Theorem 5 — Lipschitz regularization: on Borel
ProvedWassersteinDRO.Duality.lipschitz_regularization_v2Let with an arbitrary norm (a finite-dimensional real normed space) and its Borel -algebra, let be closed, let be a probability distribution on , let and , and let be a measurable, upper semicontinuous, -integrable loss function. Then the worst-case risk over the type- Wasserstein ball of radius around is bounded above by the Lipschitz-regularized nominal risk,
where and the right-hand side is formed in : if fails to be Lipschitz continuous and the bound is and the inequality holds trivially, as the paper remarks; if the product is and the bound still holds, since the radius- ball is or empty.
Formalization Note. The retired version was false because its -algebra on was a free parameter: with a non-Borel -algebra the transport cost is non-measurable, , and the radius- ball contains , which violates the bound at . The new statement does the following differently. (i) is a finite-dimensional real normed space (the paper's with an arbitrary norm) carrying its Borel -algebra ([FiniteDimensional ℝ E] [BorelSpace E]); the retired version quantified over every MeasurableSpace E, including -algebras unrelated to the norm, for which the Wasserstein cost is a lower integral and non-constant losses are integrable under no measure. (ii) The worst-case risk is the paper's with the paper's p. 2 convention for non-integrable losses (worstCaseRisk/nominalRisk v2, valued in ) instead of a supremum that silently discarded every under which is not integrable. Standing assumptions made explicit: closed (p. 6); a probability distribution; and (p. 3, 6); the loss is measurable (p. 1) and, by Assumption 1 (p. 9, "tacitly assumed to hold throughout the rest of the paper"), upper semicontinuous and -integrable. Losses are taken real-valued, a special case of the paper's extended-real-valued losses that the retired version already made. No correction to the printed source.
import Mathlib import Definitions.Def_WassersteinDRO_Duality_worstCaseRisk_v2 import Definitions.Def_WassersteinDRO_Duality_nominalRisk_v2 import Definitions.Def_WassersteinDRO_Duality_lipschitzModulus open MeasureTheory
namespace WassersteinDRO.Duality
/-- Theorem 5 (Lipschitz regularization), Kuhn, Mohajerin Esfahani, Nguyen &
Shafieezadeh-Abadeh, *Wasserstein Distributionally Robust Optimization*, INFORMS TutORials
2019 (arXiv:1908.08729v2), p. 9: the worst-case risk (6) of any fixed loss function `ℓ ∈ L` is
bounded above by the Lipschitz-regularized nominal risk, `Rε,p(PN,ℓ) ≤ R(PN,ℓ) + ε·Lip(ℓ)`.
Setting (p. 3, 6, 9): `E` is `ℝ^m` with an arbitrary norm — a finite-dimensional real normed
space — carrying its Borel σ-algebra (`[BorelSpace E]`); `Ξ ⊆ E` is closed; `PN` is a
probability distribution; `p ∈ [1,∞)`, `ε ≥ 0`. The loss satisfies the paper's standing
conventions: `ℓ` is measurable (p. 1) and, by Assumption 1 (p. 9, "tacitly assumed to hold
throughout the rest of the paper"), upper semicontinuous and `PN`-integrable. The right-hand
side is formed in `EReal`/`ℝ≥0∞`, so a non-Lipschitz `ℓ` (`Lip(ℓ) = ∞`) with `ε > 0` makes the
bound `+∞` ("trivially satisfied", p. 9); with `ε = 0` the product is `0·∞ = 0` and the bound
`R_{0,p}(PN,ℓ) ≤ R(PN,ℓ)` still holds, since the radius-`0` ball is `{PN}` or empty.
Corrected from the retired `lipschitz_regularization`: the σ-algebra was a free
`[MeasurableSpace E]` unrelated to the norm topology (so `W_p` was not a metric and the bound
failed at `ε = 0`), `E` was not finite-dimensional, measurability/upper semicontinuity of `ℓ`
were missing, and the worst-case risk dropped non-integrable `Q` instead of using the paper's
convention `E_Q[ℓ] ∈ {-∞,+∞}` (now `worstCaseRisk`/`nominalRisk` v2). -/
theorem lipschitz_regularization_v2 {E : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E]
[FiniteDimensional ℝ E] [MeasurableSpace E] [BorelSpace E]
(ε p : ℝ) (hε : 0 ≤ ε) (hp : 1 ≤ p) (Ξ : Set E) (hΞcl : IsClosed Ξ)
(PN : Measure E) [IsProbabilityMeasure PN]
(ℓ : E → ℝ) (hℓm : Measurable ℓ) (hℓusc : UpperSemicontinuous ℓ) (hℓ : Integrable ℓ PN) :
worstCaseRisk ε p Ξ PN ℓ ≤
nominalRisk PN ℓ + ((ENNReal.ofReal ε * lipschitzModulus ℓ : ENNReal) : EReal) := by sorry
end WassersteinDRO.Duality