<div dir="ltr"><div class="gmail_quote">On Wed, Mar 23, 2011 at 20:07, Matthew Knepley <span dir="ltr"><<a href="mailto:knepley@gmail.com">knepley@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
<div>It looks like this FD derivative did not give any descent directions when factored. Possibly</div><div>the FD Jacobian is too inaccurate, or the factorization was. These both mean your problem</div>
<div>is very ill-conditioned. You should try to reformulate it.</div></blockquote></div><br><div>Yes, it is often possible to reformulate to make the system better scaled. You can also run with -snes_fd_type ds which uses a more robust method of computing the step, sometimes this make a difference.</div>
<div><br></div><div>If you know the sparsity pattern of the Jacobian, you can use coloring to "efficiently" assemble a sparse Jacobian using only residual evaluation. There are many examples that do this.</div></div>