Systems Seminar
A Time-Frequency Approach to Minimax Robust Estimation of Nonstationary
Random Signals
Gerald Matz
Institut fuer Nachrichtentechnik und Hochfrequenztechnik
Technische Universitaet Wien
Abstract
We consider the estimation of nonstationary random signals
contaminated by additive nonstationary random noise. First we review the
time-varying Wiener filter which minimizes the mean square error but
requires perfect knowledge of the second-order statistics of signal and
noise. We then discuss a time-frequency formulation of the time-varying
Wiener filter that is valid for so-called underspread processes and leads
to a physically intuitive and computationally efficient alternative
time-frequency design.
Subsequently, we present minimax robust extensions of the time-varying
Wiener filter, i.e., estimators that minimize worst-case performance within
prescribed uncertainty classes for the nonstationary signal and noise
statistics. These uncertainty classes are well suited to practical
applications where due to estimation or modeling errors only imperfect
prior knowledge is available. A time-frequency reformulation of minimax
robust Wiener filters is then proposed that again simplifies their design
and interpretation.
Specializing the general theory, we finally consider so-called p-point
uncertainty models for which the resulting robust time-varying Wiener
filter has a particularly simple structure and can be implemented
efficiently using the multi-window Gabor expansion.
Time and Place: Monday, Oct. 5, 2:30-3:30 pm in 4610 Engr. Hall.
*** Note different DAY and different TIME. ***