Description

Using pyhf.infer.hypotest with poi_test < 1. for ATLAS-SUSY-2018-06 returns a NaN as observed CLs. Moreover, for poi_test > 1. observed and expected CLs seem to be very close.

Expected Behavior

Observed CLs shouldn’t be a NaN for poi_test = 0.999

Actual Behavior

Observed CLs is a NaN for poi_test = 0.999

Steps to Reproduce

Running the following code with pyhf 0.6.0 (with background only json file from ATLAS-SUSY-2018-06) :

import pyhf
pyhf.set_backend(b"pytorch")
import json
import jsonpatch

# ATLAS-SUSY-2018-06 json file
with open("./SUSY-2018-06_likelihoods/BkgOnly.json", "r") as f:
    bkg = json.load(f)
# A sample patch produced by SModelS
patch = [
    {
        "op": "add",
        "path": "/channels/2/samples/0",
        "value": {
            "data": [
                3.155344731368271
            ],
            "modifiers": [
                {
                    "data": None,
                    "type": "normfactor",
                    "name": "mu_SIG"
                },
                {
                    "data": None,
                    "type": "lumi",
                    "name": "lumi"
                }
            ],
            "name": "bsm"
        }
    },
    {
        "op": "add",
        "path": "/channels/3/samples/0",
        "value": {
            "data": [
                21.84465526863173
            ],
            "modifiers": [
                {
                    "data": None,
                    "type": "normfactor",
                    "name": "mu_SIG"
                },
                {
                    "data": None,
                    "type": "lumi",
                    "name": "lumi"
                }
            ],
            "name": "bsm"
        }
    },
    {
        "op": "remove",
        "path": "/channels/1"
    },
    {
        "op": "remove",
        "path": "/channels/0"
    }
]
#Trying to reproduce the issue
llhdSpec = jsonpatch.apply_patch(bkg, patch)
msettings = {'normsys': {'interpcode': 'code4'}, 'histosys': {'interpcode': 'code4p'}}
workspace = pyhf.Workspace(llhdSpec)
model = workspace.model(modifier_settings=msettings)
result = pyhf.infer.hypotest( .9999, workspace.data(model), model, test_stat="qtilde", return_expected=True)
print(result)

returns (tensor(nan), tensor(1.)) increasing poi_test above 1.0 seems to give very close observed and expected CLs :

mu = 1
(tensor(0.9845), tensor(0.9847))
mu = 2
(tensor(0.0070), tensor(0.0070))
mu = 3
(tensor(1.1504e-07), tensor(1.1509e-07))
mu = 4
(tensor(1.0611e-13), tensor(1.0617e-13))
mu = 5
(tensor(1.8170e-20), tensor(1.8179e-20))

P.S. : with numpy as backend, problems seem to appear near poi_test = 0.6 P.P.S : similar issues seem to appear with ATLAS-SUSY-2018-22

Checklist

• Run git fetch to get the most up to date version of master (updated to 0.6.0)
• Searched through existing Issues to confirm this is not a duplicate issue
• Filled out the Description, Expected Behavior, Actual Behavior, and Steps to Reproduce sections above or have edited/removed them in a way that fully describes the issue
• Filled out a requirements.txt