Poisson Thinning Framework

• Poisson Thinning Framework is a probabilistic operation that retains each event with a fixed probability, resulting in a Poisson process with scaled intensity.
• It unifies fundamental limit theorems such as the law of thin numbers with applications in Bayesian nonparametrics, statistical mechanics, and simulation acceleration.
• Extensions like dependent thinning and information-theoretic approaches offer unified methods for inference, cross-validation, and spatial statistics.

The Poisson thinning framework encompasses a set of probabilistic constructions and information-theoretic principles centered on the operation of "thinning" Poisson processes or, more generally, completely random measures (CRMs) by random deletion or selection of their atoms. Thinning is fundamental in the theory of point processes, with deep connections to discrete probability, statistical mechanics, information theory, and Bayesian nonparametrics. Modern developments reveal a unifying perspective, linking classic results such as the law of small numbers, information-divergence inequalities, and dependent random measure modeling.

1. Formal Definition and Core Operator

Let $X$ be an $\mathbb{N}$-valued random variable, or more generally, a point process $N$ on a space $(E, \mathcal{E})$. The $p$-thinning of $X$ (or $N$), with $p \in [0,1]$, is the random variable (or process) resulting from independently retaining each "atom" (such as a count increment, or a point) with probability $p$ and discarding it with $1-p$:

• Discrete Case: For $\mathbb{N}$0, $\mathbb{N}$1 with $\mathbb{N}$2 iid and independent of $\mathbb{N}$3.
• Point Process Case: $\mathbb{N}$4 where $\mathbb{N}$5 iid conditioned on $\mathbb{N}$6 (Johnson, 2015, Mukherjee, 2021).

This operation leaves Poisson laws invariant: if $\mathbb{N}$7, then $\mathbb{N}$8 and the retained/discarded processes are independent Poisson processes (Mukherjee, 2021).

Generating Functions: For discrete RVs, the thinning operator acts as $\mathbb{N}$9, showing an "argument shift" in the pgf (Johnson, 2015, 0906.0690).

2. Thinning, Laws of Thin Numbers, and Poisson Approximation

The "law of thin numbers" and the "law of thin processes" are foundational probabilistic theorems that clarify the analogy between thinning and Gaussian scaling:

• Law of Thin Numbers: For $N$0 i.i.d. integer-valued RVs $N$1, the thinned convolution $N$2 converges in distribution to $N$3 with $N$4, with explicit sharp rates in total variation, entropy, and Kullback-Leibler divergence (0906.0690, Harremoës et al., 2016).
• Law of Thin Processes: For i.i.d. point processes $N$5 with mean/intensity measure $N$6, the superposition $N$7 and $N$8 converges in distribution to a Poisson process with mean measure $N$9 (Aldridge, 20 Feb 2025).

Thinning commutes with convolution: $(E, \mathcal{E})$0 for independent $(E, \mathcal{E})$1, paralleling the scaling properties in the central limit regime.

3. Extensions: Dependent Thinning and Unified Framework for Random Measures

The classical (independent) Bernoulli thinning operation has been generalized to capture various forms of dependency and covariate structure, most notably in nonparametric Bayesian modeling:

• Augmented Poisson Process for CRMs: Dependent CRMs on a parameter space $(E, \mathcal{E})$2 are built by thinning a Poisson process on an augmented space $(E, \mathcal{E})$3, where $(E, \mathcal{E})$4 indexes auxiliary features (e.g., spatial or temporal covariates) and $(E, \mathcal{E})$5 is the atom's mass (Foti et al., 2012). The thinning function $(E, \mathcal{E})$6 is user-specified, allowing complex dependence on a covariate $(E, \mathcal{E})$7, leading to dependent measures $(E, \mathcal{E})$8.
  • Marginals: For each $(E, \mathcal{E})$9, $p$0 has an explicit CRM Lévy measure.
  • Covariances: Covariances across $p$1 are governed by the correlation induced by the thinning functions.
  • Conjugacy retention: Thinning only alters atom inclusion and does not affect the atom’s marginal mass distribution.
• Special Cases: By varying the base CRM and thinning function, this framework recovers the kernel beta process, dependent Dirichlet process, and other dependent CRMs (Foti et al., 2012).

4. Information-Theoretic Structure: Entropy, Inequalities, and Functional Analysis

Poisson thinning is central to discrete analogues of Gaussian information theory:

• Maximum Entropy: Within the class of ultra-log-concave (ULC) laws, the Poisson distribution maximizes entropy at fixed mean; thinning contracts entropy and Fisher information in a manner strictly parallel to scaling in the Gaussian case (Johnson, 2015, Harremoës et al., 2016, 0906.0690).
• Functional Inequalities: Model variants such as Poincaré and discrete log-Sobolev inequalities admit sharp constants under thinning, with the Poisson law saturating the bounds (Johnson, 2015).
• Monotonicity Laws: Entropy increases and Kullback-Leibler divergence to Poisson decreases under iterative thinning (the “law of thin numbers”), with sharp rates and monotonicity for ULC kernels.
• Discrete Ornstein-Uhlenbeck Analogue: The thinning Markov chain $p$2 provides a discrete semigroup mirroring the OU process, with Poisson–Charlier polynomials as eigenfunctions (0906.0690).

5. Applications: Inference, Simulation, and Bayesian Models

The Poisson thinning operation underlies diverse inferential and applied statistical paradigms:

• Model Selection and Cross-Validation: In convolution-closed families, such as Poisson or normal, data thinning allows unbiased splitting of observations for internal validation—crucial in unsupervised learning and settings where traditional sample splitting is infeasible (e.g., single-cell RNA-seq, matrix factorization) (Neufeld et al., 2023).
• Multimode Poisson Estimation: Thinning with mode-specific probabilities enables explicit and strongly consistent estimation of mixture proportions and thinning parameters via Sylvester-Ramanujan systems, with unique global MLEs and efficient computation in multilayer physical processes (e.g., neutron detection) (Anevski et al., 2018).
• Bayesian Thinning in Inverse Problems: Bayesian inference for highly structured point-source models (e.g., identification in PDEs) deploys marked Poisson process priors and a thinning step acting as a posterior filter, preserving unbiasedness and detailed balance in MCMC (Deng et al., 4 Sep 2025).
• Simulation Acceleration: Thinning algorithms, especially in nonhomogeneous/thin simulations, are foundational for rapid Monte Carlo schemes (e.g., kinetic Ising model trajectory sampling) by converting complex arrival-time inversion into efficient accept/reject sampling (Tokar et al., 30 May 2025).

6. Dependent, Deterministic, and Neighbourhood Thinning: Extensions and Non-Independence

A non-exhaustive taxonomy of advanced thinning frameworks arising in modern point process theory:

Thinning Type	Description & Key Properties	Canonical Results/Papers
Independent	Atomwise iid Bernoulli. Poisson→Poisson.	(0906.0690, Mukherjee, 2021)
Determinantal	Thinned processes via DPP on a realization to induce repulsion.	(Błaszczyszyn et al., 2018)
Neighbour-based	Retention probability depends on local neighborhood counts.	(Hlyniana, 13 Nov 2025)
Dependent Gibbs	Acceptance depends on Papangelou intensity and configuration graph.	(Hirsch et al., 26 Jan 2026, Last et al., 2021)
Thinning-stable	Point process law invariant under (possibly non-iid) thinning + scaling.	(Zuyev, 17 Apr 2025)

Neighbour-count dependent thinning produces spatial models exhibiting inhibition or clustering as dictated by retention rules $p$3 based on local counts; fine properties of intensity and correlation emerge from explicit expansions and Poisson-mixture formulas (Hlyniana, 13 Nov 2025). Determinantal thinning provides repulsive point process models with closed-form Palm, void, and Laplace functionals (Błaszczyszyn et al., 2018). Cluster-wise and Papangelou-based thinning underlie perfect simulation and coupling for spatial Gibbs measures (Hirsch et al., 26 Jan 2026, Last et al., 2021).

7. Deep Structural and Free-Probability Analogues

Poisson thinning has fundamental characterization and universality properties:

• Characterization Results: The property that Poisson processes split into independent Poissons under thinning uniquely characterizes the Poisson and Bernoulli families (Mukherjee, 2021). Analogous results hold in free probability, where free Poisson and Bernoulli variables play parallel roles, connecting to random matrix theory and high-dimensional statistics.
• Scaling, Stability, and Universality: Poisson thinning is the fixed mean, maximum-entropy, contractive operation in discrete stochastic analysis, mirroring Gaussian scaling’s role for continuous distributions (Johnson, 2015). Stable laws under thinning (thinning-stable point processes) naturally lead to Cox processes driven by strictly $p$4-stable random measures. As $p$5 the Poisson process is recovered (Zuyev, 17 Apr 2025).
• REM Universality: Poisson thinning of large random energy ensembles leads to point process convergence toward a Poisson process with exponential intensity, supporting universality analogies and convergence to Poisson–Dirichlet weights in statistical physics (Concetti et al., 5 Jun 2026).

---

In summary, the Poisson thinning framework—encompassing Bernoulli/independent and more general thinning rules—forms a canonical toolkit unifying the structure, limit theory, information-theoretic extremality, and statistical inference for discrete random variables, point processes, and completely random measures (Johnson, 2015, Foti et al., 2012, Zuyev, 17 Apr 2025, Mukherjee, 2021, 0906.0690). This machinery enables both precise theoretical analysis and practical algorithms, with applications ranging from Bayesian nonparametric modeling and cross-validation to simulation acceleration and spatial statistics.