A comparison of power spectral estimates and applications of the maximum entropy method

Abstract

A new spectral estimate, called the maximum entropy method, is described. This estimate was originated by John Parker Burg for use in seismic wave analysis. In the maximum entropy method the entropy, or information, of a signal is maximized under the constraint that the estimated autocorrelation function of the signal is the Fourier transform of the spectral power density. The spectral estimates are calculated in two ways: (1) by minimization of the error power to obtain the coefficients of the prediction error filter, as suggested by Burg, and (2) by a direct solution of the matrix equation using an algorithm due to Norman Levinson. For comparison a Blackman-Tukey technique, calculated with a Hamming window, is used also. We illustrate these three methods by applying them to a composite signal consisting of four sinusoids of unit amplitude: one each at high and low frequencies and two at moderate frequencies with respect to the Nyquist frequency, to which is added white noise of 0.5 amplitude. Results are shown to indicate that the best correspondence with the input spectrum is provided by the Burg technique. Applications of the maximum entropy method to geomagnetic micropulsations reveal complex multiplet structure in the Pc 4, 5 range. Such structure, not previously resolved by conventional techniques, has been predicted by a recent theory of magnetospheric resonances. In a period range 7 orders of magnitude longer than micropulsation periods, analysis of annual sunspot means shows that the 11-yr band is composed of at least three distinct lines. With each of these lines is associated a harmonic sequence. Long periods of the order of 100 yr also are revealed.