Patterns, Types, and Bayesian Learning

Bayesian Statisticians, decision theorists, and game theorists often use Bayesian representations to describe the probability distribution governing the evolution of a stochastic process. Generally, however, one given distribution has infinitely many different Bayesian representations. This paper identifies natural, endogenous representations whose component distributions are learnable and follow patterns. Any given distribution that satisfies an asymptotic mixing condition has a unique, up to an equivalence class, natural Bayesian representation which can be obtained by conditioning on the tail-field of the process. This result follows a parallel to de Finetti's theorem, but with exchangeability weakened to asymptotic mixing which admits many more applications.