Uncertainties on pulls

[annmwang]

Question

Is there a way to get the uncertainties on the pulls? I followed the example here:

https://github.com/scikit-hep/pyhf/blob/master/docs/examples/notebooks/pullplot.ipynb

but I think the option "return_uncertainties" is no longer accepted for the function pyhf.infer.mle.fit. I get the error

pyhf.exceptions.Unsupported: Unsupported options were passed in: ['return_uncertainties'].

I'm using pyhf version 0.5.4 and the numpy/scipy backend/optimizer.

Thanks so much!

Relevant Issues and Pull Requests

I found the notebook through: #630

[kratsg]

Use the minuit optimizer

pyhf.set_backend('numpy', 'minuit')

which returns uncertainties. You'll see in the notebook that this uses the minuit optimizer

    pyhf.set_backend("numpy", pyhf.optimize.minuit_optimizer(verbose=True))

[annmwang]

ah ok, thanks for the quick answer!

so there's no way to do this with the scipy optimizer?

[matthewfeickert]

so there's no way to do this with the scipy optimizer?

Correct. In HEP we have a goal of ascribing uncertainties to our minimizations, but this is not a shared universal goal in the task of minimization (c.f. this Twitter thread where @betatim first explained this to me). As as result, please use minuit in these circumstances.

@annmwang I'm going to close this Issue now, but please feel free to respond here if you have more questions. We're happy to try to address them.

[matthewfeickert]

@annmwang Also maybe of interest to you, in pyhf v0.6.0 you'll have access to iminuit's correlations matrix thanks to @kratsg in PR #988.

[annmwang]

Interesting... thanks for the help- I will switch to minuit then. Thanks for the pointer to the correlations matrix too!