Theorem 11.11 — irregular L&P cuts are equivalent to no intersection cut and, if unique optimal, are strictly better
DisprovedDisjunctive.GeneralDisjunctions.completeness_via_domination_v2Let be a basic feasible solution of the CGLP (11.6) with for all . Suppose does not satisfy the condition of Theorem 11.9, and there is no basic feasible solution of (11.6) with for some that satisfies it. Then no intersection cut from (from any LP basis whose basic solution lies in ) is equivalent to . Furthermore, if uniquely minimizes over (11.6), then for every lift-and-project cut (i.e. is part of a feasible solution of (11.6)) that is equivalent to an intersection cut from ,
and does not dominate on .
Formalization Note. The retired (Open) version was unfaithful in several places, one of which makes it false: the competing cuts ranged over all pairs equivalent to an intersection cut, not over L&P cuts (feasible for (11.6)), so rescaling by a large violates the strict inequality. Also corrected: and are basic (book), not merely feasible; "intersection cut from " ranges over genuine LP bases (injective, nonsingular, ) rather than arbitrary maps with junk inverses; and is the LP relaxation, not an arbitrary set.
import Mathlib import Definitions.Def_Disjunctive_GeneralDisjunctions_Cglp_v2
namespace Disjunctive.GeneralDisjunctions
/-- Theorem 11.11 (Balas, *Disjunctive Programming*, §11.5, p. 164-165; Balas–Kis 2016,
Theorem 11): let `w̄ := (ᾱ, β̄, {ūᵗ, ūᵗ₀})` be a basic feasible solution of (11.6) with `ūᵗ₀ > 0`
for all `t`. If `w̄` does not satisfy the condition of Theorem 11.9, and there is no basic
feasible solution `w̃` of (11.6) with `(α̃, β̃) = μ(ᾱ, β̄)` for some `μ > 0` that satisfies it, then
no intersection cut from `S` is equivalent to `ᾱx ≥ β̄`. Furthermore, if `(ᾱ, β̄)` uniquely
minimizes `αx̄_N - β` over (11.6), then `ᾱx̄_N - β̄ < α̃x̄_N - β̃` for every L&P cut `α̃x ≥ β̃` (an
`(α, β)`-component of a feasible solution of (11.6)) equivalent to an intersection cut from `S`,
and no such cut dominates `ᾱx ≥ β̄` on `P = {x : Ãx ≥ b̃}`.
Corrected from the retired version: `w̄` and the rescaled `w̃` are *basic* feasible solutions;
"intersection cut from `S`" ranges over genuine LP bases (injective, nonsingular, `ā0 ∈ int S`)
instead of arbitrary maps `ι`; the competing cuts `α̃x ≥ β̃` are L&P cuts, i.e. feasible for
(11.6) (otherwise rescaling `(α̃, β̃)` makes `α̃x̄_N - β̃` arbitrarily negative); and `P` is the LP
relaxation `{x : Ãx ≥ b̃}` rather than an arbitrary set. -/
theorem completeness_via_domination_v2 {n : ℕ} {M T : Type*} [Fintype M] [Fintype T]
[Nonempty T] [DecidableEq M] [DecidableEq (Fin n)]
(Atil : Matrix M (Fin n) ℝ) (btil : M → ℝ) (d : T → Fin n → ℝ) (d0 : T → ℝ)
(xbarN : Fin n → ℝ)
(alphaBar : Fin n → ℝ) (betaBar : ℝ) (u : T → M → ℝ) (u0 : T → ℝ)
(hbasic : IsBasicCGLP116Solution Atil btil d d0 alphaBar u u0 betaBar)
(hu0pos : ∀ t, 0 < u0 t)
(hnot119 : ¬ SatisfiesThm119Condition Atil u)
(hnoscaled119 : ∀ μ : ℝ, 0 < μ → ∀ u' u0',
IsBasicCGLP116Solution Atil btil d d0 (μ • alphaBar) u' u0' (μ * betaBar) →
¬ SatisfiesThm119Condition Atil u') :
¬ IsEquivalentToIntersectionCutFromS Atil btil d d0 alphaBar betaBar ∧
(IsCGLP116UniqueMin Atil btil d d0 xbarN alphaBar betaBar →
∀ alphaT : Fin n → ℝ, ∀ betaT : ℝ,
(∃ uT u0T, IsCGLP116Feasible Atil btil d d0 alphaT uT u0T betaT) →
IsEquivalentToIntersectionCutFromS Atil btil d d0 alphaT betaT →
dotProduct alphaBar xbarN - betaBar < dotProduct alphaT xbarN - betaT ∧
¬ Dominates (LPPoly Atil btil) alphaT betaT alphaBar betaBar) := by sorry
end Disjunctive.GeneralDisjunctions