Theorem 3.1, p. 12 — without sparsity information the unique minimizer of (5) is X⋆ = Σ x⋆_i v_i v_iᵀ, x⋆_i from (16a) at the positive root γ⋆ of (16b)
ProvedWassInvCov.Shrinkage.theorem_3_1Let and . Consider the distributionally robust estimation problem (5) without structural information, :
where is the set of centred normal distributions on within type-2 Wasserstein distance of . Let (possibly singular) have the spectral decomposition with orthonormal eigenvectors . Then:
- the equation
has exactly one positive solution ; 2. with
the matrix is positive definite, attains the minimum of (5) over all positive definite , and every positive definite minimizer of (5) equals .
The estimator shares the eigenvectors of and shrinks its eigenvalues by an explicit nonlinear map; it is invertible even when is singular (more dimensions than samples). This is the paper's analytical solution of the robust estimation problem.
Formalization Note The hypothesis is added: for , (16b) has no positive solution. The eigendecomposition is arbitrary (any orthonormal family and reals with ), so the statement also asserts independence of the choice of eigenbasis. Part 2 is stated for every positive root, which by part 1 is . The objective is the measure-level one of (5), with the worst-case expectation computed as an extended-real supremum over ; minimality is IsMinOn over the positive definite matrices.
import Mathlib import Definitions.Def_WassInvCov_Shrinkage_Setting open MeasureTheory ProbabilityTheory Matrix
namespace WassInvCov.Shrinkage
/-- Theorem 3.1, p. 12 (arXiv:1805.07194v1). Without sparsity information (`𝒳 = 𝕊^p_{++}`), for
`ρ > 0` and `Σ̂ ⪰ 0` with an orthonormal eigendecomposition `Σ̂ = Σᵢ λᵢ vᵢ vᵢᵀ`, equation (16b) has a
unique positive solution `γ⋆`, and `X⋆ = Σᵢ x⋆ᵢ vᵢ vᵢᵀ` with `x⋆ᵢ` from (16a) is the unique minimizer
of (5). The hypothesis `0 < p` is added (for `p = 0`, (16b) has no positive solution). -/
theorem theorem_3_1 {p : ℕ} (hp : 0 < p) (ρ : ℝ) (hρ : 0 < ρ) (Sh : Matrix (Fin p) (Fin p) ℝ)
(hSh : Sh.PosSemidef) (lam : Fin p → ℝ) (v : Fin p → EuclideanSpace ℝ (Fin p))
(hv : Orthonormal ℝ v)
(hdecomp : Sh = ∑ i, lam i • Matrix.vecMulVec (v i).ofLp (v i).ofLp) :
(∃! γ : ℝ, 0 < γ ∧ phi16b ρ lam γ = 0) ∧
∀ γs : ℝ, 0 < γs → phi16b ρ lam γs = 0 →
(shrinkageEstimator lam v γs).PosDef ∧
IsMinOn (robustObjective ρ Sh) {X | X.PosDef} (shrinkageEstimator lam v γs) ∧
∀ X : Matrix (Fin p) (Fin p) ℝ, X.PosDef →
IsMinOn (robustObjective ρ Sh) {X | X.PosDef} X → X = shrinkageEstimator lam v γs := by sorry
end WassInvCov.Shrinkage
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.