unumpy.average: init

[doronbehar]

• Addresses (already closed) combined uncertainty from multiple independent measurements? #38
• Executed ruff check with no errors related to my changes, and ran ruff format on the changed files.
• The change is fully covered by automated unit tests (Not sure how to do that without simply duplicating the code of the new unumpy.average function.
• Documented in docs/ as appropriate
• Added an entry to the CHANGES file

[doronbehar]

TBH, I'm not sure this does the correct thing when the original shape is something like (4,4,4,4,4) and axes is something like (1,3). Suggestions are more then welcome.

[doronbehar]

TBH, I'm not sure this does the correct thing when the original shape is something like (4,4,4,4,4) and axes is something like (1,3). Suggestions are more then welcome.

I fixed that using numpy.apply_along_axis for every axis recursively.

[doronbehar]

Hmm on a 2nd thought, perhaps this means that e.g for A with shape=(3,4,5,7), and axes=(1,3), if there are correlations between A[:,0,:,0] and A[:,1,:,1] are not taken into account? Because once axis=1 is averaged, the covariance_matrix call for the 1d slices of axis 1 won't take into account the correlations to the values already averaged on axis 3...

I think we can live with that, but perhaps warn the users about it in the function's doc, or elsewhere. Unless of course someone here will think of a better way to implement this.

[jagerber48]

It sounds like you are writing code to extract the standard error on the mean of a sequence of UFloats. Or maybe more generally the maximum likelihood estimator of the mean in case the individual observations are correlated? Is that correct? If so this does sound useful to me.

Maybe consider first writing code that does this and then consider extending to taking means along slices of arrays of UFloat's after that? Perhaps the former function can somehow naturally be broadcast or vectorized into the latter.

[doronbehar]

Hello and thanks for joining the discussion!

It sounds like you are writing code to extract the standard error on the mean of a sequence of UFloats. Or maybe more generally the maximum likelihood estimator of the mean in case the individual observations are correlated? Is that correct? If so this does sound useful to me.

Correct.

Maybe consider first writing code that does this and then consider extending to taking means along slices of arrays of UFloat's after that? Perhaps the former function can somehow naturally be broadcast or vectorized into the latter.

That's what I did eventually in the last attempt, but I'm pretty sure that correlations between values in different axes are not taken into consideration in that case (see my comment above). I reworded the function's __doc__ and added an implementation NOTE comment for maintainers.