On Thu, Nov 12, 2015 at 4:55 PM, Bruce Ravel <bravel@bnl.gov> wrote:
On 11/10/2015 05:54 PM, Carolyn Carr wrote:
Thanks so much for clarifying that for me! I have been basing my fitting
in R-space off of

XAFS spectroscopy; fundamental principles and data analysis DOI
10.1023/A:1019105310221

which has a section on fitting in R-space over k-space although it is
very possible I misinterpreted its analysis.

Hi Carolyn,

I believe you are referring to the text at the top left of page 150 in that paper.

Those authors say a number of things differently from how I would say them, however, there is one line that is certainly not relevant to how we do the analysis in Artemis and Ifeffit (or Artemis and Larch).  The fitting metric in R space is evaluated using the real and imaginary parts of the complex Fourier transform.  Thus, the bit at the end of that paragraph does not relate to our software.


I agree with all of this...  But on the bottom of the previous page, they do say that they fit the Imaginary and Magnitude portions of chi(R).   My recollection of XDAP is that it is really very similar to Feffit in how it does the fitting.
 
It been a long time (possibly decades!) since I have looked into the possibility of fitting magnitude and phase instead.  My vague memory is that Matt found that using real/imaginary was more stable in Ifeffit.

Yes.  I've tried this several times over the years.   Basically, using the *phase* is definitely unstable, I think basically because of the 2*pi jumps in phase (Ah, another reason to use tau!).   My recollection is that I never saw things getting better with (mag, imag) instead of (real, imag).

I can believe that Diek Koningsberger and Marius Vaarkampe decided that (mag, imag) pairs worked best for them in XDAP, and have no substantives criticisms of any they do for fitting EXAFS data.    But, I also would be concerned that using magnitude would be *more* prone to the problem of finding false minima.   That is, with oscillatory functions, false minima generally mean jumping one period.   For Imag[chi(R)], that is really, really obviously wrong, especially for heavy back-scatters. 

So, I think one has to use either Imag[chi(R)] or Real[chi(R)].  But I could believe that using pairs of (mag, imag) better (or no worse) than using (real, imag).    That's not my recollection, but it's worth revisiting.   It would be easy to make this an option in Larch (actually, I think it *was* an option in early versions of Feffit) to compare using (real, imag), (mag, imag), and even (mag, phase). and test this with something like the Feff85 test suite.  That's probably worth doing....

 
In any case, the point in that paper is not valid for our software because we use oscillatory functions when fitting in k, R, or back-transform k space.
 
I think that helps put my earlier comment into proper context for you.

Wonderful post, though!  It is very pleasant to see a post that doesn't involve someone complaining about something related to the Mac! :)


Agreed!  Thanks Carolyn for having us read more Koningsberger review articles -- that's good for everyone. 
But also, if we can test (real, imag) with (mag, imag) on a Mac, I don't see why can't we combine the two, and sort them alphabetically. ;).

--Matt