GitLab

Revamp zero-mode handling

See merge request ift/nifty!607

Treat scaling correctly

See merge request ift/nifty!616

Fix VariableCovarianceGaussianEnergy

See merge request ift/nifty!548

This was achieved by exploiting the diagonality of the Fisher metric.
In cases this is not true one needs more samples to increase sensitivity.
Tested that a factor of 2 only in the inverse covariance can now be detected.

Fixed Fisher tests. The following bugs were present:

 - Metric of VariableCovarianceGaussianEnergy was wrong (already fixed in NIFTy+7, was fixed by merging)
 - debugging factor was not used, thus leading a false 'did not raise' assertion
 - the fast Fisher test for the VariableCovarianceGaussianEnergy was totally broken
 - slightly decreasing the tolerance for the isher test leads to factors of 2
   being detected

Log invgamma

See merge request !614

The zero-mode is kind of accounted for as it is discarded in the
calculations and thus the FIXME is kind of fixed.

In most cases the exposure is unknown and must be learned from the data.
However, this corresponds exactly to learning the zero-mode of an
ordinary log-normal field. Hence, remove the normalization code from the
density_estimator call and drop the exposure parameter.

Adapt the density_estimator accordingly.

Re-introduce the FIXME on how the zero-mode is treated in the
`_Normalization` operator.

If no amplitude total offset (i.e. zero-mode of the amplitude operator)
is specified, the correlated field is still able to produce a useful
tensor-product of amplitudes. However, their overall zero-mode is of
course not determined. As convention, set the zero-mode of the final
amplitude operator to zero.

Adapt the statistics summary to not require the zero-mode.

Convert the return list of operators to a tuple as to disallow the
caller to amend or change it in other ways.

Since `noramlized_ampliudes` is not a property anymore, name it more
actively as to highlight the need for it to be called as a function.