Theorem 5.4, p. 21 — values and optimal solutions of the Moreau-regularized problems (Pλ) converge to those of (P)
ProvedMoreauChance.Consistency.theorem_5_4Let be a separable Hilbert space, a Banach space, a random vector whose law has a density with respect to Lebesgue measure, and let (a nonempty closed convex cone), , satisfy the conic setting (9)–(10) with continuous and sequentially weakly continuous. Fix and let be proper, convex, coercive and lower semicontinuous with . Consider
with values , . Then:
- and for every , and as .
- Let and let be an optimal solution of for every . If , then is an optimal solution of (P). If furthermore and is Fréchet differentiable at every point of , then in norm.
This is the consistency result of the paper: the doubly regularized problems, whose data are smooth, recover both the optimal value and the optimal solutions of the original nonsmooth chance-constrained problem.
Formalization Note Two hypotheses are added to the page. (i) is closed, used through Proposition 2.6. (ii) Sequential weak continuity of is assumed for the whole theorem; the page states it only in item 2, but the proof of item 1 uses the weak closedness of (Theorem 3.1 d), whose proof needs it) and the Mosco convergence of Proposition 5.1 b). With finite-dimensional the hypothesis is automatic. "" is stated as "neither nor ". Item 2 is vacuous when some has no optimal solution, exactly as on the page ("let be any sequence of optimal solutions"). Fréchet differentiability of the extended-valued at includes finiteness of near .
import Mathlib import Definitions.Def_MoreauChance_Consistency_Basic import Definitions.Def_MoreauChance_Consistency_Setting open MeasureTheory Filter Topology open scoped InnerProductSpace ENNReal
namespace MoreauChance.Consistency
theorem theorem_5_4
{H : Type*} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
[TopologicalSpace.SeparableSpace H]
{Y : Type*} [NormedAddCommGroup Y] [NormedSpace ℝ Y] [CompleteSpace Y]
{Ω : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ]
{m : ℕ} (ξ : Ω → EuclideanSpace ℝ (Fin m)) (hξ : Measurable ξ)
(hdens : Measure.map ξ μ ≪ volume)
(K : Set Y) (C : Set (StrongDual ℝ Y)) (h : H → ℝ) (Φ : H → EuclideanSpace ℝ (Fin m) → Y)
(hset : IsConicSetting K C h Φ)
(hΦ : Continuous (fun q : H × EuclideanSpace ℝ (Fin m) => Φ q.1 q.2))
(h_weak : SeqWeakCont h)
(p : ℝ) (hp : p ∈ Set.Icc (0 : ℝ) 1)
(ψ : H → EReal) (hψ : IsGammaZero ψ) (hψ_coercive : IsCoercive ψ)
(hfeas : (feasSet K Φ μ ξ p ∩ {x | ψ x ≠ ⊤}).Nonempty) :
(valP ψ K Φ μ ξ p ≠ ⊥ ∧ valP ψ K Φ μ ξ p ≠ ⊤ ∧
(∀ lam : ℝ, 0 < lam →
valPReg ψ C h Φ μ ξ lam p ≠ ⊥ ∧ valPReg ψ C h Φ μ ξ lam p ≠ ⊤) ∧
Tendsto (fun lam => valPReg ψ C h Φ μ ξ lam p) (𝓝[>] 0) (𝓝 (valP ψ K Φ μ ξ p))) ∧
(∀ (lamk : ℕ → ℝ) (xk : ℕ → H) (x0 : H), (∀ k, 0 < lamk k) →
Tendsto lamk atTop (𝓝 0) →
(∀ k, IsOptimalPReg ψ C h Φ μ ξ (lamk k) p (xk k)) → WeakConv xk x0 →
IsOptimalP ψ K Φ μ ξ p x0 ∧
((∀ x, ψ x ≠ ⊤) →
(∀ u, (subdiff (conj ψ) u).Nonempty → FrechetDiffAtE (conj ψ) u) →
Tendsto xk atTop (𝓝 x0))) := by sorry
end MoreauChance.Consistency
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.