Nonstationary Ornstein–Uhlenbeck Noise

• Nonstationary Ornstein–Uhlenbeck noise is defined by OU-type dynamics that incorporate time-dependent parameters, leading to explicit time-varying variance and covariance structures.
• Models use mechanisms such as non-exponential memory kernels, stochastic operator-valued volatility, and non-ergodic drift regimes to capture complex dynamic behavior.
• Advanced estimation, simulation schemes, and diagnostic methods have been developed to analyze these processes, with applications in finance, physics, and active-matter systems.

Nonstationary Ornstein–Uhlenbeck noise denotes OU-type stochastic forcing whose law, covariance, or other dependence descriptors vary with calendar time rather than depending only on lags. Across the recent literature, this label covers several distinct mechanisms: non-exponential Volterra memory kernels, stochastic operator-valued volatility, non-ergodic drift regimes, nonstationary or broad-tailed initial conditions, fluctuating damping, and time-dependent parameters. In each case, the classical OU template is retained, but time-translation invariance is lost, so quantities such as variance, autocovariance, codifference, or characteristic functionals become explicitly time-dependent (Benth et al., 2015, Stein et al., 2021, 2207.13355, Eab et al., 2016).

1. Canonical meaning and principal mechanisms

In the classical OU setting with constant damping and additive white noise, the long-time covariance is stationary. By contrast, the papers surveyed here treat models in which either the driver, the damping, the volatility, the memory kernel, or the initialization breaks that stationary structure. The resulting processes are often still mean-reverting in form, but they no longer admit a covariance depending only on lag, and in several cases they admit no invariant law at all (Eab et al., 2016, 2207.13355).

Mechanism	Representative formulation	Nonstationary signature
Non-exponential memory kernel	$V(t)=V_0\rho(t)+\int_0^t \rho(t-s)\,dL(s)$	$\gamma_V(t,t+h)$ depends on $t$
Stochastic operator-valued volatility	$dX(t)=A X(t)\,dt+\Sigma(t)\,dB(t)$	$\mathrm{Cov}(X(t))$ depends on $V(s)$
Non-ergodic drift	$dX_t=\theta X_t\,dt+dG_t,\ \theta>0$	no invariant probability measure
Random damping	$dx/dt=-\mu(t)x(t)+\chi(t)$	mean and covariance depend on $t,s$ separately

A recurring misconception is that nonstationary OU noise must come from nonstationary increments of the driving noise. The literature shows otherwise. In the Hilbert-space volatility model, the Wiener noise is standard and the nonstationarity comes from the time-varying operator-valued volatility $\Sigma(t)$ (Benth et al., 2015). In fluctuating-damping models, the additive forcing can remain white while the random damping alone generates aging and time-dependent covariance (Eab et al., 2016).

2. Volterra kernels, generalized Langevin equations, and the stationarity boundary

A broad framework is the Generalized Ornstein–Uhlenbeck Type process

$\gamma_V(t,t+h)$0

where the deterministic memory kernel $\gamma_V(t,t+h)$1 is the unique solution, whenever it exists, of

$\gamma_V(t,t+h)$2

This is the Volterra moving-average solution of the generalized Langevin equation

$\gamma_V(t,t+h)$3

with memory delegated to the deterministic problem for $\gamma_V(t,t+h)$4. If $\gamma_V(t,t+h)$5, then $\gamma_V(t,t+h)$6 and one recovers the standard OU-type Volterra representation (Stein et al., 2021).

For Gaussian noise $\gamma_V(t,t+h)$7, the process is Gaussian with

$\gamma_V(t,t+h)$8

and

$\gamma_V(t,t+h)$9

The key theorem is sharp: under this Gaussian setup, time-stationarity holds if and only if $t$0 and $t$1; the Markov property holds if and only if $t$2. Any non-exponential kernel breaks the multiplicative Cauchy identity $t$3, hence breaks Markovianity and typically produces time-inhomogeneous covariance (Stein et al., 2021).

This boundary between stationary and nonstationary behavior is illustrated by explicit kernels. The Cosine process uses $t$4, giving periodic memory and

$t$5

The paper also considers the “Quadratic OU-type” kernel $t$6 and an Airy-kernel example; both are nonstationary and non-Markov. For symmetric $t$7-stable drivers with $t$8, covariance is replaced by codifference,

$t$9

and for non-exponential kernels this quantity depends on both $dX(t)=A X(t)\,dt+\Sigma(t)\,dB(t)$0 and $dX(t)=A X(t)\,dt+\Sigma(t)\,dB(t)$1, again signaling nonstationarity (Stein et al., 2021).

A related Gaussian literature considers OU equations driven by noises with nonstationary increments, notably subfractional Brownian motion and bifractional Brownian motion. In the first-kind OU equation $dX(t)=A X(t)\,dt+\Sigma(t)\,dB(t)$2 with $dX(t)=A X(t)\,dt+\Sigma(t)\,dB(t)$3, the stationary construction available for stationary-increment drivers does not carry over. For subfBm $dX(t)=A X(t)\,dt+\Sigma(t)\,dB(t)$4, one has

$dX(t)=A X(t)\,dt+\Sigma(t)\,dB(t)$5

and for bifBm $dX(t)=A X(t)\,dt+\Sigma(t)\,dB(t)$6,

$dX(t)=A X(t)\,dt+\Sigma(t)\,dB(t)$7

Thus a stationary version need not exist even when the variance converges to a finite limit (Es-Sebaiy, 2021).

3. Operator-valued stochastic volatility in Hilbert space

A technically different mechanism arises in the operator-valued extension of Barndorff-Nielsen–Shephard stochastic volatility. Let $dX(t)=A X(t)\,dt+\Sigma(t)\,dB(t)$8 be a separable Hilbert space and $dX(t)=A X(t)\,dt+\Sigma(t)\,dB(t)$9 the Hilbert space of Hilbert–Schmidt operators. The volatility process $\mathrm{Cov}(X(t))$0 solves

$\mathrm{Cov}(X(t))$1

with bounded drift $\mathrm{Cov}(X(t))$2 and an $\mathrm{Cov}(X(t))$3-valued square-integrable Lévy process $\mathrm{Cov}(X(t))$4. The OU state process is then

$\mathrm{Cov}(X(t))$5

where $\mathrm{Cov}(X(t))$6 is an $\mathrm{Cov}(X(t))$7-valued Wiener process with covariance operator $\mathrm{Cov}(X(t))$8, independent of $\mathrm{Cov}(X(t))$9 (Benth et al., 2015).

Positivity of the stochastic volatility is central. If $V(s)$0, $V(s)$1 and $V(s)$2 are self-adjoint, and $V(s)$3 has non-decreasing paths in the Loewner order, then $V(s)$4 remains self-adjoint and non-negative definite, so the square root $V(s)$5 is well defined. The non-decreasing path condition requires that for every $V(s)$6, the real-valued process $V(s)$7 is almost surely non-decreasing; equivalently, $V(s)$8 is almost surely non-negative definite for all $V(s)$9 (Benth et al., 2015).

The volatility process has an explicit affine transform: $dX_t=\theta X_t\,dt+dG_t,\ \theta>0$0 Under a strong commutativity condition

$dX_t=\theta X_t\,dt+dG_t,\ \theta>0$1

which holds in particular when $dX_t=\theta X_t\,dt+dG_t,\ \theta>0$2 commutes with $dX_t=\theta X_t\,dt+dG_t,\ \theta>0$3 for all $dX_t=\theta X_t\,dt+dG_t,\ \theta>0$4, the characteristic functional of $dX_t=\theta X_t\,dt+dG_t,\ \theta>0$5 is exponential-affine in both $dX_t=\theta X_t\,dt+dG_t,\ \theta>0$6 and $dX_t=\theta X_t\,dt+dG_t,\ \theta>0$7 (Benth et al., 2015).

Nonstationarity follows directly from the second-order structure. In the classical OU case with constant $dX_t=\theta X_t\,dt+dG_t,\ \theta>0$8, $dX_t=\theta X_t\,dt+dG_t,\ \theta>0$9 converges to a time-independent solution of a Lyapunov equation. In the volatility-modulated model,

$dx/dt=-\mu(t)x(t)+\chi(t)$0

which is time-dependent through $dx/dt=-\mu(t)x(t)+\chi(t)$1. Even conditionally on the volatility path,

$dx/dt=-\mu(t)x(t)+\chi(t)$2

Moreover, the mean-reversion adjusted returns

$dx/dt=-\mu(t)x(t)+\chi(t)$3

satisfy

$dx/dt=-\mu(t)x(t)+\chi(t)$4

with

$dx/dt=-\mu(t)x(t)+\chi(t)$5

Hence both variance and autocovariance depend on the current state and past of $dx/dt=-\mu(t)x(t)+\chi(t)$6, and in general there is no time-invariant covariance and no classical spectral density (Benth et al., 2015).

4. Non-ergodicity, fluctuating damping, and heavy-tailed relaxation

Another usage of nonstationary OU noise concerns non-ergodic drift. In the model

$dx/dt=-\mu(t)x(t)+\chi(t)$7

with $dx/dt=-\mu(t)x(t)+\chi(t)$8 a centered Gaussian process satisfying a covariance decomposition around the fBm kernel, the solution is

$dx/dt=-\mu(t)x(t)+\chi(t)$9

Here nonstationarity is not merely second-order time-inhomogeneity: the process admits no invariant probability measure and the deterministic flow is exponentially expanding. The regularity and asymptotic behavior depend on the Hurst-like index $t,s$0, but non-ergodicity is driven solely by the sign $t,s$1 (2207.13355).

Fluctuating-damping models generate yet another mechanism. The Langevin equation

$t,s$2

is studied with $t,s$3 or $t,s$4, where $t,s$5 is either dichotomous noise or fractional Gaussian noise. Pure telegraph damping yields a mean

$t,s$6

and covariance growth controlled by

$t,s$7

so the process is nonstationary in mean and variance. In the perturbed case $t,s$8, mean stability holds if $t,s$9, and covariance stationarity in the long-time limit holds if $\Sigma(t)$0. For fractional Gaussian damping, asymptotic stationarity holds for $\Sigma(t)$1, and for $\Sigma(t)$2 when $\Sigma(t)$3; for $\Sigma(t)$4, mean and variance diverge (Eab et al., 2016).

Heavy-tailed initial data can also induce nonstationary relaxation without changing the OU coefficients. In the OU process with white $\Sigma(t)$5-stable noise, a symmetric $\Sigma(t)$6-stable initial density leads to relaxation rates

$\Sigma(t)$7

The terms $\Sigma(t)$8 are spectral, while $\Sigma(t)$9 are non-spectral. If $\gamma_V(t,t+h)$00, the leading non-spectral rate

$\gamma_V(t,t+h)$01

is smaller than the leading spectral rate $\gamma_V(t,t+h)$02 and therefore governs the long-time relaxation of the PDF and observables. This corrects the view that OU relaxation rates are always determined by the Fokker–Planck spectral ladder alone (Thiel et al., 2016).

A related displacement-level construction is the integrated OU process driven by an $\gamma_V(t,t+h)$03-stable Lévy process. Even when the OU velocity admits a stationary regime, its integral is nonstationary. In the short-memory or small-amplitude limit,

$\gamma_V(t,t+h)$04

and for first passage times,

$\gamma_V(t,t+h)$05

This identifies integrated OU noise as a nonstationary colored-noise approximation to jump processes, with convergence in the $\gamma_V(t,t+h)$06 topology rather than the $\gamma_V(t,t+h)$07 topology (Hintze et al., 2012).

5. Estimation, simulation, and diagnostics

The generalized-kernel literature provides explicit simulation schemes. For the Cosine process,

$\gamma_V(t,t+h)$08

where $\gamma_V(t,t+h)$09 are i.i.d. $\gamma_V(t,t+h)$10 with

$\gamma_V(t,t+h)$11

For the quadratic kernel $\gamma_V(t,t+h)$12, the paper gives a time-varying recursion with explicitly specified $\gamma_V(t,t+h)$13-innovations. In the Hilbert-space stochastic-volatility model, a practical choice is the compound Poisson driver with jumps $\gamma_V(t,t+h)$14, followed by mild-form propagation of $\gamma_V(t,t+h)$15, spectral computation of $\gamma_V(t,t+h)$16, and Euler–Maruyama approximation for $\gamma_V(t,t+h)$17 (Stein et al., 2021, Benth et al., 2015).

Inference reflects the specific nonstationarity mechanism. For the Cosine process, the discretized likelihood is built from innovations

$\gamma_V(t,t+h)$18

with density $\gamma_V(t,t+h)$19. Bayesian estimation uses Fox’s $\gamma_V(t,t+h)$20-function power-series approximations to symmetric $\gamma_V(t,t+h)$21-stable densities, together with priors

$\gamma_V(t,t+h)$22

and posterior sampling in JAGS. Model selection between Gaussian and $\gamma_V(t,t+h)$23-stable Cosine processes uses Kolmogorov–Smirnov, Anderson–Darling, modified KS, and McCulloch’s quantile-based statistic (Stein et al., 2021).

For continuous-time OU models driven by general Gaussian noise with the mixed-derivative decomposition

$\gamma_V(t,t+h)$24

Chen and Zhou prove strong consistency and $\gamma_V(t,t+h)$25-asymptotic normality for both the least squares estimator

$\gamma_V(t,t+h)$26

and the second moment estimator, together with Berry–Esseen bounds (Chen et al., 2020). In the non-ergodic regime $\gamma_V(t,t+h)$27, Lu shows that the continuous-time estimator

$\gamma_V(t,t+h)$28

is strongly consistent and satisfies a Cauchy-type limit,

$\gamma_V(t,t+h)$29

while discrete estimators are strongly consistent and $\gamma_V(t,t+h)$30-tight under the stated mesh conditions (2207.13355).

When nonstationarity appears as non-spectral relaxation, a different diagnostic is needed. Thiel, Sokolov, and Postnikov propose estimating the equilibrium level $\gamma_V(t,t+h)$31, forming

$\gamma_V(t,t+h)$32

complexifying via $\gamma_V(t,t+h)$33, and extracting local relaxation rates from a Morlet continuous wavelet transform. This method is designed to separate spectral and non-spectral plateaus in the relaxation dynamics (Thiel et al., 2016).

6. Domain-specific realizations and broader modeling contexts

A prominent application of operator-valued nonstationary OU noise is commodity forward-curve modeling. With $\gamma_V(t,t+h)$34, Filipović’s space on $\gamma_V(t,t+h)$35, and $\gamma_V(t,t+h)$36 in Musiela parameterization, the forward curve is

$\gamma_V(t,t+h)$37

For each $\gamma_V(t,t+h)$38, this admits a Brownian Volterra representation

$\gamma_V(t,t+h)$39

and in the commutative case $\gamma_V(t,t+h)$40. This places stochastic-volatility forward curves in direct relation to ambit fields and to nonstationary volatility patterns consistent with empirical Samuelson effects (Benth et al., 2015).

The tempo-spatial Volterra-type OU framework extends these ideas to space-time fields. The VOU equation

$\gamma_V(t,t+h)$41

admits an explicit solution in terms of the drift-measure resolvent. Strict stationarity can be recovered under the theorem’s integrability and initialization conditions, but nonstationarity appears when the drift kernel $\gamma_V(t,t+h)$42, the propagation kernel $\gamma_V(t,t+h)$43, or the initial input $\gamma_V(t,t+h)$44 breaks time-translation invariance. The second-order structure distinguishes short-range dependence when $\gamma_V(t,t+h)$45 from long-range dependence when $\gamma_V(t,t+h)$46 and has fixed sign. The framework also supplies path-regularity notions such as $\gamma_V(t,t+h)$47-càdlàg and $\gamma_V(t,t+h)$48-càdlàg versions (Pham et al., 2016).

Empirical illustrations in the GOU literature emphasize both heavy tails and seasonality. For Apple stock price log-returns aggregated at 1, 5, 10, and 15 minutes, the Cosine process yields $\gamma_V(t,t+h)$49 estimates $\gamma_V(t,t+h)$50, $\gamma_V(t,t+h)$51, $\gamma_V(t,t+h)$52, and $\gamma_V(t,t+h)$53, while EDF tests reject Gaussianity and do not reject the fitted $\gamma_V(t,t+h)$54-stable models. For weekly cardiovascular mortality in Los Angeles County from 1970 to 1979, using $\gamma_V(t,t+h)$55, the Gaussian Cosine model is favored, but the periodic kernel still implies seasonally nonstationary OU-type noise (Stein et al., 2021).

In active-matter models, nonstationary OU self-propulsion appears when the self-propulsion process $\gamma_V(t,t+h)$56 is not initialized from its stationary law or when its parameters become time-dependent. In the underdamped active OU model with inertia, the mean-squared displacement can exhibit all dynamical exponents between zero and four. After the typical inertial time $\gamma_V(t,t+h)$57, the motion reverts to overdamped behavior except in harmonically confined systems, where the plateau

$\gamma_V(t,t+h)$58

depends explicitly on inertia. With time-dependent mass accumulation, permanent superdiffusion occurs when $\gamma_V(t,t+h)$59, with exponent

$\gamma_V(t,t+h)$60

This shows that even when the OU noise itself is Gaussian and exponentially correlated, nonstationary initial conditions and time-dependent coefficients can dominate the effective long-time transport law (Nguyen et al., 2021).

Taken together, these results show that “nonstationary Ornstein–Uhlenbeck noise” is not a single model class but a family of OU-type constructions characterized by loss of time-translation invariance. The loss may be induced by memory kernels, stochastic volatility, non-ergodic drift, fluctuating damping, heavy-tailed initialization, space-time propagation, or time-dependent parameters. The common analytical theme is that the OU structure remains explicit enough to preserve affine transforms, covariance formulas, asymptotic laws, or simulation recursions, while the stationary covariance paradigm of the classical OU process is deliberately relaxed.