Conditional Dirichlet Processes and Functional Condition Models
Jaeyong Lee, Kwangmin Lee, Jaegui Lee, Seongil Jo · 2025
Read Conditional Dirichlet Processes and Functional Condition Models by Jaeyong Lee, Kwangmin Lee, Jaegui Lee, Seongil Jo (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-15932.txt.
Introduction
Jaeyong Lee and co-authors study in this 2025 statistics paper a Bayesian tool called the conditional Dirichlet process. The setting is a functional condition model, where the number of interest is defined as the answer to an equation about the whole distribution, not read off the data directly. The authors work out how beliefs about that number, and about the distribution itself, update as data arrive.
Two limit theorems anchor the theory: one as the process's total mass shrinks to zero, and one as the sample grows without bound. Worked cases cover a quantile model and a moment model, with algorithms and data examples. A side result gives the Jeffreys substitute likelihood a firm basis at last.
Catalog record
| CREATOR | Jaeyong Lee, Kwangmin Lee, Jaegui Lee, Seongil Jo |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 55 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-15932
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Conditional Dirichlet Processes and Functional Condition Models Jaeyong Lee, Kwangmin Lee, Jaegui Lee, Seongil Jo 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. 1 Introduction The Dirichlet process [ ferguson1973bayesian ] is a nonparametric prior distribution over probability distributions. First introduced by Ferguson in 1973, the Dirichlet process has since become a cornerstone of Bayesian nonparametrics due to its elegant mathematical properties and practical applicability. One of its most notable features is conjugacy: if \displaystyle F follows a Dirichlet process and \displaystyle X_{1},\ldots,X_{n}\mid F\sim F , then the posterior distribution of \displaystyle F remains a Dirichlet process. Additionally, the marginal distribution of the \displaystyle X_{i} ’s follows a Pólya urn scheme, which is central to posterior inference via Gibbs sampling in models based on the Dirichlet process. Other key theoretical results, such as the stick-breaking construction or Sethuraman’s representation, provide both mathematical insight and practical utility. With the advancement of Markov chain Monte Carlo (MCMC) methods, the Dirichlet process has been employed in a wide range of Bayesian models. Among them, the mixture of Dirichlet processes (MDP) stands out as a foundational model, widely used for density estimation and model-based clustering. In this paper, we aim to employ the Dirichlet process as part of the prior distribution for a functional condition model, which will be formally defined later. In this context, the full Dirichlet process is not required; rather, a conditional Dirichlet process, constrained by a functional of the distribution function, is necessary. In probability theory, the definition of conditional probability is neither intuitive nor entirely mathematically satisfactory. To illustrate this, consider the case where a random probability measure \displaystyle F follows a Dirichlet process. When conditioning on a functional \displaystyle g(F)=t\in\mathbb{R} , the conditional probability of the event that \displaystyle F lies in a certain set is not defined as a single number, but rather as a \displaystyle\sigma(g(F)) -measurable function, where \displaystyle\sigma(g(F)) is the \displaystyle\sigma -field generated by \displaystyle g(F) . ? and ? have made efforts to address this issue. In this paper, we follow the approach of ? . ? define conditional distributions using the concept of disintegration, which provides a more intuitive definition of conditional distributions. Disintegration decomposes a given measure into a family of measures on the level sets of a functional, where each component measure corresponds to a constant multiple of a conditional probability distribution. This formulation aligns well with the intuitive understanding of conditional distributions. In this paper, we derive the conditional distribution of the Dirichlet process based on the concept of disintegration. We now define the functional condition model and discuss its Bayesian formulation. A simplified version of the proposed functional condition model is given by \displaystyle\theta \displaystyle\sim\pi (1) \displaystyle F\mid g(F)=\theta \displaystyle\sim\mathcal{D}_{\alpha}(dF\mid g(F)=\theta) \displaystyle x_{1},\ldots,x_{n}\mid F,\theta \displaystyle\stackrel{{\scriptstyle i.i.d.}}{{\sim}}F, where \displaystyle\theta is the parameter of interest and the primary focus of the inference, \displaystyle g(F) denotes a functional of the distribution function \displaystyle F , and \displaystyle\mathcal{D}_{\alpha}(dF\mid g(F)=\theta) represents the conditional Dirichlet process given \displaystyle g(F)=\theta . The base measure \displaystyle\alpha of the Dirichlet process is a finite and non-null measure. In the literature, most existing works have considered the case where \displaystyle g(F) corresponds to a moment of \displaystyle F , leading to the term, moment condition model. In contrast, in this paper, \displaystyle g(F) is not restricted to a moment; it can be any functional of \displaystyle F . Accordingly, we refer to our framework as the functional condition model. In the existing Bayesian literature, moment condition models have primarily been studied using empirical likelihood and the exponentially tilted empirical likelihood (ETEL) methods. The empirical likelihood is obtained by maximizing the empirical likelihood ratio over the level set determined by \displaystyle g(F) . When the functional corresponds to a moment of \displaystyle F , nearly explicit expressions for the empirical likelihood can be derived. See ? for details. ? proposed using the empirical likelihood within Bayesian models in a manner analogous to the parametric likelihood. ? applied empirical likelihood to quantile regression models and established the asymptotic normality of the resulting posterior distribution. ? extended empirical likelihood methods to spatial models. ? applied the Bayesian empirical likelihood framework to generalized linear models. Furthermore, ? proposed a Hamiltonian Monte Carlo method tailored for Bayesian empirical likelihood. ? derived a likelihood function for use in moment condition models by taking the limit of the marginal distribution in a nonparametric model, and termed this likelihood the exponentially tilted empirical likelihood (ETEL). ? investigated the application of ETEL in Bayesian moment condition models and demonstrated that the Bernstein–von Mises theorem holds even under model misspecification. ? further extended the ETEL framework to conditional moment condition models. The empirical likelihood and ETEL methods offer the advantage of enabling Bayesian inference for parameters of interest in a manner analogous to that in parametric models, without requiring Bayesian computation for infinite-dimensional parameters. This is achieved by constructing the likelihood for the parameter of interest within a semiparametric model. However, these methods have certain limitations: empirical likelihood is not a true likelihood in the strict sense, and the likelihood derived from ETEL is constructed using a prior distribution that is only employed in the derivation of ETEL itself. Furthermore, the ETEL model for conditional moment condition models [ chib2022bayesian ] requires the construction of somewhat artificial conditions to approximate the conditional distributions. In this paper, we derive the posterior distribution and the marginal likelihood for the parameter of interest by employing the Dirichlet process, a widely used prior in nonparametric Bayesian modeling. Furthermore, we demonstrate that inference for the parameter of interest can be conducted in the same manner as in parametric models, using the derived marginal likelihood. Recently, ? proposed the DPGLM, a semiparametric model for the generalized linear model, where the conditional distribution of the response is specified as a covariate-dependent, exponentially tilted, normalized completely random measure. The DPGLM offers an elegant posterior sampling method for the semiparametric model. In this paper, we study the Bayesian analysis of functional condition models using the conditional Dirichlet process as part of the prior distribution. We first derive key properties of the conditional Dirichlet process using the concept of disintegration. Similar to the standard Dirichlet process, the conditional Dirichlet process serves as a conjugate prior. When used as a prior, the conditional Dirichlet process