Re: solvetoep

[David Bateman]

Gorazd Brumen wrote:
> Hi David,
> 
> I am willing to have a look at it and work on it and I think I
> can provide sth. within the next days.
> 
> Here a couple of questions:
> 
> a) What will be the means of communication? Should I post on the
> mailing list a a patch against the cvs tree?

It makes sense to keep the discussion on maintainers, to allow others to
give inputs.

> 
> b) I am not as wizardy as you are about this code, so it will take
> me a couple of days at least to get it running (or as in Blow: "I can do
> it, but it will take me at least a year").

Fine with me, my time is very very limited at the moment, and my
intention in whipping up a quick partially working patch was to allow
someone else (you) to work on it, with some ideas of what code needs to
go where.

> c) As I can see, you havent made any changes to the cvs tree yet
> (I dont know exactly how this thing goes), so, should I start from
> your patch against the cvs and work from there or is your
> patch just a reference to start from?

This is test patch, incomplete and incompletely working. No way I would
recommend committing it in its current state. In any case John reviews
all patches before they are committed, and the decision is up to him.

The patch I sent is already a pretty good starting point and I believe
all of the infrastructure for full toeplitz matrices is in place. Where
things need fixing is in toepsolve functions in the dMatrix.cc and
CMatrix.cc files, which is basically a copy of your code slightly
adapted. This is where your effort should go.. There are at least two
fixes to make; detection of singularities with an appropriate fast
return and the info flag set so that the solver falls through to
lssolve, and the reversal return values..

One other needed change is to have the priority of toeplitz detection
above that of positive definite matrices, in the MatrixType
constructors. After that the patch could probably be considered for
inclusion as it is clearly superior to the current situation. To ease
its inclusion, some test code like

/*

%!test
%! a = toeplitz(1:5,2:6);
%! A = matrix_type(a,'full');
%! assert(a\ones(5,1),A\ones(5,1),1e-10)

*/

should be added below the toepsolve functions. Some additional tests
that just this might be a good idea..

The step after that would be the sparse matrix equivalent of this patch,
but that can wait.

D.

> 
> Gorazd
>