Alan
I absolutely agree with you regarding Rietveld refinement. The same
argument is true for adjusting peak shapes to get the best fit, playing
around with different peak formulae. Often such tweaking, while it makes
the observed and calculated fit look nicer, has little effect on the
atomic positions, which in the end is what one is trying to derive. One
should not lose sight of the model.

But my question about defining SNR is more to do with the use of powder
data in general, and especially as used in industry or forensics. For
example distinguishing in a legal sense between two materials may depend
on which of two powder patterns has the higher SNR. I keep seeing this
term used in papers and books but never defined.
Mike

-----Original Message-----
From: Alan Hewat [mailto:[EMAIL PROTECTED] 
Sent: 19 February 2008 10:10
To: rietveld_l@ill.fr
Subject: RE: advice on new powder diffractometer

> So my question remains: what is the definition of signal to noise 
> ratio that is accepted for powder diffraction?

Why does it matter? A higher Bragg/Background ratio does not necessarily
mean better data if counting statistics are poor. Exaggerating slightly
:-) consider a single peak with a ratio of 100/1 compared to a peak with
a ratio 10000/1000. The second measurement will give the lowest error,
not the first which has a much higher signal/noise. And you measure lots
of points on a slowly varying background, so you have a much better
estimate of background than the normal error of a single point. Please
don't encourage people to simply maximise "signal/noise".

Similarly, low profile R-factor's can be obtained with low resolution
data and high background. That does not mean that low resolution data
produces smaller errors in structural parameters.

I worry about people treating measurement and refinement as black boxes
with simplified measures of quality such as R-factors, signal-to-noise
etc. You have to look at the physical reality of the model and the
estimated errors in its parameters, while not cheating by removing data
that doesn't fit for unknown reasons, adding too much "a priori"
information such as constraints, or throwing in extra garbage parameters
to improve the R-factors.
______________________________________________
Dr Alan Hewat, NeutronOptics, Grenoble, FRANCE
<[EMAIL PROTECTED]> +33.476.98.41.68
        http://www.NeutronOptics.com/
______________________________________________


Reply via email to