On Fri, Feb 10, 2012 at 15:58, Dmitry Karpeev <karpeev@mcs.anl.gov> wrote:

It seems to me that the sparsity pattern (and even the type of the matrices) is up to the particular DM type,

so it is, in principle, opaque.  DMDA can control it by manipulating stencil width or widths, while FEM-oriented DM 

types (e.g., libMesh) could control it via the element type, etc. The generic DM interface doesn't have a way to specify 

the sparsity pattern. As it should be, in my opinion.

Yes, it's fine for DM to have a way to allocate a sparser preconditioning matrix. The most common case in practice is MFFD for the true Jacobian and more compact stencil for the preconditioner. It just makes the interface and implementation a bit ugly/confusing to support both. It would be nice to have a more elegant way to state it.