On Mon, Nov 6, 2017 at 2:11 PM, Bikash Kanungo <[email protected]> wrote:
> Hi,
>
> I would like to access the approximate Jacobian in quasi-Newton solvers
> (e.g., BFGS) to analyze its condition number. I guess for BFGS, where the
> approximate Jacobian is built from outer products of several vectors coming
> from current and previous gradient vectors, the approximate Jacobian might
> not be stored explicitly as a matrix and instead the Jacobian times a
> vector is performed in a matrix-free manner. My reason for believing so is
> the fact that when I try to retrieve the Jacobian matrix using
> SNESGetJacobian and feed it to SLEPc, I incur segmentation fault at
> EPSSetOperators.
>
> If that's the case, is there a way to still get access to the approximate
> Jacobian and feed it to SLEPc for eigen analysis, like with the use of
> MatShell operations?
>
There is no code in SNESQN to apply the Jacobian, just its inverse. You
could wrap up that code in a MatShell and hand it to SLEPc. I am not
sure what you are looking for, but its doable. I would think there are
analytic expressions or at least bounds for the condition number.
Thanks,
Matt
> Thanks,
> Bikash
>
> --
> Bikash S. Kanungo
> PhD Student
> Computational Materials Physics Group
> Mechanical Engineering
> University of Michigan
>
>
--
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which their
experiments lead.
-- Norbert Wiener
https://www.cse.buffalo.edu/~knepley/ <http://www.caam.rice.edu/~mk51/>
