In
many engineering applications it is required to co
mpute the do
minant subspace of a
matrix
A of di
mension
me="mml103">method=retrieve&_udi=B6TYH-4N3GFH7-1&_mathId=mml103&_user=1067359&_cdi=5619&_rdoc=6&_acct=C000050221&_version=1&_userid=10&md5=94002a9231b08902772c4d692cf7d4dc" title="Click to view the MathML source" alt="Click to view the MathML source">m×n, with
me="mml104">method=retrieve&_udi=B6TYH-4N3GFH7-1&_mathId=mml104&_user=1067359&_cdi=5619&_rdoc=6&_acct=C000050221&_version=1&_userid=10&md5=8c9ef96e5b379eb02424d8379e04b591" title="Click to view the MathML source" alt="Click to view the MathML source">mmg src="http://www.sciencedirect.com/scidirimg/entities/2aa2.gif" alt="not double greater-than sign" title="not double greater-than sign" border="0">n. Often the
matrix
A is produced incre
mentally, so all the colu
mns are not available si
multaneously. This proble
m arises, e.g., in i
mage processing, where each colu
mn of the
matrix
A represents an i
mage of a given sequence leading to a singular value deco
mposition-based co
mpression [S. Chandrasekaran, B.S. Manjunath, Y.F. Wang, J. Winkeler, H. Zhang, An eigenspace update algorith
m for i
mage analysis, Graphical Models and I
mage Process. 59 (5) (1997) 321–332]. Further
more, the so-called
proper orthogonal decomposition approxi
mation uses the left do
minant subspace of a
matrix
A where a colu
mn consists of a ti
me instance of the solution of an evolution equation, e.g., the flow field fro
m a fluid dyna
mics si
mulation. Since these flow fields tend to be very large, only a s
mall nu
mber can be stored efficiently during the si
mulation, and therefore an incre
mental approach is useful [P.
Van Dooren, Gra
mian based
model reduction of large-scale dyna
mical syste
ms, in: Nu
merical Analysis 1999, Chap
man & Hall, CRC Press, London, Boca Raton, FL, 2000, pp. 231–247].