Can maximum-likelihood “threshold performance” be improved by random matrix theory tools?

Performance of maximum-likelihood estimation (MLE) is analysed in the so-called threshold region. Here, due to insufficient training sample volume and/or signal-to-noise ratio, the actual MLE performance degrades considerably with respect to the Cramer-Rao bound, because of the onset of severely erroneous estimates ("outliers"). Recently, for a limited number of training samples comparable with the observation (antenna) dimension, an improved (with respect to MLE) G-estimate of covariance matrix eigenvalues and eigenvectors have been derived by Mestre, using tools from the random matrix theory. We use these G-estimates to form the "G-likelihood function" and compare the threshold performance of the conventional ML and G-ML DOA estimation.