Re: [eigen] Divide & conquer SVD algorithm

[Gael Guennebaud <gael.guennebaud@xxxxxxxxx>]

On Tue, May 27, 2014 at 5:03 PM, Tristan Sahel <tristan.sahel@xxxxxxxxx> wrote:

Thank you for the quick and developped answers. We discussed our options and decided to start with functionality C and continue with B.

Good choice!

Actually, I also think this is a good strategy to start with the Power iteration method. Thanks to generic programming you should be able to come up with an implementation compatible with both dense and sparse matrices! So B and C for free! I'd also recommend the Inverse-Power method which is variant to extract the lowest eigenvalues. These algorithms are quite simple, but the main challenge here will be to come up with a nice API compatible with Eigen's approach.

The Power method is relevant when you want only a very few eigenvalues. When more are needed (e.g., 20 to 1000), more sophisticated algorithms like the Arnoldi/Lanczos methods are indeed preferred.

Don't miss this book: http://www-users.cs.umn.edu/~saad/eig_book_2ndEd.pdf

Gael.