Bayesian and consistent assessments (Q2447155)

The authors characterizes the set of extensive forms (games without a payoff assignment) for which the sets of Bayesian and consistent assessments coincide. The results can help to understand better how consistency brings about restrictions on beliefs and compute sequential equilibria, and derive an analogous characterization of consistency to that offered by \textit{D. Fudenberg} and \textit{J. Tirole} [J. Econ. Theory 53, No. 2, 236--260 (1991; Zbl 0717.90108)] for some multi-period games with observable actions where there are two periods or every player has at most two types. The behavior strategy of a player specifies for every information set a probability distribution. A system of beliefs specifies at each information set a probability distribution over its nodes. Hence an assessment is a behavior strategy profile for all players together with a system of beliefs. The Bayesian assessment is an assessment where behavior strategies are given by Bayesian probability distributions, and the consistent assessment is the limit point of a sequence of Bayesian assessments. \textit{D. M. Kreps} and \textit{R. Wilson} [Econometrica 50, 863--894 (1982; Zbl 0483.90092)] provided a way of selecting beliefs at zero probability information sets to define a sequential equilibrium. Every \textit{J.-F. Mertens} stable set [Math. Oper. Res. 14, No. 4, 575--625 (1989; Zbl 0687.90097)] and every \textit{J. Hillas} stable set [Econometrica 58, No. 6, 1365--1390 (1990; Zbl 0734.90125)] contains a proper equilibrium and therefore a sequential equilibrium. Section 2 sets up basic notations and definitions. With Kreps and Wilson's [loc. cit.] key results, the author characterizes the set of extensive forms where consistent assessments strictly refine the sets of Bayesian assessments in the main Theorem 1 given in Section 3. The non-consistent Bayesian assessments examples follow restrictions that capture the key feature of this extensive form to be Bayesian, and the key feature leads to Definition 3 to identify consistent assessments within Bayesian ones. Section 4 applies the result to Fudenberg and Tirole's multi-period games with observed actions. Corollary 1 shows that every Bayesian assessment is consistent for two-player two-period games with observed actions and for two-player multi-period games with observed actions with proper conditions on completely mixed probability on the finite set of types less than or equal to two. The reason why preconsistent assessments cannot be used to characterize consistency when a player has three types or more is illustrated Figure 1 of [Fudenberg and Tirole, loc. cit.] and Figure 235.1 of [Osborne and Rubinstein, loc. cit.], and preconsistency does not select beliefs at either of those information sets. The last section is an appendix, where all the equations used in the proof of Theorem 1 are derived.