Correlation matrix estimation and order selection for spectrum estimation

This paper presents a novel covariance matrix estimator for frequency estimation in time sequence analysis. A preliminary covariance matrix of size M´ is first calculated by the sample covariance matrix method, and then the final covariance of size M, with M⩽M´ is determined by employing all available correlation information in the preliminary estimate. Generally the new covariance estimator can more effectively utilize temporal correlations among the data and provides more trade-off freedom in order selection. When the orders (sizes) of the covariance matrices are properly selected, this new estimator can obtain a statistically more stable estimate of covariance matrix than the conventional approach