Tyler's Covariance Matrix Estimator in Elliptical Models with Convex Structure

Covariance matrix estimation is a fundamental problem in the field of statistical signal processing. Many algorithms for detection and inference rely on accurate covariance estimators [1, 2]. The problem is well understood in the Gaussian unstructured case. But becomes significantly harder when the underlying distribution is non-Gaussian, for example in elliptical distributions, and when there is prior knowledge on the structure. In this paper, we propose a unified framework for covariance estimation in elliptical distributions with general convex structure. Over the last years there was a great interest in covariance estimation with known structure. The motivation to these works is that in many modern applications the dimension of the underlying distribution is large and there are not enough samples to estimate it precisely without additional structure hypotheses. The prior information on the structure reduces the number of degrees of freedom in the model and allows accurate estimation with a small number of samples. This is clearly true when the structure is exact, but also when it is approximate due to the well known bias-variance tradeoff.