4260 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 57, NO. 11, NOVEMBER 2009
The QS-Householder Sliding Window
Bi-SVD Subspace Tracker
Peter Strobach
Abstract—A fast algorithm for computing the sliding window
Bi-SVD subspace tracker is introduced. This algorithm produces,
in each time step, a dominant rank- SVD subspace approximant
of an rectangular sliding window data matrix. The method
is based on the (orthonormal-square) decomposition. It uses
two row-Householder transformations for updating and one
nonorthogonal Householder transformation for downdating in
each time step. The resulting algorithm is long-term stable and
shows excellent numerical and structural properties, as known
from pure Householder-type algorithms. The dominant com-
plexity is multiplications per time update, which is
also the lower bound in dominant complexity for an algorithm
of this kind. A completely self-contained algorithm summary is
provided and a Fortran subroutine of the algorithm is available for
download from http://webuser.hs-furtwangen.de/~strobach/qsh-
bisvd.for.
Index Terms—Householder, QS-Decomposition, singular
Vectors, singular value decomposition (SVD), sliding Window
Subspace Tracking.
I. INTRODUCTION
T
HE Singular Value Decomposition (SVD) is certainly one
of the most important mathematical tools in signal pro-
cessing. There are many applications, where one wants to up-
date dominant parts of this decomposition, or unitary similar
versions thereof, in time upon the arrival of new data. Suppose
we have given a rank- SVD approximant of an data ma-
trix as follows:
(1)
where is the dominant orthonormal left singular sub-
space basis, is the dominant orthonormal right sin-
gular subspace basis, and is an square-root power ma-
trix that is two-sided orthonormally or (in the complex case)
unitary similar to the diagonal matrix of ordered dominant sin-
gular values of . In this paper, we consider the particular case
of sliding window updating of this dominant SVD subspace ap-
proximant (1) in time upon the arrival of a new data vector
on the top of the matrix as follows:
(2)
Manuscript received January 01, 2009; accepted May 24, 2009. First pub-
lished June 30, 2009; current version published October 14, 2009. The associate
editor coordinating the review of this manuscript and approving it for publica-
tion was Dr. Deniz Erdogmus.
The author is with the AST-Consulting Inc., 94133 Röhrnbach, Germany
(e-mail: peter_strobach@gmx.de).
Digital Object Identifier 10.1109/TSP.2009.2025978
whereas an old snapshot vector is dropped at the bottom
of the matrix, and an updated approximant
(3)
is computed at time step with an overall dominant operations
count of multiplications.
The algorithm of this paper is a straightforward fast imple-
mentation of the sliding window Bi-SVD subspace tracker of
Badeau, Richard, and David [1]. Their method was inspired by
the exponentially windowed bi-iteration SVD subspace tracker
of [2] and the square Hankel SVD subspace tracker of [3]. These
algorithms, in turn, can be traced back to the classical “Treppen-
iterations” of F. L. Bauer [4].
An accelerated variant of a bi-iteration type algorithm using
the so-called “Bi-LS” concept was recently described in [5]. All
these methods can be regarded as members of the larger class of
sequential orthogonal iteration algorithms.
Strobach has recently introduced a unifying theory of fast or-
thogonal iteration subspace tracking [6]. This theory is based
on the (orthonormal-square) decomposition combined with
row-Householder reductions for the effective and efficient up-
dating of such -decompositions. This method marks the top
in the development of subspace tracking algorithms, because
the resulting algorithms are clearly optimal in terms of struc-
ture, numerical accuracy and stability, and overall operations
count. This comes as no surprise since these algorithms are pure
Householder methods and Householder methods are generally
known for their superior properties. A first application is the
SVD subspace tracker of [7] for the growing-window case with
exponential forgetting.
In this paper, we take up the basic sliding window Bi-SVD
of [1] and show how one can handle this algorithm in terms
of the unifying theory of Householder subspace tracking to ar-
rive at a highly competitive algorithm with minimum domi-
nant operations count. In Section II, we summarize the basic
results of [1], which we use as a starting point for our develop-
ment. Section III introduces the key “tricks” on which our algo-
rithm development will be founded. Sections IV and V present
the main steps in the new all-Householder based fast sliding
window Bi-SVD subspace tracker. Section VI discusses aspects
of practical implementation and summarizes the algorithm in a
completely self-contained algorithm summary. A Fortran sub-
routine is available for download. Section VII shows the results
of various computer experiments using this subroutine in order
to visualize the operation of the new algorithm. Section VIII
presents the conclusions.
1053-587X/$26.00 © 2009 IEEE