Minimum-coordinate elimination on a simplex: polynomial first-step safety and coNP-complete correctness
We study a deletion rule for strictly convex quadratic minimization on a simplex: compute the affine minimizer on the current face, permanently remove a most-negative coordinate, and repeat. Correctness is required for every linear term and every resolution of ties. We characterize first-step safety by a finite maximum principle for the constrained inverse, and correctness from every initial face by principal-inverse row inequalities. Correctness from the full simplex is a different property. The two properties coincide with at most four coordinates, but an explicit rational five-coordinate example separates them. An exact reachability condition explains the separation: boundary equalities in the constrained inverse restrict the tasks that can be entered on a facet. Away from these equalities, full-start and hereditary correctness coincide. We also determine the recognition complexity. First-step safety is decidable in polynomial bit time, by combining classical strict $n$-step-vector recognition with diagonal regularization. In contrast, full-start and hereditary correctness are coNP-complete, even on a reduction family with strict column diagonal maxima on every face. The proofs include the non-strict boundary and all affine anchors; explicit rational certificates for the finite examples are given in the text.