[SciPy-dev] Tri-diagonal LAPACK Routines - Shall I interface them?

Simon Clift ssclift@gmail....
Mon Dec 7 18:50:28 CST 2009


On Saturday 05 December 2009 04:42:44 Benny Malengier wrote:
> 2009/12/5 Simon Clift <ssclift@gmail.com>:
> I do the same but with 3 components and a moving interface boundary.
> So then it is banded + an entry in the last column... that is the background 
of my request for having banded/tridiag as
> type of sparse matrices.

Ah ha... sorry, I used to T.A. a 4th-year sparse linear algebra course on a 
regular basis, the pedagogical habit dies hard on this subject. :)

> The banded matrix
> implementations don't allow for this as they are a fixed number of
> lower and upper diagonals, so you need to use csr sparse matrix also
> here, no? 

Ah, yes, in that case I'd do my own "sparse bands" structure.  My current 
problem is embarrassingly well conditioned, so as long as I can do ILU(0) and 
matrix-vector multiply the structure is adequate for Krylov sub-space 
solutions.  I don't need to add fill from LU factorization.

> I don't think a tridiag solver will be able to beat an adaptive time
> stepping BDF scheme. 

Ah, I require positive coefficients.  That requirement kind of snookers using 
anything higher order in time or space, as I have learned the hard way.

-- 
1129 Ibbetson Lane
Mississauga, Ontario
L5C 1K9       Canada


More information about the SciPy-Dev mailing list