An infinitary probability logic for type spaces (Q1932650)

This paper provides a formal wrap-up to a line of research started by \textit{John C. Harsanyi} in [Manage. Sci., Theory 14, 159--182 (1967; Zbl 0207.51102); ibid. 14, 320--334 (1968; Zbl 0177.48402); ibid. 14, 486--502 (1968; Zbl 0177.48501)]. While originally intended as a Bayesian approach to incomplete information games this research program evolved towards the type space branch of epistemic game theory. Briefly, it starts with an underlying space of states of nature \(S\) (i.e. of uncertain possibilities) and identifies each type of a player \(i\), \(t_i \in T_i\), with a probability distribution over \(S \times T_{-i}\). A series of papers [\textit{W. Armbruster} and \textit{W. Böge}, in: Game theory and related topics, Proc. Semin., Bonn/Hagen 1978, 17--28 (1979; Zbl 0428.90090); \textit{J. F. Mertens} and \textit{S. Zamir}, Int. J. Game Theory 14, 1--29 (1985; Zbl 0567.90103); \textit{A. Brandenburger} and \textit{E. Dekel}, J. Econ. Theory 59, No. 1, 189--198 (1993; Zbl 0773.90109); \textit{A. Heifetz}, Int. J. Game Theory 21, No. 4, 329--338 (1993; Zbl 0794.90009); \textit{P. Battigalli} and \textit{M. Siniscalchi}, J. Econ. Theory 88, No. 1, 188--230 (1999; Zbl 0972.91020); \textit{M. Pinter}, Econ. Theory 26, No. 1, 129--139 (2005; Zbl 1109.91315)] (among others) developed an ``explicit'' representation of types as hierarchies of beliefs (i.e. probability distributions) over beliefs over beliefs, \dots etc., starting with a set \(S\) with different properties, topological or measure-theoretic, ensuring that each of these hierarchies is identified with a single probability distribution as intended by Harsanyi. Since in applications only a few types are needed, the question arose about the existence of a universal type space that includes all the possible types over \(S\). The mathematical approach to this problem involves finding a terminal object in the category of type spaces. The existence of this object has been ensured in a variety of settings [Mertens and Zamir, loc. cit.; \textit{A. Heifetz} and \textit{D. Samet}, J. Econ. Theory 82, No. 2, 324--341 (1998; Zbl 0921.90156); \textit{L. Moss} and \textit{I. Viglizzo}, Electron. Notes Theor. Comput. Sci. 106, 279--295 (2004; Zbl 1271.91030)] etc. A related concern involves the existence, given a type space, of all possible types definable in it. This is known as the problem of belief-completeness [\textit{A. Brandenburger}, ``On the existence of a `complete' possibility structure'', in: Cognitive processes and economic behavior, Routledge Siena Stud. Polit. Econ. London: Routledge. 30--34 (2003)]. For both problems, most of the spaces obtained in the ``explicit'' analysis are universal and belief-complete. The question addressed in this paper is whether there exists a sound and complete axiomatization of the class of type spaces with respect to the class of type spaces with those features. It circumvents the impossibility of giving a finitary axiomatization by adding a class of formulas of the form \(p_{i}^{\alpha}(\phi)\), where \(\phi\) represents an event \(\lceil \phi \rceil\) in the class of states of the world (i.e. \(S \times T\)). It means that ``\(i\) assigns probability at least \(\alpha\) to \(\lceil \phi \rceil\)''. Furthermore, infinite conjunctions of those formulas are allowed, yielding a language \(\mathcal{L}\). The ensuing system of axioms gives rise to a logical system \(P\) by means of some inference rules. \(P\) is (strongly) sound and complete with respect to any class of formulas \(\Gamma\) of the language \(\mathcal{L}\) (i.e. for every \(\phi \in \mathcal{L}\), \(\Gamma \vdash \phi\) iff \(\Gamma \models \phi\)). In turn this leads to a characterization of a canonical model, in which each state of the world is a maximal consistent set of formulas of \(\mathcal{L}\). This canonical model is proven to be universal and belief-complete.