[CPMD-list] PCG MINIMIZE with ROKS MODIFIED GOEDECKER?

[Eneritz Muguruza gonzalez]

Dear CPMD users,

      Some weeks ago I sent an e-mail to the list asking about the
geometry optimization of a small molecule with TDDFT, thank you Juerg
Hutter and Irmgard for your useful answers. When trying the method
proposed by Irmgard, ROKS with the MODIFIED GOEDECKER keyword, I find
some strange behaviours in the wavefunction optimization.By performing
the minimization using different numerical recipes I find the following:

- Using ODIIS 2: it converges very slowly (more than 3000 steps) but
without problems (E=-103.6901 Ha)
 
- INITIALIZE WAVEFUNCTION RANDOM with ODIIS 2 : converges faster without
any problems. (E = -103.6901 Ha). this is the cleanest convergence

- PCG MINIMIZE goes quickly to a lower energy than the other converged
results (E = -103.6969 Ha) but doesn't converge. Changing to ODIIS (both
just after getting the minimum value or after more steps) the energy
increases giving a value E=-103.7081 Ha

- PCG MINIMZE with INITIALIZE WAVEFUNCTION RANDOM does not converged at
all.

So we can see that even if the final result is the same the PCG MINIMIZE
method reaches a lower energy value during the minimization. Here I
attach a graph in which you can see this behaviour.

Thus my question are: Is the PCG MINIMIZE method reliable in this case? 
What is the meaning of the lower energy obtained with the PCG MINIMIZE
method? Are the final results obtained with all the methods converged?


Thank you very much in advance,


          Eneritz
-------------- next part --------------
A non-text attachment was scrubbed...
Name: convergence-anomaly.jpg
Type: image/jpeg
Size: 31524 bytes
Desc: not available
Url : http://cpmd.org/pipermail/cpmd-list/attachments/20050713/611a1baf/attachment.jpg