[Paper Review] Coagulation Fragmentation Laws Induced By General Coagulations of Two-Parameter Poisson-Dirichlet Processes

Man-Wai Ho, Lancelot F. James+1 more|ArXiv.org|Jan 25, 2006

Coagulation and Flocculation Studies21 references3 citations

TL;DR

This paper introduces a novel analytical framework based on generalized Cauchy-Stieltjes transforms to characterize coagulation-fragmentation dynamics in two-parameter Poisson-Dirichlet processes, extending Pitman's duality result beyond the $PD(\beta,\theta/\alpha)$ family. It provides explicit descriptions for $PD(\alpha,\theta)$ coagulated by a broad class of power-tempered Poisson-Kingman models, including cases involving Dirichlet processes, and offers a new proof of Pitman’s original duality via distributional composition of random measures.

ABSTRACT

Pitman~(1999) describes a duality relationship between fragmentation and coagulation operators. An explicit relationship is described for the two-parameter Poisson-Dirichlet laws, with parameters {\footnotesize $(α,θ)$} and $(β,θ/α)$, wherein $PD(α, θ)$ is coagulated by $PD(β,θ/α)$ for $0

Motivation & Objective

• To generalize Pitman’s coagulation-fragmentation duality for $PD(\alpha,\theta)$ beyond the $PD(\beta,\theta/\alpha)$ family.
• To develop a non-combinatorial method for characterizing $PD(\alpha,\theta)$ coagulated by arbitrary laws $Q$ using distributional relationships of species sampling models.
• To provide explicit descriptions of the resulting random probability measures when $PD(\alpha,\theta)$ is coagulated by power-tempered Poisson-Kingman models.
• To establish connections between coagulation dynamics and Chinese restaurant franchise processes in Bayesian nonparametrics.

Proposed method

• Utilizes generalized Cauchy-Stieltjes transforms of random probability measures to derive distributional identities.
• Applies results from Vershik, Yor, and Tsilevich (2004) and James (2002) on composition of species sampling random measures.
• Employs the framework of Poisson-Kingman models to characterize $Q$-coagulated $PD(\alpha,\theta)$ for a broad class of $Q$.
• Derives finite-dimensional distributions via mixtures of Dirichlet distributions conditioned on $Q$'s law.
• Uses the structure of exchangeable random partitions and EPPF-based analysis to validate results.
• Establishes equivalence between coagulated $PD(\alpha,\theta)$ and Chinese restaurant franchise processes under certain conditions.

Experimental results

Research questions

• RQ1 Can coagulation-fragmentation duality in $PD(\alpha,\theta)$ be extended beyond the $PD(\beta,\theta/\alpha)$ family using analytical methods?
• RQ2 How can generalized Cauchy-Stieltjes transforms be used to characterize $PD(\alpha,\theta)$ coagulated by arbitrary $Q$?
• RQ3 What are the finite-dimensional distributions of $PD(0,\theta)$ coagulated by $PD(1/2,\eta)$ or $PD(0,\nu)$?
• RQ4 Is there a connection between coagulated Poisson-Dirichlet processes and Chinese restaurant franchise processes?
• RQ5 Can the combinatorial proof of Pitman’s duality be replaced by a distributional transform-based approach?

Key findings

• The paper provides a new proof of Pitman’s coagulation-fragmentation duality using Cauchy-Stieltjes transforms instead of combinatorial EPPF arguments.
• It characterizes $PD(\alpha,\theta)$ coagulated by power-tempered Poisson-Kingman models, extending beyond the $PD(\beta,\theta/\alpha)$ family.
• For $PD(0,\theta)$ coagulated by $PD(1/2,\eta)$, the finite-dimensional distribution is derived as a mixture involving $\Gamma(\theta)$ and $\Gamma(\eta+m/2)$ terms with a power-law density component.
• For $PD(0,\theta)$ coagulated by $PD(0,\nu)$, the finite-dimensional distribution is expressed as a double integral over Dirichlet and $Q$-induced densities with $\Gamma(\nu)$ normalization.
• The results establish a formal equivalence between coagulated $PD(\alpha,\theta)$ processes and Chinese restaurant franchise processes, particularly when $Q$ is a Dirichlet process.
• The framework enables explicit characterization of $R = P(Q\text{-COAG})$ for $P = PD(\alpha,\theta)$ and general $Q$, including non-Poisson-Dirichlet laws.