LHCfitNikhef / smefit_release

SMEFiT a Standard Model Effective Field Theory fitter
GNU General Public License v3.0
6 stars 1 forks source link

Ill defined relative systematic for zero central value #52

Closed jacoterh closed 11 months ago

jacoterh commented 1 year ago

Hi,

The following line throws an uncaught exception in case the central value is zero, which is the case for instance in the CMS_SSinc_RunII dataset. The fit goes on with inf in the multiplicative systematic array.

https://github.com/LHCfitNikhef/smefit_release/blob/85a6e9af62ba449b8d85161d188a3a266fb01b24/src/smefit/loader.py#L154

But it's not a problem in case the corresponding systematic is additive, it gets filtered out, but it would be nice to have some kind of fix anyway.

giacomomagni commented 1 year ago

This can be done... however I wonder how can a central value be exactly 0 ?

jacoterh commented 11 months ago

Some of the CMS RunII signal strengths have been restricted to nonnegative values and this results in a best fit value of exactly zero apparently, see page 16 in https://arxiv.org/pdf/1809.10733.pdf:

"In the case of WH, ZH, and ttH production with H → ZZ decay, as well as ZH production with H → γγ decay, the background contamination is sufficiently low so that a negative signal strength can result in an overall negative event yield. Therefore, these signal strengths are restricted to nonnegative values"

image

The gaussian approximation is not valid for these datapoints...

giacomomagni commented 11 months ago

Okay, if this is the case I think we have a problem with these points. When we symmetries the error the central value should also shift. So if the CMS result, with asymmetric error is 0, we should not have 0 in out tables... Am I correct ? @tgiani
Maybe we should remove those points @juanrojochacon ?

juanrojochacon commented 11 months ago

I would say that for such points a Gaussian likelihood is not appropriate, and hence should be discarded from the analysis

giacomomagni commented 11 months ago

shall I close @jacoterh ?