Maximum entropy spectral analysis of artificial sinusoidal signals
M. Vellante, Umberto Villante · Journal of Geophysical Research Atmospheres · 1984
We applied the Burg's algorithm to the spectral analysis of a signal that is the sum of 11 sinusoids as intermixed with different noise levels. We paid particular attention to those features that for single sinusoids represent the crucial point of this technique and found that, for a multicomponent signal, the frequency shiftings are smaller than for single sinusoids and the splitting phenomena tend to occur at much higher orders of the prediction error filter. For a noise power smaller than 17% of the power of the lowest amplitude component, all the expected peaks were easily identified and most of the power was concentrated in very narrow ranges of frequency. The behavior of the final prediction error might provide useful suggestions only for higher noise levels and the choice of the limits for the application of the Akaike's criterion is ambiguous and strongly dependent on the noise level. An analysis of individual signal components does not confirm the results of previous investigations in that the splitting hierarchy appears independent on the wave parameters. We also examined the effects of the progressive addition of new components and found that it progressively leads to smaller frequency shifts and higher orders of appearance of splitting phenomena.