Hi,
I've not yet encountered Manuel's problem on calculations "going crazy
and a huge number of threads enters the processors". This seems like
a (recent?) bug that SCM should be able to track down.

But I would like to offer some advice to those running into "problems"
with linear dependencies with ADF-BAND wrt the fit set. I've gotten
myself into "problems" on a number of occasions when I've allowed
BAND to throw away dependent fit functions, and in all the cases
it has happened when the smallest eigenvalue of the fit overlap
matrix becomes smaller than about 5E-9 to 1E-9. It seems like
the algorithm band uses to handle these problems is not able
to reliably throw away the most dependent fit function(s) in these
cases. The solution is to start "doctoring" the fit set yourself
to remove the dependent fit function(s). This is something you MUST
do when the smallest eigenvalue of the fit overlap matrix gets
down below 1E-11 to 1E-12, but I would recomend doing it in all
the cases.

Whenever I start a set of calculations for a new system with band
I do a (rather tedious) test of the fit set to control the error
in the calculated energy associated with it. In most cases I'll
set the smallest allowed eigenvalue of the fit overlap matrix to
(Dependency Fit=1E-7, Dependency Fit=5E-8, or Dependency Fit=1E-8)
and then start to remove ALL the dependent fit functions myself.
I can understand that most people will be reluctant to spend
the time needed to do so, but I opt to do this everytime to avoid
possible problems later on.

Bottom line: If the smallest eigenvalue of the fit overlap matrix
becomes smaller than say 5E-9 to 1E-9, you should start getting
worried about possible problems with your calculation. If it gets
down to say 1E-11 to 1E-12 you're definitely in trouble and MUST
modify your fit set by hand - band (in its current implementation)
will not be able to reliably remove such dependencies!

I'm afraid this is not very helpful for Manuel, but his query gave
me "the inspiration" to share this with you. It would be nice to hear
if others have similar (or contrary) experiences.

Best regards,
-Roar.

/---------------------------------------------------\
| Roar A. Olsen |
| (r.a.olsen "at" chem "dot" leidenuniv "dot" nl) |
| Leiden Institute of Chemistry |
| Gorleaus Laboratory |
| Leiden University |
| P.O. Box 9502 |
| 2300 RA Leiden |
| The Netherlands |
| |
| ph. +31 71 527 5569 |
| fax. +31 71 527 4488 |
\---------------------------------------------------/
Received on 2003-10-28 12:08:39