fix: pt variation change carried through signal region calculations

[ekauffma]

Addresses 3rd point in #162 (pt variations are carried through by overwriting jet object)

[alexander-held]

As a sanity check, could you please compare output histograms to https://github.com/eguiraud/analysis-grand-challenge-coffea/tree/correct-outputs (single file should be enough)? That one is built from the correctionlib version with back-ported W+jets scale variation fix.

[ekauffma]

Documented discrepancies here:
compare_result (1).txt

Seems to be some discrepancy due to floating point error.

[alexander-held]

That's interesting, looking at the first example

got      [43415.260890439, 90524.50293484158, 124669.85937904267, 78590.96671041285, 51739.84456109777, 36289.230414965794, 25796.08326822223, 20466.19299444227, 15611.67260972688, 12753.150585139763, 10134.699515096338, 8746.319925277914, 54751.576864877905]
expected [43415.260890439, 90524.50293484157, 124668.95084443182, 78590.93441945997, 51740.78538666148, 36289.230414965794, 25796.08326822223, 20466.19299444227, 15611.67260972688, 12753.150585139763, 10134.699515096338, 8746.319925277914, 54751.576864877905]

it looks like there are just some events migrating between bins. For example, these three bins

got 124669.85937904267, 78590.96671041285, 51739.84456109777
expected 124668.95084443182, 78590.93441945997, 51740.78538666148

are presumably caused by two events moving between the outer and inner bin. From a quick look at other cases, that looks like it generally explains the observed behavior. It should be fine to move forward with this then. Thanks for checking!

[alexander-held]

This line needs to be removed, as this update takes care of propagating things correctly.

[alexander-held]

I've looked into this a bit, code for reference is in https://gist.github.com/alexander-held/d44d089b7a71f25bae17a03665325400. This compares the correctionlib-based version from @eguiraud's branch to this PR (without correctionlib). I'm essentially only looking at nominal ttbar, single file, and a zoomed in histogram with bin edges [189.999, 190.0, 190.001]. I am printing out values of the 4j2b observable that falls within that window with 190 GeV subtracted from it.

this PR:

-0.0002548726
0.0003070348
0.0001955073
0.0003590095

resulting in bin yields of [3.69506563 3.69506563]

correctionlib version:

0.0000000000
0.0000000000
0.0000000000
0.0000000000

resulting in bin yields of [1.84753282 5.54259845]

The differences are caused by events migrating across the boundary. What is not clear to me is why all the events in the correctionlib version seem to be ending up at exactly 190 GeV.

[alexander-held]

Looking closer into types, this PR ends up with observable being

<Array [367, 245, 383, 627, ... 352, 236, 617] type='48939 * ?float64'>

while the correctionlib version has

<Array [517, 229, 192, ... 262, 3.84e+03, 198] type='48881 * ?float32'>

I guess the difference here comes down to 32 vs 64bit floats, not sure what exactly causes the difference in setups.

edit: The conversion to 64bit appears via the ak.zip(..., with_name="PtEtaPhiMLorentzVector") as far as I can tell.

[alexander-held]

Looks good, thank you!