Investigate difference between unbinned fit with PDFs versus normalization in the estimator

[redeboer]

ComPWA (C++) and TensorWaves do their unbinned negative log likelihood fit with pure functions as opposed to working with Probability Density Functions (PDFs). The normalization in each fit step therefore takes place in the estimator and not in the function. Fit fractions are then computed by integrating at the end, not by reading off coefficients.

Some questions to answer:

• Does this approach result in the same observables, like fit fractions (in our experience: yes)?
• Is there a performance win or loss?