Performance evaluation of certain ARMA models in power spectral density estimation

Abstract

For an autoregressive moving average (p, q) process the performance measures considered are the asymptotic variance of the spectral estimator and the resolution of two closely spaced sinusoids in white noise. Though the AR parameters are mainly responsible for good resolution, it is shown that proper MA parameters are also necessary in some methods. Cadzow's (Indirect) method and the singular value decomposition (SVD) method are considered for comparison. It is found that both methods have approximately the same variance of the PSD estimates in the neighborhood of the frequencies of interest. The SVD method yields a much lower model order than Cadzow's method in which the MA parameters have greater influence on resolution.<>