If you set a null space, you should make the right hand side consistent. Have you checked the _accuracy_ of the solution in the case where a null space was not used? When the system is singular, the error can be large even with a small residual, but presumably the rest of your algorithm wants a minimum norm solution.
On Thu, Nov 8, 2012 at 2:48 AM, Thomas Witkowski < thomas.witkowski at tu-dresden.de> wrote: > Maybe the following information is also useful to find the problem or/and > to understand what's going on: The inner solver solves with a Laplace > matrix (discretized with free boundaries). So I set the constant nullspace > to the KSP object. This inner solver is used inside the Schur complement > solver of a PCFIELDSPLIT object. The Schur complement solver (KSPPREONLY > and PCSHELL) is also set to have a constant null space. But the vector, on > which the PCSHELL is applied, is not orthogonal to the constant null space. > Can you help me to understand why this is still the case? > > To go back to the problem of the influence of the KSP monitor to the > solution process: When I project out the constant null space before calling > the KSPRICHARDSON with PCHYPRE, the monitor has no influence anymore on the > solution process. > -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20121108/d0c27fc7/attachment.html>
