Model theory of probability spaces

TL;DR

This paper develops the model theory of probability spaces using continuous logic, characterizing atomless probability algebras, and exploring their properties such as stability, quantifier elimination, and independence.

Contribution

It introduces the theory $Pr$ of probability algebras, identifies the model companion $APA$ for atomless cases, and analyzes their model-theoretic properties and connections to probability theory.

Findings

$APA$ is separably categorical and complete.

Probability algebras have a definable set of atoms.

The paper links probabilistic entropy with model-theoretic forking.

Abstract

This expository paper treats the model theory of probability spaces using the framework of continuous $[0,1]$-valued first order logic. The metric structures discussed, which we call probability algebras, are obtained from probability spaces by identifying two measurable sets if they differ by a set of measure zero. The class of probability algebras is axiomatizable in continuous first order logic; we denote its theory by $Pr$. We show that the existentially closed structures in this class are exactly the ones in which the underlying probability space is atomless. This subclass is also axiomatizable; its theory $APA$ is the model companion of $Pr$. We show that $APA$ is separably categorical (hence complete), has quantifier elimination, is $\omega$-stable, and has built-in canonical bases, and we give a natural characterization of its independence relation. For general probability algebras, we prove that the set of atoms (enlarged by adding $0$) is a definable set, uniformly in models of $Pr$. We use this fact as a basis for giving a complete treatment of the model theory of arbitrary probability spaces. The core of this paper is an extensive presentation of the main model theoretic properties of $APA$. We discuss Maharam's structure theorem for probability algebras, and indicate the close connections between the ideas behind it and model theory. We show how probabilistic entropy provides a rank connected to model theoretic forking in probability algebras. In the final section we mention some open problems.