Conditional Dirichlet Processes and Functional Condition Models

TL;DR

This paper explores the conditional Dirichlet process in functional condition models, deriving posterior distributions, establishing limiting theorems, and providing algorithms with practical data analysis, while linking the Jeffreys likelihood to the model.

Contribution

It introduces a novel application of the cDP to functional condition models, deriving new theoretical results and algorithms, and clarifying the Jeffreys likelihood's role in this context.

Findings

Derived posterior distributions for functional condition models.

Established limiting theorems for the cDP posterior.

Proposed algorithms and demonstrated their effectiveness with data.

Abstract

In this paper, we study the conditional Dirichlet process (cDP) when a functional of a random distribution is specified. Specifically, we apply the cDP to the functional condition model, a nonparametric model in which a finite-dimensional parameter of interest is defined as the solution to a functional equation of the distribution. We derive both the posterior distribution of the parameter of interest and the posterior distribution of the underlying distribution itself. We establish two general limiting theorems for the posterior: one as the total mass of the Dirichlet process parameter tends to zero, and another as the sample size tends to infinity. We consider two specific models, the quantile model and the moment model, and propose algorithms for posterior computation, accompanied by illustrative data analysis examples. As a byproduct, we show that the Jeffreys substitute likelihood emerges as the limit of the marginal posterior in the functional condition model with a cDP prior, thereby providing a theoretical justification that has so far been lacking.