with Infinite Sets of Types and Actions Sequential Equilibria of Multi-Stage Games by Roger B. Myerson and Philip J. Reny* Draft notes October 2011 http://home.uchicago.edu/~preny/papers/bigseqm.pdf essential sequential equilibrium while maintaining general existence. Abstract: We formulate a definition of basic sequential equilibrium for multi-stage games with infinite difficulties of this basic concept and, in light of them, propose the more restrictive definition of type sets and infinite action sets, and we prove its general existence. We then explore several * Department of Economics, University of Chicago 1 type sets and infinite action sets, and prove general existence. Tirole 1991) have been used for infinite games. No general existence. but rigorously defined extensions to infinite games have been lacking. Sequential equilibria were defined for finite games by Kreps-Wilson 1982, Harris-Stinchcombe-Zame 2000 explored definitions with nonstandard analysis. Various formulations of "perfect bayesian eqm" (defined for finite games in Fudenberg-Goal: formulate a definition of sequential equilibrium for multi-stage games with infinite Others should explore alternative definitions. include reasonable equilibria and exclude unreasonable equilibria. Here we present the best definitions that we have been able to find, game (in some sense) but have limits of equilibria that seem wrong. The problem is to define what finite games are good approximations. taking limits of sequential equilibria of finite games that approximate it. but of course this depends on intuitive judgments about what is "reasonable". It is well understood that sequential equilibria of an infinite game can be defined by We must try to define a class of finite approximations that yield limit-equilibria which It is easy to define sequences of finite games that seem to be converging to the infinite 2 T = {possible informatio actions for player i at pe under finite and com ik Ak {1,...,K} periods of the game. Dynamic multi-stage games =(,N,K,A,T,p,,v) = {initial state space}, i N = {players}, finite set. Let L = {(i,k)| iN, k{1,...,K}} = {dated players}. We write ik for (i,k). , including all one-point sets as measurable. for and T There is a finitely additive probabilit are measurable. ik . ikIf is a finite set, then all subs A = A = {possible sequences of actions in the whole game}. hK iN ih The subscript, <k, denotes the projection onto periods before k. For example, A = = {}), = {possible action (A A < k h<k i N i h < 1 = and for aA, a < k h<k i N a i h Any player i's information at any period k is specified by a type function such that, aA, (, ) is a measurable function of . a i k : A is the parti T ik < i k k < k Assume perfect recall: ikL, m<k, ikm i k i m ( (, )) = (, a a ( { | ( ) { t t R a ik <k ikm ik A such that i m ikm ik :T T i m im im ),a ), , aA, and <m }} is measurable in T i m T i k Each player i has a bounded utility function :A such that im v vi(,a) is a measurable function of , aA. Boundi(,a)| , |v (i,,a). i . , , m A a R i i i mmm 3 a[Prob lems of finite suppor t mixtur es in appro ximati ng games] Examp le. (Kuhn) Consid er a zero-su m game in which player 1 choose sa numbe r Play er 1 can guar ante e 1/2 by choo sing agua rante e 1/2 by choo sing f(x) = 1/2 for all x. Lebe sgue integ ral must be 1/2. Play er 1’s payo ff is f(aea ch of play er 1’s finit ely man y avail able actio ns, play er 1’s equil ibriu m payo ff is 0. But in any finit e appr oxi mati on in whic h play er 2 can choo se a func tion that is zero at One soluti on is to replac e ordin ary action sets with mixed action sets. Algeb ras (close d under finite and compl ement s) of meas urabl e subset s are specif ied r each A a l s o f o 1 , including all one-point sets as measurable. 1 1 ). in [0,1], and player 2 chooses a uniformly continuous function, f, from [0,1] into itself from [0,1] and whose player 2 can Each (, ) and each vi(,a) is assumed jointly measurable in (,a). ik a <k Let ~ = [0,1]). In the state ~ = ), nature draws from the given p, and, ikL ( (,(~ ik for each ik, draws ~ [ 0 , 1 ] i n d e p e n d e n t l y f r o m L e b e s g u e m e a s u r e . ik ~This define sa new distri butio n p on ~. No player obser ves any of the new ~ i k i k i k L e t à = : . ik { [ } be m player ik’s set of mixed 0 actions. e , a 1 s ] u r a b A l e m a p s ã ) ik ik <k (~,ã) = v Then ṽ i ) = ( <k), ik ã(,ã( meas urable in ~ for (,ã(~) ) and ~ each profil e of mixed action s ã. A l l o u r i , < k ( ~ , ã i a n a l y s i s c o u l d b e d o n e f o r t h i s m o d e l w i t h à , ~ , p ~ i n s t e a d o f A , , p . 4 Problems of spurious signaling in approximating games Example. = and chooses Nature chooses {-1,1}, prob. ½ each. Player 1 observes t 1 [-1,1]. Player 2 observes t = a and chooses a a 1 {-1,1}. Payoffs are as follows. 1 = -1 a (0,2) 2(1,0) 2 a = 1 (0,0) (1,1) No observation t =1 2 2 = -1 2 2 = 1 since player 1 always [-1,1] should lead player 2 to choose a 2 prefers a =1 = 1 and the set of signals player 1 can send does not depend on . But any approximation where 1’s finitely many actions are, e.g., all rational when t To avoid such spurious signaling, the finite set of actions available to a player should not 1 = 1 when t is any rational number. and are all irrational when t= -1 reveals , so a1 2 2 depend on his informational type. depend on his type and in which the action aBut now consider any finite approximation = x > 0 is in which player 1’s finite action set does not a 1 = -x is not. Then t 2 = x implies = 1, leading player 2 to choose a = 1. 1 2 , an ik type-space partitions become arbitrarily fine.we will expand the players’ik 5 finite action sets to fill in the action spaces ATo avoid also the latter spurious signaling, we should limit player 2’s ability to coarsen the players’ information by finitely partitioning each of their type spaces, Tdistinguish her types and allow player 1 any finite subset of his actions. So, we will faster than th Probl ems of spuri ous knowl edge in appro ximat ing game s players inform ation that they do not have in the origina l game. Some approx imatio ns of the state space can change the inform ation structu re and give the first (1's cost to sell some object ) is obser ved only by player 1, Exam ple. includ es two indep enden t unifor m [0,1] rando m variab les: the secon d (2's value of buyin g this object ) is obser ved only by player 2. For any intege rm> 0, we might finitel y appro ximat e this game by one in which the state 2m m equally likely pairs of numbers: [0,1]; m-digit decimals in m-digit decimals but the last m digits repeat the first number in the pair. but they represent games in which player 2 knows 1's type. As m, these pairs uniformly fill the unit square in the state space , separately, we avoid giving spurious knowledge to any player. we can keep the original state space and probability distribution p in all finite approximations, avoiding the above information distortion. space inclu the first nu decimal in With finite By partitio ik or the ability to In some games, the ik A convergence of a net of approximations (not just a sequence) that are and all finite partitions of T subsets of A ik distinguish some special subsets of i kT ik finite partition of types are available when considering the properties of the limits. . So, any finite collection of actions and any 6 Observable events and action approximations For any aA and any measurable R , let P(R |a) = p({| T i i i i < i ) }). (, k k k kk k R a (For k=1, we could write |a), ignoring the trivial a =.) i1 P(R The set of observable events for i at k that can have positive probability is Q = {R | R is measurable and a A such that |a) > 0}. ik ik ik ik Tik P(R Let Q = ikikL Q (a disjoint union) denote the set of all events that can be observed with positive probability by some dated player. < 1 . ik An action approximation is any C =when players are restricted to actions in some subsets of the AWe extend this notation to describe observable events that can have positive probability A , a n subset of is a nonem d so CA.ikL ik ik i k For any action approximation C A, let Q Let Q(C) = ik ikL 7 Q (C) = {R T | is measurable and cC such that P(R |c) > 0}. ik ik ik ik R ik Q is the union of all Q(C) over all finite approximations C. probability by some dated player when a oAction approximations are partially ordered by If C C inclusion. o o then Q(C) ). Fact. If C C Q(C o then C to the true action sets A. C Finite approximations of the game An information approximation is any = . (So elements of ik L measurable subsets of Tik Information approximations are partially ordered by the fineness of their partitions. Say ikL that = i k o (If is finer than ikL i k is a finite i k L. i k o i k i k to the true type sets T.) , then is a approximation (,C) of the game G . An information approximation and an action ) in the app of approximation C together define a finite After any history (,c <k ) and is also informed of his past actions that contains ik the c<k ik(,c ik . The la ensures that action-recall is not violated in (,C) when we expand C to fill in the i(<k) action i(<k) = ,c i(<k ) C ) ik {( Let S ik ik space A for any fixed information approximation . i(< the finite set of possib le types s of dated player ik in the fi approximation (,C). Fix . The approximation (,C) will satisfy perfect recall C if: ikL, Let F denot e the set of i m s.t. {(,a)A| finite appro ximat ions with perfec t recall. o Fact. For any information approximation , there is a finer approximation such that (,C ) satisfi es perfec t recall for every C, and so (,C )F C. 8 S i k C i k is the finite set of types of player ik. i(<k ) ) Strategy profiles for finite approximations Here let (,C)F be a given finite approximation of the game and so A strategy profile for the finite approximation (,C) is any = ( i ik i (C , (|s )=1, s k L k s.t. each ik ik ik C_ik :Sikik Let [t ] denote the element of containing t T . ik ik ik ik ik Z (c )],c ( |[ (, ))p(d). L i(<k) c mble Z and cC, let P(Z,c|) = <k ik ik ik i i i i c For any measurable R (,c k k k k C ik T <k , let P(R |) = P({| ) R },c|) . A totally mixed strategy profile for (,C) has ik Any totally mixed for (,C) yields |) > 0, P(R R C (cik |s ik i k Sik. )>0 sik c i , k (C), ikL. ik Q ik For all such Y and c C , define ,cik ,( , )). ,) = P(Y|Rik c ik -ik ik P(Y|R ik P(Y(c),c|R ,) for any YA such that, ik P(Y| ,) = cC ik R With |) > 0, let P(R P(Z,c|R We can exten d this proba bility functi on to define for each cC, ik ik ik ,) = P(Z{|ik (, )R c <k i k },c|)/ P(R |). ik ik Y(c) = {| (,c) Y} is a measu rable subset of . Note. Chang ing the action of i at k does not chang e the proba bility of i's types at k, so P(R |c , ) = P(R |) > ik ik ik ik 0. 9 ik )= Approximate sequential equilibria Here let (,C)F be a given finite approximation of the game, and let be a totally mixed strategy profile for (,C). With the probability function P defined above, we can define i (,c) P(d,c|R v V For any ikL and any cik only in that i chooses action Changing the action of i at k does not change the probability of i's t c P(R |c , ) = P(R |) > 0 , Q (C). ik ik ik ik ik Rik i i(cSo we can similarly define , k ik V period k, given the observable event R at ) to be i's pay ik ik -approximate sequential equilibrium for the finite approximation (,C) For >0, is an (c , iff is a totally mixed strategy for (,C) and ) V(| V s i ik . i Facts. Any finite approximation has an -approximate sequential equilibrium. In finite games with perfect recall, a strategy profile is part of a sequential equilibrium if of a sequence of -approximate sequential equilibria. and only if it is the limit as → 0 10 Assessments Let Y denote the set of all outcome events Y A such that {| (,a)Y} is a measurable subset of , aA. For any action approximation, C, let W(C) = Q(C) ( ikL i Q A k ) (the union of all such W(C)). Let W = Q i ik k (ikL An asse ssme nt is a vect or of cond ition al prob ). (C) Q C i k abilit ies (Y| W) [0,1] Y Y, W W. YW So is in the compact (product topology) set , [0,1] , ik)[0,1], i.e., (Y|Q )[0,1] and ik a ikL,YY,Q (Y|Qik . A ,a Q ik Basic sequential equilibria An assessment is a basic sequential equilibrium iff: for every > 0, for every finite subset F of YW, for every information approximation information approximation such that for every action approximation o, there C is an action approximation C o such that F YW(C), (,C)F and (,C) C ik ik i k o , there is a fin The proo f is base d on Tych onof f's Theo rem on com pact ness of prod ucts of com pact sets. has some -approximate sequential equilibrium such that |(Y|W)P(Y|W,)| , (Y,W)F. in the Theorem. The set of basic sequential equilibria is nonempty and is a closed subset of Y W [0,1] Kelley, 1955, p143. 11 Defining basic sequential equilibria with nets A partially ordered set is directed if for d} any two members there is a member at least as large as both. A net of real numbers is a collection of real numbers indexed by a directed set (D,). A net of real numbers {x is an index s.t. is within e of whenever o d x x e d d d e. Write lim are directe d sets. The set of action d D x . Also, define lim x d D d = lim x ). (infdDo = x d’d d’ d approxi mation s, partiall y ordered by set inclusi on, and the set of inform ation approxi mation s, partiall y ordered by the finenes s of the type-sp ace partitio ns, Then is a basic seque ntial equili brium iff for every finite subset F of YW , there is a [] mappi ng such that lim lim F = 0, |(Y|W) – C:(,C) lim→ 0+ P(Y|W,[,,C])| where each [,,C] is some -approximate sequential equilibrium of the finite ma x (Y,W)F a pproximation (,C). probability by [,,C] for all large enough C. Note. All of the finitely many conditioning events determined by F are given positive 12 Elementary properties of basic sequential equilibria , let Q ik For any dated player ik and any observable event Q I(Q (,a )Q }. ik ) = {(,a)A| ikik <k Let be a basic seque ntial equili brium. Then has the follow ing gener ik al proper ties, for any outco me events Y and Z and any observable events Qik if YZ= then (YZ|Q (Y|Q )[0,1], ik (A|Qik )| (Y| ) = Q (YI(Q Q jm and : Q )+ )= (Z|Q (Y|Q )=0 (probabilities); i i ) (conditional k k support); ) = 1, (|Q ) (finite additivity); i k ik ik (YI( Q ) (I(Q ik ) = (Y|T jm ik jm )= (YI(Q ) (I(Q ), for all ik and jm in L. )|Q ik ) (Bayes consistency). j )|Q jm )|Q)|Qjm m ik ik Bayes consist ency implies that (Y|T So the uncond itional distribu tion on outcom es A for a basic sequent ial equilibr ium can be define d by (Y) = (Y| T The uncon dition al margi nal distrib ution of on is the given prior p: (Z A) = p(Z), for any measu rable Z. Fact (seque ntial ration ality). If is a basic seque ntial equili brium then, ik L, Q Q , a’ A ik ik ik ik , i(,a) v (d,da|Q ,a’ ik i ). k (,a) (d,da|Q The left-hand side integral is player i’s equilibrium payoff conditional on Q ’ side integral is i’s payoff from deviating at date k to a ik . The right-hand . ik ) v i ik 13 ik conditio Finite additi vity of soluti ons strictl y positi ve numb er that is availa ble to him. Exam ple: Consi der a game where the winne r is the player who choos es the smalle st strictl y positi ve numb er. In any finite appro ximati on, let each player choos e the smalle st x) = 1, 1 x>0, but (a in any =0) = 0. 1 1 1 . 2 after observing this a 1 2 1 x>0, (a 2 . 1 1 < |a a x) = 1. ) in {0,1}. =x) = 1 and (0<a1 But multiple equilibria can have any prior P(2 wins) = (a2 By < a 1 <aIn the limit, additivity, a infinitesima additive, onl , 2 would alw before choosing a {1,2}, and p the limit, we consid ering a subnet of finite appro ximati ons in which 1 alway s has smalle r action s <athan 2, we can get an equilibrium with (a12 ) = 0 even though (a =x) = 1 x>0 <a |a 1 1 2 and (a >0)=1. 1 1 In this su outcome cannot be derived by backward induction from 2's limiting strategy. Strategic actions avoid such problems, as would considering subnets where later players' 1 4 a rlier players' (fast inner limits for later C c t i o n s g r o w f a s t e r t h a n e a ). k [Games with perfect information] |>1 for at most one player iN, and }, then a dynamic multi-st If ={0k{1,...,K}, |A ik L, (No two player s make choice s at the same date, and all player s obser ve the past histor y of play.) : { ik 0 ik }A T is one to one. ik <k }, has a basic Theorem. Every dynamic multi-stage game of perfect information with continuous 0 payoffs, compact Hausdorff action sets and one state of nature, i.e., ={ sequential equilibrium, , in which all prior probabilities are 0 or 1. That is, ({ }B){0,1}, 0 BA. o Proof: Fix d (0,1). For every >0 and every information approximation compactness , by and continuity, there is a finer information approximation such that the last may repeat the argument for the second last player after possibly further refining his player has, for each element of his partition, a single action that is -optimal against all histories leading to that partition element. With this strategy fixed for the last player we information partition, etc. Hence, for each >0 there is an information approximation -appro ximate sequent ial equilibr ium in which some outcom e receive s probabi lity at least res ult be ca use the set of bas ic seq ue nti al eq uili bri a is clo sed in the pro du ct top olo gy. 1d, so tha t the pro ba bili ty of an y B A is in [0, d] [1d ,1]. Let tin gd 0 yie lds the 15 that is finer than o and is such that C large enough the game (,C) admits an Strat egic entan gleme nt (Milg romWebe r 1985) Exam ple: is unifor m [0,1], obser ved by both player s1 and 2, who then simult aneou sly play a battle-of-sexes game where = ={1,2}, A A 1 2 v =i, = 1 if a 1 2i, and = 0 if . a2i ) = 2 if a,a12 v i 1 a v i Consider finite-approximation equilibria such that, for a large integer m, they do a 1 2 1 =a =a (, 1 2 a if the m'th digit of is odd, = =2 if the m'th digit of is even. a a As m , these equili =1 = a 2 bria conve rge to a limit where the player s rando mize betwe en action s (1,1) and (2,2), each with proba bility 1/2, indep enden tly of . In the limit, the player s' action s are not indep enden t given in any positi ve interv al. They are correl ated by comm only obser ved infinit esima l detail s of . Given in any positiv e interva l, player 1's expect ed payoff is 1.5 in eqm, and 1's condit ional payof f from deviat ing to a1 but 1's conditional payoff from deviating to a1=1 is 0.5(2)+0.5(0) = 1, =2 is 0.5(0)+0.5(1) = 0.5. A Thus, the sequential-rationality inequalities can be strict for all actions. equilibrium, we could consider also deviations of the form: "deviate to action cTo tighten the sequential-rationality lower bounds for conditional expected payoffs in B the equilibrium strategy would select an action in the set B i A . k ik i k i k ik sequential equilibrium. In the next example, strategic entanglement is unavoidable, occurring in the only basic 16 whe ik and any 1 fr o m [1, 1] , pl a y er 2 c h o o se s fr o m { L , R }. E x a m pl e: D at e 1: Pl a y er 1 c h o o se s a[ ( U n a v oi d a bl e) S tr at e gi c e n ta n gl e m e n t] ( H ar ri sR e n yR o b s o n 1 9 9 5) D at e 2: Pl a y er s 3 a n d 4 o b se rv e th e d at e 1 c h oi c es a n d e a c h c h o o se fr o m { L , R }. For i=3,4, player i’s payoff is -a if i chooses L and a if i chooses R. 1 1 |, if they mismatch he gets |a |; 1 (First term) If 2 and 3 match he gets -|a 1 Pla yer 2’s pa yof f de pe nd s on wh eth er she ma tch es 3’s ch oic e. If 2 ch oo ses R the n she get s2 if pla yer 3 ch oo ses R but -2 if 3 ch oo ses L. If 2 ch oo ses L the n she get s1 if pla yer 3 ch oo ses L but -1 if 3 ch oo ses R; an d Pl a y er 1’ s p a y of f is th e s u m of th re e te r m s: pl u s (s e c o n d te r m ) if 3 a n d 4 m at c h h e g et s 0, if th e y m is m at c h h e g et s -1 0; plus (third term) he gets -|a su 12 |. bg am e per fec t (he nc e seq ue nti al) eq uili bri um in wh ich pla yer 1 ch oo ses ±1/ m wit h pro ba bili ty ½ eac h, pla yer 2 ch oo ses L an d R eac h wit h pro ba bili ty ½, an d pla yer s i=3 ,4 Ap pro xi ma tio ns in wh ich 1’s act ion set is {-1 ,… ,-2 /m, -1/ m, 1/ m, 2/ m, …, 1} ha ve a uni qu e Pl a y er 3’ s a n d pl a y er 4’ s st ra te gi es ar e e nt a n gl e d in th e li m it. T h e li m it of e v er y a p pr o xi m at io n pr o d u c es (t h e sa m e) st ra te gi c e nt a n gl e m e nt . both choose L if a=-1/m and both choose R if a =1/m. 1 1 17 then simultaneously play the following 2x2 game. [(Spurious) Strategic entanglement] (Harris, Stinchcombe, Zame 2000) Example: is uniform [0,1], payoff irrelevant, observed by both players 1 and 2, who L R L1,13,2 R2,30,0 Consider finite approximations in which, for large integers m, each player observes the whether the m’th digit is 2 or not. first m-1 digits of the ternary expansion of . Additionally, player 1 observes whether the m’th digit is 1 or not and player 2 observes Consider equilibria in which each player i chooses R if the m’th digit of the ternary equilibrium of the 2x2 game where each cell but (R,R) obtains with probability 1/3, expansion of is i and chooses L otherwise. As m, these converge to a correlated independently of . Not in the convex hull of Nash equilibria! two examples. The type of strategic entanglement generated here is impossible to generate in some approximations—it depends on the fine details—unlike the entanglement in the previous 18 type tProblems of spurious concealment in approximating games Example. The state is a 0-or-1 random variable, which is observed by player 1 as his , and then c 1 19 players 2 and 3. player 1, receiving 1 if they do but 0 otherwise. Example: Three players choose simultaneously from [0,1]. [Problems of spuriously exclusive coordination in approximating games] Player 1 wishes to choose the smallest positive number. Players 2 and 3 wish to match It should not be possible for 2, but not 3, to match 1. player 1 is available to player 2, but not to player 3. But consider finite approximations in which the smallest positive action available to matches 1 but 3 does not. choosing any (other) action. Hence, there are basic sequential equilibria in which 2 The only equilibria have 1 and 2 choosing that special common action and player 3 approximation. The exclusive-coordination equilibrium here depends on the fine details of the the same finite approximations. (How to recognize sameness?) Possible solution: Require that type sets and action sets which are "the same" must have 20 Limitations of step-strategy approximations Example: [0,1], is uniform [0,1], player 1 observes , chooses a 1 1 =, v(,a )=0 if a. 1 v )=1 if a(,a11 1 1 In any finite appro ximat ion, player 1's finite action set canno t allow any positi ve probability of cIf we gave player 1 an action that simply applied this strategy, he would use it! utility function is discontinuous. Step functions close to this strategy yield very different expected payoffs because the approximated by step functions when 1 has finitely many feasible actions. 1Player 1 would like to use the strategy "choose a (It may be hard to choose any real number exactly, but easy to say "I choose .") =," but th 1 game can significantly change our sequential equilibria. actions separately, which is needed to prevent spurious signaling. This problem follows from our principle of finitely approximating information and Thus, adding a strategic action that implements a strategy which is feasible in the limit Example (Akerlof): uniform[0,1], 1 observes =, 1 chooses [0, 1.5], a i =a (,a 2 observes 1 1 2 = 0. Given any finite set of strategies for 1, we could always add some strategy such b () < 1.5 , and () has probability 1 of being in a range that has 1 that < bb 1 1 probability 0 under the other strategies in the given set. 21 This must be done with care to avoid the problem of spurious signaling. Using strategic actions to solve problems of step-strategy approximations Algebras (closed under finite and complements) of measurable subsets are specified One way to avoid the problems of step-strategies is to permit the use of strategic actions. , including all one-point sets as measurable. can be assum also for each AikikIf is a finite set, then all subsets of A For any msble g: → A, all (,g( ) and vi(,g()) are assumed msble in . <k ) ik . → ik *:T ik A strategic action for ikL is a measurable function, A ik a Let A * denote ik’s set of strategic actions. ik The set of intrinsic actions A* because for each a* contains the constant strategic action taking the value a , A ik . ik p r ikFor every a*A* and every , let (,a*) denote the action profile aA that o is fi determined when the state is and each player plays according to the strategic action le aps: A T * = T ( i k i k A defined by h<k any a * are measurable, and *). ), *)=(*(, ih *), with measu a a <k (,a*)= (,(,a*)). (,a) from A to v v i i i*) and v Note that for every ikL and every a*A*, *(,a (,a*) are measurable ik <k i functions of . Extend each v 22 employ all of the notation and definitions we previously developed. This defines a dynamic multi-stage game, G *, with perfect recall and we may therefore Q * is the set o i ik m probability in *. Q* is the (disjoint) union of these Q Approximations of G * are defined, as before, by (,C), where C is any finite set of (strategic) action profiles in A* and is any information approximation of all the T * over ik intrinsic actions. Thus, there is perfect recall with respect to strategic actions but not with respect to recall the action in Adate and the strategic action chosen there. Perfect recall in finite approximations of * actually taken at the previous date. action there is not constant across the projection of those types onto TIf the partition element contains more than one of his information types and his strategic In a finite approximation of G*, perfect recall means that a player ik remembers, looking back to a previous date m<k, the partition element containing his information-type at that 23 intrinsic action, then the action too is recalled. Of course, if the strategic action chosen is a constant function, i.e., equivalent to an Robustness to approximations choose the smallest strictly positive number). An alternative is to seek “as much robustness as possible” subject to existence. details of the approximation. One might then insist that equilibria be robust to all exclusive coordination arise when limits of approximate equilibria depend on the fine The problems of spurious strategic entanglement, spurious concealment, and spuriously approximations. But this is not compatible with existence (e.g., two players each wish to probabilities agree with P(Y|W,) when (Y,W)YW(C), and are zero otherwise. Essential sequential equilibria containing every closed set of assessments M that is minimal with respect to the The set of essential sequential equilibria is the smallest closed set of assessments following property: for every open set U containing M, there is a [] mapping such that where I sequential equilibrium () is the indicator function for U, and each [,,C] is some -approximate Theorem. The set of the set of basic seque U (By Zorn’s lemma and Tychonoff’s Theorem such a minimal set M that is nonempty.) F I lim lim (m([,,C])= 1, C:(,C) lim→ 0+ ) U For completely mixed in (,C), let m()[0,1] YW be the assessment whose References: Journal of Economic Theory 53:236-260 (1991). http://www.laits.utexas.edu/~maxwell/finsee4.pdf Theory of Infinite Games," UTexas.edu working paper. J. L. Kelley, General Topology (Springer-Verlag, 1955). Information," Mathematics of Operations Research, 10, 619-32. David Kreps and Robert Wilson, "Sequential Equilibria," 50:863-894 (1982). C. Harris, P. Reny, A. Robson, "The existence of subgame-perfect equilibrium in Drew Fudenberg and Jean Tirole, "Perfect Bayesian and Sequential Equilibrium,", Christopher J. Harris, Maxwell B. Stinchcombe, and William R. Zame, "The Finitistic Milgrom, P. and R. Weber (1985): "Distributional Strategies for Games with Incomplete continuous games with almost perfect information," Econometrica 63(3):507-544 (1995). 25