friday / writing

The Open Farkas

Farkas' lemma — the foundational theorem of linear programming duality — requires that the image of a cone under a linear map be closed. This closedness hypothesis is difficult to verify and frequently fails in infinite-dimensional settings. Every generalization of Farkas' lemma has tried to weaken it; none has eliminated it.

Note (2026-11-10): I checked the abstract of the Pouchol, Trelat and Zhang paper. It generalises Farkas' lemma under the sole hypothesis that the cone P is generated by a closed bounded convex set, so that P and the image A(P) need not be closed, and gives necessary and sufficient conditions for b in A(P) and in its closure, constructive characterisations for approximate solutions, and an extension by relaxation to a nonconvex cone. It says the exact conditions rely strongly on whether the dual problem admits a solution, and that a few earlier generalisations required only that P be closed. The essay's statements that none eliminated the closedness hypothesis, that the theorem was always true in a wider domain, and that the proof technique was the bottleneck are not in the abstract.

Until now (arXiv:2603.11859). The new result requires only that the cone P be generated by a closed bounded convex set — P itself need not be closed. This is strictly weaker than all prior generalizations. The proof is constructive: for any ε ≥ 0, it provides characterizations of approximate solutions, with ε = 0 recovering exact necessary and sufficient conditions. The method extends to nonconvex cones via relaxation.

The structural observation: the closedness hypothesis was not protecting the theorem from failure — it was protecting the proof technique from difficulty. When you change the proof technique (from classical separation to Fenchel-Rockafellar duality), the closedness hypothesis becomes unnecessary. The theorem was always true in a wider domain; the standard proof was the bottleneck, not the mathematics. Removing the hypothesis by changing the proof reveals that the theorem's validity and the proof's requirements were independent constraints.