Law of Thin Numbers in Poisson Approximation

• Law of Thin Numbers is a probabilistic limit theorem that characterizes the convergence of thinned convolutions of integer-valued random variables to a Poisson law.
• It establishes sharp convergence in metrics like total variation and relative entropy, paralleling the entropic Central Limit Theorem using Rényi’s thinning operation.
• The methodology employs semigroup properties, Poisson–Charlier expansions, and majorization to extend classical Binomial-to-Poisson results to point processes and beyond.

The Law of Thin Numbers (LTN) is a discrete probabilistic limit theorem governing the convergence of thinned convolutions of integer-valued random variables to a Poisson law, tightly paralleling the entropic Central Limit Theorem (CLT) in the information-theoretic regime. Its core lies in Rényi’s thinning operation and the interplay between convolution, entropy, and majorization, providing both sharp convergence results and lower bound rates in strong metrics like relative entropy. The LTN generalizes beyond the classical Binomial-to-Poisson scenario and extends to point processes, establishing its foundational role in the structure of Poisson approximation and discrete limit theory.

1. Thinning Operation and Definitional Framework

Thinning is a mapping on discrete probability laws defined as follows: Let $f = \{f(y): y\in\mathbb{N}_0\}$ be a probability mass function (pmf) on the non-negative integers with mean $\lambda$. For $\alpha \in [0,1]$, the $\alpha$-thinning $T_{\alpha}(f)$ (Editor’s term: "thinning operator") is the law of the random sum

$$Y_\alpha := \sum_{i=1}^{Y} X_i \qquad \text{for}\quad Y \sim f,\ X_i \mid Y = y \sim \text{i.i.d. Bern}(\alpha),$$

so that

$$T_\alpha(f)(k) = \sum_{y=k}^\infty f(y) \binom{y}{k} \alpha^k (1-\alpha)^{y-k}, \qquad k=0,1,2,...$$

The primary intuition is that, from $Y$ "particles," each is retained independently with probability $\alpha$ (0810.5203, 0906.0690, Harremoës et al., 2016).

The thinning operator satisfies crucial algebraic properties: it commutes with convolution ($T_\alpha(f * g) = T_\alpha(f) * T_\alpha(g)$) and forms a semigroup ($T_\alpha \circ T_\beta = T_{\alpha \beta}$). This operation is the canonical discrete analogue of scaling in the CLT.

2. Statement of the Law of Thin Numbers and Modes of Convergence

For a pmf $f$ with mean $\lambda<\infty$, consider the sequence

$f_n := T_{1/n}(f^{*n}), \qquad n=1,2,3,...$

where $f^{*n}$ is the $n$-fold convolution. The Law of Thin Numbers asserts the following (0810.5203, 0906.0690, Harremoës et al., 2016):

• Pointwise/weak convergence: $f_n \Rightarrow \mathrm{Poisson}(\lambda)$ as $n \to \infty$ (under minimal moment conditions).
• Total variation convergence: $\|f_n - \mathrm{Po}(\lambda)\|_{TV} \to 0$.
• Information-theoretic convergence: If $D(f^{*n}\|\mathrm{Po}(n\lambda)) < \infty$ for large $n$, then $D(f_n\|\mathrm{Po}(\lambda)) \to 0$ with $D(f_n\|\mathrm{Po}(\lambda))$ non-increasing in $n$.
• Shannon entropy monotonicity: If $f$ is ultra-log-concave (ULC), $H(f_n)$ increases monotonically towards $H(\mathrm{Po}(\lambda))$.

These results generalize the classical Binomial-to-Poisson convergence and establish the LTN as the discrete, entropic analogue of the CLT (0810.5203, 0906.0690).

3. Entropy Monotonicity, Relative Entropy, and ULC Distributions

A central feature of the LTN is monotonicity in both entropy and relative entropy. For any $f$ with mean $\lambda$:

$D(T_{1/n}(f^{*n}) \| \mathrm{Po}(\lambda))$

is finite for large enough $n$, strictly decreases in $n$, and converges to zero. For ULC $f$—that is, $i!f(i)$ forms a log-concave sequence or equivalently $i f(i) / f(i-1)$ is non-increasing—the entropy $H(f_n)$ increases monotonically, with the Poisson law achieving maximal entropy among fixed-mean ULC laws (0810.5203, 0906.0690). The proofs rely on convexity arguments, semigroup properties, and majorization orders, with crucial ingredients from size-biasing, data-processing inequalities, and de Bruijn-type representations.

4. Quantitative Rates and Moment-Based Lower Bounds

Explicit rates of convergence under the LTN have been characterized using Poisson–Charlier polynomial expansions and information-theoretic lower bounds (Harremoës et al., 2016):

• Second-moment bound: For $X \sim f$,

$$D(f\|\mathrm{Po}(\lambda)) \geq \frac12 [\mathrm{Var}(X) - \lambda]^2.$$

• Poisson–Charlier bounds: If $\mathbb{E}[C_2^{(\lambda)}(X)] \neq 0$, then

$$D(f\|\mathrm{Po}(\lambda)) \geq \frac12 \left(\mathbb{E}[C_2^{(\lambda)}(X)]\right)^2$$

with $$C_2^{(\lambda)}(k) = \frac{k^2-(2\lambda+1)k+\lambda^2}{\sqrt{2}\lambda}$$. For higher orders, rates depend on the minimal non-vanishing Charlier moment.

• Rates for Binomial thinning: $$D(\mathrm{Bi}(n,1/n)\|\mathrm{Po}(\lambda)) \asymp n^{-2}$$.

In general, for $k$ the order of the first nonvanishing Poisson–Charlier moment, $D(f_n\|\mathrm{Po}(\lambda)) = \Theta(n^{2-2k})$.

5. Information-Theoretic Analogies and Operator-Theoretic Perspectives

The trajectory of LTN mirrors that of the CLT under scaling, with thinning replacing scaling, convolutions of discrete laws replacing summation of continuous r.v.’s, and the Poisson law supplanting the Gaussian law as the limit. Key operator-theoretic parallels have been established:

• Second quantization $\Gamma(\alpha)$: Contracts densities in (Poisson) Charlier polynomial bases, modeling the action of thinning.
• Wick product: Represents convolution in the Poisson space.
• Hölder–Young-type inequalities: Provide $L^1$ contraction properties for Wick-convolved densities, supporting strong convergence in distribution (Lanconelli, 2015).

The Markov semigroup generated by the thinning operation plus free superposition—termed the “thinning Markov chain”—is directly analogous to the Ornstein–Uhlenbeck process, with spectral decay governed by Poisson–Charlier polynomials and monotonic decay of scaled Fisher information (0906.0690).

6. Extension to Point Processes: Law of Thin Processes

The Law of Thin Numbers extends to spatial point processes as the “law of thin processes” (Aldridge, 20 Feb 2025). For a locally-finite point process $\xi$ on a Polish space $𝓧$ with intensity measure $\mu$, the superposed-thinned process

$\eta_n := (1/n) \,\text{\textcircled{$\phantom{a}$}}\ (\xi_1 + \cdots + \xi_n)$

(where each point is independently kept with probability $1/n$ after aggregation) converges in distribution to a Poisson process with intensity $\mu$. The proof rests on the alternative probability generating functional $A_\xi(u) = \mathbb{E}[\prod_{x \in \xi}(1-u(x))]$, which is multiplicative under superposition and respects thinning via argument rescaling. This result highlights the universality of Poisson approximation under thinning-superposition and the invariance under higher-order dependencies, provided the first moment measure is finite.

7. Significance, Connections, and Applications

The Law of Thin Numbers provides the canonically sharp, information-theoretic Poisson limit for discrete structures, explaining the optimality of the Poisson law in entropy maximization under ULC constraints and establishing precise rates of convergence in relative entropy, total variation, and $L^1$ metrics. The theory bridges Poisson and Gaussian approximation, revealing that control over principal Poisson–Charlier moments suffices for fine quantitative Poisson approximation, with immediate applications in stochastic processes, point process theory, information theory, and operator algebraic approaches.

A significant implication is that matching mean and variance—the first two moments—often suffices for rapid Poisson approximation whenever higher-order Poisson–Charlier moments vanish, directly paralleling the two-moment sufficiency principle in Gaussian CLT theory. The extension to point processes ("law of thin processes") further broadens the reach of the LTN to infinite-dimensional settings and stochastic geometry (Aldridge, 20 Feb 2025, 0810.5203, Harremoës et al., 2016, 0906.0690, Lanconelli, 2015).