HistFactory staterror functionality in pyhf

[iokarkan]

In an effort to replicate fit results from HistFactory workspace using pyhf, I've come across the staterror configurations of HistFactory, as seen in https://root.cern/doc/v606/classRooStats_1_1HistFactory_1_1StatErrorConfig.html.

For example, I am trying to understand whether the staterror of pyhf can accommodate the
channel.SetStatErrorConfig(0.001, "Poisson")
functionality of HistFactory, with the first number being an uncertainty threshold (per bin, I imagine) and the second being the constraint type, where "Gaussian" or "Poisson" is selectable. From pyhf's documentation, the constraint type of staterror is a product of Gaussians, and I cannot find anywhere indicating that this is optional.

I have tried making an xml workspace using the pyhf CLI, and adding the
<StatErrorConfig RelErrorThreshold="0.5" ConstraintType="Poisson" />
where necessary, and trying to go back to a .json workspace, again with the CLI, but I am not noticing any changes.

Is this functionality to be expected in the future, or is it a design choice for pyhf? Thanks!

[alexander-held]

The type of constraint term can currently not be switched to Poisson, see #760. The threshold setting is also not respected at the moment, see #662.

[iokarkan]

Apologies for not finding these in the pyhf Issues! I tried searching the code but I must have missed the issue reports!

[iokarkan]

Would manually setting the staterror "data" to be 0, during the creation of the pyhf workspace, for a bin that doesn't surpass the set threshold be a work-around for the threshold part of the question?

[alexander-held]

Setting the value to 0.0 will cause another issue (#293 (comment)), but you can use some small value like 0.000001 as a workaround.