**1. Introduction**

Self-enforcing agreements—those for that no party has any incentive to break given that all others comply—should be carried out even if they are not binding in a formal sense. This is in fact the defining characteristic of the standard Nash equilibrium concept, and thus, one of the common justifications for this concept is that if players are allowed to communicate before playing a game, they could hardly reasonably agree on an outcome not satisfying this criterion. Recall that a Nash equilibrium constitutes for each player a set of strategies and beliefs (about other players' strategies), such that the strategies are the best responses to beliefs and the beliefs are correct (see e.g., Osborne 2004 [1]). We assume that there is no recourse to court-enforceable contracting, or equivalently that any such interactions have already taken place. Unfortunately, while intuitively pleasing, this justification for the use of a Nash equilibrium has been characterized by a shortage of formal models.

On a related, but distinctly different track of reasoning, it is natural to wonder why agents would ever agree on an inefficient outcome, assuming that they had the chance to talk in the first place. In other words, why would players agree ahead of time to an inefficient outcome of a game if there were another potential outcome, also an equilibrium, that gave strictly greater payoffs to all of them? Once again, the challenge has lied in constructing a realistic, but necessarily simplified, formal model of the agents' communication process. Among other problems, this inefficient result appears to be incompatible with the arguments outlined above, in which Nash equilibria in general are justified.

This type of preplay communication is often called *cheap talk* , which may be roughly defined as nonbinding, nonpayoff relevant, preplay communication. Although cheap talk has indeed received attention as a potential solution to these questions surrounding the equilibrium concept, in practice, it has been mostly used in the study of signaling games, in repeated environments (often in connection with learning), and in certain applied settings. These are of course all important applications, but these leave the original ambiguities unresolved. This paper, then, returns to the goal of constructing a

more comprehensive model of pure cheap talk and explores its relationship with equilibria and equilibrium selection.

This paper develops a model of cheap talk that involves an unlimited communication session, called a *conversation* , before the play of a standard game begins. Players announce in advance what actions they plan to take in the upcoming game, and taken together, these announcements form one possible prediction of what they may actually do. On the other hand, there is also a common prior forecast, given exogenously, of what each player will do; this forecast is updated as the conversation proceeds. An announcement is defined to be credible only if it is close to the best response to one or the other of these two predictions about the rest of the players. Otherwise, an announcement has no external justification, so it is deemed unbelievable and disregarded. The conversation continues indefinitely in this manner, possibly, but not inexorably toward some limit. Realistically, it will rarely if ever go on for very long (although, for complicated games, it may take some time), since if it is going to converge, it will do so rapidly. However, it is important not to have an artificial limit imposed externally, just as long finitely repeated games behave very differently than infinitely repeated games.

The paper's first main result is that if the conversation converges toward a limit, then this limit must be a Nash equilibrium of the underlying action game in which payoffs are determined.1 Conversely, any Nash Equilibrium forms a possible limit of the conversation. This result can be interpreted as saying that meaningful communication before a game can only lead to Nash outcomes. Since the cheap talk is the initial interaction between the players, we assume that they cannot be sure of the strategies that their opponents will follow in the communication stage. Any strategy in this phase that is weakly dominated by another is clearly not optimal; anything else is potentially the preferred choice and is therefore, given the lack of information, one possible optimal choice.2 The paper's second result then states that optimal pregame play in the conversation stage leads to an efficient outcome and that any efficient final outcome is a possible result of such a strategic conversation.3 This can be interpreted as saying that rational, or thoughtful, speech leads to efficiency. This completes the connection between cheap talk (as modeled here, i.e., in an environment where rationality and utilities are common knowledge), Nash equilibria, and Pareto optimality. The first result applies to all games (at least those with a Nash equilibrium), while the second result only has bite in games with multiple equilibria.

The conclusion derived from the second main result contrasts with previous "babbling" results, in which it is impossible to select among the set of Nash equilibria because all pregame communication is ignored. The main reason for the difference is that those previous studies looked for equilibria of the extended communication game as a whole—for instance, by assuming that the full strategies of all players are known. This allows equilibrium strategies in which no value is placed even on seemingly mutually informative communication, whereas the model below presupposes the impossibility of ignoring beneficial interchange. Thus, the present paper takes a more primitive view of pregame strategies, especially since in part it is attempting to justify the equilibrium concept in the first place. Naturally, although the model does not impose beliefs about the cheap talk stage, it still must make some assumption about beliefs held upon entering the action game. Another approach that will destroy the babbling equilibria is to assume an arbitrarily small, but positive cost to sending messages—this is a restriction on the environment rather than on the structure of equilibrium or on belief formation. While this limitation is plausible in reality, it is, strictly speaking, no longer a model of cheap talk, even if the total sum spent on sending messages is always lower than the game's smallest payoff differential.

The paper proceeds as follows. Section 2 provides a brief survey of some of the relevant literature. In Section 3, some motivation is given for the specific assumptions made in this conversational model

<sup>1</sup> The limit is an *ε*-Nash equilibrium.

<sup>2</sup> This is discussed in further detail in Section 3.

<sup>3</sup> The notion of efficiency used here is *stable efficiency*, a concept that is equivalent to Pareto efficiency in generic two person games.

of cheap talk. Section 4 lays out the formal model, stating and proving the paper's two main results. Several examples are detailed in Section 5 in order to illustrate both the cheap talk process and the implications of the theorems. Finally, Section 6 concludes the paper by summarizing the model and discussing some possible extensions of its implications.
