A Equilibria of Greedy Combinatorial Auctions

Size: px
Start display at page:

Download "A Equilibria of Greedy Combinatorial Auctions"

Transcription

1 A Equlbra of Greedy Combnatoral Auctons BRENDAN LUCIER, Mcrosoft Research ALLAN BORODIN, Unversty of Toronto We consder auctons n whch greedy algorthms, pared wth frst-prce or crtcal-prce payment rules, are used to resolve mult-parameter combnatoral allocaton problems. We study the prce of anarchy for socal welfare n such auctons. We show for a varety of equlbrum concepts, ncludng Bayes-Nash equlbrum, low-regret bddng sequences, and asynchronous best-response dynamcs, the resultng prce of anarchy bound s close to the approxmaton factor of the underlyng greedy algorthm. 1. INTRODUCTION The feld of algorthmc mechansm desgn studes systems that depend upon nteracton wth partcpants whose behavour s motvated by ther own goals, rather than those of a desgner. Relevant solutons must therefore merge the computatonal consderatons of computer scence wth the game-theoretc nsghts of economcs. The focus of ths paper s the mult-parameter doman of combnatoral allocaton problems when the goal s to assgn m objects to n agents n order to maxmze the socal welfare, subject to arbtrary downward-closed feasblty constrants. Ths class ncludes all combnatoral aucton problems that allow sngle-mnded declaratons ncludng mult-unt combnatoral auctons, unsplttable flow problems, and many others. For the goal of optmzng socal welfare, the celebrated Vckrey-Clarke-Groves (VCG) mechansm addresses game-theoretc ssues n a strong sense. In the absense of colluson, t nduces full cooperaton (e. truthtellng) as a domnant strategy. However, the VCG mechansm requres that the underlyng welfare-maxmzaton problem be solved exactly. For all but the smplest settngs, ths optmalty requrement s undesrable: exact maxmzaton may be computatonally ntractble, t may requre an unrealstc amount of communcaton from the buyers, and the resultng wnner determnaton rules may be dffcult to explan to a typcal partcpant. One way to bypass these complexty ssues s to desgn new, specally-talored mechansms for specfc assgnment problems. Indeed, there has been sgnfcant progress n desgnng domnant strategy ncentve compatble (DSIC) alternatves to the VCG mechansm. Whle ths venture has been largely successful n settngs where agent preferences are sngle-dmensonal [Archer and Tardos 2001; Brest et al. 2005; Lehmann et al. 1999; Mu alem and Nsan 2008], general settngs have proven more dffcult. It has been shown that the approxmaton ratos acheveable by DSIC mechansms and ther nonncentve compatble counterparts exhbt a large asymptotc gap for some problems [Papadmtrou et al. 2008; Dobznsk 2011; Dughm and Vondrák 2011; Dobznsk and Vondrak 2012]. Author s addresses: B. Lucer, Mcrosoft Research, brlucer@mcrosoft.com; A. Borodn, Department of Computer Scence, Unversty of Toronto, bor@cs.toronto.edu. Permsson to make dgtal or hard copes of part or all of ths work for personal or classroom use s granted wthout fee provded that copes are not made or dstrbuted for proft or commercal advantage and that copes show ths notce on the frst page or ntal screen of a dsplay along wth the full ctaton. Copyrghts for components of ths work owned by others than ACM must be honored. Abstractng wth credt s permtted. To copy otherwse, to republsh, to post on servers, to redstrbute to lsts, or to use any component of ths work n other works requres pror specfc permsson and/or a fee. Permssons may be requested from Publcatons Dept., ACM, Inc., 2 Penn Plaza, Sute 701, New York, NY USA, fax +1 (212) , or permssons@acm.org. c YYYY ACM /YYYY/01-ARTA $15.00 DOI: Journal of the ACM, Vol. V, No. N, Artcle A, Publcaton date: January YYYY.

2 A:2 A. Borodn and B. Lucer Alternatvely, one mght study classes of natural allocaton algorthms, that appear ntutve as aucton allocaton rules, wth the hope that they have desreable ncentve propertes when mplemented as mechansms. As t turns out, for many combnatoral allocaton problems, conceptually smple determnstc algorthms (e.g. greedy algorthms) meet or approach the best-known approxmaton factors subject to computatonal constrants [Lehmann et al. 1999; Mu alem and Nsan 2008; Brest et al. 2005; Babaoff and Blumrosen 2004]. These natural methods tend to be computatonally effcent and easy for bdders to understand, whch are desrable propertes n auctons. Unfortunately, such algorthms are not, n general, DSIC [Lehmann et al. 1999; Borodn and Lucer 2010]. Rather than abandonng these methods n favor of other, potentally more complex, mechansms, we are pursung an alternatve approach. Namely, rather than strvng for domnant strategy truthfulness, t may be acceptable for a system to admt strategc manpulaton, so long as the desgner s objectves are met after such manpulaton occurs. To ths end, we explore the performance of mechansms at equlbra of bdder behavor, gven an approprate model of belefs. Broadly speakng, our motvatng queston s: When can an algorthm be mplemented as a mechansm that acheves hgh socal welfare at every equlbrum? 1 And how robust are the resultng mechansms to varatons of the equlbrum concept? We demonstrate that for combnatoral allocaton problems, any greedy-lke approxmaton algorthm can be converted nto a mechansm that acheves nearly the same approxmaton factor at every equlbrum of bdder behavour. Our analyss s very general, and apples to a range of dfferent equlbrum concepts, ncludng pure and mxed Nash equlbra, Bayes-Nash (correlated) equlbra, no-regret equlbra, and terated myopc best-response. We are thus able to decouple computatonal ssues from ncentves ssues for ths class of algorthms, as one can desgn a greedy algorthm wthout consderng ts economc mplcatons, and then apply a straghtforward prcng scheme n order to acheve good performance at equlbrum. Performance of games at equlbrum has been studed extensvely n the algorthmc game theory lterature as the prce of anarchy (POA) of a gven game 2 : the rato between the optmal outcome and the worst-case outcome at any equlbrum [Papadmtrou 2001]. Put nto these terms, our goal s to convert an algorthm wth approxmaton factor c 1 nto a mechansm whose prce of anarchy s not much larger than c. Ths paper s a synthess and revson of results n [Lucer and Borodn 2010], [Lucer 2010], and results n the frst author s thess [Lucer 2011]. The paper s organzed as follows. The remander of ths secton outlnes our results and relates our work to recent papers n ths area. Secton 2 defnes the necessary concepts and applcatons for our results. Secton 3 ntroduces the concept of strongly loser-ndependence (generalzng the loser-ndependence concept from [Chekur and Gamzu 2009]) whch becomes the key property of greedy algorthms that we wll explot. Sectons 4 and 5 analyze (respectvely) prce of anarchy results for frst-prce and crtcal-prce mechansms. In Sectons 6 and 7 we consder soluton concepts for repeated games: under regret mnmzaton and best-response dynamcs, respectvely. Secton 8 concludes wth some open problems. 1 Domnant strategy truthfulness of an approxmaton mechansm s conceptually stronger as a soluton concept than that of a mechansm that approxmates the optmal socal welfare at every equlbrum. However, as noted elsewhere [Chrstodoulou et al. 2008], Bayesan Nash equlbrum s not, strctly speakng, a relaxaton of domnant strategy truthfulness. There exst truthful mechansms whose approxmaton ratos are not preserved at all Nash equlbra, such as the famous Vckrey aucton. 2 For the purpose of ths paper, we shall not consder cost mnmzaton problems. We note that the prce of anarchy concept was ntroduced n terms of cost mnmzaton games but to the best of our knowledge the only prce of anarchy results for mechansm nduced games apply to maxmzaton problems. Journal of the ACM, Vol. V, No. N, Artcle A, Publcaton date: January YYYY.

3 Equlbra of Greedy Combnatoral Auctons A: Our Results The basc queston of algorthmc mechansm desgn s ths: when can computatonally effcent algorthms be converted nto mechansms that preserve approxmaton bounds wth respect to truthfulness or POA results? We address the prce of anarchy consderatons wth respect to socal welfare maxmzaton for a broad class of allocaton problems. In the full nformaton and Bayesan settng, we study the prce of anarchy for frst prce and crtcal prce mechansms derved from greedy algorthms. Roughly speakng (and n contrast to results regardng approxmaton and truthful mechansms) we are able to show that there s often lttle or no loss from the approxmaton rato of a greedy algorthm to the correspondng mechansm prce of anarchy. We also study the long term behavor of the use of these mechansms when used n repeated games. One-Shot Auctons.. We frst consder one-shot auctons, n whch the allocaton problem s resolved only once. Followng Chrstodoulou et al. [Chrstodoulou et al. 2008], we focus our attenton on the standard (n economcs) ncomplete nformaton settng, where the approprate equlbrum concept s Bayes-Nash equlbrum. That s, we assume that agents preferences are prvate, but drawn ndependently from commonlyknown pror dstrbutons, and that players apply strateges at equlbrum gven ths partal knowledge. We pose the queston: can a gven black-box approxmaton algorthm be converted nto a mechansm that approxmately preserves ts approxmaton rato at every Bayes-Nash Equlbrum? We show that for a broad class of greedy algorthms, the answer s yes. Theorem (nformal): Suppose A s a greedy c-approxmate allocaton rule for a combnatoral allocaton problem. Then the aucton that uses A to choose allocatons, and uses a pay-your-bd payment scheme, has a Bayes-Nash Prce of Anarchy of at most c + O(c 2 /e c ). We also show that the small (and exponentally decreasng) loss n our prce of anarchy bound s necessary, by gvng an example (for every c 2) where the resultng prce of anarchy s at least c + Ω( c e ). We note that the mechansms we 4c consder are all pror-free. Thus, as n the fullnformaton case, whle we assume the exstence of type dstrbutons n order to model ratonal agent behavour, our mechansm need not be aware of these dstrbutons. In the specal case that each player s type dstrbuton s a pont mass, Bayes-Nash equlbrum reduces to standard Nash equlbrum. Our mechansms therefore also preserve approxmaton ratos at every (mxed or pure) Nash equlbrum of the full nformaton game. Our analyss also extends to the more general class of coarse correlated equlbra. For the case of pure Nash equlbrum, our prce of anarchy bound mproves to c. As s standard, our bounds on the Bayesan prce of anarchy wll assume that agent types are dstrbuted ndependently. However, we show that a weaker bound of O(c) holds when agent types are drawn from an arbtrary dstrbuton over the space of all type profles. Ths result apples to greedy algorthms that are non-adaptve, as descrbed n Secton 2.4. Thus, even f agent types are arbtrarly correlated, our mechansms yeld performance at equlbrum asymptotcally matchng that of the underlyng allocaton algorthm. A smlar bound also apples to mechansms that use the crtcal-prce payment scheme, whch s a natural extenson of second-prce payments n sngle-tem auctons. Such a payment scheme charges each bdder the mnmum bd at whch he would have mantaned hs allocaton. These bounds requre a standard no-overbddng assump- Journal of the ACM, Vol. V, No. N, Artcle A, Publcaton date: January YYYY.

4 A:4 A. Borodn and B. Lucer ton, whch s that agents are avod bddng more than ther value for any gven subset of tems. Theorem (nformal): Suppose A s a greedy c-approxmate allocaton rule for a combnatoral allocaton problem. Then, under the assumpton that agents do not overbd, the aucton that uses A to choose allocatons, and uses a crtcal-prce payment scheme, has a Bayes-Nash Prce of Anarchy of at most c + 1. We also show that the extra +1 term s necessary, by gvng an example for every c 2 n whch the resultng prce of anarchy s exactly c + 1. As wth the frst-prce results, our bounds extend to coarse correlated equlbra, and a bound of O(c) holds f agent valuatons can be correlated. Furthermore, we show that a slght modfcaton to the mechansm allows us to replace the no-overbddng assumpton wth the (conceptually weaker) assumpton that bdders avod weakly domnated strateges. Repeated Auctons.. Our bounds on effcency at equlbrum do not explctly model the manner by whch agents reach equlbrum, or mpose upon the agents any computatonal constrants whatsoever. We smply post that equlbra (or approxmate equlbra 3 ) are predctve of the behavor of ratonal agents n hgh-stakes auctons. However, n settngs where auctons are explctly repeated, one mght naturally model the dynamcs under whch bdder behavor evolves. We wll therefore also consder a repeated-game varant of combnatoral allocaton problems, n whch an aucton problem s resolved multple tmes wth the same objects and bdders. Perhaps the most well-studed modern examples of repeated auctons are auctons for advertsng spaces or slots [Edelman et al. 2005], but ths model apples also to bandwdth auctons (such as the FCC spectrum aucton), arlne landng rghts auctons [Cramton et al. 2005], etc. In these settngs a mechansm for the (one-shot) aucton problem corresponds to a repeated game to be played by the agents. Rather than vew a repeated aucton as an extensve-form game, we consder models of lmted ratonalty that attempt to capture natural bddng behavour. We we study two such models: external regret mnmzaton and asynchronous best-response dynamcs. In the frst model, agents can play arbtrary sequences of strateges for the repeated aucton, under the assumpton that they obtan low regret relatve to the best fxed strategy n hndsght. More precsely, for each bdder, the dfference between the average utlty obtaned by the bdder and the average utlty that would have been obtaned by the best sngle declaraton n hndsght must tend to 0 as the number of aucton rounds ncreases. Under the assumpton that bdders are able to mnmze external regret, our goal s to desgn an aucton mechansm that acheves an approxmaton to the optmal socal welfare on average over suffcently many rounds of the repeated aucton. Ths s precsely the problem of desgnng a mechansm wth bounded prce of total anarchy, as ntroduced by Blum et al [Blum et al. 2008]. As observed by Blum and Mansour [Blum and Mansour 2007] and Roughgarden [Roughgarden 2009; 2012], n the full nformaton settng, prce of anarchy wth respect to coarse correlated equlbrum s equal to the total prce of anarchy. Hence, the bounds stated above for coarse correlated equlbrum apply also to the total prce of anarchy. We further show that for the greedy mechansms we consder, regret-mnmzng strateges can be computed effcently, assumng a natural representaton of the bdders valuaton functons. 3 All of our bounds on socal effcency degrade gracefully when agents apply strateges n approxmate equlbrum. Namely, whenever we convert a c approxmate allocaton algorthm nto a mechansm achevng say f(c) prce of anarchy, the same proof shows that the mechansm acheves at least (and often better) an f(c + γ) approxmaton at every (1 + γ) approxmate equlbrum. Notably, f c 1 + γ, the c-approxmaton for pure equlbra of the frst prce mechansm remans a c approxmaton at every approxmate equlbrum. Journal of the ACM, Vol. V, No. N, Artcle A, Publcaton date: January YYYY.

5 Equlbra of Greedy Combnatoral Auctons A:5 Theorem (nformal): Suppose A s a greedy c-approxmate allocaton rule for a combnatoral allocaton problem. Then the aucton that uses A to choose allocatons, and charges crtcal payments, acheves a (c + 1) approxmaton to the optmal welfare when agents apply regret-mnmzng strateges. Moreover, a regret-mnmzng strategy can be mplemented n tme polynomal n the number of XOR bds 4 used to represent an agent s valuaton. In the second model, we assume that agents choose strateges that are myopc bestresponses to the current strateges of the other agents. We model ths behavour as follows: on each aucton round, an agent s chosen unformly at random, and that agent s gven the opportunty to change hs strategy to the current myopc best-response. As n the regret-mnmzaton model, our goal s to desgn aucton mechansms that acheve approxmatons to the best possble socal welfare on average over suffcently many aucton rounds, wth hgh probablty over the random choces of bdders to update. Ths s the concept of the prce of (myopc) snkng, as ntroduced by Goemans et al [Goemans et al. 2005]. We conjecture that any greedy c-approxmate allocaton rule can be mplemented as a mechansm wth prce of snkng O(c). As partal progress toward ths conjecture, we desgn mechansms talored to two partcular combnatoral allocaton problems: the unrestrcted combnatoral aucton problem and the cardnalty-restrcted combnatoral aucton. Each mechansm has prce of snkng O(c), where c s the approxmaton factor of the best-known algorthm. We recall that one method for boundng the prce of snkng s to prove that there exsts an equlbrum state that s reachable from any declaraton profle by some polynomal-length sequence of best-response steps. Ths would mply that an equlbrum state would be reached wth hgh probablty after exponentally many steps. We do not take ths approach, but rather prove that the average socal welfare obtaned after a polynomal number of steps wll approxmate the optmal welfare wth hgh probablty Related Work The semnal paper n algorthmc game theory and more specfcally algorthmc mechansm desgn s that of Nsan and Ronen [Nsan and Ronen 1999]. The basc ssue ntroduced n [Nsan and Ronen 1999] s to reconcle the competng demands for revenue and socal welfare optmzaton wth the need for computatonal effcency n the context of self nterested (.e. selfsh) agents. The two most studed soluton concepts n algorthmc game theory are truthfulness (.e. ncentve compatablty) and behavor at (all) equlbra (.e. the prce of anarchy (POA) concept). Intal POA results for games were frst ntroduced to algorthmc game theory n the semnal papers by Papdmtrou [Papadmtrou 2001] and Roughgarden and Tardos [Roughgarden and Tardos 2000]. Chrstodoulou et al. [Chrstodoulou et al. 2008] ntated the study of the prce of the prce of anarchy n the Bayesan settng. Whereas the emphass of algorthmc mechansm desgn has been to consder the approxmatons acheveable by truthful mechansms, to the best of our knowledge, our conference paper [Lucer and Borodn 2010] was the frst to consder ths constructve aspect of mechansm desgn and prce of anarchy. Snce the ntal conference verson of ths work there has been sgnfcant progress on the understandng of the prce of anarchy of mechansms n varous aucton settngs. Some examples nclude the Generalzed Second Prce aucton for sponsored 4 Equvalently, the mnmum number of (subset, value) pars (S, w ) needed so that valuaton v satsfes v(t ) = max : S T {w }. Journal of the ACM, Vol. V, No. N, Artcle A, Publcaton date: January YYYY.

6 A:6 A. Borodn and B. Lucer search ads [Caraganns et al. 2012], smultaneous sngle-tem auctons [Chrstodoulou et al. 2008; Hassdm et al. 2011; Bhawalkar and Roughgarden 2011; Feldman et al. 2013], and mult-unt auctons [Markaks and Telels 2012; de Kejzer et al. 2013]. A framework unfyng much of ths work was proposed by Syrgkans and Tardos [Syrgkans and Tardos 2013]. Chekur and Gamzu [Chekur and Gamzu 2009] defned loser-ndependent algorthms, and n the conference verson of our paper [Lucer and Borodn 2010] we argued that the basc property of greedy algorthms that we were explotng was a multparameter verson of loser-ndependence. In the frst author s thess [Lucer 2011], a strengthenng of loser-ndependence, called strong loser-ndependence, was ntroduced to smplfy the proofs and that wll be the basc property of greedy algorthms we wll use n ths paper. Loser-ndependence s conceptually related to the concept of smoothness, whch was ntroduced by Roughgarden [Roughgarden 2009] as a general way to derve prce of anarchy results for one shot and repeated games (wthout reference to mechansms that derve games). Loser-ndependence has been shown to be dfferent from ths orgnal noton of smoothness [Lucer 2011]. However, alternatve notons of smoothness defned by Lucer and Paes Leme [Lucer and Leme 2011] and Syrgkans and Tardos [Syrgkans and Tardos 2013] can also be used to derve results smlar to our results. In partcular, Syrgkans and Tardos use ther smoothness condton to derve many prce of anarchy results for allocaton mechansms, ncludng those derved from greedy c-approxmaton algorthms. Ther result for the (non correlated) mxed Bayesan and coarse correlated equlbrum mproved upon our conference results: as n our current paper, they show that the resultng prce of anarchy approaches c wth a term exponentally decreasng n c. In partcular, they show that the prce of anarchy s never worse than c As we wll show n Secton 3.1, our applcaton of strong loser-ndependence can be nterpreted as a proof of smoothness. 2. PRELIMINARIES 2.1. Feasble Allocaton Problems We consder a settng n whch there are n agents and a set M of m objects. An allocaton to agent s a subset x M. A valuaton functon v : 2 M R assgns a value to each allocaton. We assume that valuaton functons are monotone, meanng v(s) v(t ) for all S T M, and normalzed so that v( ) = 0. A valuaton functon v s sngle-mnded f there exsts a set S M and a value y 0 such that for all T M, v(t ) = y f S T and 0 otherwse. A valuaton profle v s a vector of n valuaton functons, one for each agent. In general we wll use boldface to represent vectors, subscrpt to denote the th component, and subscrpt to denote all components except, so that v = (v, v ). An allocaton profle x s a vector of n allocatons. The goal n our socal welfare maxmzaton problems s to choose an allocaton for each agent n order to maxmze the sum of agent values. A combnatoral allocaton problem s defned by a set of feasble allocatons, whch s the set of permtted allocaton profles. We further assume n combnatoral allocaton problems that ths feasblty constrant s separable, meanng that f x s feasble then (, x ) s also feasble 5 for all. Note that separablty s a weaker assumpton than the standard downward-closure property of packng problems, whch would stpulate that f x s feasble then (y, x ) s also feasble for all y x. An allocaton rule A assgns to each valuaton profle v a feasble outcome A(v); we wrte A (v) for the allocaton to agent. An allocaton rule s component-wse monotone f t satsfes the 5 We note that the combnatoral publc projects problem (CPPP) [Papadmtrou et al. 2008] s not separable. Journal of the ACM, Vol. V, No. N, Artcle A, Publcaton date: January YYYY.

7 Equlbra of Greedy Combnatoral Auctons A:7 followng property for every agent : If v (S) < ṽ (S), v (T ) = ṽ (T ) T S, and A (v, v ) = S, then A (ṽ, v ) = S We wll tend to wrte A for both an allocaton rule and an algorthm that mplements t. We wll sometmes abuse notaton and use x for an allocaton rule, rather than a specfc allocaton. Each agent [n] has a prvate valuaton functon v, hs type, whch defnes the value he attrbutes to each allocaton. The socal welfare obtaned by allocaton profle x, gven type profle v, s SW (x, v) = v (x ). We wrte SW opt (v) for max x {SW (x, v)} and say that algorthm A s a c approxmaton algorthm 6 f SW (A(v), v) 1 c SW opt(v) for all v. A type profle v and an allocaton rule A for a combnatoral allocaton problem defne crtcal values, θ (S, v ), for any agent and set S M. The value θ (S, v ) s the mnmum amount that agent could bd on set S and stll wn S, assumng the other agents have profle v. That s, θ (S, v ) = nf{z : x (z, v ) = S}. We note that ths defnton of crtcal values holds even f t s not the case that ncreasng one s value for a set necessarly ncreases the probablty of obtanng that set. However, most of the mechansms we consder n ths work do satsfy ths monotoncty property, whch motvates the termnology of a crtcal prce Mechansms A drect revelaton mechansm M(A, P ) s composed of an allocaton rule A and a payment rule P that assgns a vector of n payments to each declared valuaton profle. The mechansm proceeds by elctng a valuaton profle d from each of the agents, called the declaraton profle. It then apples the allocaton and payment rules to d to obtan an allocaton and payment for each agent. Crucally, we do not assume that d s equal to v. We wll wrte SW (d) for SW (A(d), v) when the allocaton rule and type profle are clear from context. We wll be concerned wth two dfferent payment rules, frst prce and crtcal prce. In a frst prce mechansm, an agent s charged ther declared bd d (S) for any allocated set S. For notatonal convenence, we let M 1 (A) denote the mechansm usng allocaton rule A and the frst prce payment rule. In the crtcal prce payment rule, an agent s charged hs crtcal value θ (S, d ) for any allocated set S. We wll let M 2 (A) denote the mechansm usng allocaton rule A and the crtcal prce payment rule Equlbra of One-shot Auctons The utlty of agent n mechansm M = (A, P ), gven declaraton profle d and type profle v, s u v (d) = v (A (d)) P (d). We wll often omt the dependence on v when t s clear from context, and wrte smply u (d). We say that declaraton d weakly domnates d f for all d, u (d, d ) u (d, d ), and that there exst at least one d for whch the nequalty s strct. We consder a Bayesan settng n whch the true types of the agents are not fxed, but are rather drawn from a known probablty dstrbuton F over the set of valuaton profles. We frst assume that F = F 1... F n s the product of ndependent dstrbutons, where F (v ) s the probablty that agent has type v. (Later we wll also consder correlated dstrbutons over type profles.) We wrte SW opt (F) for E v F [SW opt (v)]. A bddng strategy for agent s a functon b that maps a type v to a dstrbuton over declaratons for agent. We thnk of b (v ) as the (randomzed) bddng strategy 6 Our conventon wll be to have approxmaton ratos c 1. Journal of the ACM, Vol. V, No. N, Artcle A, Publcaton date: January YYYY.

8 A:8 A. Borodn and B. Lucer employed by agent gven that hs true type s v. We wll abuse notaton slghtly and also wrte b (v ) for the random varable representng a declaraton chosen from the correspondng dstrbuton. We wrte b(v) = b 1 (v 1 )... b n (v n ) for the (dstrbuton over) declaraton profles resultng from applyng the bd functons n b to type profle v. The strategy profle b forms a (mxed) Bayesan Nash Equlbrum (BNE) f, for every [n] and every v n the support of F, agent maxmzes hs expected utlty by makng a declaraton drawn from dstrbuton b (v ). That s, for each agent, each possble type v, and every dstrbuton ω over declaratons, E v F [u (b(v))] E v F,d ω [u (d, b (v ))]. For a mechansm M = (A, P ), we wll wrte SW M (F, b) to mean E v F [ v (A (b(v)))], the expected socal welfare gven type dstrbuton F and strategy profle b. The (mxed) Bayesan prce of anarchy (BPoA) of mechansm M s defned as sup F,b SW opt (F) SW M (F, b) where the supremum s over all type dstrbutons F and mxed BNE b for F. In other words, the BPoA of M s the worst-case rato between the expected welfare at BNE and the expected optmal welfare. We can further extend the defnton of BNE to allow a correlated dstrbuton over type profles. The defnton for correlated BNE and correlated Bayesan prce of anarchy s then the same as the above defntons, where we would no longer assume that F s a product of ndependent dstrbutons. Returnng to the case n whch F s a product dstrbuton, a number of specal cases deserve menton. When all type dstrbutons are pont masses (.e., each agent s type s determned), a BNE s referred to as a (mxed) Nash Equlbrum (NE). The Prce of Anarchy (PoA) of a mechansm M s defned analogously to the BPoA, but wth respect to fxed type profles and mxed Nash equlbra. It follows that the BPoA s always at least the PoA for a gven mechansm. A BNE (or NE) s called pure f ts consttuent bddng strateges are determnstc. In general a pure Nash equlbrum may not exst for a gven mechansm and type profle; see Appendx A. One can generalze mxed NE by relaxng the assumpton that the declaraton dstrbutons are ndependent. That s, one mght allow b(v) to be an arbtrary dstrbuton over declaratons, rather than a product dstrbuton. A dstrbuton ω over declaraton profles s a coarse correlated equlbrum (CCE) for type profle v f, for all and all declaraton dstrbutons ω, E d ω [u (d)] E d (ω,ω )[u (d)]. (1) Note that when the agent declaraton dstrbutons are ndependent, CCE s equvalent to mxed NE. We defne the analogous prce of anarchy concepts; t follows that the pure prce of anarchy s at most the mxed prce of anarchy whch n turn s at most the coarse correlated prce of anarchy Greedy Allocaton Rules We descrbe a specal type of allocaton rule, whch we wll refer to as a greedy allocaton rule. These are motvated by the prorty framework n Borodn, Nelsen and Rackoff [Borodn et al. 2002] and the monotone greedy algorthms of Mu alem and Nsan [Mu alem and Nsan 2008], extended to be adaptve as n Borodn and Lucer [Borodn and Lucer 2010]. We begn wth some defntons. A partal allocaton profle Journal of the ACM, Vol. V, No. N, Artcle A, Publcaton date: January YYYY.

9 Equlbra of Greedy Combnatoral Auctons A:9 Prorty Algorthm Input: Declaraton profle d = d 1,..., d n. 1. Fx a monotone prorty functon r. Let N =. 2. Repeat untl N = [n]: 3. Choose / N and feasble allocaton S M for gven N that maxmzes r(, S, d (S)) 4. Set x = S; add player to N 5. return x = (x 1,..., x n ) Fg. 1. The framework for a non-adaptve prorty algorthm. for agents N [n] s an allocaton profle x wth x = for all N. A partal allocaton profle s feasble f there s some feasble allocaton profle that extends t. Gven a partal allocaton profle x for subset N, some N, and allocaton y M, we say y s a feasble allocaton for gven N f the partal allocaton remans feasble when we set x = y. A prorty functon s a functon r : [n] 2 M R R. We thnk of r(, S, v) as the prorty of allocatng S M to player when v (S) = v. We say that r s monotone f t s non-decreasng n v and monotone non-ncreasng n S wth respect to set ncluson. We consder two types of greedy allocaton algorthms. A non-adaptve greedy allocaton algorthm A s an allocaton algorthm as defned n Fgure 1. We say that A s monotone when the prorty functon r s monotone. We assume that tes n step 3 are broken n an arbtrary but fxed manner (.e. we assume that the prorty functon s a 1-1 functon nducng a total orderng). A non-adaptve algorthm fxes a sngle prorty functon that s used throughout ts executon. By constrast, an adaptve greedy allocaton algorthm can change ts prorty functon on each teraton, dependng on the partal allocaton formed on the prevous teratons Applcatons We now descrbe some applcatons of greedy algorthms for partcular combnatoral allocaton problems. Combnatoral Auctons. The general combnatoral aucton problem s defned by the feasblty constrant that no two allocatons can ntersect. Lehmann et al [Lehmann et al. 1999] show that the (non-adaptve) greedy allocaton rule wth r(, S, v) = v acheves a 2m approxmaton rato for CAs. S Cardnalty-restrcted Combnatoral Auctons. In the specal case that players desres are restrcted to sets of sze at most k, the non-adaptve greedy algorthm wth r(, S, v) = v s k-approxmate assumng sngle-mnded agents. Ths translates to a (k + 1) approxmate algorthm for general (.e. mult-mnded) agents. Multple-Demand Unsplttable Flow Problem. In the unsplttable flow problem (UFP), we are gven an undrected graph wth edge capactes. The objects are the edges, and each valuaton functon s such that agent has some value v(s, t) for beng gven a path from s to t. Each agent addtonally specfes a fractonal demand d [0, 1] correspondng to a desred amount of flow to send along the gven path. An allocaton s feasble f the total allocated flow along each edge s no more than ts capacty. Let B be the mnmum edge capacty. A prmal-dual algorthm, whch s an Journal of the ACM, Vol. V, No. N, Artcle A, Publcaton date: January YYYY.

10 A:10 A. Borodn and B. Lucer adaptve greedy allocaton rule, obtans an O(m 1/(B 1) ) approxmaton for any B > 1 [Brest et al. 2005]. Convex Bundle Auctons. In a convex bundle aucton, M s the plane R 2, and allocatons must be non-ntersectng compact convex sets. We suppose that agents declare valuaton functons by makng bds for such sets. Gven such a collecton of bds, the aspect-rato, R, s defned to be the maxmum dameter of a set dvded by the mnmum wdth of a set. A non-adaptve greedy allocaton rule usng a geometrcallymotvated prorty functon yelds an O(R 4/3 ) approxmaton [Babaoff and Blumrosen 2004]. Alternatve greedy algorthms yeld better approxmaton ratos for specal cases, such as rectangles. Max-proft Unt Job Schedulng. In ths problem, each bdder has a job of unt tme to schedule on one of multple machnes. A bdder has varous wndows of tme of the form (release tme, deadlne, machne) n whch hs job could be scheduled, wth a potentally dfferent proft resultng from each wndow. The profts and wndows are prvate nformaton to each bdder. The goal of the mechansm s to schedule the jobs to maxmze the total proft. The greedy algorthm that orders bds by value obtans a 3-approxmaton, and s symmetrc wth respect to agents and objects. Unlke the prevous examples, for the case of sngle-mnded bdders, there s an optmal dynamc programmng algorthm that runs n tme O(n 7 ) [Baptste 1999]. Snce ths algorthm solves the problem optmally, t s ncentve compatble. In ths case, the resultng prce of anarchy for the greedy algorthm s appealng prmarly due to ts lnear runtme and smple allocaton rule. 3. STRONG LOSER-INDEPENDENCE Chekur and Gamzu [Chekur and Gamzu 2009] ntroduced a property known as loser-ndependence for combnatoral allocaton algorthms n sngle-parameter domans. They defne an algorthm for a combnatoral allocaton problem to be loserndependent f, whenever A (d, d ) = A (d, d ) = for some, d, d, and d, then t must be that A(d, d ) = A(d, d ). That s, f a losng agent (.e. an agent who s allocated no tems) modfes hs declaraton n such a way that he stll receves no tems, ths cannot affect the outcome of algorthm A. Note that loser-ndependence s a condton on declaraton profles, rather than on bddng functons, snce the loserndependence noton s purely algorthmc and s not a condton on equlbra. In our results we wll make use of a stronger property of greedy algorthms, whch we call strong loser-ndependence. DEFINITION 3.1. An allocaton rule A s strongly loser-ndependent f, whenever d and d satsfy A(d) A(d ), there exsts an agent and set S such that d (S) d (S) and ether A (d) = S or A (d, d ) = S. Roughly speakng, f A s a strongly loser-ndependent algorthm, then whenever a valuaton profle changes from d to d va modfcatons to losng bds (.e. an agent s declared value for sets that are not allocated to hm, when others bd accordng to d ), algorthm A wll return the same outcome on nputs d and d. We note that our defnton requres that ether A (d) = S or A (d, d ) = S, rather than A (d ) = S. The ntuton s that we thnk of losng bds as beng losers wth respect to the orgnal declaraton profle d. The property of strong loser-ndependence strengthens the defnton of loserndependence due to Chekur and Gamzu n two ways. Frst, we extend from sngleparameter settngs to multple-parameter settngs by consderng losng bds rather Journal of the ACM, Vol. V, No. N, Artcle A, Publcaton date: January YYYY.

11 Equlbra of Greedy Combnatoral Auctons A:11 than losng agents. Second, we requre that the algorthm outcome be unaffected f multple agents smultaneously modfy losng bds. It s clear from the defntons that all strongly loser-ndependent algorthms are loser-ndependent (.e. by consderng the case when d and d dffer only on the declaraton of a sngle agent). However, not all loser-ndependent algorthms are strongly loser-ndependent, even n sngle-mnded domans. For example, consder the combnatoral aucton problem and suppose that A s an algorthm that optmzes socal welfare exactly and breaks tes consstently. Then A s loser-ndependent, snce a losng agent s bd does not affect the optmal allocaton. However, A s not strongly loser-ndependent, as the followng nstance shows. Consder an aucton of two tems {a, b} to three bdders. If the (sngle-mnded) bdder declaratons are d 1 ({a, b}) = 10, d 2 ({a}) = 3, and d 3 ({b}) = 3, then the outcome s that agent 1 wns hs desred set. On the other hand, f the bdder declaratons are gven by d where d 1 ({a, b}) = 10, d 2 ({a}) = 6, and d 3 ({b}) = 6, then the outcome changes: agents 2 and 3 wn ther desred sets. However, d and d do not dffer n ther declaratons for any sets allocated by A(d), A(d 1, d ), A(d 2, d ), or A(d 3, d ), as agent 1 wns hs desred set n each of these four cases. Ths contradcts the defnton of strongly loser-ndependence. As we now show, all greedy algorthms satsfy the strong loser-ndependence property. LEMMA 3.2. Every (monotone) adaptve greedy algorthm s (component-wse monotone) and strongly loser-ndependent. PROOF. The monotoncty property follows mmedately when the prorty functon n the greedy algorthm s a monotone functon. Let A be an adaptve greedy allocaton rule, and choose any d and d such that A(d) A(d ). We wll show that there exsts some and S such that d (S) d (S) and ether A (d) = S or A (d, d ) = S. Recall the defnton of an adaptve greedy algorthm, and consder the teratons of A on nputs d and d. Let k be the frst teraton n whch the allocaton of A dffers on these two nputs. Suppose that A allocates set U to agent l on teraton k when the nput s d, and allocates T to agent j on teraton k when the nput s d. For each teraton q < k, wrte q for the agent allocated to by A (on ether nput profle) and S q for the set allocated to q. Note that f d q (S q ) d q (S q ) for any q < k then we have the desred result wth = q and S = S q. We can therefore assume that d q (S q ) = d q (S q ) for all q < k. Ths mples that the bds resolved by A are dentcal on all teratons preceedng k on nputs d and d, and therefore the values of rankng functons used n each teraton up to k must be dentcal for nputs d and d. Wrte r q for the rankng functon used n teraton q for each q k. Thus, snce the allocaton on teraton k changed from choosng set U for agent l to choosng set T for agent j, t must be that ether r k (l, U, d l (U)) r k (l, U, d l (U)) or r k (j, T, d j (T )) r k (j, T, d j (T )). Ths mples that ether d l (U) d l (U) or d j (T ) d j (T ). If d l (U) d l (U) then we have the desred result wth = l and S = U, snce A l (d) = U. We can therefore assume that d l (U) = d l (U) and d j (T ) d j (T ). Consder now the behavour of algorthm A on nput (d j, d ). We clam that A j (d j, d ) = T. Note that ths mples the desred result wth = j and S = T. To prove the clam, recall that d q (S q ) = d q (S q ) for all q < k. Thus, for each q < k and each feasble set S that could be allocated to agent j on teraton q, r q ( q, S q, d q (S q )) = r q ( q, S q, d q (S q )) > r q (j, S, d j (S)) snce A allocates S q to q on nput d. We conclude that, on nput (d j, d ), A allocates S q to agent q on each teraton q < k. On teraton k, we have r k (j, T, d j (T )) > Journal of the ACM, Vol. V, No. N, Artcle A, Publcaton date: January YYYY.

12 A:12 A. Borodn and B. Lucer r k (j, T, d j (T )) for any feasble T T (snce A allocates T to j on nput d ) and r k (j, T, d j (T )) > r k (l, U, d l (U)) = r k (l, U, d l (U)) r k (, S, d (S)) for any feasble j, due to our assumpton that d l (U) = d l (U) and the fact that A allocates U to l on teraton k for nput d. We therefore conclude that A j (d j, d ) = T as requred. We next explore an mplcaton of a strongly loser-ndependent algorthm A beng a worst-case c-approxmaton. If A s a c-approxmate algorthm, then (on any nput) the sum of the declared values for ts output profle approxmates the sum of the declared values for the optmal allocaton. We now show that t also approxmates the sum of the crtcal values of the optmal allocaton profle. LEMMA 3.3. If A s a c-approxmate strongly loser-ndependent algorthm, then for any type profle v and allocaton profle y, [n] v (A (v)) 1 c [n] θ (y, v ). PROOF. Choose any ɛ > 0. For all, let v be the sngle-mnded declaraton for set y at value θ (y, v ) ɛ. Let v be the pontwse maxmum of v and v. That s, for all S M, v (S) = max{v (S), v (S)}. By defnton of crtcal prces, we have that A (v, v ) = A (v) for all, and furthermore v (A (v)) = v (A (v)). Snce A s strongly loser-ndependent, we must therefore have A(v) = A(v ). Snce A s a c-approxmaton, we conclude that SW (x(v), v) = SW (x(v ), v ) 1 c SW (y, v ) 1 c [n] θ (y, v ) nɛ. The result follows by takng the lmt as ɛ 0. For brevty, for the remander of ths paper we wll say monotone strongly loserndependent to mean both strongly loser-ndependent and component-wse monotone Applyng Strong Loser Independence Strong loser-ndependence s a strctly algorthmc concept, devod of game theoretc consderatons. Our general approach wll be to derve prce of anarchy results for any mechansm that uses a strongly loser-ndependent c-approxmaton A as ts allocaton algorthm. To do so, we wll be usng Lemma 3.3 n conjucton wth the assumpton that a gven bd profle s an equlbrum. At a hgh level, our argument wll be as follows. For each prcng rule and equlbrum concept, equlbrum wll mply an nequalty of the form v (y ) λ θ (y, d ) + µ v (x (d)) where y s an optmal allocaton. (For Bayesan equlbrum, these terms wll be expectatons.) Ths allows us to charge the optmal gan for each agent to ts crtcal value and ts welfare from the algorthm. We then explot Lemma 3.3 to convert ths bound nto a relatonshp between the optmal welfare and the welfare at equlbrum. To make ths more specfc, n our pure Nash equlbrum result for a frst prce mechansm (Theorem 4.3), we show the followng (somewhat stronger) nequalty: v (y ) θ (y, d ) + v (x (d)) d (x (d)) 7 For pure Nash equlbrum prce of anarchy results, f we assume no over-bddng, we do not need monotoncty but t s necessary for all of our other results. Journal of the ACM, Vol. V, No. N, Artcle A, Publcaton date: January YYYY.

13 Equlbra of Greedy Combnatoral Auctons A:13 It wll then follow (gnorng an nɛ term whch dsappears as ɛ 0) that: v (y ) θ (y, d ) + v (x (d)) d (x (d)) (c 1) d (x (d)) + v (x (d)) In other words, the hgh-level approach s to charge an agent s welfare n the optmal outcome aganst hs welfare at equlbrum plus the welfare of other prce-settng agents. Ths approach s smlar to the smoothness argument as formulated by Syrgkans and Tardos [Syrgkans and Tardos 2013]. However, there s a dfference n our approaches. The smoothness condton n [Syrgkans and Tardos 2013] s talored to allocaton mechansms and asserts the exstence of some d (for each player ) satsfyng such an nequalty whereas we are assumng that d s an equlbrum. The beneft of ther mmedate reducton to smoothness s that ther prce of anarchy results for pure equlbra carry over mmedately to Bayesan prce of anarchy. However, ths prohbts establshng certan tght bounds; for example, n the frst prce mechansm we show that the pure prce of anarchy s c, whch cannot be acheved va smoothness snce ths bound does not hold for the Bayesan prce of anarchy. 4. FIRST-PRICE MECHANISMS In ths secton we analyze greedy algorthms pared wth a frst prce payment scheme. More precsely (wth the excepton of results relatng to correlated Bayesan equlbrum where we wll consder more specfc greedy allocatons), gven a strongly loserndependent algorthm A, we wll be studyng the performance of the frst-prce mechansm M 1 (A) at equlbrum. Our frst step wll be to show that a utlty-maxmzng declaraton of an agent never nvolves overbddng on a set that he may possbly be allocated. Ths wll mply that agents do not employ overbddng strateges at equlbrum. It may appear at frst glance that any strategy that recommends overbddng on sets s obvously domnated for any allocaton algorthm, snce wnnng any bd larger than one s true value leads to negatve utlty. However, we must also show that an agent cannot fnd t advantageous to overbd on some set S n order to affect hs probablty of wnnng some other set T. We wll demonstrate that such stuatons cannot occur when allocatons are chosen by a strongly loser ndependent algorthm. For a type v and a declaraton d, we wll wrte d for the declaraton defned as d (S) = mn{v (S), d (S)}. That s, d agrees wth d, except that the declared value of each set can be at most the true value for that set. Note that d = d precsely f d does not overbd on any set. We now show that any declaraton d that overbds on a set that could potentally be won s weakly domnated by strategy d. LEMMA 4.1. For any monotone strongly loser ndependent allocaton rule A, valuaton v, and declaraton profle d, we have u (d) u (d, d ). Moreover, the nequalty s strct when d (A(d)) > v (A(d)). PROOF. Let S = A (d). Suppose frst that d (S) > v (S). Then u (d) = v (S) d (S) < 0. Snce v (T ) d (T ) 0 for every set T, ths mples that u (d, d ) > u (d), as requred. Next suppose that d (S) v (S), so that d (S) = d (S). We clam that A (d, d ) = S. Suppose not, for contradcton. Then we can construct a sequence of declaratons (d 1, d 2,..., d k ), wth d 1 = d and d k = d, such that adjacent declaratons dffer only on a sngle set and declared values only decrease. Suppose j s mnmal such that Journal of the ACM, Vol. V, No. N, Artcle A, Publcaton date: January YYYY.

14 A:14 A. Borodn and B. Lucer A (d j, d ) S; such a j > 1 must exst snce, by assumpton, A (d, d ) S. Then (a) d j 1 and d j dffer only on the value assgned to some set T, (b) d j 1 (T ) > d j (T ), (c) A (d j 1, d ) = S, and (d) A (d j, d ) S. Strong loser-ndependence then mples that A (d j, d ) = T. However, the fact that d j 1 (T ) > d j (T ) then contradcts the component-wse monotoncty of A. We conclude by contradcton that A (d, d ) = S. Snce S s also A(d), we have u (d) = v (S) d (S) = u (d, d ) as requred. An mmedate corollary s that f b s a bddng strategy, and there exsts a type v and set S such that (b (v ))(S) > v (S), then b s weakly domnated by the strategy b. Moreover, b s strctly better, n terms of utlty, under any dstrbuton of declaratons n whch agent wns set S wth postve probablty. We conclude that at any BNE of mechansm M 1 (A), no player wll overbd on a set that he wns wth postve probablty. COROLLARY 4.2. For any monotone strongly loser ndependent allocaton rule A, BNE b, type v, and set S, f Pr v F [A (b(v)) = S] > 0 then (b (v ))(S) v (S) Pure Nash Equlbra We are now ready to bound the prce of anarchy of M 1 (A). We begn wth a result for pure Nash equlbra, rather than the fully general BNE case. THEOREM 4.3. Suppose A s a c-approxmate monotone strongly loser ndependent allocaton rule for a combnatoral allocaton problem. Then the prce of anarchy of M 1 (A) s at most c. PROOF. Fx type profle v and suppose that b forms a pure Nash equlbrum. Snce the Nash equlbrum s pure, we wll wrte d = b(v) for notatonal convenence. Let y be an optmal allocaton for v, and let x( ) denote the allocaton rule for A. Lemma 3.3 mples d (x (d)) 1 θ (y, d ). (2) c Choose arbtrarly small ɛ > 0 and let d be the sngle-mnded declaraton for set y at value θ (y, d ) + ɛ. Then x (d, d ) = y (from the defnton of crtcal values) and hence u (d, d ) = v (y ) θ (y, d ) ɛ. Snce d s a Nash equlbrum, t must be that v (y ) θ (y, d ) ɛ = u (d, d ) u (d, d ) = v (x (d)) d (x (d)). Summng over all and applyng (2) and Corollary 4.2 we have v (y ) θ (y, d ) d (x (d)) + v (x (d)) + nɛ (c 1) d (x (d)) + v (x (d)) + nɛ c v (x (d)) + nɛ Journal of the ACM, Vol. V, No. N, Artcle A, Publcaton date: January YYYY.

15 Equlbra of Greedy Combnatoral Auctons A:15 whch, takng ɛ 0, mples SW (x(d), v) = v (x (d)) as requred. 1 v (y ) c = 1 c SW OP T (v) The power of Theorem 4.3 s marred by the fact that, for some problem nstances, the mechansm M 1 (A) s not guaranteed to have a pure Nash equlbrum. An example s gven n Appendx A Bayes-Nash Equlbra We are now ready to bound the mxed Bayesan prce of anarchy for mechansm M 1 (A). THEOREM 4.4. Suppose A s a monotone strongly loser ndependent allocaton rule for a combnatoral allocaton problem. Then the Bayesan prce of anarchy of M 1 (A) s at most 8 c 1 e for every ndependent type dstrbuton F. c c 1 e c We note that c ( ) e = c + O(c/e c ). The remander of ths subsecton s c dedcated to the proof of Theorem 4.4. Fx a product dstrbuton F over type profles and let b( ) be a (possbly mxed) Bayes-Nash equlbrum wth respect to F. Choose some type declaraton v and let y v denote an optmal allocaton for v. Followng the proof of Theorem 4.3, we would lke to bound the expected value of θ (y v, d ) wth respect to v (y v) and u (b(v)) for each. We encapsulate ths bound n Lemma 4.6 and Corollary 4.7, below. Ths wll allow us to use Lemma 3.3 to obtan a relaton between the expected welfare of A and the expected optmal welfare; ths relatonshp s gven n Lemma 4.5. LEMMA 4.5. Suppose that A s a c-approxmate monotone strongly loser ndependent allocaton rule and that there exst constants γ 0 and σ [0, c] for [n] such that, whenever b s a Bayes-Nash equlbrum for M 1 (A), t s the case that for all, all v, and all S M, E v [θ (S, b (v ))] γv (S) σ E v [u (b(v))]. Then E v [SW (A(b(v)), v)] γ c E v[sw OP T (v)]. LEMMA 4.6. Suppose that b s a Bayes-Nash equlbrum for mechansm M 1 (A) and dstrbuton F. Then for all, all v, and all S M, ( ) v (S) E v [θ (S, b (v ))] v (S) 1 + ln E v [u (b(v))]. E v [u (b(v))] Before provng Lemmas 4.5 and 4.6, let us show how they mply Theorem 4.4. We frst note the followng smple corollary of Lemma In the ntal conference verson of ths work, we presented a bound of c + O(log c) on the BPOA. Subsequently, ths bound was ndependently mproved by Lucer [Lucer 2011] to c + O(c 2 /e c ) and by Syrgkans and Tardos [Syrgkans and Tardos 2013] to c + O(c/e c ). We present here a slghtly modfed verson of the argument from Lucer [Lucer 2011], whch yelds the mproved Syrgkans and Tardos [Syrgkans and Tardos 2013] bound of c + O(c/e c ). Journal of the ACM, Vol. V, No. N, Artcle A, Publcaton date: January YYYY.

16 A:16 A. Borodn and B. Lucer COROLLARY 4.7. Suppose that b s a Bayes-Nash equlbrum for mechansm M 1 (A) and dstrbuton F. Then for all, all v, and all S M, E v [θ (S, b (v ))] (1 e c ) v (S) c E v [u (b(v))]. PROOF. Fx agent. By Lemma 4.6, we know ( E v [θ (S, b (v ))] v (S) 1 + ln v (S) E v [u (b(v))] ) E v [u (b(v))]. (3) ( ) v Note that f 1 + ln (S) E v [u (b(v))] c then (3) mmedately mples the desred result. We can therefore assume otherwse, and choose α > 0 such that ( ) v (S) 1 + ln = c + α. E v [u (b(v))] Rearrangng, we get that v (S) = e α e c 1 E v [u (b(v))]. Applyng these two equaltes to (3), we have Snce α e α E v [θ (S, b (v ))] v (S) (c + α) E v [u (b(v))] = v (S) α e α v(s) e c 1 c E v [u (b(v))]. acheves ts maxmum value of 1/e at α = 1, we can conclude that E v [θ (S, b (v ))] v (S) 1 e c v (S) c E v [u (b(v))] as requred. Theorem 4.4 follows drectly from Corollary 4.7 and Lemma 4.5. We next complete the proof of Theorem 4.4 by provng Lemmas 4.5 and 4.6. Proof of Lemma 4.5 : Fx dstrbuton F over type profles and let b( ) be a (possbly mxed) Bayes-Nash equlbrum wth respect to F. Choose some type declaraton v and let y v denote an optmal allocaton for v. We know that for all [n] and v, E v [θ (y v, b (v ))] γv (y v ) σ E v [u (b (v ), b (v )))]. Note the dstncton between v, over whch we are takng expectatons, and v, whch s the type profle fxed to defne y v. Now, summng over and takng expectaton over all choces of v, we have ] ] ] E v [ E v [θ (y v, b (v ))] γe v [ v (y v ) E v [ We now consder each of the three terms n (4). Frst, note that ] E v [ v (y v ) σ E v [u (b (v ), b (v ))] (4) = E v [SW OP T (v)]. (5). Journal of the ACM, Vol. V, No. N, Artcle A, Publcaton date: January YYYY.

Vickrey Auction VCG Combinatorial Auctions. Mechanism Design. Algorithms and Data Structures. Winter 2016

Vickrey Auction VCG Combinatorial Auctions. Mechanism Design. Algorithms and Data Structures. Winter 2016 Mechansm Desgn Algorthms and Data Structures Wnter 2016 1 / 39 Vckrey Aucton Vckrey-Clarke-Groves Mechansms Sngle-Mnded Combnatoral Auctons 2 / 39 Mechansm Desgn (wth Money) Set A of outcomes to choose

More information

Problem Set 9 Solutions

Problem Set 9 Solutions Desgn and Analyss of Algorthms May 4, 2015 Massachusetts Insttute of Technology 6.046J/18.410J Profs. Erk Demane, Srn Devadas, and Nancy Lynch Problem Set 9 Solutons Problem Set 9 Solutons Ths problem

More information

princeton univ. F 17 cos 521: Advanced Algorithm Design Lecture 7: LP Duality Lecturer: Matt Weinberg

princeton univ. F 17 cos 521: Advanced Algorithm Design Lecture 7: LP Duality Lecturer: Matt Weinberg prnceton unv. F 17 cos 521: Advanced Algorthm Desgn Lecture 7: LP Dualty Lecturer: Matt Wenberg Scrbe: LP Dualty s an extremely useful tool for analyzng structural propertes of lnear programs. Whle there

More information

Smooth Games, Price of Anarchy and Composability of Auctions - a Quick Tutorial

Smooth Games, Price of Anarchy and Composability of Auctions - a Quick Tutorial Smooth Games, Prce of Anarchy and Composablty of Auctons - a Quck Tutoral Abhshek Snha Laboratory for Informaton and Decson Systems, Massachusetts Insttute of Technology, Cambrdge, MA 02139 Emal: snhaa@mt.edu

More information

Assortment Optimization under MNL

Assortment Optimization under MNL Assortment Optmzaton under MNL Haotan Song Aprl 30, 2017 1 Introducton The assortment optmzaton problem ams to fnd the revenue-maxmzng assortment of products to offer when the prces of products are fxed.

More information

Module 17: Mechanism Design & Optimal Auctions

Module 17: Mechanism Design & Optimal Auctions Module 7: Mechansm Desgn & Optmal Auctons Informaton Economcs (Ec 55) George Georgads Examples: Auctons Blateral trade Producton and dstrbuton n socety General Setup N agents Each agent has prvate nformaton

More information

College of Computer & Information Science Fall 2009 Northeastern University 20 October 2009

College of Computer & Information Science Fall 2009 Northeastern University 20 October 2009 College of Computer & Informaton Scence Fall 2009 Northeastern Unversty 20 October 2009 CS7880: Algorthmc Power Tools Scrbe: Jan Wen and Laura Poplawsk Lecture Outlne: Prmal-dual schema Network Desgn:

More information

COS 521: Advanced Algorithms Game Theory and Linear Programming

COS 521: Advanced Algorithms Game Theory and Linear Programming COS 521: Advanced Algorthms Game Theory and Lnear Programmng Moses Charkar February 27, 2013 In these notes, we ntroduce some basc concepts n game theory and lnear programmng (LP). We show a connecton

More information

The Second Anti-Mathima on Game Theory

The Second Anti-Mathima on Game Theory The Second Ant-Mathma on Game Theory Ath. Kehagas December 1 2006 1 Introducton In ths note we wll examne the noton of game equlbrum for three types of games 1. 2-player 2-acton zero-sum games 2. 2-player

More information

Module 3 LOSSY IMAGE COMPRESSION SYSTEMS. Version 2 ECE IIT, Kharagpur

Module 3 LOSSY IMAGE COMPRESSION SYSTEMS. Version 2 ECE IIT, Kharagpur Module 3 LOSSY IMAGE COMPRESSION SYSTEMS Verson ECE IIT, Kharagpur Lesson 6 Theory of Quantzaton Verson ECE IIT, Kharagpur Instructonal Objectves At the end of ths lesson, the students should be able to:

More information

Computing Correlated Equilibria in Multi-Player Games

Computing Correlated Equilibria in Multi-Player Games Computng Correlated Equlbra n Mult-Player Games Chrstos H. Papadmtrou Presented by Zhanxang Huang December 7th, 2005 1 The Author Dr. Chrstos H. Papadmtrou CS professor at UC Berkley (taught at Harvard,

More information

Welfare Properties of General Equilibrium. What can be said about optimality properties of resource allocation implied by general equilibrium?

Welfare Properties of General Equilibrium. What can be said about optimality properties of resource allocation implied by general equilibrium? APPLIED WELFARE ECONOMICS AND POLICY ANALYSIS Welfare Propertes of General Equlbrum What can be sad about optmalty propertes of resource allocaton mpled by general equlbrum? Any crteron used to compare

More information

Economics 101. Lecture 4 - Equilibrium and Efficiency

Economics 101. Lecture 4 - Equilibrium and Efficiency Economcs 0 Lecture 4 - Equlbrum and Effcency Intro As dscussed n the prevous lecture, we wll now move from an envronment where we looed at consumers mang decsons n solaton to analyzng economes full of

More information

NP-Completeness : Proofs

NP-Completeness : Proofs NP-Completeness : Proofs Proof Methods A method to show a decson problem Π NP-complete s as follows. (1) Show Π NP. (2) Choose an NP-complete problem Π. (3) Show Π Π. A method to show an optmzaton problem

More information

Perfect Competition and the Nash Bargaining Solution

Perfect Competition and the Nash Bargaining Solution Perfect Competton and the Nash Barganng Soluton Renhard John Department of Economcs Unversty of Bonn Adenauerallee 24-42 53113 Bonn, Germany emal: rohn@un-bonn.de May 2005 Abstract For a lnear exchange

More information

Pricing and Resource Allocation Game Theoretic Models

Pricing and Resource Allocation Game Theoretic Models Prcng and Resource Allocaton Game Theoretc Models Zhy Huang Changbn Lu Q Zhang Computer and Informaton Scence December 8, 2009 Z. Huang, C. Lu, and Q. Zhang (CIS) Game Theoretc Models December 8, 2009

More information

ON THE EQUIVALENCE OF ORDINAL BAYESIAN INCENTIVE COMPATIBILITY AND DOMINANT STRATEGY INCENTIVE COMPATIBILITY FOR RANDOM RULES

ON THE EQUIVALENCE OF ORDINAL BAYESIAN INCENTIVE COMPATIBILITY AND DOMINANT STRATEGY INCENTIVE COMPATIBILITY FOR RANDOM RULES ON THE EQUIVALENCE OF ORDINAL BAYESIAN INCENTIVE COMPATIBILITY AND DOMINANT STRATEGY INCENTIVE COMPATIBILITY FOR RANDOM RULES Madhuparna Karmokar 1 and Souvk Roy 1 1 Economc Research Unt, Indan Statstcal

More information

Module 9. Lecture 6. Duality in Assignment Problems

Module 9. Lecture 6. Duality in Assignment Problems Module 9 1 Lecture 6 Dualty n Assgnment Problems In ths lecture we attempt to answer few other mportant questons posed n earler lecture for (AP) and see how some of them can be explaned through the concept

More information

Game Theory. Lecture Notes By Y. Narahari. Department of Computer Science and Automation Indian Institute of Science Bangalore, India February 2008

Game Theory. Lecture Notes By Y. Narahari. Department of Computer Science and Automation Indian Institute of Science Bangalore, India February 2008 Game Theory Lecture Notes By Y. Narahar Department of Computer Scence and Automaton Indan Insttute of Scence Bangalore, Inda February 2008 Chapter 10: Two Person Zero Sum Games Note: Ths s a only a draft

More information

CS286r Assign One. Answer Key

CS286r Assign One. Answer Key CS286r Assgn One Answer Key 1 Game theory 1.1 1.1.1 Let off-equlbrum strateges also be that people contnue to play n Nash equlbrum. Devatng from any Nash equlbrum s a weakly domnated strategy. That s,

More information

ECE559VV Project Report

ECE559VV Project Report ECE559VV Project Report (Supplementary Notes Loc Xuan Bu I. MAX SUM-RATE SCHEDULING: THE UPLINK CASE We have seen (n the presentaton that, for downlnk (broadcast channels, the strategy maxmzng the sum-rate

More information

2E Pattern Recognition Solutions to Introduction to Pattern Recognition, Chapter 2: Bayesian pattern classification

2E Pattern Recognition Solutions to Introduction to Pattern Recognition, Chapter 2: Bayesian pattern classification E395 - Pattern Recognton Solutons to Introducton to Pattern Recognton, Chapter : Bayesan pattern classfcaton Preface Ths document s a soluton manual for selected exercses from Introducton to Pattern Recognton

More information

Pricing Network Services by Jun Shu, Pravin Varaiya

Pricing Network Services by Jun Shu, Pravin Varaiya Prcng Network Servces by Jun Shu, Pravn Varaya Presented by Hayden So September 25, 2003 Introducton: Two Network Problems Engneerng: A game theoretcal sound congeston control mechansm that s ncentve compatble

More information

CS : Algorithms and Uncertainty Lecture 17 Date: October 26, 2016

CS : Algorithms and Uncertainty Lecture 17 Date: October 26, 2016 CS 29-128: Algorthms and Uncertanty Lecture 17 Date: October 26, 2016 Instructor: Nkhl Bansal Scrbe: Mchael Denns 1 Introducton In ths lecture we wll be lookng nto the secretary problem, and an nterestng

More information

Notes on Frequency Estimation in Data Streams

Notes on Frequency Estimation in Data Streams Notes on Frequency Estmaton n Data Streams In (one of) the data streamng model(s), the data s a sequence of arrvals a 1, a 2,..., a m of the form a j = (, v) where s the dentty of the tem and belongs to

More information

Lecture 14: Bandits with Budget Constraints

Lecture 14: Bandits with Budget Constraints IEOR 8100-001: Learnng and Optmzaton for Sequental Decson Makng 03/07/16 Lecture 14: andts wth udget Constrants Instructor: Shpra Agrawal Scrbed by: Zhpeng Lu 1 Problem defnton In the regular Mult-armed

More information

Feature Selection: Part 1

Feature Selection: Part 1 CSE 546: Machne Learnng Lecture 5 Feature Selecton: Part 1 Instructor: Sham Kakade 1 Regresson n the hgh dmensonal settng How do we learn when the number of features d s greater than the sample sze n?

More information

3.1 Expectation of Functions of Several Random Variables. )' be a k-dimensional discrete or continuous random vector, with joint PMF p (, E X E X1 E X

3.1 Expectation of Functions of Several Random Variables. )' be a k-dimensional discrete or continuous random vector, with joint PMF p (, E X E X1 E X Statstcs 1: Probablty Theory II 37 3 EPECTATION OF SEVERAL RANDOM VARIABLES As n Probablty Theory I, the nterest n most stuatons les not on the actual dstrbuton of a random vector, but rather on a number

More information

More metrics on cartesian products

More metrics on cartesian products More metrcs on cartesan products If (X, d ) are metrc spaces for 1 n, then n Secton II4 of the lecture notes we defned three metrcs on X whose underlyng topologes are the product topology The purpose of

More information

U.C. Berkeley CS294: Beyond Worst-Case Analysis Luca Trevisan September 5, 2017

U.C. Berkeley CS294: Beyond Worst-Case Analysis Luca Trevisan September 5, 2017 U.C. Berkeley CS94: Beyond Worst-Case Analyss Handout 4s Luca Trevsan September 5, 07 Summary of Lecture 4 In whch we ntroduce semdefnte programmng and apply t to Max Cut. Semdefnte Programmng Recall that

More information

(1 ) (1 ) 0 (1 ) (1 ) 0

(1 ) (1 ) 0 (1 ) (1 ) 0 Appendx A Appendx A contans proofs for resubmsson "Contractng Informaton Securty n the Presence of Double oral Hazard" Proof of Lemma 1: Assume that, to the contrary, BS efforts are achevable under a blateral

More information

Lecture Notes on Linear Regression

Lecture Notes on Linear Regression Lecture Notes on Lnear Regresson Feng L fl@sdueducn Shandong Unversty, Chna Lnear Regresson Problem In regresson problem, we am at predct a contnuous target value gven an nput feature vector We assume

More information

The Minimum Universal Cost Flow in an Infeasible Flow Network

The Minimum Universal Cost Flow in an Infeasible Flow Network Journal of Scences, Islamc Republc of Iran 17(2): 175-180 (2006) Unversty of Tehran, ISSN 1016-1104 http://jscencesutacr The Mnmum Unversal Cost Flow n an Infeasble Flow Network H Saleh Fathabad * M Bagheran

More information

Maximizing the number of nonnegative subsets

Maximizing the number of nonnegative subsets Maxmzng the number of nonnegatve subsets Noga Alon Hao Huang December 1, 213 Abstract Gven a set of n real numbers, f the sum of elements of every subset of sze larger than k s negatve, what s the maxmum

More information

Structure and Drive Paul A. Jensen Copyright July 20, 2003

Structure and Drive Paul A. Jensen Copyright July 20, 2003 Structure and Drve Paul A. Jensen Copyrght July 20, 2003 A system s made up of several operatons wth flow passng between them. The structure of the system descrbes the flow paths from nputs to outputs.

More information

Endogenous timing in a mixed oligopoly consisting of a single public firm and foreign competitors. Abstract

Endogenous timing in a mixed oligopoly consisting of a single public firm and foreign competitors. Abstract Endogenous tmng n a mxed olgopoly consstng o a sngle publc rm and oregn compettors Yuanzhu Lu Chna Economcs and Management Academy, Central Unversty o Fnance and Economcs Abstract We nvestgate endogenous

More information

Lecture Randomized Load Balancing strategies and their analysis. Probability concepts include, counting, the union bound, and Chernoff bounds.

Lecture Randomized Load Balancing strategies and their analysis. Probability concepts include, counting, the union bound, and Chernoff bounds. U.C. Berkeley CS273: Parallel and Dstrbuted Theory Lecture 1 Professor Satsh Rao August 26, 2010 Lecturer: Satsh Rao Last revsed September 2, 2010 Lecture 1 1 Course Outlne We wll cover a samplng of the

More information

The Price of Anarchy in Large Games

The Price of Anarchy in Large Games The Prce of Anarchy n Large Games Mchal Feldman Tel-Avv Unversty, Israel and Mcrosoft Research, USA mfeldman@post.tau.ac.l Tm Roughgarden Stanford Unversty, USA tm@cs.stanford.edu Ncole Immorlca Mcrosoft

More information

Real-Time Systems. Multiprocessor scheduling. Multiprocessor scheduling. Multiprocessor scheduling

Real-Time Systems. Multiprocessor scheduling. Multiprocessor scheduling. Multiprocessor scheduling Real-Tme Systems Multprocessor schedulng Specfcaton Implementaton Verfcaton Multprocessor schedulng -- -- Global schedulng How are tasks assgned to processors? Statc assgnment The processor(s) used for

More information

Revenue in Resource Allocation Games and Applications

Revenue in Resource Allocation Games and Applications Revenue n Resource Allocaton Games and Applcatons by Thanh Ten Nguyen Ths thess/dssertaton document has been electroncally approved by the followng ndvduals: Tardos,Eva (Charperson) Kozen,Dexter Campbell

More information

Kernel Methods and SVMs Extension

Kernel Methods and SVMs Extension Kernel Methods and SVMs Extenson The purpose of ths document s to revew materal covered n Machne Learnng 1 Supervsed Learnng regardng support vector machnes (SVMs). Ths document also provdes a general

More information

CS294 Topics in Algorithmic Game Theory October 11, Lecture 7

CS294 Topics in Algorithmic Game Theory October 11, Lecture 7 CS294 Topcs n Algorthmc Game Theory October 11, 2011 Lecture 7 Lecturer: Chrstos Papadmtrou Scrbe: Wald Krchene, Vjay Kamble 1 Exchange economy We consder an exchange market wth m agents and n goods. Agent

More information

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.265/15.070J Fall 2013 Lecture 12 10/21/2013. Martingale Concentration Inequalities and Applications

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.265/15.070J Fall 2013 Lecture 12 10/21/2013. Martingale Concentration Inequalities and Applications MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.65/15.070J Fall 013 Lecture 1 10/1/013 Martngale Concentraton Inequaltes and Applcatons Content. 1. Exponental concentraton for martngales wth bounded ncrements.

More information

Resource Allocation with a Budget Constraint for Computing Independent Tasks in the Cloud

Resource Allocation with a Budget Constraint for Computing Independent Tasks in the Cloud Resource Allocaton wth a Budget Constrant for Computng Independent Tasks n the Cloud Wemng Sh and Bo Hong School of Electrcal and Computer Engneerng Georga Insttute of Technology, USA 2nd IEEE Internatonal

More information

Computationally Efficient Approximation Mechanisms

Computationally Efficient Approximation Mechanisms CHAPTER 12 Computatonally Effcent Approxmaton Mechansms Ron Lav Abstract We study the ntegraton of game theoretc and computatonal consderatons. In partcular, we study the desgn of computatonally effcent

More information

1 The Mistake Bound Model

1 The Mistake Bound Model 5-850: Advanced Algorthms CMU, Sprng 07 Lecture #: Onlne Learnng and Multplcatve Weghts February 7, 07 Lecturer: Anupam Gupta Scrbe: Bryan Lee,Albert Gu, Eugene Cho he Mstake Bound Model Suppose there

More information

Lecture 4. Instructor: Haipeng Luo

Lecture 4. Instructor: Haipeng Luo Lecture 4 Instructor: Hapeng Luo In the followng lectures, we focus on the expert problem and study more adaptve algorthms. Although Hedge s proven to be worst-case optmal, one may wonder how well t would

More information

Difference Equations

Difference Equations Dfference Equatons c Jan Vrbk 1 Bascs Suppose a sequence of numbers, say a 0,a 1,a,a 3,... s defned by a certan general relatonshp between, say, three consecutve values of the sequence, e.g. a + +3a +1

More information

The Order Relation and Trace Inequalities for. Hermitian Operators

The Order Relation and Trace Inequalities for. Hermitian Operators Internatonal Mathematcal Forum, Vol 3, 08, no, 507-57 HIKARI Ltd, wwwm-hkarcom https://doorg/0988/mf088055 The Order Relaton and Trace Inequaltes for Hermtan Operators Y Huang School of Informaton Scence

More information

Stanford University CS359G: Graph Partitioning and Expanders Handout 4 Luca Trevisan January 13, 2011

Stanford University CS359G: Graph Partitioning and Expanders Handout 4 Luca Trevisan January 13, 2011 Stanford Unversty CS359G: Graph Parttonng and Expanders Handout 4 Luca Trevsan January 3, 0 Lecture 4 In whch we prove the dffcult drecton of Cheeger s nequalty. As n the past lectures, consder an undrected

More information

MMA and GCMMA two methods for nonlinear optimization

MMA and GCMMA two methods for nonlinear optimization MMA and GCMMA two methods for nonlnear optmzaton Krster Svanberg Optmzaton and Systems Theory, KTH, Stockholm, Sweden. krlle@math.kth.se Ths note descrbes the algorthms used n the author s 2007 mplementatons

More information

A Lower Bound on the Competitive Ratio of Truthful Auctions

A Lower Bound on the Competitive Ratio of Truthful Auctions A Lower Bound on the Compettve Rato of Truthful Auctons Andrew V Goldberg 1, Jason D Hartlne 1, Anna R Karln 2, and Mchael Saks 3 1 Mcrosoft Research, SVC/5, 1065 La Avenda, Mountan Vew, CA 94043 {goldberg,hartlne}@mcrosoftcom

More information

Games of Threats. Elon Kohlberg Abraham Neyman. Working Paper

Games of Threats. Elon Kohlberg Abraham Neyman. Working Paper Games of Threats Elon Kohlberg Abraham Neyman Workng Paper 18-023 Games of Threats Elon Kohlberg Harvard Busness School Abraham Neyman The Hebrew Unversty of Jerusalem Workng Paper 18-023 Copyrght 2017

More information

Inner Product. Euclidean Space. Orthonormal Basis. Orthogonal

Inner Product. Euclidean Space. Orthonormal Basis. Orthogonal Inner Product Defnton 1 () A Eucldean space s a fnte-dmensonal vector space over the reals R, wth an nner product,. Defnton 2 (Inner Product) An nner product, on a real vector space X s a symmetrc, blnear,

More information

Algorithmic Game Theory. Edited by Noam Nisan, Tim Roughgarden, Éva Tardos, and Vijay Vazirani

Algorithmic Game Theory. Edited by Noam Nisan, Tim Roughgarden, Éva Tardos, and Vijay Vazirani Algorthmc Game Theory Edted by Noam Nsan, Tm Roughgarden, Éva Tardos, and Vjay Vazran Contents 1 Computatonally-Effcent Approxmaton Mechansms R. Lav page 4 3 1 Computatonally-Effcent Approxmaton Mechansms

More information

Online Appendix. t=1 (p t w)q t. Then the first order condition shows that

Online Appendix. t=1 (p t w)q t. Then the first order condition shows that Artcle forthcomng to ; manuscrpt no (Please, provde the manuscrpt number!) 1 Onlne Appendx Appendx E: Proofs Proof of Proposton 1 Frst we derve the equlbrum when the manufacturer does not vertcally ntegrate

More information

Lecture 12: Discrete Laplacian

Lecture 12: Discrete Laplacian Lecture 12: Dscrete Laplacan Scrbe: Tanye Lu Our goal s to come up wth a dscrete verson of Laplacan operator for trangulated surfaces, so that we can use t n practce to solve related problems We are mostly

More information

Ex post implementation in environments with private goods

Ex post implementation in environments with private goods Theoretcal Economcs 1 (2006), 369 393 1555-7561/20060369 Ex post mplementaton n envronments wth prvate goods SUSHIL BIKHCHANDANI Anderson School of Management, Unversty of Calforna, Los Angeles We prove

More information

Generalized Linear Methods

Generalized Linear Methods Generalzed Lnear Methods 1 Introducton In the Ensemble Methods the general dea s that usng a combnaton of several weak learner one could make a better learner. More formally, assume that we have a set

More information

Lecture Space-Bounded Derandomization

Lecture Space-Bounded Derandomization Notes on Complexty Theory Last updated: October, 2008 Jonathan Katz Lecture Space-Bounded Derandomzaton 1 Space-Bounded Derandomzaton We now dscuss derandomzaton of space-bounded algorthms. Here non-trval

More information

Abstract Single Crossing and the Value Dimension

Abstract Single Crossing and the Value Dimension Abstract Sngle Crossng and the Value Dmenson Davd Rahman September 24, 2007 Abstract When auctonng an ndvsble good wthout consumpton externaltes, abstract sngle crossng s necessary and suffcent to mplement

More information

Winter 2008 CS567 Stochastic Linear/Integer Programming Guest Lecturer: Xu, Huan

Winter 2008 CS567 Stochastic Linear/Integer Programming Guest Lecturer: Xu, Huan Wnter 2008 CS567 Stochastc Lnear/Integer Programmng Guest Lecturer: Xu, Huan Class 2: More Modelng Examples 1 Capacty Expanson Capacty expanson models optmal choces of the tmng and levels of nvestments

More information

Estimation: Part 2. Chapter GREG estimation

Estimation: Part 2. Chapter GREG estimation Chapter 9 Estmaton: Part 2 9. GREG estmaton In Chapter 8, we have seen that the regresson estmator s an effcent estmator when there s a lnear relatonshp between y and x. In ths chapter, we generalzed the

More information

Online Appendix: Reciprocity with Many Goods

Online Appendix: Reciprocity with Many Goods T D T A : O A Kyle Bagwell Stanford Unversty and NBER Robert W. Stager Dartmouth College and NBER March 2016 Abstract Ths onlne Appendx extends to a many-good settng the man features of recprocty emphaszed

More information

arxiv: v1 [cs.gt] 18 Nov 2015

arxiv: v1 [cs.gt] 18 Nov 2015 The Invsble Hand of Dynamc Market Prcng arxv:1511.05646v1 [cs.gt] 18 Nov 2015 Vncent Cohen-Addad vcohen@d.ens.fr Alon Eden alonarden@gmal.com Amos Fat fat@tau.ac.l November 19, 2015 Abstract Mchal Feldman

More information

Solutions HW #2. minimize. Ax = b. Give the dual problem, and make the implicit equality constraints explicit. Solution.

Solutions HW #2. minimize. Ax = b. Give the dual problem, and make the implicit equality constraints explicit. Solution. Solutons HW #2 Dual of general LP. Fnd the dual functon of the LP mnmze subject to c T x Gx h Ax = b. Gve the dual problem, and make the mplct equalty constrants explct. Soluton. 1. The Lagrangan s L(x,

More information

On the Multicriteria Integer Network Flow Problem

On the Multicriteria Integer Network Flow Problem BULGARIAN ACADEMY OF SCIENCES CYBERNETICS AND INFORMATION TECHNOLOGIES Volume 5, No 2 Sofa 2005 On the Multcrtera Integer Network Flow Problem Vassl Vasslev, Marana Nkolova, Maryana Vassleva Insttute of

More information

Bayesian Incentive Compatibility via Fractional Assignments

Bayesian Incentive Compatibility via Fractional Assignments Bayesan Incentve Compatblty va Fractonal Assgnments Xaohu Be Zhy Huang Abstract Very recently, Hartlne and Lucer [14] studed sngleparameter mechansm desgn problems n the Bayesan settng. They proposed a

More information

Sponsored Search Equilibria for Conservative Bidders

Sponsored Search Equilibria for Conservative Bidders Sponsored Search Equlbra for Conservatve Bdders Renato Paes Leme Department of Computer Scence Cornell Unversty Ithaca, NY renatoppl@cs.cornell.edu Éva Tardos Department of Computer Scence Cornell Unversty

More information

Calculation of time complexity (3%)

Calculation of time complexity (3%) Problem 1. (30%) Calculaton of tme complexty (3%) Gven n ctes, usng exhaust search to see every result takes O(n!). Calculaton of tme needed to solve the problem (2%) 40 ctes:40! dfferent tours 40 add

More information

find (x): given element x, return the canonical element of the set containing x;

find (x): given element x, return the canonical element of the set containing x; COS 43 Sprng, 009 Dsjont Set Unon Problem: Mantan a collecton of dsjont sets. Two operatons: fnd the set contanng a gven element; unte two sets nto one (destructvely). Approach: Canoncal element method:

More information

U.C. Berkeley CS278: Computational Complexity Professor Luca Trevisan 2/21/2008. Notes for Lecture 8

U.C. Berkeley CS278: Computational Complexity Professor Luca Trevisan 2/21/2008. Notes for Lecture 8 U.C. Berkeley CS278: Computatonal Complexty Handout N8 Professor Luca Trevsan 2/21/2008 Notes for Lecture 8 1 Undrected Connectvty In the undrected s t connectvty problem (abbrevated ST-UCONN) we are gven

More information

Market structure and Innovation

Market structure and Innovation Market structure and Innovaton Ths presentaton s based on the paper Market structure and Innovaton authored by Glenn C. Loury, publshed n The Quarterly Journal of Economcs, Vol. 93, No.3 ( Aug 1979) I.

More information

THE CHINESE REMAINDER THEOREM. We should thank the Chinese for their wonderful remainder theorem. Glenn Stevens

THE CHINESE REMAINDER THEOREM. We should thank the Chinese for their wonderful remainder theorem. Glenn Stevens THE CHINESE REMAINDER THEOREM KEITH CONRAD We should thank the Chnese for ther wonderful remander theorem. Glenn Stevens 1. Introducton The Chnese remander theorem says we can unquely solve any par of

More information

Simultaneous Optimization of Berth Allocation, Quay Crane Assignment and Quay Crane Scheduling Problems in Container Terminals

Simultaneous Optimization of Berth Allocation, Quay Crane Assignment and Quay Crane Scheduling Problems in Container Terminals Smultaneous Optmzaton of Berth Allocaton, Quay Crane Assgnment and Quay Crane Schedulng Problems n Contaner Termnals Necat Aras, Yavuz Türkoğulları, Z. Caner Taşkın, Kuban Altınel Abstract In ths work,

More information

Lectures - Week 4 Matrix norms, Conditioning, Vector Spaces, Linear Independence, Spanning sets and Basis, Null space and Range of a Matrix

Lectures - Week 4 Matrix norms, Conditioning, Vector Spaces, Linear Independence, Spanning sets and Basis, Null space and Range of a Matrix Lectures - Week 4 Matrx norms, Condtonng, Vector Spaces, Lnear Independence, Spannng sets and Bass, Null space and Range of a Matrx Matrx Norms Now we turn to assocatng a number to each matrx. We could

More information

6.854J / J Advanced Algorithms Fall 2008

6.854J / J Advanced Algorithms Fall 2008 MIT OpenCourseWare http://ocw.mt.edu 6.854J / 18.415J Advanced Algorthms Fall 2008 For nformaton about ctng these materals or our Terms of Use, vst: http://ocw.mt.edu/terms. 18.415/6.854 Advanced Algorthms

More information

Chapter 5. Solution of System of Linear Equations. Module No. 6. Solution of Inconsistent and Ill Conditioned Systems

Chapter 5. Solution of System of Linear Equations. Module No. 6. Solution of Inconsistent and Ill Conditioned Systems Numercal Analyss by Dr. Anta Pal Assstant Professor Department of Mathematcs Natonal Insttute of Technology Durgapur Durgapur-713209 emal: anta.bue@gmal.com 1 . Chapter 5 Soluton of System of Lnear Equatons

More information

Lecture 10 Support Vector Machines II

Lecture 10 Support Vector Machines II Lecture 10 Support Vector Machnes II 22 February 2016 Taylor B. Arnold Yale Statstcs STAT 365/665 1/28 Notes: Problem 3 s posted and due ths upcomng Frday There was an early bug n the fake-test data; fxed

More information

Non-bossy Single Object Auctions

Non-bossy Single Object Auctions Non-bossy Sngle Object Auctons Debass Mshra and Abdul Quadr January 3, 2014 Abstract We study determnstc sngle object auctons n the prvate values envronment. We show that an allocaton rule s mplementable

More information

Lecture 4: November 17, Part 1 Single Buffer Management

Lecture 4: November 17, Part 1 Single Buffer Management Lecturer: Ad Rosén Algorthms for the anagement of Networs Fall 2003-2004 Lecture 4: November 7, 2003 Scrbe: Guy Grebla Part Sngle Buffer anagement In the prevous lecture we taled about the Combned Input

More information

Introduction. 1. The Model

Introduction. 1. The Model H23, Q5 Introducton In the feld of polluton regulaton the problems stemmng from the asymmetry of nformaton between the regulator and the pollutng frms have been thoroughly studed. The semnal works by Wetzman

More information

Lecture 2: Gram-Schmidt Vectors and the LLL Algorithm

Lecture 2: Gram-Schmidt Vectors and the LLL Algorithm NYU, Fall 2016 Lattces Mn Course Lecture 2: Gram-Schmdt Vectors and the LLL Algorthm Lecturer: Noah Stephens-Davdowtz 2.1 The Shortest Vector Problem In our last lecture, we consdered short solutons to

More information

NUMERICAL DIFFERENTIATION

NUMERICAL DIFFERENTIATION NUMERICAL DIFFERENTIATION 1 Introducton Dfferentaton s a method to compute the rate at whch a dependent output y changes wth respect to the change n the ndependent nput x. Ths rate of change s called the

More information

Axiomatizations of Pareto Equilibria in Multicriteria Games

Axiomatizations of Pareto Equilibria in Multicriteria Games ames and Economc Behavor 28, 146154 1999. Artcle ID game.1998.0680, avalable onlne at http:www.dealbrary.com on Axomatzatons of Pareto Equlbra n Multcrtera ames Mark Voorneveld,* Dres Vermeulen, and Peter

More information

Eigenvalues of Random Graphs

Eigenvalues of Random Graphs Spectral Graph Theory Lecture 2 Egenvalues of Random Graphs Danel A. Spelman November 4, 202 2. Introducton In ths lecture, we consder a random graph on n vertces n whch each edge s chosen to be n the

More information

Infinitely Split Nash Equilibrium Problems in Repeated Games

Infinitely Split Nash Equilibrium Problems in Repeated Games Infntely Splt ash Equlbrum Problems n Repeated Games Jnlu L Department of Mathematcs Shawnee State Unversty Portsmouth, Oho 4566 USA Abstract In ths paper, we ntroduce the concept of nfntely splt ash equlbrum

More information

On Random Sampling Auctions for Digital Goods

On Random Sampling Auctions for Digital Goods On Random Samplng Auctons for Dgtal Goods Saeed Alae Department of Computer Scence Unversty of Maryland College Park, MD 07 alae@cs.umd.edu Azarakhsh Malekan Department of Computer Scence Unversty of Maryland

More information

EFFICIENCY OF MECHANISMS IN COMPLEX MARKETS

EFFICIENCY OF MECHANISMS IN COMPLEX MARKETS EFFICIENCY OF MECHANISMS IN COMPLEX MARKETS A Dssertaton Presented to the Faculty of the Graduate School of Cornell Unversty n Partal Fulfllment of the Requrements for the Degree of Doctor of Phlosophy

More information

An Interactive Optimisation Tool for Allocation Problems

An Interactive Optimisation Tool for Allocation Problems An Interactve Optmsaton ool for Allocaton Problems Fredr Bonäs, Joam Westerlund and apo Westerlund Process Desgn Laboratory, Faculty of echnology, Åbo Aadem Unversty, uru 20500, Fnland hs paper presents

More information

Edge Isoperimetric Inequalities

Edge Isoperimetric Inequalities November 7, 2005 Ross M. Rchardson Edge Isopermetrc Inequaltes 1 Four Questons Recall that n the last lecture we looked at the problem of sopermetrc nequaltes n the hypercube, Q n. Our noton of boundary

More information

U.C. Berkeley CS294: Spectral Methods and Expanders Handout 8 Luca Trevisan February 17, 2016

U.C. Berkeley CS294: Spectral Methods and Expanders Handout 8 Luca Trevisan February 17, 2016 U.C. Berkeley CS94: Spectral Methods and Expanders Handout 8 Luca Trevsan February 7, 06 Lecture 8: Spectral Algorthms Wrap-up In whch we talk about even more generalzatons of Cheeger s nequaltes, and

More information

Global Sensitivity. Tuesday 20 th February, 2018

Global Sensitivity. Tuesday 20 th February, 2018 Global Senstvty Tuesday 2 th February, 28 ) Local Senstvty Most senstvty analyses [] are based on local estmates of senstvty, typcally by expandng the response n a Taylor seres about some specfc values

More information

Approximate Smallest Enclosing Balls

Approximate Smallest Enclosing Balls Chapter 5 Approxmate Smallest Enclosng Balls 5. Boundng Volumes A boundng volume for a set S R d s a superset of S wth a smple shape, for example a box, a ball, or an ellpsod. Fgure 5.: Boundng boxes Q(P

More information

Min Cut, Fast Cut, Polynomial Identities

Min Cut, Fast Cut, Polynomial Identities Randomzed Algorthms, Summer 016 Mn Cut, Fast Cut, Polynomal Identtes Instructor: Thomas Kesselhem and Kurt Mehlhorn 1 Mn Cuts n Graphs Lecture (5 pages) Throughout ths secton, G = (V, E) s a mult-graph.

More information

Lecture 4: Universal Hash Functions/Streaming Cont d

Lecture 4: Universal Hash Functions/Streaming Cont d CSE 5: Desgn and Analyss of Algorthms I Sprng 06 Lecture 4: Unversal Hash Functons/Streamng Cont d Lecturer: Shayan Oves Gharan Aprl 6th Scrbe: Jacob Schreber Dsclamer: These notes have not been subjected

More information

Finding Dense Subgraphs in G(n, 1/2)

Finding Dense Subgraphs in G(n, 1/2) Fndng Dense Subgraphs n Gn, 1/ Atsh Das Sarma 1, Amt Deshpande, and Rav Kannan 1 Georga Insttute of Technology,atsh@cc.gatech.edu Mcrosoft Research-Bangalore,amtdesh,annan@mcrosoft.com Abstract. Fndng

More information

Tornado and Luby Transform Codes. Ashish Khisti Presentation October 22, 2003

Tornado and Luby Transform Codes. Ashish Khisti Presentation October 22, 2003 Tornado and Luby Transform Codes Ashsh Khst 6.454 Presentaton October 22, 2003 Background: Erasure Channel Elas[956] studed the Erasure Channel β x x β β x 2 m x 2 k? Capacty of Noseless Erasure Channel

More information

Linear Approximation with Regularization and Moving Least Squares

Linear Approximation with Regularization and Moving Least Squares Lnear Approxmaton wth Regularzaton and Movng Least Squares Igor Grešovn May 007 Revson 4.6 (Revson : March 004). 5 4 3 0.5 3 3.5 4 Contents: Lnear Fttng...4. Weghted Least Squares n Functon Approxmaton...

More information

1 Derivation of Rate Equations from Single-Cell Conductance (Hodgkin-Huxley-like) Equations

1 Derivation of Rate Equations from Single-Cell Conductance (Hodgkin-Huxley-like) Equations Physcs 171/271 -Davd Klenfeld - Fall 2005 (revsed Wnter 2011) 1 Dervaton of Rate Equatons from Sngle-Cell Conductance (Hodgkn-Huxley-lke) Equatons We consder a network of many neurons, each of whch obeys

More information