root-project
[RF] Possibility to automatise creation of simplified models for running toys
Feature description
Several analyses are required to run toys to estimate the final physics results. As a prime example, see the Effective Field Theory (EFT) fits that do not fulfill the requirements of the Wilk's approximation: https://arxiv.org/abs/2207.01350 Thus, the asymptotic approximation, in general, does not hold for the EFT fits.
Since these EFT fits often have many parameters running toys with the full model is prohibitively expensive. Thus, one often approaches the problem by simplifying the likelihood.
As an example, consider this approach:
- Build the full model for your analysis, e.g. using HistFactory
- Provide unconstrained parameters per bin (i.e. ShapeFactors in HistFactory) and run the minimisation
- Run Hesse() to obtain the full covariance matrix
- Build a new, chi^2 model using the full covariance matrix and replace the bin yields with the externally provided parameterisation (in case of EFT this is quadratic parameterisation).
- Use the new chi^2 reparametrised model for toys
The main advantage here is that the new chi^2 model will have O(10) parameters, while the original one can have O(100) parameters and much more complicated model so the new simplified model should be significantly faster to evaluate.
Of course, this relies on the gaussian approximation for the original, non-reparameterised, model to be valid.
It seems this function: https://root.cern.ch/doc/master/classRooFitResult.html#aec39da59156d47b9b4a0c66d1e4a63af Actually does big part of what is needed?
The procedure above is probably implemented in custom scripts (and likely in CMS combine?), the request here is if we could somehow automate this directly in RooFit.
Alternatives considered
No response
Additional context
No response
