Allow fractional uncertainties?

[shunt16]

From Zhav in #1:

I am a bit unsure about this choice. In an error-covariance matrix, the values in the matrix are covariances and variances which have the unit of the measurand squared, while in an error-correlation matrix, the entries are unitless.

[shunt16]

Are you proposing we don't allow fractional uncertainties at all? It is very typical to evaluate/store fractional uncertainties - both to enable interpretation and minimise size on disc. You can always evaluate the uncertainty in the units of the measurand from the fractional uncertainties (which is what the CoMet Toolkit tools do for example).

[EmmaWoolliams]

This all depends on how we handle units. If we allow different units for the uncertainty and the measured value, then we need a way of converting between them - that could include treating them as relative uncertainties rather than absolute uncertainties - but surely they are not "fractional" - i.e. more likely to be described by a decimal (0.013) than as 13/1000? Is "fractional the right word"? Then we could have % as a unit, but also mK for a K quantity etc. Not sure what's done in current NetCDF for how much the units have "meaning" vs just being a label - i.e. whether the computer can use them... If we link to the SI-brochure activity that could work...

Maybe something to document as needing to be thought about later.

[EmmaWoolliams]

Related is Peter Harris's comment about whether we include sensitivity coefficients explicitly. That also links to the question of what the uncertainty represents - and when we talk about there being multiple uncertainties associated with a quantity variable .. .

[EmmaWoolliams]

It's super dangerous from a bug-finding perspective to assume that if you don't give units it's a relative uncertainty! I would much prefer to see you explicitly stating the unit for all and having some representative of it being relative.