Fractional Poisson Process

• Fractional Poisson process is a renewal process defined by heavy-tailed Mittag–Leffler waiting times, where the parameter α controls memory and scaling characteristics.
• It is governed by fractional differential equations with Caputo derivatives and possesses explicit generating functions that capture overdispersion and long-range dependence.
• The process has broad applications in anomalous diffusion, finance, and risk analysis, with extensions including state-dependent and compound models offering enhanced modeling flexibility.

A fractional Poisson process (FPP) is a non-Markovian, renewal-type extension of the classical Poisson process, replacing exponential inter-arrival times with heavy-tailed Mittag–Leffler distributed waiting times. The FPP is indexed by a real parameter $0<\alpha\le1$ (often denoted $\beta$ in the literature), which governs the memory and scaling properties of the process. This generalization is foundational in modeling event-driven phenomena with anomalous (non-Markovian, non-Lévy, long-range dependent) dynamics, and appears in fractional diffusion, insurance risk, finance, transport, and statistical anomaly detection (Vellaisamy et al., 2013, Politi et al., 2011, Meerschaert et al., 2010).

1. Mathematical Definition and Governing Equations

The FPP $\{N_\alpha(t)\}$ is constructed as a renewal process with i.i.d. inter-arrival times $(T_1, T_2, \ldots)$, each distributed according to the Mittag–Leffler law with density

$$f_\alpha(t) = \lambda t^{\alpha-1} E_{\alpha,\alpha}(-\lambda t^\alpha)$$

where $$E_{\alpha,\beta}(z) = \sum_{k=0}^{\infty} \frac{z^k}{\Gamma(\alpha k + \beta)}$$, $\lambda > 0$, $t>0$ (Politi et al., 2011, Meerschaert et al., 2010). The cumulative distribution and survival function are

$$F_\alpha(t) = 1 - E_\alpha(-\lambda t^\alpha), \qquad S_\alpha(t) = E_\alpha(-\lambda t^\alpha)$$

where $E_\alpha(z) := E_{\alpha,1}(z)$.

The counting process is

$\beta$0

State probabilities $\beta$1 solve the fractional Kolmogorov forward (master) equation with Caputo derivative: $\beta$2 with $\beta$3, $\beta$4 for $\beta$5 (Vellaisamy et al., 2013, Crescenzo et al., 2015, Kreer et al., 2013, Meerschaert et al., 2010). This generalizes the classical Poisson evolution via time-fractional calculus.

2. Explicit Distributions, Generating Functions, and Transforms

Closed-form expressions for $\beta$6 are available via series of Mittag–Leffler-type functions: $\beta$7 Alternatively, the probability generating function (PGF) and Laplace transforms are: $\beta$8

$\beta$9

These compact forms allow recovery of moments and enable analytic/numerical evaluation (Meerschaert et al., 2010, Kreer et al., 2013).

3. Subordination and Stable Inverse Representation

The FPP admits an equivalent construction via subordination: if $\{N_\alpha(t)\}$0 is a standard strictly increasing $\{N_\alpha(t)\}$1-stable subordinator,

$\{N_\alpha(t)\}$2

where $\{N_\alpha(t)\}$3 is the inverse stable subordinator, and $\{N_\alpha(t)\}$4 is an ordinary Poisson process (Meerschaert et al., 2010, Gorenflo et al., 2013, Kumar et al., 2018). The Laplace transform of the waiting times matches the Mittag–Leffler density. This time-change viewpoint is crucial in connecting the FPP to fractional diffusion equations and modeling subordinated Lévy processes.

Generalizations include tempered and distributed-order FPPs, employing subordinators with Laplace exponents $\{N_\alpha(t)\}$5: $\{N_\alpha(t)\}$6 enabling broader modeling scope (Meerschaert et al., 2010, Maheshwari et al., 2017).

4. Moments, Scaling, and Dependence Structure

The FPP exhibits anomalous, sublinear scaling of moments: $\{N_\alpha(t)\}$7 Variance exceeds the mean for $\{N_\alpha(t)\}$8, indicating overdispersion and event clustering due to heavy tails (Politi et al., 2011, Vellaisamy et al., 2013).

The process is non-Markovian and exhibits long memory:

• Increments are neither stationary nor independent: the transition probabilities depend on age and current state.
• Covariance decays algebraically, and the process can show long-range dependence (LRD) (Vellaisamy et al., 2013, Maheshwari et al., 2016, Kataria et al., 2020).

5. Extensions: Generalized, State-Dependent, Space-Time, and Compound FPP

Several extended models have been developed:

• Generalized FPP (GFPP): Waiting times follow Prabhakar–Mittag–Leffler laws parameterized by real $\{N_\alpha(t)\}$9, $(T_1, T_2, \ldots)$0; unifies Laskin FPP, Erlang, and Poisson (Michelitsch et al., 2019).
• State-dependent FPP: The fractional parameter $(T_1, T_2, \ldots)$1 is allowed to vary with state $(T_1, T_2, \ldots)$2, yielding unsolved waiting-time distributions and more complex Laplace transforms (Garra et al., 2013).
• Space-time FPP and non-homogeneous FPP: Incorporates spatial and temporal fractionalization, double subordination, and time-varying rates (Maheshwari et al., 2016).
• Compound FPP: Event magnitudes are random ($(T_1, T_2, \ldots)$3), with sums $(T_1, T_2, \ldots)$4; its characteristic function is $(T_1, T_2, \ldots)$5 (Scalas, 2011, Politi et al., 2011).
• Convoluted FPP (CFPP): Incorporates space variable convolution, generalized intensity $(T_1, T_2, \ldots)$6, modifies overdispersion and dependence (Kataria et al., 2020).

6. Estimation, Simulation, and Applications

Parameter estimation for the FPP is nontrivial due to lack of finite moments for $(T_1, T_2, \ldots)$7. Methods include:

• Method-of-moments (MOM): Estimation is based on sample mean and variance of $(T_1, T_2, \ldots)$8, yielding closed-form estimators for $(T_1, T_2, \ldots)$9 (Cahoy et al., 2018, Mendel et al., 11 Nov 2025).
• Maximum Likelihood (ML) and quantile-based (QB) methods provide better efficiency at higher computational cost (Mendel et al., 11 Nov 2025).
• Neural approaches: LSTM architectures outperform MOM estimators in MSE and computational speed, tracking time-varying parameter dynamics (Gupta et al., 5 Dec 2025).

Simulation algorithms exploit:

• Mixture representations of Mittag–Leffler variables (e.g., Kanter–Wiktorsson algorithm).
• Subordination via simulated α-stable processes (Maheshwari et al., 2017, Cahoy et al., 2018).

Applications span:

• Anomalous transport and diffusion: CTRW models with subdiffusive MSD scaling $$f_\alpha(t) = \lambda t^{\alpha-1} E_{\alpha,\alpha}(-\lambda t^\alpha)$$0 (Gorenflo et al., 2013, Scalas, 2011).
• Risk process and insurance: Surplus/risk evolution under heavy-tailed claims; ruin probability is unchanged from classical case (Kumar et al., 2018).
• Meteorology: Extreme event clustering and seasonal adaptation (e.g., extratropical cyclone return times) (Mendel et al., 11 Nov 2025).
• High-frequency finance, telecom, neurospike trains, reliability: Event bursts and heavy tail modeling.

7. Infinite Divisibility, Limit Theorems, and Analytical Structure

The FPP one-dimensional distributions are not infinitely divisible for $$f_\alpha(t) = \lambda t^{\alpha-1} E_{\alpha,\alpha}(-\lambda t^\alpha)$$1 (Vellaisamy et al., 2013). This discontinuity with classical Lévy processes results from the Mittag–Leffler subordinator properties. Limit theorems: $$f_\alpha(t) = \lambda t^{\alpha-1} E_{\alpha,\alpha}(-\lambda t^\alpha)$$2 where $$f_\alpha(t) = \lambda t^{\alpha-1} E_{\alpha,\alpha}(-\lambda t^\alpha)$$3 is the limiting inverse-stable random variable. Functional convergence and ergodic properties are rigorously characterized via Laplace and renewal theoretic methods (Vellaisamy et al., 2013, Kreer et al., 2013, Meerschaert et al., 2010).

Analytic structures for finite- and infinite-dimensional distributions are available, including multiple-integral formulas for joint laws, and system representations as infinite convergent ODE hierarchies in transformed time (Kreer et al., 2013, Politi et al., 2011).

8. Connections to Fractional Calculus and Diffusion

The FPP's evolution is governed by fractional differential equations—specifically, the Caputo or Riemann-Liouville derivatives. These underpin anomalous diffusion PDEs and fractional birth–death processes, and establish its foundational role in continuous-time random walks with memory (Vellaisamy et al., 2013, Michelitsch et al., 2019, Crescenzo et al., 2015).

Key foundational and recent works include (Vellaisamy et al., 2013, Politi et al., 2011, Meerschaert et al., 2010, Crescenzo et al., 2015, Cahoy et al., 2018, Mendel et al., 11 Nov 2025, Gupta et al., 5 Dec 2025, Kataria et al., 2020, Gorenflo et al., 2013, Garra et al., 2013, Maheshwari et al., 2016, Maheshwari et al., 2017, Scalas, 2011, Kreer et al., 2013, Kumar et al., 2018, Michelitsch et al., 2019).

---

The fractional Poisson process is thus a mathematically rigorous, parameter-rich, and analytically tractable generalization of the Poisson process, instrumental for modeling and analysis of systems exhibiting heavy tails, memory, and clustering in event timings.