UNIVERSITY OF ILLINOIS LIBRARY AT URBANA-CHAMPAIGN BOOKSTACKS be charged a minimum # ^T" You may JAN 4 2000 NOVi 1999 When renewing hv «i ""o- Previa,/^ £•«. ■*. new «,„. date L162 Digitized by the Internet Archive in 2011 with funding from University of Illinois Urbana-Champaign http://www.archive.org/details/mechanismswithno1527chak BEBR FACULTY WORKING PAPER NO. 89-1527 Mechanisms with No Regret: Welfare Economics and Information Reconsidered FEB 2 3 1989 UK "■UNJiS Bhaskar Chakravorti WORKING PAPER ON THE POLITICAL ECONOMY OF INSTITUTIONS NO. 25 College of Commerce and Business Administration Bureau of Economic and Business Research University of Illinois Urbana-Champaign BEBR FACULTY WORKING PAPER NO. 89-1527 College of Commerce and Business Administration University of Illinois at Urbana- Champaign January 1989 Mechanisms with No Regret: Welfare Economics and Information Reconsidered Bhaskar Chakravorti, Assistant Professor Department of Economics I have benefitted from comments and discussions in seminars at the University of Illinois at Urbana -Champaign, Penn State University and the Game Theory Conference in honor of Lloyd Shapley in Columbus, Ohio, where this paper was presented. I thank the participants. Conversations with Lanny Arvan were especially helpful. The usual disclaimer applies. MECHANISMS WITH NO REGRET: Welfare Economics and Information Reconsidered by Bhaskar Chakravorti Department of Economics University of Illinois at Urbana-Champaign Champaign, IL 61820 February 1988 Revised: December 1988 Abstract This paper achieves four objectives relating to the design of efficient mechanisms in economies with asymmetric information. First, we show that it is impossible to design an individually rational and efficient mechanism using Bayesian equilibrium as the solution concept. Under complete information, however, such mechanisms can be designed. Second, we attempt to bridge this gap between complete information economies and an important sub-class of asymmetric information economies by exploring the possibilities with "mechanisms with no regret" that leak information. A complete characterization is given of "No regret-implementation" using such mechanisms. These mechanisms are played in four stages including a "regret phase" and yield a refinement of Green and Laffont's (1987) posterior implementability concept. Third, it is shown that though such mechanisms cannot implement interim individually rational and efficient performance standards, they can implement ex post individually rational and efficient standards. Finally, it is shown that even under asymmetric information such mechanisms implement any Nash implementable performance standard, such as the core or the Walrasian correspondence. Key Words: Implementation, Bayesian equilibrium with no regret, individual rationality-efficiency, asymmetric information. JEL Code: 026, I have benefited from comments and discussions in seminars at the University of Illinois at Urbana-Champaign, Penn State University and the Game Theory Conference in honor of Lloyd Shapley in Columbus, OH., where this paper was presented. I thank the participants. Conversations with Lanny Arvan were especially helpful. The usual disclaimer applies. 1. Introduction In economies with asymmetric information, a performance standard is a non-empty set of socially "desirable" state-contingent allocations. Given the incompleteness of information about the realized state of the world, a mechanism is required to implement a given performance standard. An implementing mechanism is a game of incomplete information with a non-empty set of equilibrium outcomes that is contained within the standard. {Full implementation requires coincidence of the two sets.) Most economies aspire towards the two basic objectives of individual rationality and efficiency. This paper reports a fundamental difficulty with any attempt to design mechanisms for implementing individually rational-efficient performance standards, and then proposes conditions under which there exists a solution to this crucial problem. In light of the difficulties that were just alluded to, this is the only result that demonstrates the possibility of designing individually rational-efficient mechanisms in a wide class of asymmetric information economies. Moreover, we show that despite the information asymmetry it is possible to implement any standard that is Nash-implementable (i.e. with complete information), such as the core or the Walrasian correspondence. By the Revelation Principle (Myerson (1979), Dasgupta, Hammond and Maskin (1979), Harris and Townsend (1981), etc.), any performance standard that has a non-empty intersection with the set of Bayesian equilibrium outcomes of a game must satisfy an incentive comaptibility (or self-selection) condition. We know of no satisfactory performance standard that meets this condition. Thus, if we are interested in questions of mechanism design in a wide class of environments, the conclusions are negative right away. To surmount this initial hurdle, Palfrey and Srivastava (1987a) invoke Postlewaite and Schmeidler's (1986) restriction on the informational structure — non-exclusivity of information (NEI) — and then ask whether performance standards that were fully implementable under complete information can be fully implemented when information is asymmetric. The NEI restriction essentially requires that if all agents except one pool their private information they can deduce the information of the remaining agent. It ensures incentive compatibility but is clearly restrictive. Unfortunately, some restriction of this nature is necessary to evaluate the implementability of natural economic performance standards. The striking message of Palfrey and Srivastava is: despite the restriction and despite the fact that incentive compatibility is satisfied, full Bayesian implementation of either the interim or the ex post efficiency standard (as defined in Holmstrom and Myerson (1983)) is impossible. This is in sharp contrast to the positive results obtained under complete information. Palfrey and Srivastava's (1987a) negative results provide the starting point for this paper. We extend their arguments and show that the situation is even more grim than is implied by their results. Define an efficient Bayesian mechanism as a game such that all of its Bayesian equilibrium outcomes are efficient (in either an interim or ex post sense). However disappointing the Palfrey and Srivastava conclusions may be, they do not imply that we can never hope to design an efficient mechanism since there still exists the possibility of Bayesian-implementing (as opposed to fully Bayesian-implementing) the efficiency standard or one of its subsets. The first part of this paper argues that, indeed, even under NEI, we cannot hope to find an efficient Bayesian mechanism if we impose an additional restriction of individual rationality on the standard. The latter rules out uninteresting outcomes that assign all resources to one individual. Such outcomes are efficient and trivially implementable. To summarize, both the interim and the ex post individual rationality-efficiency standard (and their respective subsets) are non-implementable in Bayesian equilibria. An additional implication of this part of the paper is: when one is interested in checking for impossibility results, "implementation" (as opposed to "full implementation") is the notion that one should focus on; a standard that is not fully implementable may be implementable whereas one that is not implementable can never be fully implementable. These impossibility results indicate the cost, in terms of social welfare, that is imposed on society by the presence of information asymmetry. Under complete information, the literature on Nash-implementation, (Maskin (1977), Hurwicz, Maskin and Postlewaite (1984), etc.) has shown that individually rational-efficient mechanisms do indeed exist. Can this gap be filled by a mechanism that not only allocates resources but also "leaks" information? The next step in the agenda is to tackle the problem of non-existence of individually rational-efficient Bayesian mechanisms by addressing the question posed in the previous paragraph. We take advantage of the fact that the mechanism designer has the flexibility to not only choose the game but also how it is played. Following the lead of Green and Laffont (1987), we study the case of mechanisms with a "regret" phase. Green and Laffont do not provide any specific structure to the changes in the rules of the game. In this paper, we adopt a particular sequence of moves that allows for a regret phase in a simultaneous-move game. This sequence is a slight modification of the structure that is implicit in Green and Laf font's model. The resulting equilibrium that we obtain is a refinement of Bayesian equilibria and of posterior implementable equilibria (due to Green and Laffont). The structure that we employ evolves in four stages and is informally described with the help of an example in the following paragraphs. Example 1: Consider a game with two players and three states of the world: "Rain", "Shine" and "Clouds". Player 1 chooses T or B and is completely uninformed. Her prior beliefs are that there is an equal chance of any one of the states occurring. Player 2 chooses L, M or R and is completely informed. Their payoffs are given in Figure 1. This is the original game. The pure strategy Bayesian equilibria of the original game are given in Figure 2. A game designer can introduce a regret phase in rules of play of this game. [Insert Figures 1 and 2 here.] The game with a regret phase is played in four stages. In the first stage, the players and the game designer calculate the set of Bayesian equilibria in the original game. This constitutes the maximal set of possible outcomes. Next, Nature selects a state of the world. In the second stage, the players submit to the designer the actions they intend to take in every equilibrium. These submissions are not commitments and are made public knowledge by the designer. In the third stage, the agents are given the opportunity to entertain the desire to unilaterally deviate from one or more of the Bayesian equilibria. In the final stage, (as is usually assumed in game theory) by some unbiased method a particular Bayesian equilibrium, say s, is selected from the set comprising equilibria from which no player wants to unilaterally deviate. s becomes common knowledge among the players and the designer. The game designer distributes the outcome/payoffs to the players by checking the actions that each player had submitted for s and using the payoff matrices in Figure 1. It can be observed that this structure is different from the one that underlies Green and Laffont's notion of posterior implementable equilibria. In their case, the unbiased selection of the equilibrium is done before the second stage. Thus, they would require that in the second stage the players submit the actions that they intend to take in the particular equilibrium that is chosen. To return to the example, a Bayesian equilibrium will survive the third stage if and only if the information conveyed by the second stage public submissions is such that neither player has an incentive to unilaterally deviate in stage three from the action he/she had submitted for that equilibrium in the previous stage. Such an equilibrium is referred to as a Bayesian equilibrium with no regret. Since equilibria are pairs of strategies that are common knowledge functions, Player 1 could conceivably acquire some information if Player 2's action were associated with a unique state in some Bayesian equilibrium. However, not all Bayesian equilibria convey information in this manner. Let the set of Bayesian equilibria with no regret be denoted E(D. Every Bayesian equilibrium with no regret must be immune to information revealed by any other Bayesian equilibrium with no regret. This follows from the fact that in stage four a Bayesian equilibrium with no regret will be chosen to determine the final outcome. Given that in stage two the players do not know which one will be chosen, they will not have an incentive to unilaterally make arbitrary submissions for any equilibrium in E(D. This rules out unilateral manipulation on the part of the players by the submission of messages that could mislead others. An equilibrium that is not in E(D does not convey information since the players may submit messages corresponding to such an equilibrium to mislead others about the state of the world knowing that such equilibria shall never be used to determine the final outcome. The calculation of the set of Bayesian equilibria with no regret may appear somewhat circular. We shall show how it is determined for the example at hand. * Observe that the functions s', s" and s carry information if the state is "Clouds". The equilibria s' and s remain self-enforcing even if the information were revealed to Player 1 in the "Clouds" state by s', s" * * or s . This gives us a sufficiency condition: s' and s are Bayesian equilibria with no regret. Are there any others? A necessary condition is that a Bayesian equilibrium with no regret must not be destroyed by information conveyed by another Bayesian equilibrium with no regret. It can be checked that in s and s" Player 1 has an incentive to deviate in the "Clouds" state given the information conveyed by s' and s . Thus, s' and s constitute the set of Bayesian equilibria with no regret, E(D. We conclude this example with two observations. First, note that the * Bayesian equilibria with no regret (e.g. s ) are not necessarily complete information Nash equilibria of the game. Second, if we had employed Green * and Laffont's criterion, s, s' and s would have survived and s" would have been eliminated. In sum, our criterion of regret requires greater robustness from an equilibrium. While this is a strength from the viewpoint of stability, it makes it more difficult to guarantee existence of such equilibria. In this paper, we shall, however, prove the existence of the set of Bayesian equilibria with no regret for any mechanism that we use for our implementation result. The positive results in the paper make use of mechanisms with a regret phase as outlined in the example above. A complete characterization is provided for performance standards that are No-regret-( N R)-implementable (i.e. we use Bayesian equilibrium with no regret as the appropriate equilibrium concept). The characterization results are used to prove the following arguments. The interim individual rationality-efficiency standard cannot be NR-implemented. However, the ex post individual rationality-efficiency standard is NR-implementable. In general, any standard that is Nash-implementable under complete information is also NR-implementable under asymmetric information. The sub-class of environments in which we demonstrate these results satisfy the NEI condition. Thus, we are successful in bridging the gap (in terms of efficient mechanism design) between NEI environments and complete information environments. As mentioned earlier, NEI is clearly restrictive. However, given the constraints imposed by the Revelation Principle, some condition "close" to NEI is also necessary to obtain any positive result (Blume and Easley (1987) prove the necessity of NEI for implementability of the Rational Expectations Equilibrium standard). Moreover, NEI is satisfied in a large variety of environments involving informational asymmetry. A few of these are listed below: (i) In large enough economies the private information of any single individual can be fairly accurately predicted by pooling the information observed by the rest of the population. (ii) Individuals can be one of several "types". The proportions of each type in the population is known. (iii) Information may depend on physical possession of some commodity whose different varieties and total quantity is known. Each individual's private holding is not common knowledge. (iv) There are at least two fully informed individuals in the economy, e.g. in certain markets with several sellers and buyers, all sellers are presumed to be informed and all buyers are uninformed. (v) Each individual has at least one other person who can verify his/her information, e.g. close relatives, witnesses, etc. (vi) Information is held by coalitions of individuals with the minimum size of each coalition being two, e.g. family units with common preferences over goods, preferences are determined by the school one attends, collaborators who know each other's information, etc. In sum, given that most individuals exist within certain social institutions, truly exclusive relevant information may indeed be a rarity. The success of real-world information collection mechanisms such as courts, inquiry commissions, etc., which rely on gathering enough individuals to be able to deduce the private information held by one individual, indicates that NEI environments constitute a significant class. Thus, for a vast number of realistic applications, a positive result that relies on the NEI condition represents a significant improvement over impossibility results or positive results under complete information. Juxtaposed with related literature, our findings have several interesting implications: (i) Green and Laffont's (1987) conclusions regarding the application of games with a regret phase to the problem of mechanism design were largely negative. We employ a stronger definition of implementability and equilibrium and show that such games do, indeed, have important applications towards deriving positive results. (ii) The most widely studied concept of implementation is Nash-implementation whose origins lie in the classic work of Maskin (1977). Maskin's characterization of Nash-implementable standards (also see Saijo (1988)) has been criticized on the grounds that it applies only to complete information settings. Our results show that the class of standards that Maskin had identified as being implementable can be implemented (by mechanisms with no regret) even when information is asymmetric. (iv) Finally, we find that appropriate refinements of Bayesian equilibria broadens the scope mechanism design — a fact discovered by Palfrey and Srivastava (1987b) for economies with private values and by Moore and Repullo (1988) for economies with complete information. Our results hold for more general economies including those with common values. In addition, the refinement is generated by changes made by the mechanism designer and we do not rely on agents employing a specific refinement criterion — observe that the Palfrey and Srivastava (1987b) mechanism is based on elimination of dominated strategies and not successive elimination of dominated strategies. The latter criterion would be more appealing from a game-theoretic standpoint. The following section presents the basic model that we shall use — an exchange economy with privately informed agents. Section 3 introduces Bayesian equilibria with no regret. Sections 4, 5 and 6 present results on Bayesian-implementation and NR-implementation respectively and their welfare inplications. The final section discusses possible extensions. 2. Preliminaries 10 An asymmetric information economy, e, is a triple {L, N, 8}. L is a set of goods, N is a set of agents and 6 is a set of states of the world. All of these sets are non-empty and finite and the cardinalities of L and N are given by I and n, respectively. & is the domain of all asymmetric information economies. In the definitions that follow, we focus on a given e e &. An explicit reference to e is dropped to minimize notational burden. Let e = {L, N, 9} be given. Every agent i e N is completely * I characterized by a list (u , w , IT , q ), where u : R x 6 -» R is agent i's l 1 1 M I + von Neumann-Morgenstern utility function; q (* 0) e R is agent i's initial endowment of goods; IT is agent i's natural information partition * of 9 and q : 9 -> (0, 1) is agent i's prior probability distribution on 9. Each constituent of this list is assumed to be given exogenously, and is common knowledge in the sense of Aumann (1976). Let the function I : 8 -> i IT be defined by I°{6) = ie' € 9: there exists n € TT such that 6, Q' € i i ii it h The latter is agent i's natural information set in state B. By "natural" information we refer to the information structure that the agent is exogenously endowed with. This distinguishes it from the information that can be acquired endogenously. In the sequel, let TT = x TT and Q = Y u) . In addition, unless specified otherwise, x = [x ) and x = M€N 1 r 1 1€N -i (xr) J j€N\(l> A = iz e R : Y z < Q} is the set of feasible allocations. A + ^1<=N 1 - state-contingent allocation is a random variable /: 9 -» A. F is the domain of such functions. A performance standard

[0, 1] defined by Bayes' Law, i.e. for all 9 e 6 and for all $ e P(8), * q (0) if € ?; q (0. ? ) = i V e » *, (8 '> 0, otherwise. /Igent i's expected utility from f e F, given } e PCI (Q)) is £ q (8', 1 0'e?' ?)u (/ O'), 0'), and is written more compactly as EU (/ 9). Agent i's ^-expected lower contour set at f is given by EL (f 9) = {g € F: Et/ C/ 9) > EU(g I m The domain under consideration, S is defined as the collection of all economies e ={L, N, Q} € g, that satisfy the following: [Al] {strict monotonicity of preferences) Vi € N, V0 € 0, u (., .) is t strictly increasing in z 6 R , and [A2] (non-exclusivity of information) Vi € N, V0 € 0, n I (0) = ' J€N\(1> j . [A3] | J\T | > 2. A mechanism is a game T = {JV, M, £}, where, given that M is agent i's message (or action) space, M = x M ; and £: M -» i4 is an outcome function. Agent i's strategy is a random variable s : Q ^ M such that s. is IT -measurable. Let S be the domain of such functions. Let S = X S . 11 1€N 1 3. Bayesian Equilibria with No Regret The fundamental solution concept for games with asymmetric information is that of Bayesian equilibrium due to Harsanyi (1967). This concept is 12 defined as follows: s € S is a Bayesian equilibrium of T = {N, M, £,) if Vi e N, V0 e 0, Vs' € S , i 1 £«(s\ s ) € EL (£os I I°(Q)). 1-1 1 ' 1 Let £°(r) denote the set o/ Bayesian equilibria of r and £°(D = {£°s e F: s € E (D, T = {N, M, £». We shall use the following structure to formally specify the changes in the rules of play introduced by a regret phase in a game. The construction given below shall be used throughout the paper. Alternative constructions are, of course, possible. The eventual objective of this paper is to provide one method of construction that a mechanism designer can employ. An original game T played with a regret phase is denoted T . Suppose T is such that E°(D * . Let K = {k : £°(D -> M defined ^ ill by: 39 € such that Vs € E°(D, s (0) = k (s)K Let I (0, s) = {0' € ill J (0): s (0') = s (0)} be the information set revealed by s in 0. l -l -I T is played in four stages. STAGE 1: The agents and the mechanism designer compute E (D. BETWEEN STAGES 1 AND 2: Nature selects e 8. STAGE 2: Each agent i submits k € K to the designer. The designer makes k public knowledge. STAGE 3: Each agent i is given the opportunity to entertain a desire to deviate from any »c(s) for s e E°(D. Let the set of Bayesian equilibria 13 that have not induced a desire for unilateral deviation by any i e N be denoted £(D. STAGE 4: By some unbiased method, s € £(D is selected and is common knowledge among the agents and the designer. The designer chooses the outcome £(k(s)). £(D is the set of Bayesian equilibria with no regret. Let E (D = P {£°s € F: s e £(r), r = iN, M, £}}. A game with a regret phase, T , for which E(D * is a mechanism with no regret. For the purposes of this paper, the following conditions shall suffice for determining E(D. (A) is a sufficient condition for £(D * and (B) is a necessary condition. Condition (A): If 3s e E°(X; such that Vs' € E°(T;, Vi e JV, Vs" € S , 1 i ^o( s , ' ( S ) e el^os | r ce, s';;, then s € E(D. Condition (B): If E(D * 0, then Vs, s' e E(D, Vi € A/, Vs" € S , €o(s", s ) € ELC^os I ire, s'U l -i i ^ ' i Condition (A) provides a method for determining at least one Bayesian equilibrium with no regret. If there exists an equilibrium in E (D that is immune to information revealed by all the Bayesian equilibria, it must be a Bayesian equilibrium with no regret. Condition (B) provides an internal consistency condition for the set E(D. Every Bayesian equilibrium with no regret must be immune to information revealed by any other Bayesian equilibrium with no regret. This follows from the fact that in stage four a Bayesian equilibrium with no regret will be chosen to determine the final outcome. Given that in stage two the agents do not 14 know which one will be chosen, they will not have an incentive to unilaterally make arbitrary submissions for any equilibrium in £(D. This rules out unilateral manipulation on the part of the agents by the submission of messages that could mislead other agents. An equilibrium that is not in £(D does not convey information since agents may submit messages corresponding to such an equilibrium to mislead others about the state of the world knowing that such equilibria shall never be used to determine the final outcome. 4. Bayesian-Implementation and Welfare Implications The classical approach to welfare economics has been to identify the subset of allocations that are Pareto-efficient within the set of all physically and technologically feasible allocations. Among these efficient allocations, attention is generally focused on ones that are individually rational. However, in asymmetric information economies, these welfare evaluations must also take account of informational constraints — an uninformed social planner cannot identify the set of individually rational-efficient allocations in the absence of complete information about the agents' preferences. The notion of an individually rational-efficient allocation would vary depending on the extent of insurance that the allocation provides each agent. Individual rationality and Pareto-efficiency are thus extended to take account of different levels of insurance and the appropriate notion depends on the timing of the welfare analysis (see Holmstrom and Myerson (1983) for a detailed discussion). The two primary concepts of efficiency are: 15 Interim-efficiency: A state-contingent allocation / is interim-efficient if there is no g e F such that Vi e N, V9 e 6, / € £L (g i | r°C8)) and / € int(EL (g | f°(9)) for some i e A/ and some 9 e 8. Ex post-efficiency: A state-contingent allocation / is ex post-efficient if there is no g € F such that Vi € N, V9 e 0, / e EL (g i ' {9}) and / e int(EL (g | {9}) for some i € N and some 9 e 8. Given w e E defined by w(9) = u for all 9 € 8, let T l = (f e F: f is interim efficient and Vi e N, V9 € 8, w € EL C/ | I°C9;;} and P e h {/ <= F: / is ex post efficient and Vi 6 N, V9 € 8, w € EL C/ {9}} denote, respectively, the sets of interim individually rational-efficient and ex post individually rational-efficient performance standards. Once a social planner decides on the appropriate notion of efficiency, the question of implementing an efficient performance standard arises. Once again, the informational asymmetry poses a constraint. A naive mechanism in which each agent is asked to report his/her private information to the planner will generally not ensure truth-telling as the unique equilibrium strategy profile. Thus, we need to be guaranteed the existence of a mechanism that implements the given standard. This is defined (for the case where Bayesian equilibrium is the solution concept) as follows: A performance standard

. 1 ' 1€N 1 " i€N 1 1 Let 8 : 8 -» 8 be the deception induced by a and defined by 6 (8) = H a (I (0)). By assumption A2, 8 is a well-defined function. A performance standard

fgoe a e ELff»e* I Ae)j;, then /°8 € such that

is 17 fully Bayesian-implementable by a game in e if and only if it satisfies Bayesian monotonicity. Proof: See Palfrey and Srivastava (1985). ■ The following proposition provides a parallel characterization of Bayesian-implementability. Proposition 1: Let e = {L, N, Q} be an economy in 8. A performance standard

as follows. Let L = {X, Y) with the quantities of the two goods being denoted by the corresponding lower case letters. Let N = {1, 2, 3, 4} and let 9 = {a, b, c). IT = TT = {(a), (b), 1 2 (c)> and TT = IT = {(a), (b, c)>, i.e. agents 1 and 2 are fully informed and agents 3 and 4 cannot distinguish between states b and c when one of them occurs. An allocation z is written as (x , y ) u> = ((0, 1), (0, 1 1 16N 1), (1, 0), (1, 0)). Each state is equally likely, i.e. q (a) = q (b) = i i * 1 I q (c) = - for all i € N. u : \R x 8 -» IR is given as follows: Vi € {1, 2), V9 € 9, a (z , 9) = x + l.ly , i J r u (z . e) = 3 3 ( *3 + y 3 ] ' if 9 = a 0.25U + y ) if 9 = b 3 3 0.75(x + y ) if 9 = c. 3 J 3 u „ (z „' 8) = 4 4 4 4 if G = a 0.75(x + y ) if 9 = b 4 4 0.25(x + y ) if 9 = c. 4 4 It can be checked that T * 0. Consider the following state-contingent * allocation, denoted / : 19 z = 1 z = 2 z = 3 Z = 4 a b c (0, 1) (0, 1) (0, 1) (0, 1) (0. 1) (0, 1) (1, 0) (0, 0) (5,0) (1, 0) (2, 0) (i 0) 3 * f 6 p\ Also check that assumptions A1-A2 are It can be checked that / e T satisfied. Choose any

and g 6 F. Write /(a) as (x, y) and g(a) as (x\ y'). By construction, the utility functions of agents 1 and 2 are state-independent. Therefore, given that they are completely informed, the hypothesis of Property M is trivially satisfied for these two agents. The following relationships imply that the hypothesis of Property M is met for the remaining agents: x + y fc x' + y' >* 0.5[0.25(x + y ) + 0.75(x + y )] £ 0.5[0.25(x' + 3333 33 33 3 y') + 0.75(x' + y')]. 3 3 ^3 x + y > x' + y' => 0.5[0.75(x + y ) + 0.25(x + y )] > 0.5(0. 75(x' + 4444 44 4^4 4 y') + 0.25(x' + y')]. 4 44 For P to satisfy Property M, we must have f°Q e and x > 0. 3 4 20 Choose c such that min{x , x ) > c > 0. Consider an alternative rule h € F 3 4 defined by h(a) = /(a), Vi e {1, 2}, V0 e (b, c), h(8) = /(a), i l h (b) = (x -e, y ), 3 3 3 h (b) = (x +e, y ), 4 4 4 h (c) = (x +-, y ), 3 3 3 3 h (c) = (x --, y ). 4 4 3 4 Thus, El/ (h 0.5[0.25(x + y ) + 0.75(x + 4 • / 4 434 4" , 4 4 y )] = EU (/oG a I {b, c». Thus, /oe a 0 g IT as in the previous proof, i.e. for all i e N, for all n e TT , a (ir ) = {a). Thus, a is a 111! CCMO and for all 9 e 9, 9 a (9) = a. Choose any

x' + y' => 0.5[(x: + 1.5y ) + (x + 0.5y )] £ 0.5[(x' + 3 J 3 3 J 3 3 3 3 ^ 3 3 1.5y') + W + 0.5y')]. ^3 3 3 Thus, the hypothesis of Property M is satisfied. For T to satisfy oc oc Property M, we must have f°0 e and y > 0. Choose c such that min{— x , y } > e > 0. Consider an alternative rule h e n 3 M F defined by: h (b) = (x +l.le, y -e) l 11 h (b) = (x -1.1c, y +e) 3 3 J 3 Vi € {1, 3>, V9 € {a, c>, h (9) = /(a) Vi € {2, 4}, h = / oe a . Thus, EC/ (h {b» = x - l.le + 1.5c + 1.5y = x + 0.4c + 1.5y > x + 3 ' 3 y 3 3 ^3 3 l.5y 3 = EI/ (/«e a | {b» and £1/ (h | {b» = x + l.le + l.ly - 1.1c = x + l.ly = EU (/oe a I ). For all i € N and all 9 e 6, /<>e a € EL (h I {9}). li 1 i ' Thus, foQ <£

e . Given that this holds for all

°s. By Condition (B) the following holds for all i € N, for all 9 e 8, for all s' e S , for all s" <= E(D: ^o( S ', s ) € el (/ I i°(e)) n eu/ I J( . s "» (i: Choose a CCMO a. Next, suppose that for all i € N, for all 9 € 8, for all 9' = 9%), for all s" € £(D, for all g & EL(f \ I°(e')) fl EL (/ I * 7 (9', s")), the following holds for all s € S: go9 a € el (foe* | I°(e)) f) ELifoe" | i (e, s*); (2) Given (1) and (2), for all i e N, for all 9 e 8, for all s' e S , for all * s e S, the following holds: €•($;, s^jce" € ELCfoe* | i°(9)) H el (foe* \ lie, s*)) (3) By definition of a, s »8 is IT -measurable for all i € N. By Condition (A), (3) implies that s°9 a € E(D and /°9 a € E (D. By definition of full NR-implementation of 0.5[0.25(x + y ) + 0.75(x + y )] > 0.5[0.25(x' + 3333 33 33 3 y') + 0.75(x' + y')]. 3 3 3 x + y > x' + y' =» 0.5[0.75(x + y ) + 0.25(x + y )] > 0.5[0.75(x' + 4444 4" / 4 4 y 4 4 y') + 0.25(x' + y')] and 4 44 x + y > x' + y' =* 0.25(x + y ) £ 0.25(x' + y') 3 y 3 3 y 3 3 J 3 3 J 3 x + y > x' + y' => 0.75(x + y ) £ 0.75(x' + y') 3 ^3 3 y 3 3 3 3 "^3 x + y > x' + y' => 0.75(x + y ) > 0.75(x' + y') 4 y 4 4 y 4 4 4 4 J 4 x + y >: x' + y' =» 0.25(x + y ) £ 0.25(x' + y') 4 4 4^4 4"'4 4" , 4 Given these observations, the hypothesis of Property Ml is met for any game T. The rest of the proof of Theorem 5 follows the proof of Theorem 1 and it is established that f°B 4 ;; => {go9 a e ELffoQ* | ;>, then f°Q e

. Definition: Vi <= N, m satisfies Property y\i if the following conditions hold: (i) H tt (j) * 0. 1 I j€N\(i> j J (ii) 3/ e ? such that Vj e JV\{i>, /(j) = /. (iii) Vj e W\U>, 5(j) = 0. * * Definition: Vm € M, 9 (m) is defined such that f\ n (i) = {8 (m)}. 1 ' 1€N i Definition: Vi 6 N, Vm € M , (m ) is defined such that tt ( j) -1 -1 1 -1 ' ' j€N\(i> j J = {9(m )>. l -l Definition: Vm € M, Kim) = ii e W: Vj € JVMO, 5(i) 2: 6{j)>. (II) £: M -> A is given by the schematic diagram in Figure 3. At Proof of Lemma h_ Choose / e . If possibility (b) occurs, by Case 3, we have £©(s\ s ) € ELif {9}) which implies that for all s" € S, %°{s\ s) € EL(/ | I°(9)) f) ELC/ | 7(9, s")). By Condition (A), s e E(^()) c , for all 0' e 6, s. (0') < 13 13 J3 s' (6'). Let m = s(e) and m* = (s'(0), s (e)). We shall show that 13 1-1 there exists i € N such that the following hold: € (m') > ^(m) (*) V8' 6 8, ^(sjO*), s (9*)) ^ ^(s(9*)) (**) There are two possibilities: (i) Possibility 1: There exists j e N such that K{m) = {j}. Therefore, for all i € N\{j), m does not satisfy Property y\i. Choose i € N\{j}. By definition, Property y|£ is not met even if i deviates to s'_. By the outcome rule associated with Case 4A, we have £, (m') = Q. Since |AT\{j}| > 1, there exists i € N\{j} such that £ (m) * Q. Thus, (*) holds. (ii) Possibility 2: There does not exist any J 6 N such that Kim) = {j}. In this case, by construction, £(m) = 0. By the outcome rules associated with Cases 2A, 3A and 4A, and given that for all 0' e 8, for all / € 0. Thus, (*) holds. To check that (**) is true for the agent i for whom (*) holds, choose 0" e 8 with 0" * 0. There are, again, two possibilities: (a) Possibility 1: There exists j e N such that Kisie")) = (J}. The arguments given in (i) above apply. Thus, for all k e N\{j), £ is'ie"), k k Kb s (9")) = fl. Also, £(s'(9"), s (9")) = C(s(8")). Thus, (**) holds for -k J -J i. (b) Possibility 2: There does not exist j € N such that K(s(9")) = {j}. By construction, for some / € s = /"°e a . We need to show that /°9 a € ). Suppose i unilaterally deviates to s' e S where for all 9 6 9, s'(9) = M (s (9), g, 5'U)). S'U) is such that for all J € N\{i), for all 9 e 0, 5'(i) > s (0). By construction, for all 9 € 9, 6 (s (9)) = 9 (s(9)). J3 1-1 Thus, for all 9 € 0, (s'(9), s (9)) satisfies Case 3A and £((s'(9), i -l i * a s (9)) = g(6 (s (9))) = g(Q (s(9)) = g(9 (9)). Observe that, by -l ° i -l * Condition (B), s € £(S*(9 a | {9}) for all 9 € 8 and we conclude that for all s e S, goe a e EHfoe" | r°(e)) fl EUf°e a | r(e, s)). Given that the last conclusion holds for all i 6 W, by the fact that

9 g

) for e" = e a (e) (+) g * EL (f I , (ii) 3/ €

Case 3: 31 € N such that (i) Vj e N\{i}, f(i) * f(j) and (ii) m_ satisfies Property y I i. Case 3A (i) Kim) = {i} (ii) /(i) € EL(/(J)| {9(m )}), j * L €(m) = /(i)(9(m )) 1 -i Case 3B Otherwise 1 €(m) = Case 4: Otherwise Case 4A: 3i € N with K(m) =