comp.lang.idl-pvwave archive: archive » Re: Sparse matrix algorithms

Home » Public Forums » archive » Re: Sparse matrix algorithms

Show: Today's Messages :: Show Polls :: Message Navigator
E-mail to friend

Re: Sparse matrix algorithms [message #28361 is a reply to message #28281]

Mon, 03 December 2001 06:56

the_cacc
Messages: 104
Registered: October 2001

Senior Member

> did they say when a new edition can be expected?

No, but I didn't ask. Email nr@nr.com and they'll write back
promptly (in my experience).

> ... Some of my matrices are
> extremely skinny, with an extremely large "long" dimension (on
> the order of millions), and I have to compute the matrix
> multiply transpose(A)##A which in the sparse algorithm loops
> over the smaller dimension. Filling this out might be costly
> (timewise) for me, though I'm just speculating.

I think you may be right - storing millions of zeros does defeat
the purpose. In my case, 10^4 zeros is not such a burden. From
what you say it seems you may be attempting to solve Ax = b
by using A^T A x = A^T b. If so, don't forget that matrix algebra
is associative so you can do A^T (Ax) rather than (A^T A)x and
save some CPU cycles.

> ...the sparse
> bi-conjugate gradient solver implemented (since they could rip
> it directly out of NR).
>

I found this to be slower than the conjugate gradient solver, which
you'll have to implement yourself (v. easy by the way) and which
is guaranteed to converge for the problem A^T A x = A^T b.

Ciao.

Report message to a moderator

[Message index]

		Re: Sparse matrix algorithms By: Ralf Flicker on Fri, 30 November 2001 11:28
		Re: Sparse matrix algorithms By: the_cacc on Fri, 30 November 2001 03:54
		Re: Sparse matrix algorithms By: the_cacc on Mon, 03 December 2001 06:56

Previous Topic:	Re: How to know where IDL found the source of one procedure?
Next Topic:	Re: IDL Shapefile Object

-=] Back to Top [=-

[ Syndicate this forum (XML) ] [

] [

]

Current Time: Sun Oct 12 00:56:08 PDT 2025

Total time taken to generate the page: 0.32245 seconds