§2, display (7), p. 364 — the optimal MSE equals the kriging MSE plus a positive definite form in Σ_M(x₀,·)
ProvedStochKriging.OptimalMSE.display_7Fix the number of design points. There is a rule that assigns to every pair of positive definite matrices a positive definite matrix , such that the following holds in every second-order stochastic kriging model.
Let with square-integrable, mean-zero field and noise , field and noise uncorrelated, and noise correlation across design points allowed (CRN). Let be a design with every , let be the vector of sample means, and assume the covariance matrices and of §2 are both positive definite. Then for every prediction point the optimal MSE over the linear predictors (5),
satisfies
and, whenever , is strictly larger than the bracketed term.
The bracketed term is the usual kriging MSE, the error of predicting from noise-free observations of the field. The result says that intrinsic simulation noise inflates the optimal MSE by a positive definite quadratic form in , whose matrix depends only on and , not on or on the model beyond these matrices.
Formalization Note " is a positive definite matrix that depends on and " is rendered as a function of the two matrices, chosen before the model and the point ; it is not fixed by the statement. The optimal MSE is the infimum over all . The hypotheses and are not written on the page: the first makes genuine (Lean's inverse of a singular matrix is ), the second is necessary for to be positive definite (with the two lines coincide). The uncorrelatedness of field and noise and the finite second moments are added as in the model. The settings form an arbitrary type, design points are indexed by Fin k, and the statement is uniform over all probability spaces and setting types (in fixed universes).
import Mathlib import Definitions.Def_StochKriging_OptimalMSE_Model open MeasureTheory ProbabilityTheory Matrix
namespace StochKriging.OptimalMSE
universe u v
/-- Display (7), p. 364. For `k` design points there is a rule `Ξ` assigning to every pair of
positive definite `k × k` matrices `(ΣM, Σε)` a positive definite matrix `Ξ(ΣM, Σε)` such that, in
every model (3) whose covariance matrices `ΣM` and `Σε` are positive definite, and at every
prediction point `x₀`, the optimal MSE over the linear predictors (5) is
`ΣM(x₀,x₀) − ΣM(x₀,·)ᵀ[ΣM + Σε]⁻¹ΣM(x₀,·)`, equals the kriging MSE
`ΣM(x₀,x₀) − ΣM(x₀,·)ᵀΣM⁻¹ΣM(x₀,·)` plus `ΣM(x₀,·)ᵀ Ξ(ΣM, Σε) ΣM(x₀,·)`, and exceeds the kriging
MSE whenever `ΣM(x₀,·) ≠ 0`. -/
theorem display_7 (k : ℕ) :
∃ Ξ : Matrix (Fin k) (Fin k) ℝ → Matrix (Fin k) (Fin k) ℝ → Matrix (Fin k) (Fin k) ℝ,
(∀ A B : Matrix (Fin k) (Fin k) ℝ, A.PosDef → B.PosDef → (Ξ A B).PosDef) ∧
∀ (Ω : Type u) (X : Type v) [MeasurableSpace Ω] (P : Measure Ω) [IsProbabilityMeasure P]
(β₀ : ℝ) (M : X → Ω → ℝ) (ε : ℕ → X → Ω → ℝ) (x : Fin k → X) (n : Fin k → ℕ),
IsModel P M ε → (∀ i, 1 ≤ n i) →
(SigmaM P M x).PosDef → (SigmaEps P ε x n).PosDef →
∀ x₀ : X,
optimalMSE P β₀ M ε x n x₀ =
SigmaMFun P M x₀ x₀ - SigmaMCross P M x x₀ ⬝ᵥ
((SigmaM P M x + SigmaEps P ε x n)⁻¹ *ᵥ SigmaMCross P M x x₀) ∧
optimalMSE P β₀ M ε x n x₀ =
(SigmaMFun P M x₀ x₀ - SigmaMCross P M x x₀ ⬝ᵥ
((SigmaM P M x)⁻¹ *ᵥ SigmaMCross P M x x₀)) +
SigmaMCross P M x x₀ ⬝ᵥ
(Ξ (SigmaM P M x) (SigmaEps P ε x n) *ᵥ SigmaMCross P M x x₀) ∧
(SigmaMCross P M x x₀ ≠ 0 →
SigmaMFun P M x₀ x₀ - SigmaMCross P M x x₀ ⬝ᵥ
((SigmaM P M x)⁻¹ *ᵥ SigmaMCross P M x x₀) <
optimalMSE P β₀ M ε x n x₀) := by sorry
end StochKriging.OptimalMSE
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.