Cramér-Rao Bounds for Blind Multichannel Estimation

Abstract

In some estimation problems, not all the parameters can be identified, which results in singularity of the Fisher Information Matrix (FIM). The Cramér-Rao Bound (CRB), which is the inverse of the FIM, is then not defined. To regularize the estimation problem, one can impose constraints on the parameters and derive the corresponding CRBs. The correspondence between local identifiability and FIM regularity is studied here. Furthermore the number of FIM singularities is shown to be equal to the number of independent constraints necessary to have a well-defined constrained CRB and local identifiability. In general, many sets of constraints can render the parameters identifiable, giving different values for the CRB, that are not always relevant. When the constraints can be chosen, we propose a constrained CRB, the pseudo-inverse of the FIM, which gives, for a minimum number of constraints, the lowest bound on the mean squared estimation error. These results are applied to two approaches to blind FIR multichannel estimation which allow identification of the channel up to a scale or phase factor. These two approaches correspond to deterministic and Gaussian models for the unknown channel inputs. The singularities of the FIMs and local identifiability are studied and the corresponding constrained CRBs are derived and interpreted.