Assigned errors must be positive, in hypotest

[ssaha1234]

Dear pyhf members,
I am running a simple hypotest for producing limit plots for my analysis. I am using pyfh 0.6.3, following is a short snippet of the code

pyhf.set_backend("JAX", pyhf.optimize.minuit_optimizer(maxiter=int(1e9), tolerance=10000000))
wspace = pyhf.Workspace(json.load(open("LH_CLs_{}.json".format(int(sys.argv[2])))))
model = wspace.model()
data = wspace.data(model)

result = pyhf.infer.hypotest(float(sys.argv[3]), data, model, return_expected_set=True)
retval = [result, float(sys.argv[2]), float(sys.argv[3])]
pickle.dump(retval, open("CLs_result_{}.pkl".format(int(sys.argv[1])), "wb"))

I am getting the following user warning -
Assigned errors must be positive. Non-positive values are replaced by a heuristic.

after which the job goes into hold status on Condor. I do not have any negative values in the json file, I do not understand the source of this warning, is there any hint for solving this issue. I do see something related to the error here - #1919 but I am not sure I understand it completely. Any help would be welcome !

P.S. - I released the jobs on Condor, so I am hoping I will get results soon.

Thank you !
Best,
Shreya

[kratsg]

Why is your tolerance very large? This seems wrong.

[ssaha1234]

Hi @kratsg,
It is because the fit was not converging with a smaller tolerance :( Ofcourse, I can reduce it a little and check, what is the accepted value of tolerance, in general ?
Thanks !
Best,
Shreya

[alexander-held]

The default of 0.1 is good, going lower to 1e-3 or 1e-4 can be useful in some cases. If it does not converge with these settings, there is likely a problem in your model.

[matthewfeickert]

Yes, that is indicative of a problem with the model.

[ssaha1234]

Hi @alexander-held and @matthewfeickert,
Okay I understand, I will try with a smaller tolerance and check if the fit converges or not.
Thanks !
Best,
Shreya

[matthewfeickert]

I am getting the following user warning - Assigned errors must be positive. Non-positive values are replaced by a heuristic.

after which the job goes into hold status on Condor. I do not have any negative values in the json file, I do not understand the source of this warning, is there any hint for solving this issue.

PR #1919 resolved Issue #1918 and will go into pyhf v0.7.0. To avoid getting this warning during rumtime now with pyhf v0.6.3 just use iminuit<2.12.2. So assuming you're installing pyhf v0.6.3 with the minuit extra:

# requirements.txt
pyhf[minuit]==0.6.3
iminuit<2.12.2

$ python -m pip install -r requirements.txt

[ssaha1234]

Hi @matthewfeickert,
Thanks for the quick reply, I will check the version of iminuit I am using currently, and make sure it is below 2.12.2. Thanks for linking #1918, I understand it a bit better now !
Thanks.
Best,
Shreya