Go. Glentis et N. Kalouptsidis, EFFICIENT MULTICHANNEL FIR FILTERING USING A SINGLE-STEP VERSATILE ORDER RECURSIVE ALGORITHM, Signal processing, 37(3), 1994, pp. 437-462
In this paper two highly efficient, order recursive algorithms for lea
st squares multichannel FIR filtering and multivariable system identif
ication are developed. Multichannel FIR filters with different number
of delay elements for each input channel are allowed. The first algori
thm uses two terms Levinson type recursions. The later utilizes Schur
type formulas for updating the pertinent parameters, thus being suitab
le for parallel implementation. Multichannel FIR filters are described
by a multi-index [m1, m2,...,m(k)] where mi equals the number of dela
y elements associated with the i-input channel, i = 1, 2,...,k. The no
vel feature of the proposed algorithms is that they employ updates of
the form [m1, m2,...,m(i),...,m(k)] --> [m1, m2,...,m(i) + 1,...,m(k),
]. Therefore, and in contrast to existing methods, they offer the grea
test possible maneuverability in the index space. This flexibility can
be taken into advantage when the true index is not known, except from
being an element of a set. Computationally efficient paths that searc
h for the true index are described. If the true filter order [p1, p2,.
..,p(k)] is known, the filter coefficients are computed at P = (p1 + P
2 + ... + P(k)) steps, by a repetitive application of single step recu
rsions. The computational complexity of the method is O(kP2), while ex
ecution time could be reduced to O(1) or O(P) if the Schur type algori
thm is implemented in a parallel processing environment on a rectangul
ar or on a linear array, respectively. The final filter can be approac
hed by p!/(p1!p2!...p(k)!) distinct order updating paths, each time pa
ssing through different lower dimension filters. This feature can be u
tilized for the efficient determination of the order of a multichannel
process, accelerating the exhaustive searching procedure required by
most of the order determination criteria. Finally, the mean squared er
ror is considered with potential applications to the optimal two-dimen
sional (2-D) FIR filtering and 2-D system identification.