Deconvolution and Spectral Estimation Using Final Prediction Error.

摘要

Least-squares, zero-lag inverse filters may be used for predictive deconvolution of stationary time series and for obtaining autoregressive or maximum entropy spectral estimates. The greatest problem in finding such an inverse filter is determining the optimum operator length for a given finite length of data. The identical problem of determining the correct order of an autoregressive model for the data has been solved by Akaike, whose final prediction error (FPE) statistic is a minimum for the optimum length model. This minimum FPE criterion may be applied to both single and multiple time series.