Digitized by the Internet Archive in 2011 with funding from Boston Library Consortium Member Libraries http://www.archive.org/details/communicationequOOfarr J^-. ?•''< worKirrg p f' department of economics ^^1 '*^4r \< ^^^g^mMg ^B 1 Coiranunication Eq uilibrium in Games ! i Joseph Farrell Number 38.2 -%v June 1985 massachusetts institute of technology 50 memorial drive Cambridge, mass. 02139 Communication Equilibrium in Games Joseph Farrell Number 382 June 1985 Communication Equilibrium in Games Joseph Farrell GTE Laboratories and MIT; June ] 985 1. INTRODUCTION simultaneous-move game of incomplete information, the standard solu- In a This is the natural tion concept is Bayesian equil ibrium. generalization (to incomplete-information games) of the notion of Nash equilibrium: the modification instead that, is choosing of each player chooses such a probability distribution a over actions, probability distribution for each of his possible The different players may be thought of as making their choices in "types." separate rooms: there is no communication among them. Myerson (1983) has argued for the possibility of a different solution concept, allowing for centralized communication. In correlated equil ibrium, a each player secretly reveals his private information to worthy mediator who committed in advance to is a disinterested trust- known behavioral a mediator collates this information, and sends each player secretly dation for his requirement for equilibrium is The move. that rule. a no The recommen- player be tempted to lie and/or to cheat, given that he believes that all others report honestly and act obediently. Correlated equilibrium allows for communication among players, about Accordingly, to it make is a Often there is no disinterested mediator available. somewhat artificial way. Agents but in strategic choices often talk directly to one another. important to ask whether communication without a mediator is the same as communication with one. In an this paper, outcome that specified way. we introduce the concept of a communication equi ibrium: equilibrium when players can talk to one another is an We show that every Bayesian equilibrium is a in communication equilibrium, but not vice versa, and that every communication equilibrium is correlated equilibrium, but not direct communication can count; tive value. vice on versa. Thus on the other hand, a the one a hand a costless mediator can have posi- We then ask what kinds of information can be revealed by unmediated com- munication. the standard case of In independently distributed types, we show that there are invariants that are always preserved under unmediated communi- cation, sents but need not be preserved when the fact that, without a a mediator mediator, it is available. This repre- impossible for is each of two players to make his revelation contingent on what the other says. Possibly because of game theory's beginnings case (von Neumann the Prisoners' and Morgenstern, 1944) in as strategic relationships were Many theorists informally essentially therefore (absent differential costs of sending messages, etc.) communication can be realize how to Dilemma since the subject turned to the more general case, com- all if two-person zero-sum and the excessive emphasis given munication has been badly neglected in game theory. think the important ignored. It is also payoff-irrelevant moves adversarial, as in something of may be for Spence, 1970, surprise a the and outcome to of a game. Roughly, payoff-irrelevant opportunities for coordination in two-person zero-sum games, in but effect to delegate some control "his opponent" ing even is misleading). moves matter the game. in over general his to the extent that there are There are no such opportunities each player may be prepared move to the other player (the in term This can be accomplished through message-send- if messages do not directly affect payoffs. DEFINITIONS 2. consider We indexed by which move TT.; X. E •) (*, can .-1 We write X.. is given by = t (x u t ) for all t a Players random variable ) t T.. e -1 and = x n Player t. (x,, ..., also chooses i x n ). a Player i's -^ The priors tt.(*) and the payoff . are functions u are common knowledge. We write T T,x ... = x and X T in = XiX n i probability distributions on any set w follows: as according to the probability distri- T. .... (t,, described be observes the value of i we assume that n.(t) > 1 1 payoff Player n. ..., 2, which G distributed on the finite set is bution 1, game a A. ...xX .We write A(A) for the set of An outcome of the game is a function T ^ A(X). : An outcome w: T A(X) -* stratagies, functions s-: 1. The strategies s. 2. The strategies s. T. is a Bayesian equil ibrium outcome if there exist ^ A(X.), constitute a such that Bayesian equilibrium: lead to the outcome w: weight assigned by w(t) to the point x e that is, X is equal for each t e the T, to n S.(t.)(x.) 7T It is a standard result that a Bayesian equilibrium will exist in a finite game such as ours. An outcome w: 1983) if, T -> A(X) is a correlated equil ibrium outcome (see e.g., Myerson, when a trusted mediator asks for reports of t, performs any neces- sary randomization, and then instructs players secretly as to the value of they should choose, there are no incentives for vate information or to disobey his x. a player to misreport his pri- instructions, assuming that other players are reporting truthfully and acting obediently. We say that a game G' communication version of G if is a T. (just as in G) G' consists of three phases: 1. Player 2. There learns i t opportunities are speak. t- assume (We unrestrictive if functions, but n that = may 2.) announce for players to all players hear messages, i.e., to this is announcements; all These messages do not directly enter payoff convey information and influence others' final choices. The players simultaneously choose x. 3. We emphasize that payoffs are given by there is payoffs. no need to w' (x, u X. t) (just as in G). just as in know what messages were announced However, as we will Every outcome z of G' forget the information see below, G in a is a communication equil ibrium outcome of For instance, if we make We will natural in G', G is x. way (simply If w' is the then we say G. the communication opportunities vacuous, that any Bayesian equilibrium outcome w of rium outcome. and x, order to calculate concerning the messages announced). outcome corresponding to some perfect Bayesian equilibrium that w t sometimes the messages will affect determines an outcome w of in w' in given G: also a we see communication equilib- show below that there can be others. RELATIONSHIPS OF THE THRtE EQUILIBRIUM CONCEPTS. 3. Theorem 1: Every Bayesian equilibrium outcome is 1. communication equilibrium out- a come, but not vice versa. equilibrium communication Every 2. outcome is correlated a equilibrium outcome, but not vice versa. Proof: First, we can always set 1. come is 1: < a < dent. 1 This < b. Hence every Bayesian equilibrium out- G. communication equilibrium outcome. a fails, consider Example Example = G' is a 1: symmetric = {y, z}. t. each player, For game. The probability that For each player, X To see that the converse = a is and p, Player indepen- z t (-tp -t^) -t2) lj_ If p < a, 0) (0, 1 is for (z, z) better move if the unique Bayesian to be chosen for all Let q denote the probability that player player I's choice. 0) then the unique rational izable (hence, equilibrium) outcome of Example Proof: and t^ are t, where 1 (0, Lemma tp b}, 2 y - {a, Payoffs are given, by: Player (1 T = If t-, = b, 2 then z dominates y. will If play y, t, = a, t^ and consider then y is the q(l + (1 a) if q > a. i.e., know - < q - q)(-a) > However, since z dominates y for player Hence if p < p. wise, nor can player 2. then player described b, we Like- 1. more striking example with infinite T, a Farrell, 1983.) in The outcome which assigns probability 2j_ = cannot rationally play y. 1 This proves Lemma This is an adaptation of (Note: Lemma a, whenever t^ 2 and assigns probability to 1 y) (y, when t = (a, a), for other values of t^ is a communication z) (z, to 1 equilibrium outcome. Proof: Define as follows. G' Stage "A" or "B". If t = b, announce announces. If t = a, they do so, If Lemma is not Theorem a 1 Next, 2 outcome w, complete. in 6' "B". for both players to use this strat- This proves Lemma 2. a communication equilibrium outcome that (in G); thus the proof of part (1) of the positive statement is Costless communication can make we prove part a whatever the other player the outcome is as described above. displays an example of easy to prove: z if he announced z Bayesian equilibrium outcome is play strategy: announce "A"; then play y if the other player perfect Bayesian equilibrium is a egy. and then "B", announced "A", and play It each player simultaneously announces I, Consider the following is G. II Stage In (2) of Theorem 1. whatever Again, the a process difference. in G' that yields an mediator can commit himself to imitating that process internally. The fact that w was supported as an equilibrium in G' implies that it will be an equilibrium for players to report truthfully and to obey instructions. To see that not every correlated equilibrium outcome is equilibrium outcome, consider Example Example 2^ {z, sible t E T = {c, d) X {c, we use four bi-matrices, d). Player 2 t = (c,c): z Player t = w 2 (1, 1) (0, 0) W (0, 0) (0, 0) 1 (c,d): w z communication 2. This too is a symmetric two-player game, with T. To represent the payoffs, w). a (0, 0) (-2, 0) (1, -2) (-2, -3) = {c, d) and X. = one for each pos- t = (d, c) w z t = (d, (0, 0) (-2, 1) (0, -2) (-3, -2) d) w z z (0, 0) (0, 0) w (0, 0) (1, 1) The types are independently distributed, with Lemma 3: In if t = Example 2, the following > n.(c) 1 outcome is > 1/2. correlated equilibrium out- a come: Otherwise, Proof ior, (z, z, x = (w, w); z). Suppose 3: player to play he will = x Lemma of then (d,d), a mediator announces unless both players report instruct both to play w. If this that their types as d, a asked to play play Thus, z if c. If player 1 = d, 1 reports truthfully when t, = 2 honest and obe- is reports truthfully, he will be and he knows that player whatever player 2's report, Since np (c) > 1/2, c, he then will he 2 will would not. obey the mediator's His expected payoff is TTp(c). Would player ti = Would he then prefer to play w? z. instruction. that t, which case correlated equilibrium. Consider player I's incentives, assuming that player Suppose first that in each induces honest and obedient behav- then we have shown that the outcome given is dient. instruct will he 1 benefit from lying when then what happens depends on t^: t, = c? If he falsely reports probability With player (c), -np mediator then asks each player to play that Xp = truthfully 2 Player z. sees that he should play z, and gets payoff now knows that tp The = and c probability 1 - player (c), iTp 1 1. truthfully reports 2 mediator then asks each player to play w. Xp = w. t^ = c. Accordingly, consulting the first of our payoff tables, player 2. With 1 reports Player 1 tp = d. now knows that tp Consulting the second payoff table, second column, = The d and sees that what- 1 ever he does he gets payoff (-2). Hence Tip (c) + (1 Next, expected the - payoff iT2(c))(-2) < consider a to player of type d. 1 c will who c lies is not choose to lie. tells the truth, he If type of A player of type (c). ITp player a one of two things may happen. With probability player (c), iTp mediator asks both players to play of the third payoff table, 2 truthfully reports Player z. = Then c. the thus consults the first column 1 that he might as well and sees tp obey. This gives him payoff 0. With probability players to play [1 - •np(c)], Player w. reports tp 2 now knows tp 1 = Thus if t, = d, player he tells the truth he will If t, = d and player 1 1 Whatever tp may be, it gives 1 t, = d. payoff of 0. 1 - d. = w. The mediator asks both Consulting the fourth 1. 'np(c) by telling the truth; and if obey instructions. Xp = z. a can get Xp d, table, he sees he should obey, and gets payoff = he will lies, x, Since = z < not learn tp, and will is at least as good a move as x, 1 - TTp(c), This concludes the proof of Lemma 3. know that = and w; lying is unprofitable for 1 if 10 We complete our proof that a a correlated equilibrium mechanism need not be communication equilibrium mechanism with Lemma Lemma The outcome described in Lemma 3 4j_ is 4: not communication equilibrium a outcome. Proof of Lemma Lemma corollary of Theorem below. However, 4 is a apparatus is required only in order to allow for mixed Bayesian strategies in G'. It 4j_ 5 very straightforward to show that neither player is that Example in 2 is willing to reveal his type before knowing the other's. notice that, To see this, he will set Xp sirable otherwise. = n2(c)(-2) + (1 (since of choose Xp = Lemmas If = d, •iT2(c))(l) = 1-3 also 1-3 Since z). and 4 not TTp(c), reveal tiplc) if 1 = t the fact would revealing t, will 2 3 This is desirable for player w. - knows or believes that if player 2 while concealing and tr,, < 0, player then 1 each will essarily be properly modeled by assuming mediator. a mark case in which communication cannot matter: if of payoff a 1 player will want to conceal. together complete the proof of Theorem that decentralized costless communication can matter, gives it payoff a 1 then d, but unde- d), (d, give = t, Theorem 1. 1 shows and that it cannot necWe next examine there so is a bench- much conflict that players can never trust one another. Theorem 2j_ In two-person zero-sum game, every correlated equilibrium outcome a gives the same payoffs in each state affect payoffs not tively, is the reason is clear. affect the action rule bility in with report a as Bayesian equilibrium. Another way to say this is that only "inessential" communication that Remark: does t x,,: possible that means that, when if 2 T^ ^ A(Xp) 1 such in means that t, . he But a game. each player (say, used by the other. his true type is is equilibrium By choosing a message, reporting for player t, in prefers since Intui1) may Incentive-compatithe rule the game is associated zero-sum, asked to follow one of the rules used in equilibrium, 11 he would do better to follow any other. This proves Theorem 2 for the case where the rule can be inferred from the mediator's message. order to allow In for the more general case, we give the following proof. Theorem of Proof Although 2j_ general in a player in zero-sum game whose a opponent pays attention to his claims can profit by lying, write down instead a a lying strategy for the general successful deviation from a it difficult to is so we case, consider proposed correlated-equil ibrium mechanism to ran- domized uninformative messages. Consider a correlated-equil ibrium, in which each player sends messages to the mediator revealing his private information, Suppose that player appropriate moves. 1 and then the mediator suggests begins randomizing his messages so and playing that they convey no information about his type, Bayesian equi- a librium strategy independent of what the mediator suggests for him. 2 knew this was strategy, happening, his and both would get Bayesian response, best best response would be to play his his not know about I's deviation, and equilibrium payoffs. on would If player equilibrium payoffs. the other hand, therefore do Hence player 1 no better than player If then player 2 Bayesian 2 did would not use Bayesian his get would do at least as well as his Bayesian payoffs. Thus, least can his do no in any correlated equilibrium, each type Bayesian equilibrium expected payoff. better than his. The same of This argument, player implies reversing 1 must get at that player players 1 and 2 2, gives the bounds the other way, thus proving Theorem 2. Remark: We cannot quite say that communication will not occur. For example, consider the game which gives payoffs of zero to both players whatever happens. Theorem Many correlated equilibria will exist in which the players communicate. 2 expresses a sense in which such uninteresting communication only kind that can occur in equilibrium in a two-person zero-sum game. is the 12 Corollary to Every communication equilibrium outcome gives the same Theoreir. 2j_ payoffs (to each type) as Bayesian equilibrium. Discussion: The proof of Theorem when communication will in be 2 suggest the following heuristic analysis of important. When there is essential communication equilibrium, each player is allowing the other to choose his move to some since extent, rule, it is common knowledge that one message will another another. tions are sufficiently (u 12 u A general = mixed; ) but is is wise precisely when payoff func- "positively correlated." A has some aspects of each type. some games they are separable. Often they An example group of risk-lovers wishing to meet to play poker. 12 game of pure coordination the polar case, opposite to that of the zero-sum game (u game in Such delegation induce one action If is are = -u ). inextricably the problem of a they coordinate suc- cessfully on the meeting-place, they will play an enjoyable game of pure conflict; otherwise they will drink estly concerning where to meet, knowledge of odds, etc. alone. We expect them to communicate hon- but not concerning their playing strategies, 13 4. AIMNOUIMCEABLE INFORMATION We now turn to a limitation of decentralized communication that has nothing to do with incentives to tell possible both for I's revelation of 2's revelation of tp to depend on is possible with the value of We show that a mediator can absence of a t, the value of tp, Such . dependence mutual this section, we define the In structure having the right 5) a and for (if information is verifiable ex-post) gen- mediator various apply this result (Theorem it is not structure and announceable information struc - information erate any public information the to depend on t, mediator, as in Example 2. a concepts of publ ic ture. As in Example 2 above, the truth. constraints other "mean"; must satisfied. be the proof of Theorem to complete that but in We and to dis- 1 cuss bargaining with two-sided uncertainty. In this we are concerned with constraints on what section, ally possible to reveal. tives, punishment is controlled (e.g., if quate it may still available there is a information is the subject of this for liars). lot of noise is in actions each (and ade- cannot be the observation of actions), revelation of Just what revelations can be achieved without a mediator section. publicly heard by all. equilibrium, private If We assume (with some loss of generality if ken verifiable ex-post is be desirable to have some, but less than complete, private information. logic- Accordingly, we suppress for now the role of incen- that private by assuming is it player's Therefore, information together with the information inherent of course, reduplicate some or all at n each consists in > 2) stage of his of a communication private information what has been announced (which may, of the player's private information). latter information we call the publ ic cation happens, information changes (in the public that every message spo- information at that stage. a This As communi- way that depends on the true state t) until, just before moves in G are actually selected, each player sees the final publ ic information together with his own private information. 14 We can represent the final public information as a posterior probability distribution p on T. Each player i's information is then the conditional dis- tribution generated by restricting p to T,x ...xT. ,x {t.} A information publ ic structure is function a a: x T.,, 1 -* x...x T A(a(T)), . where o(t)(p) describes how likely the posterior p on T is to become the final public information, given that t plicity that each o(t) T e is support finite has the true state. in We will A(T), suppose for sim- that o maps so into A. T (A(T)), where Ar(*) represents the set of distributions with finite support. Write over That T. ..., {p,, is, p-, for the Pm] public information. as final while if p Thus, for each is not one of the p. A(T) £ The mean of o is the distribution = via) ^(t) I UT i, o(t)(p.) as t then o(t)(p) vi(o) for some > for all = varies t e t T, e T. on T given by o(t)(p) pdp / " supports of o(t) the p^ are the distributions on T which can arise under ..., , union of (4.1) ' A(T) N = I I o(t) TT{t) (p.) ttT i=l so that, for any t' v(o)(t') I I UT i -IT ^(t) o (t)(p.) p.(t'). =l Every public information structure has mean Theorem 3: Proof: Let m, t ^ T, z N = (4.2) p. ^ , ..., mj, final public information is ^ I p(m.iT) teT ^ " it. be all possible message histories. leads to message-history m. q(t|m.) (4.3). tt(t) ' q('' = with probability p(m.|t). | Suppose that state Then, after m. , the m.), where by Bayes' rule p(m.|t) " ^ TT(t). (4.4) 15 Now in our previous notation while o(t)(q(' ... p, = p(m.|t). ..., M), ]x{a) into our present notation gives us |m. )) are p^ the distributions q (•im.)(i = l, Hence translating the definition of M y(o) = (t) I ^(t) p(m.|T) q(t|ni.) I teT i=l ^ (4.5) ^ M q(t|m.) I ' i=l p(m. |t) I it (t) rearranging (4.6) by (4.4) (4.7) teT ' M p(m.|t) ^(t) " I =l ^ -n(t) since i = p(m. Z This proves Theorem 3. 1 1 ) = 1. Next we show that \i{a) = is also it sufficient for centralized announceabil ity. Theorem If via) = T ^ Aj: Let o: 4j_ then IT, (A(T)) be any function. mediator who knows a t can induce public information struc- ture o by public announcement. Proof: As before, let weight in some o(t). ..., |T|; j = 1, p, , .... p^ be The function ..., 2, o (Here N. the distributions is thus we ignore the Cartesian product structure.) 1 distribution q(*|m.) on T T given positive described by o(t )(?•)> simply the elements number i = of 2, 1, T and Suppose that our mediator announces message m. with probability p^m-lt'') when the state J on is t . Then the posterior J is given by w i P(^.i!t^) teT J ^(t^) (4.8) 16 Simple algebra shows this posterior distribution will that be p. if we set p(ni.|t^) = 7T(t^'Mp-(t^) J remains to check that, It ^(t) o(t)(p.)] I ^sT J (4.9) J' with this specification, Tr(T) I o (t) (p.) • p.(t') J teT j = (a)(t') y (4.10) J and by assumption this is equal to iT(t This proves Theorem ). T ^ Thus we have seen that every function o: tor can publicly reveal This means that exactly as much or as little as is desirable, only to the condition on the mean. tt can be a mediasubject Without a mediator, controlled revelation independent of T harder: is Theorem Suppose that 5j_ be elements of T., t^j t. = i z 1, T. public information P(t') p(t") Proof of Theorem = (in because Ti t- in a a player from T, 1, 1 ' s i, let t'. . , , i j ?* iV be Let . t'. , First, 5: is is t'. t'.' Then if the distribution p(t) can occur as (4.11) note that (4.11) independent of t., TT(t,) TT(t,) i ?! -"(t^) is true of the prior distribution j , tt so that each side of (4.11) (tp) ). preserved by any revelation by a ... a rT(t player. Next, Suppose is we 1 makes point p of public information satisfying (4.11). possible public posterior after I's revelation. I I Suppose For each e . p(t') p(t"). = revelation, beginning at Let p' be n. t decentralized revelation system, we have the obvious notation) show that (4.11) is ..., but not necessarily respectively. a 4. Ar(A(T)) with mean realized through (randomized) central announcements. p see To 1. ^ this, we notice that I = p(m.it'') Z = tp no new II T, II = t,. Then since t' and t_' ^^^ indistinguishable to information on the likelihood ratio between them can emerge revelation: 17 P'(t') P'(t') P(t') P(t') _ " Likewise, since P'(t") 1 (4.12) cannot distinguish between t" and t" P(t") _ (4.13) Since (4.11) can be rewritten as p(t" P(t' equations (4.14) p(' (4.12) and (4.13) assure us that, information p, then it also holds for p'. if (4.11) holds for the public This concludes the proof of Theorem 18 APPLICATION 5. A gap in the proof of Theorem mixed revelation strategies, information. Conceivably slightly that t, with mutual = then x; circumspection Theorem information. 2 until that have final as point with p(x, x) p(x, x) and Theorem one = p(y, y) 5 0, i of its (e.g.) for they had mutually revealed t have, possible be stochastically convey 1 hint to very might respond cautiously, and they could proceed information is "whether or not to which for messages might ensures 5 1. given above was that we did not allow for 1 i.e., it PROOF OF THEOREM 1: = (x, cannot x)." public and p(x, y) > p(x, y) p(y, this 0, p(y, x) guarantees that this cannot be. happen. appropriate the For the required Thus the revelation system would information x) > 0. states, at least one Hence we would have (4.15) 19 APPLICATION 6. 2: BARGAINING: Consider two players bargaining over an object. the object is private information, E {I's value exceeds 2's) and F = If player whether 2 or the object, owns event the not The events {2's value exceeds I's) are both possible. natural is for the players to want to know We occurred. has E and the values are independent. = it Each player's value for whether ask this information could conceivably be revealed by communication between them. If we write for u value and I's either of two values (u,, u^; v, for v value, 2's and each can take on Vp), and if , u^ > v^ > U2 > v^ (4.16) then we need to reveal whether or not the event course, might revelation would tell full necessary be us whether subsequent for incentive = F or (up, not reasons F v, ) has occurred. has occurred, not to reveal Of but (in it the course of discovering whether there are gains from trade) just what the values are. Hence we want system that reveals whether a without leaking further information if With for his a mediator, and type, such then a players will bargain if "E" "E" or F has occurred, The mediator asks each player or "F". If everyone expects that announced, and will walk away if "F" is but has occurred. E system is simple. announces E is, the then truth-telling is an equilibrium. Without a mediator, this, apply Theorem p(Up v^ Equation (4.17) 5 the to get, p(u2, v^) = revelation cannot be achieved. partial see for any public information system, p p(up v^) p(u2, v^) (4.17) cannot be satisfied with p(u,, strictly positive but p(up, To v, ) zero. Hence v,), p(u2, v^) and p(u,, Vp) any system of public communica- tion that establishes whether or not there are gains from trade must also leak 20 further information. It is not possible, without a mediator, for the players simply to find out whether or not there are gains from trade. 21 CONCLUSIONS 7. Costless communication Outcomes matters. games in which would not be equilibria without communication. two-person zero-sum games, but in general become equilibria This does not happen in important. it can be Since players our equilibrium notion should reflect the very often can talk to one another, fact. Players may be prepared to reveal Mediators matter. tion to a trusted mediator (who pass will only on it important if it informa- will not be exploited against the player), when they would keep the information secret in the absence of a mediator. There has been very little work on communication ing the set the set of Bayesian equilibria. algorithm coordination to 5 aspects) gives 5 that game a aspects" enough conflict has communication Theorem 2 is a would be exciting to know how to decompose tion necessary conditions, but will be complicated. determine when would also be very desirable. It Theorem The interaction of incentives with the logical restrictions analyzed in Theorem its Characteriz- games. of communication equilibria may be no easier than characterizing they are far from sufficient. An in and "conflict aspects," as happen cannot first a step general in (relative to in equilibrium this direction. game into "coordina- hinted on page 12. However, I have made no progress on this. Some communication structures are more direction of work on this, see Farrell focus on the trivial messages are ignored. one. In my opinion, munication, which is communication (1985). structure plausible than others. For one But there is no good reason to in which all payoff-irrelevant That is just one among many, and often not a plausible the literature should begin to take more account of com- one of the salient features of the human world. 22 8. 1. 2. REFERENCES Farrell, J., "Communication in Games: Mechanism Design Without a Media- tor," M.I.T. Economics Working Paper 334, Farrell, "Credible Neologisms J., in 1983. Simple Communication Games," GTE Laboratories Technical Note 85-407.5, 1985. 3. Myerson, "Incentive R., Compatibility and Bayesian Equilibrium: An Introduction," Mimeo, Northwestern University, 1983. 4. von Neumann, J., and 0. Morgenstern, The Theory of Games and Economic Behavior, Princeton University Press, 1944. 5. Spence, A. M., Market Signal ing. Harvard University Press, 1970. S85k -^'^ -^^^^o Date Due Lib-26-67 MIT LIBRARIES DUPL 3 TDflO 1 0D562fl31 1