Bayesian Rationality and Social Norms
Herbert Gintis∗
November 13, 2007
1 Introduction
There is a glaring contrast between economics, which models social interaction as a Nash equilibrium of a game played by rational decision-makers (Mas-Colell, Whinston and Green 1995), and sociology, which models social interaction as the role-playing of individuals guided by social norms (Durkheim 1933[1902], Parsons 1967).1 The incentive compatibility requirements of Nash equilibrium are virtually ignored by sociologists, who argue that humans are prosocial by nature, and gener- ally carry out the duties and obligations associated with the social roles they assume rather than maximizing personal gain. The commonality of beliefs stressed by so- ciologists as key to social efficiency have traditionally been ignored by economists, who typically celebrate the irreducible heterogeneity of rational agents (Becker and Stigler 1977). The remarkable body of empirical evidence supporting the sociological model has made little impression on economists, who favor deriving social norms from the interaction of rational actors. The equally remarkable body of evidence supporting the rational actor model is rejected by sociologists on grounds that the standard sociological model incorporates all the strengths of the rational actor model, and in addition can deal with the socially determined nature of beliefs (Boudon 2003). The discrepancy between these two firmly held positions has been a major factor sustaining theoretical disarray in the behavioral sciences (Gintis 2007). Recent findings suggest, however, that economic and sociological theory can be rendered consistent and mutually supportive. First, epistemic game theory has ∗ Santa Fe Institute and Central European University. I am grateful for the comments of Masahiko Aoki, RobertAumann, Adam Brandenburger, Gerry Mackie,Yusuke Narita, Robert Sugden, Giacomo Sillari, Luca Tumolini, and Peter Vanderschraff. I would like to thank the John D. and Catherine T. MacArthur Foundation for financial support. 1A Bayesian rational agent maximizes expected utility, given a subjective prior over the various possible outcomes (Savage 1954, Anscombe and Aumann 1963, Kreps 1988).
1 demonstrated the centrality of common priors and common knowledge in ensuring that rational agents play a Nash equilibrium (Aumann and Brandenburger 1995, Polak 1999). Second, behavioral game theory has demonstrated that normative behavior need not, and usually does not, violate Bayesian rationality (Fehr and Gächter 2002, Andreoni and Miller 2002, Gintis, Bowles, Boyd and Fehr 2005, Gneezy 2005). Third, Aumann (1987) has shown that the natural equilibrium con- dition for game theory is the correlated equilibrium, and in this paper I will show that the most promising candidate for a correlating device—which epistemic game theory does not provide, even in principle—is the social norm. Finally, epistemic game theory has shown how a commonality of beliefs can be maintained by a set of rational actors facing asymmetric information. The need for such commonality was stressed by Durkheim (1933[1902]), who observed that the division of labor in modern society fosters a heterogeneous citizenry and an ethic of individual autonomy, yet social efficiency requires organic solidarity based on collective representations that harmonize the mental constructs of individuals.2 Never answered in the sociological literature is how a common culture can be maintained by rational actors facing asymmetric information. Aumann (1976) an- swered this question by showing that if rational agents have a common prior and if their posteriors following an event are common knowledge, then their posteriors agree as well.3 This analysis shows that collective representations can be main- tained by rational actors, so the ‘suspension of disbelief’ prima facie required by Durkheim’s organic solidarity is in fact unnecessary. If Bayesian rational agents have common prior probabilities or expectations, the receipt of asymmetric in- formation conserves the commonality of posteriors, provided these are common knowledge. However, game theory fails to supply conditions under which pos- teriors will be common knowledge, and supplies no reasons for the existence of common priors. If game theory could explain collective representations (common priors and common knowledge), even in principle, as an outcome of the interaction of rational agents, it would be plausible to view game theory as the unique foundation for all of the behavioral sciences (economics, sociology, anthropology, social psychology, and political science). While the aim of recent game-theoretic research has been exactly this, I will show that this task cannot be achieved, even in principle. The foundations of behavioral science must therefore include analytical constructs in addition to game theory, and the minimum set of human characteristics needed to
2To the extent that there are ethnic, regional, cultural, and social differences within a society, collective representations may differ across groups. Moreover, organic solidarity is obviously an equilibrium conditions that fails in periods of social change. 3Geanakoplos (1992) provides an overview of the literature inspired by Aumann’s original Agree- ment Theorem.
2 model strategic interaction must include more than Bayesian rationality. Game theory would provide a mechanism for the formation of common priors if the Harsanyi doctrine (Harsanyi 1967–68) were correct. The Harsanyi doctrine holds that rational individuals can have divergent beliefs only if they have different information. Asymmetric information, according to Harsanyi, can be modeled by assuming common prior beliefs and formally incorporating the mechanisms of informational asymmetry in the model itself. This argument may be plausible under favorable observational conditions when the events in question are natural occurrences for which causal covering laws and probabilistic frequencies exist. However, it is not plausible when the events involve the subjective priors of other agents (Morris 1995, Gul 1998). Aumann’s (1998) classic response to Gul’s (1998) critique of the common prior assumption is that the common prior assumption, “embodies a reasonable and use- ful approach to interactive decision problems, though by no means the only such approach.” (p. 929) Aumann’s position, then, is that the concordance of beliefs across individuals (common priors) is sufficiently widespread in social groups that it may be taken as a starting-point for analytical purposes, although it does not derive from principles of Bayesian rationality. While this position is pragmatically defensible as a provisional step in the development of social theory, it is clearly not the final word on the topic. Indeed, game theory is rejected by many behavioral scientists who, like Durkheim and Parsons, consider the commonality of beliefs a highly problematic yet key aspect of human sociality, and study the mechanisms that foster such commonality. We say an event E is common knowledge for agents i = 1,...,n if the each agent i knows E, each i knows that each agent j knows E, each agent k knows that each j knows that each i knows E, and so on. So, for instance, we say that common priors are common knowledge if each agent knows that each agents has the same priors, each agent knows that each agent knows that each agent has the same priors, and so on. The importance of common knowledge for interactive epistemology was first stressed by Lewis (1969) and Aumann (1976), its critical role among the conditions for correlated and Nash equilibrium being established in Aumann (1987) and Aumann and Brandenburger (1995). Aumann’s (1976) epistemological analysis of common knowledge derives col- lective representations from axioms of epistemic logic, in the form of a theorem asserting that if E is a public event, meaning that it is self-evident for each of a set of agents, then at every state ω ∈ E, E is necessarily common knowledge (precise definitions are given later). However, as we shall see, if the knowledge system is constructed without specific inter-agent epistemological assumptions, the only public event that can be shown to follow from the tautologies of the system and as- sumptions concerning the distribution of first-order knowledge, is the trivial event
3 , the whole event space. We conclude that social norms and conventions, which do assert the common- ality of mental representations, are emergent properties of human social systems that must be posited along side of Bayesian rationality to explain human strate- gic interaction. The capacity to share beliefs is also an evolved characteristic of the human brain (Premack and Woodruff 1978, Heyes 1998, Tomasello, Carpen- ter, Call, Behne and Moll 2005). This capacity doubtless arose during the gene- culture coevolutionary process that gave rise to our species (Cavalli-Sforza and Feldman 1981, Boyd and Richerson 1985, Durham 1991). Indeed, human culture is only the latest in a series of emergent transitions characteristic of the biological evolution of complexity from the most primitive microbes to the present (Maynard Smith and Szathmary 1997, Morowitz 2002). Affirming the emergent character of social norms means they have the status of fundamental principles that cannot be derived analytically from other funda- mental principles. Affording this status to social norms requires that we rethink certain basic theoretical presuppositions, because it entails denying the principle of methodological individualism, which holds that all social phenomena can be ex- plained purely in terms of the characteristics of the agents involved. Methodological individualism is a deeply held, but unsupported, precept of game theory. Afford- ing explanatory power to social norms, however, will allow us to solve heretofore insoluble problems of epistemic game theory.4 The sociological and economic arguments, properly cast, are thus in fact com- plementary, not alternative, explanations. Sociological models explain why there are common priors, and common knowledge: only arguments taken from psy- chology and sociology can make scientifically grounded assertions as to when the conditions for the transformation of mutual knowledge into common knowledge might obtain. Economics provides the basic model of rational choice, shows how informational heterogeneity can be consonant with common posteriors, and iden- tifies the correlated equilibrium as the natural equilibrium condition for rational actors with common priors. Sociology provides the analysis of correlating devices, in the form of social norms and conventions, and psychology elucidates the brain mechanisms that permit and foster the commonality of beliefs. The framework for integrating social norms and Bayesian rationality proposed in this paper recognizes three inferential stages. First, there are natural occurrences
4Game theorists often take methodological individualism as synonymous with the scientific method, on the grounds that scientific explanation consists precisely in modeling the whole in terms of the interaction of its parts. However, in a complex system, the modeler cannot abandon higher-level constructs. For instance, it is rarely fruitful to analyze the behavior computer software in terms of the interactions of individual transistors, or to model the immune system of an organism in terms of the quantum states of its constituent chemical molecules.
4 N, such as “the ball is yellow,” or “it is raining in Paris” that are immediately perceived by individuals as first-order sensory experience. Under some conditions these natural occurrences are public signals that are mutually accessible to members of a group, meaning that if one member knows N, then he knows that each other member knows N. For instance, if i and j are both looking at the same yellow ball, the ball’s color may be mutually accessible: i knows that j knows that the ball is yellow. Second, there are higher-order socially defined events which we call games G, which specify the type of strategic interaction appropriate to the social situation at hand. Games are not mutually accessible, but social conventions may specify that a mutually accessible event F indicates G. We call F a frame, we write G = φ(F), and we think of the relation “F indicates G to agent i” as asserting that when i knows F , he proceeds through a series of mental steps involving the consideration of known social regularities, such as norms and conventions, at the conclusion of which i knows G (Lewis 1969, Cubitt and Sugden 2003). Assuming F is a public indicator of G, and that the individuals involved are symmetric reasoners (precise definitions are left for later), then G will be common knowledge. Third, given epistemic game G, certain social processes that transform private into public information may justify the assumption of common prior for G, which in turn determines a correlated equilibrium of G. The correlating device is a social norm or convention N = ψ(G), which specifies exactly how the game will be played by rational agents. Of course, in the real world, at any stage there may be irregularities that produce non-equilibrium outcomes.
2 Informational Asymmetry and Common Knowledge
Aumann (1976) proved that if Bayesian rational agents have a common prior con- cerning the expected value of a variable or the probability of an event, the receipt of asymmetric information conserves the commonality of posteriors, provided these are common knowledge. This theorem is usually treated as a curiosum, because there is considerable “agreement to disagree” on asset markets. In fact, it is a very general argument explaining why common knowledge can be updated via asymmet- ric information and yet remain common knowledge among rational social agents. We present an elegant generalization of Aumann’s argument. Consider a finite state space with typical element ω ∈ . Subsets of are called events. Each agent i has a subjective prior pi(·) over . We represent asymmetric information by endowing each agent i with a partition Pi of such that in state ω ∈ P ∈ Pi, i knows only that the true state is in P . This implies that pi(·) must be constant on the elements P ∈ Pi. We call the elements P ∈ P the cells of P. We write the cell of i’s partition containing ω as Piω ∈ Pi, so the conditional probability of an event
5 E ⊆ for i in state ω is pi(E|Piω). We assume pi(Piω) > 0 for all ω ∈ . The common knowledge partition P∗ corresponding to individual partitions P1,...,Pn ω is the finest common coarsening of the individual partitions. That is, the cell P∗ of P∗ containing ω is the union of cells of Pi for each i = 1,...,n. We call this construction in modal logic the semantic model of interactive epistemology.5 P Let f :2 i →. We say a f satisfies the sure thing principle on if for all P,Q ⊆ with P ∩ Q =∅,iff(P) = f (Q) = a, then f(P ∪ Q) = a.For instance, if p is a probability distribution on and E is an event, then the posterior probability f(X)= p(E|X) satisfies the sure thing principle, as does the expected value f(X) = E[x|X] of a random variable x given X ⊆ . We then have the following Agreement Theorem (Collins 1997):
Theorem 1. Suppose for agent i = 1,...,n, fi satisfies the sure thing principle ω on , and suppose it is common knowledge at ω that fi = ai. Then fi(P∗ ) = ai.
ω Proof: Because P∗ is the disjoint union of i’s partition cells Piω , and fi = ai on ω each of these cells, by the sure thing principle, fi = ai on P∗ .
Corollary 1. Suppose agents i = 1,...,nshare a common prior on , indicat- ing an event E has probability p(E). Suppose each agent i now receives private information that the actual state ω is in Piω. Then, if the posterior probabilities ai = p(E|Piω) are common knowledge, then a1 = ...= an.
ω Proof: Let fi = p(·|Piω). Applying Theorem 1, we have ai = p(E|P∗ ) for all i.
3 The Agreement Theorem and the Conditions for Nash Equi- librium
Agreement theorems are at the heart of game theory because the conditions under which rational agents play a Nash equilibrium are precisely those that create con- ditions for an Agreement Theorem (Aumann and Brandenburger 1995). We define an epistemic game G to be a normal form game with players i = 1,...,n and a finite pure strategy set Ai for each player i,soA = A1 × ...× An is the set of pure strategy profiles for G, with payoffs πi :A→R. In addition, G includes a set of possible states of the game, a knowledge structure (Pi, Ki, Pi) defined over for each player i, and a subjective prior pi(·; ω) over the actions A−i for the other players that is a function of the state ω. Because player i cannot distinguish among states in Piω, pi(·; ω) = pi(·; ω ) for all ω ∈ Piω. Finally, we specify that each state ω specifies the players’pure strategy choices a(ω) = (a1(ω),...,αn(ω)) ∈ A
5I have found no uniform terminology for this construction in the literature.
6 ω ω and their beliefs φ1 ,...,φn concerning the choices of the other players. We call ω φi i’s conjecture concerning the behavior of the other players at ω. Player i’s ω = ; conjecture is derived from i’s subjective prior by φi (a−i) pi(a−i Piω). Thus, at state ω ∈ , each player i takes the action ai(ω) ∈ Ai and has the probability ω ω distribution φi over A−i. Note that φi is constant on Piω, so we could write it as Pi ω φi , but we will generally use the more concise notation. A player i is Bayesian [ ω ] rational at ω precisely if ai(ω) maximizes E πi(ai,φi ) , which is defined in the usual way as [ ω ]= ω E πi(ai,φi ) φi (a−i)πi(ai,a−i). (1) a−i ∈Ai
We say that “player i knows the other player’s actions a−i at state ω”ifPiω ⊆ { | = } ω = ∈ ω ω (a−i) a−i , which means φi (a−i(ω )) 0 for all ω / Piω. We assume { | ω = }⊆{ | = } that ω φi (a−i) 1 ω α−i(ω ) a−i ; i.e., if agent i knows something, ω = ω ∈ then what he knows is true. Note that φi φi for all ω Piω, because i cannot distinguish among states is Piω. Aumann and Brandenburger (1995) and Polak (1999) are responsible for the following theorem, in which knowledge is interpreted as belief with probability one, and where “mutually known” means “known by all players”: Theorem 2. Let G be an epistemic game with n>2 players, and let φ = φ1,...,φn be a set of conjectures. Suppose at ω ∈ it is mutually known that the game is G, mutually known that players are rational, and commonly known that φ is the set of conjectures for the game. Then for each j = 1,...,n, all i = j induce the same conjecture σj (ω) about j, (σ1(ω),...,σn(ω)) form a Nash equilibrium of G, and it is commonly known that each player is Bayesian rational. Because the assumptions that the players know the game and are rational are relatively weak, Theorem 2 indicates that common knowledge of conjectures is the key problematic condition we need to conclude that rational agents will implement a Nash equilibrium. The question, then, is under what conditions is common knowl- edge of conjectures likely to be instantiated in real-world strategic interactions?
4 When Mutual Knowledge becomes Common Knowledge
Let P be the knowledge partition of the state space for an agent. The cell of P containing ω is written Pω. We say the agent knows an event E at state ω if Pω ⊆ E. If the agent knows E for all ω ∈ E, we say E is self-evident.6 It is easy to see that a self-evident event is the union of cells of P. If two events are
6This usage conforms to Bacharach (1992). Binmore and Brandenburger (1990) call a self-evident event a “truism,” and refer to an event that is a truism for all agents as a “common truism.” Following
7 self-evident, so is their union. Hence for any event E, it we can define KE as the largest self-evident event contained in E. Clearly, KE ={ω|Pω ⊆ E}. An event E is thus self-evident if and only if KE = E. Suppose we have a set of n of agents, each of whom has a partition Pi of and a corresponding knowledge operator Ki, i = 1,...,n. We say an event E is a public event if E is self-evident for all i = 1,...,n. If two events are self-evident for an agent, so is their intersection. Hence, the intersection of public events is also a public event. Thus, for any ω ∈ , there is a minimal public event P∗ω containing ω. It is easy to check that the events P∗ω are the cells of the common knowledge partition P∗ corresponding to P1,...,Pn. We say an event E is common knowledge at ω if P∗ω ⊆ E. We define K∗E ={ω ∈ E|P∗ω ⊆ E},soK∗ is the knowledge operator corresponding to P∗. It might be thought that an event E being common knowledge implies nothing concerning what agents know about the knowledge of other agents. This, however, is not the case. Indeed, Aumann (1976) showed that an event E is common knowledge precisely when everyone knows E, everyone knows that everyone else knows E, and so on.7 This shows that the self-evidence of an event must incorporate in some deep sense the structure of mutual interrelatedness of beliefs. A weakness of the semantic model is its failure to elucidate this sense. The idea behind this result is illustrated in Figure 1, where E is a public event for two agents, Alice and Bob. Here, PAω ⊆ E, so Alice knows E at ω. For each ω ∈
⊆ ⊆∪ ⊆ PAω, PB ω E, because E is self-evident for Bob. But PAω ω ∈PAωPB ω E, so Bob knows that Alice knows E at ω. In Figure 1, two of Bob’s partition cells cover Alice’s partition cell at ω. A similar argument shows that Alice knows that Bob knows that Alice knows E at ω, and so on. The conclusion that when a public event occurs, each agent knows that the others know that the event occurred, is quite striking, for it asserts that self-evidence, which is ostensibly a characteristic of a single agent, when shared among agents, permits agents to know the content of the minds of other agents. Of course, the term “to know” implies that the object of knowledge is true, and therefore tautologically, if i knows E, then i knows that j = i does not know that ¬E; i.e., by definition, if an agent knows E, he knows that other agents know that E is “possible.” But, this tautological implication cannot be extended to a determination of what other agents
Milgrom (1981), we prefer the term “public event” to the term “common truism.” Aumann (1987) has no term for what we call “self-evident,” and uses the term “self-evident” for what we call “public.” Because there is no accepted usage, I will retain the self-evident/public terminology on grounds that these terms best describe the situation under investigation. 7In fact, Aumann (1987) defines E to be common knowledge if each agent knows E, each knows the others know E, and so on. He then proves that common knowledge is equivalent to KiE = E for each agent i; i.e., using our terminology, E is self-evident for each agent.
8 E