A comment on this thread, started by Bingjie: On Nov 26, 2012, at 11:01 AM, Bruce Ravel wrote: 2, In Artemis, since I was told that each shell in R-space can be fitted separatly, is it OK to fit each shell in the R-range respectively, and then join up the fitted curve and just ignore the not match part in other R-ranges? And can a shell be comprised of several different paths (like Fe-O and Fe-S)? There is, currently, no way to exclude an interior region from a fit in Artemis (i.e. there is no way to fit from 1 to 2 and from 2.5 to 3.5, but to exclude the region from 2 to 2.5). I actually think that's a bad idea and doubt that I would ever implement such a thing. Sure there is. Just make a copy of the data set and treat it as a multiple dataset fit, with the two epsilons forced to the same value. I've considered using that trick in the past, but never done so in earnest, in part because of the very important considerations Bruce describes. So why did I consider it? To get around a really messy forest of glitches in data. While a single glitch can be removed, replacing a forest of glitches with zeroes is creating data that isn't there. Better to somehow exclude the data without fitting it, working only with the non-glitchy data above and below. As I've said, I never got quite desperate enough to do that on a real fit, though--I always found some better way out of the problem. But I agree strongly with Bruce that it's problematic to use this trick to avoid data, as opposed to glitches. The reason I don't like this idea is that, in general, shells in the EXAFS signal are not completely isolated. Because of the limited data range and other reasons, the signals form the various shells have long tails and overlap considerably. That overlap is a very important part of the problem. Getting the small parts of the chi(R) spectrum right is just as important as getting the large parts right. Excluding an interior region could artificially remove this overlap from the evaluation of the fit. That would have a very serious impact on the statistical quality of the analysis. I also worry that an interior exclusion would be used in an attempt to minimize the impact of multiple scattering paths on the analysis. That, too, would be a statistics disaster. In short, you should embrace the correlations between the parts of the problem. --Scott Calvin Sarah Lawrence College