Re: [ublas] How to solve systems of linear equations?

Subject: Re: [ublas] How to solve systems of linear equations?
From: Thomas Klimpel (Thomas.Klimpel_at_[hidden])
Date: 2010-03-11 17:02:13

Rui Maciel wrote:
> Thomas Klimpel wrote:
> > What has "boost::ublas" done to you that you compare it with TNT?
> 
> I don't understand. What do you mean by that?

To put it straight, libraries like "boost::ublas" or "eigen2" play in a differnt league than "TNT". Even the idea to compare them to TNT is strange, and the conclusion that the documentation of TNT is better than the documentation of "boost::ublas" or "eigen2" is just not comprehensible.

> > (What I hate most about TNT is the choice between a really old stable version
> > 1.26 and an old beta version 3.0.12. If they would at least manage to
> > release a stable 3.0 version.)
> 
> Well, they did put up a decent documentation which also encompasses  a
> hand full of decomposition routines.

You are talking about JAMA here, I guess. Do you know "linpack" and "eispack"? JAMA is based on these instead of "lapack". Did you know that "lapack" is the successor of "linpack" and "eispack"? And now "lapack++" tells you that "TNT" is its successor? Looks like "lapack" and "lapack++" don't have much in common.

But still, this really helped me to understand what you don't like about "boost::ublas".

> > eigen2
> 
> I was also considering eigen along with TNT.

>From what I understand now ("a hand full of decomposition routines", "newbies"...), I would say "eigen" is what you want. But please, check whether the version available through Ubuntu's repositories is really the version that the "eigen" developers recommend (same check for the documentation...). Not that you end up using eigen1 and then start telling the "eigen" developers that TNT is much better than "eigen".

> > and many other (so add
> > blitz++
> 
> No support for sparse matrices.

JAMA doesn't support sparse matrices either, so the sparse matrix support of TNT won't help you much here.

> > There was a
> > really "huge" table in an older version of the documentation for
> "glas"
> > with many, many numerical libraries and some of the features they
> > supported at that point in time. It was probably a wise decision to
> remove
> > this table from the documentation, since it would have been an
> impossible
> > task to keep it up to date.
> 
> This is straying a bit away from the subject but when faced with the
> choice of
> having a documentation that may not be quite up to date and not having
> one at
> all, I believe that no one would choose to not have any documentation.

I'm not sure you understood what I wanted to say. I found the page I was reffering to (see attachment). The "huge" table with many, many numerical libraries is actually just a link on this page:
http://www.netlib.org/utk/people/JackDongarra/la-sw.html

As you see, even so the page was last updated on 4/13/2009, it misses eigen2, for example.
What I wanted to make clear is that there are so many numerical packages out there that anyone of us who would be provoked by your strange enthusiasm for "TNT" to point out that there are better numerical libraries than "TNT" out there would be bound to omit some excellent packages (and so did I, take "PETSc" for example).

Regards,
Thomas