[Ifeffit] Fwd: Re: why ss_2 is negative?

Scott Calvin SCalvin at slc.edu
Sun Mar 4 12:01:47 CST 2007


Hi Hao,

I haven't had time to look at your fits, but I have some generic 
responses to what you say below.

It is typical of someone new to the field to chase r-factors. Don't 
do it! I could take any spectrum and any model and by floating enough 
variables, come up with a terrific looking r-factor...for a fit 
that's utter nonsense. At the risk of having another set of 
completely arbitrary criteria named after me <grin>, here's the 
guidelines I give undergraduates when I'm teaching the technique. 
(Note: these guidelines apply to decent-quality data. For cases like 
very dilute fluorescence, it's reasonable to expect statistical 
effects to inflate the R-factor a bit.)

R-factor > 0.10: Serious problems with the fit. The underlying model 
may be incorrect. It's best at this stage to look at the spectrum for 
clues. Maybe the wiggles are qualitatively right, but shifted over or 
the wrong amplitude. Then the model may be OK, but things like the 
free parameters and constraints may need to be adjusted. On the other 
hand, the wiggles may be qualitatively wrong, in which case the 
underlying model must be seriously questioned.

R-factor in the range 0.05 to 0.10: Underlying model may be correct, 
but there is likely some effect not being taken into account (for 
example: phase impurity, oversimplified sigma2 constraints, 
vacancies, etc., etc.). Alternatively, perhaps it's too wide a 
k-range, problems with background subtraction, or the like. I have 
occasionally published fits in this range, although always with an 
explanation of possible factors in the text of the article.

R-factor in the range 0.02 to 0.05: Decently good match between 
fitted and actual spectra. There's still enough of a mismatch that, 
if the data is good quality, there are probably some issues with 
details of the model. At this point, R-factors are becoming less of a 
concern than the plausibility of the constraint scheme and the fitted 
parameters, the number of degrees of freedom, agreement with other 
source of information about the system, etc..

R-factor less than 0.02. Good match between fitted and actual 
spectra. Unless you're doing technical work on a very 
well-characterized sample (say, a piece of copper foil), there's no 
point in trying to reduce the R-factor any further. You're a lot 
better off with a constraint scheme that can be explained on physical 
grounds and an R-factor of 0.019, than, for example, introducing a 
parameter for the third cumulant of a fourth-nearest-neighbor in a 
metal to get an R-factor of 0.005.

These are broad guidelines only; the R-factor has no meaning in a 
statistical sense, so what to expect is highly dependent on data quality.

* * *

OK--I've taken a very quick look at your fits to help answer your 
question about the k-weights.

Your k-weight 2 and k-weight 3 fits are consistent, in that the error 
bars of corresponding parameters overlap. But the k-weight 2 fit has 
enormous uncertainties, and is thus pretty much useless. What good 
does it do you to find that n1 is 3.4 +/- 3.5? Presumably you knew 
that already. :)

So in that sense the k-weight 3 fit is "better." But there are other problems:

--the S02 is a bit high. Ideally it shouldn't be higher than 1.0.

--You seem to be fitting too high an R-range. Going to 3 when your 
most distant path has an Reff of 2.4 is dangerous...depending on your 
substance there may be other stuff out there you're not accounting for.

--The negative sigma2 that you're worried about is NOT a big problem, 
however. It is given as -0.003 +/- 0.008. So it could be positive 
according to the fitted results.

--It looks like you set n2 to 0.175 based on some previous fit. Do 
you believe that's physically reasonable for your system? This 
business of running a fit, finding some parameters, and then running 
a fit on the same data fixing some parameters to values from a 
previous fit is at best dangerous, and at worst nonsense. That's 
different than the advice often given on this list, where you compare 
a fit with a parameter fixed to a suspected-prior-knowledge value to 
one where it is allowed to float. For example, saying "the chemists 
tell me this coordination number should be 6. I'll fix it at 6, and 
I'll let it float, and if the floated case doesn't seem better, I'll 
go back to fixing it at 6" is very different from "my first fit on 
this sample gave me a coordination number of 5.47 +/- 4.8. That's a 
big uncertainty, so I'll set the coordination number to 5.48 and 
proceed." Using the first procedure in a final, published fit is 
defensible, the second one is not.

Hope that helps.

--Scott Calvin
Sarah Lawrence College


At 01:48 PM 2/27/2007, you wrote:

>I took your suggestion and have the amp and e0 the same for each path. In
>addition, I fixed the number axial U-O as 2. For the equatorial ligand, since
>the total coordination number is between 4 and 6, I guess the number 
>of U-F as
>2(that is n1), and def the number of U-O as (5-n1)(that is n2). Then 
>I did the
>fit. The result showed the ss_2 is still negative. So I set the ss_2 
>as 0.003,
>did the fitting again.   I don't know if it is ok to fix the ss_2=0.003. It
>seems the dr_1 change a lot.
>
>In addition, I tried to do fitting in Kw=2 in stead of Kw=3. At this 
>time, the
>ss_2 is positive, and from my understanding, it seems the modeling is more
>resonable for dr_1 and dr_2, but the R-factor is 0.023, a little bit larger
>than 0.02. How can I decide which one I should use, the Kw=3 or 
>Kw=2? Is there
>anything I can do to improve the modeling?





More information about the Ifeffit mailing list