App. A, proof of Lemma 5 — with ρ = 0 the step-1 output satisfies 0 = ∇fᵢ(yᵢ) + Aᵢᵀλ + (C*ᵢ)ᵀκᵢ = gᵢ + Aᵢᵀλ + Cᵢᵀκᵢ
ProvedALADIN.DualDecomp.stationaritydual-decompositionkkt-conditionsp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1
Let be an instance of problem (1.1) and fix a block . Suppose and are twice continuously differentiable, and is a step-1 output of Algorithm 2 for block with penalty parameter (any scaling matrix , any , dual iterate ), so that solves (A.2) with multiplier . Then for every matrix ,
where is the active-constraint Jacobian at and is the modified gradient of step 3.
This is the stationarity condition of (A.2) rewritten with the modified gradient; in the proof of Lemma 5 it eliminates from the dual of the QP (3.3).
Preamble
import Mathlib import Definitions.Def_ALADIN_DualDecomp_Problem import Definitions.Def_ALADIN_DualDecomp_Step
Formal statement
namespace ALADIN.DualDecomp
open Matrix
/-- App. A, proof of Lemma 5, p. 1123: if `(yᵢ, κᵢ)` is the output of Algorithm 2, step 1 with `ρ = 0`
(so (3.2) is (A.2)), and `fᵢ`, `hᵢ` are twice continuously differentiable, then for every
choice of `Cᵢ`,
`0 = ∇fᵢ(yᵢ) + Aᵢᵀ λ + (C*ᵢ)ᵀ κᵢ = gᵢ + Aᵢᵀ λ + Cᵢᵀ κᵢ`. -/
theorem stationarity {N n m nh : ℕ} (P : Problem N n m nh) (Sig : Matrix (Fin n) (Fin n) ℝ)
(x : Fin n → ℝ) (lam : Fin m → ℝ) (i : Fin N) (yi : Fin n → ℝ) (κi : Fin nh → ℝ)
(Ci : Matrix (Fin nh) (Fin n) ℝ)
(hf : ContDiff ℝ 2 (P.f i)) (hh : ContDiff ℝ 2 (P.h i))
(hstep1 : Step1Solution P 0 Sig x lam i yi κi) :
grad (P.f i) yi + (P.A i)ᵀ *ᵥ lam + (Cstar P i yi)ᵀ *ᵥ κi = 0 ∧
modGrad P i yi κi Ci + (P.A i)ᵀ *ᵥ lam + Ciᵀ *ᵥ κi = 0 := by sorry
end ALADIN.DualDecomp
Source
Houska, Frasch, Diehl, An augmented Lagrangian based algorithm for distributed nonconvex optimization, SIAM J. Optim. 26 (2016), pp. 1123–1124, App. A, proof of Lemma 5, stationarity display
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.