Infinitely Divisible Processes Overview

• Infinitely divisible processes are stochastic processes defined by finite-dimensional distributions that decompose into sums of i.i.d. components, characterized by Lévy–Khintchine representations.
• They encompass a broad class including Lévy, moving average, and selfsimilar processes, facilitating advanced analysis in path-space measures and semimartingale decompositions.
• Practical simulation techniques, such as shot-noise series and stochastic integrals, provide explicit error bounds and effective methods for modeling heavy-tailed and dependent structures.

An infinitely divisible process is a stochastic process such that for every integer $n\geq 1$, every finite-dimensional marginal admits a decomposition as the sum of $n$ independent identically distributed processes. This property connects the analysis of path-space measures, Lévy–Khintchine representations, series and integral decompositions, simulation techniques, and links to semimartingales, stable processes, and extremes of dependence.

1. Definition and General Structure

A random vector $X\in\mathbb{R}^d$ is infinitely divisible if, for each $n\in\mathbb{N}$, there exist i.i.d.\ random variables $X_{1,n},\dots,X_{n,n}$ such that $X\overset{d}=X_{1,n}+\cdots+X_{n,n}$. Every infinitely divisible law admits the Lévy–Khintchine representation: its characteristic function is

$$\varphi(\theta)=\mathbb{E}[e^{i\langle\theta,X\rangle}] =\exp\left( i\langle\theta,a\rangle - \tfrac12 \langle\theta,S\theta\rangle + \int_{\mathbb{R}_0^d} \left(e^{i\langle\theta,z\rangle} - 1 - i\langle\theta,z\rangle 1_{\{\|z\|\le 1\}}\right) \nu(dz) \right)$$

with drift $a\in\mathbb{R}^d$, covariance $S\ge0$, and Lévy measure $\nu$ satisfying $\int (1\wedge\|z\|^2)\nu(dz)<\infty$ [(Yuan et al., 2021), Section 1].

A stochastic process $\{X_t: t\ge0\}$ is infinitely divisible if each finite-dimensional distribution is infinitely divisible. Lévy processes form a central subclass of infinitely divisible processes but the concept encompasses a much broader class including moving averages, selfsimilar processes, and time-changed or subordinated structures [(Yuan et al., 2021); (Hakassou et al., 2012); (Fonseca-Mora, 2017)].

2. Lévy–Khintchine and Lévy–Itô Representations

The infinite divisibility property extends the Lévy–Khintchine formula to path spaces and duals of nuclear spaces. For a process $X=(X_t)_{t\in T}$, the path-space law is determined by a triplet $(E,v,b)$:

• $E:T\times T\to\mathbb{R}$: Gaussian covariance.
• $v$: path-space Lévy measure on $\mathbb{R}^T$ satisfying integrability and regularity conditions.
• $b\in\mathbb{R}^T$: drift function.

The finite-dimensional characteristic function is then

$$\mathbb{E}\exp \left(i \sum_{j=1}^n a_j X_{t_j}\right) = \exp\left(-\tfrac12 (a,E_I a) + \int_{\mathbb{R}^T} \left(e^{i \sum_j a_j x(t_j)} - 1 - i \sum_j a_j [x(t_j)]\right) v(dx) + i\sum_j a_j b(t_j)\right)$$

[(Rosinski, 2016), Sec. 1].

In the dual of a nuclear space $\Phi'_\beta$, one has the Lévy–Khintchine formula for infinitely divisible measures on $\Phi'_\beta$ with drift $m\in\Phi'_\beta$, covariance form $Q$ on $\Phi$, and Lévy measure $\nu$: $$\Psi(\varphi) = i m[\varphi] - \tfrac12 Q(\varphi) + \int_{\Phi'\setminus\{0\}} \left(e^{i f[\varphi]} - 1 - i f[\varphi] 1_{\{\|f\|\le 1\}}\right)\nu(df)$$ [(Fonseca-Mora, 2017), Sec. 3].

The Lévy–Itô decomposition gives

$$L_t = t m + W_t + \int_{\|f\| \leq 1} f\, \widetilde{N}(t, df) + \int_{\|f\| > 1} f\, N(t, df)$$

where $W_t$ is a Gaussian process, $N$ is a Poisson random measure, and $\widetilde{N}$ its compensated version [(Fonseca-Mora, 2017), Sec. 4].

3. Series and Integral Representations

Shot-noise and Ferguson–Klass–Rosiński Series

Any infinitely divisible law with Lévy measure $\nu$ (after suitable decomposition) admits a shot-noise series

$$X \overset{d}{=} \sum_{k=1}^\infty \left[ H(\Gamma_k, U_k) - c_k \right]$$

where $\Gamma_k$ are Poisson points, $U_k$ are i.i.d.\ marks, and $c_k$ are centering shifts chosen for convergence. This representation is robust in multivariate and path-space settings, greatly facilitating simulation [(Yuan et al., 2021), Sec. 2].

Truncation after $n$ terms yields a residual tail $R^{(n)}$, whose mean-square (or higher moment) error can be explicitly evaluated: $$\mathbb{E}\|R^{(n)}\|^2 = \int_{\mathbb{R}_0^d} \|z\|^2 [\nu-\nu_n](dz)$$ where $\nu_n$ is the truncated Lévy measure [(Yuan et al., 2021), Eq. (2)].

Stochastic Integrals and Moving Averages

General process representations are given as

$X_t = \int_{\mathcal{T}} f(t,s)\, dL_s$

where $L_s$ is a Lévy process and $f$ is a suitable kernel. For stationary increment mixed moving averages (SIMMA),

$X(t) = \int_{\mathbb{R}\times V} f(t,s,v) W(ds,dv)$

where $W$ is an ID independently scattered random measure and $f$ is typically a deterministic kernel (Basse-O'Connor et al., 2012, Basse-O'Connor et al., 2014).

Time-changed and selfsimilar structures arise in additive processes with dilative stability: $X_t = \int_{-\infty}^{\log t} e^{u(a-d/2)} dY_u$ with $Y$ itself constructed from an integral with respect to a background process and ultimately leading to a time-changed Lévy process representation [(Bhatti et al., 2016), Sec. 2].

4. Semimartingale and Path Regularity Criteria

A general result for infinitely divisible semimartingales $X$ representable as stochastic integrals against infinitely divisible random measures $A$: $$X_t = \int_{(-\infty,t]\times V} \phi(t,s,v) A(ds,dv)$$ satisfies that $X$ is a semimartingale if and only if it admits a unique decomposition

$X_t = X_0 + M_t + A_t$

where $M$ is càdlàg with independent increments (the “martingale” part) and $A$ is a predictable, finite variation process,

$M_t = \int_{(0,t]\times V} \phi(s,s,v) A(ds,dv)$

$$A_t = \int_{(-\infty,t]\times V} [\phi(t,s,v) - \phi(s+,s,v)] A(ds,dv)$$

(Basse-O'Connor et al., 2012, Basse-O'Connor et al., 2014).

For stationary increment cases ($X$ a SIMMA), absolute continuity of the kernel in time (with suitable moment and integrability conditions for the jump and Gaussian parts) is necessary and sufficient for finite variation paths (Basse-O'Connor et al., 2012).

5. Sampling and Simulation Techniques

Shot-noise truncation provides efficient, practical sample-path generation:

• Fix truncation parameter $n$ (on Poisson intensity or absolute jump size).
• Simulate Poisson number of jumps and their marks.
• For SDEs driven by Lévy noise, a jump-adapted discretization combines exact simulation of jump times with strong-order schemes for drift and diffusion between jumps.

Error control is explicit: for $\alpha$-stable processes, the number of jumps $n$ required for mean-square error $\le \varepsilon^2$ satisfies $n \asymp \varepsilon^{-2\alpha/(2-\alpha)}$; for tempered or gamma processes, error decays exponentially allowing smaller $n$ (Yuan et al., 2021, Kawai, 2021).

Small jumps (with controllable second moment) can be replaced by a Gaussian approximation, reducing computational cost [(Yuan et al., 2021), Sec. 6].

6. Infinitely Divisible Processes with Respect to Time and Extensions

A process $\{X_t\}_{t\geq 0}$ is infinitely divisible with respect to time (IDT) if for every $n\geq 1$,

$$\{X_{nt}\}_{t\geq0} \overset{d}{=} \left\{\sum_{i=1}^n X_t^{(i)}\right\}_{t\geq 0}$$

where $X^{(1)},\ldots,X^{(n)}$ are i.i.d. copies. Each one-dimensional marginal $X_t$ is then classically infinitely divisible, and there exists a unique associated Lévy process with matching marginals [(Hakassou et al., 2012), Sec. 1].

Strong IDT processes admit the stronger property of path decomposition: $$\{X_t\}_{t\geq0} \overset{d}{=} \left\{ \sum_{i=1}^n X^{(i,n)}_{\,t/n} \right\}_{t\geq 0}$$ [(Mai et al., 2018), Sec. 1].

Multiparameter extensions (IDT of type 1 or 2) and weak IDT (equality in law only for fixed $t$) extend the theory to a broader class, including classical Gaussian random fields and operator-stable fields (Hakassou et al., 2012).

7. Applications, Examples, and Extreme Models

Representative infinitely divisible processes:

• Fractional Lévy processes: moving averages with fractional kernels driven by Lévy processes, semimartingale if and only if suitable kernel regularity and small-jump moments are satisfied [(Basse-O'Connor et al., 2012); (Basse-O'Connor et al., 2012)].
• Stable and tempered stable Lévy motions: synthesized via LePage or shot-noise series, with simulation and error bounds explicit for heavy-tailed increments (Yuan et al., 2021, Kawai, 2021).
• Non-decreasing (subordinator) ID processes: Min- and max-infinite divisibility, exchangeable sequences with extreme-value and copula structure, constructed using stochastic process analogues of Marshall–Olkin or Archimedean models (Brück et al., 2020, Mai et al., 2018).
• Long-memory stationary infinitely divisible sequences: Ergodic-theoretic representations for heavy-tailed processes, with functional large deviation principles governed by null-recurrence and conservative flows [(Ghosh, 2010); (Owada, 2013)].

Extremal dependence structures, copulas, and the de Finetti-type correspondences further extend the reach of infinitely divisible processes in probabilistic dependence modeling (Brück et al., 2020).

---

The theory of infinitely divisible processes underpins a wide spectrum of stochastic modeling, simulation, and limit theory. Their path properties—finite variation, semimartingale structure, and regularity—can be characterized through their Lévy–Khintchine triplets and explicit series/integral representations, with robust tools for simulation and broad applicability from stochastic integration to extreme-value analysis and multivariate dependence [(Yuan et al., 2021); (Rosinski, 2016); (Fonseca-Mora, 2017); (Kawai, 2021); (Hakassou et al., 2012); (Brück et al., 2020); (Mai et al., 2018)].