Re: [Ifeffit] running ifeffit under 64-bit windows7

Hi Kicaj, OK, sorry I misunderstood then. And, despite my laziness, I think that Sameh is right that it's probably time for a more complete update.... Such a thing should probably feature Bruce's newer codes of course. --Matt On Sat, Mar 10, 2012 at 10:16 AM, Matt Newville wrote:

Hi Kicaj,

2012/3/10 "Dr. Dariusz A. Zając" :

...

Hi, maybe these below clarify a little bit the problem, but the problem sounds very intriguing http://millenia.cars.aps.anl.gov/pipermail/ifeffit/2004-July/005729.html http://millenia.cars.aps.anl.gov/pipermail/ifeffit/2005-October/006613.html http://cars9.uchicago.edu/ifeffit/FAQ/FeffitModeling

I am waiting also for the answer from authors

I would have said these questions have been answered, but maybe I misunderstand... What is the question you are waiting to be answered?

All of chi-square, reduced chi-square, and R factor express the sum of squares of the residual (data-model) after a fit has finished.  The difference between these statistics is how they are scaled.

In particular, chi-square is scaled by the estimated error in the data. If you look at a (naive?) introduction to statistics, you will see it stated that this should be approximately the number of degrees of freedom in the fit.  Reduced chi-square is then defined to be chi-squared / (the number of degrees of freedom in the fit), so that it should be 1 (according to statistics 101).   This presupposes a couple of things that aren't very true for us:  a) it assumes we actually know the uncertainty in the data -- the automated estimate in ifefit is pretty simplistic.  b) it assumes our model of the data is much better than that data uncertainty. Many people describe these as "systematic errors" and include alll sorts of data processing artifacts as well as errors in the Feff calculations.

For us, reduced chi-square is almost always >> 1, unless the data is very noisy.

R-factor scales the fit residual by the magnitude of the data itself, for some estimate of "fractional misfit".   This gives a convenient measure that is independent of the scale of the data (and so also independent of data k-range and k-weight for fits in R-space), and can more easily be made into a "rule of thumb", say "If R-factor > 0.05, then you should  be wary of the results".

Hope that helps,

--Matt

-- --Matt Newville <newville at cars.uchicago.edu> 630-252-0431