MCT - model document delivery.

[Markus Demleitner]

Hi Mark,

On Fri, Sep 04, 2020 at 10:52:27AM -0400, CresitelloDittmar, Mark wrote:
> Meas-1.0 <http://www.ivoa.net/documents/Meas/20200413/index.html> (
> 13/Apr/2020 )
>   * Updated to RFC comments
>   * Topics:
>      o I don't believe there are open topics here..

Hm.... :-)

So, from my RFC comments: I still don't understand why you have
different types for Generic, Time, etc.  Am I missing the rationale?
For what we're doing in Meas (essentially, associating values and
errors), how is just GenericMeasure not enough?

Even more importantly, I'm still rather strictly against using
anything from coords in meas.  Having a value and an error is a lot
more fundamental than having coordinates (which more or less imply a
vector space if the word is to mean anything).  What we do now, on
the other hand, links meas to coords without any profit I can discern.

Then, while it's some progress to state

  The current model assumes Gaussian distributions with shpes defined
  at the 68% confidence level,

that claims *a lot* more than you usually can, and thus I couldn't
usually use this to annotate my value-error pairs; and it would
probably require a bit of explanation with the asymmetrical error
models, as the "half-Gaussians" will be discontinuous at the center
if their widths are diffrent.  That's not a disaster for a
distribution, but it is still a bit funky.

But really, I don't think we should try to be that precise at all.
If we don't speak about distributions a lot more, we should confess
up and simply derive

NaiveError

from the (abstract) Uncertainty.  It would say something like

  This error does not really specify anything about the underlying
  distributions or confidence levels.  Essentially, it is intended to
  support plotting of error bars with mild semantics.  For further
  analysis, human interpretation or additional metadata (probably
  from later versions of this model) is required.

And I still think all the effort on the various multi-D errors should
not be made at this point.  They're all special cases of correlated
errors, and if we tackle correlations at this point at all, they
should talk about individual errors.

Hence, I'd still say the model should just have:

Measurement:
  value: float
  error: Uncertainty

Uncertainty (abstract class)

NaiveError:
  value

and perhaps

Correlation:
  coefficient: float
  err1, err2: Uncertainty


As said in my original RFC comment, in contrast to the current model,
that would allow a meaningful annotation of the Gaia result catalogue
(and, I claim, many, many others).  Which is something I'd dearly
like to do.

And that step will also break the dependency of Meas on Coords.  Less
entangled standards are always a big win, as they allow for
independent development.  Which, given that we pretty certainly will
want errors with more precise machine-readable semantics
(distributions, their parameters, confidence levels...) once we can
do the naive, simple things, is particularly important for Meas.

        -- Markus