Work

The Potapov–Grishin Principal Hyperface Method, Completed and Machine-Checked

Year: 2026
Type: preprint
Abstract: The principal hyperface method of G. G. Potapov and I. Yu. Grishin (2007) solves a linear program by a sequence of geometric decisions. It picks the constraint hyperplane whose normal is closest in direction to the objective vector, turns that inequality into an equality, eliminates one variable, and repeats, so that the dimension drops by one at each step. The method is short and easy to picture, but it was published as a heuristic, without a proof of when it is correct. This paper completes the method and verifies every result in the Lean 4 proof assistant. We prove that a run of the method returns an optimal point if and only if every hyperface it selects is principal, that is, active at some optimum. We prove that a principal hyperface always exists when the objective is non-zero. We prove that in two dimensions the projection rule of the method never errs on systems without redundant constraints, and we give a machine-checked counterexample in three dimensions. Finally, we turn the method into a certified algorithm: it follows the projection rule, checks a dual certificate after each step, and backtracks when the check fails. We prove that it always returns an optimal point when one exists, and every certified answer can be checked independently. Along the way we formalise Farkas' lemma over an arbitrary ordered field by Fourier-Motzkin elimination. The worked example of the 2008 paper is confirmed exactly. A Python implementation of the same algorithm in exact rational arithmetic accompanies the paper. Code, Lean formalisation and figures: https://github.com/albond/principal-hyperfaces
Language: English
Open Access status: green
OpenAlex ID: W7233726485
Cites:
Cited by:
Related to: 10