[RF] Regression in ranged fits with RooSimultaneous due to new likelihood evaluation #18718
Description
Description
As we learned from a user report on the forum, ranged fits with RooSimultaneous are currently broken with the RooFit::EvalBackend("cpu"):
https://root-forum.cern.ch/t/problems-fitting-simultaneous-pdf-fit-in-subranges/63537
It only works with RooFit::EvalBackend("legacy").
@will-cern, I think this is also the issue that you faced some time ago, and now we have a simple reproducer so I think I can fix it soon.
Reproducer
using namespace RooFit;
const double nBkgA_nom = 9000;
const double nBkgB_nom = 10000;
RooRealVar x("x", "Observable", 100, 150);
x.setRange("fullRange", 100, 150);
x.setRange("fitRange", 100, 130);
// For SplitRange() support
x.setRange("fullRange_RunA", 100, 150);
x.setRange("fitRange_RunA", 100, 130);
x.setRange("fullRange_RunB", 100, 150);
x.setRange("fitRange_RunB", 100, 130);
RooRealVar nBkgA("nBkgA", "", nBkgA_nom, 0.8 * nBkgA_nom, 1.2 * nBkgA_nom);
RooRealVar nBkgB("nBkgB", "", nBkgB_nom, 0.8 * nBkgB_nom, 1.2 * nBkgB_nom);
RooExponential expA("expA", "", x, RooConst(-0.06));
RooAddPdf modelA("modelA", "", {expA}, {nBkgA});
RooExponential expB("expB", "", x, RooConst(-0.09));
RooAddPdf modelB("modelB", "", {expB}, {nBkgB});
RooCategory runCat("runCat", "", {{"RunA", 0}, {"RunB", 1}});
RooSimultaneous simPdf("simPdf", "", {{"RunA", &modelA}, {"RunB", &modelB}}, runCat);
std::unique_ptr<RooDataSet> combData{simPdf.generate(RooArgSet(x, runCat), Extended())};
using Res = std::unique_ptr<RooFitResult>;
Res fitResult{simPdf.fitTo(*combData, Save(), Range("fitRange"),
// SumCoefRange("fullRange"), // doesn't help...
// SplitRange(), // also doesn't help...
EvalBackend("cpu") // works only with the "legacy" backend
)};
fitResult->Print("v");The output is:
Floating Parameter InitialValue FinalValue +/- Error GblCorr.
-------------------- ------------ -------------------------- --------
nBkgA 9.0000e+03 7.2000e+03 +/- 2.05e-02 <none>
nBkgB 1.0000e+04 8.0000e+03 +/- 6.79e-04 <none>Looks like there is not even a fit happening, but the parameters are set to some constant values.
With EvalBackend("legacy"), the output is correct:
Floating Parameter InitialValue FinalValue +/- Error GblCorr.
-------------------- ------------ -------------------------- --------
nBkgA 9.0000e+03 8.9486e+03 +/- 1.01e+02 <none>
nBkgB 1.0000e+04 1.0103e+04 +/- 1.03e+02 <none>