<div dir="ltr"><div><div><div><div>Jed,<br><br></div>1) What exactly are the PETSc options for CGNE? Also, would LSQR be worth trying? I am doing all of this through Firedrake, so I hope these things can be done directly through simply providing command line PETSc options :)<br><br></div>2) So i spoke with Matt the other day, and the primary issue I am having with LSFEM is finding a suitable preconditioner for the problematic penalty term in Darcy's equation (i.e., the div-div term). So if I had this:<br><br></div>(u, v) + (u, grad(q)) + (grad(p), v) + (grad(p), grad(q)) + (div(u), (div(v)) = (rho*b, v + grad(q))<br><br></div>If I remove the div-div term, I have a very nice SPD system which could simply be solved with CG/HYPRE. Do you know of any good preconditioning strategies for this type of problem? <br><div><br></div><div>Thanks,<br></div><div>Justin<br></div></div><div class="gmail_extra"><br><div class="gmail_quote">On Thu, Dec 10, 2015 at 9:13 PM, Jed Brown <span dir="ltr"><<a href="mailto:jed@jedbrown.org" target="_blank">jed@jedbrown.org</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><span class="">Justin Chang <<a href="mailto:jychang48@gmail.com">jychang48@gmail.com</a>> writes:<br>
<br>
> So I am wanting to compare the performance of various FEM discretization<br>
> with their respective "best" possible solver/pre conditioner. There<br>
> are saddle-point systems which HDiv formulations like RT0 work, but then<br>
> there are others like LSFEM that are naturally SPD and so the CG solver can<br>
> be used (though finding a good preconditioner is still an open problem).<br>
<br>
</span>LSFEM ensures an SPD system because it's effectively solving the normal<br>
equations. You can use CGNE if you want, but it's not likely to work<br>
well. Note that LS methods give up local conservation, which was the<br>
reason to choose a H(div) formulation in the first place. You can use<br>
compatible spaces/Lagrange multiplies with LS techniques to maintain<br>
conservation, but you'll return to a saddle point problem (with more<br>
degrees of freedom). There's no free lunch.<br>
<span class=""><br>
> I have read and learned that the advantage of LSFEM is that it will always<br>
> give you an SPD system, even for non-linear problems (because what you do<br>
> is linearize the problem first and then minimize/take the Gateaux<br>
> derivative to get the weak form). But after talking to some people and<br>
> reading some stuff online, it seems one could also make non SPD systems SPD<br>
> (hence eliminating what may be the only advantage of LSFEM).<br>
<br>
</span>(Symmetric) saddle point problems have an SPD Schur complement. Schur<br>
complements are dense in general, but some discretization choices (e.g.,<br>
quadrature in BDM) can make the primal block diagonal or block-diagonal,<br>
resulting in a sparse Schur complement. If the Schur complement is<br>
dense, you might be able to approximate it by a sparse matrix. The<br>
quality of such an approximation depends on the physics and the<br>
discretization.<br>
</blockquote></div><br></div>