Digitized by the Internet Archive in 2011 with funding from Boston Library Consortium Member Libraries http://www.archive.org/details/onrobustnessofeqOOfude working paper department of economics ON THE ROBUSTNESS OF EQUILIBRIUM REFINEMENTS Drew Fuderiberg David M. Kreps David K. Levine No. 454 July 1987 massachusetts institute of technology 50 memorial drive Cambridge, mass. 02139 ON THE ROBUSTNESS OF EQUILIBRIUM REFINEMENTS Drew Fuderiberg David M. Kreps David K. Levine No. 454 July 1987 On the Robustness of Equilibrium Refinements* Drew Fudenberg, David M. Kreps, and David K. Levine < July 1987 Abstract. The philosophy of equilibrium refinements knows things about the structure The word "know" unreasonable. phasis. If in a fixed game game, can of the in the that the analyst, is reject some Nash equilibria as the analyst can reject a particular equilibrium outcome, thoughts about rejecting the outcome. between games, and we characterize We equilibrium, perfection, stability. may have second consider several notions of distance their implications for the robustness of equi- librium refinements. (Keywords: noncooperative JEL Numbers: game theory, refinements of Nash 026, 213) Introduction 1. effort he preceding sentence deserves special em- but he cannot do so for games arbitrarily "close by", then he Much if has been devoted recently to refining the notion of a Nash equilibrium. Beginning with Selten's [11,12] notions of perfection, concepts such as properness (Myerson [8]), sequentially (Kreps and Wilson [7]), often invoked or discussed in the literature. and stability The guiding philosophy that the analyst knows things about the structure of the some of the Nash know in equilibria as unreasonable. The game [5]) are of these refinements is that enable him to reject point of this paper is that the word the preceding sentence deserves emphasis. Specifically, the analyst creates a model of the situation that in the (Kohlberg and Mertens is a simplification and (he hopes) an approximation. Suppose that, model, the analyst can reject particular equilibrium a. outcome using the various refinements, but he cannot do so for models that are arbitrarily "close" to the one created. Unless the analyst's faith in the model about rejecting To study this is absolute, he may be wise to have second thoughts outcome. this "robustness" issue, we carry out the following program. Fix a space of We would like to thank Elon Kohlberg, Jean-Francois Mertens, Lloyd Shapley, two and especially Eddie Dekel for helpful discussion and comments. Financial support from National Science Foundation Grants SES84-08586, SES85-09484 and SES85-09697 and from the Alfred P. Sloan Foundation is gratefully acknowledged. * referees MIT Department of Economics, Stanford U. Graduate School of Business, and UCLA Department of Economics and U. of Minnesota Department of Economics, respectively. * 2009745 games and Nash a notion of "closeness" for the equilibria, is strict games in for each player, the strategy prescribed if, to the other players' strategies. This is equilibria of a given game such equilibria, near strict. equilibrium that is game in [5] the space, a a unique best response is we can think as formidable a refinement criterion as implying, for example, Kohlberg and Mertens' of, the space. For each Now hyperstability. ask: which Nash are limit points of strict equilibria, for nearby games? We call Following the discussion above, we would hesitate to reject any near strict, insofar as the sense of closeness specified captures our doubts about the exact specification of the game. In section 2, we take for the space of The result is all we measure player and action) normal form, and payoffs. games normal form games over a. fixed (finite "closeness" with the Euclidean metric on that every pure strategy Nash equilibrium near is strict. (We also give a weakening of the strictness requirement that accomodates every equilibrium, pure or mixed.) This motivated is hardly surprising, but then the various refinements of Nash equilibrium are for the most part by the analysis of extensive games. turn to extensive games with the following philosophy: may in repeated in roughly, a (physical) extensive form is, is information about earlier is But the analyst given. games with incomplete information. (See Kreps, Milgrom, Roberts and Wilson and Fudenberg and Maskin Our theory [3].) differs from one important way. The previous work considers the sufficiently long, are certain (in his model) is that arise from such doubts are considered in the work on reputation family of repeated games of varying lengths. is we entertain doubts about the players' payoffs and/or their knowledge of the payoffs. The perturbations [6] analyst who moves when, with what of the "physical" rules of the game: moves, and so on. That The In the rest of the paper, Nash (and even, fixed, The this effect of work on reputation a fixed perturbation on a typical result there see below, near strict) in the perturbed and the perturbation is allowed to vary (and game. In made players' payoffs himself, but is, that game can be near knows that these payoffs an extensive form the extensive form, and "closeness" is is fixed, the space of this to vanish). if the horizon paper the game Hence only Nash strict. Section 3 concerns the program for cases in which the analyst That is even outcomes that are not Nash equilibria of the unperturbed game equilibrium outcomes of the original players. effects are may be unsure of the common knowledge among games is the space of measured by the Euclidean metric on all the payoffs for payoffs. We do not get very clean results here. There do exist near strict equilibria that are not strict, but the class of near strict equilibria has no apparent simple characterization. Sections 4 and 5 are the heart of the paper. Here we imagine that the analyst has no doubts about the physical rules of the game, but willing to is not quite certain of payoffs, and is admit the possibility that the players themselves are not quite certain of each others' payoffs. Put another way, close to a given extensive game are entertain slight doubts about each others' payoffs, and our analyst an equilibrium that cannot be rejected games that are nearby in games is in which players not prepared to reject in this sense. Section 4 deals with a definition of "closeness" under which every (pure) Nash equilib- rium of the original game 4.1. Technical details is near strict. We begin with motivating example and the definition of closeness are given ness allows each player to be uncertain of his may a. own payoff, as a small ex ante uncertainty can In 4.3 we show large that, with this notion of closeness, (Once again, mixed strict. become equilibria, all it subsection This version of close- and to believe that (with small probability) have better information about an unexpected deviation by one player can signal that in 4.2. in his than he does. In opponents this setting, players should change their play, upon observing an "unexpected" all action. pure strategy Nash equilibria are near can be accomodated with a minor weakening of the definition.) In section 5 we consider a second notion of closeness, which requires that each player's additional private information relate only to his own payoffs. If each player's additional information must also be independent of the information of the others, then an unexpected own deviation by a player signals only that his (In the literature ipated. on reputation bations used satify this restriction.) (pure) Nash in With equilibria are near strict. payoffs are different than had been antic- repeated games referenced above, the pertur- this However more all restrictive notion of closeness, not all pure strategy equilibria that are trembling hand perfect in the normal form are near subgame perfect. perfection. tion We Finally, may be is like correlated. including some that are not obtain a converse by weakening slightly the notion of trembling-hand we consider must be independently equilibria that strict, strictly relaxing the assumption that players' additional informa- distributed. Here we obtain a characterization of near strict trembling hand perfection, but where the "trembles" of various players The idea that small otherwise imperfect, equilibria among make others, we consider can render uncertainties of the kind this point. is Van not original to us. Our contribution to define is Damme some perfect and Myerson [13] and characterize more [9], precisely the effects of various kinds of uncertainties corresponding to various measures of closeness. 2. Fix a finite player, finite action games with i = finitely many by S where The does E,- set j{i,s) is a Two strategies ct_,- = utility whether ?' cr,- is 's of the other players, i a, ?' this throughout to strategies.) under strategy s We i take T . Let Denote = R IxS , . natural topology, namely the Euclidean topology. So and ...,ctj) E a\ (ai,...,aj) i , and E = Yli=i ^« are chosen by use the ? 's is, opponents, no matter what all players receive l for a given mixing Strict equilibria satisfy all the is wm . up to equivalent strategies strict (unless the (^?e • .) are said to be equivalent , chooses a, or o\ equilibrium, taken as a singleton set, many normal form: game 7 unique best response to the strategies pure strategies can be this definition. , (cj, ...,o",_i,<7i+i, Nash equilibrium a = player i, i possesses finitely games over profile to refer to elements of strategies for player restrict attention £ 5, index the pure strategies of player a, the payoff to player T comes endowed with term strategy A , whom the space of mixed strategies for player , same let Let T be the space of . 7 £ T for and (We normal form. players, each of 1,...,/ index the players, Yii=i ^i the Normal form games and payoff perturbations for is 2 i cr_,- is = called strict if, for each (01, ...,ct,_i,ct, + i, ...,07) Note that only an equilibrium in over equivalent strategies), according to standard refinements; in particular, a strict hyperst&ble in the sense of Kohlberg and Mertens 3 [5]. 1 Equivalent strategies arise naturally when the normal form game is constructed from an extensive form game, and one player has multiple information sets in sequence. Then if the player takes an action at an earlier information set that precludes reaching a later information set, the choice of the player at that later information set is irrelevant. N.B., it is the player himself who precludes the later information set - equivalence of strategies does not depend on the strategy choices a_,- of other players. 2 That is, any equally good best response a\ must be equivalent to C{ We could equally normal form, in which all equivalent strategies are . well speak of strictness in the reduced identified. 3 Recall that Kohlberg and Mertens work with the reduced normal form. Conversely, 4 A Definition. is £ strategy profile a G a sequence of games 7™ (over the {ct"} such that lim n a n = a form for A 1. if it is The proof is 7 7 same normal form n for each (ii) , normal form strict in the un , is as 7 a for a game 7 if there and a sequence of strategies ) strict equilibrium of 7" , and (iii) . Proposition and only = lim n 7" (i) near is a pure strategy To quite simple. is a £ E strategy combination is Nash equilibrium near strict for 7 in the for 7 if . see that any strategy profile that a Nash equilibrium of 7 normal form is near strict in the Normal use the upper hemi-continuity of the Nash equilib- , rium correspondence. For the converse, consider the perturbation that adds a small amount to the players' utilities at the It is outcome prescribed by the strategy easy to deal with mixed strategies, nition of near strictness. Recall if one permits a small extension to the from Kohlberg and Mertens form game, an equivalent normal form game is in question. one in defi- that for any given normal [5] which pure strategies that are con- vex combinations of other pure strategies are added to or deleted from the original game. Corresponding to every strategy profile in an original game are (possibly many) strategy a given equivalent game. Consider the following modification of the definition profiles for of near strictness in the normal form. A Definition. normal forms that is near Proposition only if it is strategy profile for game if it strict in the 2. A for a normal form strategy profile is depicted easy. A o for 2 near strict in equivalent strict in equivalent- of an equivalent normal forms singleton set" instead of "strict". in 7' for 7 if and Consider which player 2 randomizes we have an equivalent game added as a new strategy restrict attention to game . and the particular equilibrium any pure strategy equilibrium that we 1 near for 7' is equally between L and R. In figure 2 as long as is simple example will illustrate the method of proof. in figure 1, equibbrium strategy normal form game 7 corresponds to some strategy profile a a Nash equilibrium of 7 Again the proof the 7 a for player 2. 7' , with the particular In this game, this added is, as a singleton set, hyperstable, will be strict. Hence pure strategy equilibria, we could use "hyperstable as a M strategy corresponds to the mixed strategy of player 2 equilibrium in the original game corresponds game, and we can apply the The spirit of this first in the original game; the Nash to a pure strategy equilibrium in the equivalent proposition. extended definition that players is may derive a little extra utility from playing a particular randomized strategy. This technique can be applied at any point development to follow, to extend results from pure to mixed equilibria. However, as in the both referees emphatically observed, this what will, in hence, is an device of artificial follows, discuss only the near strictness of 3. The is little pure strategy content. Hence we profiles. Extensive games and payoff perturbations result that every pure Nash equilibria of 7 near hyperstable as a singleton set) is near strict in the normal form (and, hardly surprising, since is sequence of (vanishing) payoff perturbations to the normal form game. most normal form games, every Nash equilibrium refinements were created initially to deal with we might wish form. Hence 4 satisfies games that we aUow any for Recall that, for the standard refinements; the arise from a specified extensive game model, to permit, as perturbations to an initial extensive only perturbations that respect the structure of that extensive form. We E form still can imagine that our outside analyst , concerned with a game over an extensive is and he entertains no doubts about the physical our analyst may have doubts about the full rules of the game so specified. specification of the situation. But He may, for example, be slightly unsure of the payoffs to the players or their probability assessments concerning nature's moves. 5 assessments. for nature's 7 6 T 4 , Thus we can moves in which strategy The We E . will ask: suppose that there is no question about the probability Fix an extensive form Let T be the space of all E and probability assessments payoffs for players in E . For a given profiles are near strict in the following sense? somewhat more satisfactory treatment for mixed strategy which one considers as a perturbation of a given game the same game in which a player is "replaced" by a continuum of identical copies, as per Harsanyi's [4] classic treatment of purification. We will not follow that route here, except to suggest to the interested reader that, when one comes to general elaborations in section 4, one will have to be careful with this if one is not to get as near strict all correlated equilibria, precisely referees suggested a equilibria, in in the spirit of 5 Myerson [9]. The usual accounting tricks make this without loss of generality, about the support of those assessments. 6 if there is no question Fixing Definition. the extensive form 7 6 T, a (pure) strategy if We know n -+ <r from Section 2 that, <7 strict extensive form but that are not strict. strict in the extensive form It is is taken from Kreps and Wilson equilibrium (L,D) will be profiles that are near (Our system 3. [7].) It Indeed, for some extensive games, no equilibrium The game generally, is near depicted in figure 4 But not every diagramming extensive for should be easy to see that the Nash is near strict in is. the extensive form. an example: While (R,Uu) and (R,Ud) (and, more any randomization between these two strategies) are very nice equilibria, neither strict. Normal form, In fact, for an equilibrium of an extensive player whose game necessary that any information set h that it is probability either is is strict in the near strict in the extensive is not near strict in the extensive form; only (R,U) is (i) is move it is is to be near strict in the not reached with positive precluded solely by the action at some earlier information set of the at h or a (ii) is irrelevant to everyone's payoff. This "dummy" is information set, where the player's choice a very simple manifestation of the type of problem that leads Kohlberg and Mertens to define stable components of equilibria, and similarly strict easy to show that every strategy profile that Consider the extensive game in Figure form games in themselves. Viewing a given normal form as pure strategy Nash equilibrium over a given extensive form form. strict are equilibria that are pure strategy Nash equilibrium. a is near 7 and (normal form) > some games, there a simple extensive form, we see then that there near — is . for at least near strict in the normal form but not is a of the game 7 there exist a sequence of payoffs 7" a" of 7" such that equilibria profile attempt to define near we could components. However, we are more interested strict in the implications of the less restrictive notions of "closeness" that correspond to the players not being quite certain of each others' payoffs. 4. 4.1. General Elaborations An example To motivate found in Van this Damme development, we provide a simple example. [13].) Consider first the game Imagine an outside analyst whose thought about in figure 3 (A similar example is and the equilibrium (L,D) this situation run as follows: know "I the particular physical rules of this game. R between L and And Player I'm fairly sure that the payoffs are as in figure certain of this. am I U and player 2 must choose between , willing to admit that there are quite different from those 6hown. Moreover, between the players what those payoffs information, in that players do not is That are. know the D and first R if is choose chosen. But I'm not completely 3. a small chance that the payoffs might not be it must 1 is, the common knowledge game may have incomplete precise payoffs, although all the players will have priors that give probability close to one for the payoffs in figure 3." To model we introduce a game that Nature moves follows. is this, first, more elaborate than that is selecting one of several "versions" of the distinguished by having the same extensive form game. Each version may have as in figure 3 but Nature chooses version having payoffs close to those payoffs. in figure 3, as in figure 3 different with probability Each player has a partition over versions of the game, with each respective close to one. player being told in which One such of his partition the true version cell elaboration is given in figure which occurs with probability occurs with probability picks, while player 2 is from the original game e , — €, 1 5. Nature picks one of two versions. The has the same payoffs as has very different payoffs. Player not told. for small Is this f ? elaboration of the Our lies. in figure 3; 1 first, the second, which knows which version nature game in figure 3 very different outside analyst, plagued by the sort of doubts expressed above, might not be willing to rule out the possibility that the players perceive the situation as being that in figure The (Zi-fto, we D point of this example D) is decrease , strict in the normal form, as shown in figure player l's choices at his given the strategy of player If game near 1, player 2 is given the D to measure distance so that, as c optimal the choice then we would conclude, in the strict. This is see: strictly remains if strict as will nature picks version 2, game approaches the of earlier discussion, that (£, (L,D) play optimal choices. And, goes to zero, this spirit it Given that player 2 move only so even though, in figure 3, 8 figure 5, the equilibrium Moreover, 6. both sets are 2. in figure 3, is two information by strictly we were figure 3 now easy to make. For the game in towards zero. This should not be bard to e which renders is 5. is D) in subgame imperfect. (Note that a nontrivial step as goes to zero? e With We is implicit here: In give a formal criterion below.) we now develop this as a prelude, We begin with a precise definition concerning when one of another. We games game Fix an /-player extensive (with nodes denoted by various players, initial some t ) which W ; The kind game And E game will w / that W £ H game tree , and terminal nodes , Z z £ T we have in an ; with H{i) denoting the information actions a £ A(h) at each information set; ; assigning utilities to E This prescribes a . partitioned into sets of nodes T, that "belong" to the node all and a players. mind is one which one of TV possible in selected by nature at the outset, where each version has is game above and, except for initial uncertainty as to nature's choice of structure. Versions are distinguished by the players' payoffs. be called elaborations of N this, fold at (i, n) for copy of utility to player sets are W i composed , t . all E . N fold If player n ; i.e., copy of T, i T moves = T X - t , T or {1, ..., . r ; For each player i, in set of initial a partition Pi Actions at information sets are inherited set in in scrutiny. Player 9 t (Z, E the , n) to denote the (if t £ then T,- ), nodes consists of the node (w,n) given by p(w)fi(n) and the information The information structure bears some node use E . The denoted by u(i,z,n). Finally, information as follows. is positive on {l,...,A }. In /j N] we ..., N} The initial node (z,n) given, X {1, at (non-terminal) at terminal is A of perfect recall built up as follows. with the probability of denoted by Pi(n)) H(t) X Pi(n) . imagine a game tree consists of an moves E given, together with a probability distribution is n th copy of the node cells a small perturbation players are unsure (in a general sense) as to which version prevails. Such perturbations integer N R same information To formalize i —> game above tree of the version, the Z I x of perturbation of "versions" of the is E of perfect recall, information sets h £ set containing the (nonterminal) : is to be close. of which are initial nodes assessment p over payoff function u game not develop a formal topology on the space of games, but instead we will give a simple sufficient condition for of a general treatment. Elaboration perturbations 4.2. the what sense does {L\R.2,D) approach (L,D) E of {1,2,..., A} (with with node (t,n) £ T{ is the obvious fashion. i 's exogenously given informa- tion concerning which version That is, if it is H(t) and his construction. player i 's n turn to move knowledge about n E The game chosen at the outset is is node (t,n) at given by Pi(n) in figure 5 consists of . P\ is tions of E in E consider such elaborations E . For E of a E game to be a "small" perturbation, t is . given by Figures 3 and 5 illustrate the basic E in figure 3. In the first In the second (bottom) in figure 3. learns at the outset which version 1 the discrete partition. Player 2 learns nothing; We knowledge about his , two versions of (upper) version, the payoffs are identical with those version, they are quite different. Player described by the partition P; is Pi i.e., as being we take it is is chosen; i.e., the trivial partition. among the possible perturba- to be sufficient that the payoffs should be, with high probability, approximately those in E . Formally, we pose the following criterion. Convergence Criterion. For a given game E To say that the sequence approaches . (i) there ration (ii) (iii) , {E k } let be a sequence of elaborations of sufficient that: it is and a uniform bound on the absolute values of the u k for each k, i and hnu-^oo^l) = Conditions , , a uniform bound on the number of versions of the original is Ek E E (ii) and z , lim/.—^, u k (i, z, 1) = u(i, z) ; game ; and (iii) state that, along the sequence, the probability that nature picks a same as in the original game approaches Note that a convergent sequence of extensive form payoff perturbations of (trivial) convergent sequence of elaborations. 6 The first it cannot hurt. The second part of the probability of other versions of those versions, then we can is going to zero, if (i) is there E part of condition unnecessary to support a notion of closeness, but since we are giving a for convergence, each elabo- 1. version in which payoffs are asymptotically the a in corresponds to (i) is probably sufficient condition quite important, however: is one. Even if no uniform bound on the payoffs inflate the payoffs in other versions so that, in expected utility 6 Strengthening condition (ii) to require that payoffs in the first version exactly equal those in the original game is vacuous: Given a sequence of elaborations satisfying (ii), we can replace version 1 by two versions, called 1' and 1" with payoffs in 1' exactly as in the original game, and specify that no player can distinguish between 1' and 1" With appropriate care in picking payoffs in version 1" and the relative prior probabilities of the two, the new elaborations would be strategically equivalent to the old ones. , . 10 terms, the players' expected payoffs over With the uniform bound on expected payoffs, only the will loom large Is this The in the versions are anything all payoffs, however, (iii) to be. implies that in ex ante calculations of version (which gives nearly the payoffs of the original game) first the limit. games a reasonable sufficient condition for closeness of considerations put into the would argue that we wish them it mouth of our outside analyst at the top of this section may although the reader is, form? in the extensive already see how broad a class of small perturbations this allows. Insofar as the refinements of Nash equilibrium deal with out-ofequilibrium behavior, a small probability ex ante of very different payoffs large ex post. Near As strictness follows, we will see that this E do we must we wish specify a mode given by is E game. This is not completely trivial, in the original because strategies game. Since in we take only able to view actions in the physical is (and not those actions contingent on which version nature selected), one to look at the (marginal) distribution induced by strategies on endpoints of the "physical" game. That one could compute the distributions induced on is, Note that even that they converge. has a single version of T , if all this gives us of entire strategies. We Definition. An (t, , and require convergence of strategies only at information sets is weaker than the usual convergence therefore use a slightly sharper convergence criterion: convergence of behavior prescribed in an elaborated contain the nodes Z elaborations are trivial, in the sense that each that are reached with positive probability. Hence this is To of convergence of a sequence of strategies for elaborations the point of view that the outside analyst there this. to identify (equilibrium) strategies for a given an elaborated game are much richer than strategies answer and the reader should consider that are near strict, in this case under the notion of closeness developed above. to a strategy for the original game so, is quite under general elaborations the previous two sections, in game so, what whether the uncertainty of an outside analyst may be so great as carefully 4.3. In may loom game at all We ask for information sets h that 1) strategy a for the game a sequence of elaborations of E E {E k } , near is , under general elaborations that converges to convergence criterion given above, and a sequence of 11 strict strict E in the sense of (normal form) equilibria if the {cr k } for the respective elaborations, such that the behavior prescribed games converges in their respective A 3. elaborations if Proof. First we show that any strategy is profile that a Nash equilibrium. Suppose that some a Then there some player is i where u(i,a) denotes the the strategy profile o for all j 5",- , (<7,,<7_i) is > e 1) . near strict under general is in strict but E . not a Nash equilibrium. is such that u(i, in the i (/., near strict under general elaborations is near is and utility to player and , strategy , E nodes at prescribed by a Nash equilibrium a (pure strategy) if it is t a for the game pure strategy profile Proposition and only node to behavior at by the a k game if > (<7,-,<7_j)) the outcome +€ u(i,a) determined by is the strategy profile composed of a, for and Oj i 7^ i {E k } and Fix a sequence of elaborations converge to a the sense in payoffs in version we {v k } which strict equilibria are using here. Payoffs are uniformly bounded in the prior probability of version approaches 1 all versions, k approach payoffs in the originally given game, and of elaboration 1 corresponding 1 Therefore . overall expected utility in i 's elaboration k approaches u(i,a). Suppose that there were some &{ as above. Construct d k so that, in elaboration prescription of d{ , k, at information sets and otherwise (/i,P,-(l)) follows the prescription of i coverges to u(i, (ai,a_,)) so that for large enough k in elaboration k , Ek strictly node z € dominant T , ask i i 'S i , fix follows the i , As k grows, u(i,(a k )) b k would be better for i than a\ a (pure strategy) Nash equilibrium a and construct an elab- and pure strategy for player 777,(2, Si) payoffs at i = .) when To give — T7l;(2,S;) Letting #5,- be the 771,(2, s^) z lies along a i for i , design payoffs for player i i that make this, at had to deviate from Si strategy each terminal if 2 is to be path that passes through no information the strict incentive to follow sets at every opportunity, set s,- . number are Yli=i(i^^i + 1) where each is drawn from the r, S{ throughout the course of play. To do how many times reached. (For example, of . l as follows. For each player S{ dk H a contradiction. For the converse, we oration , h £ for of pure strategies for player versions of the original game. set We , in each elaboration there label these versions by 5ilJ{0}. Inversion 12 i (r : , ....,r,) (r l5 ..., of elaboration k 77) , the payoffs to player £ Si (i) If r; , i are as follows: then player ? assigned the payoffs described above that is make a strictly r, dominant strategy. (ii) If = r,- and, for some j ^ i rj , £ Sj then player , make assigned payoffs that is i playing Oi a strictly dominant strategy. If (iii) = T{ to u(i,-) — and, moreover, mi(-,<7i)/k where u , = r$ for all j , then player is i assigned utility equal the originally specified utility function and is m,- the is number-of-deviations function defined above. The oration has prior probability k = M«( r i) prior distribution on these versions (^ ~ l)/'c selected, at version ^ i crazy as Ijk he , is is i = ' Mi( r !') player , may seem The has pure strategies. among crazy or sane and, if where /x;(r,) the ?' is = told the value of quite complex, but either "crazy" or "sane", divided equally ? Version set as follows. 1/(A: • (rj, ...,?"/) #5,) if r; G of elab5"; and Finally, for his initial information concerning the version • (rj, ....,77) This construction player r [,- I is many crazy, the it has 7-,- simple interpretation. a. and there are as many overall chance that different i . different A his craziness takes. selected independently. (Note that the payoffs of one player will other, so the payoff perturbations are not independent.) If for i crazy in elaboration is forms of craziness. form that ways player Each to be k is told whether is Players' "types" are depend on the type of the a player is crazy according to a certain pure strategy, then the player has payoffs that cause him to play that strategy at every available opportunity. ' If a player is sane and some other player type), then the sane player wants to follow the fixed available opportunity. It in favor of the fixed crazy (of any Nash equilibrium strategy Finally, if all players are sane, game, with a small "kick" is a,- at every then payoffs are as in the original Nash equilibrium strategies a; . should be evident that this sequence of elaborations converges to the originally given game, according to the criterion ration, the following strategies d{ we have given. We are strict: Player 7 claim, moreover, that in each elaboi follows a,- if his initial information This sort of construction, in which every strategy is played with positive probability, used by Fudenberg and Maskin [3] to obtain a robust folk theorem for repeated games with long but finite horizons and small levels of incomplete information. is 13 is = r,- and he follows , we established, will 7"j if information his initial {dj} of his opponents cr; perfect recall, the information sets of h for that i from different &i is b{ and cr,- earliest in is and cr; because this information sets of is We ). own i 's £', . Once this claim is wish to establish that a actions preclude. terms of precedence among in which best responses to the strategies 's i By is l precisely the assumption of are ordered by precedence. Take any information i all of i 's information sets in which previous actions (which would be the same under i 's an earliest information set) do not themselves preclude h we satisfy these conditions, i one of is some elaboration k (in except possibly at information sets that set e v, have the proposition. Suppose that some (pure) strategy a, some is . If no are done. There are three possibilities to consider: The information (i) h preclude , strategies in h set and because every strategy And at that information set, does strictly worse than (ii) <7; , information set that equilibrium. is i strict best strictly better is latter in is h has positive ?' strictly better . i a with positive probability which everyone , and it corresponds to an in the original in game and one of two ways: Either sane, in which case foDowing d{ is a strict best (recall the "kicker" in defining utilities for this the all-sane version), or someone else is crazy and i is than a, , we sane, in which case following response for the rest of the game. In either case, there is a strategy for ct; i and again we have a contradiction. more than one weakly best hit himself doesn't possible under their a strategy for with positive probability under (The kicker perturbation to as there a, is hit it is is i any action other than that prescribed by d{ belongs to the sane variety of Then under a Because . the information set h are in a version in which a i ), i opponents (from then on). Hence there CTj set in the version in response for is is which contradicts the assumed optimality of The information we are profile of his a (conditional on whatever type prior probability. than belongs to a crazy variety of utility in the all-sane version would be unnecessary as long player, since then we would have that following a,- is at least the all-sane version, strictly best in every someone-crazy version, and the would have strictly positive probability. has zero probability.) 14 For one player games, however, the latter (iii) The information set that is h belongs to the sane variety of set off the equilibrium path in does not, by his actions, preclude h strategy combination by i 's . is Since there opponents, there reached. Moreover, Bayes' rule forces must be crazy the original i is i to follow a,- , corresponds to an information it game under the strategy a , and player h positive prior probability that to conclude that at this point i positive prior probability of every other is at this point, at least insofar as they strictly better for i all some one or more of his a follow their parts of will be opponents . Hence it and henceforth, and we have the same contradiction as in the previous two steps. Q.E.D. Proposition 3 shows that any Nash equilibrium can be "justified" the right kinds of doubts about each other's payoffs. this deviation. However all his way in of the Na,sh equilibria are "robust" in that sense that, given a we can consider any "small enough" same conclusion. justifies a given equibbrium, Loosely speaking, every (pure) Nash equibbrium 5. games and get further perturbations of the elaborated "open cone" of elaborations whose vertex 5.1. which a player responds to opponents depends on the inferences that he draws from sequence of elaborations that approaches the original game and the the players entertain Naturally, different forms of payoff uncertainty will justify different equibbria, because the an unexpected deviation by if is is justified by an the original game. Elaborations with Types Independent types In terms of our outside analyst story, section 4 says that unless one players wiU not draw unmodelled inferences about their of their opponents, one should not reject any (pure) that our outside analyst is own is certain that payoffs from the deviations Nash equibbrium. Suppose, though, prepared to assert that the model as written captures ab of the information that players have about each other; he entertains (only) the possibibty that each player has unmodelled private information about his own payoffs. from the bterature, the analyst entertains the possibibty that each player types (specified by payoffs), and each player knows his own type. If it is Using language is one of several also true that no player has information (beyond the information possessed by the analyst, which ab players have) about the types of the others, then we will say that the types are independent. 15 In this case the class of small perturbations that the analyst and the 4.2, set of near strict equilibria that result (Note that would consider might be smaller as is smaller than well. considered in section 4, a player's payoffs were in the general elaborations not wholely determined by the player's "type". In particular, in the proof of proposition the utility function of the "sane" type of player player was "crazy." Hence we payoffs; now will distributed; that To (L,D) that the prior we 1 lay with D . on whether any other crucially game was pertinent to player (This, of course, supposes that L chooses 1 dominant This below) is will 2's best choice the top version of the game.) in could not be given information about 2's payoffs superior 1 moving is not render near In figure 8 informed about sequential; a 8. some strict know 8 that In figure 7, equilibria that are not themselves strict. Consider we have an equilibrium (Lu,D) which we have an elaboration 1 of this only. (There are game in which there no information. In his payoffs; player 2 gets this more involved elaboration with independent types bottom version of the game, player (Li^^i^, D) 1 plays is is 8 The R% with subgame some uncer- two possible types of player renders (Lu,D) near strict under such elaborations. native, in the not to say that "elaborations with independent types" (to be defined formally tainty as to the payoffs to player 1 is in for him. the games in figures 7 and imperfect. game Hence, 2's payoffs. information set that contains the node from the high probability version, would is a in figure 5, to the information that 2 receives, this would be impossible, as player 2, U own and the Nash equi- in figure 3 by choosing R, was communicating to 2 that player his also require that the types be independently near strict by formulating, received information that we supposed that 3, have a product structure.) will this equilibrium this elaboration, player 1, if f.i depended meet the requirement that each player knows require this. Initially, We made . which player But did not i see the force of this restriction, consider the librium in is, we in 1.) Player game, (L^R^u-ido, D) will intuition make is it strict, which that in the alter- positive probability and will not strict because player l's choice between u^ and di To make the choice matter, imagine that there is independent uncertainty about 2's payoffs, where, with small prior probability, 2 has A as a dominant strategy. Think of an elaboration with four versions, two (for each type of player 1) by two (for each type of player 2), where each player knows his own type but not the type of the other. If this is too sketchy, the general definition and construction to be given in a moment will clarify matters. is In figure 8, is irrelevant given player 2's choice of D . 16 continue with d 2 R 2 d2 because in the bottom version, , dominant. Player is 2, then, given the move, assesses probability one (at the L\ action equilibrium) that the bottom version being played. As is With must modify the An ration elaboration is Ni Yii i there said to be an elaboration with independent types is an integer is for each (iii) there is and version (iv) for each (j and j G i u ...,'jj) , u(i, TV,- is simply i /z,- {1, ...,TV,} ;(ju the information given to prevails we such that the number if: of versions in the elabo- i a probability distribution i First, an allowable elaboration. writing (ji,...,ji) as the index of one of the elaborations, where (ii) is definition of for each player (i) therefore continue with d 2 given the chance, 2 chooses D. example to provide motivation, we turn to a formal development. this Definition. 1 will if ...,j/)) = on {1,...,TV,} with a payoff function 6 ji //((ji, {1,...,TV,-} ji)) ...., v(i, z;j) , there — [} m(ji) such that for each ; i v(i, --Ji) the true version is (j\,.--,ji) concerning which version j,- The same convergence criterion as in section 4.2 is we now used, although restrict to elaborations with independent types for perturbations. Definition. A strategy a for an extensive independent types {E k } if there that converges to , k strict equilibria {cr } for the by the d k at nodes To give our result, Definition. A each player i 1) if converges to behavior at node we need a there and index k is , under elaborations with t prescribed by a Damme , further definition. game 7 a sequence of totally mixed strategies a k o± is . a strict best o — is > strictly perfect a such that, response (up to equivalent strategies for k [14]. 17 for i ) 9 . The term "strictly perfect" is used by Okada [10] to mean something quite same vein, we later use the term "quasi-perfect" in a sense different from van E corresponding normal forms, such that the behavior prescribed In the in strict convergence criterion, and a sequence of in the sense of the to the other players' strategies prescribed by 9 near (pure strategy) equilibrium a of a normal form the normal form in (i, is a sequence of elaborations with independent types of is E E game different. its usage Remark. Figure 9 shows that (Lu D) is 7 normal form. The num- strictly perfect in the bers in parentheses are "trembles" corresponding to a sequence of totally mixed strategies converging to (Lx,D) such that u. Ru , are strict best responses for players normal form. The key respectively, in the reduced often than onto D Lx and even though, having played is that player R trembles onto 2, Rd more player one "should" continue with , . 1 and 1 10 Proposition For a given extensive form 4. normal form in the Proof. E , every equilibrium o that near strict under elaborations with independent types. is — Fix a sequence of totally mixed strategies u k o best response to each In the . of pure strategies of player original that type Sj i if s; = 0{ for i is chosen game strictly perfect is , . k th elaboration, N, Payoffs for and they make is > i sl in versions o such that each is set equal to a,- is #5,-, the corresponding to a strict number are as in the S{ dominant otherwise. The marginal probability exactly that assigned to s; in a\ . Hence these independent elaborations do converge in the sense of the convergence criterion to the originally specified game. By play a, construction, it and to play S{ , is a strict Nash equilibrium when crazy of type s, . in Ek for each player i , if sane, to This gives the result. Q.E.D. Given our previous results, Proposition 3 should not be surprising. Imperfect equilibria that are perfect in the normal form correspond to players inferring that, deviated from equilibrium play once, he justified if is likely to do so again. if an opponent has Such inferences can be each player believes that with small probability his opponents' payoffs may be quite different from payoffs that have high prior probability. The converse strict to Proposition 4 is, however, false: with respect to elaborations with independent types. The problem equilibrium correspondence is L both receive 1. If either is that the perfect poorly behaved with respect to extensive form payoff pertur- bations. Consider the foDowing two person simultaneous , Imperfect equilibria can be near chooses R , move game. If both players choose then both receive zero. Only (L,L) 10 is perfect, but Recall that Selten introduced perfectness in the agent-normal form precisely to rule out such correlated trembles. 18 (R,R) near is under payoff perturbations strict payoff to (R,R) was and approach zero. let ( One way e in the extensive form: for each player, holding the payoffs for to obtain a partial converse to Proposition 4 is Imagine thai the {L,R) and (R, L) at zero, assume that each players' to "main type" (the type corresponding to the main version of the game) has exactly the payoffs given in the original game. That is, our outside analyst assesses small probability that he has the payoffs off by (possibly) a large amount, but he assesses high prior probability that he has precisely the correct payoffs, and each player knows is not particularly interesting. Instead, taking its we modify the E an extensive form m for the extensive form, {u } , is with hmit u normal forms corresponding to the u m , 12 um let - 5. . a sequence of payoffs if there such that a is strictly perfect in is u quasi-perfect. is game E and only if Suppose that o be the payoff function 11 quasi-perfect. Let u be the extensive form payoffs such that a From Proposition strategy profiles dm ' k 4, . Write Em for those elaborations are such that & m,k — a in Em ' k are uniformly take m;v large enough so that u mN bounded is within 11 in k . 1/2N E for , and normal form game with payoffs ' k E m,k of Em and are strict equilibria in Examination of the construction number m am is quasi-perfect. strictly perfect in the is for the extensive Proposition 4 shows that we can assume that the the elaborations if it is there are elaborations with independent types the corresponding normal forms, and a a pure strategy equilibrium , under elaborations with independent types is each of the . For a given extensive form forms corresponding to the u m um (pure strategy) equilibrium quasi-perfect Remark. Every perfect (pure strategy) equilibrium Proposition A with payoff function u. a of the corresponding normal form game Proof. definition of a perfect equibbrium by closure with respect to payoff perturbations in the extensive form. Definition. Fix near-strict This class of perturbation this. of versions and size of payoffs in and k of u in , . For each positive integer and then take k^j large N , enough Eddie Dekel has pointed out that this definition of quasi-perfection is equivalent to a definition in which a need be only (weakly) perfect in each of the normal forms corresponding to the u m 12 We thank Eddie Dekel for finding an error in an earber version of this proof. . 19 so that the (iii) , ' a mN with Hence a To is ' is N a. near E on strategies of game E in more are within 1/27V of u mN k» in this elaboration. It is ; and immediate N (in approaches a strict equilibria for those elaborations that E game a strategy for is ct, 1)/7V or we begin with a establish the converse, If . — > ) . strict. vation. Fix an extensive of E mN 1 sequence of elaborations with independent types approaching sequence of a. of the payoffs in version (ii) ; version has probability (TV first {E mN k *} that E N > kN (i) i E in and , E in i induced by cr,- let E , we . That and a preliminary obser- piece of notation be an elaboration (with independent types) Ma, will write averaging over is, according to ^ (or averaging over of all marginal distribution for the versions of the original all personal types according to i 's /x, , which gives the same result because types are independent) we compute the probability with which Now strict to i will assume that M<7; in a is totally and a pure strategy E We claim that . mixed, for each i E in This . is , we Sj i , has can, in the s; as a for a,- is , a each player and follow the dominant strategy i : "Follow for the still appended types," a is composite strategies for the others. (The constraint on the a where to each in the strict best extra response prior strictly positive) so that the from the for types a,- is loss of generality augmented elaboration (with the extra types) choose the probability of those extra types to be sufficiently small (but composite strategy E in i a which we can append so because dominant strategy. Because profile without it is of types a further type for each pure strategy of list type for strategy i E suppose that we have an elaboration equilibrium for the normal form of player's for be "observed" to play various strategies original elaboration, strict best response to the size of the probability of i 's extra types depends on the extent to which the other players' strategies were strict best responses.) Suppose then that a dependent types) just above, Ek E we can assume For each player tree into —y two sets: i Let is near strict for E and corresponding that WLa k is , and fix a sequence of elaborations (with strict equilibria ak — * a . By the observation totally mixed. (holding a fixed), partition the terminal nodes of the original A,- game be those nodes which, for some combination of strategies by opponents, can be reached when i uses v (i,z; 1) to denote the payoffs to player in- a,- i , and in 20 let B{ be the complementary the main version of the game set. i 's Using in elaboration Ek node at z consider the payoffs , w k (i,z) = supu'(?. j>k w '(i,z) = We assert that Ma- sequence To z for z; 1), game with and £ A, for form given by £ Bi. z wk payoffs given by against the test , 7 . we compare see this, strategy against Si u J (-, payoffs are is for the original extensive inf u J'(i, z\ 1), strictly perfect in the is <r w k (i, z) the difference to wk Mai,- when payoffs are given by •;].), for j > reached when strategies are strategies are (s;,Mai,-) a= between playing i . We k a; and some other (pure) with the difference when the Write 11(2) for the probability that terminal node z . (a,-, Mai,-) and write , for this probability II' (z) when the wish, then, to sign ]T(n(z) - n'(z)K(», - £(n(*) - n'(z)y(i, z) z- 1). Z 2 Partition the terminal nodes z into four classes: (1) nodes that are reached with positive probability under both a; and U(z) up A, 11(2) = > and , = IL'(z) ; (4) s; ; nodes with classes (2) and II'(z) in class (1). (4) nodes such that (2) make up B l Now . - zeClass (3) But a,- ativity of perfect Thus a is A when is a II'(z) z in class (3) are in A{ conclude, then, that II'(z) Note that . since A a,- = ; (3) and classes (1) and Sj nodes with (3) make are pure strategies, for z in class (1) net to zero. For k ]T U'(z){w (i,z)-v^i,z-l)), ^eClass (4) where we have used the facts that class (4). Since = we can rewrite n(z)(™%z)-v j (i,z;l)) J2 U(z) Thus the terms making up z in class (2), the terms are zero. Thus A= = > Il'(z) Jl(z) A is strict best , = for wk > z v^ in class (3) there, and = II(z) and conversely for z in We for class (4). nonnegative. response to Mai,- shows that the same is payoffs are given by when payoffs are given by v J true when payoffs are given by wk And, since . vi(-,-;l) . w k Thus . The nonnega converges to u, is strictly wk — > u. quasi-perfect. Q.E.D. 21 5.2. Personal types The perturbations with independent types effect of and tensive form payoff perturbations different information sets of the same is essentially We player. can, and to allow for ex- between the "trembles" at to allow correlation (ii) (i) now loosen the definition will, of elaborations with independent types to allow for the trembles to be correlated across different players. An elaboration has personal types if each player's payoffs depend only on his own infor- mation about Nature's choice of version; the players' information need not be independent. must Formally, an elaboration with personal types satisfy (i), of an elaboration with independent types, but the overall and (iii) (iv) of the definition measure n on versions needn't have the product form. Consider the R game and r, players 1, and 2 in this a 3 play by pair of independent strategies least 13 in figure 10. 2x2x2 and 1 subgame where In the players and 3 choose 3 have played Against any player 3 can guarantee himself a payoff of at 2, Hence any (normal form) perfect equilibrium of the game. 1 and simultaneous move game. B subgame by playing a 50-50 mixture of A and equilibrium of the game, 1 R and r , in , so d is never part of any normal form perfect The subgame must respectively. along the equilibrium path in any perfect equilibrium, and in the subgame 1 and 2 lie have unique equilibrium strategies. (They play matching pennies no matter what 3 does.) Hence 3's best choice is A namely (R, r,A, .577 this, hence this is . That + is, -5T, .5h game has a unique (Normal form) this + .5t) . perfect equilibrium, Perturbing the extensive form payoffs the unique equilibrium that is strict will not affect under elaborations with independent types. But now consider elaborations with personal types. In particular, imagine that there are two other types of players (We dominant strategy. 1, RT for which is dominant. t is 13 [9] This game For player is a all. and 2. refer to this as dominant. 3 learns nothing at 1 2, there type is RE is .) his own h is game suggested making the same point. 22 is a dominant and a type type, so, in particular, player crucial to our construction, the probability variation on a similar RE In the second other type of player a type for which Each player learns only And, what gives another example, In the first other type of player 1, to us by J.-F. Mertens. measure Myerson such that there is /i is high probability, say RH type - 1 and 2 2<5 that , players are of the "standard" all types, probability 6 that 1 RT and 2 the reader doesn't like this perfect correlation and 2, it will suffice type is / (If . is L by 14 equilibrium path. 1, and + .5T, .5/i+ .5<) in the original 3 can L plays This equilibrium made near is The been selected. Given the prior H with and 2 with h , and correlated continuation by -.5 and so , in the main We advance may seem . nor (Of course, B 2 with will this choice Nash puts him off the as follows: Since some other t . version has (i) 1 When will follow 3 assesses this guarantee him a payoff above by 3 validates l's choice of D definitions to follow in the following spirit. to the reader to be too large a class of perturbation, because know more about a second player's payoffs than that second player constraining ourselves to elaborations with the "type" structure, suppose that each player has all the information about his own we payoffs that any other player further restriction to independent types implies that no player learns anything about others' types from in their past learned d A neither example and the general this By herself. The 2, and 1 Nash because It is as equally likely that T follow with consider the 1 type version.) they allow one player to has. and 1 it 1 is the types of in argument runs signals to 3 that posited, 3 sees 1 will (ii) 3 optimally chooses General elaborations knows we \i game. that /> we adopt the personal type strict if intuitive R the main version, the choice of in Now do anything he wishes, as L by elaborations above, for 6 approaching zero. 1 and probability , to have high but less than perfect correlation.) equilibrium outcome (L,d,a,.5H d by 3 engenders type h is his own. One might instead imagine that players have "links" - perhaps they went to the same school, or had similar teachers when they game theory, or whatever. Knowing that these links are possible, one suspects that knowing something about the payoffs of one player might of another. This is true, pari pasu, of players within a tell you something about payoffs game. Which leads to personal, but not independent, types. Although on it doesn't lead to quite the this sort of elaboration. Nash equilibria, one is By analyzing same the thing, we can give another perspective game noncooperatively and by restricting to supposing that players cannot access correlating devices, as defined in 14 Subsequent moves can be fiUed in any way the reader wishes, and we will still have a Nash equilibrium. We have filled them in with the subgame perfect equilibrium strategies for a reason that will become apparent momentarily. 23 Aumann (See also Forges [1]. Myerson arid [2] of the outside analyst's state of knowledge have access to correlating devices, he perturbation of the Suppose that a more accurate statement [9].) that while he is is not completely certain of is game would be an do not fairly certain that players Then an appropriate this. elaboration with personal types, where the utility functions of the players in each version are close to (or perhaps even identical with) their game modelled. payoffs in the had access to a correlating device, the means that the is a game And device will be of his information set, afraid that have obtained access to the correlating device and is if if matching pennies fact that they play no use to them. But even players if one 2 subgame in the we change the subgame and 1 so that is it available. unable to observe the outcome generated by the correlating device, then is he might choose d at point in figure 10, of coordination, then they might well use a correlating device player 3 if example given (In the l's will R choice of indicates that if and 2 The play a correlated equilibrium.) that a low prior probability that players can access correlating devices a high posterior probability, 1 may become one observes an out of equilibrium action from one or more of the players. To characterize the set of near strict equilibria under elaborations with personal types, we introduce the may following definition, which modify trembling hand perfection, to allow Definitions. Fix a pure strategy profile where for each i and k of players other than pure strategy profile i <f>\ , , is a is a be of independent interest. . The equilibrium for a normal form game. is to A sequence {{<f>i}{-i}kLi a totally mixed distribution over the pure strategy profiles c- perfect test sequence for a a of a normal form game 7 is idea for "correlated trembles." is sequence {{^f}} for a such that for each player 4>i The strictly c-perfect if each a i - t for each if, c-perfect if a , $ (a) — > 1 . A there exists a c-perfect test and index k is i , strict best a is a best response to response in the normal form. Remarks. strategies Note that the distributions by the players. <f>^ need not correspond to independently mixed Also perfection and c-perfection are identical for the case of two-player games. Proposition 6. If an equilibrium o is strictly c-perfect, to elaborations in personal types. 24 then it is near strict with respect , Proof. Let k be the c-perfect test sequence that makes a <f> = elaborations (for k game: There will where runs over i for the other players. And all In the in cr'_ i version, player (i, a'_i) will which players are construct of the original sane, having the all game there will be one version of the the players and game, and the other players original game be a "main" version of the payoffs given in the original game. (i,a'_ { ), composed of two kinds of versions 1,2,...) that are We strictly c-perfect. for every pair over the pure strategy combinations will i have the payoffs given in the have payoffs that make the constituent, parts of a'_ x dominant is as in our previous constructions. In each elaboration, player "sane" or not and, for an elaboration To version of is We 1 is not, what are k th elaboration, his payoffs. This it — ne k then meets the requirements remains to specify the probabilities of the various will assign probabilities so that, in (for some versions in which all he in private types. specify the versions. if told whether he is i i > e^ is elaboration k , and the marginal probability to be determined in a bit), sane and the others not will be k use p (i,c'_i)tk to denote the probability of the (i,a'_ 1 u_ conditional probability of crazy types { , { ) the probability of the main ek More . version; that conditional on i precisely, is, p k we will (i,a'_ i ) is the being sane and the others crazy. Suppose that, and according to in elaboration their the probability that i bility I assesses that his - that everyone k (i, a'_i) and, for play according to a when they is + tkP k {i, di) l-(n-l)£* a'_ i , ^ i is is sane. ' We will want = €k /(l-{n-l)e k ) 25 is the joint proba- non-zero only this to equal , p (t,a_i) is sane and the others play yields <r_, being sane, ._,_. sane and the others play a'_ { (which i i are sane will play a'_ i is denotes the indicator function. The numerator in this expression the marginal probability that p all opponents n€fc)V plus the probability that only player for players , dominant strategies when not. Then conditional on (1 where k a'_ i ; if cr'_ i — cr_,- the denominator ^H^'-i) > ), is which, solving To ensure that ;/"(?', CTj) > , we must have that 1 That is, 1 we choose if hand k . > e^ By And since k construction, is <f> -#(*-,-) n-(n-l)#(<7_0 to be less or equal to the side of the previous inequality, well defined. is fc we must have €fc< So - ne k - (n - l)e 1 m°-i) > and use we k <f> minimum (over to specify the see immediately that a c-perfect test sequence, fc : (a_;) — » 1 i ) p a k is of the terms in the right as above, the elaboration strict in each elaboration implying that , ejt — , so the elaborations converge to the original game. Q.E.D. Finally, we have a converse to Proposition 6, exactly parallel to our converse to Propo- sition 4. An Definition. perfect if equilibrium a of an extensive form game k there exists a sequence of payoffs u — E with payoffs u u such that a is is quasi-c- strictly c-perfect in the corresponding normal form games. Proposition if it is , An near strict The proof i 7. is equilibrium in a of an extensive form game is quasi-c-perfect and only elaborations with independent types. just as in the proof of Proposition 5, except that Mai - 2 is replaced, for each by the conditional distribution of the actions of the others, conditional on first if type. 26 i being the References 1. Aumann, R. Subjectivity and correlation in randomized strategies, J. Math. Econ. 1 (1975), 67-96. 2. An approach F. Forges, communication equilibrium, Econometrics 54 (1986), 1375- to 1386. 3. D. Fudenberg and E. Maskin, Folk theorems for repeated games with discounting or with incomplete information, Econometrica 54 (1986), 533-554. Harsanyi, 4. J. Games with randomly equilibrium points, Int. 5. Kohlberg and E. J. J. disturbed payoffs: Game Theory F. Mertens, On A new rationale for mixed-strategy 2 (1973), 1-23. the strategic stability of equilibria, Econometrica 54 (1986), 1003-1038. 6. D. Kreps, P. Milgrom, J. repeated prisoners' dilemma, Roberts and R. Wilson, Rational cooperation in the J. finitely- Econ. Theory 27 (1982), 245-252. 7. D. Kreps and R. Wilson, Sequential equilibria, Econometrica 50 (1982), 863-894. 8. R. Myerson, Refinements of the nash equilibrium concept, Int. J. Game Theory 7 (1978), 73-80. 9. 10. R. Myerson, Multistage games with communication, Econometrica 54 (1986), 323-358. Okada, On the stability of perfect equilibrium points, Int. J. Game Theory 10 (1981), 67-73. 11. R. Selten, Spieltheoretische behandlung eines oligopmodells mit nachfragetragheit, Z. fur die Gesamte Staatswissenschaft 121 12. R. Selten, games, 13. Int. J. E. van (1965), 301-324. Re-examination of the perfectness concept Game Theory 4 Damme, for equilibrium points in extensive (1975), 25-55. "Refinements of the Nash Equilibrium Concept," Springer- Verlag, Berlin/Heidelberg/New York, 1983. 14. E. van Damme, A relation between perfect equilibria in extensive proper equilibria in normal form games, Int. J. 27 Game Theory 13 form games and (1984), 1-13. L R 0,1 2,1 FIGURE 1 2 L M R 0,1 1,1 2,1 FIGURE 2 (2 0) ' u L (1,1) -< 1 - • R 2 o (-1,-1) FIGURE 3 (1.1) u (0,0) i I i 1 R 2 (0,0) u 0,1) FIGURE 4 u U (1.1)-* (2,0) R. 1 M -*-• (-1.-1) u (0,-1) 1 (-1,0)^ -o- w (0,0) FIGURE 5 D 1 n-2 Lu = Ld 2,1 2,1 Ru 3,3 0,2 Rd 1,0 0,2 (*) (£) FIGURE C (tt) (i) (2,1) -* o • •^ FIGURE 7 (3,3) 1 (2,1)-* L (-1.1) R 1 •o 1 (3,3) R. *+ FIGURE S D u L 1 L2 1-26 hR 2 1-6 1-6 , , 1-26, 1-6 1-26 R1L2 2-36 R R2 2-26,-6 1-6 , 1-6 -1,-1+6 ,0 -1+6 ,-1+6 1 FIGURE 9 361 I C83 ^(1,-1,-2.9) --(-1,1,1.1) (-V2.2)-* (-2-3,-V2) FIGURE 10 MIT LIBRARIES 3 TDflD ODS 13 D7E