How it works

These pages give the mathematics behind binference, for readers who want to know exactly what the code computes. They use the symbols of the notation page, and the defaults page holds every threshold they mention.

• The model: the model, its form with one latent label per object, and every assumption in one list.

• Point estimates: EM as iterated stacking, mix-SQP, the optimality certificate, missing information, and why the mode depends on the coordinates.

• Priors in detail: the Dirichlet family as trees, the logistic family, the unconstrained coordinates and the hyperpriors.

• Engines in detail: NUTS, Gibbs, Laplace with importance sampling, mean-field VB and the bootstrap.

• How auto chooses: the rules of method="auto" and the benchmark behind them.

• Validation: diagnostics, simulation-based calibration, coverage, and the statistical checks of the test suite.