App. A, proof of Lemma 5 — explicit minimization of ½ΔᵀHΔ + wᵀΔ over {CΔ = 0}
ProvedALADIN.DualDecomp.explicit_min_dylinear-algebrap2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1quadratic-programming
Let be symmetric positive definite, of full row rank , and . Put
Then
- is invertible;
- satisfies ;
- minimizes over ;
- the minimum value is ;
- .
In the proof of Lemma 5 this is applied blockwise with , , : summing the minimum values gives with from (the corrected) (A.6), and item 5 shows that is negative semidefinite.
Preamble
import Mathlib
Formal statement
namespace ALADIN.DualDecomp
open Matrix
/-- App. A, proof of Lemma 5, p. 1124 ("we solve the remaining minimization problem over Δy
explicitly"): for `H ∈ ℝ^{n×n}` positive definite, `C ∈ ℝ^{nh×n}` of full row rank and `w ∈ ℝⁿ`, put
`P = H⁻¹ − H⁻¹ Cᵀ (C H⁻¹ Cᵀ)⁻¹ C H⁻¹`. Then `C H⁻¹ Cᵀ` is invertible, `Δ⋆ = −P w` satisfies `C Δ⋆ = 0`
and minimizes `½ Δᵀ H Δ + wᵀ Δ` over `{Δ | C Δ = 0}`, the minimum value is `−½ wᵀ P w`, and `wᵀ P w ≥ 0`. -/
theorem explicit_min_dy {n nh : ℕ} (H : Matrix (Fin n) (Fin n) ℝ) (C : Matrix (Fin nh) (Fin n) ℝ)
(hH : H.PosDef) (hC : C.rank = nh) (w : Fin n → ℝ) :
let P := H⁻¹ - H⁻¹ * Cᵀ * (C * H⁻¹ * Cᵀ)⁻¹ * C * H⁻¹
IsUnit (C * H⁻¹ * Cᵀ).det ∧
C *ᵥ (-(P *ᵥ w)) = 0 ∧
IsMinOn (fun Δ : Fin n → ℝ => 1 / 2 * (Δ ⬝ᵥ (H *ᵥ Δ)) + w ⬝ᵥ Δ) {Δ | C *ᵥ Δ = 0}
(-(P *ᵥ w)) ∧
1 / 2 * ((-(P *ᵥ w)) ⬝ᵥ (H *ᵥ (-(P *ᵥ w)))) + w ⬝ᵥ (-(P *ᵥ w)) = -(1 / 2) * (w ⬝ᵥ (P *ᵥ w)) ∧
0 ≤ w ⬝ᵥ (P *ᵥ w) := by sorry
end ALADIN.DualDecomp
Source
Houska, Frasch, Diehl, An augmented Lagrangian based algorithm for distributed nonconvex optimization, SIAM J. Optim. 26 (2016), p. 1124, App. A, proof of Lemma 5, explicit minimization over Δy
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.