Adaptive M-estimators for use in structured and unstructured robust covariance estimation

Covariance estimation is necessary in many applications such as source detection in array processing. Unfortunately, the sample covariance estimator is not robust. Here we investigate two broad approaches to robust covariance matrix estimation. The first is a model-free element-wise procedure, while the second is a structured approach based on pre-whitening. Both approaches utilize a robust one-dimensional scale estimator. It is the choice of this scale estimator and its effect on the overall covariance estimator that is the main purpose of this study. An adaptive M-estimator of scale is shown to have several advantages. Depending on the final comparison criterion, its use in a structured or element-wise covariance matrix estimator can lead to improved, robust performance