Rough Ornstein–Uhlenbeck Process

• The rough Ornstein–Uhlenbeck process is a mean-reverting stochastic model enhanced by fractional noise (H<1/2) that produces locally irregular, non-Markovian dynamics.
• It generalizes classical and fractional variants by incorporating multifractal intermittency via Gaussian multiplicative chaos while maintaining stationarity and finite variance.
• Its structure is pivotal in asymptotic statistics, homogenization, and rough volatility modeling, enabling efficient estimation and discrete-time approximations.

Searching arXiv for papers on rough and fractional Ornstein–Uhlenbeck processes to ground the article. A rough Ornstein–Uhlenbeck process is an Ornstein–Uhlenbeck-type mean-reverting process whose local regularity is rough in the sense associated with fractional Brownian motion, typically through a fractional-noise forcing with Hurst parameter $H<1/2$. In the canonical construction, the classical stationary Ornstein–Uhlenbeck process is generalized to the fractional Ornstein–Uhlenbeck process, which preserves mean reversion and stationarity while replacing Markovian white-noise forcing by non-Markovian fractional input; further extensions add multifractal intermittency through Gaussian multiplicative chaos, producing causal, finite-variance, non-Gaussian models (Chevillard et al., 2020). Rough fractional Ornstein–Uhlenbeck processes also appear as central objects in asymptotic statistics, homogenization, and rough-volatility approximation, especially in the ergodic regime under continuous observation and in discrete long-memory score-driven limits (Chiba et al., 2022).

1. Classical origin and the meaning of roughness

The classical Ornstein–Uhlenbeck process is the unique solution of

$dX(t)=-\frac{1}{T}X(t)\,dt+dW(t),$

where $W(t)$ is a Wiener process and $T>0$ is the relaxation timescale. Its stationary version admits the causal representation

$X(t)=\int_{-\infty}^t e^{-(t-s)/T}\,dW(s).$

In this form, the process is explicitly causal and stationary, has finite variance, and has rough sample paths: like Brownian motion, its paths are almost surely nowhere differentiable and are Hölder continuous of any order $<1/2$ (Chevillard et al., 2020).

The expression “rough Ornstein–Uhlenbeck process” is used in a more specific sense when the local regularity is tuned away from the Brownian benchmark. The natural generalization is the fractional Ornstein–Uhlenbeck process, whose local roughness matches that of a fractional Brownian motion with Hurst exponent $H\in(0,1)$. In the genuinely rough regime,

$H<\frac12,$

the process has rougher-than-Brownian sample paths and negatively correlated increments; this is the regime treated in the ergodic statistical theory of rough fractional Ornstein–Uhlenbeck processes (Chiba et al., 2022).

The parameter $H$ separates the three canonical regularity classes. The case $H=1/2$ recovers the classical Ornstein–Uhlenbeck roughness. The case $dX(t)=-\frac{1}{T}X(t)\,dt+dW(t),$0 yields rougher paths. The case $dX(t)=-\frac{1}{T}X(t)\,dt+dW(t),$1 yields smoother paths, though still not differentiable in general (Chevillard et al., 2020).

2. Fractional Ornstein–Uhlenbeck construction

The fractional Ornstein–Uhlenbeck process is formally defined by

$dX(t)=-\frac{1}{T}X(t)\,dt+dW(t),$2

where $dX(t)=-\frac{1}{T}X(t)\,dt+dW(t),$3 denotes the increment of a fractional Brownian motion. Because fractional Brownian motion is not a semimartingale when $dX(t)=-\frac{1}{T}X(t)\,dt+dW(t),$4, the forcing must be regularized at scale $dX(t)=-\frac{1}{T}X(t)\,dt+dW(t),$5 to make the dynamics pathwise meaningful. For the regularized model, the unique stationary solution is

$dX(t)=-\frac{1}{T}X(t)\,dt+dW(t),$6

and the canonical fractional Ornstein–Uhlenbeck process $dX(t)=-\frac{1}{T}X(t)\,dt+dW(t),$7 is obtained in the limit $dX(t)=-\frac{1}{T}X(t)\,dt+dW(t),$8 (Chevillard et al., 2020).

Its short-scale behavior is governed by the increment asymptotic

$dX(t)=-\frac{1}{T}X(t)\,dt+dW(t),$9

This identifies the local Hölder exponent as $W(t)$0, so the fractional Ornstein–Uhlenbeck process and fractional Brownian motion share the same small-scale roughness (Chevillard et al., 2020).

In the ergodic drift-parameterization used in continuous-observation statistics, the stationary process is written as

$W(t)$1

with $W(t)$2. The nonstationary solution started from $W(t)$3 is

$W(t)$4

For $W(t)$5, the process is ergodic and converges exponentially to the stationary one. Its stationary covariance satisfies

$W(t)$6

and

$W(t)$7

a decay rate that plays a central role in the asymptotic analysis of estimation (Chiba et al., 2022).

3. Multifractal fractional Ornstein–Uhlenbeck processes

A further extension introduces intermittency by multiplying the fractional driving noise by a log-normal Gaussian multiplicative chaos factor. The regularized chaos field is

$W(t)$8

where $W(t)$9 is a stationary Gaussian field independent of the white noise $T>0$0. Its limiting covariance is

$T>0$1

where $T>0$2 is bounded, continuous, even, decays at large scales, and satisfies

$T>0$3

This is the standard log-correlated input that produces Gaussian multiplicative chaos (Chevillard et al., 2020).

The regularized multifractal fractional Ornstein–Uhlenbeck process $T>0$4 is the unique stationary solution of

$T>0$5

with driving random measure

$T>0$6

and causal drift-like term

$T>0$7

The stationary solution has the integral form

$T>0$8

and the limit as $T>0$9 is denoted $X(t)=\int_{-\infty}^t e^{-(t-s)/T}\,dW(s).$0, the multifractal fractional Ornstein–Uhlenbeck process (Chevillard et al., 2020).

This construction preserves statistical stationarity and finite variance while changing the higher-order statistics. A key identity is that the covariance of $X(t)=\int_{-\infty}^t e^{-(t-s)/T}\,dW(s).$1 is the same as that of the underlying fractional Ornstein–Uhlenbeck process $X(t)=\int_{-\infty}^t e^{-(t-s)/T}\,dW(s).$2, and in particular

$X(t)=\int_{-\infty}^t e^{-(t-s)/T}\,dW(s).$3

Thus multifractality changes higher-order moments but not the variance (Chevillard et al., 2020).

The model is explicitly causal and non-anticipative, because all kernels are supported on the past and the process is adapted to the underlying filtration. At the same time, it is non-Markovian, since the forcing is a convolution against a long-memory kernel and a correlated chaos field. For integer $X(t)=\int_{-\infty}^t e^{-(t-s)/T}\,dW(s).$4,

$X(t)=\int_{-\infty}^t e^{-(t-s)/T}\,dW(s).$5

and the small-scale increment moments obey the multifractal scaling law

$X(t)=\int_{-\infty}^t e^{-(t-s)/T}\,dW(s).$6

The parameter $X(t)=\int_{-\infty}^t e^{-(t-s)/T}\,dW(s).$7 controls the baseline roughness, while the intermittency parameter $X(t)=\int_{-\infty}^t e^{-(t-s)/T}\,dW(s).$8 lowers the scaling exponents of higher moments and thereby introduces multifractal corrections (Chevillard et al., 2020).

4. Functional limits, homogenization, and rough creation

The fractional Ornstein–Uhlenbeck process is also a fast noise model whose nonlinear functionals generate rough limits. In the scaling

$X(t)=\int_{-\infty}^t e^{-(t-s)/T}\,dW(s).$9

the stationary fast process $<1/2$0 is used to study integrated functionals

$<1/2$1

where the normalization depends on the Hermite rank $<1/2$2 through

$<1/2$3

The limiting behavior splits into two regimes: a Gaussian regime converging to Brownian motion and a non-Gaussian regime converging to Hermite processes. In vector form, the joint limit may contain both types simultaneously, and the Brownian and Hermite blocks are asymptotically independent (Gehringer et al., 2020).

For general centered $<1/2$4, the convergence holds in finite-dimensional distributions; under stronger $<1/2$5 assumptions it holds in the Hölder topology. The limiting Hermite process of rank $<1/2$6 has covariance

$<1/2$7

and sample paths Hölder continuous of every order $<1/2$8. For $<1/2$9, it reduces to fractional Brownian motion (Gehringer et al., 2020).

This is the mechanism behind the paper’s “rough creation” result: a family of random smooth curves converges weakly to a non-Markovian random process with rough sample paths. In the slow–fast equation

$H\in(0,1)$0

the limit may be a Stratonovich SDE when $H\in(0,1)$1, or a Young differential equation driven by a Hermite process when $H\in(0,1)$2. The second-order problem

$H\in(0,1)$3

yields, for $H\in(0,1)$4, a fractional Brownian-driven limit and includes the kinetic fractional Brownian motion model (Gehringer et al., 2020).

A plausible implication is that rough Ornstein–Uhlenbeck noise is not only a stationary model in its own right but also an effective microscopic source for rough non-Markovian macroscopic dynamics.

5. Statistical inference and discrete-time approximation

For the ergodic rough fractional Ornstein–Uhlenbeck process under continuous observation on $H\in(0,1)$5, with

$H\in(0,1)$6

and

$H\in(0,1)$7

the asymptotic estimation theory is markedly nonclassical. The maximum likelihood estimator is asymptotically efficient, the model is LAN under suitable scaling, and the natural rates differ across parameters: $H\in(0,1)$8 and $H\in(0,1)$9 are estimated at rate $H<\frac12,$0, whereas $H<\frac12,$1 is estimated at the faster rate $H<\frac12,$2. In the $H<\frac12,$3-parameterization,

$H<\frac12,$4

and

$H<\frac12,$5

where

$H<\frac12,$6

The roughness $H<\frac12,$7 is therefore not merely a path property; it changes the information geometry of the model and produces a faster rate for estimating the stationary mean level $H<\frac12,$8 (Chiba et al., 2022).

A discrete-time route to rough Ornstein–Uhlenbeck limits is provided by long-memory score-driven models. In the infinite-lag recursion

$H<\frac12,$9

the lag coefficients satisfy the heavy-tail condition

$H$0

and the nearly unstable scaling

$H$1

Under the appropriate normalization, the parameter process converges weakly to the rough fractional diffusion

$H$2

with

$H$3

The short-time singularity of the Mittag–Leffler kernel is the mechanism producing the rough limit (Wu et al., 11 Sep 2025).

When these score-driven recursions are used for volatility, they become discrete-time approximations to rough volatility. The paper gives EGARCH($H$4) and quasi-score-driven constructions whose limits generate rough-volatility models with leverage, and reports simulations for European, Asian, Lookback, and Barrier options. The long-memory convolution is computed using FFT, reducing complexity from

$H$5

to

$H$6

This places rough Ornstein–Uhlenbeck dynamics at the intersection of continuous-time rough models and implementable discrete-time statistical recursions (Wu et al., 11 Sep 2025).

6. Conceptual boundaries and related Ornstein–Uhlenbeck variants

The rough Ornstein–Uhlenbeck family is best understood by contrast with other Ornstein–Uhlenbeck generalizations that alter dependence structure without producing rough sample paths. The following comparison summarizes the three canonical levels in the fractional and multifractal hierarchy.

Model	Driving mechanism	Properties highlighted
OU	White noise	Gaussian, stationary, Markovian, rough like Brownian motion
fOU	Fractional-noise-like forcing	Gaussian, stationary, non-Markovian, roughness $H$7
MfOU	Gaussian multiplicative chaos modulation of fractional causal noise	Stationary, non-Gaussian, non-Markovian, causal, multifractal intermittency

This hierarchy follows the progression

$H$8

with mean reversion retained throughout and increasing richness in roughness, memory, and higher-order scaling (Chevillard et al., 2020).

Two common misconceptions concern models that are nonstandard but not rough in the modern sense. The bounded-variation Ornstein–Uhlenbeck process replaces Brownian motion by an integrated telegraph process, producing trajectories that are piecewise deterministic, continuous, and of bounded variation rather than rough in the Brownian sense; it is therefore a finite-speed, non-rough precursor of the classical Ornstein–Uhlenbeck process (Ratanov, 2020). The well-balanced Lévy-driven Ornstein–Uhlenbeck process

$H$9

is likewise not rough in the sample-path sense: it has continuous sample paths, is locally Lipschitz continuous, and is of finite variation on compacts, even though its dependence structure differs substantially from the classical one (Schnurr et al., 2010).

A broader generalization is the generalized Ornstein–Uhlenbeck type process

$H=1/2$0

with deterministic memory kernel $H=1/2$1. In the Brownian case, time-stationarity and the Markov property both force the exponential kernel $H=1/2$2. Non-exponential kernels therefore produce non-Markovian variants, and stable or Lévy drivers can yield heavy-tailed and irregular trajectories, but the framework itself does not define rough Ornstein–Uhlenbeck processes as a separate class (Stein et al., 2021).

A different mechanism for irregularity appears in the Ornstein–Uhlenbeck process with fluctuating damping,

$H=1/2$3

where the damping is random. When $H=1/2$4 is fractional Gaussian noise, the model becomes non-Markovian and the Hurst index controls memory and irregularity. In the perturbed case $H=1/2$5, the mean is stable for $H=1/2$6, unstable for $H=1/2$7, and critical at $H=1/2$8. This suggests a distinct route to roughness, through multiplicative random damping rather than additive fractional Ornstein–Uhlenbeck forcing (Eab et al., 2016).

Within this landscape, the phrase “rough Ornstein–Uhlenbeck process” is most precisely reserved for the fractional regime in which the Hurst parameter governs local Hölder roughness and, in the rough case $H=1/2$9, yields rougher-than-Brownian stationary mean-reverting dynamics.