Hi Riti,
I hope it's OK that I'm sending this to the ifeffit mailing list, as
there might be other interesting perspectives on your questions.
On Tue, Oct 11, 2011 at 3:47 PM, Ritimukta Sarangi
Hi Matt, This is Riti again. This time I have a question about EXAFS, which you are an expert on :-) I would like to get your view on systematic errors in EXAFS. Say the data were collected on a standard EXAFS beamline like 7-3 at SSRL. The data have been collected using fluorescence mode and a 30 element Ge detector. The data are modest, between k=2-15 A.
I used to quote in my papers that typical statistical and systematic errors lead to a total error in first shell distances of +-0.02 A. I know it can be more than that in some circumstances but then I tried to look at it a little more closely to separate out the systematic error from all sources (instrument, measurement, detector, Analysis), but I could not find a good reference. I understand that this is a difficult number to estimate. I was hoping that you could shed some light on it. Thanks! -Riti
I'm not sure of a good general reference for that either, though perhaps some of the papers from Corwin Booth are a good place to start? Quoting +/- 0.02 Ang seems like a reasonable estimate for a typical value, but that's hardly rigorous. My sense is that it's pretty challenging to separate "systematic" and "random" noise for most data sets. Fluorescence data from a solid-state detector is probably the best-case for such a separation, because you have multiple detector elements or repeated scans, and usually the individual detectors or scans really are dominated by statistical fluctuations. If you add detector elements or scans together until the noise in the data (or variations in estimated parameters or estimated error bars) stops getting better, systematic errors are starting to be noticeable. Eventually, as you increase the fluorescence counts, the fits will stop improving, which should be when systematic errors dominate. --Matt
On Wednesday, October 12, 2011 02:35:21 pm Matt Newville wrote:
I used to quote in my papers that typical statistical and systematic errors lead to a total error in first shell distances of +-0.02 A. Quoting +/- 0.02 Ang seems like a reasonable estimate for a typical value, but that's hardly rigorous.
I don't have a better answer for Riti's question than the one Matt gave. I do, however, want to get up on a soapbox about the practice of quoting typical error bars. While I certainly agree that something like 0.02 is an unsurprising uncertainty in a distance, there is a very good reason to always report measured uncertainties. The problem with sweeping measured uncertainties under the rug by asserting a typical value is that it is probably not true that for a real-world, interesting, research project that *all* the uncertainties are of a typical size. Because of issues of experimental error and correlation between parameters, it is likely that some of your parameters have uncertainties that are larger than the typical value. You may even have some that are smaller. It is probably also true that some of the most interesting parameters in the fit -- the ones that are most salient to the argument being made in the manuscript -- are the ones with non-typical uncertainties. If you fail to report that part of your analysis, you are depriving your reader of information that is important to the evaluation of your results. It is for that reason that I discourage the practice of citing typical uncertainties. B P.S. This paper by Corwin Booth about error analysis was one of my favorite presentations at the last XAFS conference http://dx.doi.org/10.1088/1742-6596/190/1/012028 -- Bruce Ravel ------------------------------------ bravel@bnl.gov National Institute of Standards and Technology Synchrotron Methods Group at NSLS --- Beamlines U7A, X24A, X23A2 Building 535A Upton NY, 11973 My homepage: http://xafs.org/BruceRavel EXAFS software: http://cars9.uchicago.edu/ifeffit/Demeter
Hi Matt, Thank you. I am looking forward to comments from members from the forum. Yes, I have found some of Corwin's references. Perhaps it might help current and future EXAFS folks to have a page on the Ifeffit website discussing the causes of such systematic and random errors that you point out and listing some of these references? Best, -Riti On Oct 12, 2011, at 11:35 AM, Matt Newville wrote:
Hi Riti,
I hope it's OK that I'm sending this to the ifeffit mailing list, as there might be other interesting perspectives on your questions.
On Tue, Oct 11, 2011 at 3:47 PM, Ritimukta Sarangi
wrote: Hi Matt, This is Riti again. This time I have a question about EXAFS, which you are an expert on :-) I would like to get your view on systematic errors in EXAFS. Say the data were collected on a standard EXAFS beamline like 7-3 at SSRL. The data have been collected using fluorescence mode and a 30 element Ge detector. The data are modest, between k=2-15 A.
I used to quote in my papers that typical statistical and systematic errors lead to a total error in first shell distances of +-0.02 A. I know it can be more than that in some circumstances but then I tried to look at it a little more closely to separate out the systematic error from all sources (instrument, measurement, detector, Analysis), but I could not find a good reference. I understand that this is a difficult number to estimate. I was hoping that you could shed some light on it. Thanks! -Riti
I'm not sure of a good general reference for that either, though perhaps some of the papers from Corwin Booth are a good place to start? Quoting +/- 0.02 Ang seems like a reasonable estimate for a typical value, but that's hardly rigorous.
My sense is that it's pretty challenging to separate "systematic" and "random" noise for most data sets. Fluorescence data from a solid-state detector is probably the best-case for such a separation, because you have multiple detector elements or repeated scans, and usually the individual detectors or scans really are dominated by statistical fluctuations. If you add detector elements or scans together until the noise in the data (or variations in estimated parameters or estimated error bars) stops getting better, systematic errors are starting to be noticeable. Eventually, as you increase the fluorescence counts, the fits will stop improving, which should be when systematic errors dominate.
--Matt
On Wednesday, October 12, 2011 04:47:28 pm Ritimukta Sarangi wrote:
Hi Matt, Thank you. I am looking forward to comments from members from the forum. Yes, I have found some of Corwin's references. Perhaps it might help current and future EXAFS folks to have a page on the Ifeffit website discussing the causes of such systematic and random errors that you point out and listing some of these references? Best, -Riti
Anyone can ask Matt for an account at xafs.org and contribute interesting and useful things. I think a wiki page about errors in XAS measurement would be the bomb. B -- Bruce Ravel ------------------------------------ bravel@bnl.gov National Institute of Standards and Technology Synchrotron Methods Group at NSLS --- Beamlines U7A, X24A, X23A2 Building 535A Upton NY, 11973 My homepage: http://xafs.org/BruceRavel EXAFS software: http://cars9.uchicago.edu/ifeffit/Demeter
Hi Riti,
On Wed, Oct 12, 2011 at 4:54 PM, Bruce Ravel
On Wednesday, October 12, 2011 04:47:28 pm Ritimukta Sarangi wrote:
Hi Matt, Thank you. I am looking forward to comments from members from the forum. Yes, I have found some of Corwin's references. Perhaps it might help current and future EXAFS folks to have a page on the Ifeffit website discussing the causes of such systematic and random errors that you point out and listing some of these references? Best, -Riti
Anyone can ask Matt for an account at xafs.org and contribute interesting and useful things.
Actually, anyone can create an account on xafs.org and create pages there. Do to some inevitable abuse, many pages there are restricted so that only certain users can edit them, and some are even restricted for who may see them. (For the ifeffit website, I do restrict accounts -- oddly there was much more abuse there).
I think a wiki page about errors in XAS measurement would be the bomb.
I agree, that would be really nice. --Matt
participants (3)
-
Bruce Ravel
-
Matt Newville
-
Ritimukta Sarangi