducedbytheotherplayersbyanextentofρ≥0,whilethesystemdesigner 6 estimates that each player perceives the load of all others by an extent of σ ≥0. The above 2 mentionedextensionsreducetospecialcasesofourmodelbychoosingtheparametersρandσ accordingly.Asinmostrelatedworks,weconcentrateoncongestiongameswithaffinelatency ] T functionshere. G Despite the fact that we deal with a more general class of congestion games, we manage to derivetightboundsonthepriceofanarchyandthepriceofstabilityforalargerangeofpa- . s rameters. Our bounds provide a complete picture of the inefficiency of equilibria for these c perception-parameterizedcongestiongames.Asaresult,weobtaintightboundsontheprice [ ofanarchyandthepriceofstabilityfortheabovementionedextensions.Ourresultsalsoreveal 1 howoneshould“design”thecostfunctionsoftheplayersinordertoreducethepriceofanar- v chy.Somewhatcounterintuitively,ifeachplayercaresaboutallotherplayerstotheextentof 4 0.625only(insteadof1inthestandardsetting)thepriceofanarchyreducesfrom2.5to2.155 1 andthisisbestpossible. 6 7 0 . 1 0 7 1 : v i X r a 1 Introduction Congestion games constitute an important class of non-cooperative games which was introduced by Rosenthal in 1973 [13]. In a congestion game, we are given a set of resource from which a set ofplayerscanchoose.Eachresourceisassociatedwithacostfunctionwhichspecifiesthecostof this resource depending on the total number of players using it. Every player chooses a subset of resources(fromasetofresourcesubsetsavailabletoher)andexperiencesacostequaltothesumof thecostsofthechosenresources.Congestiongamesareboththeoreticallyappealingandpractically relevant.Forexample,theyhaveapplicationsinnetworkrouting,resourceallocationandscheduling problems. Rosenthal[13]provedthateverycongestiongamehasapureNashequilibrium,i.e.,astrategy profile such that no player can decrease her cost by unilaterally deviating to another feasible set of resources. This result was established through the use of an exact potential function (known as Rosenthal potential) satisfying that the cost difference induced by a unilateral player deviation is equal to the potential difference of the respective strategy profiles. In fact, Monderer and Shapley [11]showedthattheclassofgamesadmittinganexactpotentialfunctionisisomorphictotheclass ofcongestiongames. One of the main research directions in algorithmic game theory focusses on quantifying the inefficiencycausedbyselfishbehavior.TheideaistoassessthequalityofaNashequilibriumrelative toanoptimaloutcome.Herethequalityofanoutcomeismeasuredintermsofagivensocialcost objective(e.g.,thesumofthecostsofallplayers).KoutsoupiasandPapadimitriou[10]introduced thepriceofanarchyastheratiobetweentheworstsocialcostofaNashequilibriumandthesocial costofanoptimum.Anshelevichetal.[1]definedthepriceofstabilityastheratiobetweenthebest socialcostofaNashequilibriumandthesocialcostofanoptimum. Inrecentyears,severalextensionsofRosenthal’scongestiongameswereproposedtoincorporate aspects which are not captured by the standard model. For example, these extensions include risk sensitivityofplayersinuncertainsettings[12],altruisticplayerbehavior[4,5]andcongestiongames withtaxes[3,14].WeelaborateinmoredetailontheseextensionsinSection2.Thesegameswere studiedintensivelywiththegoaltoobtainapreciseunderstandingofthepriceofanarchy. In this paper, we introduce a new model of congestion games, which we term perception- parameterized congestion games, that captures all these extensions (and more) in a unifying way. Thekeyideahereistoparameterizeboththeperceivedcostofeachplayerandthesocialcostfunc- tion.Intuitively,eachplayerperceivestheloadinducedbytheotherplayersbyanextentofρ ≥0, while the system designer estimates that each player perceives the load of all others by an extent of σ ≥0. The above mentioned extensions reduce to special cases of our model by choosing the parametersρ andσ accordingly. Despite the fact that we deal with a more general class of congestion games, we manage to derive tight bounds on the price of anarchy and the price of stability for a large range of parame- ters. Our bounds provide a complete picture of the inefficiency of equilibria for these perception- parameterizedcongestiongames.Asaconsequence,weobtaintightboundsonthepriceofanarchy andthepriceofstabilityfortheabovementionedextensions.Whilethepriceofanarchyboundsare (mostly)knownfrompreviousresults,thepriceofstabilityresultsarenew.Asin[3,4,5,12,14],we focushereoncongestiongameswithaffinecostfunctions.3 We illustrate our model by means of a simple example; formal definitions of our perception- parameterizedcongestiongamesaregiveninSection2.Supposewearegivenasetofmresources andthateveryplayerhastochoosepreciselyoneoftheseresources.Thecostofaresourcee∈[m]4 is given by a cost function c that maps the load on e to a real value. In the classical setting, the e loadofaresourceeisdefinedasthetotalnumberofplayersx usinge.Thatis,thecostthatplayer e i experiences when choosing resource e is c (x ). In contrast, in our setting players have different e e perceptionsoftheloadinducedbytheotherplayers.Moreprecisely,theperceivedload ofplayeri choosingresourceeis1+ρ(x −1),whereρ ≥0isaparameter.Consequently,theperceivedcost e ofplayeriforchoosingeisc (1+ρ(x −1)).Notethatasρ increasesplayerscaremoreaboutthe e e 3Conceptually,ourmodelcaneasilybeadaptedtomoregenerallatencyfunctions;onlythederivationofthe respectiveboundsisanalyticallymuchmoreinvolved. 4Givenapositiveintegerm,weuse[m]torefertotheset{1,...,m}. 1 Model Parameters PoA Reference Classical ρ=σ=1 5/2 [6] Altruism(1) σ=1,1≤ρ≤2 (4ρ+1)/(1+ρ) [4,5] Altruism(2) σ=1,2≤ρ≤∞ ρ+1 [5] Uncertainty:riskneutral-players σ=ρ=1/2 5/3 [12] Uncertainty:Wald’sminimaxprinciple σ=1/2,ρ=1 2 [2,12] Constantuniversaltaxes σ=1,ρ=h(1). 2.155 [3] Generalizedaffinecongestiongames − ∞ [∗] Model Parameters PoS Reference Classical ρ=σ=1 1.577 [3] √ √ Altruism(1) σ=1,1≤ρ≤2 ( 3+1)/( 3+ρ−1) [∗] Uncertainty:riskneutral-players σ=ρ=1/2 1.447 [∗] Uncertainty:Wald’sminimaxprinciple σ=1/2,ρ=1 1 [∗] Constantuniversaltaxes σ=1,ρ=h(1) 2.013 [∗] Generalizedaffinecongestiongames − 2 [∗] Table1.Anoverviewof(tight)priceofanarchyandpriceofstabilityresultsforcertainvaluesofρandσ.Here h(1)≈0.625(seeTheorem1foraformaldefinition).Anasteriskindicatesthatthisresultisnew. presence of other players.5 In addition, we introduce a similar parameter σ ≥0 for the social cost objective.Intuitively,thiscanbeseenasthesystemdesigner’sestimateofhoweachplayerperceives theloadoftheotherplayers.Inourexample,thesocialcostisdefinedas∑e∈[m]ce(1+σ(xe−1))xe. OurResults. InSection3,weproveaboundof (cid:26) (cid:27) 2ρ(1+σ)+1 max ρ+1, (1) ρ+1 onthepriceofanarchyofaffinecongestiongamesforalargerangeorparameters(ρ,σ)(seealso Figure1foranillustration).Forgeneralaffinecongestiongamesweprovethatthisboundistight. Further,evenforthespecialcaseofsymmetricnetworkcongestiongamesweshowthatthebound (2ρ(1+σ)+1)/(ρ+1)isasymptoticallytight(ontherangeforwhichitisattained). InSection4,wegiveaboundof (cid:112) σ(σ+2)+σ (2) (cid:112) σ(σ+2)+ρ−σ on the price of stability of affine congestion games for a large range or pairs (ρ,σ) (see below for details). For general affine congestion games we show that this bound is tight. For symmetric networkcongestiongameswegiveabetter(tight)boundonthepriceofstabilityforthecaseσ =1 andρ ≥0arbitrary(detailsaregiveninSection5). An overview of the price of anarchy and the price of stability results that we obtain from (1) and(2)forseveralapplicationsknownintheliteratureisgiveninTable1.Therespectivereferences wheretheseboundswereestablishedfirstaregivenintherightmostcolumn.Theconnectionbetween theseapplicationsandourmodelisdiscussedindetailattheendofSection2. Inlightoftheabovebounds,weobtainan(almost)completepictureoftheinefficiencyofequi- libria(parameterizedbyρandσ);forexample,seeFigure3foranillustrationofthepriceofanarchy ifσ =1.Notethatthepriceofanarchydecreasesfrom 5 forρ=1to2.155forρ=h(1)≈0.625.6 2 WereferthereadertoSection5forfurtherremarksanddiscussionsoftheresults. 5In general, the parameter ρ can be player- or resource-specific, but in this work we concentrate on the homogenouscase(i.e.,allplayershavethesameperceptionparameter). 6Thepriceofanarchyforρ=h(1)wasfirstestablishedbyCaragiannisetal.[3].However,ourboundsreveal thatthepriceofanarchyisinfactminimizedatρ=h(1)(seealsoFigure3). 2 2 OurModel,ApplicationsandRelatedWork Wefirstformallyintroduceourmodelofcongestiongameswithparameterizedperceptions.Wethen show that our model subsumes several other models that were studied in the literature as special cases. CongestionGames. AcongestiongameΓ isgivenbyatuple(N,E,(S) ,(c ) )whereN=[n] i i∈N e e∈E isthesetofplayers,E thesetofresources(orfacilities),S ⊆2E thesetofstrategiesofplayeri,and i c :R →R the cost function of facility e. Given a strategy profile s=(s ,...,s )∈×S, we e ≥0 ≥0 1 n i i definex asthenumberofplayersusingresourcee,i.e.,x =x (s)=|{i∈N:e∈s}|.IfS =S for e e e i i j alli,j∈N,thegameiscalledsymmetric.ForagivengraphG=(V,E),wecallΓ a(directed)network congestiongameifforeveryplayerithereexists,t ∈V suchthatS isthesetofall(directed)(s,t)- i i i i i pathsinG.Anaffinecongestiongamehascostfunctionsoftheformc (x)=a x+b witha ,b ≥0. e e e e e Ifb =0foralle∈E,thegameiscalledlinear. e OurModel. Weintroduceourunifyingmodelofperception-parameterizedcongestiongameswith affinelatencyfunctions.Forafixedparameterρ ≥0,wedefinethecostofplayeri∈N by Cρ(s)= ∑c (1+ρ(x −1))=a [1+ρ(x −1)]+b (3) i e e e e e e∈si foragivenstrategyprofiles=(s ,...,s ).Forafixedparameterσ ≥0,thesocialcostofastrategy 1 n profilesisgivenby Cσ(s)=∑Cσ(s)= ∑x (a [1+σ(x −1)]+b ). (4) i e e e e i e∈E Werefertothecaseρ=σ =1astheclassicalcongestiongamewithcostfunctionsc (x)=a x+b e e e foralle∈E. InefficiencyofEquilibria. AstrategyprofilesisaNashequilibriumifforallplayersi∈N itholds thatCρ(s)≤Cρ(s(cid:48),s )foralls(cid:48)∈S,where(s(cid:48),s )denotesthestrategyprofileinwhichplayeri i i i −i i i i −i playss(cid:48)andalltheotherplayerstheirstrategyins.Thepriceofanarchy(PoA)andpriceofstability i (PoS)ofagameΓ aredefinedas max Cσ(s) min Cσ(s) s∈NE s∈NE PoA(Γ,ρ,σ)= and PoS(Γ,ρ,σ)= , mins∗∈×iSiCσ(s∗) mins∗∈×iSiCσ(s∗) where NE = NE(ρ) denotes the set of Nash equilibria with respect to the player costs as de- fined in (3). For a collection of games H we define PoA(H,ρ,σ) = sup PoA(Γ,ρ,σ) and Γ∈H PoS(H,ρ,σ)=sup PoS(Γ,ρ,σ). Rosenthal [13] shows that classical congestion games (i.e., Γ∈H ρ =σ =1)haveanexactpotentialfunction:Φ :×S →Risanexactpotentialfunctionforacon- i i gestiongameΓ ifforeverystrategyprofiles,foreveryi∈Nandeverys(cid:48)∈S: Φ(s)−Φ(s ,s(cid:48))= i i −i i Ci(s)−Ci(s−i,s(cid:48)i).TheRosenthalpotentialΦ(s)=∑e∈E∑xke=1ce(k)isanexactpotentialfunctionfor classicalcongestiongames. Applications. Wereviewvariousmodelsthatfallwithin,orarerelatedto,theframeworkproposed above(forcertainvaluesofρ andσ).Thesemodelssometimesinterprettheparametersdifferently thanexplainedabove. Altruism[4,5].Wecanrewritethecostofplayerias Cρ(s)= ∑(a x +b )+(ρ−1)a (x −1). i e e e e e e∈si The term (ρ−1)a (x −1) can be interpreted as a “dynamic” (meaning load-dependent) tax that e e players using resource e have to pay.7 For 1≤ρ ≤∞ and σ =1, this model is equivalent to the altruisticplayersettingproposedbyCaragiannisetal.[4].Chenetal.[5]alsostudythismodelof 7Infact,forρ=2thiscorrespondstothedynamictaxesasproposedinatechnicalreportbySingh[14]. 3 altruismfor1≤ρ ≤2andσ =1.(Theequivalencebetweenthealtruisticmodelandourmodelis provedinLemma3intheappendix.) Constanttaxes[3].Wecanrewritethecostofplayerias Cρ(s)= ∑ρa x +(1−ρ)a +b . i e e e e e∈si NowastrategyprofilesisaNashequilibriumifforeveryplayeriandeverys(cid:48)∈S i i ∑ρa x +(1−ρ)a +b ≤ ∑ ρa x +(1−ρ)a +b + ∑ ρa (x +1)+(1−ρ)a +b . e e e e e e e e e e e e e∈si e∈si∩s(cid:48)i e∈s(cid:48)i\si Dividingbyρ givesthatsisalsoaNashequilibriumforthecostfunctions b 1−ρ Tρ(s)= ∑a x + e +∑ a . i e e ρ ρ e e∈si e∈si That is, s is a Nash equilibrium in a classical congestion game in which players take into account constant resource taxes of the form (1−ρ)/ρ·a . Caragiannis, Kaklamanis and Kanellopoulos e [3] study this type of taxes, which they call universal tax functions, for ρ satisfying (1−ρ)/ρ = √ 3/2 3−2. They consider these taxes to be refundable, i.e., they are not taken into account in the socialcost,whichisequivalenttothecaseσ =1.Notethatthefunctionτ :(0,1]→[0,∞)defined byτ(ρ)=(1−ρ)/ρ isbijective.8 Risksensitivityunderuncertainty[12].Nikolova,PiliourasandShamma[12]considerconges- tiongamesinwhichthereisa(non-deterministic)orderoftheplayersoneveryresource.Aplayer isonlyaffectedbyplayersinfrontofher.Thatis,theloadonresourceeforplayeriinastrictorder- ingr,wherer (i)denotesthepositionofplayeri,isgivenbyx (i)=|{j∈N:r (j)≤r (i)}|.The e e e e costofplayeriisthenCi(s)=∑e∈sice(xe(i)).Notethatxe(i)isarandomvariableiftheorderingis non-deterministic.Thesocialcostofthemodelisdefinedbythesumofallplayercosts, C12(s)= ∑aexe(xe+1)+be 2 e∈E which is independent of the ordering r. (This holds because in every ordering there is always one playerfirst,oneplayersecond,andsoon.)Notethatthesocialcostcorrespondstothecaseσ = 1 2 in our framework. Nikolova et al. [12] study various risk attitudes towards the ordering r that is assumed to have a uniform distribution over all possible orderings. The two relevant attitudes are that of risk-neutral players and players applying Wald’s minimax principle. Risk-neutral players define their cost as the expected cost under the ordering r, which correspond to the case ρ = 1 in 2 (3).Thiscanroughlybeinterpretedasthatplayersexpecttobescheduledinthemiddleonaverage. Wald’s minimax principle implies that players assume a worst-case scenario, i.e., being scheduled lastonalltheresources.Thiscorrespondstothecaseρ =1. ApproximateNashequilibria[7].SupposethatsisaNashequilibriumunderthecostfunctions definedin(3).Then,inparticular,wehave C1(s)≤Cρ(s)≤Cρ(s(cid:48),s )≤ρC1(s(cid:48),s ) i i i i −i i i −i foranyplayeriands(cid:48)∈S andρ ≥1.Thatis,wehaveC1(s)≤ρ·C1(s(cid:48),s )whichmeansthatthe i i i i i −i strategyprofilesisaρ-approximateequilibrium,asstudiedbyChristodoulou,KoutsoupiasandSpi- rakis[7].Inparticular,thisimpliesthatanyupperboundonthepriceofanarchy,orpriceofstability, inourframeworkyieldsanupperboundonthepriceofstabilityforρ-approximateequilibriaforthe sameclassofgames. Generalized affine congestion games. Let A(cid:48) denote the class of all congestion games Γ for whichallresourceshavethesamecostfunctionc(x)=ax+b,wherea=a(Γ)andb=b(Γ)satisfy a≥0anda+b>0.Theclassofaffinecongestiongameswithnon-negativecoefficientsiscontained 8Thisrelationbetweenaltruism(orspite)andconstanttaxesisalsomentionedbyCaragiannisetal.[4]. 4 ρ ρ=2σ ρ+1 2ρ(1+σ)+1 ρ=σ ρ+1 ρ=h(σ) ? 0 1 1 σ 2 Fig.1.Theboundρ+1holdsforρ ≥2σ ≥1.Thebound(2ρ(1+σ)+1)/(1+ρ)holdsforσ ≤ρ ≤2σ. Roughly speaking, this bound also holds for h(σ)≤ρ ≤σ, but our proof of Theorem 1 only works for a discretizedrangeofσ (hencetheverticaldottedlinesinthisarea).ThefunctionhisgiveninTheorem1. inA(cid:48) sinceeverysuchgamecanalwaysbetransformed9 intoagameΓ(cid:48) witha =1andb =0for e e all resources e∈E(cid:48), where E(cid:48) is the resource set ofΓ(cid:48). Without loss of generality we can assume that a+b=1, since the cost functions can be scaled by 1/(a+b). The cost functions ofΓ ∈A(cid:48) canthenequivalentlybewrittenasc(x)=ρx+(1−ρ)forρ ≥0.Thisispreciselythedefinitionof Cρ(s)(witha =1andb =0takenthere).Inparticular,ifwetakeσ =ρ,meaningthatCρ(s)= i e e ∑ Cρ(s),wehave i∈N i PoA(A(cid:48))=sup PoA(A,ρ,ρ) and PoS(A(cid:48))=sup PoS(A,ρ,ρ), ρ≥0 ρ≥0 whereAdenotestheclassofaffinecongestiongameswithnon-negativecoefficients. Duetopagelimitationssomematerialisomittedfromthemaintextbelowandcanbefoundinthe appendix. 3 PriceofAnarchy Inthissectionwederiveanupperboundonthepriceofanarchyforaffinecongestiongamesof (cid:26) (cid:27) 2ρ(1+σ)+1 max ρ+1, ρ+1 forawiderangeofpairs(ρ,σ)thatcapturesallknown(tous)priceofanarchyresultsintheliter- aturethatfallwithinourmodel.Weshowthatthisboundistightforgeneralcongestiongames.We alsoshowthatthebound(2ρ(1+σ)+1)/(ρ+1)is(asymptotically)tightforsymmetricnetwork congestiongames,fortherangeof(ρ,σ)onwhichitisattained. WeneedthefollowingtechnicalresultfortheproofofthemainresultinTheorem1: Lemma1. LetsbeaNashequilibriumunderthecostfunctionsCρ(s)andlets∗ beaminimizerof i Cσ(·).Forρ,σ ≥0fixed,ifthereexistα(ρ,σ),β(ρ,σ)≥0suchthat (1+ρ·x)y−ρ(x−1)x−x≤−β(ρ,σ)(1+σ(x−1))x+α(ρ,σ)(1+σ(y−1))y forallnon-negativeintegersxandy,thenβ(ρ,σ)Cσ(s)≤α(ρ,σ)Cσ(s∗). Theorem1. LetsbeaNashequilibriumunderthecostfunctionsCρ(s)andlets∗beaminimizerof i Cσ(·).Then Cσ(s) 2ρ(1+σ)+1 ≤ (5) Cσ(s∗) ρ+1 9ThistransformationcanbedoneinsuchawaythatbothPoAandPoSofthegamedonotchange.Foraproof thereaderisreferredto,e.g.,[5,Lemma4.3]. 5 i) if 1 ≤σ ≤ρ ≤2σ, 2 ii) ifσ =1andh(σ)≤ρ ≤2σ, (cid:112) whereh(σ)=g(1+σ+ σ(σ+2),σ)istheoptimumofthefunction σ(a2−1) g(a,σ)= . (1+σ)a2−(2σ+1)a+2σ(σ+1) Furthermore,thereexistsafunction∆ =∆(σ)(specifiedintheappendix)satisfyingforeveryfixed σ ≥1/2:if∆(σ )≥0,then(5)istrueforallh(σ )≤ρ ≤2σ . 0 0 0 0 Proof (Sketch).Forthefunctionsα(ρ,σ)=(2ρ(1+σ)+1)/(1+2σ)andβ(ρ,σ)=(1+ρ)/(1+ 2σ),weprovetheinequalityinLemma1.Weshowthat,forcertainfunctions f and f ,thesmallest 1 2 ρ satisfyingtheinequalityofLemma1isgivenbythequantity f (x,y,σ) h(σ)= sup − 2 . x,y∈N:f1(x,y,σ)>0 f1(x,y,σ) Wedividetheset(x,y)∈N×Ninlinesoftheformx=ayanddeterminethesupremumoverevery line. After that we take the supremum over all lines, which then gives the desired result. We first showthatthecasex≤ywhichistrivial.Wethenfocusony>x.Inthiscasewewanttodetermine f (ay,y,σ) h(σ)= sup sup− 2 . a∈R>1y>1 f1(ay,y,σ) We show that h(σ)=max{γ (σ),γ (σ)} for certain functions γ and γ . Numerical experiments 1 2 1 2 suggestthat∆(σ):=γ (σ)−γ (σ)≥0,thatis,themaximumisalwaysattainedforγ (whichisthe 1 2 1 definitionofhgiveninthestatementofthetheorem).Inparticular,thismeansthatif,forafixedσ, thenon-negativityof∆(σ)ischecked,thentheproofindeedyieldstheinequalityofLemma1for h(σ)≤ρ ≤2σ. (cid:116)(cid:117) Numerical experiments suggest that ∆(σ) is non-negative for all σ ≥1/2. We emphasize that forafixedσ,with∆(σ)≥0,theproofthattheinequalityholdsforallh(σ)≤ρ ≤2σ isexactin theparameterρ.ThefirsttwocasesofTheorem1captureallthepriceofanarchyresultsfromthe literature. WenextshowthattheboundofTheorem1isalsoan(asymptotic)lowerboundforlinearsym- metricnetworkcongestiongames.10Thisimprovesaresultintherisk-uncertaintymodelofPiliouras etal.[12],whoonlyproveasymptotictightnessforsymmetriclinearcongestiongames(fortheirre- spectivevaluesofρ andσ).ItalsoimprovesaresultinthealtruismmodelbyChenetal.[5],who showtightnessonlyforgeneralcongestiongames.Theproofisageneralizationoftheconstruction of Correa et al. [8], who proved that for the classical case ρ =σ =1, the Price of Anarchy upper boundof5/2,shownin[6],isasymptoticallytightforsymmetricnetworkcongestiongames. Theorem2. Forρ,σ>0fixed,andplayerswithcostfunctionsCρ(s),thereexistsymmetricnetwork i linearcongestiongamessuchthat Cσ(s) 2ρ(1+σ)+1 ≥ −ε Cσ(s∗) ρ+1 foranyε >0,wheresisaNashequilibrium,ands∗asociallyoptimalstrategyprofile. Proof. Weconstructasymmetricnetworklinearcongestiongamewithnplayers.Wefirstdescribe thegraphtopologyusedintheproofofTheorem5in[8](usingsimilarnotationandterminology). ThegraphGconsistsofnprincipaldisjointpaths,calledP,...,P fromstot (horizontalpathsin 1 n Figure2),eachconsisting2n−1arcs(andhence2nnodes).Withe the j-tharconpathiisdenoted i,j fori=1,...,nand j=1,...,2n−1.Also,v denotesthe j-thnodeonpathifori=1,...,nand i,j j=1,...,2n.Therearealson(n−1)connectingarcs:foreverypathithereisanarc(v ,v ) i,2k+1 i−1,2k 6 ρx x ρx x ρx x ρx (1+ρ)x (1+ρ)x v w 1 1 s t v 2 w 2 0 1 2 3 4 5 6 7 8 Fig.2.Illustrationoftheinstanceforn=5.Thedashed(blue)pathindicatesthestrategyofplayer2intheNash equilibrium.Foreveryprincipalpath,thefirstandlastarchavecost(1+ρ)x,andinbetweenthecostsalternate betweenρxandx(startingandendingwithρx).Thediagonalconnectingarcshavecostzero.Thenumbersat thebottomindicatethelayers.Thebold(red)subpathsindicatethetwodeviationsituationsthatareanalyzedto provethatsisindeedaNashequilibrium. fork=1,...,n−1,wherei−1istakenmodulon(thediagonalarcsinFigure2).For j≥1fixed, wesaythatthearcse fori=1,...,nformthe(j−1)-thlayerofG(seeFigure2). i,j The cost functions are as follows. All arcs leaving s (the arcs e for i=1,...,n) and all arcs i,1 enteringt (thearcse fori=1,...,n)havelatencyc (x)=(1+ρ)x.Foralli=1,...,n,thearcs i,2n e e fork=1,...,n−1havecostfunctionc (x)=ρx,whereasthearcse fork=1,...,n−2 i,2k−1 e i,2k havecostfunctionc (x)=x.Allotherarcs(thediagonalconnectingarcs)havecostzero. e ThefeasiblestrategyprofiletinwhichplayeriusesprincipalpathP,foralli=1,...,nhassocial i costCσ(t)=n(2(1+ρ)+(n−1)ρ+(n−2))=n((1+ρ)n+ρ).ANashequilibriumisgivenby the strategy profile in which every player k uses the following path: she starts with arcs e and k,1 e ,thenusesallarcsoftheforme ,e ,e for j=1,...,n−1,andendswitharcs k,2 k+j,2j k+j,2j+1 k+j,2j+2 e ,e (andusesallconnectingarcsinbetween).11Notethatallthe(principal)arcs k+n−1,2n−2 k+n−1,2n−1 oflayer j haveload1is j iseven,andload2if j isodd.Thesocialcostofthisprofileisgivenby Cσ(s)=n(2(1+ρ)+(n−1)·2·ρ(1+σ(2−1))+n−2)=n((1+2ρ(1+σ))n−2ρσ).Itfollows that Cσ(s)/Cσ(t)↑(1+2ρ(1+σ))/(1+ρ) as n→∞. We now show that the above mentioned strategyprofilesisindeedaNashequilibrium. Fixsomeplayer,sayplayer2,asinFigure2,andsupposethatthisplayerdeviatestosomepath Q. Let j be the first layer in which P and Q overlap. Note that j must be odd. The costCρ(s) of 2 2 player2,onthesubpathofP leadingtothefirstoverlappingarcwithQ,isatmost 2 j−1 j−1 (1+ρ)+ ·[2·[ρ(1+ρ(2−1))]+1]+ρ(1+ρ(2−1))=(1+ρ)2+ (1+2ρ(1+ρ)) 2 2 ThesubpathofQleadingtothefirstoverlappingarcwithP hasCρ(Q,s )asfollows.Sheusesat 2 i −i leastonearcineveryoddlayer(beforetheoverlappinglayer)withaloadof3andonearcofevery evenlayer(beforetheoverlappingarc)withload2,meaningthatthecostofplayerionthesubpath ofQisatleast j−1 j−1 (1+ρ)(1+ρ(2−1))+ ·[(ρ(1+ρ(3−1)))+(1+ρ(2−1))]=(1+ρ)2+ ·(2ρ+1)(1+ρ) 2 2 Since1+2ρ(1+ρ)<(2ρ+1)(1+ρ)forallρ≥0,iffollowsthatthecostofplayerionthesubpath ofP isnoworsethanthatofthesubpathofQ,whenplayer2deviatesfromP toQ.Iffollowsthat 2 2 itsufficestoshowthatP isanequilibriumstrategyinswithrespecttodeviationsQthatoverlapon 2 the first arc e with P. A similar argument shows that it also suffices to look at deviations Q for 2,1 2 whichQandP overlaponthelastarce ofP. 2 2,2n−1 2 10Intheappendix(Theorem8)weshow(non-asymptotic)tightnessforgeneralcongestiongames. 11ThisissimilartotheconstructioninTheorem5[8]. 7 NowsupposethatP andQdonotoverlaponsomeinternalpartofP.Notethatthefirstarcof 2 2 Q that is not contained in P, say (v ,w ) must be in an even layer, and also that the last arc, say 2 1 1 (v ,w )(whichisaconnectingarc)isinanoddlayer(notethatv (cid:54)=sandw (cid:54)=tw.l.o.g.bywhatis 2 2 1 2 saidinthepreviousparagraph).ItisnothardtoseethatthesubpathofQfromv tow containsthe 1 2 same number of even-layered arcs as the subpath of P, and the same number of odd-layered arcs 2 asthesubpathofP.However,theloadonalltheodd-layeredarcsonthesubpathofdeviationQis 2 3,whereastheloadonodd-layeredarcsinthesubpathofP betweenv andw (instrategys)is2. 2 1 2 Similarly,theloadoneveryeven-layeredarconthesubpathofdeviationQis2,whereastheloadon evereven-layeredarcinthesubpathofP is1.HencethesubpathofdeviationQbetweenv andw 2 1 2 canneverbeprofitable. (cid:116)(cid:117) Forρ≥2σ,wecanobtainatightboundofρ+1onthepriceofanarchy.Remarkably,thebound itselfdoesnotdependonσ,onlytherangeofρ andσ forwhichitholds.Fortheparametersσ =1 andρ ≥2inthealtruismmodelofCaragiannisetal.[4],thisboundisknowntobetightfornon- symmetric singleton congestion games (where all strategies consist of a single resource). We only providetightnessforgeneralcongestiongames,buttheconstructionissignificantlysimpler. Theorem3. LetsbeaNashequilibriumunderthecostfunctionsCρ(s)andlets∗beaminimizerof i Cσ(·).ThenCσ(s)/Cσ(s∗)≤ρ+1for1≤2σ ≤ρ.Furthermore,thisboundistight. 4 PriceofStability Inthissectionwegiveatightboundforthepriceofstabilityof (cid:112) σ(σ+2)+σ (cid:112) σ(σ+2)+ρ−σ foralargerangeofpairs(ρ,σ).Weshowthisboundtobe(asymptotically)tight. Weneedthefollowingtechnicallemma. Lemma2. Forallnon-negativeintegersxandy,andσ ≥0arbitrary,wehave (cid:18) 1(cid:19)2 1 (cid:112) x−y+ − +2σx(x−1)+( σ(σ+2)+σ)[y(y−1)−x(x−1)]≥0. 2 4 Theorem4. LetsbeabestNashequilibriumunderthecostfunctionsCρ(s)andlets∗ beamini- i mizerofCσ(·).Wehave (cid:112) Cσ(s) σ(σ+2)+σ ≤ Cσ(s∗) (cid:112)σ(σ+2)+ρ−σ forallσ >0and 2σ ≤ρ ≤2σ. (cid:112) 1+σ+ σ(σ+2) AnoverviewoftheimplicationsofthisboundisgiveninTable1.Theboundof2forgeneralized affinecongestiongamesrequiressomeadditionalarguments(seeappendix). Proof. OurproofissimilartoatechniqueofChristodoulou,KoutsoupiasandSpirakis[7]forupper bounding theprice of stabilityof ρ-approximate equilibria(we elaborate onthe conceptual differ- ence in Section 5). However, for a general σ the analysis is more involved. The main technical contributioncomesfromestablishingtheinequalityinLemma2. WecanwriteCρ(s)=a x +b +(ρ−1)a x ,which,byRosenthal[13],impliesthat i e e e e e x (x +1) (x −1)x Φρ(s):= ∑a e e +b x +(ρ−1)∑a e e e e e e 2 2 e∈E e∈E 8 isanexactpotentialforCρ(s). i TheideaoftheproofistocombinetheNashinequalities,andthefactthattheglobalminimum ofΦρ(·)isaNashequilibrium(becauseitisanexactpotential).Letsdenotetheglobalminimum ofΦρ,ands∗ asociallyoptimalsolution.Wecanwithoutlossofgeneralityassumethata =1and e b =0.TheNashinequalities(asinthepriceofanarchyanalysis)yield e ∑x (1+ρ(x −1))≤ ∑(1+ρx )x∗ e e e e e∈E e∈E whereasthefactthatsisaglobaloptimumofΦρ(·)yieldsΦρ(s)≤Φρ(s∗),whichreducesto ∑ρx2+(2−ρ)x ≤ ∑ρ(x∗)2+(2−ρ)x∗. e e e e e∈E e∈E Ifwecanfindγ,δ ≥0,andsomeK≥1,forwhich (cid:0)0≤(cid:1) γ(cid:2)ρ(x∗)2+(2−ρ)x∗−ρx2−(2−ρ)x (cid:3)+δ[(1+ρx )x∗−x (1+ρ(x −1)] e e e e e e e e ≤K·x∗[1+σ(x∗−1)]−x [1+σ(x −1)], (6) e e e e thenthisimpliesthatCσ(s)/Cσ(s∗)≤K.Wetake K−1 (ρ−1)K+1 δ = and γ = . ρ 2ρ It is not hard to see that δ ≥0 always holds, however, for γ we have to be more careful. We will later verify for which combinations of ρ and σ the parameter γ is indeed non-negative. Rewriting theexpressionin(6)yieldsthatwehavetofindK satisfying f (x ,x∗,σ) (x∗)2−2x x∗+(1+2σ)x2−x∗+(1−2σ)x K≥ 2 e e := e e e e e e . f (x ,x∗,ρ,σ) [(1−ρ+2σ)(x∗)2−2x x∗+(1+ρ)x2+(ρ−1−2σ)x∗−(ρ−1)x ] 1 e e e e e e e e Notethatthisreasoningisonlycorrectif f (x ,x∗,ρ,σ)≥0.Thisistruesince 1 e e (cid:18) 1(cid:19)2 1 f (x ,x∗,ρ,σ)= x −x∗+ − +(2σ−ρ)x∗(x∗−1)+ρx (x −1) 1 e e e e 2 4 e e e e whichisnon-negativeforallx ,x∗∈N,σ ≥0and0≤ρ ≤2σ.Furthermore,theexpressioniszero e e ifandonlyif(x ,x∗)∈{(0,1),(1,1)},butforthesepairsthenominatorisalsozero,andhence,the e e expressionin(6)isthereforesatisfiedforthosepairs.Wecanwrite (cid:18) 1(cid:19)2 1 f (x ,x∗,σ)= x −x∗+ − +2σx (x −1) 2 e e e e 2 4 e e andtherefore f /f = A ,where 2 1 A+(2σ−ρ)B (cid:18) 1(cid:19)2 1 A= x −x∗+ − +2σx (x −1), B=x∗(x∗−1)−x (x −1). e e 2 4 e e e e e e Notethatifρ =2σ,wehave f /f =1,andhencewecantakeK=1.Otherwise, 2 1 (cid:112) A σ(σ+2)+σ (cid:112) ≤ =:K ⇔ A+( σ(σ+2)+σ)B≥0 (cid:112) A+(2σ−ρ)B σ(σ+2)+ρ−σ TheinequalityontherightistruebyLemma2. Tofinishtheproof,wedeterminethepairs(ρ,σ)forwhichtheparameterγisnon-negative.This holdsifandonlyif (cid:112) σ(σ+2)+σ (ρ−1)K+1=(ρ−1) +1≥0. (cid:112) σ(σ+2)+ρ−σ Rewritingthisyieldstheboundonρ inthestatementofthetheorem. (cid:116)(cid:117) 9