[Pw_forum] BFGS convergence

PRASENJIT GHOSH prasenjit at jncasr.ac.in
Tue Aug 29 13:00:42 CEST 2006


Dear Paolo,

Thanks for the clarification.


>> I'm keeping the bottom 2 layers fixed & relaxing rest of the layers
>> along
>> all directions. I have set forc_conv_thr as 5.0D-4. When the BFGS
>> optimization is complete the total force printed out is 0.001184 Ry/au.,
>> which is higher than the threshold.
>
> this is not the quantity that is checked for convergence, though:
> ---
> forc_conv_thr  REAL ( default = 1.0D-3 )
>                convergence threshold on forces (a.u) for ionic
>                minimization: the convergence criterion is satisfied
>            javascript:sqspellIgnoreAll()    when all components of all
forces are smaller than
>                forc_conv_thr.
> ---

That means grad_thr and  forc_conv_thr are two different quantities. Am I
correct? what is then grad_thr?

Moreover how important is it to satisfy Wolfe's condition? What I
understood is that Wolfe's condition ensures that one always goes
downhill. But under quadratic approx. grad.step < 0.0 already ensures that
one is going downhill
.

With regards,
Prasenjit.




PRASENJIT GHOSH,
Ph. D STUDENT,
THEORETICAL SCIENCES UNIT,
JAWAHARLAL NEHRU CENTRE,
JAKKUR P. O.,
BANGALORE - 560064,
INDIA.

PHONE:+91-80-22082835 (OFFICE)
            +91 9880519401 (MOBILE)



More information about the Pw_forum mailing list