On Sun, 14 Aug 2005, Michael Povolotsky wrote:
>> I assume these "global" DoFs are coupled to every other DoF in the
>> mesh, right? In that case:
>>
>> Michael, Would it be reasonable to pre-eliminate them from the matrix?
>> You could just build a std::vector or two for each DoF that way, solve
>> for just the global DoFs, and use them to modify the right hand sides
>> for the local DoFs.
>
> Hello Roy,
> I see one weak point in your suggestion. Due to the fact that "global" DOFs
> are coupled
> to almost every other DOF, after the pre-elimination of them the system
> matrix becomes very dense. I wonder if this method can be realized
> effectively.
> What do you think?
I think you're probably right. I made this suggestion based on having
seen an interesting solver algorithm for such a "sparse except for a
couple dense rows" matrix, but the problem was 1-D and so the sparse
part was actually a narrow band. I don't know if there's anything
similar you can do for more general sparse matrices.
---
Roy
|