R/selection-likelihood-plan.R
selection_event_support.RdChecks whether a Gaussian candidate law can reach a positive-weight event for almost every retained latent context. The complete candidate-plus-retained covariance must be positive definite. Retained covariance itself may be singular. This is a support check, not a numerical integration diagnostic.
selection_event_support(prior, candidate_basis)a weightfunction prior specifying one publication group's rule.
finite numeric matrix whose column space is exactly the candidate variation in that group. Zero columns are allowed, including a matrix with no columns. Use compiler-declared source geometry, not posterior covariance estimates or numerical rank truncation.
A list with feasible (TRUE, FALSE, or NA) and a stable
character reason. FALSE proves that some positive-probability retained
contexts have zero acceptance probability. NA means that a positive
direction could not be certified. Reasons are positive_weights,
varying_best_result, varying_two_tails, positive_direction,
deterministic_candidate, deterministic_candidate_rows,
opposed_candidate_direction, or positive_direction_unavailable.
Positive weights require no candidate variation. With hard-zero bins, the best-result rule needs at least one varying result. Two-sided product selection, and one-sided product selection with a positive least-significant weight, need every result to vary. One-sided product selection with a zero least-significant weight instead needs a strictly positive vector in the candidate column space.
The positive-direction check uses sign-definite columns, then base-R QR and optimization as candidate finders. A computed direction is accepted only when its positive entries exceed a standard dot-product roundoff bound. Failed searches return unknown rather than proving infeasibility.
The caller must check every candidate span having positive prior probability, including zero-SD and inclusion branches. When only a guaranteed subspace is supplied, only a positive result is conclusive; other results cannot classify the complete candidate law. This helper does not derive scale guarantees or change weights, priors, or source geometry.