Hi,
according to PETSc user manual [ch 4.6 Solving singular systems],
nullspace for
the Krylov Solver should be provided as an array of orthonormal vectors.
I think
it is then a good idea to make the constructor of VectorSpaceBasis check
the
orthogonality. Or is the orthogonality property not necessary?
I played with linear elasticity with no bcs on displacement, which is a
singular
problem and its solution requires eliminating the rigid motions. The
basis vectors
for rotation that I provided at first were describing rotation around
origin but that
was not center of gravity of my body. While these vectors were all in
the nullspace,
the basis was not orthogonal and I suspect that's why the solver would
not converge.
Once I shifted the domain, all was fine. So in my experience
orthogonality is crucial.
I'd be interested if somebody else has different/similar experience. Thanks.
Regards, Miro
_______________________________________________
fenics mailing list
[email protected]
http://fenicsproject.org/mailman/listinfo/fenics
