[Ifeffit] fitting questions

Matt Newville newville at cars.uchicago.edu
Wed Jan 7 17:20:20 CST 2004

Hi Kristine, Bruce, Shelly, everyone,

I think Bruce and Shelly gave most of the answers to your 
questions, but I thought I'd add a few comments:
> (1)  I was wondering (this might be a silly question) whether it was
> considered inappropriate to fit over a range larger than your spectra?
> (i.e. k-range of fit set to 2-15 when in actuality the spectra data only
> goes to ~12.5?)

I think it's not inappropriate to go further than the data 'really
goes', but rarely helps in the end.  

The common sense approach would be to set kmax to a value where the
signal and noise look to be about the same size.  Beyond that,
you're adding more noise than signal, and so not doing much good.
Common sense is sometimes wrong and is difficult to automate.
> (2) chi-square, r-factors from artemis

I think those are all reporting issues.  The feffit() command could,
in principle, write the partial chi-square for each data set.  
Would this be useful?

> (3)  Also is there anyway to have Artemis deternine the best k-range (or
> perhaps a way to have it step through different values of kmin and kmax)
> ?  Just wishful thinking on my part probably...
> Please let me know if you can help me understand these issues a bit better.

Automatically setting kmin wouldn't be too hard to automate (say, a
point near k=2.5 where chi(k) is close to zero??), but automating
the setting of kmax is more challenging.  

One approach would be to use the chi_noise() command, which
estimates the noise in chi(k) and saves it as epsilon_k.  A
recommended kmax might be set to a value such that a "heavily
smoothed" chi(k) spectra is always below the estimated noise level.  
The chi_noise() command could do something like this to set a
recommended kmax, though I'm not sure of a precise definition for
"heavily smoothed" or how well this would work in practice.   This
does seem like it's worth looking into.


More information about the Ifeffit mailing list