Proposition 1 — strong duality for the worst-case expected loss over an optimal-transport ball, with the minimum attained
ProvedRWPI.SqrtLasso.proposition_1Let be training data, , with empirical distribution . Let be a lower semicontinuous cost with for every , and let be an upper semicontinuous loss. For let
(terms with infinite cost do not contribute). Then for every
and the minimum is attained: there is achieving equality, and the left side is at most the bracket for every .
The proposition turns the supremum over uncountably many probability measures into a one-dimensional convex minimization, in which is often explicit. The paper quotes it from Blanchet and Murthy (Math. Oper. Res., 2019, Theorem 1). Taking the infimum over the regression parameter on both sides gives the paper's identity (12); in the paper the loss is , and the statement here holds for each fixed .
Formalization Note Two hypotheses are added to the page and recorded: (every loss in the paper is nonnegative; the expectation is a lower Lebesgue integral), and (assumed in Blanchet–Murthy; at the minimum need not be attained). Both sides take values in ; "min" is stated as attainment plus a lower bound. Identity (12) is not stated separately. The data space is with the product topology.
import Mathlib import Definitions.Def_WassersteinDRO_Regularization_empiricalDistribution import Definitions.Def_RWPI_SqrtLasso_worstCase import Definitions.Def_RWPI_SqrtLasso_phi open MeasureTheory
namespace RWPI.SqrtLasso
/-- Proposition 1 (Blanchet, Kang & Murthy, p. 10; quoted from Blanchet & Murthy, Math. Oper. Res.
2019, Theorem 1), first identity: strong duality for the worst-case expected loss over the
optimal-transport ball `{P : D_c(P, P_n) ≤ δ}` around the empirical distribution `P_n` of the data
`(X_i, Y_i)`, with the minimum over `γ ≥ 0` attained.
The cost `c` is lower semicontinuous, `[0, ∞]`-valued and vanishes on the diagonal; the loss `l` is
upper semicontinuous. Added hypotheses (recorded): `l ≥ 0` (every loss of the paper) and `δ > 0`
(assumed in Blanchet–Murthy; at `δ = 0` the minimum need not be attained). -/
theorem proposition_1 {d n : ℕ} (hn : 0 < n) (X : Fin n → Fin d → ℝ) (Y : Fin n → ℝ)
(c : (Fin d → ℝ) × ℝ → (Fin d → ℝ) × ℝ → ENNReal)
(hc : LowerSemicontinuous (Function.uncurry c)) (hc0 : ∀ z, c z z = 0)
(l : (Fin d → ℝ) × ℝ → ℝ) (hl : UpperSemicontinuous l) (hl0 : ∀ z, 0 ≤ l z)
(δ : ℝ) (hδ : 0 < δ) :
∃ γ : ℝ, 0 ≤ γ ∧
worstCase c δ (WassersteinDRO.Regularization.empiricalDistribution
(fun i : Fin n => (X i, Y i))) l =
ENNReal.ofReal (γ * δ) + (n : ENNReal)⁻¹ * ∑ i, phi c l γ (X i, Y i) ∧
∀ γ' : ℝ, 0 ≤ γ' →
worstCase c δ (WassersteinDRO.Regularization.empiricalDistribution
(fun i : Fin n => (X i, Y i))) l ≤
ENNReal.ofReal (γ' * δ) + (n : ENNReal)⁻¹ * ∑ i, phi c l γ' (X i, Y i) := by sorry
end RWPI.SqrtLasso
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.