Skip to Main Content
In much of modern radar, sonar, and wireless communication, it seems more reasonable to model "measurement noise" as subspace interference-plus-broadband noise than as colored noise. This observation leads naturally to a variety of detection and estimation problems in the linear statistical model. To solve these problems, one requires oblique pseudo-inverses, oblique projections, and zero-forcing orthogonal projections. The problem is that these operators depend on knowledge of signal and interference subspaces, and this information is often not at hand. More typically, the signal subspace is known, but the interference subspace is unknown. We prove a theorem that allows these operators to be estimated directly from experimental data, without knowledge of the interference subspace. As a byproduct, the theorem shows how signal subspace covariance may be estimated. When the strict identities of the theorem are approximated, then the detectors, estimators, and beamformers of this paper take on the form of adaptive subspace estimators, detectors, and Capon beamformers, all of which are reduced in rank. The fundamental operator turns out to be a certain reduced-rank Wiener filter, which we clarify in the course of our derivations. The results of this paper form a foundation for the rapid adaptation of receivers that are then used for detection and estimation. They may be applied to detection and estimation in radar, sonar, and hyperspectral imaging and to data decoding in multiuser communication receivers.