The linear solve computed a search direction that was not a descent direction. It determined this by attempting a line search. On Apr 23, 2012 8:23 AM, "Dominik Szczerba" <dominik at itis.ethz.ch> wrote:
> Hi, > > I am implementing both FormFunction and FormJacobian for the SNES > solver. It seems to start off correctly, with the 2 functions being > called one after another in a loop, and with ||Ax-b|| decreasing about > 3 orders of magnitude. But then only the FormFunction is being called > several times, without subsequent calls to FormJacobian and obviously > the residue is not being reduced any further, because there are no > calls to the underlying ksp. If I inspect the solution it looks > qualitatively "on the way" to the expected one, but the values are not > correct, just as if the iterations were prematurely stopped. What sort > of a problem can such a behavior indicate? > > Regards, > Dominik > -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20120423/a10d7ad6/attachment.htm>
