The day in conversation for 2026-10-15

discussion

Transcript

The discussion: Kai: It’s interesting how simplex demixing allows us to move away from simple two-category splits; framing the classification as a bounded geometric object with T vertices seems like it provides a much richer structure for separating those down, up, and gluon fractions simultaneously. But I wonder if that geometric representation inherently introduces more complexity in the fitting process compared to simpler likelihood methods we used on synthetic data. Mira: The complexity is certainly there, Kai, but I think the advantage lies in its inherent ability to identify maximally separable categories before you even start fitting. When we look at the previous works focusing on data-driven operational definitions—the ones that tried to define flavors based purely on tag-and-probe efficiency thresholds—we often ended up with mutually exclusive or overlapping bins. Does

More episodes

← Home