Catch 100% normalization uncertainty modifiers

[alexander-held]

Summary

I somewhat frequently run into workspaces with 100% normalization systematics, which result in NaN yields and likelihoods in pyhf. They have modifiers that look like the following:

{
    "data": {
        "hi": 2.0,
        "lo": 0.0
    },
    "name": "some_name",
    "type": "normsys"
},

Extrapolation is not defined if the down variation is exactly zero, so fits with pyhf failing is not surprising. These workspaces usually were produced in ROOT, where they seem to work fine. There must be some protection to handle this case, I am not sure how it works.

In such cases it is not very obvious to users of pyhf where the problem lies. I think it would be good to catch this with an informative error message. It might also be good to investigate the ROOT treatment and possibly adopt it.

Additional Information

n/a

Code of Conduct

• I agree to follow the Code of Conduct