Overdetermined noise compensation without eigendecomposition for frequency estimation

The authors consider the estimation of the frequencies of noisy sinusoids, without eigendecomposition or correlation lags computation. The algorithm based on an iterative noise compensation on the predictors matrix proposed by A. Ferrari et al. (IEEE Trans. vol.ASSP-40, no.5, p.1263-6 of 1992) is extended to the case of higher order 1-step predictors (larger than twice the number of sinusoids), to lower the variance. The procedure is made more efficient by using the split Levinson algorithm. A noise subspace projection operator is derived from the noiseless predictors, which allows a multichannel signal classification (MUSIC)-like frequency estimation, at low computational cost. Simulations and comparisons with the classical root-MUSIC estimator results are presented.<>