We propose a time-varying Wiener filter for nonstationary signal estimation
that is robust in a minimax sense. This robust Wiener filter optimizes wor
st-case performance within novel "p-point" uncertainty classes of nonstatio
nary random processes. Furthermore, it features constant performance within
these uncertainty classes and requires less detailed prior knowledge than
the ordinary time-varying Wiener filter. We also propose a time-frequency f
ormulation that is intuitively appealing since signal subspaces are replace
d by time-frequency regions, and an efficient on-line implementation using
local cosine bases. Our theory is illustrated by numerical simulations and
a real-data example. (C) 2000 The Franklin Institute. Published by Elsevier
Science Ltd. All rights reserved.