Frequency superresolution performance of the constrained total least squares method

The constrained total-least-squares (CTLS) method is a constrained quadratic least-squares recipe for solving an overdetermined linear system of equations whose coefficients are perturbed by noise components that are linearly dependent. When a time series consists of a superposition of undamped sinusoids, it is then possible to impose a forward-backward linear prediction structure on the series and solve for the linear prediction coefficients by using the CTLS or the modified forward-backward linear prediction (MFBLP) techniques. Extensive simulations are presented that show the CTLS technique to be more accurate than the MFBLP for estimating the closely spaced frequencies of the sinusoids, especially at low signal-to-noise ratios and when the initial phases of the sinusoids are random