<div class="gmail_quote">On Mon, Aug 13, 2012 at 9:14 AM, Ataollah Mesgarnejad <span dir="ltr"><<a href="mailto:amesga1@tigers.lsu.edu" target="_blank">amesga1@tigers.lsu.edu</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
You see this is really bad for a case like mine where essentially I am solving a quadratic problem with box constraints (Jac and Constraints are constant during each solve). Is there anything we can do to use this structure shouldn't we be able to get a quadratic convergence ?</blockquote>
</div><br><div>Your problem is not smooth, thus you cannot expect quadratic convergence in a general setting. The active set/reduced space methods (SNESVIRS) change in a discrete way as the bound constraints are activated or deactivate, therefore they do not generally converge quadratically until the constraint set has been identified. You can try semismooth Newton methods (SNESVISS) for a method that has a better chance of converging quadratically.</div>