Backward Induction, Normal Form Perfection and Explicable Equilibria Philip J. Reny Econometrica, Volume 60, Issue 3 (May, 1992), 627-649. Stable URL: bttp//links,jstor.org/sic¥2sici=0012-9682%28199205%2960%3A3%3C627%3ABINFPAG%3E2.0,CO%3B2-A ‘Your use of the ISTOR archive indicates your acceptance of JSTOR’s Terms and Conditions of Use, available at hhup:/www.jstor org/about/terms.html. JSTOR’s Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the sereen or printed page of such transmission. Econometrica is published by The Econometric Society. Please contact the publisher for further permissions regarding the use of this work, Publisher contact information may be obtained at hutp:/www jstor-org/journalsveconosoc.html. Econometrica (©1992 The Econometric Society ISTOR and the ISTOR logo are trademarks of ISTOR, and are Registered in the U.S. Patent and Trademark Office For more information on ISTOR contact [email protected], (©2003 JSTOR hupslwww jstor.org/ Thu Jul 31 12:12:38 2003,Econometrica, Vo. 60, No. 3 (May, 1992), 627-649, BACKWARD INDUCTION, NORMAL FORM PERFECTION AND EXPLICABLE EQUILIBRIA By Pup J. Rey! ‘of Kreps and Wilson's (1982) notion of sequential rationality for exten sive form games is presented. The motivation behind this weakening stems from the ly rational behavior in subgames reachable only through lation of sequential fationaiy. Although the notion of weak sequential rationality ‘developed here is based upon extensive form consideration, it bears a close relation to Selte's (1975) normal form perfect equilibria. Also, a class of restrictions one might impose upon out of equilibrium beliels it suggesied, The weakest ofthese results from the sssumplion that deviatrs are expected utility maximizes. The strongest yields backward induction outeomes in generic games with perfect information. Lastly, iis argued, with the aid ofan example having imperfect information, that sequeatial rationality not the consequence of equilibrium play and the absence of incredible threats. Keywonos: Backward induction, sequential rationality, normal form perfection 1. INTRODUCTION ‘Tue THEORY oF GaMEs has evolved steadily since its beginnings as expressed in von Neumann and Morgenstern’s “Theory of Games and Economic Behavior” (1944), Among the major subsequent developments are Nash's (1951) equil- brium existence result and Selten's (1965) work on subgame perfection. Since then, even sharper refinements of Nash’s equilibrium concept have appeared (elten (1975), Kreps and Wilson (1982), Kohlberg and Mertens (1986)). ‘A common thread among these has been the backward induction. (We use this term broadly. In particular, it includes Kreps and Wilson's (1982) notion of sequential rationality.) This notion has had a long. history in game theory and dates back at least to Zermelo (1912) who used it to prove the existence of a non-losing strategy in chess. More recently, however, the arguments supporting backward induction have come under closer scrutiny. See, for instance, Rosenthal (1981), Basu (1985), Binmore (1985), and Reny (1985), Although their approaches are distinct, each concludes that backward induction logic occasionally (even often) runs into difficulty. Consider the definition of sequential rationality proposed by Kreps and Wilson (1982), Its most natural interpretation suggests that at any point in the ‘game, all players believe that all will play as sequential rationality dictates from that point on. In particular, suchis the case after a deviation (or any number of. deviations) by any player. What seems unreasonable about this requirement is, that it asks player 1, say, after having seen player 2 deviate from equilibrium "this work isa revised version of the second chapter of my Ph.D. dissertation. 1 am deeply grateful tomy advisor, Hugo Sonnenschein, for his constant encouragement and valuable comments, wish also to thank Andy McLennan for his detailed suggestions on an earlier draft. I received a8 well helpful comments from Jeff Banks, George Mailath, Martin Osborne, Ariel Rubinstein, Jorgen Weibul two anonymous referees, and a co-editor. Remaining errors are of course my own, ar628, PHILIP J. RENY play, to believe that 2's beliefs about what the others are playing are the same as player 1 had perceived them to be before observing 2's deviation, even though such a deviation would then indicate irrational behavior on the part of player 2 given 2's beliefs. Let us call this the “inference assumption” of sequential rationality. It would seem more reasonable for player 1 in this position, if he remains convinced that player 2 is a maximizer, to revise his opinion about what player 2 believes the others are playing in such a way that 2°s deviation from the ‘equilibrium strategy profile is a best response given these new beliefs? In this case, after 2's deviation, player 1 will no longer in general believe that player 2’s choices thereafter will be optimal against the equilibrium strategy profile. They will instead be optimal against 1's newly hypothesized beliefs about 2’s beliefs about what the others are playing. If, on the other hand, the deviation by 2 raises doubts as to whether 2 is a maximizer, then there is no reason for 1 to believe that 2 will be optimizing against any strategy profile, let alone the equilibrium profile. Hence, after a player has deviated from equilibrium play there is no reason for the others to believe (and similarly no reason for the theory to insist) that that player will play optimally against the equilibrium strategy profile in the continuation. In the sequel then, once a player is known to have deviated from equilibrium, we shall not require that the equilibrium strategy profile be common belief among the players.’ Indeed, how can it be when one of the players has demonstrated by a past choice that he is not playing as the equilibrium dictates. The player who has deviated shall be treated as a nonconformist (to steal a phrase from Luce and Raiffa (1957)) whose future behavior will be predicted by the theory for the benefit of remaining conformists, (or bystanding theorists). This predicted behavior need not be optimal when one takes the theory into account, but since by this time, the player who has deviated is believed not to know the theory, this is not relevant. The equilibrium strategies (which include predicted behavior of nonconformists) are assumed to be common belief only among remaining conformists. Allowing past behavior to categorize players as conformists and nonconform- ists runs the risk of creating opportunity for strategic manipulation, It must never be the case that knowledge of the theory gives any player an incentive to become a nonconformist. Thus, the theory developed here will be constrained so that first-time unilateral deviations are never in one’s best interests. (This property is shared by all Nash equilibria.) This differs from a theory like 2 this idea isnot new; see for instance Bernheim (1984), McLennan (1985), or Pearce (1988). [Note that it might be equally reasonable for player 1 to simply believe that 2s payoffs are eiflerent ‘than he thought to begin with. We assume, however, that any poteatial uncertainty over pasos Is already modelled inthe extensive form so thatthe payoffs, atleast at the level of the extensive form, ‘are not subject to uncertainty. The same goes as well for al other aspect of the extensive form (he ‘Bupber of players, the order of pay, the avalable choices, ete) “An event £ is common belief if everyone believes E, everyone believes that everyone believes E,.... Defined in this manne (a formal treatment can be found, for instance, in Tan and Werland 1985), may be common belief and false simultaneously. in 1976) definition of ‘common knowiedge, this is not possible. Once an event is common knowledge, i remains so since it ‘must be true. With the deGition of Tan and Werlang, an event may at some point be common belief and later be discovered false, thereby breaking the common belieBACKWARD INDUCTION 9 Rosenthal’s (1981) where players’ beliefs about their opponents behavior is potentially inconsistent with expected utility maximization. In such a situation, strategic manipulation may be worthwhile. ‘A second assumption included in the usual justification of sequential rational- ity is that deviations are independent across players and independent of the actions of nature: a deviation by i does not influence the likelihood of a deviation by j from k's point of view. Although one could entertain the possibility of correlated deviations, we are principally concerned with the effect of relaxing the inference assumption described above. Thus, we shall throughout continue to assume that deviations from equilibrium are independent across players. For a discussion of the appropriateness and possible consequences of correlated deviations, we refer the reader to Binmore (1985), Fudenberg et. al (1988), Hillas (1986), Kreps (1987), and Kreps and Ramey (1987). Note however, that for games with two players, the correlated deviations issue vanishes whereas the inference issue does not. Related to our assumption of independence is the employment of consistent assessments (see Kreps and Wilson (1982)) throughout the paper. justification for consistent assessments is thoroughly explored in Kohlberg and Reny (1991), where consistency (and therefore independence) is shown to be a consequence of any “complete” theory. ‘The paper is outlined as follows: Section 2 contains a list of the symbols used throughout. In Section 3, the inference assumption implicit in sequential ratio- nality is fully relaxed, leading to the definition of weak sequential rationality. A relation between this extensive form notion and Selten’s (1975) normal form perfect equilibrium is provided. Section 4 provides some examples which moti vate the restrictions on out of equilibrium inferences that follow in Section 5. Rather than attempting to provide a definitive answer to the question “What are reasonable out of equilibrium inferences?”, we instead suggest a class of restrictions one might impose. The restrictions in this class share the property of being “best response motivated,” in a sense that will be explained fully below. ‘The strongest restrictions in this class lead to backward induction outcomes in generic games with perfect information (Proposition 3). Backward induction does not necessarily result in games with imperfect information, however, and this is illustrated with an example in Section 6. More importantly, this example suggests that subgame perfection (and hence sequential rationality) is stronger than it was intended to be since (as we argue) it eliminates equilibria involving no “incredible threats.” ‘We shall restrict our analysis to finite extensive form games with perfect recall (Kuhn (1953)). The formal description of an extensive form game is by now standard (see for instance Kreps and Wilson (1982), Kuhn (1953), van Damme (1983). To the extent necessary, we shall use for the most part the notation given in Kreps and Wilson (1982), We list below with very terse explanations the630 PHILP J. RENY symbols we will require. This list also contains symbols not contained in Kreps and Wilson (1982) for which we will have need. These additional symbols are carefully defined in the text when needed. Symbol Meaning finite set of nodes < partial order on T such that Wre T(x such that a! is a best response against =, for all n and all i€1 Also, for i €1, s€S', and ' EIT’, define the probability of the pure strategy s nm! by: p(slar') = TT, «,-4y7(a), where s~'(1) denotes those choices in 4’ in weight 1 by s. Say that seS" is given positive weight by meII' it Pll’) >0. We now establish a relationship between weakly sequentially ratio nal assessments and normal form perfect equilibria, Prorostrion 1: Given a finite extensive form game, if m€II is a perfect equilibrium of the associated normal form game, then there is a system of beliefs such that (7) is weakly sequentially rational. 1 wish to thank Avil Rubinstein for encou strategy implied in the definition of weak sequent poipt. Min an earlier version of this paper, the meaning of “x € IT is a perfect equilibrium of the ‘sociated normal form game” was n0t provided. T wish t thank Jorgen Weibull for pointing out ted room for ambiguity. One might also consider perfect if a realization equivalent fect in the usual sense. With this definition, w might not be part ofa weakly re to emphasize the interpretation of tionality. Rubinstein (1991) makes a siiBACKWARD INDUCTION 633, Proor: Since w is perfect, there is a sequence (iq, 7)Ja1S¥° with Tq2 and 4, > for some # ©A, such that any pure Strategy s € S* given positive weight by 7! is a best response against 7, forall n= 1,2,.... We claim that for any such pure strategy 5 © S' and for all h © H' s-relevant, s is a best response against (u,77) at h. For suppose not. Then s is not a best response against (i, 7) at some h € HT s-relevant. Continuity of i's payoff in (y, 7) then implies that for n large enough, s is not a best response against (j1,,7,) at H. Hence, there is a #! € 1’ differing from s only on choices available at informa- tion sets belonging to i (weakly) following /, such that Evers ul(z)|h] > BM = 9[ul(z)Ih] But since both (4,27; and (s,77;") reach h with positive probability (1 is both sand #4-relevant and 7, is completely mixed), the inequality implies that # sives i a higher expected payoff than s in the associated normal form game against 7," Hence, s is not a best response against =, for all n large enough, a contradiction. This establishes the claim. Now observe that h eH" is w-rele- vant if and only if © H" is s-relevant for some pure strategy s € 5° given Positive weight by =. Together with the claim, this implies that (1, ) € W. OED. Note that one cannot substitute “Nash equilibrium” for “perfect equilibrium” in the statement of Proposition 1, since in games where no player moves twice along any path, W coincides with the set of sequential equilibria. Also, since in simultaneous-move games every Nash equilibrium is weakly sequentially ratio- nal, the converse of Proposition 1 is false.* However, analogous to the relation established by Kreps and Wilson (1982) between sequential and extensive form perfect equilibria (see their Theorems 1 and 3), it is the case that fixing an extensive form, for “almost all” payoff vectors, “almost all” weakly sequentially rational assessments are normal form perfect. A precise statement of this result (Proposition A) as well as its proof can be found in the Appendix. 4, FURTHER RESTRICTIONS If one is unwilling or unable to place any restrictions upon the future behavior of a player who has deviated in the past, then the previous section suggests that normal form perfect equilibria will produce entirely reasonable solutions in extensive form games. On the other hand, one feels that at least on ‘occasion certain restrictions upon inferences concerning the future behavior of, deviators suggest themselves. The converse holds if(+) is strengthened to Cte aremax B= ui(2)A] forall n, js any sequence in ¥” converging to (su, ). Thus, supplementing the notion of Tationality with form of cautious behavior (ak embodied in Selten-like trembles), provides an extensive form characterization of Selten’s (1975) normal form perfect equilibria,634 PHILP 3, RENY Consider, for example, the game in Figure 1. Here, there is no conflict. The players’ payoffs coincide at every endpoint and are given by the single number there. Now, if player 1 plays ay, it seems entirely reasonable that player 2 should play 4 anticipating that 1 on his next turn will choose a yielding both the highest possible payoff of 3. Indeed, ((a,,a,), A) is weakly sequentially rational. However, ((d,,d,), D) is also weakly sequentially rational—it is also perfect in the corresponding normal form game. The reason for this is that once player 1 deviates from ((d,,d,), D) by playing a, player 2 is allowed to believe, for instance, that player 1 is “crazy” and if given the chance will choose d3. Indeed, this is the prediction of the theory in this case as embodied in ((dy,d;), D). Believing this, player 2 would certainly do well to choose D, and believing all of this, it is indeed best for player 1 to choose dy. ‘Now, inferring from 1's choice of a, that 1 is crazy, when a, is consistent with the perfectly sensible continuation by 1 of a, is, at best, extreme. In this instance one feels as though the only reasonable inference upon seeing a, is to presume that a)—1’s unique best response at his second node—will subse- quently be played by 1. In this sense, inferring that player 1 will next choose a3, and not ds, is a “best response motivated” inference. Holding fast to the conviction that one’s opponent is playing some best response provides opportunities for forward induction as the (well-known) example of Figure 2 illustrates. In this example there is a proper (and hence sequential) equilibrium in which player 1 chooses OUT, yielding the outcome (2,2), However, suppose that player 2 explains having been reached in terms of player 1 playing some best response. Then player 2, if reached, would believe that 1 played —the only strategy for I reaching 2 that is a best response for 1 against something 2 might do. Consequently 2 will play /, the unique best response given these beliefs. But believing all of this, player 1 does best to choose L. ‘Thus, if 2's inferences about 1’s behavior are best response motivated and 1 believes this, then the outcome of the game of Figure 2 is unique, and must be GD. ‘Consider now modifying the game of Figure 2. Replace the outcomes (3, 1) and (1,3) by games G, and G, respectively, where G, and G are each games of matching pennies between two new players, 3 and 4. The payoffs to players 1 and 2 depend upon the outcome of these matching pennies games as follows. In Gy, player 1 obtains a payoff of 3 regardless of whether or not 3 and 4 match, WR At Fioune |[BACKWARD INDUCTION 635 ©) O @ @ Ficune 2 while player 2 obtains a payoff of 3 if they match and —1 if they do not. In Gy, the payofis to players 1 and 2 are reversed. Thus, the unique equilibria in G, and G, yield the outcomes (3, 1) and (1,3), respectively for players 1 and 2. Let us suppose that player 2's inferences about 1’s behavior remain best response motivated, and that player 1 believes this. How does this bear upon the analysis of the modified game? If, by best response motivated, it is meant that player 2, when reached, believes that player 1 chose a strategy which best response to something the other players (2, 3, and 4) might do, then # places no additional restrictions on beliefs beyond those of weak sequential rationality. The reason is that in the modified game, both L and R are best responses against some strategy choice for players 2, 3, and 4. As a result, the equilibrium in which player 1 plays OUT remains. (On the other hand, 2 may be unwilling to suppose that 1 chose a strategy that is a best response against just anything the others might do. Indeed, it may be much more likely, in player 2's eyes, that player 1 believes that the others are playing some weakly sequentially rational strategy. Accordingly, 2's inferences being ‘“best response motivated” may mean that 2 believes that 1 is playing a best response to some prior over the set of weakly sequentially rational strategies. If this is the case, and 1 believes this, then the only surviving strategy-pair for players 1 and 2 is (L,/), with the associated expected payoff of (3,1). The reason for this is that every weakly sequentially rational assessment involves players 3 and 4 cach choosing heads and tails with probability one-half. Consequently, R is not a best response to any prior over the set weakly sequentially rational strategies, whereas L is. The result now follows as before. ‘The point of all of this is to illustrate that assumptions regarding the beliefs ‘that deviators hold about the play of others determine whether or not one strategy or another is a best response for the deviator. And this in turn, according to our eventual definition, will determine whether or not one infer- ence of another concerning the behavior of a deviator is “best response motivated.” ‘One way of restricting the beliefs that a deviator (or any other player) holds about the play of others is illustrated in Pearce’s (1984) notion of extensive form636 PHILIP J, RENY rationalizability. On a first pass, deviators, indeed all players, are assumed to be playing a best response given some prior over the strategies of the others. Each player maintains this assumption so long as the information he possesses about the actual history of play does not contradict it. That is, this assumption is ‘maintained whenever possible. This restricts the set of strategies to B, CI say. But if it is clear that players are playing some element of B,, then whenever Possible shouldn't all players be assumed to be playing a best response given some prior over elements of B;? If the answer is yes, and it is from the point of iew of rationalizability, then this further restricts the collection of strategies to B,, and so on. Now, suppose that for cach k> 1, it is clear that players’ strategies are restricted to elements of B,, and itis also clear that as a result it should be assumed (whenever possible) that all players are playing a best response given some prior on B,. Then, for each k > 1, the future behavior of a player who has taken an unexpected action should then always frst be explained in terms of a best response given some prior on B,.,., if possible, rather than on B,, The basic philosophy is that even though player j, say, may have played in an unexpected manner, the others bestow upon j as “high a level of sensible behavior” as is consistent with the history of play. Consequently, if playing a best response to some prior on By... is deemed more sensible than playing a best response only to some prior on B,, then inferences about j’s behavior will be in terms of the former, whenever possible, rather than the later. Similar iterative approaches to restricting out-of-equilibrium beliefs have also been employed by Banks and Sobel (1987), Bernheim (1984), and McLennan (1985). The first and last of these differ sharply in spirit from Pearce (1984) in that the standards of sequential equilibrium are taken as basic and their iterative procedures are employed to provide a refinement of it. Moreover, in the case of Banks and Sobel, attention is restricted entirely to signalling games. In order to refine weak sequential rationality, we too shall employ an iterative procedure to suggest restrictions on beliefs regarding the play of deviators. One manner in which this procedure differs from each of the others above is that there is some flexibility in the first iteration. This fist iteration may be as weak as Pearce’s and ask only that players believe that deviators are choosing a best response to some prior over the entire set of strategies; or as strong as asking that players believe that deviators are choosing a best response to some prior over the set of weakly sequentially rational strategies. As the modified Figure 2 game illustrates, the difference can be striking. Subsequent iterations are then sly described (a formal treatment appears in the next Before proceeding, it is worthwhile to point out a tension between the backward induction method of solution, and the forward induction logic that is central to the most basic of the restrictions discussed above. Consider the game of Figure 3. The unique backward induction strategies are (d,,d3) and (Dj, D;) for players 1 and 2, respectively. Let us assume that the players believe, whenever possible, that their opponent is playing some best response. In the iterative procedure described in the next section, at least this much is always assumed.BACKWARD INDUCTION 637 () () () () () Froune 3 Suppose now that player 2's first decision node is reached. By assumption, 2 must then believe that 1 is playing the strategy (a,,a,). For the only other possible pure strategy is (a,,d;), and this is strictly dominated by (d,,*). Consequently, 2's unique best response, if reached, is (4), D,). Of course, if player 1 believed all of this, he would choose d,. Note then that because player 2 explains 1’s behavior in terms of expected utility maximization, 2's backward induction strategy, (D,,D,), is rendered “irrational.” It does not provide 2, when reached, with a best response. Note also that although the backward induction outcome happens to obtain, it does not obtain by virtue of the usual backward programming argument—after all, backward programming would dictate (D,, D3), not (4,, D3), for player 2. Despite the inappropriateness, indeed the inapplicability of the usual back- ward programming argument in the presence of best response motivated infer- ences, it is shown below that in games with perfect information and no indifference among terminal nodes, the iterative procedure described in the next section—which incorporates such restrictions—nonetheless isolates the unique backward induction outcome. 5. BACKWARD INDUCTION AND EXPLICABLE EQUILIBRIA For ACY, let A" denote the projection of A onto II, and for A CII let co A denote the convex hull of A, Let B be a subset of the set of behavior strategies, IT. We shall say that a strategy m‘ €II' is a best response relative to BCI/ if there is a » © and an element, y, of co B with w derived from y using Bayes rule when possible, such that for every h € H' srrelevant, given 1, cris a best response against y at A. In keeping with the desire to provide inferences that are “best response motivated,” we shall suppose that at each information set, whenever possible (ie. so long as the information provided there allows it), each player believes that the others are playing a best response relative to B. Thus, B represents, in a sense, a common standard of behavior recognized by all players—deviators and nondeviators alike. Now, if the particu- lar (modelling) circumstances indicate that a single deviation by any player renders the future behavior of that player entirely unpredictable, save for expected utility maximization, then setting B = IT is appropriate. On the other hand, one may wish to consider environments where a deviation is not consi cred an indication that the deviator does not conform in any way to the638, PHILIP J. RENY standards, say of Nash, but that the deviator is simply mistaken as to which Nash equilibrium is “in effect” (similar to McLennan (1985)) and is perhaps playing a best response against some other Nash equilibrium. Here, one would restrict B to Nash equilibrium strategies. Similar considerations upon replacing Nash behavior by weakly sequentially rational behavior would warrant putting B equal to the set of weakly sequentially rational strategies. Thus, a choice of a common standard of behavior BCIT is a modelling decision which captures part of the nature of a deviation. A deviation can in principle be a calamitous event leading to almost complete unpredictability concerning the behavior of the deviator (in this case put B =/T) or it may be viewed less severely and as such a more parsimonious explanation for it may be sought (by putting B = W" say). Recognizing that there is no “correct” choice of B, only different ones, each contributing to the interpretation of a deviation, we allow the analysis that follows to depend upon the particular choice of B.’ So, suppose that (1,77) is weakly sequentially rational, + €B, and that for J#i (@) hEH! is not relevant (hence if reached, j must have deviated), (b) A is both s and s'relevant for distinct pure strategies s,s’ ©’, and (6) 5 is a best response relative to B while s' is not. Under these circumstances, if A is reached we shall restrict ’s inference about j’s behavior to give zero weight to 4’ and all realization equivalent strategies, the set of which we denote by [5] ‘The reason for this is that by hypothesis i believes, whenever possible, that jis playing a best response relative to B. Since s, and hence every member of [3], is consistent with h being reached, it is indeed possible for i to believe this. Consequently, # must place zero weight on j having played any member of [s'] Now, according to the weakly sequentially rational assessment (4,77), infer- fences about the future behavior of our deviator j above, from i's (a nondeviator's) point of view at he H', are contained in 2! restricted to those actions at information sets f' ¢ H? following h. Player i's inference about j's (as well as others’) past behavior are contained in i's beliefs over the nodes in H’ as described by .. Since h and any information set of ;'s following it are reached with probability zero given 7, the weight placed upon [s’] relative to [3] is determined by the particular sequence {(y1,,7,)I;-1¥° converging to (4,7) and is given by (20) tim lees) noe ((s]m) If this limit is zero, we'll say that the inferences determined by 7 €B are best response motivated relative to B. Let B, denote the set of all such behavior strategies. That is, + © B, if + €B and zis the limit of some completely mixed "In fact, in what follows we allow any B satisfying W"" CB CIT. This strengthens Propositions 2 and 3 since the set of explicable equilibria is nondecreasing (ordered by set inclusion) in B (ke insisting that B= 17 weakens Proposition 2 and yields the strongest version of Proposition 3 ‘whereas insisting that B= 17" has precisely the opposite effect). Also, the discussion in Section 5 ray be somewhat less convincing if t relied on B= IT. t does ot. In an earlier version of this paper, B= 17" was imposed. I owe thanks to an anonymous referee who strongly encouraged that {his be weakened,BACKWARD INDUCTION 69 sequence of behavior strategies (n,) such that whenever for two distinct players i, j and any hE H',s, s' €5!, conditions (a)-(c) in the preceding paragraph hold, the limit in (+ *) is zero. Thus, in the game depicted in Figure 1, the inferences determined by = are best response motivated relative to IT if and only if a4a3)=1. But one may wish now to go even further. If it is reasonable to presume that all players’ inferences are described by some element of B, CB, then perhaps it is also reasonable to presume that this is common belief. One is now tempted to treat B, exactly as B earlier. B, represents (as did B) a common standard of behavior. Again (as before) there will be those s € S' which are best responses relative to B, and those which are not, and one is tempted to insist that inferences determined by mB, be best response motivated relative to By. This will produce a B,0, and ie 1 define Cy, ,, Sj as follows: Given for all J and k’=0,1,2,...,k Cy and Shay let 7 0 aefes\ @) Gy {om €C,3{(uq,7,)} © Y°, such that: I is not a best response relative to Ci! (8) (Het) > (He) () We(0.,...k),ienitses'\ Y sjand ses} (Ls) | . then ie To see that this does indeed summarize the iterative procedure described above, note that C,,, by virtue of conditions (jiXa) and (b), contains those consistent assessments whose associated strategies are best response motivated relative to those of C,. These strategies are identified by condition (i) It might bbe wondered why the iterative definition above involves sets of assessments (belief-strategy pairs), rather than simply strategies alone. After all, the infer- ‘ences we wish to restrict are merely strategies. The reason for this is that, in the end, we are interested in the assessment which results subsequent to the implied restrictions upon inferences. Hence, we must ensure that inferences about future behavior, as expressed in 2, are compatible with inferences about past behavior, as expressed in y2. This is accomplished by consideration of the full assessment, and in particular by condition (iiXa).. ‘When is it reasonable to apply successive stages of this iterative procedure to refine inferences? We only wish we could offer a definitive answer. Note,640 PHILIP 3. RENY however, putting By ~ Ci!, that at the kth stage, condition (ii) is an expression of the following situation: Vj € (0,1,...,), everyone believes that every ele- ment in B, is considered by all players to be infinitely mote likely than every element in B,_;\B;. Consequently, a deviator’s behavior is always explained first as a best response relative to B, and only if this is impossible explained as a best response relative to B,_, and only if both of these are impossible...and only as a last resort explained as a best response (if possible) relative to By. Now, for larger and larger values of k, one may be less convinced that this situation actually holds. In particular, the very careful lexicographic ordering of the likelihood of elements in the sets B, and B,_\B, for all k > 1 shared by the players regardless of the history implies that deviators and nondeviators alike share a fairly detailed theory of behavior; one that becomes more detailed as k increases. Nevertheless, one can in principle choose k to be arbitrarily large, even infinite, The next proposition establishes that there is always some descrip tion of behavior that is compatible with any number of iterations. Fix B CII. Call (u,7) kth-order explicable (relative to B) if (uy7) WC and explicable (relative to B) if this holds for all k>0. Call B admissible, if W'CBcIL. Prorostmion 2: For all admissible B, there is at least one explicable equilid- rium. Moreover, there is a finite N so that every Nth-order explicable equilibrium is explicable. Proor: See the Appendix. To see how explicability differs from weak sequential rationality, one need only note that every Nash equilibrium is weakly sequentially rational in the game of Figure 2, yet only (L,1) is first-order explicable (this is clear from the discussion in Section 2) and hence, by Proposition 2, explicable—and this for any admissible B. In addition, for the modified Figure 2 game, both (L,1) and (OUT, r) are (among others) explicable equilibria when B = IT—we're ignoring player 3 and 4’s unique matching pennies equilibrium strategies. However, only (L,D is first-order explicable, hence explicable, when B= W Tis not our intent to convince the reader that the iterative application of best response motivated restrictions upon inferences described above is a basic ingredient of an extensive form solution concept. The principles of expected utility maximization and equilibrium behavior lead only to weakly sequentially rational behavior in extensive form games, and it is this that we consider basic. On the other hand, the example of Figure 1 leads one very naturally toward best response motivated inferences. But if it is natural to presume that all players (even deviators) are expected utility maximizers (relative to B), itis difficult to ly that under no circumstances is it the case that in addition, all players believe this (including deviators). Thus, once one considers the possii ity that players’ inferences are best response motivated relative to B (and without this we must come to grips with the example of Figure 1), leading to theBACKWARD INDUCTION 641 standard of behavior B,, one is compelled to consider the consequences of allowing players’ inferences to be best response motivated relative to B, (ielding B,, etc.) Just as the choice of B is a reflection of how one interprets deviations from equilibrium, so too is the level at which one ceases to apply the iterative process described above a reflection of the beliefs and behavioral characteristics all players are considered to have in common. In our view, itis equally inappropri- ate to say that deviators and nondeviators always have as much in common as is needed to support only explicable equilibria, as it is to say that they have as little in common so that one must always consider every weakly sequentially rational assessment. There is, however, a connection between the behavioral and belief characteristics shared by deviators and nondeviators, and backward induction outcomes. As the following proposition shows, the best response ‘motivated restrictions upon inferences embodied in explicable equilibria lead to backward induction outcomes in most games with perfect information, Proposrrion 3: Given any finite extensive form game with perfect information and no indifference among terminal nodes, if B is admissible, then the unique explicable equilibrium outcome is the (unique) backward induction outcome PRooF: See the Appendix* At first glance, this result may seem hardly a surprise. There is, however, something to show here. For recall that (as in the example of Figure 3) the choices dictated by the usual backward programming argument need not always be among those dictated by explicable equilibrium. And in general, discrepan- cies with backward induction play off the equilibrium path may well lead to discrepancies along it. Thus, it must be shown that the discrepancies with backward induction play inherent in explicable equilibrium are only off the backward induction path and do not lead to discrepancies along it. One should note that without perfect information, explicable equilibrium outcomes need rot be subgame perfect (we provide an example below). Also, the number of iterations involved in generating the unique outcome alluded to in Proposition 3 grows with the number of available pure strategies and hence with the game's length. Thus, if one feels that Wy is less plausible for larger values of NY, then at least from the point of view expressed here (in terms of best response motivated restrictions on inferences) the backward induction outcome becomes less plausi- ble in particular, as the length of a game increases. "The only proof of Proposition 3 the author has managed to come up with appears to employ tools that are much more powerful than necessary. In particular, a property of fuly stable sets (see ‘Kohlberg and Mertens (198) is used. One would expect that'a rather straightforward induction argument on the length ofthe tee ought todo the irik. Tt does not. The reason is that as shown by the example of Figure 3, the continuation payofls corresponding to nodes not along a particular explicable equlbum path do not generally coincide with the backward induction continuation payotls. With alittle thought the reader willbe convinced that this eflectively precludes carrying out the desired induction step.642 PHILIP 3. RENY 6. THE RESTRICTIVENESS OF SUBGAME PERFECTION: AN EXAMPLE Consider the example given in Figure 4. The unique subgame perfect (and hence sequential) equilibrium is (IN, (1/2)L),(1/2)). Thus, sequential ratio- nality leads to @ unique equilibrium. We wish to argue however that other (weakly sequentially rational) assessments are also entirely reasonable. Let a denote the probability weight that player 2's beliefs place upon the node labelled x. Now, consider the following conditions on any refinement of Nash ‘equilibrium applied here: (i) a is consistent with Bayes rule when possible, and Gi player 2 maximizes expected utility at his information set given a. First, one should note that both (i) and (i) are implied by sequential equilibrium (and therefore any of its refinements). Second, it is easy to see that i) and (i) yield a> (1/2). For suppose, to the contrary, that a <(1/2). Then by (i) player 2 must choose / with probability one. Since in any Nash equilibrium, player 1 must play a best response against 2's strategy, player 1 will then choose (IN, L), the unique best response against /. But (i) then implies that a= 1>(1/2), a contradiction. To recap, any refinement of Nash equilibrium satisfying (i) and (i) yields « > (1/2). But, whenever, a >(1/2), r is a best response for player 2 when reached. Hence, r is a perfectly credible choice for 2 within any refine- ment of Nash equilibrium satisfying the two basic requirements listed above, But if 2 can credibly play r as a best response given any admissible beliefs he ‘may have when reached, then player 1 has no reason to doubt this from the start and with such beliefs about 2, 1 should respond optimally with OUT. Hence, we conclude that any assessment of the form (OUT, aL),r) with a >(1/2) Q's beliefs as well as 1's probability of choosing L at his second information set are given by a) is entirely reasonable. It is easily verified that all of these, in addition to the unique subgame perfect equilibrium, are explicable. So, although subgame perfection and sequential equilibrium do indeed rule out incredible threats, they (and hence all other standard refinements) rule out perfectly credible choices as well. For instance, r chosen with probability one isBACKWARD INDUCTION 643 ruled out in the game of Figure 3. Consequently, reasonable equilibria are unintentionally eliminated. If one objects to (COUT, aL),r) a > (1/2) on the grounds that once IN is in fact played, L is not a best response against r, we remind the reader of our interpretation of a strategy. In this instance, OUT represents 1's rational plan in light of the theory. Since 1's second information set is precluded by this plan, aL is a prediction supplied by the theory to render a decision by 2 possible if reached. (This prediction also ultimately renders a decision by 1 possible at the start since 1 must know how 2 would react if reached.) Now, since 2's plan is to play r if reached, then if he is reached, either 1 is unaware of 2's plan, or 1 is not an expected utility maximizer. In either case, there is now no reason to expect 1's second choice to be a best response against r, and therefore no reason for the theory not to predict that 1 might play L. Hence this objection is ultimately untenable. ‘One might also object to this example as a criticism of subgame perfection and sequential rationality because of the fact that the equilibria ((QUT, aL),r) a> (1/2) become sequential (and therefore subgame perfect) when player 1's two information sets are coalesced into one so that the extensive form has the ‘same shape as that of Figure 2. This line of reasoning itself, however, leads to a rejection of both subgame perfection and sequential rationality. For it implicitly ts that coalescing player 1's information sets makes no strategic difference, ‘a property that this example shows sequentially rational assessments do not have? Finally, the example indicates that from the point of view of explicable ‘equilibria, substituting independent agents for a player at each of his informa- tion sets (as in Selten’s (1975) agent normal form construction) does not leave the strategic nature of the game unchanged. Replacing player 1 with agents 1a and Ib at 1’s two information sets and treating these as separate players who happen to share the same payoffs, yields the unique explicable equilibrium {IN,/2)L,1/2!). The reason (OUT, aL,r) a> (1/2) now fails to be explicable is that if 1a were to deviate and play IN, we cannot conclude that 1b is unaware of 2's intended choice of r. On the contrary, since 1b has not deviated he is assumed to be aware of 2's intended choice. Hence, 1b should ° That sequently rational assessments are not “invariant” to coalescing information sets in this way was pointed out by Kohlberg and Mertens (1986). Teit “stable” set of equilibria are invariant not only with respect to coalescing of information sets, but any transformation leaving the game's fedaced normal form unchanged. The unique stable set of the game in Figure 4 isthe singleton set ((UN;1/230,1/2L)). Hence, stabiliy also rules out equilibria we consider entirely reasonable, land does so in particular whether or not we coalesce 1's information sts. Although we have not proven tis, we conjecture that explicable equilibria are invariant to Thompson's (1952) "coalescing ‘of moves information sets)" and "interchange of moves” transformations (Thompsons transforma- tions are described in Kohlberg and Mertens (1986). Simple examples involving weakly dominsted rtegies show that explicable equilibria are not invariant to Thompson's “adltion ofa superftous move" transformation, However, one can modify the definition of explicabily in a way that Incorporates a form of cautious behavior to obtain invariance in Kolberg and Mertens” (1986) Sense. This is done in Reay (198).644 PHILIP J. RENY choose R with weight one, his unique best response against r. But this of course upsets the candidate equilibrium.” Dept. of Economics, The University of Westen Ontario, London, Ontario, Canada N6A SC2, Manuscript recived June, 1988; final revision received October, 1991 APPENDIX ‘This appendix contains the proofs of Propositions 2 and 3 as well a both a statement and proof of partial converse to Proposition 1. We sall begin with the later. ‘We first remind the reader of some definitions from Kreps and. Wilson (1982). A basis for an extensive form i collection 6, of nodes x EX and actions a € A. Define Vy = (4) € Wiqu(x) > 0 If and only if xb, and z(a)>0 if and only if a'€D), and Ml, (x € IH|(a)>0 if and only if 4&6), We wil need to consider basis pairs (8), each pair being denoted by c. Given basis pit = (0,8) define “A, = (jr:r,)= Vs x lly I is realization equvalent to”). The finite set of consistent basis pars contains all bass pais c such that A, is nonempty, and is denoted by C. INote that for any r,# ©, REL" ie relevant if and only if it i also relevant. Let Hj=(heH\h is esclevant for some (al) + ET, and let TH, = Uje rH Also, note that Ar#2 with ¢= (6,0) implies Hy = Hy. For a fixed extensive form, W the set of weakly sequential rational assessments depends upon the vector of joint payors u © RY *!.We denote this dependence by Wu). We'll say that a © ACs) is. best response agninst (a,x) at = 1" there isa #"€ IT" a best response against (i,m) at with #(a) > 0. Now for © Cle Wi(u) = ue) = GW) Wy GD if he Hy Hy, then @ ACH) best response against (47) at hy Git h © H\ My, then a ACH) Hf and only if a Is 3 best response agains (uy) at A), Note thatthe collsction of W-(0)'s partitions IM). Further partition Wu) into Wu) and We"), where (yx) Welw) if and only if for all 7 and al Wed" rérelevant, any action a © A(H) that isa best response against (1,2) ath is taken with postive probability, and where Wu) consists ofthe femaning elements Of (0. Prorosimon A: For all payoff ue R*2'*” except posibly a et of Lebesgue measure zero, and for even e=(b,BVEC, W(u) iether empry or ia manifold of dimension n(c) = dim ¥y ~ #(W OA) + #11, W(1) is contained in the Yorlative frontier of Wu), (Gi) every (a2) © HE(u) is nomatform perfec. Poor: By Lemma 2 of Kreps and Wilson (1982), henceforth KW, and the definition of A., for every c EC, A, i manifold. We now follow Section A.2of the Appendix in KW modifying where necessary. As there, each player 11 is thought of as having his own decision tree (T", <), and D! fs ast of functions, each assigning a number to every node in 's decision tree, with the constraint that ze is assigned to certain prespecified nodes (See KW, p. 891, for the detail). For each ceCy(u.z,7)eA,, and weU" (UR? and’ U= XU, define a map Fe ALxU! oD? 8 follows Fu. Bw) = (ae Ny) forevery yer’, where >i exaetly as defined in KW, p. 891, equations (A.4Xa) and (6). By KW's Lemma AS, fixing 1 am grateful to Jeff Banks for suggesting this comparison with agent normal form ideas Indeed, the comparison goes beyond explicable equilibria, Any equlbrium concept incorporating some form of forward induction (see Kohlberg and Mertens (1986) will be sensitne to the introduction of independent agents. The independence severs the link between past and future ‘behavior, and hence fenders impotent the argument upon which forwaed induction Is based.BACKWARD INDUCTION 645, (ir. 3) 6A, 7 isa bijection fom U! to D! Detine BI: A, xD! U! to be the asocated invene” of F Let D= X,qyD! and let $ DXA,—+U be given by dsu.,2)= X 0 Bd asr:#) We now mimic KW from Lemma AS onward, replacing where necessary by 5 $y, af, and 5) by Wu, Wu, and W."(u), respectively, and Y, by Aq. partial ‘re obtain (analogous to KW's Lemma A7) the following (a2) © ¥ isn Wu) i and only ome BOB'NA, (is) © pry, (Mu) [Dy % cD. Moreover, WEG) = pry, Mu) > [Dyia% AD, where for BCA, Dy is defined exaety asin KW, p, 892. ‘Now, as in KW's Section A3! we shall call we U nice if for exery consistent basis ps c= (byb)EC and every subset of A such that 2b’ OVA the map @ restricted to the Joma ‘Dy xc has was a regular value. Otherwise 1 is called notice. Since the number of such pats (2p) is Site, Sends theorem guarantees that the set of not ice w has Lebesgue measure 2270 Ficow any nce w €U. As mentioned, (0) = pry, 8" UDp na AD. Because ui nice and therefore regular value of # rested 19 Dyna Aos $ GI OUDyrn4% Ac) is ether Empty or is a manifold of dimension d= dim Dy,» dim Ac dimU = #2-#1~ 0 A) ~ $i) dim A, #2: #1 = die A, MU 0A) + HEE . ‘Now, fom the definitions of 8: Dyna and Aq if (dn, €B-MwdOLDy na XAG then 90 fs Gdyjez,#) for any My realization equivalent to . Hence, for every such fined (d, 1.) wwe have precisely Eye yy,y{®8'O.ACh))~ 1) degrees of freedom remaining in Bu) [Dyna Aeh and Bu) ALDy a 4% Ac 1 8 product space. Its projection onto DX Vp is @ rmanlidls of dimension. d ~ Eqciyysl@(®' AU) ~ 1) But since dim A, = dim Vy + pcan OA) hand Ao ins Hy Hy We Bd Enc lA A) =I) Fix # € My, For 3 My let € My denote the joint strategy that is realization equivalent 10 ‘x and otherwise identical with 2, (This well defined since p= Hy.) Now, to prove (of the Proposition, not that difeomorphism carrying some neighborhood of Bojy xg, @"w) [Dying XA into RYO, then "(uy ) = (u,2, 1), 8 a dilleomerphism carrying the ‘Yoepjecton of ths neighborhood iio R Part (i) follows precisely sin KW wsing our analogue of ther Lemma 47. Par Gi follows KW proot of Theorem 3p, 89, Wherein addition to the substitutions already mentioned, A? is used instead of V9, where ALC (uy, 9)€P° IT°| forall AC Hy = Hy and Alla a(t), 0%q)= 74a) ia manifold contracted (asin KWs Lemma A3) 50 that APU A, is a manifold with boundary equal to A. OED Pnoor of Prorostrion 2: The method of proof here employs in a fashion similar to Kreps and Wilson (1982), “trembles” as introduced by Selten (1975). On the other hand, the sequence of trembles must be chosen carefully, 30 that our existence proof i closer sil to those of MeLennan (2985) and Myerson (1979). Let M° be the set of joint completely mited strategies, Then m G.Af° gives we 468" for 1€f, Ir eI is realization equivalent to me M? we have (At) p(ls}a)=m(s), Wer ses Now, define for integers m,k> 1 MZGM? as follows: Mp= (meMIViE!, seS',mi(s) 0-4) (Next, given Sj forall /< 1, j>0, define for every k,1 > 1 the correspondence Gf: Mf —+ Mj by saying th € Gm) iff for every 4 0 iter sof (7h 1.405} and nt et repo ese Gi) 55" \ UL Sfand sis not a best response against m implies si(s) =n =} cay ses \ UL ssimptin w'(2) 20 ro646 PHILIP J. RENY Cleary, form large enough (say, above 2) Gx(m) for all k> 1 and m Mf is nonempty. For the rest of the proof we restrict attention to m >. Its straightforward 10 establish that for k> 1 and n> Mf is compact, convex and nonempty, and Gf is upper hemicontinuows, convervaled and ‘nonempy-valued. Hence by Kakutan’'s (1941) fixed point theorem, fing & > 1, for every n>", there exists m, © ME with m, = Gi(m,) Now, since'm, =M° for all n>, there is a unigue realization equivalent €I1 (this uses perfect recall and’ Kuhn's (1953) theorem), and a Bayes rule induced , © 4. With this construction in mind we shall write m, = xq. Without ioss we assume that (jy) "> Gu.) € ¥. We now show that (u's) €Wy, Since my € GIy), ifs © hot a est response against my (or equivalently), then for every in (sf, mids) 0 implies that si a best response against! for all large enough. Since this holds for fll /e/ and 8%, is perfect equilibrium of the extensive game's asocated normal form. Proposition 1 implics that iu.) © W'= WC Cy= Wo, where the ist equality follows since W!™ CB and second by definition of ‘We now show inductively that (17) € Cy for €(0,1,2,.orK) $0 that (n,m) € WAG, = for all R> 1 Suppose that for "0,1... Mh (ras fom above sais G pty Is) sitesi, ATED 9 OT (LF (ie. Gu, 9) €Cp). We'd ke to show that (A.2) alo olds for '=/+ 1,30 that (uy) € Cp. 80, choose = #25"\ Ufa, and st tis being the only nontrivial ease to check. ‘Since (2) yet) © Ch, and s €S1, there isan h =H selevant such that sis nota best response ath against (2) and hence against (iq, for m large enough (sy'n >M (> 7) Consequently, #5 not a best response agaist for'n >A Bul m= pis thats © Sf not 2 best Fesponse against my © GfCm,) for elds mits) = 0-990) and similarly #]©5"\ US} implies mifs)>n-O-P, for nf. Since Gp places identical restietons upon realization equlvalent strategies, we therefore have (eclling 1) fot (2) we (o,1, (Lean) _ midleD) - (Use) macCD) ~ cn hard th coma fied hat fe. Henn (2). Stace 1 ‘was arbitrary and B, ©, for all K> 1, we've shown that 7» @ forall > 0. “To show that 14) » © and fo establish the second Clim ofthe Proposition it suices to show that there i finite N’ such that 7-H = 2-H, = Bry (where the lst equality follows since Wk, GW, forall k> 0). But the definition of Wy assures us ofthis if for some finite N, 2 forall (1. Finally, SiS" and S}e\Si= 0 for (€1, j #7, together withthe finiteness of 5" implies that such an N exists OED. Proor o Prorostrio 3: The proot i split into two parts, Part I shows that for any extensive 1024) can be supported by some equi 10 indifference among terminal nodes Ghencetorth ind Mertens" (1986) sense, a fully stable ‘outcome exists. Since every fully stable set contains a subgame perfect (in fact sequential) equilibrium outcome, the only lly table outcome in simple games is the unique backward induction (subgame perfect equilibrium) one. Part I thus implies that in simple games, atleast one explicable equilibrium generates the backward induction outcome. Part Il completes the proot by Showing that in simple games there isa unique explicable equilibrium outcome. (Fully stable outcomes are supported by explicable equilibria Let F be a fully stable set generating a fully stable outcome. By the proof of Proposition 2 we may choose NV so that Sy = for all 17. Consequenty, Wy is the set of explicable equilibria. Consider now the following compact, convex polyhedron P* contained in the interior of the simplex of joint mixed sirateses meP" itor allel, for some j€ (0,1,....N G ——se8!\ YU sfimpties mi(s) > 0— } 55}, then mi(s) an OF,BACKWARD INDUCTION 7 Cleary, P* converges in Hausdorff distance tothe unit simplex as >, and comparing P* with the definition of GF fom the proof of Proposition 2 indicates that my isa Nash equlrim ofthe fame restricted to P" if and only if mye Gn). Now, F fully sable implies that there is a Sequence (mr of equilibria ofthe game tesrted {0 P* converging to some me F. AS noted, ‘my Gin) forall n Bat mg my fr some wy © M1, and (i, tne © W maybe assumed to converse 10"(u, EW. But ‘ance’ m, © GRCrny) for all, the proot of Proposition 2 yields Ghim)e Wy. Hence (1,2) isan explicable equitium yielding the same outcome at m © F, which ‘8 ofcourse he fully stable outcome generated by F. 1 Gimple games have uniave explicable equilibria): Asin art, choose WN so that Si= 9 for all fe Hence, Wy isthe set of explicable quia. Letting Fr denote the given simple ume and ‘ecaling that denotes the st of nodes, conser the subset of nodes: tT {re Tim > A with p,(t) defined, p,_(¢) is s-rlevant for some 5 « siPa"? \ U sn Using the restriction of < to 7, iti clear thatthe origin of 7” isthe origin of Alo, if 2 isa terminal node of Pt sa terminal node of. For suppose not, Consider the path irom the origin tog. Lety #2’ be the largest (with respect to <) node on this ath such that o(p(Q)) = 2) It Sich ay exists, then by definition of 1 there is an 9 SHH YNGISHOAD ao that y is “relevant. But 2” not terminal in T then implies that ther sav € T immediately following 2” with © elvan, implying 0 7, a contradiction. If no such y exits, then cleariy the same contradic tion arses, since S*°\ U #353 w 2 by Propasition 2, Hence, terminal nodes of "are terminal nodes of 7. ‘Now, suppose (by way of contradiction) that there is more than one explicable equilibrium outcome of Since forall 7 every F©5" shen postive weight in some expicable equim in'S*\ Ut this implies that T”eontains atleast two terminal nodes. Hence, thee is a tT” having at fest two unordered successors, u and sy, and such that all successors of have at most ‘ne successor (ce Figure AI below). Moreover, we claim that this implies that for aay © CU. the induced probabiliy that 2, (the terminal node following u) is reached given that u has been reached, fone. The same f te about, (he terminal node follwing 2) shen that hasbeen Feached. To show this we sallbe content to prove that a(x) (he last action taken to reach 2) is ten weight one in every w= Cif (se the igure). The argument ithe same fr all actions alon the paths ftom u to z, and ¢ to z,. Take Cf and is corresponding sequence in ¥ (Gramlgay. Now, since 7, is complciely mixed we have (where all summations are over pure648, PHILIP J, RENY ma EL w(sinite) / E o( simi) sso Seon E__ vl stes) PSone E_ vlsteg)s EL wl stas) abide nots 1 since 0 es (shi) Fn MND LE MENS | Finn EY were the inequality follows sine 5, €.% Uj chosen such hat «i eevee {Ext y detnton oT) and een’ s Ssh at wis rvelevant and nor US'S" by detnton of 7 andthe parca subgame we have cone. Hence the ‘ithe term term it ofthe sum er by dfn of) Thus as is gen weit one ‘in every 7 € Cj, which establishes our claim. Nov by asion of 7s a,c'© 7 py that there ei 5, 5! U 5154 tc tha sefeleant snd hrecrant, Since tre no inference among tml Hoes we mm Mihou ls of general thats) sty prefers, fo = Now, nee, €5%0\ Uf gS? atch thar sreevant, the detion of 1 and the Sugai we've chosén impli, i relevant, But since'cvery se Cf leat tos) shen 0 and sy. gen and ty i peferied ts, by (ys, not abet response relate 0 Cf Hence,» ©" coadcting S{?™= 2. QED, REFERENCES ‘Aun, R, (1976: “Agreeing to Disagree,” The Amal of Statics, 4, 1236-1238, Basis, JS, ano J, Sonet (1987: “Equilibrium Selection In Signaling Games,” Econometrica, $5, ‘47-962 Basu, K. (1985p °S Princeton (mimeo). Busan, B. D. (1984): “Rationalizable Strategic Behavior,” Econometrica, 52, 1007-1028 Bone, K. G. (985): “Modeling Rational Payers,” London School of Eeonomics and University of Pennsyivania (mimeo). Danaus, E. van (1983): Refinements of the Nash Equilibrium Concept. Berlin: Springer-Verlag. Fupexaexo, D., D. KRErS, an D. Levine (1988). “On the Robustness of Equilibrium Refine: ‘ens," Journal of Economie Theory, 4, 354~380 Huttas J. (1986): “Sequential Equilibrium and Stable Sets of Beli,” mimeo, Stanford University Kacutan, S. 1940; “A Generalization of Brouwer's Fixed Point Theorem,” Duke Mathematical Journal, 8, 457-889. Kowsenc, E, ano JF. Mentens (1986: “On the Strategic Stbil tegic Irtionality in Extensive Games,” Insitute for Advanced Studies, of Equilibria” Econometrica, Koitnexo, E. AND P. J. Rewy (1991): “A Rationale for Consistent Assessments,” in preparation. Kars, D. (1987) "Out of Equilibrium Beliefs and Out of Equilrium Behavior” mimeo, Stanford ‘University Kars, D, axo G, Rasy (1987): “Structural Consistency, Consistency and Sequential Rationai ‘Econometrica, 55, 1381-1348, Kars, D, avo R. Witson (198): “Sequential Equilibria,” Econometrica, 50, 863-894 Kum, H. (1953 “Extensive Games and the Problem of Information,” in Conmbutions to the ‘Theory of Games, Vol. 2. Princeton: Princeton University Press, pp. 193-216,BACKWARD INDUCTION 649 Luce, R. D., ap H. Rata (1957): Games and Decisions. New York: John Wiley and Sons, “McLennan A. (1985) “ustiiable Beliefs in Sequential Equilibrium,” Econometrica, 53, 889-904, ‘Myerson, R. B. (1978): “Refinements ofthe Nash Equilibrium Concept,” Dnerational Journal of ‘Game Theory, 7, 73-80 [Nast J. F. (095i: “Non-Cooperative Games;" Annas of Mathematics, $4, 86-295, [Neunians, J. voN, aND O. MORGENSTERN (1953): Theory of Games and Economic Behavior, 3rd ition, Princeton: Princeton University Press Peance, D. G. (1984): “Rationalizable Strategic Behavior and the Problem of Perfection,” Econo ‘metric, $2, 1029-1050, ‘Rex, P. J. (1985): “Rationality, Common Knowledge and the Theory of Games,” mimeo, Depat: ment of Economics, Princeton Univers (987): “Explicable Equilibria,” mimeo, Department of Economics, The University of ‘Western Ontario Rosen, R. W. (1981): “Games of Perfect Information, Predatory Pricing and the Chain Store Parades,” Jounal of Econamie Theor, 25, 92-100. Ruanssten, A. (991): “Comments on the Interpretation of Game Theory,” Econometrica, $9, 909-924, Serres, R. (1965): “Spieltheoretische Behandlung eines Oligopolmodells mit Nachfragetragheit” “Zetschit fur die Gesamte Statevistenschaft, 12 301-326 and 687-889. (075): "Reexamination of the Perfectness Concept for Equilibrium Points in Extensive Games,” Iernational Journal of Game Theory, 4, 25-55. ‘Tax, T, ano S. WERLANG (1985). “On Aumann's Notion of Common Knowledge—An Alternative ‘Appfoach,” mimeo, University of Chicago Business School and Princeton Univers ‘Twoursox, F (1982) “Equivalence of Games in Extensive Form,” RM 758, The Rand Corporation, Zewnnxo, E. (1912): “Uber eine Anwendung der Mengenlehre au die Theorie des Schachspiels,” Proceedings of the Fifth Inemational Congress of Mathematicians, Vol. 2, ed. by E. W. Hobson and A. E. H, Love. Cambridge: Cambridge Univesity Press, 501-504,