JDQR ?

[David Bateman]

Søren Hauberg wrote:
> Hi All,
>   This mail serves the purpose of show-casing my ignorance.
> I'd like to play a bit around with the "Locally Linear Embedding
> (LLE)" [1], but that code requires 'eigs'. While I can get 'eigs' (and
> already have) on my machine, its license situation is annoying. At the
> LLE page, they say
>
>   "Actually, eigs doesn't really work at all for large problems. You
>    need to use a more sophisticated solver.
>    We have had good results with the JDQR package, from here."
>
> where "here" refers to [2]. So, my question is: does anybody know if the
> JDQR could be used to provide an alternative to 'eigs' that we could
> use. JDQR doesn't come with any license information, but if it's
> suitable for providing 'eigs' than we at least ask the JDQR developers
> to relicense under GPL.
>
> Søren
>
> [1] http://www.cs.toronto.edu/~roweis/lle/
> [2] http://www.math.uu.nl/people/vorst/JDQR.html
>
>
>   
Yes JDQR seems to have the capabilities needed to implement the eigs 
function. In fact it uses an Arnoldi technique to initialize the problem 
in any case and then uses a partial Schur decomposition in the 
iterations rather than continuing with the Arnoldi technique.

The downside is that jdqr is a rather long and messy m-file, with lots 
of functions for iterative sparse solvers that should be in Octave 
independently of jdqr itself. You also should try and get a better 
preconditioner into Octave if you want to go this way. So its not a 
trivial amount of work.. The code was written in 1998 and doesn't seem 
to have been changed since and so getting a license change will also be 
fun.. You'd probably be better trying to get Rice Uni to change the 
license on ARPACK than jdqr as they seems to be fairly open to the idea, 
but hesitate due to amount of work implied with their lawyers to do it.

Cheers
David





-- 
David Bateman                                David.Bateman at motorola.com
Motorola Labs - Paris                        +33 1 69 35 48 04 (Ph) 
Parc Les Algorithmes, Commune de St Aubin    +33 6 72 01 06 33 (Mob) 
91193 Gif-Sur-Yvette FRANCE                  +33 1 69 35 77 01 (Fax) 

The information contained in this communication has been classified as: 

[x] General Business Information 
[ ] Motorola Internal Use Only 
[ ] Motorola Confidential Proprietary