<p>The renormalization will only matter if the inverse of betamix is regularized. I am not that familiar with how the Johnson algorithm behaves for bad problems to know if regularization helps in practice, or does matter for a multisecant method.</p>
<div class="gmail_quote">On Apr 12, 2011 1:41 AM, "Stefano de Gironcoli" <<a href="mailto:degironc@sissa.it">degironc@sissa.it</a>> wrote:<br type="attribution">> It has been there for some time, didn't seem to have any relevant effect.<br>
> do you have evidence of the contrary?<br>> stefano<br>> <br>> On 04/11/2011 05:25 PM, Laurence Marks wrote:<br>>> Thanks.<br>>><br>>> Another question if I may. From the looks of PW/mix_rho.f90 you do not<br>
>> use the weights in the Johnson paper, just a straight inverse of<br>>> betamix (what would be called Y^T Y in the optimization literature) at<br>>> lines 289-295? Have you considered a regularization, e.g. adding after<br>
>> line 278 something like<br>>><br>>> betamix(i,i)=betamix(i,i) + 1.D-7/iter_used<br>>><br>>> which is about right for your Al (001) example. The regularization<br>>> term (e.g. PRB 78, 075114 (2008)) is a bit empirical, although I might<br>
>> be able to change this.<br>>><br>>> This is safe, so long as mix_rho.f90 is only used for mixing densities<br>>> during the scf iterations -- is it used elsewhere?<br>>><br>>> On Mon, Apr 11, 2011 at 5:12 AM, Stefano de Gironcoli<<a href="mailto:degironc@sissa.it">degironc@sissa.it</a>> wrote:<br>
>>> in my previous post<br>>>><br>>>> reminder -> remainder<br>>>><br>>>> stefano<br>>>><br>>>> On 04/11/2011 11:57 AM, Stefano de Gironcoli wrote:<br>>>>> dear Laurence Marks<br>
>>>><br>>>>> thank you for contributing this patch for bfgs !<br>>>>>> A quick question. In the ion optimization it looks like you are<br>>>>>> starting from some iterpolation of the new density (i.e. "NEW-OLD<br>
>>>>> atomic charge density approx. for the potential"), what is it?<br>>>>> if i remember correctly the charge density of the new positions is<br>>>>> written<br>>>>> as the superposition of atomic charges plus a reminder which is computed<br>
>>>> as the scf rho minus the superposition of atomic charges at the old<br>>>>> positions.<br>>>>><br>>>>> rho_trial_new = rho_atomic_newR + (rho_scf_oldR - rho_atomic_oldR)<br>
>>>><br>>>>> this is done for the first ionic iteration and assumes that the main<br>>>>> change in the density is due to rigid displacement of atomic-like<br>>>>> contributions.<br>
>>>><br>>>>> At subsequent iterations the reminder (rho_scf_oldR-rho_atomic_oldR)<br>>>>> can be extrapolated on the basis of its change in the previous couple<br>>>>> of iterations<br>
>>>><br>>>>> Stefano de Gironcoli<br>>>>><br>>>>> On 04/11/2011 01:11 AM, Laurence Marks wrote:<br>>>>>> For completeness, added proper comments.<br>>>>>><br>
>>>>> On Sun, Apr 10, 2011 at 4:13 PM, Laurence Marks<br>>>>>> <<a href="mailto:L-marks@northwestern.edu">L-marks@northwestern.edu</a>> wrote:<br>>>>>>> A very minor bug that you probably known: some of the routines in<br>
>>>>>> S3DE/iotk/src have lines such as "# 1 "iotk_write_interf.spp" ". Most<br>>>>>>> sensible preprocessors will ignore these and just give warnings.<br>>>>>>><br>
>>>>>> A more serious bug. Your bfgs code does not have curvature failure<br>>>>>>> conditions trapped. Not to get too technical here (contact me offline<br>>>>>>> if needed), unless one is close to the minimum bfgs fails unless this<br>
>>>>>> is done. The failure is well documented, less well known, as is the<br>>>>>>> change needed (at least the standard form). I am attaching a modified<br>>>>>>> version with the standard fix. It gives a slightly lower energy with<br>
>>>>>> smaller forces in about the same number of iterations -- due to<br>>>>>>> numerical limitations I cannot compare exactly with your reference<br>>>>>>> directory. I am a newbie with this code so there could be other<br>
>>>>>> repercussions of this change if it is used for something except<br>>>>>>> optimizing the atomic positions, so perhaps a few tests are<br>>>>>>> appropriate for harder problems.<br>
>>>>>><br>>>>>>> A quick question. In the ion optimization it looks like you are<br>>>>>>> starting from some iterpolation of the new density (i.e. "NEW-OLD<br>>>>>>> atomic charge density approx. for the potential"), what is it?<br>
>>>>>><br>>>>>>> Another quick one: line 1766 of install/<a href="http://configure.ac">configure.ac</a> nulls out<br>>>>>>> scalapack_libs and the lines below look like they are special tests,<br>
>>>>>> which seems to be inconsistent with line 150 and standard protocols of<br>>>>>>> letting the user define input variables. (OK, while scalapack is<br>>>>>>> probably only useful for large problems I might want to do some.)<br>
>>>>>><br>>>>>>> --<br>>>>>>> Professor Laurence Marks<br>>>>>>> Department of Materials Science and Engineering<br>>>>>>> MSE Rm 2036 Cook Hall<br>
>>>>>> 2220 N Campus Drive<br>>>>>>> Northwestern University<br>>>>>>> Evanston, IL 60208, USA<br>>>>>>> Tel: (847) 491-3996 Fax: (847) 491-7820<br>>>>>>> email: L-marks at northwestern dot edu<br>
>>>>>> Web: <a href="http://www.numis.northwestern.edu">www.numis.northwestern.edu</a><br>>>>>>> Chair, Commission on Electron Crystallography of IUCR<br>>>>>>> <a href="http://www.numis.northwestern.edu/">www.numis.northwestern.edu/</a><br>
>>>>>> Research is to see what everybody else has seen, and to think what<br>>>>>>> nobody else has thought<br>>>>>>> Albert Szent-Gyorgi<br>>>>>>><br>
>>>>><br>>>>>><br>>>>>> _______________________________________________<br>>>>>> Pw_forum mailing list<br>>>>>> <a href="mailto:Pw_forum@pwscf.org">Pw_forum@pwscf.org</a><br>
>>>>> <a href="http://www.democritos.it/mailman/listinfo/pw_forum">http://www.democritos.it/mailman/listinfo/pw_forum</a><br>>>>><br>>>> _______________________________________________<br>
>>> Pw_forum mailing list<br>>>> <a href="mailto:Pw_forum@pwscf.org">Pw_forum@pwscf.org</a><br>>>> <a href="http://www.democritos.it/mailman/listinfo/pw_forum">http://www.democritos.it/mailman/listinfo/pw_forum</a><br>
>>><br>>><br>>><br>> <br>> _______________________________________________<br>> Pw_forum mailing list<br>> <a href="mailto:Pw_forum@pwscf.org">Pw_forum@pwscf.org</a><br>> <a href="http://www.democritos.it/mailman/listinfo/pw_forum">http://www.democritos.it/mailman/listinfo/pw_forum</a><br>
</div>