[molpro-user] No convergence in Boys Localization after 200 iterations

[Susi Lehtola]

On Mon, 21 Oct 2013 13:12:52 +0200
Muammar El Khatib <muammarelkhatib at gmail.com> wrote:
> Dear *.*,
> 
> 
> I'm trying to do orbital localization in a conjugated molecule using
> the Boys method. In my calculation, I turned off the symmetry by
> specifying so using the NOSYM keyword as suggested in § 32.9.2 in the
> manual. I also used the NOORIENT option in the geometry block. When
> reading, I found that a poor localization is likely to happen when
> you have a strongly conjugated system (§ 32.9.4) which is my case.

(clip)

> The output prints the following message:
> 
> ?NO CONVERGENCE IN BOYS LOCALIZATION AFTER 200 ITERATIONS. SYMMETRY=
> 1  DR= 0.4497D-08
> 
> In spite of the convergence problem, MOLPRO gives me a set of
> localized orbitals that, as said in the manual, are poorly localized
> (no convergence in the process).
> 
> What would you suggest me to try?. I checked if there existed some
> card to change the number of iterations, but I think that if in 200
> it didn't do it so it's not a good idea to increase it. Then, there
> is the option of localization thresholds but I'm not sure that it'll
> make the localization to converge or what values are sensibles to do
> it.

In my recent experience, Boys localization can be pretty hard to do with
first-order methods, because the minimum is often not well defined. 200
iterations is not a very large number, you might have to increase it to
2000 iterations or so. But it still might not converge, especially if
you use a diffuse basis set.

> PS. I tried using the Pipek-Mezey criterion and the calculation
> converged, but I'm interested to know how to troubleshoot this
> problem in Boys. This is because, even the Boys localization didn't
> converge, the orbitals seem better localized.

Pipek-Mezey usually converges much faster than the Boys criterion, so
this is expected. Molpro also seems to have a second-order localization
method implemented, which can help in tough cases. But this only seems
to have been implemented for Pipek-Mezey, which usually doesn't have
convergence problems in the occupied space. You may need it in the
virtual space, though.

PS. I didn't find any reference to the second-order method in the
molpro manual. I guess it's the method by Leonard and Luken [Theor.
Chim. Acta 62, 107 (1982)]; a citation should really be added..
-- 
---------------------------------------------------------------
Mr. Susi Lehtola, PhD             Postdoctoral researcher
susi.lehtola at alumni.helsinki.fi   Department of Applied Physics
http://www.helsinki.fi/~jzlehtol  Aalto University
                                  Finland
---------------------------------------------------------------
Susi Lehtola, FT                  Tutkijatohtori
susi.lehtola at alumni.helsinki.fi   Fysiikan laitos
http://www.helsinki.fi/~jzlehtol  Aalto-yliopisto
---------------------------------------------------------------