Table Of Content

Unit Vector Games Rahul Savani∗ Bernhard von Stengel† February 15, 2016 6 Publishedin: InternationalJournalofEconomicTheory12(2016),7–27. doi: 10.1111/ijet.12077 1 0 2 Abstract b e McLennan and Tourky (2010) showed that “imitation games” provide a new view of the F computation of Nash equilibria of bimatrix games with the Lemke–Howson algorithm. In 4 animitationgame,thepayoffmatrixofoneoftheplayersistheidentitymatrix. Westudy 1 the more general “unit vector games”, which are already known, where the payoff ma- ] trix of one player is composed of unit vectors. Our main application is a simplification T G of the construction by Savani and von Stengel (2006) of bimatrix games where two basic . equilibrium-finding algorithms take exponentially many steps: the Lemke–Howson algo- s c rithm,andsupportenumeration. [ 3 Keywords: bimatrixgame,Nashequilibriumcomputation,imitationgame,Lemke–Howsonalgorithm, v 3 unitvectorgame 4 2 2 1 Introduction 0 . 1 0 A bimatrix game is a two-player game in strategic form. The Nash equilibria of a bimatrix 5 game correspond to pairs of vertices of two polyhedra derived from the payoff matrices. These 1 : vertex pairs have to be “completely labeled”, which expresses the equilibrium condition that v i every pure strategy of a player (represented by a “label”) is either a best response to the other X player’smixedstrategyorplayedwithprobabilityzero. r a This polyhedral view gives rise to algorithms that compute a Nash equilibrium. A classical method is the algorithm by Lemke and Howson (1964) which follows a path of “almost com- pletelylabeled”polytopeedgesthatterminatesatNashequilibrium. TheLemke–Howson(LH) algorithm has been one inspiration for the complexity class PPAD defined by Papadimitriou (1994) of computational problems defined by such path-following arguments, which includes more general equilibrium problems such as the computation of approximate Brouwer fixed ∗Department of Computer Science, University of Liverpool, Liverpool L69 3BX, United Kingdom. Email: [email protected] †Department of Mathematics, London School of Economics, London WC2A 2AE, United Kingdom. Email: [email protected] 1 points. An important result proved by Chen and Deng (2006) and Daskalakis, Goldberg, and Papadimitriou (2009) states that every problem in the class PPAD can be reduced to finding a Nash equilibrium of a bimatrix game, which makes this problem “PPAD-complete”. (The problem of finding the Nash equilibrium at the end of a specific path is a much harder, namely PSPACE-complete,seeGoldberg,Papadimitriou,andSavani2013.) If an algorithm takes exponentially many steps (measured in the size of its input) for certain problem instances, these are considered “hard” instances for the algorithm. Savani and von Stengel (2006) constructed bimatrix games that are hard instances for the LH algorithm. Their construction uses “dual cyclic polytopes” which have a well-known vertex structure for any dimension and number of linear inequalities. Morris (1994) used similarly labeled dual cyclic polytopes where all “Lemke paths” are exponentially long. A Lemke path is related to the path computedbytheLHalgorithm,butisdefinedonasinglepolytopethatdoesnothaveaproduct structure corresponding to a bimatrix game. The completely labeled vertex found by a Lemke path can be interpreted as a symmetric equilibrium of a symmetric bimatrix game. However, as in the example in Figure 4 below, such a symmetric game may also have nonsymmetric equilibriawhichhereareeasytocompute,sothattheresultbyMorris(1994)seemedunsuitable todescribegamesthatarehardtosolvewiththeLHalgorithm. The “imitation games” defined by McLennan and Tourky (2010) changed this picture. In an imitationgame,thepayoffmatrixofoneoftheplayersistheidentitymatrix. Themixedstrategy ofthatplayerinanyNashequilibriumoftheimitationgamecorrespondsexactlytoasymmetric equilibriumofthesymmetricgamedefinedbythepayoffmatrixoftheotherplayer. Inthatway, an algorithm that finds a Nash equilibrium of a bimatrix game can be used to find a symmetric Nash equilibrium of a symmetric game. (The converse statement that a bimatrix game can be “symmetrized”, see Proposition 2 below, is an earlier folklore result stated for zero-sum games byGale,Kuhn,andTucker1950.) In one sense the two-polytope construction of Savani and von Stengel (2006) was overly com- plicated: the imitation games by McLennan and Tourky (2010) provide a simple and elegant waytoturnthesingle-polytopeconstructionofMorris(1994)intoexponentially-longLHpaths forbimatrixgames. Inanothersense,theconstructionofSavaniandvonStengelwasnotredun- dant. Namely,thesquareimitationgamesobtainedfromMorris(1994)haveasinglecompletely mixedequilibriumthatiseasilycomputedbyequatingallpayoffsforallpurestrategies. Savani and von Stengel (2006) extended their construction of square games with long LH paths (and a single completely mixed equilibrium) to non-square games that are simultaneously hard for theLHalgorithmand“supportenumeration”,whichisanothernaturalandsimplealgorithmfor finding equilibria. The support of a mixed strategy is the set of pure strategies that are played withpositiveprobability. Givenapairofsupportsofequalsize,themixedstrategyprobabilities arefoundbyequatingallpayoffsfortheotherplayer’ssupport,whichthenhavetobecompared withpayoffsoutsidethesupporttoestablishtheequilibriumproperty(seeDickhautandKaplan 1991). In this paper, we extend the idea of imitation games to games where one payoff matrix is arbitrary and the other is a set of unit vectors. We call these unit vector games. An imitation game is an example of a unit vector game, where the unit vectors form an identity matrix. The mainresultofthispaperisanapplicationofunitvectorgames: weusethemtoextendMorris’s 2 constructiontoobtainnon-squarebimatrixgamesthatuseonlyonedualcyclicpolytope,rather than the two used by Savani and von Stengel, and which are simultaneously hard both for the LHalgorithmandsupportenumeration. Thisresult(Theorem11)wasfirstdescribedbySavani (2006,Section3.8). Before presenting this construction in Section 3, we introduce in Section 2 the required back- groundonlabeledbestresponsepolytopesforbimatrixgames,inanaccessiblepresentationdue to Shapley (1974) that we think every game theorist should know. We define unit vector games and the use of imitation games, and their relationships to the LH algorithm. We will make the casethatunitvectorgamesprovideageneralandsimplewaytoconstructbimatrixgamesusing asinglelabeledpolytope. To our knowledge, unit vector games were first defined and used by Balthasar (2009, Lemma 4.10) in a different context, namely in order to prove that a symmetric equilibrium of a nondegenerate symmetric game that has positive “symmetric index” can be made the unique symmetric equilibrium of a larger symmetric game by adding suitable strategies (Balthasar 2009, Theorem4.1). 2 Unit vector games Inthissection,wefirstdescribeinSection2.1howlabeledpolyhedracapturethe“best-response regions” of mixed strategies where a particular pure strategy of the other player is a best response, and how these are used to identify Nash equilibria. In Section 2.2 we introduce unit vector games, whose equilibria correspond to completely labeled vertices of a single labeled polytope. In Section 2.3 we discuss the role of imitation games for symmetric games and their symmetric equilibria. Finally, in Section 2.4, we show how Lemke paths defined for single la- beledpolytopesare“projections”ofseeminglymoregeneralLHpathsinthecaseofunitvector games(Theorem5). 2.1 Nashequilibriaofbimatrixgamesandpolytopes Consider an m×n bimatrix game (A,B). We describe a geometric-combinatorial “labeling” method, due to Shapley (1974), that allows an easy identification of the Nash equilibria of the game. Ithasanequivalentdescriptionintermsofpolytopesderivedfromthepayoffmatrices. Let 0 betheall-zerovectorandlet 1 betheall-onevectorofappropriatedimension. Allvectors are column vectors and C(cid:62) is the transpose of any matrix C, so 1(cid:62) is the all-one row vector. Let X andY bethemixed-strategysimplicesofthetwoplayers, X ={x∈Rm |x≥0, 1(cid:62)x=1}, Y ={y∈Rn |y≥0, 1(cid:62)y=1}. (1) Itisconvenienttoidentifythe m+n purestrategiesofthetwoplayersbyseparatelabelswhere the labels 1,...,m denote the m pure strategies of the row player 1 and the labels m+1,..., m+n denotethe n purestrategiesofthecolumnplayer2. Consider mixed strategies x∈X and y∈Y. We say that x has label m+ j for 1≤ j ≤n if j is a pure best response of player 2 to x. Similarly, y has label i for 1≤i≤m if i is a pure best responseofplayer1to y. Inaddition,wesaythat x haslabel i for 1≤i≤m if x =0,andthat i 3 y haslabel m+ j for 1≤ j≤n if y =0. Thatis,amixedstrategysuchas x haslabel i (oneof j theplayer’sownpurestrategies)if i isnotplayed. In a Nash equilibrium, every pure strategy that is played with positive probability is a best response to the other player’s mixed strategy. In other words, if a pure strategy is not a best response, it is played with probability zero. Hence, a mixed strategy pair (x,y) is a Nash equilibriumifandonlyifeverylabelin {1,...,m+n} appearsasalabelof x orof y. TheNash equilibria are therefore exactly those pairs (x,y) in X ×Y that are completely labeled in this sense. Asanexample,considerthe 3×3 game (A,B) with     1 0 0 0 2 4 A=0 1 0, B=3 2 0. (2) 0 0 1 0 2 0 The labels 1,2,3 represent the pure strategies of player 1and 4,5,6 those of player 2. Figure 1 shows X andY with theselabels shown ascircled numbers. The interiors ofthese triangles are coveredbybest-responseregionslabeledbytheotherplayer’spurestrategies,whichareclosed polyhedral sets where the respective pure strategy is a best response. For example, the best- response region in Y with label 1 is the set of those (y ,y ,y ) such that y ≥y and y ≥y , 1 2 3 1 2 1 3 due to the particularly simple form of A in (2). The outsides of X and Y are labeled with the players’ own pure strategies where these are not played. These outside facets are opposite to the vertex where only that pure strategy is played; for example, label 1 is the label of the facet of X opposite to the vertex (1,0,0). In Figure 1 there is only one pair (x,y) that is completely labeled,namely x=(1,2,0) withlabels 3,4,5 and y=(1,1,0) withlabels 1,2,6,sothisisthe 3 3 2 2 onlyNashequilibriumofthegame. (0,0,1) (0,0,1) 4 2 1 3 5 5 1 2 6 4 x y (1,0,0) 3 (0,1,0) (1,0,0) 6 (0,1,0) Figure1 Labeledmixedstrategysets X andY forthegame(2). The subdivision of X and Y into best-response regions is most easily seen with the help of the “upper envelope” of the payoffs to the other player, which are defined by the following polyhedra. Let P={(x,v)∈X×R|B(cid:62)x≤1v}, Q={(y,u)∈Y ×R|Ay≤1u}. (3) Fortheexample(2),theinequalities B(cid:62)x≤1v statethat 3x ≤v, 2x +2x +2x ≤v, 4x ≤v, 2 1 2 3 1 which say that v is at least the best-response payoff to player 2. If one of these inequalities 4 is tight (holds as equality), then v is exactly the best-response payoff to player 2. The left- handdiagraminFigure2showsthese“best-responsefacets”of P,andtheirprojectionto X by ignoringthepayoffvariable v,whichdefinesthesubdivisionof X intobest-responseregionsas intheleft-handdiagraminFigure1. x 3 4 v 3 2 1 4 0 (0,0,1) 4 5 3 5 3 (0,0,0) 2 2 6 1 6 1 4 4 0 0 (1,0,0) (0,1,0) x1 x2 Figure2 Best-response facets of the polyhedron P in (3), and the polytope P in (4), for the gamein(2). Throughoutthispaper,assume (withoutlossofgenerality)that A and B(cid:62) arenon-negativeand have no zero column. Then v and u in B(cid:62)x≤1v and Ay≤1u are always positive. By dividing theseinequalitiesby v and u,respectively,andwriting x insteadof x/v and y insteadof y /u, i i j j thepolyhedra P and Q arereplacedby P and Q, P={x∈Rm |x≥0, B(cid:62)x≤1}, Q={y∈Rn |Ay≤1, y≥0}, (4) whichareboundedandthereforepolytopes. For B in(2), P isshownontherightinFigure2. Both polytopes P and Q in (4) are defined by m+n inequalities that correspond to the pure strategiesoftheplayer,whichwehavedenotedbythelabels 1,...,m+n. Wecannowidentify the labels, as pure best responses of the other player, or unplayed own pure strategies, as tight inequalities in either polytope. That is, a point x in P has label k if the kth inequality in P is tight,thatis,if x =0 for 1≤k≤m or (B(cid:62)x) =1 for m+1≤k≤m+n. Similarly, y in Q k k−m haslabel k if (Ay) =1 for 1≤k≤m or y =0 for m+1≤k≤m+n. Then (x,y) in P×Q k k−m is completely labeled if x and y together have all labels in {1,...,m+n}. With the exception of (0,0),thesecompletelylabeledpointsof P×Q represent(afterrescalingtobecomepairsof mixedstrategies)exactlytheNashequilibriaofthegame (A,B). Thepair (x,y) in P×Q iscompletelylabeledif x =0or(Ay) =1foralli=1,...,m, y =0or(B(cid:62)x) =1forall j=1,...,n. (5) i i j j Because x, 1−Ay, y, and 1−B(cid:62)x are all non-negative, the complementarity condition (5) can alsobestatedastheorthogonalitycondition x(cid:62)(1−Ay)=0, y(cid:62)(1−B(cid:62)x)=0. (6) The characterization of Nash equilibria as completely labeled pairs (x,y) holds for arbitrary bimatrix games. For considering algorithms, it is useful to assume that the game is nondegenerate in the sense that no point in P has more than m labels, and no point in Q has more than 5 n labels. Clearly, for a nondegenerate game, in an equilibrium (x,y) each label appears exactly onceeitherasalabelof x orof y. Nondegeneracy is equivalent to the condition that the number of pure best responses against a mixed strategy is never larger than the size of the support of that mixed strategy. It implies that P is a simple polytope in the sense that no point of P lies on more than m facets, and similarly that Q isasimplepolytope. Afacetisobtainedbyturningoneoftheinequalitiesthatdefinethe polytope into an equality, provided that the inequality is irredundant, that is, cannot be omitted without changing the polytope. A redundant inequality in the definition of P and Q may also give rise to a degeneracy if it corresponds to a pure strategy that is weakly (but not strictly) dominated by, or payoff equivalent to, a mixture of other strategies. For a detailed discussion ofdegeneracyseevonStengel(2002). 2.2 Unitvectorgamesandasinglelabeledpolytope The components of the kth unit vector e are 0 except for the kth component, which is 1. In k an m×n unit vector game (A,B), every column of A is a unit vector in Rm. The matrix B is arbitrary,andwithoutlossofgenerality B(cid:62) isnon-negativeandhasnozerocolumn. In this subsection, we consider such a unit vector game (A,B). Let the jth column of A be the unit vector e , for 1≤ j ≤n. Then the sequence (cid:96)(1),...,(cid:96)(n) together with the payoff (cid:96)(j) matrix B completelyspecifiesthegame. Forthisgame,thepolytope Q in(4)hasaveryspecialstructure. For 1≤i≤m,let N ={ j|(cid:96)(j)=i, 1≤ j≤n} (7) i sothat N isthesetofthosecolumns j whosebestresponseisrow i. Thesesets N arepairwise i i disjoint,andtheirunionis {1,...,n}. Thenclearly Q={y∈Rn | ∑ y ≤1, 1≤i≤m, y≥0}. (8) j j∈N i That is, except for the order of inequalities, Q is the product of m simplices of the form {z∈ RNi |∑j∈N zj ≤1, z≥0},for 1≤i≤m. Ifeach Ni isasingleton,then,by(7), A isapermuted i identitymatrix, n=m,eachsimplexistheunitinterval,and Q isthe n-dimensionalunitcube. Inanybimatrixgame,thepolytopes PandQin(4)eachhave m+ninequalitiesthatcorrespond to the pure strategies of the two players. Turning the kth inequality into an equality typically definesafacetofthepolytope,whichdefinesthelabel k ofthatfacet, 1≤k≤m+n. Inourunitvectorgame (A,B) wherethe jthcolumnof A istheunitvector e ,for 1≤ j≤n, (cid:96)(j) weintroducethelabeledpolytope P(cid:96), P(cid:96) ={x∈Rm |x≥0, B(cid:62)x≤1}, (9) wherethe m+n inequalitiesof P(cid:96) havethelabels i forthefirst m inequalities x ≥0, 1≤i≤m, i andthe jthinequalityof B(cid:62)x≤1 haslabel (cid:96)(j),for 1≤ j≤n. Thatis, P(cid:96) isjustthepolytope P in (4) except that the last n inequalities are labeled with (cid:96)(1),...,(cid:96)(n), each of which is a number in {1,...,m}. A point x of P(cid:96) is completely labeled if every number in {1,...,m} 6 appears as the label of an inequality that is tight for x. In particular, if P(cid:96) is a simple polytope with one label for each facet, then x is completely labeled if x is a vertex of P(cid:96) so that the m facetsthat x liesontogetherhavealllabels 1,...,m. Thefollowingpropositionshowsthatwiththeselabels, P(cid:96) carriesalltheinformationaboutthe unit vector game, and the polytope Q is not needed. The proposition was first stated in a dual versionbyBalthasar(2009,Lemma4.10),andinessentiallythisformbyVe´ghandvonStengel (2015,Proposition1). ItsproofalsoprovidesthefirststepoftheproofofTheorem5below. Proposition1 Consider a labeled polytope P(cid:96) with labels as described following (9). Then x is a completely labeled point of P(cid:96)−{0} if and only if for some y∈Q the pair (x,y) is (after scaling)aNashequilibriumofthe m×n unitvectorgame (A,B) where A=[e ···e ]. (cid:96)(1) (cid:96)(n) Proof. Let (x,y)∈P×Q−{(0,0)} beaNashequilibrium,soithasalllabelsin {1,...,m+n}. Then x isacompletelylabeledpointof P(cid:96) forthefollowingreason. If x =0 then x haslabel i. i If x >0 then y has label i, that is, (Ay) =1, which requires that for some j we have y >0 i i j andthe jthcolumnof A isequalto e ,thatis, (cid:96)(j)=i. Because y >0,and (x,y) iscompletely i j labeled, x has label m+ j in P, that is, (B(cid:62)x) =1, which means x has label (cid:96)(j)=i in P(cid:96), as j required. Conversely,let x beacompletelylabeledpointof P(cid:96)−{0}. Thenforeach i in {1,...,m} with x >0,label i for x comesfromabindinginequality (B(cid:62)x) =1 withlabel (cid:96)(j)=i,thatis,for i j some j∈N in(7). Let y =1 and y =0 forall h∈N −{j},and dothisforall i with x >0. i j h i i Itiseasytoseethatthepair (x,y) isacompletelylabeledpointof P×Q. The game in (2) is a unit vector game. For this game, the polytope P in (4) is shown on the right in Figure 2, where we have shown only the labels 4,5,6 for the “best response facets”. In addition,thefacetswithlabels 1,2,3 where x =0, x =0, x =0 arethefacets,hiddeninthis 1 2 3 picture, at the back right, back left, and bottom of the polytope, respectively. In the polytope P(cid:96),thelabels 4,5,6 arereplacedby 1,2,3 becausethecorrespondingcolumnsof A aretheunit vectors e ,e ,e . Figure3showsthispolytopeinsuchawaythatthereisonlyonehiddenfacet, 1 2 3 with label 1 where x =0. Apart from the origin 0, the only completely labeled point of P(cid:96) is 1 x asshown,whichispartofaNashequilibrium (x,y) asstatedinProposition1. A polytope like in (9) that has a label for each facet provides a particularly natural way to describeequilibrium-findingalgorithms,asdescribedinSection2.4below. 2.3 Reductionsbetweenequilibriaofbimatrixandunitvectorgames A method that “solves” a unit vector game in the sense of finding one equilibrium, or all equilibria, of the game, can be used to solve an arbitrary bimatrix game. The first step in seeing this is the fact that the equilibria of a bimatrix game correspond to the symmetric equilibria of a suitable symmetric game. This “symmetrization” has been observed for zero-sum games by Gale,Kuhn,andTucker(1950)andseemstobeafolkloreresultforbimatrixgames. Proposition2 Let (A,B) be a bimatrix game, and (x,y)∈P×Q−{(0,0)} in (4). Then (x,y) (suitably scaled) is a Nash equilibrium of (A,B) if and only if (z,z) (suitably scaled) is a symmetricequilibriumof (C,C(cid:62)) with z=(x,y) andC=(cid:0)0 A(cid:1). B(cid:62) 0 7 2 2 3 1 1 x 3 0 Figure3 The polytope P(cid:96) for the unit vector game (2). The hidden facet at the back has label 1,writtenontheleft. Proof. This holds by (6) because (z,z) is an equilibrium of (C,C(cid:62)) if and only if z(cid:54)=0, z≥0, Cz≤1,and z(cid:62)(1−Cz)=0. By Proposition 2, finding an equilibrium of a bimatrix game can be reduced to finding a symmetric equilibrium of a symmetric bimatrix game. The converse follows from the following proposition, due to McLennan and Tourky (2010, Proposition 2.1), with the help of imitation games. Theydefineanimitationgameasan m×m bimatrixgame (A,B) where B istheidentity matrix. Here,wedefineanimitationgameasaspecialunitvectorgame (A,B) where A (rather than B) is the identity matrix I. The reason for this (clearly not very material) change is that this game is completely described by the polytope P(cid:96) in (9), which corresponds to P in (4) andcomparedto Q hasamorenaturaldescriptionbecausethe m inequalities x≥0 withlabels 1,...,m arelistedfirst. Proposition3 Thepair (x,x) isasymmetricNashequilibriumofthesymmetricbimatrixgame (C,C(cid:62)) ifandonlyif (x,y) isaNashequilibriumoftheimitationgame (I,C(cid:62)) forsome y. (0,0,1) (0,0,1) 2 1 5 4 b b 5 2 6 3 4 1 a x a x (1,0,0) 3 (0,1,0) (1,0,0) 6 (0,1,0) Figure4 Labeledmixed-strategysets X andY forthesymmetricgame (C,C(cid:62)) in(10). Asanexample,considerthesymmetricgame (C,C(cid:62)) with     0 3 0 0 2 4 C=2 2 2, C(cid:62) =3 2 0, (10) 4 0 0 0 2 0 8 so that C(cid:62) = B in (2). Figure 4 shows the labeled mixed-strategy simplices X and Y for this game. In addition to the symmetric equilibrium (x,x) where x = (1,2,0), the game has two 3 3 non-symmetric equilibria (a,b) and (b,a) where a = (1,1,0) and b = (0,2,1). A method 2 2 3 3 that just finds a Nash equilibrium of a bimatrix game may not find a symmetric equilibrium when applied to this game, which shows the use of Proposition 3. The corresponding imitation game (I,C(cid:62)) is just (A,B) in (2), which has the unique equilibrium (x,y) where (x,x) is the symmetricequilibriumof (C,C(cid:62)). (0,0,1) (0,0,1) 2 1 2 1 2 2 3 3 1 1 x x (1,0,0) 3 (0,1,0) (1,0,0) 3 (0,1,0) Figure5 (Left) Best-responseregions for identifying symmetric equilibria. (Right) Degener- atesymmetricgame(11)withauniquesymmetricequilibrium. Theleft-handdiagraminFigure5showsthemixedstrategysimplex X subdividedintoregions of pure best responses against the mixed strategy itself, which corresponds to the polytope P(cid:96) inFigure3. The(inthiscaseunique)symmetricequilibriumisthecompletelylabeledpoint x. Theright-handdiagraminFigure5showsthissubdivisionofX foranothergame(C,C(cid:62))where     0 4 0 0 2 4 C=2 2 2, C(cid:62) =4 2 0. (11) 4 0 0 0 2 0 Thisgameisdegeneratebecausethemixedstrategy x=(1,1,0) hasthreepurebestresponses. 2 2 This mixed strategy x also defines the unique symmetric equilibrium (x,x) of this game. How- ever, the corresponding equilibria (x,y) of the imitation game (I,C(cid:62)) are not unique, because duetothedegeneracyanyconvexcombinationof (1,1,0) and (1,1,1) canbechosenfor y,as 2 2 3 3 3 showninFigure6. Hencethereductionbetweensymmetricequilibria(x,x)ofasymmetricgameandNashequilib- ria (x,y) ofthecorrespondingimitationgamestatedinProposition3doesnotpreserveunique- nessifthegameisdegenerate. 2.4 LemkepathsandLemke–Howsonpaths Consider a labeled polytope P(cid:96) as in (9). We assume throughout that P(cid:96) is nondegenerate, that is, no point of P(cid:96) has more than m labels. Therefore, P(cid:96) is a simple polytope, and every tight inequalitydefinesaseparatefacet(wecanomitinequalitiesthatarenevertight),eachofwhich hasalabelin {1,...,m}. Thepath-followingmethodsdescribedinthissectioncanbeextended todegenerategamesandpolytopes;foranexpositionseevonStengel(2002). 9 (0,0,1) (0,0,1) 4 2 1 3 5 5 1 y 2 6 4 x (1,0,0) 3 (0,1,0) (1,0,0) 6 (0,1,0) Figure6 Labeled mixed-strategy sets for the imitation game (I,C(cid:62)) for the degenerate sym- metricgame(11)wheretheequilibria (x,y) arenotunique. A Lemke path is a path that starts at a completely labeled vertex of P(cid:96) such as 0 and ends at another completely labeled vertex. It is defined by choosing one label k in {1,...,m} that is allowed to be missing. After this choice of k, the path proceeds in a unique manner from the starting point. By leaving the facet with label k, a unique edge is traversed whose endpoint is anothervertex,whichliesonanewfacet. Thelabel,say j,ofthatfacet,issaidtobepickedup. Ifthisisthemissinglabel k,thenthepathterminatesatacompletelylabeledvertex. Otherwise, j is clearly duplicate and the next edge is uniquely chosen by leaving the facet that so far had label j, and the process is repeated. The resulting path consists of a sequence of k-almost complementary edges and vertices (so defined by having all labels except possibly k, where k occurs only at the starting point and endpoint of the path). The path cannot revisit a vertex because this would offer a second way to proceed when that vertex is first encountered, which is not the case because P(cid:96) is nondegenerate. Hence, the path terminates at another completely labeled vertex of P(cid:96) (which is a Nash equilibrium of the corresponding unit vector game in Proposition1ifthepathstartsat 0). Figure7showsanexample. 2 2 3 1 1 x 3 0 Figure7 Lemkepathformissinglabel 1 forthepolytopeinFigure3. For a fixed missing label k, every completely labeled vertex of P(cid:96) is a separate endpoint of a Lemke path. Because each path has two endpoints, there is an even number of them, and all of these except 0 are Nash equilibria of the unit vector game, so the number of Nash equilibria is odd. 10