Asymptotic behavior of clusters in hierarchical species sampling models

Abstract: Consider a sample of size $N$ from a population governed by a hierarchical species sampling model. We study the large $N$ asymptotic behavior of the number ${\bf K}N$ of clusters and the number ${\bf M}{r,N}$ of clusters with frequency $r$ in the sample. In particular, we show almost sure and $L^p$ convergence for ${\bf M}{r,N}$, obtain Gaussian fluctuation theorems for ${\bf K}_N$, and establish large deviation principles for both ${\bf K}_N$ and ${\bf M}{r,N}$. Our approach relies on a random sample size representation of the number of clusters through the corresponding non-hierarchical species sampling model.

• The paper analyzes asymptotic properties of cluster counts in hierarchical species sampling models, establishing convergence, Gaussian fluctuations, and large deviations.
• The analysis uses a random index representation for HSSMs and extends known results from single-level models via probabilistic arguments, including martingale approaches.
• This work has implications for Bayesian nonparametrics and estimating unseen species, applying to various hierarchical configurations like the Pitman-Yor process.

Asymptotic Behavior of Clusters in Hierarchical Species Sampling Models

The paper by Favaro, Feng, and Paguyo addresses the asymptotic properties of hierarchical species sampling models (HSSMs), particularly focusing on the number of clusters, denoted as $K_N$, and the number of clusters of a specific frequency $r$ in the sample, denoted as $M_{r,N}$. The results reveal fundamental asymptotic behaviors characterized by almost sure convergence, Gaussian fluctuations, and large deviation principles for these quantities.

Central to the analysis is the random index representation for HSSMs, an extension of species sampling models (SSMs), which introduces hierarchical dependencies by considering multiple levels of random measures. Among the various types of SSMs considered, the hierarchical Pitman-Yor process (HPYP) is prominent. The HPYP generalizes several known processes, including the hierarchical Dirichlet process, and is crucial in this paper's exploration of the hierarchical structures' effects on the asymptotic behavior of sampling models.

Key Results

1. Almost Sure and $LP$ Convergence:
   • The authors establish almost sure and $LP$ convergence results for $M_{r,N}$, building on existing properties of single-level species sampling models. These results are crucial for understanding long-term characteristics of $M_{r,N}$ as the sample size $N$ becomes large.
2. Gaussian Fluctuations:
   • The paper presents Gaussian fluctuation results for the number of clusters $K_N$ in both single and hierarchical settings through rigorous probabilistic arguments. Using martingale approaches and exploiting the hierarchical nature of the model, the authors extend known Gaussian fluctuations to the hierarchical context, which signifies a step forward in quantifying the variances inherent in such models.
3. Large Deviation Principles:
   • The large deviation principles for $K_N$ and $M_{r,N}$ provide insight into the probability of deviations from expected behavior. The analysis extends known results from single-level frameworks and illustrates the impact of hierarchical structures on these principles.
4. Hierarchical Extensions:
   • By considering different combinations and mixtures of the Dirichlet and Pitman-Yor processes, the paper broadens the applicability of the results to a range of hierarchical configurations. This versatility is particularly useful for applications in Bayesian nonparametrics and related fields where hierarchical structures are prevalent.

Implications and Future Work

The findings have substantial implications for theoretical advancements in Bayesian nonparametric methods, specifically in the study of random partitions and sampling models. Projects such as estimating unseen species numbers, developing consistent estimators for hierarchical model parameters, and characterizing convergence rates in hierarchical settings can directly benefit from this work.

Researchers are encouraged to explore open problems mentioned, such as Gaussian fluctuations for $M_{r,N}$, and the non-asymptotic analysis of $K_N$ and $M_{r,N}$. Additionally, Bayesian inference for species sampling problems, especially in estimating the number of unseen species, as discussed in the paper, presents a rich area for further investigation. The pursuit of practical and efficient computational methods for posterior distributions in hierarchical models also remains promising.

In conclusion, this paper crucially enhances our understanding of the asymptotic behavior of hierarchical species sampling models, providing a solid foundation for further theoretical and applied research within this domain.