[Paper Review] Characterisation of exchangeable sequences through empirical distributions

Martin Bladt, Shaiderman, Dimitry|arXiv (Cornell University)|Mar 19, 2019

Stochastic processes and financial applicationsEconomics, Econometrics and Finance11 references3 citations

TL;DR

This paper provides the first complete proof of the converse to de Finetti's theorem: an exchangeable sequence is characterized by its empirical distributions forming a reverse-measure-valued martingale. Using a Markov chain approach for the binary case and a discretization-combinatorial argument for the general case, the authors establish that the reverse-martingale property of empirical measures fully characterizes exchangeability, resolving an incomplete proof by Kallenberg (2005).

ABSTRACT

It is a well-known fact that an exchangeable sequence has empirical distributions that form a reverse-martingale. This paper is devoted to proof of the converse statement. As a byproduct of the proof for the binary case, we introduce and discuss the notion of two-coloring exchangeability.

Motivation & Objective

• To establish the converse of de Finetti's theorem: that the reverse-martingale property of empirical distributions implies exchangeability.
• To address the incompleteness in Kallenberg's 2005 proof of this converse statement.
• To introduce and analyze the novel concept of two-coloring exchangeability in the binary case.
• To generalize the result to arbitrary standard Borel spaces using discretization and combinatorial techniques.
• To place the main result in the broader context of random probability measures, martingales, and exchangeability.

Proposed method

• For the binary case, the authors use a Markov chain approach on the partial sum process $Y_n = \sum_{i=1}^n \xi_i$ to analyze exchangeability under the reverse-martingale condition.
• They define a transition kernel based on the conditional distribution of $\xi_n$ given $Y_n = y$, leveraging symmetry and exchangeability to derive the reverse-martingale property.
• The proof introduces the notion of two-coloring exchangeability, which characterizes binary sequences where the empirical measure process forms a reverse-martingale.
• For the general case, the authors employ a discretization argument to approximate the state space $S$ and reduce the problem to finite-dimensional combinatorial structures.
• Combinatorial identities involving permutations and exchangeable sequences are used to show that the reverse-martingale condition forces exchangeability across all finite-dimensional distributions.
• The proof relies on the fact that the conditional expectation $E(\eta_n f \mid \mathcal{T}_{n+1}) = \eta_{n+1} f$ for all bounded measurable $f$ implies symmetry across all indices, hence exchangeability.

Experimental results

Research questions

• RQ1 Is the reverse-measure-valued martingale property of empirical distributions sufficient for a sequence to be exchangeable?
• RQ2 Can the converse of de Finetti's theorem be fully proven using methods distinct from Kallenberg's original (incomplete) approach?
• RQ3 What is the significance of the two-coloring exchangeability concept in the context of binary exchangeable sequences?
• RQ4 How can the reverse-martingale condition on empirical measures be used to reconstruct the exchangeability of a sequence in general state spaces?
• RQ5 What is the relationship between the predictive distributions $p_n$, the empirical measures $\eta_n$, and the exchangeability of the underlying sequence?

Key findings

• The paper provides a complete and rigorous proof that a sequence is exchangeable if and only if its empirical distributions form a reverse-measure-valued martingale, confirming Theorem 1.3.
• For the binary case, the proof establishes that exchangeability is equivalent to the reverse-martingale property of the empirical measure process, and introduces the concept of two-coloring exchangeability as a new characterization.
• The authors demonstrate that the reverse-martingale condition on $\eta_n$ implies that the joint distribution of any finite subsequence is invariant under permutation of indices, thus enforcing exchangeability.
• The general case is proven via a discretization argument that reduces the problem to finite-dimensional combinatorial structures, showing that the reverse-martingale condition forces symmetry across all finite-dimensional marginals.
• The paper identifies and corrects an incompleteness in Kallenberg's 2005 proof of the same result, showing that his original approach cannot be repaired along the same lines.
• The result places the empirical measure process $\eta_n$ in a central role in characterizing exchangeability, distinct from the predictive distribution $p_n$, which alone is not sufficient for exchangeability.