RIX PWb : (Co)Variance propagation while resampling

[astronomyk]

https://jira.eso.org/browse/MET-2076

The cube reconstruction for the IFU suggests that it will handle variances and possibly also take into account covariances. However, because p. 376 says "As an approximation, the pixel noises could be treated as uncorrelated" it is not clear, to what extent this will actually happen, and I do not find anywhere how covariances would be saved (data structure and format).
Related to this, the combination of multiple exposures for the IFU heavily relies on resampling (metis_ifu_postprocess and metis_ifu_adi_cgrph). To propagate the variance through that process in a meaningful way would require taking covariances into account.

@oczoske @sesquideus @hugobuddel - I have no idea where to start here.
@sesquideus, I assigned you as it seems up your alley.

[oczoske]

The quoted formulation is mine and dates back to when I wrote down my initial thoughts on where this might be going. It's a bit of a trade-off between the complexities of the model matrix A and the covariance matrix C (if I remember the notation correctly). If you go back to the detector images, then C is diagonal and easy, but A is very complicated (I would consider this the ideal solution, if it's practicable). If you use the rectified cubes, then A is comparatively simple, but C is non-diagonal (but fairly sparse I guess). As I compromise I suggested treating pixel noise as uncorrelated, which means only taking the diagonal of C into account. I cannot tell how whether that would be a decent approximation.
As a response I would suggest admitting openly that you don't know yet what the optimal strategy will be (unless you do?). At the very least say that a bad-pixel mask will be used to prevent extreme pixel failures from spoiling the result.

[sesquideus]

Compiled an answer.