A NON-UNITARY EXTENSION TO SPECTRAL ESTIMATION

This paper introduces a new parametric frequency estimation technique, ROCK MUSIC, for uncovering frequency components of a sum of complex exponentials in noise. The ROCK MUSIC method is similar to conventional frequency estimation methods, like MUSIC, that decompose the autocorrelation matrix into signal and noise subspaces. The distinguishing feature of this new method is that it does not use an eigenvector decomposition but a decomposition based on a Reduced Order Correlation Kernel. This new non-unitary basis compresses the signal space, thereby making it much more robust to incorrect subspace partitioning than previous subspace frequency estimation methods. Simulation of the ROCK MUSIC technique indicates that large signal spaces may be represented by just a few basis vectors. The signi cance of this nding is that one does not need to know the correct number of signals present in order to obtain classi cation of the frequencies. Thus the new ROCK MUSIC method resolves a fundamental limitation of conventional subspace frequency estimation techniques. Finally, a signi cant savings is possible, both in terms of computation and sample support, because only a small number of basis vectors are required.