Species Sampling Process

• Species Sampling Process is a framework for modeling exchangeable discrete distributions via random probability measures with atomic support on Polish spaces.
• It utilizes latent partition representations and exchangeable partition probability functions (EPPFs) to depict the clustering structure and asymptotic behavior.
• The approach bridges classical models like the Dirichlet and Pitman–Yor processes, highlighting how base measure compositions influence sparsity and cluster fusion.

A species sampling process (SSP) is a framework for modeling random discrete distributions that arise through the assignment of observations to clusters (species) according to an exchangeable law. In its standard construction, an SSP consists of a random probability measure on a Polish space $X$ with atomic support, where both the locations (species labels) and the weights are random, and observables are i.i.d. from this draw. The theory of SSPs and their clustering structure, as developed by Bassetti and Ladelli, provides a unified representation for exchangeable species sampling sequences, general base measures, explicit partition laws, asymptotic count formulas, and implications for Bayesian sparsity and clustering (Bassetti et al., 2019). This article examines the mathematical structure, representation theory, asymptotics, and applied relevance of SSPs.

1. Formal Definition and de Finetti-Type Representation

The canonical SSP is a random discrete probability measure

$P = \sum_{j\ge 1} p_j\,\delta_{Z_j},$

where:

• $(p_j)_{j\ge 1}$ are random, nonnegative weights summing to 1,
• $(Z_j)_{j\ge 1}$ are i.i.d. points sampled from a general base measure $H$ on $X$, independent of $(p_j)$.

A sequence $(\xi_n)_{n\ge1}$ is called a generalized species-sampling sequence (gSSS$(q,H)$) directed by $P$ if, conditional on $P$, $\xi_n \overset{iid}{\sim} P$. The resulting sequence is exchangeable and satisfies a de Finetti representation: $$P\{\xi_1 \in A_1, \ldots, \xi_n \in A_n\} = \int \prod_{i=1}^n P(A_i)\; d\mathcal{L}(P),$$ where $\mathcal{L}(P)$ is the law of $P$. Equivalently, each observation may be written as

$\xi_n = Z_{I_n},$

for random assignment variables $I_n$ drawn i.i.d. according to $(p_j)$.

2. Latent Partition Representation and the Associated Exchangeable Partition Probability Function (EPPF)

The clustering structure of a species sampling sequence is governed by a latent exchangeable partition $\Pi$ of $\mathbb{N}$ with an EPPF $q$. By Kingman's correspondence, $q$ is determined by the ranked mass partition of $P$. There exists a representation: $$(\xi_n)_{n\ge 1} = \big(Z'_{C_n(\Pi)}\big)_{n\ge 1},$$ where

• $\Pi\sim\mathrm{EPPF}(q)$,
• $Z'_k \overset{iid}{\sim} H$ (independent of $\Pi$),
• $C_n(\Pi)$ gives the index of the block of $\Pi$ containing $n$.

This two-level construction is fundamental: first, an exchangeable partition $\Pi$ is generated according to $q$; then, each block draws an atom from $H$, producing possible block mergers when $H$ is not diffuse.

3. Partition Law Induced by the Observations and General Base Measure Effects

When $H$ is not purely diffuse, distinct blocks of $\Pi$ may receive the same atom, causing the observed partition to be coarser. The EPPF $q^*$ for the induced partition is expressible in terms of $q$ and multi-table tie probabilities determined by $H$. For a putative partition $T_n$ with block sizes $n_1,\dots,n_k$: $$P\{\text{induced partition} = T_n\} = \sum_{m \in M(n)} H^{\#}(m) \sum_{A\in A(m)} c(A)\,q(\text{table sizes under }A),$$ where:

• $M(n)$ indexes feasible subtable assignments,
• $H^{\#}(m)$ is the probability for $H$ that the prescribed table atom multiplicities occur,
• $A(m)$ enumerates assignments of subtables to final clusters,
• $c(A)$ provides combinatorial weights.

This formula quantifies the effect of base measure atoms, producing extra clustering (sparsity) due to coincident draws.

4. Asymptotic Behavior of Cluster Counts and Fixed-Size Frequencies

Let $K_n$ denote the number of clusters (blocks) induced by $(\xi_n)_{n=1}^n$, and $K_{n,r}$ the number with size exactly $r$. Suppose the directing partition $\Pi$ has asymptotic diversity $S>0$,

$K_n(\Pi)/C_n \to S\quad\text{a.s.,}$

for $C_n\to\infty$ (e.g., $C_n=n^o \ell(n)$ for Gibbs-type PRMs). With $H = c H_d + (1-c)H_c$ (atomic/diffuse mixture):

• If $H_d$ is finite-atomic, $K_n/C_n \to (1-c)S$ almost surely.
• If $H_d$ has infinitely many atoms, $K_n/b(C_n) \to (1-c)S$ for $b(x) = x^{o_0}\ell_0(x)$ as the growth rate of observed distinct atoms, or diverges if $c=1$.

For $r\ge 1$, in the Gibbs-type case: $$\frac{K_{n,r}}{n^o \ell(n)} \to (1-c)S \frac{\Gamma(r-o)}{r! \Gamma(1-o)}.$$

These laws quantify the effect of both the partition structure and base measure mixture on the proliferation of clusters and cluster sizes, with "atomic mass" $c$ controlling the dilution.

5. Consequences for Bayesian Nonparametrics and Sparsity Induction

The general base measure $H$—particularly in spike-and-slab or point-mass-support formulations—enables prior incorporation of sparsity and structural hypotheses. In applications (e.g., regression with sharp nulls, variable selection) one sets $H$ to have discrete atoms at preferred hypotheses, thereby merging latent partition blocks associated with these atoms. The induced random partitions can be interpreted by the metaphor: "random seating plan $\to$ table choices $\to$ dish assignments (atoms from $H$)" with merging when different tables draw the same atom.

This facilitates computation of predictive laws, EPPFs, and the explicit mechanism by which mixture bases in $H$ control clustering behavior, cluster merging, and exchangeability retention.

6. Classical Model Reductions: Dirichlet and Pitman–Yor Specializations

Specific choices for $q$ and $H$ recover standard models:

• For a Dirichlet process ($q$ Ewens–Pitman EPPF with $o=0$, $\theta$), integration yields the known spike-and-slab DP formulas.
• For the Pitman–Yor process ($0$\theta> -o$), the induced partition law specializes to the spike-and-slab PY formulas of Canale–Lijoi–Prünster, e.g. for $H(x)=a\delta_{x_0} + (1-a)H_c$,

$$P\{\Pi^*_n=(n_i)\} = (1-a)^k \frac{(\theta+o)_{(k-1); (1-o)}}{(\theta+1)_{(n-1)}} \prod_{i=1}^k (1-o)_{(n_i-1)} + \text{corrections of order }a(1-a)^{k-1}.$$

These reduce the general SSP partition laws to closed-form EPPFs associated to widely used BNP priors.

References and Further Directions

All foundational claims, representations, and formulas in this article are from Bassetti & Ladelli (Bassetti et al., 2019). The work subsumes diverse directions in Bayesian nonparametrics, mixture modeling, sparsity priors, and random partition theory, supporting practical posterior inference and prior elicitation in advanced clustering scenarios with general base measures. The two-level representation is especially powerful for understanding the interplay between latent partition structure and observed clustering, and the use of mixture base measures to encode prior information and sparsity in Bayesian analysis.