[Ifeffit] Fitting procedure

[Matt Newville]

HI Abhijeet,

>        I think there is a lot of confusion over the method that i used for fitting, So I should make it clear here.
> In this method I fit the first shell by taking into account only first shell R range i.e. 1.5 - 2.7. Then after getting
> a nice fit for the > first shell, I extract its parameters and make them set in Artemis. Then I include the second
> shell by increasing R range to 1.5 - 3.6. In the same manner I fit the first three shells of copper foil and the
> parameters I got for them are reasonably good.
> The problem of high SO2 only lies with one of my foil data which is not as good as other data. For other good
> data I got SO2 value for second shell within the exceptable limit.
> One more thing I want to say is that we can check the corelation between different parameters of shells by
> making them guess during any step of the fitting e.g. during the fitting of second shell we can make any parameter
> of the first shell to vary and check its corelation with them.
> But if I am counting the first shell data twice, I will include only second shell R range after first shell fit.I think it will
>  not make any difference to results. If yes please tell me the reasons.

> I want to add some more things I have done
> 1) E0 is kept same for all the shells
> 2) All other parameters like SO2 , del r , SS2 are calculated separately for
> different shells.
> So. is there problem with taking different SO2  for different shells. Surely
> del r and SO2 will be different for different shells , If I got it right.
>

I'll assume that you've read all the posts about this.

Since you fit the first shell, then fix these parameters to fit the
second and third shells, why is it that you use an R-range that
includes the first shell?  That's not fitting the second and third
shells, that's fitting the first, second, and third shells, only your
not letting the fit adjust the model for the first shell.  I wouldn't
call this wrong, exactly, but you're not using all you know about the
spectra (that is, that the first shell is well separated from the
second and third, and that you know the R-ranges over which the
different parameters will influence the modeled spectra).      Since
you're choosing variables that can only change a portion of the
spectra (and, as it turns out, the more subtle portion of the
spectra),  the mis-fit to the first shell could easily dominate the
fit to the higher shells.  I think you'd be better off changing the R
range to ignore most of the first shell.

You say the problem with S02 lies only with one data set.  The concern
from the people more experienced in analysis here is not that your
data is noisy, but that having multiple S02 makes a model that is
either non-physical or very hard to understand.   My guess is that
you're varying S02 per shell, but leaving the coordination number N
(path degeneracy) fixed at the nominal values for an fcc solid.  It is
more common (and generally preferred) to use exactly 1 S02 per spectra
(or even the same S02 for a set of spectra on the same edge) and allow
the coordination number to vary.  This is generally easier to
interpret.     Since the amplitude for a path is N*S02, the two
approaches are mathematically equivalent.

> One more thing I want to say is that we can check the corelation between different parameters of shells by
> making them guess during any step of the fitting e.g. during the fitting of second shell we can make any parameter
> of the first shell to vary and check its corelation with them.

Yes, that is right.  This is generally a good thing to do.

Cheers,

--Matt