3084 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO. 12, DECEMBER 1997
Subspace Methods for Blind Estimation
of Time-Varying FIR Channels
Michail K. Tsatsanis and Georgios B. Giannakis, Fellow, IEEE
Abstract—Novel linear algorithms are proposed in this paper
for estimating time-varying FIR systems, without resorting to
higher order statistics. The proposed methods are applicable to
systems where each time-varying tap coefficient can be described
(with respect to time) as a linear combination of a finite number
of basis functions. Examples of such channels include almost
periodically varying ones (Fourier Series description) or channels
locally modeled by a truncated Taylor series or by a wavelet
expansion. It is shown that the estimation of the expansion
parameters is equivalent to estimating the second-order param-
eters of an unobservable FIR single-input-many-output (SIMO)
process, which are directly computed (under some assumptions)
from the observation data. By exploiting this equivalence, a
number of different blind subspace methods are applicable, which
have been originally developed in the context of time-invariant
SIMO systems. Identifiability issues are investigated, and some
illustrative simulations are presented.
I. INTRODUCTION
L
INEAR parametric models have found widespread use
as basic tools in the analysis of physical phenomena
and engineering systems. ARMA models have been employed
in the mathematical analysis of such diverse signals and
systems as seismic responses [14], speech recordings [9], and
communication links (e.g., [10] and [17]). When only output
information is available, however, the identification of MA
models (or MA parts of ARMA models) imposes challeng-
ing theoretical questions and/or considerable computational
burden.
A great deal of research effort has focused in the past on the
problem of estimating the parameters
of an MA model
(1)
of order
driven by an i.i.d. sequence when only
measurements of
are available [3], [5], [15], [28]. In
fact, the renewed interest in higher order statistics (HOS) is
to a great extend motivated by their ability to provide output-
only (linear and nonlinear) solutions to the MA identification
problem [15].
Manuscript received April 20, 1996; revised September 11, 1997. This work
was supported by the National Science Foundation Grants NSF-MIP 9424305
and NSF-NCR 9706658. The associate editor coordinating the review of this
paper and approving it for publication was Dr. Zhi Ding.
M. K. Tstasanis is with the Department of Electrical Engineering and
Computer Science, Stevens Institute of Technology, Hoboken, NJ 07030 USA
(e-mail: mtsatsan@stevens-tech.edu).
G. B. Giannais is with the Department of Electrical Engineering, Uni-
versity of Virginia, Charlottesville, VA 22903-2442 USA (e-mail: geor-
gios@virginia.edu).
Publisher Item Identifier S 1053-587X(97)08554-1.
While in many cases the linear time-invariant (TI) model
of (1) may provide satisfactory description of the underlying
system, in certain applications, the time-invariance assumption
is not warranted. For example, in wireless communications,
the multipath environment changes with time, as the mobiles
move. As a consequence, the system parameters
change
with time [26]. Other applications where nonstationary signals
appear include econometrics [18] and speech processing [9],
to name but a few examples.
The situation is even more pronounced when HOS-based
methods are employed. Due to their notoriously slow con-
vergence, these methods require very long data records and,
hence, are prone to a violation of the TI assumption even for
very slowly changing systems. If the TI description of (1)
does not provide a satisfactory approximation to the system’s
behavior, a time-varying (TV) model has to be considered
(2)
where the impulse response
now depends explicitly
on time
.
One way of handling such system variations in the iden-
tification procedure is by employing adaptive algorithms. If
the system variations are slow (when compared with the
algorithm’s convergence time), then the adaptive algorithm
will track the TV system parameters. The popularity of adap-
tive signal processing is an indication of the importance
of TV systems in engineering applications. When output-
only identification is considered, however, the very long
convergence time of adaptive algorithms (e.g., [7]) renders
them impractical for all but the most slowly changing systems.
For this reason, the Godard algorithm [7] (a popular blind-
channel equalization technique) is, in most cases, studied only
for TI channels.
It appears that a basis expansion approach would be the
natural way of modeling the TV parameters
. Prompted
by the general applicability of basis expansion approximations
(e.g., Fourier and Taylor series expansion, polynomial approx-
imation, etc.), we consider in this paper the parameters
to be given by a linear combination of some known sequences.
Hence, the system identification problem is equivalent to
estimating the coefficients of this expansion.
Basis expansion ideas have been used in the past to estimate
TV-AR systems in the context of speech processing [9], [12],
[25]. Similarly, Taylor expansion of the TV parameters were
used in [11] and [18] in the context of economic time series
analysis. However, only I/O cases or AR linear prediction
1053–587X/97$10.00 1997 IEEE