Linear minimal variance estimation in target detection and location

Abstract In automatic target recognition, the determination of the observationally conditioned average of a specified target property frequently involves severe implementation difficulties. It is thus of interest to consider approximations that are linear in the input pattern, since these are simply implementable. In this paper we examine the domain of validity of such approximations. We show that they are valid in the regions of target parameter space, where either 1. (a) the assumed target is a close match to the input pattern, or 2. (b) the assumed target is weakly correlated with the input pattern and the norm (to be defined later) of the target is weakly dependent on the target parameters.