[Ifeffit] Fwd: Re: why ss_2 is negative?
Carlo Segre
segre at iit.edu
Tue Feb 27 17:45:42 CST 2007
Hello Hao:
I took a look at your fits _very_ briefly. Just a couple of things to
note.
The range of chi(k) which you use in the fit is probably a bit large.
You are starting at 2.1 and the data at that point is really not usable.
I would start at 3 or so. On the high end, it seems OK.
The estimated standard deviations on lots of your parameters are very
large, on the order of the parameter values themselves. This is a red
flag and it generally indicates that the fit has too many variables. This
is the biggest concern that I have. Look at the values:
dr = 0.0010050 +/- 0.0111490
ss = 0.0031940 +/- 0.0021850
dr_1 = 0.0464110 +/- 0.1100960
ss_1 = 0.0130090 +/- 0.0188820
dr_2 = 0.0896690 +/- 0.0765040
ss_2 = 0.0043400 +/- 0.0106960
n1 = 3.3690270 +/- 3.5237430
most of these values are indistinguishable from zero. (As an aside, are
you getting an error box when you try to fit? The parameter "dr" is not
allowed in newer versions of Artemis.) Here is my suggestion:
Change the chi(k) range to 3-13.2. Change the fitting range to 1-2.65 or
so. Link the ss_1 and ss_2 parameters together and set dr to 0.0 as it is
basically there already and has a large uncertainty.
By doing this you will have reduced significantly the number of parameters
and the uncertainties will be more manageable. More importantly, when you
work on a model, start with a more highly constrained one (fewer paths,
for example) and try to work your way up while always making sure that the
results make sense. You will be able to reduce the parameter space that
you have to cover this way. Keep a close eye on the "Reduced Chi-square"
as you add more "guess" parameters. If it goes up, you may be
overreaching on your model even though the "R-factor" goes down!
HTH
Carlo
--
Carlo U. Segre -- Professor of Physics
Associate Dean for Special Projects, Graduate College
Illinois Institute of Technology
Voice: 312.567.3498 Fax: 312.567.3494
segre at iit.edu http://www.iit.edu/~segre segre at debian.org
More information about the Ifeffit
mailing list