[Ifeffit] Large Amplitude Values

[Scott Calvin]

Hi Gavin,

The problem isn't the value of S02; it's the uncertainty. I've noticed  
that Ifeffit has a tendency to push up the best fit value of S02 when  
it's very uncertain, although I'd have to think more about how it  
determines error bars to confirm that. But it does make sense, in a  
"Price is Right" kind of way--negative S02's presumably fit horribly,  
because they turn chi(k) upside down, and so that biases uncertain  
S02's to move the whole range up.

In any case, your focus should be on reducing that uncertainty.

1) One way to do that is to fit with multiple k-weights, assuming  
you're not doing that already. Check the kw 1, kw 2, and kw 3 boxes  
all together and run a fit. The reason this works is that the EXFAS  
equation shows no k-dependence for S02, but a k^2 dependence for  
sigma^2, which often shows a high correlation with it in fits. Fitting  
multiple k-weights sometimes helps break that correlation.

2) Along the same lines, if you can squeeze out any additional k-range  
that may help.

3) If you're fitting coordination numbers, then adding additional  
scattering shells with some physically defensible scheme for  
constraining coordination numbers to a small number of parameters can  
help a lot.

4) Another good technique is to fit multiple samples simultaneously,  
constraining S02 to be the same for all of them. Or fit the sample and  
a standard measured in a similar way simultaneously, again  
constraining S02 to be the same.

5) Along the same lines, you could fit a standard measured in a  
similar way to determine S02, and then constrain the fit of your  
sample to take on that value.

4 and 5 are similar, so you may wonder if I have a preference. I'd say  
that if the samples, beam, detectors, data, and data reduction are all  
well behaved, then #5 is probably best, and has the benefit of being a  
technique with a long pedigree. If you're a little suspicious of  
something in the chain, though (for example, it's difficult to tell if  
you've been consistent in normalizing your standard and sample,  
because one has a big white line and the other doesn't), then #4 has  
the benefit that it distributes the error in the parameters you are  
fitting between sample and standard. This is good both because your  
sample has less error than otherwise, and because the values for the  
standard act as a "canary in a coal mine," warning you by their  
deviation from known values as to the magnitude of the errors you're  
looking at.

--Scott Calvin
Sarah Lawrence College

On Jul 31, 2010, at 11:56 AM, Gavin Garside wrote:

> Scott,
>
> Thank you for a quick response.  The value I am getting for SO2 in  
> the fit most of interest is 2.95 plus/minus 3.72.  So with the error  
> bar I am in range, but I was just suspicious of it before I make any  
> claims about it.  All my experiments were done in florescence  
> because we have ordered bulk material.  By creating a sample that  
> would work in fluorescence I may have introduced dislocations or  
> imperfections that would have effected the physical properties of  
> interest in this sample.
>
> Gavin Garside
> University of Utah
>
> From: Scott Calvin <SCalvin at slc.edu>
> To: XAFS Analysis using Ifeffit <ifeffit at millenia.cars.aps.anl.gov>
> Sent: Sat, July 31, 2010 4:52:35 AM
> Subject: Re: [Ifeffit] Large Amplitude Values
>
> Hi Gavin,
>
> What are the uncertainties on the high S02 values?
>
> Fluorescence is unlikely to be the culprit. While it can affect your  
> ability to normalize properly, you're unlikely to account for a  
> factor of 2 by normalization if the data is relatively decent. And  
> self-absorption tends to suppress S02, not exaggerate it.
>
> Why did you switch to fluorescence on just the handful of data sets?  
> That might provide us a clue.
>
> --Scott Calvin
> Sarah Lawrence College
>
> On Jul 30, 2010, at 10:47 PM, Gavin Garside wrote:
>
>> Fellow X-Ray Absorption Enthusiasts,
>>
>> I have recently compiled a model that gives excellent visual fits  
>> in R, q, and k space for bond spacing in a BCC structure.  This  
>> model gives bond spacings that make sense, and are very close to  
>> what would be expected from this set.  The R factors are very low,  
>> and the enot values correspond quite well to the edge.  However,  
>> our amplitude values are much larger than typically expected.  They  
>> come in at the range of 1.8 up to 5.0, but only on a few data  
>> sets.  On all the rest the amplitude values are 0.4 to 1.0.  Could  
>> this increase in amplitude be attributed to the fact that we ran  
>> florescence measurements instead of transmission, and have a weaker  
>> signal coming to the detector?  What else could be causing this in  
>> only one data set? All samples used in this model have the same  
>> structure.  Thanks in advance to any replies, your help and time is  
>> appreciated.
>>
>
>
> _______________________________________________
> Ifeffit mailing list
> Ifeffit at millenia.cars.aps.anl.gov
> http://millenia.cars.aps.anl.gov/mailman/listinfo/ifeffit

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://millenia.cars.aps.anl.gov/pipermail/ifeffit/attachments/20100731/f33db2c5/attachment.html>