Improved spectrum performance via a data-adaptive weighted Burg technique

A new method for generating autoregressive (AR) parameters for spectral estimation is introduced and shown to be a data-adaptive weighted Burg (DAB) technique. The respective AR coefficient estimates minimize a weighted measure of the squared forward and backward linear predictive errors with weights that are proportional to the common energy of the forward and backward process realizations and result in a stable filter representation. Spectra are computed comparing the new method to other weighted Burg techniques (e.g., rectangular, Hamming, and 'optimum' parabolic weighting) and demonstrate performance improvements in frequency bias and resolution. No evidence of a line-splitting tendency is exhibited. However, somewhat greater spectral variability is exhibited by the DAB method in processing nonsinusoidal signals. The provision of asymptotically unbiased AR parameter estimates is presented.