Spectrum data model: accuracy

[Alberto Micol]

TOC:

  - Accuracy: Bounds vs Errors
  - Accuracy is required
  - Inconsistent Accuracy default values


BOUNDS vs ERRORS
----------------

Let me start with a recent email exchange:

>>
>> Spectrum.Char.TimeAxis.Coverage.Location.Value and .Bounds.Extent
>> Unless the data are time series, these should be 'should' not 'must'
>> as not all spectra etc. have recorded times.
>
> You usually know the date to within at least a decade.
>
> Location.Value =  1975 Jan 1
> Bounds.Extent = 10 years

This is very confusing.
If the date is not know very precisely, the accuracy must be not zero.
The accuracy, in this sense, is linked to the concept of "how
well I know this measurement", hence it is linked to the Error.

The bounds are not errors. The bounds tell the starting point and the
ending point (and each with its own error) of a property of an  
observations,
and have nothing to do with the precision with which we know the  
location.

Bounds.Extent = 10 years means that the observation spanned across 10  
years,
and the user will think that the spectrum is a co-addition or anyway  
an assembly
of spectra taken within 10 years.


In 4.6.1 "Coordinate bins", the sentences:

[...] width values are suitable for Spectrum.Char (the global  
accuracy)[...]
[...] if absent, the bin limits are assumed to be [...] and bounded  
by range
       given in Char.*.Coverage.Extent

gives in fact the impression that the bins (a sampling concept) and  
the bounds
(a coverage concept) are intermixed, and that they are related to the
global accuracy.

I think that should be clarified, and the concepts disentangled.


REQUIRED FIELDS
---------------

To go back to the required/optional fields:
the initial paragraph in 4.6 "Accuracy Fields"
states that all the Accuracy fields are optional.

I think this is simply bad. A data provider for sure
does know the accuracy, at least at a course level,
and if s/he doesn't, s/he should admit it, frankly, not silently.

An example (cfr example above):
What could happen if an astronomer downloads the spectrum taken 1- 
Jan-1975
without knowing that the uncertainty on the date is 10 years?
the answer is: wrong science.

Physics requires measurement errors, the VO MUST provide them.
And if it is impossible to know the errors, the VO MUST state that.
Error = UNKNOWN is much better than silence.

Especially given that the SpectralAxis.Accuracy.StatError (and SysError)
defaults to zero (that is, by default the measurement is perfect!).  
(page 16)


Inconsistent default values:
----------------------------

BTW, some Stat/Sys Errors default to zero some other to unkown some  
to undefined.
I think we have to consistently make them defaulting to the less  
risky value: unkown.


Alberto