MAT. UBRAR1E8 - DEWEY Digitized by the Internet Archive in 2011 with funding from Boston Library Consortium Member Libraries http://www.archive.org/details/necessarysuffici00smit2 HB31 .M415 no. working paper department of economics NECESSARY AND SUFFICIENT CONDITIONS FOR THE PERFECT FINITE HORIZON FOLK THEOREM LONES SMITH 94-17 (93-6 Rev.) May 1994 Massachusetts institute of technology 50 memorial drive Cambridge, mass. 02139 NECESSARY AND SUFFICIENT CONDITIONS FOR THE PERFECT FINITE HORIZON FOLK THEOREM LONES SMITH 94-17 (93-6 Rev.) May 1994 M.I.T. LIBRARIES JUL 1 2 1994 RECEIVED Forthcoming in Econometrica Necessary and Sufficient Conditions for the Perfect Finite Horizon Folk Theorem by Lones Smith* Department of Economics Massachusetts Institute of Technology Cambridge, May MA 02139 23, 1994 INTRODUCTION 1. For a given n-player normal form game G, where the objective function finitely-repeated let G(8,T) be the T-fold repeated game the average discounted is game, the perfect folk theorem is sum said to hold For this of stage payoffs. if the set of subgame perfect equilibrium (SPE) payoffs includes any feasible and strictly individually rational payoff vector of G for large enough T < oo and 8 < Benoit and Krishna (1985) [hereafter, BK] 1. produced a perfect folk theorem for G(6, T) that obtains when — namely, the full-dimensionality condition conditions for the infinite-horizon folk theorem and Maskin (1986), and of Fudenberg (ii) G satisfies the sufficient (?) each player has distinct Nash payoffs in G. Recently, Abreu, Dutta and Smith (1993) discovered an (essentially) — non-equivalent (NEU) utilities I Nash (i) NEU, and proved a revised general 'folk provide a simple, easily verifiable, and necessary and sufficient condition theorem that replaces for the finite-horizon folk distinct condition for — that neatly supplants full-dimensionality. Wen (1993) subsequently examined games possibly violating theorem'. Here iff The (ii): payoffs, only insists that players' behavior intuitive condition, recursively be iteratively leveraged near the end of the repeated game. The primary of contribution of this note 1 BK, and thus complete NEU is purely conceptual: finish the work only differed from full-dimensionality by a nongeneric class of games, so too, within Nash I wish to the perfect information folk theorem program. Indeed, just as the class of stage games with recursively distinct I I payoffs for all players later provide also recast a necessary and BK in simple new proof means in The Wen's Nash is But conceptual clarity is not of the necessity of tiered 'folk' BK without sufficient condition for the finite-horizon Wen's more encompassing framework. logic of the (1990/92) zero. is Nash payoffs, the measure without distinct NEU in Abreu 2 its Nash In so doing, et al. (1993), own reward, I as folk theorem. happen upon a and of what 'necessity' theorems more generally. folk theorem, as best exemplified in Krishna (1989) or Smith rather simple: Late in the repeated game, because payoffs, the behavior of all players have distinct any one of them can be leveraged by threatening to finish off with a fixed number (say S) of plays of that player's worst Nash payoff, rather than cycle through all his best Nash ishments work perfectly well. effect Away from the end of the game, the infinite-horizon punAnd because S is fixed independent of the horizon length, the payoffs. on the average payoff can be made arbitrarily small. This essentially Rather than formally define player example *It I am condition, I first motivate it their proof. with the following three- game G. BK conjecture under weak conditions that only one player need have distinct paper might in part be seen as proving a precise and rigorous statement of this conjecture. grateful to a referee for pressing me on this point. should be noted that payoffs. This 2 my is In this game, 2,2,3 2,2,2 2,2,2 2,1,-1 0,-1,-1 0,-1,-1 2,2,2 2,2,2 2,2,2 -1,0,-1 -1,-1,-1 -1,-1,-1 2,2,2 2,2,2 -1,-1,0 -1,0,-1 -1,-1,-1 -1,-1,-1 chooses rows (actions 1 [/, M,D), 2 chooses columns (actions chooses matrices (actions L,R). Player 3 strictly prefers is weakly dominated (pure or mixed) payoffs are minimax payoff the all convex combinations of for all three players. is Because players and 2 have unique Nash 1 payoffs of 2 and 3, so that his behavior only threaten to switch from (U, large enough, 3 is induced for players is (2, 1). So player G(R) (i.e. 2's when play (r, 1 and 2. It player 3 plays L). payoff of 2 in G(£, L). the game! That I £, G I is 1, which what player I NEU, satisfies And my new Namely, theorem'. behavior of The all if it is condition for 1 3 wish just prior to this and 2 do, a new game By choosing S' (and by implication S) large many periods, say 5", as In particular, players 2 0, which new differs I wish and 3 are willing (one-player) from G(r, R) unique optimal l's 1 game to near the end of sufficient for the is G has recursively distinct Nash payoffs. BK result: So long as such a recursive all players, then a perfect folk theorem necessary too for the general perfect finite-horizon any such chain of recursive reductions, I SPE payoffs is is also rather simple. For 'folk cannot leverage the either point-valued, or his behavior SPE it is any given horizon length T, not. Evidently, if the behavior of payoffs are multivalued for large enough cannot be leveraged, then he entertains a unique equilibrium But observe that no one player can simultaneously minimax the other two, a on in my remarks on the necessity of NEU. later S players, then a 'folk theorem' does not obtain. set of if choosing from his unique Nash payoff differs does, yielding a a player can be leveraged as above, then his T, while need a folk theorem now follows by the same proof as before. intuition behind the necessity a player's I have now also leveraged the behavior of player satisfies By I have now leveraged the behavior of player 2 near procedure eventually leverages the behavior of obtains. and thus the standard L) for the 5-period phase. summarize the above procedure by saying that game a G Furthermore, leveraged near the end of the game: has the unique optimal payoff of It (2, 2, 3). has the unique Nash equilibrium payoff vector Nash payoff from G(R) irrespective of 1. and G satisfies NEU, , R) (resp. r) G does not satisfy the BK condition can induce player 2 to play £ m, or r for as I for player (2, 2, 2) R for as many periods, say 5', as plays R irrespective of what players 1 just prior to this penultimate 'Nash' phase. If payoffs, is L) to (M, £, the end of the game! Iterate this process. enough, D easy to see that the only Nash willing to play Nash phase. When player 3 G(R) to R, while action and 3 r), theorem applies to G. Next, player 3 enjoys the distinct (extremal) infinite-horizon folk Nash it is m, 3 Nonetheless, a folk theorem does obtain! For (ii). in Thus for player 1 (resp. player 2). L £, fact of some importance payoff for all T. Clearly, in the latter case, G instance, in 1 game has T> 5 and T> all satisfying T, 19, respectively. one player has distinct Nash payoffs, then the if two pure and generically a third mixed Nash equilibrium. Thus, at least games generic payoffs for payoff set for G speak to the genericity of the distinct Nash payoff requirement theorem obtains. Indeed, folk SPE only receive distinct special payoffs in when the SPE with no discounting, player 3 has a multivalued while players 2 and The cannot possibly hope for anything more. For I my condition, Nash players enjoy distinct all my Section 2 focuses on the stage game, and describes payoffs. iterative procedure. and necessity sufficiency (the folk theorem) in section 3.1, for I dwell on (in all respects) in section 3.2. THE STAGE GAME 2. 2.1 Basic Definitions Let G= (Ai, = i 7r,-; i's finite set of actions, set 7r(a) = (tti(o), Simply write . . , . G . . n) be a finite normal form n-player game, where Ai , . A= and " x t =1J4,-. Let player i's utility function n n (a)). Let Mi be player 7r,(^) for The game 1, utilities genstern utility functions are equivalent, (1993), let {lb in the to i. C Denote the nj(-). X, b = 1, . . . = = > 0, for all i} F* ^ 0. Note that under Wen Kj(a) and NEU, w = max a < j for all i and j, 7r,(-) is 1 = = denote by X{i) ttj(-) 7r,(aj, € for all j all a_j) of player 1(i). l To it i is co{7r(//) : /j, reduces to the standard € I. (/ii, l j€I . //„) € M. Following I Wen such that any two players The effective minimax payoff the best payoff that anyone in all 4 players in I(i)- e M}, and set. To call F* : sidestep trivialities, let minimax strategy payoff = {w G F for player z, set. it is true. Indeed, w' is certainly a minimax profile weakly exceeds any minimax payoff for all j £ Z{i), let ) Analogous to the fact that the minimax = minmax7Tj(a) > maxmin Jrj(a) = max a . not a positive affine > maximin games, Tr^u;') . M,-. as such, but see that iri(w nj(a'j,a.-j) for j = )- : M = x"=1 with {1,2, ... ,n}. minimax payoff over F= fi , players with equivalent utilities 1{i). (strictly) individually rational (1993) does not describe € set, ir, player no two players' von Neumann Mor- the coarsest partition of Define i. if the feasible and (strictly) rational payoff feasible some player i.e. the greatest i.e. for all Wi and F* the 7r,-(-) (NEU) set of players as utilities; min a maXj 6 i(,) max„ Normalize n^w') for B} be Normalize payoffs so that X(i) can guarantee himself, 4 , same X& have equivalent level 7r,(w') mixed strategy Vs expected payoff under the mixed strategy has non-equivalent transformation of i's be is A — M, and j€/ a j€/ minmax7r,(a,-,a_,) ) \ a -j a i J [ . in constant sum 2.2 Recursively Distinct Nash Payoffs Given a subset of players 5 J' = a j, let = {ji, j%, (ah ,ah ,..., a jm ) G(aj>) be the induced (n . . , C X and their jm } Mh x Mh x € — m)-player game • • x • (possibly mixed) actions M jm X\J' for players = Mj>, (1) obtained from G when the actions of players J' are fixed to aj> G as an increasing sequence of h > Define a Nash decomposition of {0 = y{ e Jg -i) of 1, nonempty subsets namely of players from I, so that for g 1 . . h, actions , . = JoC JiC J2 C---C JhQX}, G(e^_,) and y{fjg _ ejg _ x ) x of , fjg _ G y G{fJg _ Mj _ g a pair of Nash payoff vectors different exacta/ for players in ) 1 exist with Y (2) Jg \ Jg _ u Viej^i * V(fs^)i i.e. (3) E Jg \ Jg -\- For instance, players in J\ have distinct Nash payoffs in G. The game G has recursively distinct Nash payoffs if there is a Nash decomposition with for all = Jh i X. So holds. if, BK, as in The converse is all players have distinct not true, as illustrated not have recursively distinct Nash payoffs, then arise, since union of distinct There Nash t It is An G that do ambiguity may not inconceivable that the player. In fact, this cannot occur. J* C X who have 6 recursively FINITELY-REPEATED GAMES Nash games with (a,i , . . . , oj.t) all players. be a behavior strategy for player with the strategy profile a. Player pected discounted Payoffs perfect monitoring, allowing each player to condition his current actions on the past actions of in period impossible. games payoffs. shall analyze finitely-repeated = is a well-defined maximal set of players is 3.1 Sufficiency of Recursively Distinct Let a, later consider I =X Jh payoffs in G, then this condition Nash decompositions need not be unique. 3. I earlier. Nash decompositions could include every all Lemma i.e. Nash sum of his payoffs: -—^ Y.J 5 Throughout, ACB 6 The proof of this result is omitted; see means that A is a strict my i's and nu (a) his expected payoff objective function in G(S, T) (-1 <5 i, 7r, ( (a). subset of B, i.e. The set of SPE payoffs AC B and A / B. (1994) working paper. is is the ex- V(6, T). Beyond replacing the Theorem Nash payoff requirement, Theorem distinct 3.7 in three ways. First, it admits payoff discounting, where 5 and Second, independently over the relevant range. Wen pursued by BK's (1993), which will imply period, players can condition on the random variable. 7 BK, Just as in with NEU. T can vary theorem', as 'folk Third, unlike the (more public randomization is used: In every outcome of a publicly observed exogenous continuous assume that players can observe deviations within shall I result BK, intuitive) use of long deterministic cycles in a more general it is from BK's 1 differs the support of strictly mixed strategies. Alternatively, just assume that the folk theorem pure strategy minimax payoff refers to the level. 8 G has recursively distinct Nash payoffs. Then for the finitely-repeated game G(5,T), V« € F* and Ve > 0, 3v G V(S, T) with \\v - u\\ < e. 3 T < oo and S < 1 so that T>T and S G [5 1] Theorem (The 'Folk Theorem') 1 Suppose that the stage game = , Fix a Nash decomposition Proof: c9 g for = 1, 2, . . . , = .*W£ *€ Jg\ J 9-1 Let the p h. (2) for > \\y( which inequality e J.-i)i ~ y(fjg -i)i\\ > be the largest payoff range best and worst payoffs) for any player in G, and 2kp/cg ones. Saving on notation, . let y 9 Further, let z 9A be the less y(fjg -i) for player i € Jg \ Jg -\. ip g and define ° (i.e. the difference between the least even (k) number above and y{fjg _, preferred Nash payoff vector amongst y(ej denote y{ejg _^ Since for -kp + it (3) obtains, all i ) in even periods ?>i>g (k)z?> follows that without payoff discounting, any player m > 0, For any for 7 g = h — One may l,h — 2, . . . , 1. = ijj g (m Define = ip h and i (4) , has a strict incentive to conform , if deviations + s g (m) for (m) and + s g +i(m) £o( ) . recursively define Sh(m) s g (m) _, odd i to k consecutive periods of any action profile followed by ipg{k) periods of y9 are punished by switching each y9 to z 9,t in Jg \ Jg -u e g (k) y ii> ) m = ) 0> H and also take advantage of the correlating device to 1- t g s h (m)) (m) = si(ra) + ••• produce a nearly exact, rather than an one can use the correlation device to do away with the ^approximation for all payoff vectors not on the boundary of F. I omit such a laborious exercise. This assumption is not necessary to detect deviations from the minimax phase (step 3, below), as noted in my (1994) working paper. But for the two recursively Nash phases, no such workaround exists. It would thus be interesting to know whether Theorem 2 goes through when J* is defined given pure actions (1), in which case this would not be a concern. approximate, folk theorem. That is, = g 1, . . , . The T-period equilibrium outcome sequence h. is h h a,...,a;y ,...,y ;...;y\...,y where a If v is I 9 and y9 ^ j, Ui (target payoff now x x 3> (strict IR), x\ < \\v — < u\\ e for big enough 5 £ x? for all j periods of the repeated game; some player 2. Jg < . Reward Phase. Play xJ for r periods. Then Bad Recursive Nash Phase. deviates, where g" < g' set g' «— , straightforward (or see SPE for a game rather stark. 9 10 [If some my q, r G [If £ any i + r + th(q + r) i deviates early, start 3; if i -f- j g (q and + 5.] i. some or 2. 1 Play z 9 '^ until period g" and t T - V-i(? + r restart 5.] Then go )- pfj e to step <?g" 2. (1994) working paper) to verify that these strategies and big enough 5q Nash and To. 11 Payoffs does not have recursively distinct Nash payoffs, the consequences are The following result has a flavour of the "zero-one" laws of probability Note that q and r are implicitly defined in steps 3 and 4 below. Below, i and j denote arbitrary players, and g, g' and g" arbitrary shall use the simple notation j <— i to mean "assign j the value i." indices in {1, 2, . . . , h). For clarity, Also, steps always follow sequentially, unless otherwise indicated. Bracketed remarks refer to off-path play, 11 any deviates early, restart 3; if , I n For X(i) deviates, start 4.] Set j 4- return to step 3.2 Necessity of Recursively Distinct If ,x 10 this equilibrium. > w' for q periods. [If j Jg> deviates late, start 5.] constitute an . 5.] i 4. is . and which support T — th(q + r). . Minimax Phase. Play G . others are called early deviations. deviates late, start 3. i It , Good Recursive Nash Phase. For g = h,...,l\ Play y 9 in periods T r) + 1, T — t g -i(q + r). [If some G Jg deviates late, where g' < g, start . 5. all Play a until period in and T. 1 domination). explicitly describe the players' strategies Main Path. 0. Z(i) (payoff asymmetry), ease of exposition, late deviations are those occurring during the final q 1. > played for s g (q+r) periods. Fix £ is have established the existence of feasible payoff vectors x et al. (1993) such that for alii < periods, the (5-discounted average payoff vector, then is Abreu x\ T—th(q+r) played for 1 i.e. following deviations. Absent public randomization, the simple direct proof presented above would require two modifications: First, the target outcome a would have to be replaced by an approximating finite outcome profile. Second, ditto for each x' vector. Finally, to my knowledge, public randomization is essential if one wishes to work with the mixed minimax strategy, which I noted earlier was possible. See for instance, Abreu et al. (1993). T — Namely, as theory. payoff set F*, or some oo, either > V(S,T) tends to the players receive a unique SPE payoff. strictly individually rational There is no middle ground. Theorem 2 For any T < oo and any S £ (0, 1], players inX\J* receive SPE of G(S,T), equal to their unique Nash equilibrium payoff of G. 12 X \ J* has a in I \ J* has Proof: Every player in unique Nash and hence Assume that everyone a unique payoff in G, for T= 1, . . . ,T Then . SPE SPE a payoff in any payoff in G(S, = 1) Nash payoff in G(8, T) equal to his in the first period of + G(S,T 1), players in G. X \ J* must play a Nash equilibrium of G(ej.), for any ej- € Mj., because they have a unique SPE By induction, the result obtains for continuation payoff by assumption. all T. Remarks: 1. This stark necessity result contrasts markedly with the partial failure of NEU, where we need only Let F* 2. = {Xf,}. profile separately holding , 13 all with the effective minimax payoff If (+) 7t i( aji a -j) V«}, where the partition for no such partition {lb} does there exist any strategy SPE > mina max € f,^ max^ 7r,-(oy,o_j-), for By corollary, the necessity result of (a particularization of (*)), X(i) payoffs Pareto dominate the effective This yields 'necessity' in another respect: The folk theorem F*. level. two or more players in any Xb to their worst feasible payoff in necessarily 7r,(w') (1993) proves that level > mina maxj€ x(t) maxaj {v € F*\v{ {lb} strictly refines F* then minimax payoff replace the if NEU fails, Abreu et al. fails X(i). But Wen minimax outcome. 14 when F* (1993) follows! then {It} admits a C is replaced by For under strict refinement, NSM and the (standard) finite- or infinite-horizon folk theorem cannot obtain! 3. Recursively distinct Nash payoffs the finite horizon 12 13 14 Note that That is, this Nash folk theorem, is also the necessary for all i, and sufficient condition for due to Benoit and Krishna (1987). 15 theorem says nothing about the payoff £ l(i) C l(i) and X(i) C I(») for set of players in some J*. i. Unlike his folk theorem, this result of his actually doesn't rely on observable mixed strategies. And it holds for finitely-repeated games too. 15 I am grateful to a refereefor suggesting this. For a proof of sufficiency, consider the following strategies: Let the equilibrium path follow steps 1 and 2. Off-path, retain the minimax phase, step 3, but allow it to last until the end of the game, and omit any reference to play after deviations, including step 4 (as it is not needed in a Nash equilibrium); finally, retain step 5, but omit the references to off-path play. To see necessity, just observe that Theorem 2 still obtains if SPE is replaced by Nash equilibrium. References Abreu, Dilip, and Prajit Dutta (1991) 'Folk Theorems for Discounted Repeated Games: a New Condition.' University of Rochester mimeo Abreu, Dilip, Prajit Dutta, and Lones Smith (1993) 'The Folk Theorem for Repeated Games: a NEU Condition.' Forthcoming in Econometrica Benoit, Jean-Pierre, and Vijay Krishna (1985) 'Finitely Repeated Games.' Econometrica 53, — 905-922 (1987) 'Nash Equilibria of Finitely Repeated Games.' International Journal of Theory 16, Game 197-204 Fudenberg, Drew, and Eric Maskin (1986) 'The Folk Theorem in Repeated Games with Discounting or Incomplete Information.' Econometrica 54, 533-554 Krishna, Vijay (1989) 'The Folk Theorems for Repeated Games.' mimeo 89-003, Harvard Business School Smith, Lones (1990/92) 'Folk Theorems: Two-Dimensionality mimeo Wen, Quan (1993) 'The 'Folk Theorem' for Repeated Information.' Forthcoming in Econometrica is (Almost) Enough.' Games with Complete MIT APPENDIX: OMITTED MATERIAL 4. Proof of 4.1 If for Lemma some Nash decomposition I = X, have Jh then unambiguously J* = 2, and I am done. Otherwise, let £,(G) be the set of Nash all the alternative unique nested sequence of {0 =J C h! C Ji payoffs for player nonempty J2 C • • • X\J for whom {£i(G(aj aj players Jg+ \ J exist, set h! = g. € G )), g \ Note that Mj {Jg } in the game G. Consider satisfying C Jh C J}, > given by the following (well-defined) recursion: Given i sets i Jg , the set Jg +i \ Jg is all does not consist of a singleton. } If players no such g it may not be possible to find only two Nash equilibria providing distinct pay- Jg +\\Jg Still, can show that Jh = Jh for all Nash Moreover, for any such nested sequence decompositions (1). For is clear that Jh C Jh For {£i(G(a,j,)),aj, € (1), all players in any Jg are eventually included in some Jg simultaneously for offs players all i G I - it <. > Mj,} does not if . consist of a singleton, then neither does {£i(G(aj„)),aj„ G Mj,,} if J'CJ". Omitted Parts of Proof of Theorem 4.2 Step Observable Mixed Strategies 1: proceed recursively as I I ensure subgame perfection. some that all deterrents are strict by continuity of discounted S G 1 [5o,l], for some 5q sums < 1. positive margin, say in 5, they will thus I remain shall check that after First note that J only check 1, without discounting. Given strict for any level of discounting no history consisting of isolated deviations does a one-shot deviation yield a higher payoff than the strategy profile; by the "unimprovability" criterion, that profile will be a I now select q and r. In light of x\ w\ for all x\ < 16 i, where U{ too, (4) The /?' + > 0, let qx) > subgame q satisfy PI + perfect equilibrium. 16 1 (4) (resp. w') is the best (resp. worst) payoff vector for player i in G. Since simultaneously renders steps 3 and 4 a strict deterrent to early deviations inequality (4) — in particular, why the left-hand side is not (q + l)x\ obeying the random correlating device generating x' might sometimes require outcome. i — reflects the fact that to play his worst possible from step for any 1, Next, step 4 deters deviations from step 3 r. + rxj > Pj+ rx3] + (q- qw'j for all j 7^ Next, This i. possible because Xj is [q i lg G i>; (s 9 (Q > s g (q 1 = . . + r) • + • s h {q T > T T < oo for — \\u Step v\\ < , . , 1 + 1 h. each deterrent remains < and 5 j i for whom an adaptation of Abreu is k, F , or x\ c !J to = c\ xJ or , 3 l cj \ (b) > onto the i-j j arj: is 6 F* such c'-' € F* — to which the reader coordinate plane has l. satisfies is referred After step that: (a) payoff x' in step 5 4, minimaxing players j ^ full dimension; for 17 not indifferent between c,J and x (via the public randomization device) incentives of et al. (1993) play proceeds just as in step There must exist payoff vectors c f° r [<$o,l], To need only (below) consider deviations from m' by those players I the projection of any other player IJ > q + r + th{q + r). some 8 < 1. Finally, if r are fixed. a simple argument that 7^ S € feasible Unobservable Mixed Strategies 2: The argument for strict for all is large enough, the resulting average payoff vector v and since q e, 1 an immediate is + r)](-p) + s g (q + r)yf Given q and r as defined above, the above program For Indeed, this + r))(-p) + s g (q + r)y? + r)zf' 1, 2, 1 which upon substitution yields + r + s g+l (q + = and g , (3), enough that x^. verify that step 5 deters all late deviations. I consequence of inequality for all > + 1)uj r is large if i is 1 , or indifferent c- ^ between xj\ and l (c) x' j prefers probabilistically replaced is and l by c i where the probabilities are chosen to maintain the , i, and player i is unaffected by the transitions given (a). Let Pt'(aj) satisfy for all aj,a'j in the support of m}, and Kj {a, raLj ) , *-h-i That such real numbers (1 ttj (a'j , < q, wLj = ) +S + -..+ S^Mia'j) -p^ajW/ - x)). 3 p\ - t {-) exist is a consequence of (b). That they can be chosen be (non-unique) probabilities follows from this construction, for instance, when cj 3 (6): 17 Let p\ {aj) Below, if = such k 0, where action a,- exist, the expression yields the highest u j ^ *" > to x^ in myopic payoff to player j (when the actually denotes j 10 (5) ^ i, k. other players play m* 3 substituting p\ {aj) 8 < 1 and = in (5), r. m 1 a'j will follow that pJ (Oj) < the realized sequence of action — 1) for each j ^ l/(n Also observe cj -. p i, x i{a q ) myopic payoff to player yields the lowest J it Furthermore, p t{a") step 4 with probability l/(n action l a" in the support of If If •). — > < l/(n \)q for any larger 8 x^ implies that profiles is a q = £? p^ (ajt) = (a,i,a,2, for all j ^ i. p . tj . By and (a") , — then after l)q for large enough r, > j, or any other action for all a g ), then 7 c'- a'-. replaces construction, < yielding a well-defined probabilistic transition function. independent of actions by players j ^ i and of j 's own ^t, actions in periods s (5) a;* in ij p (-) < As p? is ensures that players are indifferent across the support of their minimaxing strategies. Finally, in light of (c), the punishments that previously supported step 5 11 still obtain. uou MIT LIBRARIES 3 TOflO D0flSb7MD 3