On Thu, Apr 20, 2017 at 5:02 PM, Christopher Thomas Chantler <chantler@unimelb.edu.au> wrote:
Note that in general any and every smoothing operation reduces the information content of the data and its ability to reveal structure.
Well, maybe. If one has mu(E) measured every 0.01 eV over an 600 eV EXAFS scan to k=12A^-1, one does not really 60,000 independent measures of the structure. Yuji's data wasn't that absurd, but it did have close to 4000 measurements for the full EXAFS spectrum out to k=18^-1. And to be clear, there is not anything wrong with that, it's just a matter of how you decide to treat it.
As the plots attached in Yuji's original message and my replies show, the resulting chi(k) definitely has is amplitude suppressed when doing a simple boxcar average of data onto the 0.05 Ang^-1 grid (Athena with Ifeffit). But when using cubic spline interpolation (Larch), the amplitude of the EXAFS oscillations are not noticeably suppressed, though the high frequency noise is also much higher. Applying a Savitzky-Golay filter prior to the cubic spline interpolation did not appreciably suppress the EXAFS oscillations though the high frequency noise was reduced.
Very finely-spaced energy data might reveal is at much higher R than we can hope to model with EXAFS. A k-grid of 0.05 Ang^-1 can give frequencies to 31Ang, and so is probably accurate to 16Ang without significant signal loss. That is, out to k=18Ang^-1, you really only need 360 samples, but you'd like these as noise-free as possible. Having 3600 measurements on a grid of 0.005 Ang^-1 might give you data out to 160Ang in principle, but the photo-electron tends to not cooperate.