(A.2), proof of Theorem 3.7 — cluster points of -minimizers of epi-convergent functions minimize the limit
ProvedRockafellianRelax.eq_A2_clusterPt_mem_argminoptimizationp2o-gran-per-chapterp2o-plan-paperp2o-strengthenedp2o-v1stochastic-optimizationvariational-analysis
Let be a first-countable topological space, and let with epi-converging to (in the sense of (3.8)–(3.9)) and . Let with , and
Then every cluster point of satisfies
in particular . No properness of is assumed: may take the value . In the paper this is applied with , , , where follows from .
Differences from the source:
- generalization: ℝᵐ × ℝⁿ, the tilted functions f̃^ν, f̃ of the proof, with f̃ ≢ ∞ deduced from dom φ ≠ ∅ → any first-countable topological space α, any ψ^ν epi-converging to ψ, with dom ψ ≠ ∅; reason: ℝᵐ × ℝⁿ is first countable and f̃(0, x) = φ(x), so the paper's case is an instance
- generalization: cases (i) f̃ proper ([59, Theorem 5.5(b)]) and (ii) f̃ takes −∞ somewhere, treated separately → one statement with no properness assumption; reason: cases (i) and (ii) together cover every f̃ with dom f̃ ≠ ∅; the conclusion (A.2) is the same in both
- generalization: ε^ν ↘ 0 → ε^ν ≥ 0 and ε^ν → 0 (not necessarily monotone); reason: a nonincreasing sequence tending to 0 is such a sequence
- encoding: R̄, dom ψ = {ψ < ∞}, cluster point of {z^ν, ν ∈ ℕ} → EReal, AubinLipschitz.effDom ψ = {ψ ≠ ⊤}, MapClusterPt ẑ atTop z, ν ∈ Lean's ℕ from 0; reason: Lean convention; same instances
Preamble
import Mathlib import Definitions.Def_AubinLipschitz_effDom import Definitions.Def_RockafellianRelax_argmin import Definitions.Def_PerspEpi_SeqEpiConv open Filter Topology
Formal statement
namespace RockafellianRelax
/-- **Cluster points of approximate minimizers of epi-convergent functions minimize the limit**,
display (A.2). Royset, Chen & Eckstrand, arXiv:2204.04762v4, Proof of Theorem 3.7, display (A.2),
p.37: "We consider three cases. (i) If f̃ is proper, then [59, Theorem 5.5(b)] applies and we
conclude that (û, x̂) ∈ argmin_{u,x} f̃(u, x). (A.2) (ii) If f̃ takes the value −∞ at one or more
points, then the arguments in the proof of [59, Theorem 5.5(a,b)] carry over and we again conclude
that (A.2) holds. (iii) The case f̃ being identical to ∞ is ruled out by the assumption dom φ ≠ ∅."
Let `ψν` epi-converge to `ψ`, with `dom ψ ≠ ∅`, let `εν ≥ 0` with `εν → 0` and
`zν ∈ εν-argmin ψν`. Then every cluster point `ẑ` of `(zν)` lies in `argmin ψ`.
**Formalization Note.** Stated on any first-countable topological space `α` (the paper's case is
`α = ℝᵐ × ℝⁿ` with `ψν = f̃ν`, `ψ = f̃`); cases (i) and (ii) are covered at once, without
properness, and case (iii) is the hypothesis `dom ψ ≠ ∅`. `argmin` requires `ψ(ẑ) < +∞`. A cluster
point is `MapClusterPt ẑ atTop z`. -/
theorem eq_A2_clusterPt_mem_argmin {α : Type*} [TopologicalSpace α] [FirstCountableTopology α]
(Ψ : ℕ → α → EReal) (ψ : α → EReal) (hepi : PerspEpi.SeqEpiConv Ψ ψ)
(hdom : (AubinLipschitz.effDom ψ).Nonempty) (ε : ℕ → ℝ) (hε0 : ∀ ν, 0 ≤ ε ν)
(hε : Tendsto ε atTop (𝓝 0)) (z : ℕ → α) (hz : ∀ ν, z ν ∈ epsArgmin (Ψ ν) (ε ν))
(zhat : α) (hzhat : MapClusterPt zhat atTop z) :
zhat ∈ argmin ψ := by sorry
end RockafellianRelax
Source
Royset, Chen & Eckstrand, Rockafellian Relaxation and Stochastic Optimization under Perturbations, arXiv:2204.04762v4, Proof of Theorem 3.7, display (A.2), p.37
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.