Theorem 2.5 — the minimax MMSE problem (2) over the Gaussian Wasserstein ball equals program (5); ψ⋆ is affine and N(µ, S⋆) is least favorable
ProvedWassKF.Reform.theorem_2_5Let with , and . Let with and , let , let , and let be the Wasserstein ambiguity set (3). Let be the family of all measurable functions . Consider the finite program
Then the minimax problem (2) is equivalent to (5) in the following sense.
- The optimal value of (2), , equals the optimal value of (5).
- If is optimal in (5), then the affine function
is measurable and attains the infimum in (2): it is a distributionally robust minimum mean square error estimator. 3. For the same , the normal distribution belongs to and attains the supremum in nature's problem : it is the least favorable prior.
The theorem reduces an infinite-dimensional minimax problem over a nonconvex set of distributions to a finite program over covariance matrices, and recovers both the robust estimator and the least favorable prior from its optimal solution.
Formalization Note "Equivalent" is not defined on the page; it is read as items 1–3, which cover both sentences of the theorem. Values of (2) and of nature's problem are in (lower Lebesgue integrals); the value of (5) is the published sdpValue in EReal, compared through the coercion EReal. Program (5) is the published sdpFeasibleSet ρ Σ σ with objective sdpObjective; holds by symmetry of . is passed as the least eigenvalue of . The word "convex" in "finite convex program" is not part of the formal statement. Neither existence nor uniqueness of an optimal is asserted: items 2 and 3 hold for every optimal . On the feasible set with , so is invertible and the matrix inverse has no junk value. is the published affineEstimator.
import Mathlib import Definitions.Def_WassersteinDRO_Shrinkage_sdpFeasibleSet import Definitions.Def_WassersteinDRO_Shrinkage_sdpObjective import Definitions.Def_WassersteinDRO_Shrinkage_sdpValue import Definitions.Def_WassersteinDRO_Shrinkage_affineEstimator import Definitions.Def_WassKF_Reform_GaussianBall import Definitions.Def_WassKF_Reform_expectedLoss import Definitions.Def_WassKF_Reform_minimaxValue import Definitions.Def_WassKF_Reform_maximinValue open MeasureTheory ProbabilityTheory
namespace WassKF.Reform
/-- Theorem 2.5 (Tractable reformulation), Shafieezadeh-Abadeh et al., arXiv:1809.08830v3, p. 3.
Let `ℙ = 𝒩_d(μ, Sig)` with `Sig ≻ 0`, `ρ ≥ 0`, and `σ = λ_min(Sig)`. Then
1. the optimal value of the minimax problem (2) over the ambiguity set (3) equals the optimal value
of program (5), `sup Tr[S_xx − S_xy S_yy⁻¹ S_yx]` over `sdpFeasibleSet ρ Sig σ`;
2. for every optimal solution `S⋆` of (5), the affine estimator
`ψ⋆(y) = S⋆_xy (S⋆_yy)⁻¹ (y − μ_y) + μ_x` is measurable and attains the outer infimum of (2);
3. for the same `S⋆`, `Q⋆ = 𝒩_d(μ, S⋆)` lies in the ambiguity set and attains the outer supremum
of nature's problem `sup_{Q ∈ 𝒫} inf_{ψ ∈ ℒ} E^Q[‖x − ψ(y)‖²]` (it is the least favorable prior).
Here `μ_x = (μ_i)_{i ∈ inl}` and `μ_y = (μ_j)_{j ∈ inr}`. -/
theorem theorem_2_5 {n m : ℕ} (μ : EuclideanSpace ℝ (Fin n ⊕ Fin m))
(Sig : Matrix (Fin n ⊕ Fin m) (Fin n ⊕ Fin m) ℝ) (hSig : Sig.PosDef)
(ρ : ℝ) (hρ : 0 ≤ ρ) (σ : ℝ) (hσ : IsLeast (Set.range hSig.isHermitian.eigenvalues) σ) :
((minimaxValue μ Sig ρ : ENNReal) : EReal) = WassersteinDRO.Shrinkage.sdpValue ρ Sig σ ∧
∀ Sstar ∈ WassersteinDRO.Shrinkage.sdpFeasibleSet ρ Sig σ,
IsMaxOn WassersteinDRO.Shrinkage.sdpObjective
(WassersteinDRO.Shrinkage.sdpFeasibleSet ρ Sig σ) Sstar →
(Measurable (WassersteinDRO.Shrinkage.affineEstimator Sstar.toBlocks₁₂ Sstar.toBlocks₂₂
((EuclideanSpace.equiv (Fin n) ℝ).symm (fun i => μ (Sum.inl i)))
((EuclideanSpace.equiv (Fin m) ℝ).symm (fun j => μ (Sum.inr j)))) ∧
(⨆ (Q : Measure (EuclideanSpace ℝ (Fin n ⊕ Fin m))) (_ : Q ∈ GaussianBall μ Sig ρ),
expectedLoss (WassersteinDRO.Shrinkage.affineEstimator Sstar.toBlocks₁₂
Sstar.toBlocks₂₂
((EuclideanSpace.equiv (Fin n) ℝ).symm (fun i => μ (Sum.inl i)))
((EuclideanSpace.equiv (Fin m) ℝ).symm (fun j => μ (Sum.inr j)))) Q) =
minimaxValue μ Sig ρ) ∧
(multivariateGaussian μ Sstar ∈ GaussianBall μ Sig ρ ∧
(⨅ (ψ : EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n)) (_ : Measurable ψ),
expectedLoss ψ (multivariateGaussian μ Sstar)) =
maximinValue μ Sig ρ) := by sorry
end WassKF.Reform
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.