Scope of pyhf.infer

[alexander-held]

Question

What is the planned scope of pyhf.infer? For fit model studies, there is a range of useful functionality that could be implemented either there or manually by the analyzer in need of the functionality. I believe there might be a set of features needed by sufficiently many people that a central implementation would avoid duplicate work.

This is an incomplete list that can probably be expanded:

• fit() with specific nuisance parameters fixed to specific values (with a central implementation instead of going through the optimizer API)
• non-Asimov pseudodata generation (dataset generated with list of nuisance parameters set to non-nominal values)
• nuisance parameter re-centering (e.g. to make the above dataset into an Asimov dataset again, where the best-fit and the nominal nuisance parameter values agree, but those new nominal values are not the same as what's used in the "normal" Asimov dataset

Relevant Issues and Pull Requests

500, #653, #667

[lukasheinrich]

Hi @alexander-held

I think all of these points can be covered. The idea is ideally to factor our pyhf.infer from pyhf and as such the API should be kept generic and not touch on model detail specific to HistFactory. But as far as I understand you're not requesting anything HF specific.

[alexander-held]

This is a good point, there is nothing HistFactory specific here, it's something more in the RooStats world I guess.

[kratsg]

This should be addressed on some level by #948 as well.