Hello everyone,
When asking to display asymmetric error bars, TH1 and RooDataHist give the same intervals except for a bin with 0 event.
The frequentist approach states that there is 68% coverage and 32% in the upper and lower tails (16% each). For the case n=0 there is no lower bound, so the upper bound could correspond to 32%. RooDataHist makes this distinction, while TH1 does not.
For RooDataHist, it returns the one-sided upper limit [0, 1.35].
For TH1, it returns the double-sided upper limit [0, 1.84]. As you can see in the link, the one-sided upper limit is also present l.9160 , but it is commented out without any explanations.
I wanted to know the reason behind this decision for TH1, why the RooDataHist uses another definition, and which one we should choose when using asymmetric uncertainties for empty bins?