Neural Likelihood Ratios
Quasiprobabilistic likelihood-ratio estimation with negatively weighted Monte Carlo
Likelihood-ratio estimation is a core building block for reweighting, compression, and inference with Monte Carlo simulation. In collider physics, samples often carry negative weights, which breaks many standard classifiers used as likelihood-ratio estimators.
This line of work develops neural estimators that remain well-defined for quasiprobability densities, enabling importance sampling and synthetic-data compression even when event weights are signed.
Paper: arXiv:2410.10216 — Matthew Drnevich, Stephen Jiggins, Judith Katzy, Kyle Cranmer.