Time-varying autoregressive adaptive filtering for airborne radar applications

Multivariate time-varying autoregressive models of order m (TVAR(m)) are introduced based on the Dym-Gohberg band extension technique for finite operator-valued matrices. For particular side-looking airborne radar scenarios (based on the KASSPER-II data set) and novel optimal time-varying transmit antenna pattern control, we demonstrate that model mismatch losses associated with relatively small TVAR order m (m ~ 4 to 8), are quite low (1.5 to 3 dB) inside the azimuth-Doppler range of interest, and allow for significant reduction in training sample support.