search

LSSTDESC / qp

Quantile Parametrization for probability distribution functions module
MIT License
10 stars 3 forks source link

Extension of qp.Ensemble for non-uniform parameterizations (especially number of parameters) #27

Open aimalz opened 3 years ago

aimalz commented 3 years ago

@johannct reminded me today that neither the original nor current versions of qp support ensembles where PDFs may not have the same parameterization or even number of parameters. There are certainly use cases for both kinds of mixed-membership ensemble situations, but the latter is an immediate issue for the SparsePZ representation. It seems like this could be solved with some kind of composite ensemble containing multiple qp.Ensemble objects, each of whose members share parameterization and number of parameters with one another but differ from the other ensembles in the composite. I guess its biggest difference from a regular qp.Ensemble would be some sort of metadata for mapping the members of each ensemble within the composite to some global membership ordering that could be used for indexing. (Also, composite might not be the best word for this idea!)

aimalz commented 3 years ago

I just happened upon this old LSST Community post mentioning a Python library for non-uniformly shaped arrays, which might be applicable here?

eacharles commented 11 months ago

Just to be clear, this is more about jagged arrays (i.e., different number of parameters) , not about different parameterizations (i.e., having gaussians and histogram representations in the same ensemble.)