Generalized Poisson Random Field

• Generalized Poisson Random Field is a probabilistic model that extends classical Poisson fields to capture overdispersion, clustering, and structured spatial dependence while retaining analytic tractability.
• It can be constructed via tree-structured Markov random fields, compound Poisson measures, and fractional models, each offering distinct mechanisms for modeling count data.
• Its adaptable framework supports applications in spatial statistics, risk aggregation, and stochastic geometry, with efficient sampling and closed-form likelihoods enabling practical inference.

A Generalized Poisson Random Field (GPRF) is a probabilistic model that extends the classical Poisson random field by allowing for overdispersed, clustered, or more richly dependent spatial or indexed count structures, while retaining analytic tractability and, frequently, exact Poisson marginals. GPRFs encompass several distinct yet related constructions, including Markov random fields with Poisson marginals, compound and spatially indexed Poisson fields with generalized increment structures, and fractional generalizations governed by fractional differential equations. These frameworks are widely applicable in spatial statistics, multivariate count modeling, stochastic geometry, and applied probability, offering flexibility for both spatial independence and complex dependence settings.

1. Core Definitions and Model Classes

A GPRF is defined according to the geometric or index structure of interest:

• Tree-Structured Markov Random Fields: Let $\mathcal{T} = (V, E)$ be an undirected tree with $|V| = n$ vertices, and fix $\lambda > 0$. Each vertex $v$ hosts a count $N_v$, constructed recursively using a root $r\in V$ and parameters $\{\alpha_e \in [0,1]\}_{e\in E}$:

$$\begin{aligned} &L_r \sim \text{Pois}(\lambda), \ &L_v \sim \text{Pois}(\lambda(1-\alpha_{(\text{pa}(v),v)})) \text{ for } v \neq r, \ &N_r = L_r, \ &N_v = (\alpha_{(\text{pa}(v),v)} \circ N_{\text{pa}(v)}) + L_v, \;\; v\neq r, \end{aligned}$$

where $\alpha \circ X$ denotes binomial thinning. All marginals satisfy $N_v \sim \text{Pois}(\lambda)$ and the vector $|V| = n$0 is a Markov random field (MRF) (Côté et al., 2024).

• Compound Poisson Spatial Random Fields: On $|V| = n$1, let $|V| = n$2 denote all bounded rectangles. For $|V| = n$3 and rates $|V| = n$4, define $|V| = n$5, where $|V| = n$6 are independent Poisson random measures of intensity $|V| = n$7. The resulting random measure $|V| = n$8 is a GPRF, admitting the infinitesimal characterization:

$|V| = n$9

This allows for the possibility of multiple points (of type $\lambda > 0$0) in a vanishingly small region, supporting overdispersed or batch arrival phenomena (Vishwakarma, 23 Jan 2026).

• Fractional and Time-Changed GPRFs: Fractional GPRFs replace the index (e.g., area or volume) with an independent random process such as an inverse stable subordinator, leading to fractional order governing equations. For example, on $\lambda > 0$1, for $\lambda > 0$2, $\lambda > 0$3:

$\lambda > 0$4

where $\lambda > 0$5 is the rising Pochhammer symbol and $\lambda > 0$6 is the generalized Mittag-Leffler function (Kataria et al., 2024).

• Renewal-Process-Driven Spatial GPRFs: Here, at any location $\lambda > 0$7, $\lambda > 0$8 is the number of arrivals by time $\lambda > 0$9 with spatially varying, non-independent inter-arrival times, constructed from an underlying Gaussian random field, ensuring $v$0 and providing nontrivial spatial correlation (Morales-Navarrete et al., 2021).

2. Marginal and Joint Distributions

GPRFs are designed so that single-site marginals are either exactly Poisson or involve fractional or compound generalizations:

• Exact Poisson Marginals: All tree-structured GPRFs and certain spatial compound GPRFs achieve $v$1 or $v$2 under specific parameterizations, independent of dependence structure or thinning parameters (Côté et al., 2024, Vishwakarma, 23 Jan 2026).
• Joint Probability Mass Function (PMF): For tree-structured MRFs, the joint PMF has a recursive factorization:

$v$3

(Côté et al., 2024).

• Probability Generating Functions (PGF): For the random measure $v$4 on a region $v$5,

$v$6

for the spatial compound-Poisson GPRF (Vishwakarma, 23 Jan 2026), or through recursive formulas involving tree factorization for MRF GPRFs (Côté et al., 2024).

• Fractional PGFs: Fractional generalizations satisfy differential equations of Caputo type and their PGFs are expressed in generalized Mittag–Leffler or Wright series,

$v$7

(Kataria et al., 2024).

3. Dependence, Thinning, and Stochastic Ordering

Dependence in GPRFs is governed by explicit parameters or structural transformations:

• Binomial Thinning and Markov Structure: In tree-structured GPRFs, dependence is introduced through edge parameters $v$8, regulating the transmission of counts via binomial thinning. Increasing $v$9 strengthens dependence, interpolating from independence ($N_v$0) to comonotonicity ($N_v$1) (Côté et al., 2024).
• Superposition and Compound Representations: Spatial GPRFs admit construction via superposition of independent Poisson random fields or, equivalently, as compound Poisson random measures with cluster sizes (or marks) distributed as a finite mixture (Vishwakarma, 23 Jan 2026).
• Thinning Operations: Both classical ($N_v$2-thinning) and general GPRF-thinning (mark-dependent) lead to independent GPRFs with reduced intensity parameters, preserving the field structure, and providing natural tools for modeling selection or loss processes (Vishwakarma, 23 Jan 2026).
• Stochastic Ordering: On a fixed tree and $N_v$3, increasing any $N_v$4 results in a supermodular ordering; sums become larger in the convex order as dependence increases. This enables rigorous comparative statics and informs risk aggregation analyses (Côté et al., 2024).

4. Fractional and Time-Changed Extensions

Fractional GPRFs generalize classical fields via random time changes and fractional calculus:

• Time-Change Representations: Fractional GPRFs are realized by running a classical GPRF on an independent inverse stable subordinator, introducing long-range dependence and heavy-tailed growth in each index. The random time process $N_v$5 has Laplace transform

$N_v$6

(Kataria et al., 2024).

• Governance by Fractional PDEs: The PGF satisfies a Caputo-type fractional differential equation with respect to volume,

$N_v$7

with $N_v$8, and reduces to the classical equation for $N_v$9 (Kataria et al., 2024).

• Moments and Covariance: For a set $r\in V$0, \begin{align*} \mathbb{E}[N_{\alpha, \gamma}(B)] &= \frac{\gamma \lambda |B|^\alpha}{\Gamma(\alpha+1)} \ \mathrm{Var}[N_{\alpha, \gamma}(B)] &= \frac{\gamma \lambda |B|^\alpha}{\Gamma(\alpha+1)} + \frac{\gamma(\gamma+1)\lambda^2 |B|^{2\alpha}}{\Gamma(2\alpha+1)} - \left(\frac{\gamma \lambda |B|^\alpha}{\Gamma(\alpha+1)}\right)^2 \end{align*} (Kataria et al., 2024).

5. Estimation, Sampling, and Computational Properties

The unified analytic structure of GPRFs yields practical statistical and computational methods:

• Exact Sampling: Tree-structured GPRFs admit a one-pass $r\in V$1 sampler utilizing $r\in V$2 Poisson and $r\in V$3 binomial draws in topological order (Côté et al., 2024). For compound-Poisson GPRFs, standard compound-Poissonization applies.
• Closed-Form Likelihoods: Analytically available PMFs and PGFs enable likelihood-based inference. For spatial GPRFs as renewal-driven fields, explicit bivariate series allow for composite likelihood or pairwise likelihood estimation, while zero-inflated models introduce further flexibility (Morales-Navarrete et al., 2021).
• Computational Efficiency: Tree-structured models obviate the need for normalization constants, support fast sum-aggregation (e.g., via Panjer recursions, FFT), and yield efficient allocation formulas for conditional expectation calculations (Côté et al., 2024).
• Comparison with Poisson–Log-Gaussian and Copula Models: Only the GPRF maintains (in certain constructions) exact Poisson marginals, a smooth (nugget-free) covariance structure, and interpretable dependence via model parameters. Other models may induce discontinuities or complicate physical interpretation (Morales-Navarrete et al., 2021).

6. Applications and Practical Interpretations

GPRFs allow for direct modeling of diverse, real-world spatial and multivariate count data exhibiting nontrivial dependence, overdispersion, or clustering:

• Overdispersed and Clustered Point Patterns: In spatial settings, the possibility of multiple (batch or clump) points in infinitesimal regions differentiates the GPRF from the Poisson field, capturing natural clustering in ecology (e.g., tree species aggregation), epidemiology (multiple infections), telecommunications (batch arrivals), and materials science (defect clustering) (Vishwakarma, 23 Jan 2026).
• Risk Aggregation and Allocation: For financial or insurance contexts, explicit formulas for sums $r\in V$4 and expected allocations $r\in V$5 facilitate loss aggregation, capital allocation, and risk measurement calculations (Côté et al., 2024).
• Integral and Path-Functionals: Integrals of GPRFs over space (or trajectories) admit compound-Poisson representations and allow variance calculations and functional central limit theorem analyses (Vishwakarma, 23 Jan 2026, Kataria et al., 2024).

7. Special Cases, Reductions, and Extensions

GPRFs unify and extend several classical and modern constructions:

• Reduction to Poisson and Fractional Poisson Fields: For $r\in V$6, $r\in V$7, and $r\in V$8, all GPRF expressions reduce to the classical Poisson random field (homogeneous or inhomogeneous, as parameterized). Fractional GPRF reduces to the Beghin–Orsingher fractional Poisson process when $r\in V$9, $\{\alpha_e \in [0,1]\}_{e\in E}$0 (Kataria et al., 2024, Vishwakarma, 23 Jan 2026).
• Skellam-Type and Signed Point Processes: By superposing independent GPRFs with possibly negative weights, one obtains Skellam-type fields with explicit moment-generating functions and compound-Poisson representations (Vishwakarma, 23 Jan 2026).
• Order Statistics and Path Integrals: Conditioning on random counts, distributional results for order statistics, and path integrals are tractable, supporting quantitative analysis in stochastic geometry and statistical physics (Kataria et al., 2024).

A plausible implication is that the analytical manageability, sampling tractability, and capacity for nuanced dependence modeling poise GPRFs as a versatile foundation for theory and applications in spatial statistics, risk aggregation, and the study of complex count processes.