Generalized Ornstein–Uhlenbeck Process

• Generalized Ornstein–Uhlenbeck process is a family of continuous-time mean-reverting models extending the classical OU by incorporating non-Brownian noise and complex state-space dynamics.
• It encompasses variants such as Lévy-driven, Volterra, and fractional models, each altering drift, memory kernels, or noise input to capture intricate system behaviors.
• These processes are crucial for analyzing equilibrium, ergodicity, and resetting in various applications including finance, turbulence, and geophysical systems.

The generalized Ornstein–Uhlenbeck process is a family of continuous-time mean-reverting stochastic models built by extending the classical Ornstein–Uhlenbeck process beyond Brownian forcing on a one-dimensional Euclidean state space. The classical benchmark is the Gaussian Markov diffusion

$dX_t=-\gamma(X_t-\mu)\,dt+\sqrt{2D}\,dW_t,$

with mean

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$

variance

$$\operatorname{Var}[X_t]=\frac{D}{\gamma}\big(1-e^{-2\gamma t}\big),$$

and stationary Gaussian density

$$P_{\mathrm{st}}(x)=\sqrt{\frac{\gamma}{2\pi D}}\exp\!\left[-\frac{\gamma(x-\mu)^2}{2D}\right].$$

In the literature, the same expression “generalized Ornstein–Uhlenbeck process” is used for several non-equivalent constructions: multiplicative Lévy SDEs, Langevin equations with general stationary-increment noise, Volterra memory-kernel models, fractional and multifractal extensions, indexed and infinite-dimensional processes, state-space constrained diffusions, and history-valued random equilibria (Trajanovski et al., 2023, Kevei, 2016, Voutilainen et al., 2019, Applebaum, 2014).

1. Classical benchmark and principal definitions

The common structural core is linear mean reversion combined with a nontrivial noise or state-space mechanism. In the standard diffusion case, the restoring drift is linear and the process may be viewed as Brownian motion in a harmonic potential; bounded variance, an equilibrium Gaussian, and exponential relaxation are the signature properties (Trajanovski et al., 2023).

Several widely used generalizations are summarized by the following defining relations.

Variant	Defining relation	Distinguishing feature
Multiplicative Lévy GOU	$dV_t=V_{t-}\,dU_t+dL_t$	bivariate Lévy input
Vector-valued Langevin GOU	$dU_t=-\Theta U_t\,dt+dG_t$	matrix drift, stationary increments
Volterra GOU type	$V(t)=V_0\rho(t)+\int_0^t \rho(t-s)\,dL(s)$	memory kernel
Delay/SFDE GOU	$dx(t)=L(x_t)\,dt+\Sigma\,dB(t)$	history-space dynamics
Set-indexed OU	$$\mathbb{E}[X_UX_V]=\sigma' e^{-\lambda m(U\triangle V)}$$	index by sets

In the multiplicative Lévy formulation, the state enters the noise coefficient itself; in the Langevin formulation, the linear drift matrix is retained but the driver need only have stationary increments; in the Volterra formulation, exponential memory is replaced by a general kernel; in the delay formulation, the state is a segment in $C([-\tau,0];\mathbb{R}^n)$; and in the set-indexed formulation, “time” is replaced by an indexing collection of sets with covariance controlled by the symmetric-difference measure (Kevei, 2016, Voutilainen et al., 2019, Stein et al., 2021, Lv, 13 Aug 2025, Balança et al., 2012).

This suggests that the term is best understood as a class label rather than a single canonical model.

2. Lévy-driven and multiplicative generalized OU dynamics

A central definition in probability theory is the one-dimensional GOU driven by a two-dimensional Lévy input,

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$0

where $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$1 is a bivariate Lévy process with Lévy–Khintchine triplet $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$2. This process is the unique strong solution, and its explicit representation is given through the Doléans–Dade stochastic exponential:

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$3

with

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$4

If $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$5, one may write $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$6 and obtain the familiar exponential-functional form

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$7

Lévy-driven OU appears as the specialization $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$8, and the classical OU as the further specialization in which $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$9 is Brownian (Kevei, 2016).

The generator makes the affine structure explicit:

$$\operatorname{Var}[X_t]=\frac{D}{\gamma}\big(1-e^{-2\gamma t}\big),$$0

This formula is the basis for Foster–Lyapunov drift inequalities and for the asymptotic theory of invariant measures (Kevei, 2016).

The same multiplicative architecture persists under regime switching. In the Markov-modulated setting, the process is embedded in continuous time through a bivariate Markov-additive process, and in the positive case has the explicit pathwise form

$$\operatorname{Var}[X_t]=\frac{D}{\gamma}\big(1-e^{-2\gamma t}\big),$$1

Strict stationarity is characterized by convergence of an exponential functional of the dual Markov-additive process, and the stationary law is the law of $$\operatorname{Var}[X_t]=\frac{D}{\gamma}\big(1-e^{-2\gamma t}\big),$$2. This extension is used in a Markov-modulated risk model with stochastic investment, where ruin probabilities are expressed through the distribution of an exponential functional of a Markov-additive process (Behme et al., 2020).

A narrower but influential inference-oriented specialization sets $$\operatorname{Var}[X_t]=\frac{D}{\gamma}\big(1-e^{-2\gamma t}\big),$$3, so that

$$\operatorname{Var}[X_t]=\frac{D}{\gamma}\big(1-e^{-2\gamma t}\big),$$4

In that form, low-frequency observations of the stationary law are linked to the Mellin transform of the exponential functional $$\operatorname{Var}[X_t]=\frac{D}{\gamma}\big(1-e^{-2\gamma t}\big),$$5, which enables recovery of the Lévy triplet of $$\operatorname{Var}[X_t]=\frac{D}{\gamma}\big(1-e^{-2\gamma t}\big),$$6 (Belomestny et al., 2015).

Tractable Lévy subclasses include additive OU models

$$\operatorname{Var}[X_t]=\frac{D}{\gamma}\big(1-e^{-2\gamma t}\big),$$7

driven by compound Poisson, gamma-related, or double-exponential jump inputs. For gamma and bilateral gamma stationary laws, the background driving Lévy process is compound Poisson with exponential or double-exponential jump distribution, and exact transition laws can be written through self-decomposable $$\operatorname{Var}[X_t]=\frac{D}{\gamma}\big(1-e^{-2\gamma t}\big),$$8-remainders and Polya mixtures of Erlang laws (Zhou et al., 2016, Petroni et al., 2020).

3. Memory kernels, fractionalization, and Gaussian generalizations

Another major branch replaces the exponential OU kernel by a general memory kernel. The resulting “GOU type process” is defined by

$$\operatorname{Var}[X_t]=\frac{D}{\gamma}\big(1-e^{-2\gamma t}\big),$$9

where $$P_{\mathrm{st}}(x)=\sqrt{\frac{\gamma}{2\pi D}}\exp\!\left[-\frac{\gamma(x-\mu)^2}{2D}\right].$$0 satisfies

$$P_{\mathrm{st}}(x)=\sqrt{\frac{\gamma}{2\pi D}}\exp\!\left[-\frac{\gamma(x-\mu)^2}{2D}\right].$$1

If $$P_{\mathrm{st}}(x)=\sqrt{\frac{\gamma}{2\pi D}}\exp\!\left[-\frac{\gamma(x-\mu)^2}{2D}\right].$$2 with total mass $$P_{\mathrm{st}}(x)=\sqrt{\frac{\gamma}{2\pi D}}\exp\!\left[-\frac{\gamma(x-\mu)^2}{2D}\right].$$3, then $$P_{\mathrm{st}}(x)=\sqrt{\frac{\gamma}{2\pi D}}\exp\!\left[-\frac{\gamma(x-\mu)^2}{2D}\right].$$4 and the classical OU kernel is recovered. When $$P_{\mathrm{st}}(x)=\sqrt{\frac{\gamma}{2\pi D}}\exp\!\left[-\frac{\gamma(x-\mu)^2}{2D}\right].$$5 is absolutely continuous, $$P_{\mathrm{st}}(x)=\sqrt{\frac{\gamma}{2\pi D}}\exp\!\left[-\frac{\gamma(x-\mu)^2}{2D}\right].$$6 solves the Sturm-type ODE

$$P_{\mathrm{st}}(x)=\sqrt{\frac{\gamma}{2\pi D}}\exp\!\left[-\frac{\gamma(x-\mu)^2}{2D}\right].$$7

Special kernels include cosine, quadratic, Airy, Hill, and Mathieu types (Stein et al., 2021).

A common misconception is that any linear mean-reverting Volterra model remains “OU-like” in the Markov and stationary senses. In the Brownian-noise framework above, that is false: time-stationarity holds if and only if $$P_{\mathrm{st}}(x)=\sqrt{\frac{\gamma}{2\pi D}}\exp\!\left[-\frac{\gamma(x-\mu)^2}{2D}\right].$$8 for all $$P_{\mathrm{st}}(x)=\sqrt{\frac{\gamma}{2\pi D}}\exp\!\left[-\frac{\gamma(x-\mu)^2}{2D}\right].$$9 and $dV_t=V_{t-}\,dU_t+dL_t$0, while the Markov property holds if and only if $dV_t=V_{t-}\,dU_t+dL_t$1 for some $dV_t=V_{t-}\,dU_t+dL_t$2. Non-exponential kernels therefore produce genuinely non-Markov and, in general, non-stationary extensions (Stein et al., 2021).

Fractional and multifractal Gaussian extensions modify the noise rather than the drift kernel. The fractional OU is written formally as

$dV_t=V_{t-}\,dU_t+dL_t$3

and the multifractal fractional OU further replaces the Gaussian input by a causal random measure weighted by Gaussian multiplicative chaos:

$dV_t=V_{t-}\,dU_t+dL_t$4

For each $dV_t=V_{t-}\,dU_t+dL_t$5, the stationary causal solution is

$dV_t=V_{t-}\,dU_t+dL_t$6

Its covariance coincides with that of the fractional OU, while higher-order increments exhibit multifractal scaling,

$dV_t=V_{t-}\,dU_t+dL_t$7

under the admissibility condition $dV_t=V_{t-}\,dU_t+dL_t$8 (Chevillard et al., 2020).

A different Gaussian generalization keeps the linear SDE

$dV_t=V_{t-}\,dU_t+dL_t$9

but allows $dU_t=-\Theta U_t\,dt+dG_t$0 to be a centered Gaussian process whose covariance satisfies

$dU_t=-\Theta U_t\,dt+dG_t$1

with $dU_t=-\Theta U_t\,dt+dG_t$2. This includes subfractional Brownian motion and bifractional Brownian motion and yields a non-Markovian Gaussian OU with consistent continuous-time estimators for the drift parameter (Chen et al., 2020).

The generalized Fokker–Planck approach proceeds yet differently. In Fourier space, the time derivative is replaced by

$dU_t=-\Theta U_t\,dt+dG_t$3

leading to

$dU_t=-\Theta U_t\,dt+dG_t$4

The solution remains Gaussian,

$dU_t=-\Theta U_t\,dt+dG_t$5

and can be written as a deterministic time-change of the classical OU via

$dU_t=-\Theta U_t\,dt+dG_t$6

The stationary Gaussian version built in that paper has covariance proportional to $dU_t=-\Theta U_t\,dt+dG_t$7 and exhibits long-range dependence for $dU_t=-\Theta U_t\,dt+dG_t$8 (Beghin, 2018).

A further operator-theoretic construction iterates the OU operator itself. Writing

$dU_t=-\Theta U_t\,dt+dG_t$9

one defines

$V(t)=V_0\rho(t)+\int_0^t \rho(t-s)\,dL(s)$0

obtaining an OU$V(t)=V_0\rho(t)+\int_0^t \rho(t-s)\,dL(s)$1 process. The iteration collapses to a linear combination of basic OU components,

$V(t)=V_0\rho(t)+\int_0^t \rho(t-s)\,dL(s)$2

with

$V(t)=V_0\rho(t)+\int_0^t \rho(t-s)\,dL(s)$3

For $V(t)=V_0\rho(t)+\int_0^t \rho(t-s)\,dL(s)$4, the sampled covariance structure is in general different from that of a discrete AR$V(t)=V_0\rho(t)+\int_0^t \rho(t-s)\,dL(s)$5 process (Arratia et al., 2012).

4. Geometry, dimension, indexing, and state-space structure

Generalization also occurs through the state space itself. On the two-dimensional comb, motion along the $V(t)=V_0\rho(t)+\int_0^t \rho(t-s)\,dL(s)$6-backbone is allowed only at $V(t)=V_0\rho(t)+\int_0^t \rho(t-s)\,dL(s)$7, while motion along the $V(t)=V_0\rho(t)+\int_0^t \rho(t-s)\,dL(s)$8-fingers is diffusive. The OU dynamics on this geometry are governed by

$V(t)=V_0\rho(t)+\int_0^t \rho(t-s)\,dL(s)$9

where

$dx(t)=L(x_t)\,dt+\Sigma\,dB(t)$0

Eliminating the finger coordinate yields a fractional backbone equation of order $dx(t)=L(x_t)\,dt+\Sigma\,dB(t)$1,

$dx(t)=L(x_t)\,dt+\Sigma\,dB(t)$2

and the first two moments involve Mittag–Leffler functions. With stochastic resetting,

$dx(t)=L(x_t)\,dt+\Sigma\,dB(t)$3

the backbone dynamics become tempered fractional, and the stationary state is a non-equilibrium stationary state shaped jointly by mean reversion, geometric trapping, and resetting (Trajanovski et al., 2023).

In finite dimension, the multidimensional Langevin equation

$dx(t)=L(x_t)\,dt+\Sigma\,dB(t)$4

with $dx(t)=L(x_t)\,dt+\Sigma\,dB(t)$5 and a vector-valued driver $dx(t)=L(x_t)\,dt+\Sigma\,dB(t)$6 with stationary increments provides another canonical extension. A stationary solution exists when the spectrum of $dx(t)=L(x_t)\,dt+\Sigma\,dB(t)$7 lies in the open right half-plane and

$dx(t)=L(x_t)\,dt+\Sigma\,dB(t)$8

The same work proves that every stationary $dx(t)=L(x_t)\,dt+\Sigma\,dB(t)$9-valued continuous-time process admits such a representation for a unique $$\mathbb{E}[X_UX_V]=\sigma' e^{-\lambda m(U\triangle V)}$$0, and that $$\mathbb{E}[X_UX_V]=\sigma' e^{-\lambda m(U\triangle V)}$$1 satisfies a continuous-time algebraic Riccati equation involving the stationary cross-covariance matrix $$\mathbb{E}[X_UX_V]=\sigma' e^{-\lambda m(U\triangle V)}$$2 (Voutilainen et al., 2019).

A specialized bivariate generalization is the elliptical OU process. In complex notation $$\mathbb{E}[X_UX_V]=\sigma' e^{-\lambda m(U\triangle V)}$$3, it obeys

$$\mathbb{E}[X_UX_V]=\sigma' e^{-\lambda m(U\triangle V)}$$4

The conjugate term $$\mathbb{E}[X_UX_V]=\sigma' e^{-\lambda m(U\triangle V)}$$5 generates anisotropy and ellipticity, while $$\mathbb{E}[X_UX_V]=\sigma' e^{-\lambda m(U\triangle V)}$$6 is the stationarity condition. The process remains Gaussian and Markov and admits a frequency-domain formulation through $$\mathbb{E}[X_UX_V]=\sigma' e^{-\lambda m(U\triangle V)}$$7 and the complementary spectrum $$\mathbb{E}[X_UX_V]=\sigma' e^{-\lambda m(U\triangle V)}$$8 (Sykulski et al., 2020).

In infinite dimension, the natural extension is the Hilbert-space OU equation

$$\mathbb{E}[X_UX_V]=\sigma' e^{-\lambda m(U\triangle V)}$$9

where $C([-\tau,0];\mathbb{R}^n)$0 generates a $C([-\tau,0];\mathbb{R}^n)$1-semigroup $C([-\tau,0];\mathbb{R}^n)$2 on a separable Hilbert space $C([-\tau,0];\mathbb{R}^n)$3, and $C([-\tau,0];\mathbb{R}^n)$4 is an $C([-\tau,0];\mathbb{R}^n)$5-valued Lévy process. The mild solution is

$C([-\tau,0];\mathbb{R}^n)$6

Its transition semigroup has generalized Mehler form

$C([-\tau,0];\mathbb{R}^n)$7

and invariant measures are characterized by operator self-decomposability (Applebaum, 2014).

The indexing set may also be generalized. For a set-indexed OU process on an indexing collection $C([-\tau,0];\mathbb{R}^n)$8, the stationary covariance is

$C([-\tau,0];\mathbb{R}^n)$9

The process is completely characterized by $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$00-monotone inner- and outer-continuity, $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$01-stationarity, and the $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$02-Markov property. In the multiparameter rectangle case, it admits an integral representation against a Brownian sheet (Balança et al., 2012).

A history-valued extension arises for affine stochastic functional differential equations,

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$03

where $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$04. If the characteristic spectral bound $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$05, the generalized OU random variable is the stationary stochastic convolution

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$06

and it is a tempered random equilibrium attracting all pull-back trajectories in the history norm (Lv, 13 Aug 2025).

5. Invariant laws, ergodicity, duality, resetting, and path functionals

For multiplicative Lévy GOU processes, ergodicity is governed by the explicit generator and Foster–Lyapunov inequalities. Sufficient conditions are formulated in terms of integrability of the joint Lévy measure and negativity of a “log-drift” or $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$07-drift expression. Depending on the Lyapunov function, one obtains ergodicity, polynomial subexponential convergence,

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$08

stretched-exponential convergence,

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$09

or exponential ergodicity in weighted norm,

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$10

If $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$11, regeneration at $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$12 implies

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$13

In the Brownian special case, positive recurrence occurs when $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$14 (Kevei, 2016).

For infinite-dimensional OU processes, invariant laws are characterized by weak convergence of the stochastic-convolution laws $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$15 and by operator self-decomposability. If $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$16 is exponentially stable, sufficient conditions include

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$17

together with existence of the appropriate drift limit. In the Gaussian case, the invariant covariance solves

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$18

in the operator sense (Applebaum, 2014).

Duality theory shows that the class is closed under Siegmund duality and under reversal of the stochastic flow. For

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$19

with $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$20, the dual is again a GOU,

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$21

where

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$22

The stationary distribution of the dual provides information about the hitting time of zero of the original process: under the assumptions of the hitting-time theorem,

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$23

If $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$24 is a subordinator, then $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$25 (Behme et al., 7 Apr 2025).

Resetting produces a different stationary mechanism. In one dimension, the OU process with resetting to $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$26 satisfies

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$27

with renewal representation

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$28

Its non-equilibrium stationary state is

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$29

and when $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$30 it exhibits a derivative discontinuity at the reset point and skewness toward $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$31. On the comb, resetting tempers the fractional memory but does not restore equilibrium (Trajanovski et al., 2023).

Path functionals of generalized OU processes have been analyzed in several directions. For the positive stationary genOU

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$32

the integrated increments

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$33

are regularly varying, and the asymptotic theory of extremes, sample autocovariances, and point-process convergence is governed by Kesten–Goldie mechanisms and extremal clustering. This framework includes continuous-time versions of $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$34, $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$35, the $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$36, and Nelson’s diffusion model (Fasen, 2010).

Occupation times admit explicit Laplace transforms in some tractable jump models. For the OU process driven by a double-exponential jump diffusion, the transform

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$37

is reduced to a piecewise form with coefficients determined by a $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$38 linear system built from exit-functionals. The smooth-fit identity $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$39 remains valid and the method extends to more general Lévy drivers (Zhou et al., 2016).

6. Inference, simulation, and application domains

Because “generalized OU” denotes several model classes, the inferential toolkit is similarly heterogeneous. In the multidimensional Langevin setting, $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$40 is recovered from the stationary cross-covariance via the continuous-time algebraic Riccati equation

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$41

and the estimator $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$42 is defined as the positive semidefinite solution of the empirical CARE. Consistency follows from uniform convergence of the empirical cross-covariance, and Gaussian CLTs hold, for example, when the driver has independent fractional Brownian components with $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$43 (Voutilainen et al., 2019).

For OU processes driven by a general centered Gaussian noise, two continuous-observation estimators of the drift were analyzed:

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$44

Both are strongly consistent and asymptotically normal, with Berry–Esseen bounds whose rate is piecewise in $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$45. The proofs rely on the Hilbert space $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$46 associated with $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$47, second-chaos expansions, and Malliavin–Stein estimates (Chen et al., 2020).

In the multiplicative Lévy setting with deterministic $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$48, low-frequency stationary observations support a Mellin-transform method. If

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$49

then

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$50

with $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$51 the Laplace exponent of $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$52. The empirical Mellin transform

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$53

yields

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$54

from which weighted least-squares estimators of $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$55 and $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$56 and a regularized Fourier inversion estimator of the Lévy density are obtained. The resulting rates are minimax optimal up to logarithmic factors (Belomestny et al., 2015).

Simulation methods are equally model-specific. For gamma and bilateral gamma OU processes, exact generation exploits the self-decomposable $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$57-remainder with $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$58. In the gamma case, one draws $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$59, samples an Erlang law with rate $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$60, and updates

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$61

For integer shape, $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$62 becomes binomial. The bilateral gamma case is simulated by differencing two independent gamma remainders. These algorithms are exact and, in the numerical study, are significantly faster than shot-noise alternatives (Petroni et al., 2020).

The multifractal fractional OU process is simulated on a periodic interval by a DFT-based scheme: the causal kernel $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$63 is built in Fourier space, the regularized log field $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$64 is generated by convolution, the chaos weight $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$65 is formed, and the process is obtained through another Fourier-space convolution. The same paper proves convergence of second-order structure and higher-order multifractal scaling (Chevillard et al., 2020).

For the cosine GOU type process,

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$66

maximum-likelihood estimation is based on the exact AR$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$67-type discretization

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$68

while Bayesian inference uses power-series representations of Fox’s $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$69-function to approximate stable densities. The same study applies four goodness-of-fit tests—KS, AD, MKS, and McCulloch—to choose between Gaussian and $\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$70-stable specifications, with applications to Apple stock-market data and cardiovascular mortality in Los Angeles County (Stein et al., 2021).

Frequency-domain inference is especially natural in oscillatory generalizations. The elliptical OU process is estimated by Whittle likelihood using the spectral matrix

$\mathbb{E}[X_t]=\mu+(x_0-\mu)e^{-\gamma t},$71

which supports semiparametric estimation of damping, frequency, eccentricity, and orientation. The method was applied to Earth’s polar motion, where the annual oscillation appeared significantly elliptical while the Chandler wobble did not provide clear evidence of strong ellipticity (Sykulski et al., 2020).

The application range is correspondingly broad: branched or porous transport, search with restart, confined dynamics in complex media, stochastic volatility and market inactivity, risk theory with stochastic investment, turbulence, geophysics, coupled oscillatory systems, set-indexed random fields, and stochastic delay systems with random equilibria (Trajanovski et al., 2023, Behme et al., 2020, Chevillard et al., 2020, Lv, 13 Aug 2025).