Ratio of normsys modifiers

[itsaklid]

Hello,

If I have a normsys modifier named modifierA, attached to a Sample A, is it possible to have the ratio $\frac{modifierB}{modifierA}$ attached to a Sample B , where modifierB is another normsys modifier with the current pyhf implementation ?

Thanks a lot in advance!

[alexander-held]

For ratios there is a trick where you use the ratio itself as a modifier and then the denominator in addition as the second modifier. See #2399 (comment) for this approach when using normfactor modifiers.

I have not seen seen this used or thought about it for normsys modifiers, so not sure spontaneously if the exponential extrapolation (compared to the linear case for normfactor) makes this unsuitable. The modifiers still both multiply the effect of samples, which is what this approach mainly relies on.

[itsaklid]

Hi @alexander-held , thanks a lot for your answer!

Indeed I'm doing something along the lines of #2399 for the normfactor modifiers which appear as ratios in my fit.

I tried to keep my question as generic as possible which hid some of the complexity of the situation I'm in. The subtle point is that the correlation of modifierA and modifierB with respective histosys modifiers might need to be taken into account. This is of course similar to what cabinetry implements with the NormPlusShape systematic, but due to case-specific reasons I'm disentangling the normalization from the shape effects at first, similar to what is envisioned in the #477 cabinetry issue and then manually correlating only certain shape to certain normalization effects (with some of the latter ones appearing in ratios).

If I define the ratio of the normsys modifiers beforehand, propagate their uncertainties correctly and introduce a new modifier for it, then it seems to me that the correlations with the respective histosys modifiers would not be taken into account properly. Would you agree with such a statement ?

Taking this a step further, the experimental implementation of custom modifiers in #1991 can handle only normfactor modifers as of now, is that right ? Or am I perhaps misinterpreting something ?

Thanks a lot for you time!

Cheers,
Ilias

[alexander-held]

This sounds like a very unusual setup to me and I think I would need to understand it in more detail to offer useful advice. At a high level I wonder if there might be a better way to parameterize the uncertainties for your model to simplify the setup. It is not obvious to me that histosys contributions would be problematic at the moment but that may just be my lack of understanding of what you are looking for.

Regarding #1991: I am not sure myself but I believe the scope was broader than just normfactor modifiers, @kratsg @lukasheinrich @matthewfeickert could you confirm? I have only tried it with normfactor myself thus far.