What is Wrong with MEM?

Abstract

The maximum-entropy method (MEM) is a typical method, where the power spectral estimator is claimed to be proportional to the power of the transfer function of the autoregression (AR) model filter. The AR power spectral estimator is shown theoretically and with practical examples to be a very bad approximation of the true power spectrum. The estimator loses almost all the intensity (line strength) information of the power spectrum, and there can be very large distortions in the estimated power spectrum. The reason for the malfunction of the MEM is the approximation where the autocorrelation function of any random white noise is assumed to be the Dirac delta function. However, this is not true in practice.