Dealing with low count signals

[Ben Buse]

What's the best approach to dealing with low count eds signals?

Is it best to keep spectra at full resolution or reduce spectra resolution so instead of 10eV per channel have 40eV per channel?

Or does it make no different for eds fitting algorithms?

[Nicholas Ritchie]

Statistically, there is no advantage to increasing the EDS bin size.  In short, either way you have N counts over W channels narrow channels or W/4 channels broader channels and the count statistics remain determined by the sqrt(N). 

There is some advantage to having narrower bins when it comes to distinguishing peak overlaps.  The shape of the peak is critical for resolving interferences. In the limit that we integrate over the "full peak width", essentially we are reducing the peak to a single value with zero shape information.  If two peak regions overlap, there is essentially no information to distinguish which peak is which.  However, if you retain peak shape information by using a narrower channel width, we can use this information as a basis to fit the measured peak shape and distinguish the fraction of counts to attribute to each peak. 

There is however a point of diminishing returns because of the underlying resolution of the detector.  I find this to be about 10 eV/channel.  In my testing, there is little advantage to using 5 eV/ch or narrower bin widths.