Rationalizability in Games with a Continuum of Players∗
Pedro Jara-Moroni †
Instituto Cient´ıficoMilenio Sistemas Complejos de Ingenier´ıa,Fac. de Scs. F´ısicasy Matem´aticas,Universidad de Chile, Santiago, Chile. ‡ This version March 2010
Abstract The concept of Rationalizability has been used in the last fifteen years to study stability of equilibria on models with a continuum of agents such as competitive markets, macroeconomic dynamics and currency attacks. However, Rationalizability has been formally defined in a gen- eral setting only for games with a finite number of players. We propose then an exploration of Rationalizability in the context of Games with a Continuum of Players. We deal specifically with a special class of these games, in which the payoff of a player depends only on his own strategy and on an aggregate value that we call the state of the game, which is obtained from the complete action profile. In this paper we define the sets of Point-Rationalizable States and Rationalizable States. For the case of continuous payoffs and compact strategy sets we characterize these sets and prove their convexity and compactness. We provide as well results on equivalence of Point and standard Rationalizability.
Keywords: Rationalizable Strategies, Non-atomic Games, Expectational Coordination, Rational Expectations, Eductive Stability, Strong Rationality. JEL Classification: D84, C72, C62.
1 Introduction
The concept of Strong Rationality was first introduced by Guesnerie (1992) in a model of a standard market with a continuum of producers. There, an equilibrium of the market is said to be Strongly Rational, or Eductively Stable, if it is the only Rationalizable Solution of the economic system. Inspired by the work of Muth (1961), the purpose of such an exercise was to give a rationale for the Rational Expectations Hypothesis. Strong rationality has been studied as well in macroeconomic models in terms of stability of equilibria. See for instance Evans and Guesnerie (1993), where they study Eductive Stability in a general linear model of Rational Expectations or Evans and Guesnerie (2003, 2005) for dynamics in macroeconomics. More examples of applications of Strong Rationality can be found in Chamley (2004), where the author presents models of Stag Hunts in the context of coordination in games with strategic complementarities.1
∗This article is part of my dissertation, prepared at Paris School of Economics (PSE) under the supervision of Roger Guesnerie and Alejandro Jofr´e. †I would like to thank Ali Khan, Thomas Boulogne, for helpful conversations; Bruno Strulovici and Nicol´asFigueroa for pointing me out relevant references; and the audiences at the seminars at PSE, Department of Mathematical En- gineering (DIM) of the University of Chile and Economics at Johns Hopkins University; as well as the audiences at Franco-Chilean Congress on Optimization, Chilean Economic Society Meeting, Chilean-Brazilean Workshop on Eco- nomic Theory, Third World Congress of the Game Theory Society and of the European Meeting of the Econometric Society for their valuable questions and remarks. I am especially grateful to two anonymous referees whose constructive criticism and suggestions contributed significantly to improve the presentation of the main contributions of the paper. ‡Postal address: Departamento de Ingenier´ıaCivil, Blanco Encalada No 2002, 5o piso, Santiago, Chile. Phone: +56 2 9784376. Fax +56 2 6718788 E-mail: [email protected]. 1See as well Morris and Shin (1998), Chamley (1999), Desgranges (2000), Guesnerie (2002), Gauthier (2002), Chatterji and Ghosal (2004), Desgranges and Heinemann (2005) and Ghosal (2006) although not all these models fit necessarily the framework that is presented below.
1 The Rationalizable Solution of the economic system used by Guesnerie in the definition of Strong Rationality, refers to the concept of Rationalizable Strategies as defined by Pearce (1984) in the context of games with a finite number of players and finite sets of strategies. Rationalizable Strategies were formally introduced by Bernheim (1984) and Pearce (1984) as the “adequate” solution concept under the premises that players are rational utility maximizers that take decisions independently and that rationality is common certainty. In both papers and later treatments, however, the definition and characterization of rationalizable “solutions” were developed for games with a finite number of players. 2 On the other hand, several economic models that deal with strong rationality feature a continuum of agents, including the seminal work by Guesnerie (1992) and the works that are mentioned in the first paragraph above. In those models, the definition of the Rationalizable Solution is usually intu- itive and/or context-specific, adapting the original definitions and characterizations of Rationalizable Strategies to such models. To our knowledge, there is no established definition for Rationalizable Strategies, or Rationalizability for what matters, in a general framework with a continuum of players. The present paper addresses this gap between the established theory and its economic applications exploring the ideas of rationalizability in a class of non-atomic non-cooperative games with a continuum of players. One part of the task is to provide a suitable model of such games, that allows to define and characterize Rationalizable Outcomes. We will work in a framework where payoffs depend both on players’ own actions and on the average of the actions taken by all the players. We will call this average the state of the game, and we will define the sets of Point-Rationalizable States and Rationalizable States. This last approach is not evident nor a generalization of finite player games, since in “small” games, and opposed to what we do here, players can actually affect directly and unilaterally the payoff of other players. Our approach will be that of an iterative process of elimination of non-best-responses (or non- expected-utility-maximizers). This process is proven to deliver the set of Rationalizable Strategies in games with a finite number of players, compact strategy sets and continuous payoff functions (Pearce, 1984; Bernheim, 1984).3 In a more general context, however, the Rationalizable Set may fail to be the result of the iterated elimination of non-best-responses since it should be required to be as well a fixed point of the process. Recent papers address this issue providing examples in general normal form games with finite number of players, where the set obtained as the limit of this process is not a fixed point of the process itself. 4 The authors go beyond to explore more complex iterated processes of elimination of strategies (see for instance Dufwenberg and Stegeman, 2002; Apt, 2007; Chen et al., 2007, 5). The question emerges on how should this process be defined in the context of a continuum of players? When can we claim that the result of the iterative elimination process gives a set that we may call Rationalizable? Our main results are Theorems 3.8 and 3.11 where we characterize these sets as the results of iterated elimination of states. More precisely, we extend Propositions 3.1 and 3.2 in Bernheim (1984) to (Point-)Rationalizable States in the context of games with compact strategy sets, continuous utility functions and a continuum of players. The need for these two Theorems comes from the proof of Proposition 3.2 in Bernheim (1984), where a convergent subsequence extraction argument is used, argument that is no longer valid in the context of a continuum of players. A different limit concept is needed to conclude. Moreover, certain measurability properties must be required to have a well defined process of iterated elimination. The remainder of the paper is as follows: in section 2 we introduce games with a continuum of players and some notation; in section 3 we discuss how the rationalizability arguments may fit in this context. For the particular class of games with an aggregate state, in subsections 3.1 and 3.2 we define Point-Rationalizable and Rationalizable States and we characterize them in subsection 3.3. As Corollaries we obtain their convexity and compactness. In section 4 we explore the relation between Point-Rationalizable and Rationalizable States. We close the presentation with comments and conclusions in section 6. 2See as well Tan and da Costa Werlang (1988). 3See Propositions 3.1 and 3.2 in Bernheim (1984). Proposition 3.2 proves as well, as Bernheim and Pearce claim in their articles, that their definitions are indeed equivalent. 4 The problem rises when the assumptions on utility functions and strategy sets are relaxed (namely the cases for unbounded strategy sets and/or discontinuous utility functions). 5Dufwenberg and Stegeman (2002) and Chen et al. (2007) put emphasis in Reny’s (1999) better-reply secure games.
2 2 The model
We consider games with a continuum of players. Schmeidler (1973) introduced a concept of equi- librium and provided existence results in games where a strategy profile is an equivalence class of measurable functions from the set of players into a strategy set, and the payoff function of a player depends on his own strategy and the strategy profile played. A different approach was presented later by Mas-Colell (1984) 6 and more general frameworks can be found in Khan and Papageorgiou (1987) and Khan et al. (1997) as well. For a comprehensive review of games with many players see Khan and Sun (2002). We will focus mainly in Schmeidler’s general setting and specially in games where payoffs depend on an “average” of the actions taken by all the players (Rath, 1992). In this setting, the set of players I is the unit interval in R endowed with the Lebesgue measure. We assume that all the players have the same set of available actions, denoted A, which is a compact subset of Rn. An action profile is a measurable function from I to A and the set of action profiles is denoted by AI . Our aim is to capture the relevant features of a class of models where there is a set S ⊆ RK and a variable s ∈ S that represent, respectively, the set of states and the state of an economic system. Payoff for each player depends on his action and the value of this state. To obtain the value of the state we have an aggregation operator Ag : AI → S that associates to each action profile a, the state s = Ag(a). Players are negligible in the sense that they are unable to affect the state of the system unilaterally. For each agent i, there is a payoff function defined on the product of the set of available actions A and S, u(i, · , · ): A × S → R. For simplicity, we take the aggregation operator mentioned above as the integral of the strategy profile with respect to the Lebesgue measure: Z Ag(a) ≡ a(i) di I so that AI , the set of measurable functions from I to A, is contained in dom Ag, the set of integrable functions from I to Rn, and the set of states of the game is S ≡ co {A}.7 8 We denote the set of real valued bounded continuous functions defined on a space X by Cb(X). Let UA×S denote the topological space of Cb(A × S) endowed with the sup norm topology. Throughout the document we will be using the following assumption:
HM The function u : I → UA×S that associates players with their payoff functions is measurable (Rath, 1992). This formulation leaves us in Rath’s extension of Schmeidler’s formulation of games with a continuum of players. To denote games with a continuum of players that have an aggregate state as above, we will use the notation u. We will denote then equivalently u(i) and u(i, · , · ). The definition of Nash Equilibrium is simply re-stated for a game u as a strategy profile a∗ ∈ AI such that λ-almost-everywhere in I: Z Z u i, a∗(i) , a∗(i) di ≥ u i, y, a∗(i) di ∀ y ∈ S, I I In this framework Rath shows that for every game there exists a pure strategy Nash Equilibrium.
6In Mas-Colell (1984) what matters is not strategy profiles but a distribution on the product set of payoff functions and strategies. 7If the strategy sets differed across agents, strategy profiles could be identified with measurable selections of the set R valued mapping i ⇒ A(i). Hypothesis over the correspondence i ⇒ A(i) that assure that the set S ≡ I A(i) di is well defined can be found in Aumann (1965) or in Chapter 14 of Rockafellar and Wets (1998). 8The aggregation operator can as well be the integral of the strategy profile with respect to any measure that is absolutely continuous with respect to the Lebesgue measure, or the composition of this result with a continuous function. That is, Z Ag(a) ≡ G a(i) dλ¯(i) I where λ¯ is absolutely continuous with respect to the Lebesgue measure and G : co {A} → S is a continuous function; the results in this work could well be extended to this setting. See footnote 20
3 3 Rationalizability
Our approach to Rationalizability follows from Guesnerie (1992) who is in turn inspired by Pearce (1984), where the rationalizable set is obtained as the result of the iterated elimination of unreasonable outcomes. This approach is a natural way to grasp the rationalizability argument. Nevertheless, from the characterizations of rationalizability presented by Bernheim and Pearce in their seminal papers and the literature that follows we require that the rationalizable set must not only be the the result of the iterated elimination process, but it should as well be a fixed point of this process.9 In the light of Bernheim’s Proposition 3.2, care is needed to claim that the limit of the process of iterated elimination is in fact a set that we could call of rationalizable outcomes. In particular, when we pass from the finite to the continuous player sets there are three main issues to take into account. The first one is how to address forecasts. In the finite player case it is direct to use (product) measures as forecasts and take expectation over payoff functions to make decisions. This is not evident in the context of continuum of players. A difficulty relates with the continuity or measurability properties that must be attributed to subjective beliefs, as a function of the player’s name. The second issue stems from the first one and is related to the space in which one should seek the rationalizable set. The set of action profiles may not be appealing in contexts where the set of players is a continuum. Finally, the process of iterated elimination of outcomes itself could well be undefined without proper assumptions. Consequently it is necessary to provide conditions . Following Pearce (1984) and Guesnerie (1992) then, consider the following line of reasoning. Given the strategy profile set AI , all players know that each player will only play an action that is a best response to some forecast over strategy profiles. Actual strategy profiles then must be selections of a best response correspondence constructed, for each player, as the image through their best response mapping of the complete set of forecasts over AI . This will give a new (hopefully strict) subset of AI over which it is possible to apply the same reasoning. This process may continue ad-infinitum. In the present framework it is natural to model agents as having forecasts on the set of states rather than on the set of action profiles, since the relevant information that agents need to take a decision is the value of the state s.10 Our first example, taken from Guesnerie’s (1992) paper, gives clear insight on how to treat rationalizability in the context of a continuum of players. For more details, the reader is referred to the original paper.
Example 1. Consider a group of farmers, represented by the [0, 1] ≡ I interval, that participate in a market in which production decisions are taken one period before production is sold. Each farmer i ∈ I has a cost function ci : R+ → R. The price p at which the good is sold is obtained from the (given) R inverse demand function P : R+ → [0, pmax] evaluated in total aggregate production p = P q(i) di where q(i) is farmer i’s production. Since an individual change in production does not change the value of the price, the product is sold under price-taking behavior, so each farmer i ∈ I maximizes his payoff function u(i, · , · ): R+ × [0, pmax] → R defined by u(i, q(i) , p) ≡ pq(i) − ci(q(i)). An equilibrium of this system is a price p∗ such that p∗ = P R q∗(i) di and u(i, q∗(i) , p∗) ≥ u(i, q, p∗) ∀ q ∈ R, ∀ i ∈ I. At the moment of taking the production decision, farmers do not actually know the value of the price at which their production will be sold. Consequently they have to rely on forecasts of the price or of the production decision of the other farmers. The concept in scrutiny in our work is related to how this (these) forecast(s) is (are) generated. Forecasts of farmers should be rational in the sense that no unreasonable price should be given positive probability of being achieved. It is in this setting that Guesnerie introduces the concept of strong rationality or eductive stability 11 as the uniqueness of rationalizable prices which are obtained from the elimination of the unreasonable forecasts of possible outcomes. To obtain these rationalizable prices, Guesnerie describes, in what he calls the eductive procedure, how the unreasonable prices can be eliminated using an iterative process of elimination of non-best-response actions. From the farmers problem we can obtain for each farmer his supply function s(i, · ) : [0, pmax] → R+. The structure of the payoff function implies that for a given forecast µ of a farmer i over the value of the price, his optimal production is obtained evaluating his supply function in the expectation
9This pertains to some type of best response property that the rationalizable set must satisfy. 10See as well Guesnerie (2002) for a discussion on this matter. 11An equilibrium of an economic system is said to be strongly rational or eductively stable if it is the only Rationalizable outcome of the system. We will refer equivalently to outcomes as being strongly rational or eductively stable.
4 under µ of the price, Eµ [p]: s(i, Eµ [p]). Farmers know that a price higher than pmax gives no demand and so prices higher than those are unreasonable. Since all farmers can obtain this conclusion, all farmers know that the other farmers should not have forecasts that give positive weight to prices that are greater than pmax. The expectation of each of the farmers’ forecasts then cannot be greater than pmax and so under necessary measurability hypothesis we can claim that aggregate supply can not R 1 be greater than S(pmax) = 0 s(i, pmax) di. Since all farmers know that aggregate supply can not be greater than S(pmax), they know then that the price, obtained through the inverse demand function, 1 can not be smaller than pmin = P (S(pmax)). All farmers know then that forecasts are constrained by 1 1 the interval pmin, pmax . They have discarded all the prices above pmax and below pmin. This same reasoning can be made now starting from this new interval.
q 6
@ P −1(p) @ @ © S(p) @ ? @ !! @ ! !! @ !! ∗ @ ! q !! !! @ ! @ !! !! @ @ @ @@ - p1 p∗ pmax min 6 p 2 ? pmax
Figure 1: The eductive process
In Figure 1 we can see the aggregate supply function depicted along with the demand function. We have seen that to eliminate unreasonable prices in this model we only need these two functions. The process, as described in the Figure, continues until the farmers eliminate all the prices except the unique equilibrium price p∗. We say then that this price is (globally) eductively stable. Note that the eductive process could give more than only the equilibrium point. This could happen for instance if −1 S(pmax) ≥ P (0). In this situation the rationalizable set would be the whole interval [0, pmax], since farmers would not be able to eliminate prices belonging to this interval.
The particular structure of Example 1 allows us to look at outcomes on the set of prices (or aggregate production), instead of the set of action profiles (production profiles). The eductive procedure consists in eliminating the prices that do not emerge as a consequence of farmers taking productions decisions that are best responses to the remaining action profiles, or equivalently remaining values of aggregate production or prices, and delivers the rationalizable set of prices.12 The intuition described above will allow us to rigorously formulate the processes of elimination of states of the game that are not consequence of best response to any possible (subjective probability) forecast profile over a given set of states. Following the terminology of Bernheim we will make a difference between Rationalizability and Point-Rationalizability. Rationalizability differs from Point-Rationalizability in that forecasts are
12A second characteristic of this setting is that the eductive procedure can be done by simply eliminating prices that are beyond the upper and lower bounds that are obtained in each iteration. However, this comes from the monotonicity properties of the aggregation operator (the integral) and the supply function of the farmers. It is not always the case that the eductive process works this way. This second feature of Example 1 is related to the ordered structure of the games studied in Milgrom and Roberts (1990). We give a result regarding these games in section 4 (see as well Guesnerie and Jara-Moroni, 2007).
5 probability distributions whose supports are contained in the sets of outcomes; while for Point- Rationalizability, forecasts are points or Dirac probability measures on these sets. We will continue now by giving formal definitions for Point and Standard Rationalizability for games with a continuum of players and an aggregate state.
3.1 Point-Rationalizable States
Consider the mapping B(i, · ): S ⇒ A that gives, for each player, the optimal action set given a state of the system,
B(i, s) := argmax {u(i, y, s): y ∈ A} .
A feature that distinguishes this approach from a more general assessment of the subject as that the mapping B(i, · ) goes from S ⊂ Rn, instead of AI for instance, to A ⊂ Rn. It is direct to see, however, that for a given action profile a, in the context of a game u, the best response of a player i to a is given by B i, R a. For each i ∈ I and a set X ⊆ S, consider the image through B(i, · ) of the set X [ B(i, X) := B(i, s) . s∈X Given X ⊆ S consider the set of all the measurable selections taken from the correspondence i ⇒ B(i, X) that associates the set B(i, X) to each player i ∈ I. Then, take all the possible images through the aggregation mapping of such functions. Denoting by P(X) the family of subsets of X, we define the mapping P r : P(S) → P(S) that - to each set X ⊆ S - associates the set P r(X) ⊆ S defined by:
a is a measurable selection of the P r(X) := s ∈ S : s = Ag(a), . (3.1) correspondence i ⇒ B(i, X) Our assumptions on the aggregation operator Ag allow us to re-write definition (3.1) as the integral of a set valued mapping:13 Z P r(X) ≡ B(i, X) di. I The set P r(X) gives the set of states that are obtained as consequence of optimal behavior when forecasts are: heterogeneous, constrained to the set X and in the form of points in this set. Before continuing, we state a relevant property associated to the mapping B.
Lemma 3.1. In a game u, for a non-empty closed set X ⊆ S the correspondence i ⇒ B(i, X) is measurable and has non-empty compact values. The proof is relegated to the Appendix. Lemma 3.1 above and Theorem 2 in Aumann (1965) assure that P r(X) is non empty and closed whenever X is non empty and closed. This set to set mapping allows us to rigorously define a process of iterative elimination of states and a set of point rationalizable states. Consider the process given by iterations of P r. That is,
P r0(S) := S and P rt+1(S) := P r P rt(S) for t ≥ 1.
As noted above, the set of Point-Rationalizable States, noted PS, must then satisfy: ∞ \ t PS ⊆ P r (S) . (3.2) t=0 13 Rn The integral of a correspondence F : I ⇒ is calculated, following Aumann (1965), as the set of integrals of all the integrable selections of F . This is, Z Z F (i) di ≡ f(i) di : f is an integrable selection of F I I R R R where fdi := f1(i) di, . . . , fn(i) di .
6 The right hand side of (3.2) represents the iterative elimination of non reachable states. At each step of this process, players only keep in mind the states that could be reached following rational actions based on point expectations given by the set of the previous step. If a state is not reached by actions following forecasts constrained at a certain step of the process, then it is not rationalizable. Since t +∞ Rn the family of sets P r (S) t=0 is a nested (decreasing) family of closed subsets of , the infinite intersection in expression (3.2) turns out to be the exact Painlev´e-Kuratowski limit of the sequence of sets. The second condition that the set of Point-Rationalizable States must satisfy is:
PS ⊆ P r(PS) . (3.3) T∞ t Note that if a set satisfies condition (3.3), then it is a subset of the set t=0 P r (S). We define the set of Point-Rationalizable States, as follows:
Definition 3.2. The set of Point-Rationalizable States, PS, is the maximal subset X ⊆ S that satisfies condition (3.3).
3.2 Rationalizable States
Before we enter into context we need to introduce some concepts and some notation. For a Borel set X ⊆ Rn we denote by P(X) the set of probability measures on the Borel subsets of X, B(X). Equivalently this is the set of probability measures on the Borel sets of Rn whose support is in X. ∗ We will endow the set P(X) with the weak* topology w = σ(P(X) ,Cb(X)). For a Borel subset Z of X in Rn, P(Z) can be considered to be a subset of P(X) and the weak* topology on P(Z) is the relativization of the weak* topology on P(X) to P(Z). The set X can be topologically identified with a subspace of P(X) by the embedding x 7→ δx. An important property that we will use is that X is compact if and only if P(X) is compact (and metrizable, since we use the norm in Rn).14 As we said before, in the setting of Rath there is a simple way to get through the inconvenience of defining an induced probability measure over the set of action profiles, using the presence of the state variable of the game over which players have an infinitesimal influence. We define then for each player the set valued mapping that gives the actions that maximize expected utility given a probability measure µ over the set of states S, µ ∈ P(S), B(i, · ): P(S) ⇒ A: B E (i, µ) : = argmaxy∈A µ [u(i, y, s)] Z ≡ argmaxy∈A u(i, y, s) dµ(s) . S The process of elimination of unreasonable states, taking into account that players could now use probability forecasts is described with the mapping R : B(S) → P(S):
Z a ∈ AI , a is a measurable selection R(X) := a(i) di : B I of i ⇒ (i, P(X)) Z (3.4) ≡ B(i, P(X)) di, I where B(i, P(X)) is the image through B(i, · ) of the set P(X). The set R(X) gives the set of states that are obtained as consequence of optimal behavior when forecasts are: heterogeneous, constrained to the set X and in the form of probability measures whose supports are contained in this set. As we stated earlier, the difference between P r and R is that the second process considers expected utility maximizers. For a given Borel set X ⊆ S, X can be embedded into P(X). This means that B(i, X) ⊆ B(i, P(X)) and so we have that P r(X) ⊆ R(X). Proposition 3.3. In a game u, if X ⊆ S is nonempty and closed, then R(X) is nonempty, convex and closed. For the proof we will use the following Lemma (see proof in appendix).
14See Aliprantis and Border (1999) for detailed treatments of these and other results.
7 n Lemma 3.4. Let Y and X be compact subsets of R . Given a function u ∈ Cb(Y × X), the function U : Y × P(X) → R, which to each (y, µ) ∈ Y × P(X) associates the expectation Z U(y, µ) ≡ u(y, x) dµ(x) , X is continuous when P(X) is endowed with the weak* topology. Proof of Proposition 3.3. n From our assumptions, u(i) belongs to Cb(A × S), and A and S are compact sets in R , so Lemma 3.4, along with Berges’ Theorem, imply that for each i ∈ I the correspondence B(i, · ): P(S) ⇒ A is upper semi continuous. Since X is closed, it is compact, therefore P(X) is a compact subset of P(S) and so the set B(i, P(X)) is closed for every i ∈ I. From Theorem 4 in Aumann (1965) we get that R B I (i, P(X)) di is closed. On the other hand, Lemma 3.1 states that the correspondence i ⇒ B(i, X) is measurable, which means that it has a measurable selection a. Since B(i, X) ⊆ B(i, P(X)), a is also a selection of B R B i ⇒ (i, P(X)). This implies that I (i, P(X)) di is nonempty. Convexity comes from the fact that R(X) is obtained as an integral of a set valued mapping.
As in the previous case, the set of Rationalizable States RS must satisfy
RS ⊆ R(RS) (3.5) and so we give an analogous definition. Definition 3.5. The set of Rationalizable States is the maximal subset X ⊆ S that satisfies: X ⊆ R(X) and we note it RS. Remark 3.6. Conditions (3.3) and (3.5) are related to the definition of Tight Sets Closed Under Rational Behavior (Tight CURB Sets) given in Basu and Weibull (1991). Indeed Basu and Weibull make the observation that the set of rationalizable strategy profiles in a finite game with compact strategy sets and continuous payoff functions, is in fact the maximal tight curb set, which is analogous to our definitions of (Point-)Rationalizability. Remark 3.7. At this point, it is appropriate to make the connection between the states set approach and the strategic approach and see that the latter provides a fairly convincing view of standard Ratio- nalizability in our context. Let us note that players do not care to differentiate between action profiles that give the same aggregate state. In mathematical terms, action profiles that have the same inte- gral. Looking at the integral as a function from AI to S, we see that the measurable space structure considered in S may induce a measurable space structure on AI , AI , AI . The σ-field AI , is then the “integral induced” σ-field.15 Any probability measure, ν, defined on AI induces a probability measure in P(S) as usual: µ( · ) ≡ ν Ag−1( · ) . Thus, when applying the process defined in (3.4), we are in fact considering at least all probability measures that can be defined over AI . Reciprocally, for a given probability measure in P(S) the I 16 function ν : A → R+ defined by: ν(H) := µ(Ag(H)) is in fact a probability measure on AI . The question rises on whether a richer σ-field would make a difference in terms of the Rationalizable sets. The answer is given by the structure of the game. Payoff functions are defined on the values of the integrals of the action profiles. Consequently, when taking expectation, agents precisely only care about the probabilities of events of the form a ∈ AI : R a = s , for each s ∈ S which brings us back to the induced σ-field AI . We may point out, however, in light of this remark, that the relevant strategic concept associated to our approach would be that of Correlated Rationalizable Strategies (Brandenburger and Dekel, 1987).
15The σ-field of the pre-images of the Borel subsets of S through the integral. AI ≡ Ag−1(X): X ∈ B(S) .
16By construction, if H ∈ AI then Ag(H) ∈ B(S).
8 3.3 Characterization
We now give an answer to the question of whether we can obtain, in our context, the same conclusion of Bernheim’s Propositions 3.1 and 3.2 for (Point-)Rationalizable Strategies in the case of finite player games with compact strategy sets and continuous utility functions. Our first main result, Theorem 3.8, states that when passing to the continuous the set of Point- Rationalizable States can actually be obtained from the eductive process providing an appealing char- acterization of this set.
0 T∞ t Theorem 3.8. Let us write PS := t=0 P r (S). The set of Point-Rationalizable States of a game u can be calculated as
0 PS ≡ PS Proof. We will show that:
0 0 P r(PS) ≡ PS
0 0 0 Let us begin by showing that P r(PS) ⊆ PS. Indeed, if s ∈ P r(PS) then, by the definition of P r, 0 R 0 t there exists a measurable selection a : I → A of i ⇒ B(i, PS), such that s = a. Since PS ⊆ P r (S) 0 t t ∀ t ≥ 0, we have that B(i, PS) ⊆ B i, P r (S) ∀ t ≥ 0 ∀ i ∈ I. So a is a selection of i ⇒ B i, P r (S) t+1 0 and then s ∈ P r (S) ∀ t ≥ 0, which means that s ∈ PS. 0 0 t Now we show that PS ⊆ P r(PS). For this inclusion, consider the following sequence F : I ⇒ A, t ≥ 0, of set valued mappings:
F 0(i) := A ∀ i ∈ I ∀ i ∈ IF t(i) := B i, P rt−1(S) t ≥ 1
We have that Z P rt(S) ≡ F t(i) di. I Since u(i) ∈ UA×S, then ∀ i ∈ I the mapping B(i, · ): S ⇒ A is u.s.c. and, as a consequence, the R 0 set B(i, X) is compact for any compact subset X ⊆ S (Berge, 1997). Since S ≡ I F , from Aumann (1965) we get that S is non empty and compact.17 From Lemma 3.1 we get that F 1 is measurable and compact valued and by induction over t, we get that for all t ≥ 1, P rt−1(S) is non empty, convex and compact, and F t is measurable and compact valued. Consider then the set valued mapping F : I ⇒ A defined as the point-wise lim sup of the sequence t t F , noted p-lim supt F , obtained as: t t F (i) := p-lim supt F (i) ≡ lim sup F (i) t
t t t where the right hand side is the set of all cluster points of sequences {y }t∈N such that y ∈ F (i). From Rockafellar and Wets18 we get that F is measurable and compact valued. 0 R t So now let us take a point s ∈ PS. That is, s ∈ I F for all t ≥ 0. This gives a sequence of t R t measurable selections {a }t∈N, such that s = a . From the Lemma proved in Aumann (1976) we get R t that s ∈ I F , since for each i ∈ I the cluster points of {a (i)}t∈N belong to F (i) and a is the trivial limit of the constant sequence R at. 0 Now it suffices to check that F (i) ⊆ B(i, PS), since then we would have Z Z 0 0 s ∈ F di ⊆ B(i, PS) di ≡ P r(PS) . I I This comes from the upper semi continuity of B(i, · ). Take y ∈ F (i). From the definition of F , y is t t t a cluster point of a sequence {y }t∈N such that y ∈ F (i). That is, there is a sequence of elements
17We have already noted that in fact S ≡ co {A} and so in particular it is also convex, which is of no relevance in this proof, but is the key property in Corollary 3.9. 18See Rockafellar and Wets (1998) Ch. 4 and 5 and Theorem 14.20 .
9 k k tk−1 tk k tk of S, s k∈N such that s ∈ P r (S), y ∈ B i, s and limk y = y. Through a subsequence of k tk 0 k s k∈N we get that the limit of {y }k∈N must belong to B(i, PS), since all cluster points of s k∈N 0 are in PS, being the intersection of a nested family of compact sets. The main challenge in the proof of Theorem 3.8 is to identify the adequate convergence concept for the eductive process. In the case of finite player games there is no such a question, since in that case the technique is simply to take a convergent subsequence of points (in the finite dimensional strategy profile set) from the sequence of sets that participate in the eductive procedure and using (semi) continuity arguments of the best response mappings obtain the result. 19 However, in our setting these arguments fail to prevail. From the proof of the Theorem, we see that the set of Point-Rationalizable States is obtained as the integral of the point-wise upper limit of a sequence of set valued mappings. So the relevant improvement in the proof (besides measurability requirements) is to give the adequate limit concept. From Theorem 3.8 we get as a Corollary that in the games that we are considering, the set of Point-Rationalizable States is “well behaved”. 20 Corollary 3.9. The set of Point-Rationalizable States of a game u is well defined, non-empty, convex and compact. Proof. From Theorem 3.8, PS is the intersection of a nested family of non-empty compact convex sets. From Rath’s existence Theorem we get that there is point s∗ ∈ S such that s∗ ∈ P rt(S) ∀ t. These two facts lead us to conclude that this intersection is compact, convex and non empty. Regarding Rationalizability, Proposition 3.3 allows us to define the Eductive Process. As before, we consider the iterative elimination of non generated states, but now allowing for probability forecasts of players. The iterative process begins with the whole set of outcomes R0(S) := S. Then, on each iteration, the states that are not reached by the process R are eliminated
Rt+1(S) := R Rt(S) .
Recall that since we start the process at S, what we get is a nested family of sets that, following Proposition 3.3, are nonempty convex and closed. The Eductive Process delivers then the set,
∞ 0 \ t RS := R (S) . t=0
0 Theorem 3.10. In a game u, the set RS is non empty, convex and compact. Proof. 0 RS is the intersections of closed convex sets, so it is convex and closed. Again, Rath’s existence 0 t Theorem assures that RS is nonempty, since equilibria belong to every set R (S). 0 From the definition of the set of Rationalizable States, we have RS ⊆ RS. The following Theorem, analogous to Theorem 3.8, shows that these two sets are actually the same. Theorem 3.11. The set of Rationalizable States of a game u can be calculated as
0 RS ≡ RS Although there are certain details where attention must be put, the proof of Theorem 3.11 follows the proof of Theorem 3.8 and is therefore relegated to the appendix.
19See the proof of Proposition 3.1 in Bernheim (1984). 20Theorem 3.8 holds as well for an aggregation operator defined using a function G as in Footnote 8 and if G is affine, Corollary 3.9 holds as well.
10 An important property stated in Corollary 3.9 and Theorem 3.10 is the convexity of the Point- Rationalizable and Rationalizable States sets. It is crucial to obtain reduction procedures from Ra- tionalizable to Point-Rationalizable Sets (see Section 4) and more simple iterative procedures as well (Guesnerie and Jara-Moroni, 2007). Moreover, we know that the convex hull of the set of equilibria is also contained in the set of Point-Rationalizable States and this inclusion may be strict. To see that the Theorems are not vacuous consider the following example.
Example 2. Consider the game where I ≡ [0, 1], A ≡ [0, 1] and u(i) ≡ u : [0, 1]2 → R for all i ∈ I, such that it generates the following best reply correspondence, depicted in Figure 2a:
s∗ if s ≤ s¯, B(s) = {0, s¯(1 − α) + sα} if s > s¯, where s∗,s ¯ and α are in ]0, 1[ and s∗ < s¯. It is clear that this game does not satisfy the hypothesis of Theorem 3.8 since no continuous utility function may give rise to such a best response correspondence.
6 6 6 1 1
s¯(1 − α) + α ¨ s1 ¨ .. ¨ s2 ¨ ...... ¨ s ¨ ... ¨ 3 ... s¯ ¨ ...... b b ... S ...... B(S) ...... ©.. ∗ 0 .. s .. PS ...... r r r ...... - ..-. ∗ 0 ss ¯ 1 z ∗ }| { s¯ 10 b {s } ≡ PS rr (a) The best reply correspondence (b) The elimination of non-generated-states with B : [0, 1] ⇒ [0, 1]. P r(X) ≡ co {B(X)} and the Point-Rationalizable State set.
0 Figure 2: Example 2: The set of Point-Rationalizable States is not the set PS.
The only equilibrium of the game is s∗. Since all the players have the same best reply correspondence, the process of elimination of non- generated-states is obtained by: P r(X) ≡ co {B(X)} The image through the best reply correspondence of the state set is
B(S) ≡ {0, s∗} ∪ ]¯s, s¯(1 − α) + α] , then P r(S) ≡ co {B(S)} ≡ [0, s¯(1 − α) + α]. We see from the form of the best reply correspondence that on each iteration of P r we get an +∞ interval of the form [0, st] where the sequence {st}t=0 satisfies:
st+1 =s ¯(1 − α) + stα, with s0 = 1, which gives for t ≥ 1,