Ornstein–Uhlenbeck–Lévy Noises

• Ornstein–Uhlenbeck–Lévy noises are defined as solutions to Langevin-type SDEs driven by Lévy processes, capturing both diffusion and jump behaviors.
• They exhibit both spectral and non-spectral relaxation modes, enabling precise modeling of heavy-tailed phenomena and anomalous decay rates.
• These processes offer a robust framework for invariant measure estimation and law equivalence analysis, with applications in finance, physics, and spatial-temporal modeling.

Ornstein–Uhlenbeck–Lévy noises are a central class of stochastic processes arising as solutions to stochastic differential equations (SDEs) in which the driving noise is a Lévy process, generalizing the classic Ornstein–Uhlenbeck (OU) process beyond Brownian drivers. These processes serve as canonical models for jump-diffusion and heavy-tailed temporal and spatial stochastic phenomena and form the backbone of modern stochastic analysis, statistical physics, finance, SPDEs, and high-dimensional probability theory. The mathematical structure, regularity, relaxation, invariant measures, law equivalence, and estimation theory for OU–Lévy models have been thoroughly investigated, both in finite and infinite-dimensional contexts.

1. Definition and Core Structure

The OU–Lévy process is defined as the solution to a Langevin-type SDE driven by a Lévy process. In the finite-dimensional setting, the classical version is given by

$dX_t = -\nu X_t\,dt + dL^\mu_t$

where:

• $X_t$ is the process (scalar or vector-valued);
• $\nu>0$ is the friction or drift coefficient;
• $L^\mu_t$ is a symmetric $\mu$-stable Lévy process with characteristic function $\mathbb E[e^{ikL^\mu_t}] = \exp(-K|k|^\mu t)$, for some $K>0$, with $\mu \in (0,2]$ controlling the tail index and jump activity.

The dynamics can be equivalently described by the associated fractional Fokker–Planck (master) equation for the probability density $p(x,t)$: $$\partial_t p(x,t) = \nu \partial_x[x p(x,t)] + K D_x^{\mu} p(x,t)$$ where $X_t$0 is the Riesz–Weyl fractional derivative, acting as $X_t$1, and the process admits jumps for $X_t$2 (Thiel et al., 2016).

Generalizations to Banach or Hilbert spaces, or even to operator-valued and tempo-spatial systems, are formulated via

$X_t$3

where $X_t$4 is the generator of a $X_t$5-semigroup and $X_t$6 is an $X_t$7-valued (or cylindrical) Lévy process. The mild solution is given by the stochastic convolution

$X_t$8

(Applebaum, 2014, Riedle, 2012, Pham et al., 2016).

2. Existence, Uniqueness, and Path Regularity

For tempo-spatial models (e.g., Volterra-type OU equations),

$X_t$9

well-posedness is secured under assumptions on the drift measure $\nu>0$0, kernel $\nu>0$1, and Lévy basis $\nu>0$2. Uniqueness and measurability in the appropriate Banach context follow when the resolvent of $\nu>0$3 exists and the Lévy measure $\nu>0$4 satisfies moment conditions (e.g., $\nu>0$5 for suitable choices of $\nu>0$6 and $\nu>0$7).

If $\nu>0$8 is sufficiently regular (e.g., $\nu>0$9 with derivatives in $L^\mu_t$0), and the Gaussian component is absent, then the process has a càdlàg version in time, with continuous spatial paths (Pham et al., 2016). For Gaussian noise and $L^\mu_t$1 square-integrable with suitable increment bounds, $L^\mu_t$2 admits locally Hölder-continuous paths.

In Hilbert/Banach space, stochastic integrals with respect to Lévy or cylindrical Lévy noise are well-defined under regulated operator-valued integrands whose range and jump measures satisfy tightness and covariance criteria (Riedle, 2012).

3. Spectral Relaxation and Non-spectral Modes

The relaxation behavior of Ornstein–Uhlenbeck–Lévy processes is governed by the interplay between the drift and the Lévy driving law:

• For certain initial densities (domain of attraction: sufficiently light tails), relaxation rates coincide with the spectral points of a Hermitian operator associated via similarity transformation to the Fokker–Planck generator. For the classic OUP, this yields relaxation rates at $L^\mu_t$3 ($L^\mu_t$4 for scalar case with noise index $L^\mu_t$5) (Thiel et al., 2016, Toenjes et al., 2012).
• When the initial condition is broader-tailed (e.g., an $L^\mu_t$6-stable law, $L^\mu_t$7), “non-spectral” modes emerge, leading to anomalously slow relaxation at rates $L^\mu_t$8. The slowest non-spectral rate $L^\mu_t$9 governs asymptotic decay. In generalized OUPs with heavy-tailed noise, non-spectral relaxation is generic (Toenjes et al., 2012).
• Observables $\mu$0 admit expansions $\mu$1; the slowest (smallest) $\mu$2 dominates long-time decay.

Continuous wavelet transformation provides an algorithmic way to extract spectral and non-spectral rates from data (Thiel et al., 2016).

4. Stationarity, Invariant Measures, and Heavy Tails

For the Markovian case (exponential memory kernel), stationary and invariant laws are derived via convolution with the driving noise and propagator. The invariant measure for OU–Lévy processes is infinitely divisible, with parameters determined by the drift, covariance, and Lévy measure of the driving noise, as well as the time-integrated propagator: $\mu$3

$\mu$4

$\mu$5

(Applebaum, 2014). The law is operator self-decomposable and, under exponential stability, the convolution law converges weakly to the invariant measure.

OU dynamics retain the heavy-tailed properties of the driving Lévy process, and in models with regime-switching (i.e., random switching of drift coefficient and volatility via a Markov chain), the stationary law may exhibit heavy tails even if each regime is individually light-tailed. The tail index is determined via the spectral properties of the Markov chain and drifts (Liao et al., 2019).

5. High-Dimensional, Operator-Valued, and Cylindrical Structures

In infinite dimensions, the OU–Lévy equation

$\mu$6

has unique solutions in both mild and strong senses, provided the operator $\mu$7 generates a strongly continuous group and the stochastic integral $\mu$8 is defined (using the Lévy–Itô decomposition, or cylindrical integration for generalized noise) (Applebaum, 2014, Riedle, 2012).

Stochastic volatility and ambit-field models generalize the basic OU–Lévy theory to operator-valued and space–time settings via noise $\mu$9 and volatility processes $\mathbb E[e^{ikL^\mu_t}] = \exp(-K|k|^\mu t)$0 in Hilbert–Schmidt or operator spaces. Affine structures facilitate explicit expressions for the characteristic function of the solution, enabling direct calculation of marginal and transition laws (Benth et al., 2015, Pham et al., 2016).

Graph-based network time series models, such as the grOU framework, adapt the OU–Lévy system to complex networked systems, capturing jump and heavy-tail phenomena in both node- and edge-indexed time series (Chen et al., 15 May 2026).

6. Law Equivalence and Rigidity Phenomena

The equivalence of path laws for OU–Lévy systems with differing drift operators but the same driving noise is characterized by:

• For non-degenerate Gaussian parts (eigenvalues of noise covariance strictly positive), laws are equivalent if the difference of drift operators satisfies suitable integrability (Hilbert–Schmidt) conditions and directional Cameron–Martin hypotheses. The Radon–Nikodym derivative can be expressed via a Doleans–Dade exponential (Kania, 20 Oct 2025, Bartosz et al., 2018).
• For pure-jump Lévy noise (vanishing Gaussian part and drift), the only way the path laws can be equivalent is if the processes coincide pathwise; “rigidity” prohibits nontrivial Girsanov changes of measure in this setting (Kania, 20 Oct 2025, Bartosz et al., 2018).

Necessary and sufficient conditions for integrability, invariant measure existence, and law equivalence are provided in the infinite- and cylindrical-dimensional context (Riedle, 2012, Applebaum, 2014).

7. Practical Estimation and Applications

Maximum likelihood and Bayesian estimation procedures for OU–Lévy models and their generalizations (including the “Cosine process” and network time series) are facilitated by explicit forms for the likelihood (often quadratic under Gaussian assumptions) and approximations for stable densities (e.g., Fox H-function representations) (Stein et al., 2021, Chen et al., 15 May 2026).

Applications are diverse, including modeling of financial time series with heavy tails, volatility forecasting in commodity markets via stochastic volatility and ambit-field approaches, as well as spatial–temporal modeling in meteorology, turbulence, and SPDEs with jump noise.

Simulation studies and real-data fits (e.g., Apple Inc. high-frequency returns, cardiovascular mortality rates) indicate that Lévy-driven OU models accurately reproduce heavy-tailed, jump-diffusive behavior, and model selection can be guided by KS, AD, modified KS, and quantile-based statistics (Stein et al., 2021).

---

Key references:

• (Thiel et al., 2016) Detailed analysis of spectral and non-spectral relaxation for OUPs with stable noise.
• (Toenjes et al., 2012) Rigorous foundations for non-spectral modes in generalized OU systems.
• (Liao et al., 2019) Long-time behavior, heavy tail emergence through regime-switching.
• (Pham et al., 2016) Volterra-tempo-spatial OU processes.
• (Srokowski, 2011) Stationary distributions and regularity for OUL processes with inertia and colored noise.
• (Stein et al., 2021, Chen et al., 15 May 2026) Parameter estimation and real-world applications.
• (Applebaum, 2014, Riedle, 2012) Infinite and cylindrical dimensional generalizations, stochastic integration.
• (Kania, 20 Oct 2025, Bartosz et al., 2018) Law equivalence and rigidity in the presence/absence of a Gaussian component.
• (Benth et al., 2015) Operator-valued stochastic volatility and ambit-field links.