Ergodic Fractional Ornstein-Uhlenbeck Process

• Ergodic fOU processes are stationary solutions to Langevin-type SDEs driven by fractional Brownian motion, characterized by mean reversion and self-similarity.
• The process is defined by rigorous spectral representations and covariance structures, accommodating both univariate and multivariate models with long-range dependence.
• Parameter estimation techniques, including generalized moments and maximum likelihood, are bolstered by Malliavin calculus and CLTs to ensure precise inference.

An ergodic fractional Ornstein-Uhlenbeck (fOU) process is a stationary solution to the Langevin-type stochastic differential equation driven by fractional Brownian motion (fBm), exhibiting mean-reverting and self-similar properties. The ergodic fOU process generalizes the classical Ornstein-Uhlenbeck process by permitting noise with long-range or short-range dependence parameterized by the Hurst index. Recent advances rigorously clarify existence, uniqueness, covariance and mixing properties, invariant measures, and parameter estimation methodologies in both uni- and multivariate settings, considering discrete and continuous sampling, as well as the role of regularity in the underlying Hurst parameter.

1. Mathematical Formulation and Stationary Solutions

Let $B^H = (B^H_t)_{t \in \mathbb{R}}$ be a two-sided fractional Brownian motion with Hurst parameter $H \in (0,1)$, with covariance

$$\mathbb{E}[B^H_t B^H_s] = \frac{1}{2}(|t|^{2H} + |s|^{2H} - |t-s|^{2H}).$$

The canonical (univariate) fOU SDE is

$dX_t = -\theta X_t \,dt + \sigma \,dB^H_t,$

with drift $\theta > 0$ and volatility $\sigma > 0$. For $H > 0$, there exists a unique pathwise solution possessing a strictly stationary version

$$Y_t = \sigma \int_{-\infty}^t e^{-\theta (t-s)}\,dB^H_s,$$

which is mean-zero Gaussian.

Multivariate generalization involves a $d$-variate process $X = (X_t)_{t \in \mathbb{R}} \in \mathbb{R}^d$ solving

$dX_t = A X_t dt + \Sigma dB^H_t,$

where $A$ is a $d\times d$ real matrix (with eigenvalues having positive real part), $\Sigma$ is diagonal, and each $B^{H_i}$ is a marginal fBm, possibly with different Hurst indices $H_i$ and cross-covariance parameters $\rho_{ij}, \eta_{ij}$. The stationary solution for $i=1,\dots,d$ is

$$X^i_t = \nu_i \int_{-\infty}^t e^{-\alpha_i (t-s)}\,dB^{H_i}_s.$$

2. Covariance Structure and Regularity

The stationary variance and autocovariance functions admit spectral representations, e.g.,

$$\mathbb{E}[Y_0^2] = \sigma^2 \frac{\Gamma(2H+1)\, \sin(\pi H)}{2\pi} \int_{-\infty}^\infty |x|^{1-2H}/(\theta^2 + x^2)\,dx.$$

The autocovariance $\mathbb{E}[Y_0 Y_h]$ has the same spectral kernel but with $e^{i x h}$. For each multivariate coordinate,

$$r_{ii}(\tau) = \nu_i^2\,\frac{\Gamma(2H_i+1)\, \sin(\pi H_i)}{2\pi} \int_\mathbb{R} e^{i\tau x} |x|^{1-2H_i}/(\alpha_i^2 + x^2)\,dx.$$

Cross-covariance for $i \neq j$ is fully characterized by $H_{ij}$, $\rho_{ij}$, $\eta_{ij}$ and involves integral kernels $I_{ij}(s)$.

For small $s$, path regularity satisfies: $$r_{ii}(s) = \mathrm{Var}(X^i_0) - \frac{\nu_i^2}{2} |s|^{2H_i} + o(|s|^{2H_i}),$$ showing $(2H_i)$-Hölder continuity; cross-covariance differences scale as $|s|^{H_{ij}}$.

3. Ergodicity and Mixing

The process is ergodic provided the drift is strictly contractive ($\theta > 0$ or $\mathrm{Re}(\mathrm{eigs}(A)) > 0$). Covariance decay is exponential for univariate and second kind cases (Azmoodeh et al., 2013), or algebraic for general fOU with $|r_{ij}(s)| = O(s^{H_i+H_j-2})$ in the multivariate case (Dugo et al., 2024). Ergodicity means time averages converge almost surely or in probability to their stationary expectations, i.e.,

$$\lim_{T\to\infty} \frac{1}{T} \int_0^T f(X_s)\,ds = \mathbb{E}[f(X_0)],$$

for each bounded integrable $f$. Exponential mixing rates (OU of second kind) yield

$$|c(t)| \leq C e^{-\alpha t}, \quad\text{implying} \quad P\left(\left|\frac{1}{T} \int_0^T f(U_t)\,dt - \mathbb{E}[f(U_0)]\right| > \epsilon\right) = O(e^{-CT}).$$

For fOU processes, ergodic means and trajectories are almost surely jointly Hölder continuous in $H$ (Haress et al., 2022), enabling parameter estimation with explicit regularity control.

4. Inference and Asymptotic Theory

Parameter estimation is typically performed via generalized method of moments, ergodic averages, or maximum likelihood.

• Generalized Moment Approach: Estimate parameters by matching empirical and theoretical moments computed from stationary correlations at one or more time-lags, e.g.,

$$f(\theta,H,\sigma) = (\mathbb{E}[Y_0^2], \mathbb{E}[Y_0 Y_h], \mathbb{E}[Y_0 Y_{2h}]),$$

and solving $$f(\hat\theta_n, \hat H_n, \hat \sigma_n) = (\hat\eta_n, \hat\eta_{h,n}, \hat\eta_{2h,n})$$, where sample moments are observed at fixed step $h$. Asymptotic normality is established via Malliavin calculus: centered quadratic forms are represented as double Wiener-Itô integrals, and CLTs follow from the fourth-moment theorem (Haress et al., 2020, Dugo et al., 2024).

• Maximum Likelihood Estimation: For continuous observation, MLE for $\theta$ is

$$\hat\theta_T = -\frac{\int_0^T X_t\,dX_t}{\int_0^T X_t^2\,dt},$$

which is $\sqrt{T}$-consistent and asymptotically normal, achieving minimax efficiency. The Local Asymptotic Normality (LAN) property is rigorously established with Gaussian log-likelihood expansion, and Fisher information is

$I(\theta) = \Gamma(2H+1)\theta^{-2H}.$

The drift parameter $\alpha$ and level parameter $\mu$ in fOU with $H < 1/2$ have differing rates, i.e., $\sqrt{T}$ for $\alpha$ and $T^{1-H}$ for $\mu$ (Chiba et al., 2022). For $H > 1/2$, the usual $\sqrt{T}$ rate applies, and LAN holds with explicit forms (Liu et al., 2015).

• Multivariate Estimation: For cross-correlation, two types of estimators appear:
  • Low-frequency: Based on fixed lag covariances, explicit combinations of products over lags $0$ or $\pm s$.
  • High-frequency: Based on increments, normalized by $n \Delta_n^H$.

Consistency and CLTs for estimators depend critically on $H$. For $H < 3/2$, limit laws are Gaussian; for $H > 3/2$, limiting distributions are non-Gaussian and involve quadratic functionals in the Wiener chaos (Dugo et al., 2024).

5. Regularity with Respect to Hurst Parameter

Recent work establishes pathwise and ergodic-means regularity in the Hurst parameter $H$, with Hölder continuity uniformly in time (Haress et al., 2022). Invariant measures $\mu_H$ are $O(|H-H'|^{1-\varepsilon})$-close in Wasserstein or bounded-Lipschitz distance. Multiparameter Garsia–Rodemich–Rumsey lemmas, precise variance estimates, and combinatorial Gaussian moment expansions are used to upgrade $L^2$ bounds to almost sure continuity, providing robust tools for statistical estimation and sensitivity analysis.

An application is ergodic estimation of $H$ from discretely observed paths, exploiting the stable dependence of empirical ergodic averages and invariant measures on $H$.

6. The Fractional Ornstein-Uhlenbeck Process of Second Kind

The fOU process of the second kind is driven by $dY_t^{(1)} := \int_0^t e^{-s}\,dB_{a_s}$, with $a_t = H e^{t/H}$ and $H > 1/2$ (Azmoodeh et al., 2013). Its unique stationary solution is

$$U_t = \int_{-\infty}^t e^{-\theta(t-u)}\,dY_u^{(1)},$$

with variance

$$V(\theta) = \frac{1}{2}(2H-1)H^{2H}\theta^{-2H} B((\theta-1)H+1, 2H-1),$$

where $B(\cdot,\cdot)$ is the Beta function. Exponential decay in the covariance yields strong mixing and ergodicity. CLTs for parameter estimators, based on discrete data, hold throughout $H \in (1/2,1)$ via Malliavin calculus techniques.

7. Statistical and Spectral Methods

Spectral representations and Wiener chaos expansions are fundamental for analyzing moments and limit theorems. Malliavin–Stein techniques underpin CLTs for quadratic forms, with the fourth-moment theorem as a critical criterion for Gaussian limits. Cumulant methods are invoked for non-Gaussian asymptotics.

For multivariate processes, covariance structure is dictated jointly by marginal parameters $(\nu_i,H_i)$, the drift matrix $A$, and cross-correlation parameters $(\rho_{ij},\eta_{ij})$. The process is strictly stationary and time-reversible under diagonal $A$ and vanishing $\eta_{ij}$.

8. Summary Table: Key Properties Across Model Classes

Process Type	Stationarity & Mixing	Invariant Law
Univariate fOU	Strict, exponential	$$\mathcal{N}\left(0, \Gamma(2H+1)\theta^{-2H}\right)$$
Multivariate fOU	Strict, algebraic	Multivariate Gaussian (explicit covariance)
OU of Second Kind	Strict, exponential	Centered Gaussian, Beta function variance
Ergodic Means in $H$	Hölder in $H$, uniform	Wasserstein-Hölder regularity

All estimators are strongly consistent under ergodic sampling; CLTs and minimax optimality are established via spectral/Malliavin techniques.

9. Context and Impact

The rigorous characterization of ergodic fOU processes, spanning invariant law, regularity, and parameter estimation, enables robust modeling of time series with memory, roughness, or long-range dependence. The methods developed, such as Malliavin calculus, Garsia–Rodemich–Rumsey lemmas, and spectral representations, are influential in statistical inference, model selection, and applications ranging from finance to statistical physics. The sensitivity results in $H$ support stability analysis and adaptive inference. In multivariate and high-frequency settings, explicit covariance and limit theorems allow practical and theoretically sound estimation, including regimes with non-Gaussian asymptotics that arise for strong long-memory effects.