§3.2, p. 12 — the μth equality constraint of 𝒮𝒫^u holds P-a.s. iff H_μ = 0 (Ξ with relative interior)
ProvedPrimalDualLDR.RandomRecourse.claim_3_13Assume the standing assumptions of §3: is the support of the probability measure , it is nonempty and bounded, it spans , and every is symmetric. Assume in addition that has a relative interior point: some and satisfy whenever and .
Let , let be a symmetric matrix and , and set
Then
This turns each semi-infinite equality constraint of into a finite linear matrix equation, which is the equality constraint of the SDP (3.14).
Formalization Note The relative-interior hypothesis is a repair. The paper deduces "the interior of is nonempty" from " spans ", which needs to have dimension ; for a nonconvex of type (3.11) with this fails (e.g. and cut out by and , where vanishes on ). The paper states on p. 10 that spanning is equivalent to "dimension ", and the added hypothesis is that reading. For it follows from the standing assumptions. is assumed symmetric, as in ; without symmetry only the symmetric part of would be determined.
import Mathlib import Definitions.Def_PrimalDualLDR_RandomRecourse_Problems open MeasureTheory Matrix
namespace PrimalDualLDR.RandomRecourse
/-- **Kuhn, Wiesemann, Georghiou (preprint 2009), §3.2, p. 12 (unnumbered, around (3.13)).**
Under the standing assumptions of §3, for a decision `X ∈ ℝ^{n×k}`, a symmetric slack matrix
`S_μ ∈ 𝕊` and a row `μ`, the `μ`th equality constraint of `𝒮𝒫^u`,
`ξᵀA_μXξ + ξᵀS_μξ = b_μᵀξ` `P`-almost surely, holds if and only if
`H_μ = ½(A_μX + XᵀA_μᵀ − e_1b_μᵀ − b_μe_1ᵀ) + S_μ = 0`.
Repair: the hypothesis `hrelint` (some point of `Ξ` has a neighbourhood, relative to the
hyperplane `{e_1ᵀξ = 1}`, contained in `Ξ`) is added. The paper derives "the interior of
`cone(Ξ)` is nonempty" from "`Ξ` spans `ℝ^k`", which fails when `Ξ` is not convex (for `l ≥ 2`,
`Ξ` can be two points); `hrelint` is the paper's "`Ξ` has dimension `k − 1`" (p. 10). For
`l = 1` it follows from the standing assumptions. -/
theorem claim_3_13 (σ : Setting) (hσ : σ.Standing)
(hrelint : ∃ ξ₀ ∈ σ.Xi, ∃ r : ℝ, 0 < r ∧
∀ ξ : Fin σ.k → ℝ, σ.e1 ⬝ᵥ ξ = 1 → dist ξ ξ₀ < r → ξ ∈ σ.Xi)
(X : Matrix (Fin σ.n) (Fin σ.k) ℝ) (Sμ : Matrix (Fin σ.k) (Fin σ.k) ℝ) (hS : Sμ.IsSymm)
(μ : Fin σ.m) :
(∀ᵐ ξ ∂σ.P, ξ ⬝ᵥ ((σ.A μ * X) *ᵥ ξ) + ξ ⬝ᵥ (Sμ *ᵥ ξ) = σ.B μ ⬝ᵥ ξ) ↔
σ.H X Sμ μ = 0 := by sorry
end PrimalDualLDR.RandomRecourse
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.