Future developments

[ajgilbert]

Have a few developments that I'm hoping to work on in the next year, just noting them here for future reference:

1. Rewrite of the CMSHist* classes. For the last combination quickly hacked together CMSHistSum, which encapsulates sum over processes without individual RooFit objects for these (e.g. compared to CMSHistErrorPropagator with a list of CMSHistFunc as input). Don't want to maintain to parallel codes going forward. Prefer to migrate to CMSHistSum, making sure it can generate the per-process objects (CMSHistFuncWrappers) on the fly in the limited cases where we need these (e.g. post-fit shapes in FitDiagnostics)
2. Related to 1., plan to fully support analytic gradients within the "standard" binned combined likelihood. In a first iteration, this means: lnN (symmetric and asymmetric), gmN(?), shape and shapeN, and simple multiplicative rateParams. All of the above will account for the fact that the autoMCStat parameters may be analytically profiled.
3. Built-in efficient per-bin EFT-style scaling, with possibility to factorise shape systematics (i.e. providing them only for the SM template, an assuming Up/Down ratios hold for all EFT points). Option to inject this scaling via physics models. Should also support analytic gradient for this.
4. Full support for RooParametericHist-style behaviour in CMSHistSum. Make it easy to mix and match a parameterised hist with normal ones, supporting autoMCStats consistently. Can support analytic gradient for simple multiplicative factors.
5. For easier unfolding: allow signal processes to be specified as a TH2, with one axis implicitly taken to be the observable, and the other the set generator-level bins. Thus avoids the case where user has to be specify very large number of signal processes in the text datacard, e.g. CMS jet cross section analyses with O(100)s of bins (make transition from TUnfold more natural)
6. Explore feasibility of providing analytic Hessian (will soon be supported in Minuit2).