Matrix/Linear Algebra Library?

[Fawzi Mohamed]

On 2008-10-03 11:46:11 +0200, Fawzi Mohamed <fmohamed at mac.com> said:

> On 2008-10-03 10:27:27 +0200, Don Clugston <nospam at nospam.com> said:
> 
>> Bill Baxter wrote:
>>> Oh, and Blade is a very low-level library aimed at optimizing the very
>>> simplest linear algebra expressions like vector+vector.  Last I
>>> checked it didn't contain any matrix ops at all.  And it seems like
>>> Don is frying different fish these days (like Tango BigInt).
>> 
>> It's pretty much obselete now that array ops are in the language. That 
>> always happens to my most interesting, sophisticated code -- Walter 
>> turns it into a one-liner.
>> 
>> BTW BigInt will eventually get to Phobos, too.
>> 
>> 
>>> 
>>> --bb
>>> 
>>> On Fri, Oct 3, 2008 at 6:50 AM, Benji Smith <dlanguage at benjismith.net> wrote:
>>>> Does anyone have a recommendation for a linear algebra library in D?
>>>> 
>>>> I need to do a singular value decomposition on large sparse matrices (approx
>>>> 100,000 x 100,000).
>>>> 
>>>> I've glanced at the Blade library, on DSource, but my linear algebra is a
>>>> bit rusty, and I can't tell whether the library simply lacks the ability to
>>>> do an SVD or whether it can be done, but only by composing other operations.
>>>> 
>>>> Thanks!
>>>> 
>>>> --benji
> 
> I have written dense multidimensional arrays in D.
> They provide some wrapper to various Lapack functions (through Bill's 
> wrappers).
> The library is available at
> 	http://github.com/fawzi
> SVD is also there, but just for *dense* matrixes.
> Bill's wrappers and his dflat also work with sparse matrixes.
> 
> Actually I think that you should think more about your problem
> If you need full SVD U and V matrixes are *full* even if A is sparse.
> 100'000x100'000 double matrix ~74GB, floats is half of it, but you have 
> at least two matrixes.
> 
> Either you have really big resources, or you switch to a cluster and 
> scalapack, or (better) you reformulate your problem so that you either 
> need just a partial decomposition, or you can solve it iteratively.
> 
> Fawzi

Probably if you want least square approximation (that is what you meant 
with LSA?) some kind of direct minimizer is the way to go (but be 
careful the problem might ill conditioned).
Fawzi