[obs-tap]:updates on the Proposed recommendation + new document

[Rob Seaman]

On Jul 12, 2011, at 6:40 AM, Arnold Rots wrote:

> So, which of the three options should I choose for o_stat_error:
> 
> 1. Return "POISSON"
> 2. Return -1 (or some other negative value) that by convention will mean "POISSON"
> 3. Do not return o_stat_error (i.e., don't provide any information on it)

Number 3 is always an option, but the VO should provide interfaces that can do better.

If the noise model is purely poisson then some token that can't be misunderstood should be designated.  Others will have to comment on whether a non-numeric string is "legal".  Otherwise probably not -1 since such a value might be middle-of-range depending on the units.  Either a very small or very large negative number.

If the noise model is poisson+gaussian, what I was trying to suggest was that the value should be a negative number whose absolute value is the gaussian part, eg, -13.7 would mean a gaussian "background" (whatever that means for these holdings) of 13.7 DNs (or whatever the unit is), but that it is to be understood that there is also a signal dependent poisson component.  The handling for the compound noise model would be application dependent; the question here is how to express it.  Applications that don't know poisson from "fish" could just take the absolute value.  Applications that don't know gaussians would see the negative and know to take the square root of the signal.  Those that know about both could do something more nuanced.

There should also be a clear consensus of whether "error" means variance or rather "that thing that one would call standard deviation if the noise were interpreted as purely gaussian".  How do other interfaces and statistical packages handle this issue?

Rob