[XRF] Remove the silent change of the asimov null hypothesis

[Phmonski]

This Pull request:

XRooFit changed the asimov null hypothesis, disallowing a value of 0.0. This is done because bins with signal but no background contributions result in infinities when the asimov data set is created at mu=1.0 and tested at mu=0.0.

I suggest we omit the change of this value, as this changes the result of the expected p-value without alerting the user behing the scenes. Furthermore, the issue lies with the workspaces, and not the calculation of the pvalue in xroofit.

@will-cern Do you aggree with this change?

[will-cern]

Changes to xRooFit should really be made in the xRooFit repo and then we sync them into ROOT. In xRooFit the value was already changed to 1e-3 instead of 0.5: https://gitlab.cern.ch/will/xroofit/-/blob/master/src/xRooNLLVar.cxx?ref_type=heads#L1798 ... was this small enough in your tests to reproduce results? Otherwise how are you dealing with the models with the pure-signal bins?

[Phmonski]

We guard against the signal only bins, by adding a small yield (1e-9) to the background templates in the SR. This way we can still perform the null hypothesis at mu=0.0 and create the asimov dataset at mu=1.0. This is one of the standard measures to guard against issues with signal-only bins when creating the workspaces.

Either way, would it be possible to change the value of the default away from 0.5 and backport the change to v6.40?

[github-actions]

Test Results

    23 files      23 suites   3d 14h 56m 40s ⏱️
 3 878 tests  3 878 ✅ 0 💤 0 ❌
79 906 runs  79 906 ✅ 0 💤 0 ❌

Results for commit 430d180.