On Thu, Oct 24, 2013 at 1:15 PM, Jed Brown <[email protected]> wrote:
> Matthew Knepley <[email protected]> writes: > > I would argue that Saad's implementation suggestions (like incremental > > QR) are much better than the GCR and justify an independent citation. > > The real difference is that GCR keeps two sets of vectors. It does not > have any "brute-force QR". But GCR allows nonlinear preconditioners and > provides the true residual at each iteration at no extra cost. These > weaknesses were not pointed out in the 1986 GMRES paper. The 1993 > FGMRES paper did not cite the GCR paper, though it has pretty much the > same attributes, minus GCR's ability to produce the true residual. I consider this level of dissection overkill. > >> "In practical implementation it is usually more suitable to replace > >> the Gram-Schmidt algorithm of step 2 by the modified Gram-Schmidt > >> algorithm" > >> > >> If someone uses LGMRES, would we produce a citation only to Baker et al, > >> > > > > Only to Baker. This should be easy since SS would be associated with > GMRES. > > What about CG with the single reduction or with Bill's trick? Does that > tweak mean that Hestenes and Stiefel don't get cited, where as they > would be otherwise? > I would cite H&S. > Who gets cited for PCFIELDSPLIT? I think no one. Breaking stuff into pieces is simply too elementary. If we can attach options to citations, we could possible cite things. > >> or also to Saad & Schultz? What about the BiCG family, containing many > >> more variants that are slight variations on existing methods? Or > >> > > > > We need to build in support for selection with options I think. > > Okay, to do this, we need to extend the interface to include one or more > classification labels. Should those labels be extensible (dynamically > registered) or static (enum)? > Do you have to ask? Matt -- What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. -- Norbert Wiener
