Particle identification and misidentification

Let x collect the measured identification variables. A likelihood ratio compares their joint density under an electron hypothesis with the density under a pion hypothesis. A cut on that ratio can define an electron-like sample. This is a response-based decision, not proof that each accepted object is an electron. The optimal threshold depends on the analysis objective, the relevant competing species, and their expected abundances; a working point optimized for one search need not suit another.

Combining several measurements requires their joint distribution or a validated approximation to it. Multiplying one-dimensional likelihoods while ignoring strong correlations can double-count information. A neural-network score can capture correlations but is not automatically a posterior species probability, even if its output lies between zero and one. Training class fractions, loss choice, calibration, and changes in detector conditions affect that interpretation. Validation should inspect the populations and kinematic region in which the score will be used.

Bayes’ rule converts pass probabilities into the composition after selection. If the preselection electron fraction is pi e and only electrons and pions are present, accepted electron weight is that fraction times electron efficiency. Accepted pion weight is the complementary fraction times the fake probability. Dividing electron weight by their sum gives purity. A small fake rate can still dominate when the background species is sufficiently more common.

The receiver-operating curve describes efficiency versus fake rate as the cut varies, but it does not specify an experiment’s base rates. Moving along that curve trades accepted signal against contamination. With multiple background species, the denominator contains a sum of their abundance-weighted pass rates. An analysis that changes its event selection can change those abundances without changing the classifier itself, so its selected purity may move even when detector-level efficiency and misidentification probabilities remain stable.

Control decays with well-constrained species content allow identification efficiencies to be measured in data. Tag-and-probe methods constrain one object strongly and test another with a looser identification requirement, while fitting or subtracting residual backgrounds. Such samples must not impose the tested decision on every probe. Momentum, direction, track multiplicity, and detector period can all change response; reweighting or binning can help transfer measured rates to a different analysis population.

For an event containing two candidates, multiplying single-object efficiencies is valid only after conditioning on variables that make their decisions sufficiently independent, or after explicitly modeling their correlation. Shared occupancy, common calibration shifts, or overlapping showers can violate that assumption. Data-to-simulation scale factors also carry uncertainty and correlations between bins. Reporting only a global purity hides these dependencies; a usable measurement identifies the selection, true-species definition, calibration sample, and response uncertainties.