Re: [eigen] reduces svd

[ Thread Index | Date Index | More Archives ]

2010/7/30 Hauke Heibel <hauke.heibel@xxxxxxxxxxxxxx>:
> Hi,
> At the moment we don't support reduced SVD, right? I am wondering how
> difficult it would be to adapt our method to support this.
> The reduced SVD of A (being m-by-n with m >> n) results in a
> decomposition [U, S, V] = svd(A) where U is m-by-n, S is n-by-n and V
> is n-by-n.
> I understand that we don't do this right now because U and V are
> unitary implying "squareness" but I am wondering if it were difficult
> to offer the latter. Maybe this is not even required at all with the
> new implementation.
> This whole thing is referred to by Matlab as "economy size
> decomposition", see here:

One of the benefits of the divide-and-conquer SVD i'm doing (erm...
yes I know :( ) is that it allows to do that and much more. Basically,
it is a single decomposition from which you can compute on demand any
part of U or V or even multiply a matrix by them without computing
them at all. So it is much more powerful than anything the
Givens-based SVD APIs can offer. See my mail "plan 9....". Details
(link to the research paper) are given there if anyone wants to help.
(Also I'll answer any question...) I'm definitely doing it, i just
have actually less time than usual during july-august for ... personal
reasons :-)


> - Hauke

Mail converted by MHonArc 2.6.19+