dolfin team mailing list archive

[kent-and@xxxxxxxxx]

> On Mon, Jul 07, 2008 at 09:38:15AM +0100, Garth N. Wells wrote:
>> I've had a look at the code for applying Dirichlet boundary conditions
>> by eliminating constrained dofs, and I think that we need to do it
>> differently. A constrained dof should be indicated by a negative index
>> in the DofMap, and then eliminated during the main assembly loop. We
>> then need to do something for DiscreteFunction so that it can pick up
>> the boundary values, which should be pretty easy.
>
> PETSc handles this (skipping insertion for negative indices).
>
> http://www-unix.mcs.anl.gov/petsc/petsc-as/snapshots/petsc-current/docs/manualpages/Mat/MatSetValues.html
>
> Does Epetra do the same?
>
> --
> Anders


Sorry if I am stressing the wrong points here. I dont mind negative indices.
The way I understand it is, to simplify it to a 2x2 system:

First, after element matrix and vector calculations are done we have

[ a00   a01 ] [x0] = [b0]
[ a10   a11 ] [x1] = [b1]

enforcing x0 = u0 gives

[  1         0 ] [x0] = [u0]
[ a10   a11 ] [x1] = [b1]

then the symmetric variant of this system:

[  1         0 ] [x0] = [u0]
[  0      a11 ] [x1] = [b1 -  a10 u0 ]

Assembling only negative indices in this example
would give the system:

[ a11 ] [x1] = [b1 -  a10 u0 ]

Still a10 and u0 is needed for calculating the right hand side.

Hence, negative indicies only remove the identity part, which
seems harmless enough.

Using only negative indicies will however result in some work during post
processing if one like to see the solution also on the boundary.

Kent