On a simple estimate of correlations of stationary random sequences

Hurt, Jan

Applications of Mathematics, Tome 18 (1973) no. 3, pp. 176-187.

Résumé

Suppose that $\{X_t\}_{t\in T}$ is a stationary Gaussian discrete random process where $T$ is the set of integers. Assume $EX_t=0,\ t\in T$, and denote $Z_t=sign X_t,\ T_{tj}=Z_tZ_{t+j}$ for $j$ natural. It is shown that $ET_{tj}=2\ arcsin\ \rho_j/\pi$ so that the quantities $T_{tj}$ may be used to estimate the correlation function $\{\rho_j\}_{j \in N}$. (Here $\rho_j$ denotes the correlation between $X_t$ and $X_{t+j}$.) Further, the formula for $cov(T_{0j},T_{kj})$ in terms of $\rho$'s is given. Asymptotic properties of the mean $\bar{T}_j=\sum^{N-j}_{t=1} T_{tj}/(N-j)$ are studied under the asumption that the spectral density of $\{X_t\}_{t\in T}$ is nonzero and possesses bounded second derivative. Particularly, the derived results hold for stationary autoregressive Gaussian random sequences which is the most important case in practice. It is proved that $\bar{T}_j}$ is asymptotically normally distributed and that the sequence $\{T_{tj}\}_{t\in T}$ satisfies the law of large numbers. Finally, some numerical examples and Monte-Carlo studies are given.

Détail

DOI : 10.21136/AM.1973.103468

Classification : 62E20, 62M10