Lemma 17 — every feasible plan is improved by one concentrated on {f(y) ≥ f(x)}
ProvedModelRiskOT.PrimalOpt.lemma_17distributionally-robust-optimizationoptimal-transportp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1
Let be a Polish space, a probability measure on , , and let and satisfy (A1) and (A2). Then for every there is concentrated on , i.e. , such that
and moreover whenever is well defined.
The lemma shows that the primal problem may be restricted to monotone plans, and that its value is never an ambiguity.
Formalization Note "Concentrated on" means measure one (p. 4). The last inequality is stated for every , as in EReal: when is not well defined (), by the EReal convention, so the unconditional statement is equivalent to the paper's.
Preamble
import Mathlib import Definitions.Def_ModelRiskOT_PrimalOpt_PrimalFeasibleMono
Formal statement
namespace ModelRiskOT.PrimalOpt
open MeasureTheory
/-- **Lemma 17** (App. B.4, p. 43). Under (A1) and (A2), for every `π ∈ Φ_{μ,δ}` there is
`π′ ∈ Φ_{μ,δ}` concentrated on `{(x, y) : f(y) ≥ f(x)}` (i.e. `π′ ∈ Φ′_{μ,δ}`) with
`∫ f⁺(y) dπ′ ≥ ∫ f⁺(y) dπ` and `∫ f⁻(y) dπ′ ≤ ∫ f⁻ dμ` (55), and `I(π′) ≥ I(π)`. The last
inequality is stated for every `π`: when `∫ f(y) dπ` is not well defined (`∞ − ∞`), `I(π) = ⊥`. -/
theorem lemma_17 {S : Type*} [TopologicalSpace S] [PolishSpace S] [MeasurableSpace S]
[BorelSpace S] (c : S → S → ℝ) (f : S → ℝ) (μ : Measure S) [IsProbabilityMeasure μ]
(δ : ℝ) (hδ : 0 < δ) (hA1 : ModelRiskOT.Duality.AssumptionA1 c) (hA2 : AssumptionA2 f μ)
(π : Measure (S × S)) (hπ : π ∈ ModelRiskOT.Duality.primalFeasible c μ δ) :
∃ π' ∈ primalFeasibleMono c f μ δ,
∫⁻ p, ENNReal.ofReal (f p.2) ∂π ≤ ∫⁻ p, ENNReal.ofReal (f p.2) ∂π' ∧
∫⁻ p, ENNReal.ofReal (-f p.2) ∂π' ≤ ∫⁻ x, ENNReal.ofReal (-f x) ∂μ ∧
ModelRiskOT.Duality.primalObj f π ≤ ModelRiskOT.Duality.primalObj f π' := by sorry
end ModelRiskOT.PrimalOpt
Source
Blanchet & Murthy, Quantifying Distributional Model Risk via Optimal Transport, arXiv:1604.01446v2, p. 43, Appendix B.4, Lemma 17, Eq. (55)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.