Searching my mail archive for something else turned up this unanswered
question, so I'll reply.

Patrick Farrell <[email protected]> writes:

> Hi,
>
> As I mentioned in a question
>
> (http://fenicsproject.org/qa/321/is-it-possible-to-solve-a-nonlinear-problem-matrix-free)
>
> I'd like to solve a problem where the Jacobian of my nonlinear problem
> is dense, but its action is easy to compute.

This is the setting where classical optimization methods like nonlinear
CG, nonlinear GMRES, and quasi-Newton/BFGS are commonly used.  If your
nonlinear problem is actually the first-order optimality conditions for
an optimization problem, then all the theory applies.  If you don't have
an objective functional or your residual is not the gradient of
something, then I would recommend starting with nonlinear GMRES.  There
are a bunch of examples in PETSc, but it could be as simple as
"-snes_type ngmres".  How to expose this through Dolfin is something for
you all to decide.

If the spectrum of your operator decays rapidly, then the above should
converge well, similar to an unpreconditioned Krylov method.  This is
common for many regularized inverse problems because they have a
spectrum that is spectrally equivalent to a second kind Fredholm
integral operator.  But if you do not have such decay or if you have a
great deal of geometric complexity, you may have to put some effort into
preconditioning.  I'd have to know a lot more about your problem to
suggest preconditioning strategies, but you should try the
unpreconditioned method first.

Attachment: pgpsncYqT_K5q.pgp
Description: PGP signature

_______________________________________________
fenics mailing list
[email protected]
http://fenicsproject.org/mailman/listinfo/fenics

Reply via email to