<div dir="ltr"><div class="gmail_extra"><br><div class="gmail_quote">On Sun, Mar 3, 2013 at 10:52 PM, Jie Chen <span dir="ltr"><<a href="mailto:jiechen@mcs.anl.gov" target="_blank">jiechen@mcs.anl.gov</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">Jed, I guess you can stick to Z as eigenvectors or you can allow Z to be anything a user defines.</blockquote></div><br>
If we're working with something user-defined, then we can just use PCMG with Galerkin and a user-provided interpolation matrix. When using sparse storage for the matrix, this also takes advantage of sparsity. It seems like the only place where we would need special support is if the Krylov method needs to populate the deflation space (interpolation matrix) as the iteration goes along (e.g., updating on restart).</div>
</div>