Abstract
•An extension of the matrix Slepian-Bangs formula to elliptical symmetric distributions under the assumption that the arbitrary density generator depends on unknown parameters is given.•This new matrix Slepian-Bangs formula involves a common structure to the existing one with a simple corrective coefficient.•Closed-form expressions of this coefficient are given for Student’s t and generalized Gaussian distributions and are each compared according to different knowledge of the density generator.•This allows us to conclude that for an arbitrary parameterization, the Cramér-Rao bound may be very sensitive to the knowledge of the density generator for super-Gaussian distributions contrary to sub-Gaussian distributions.
This paper mainly deals with an extension of the matrix Slepian-Bangs (SB) formula to elliptical symmetric (ES) distributions under the assumption that the arbitrary density generator depends on unknown parameters, aiming to rigorously quantify and understand the impact of this assumption on ES distributed parametric estimation models. This matrix SB formula is derived in a unified way within the framework of real (RES) and circular (C-CES) or noncircular (NC-CES) complex elliptically symmetric distributions, and then compared to the matrix SB formula obtained with fully known or completely unknown density generators. This new matrix SB formula involves a common structure to the existing one with a simple corrective coefficient. Closed-form expressions of this coefficient are given for Student’s t and generalized Gaussian distributions and are each compared according to different knowledge of the density generator. This allows us to conclude that for an arbitrary parameterization, the Cramér-Rao bound (CRB) may be very sensitive to the knowledge of the density generator for super-Gaussian distributions contrary to sub-Gaussian distributions. Finally, we prove that for the parametrization with an unknown scale factor, the CRB for the estimation of the other parameters of the scatter matrix does not depend on the type of knowledge of the density generator. This latter result remains true for the specific noisy linear mixture data model where the parameter of interest is characterized by the range space of the mixing matrix.