[petsc-users] Using KSPLSQR without explicitly computing U(transpose)U

[Jed Brown]

On Sun, Jan 23, 2011 at 03:39, Gaurish Telang <gaurish108 at gmail.com> wrote:

> But for this I had to compute U(transpose)U explicitly. The algorithm
> mentioned in the paper does NOT involve an explicit computation of
> U(transpose)U.
>
> Is there a way to avoid the explicit and expensive computation of
> U(transpose)U and use KSPLSQR?  The details on this page
>

The need for explicit U^T U is for algebraic preconditioners (like
incomplete factorization).  You can try using cheaper approximations to this
product or use a PCShell if you want to avoid explicitly forming the
product.  Chances are that unpreconditioned LSQR will converge too slowly.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20110123/ae89f6f5/attachment.htm>