Maximum likelihood autoregressive model parameter estimation with noise corrupted independent snapshots

Abstract

Maximum likelihood autoregressive (AR) model parameter estimation problem with independent snapshots observed under white Gaussian measurement noise is studied. In addition to the AR model parameters, the measurement noise variance is also included among the unknowns of the problem to develop a general solution covering several special cases such as the case of known noise variance, noise-free snapshots, the single snapshot operation etc. The presented solution is based on the expectation-maximization method which is formulated by assigning the noise-free snapshots as the missing data. An approximate version of the suggested method, at a significantly reduced computational load with virtually no loss of performance, has also been developed. Numerical results indicate that the suggested solution brings major performance improvements in terms of estimation accuracy and does not suffer from unstable AR filter estimates unlike some other methods in the literature. The suggested method can be especially useful for small-dimensional multiple-snapshot noisy AR modeling applications such as the clutter power spectrum modeling application in radar signal processing.