Re: [Ifeffit] normalization methods
Hi All,
We did an analysis comparing on order of 100 XANES spectra and found that a normalization producing the most stereotypically "correct" spectrum did not always produce the best linear combinations between them. That is, we tend to think a XANES spectrum should be flat (derivative=0) before the edge, and then there is a step, and wiggles, and the wiggles are centered around another flat line. Certainly, this form of spectrum communicates well in publications, and probably it is best for publications given that different researchers use different normalization algorithms. However, when comparing XANES spectra against each other quantitatively, it is more important that the subtraction method be the same. What we did that worked well was write a normalization routine in MATLAB (derived from Matthew Marcus' description since we acquired XANES on his beamline). Then we were better able to compare spectra and fit them against each other using linear combination. When using spectra that were individually normalized, the linear combinations were never as good. Incidentally, this is an argument for always including raw spectra in supplementary materials even though you would want to use the normalized spectrum in a publication.
Cheers,
Zack
On May 15, 2013, at 9:58 AM, ifeffit-request@millenia.cars.aps.anl.gov wrote:
Send Ifeffit mailing list submissions to
ifeffit@millenia.cars.aps.anl.gov
To subscribe or unsubscribe via the World Wide Web, visit
http://millenia.cars.aps.anl.gov/mailman/listinfo/ifeffit
or, via email, send a message with subject or body 'help' to
ifeffit-request@millenia.cars.aps.anl.gov
You can reach the person managing the list at
ifeffit-owner@millenia.cars.aps.anl.gov
When replying, please edit your Subject line so it is more specific
than "Re: Contents of Ifeffit digest..."
Today's Topics:
1. normalization methods (Matt Newville)
2. Re: normalization methods (Matthew Marcus)
3. Re: normalization methods (Matt Newville)
4. Re: normalization methods (Matthew Marcus)
5. Re: normalization methods (George Sterbinsky)
----------------------------------------------------------------------
Message: 1
Date: Wed, 15 May 2013 09:35:41 -0500
From: Matt Newville
Hi Folks,
Over on the github pages for larch, Mauro and Bruce raised an issue about the "flattening" in Athena. See https://github.com/xraypy/xraylarch/issues/44
I've added a "flattened output" from Larch's pre_edge() function, but the question has been raised of whether this is "better" than the simpler normalized spectra, especially for doing PCA and/or LCF for XANES.
Currently, the "normalized" spectra is just "(mu - pre_edge_line)/edge_step". Clearly, a line fitted to the pre-edge of the spectra is not sufficient to remove all instrumental backgrounds. In some sense, flattening attempts to do a better job, fitting the post-edge spectra to a quadratic function. As Mauro, Bruce, and Carmelo have pointed out, it is less clear that it is actually better for XANES analysis. I think the main concerns are that a) it is so spectra-specific, and b) it turns on at E0 with a step function.
Bruce suggested doing something more like MBACK or Ifeffit's bkg_cl(). It would certainly be possible to do some sort of "flattening" so that mu follows the expected energy dependence from tabularized mu(E).
Does anyone else have suggestions, opinions, etc? Feel free to give them here or at the github page....
--Matt _______________________________________________ Ifeffit mailing list Ifeffit@millenia.cars.aps.anl.gov http://millenia.cars.aps.anl.gov/mailman/listinfo/ifeffit
------------------------------
Message: 3
Date: Wed, 15 May 2013 10:25:23 -0500
From: Matt Newville
What I typically do for XANES is divide mu-mu_pre_edge_line by a linear function which goes through the post-edge oscillations. This division goes over the whole data range, including pre-edge. If the data has obvious curvature in the post-edge, I'll use a higher-order polynomial. For transmission data, what sometimes linearizes the background is to change the abscissa to 1/E^2.7 (the rule-of-thumb absorption shape) and change it back afterward. All this is, of course, highly subjective and one of the reasons for taking extended XANES data (300eV, for instance). For short-range XANES, there isn't enough info to do more than divide by a constant. Once this is done, my LCF programs allow a slope adjustment as a free parameter, thus muNorm(E) = (1+a*(E-E0))*Sum_on_ref{x[ref]*muNorm[ref](E)}. A sign that this degree of freedom may be being abused is if the sum of the x[ref] is far from 1 or if a*(Emax-E0) is large. Don't get me started on overabsorption :-) mam
Thanks -- I should have said that pre_edge() can now do a
victoreen-ish fit, regressing a line to mu*E^nvict (nvict can be any
real value).
Still, it seems that the current flattening is somewhere between
"better" and "worse", which is unsettling... Applying the
"flattening" polynomial to the pre-edge range definitely seems to give
poor results, but maybe some energy-dependent compromise is possible.
And, of course, over-absorption is next on the list!
--Matt
------------------------------
Message: 4
Date: Wed, 15 May 2013 08:41:55 -0700
From: Matthew Marcus
Hi Matthew,
On Wed, May 15, 2013 at 9:57 AM, Matthew Marcus
wrote: What I typically do for XANES is divide mu-mu_pre_edge_line by a linear function which goes through the post-edge oscillations. This division goes over the whole data range, including pre-edge. If the data has obvious curvature in the post-edge, I'll use a higher-order polynomial. For transmission data, what sometimes linearizes the background is to change the abscissa to 1/E^2.7 (the rule-of-thumb absorption shape) and change it back afterward. All this is, of course, highly subjective and one of the reasons for taking extended XANES data (300eV, for instance). For short-range XANES, there isn't enough info to do more than divide by a constant. Once this is done, my LCF programs allow a slope adjustment as a free parameter, thus muNorm(E) = (1+a*(E-E0))*Sum_on_ref{x[ref]*muNorm[ref](E)}. A sign that this degree of freedom may be being abused is if the sum of the x[ref] is far from 1 or if a*(Emax-E0) is large. Don't get me started on overabsorption :-) mam
Thanks -- I should have said that pre_edge() can now do a victoreen-ish fit, regressing a line to mu*E^nvict (nvict can be any real value).
Still, it seems that the current flattening is somewhere between "better" and "worse", which is unsettling... Applying the "flattening" polynomial to the pre-edge range definitely seems to give poor results, but maybe some energy-dependent compromise is possible.
And, of course, over-absorption is next on the list!
--Matt _______________________________________________ Ifeffit mailing list Ifeffit@millenia.cars.aps.anl.gov http://millenia.cars.aps.anl.gov/mailman/listinfo/ifeffit
------------------------------
Message: 5
Date: Wed, 15 May 2013 12:58:53 -0400
From: George Sterbinsky
The way I commonly do pre-edge is to fit with some form plus a power-law singularity representing the initial rise of the edge, then subtract out that "some form". Now, that form can be either linear, linear+E^(-2.7) (for transmission), or linear+ another power-law singularity centered at the center passband energy of the fluorescence detector. That latter is for fluorescence data which is affected by the tail of the elastic/Compton peak from the incident energy. Whichever form is taken gets subtraccted from the whole data range, resulting in data which is pre-edge-subtracted but not yet post-edge normalized. The path then splits; for EXAFS, the usual conversion to k-space, spline fitting in the post-edge, subtraction and division is done, all interactively. Tensioned spline is also available due to request of a prominent user. For XANES, the post-edge is fit as previously described. Thus, there's no distinction made between data above and below E0 in XANES, whereas there is such a distinction in EXAFS. mam
On 5/15/2013 8:25 AM, Matt Newville wrote:
Hi Matthew,
On Wed, May 15, 2013 at 9:57 AM, Matthew Marcus
wrote: What I typically do for XANES is divide mu-mu_pre_edge_line by a linear function which goes through the post-edge oscillations. This division goes over the whole data range, including pre-edge. If the data has obvious curvature in the post-edge, I'll use a higher-order polynomial. For transmission data, what sometimes linearizes the background is to change the abscissa to 1/E^2.7 (the rule-of-thumb absorption shape) and change it back afterward. All this is, of course, highly subjective and one of the reasons for taking extended XANES data (300eV, for instance). For short-range XANES, there isn't enough info to do more than divide by a constant. Once this is done, my LCF programs allow a slope adjustment as a free parameter, thus muNorm(E) = (1+a*(E-E0))*Sum_on_ref{x[ref]***muNorm[ref](E)}. A sign that this degree of freedom may be being abused is if the sum of the x[ref] is far from 1 or if a*(Emax-E0) is large. Don't get me started on overabsorption :-) mam
Thanks -- I should have said that pre_edge() can now do a victoreen-ish fit, regressing a line to mu*E^nvict (nvict can be any real value).
Still, it seems that the current flattening is somewhere between "better" and "worse", which is unsettling... Applying the "flattening" polynomial to the pre-edge range definitely seems to give poor results, but maybe some energy-dependent compromise is possible.
And, of course, over-absorption is next on the list!
--Matt ______________________________**_________________ Ifeffit mailing list Ifeffit@millenia.cars.aps.anl.**gov
http://millenia.cars.aps.anl.**gov/mailman/listinfo/ifeffithttp://millenia.cars.aps.anl.gov/mailman/listinfo/ifeffit ______________________________**_________________ Ifeffit mailing list Ifeffit@millenia.cars.aps.anl.**gov
http://millenia.cars.aps.anl.**gov/mailman/listinfo/ifeffithttp://millenia.cars.aps.anl.gov/mailman/listinfo/ifeffit
participants (1)
-
Zack Gainsforth