ClarkMcGrew
commented
1 month ago This would be good for the MCMC like calculations that don't assume the likelihood is smooth and continuous, but it has fairly deep implications for the properties of the likelihood. I think the statistical theory is also a little fraught because it introduces similar model dependent conditional probabilities between sample bins that we see in the current reweighting techniques (bin to bin migration is still coupled to the model dependent location of the events inside the bins).
Finally, from my reading of the code it would also involve a fairly significant restructuring of the data structures associating events to bins. Since it's a design change, I think it falls into the "next time we write a fitter" category.
Some parameters like energy scale can be defined as a distorsion of event observables.
This means at each evaluation of the Propagator, some event would request a new bin search wrt sample or parameter dials.
In current statistical analysis, these parameters are treated as an effective reweight factor that simulate those migrations. This technique has historically been chosen for improving performances and insuring the continuity of the second derivative of LLH. One of the use is the lack of accuracy for propagating the systematics since this effective method uses an a priori shape of the involved spectra.
Having this feature would allow the user to have a more accurate propagation of systematics or allow to build the appropriate reweight dials to be used in a given analysis.