Parameter estimation of non-ergodic Ornstein-Uhlenbeck

TL;DR

This paper develops statistical methods for estimating the drift parameter of a non-ergodic Ornstein-Uhlenbeck process driven by a broad class of Gaussian processes, ensuring consistency and asymptotic normality of estimators.

Contribution

It extends parameter estimation techniques to non-ergodic OU processes driven by general Gaussian noises satisfying specific regularity conditions.

Findings

Establishes strong consistency of the estimators.

Proves asymptotic normality of the estimators.

Validates the approach for a wide class of Gaussian processes.

Abstract

In this paper, we consider the statistical inference of the drift parameter $\theta$ of non-ergodic Ornstein-Uhlenbeck~(O-U) process driven by a general Gaussian process $(G_t)_{t\ge 0}$. When $H \in (0, \frac 12) \cup (\frac 12,1) $ the second order mixed partial derivative of $R (t, s) = E [G_t G_s] $ can be decomposed into two parts, one of which coincides with that of fractional Brownian motion (fBm), and the other of which is bounded by $|ts|^{H-1}$. This condition covers a large number of common Gaussian processes such as fBm, sub-fractional Brownian motion and bi-fractional Brownian motion. Under this condition, we verify that $(G_t)_{t\ge 0}$ satisfies the four assumptions in references \cite{El2016}, that is, noise has H\"{o}lder continuous path; the variance of noise is bounded by the power function; the asymptotic variance of the solution $X_T$ in the case of ergodic O-U process $X$ exists and strictly positive as $T \to \infty$; for fixed $s \in [0,T)$, the noise $G_s$ is asymptotically independent of the ergodic solution $X_T$ as $T \to \infty$, thus ensure the strong consistency and the asymptotic distribution of the estimator $\tilde{\theta}_T$ based on continuous observations of $X$. Verify that $(G_t)_{t\ge 0}$ satisfies the assumption in references \cite{Es-Sebaiy2019}, that is, the variance of the increment process $\{ \zeta_{t_i}-\zeta_{t_{i -1}}, i =1,..., n \}$ is bounded by the product of a power function and a negative exponential function, which ensure that $\hat{\theta}_n$ and $\check{\theta}_n $ are strong consistent and the sequences $\sqrt{T_n} (\hat {\theta}_n - \theta)$ and $\sqrt {T_n} (\check {\theta}_n - \theta)$ are tight based on discrete observations of $X$