Autoregression and irregular sampling: Spectral estimation

Abstract

We describe how to model an irregularly sampled data sequence as a set of observations from a linear stochastic process, permitting data classification and spectral estimation. The technique is a generalisation of the prediction error approach taken by the Burg and Covariance methods, in that a generalised prediction error energy is minimised with respect to the AR coefficients. Various tests on synthetic data, using line and broad-band spectra, show that the method is robust. In a companion paper we show that the generalised prediction error approach permits filtering or signal separation.