Poisson Process Extensions Overview

• Poisson Process Extensions are advanced models that generalize classical Poisson processes to include non-exponential waiting times, batch arrivals, and state- or time-dependent intensities.
• Time-changed and subordinated processes use stochastic clocks to capture overdispersion and long-range dependencies, enhancing applicability in non-stationary and complex systems.
• Renewal-based, cluster, and noise-augmented extensions provide flexible modeling frameworks for applications in reliability, queueing, and risk analysis.

A Poisson process is a prototypical stochastic counting process with stationary, independent increments and a fixed jump size. Various extensions have been proposed to address non-standard dynamics observed in empirical systems, to allow for non-exponential waiting times, batch arrivals, or state- or time-dependent intensities. These extensions yield a broad and technically rich landscape of point process models critical for contemporary probability theory, applied stochastic modeling, and mathematical statistics. Key variants include time-changed and subordinated Poisson processes, Poisson processes of order $i$, compound and fractional Poisson processes, filtered and noise-augmented Poisson processes, geometric-, random-, and fixed-replacement Poisson systems, Poisson cluster processes, and functional invariance principles.

1. Weighted Sums: Poisson Processes of Order $i$ and Compound Generalizations

A fundamental extension is the Poisson process of order $i$, defined as

$Y(t) = \sum_{j=1}^i j\,N_j(t),$

where $N_1, \dots, N_i$ are independent homogeneous Poisson processes each of rate $\lambda$. This is equivalently a compound Poisson process with uniform integer jumps on $\{1,\dots,i\}$ and parent rate $i\lambda$, i.e., for $N(t)\sim$ Pois$(i\lambda t)$,

$$Y(t) \stackrel{d}{=} \sum_{\ell=1}^{N(t)} X^i_\ell,\quad P\{X^i_\ell = j\}=1/i.$$

The probability generating function is explicit:

$$E[u^{Y(t)}] = \exp\{-i\lambda t + \lambda t \sum_{j=1}^i u^j\}.$$

The Kolmogorov forward equation for the marginals $p_n(t) = P\{Y(t)=n\}$ reads

$$\partial_t p_n(t) = -i\lambda p_n(t) + \lambda \sum_{j=1}^i p_{n-j}(t).$$

Higher moment structure is determined directly:

$$E[Y(t)] = \lambda t\, \frac{i(i+1)}{2},\quad \mathrm{Var}[Y(t)] = \lambda t\, \frac{i(i+1)(2i+1)}{6}.$$

This framework generalizes further to arbitrary weighted combinations $Z(t) = \sum_j g(j) N_j(t)$ by selecting integer-valued $g(\cdot)$. Such processes unify the classical Poisson, batch Poisson, and ordered Poisson distributions (Maheshwari et al., 2019, &&&1&&&).

2. Time-Changed and Subordinated Poisson Processes

Time-changing a Poisson process, i.e., replacing physical time $t$ by a stochastic time $H^f(t)$ determined by an independent subordinator (Lévy process with increasing paths), yields a Cox process with highly non-trivial marginal and trajectory structure. For a Bernstein subordinator $H^f(t)$ with Laplace exponent $f(\mu)$, the time-changed process

$W(t) := Y(H^f(t)) = \sum_{j=1}^i j\,N_j(H^f(t))$

admits a probability generating function

$$E[u^{W(t)}] = \exp\{-t f[i\lambda - \lambda \sum_{j=1}^i u^j]\},$$

and is a Cox process ("doubly stochastic Poisson") conditional on $H^f(t)$ (Maheshwari et al., 2019). The mixture and compound-sum representations hold:

$$W(t) \stackrel{d}{=} \sum_{\ell=1}^{M(t)} X^i_\ell,$$

with $M(t)$ a Cox process as above.

For geometric subordinated Poisson processes (GSPP), the stochastic clock is a discrete geometric counting process $G_\mu(t)$, parameterized by $\mu>0$. The process $N_G(t) = N(G_\mu(t))$ is tractable in its pgf, distributional characteristics, and higher-order joint moment structure. The GSPP encompasses the mixed Poisson, Poisson, and batch-counting special cases as particular parameter regimes (Gupta et al., 26 Feb 2025). The compound and multiplicative versions—GSCPP and GSMPP—extend this framework to general jump distributions and products of positive random variables, respectively.

Additionally, time-fractional generalizations use subordinators such as the inverse stable process $E_\beta(t)$, yielding the time-fractional compound Poisson process (TFCPP), in which marginal distributions satisfy a Caputo fractional differential equation and display long-range dependence and overdispersion (Vellaisamy et al., 2024).

3. Hitting Times, State Skipping, and Recurrences

Extensions with jumps of size greater than one introduce the phenomenon of "skipping states": the process can make transitions that skip certain integer levels entirely with positive probability. For $Y(t)$ as above, the probability of ever hitting state $k$ is

$$P\{T_k < \infty\} = \begin{cases} k/i, & 1 \leq k < i, \ 1, & k \geq i. \end{cases}$$

and $P\{T_k = \infty\} = 1 - k/i > 0$ for $k < i$ (Maheshwari et al., 2019). In the time-changed Cox process case, the heavy-tailed nature of the subordinator further increases the range of levels that may be skipped, with $P\{T_k < \infty\}<1$ for all $k$.

Hitting time densities are represented via convolution-type formulas:

$$P\{T_k\in dt\} = \lambda \sum_{h=1}^k P\{Y(t) = k - h\} dt$$

and more generally, for time-changed processes, with jump distributions reflecting the randomization of the clock.

Kolmogorov forward equations and recurrence relations for marginals extend, with higher-order difference-differential operators (including, in the fractional-time case, replacing $\partial/\partial t$ by $f(\partial_t)$).

4. Statistical Structure, Overdispersion, and Long-Range Dependence

Compound, subordinated, and fractional Poisson extensions all exhibit overdispersion relative to the standard Poisson process (variance exceeding the mean), with explicit formulas for mean, variance, covariance, and higher moments. For instance, the generalized compound Poisson process with jumps $Y_j$ and total rate $\delta = \sum_j \lambda_j$ yields:

$$E[H(t)] = t \sum_j j \lambda_j, \qquad \mathrm{Var}(H(t)) = t \sum_j j^2 \lambda_j > E[H(t)],$$

with long-range dependence evident in the correlation decay:

$$\operatorname{Corr}(H(s), H(t)) \sim \sqrt{\frac{s}{t}}, \quad t\to\infty.$$

Time-fractional extensions with index $\beta$ further slow the correlation decay (as $t^{-\beta}$), break infinite divisibility, and display overdispersion persistently across scales (Vellaisamy et al., 2024).

In geometric subordinated Poisson process models, the covariance of $(N_G(s), N_G(t))$ decays as $t^{-1}$, confirming upper-orthant dependence and long-memory characteristics (Gupta et al., 26 Feb 2025).

5. Renewal-Based and Replacement Extensions

The replace-after-fixed-time (RaFT) process constrains each interarrival time $X_k$ (Exponential($\lambda$)) to a maximum of $r>0$, i.e., $Y_k = \min\{X_k, r\}$, forming $N(t)$ as the maximal $n$ s.t. $\sum_{k=1}^n Y_k \leq t$. Probabilities, moments, and generating functions are computed explicitly using renewal theory and convolution identities. The model further generalizes to the replace-after-random-time (RaRT) process where $r$ is randomized with continuous law $f_R(r)$. Both models can be viewed as batch-truncated or censored renewal processes, capturing behaviors relevant in reliability, queueing, and maintenance applications (Marengo et al., 2018).

As $r \to \infty$, these processes revert to the standard Poisson structure. Randomizing $r$ increases over-dispersion, with the mean rate given asymptotically by $\lambda/[1-e^{-\lambda r}]$ (RaFT) or $\lambda/\mathbb{E}[1-e^{-\lambda R}]$ (RaRT), always exceeding the homogeneous Poisson rate.

6. Extensions via Perturbation Formulas and Invariance Principles

Margulis–Russo-type perturbation formulas for Poisson processes enable unified derivations of integral representations for distributions such as Poisson, binomial, negative-binomial, and compound Poisson, and facilitate variational characterizations, including Crofton-type derivative formulas in integral geometry, and new integro-differential equations for strictly $\alpha$-stable and related distributions (Last et al., 2019). Functional invariance principles in spaces $\ell^\infty$ of bounded functions allow Poisson limit theorems (both classical and non-stationary) to be upgraded to process-level convergence, characterizing the weak limit of partial sums of (possibly dependent) arrays by time-changed Poisson processes, with covariate-adapted scaling and time-transformations (Niang et al., 2022).

7. Poisson Cluster, Filtered, and Noise-Augmented Models

Cluster models introduce a center Poisson process $N(t)$ (possibly non-homogeneous) whose events trigger random-size or random-duration "clusters" or secondary processes $L_j(t)$ (additive, with independent increments). The resultant process accommodates batch arrivals and non-stationary dynamics:

$M(t) = \sum_{j:\,T_j \leq 1} L_j(t - T_j).$

Conditional prediction and mean-square error formulas are obtainable via Laplace transforms, recursions, and renewal-theoretic reductions, supporting actuary and risk applications with non-stationary payment or claim structure (Matsui, 2013).

Filtered Poisson processes pass Poisson pulses through a convolution kernel (typically exponential), often with additional additive or dynamical noise. These models are foundational for descriptions of intermittent turbulence, electronic shot noise, and related systems. Moments, autocorrelation, and spectral density are given analytically, with clear diagnostic signals for distinguishing between purely additive versus dynamical noise via spectral slope and autocorrelation discontinuity at zero lag (Theodorsen et al., 2016).

---

References:

• "Superposition of time-changed Poisson processes and their hitting times" (Maheshwari et al., 2019)
• "A New Compound Poisson Process and Its Fractional Versions" (Vellaisamy et al., 2024)
• "Geometrical subordinated Poisson processes and its extensions" (Gupta et al., 26 Feb 2025)
• "Replace-after-Fixed-or-Random-Time Extensions of the Poisson Process" (Marengo et al., 2018)
• "Prediction in a non-homogeneous Poisson cluster model" (Matsui, 2013)
• "Statistical properties of a filtered Poisson process with additive random noise: Distributions, correlations and moment estimation" (Theodorsen et al., 2016)
• "Applications of the perturbation formula for Poisson processes to elementary and geometric probability" (Last et al., 2019)
• "$\ell^{\infty}$ Poisson invariance principles from two classical Poisson limit theorems and extension to non-stationary independent sequences" (Niang et al., 2022)