More on optimality of EVD methods for bearing estimation

A general expression is developed for the variance of bearing estimates of M sources obtained by EVD and the MUSIC algorithm (or an extension of it already introduced by the authors in the case of partially correlated sources). Comparison with a suitable form of the Cramer-Rao bounds demonstrates that optimality is reached only if all signal eigenvalues are much greater than the noise level. The analysis is extended to specific improvements such as spatial averaging or multiple frequency averaging, the effect of which is to avoid the occurrence of very small eigenvalues in the case of correlated sources.<<ETX>>