CCL:G: Solving linear dependence in Gaussian with IOp(3/59)

 Sent to CCL by: Mikael Johansson [mikael.johansson~!~iki.fi]
 Hello Ron and All!
 On Wed, 26 Jan 2011, Ron Bakus rbakus+*+chem.ucsb.edu wrote:
Sometimes in these large systems I encounter situations where there is some near linear dependence and a resulting oscillation in the SCF with no converegence.
The only reliable fix I have found is to use IOp(3/59) (threshold for throwing away eigenvectors of S) and change it from the default 6 to some lower number, usually 5 (I can see that this results in an extra 10 or so [out of a 1000] basis sets being culled from the calculation). I was hoping someone could detail some of the consequences and possible pitfalls of changing the default value.
Not Gaussian specific, as I don't know the details of its IOps: In general, there shouldn't be much of a problem with removing linearly dependent basis functions from the set, as the reason they are linearly dependent (in almost all cases, I believe) is that some other function describes practically the same parts of the density. As long as you are sure that the reason for non-convergence is linear dependency, you should be safe by throwing out some functions. But the reason for the convergence problems could of course be something else, and the "fix" of throwing out diffuse functions might just be a side-effect.
One problem that you could end up with is if you are comparing energies, and then due to the culling of bf's end up with essentially a different basis set for different systems/geometries. You could check how serious this is by changing the threshold for discarding bf's for a similar system without convergence problems, and see how much the total energy differs between different settings. 
 If I've missed something, I'm sure someone will comment :-)
 Have a nice day,
     Mikael J.
     http://www.iki.fi/~mpjohans/