Hello.

I noticed a weird feature in root 6.24.00, when trying to fit a model using a RooSimultaneous pdf.

Am I doing something forbidden in root 6.24.00 ?

Best,

Jean-Baptiste

Describe the bug

I have a model with a single category, describing the shape of a distribution, with three unconstrained nuisance parameters, one constrained nuisance parameter and one parameter of interest. The three unconstrained NP describe the shape of the background and the background yield. The poi is proportional to the signal yield. The dataset is a background only dataset.

I can do the fit with as pdf : model_had1A = RooProdPdf (fsb_had1A, constbias_had1A) where
fsb = ns x fs + nb x fb, fs and fb are signal and background pdf, ns contains the poi, nb is free floating. It runs
smoothly even if we are close to the unphysical region (poi < 0). (Probably many error messages from the exploration of the unphysical region have been removed from the output.)
I can also try a fit with the pdf : simPdf = RooSimultaneous (indexCat=had1A, had1A=model_had1A) : this is virtually the same pdf, but this time embeded in a RooSimultaneous object. The fit fails.

Between the two root versions, one things that appeared weird to me is that fsb_had1A does not seem to be
normalized when embeded in the RooSimultaneous in root 6.24.00. Maybe that is fine, but this is different from
what I see in root 6.22.02.

Expected behavior

I would expect exactly the same results in both fits, with a best fit poi = 0 and a reasonable uncertainty.
This is what I see in root-6.22.02. In root 6.24.00 the fit with a RooSimultaneous fails.

To Reproduce

I put the code here /afs/cern.ch/user/j/jdevivi/public/ISSUEROOFIT

In root-6.24.00, I just do
root.exe testWSsimulvsprod.C

In root-6.22.02, I do
root.exe load.C testWSsimulvsprod.C
since I use a RooCrystalBall from root-6.24 and did not put the code in the workspace.

Log files can be found in the same directory.

Setup

ROOT version : 6.24.00
Operating system : macos 10.15.7 clang-12
I built root myself. But I also tried via a docker image built for ATLAS