Theorem 4.9, p. 20 — under Assumptions 1–3, f_k^primal ↘ f⋆ and f_k^dual ↘ f⋆, and limits of dual minimizers give global minimizers
ProvedRobustPMI.Hierarchy.theorem_4_9This is the asymptotic convergence theorem for the moment-SOS hierarchy (4.4)–(4.5) of the robust polynomial matrix inequality problem
Suppose (RPMIO) has an optimal solution and Assumptions 1–3 hold: , are SOS-convex, is PSD-SOS-convex for every , is compact, the quadratic module of is Archimedean in the sense of Assumption 2, and Slater's condition holds. Then:
- and as : both sequences are non-increasing for and converge to .
- Let be a minimizer of (4.5) at order . For any convergent subsequence with limit , the point is a global minimizer of (RPMIO).
- Consequently, if (RPMIO) has a unique global minimizer , then .
The theorem says that the hierarchy solves the robust problem in the limit, both in value and in the minimizer, although (RPMIO) has infinitely many matrix constraints and is in general hard.
Formalization Note The values are extended reals and the limits are taken in the extended reals with coerced. The paper presupposes that each exists; the Lean statement takes a sequence that is a minimizer of (4.5) for all sufficiently large . The moment vectors of (4.5) are indexed by , so convergence of the subsequence is required on exactly those coordinates. The standing assumption of p. 2 (an optimizer exists) and the symmetry of and are hypotheses.
import Mathlib import Definitions.Def_RobustPMI_Hierarchy_Model
namespace RobustPMI.Hierarchy
open Filter Topology
/-- Theorem 4.9 (p. 20): under Assumptions 1–3 (and the standing assumption that (RPMIO) has an
optimizer),
(i) `f_k^primal ↘ f⋆` and `f_k^dual ↘ f⋆` as `k → ∞`;
(ii) if `s^{(k,⋆)}` is a minimizer of (4.5) for all large `k`, then for every convergent subsequence
`s^{(k_i,⋆)} → s⋆`, `s⋆_e` is a global minimizer of (RPMIO); consequently, if (RPMIO) has a unique
global minimizer `y⋆`, then `s_e^{(k,⋆)} → y⋆`. -/
theorem theorem_4_9 {ℓ n m q s : ℕ} (f : MvPolynomial (Fin ℓ) ℝ)
(θ : Fin s → MvPolynomial (Fin ℓ) ℝ)
(P : Matrix (Fin m) (Fin m) (MvPolynomial (Fin ℓ ⊕ Fin n) ℝ))
(G : Matrix (Fin q) (Fin q) (MvPolynomial (Fin n) ℝ))
(hP : P.IsSymm) (hG : G.IsSymm)
(h1 : Assumption1 f θ P G) (h2 : Assumption2 G) (h3 : Assumption3 θ P G)
(hmin : ∃ y, IsGlobalMin f θ P G y) :
(AntitoneOn (fprimal f θ P G) (Set.Ici ((kx P G + 1) / 2)) ∧
AntitoneOn (fdual f θ P G) (Set.Ici ((kx P G + 1) / 2)) ∧
Tendsto (fprimal f θ P G) atTop (𝓝 (fstar f θ P G : EReal)) ∧
Tendsto (fdual f θ P G) atTop (𝓝 (fstar f θ P G : EReal))) ∧
(∀ sq : ℕ → (Fin ℓ →₀ ℕ) → ℝ, (∀ᶠ k in atTop, IsDualMinimizer f θ P G k (sq k)) →
∀ φ : ℕ → ℕ, StrictMono φ → ∀ sstar : (Fin ℓ →₀ ℕ) → ℝ,
(∀ α ∈ monos (σ := Fin ℓ) (2 * ((ky f θ P + 1) / 2)),
Tendsto (fun i => sq (φ i) α) atTop (𝓝 (sstar α))) →
IsGlobalMin f θ P G (sE sstar)) ∧
(∀ sq : ℕ → (Fin ℓ →₀ ℕ) → ℝ, (∀ᶠ k in atTop, IsDualMinimizer f θ P G k (sq k)) →
∀ ystar : Fin ℓ → ℝ, (∀ y, IsGlobalMin f θ P G y ↔ y = ystar) →
Tendsto (fun k => sE (sq k)) atTop (𝓝 ystar)) := by sorry
end RobustPMI.Hierarchy
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.