§2, display (6), p. 364 — the stochastic kriging predictor is the unique MSE-optimal predictor of the form (5)
ProvedStochKriging.OptimalMSE.display_6Work in the second-order stochastic kriging model: outputs 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 let , and be the covariance matrices and vector of §2. Assume is positive definite, and put
Then, for every prediction point :
- the stochastic kriging predictor is the linear predictor of the form (5);
- it has the smallest mean squared error for among all predictors of the form (5):
- it is the only one: any weights achieving an MSE no larger than that of equal .
This is the paper's display (6), which defines stochastic kriging: the classical kriging predictor with the covariance of the data enlarged by the intrinsic covariance .
Formalization Note The weights range over all of , with no unbiasedness constraint. No Gaussianity, no independence across replications and no "no CRN" assumption is made; is a general covariance matrix. The uncorrelatedness of field and noise and the finite second moments are added (§2 uses them implicitly). Positive definiteness of makes the inverse in (6) genuine; the sum is always positive semidefinite, being the covariance matrix of , so the hypothesis says exactly that it is invertible.
import Mathlib import Definitions.Def_StochKriging_OptimalMSE_Model open MeasureTheory ProbabilityTheory Matrix
namespace StochKriging.OptimalMSE
/-- Display (6), p. 364: the stochastic kriging predictor `Ŷ(x₀)` is a predictor of the form (5),
with weights `λ = [ΣM + Σε]⁻¹ ΣM(x₀, ·)` and `λ₀ = β₀ − β₀ λᵀ1_k`; it has the smallest MSE for
`Y(x₀)` among all predictors of the form (5); and these weights are the only minimizers. -/
theorem display_6 {Ω X : Type*} [MeasurableSpace Ω] {P : Measure Ω} [IsProbabilityMeasure P]
{k : ℕ} (β₀ : ℝ) (M : X → Ω → ℝ) (ε : ℕ → X → Ω → ℝ) (x : Fin k → X) (n : Fin k → ℕ)
(hmodel : IsModel P M ε) (hn : ∀ i, 1 ≤ n i)
(hSig : (SigmaM P M x + SigmaEps P ε x n).PosDef) (x₀ : X) :
(∀ ω, skPredictor P β₀ M ε x n x₀ ω =
linPred
(β₀ - β₀ * (((SigmaM P M x + SigmaEps P ε x n)⁻¹ *ᵥ SigmaMCross P M x x₀) ⬝ᵥ
fun _ => 1))
((SigmaM P M x + SigmaEps P ε x n)⁻¹ *ᵥ SigmaMCross P M x x₀)
(sampleMean β₀ M ε x n) ω) ∧
(∀ (l₀ : ℝ) (l : Fin k → ℝ),
mse P (skPredictor P β₀ M ε x n x₀) (target β₀ M x₀) ≤
mse P (linPred l₀ l (sampleMean β₀ M ε x n)) (target β₀ M x₀)) ∧
(∀ (l₀ : ℝ) (l : Fin k → ℝ),
mse P (linPred l₀ l (sampleMean β₀ M ε x n)) (target β₀ M x₀) ≤
mse P (skPredictor P β₀ M ε x n x₀) (target β₀ M x₀) →
l = (SigmaM P M x + SigmaEps P ε x n)⁻¹ *ᵥ SigmaMCross P M x x₀ ∧
l₀ = β₀ - β₀ * (((SigmaM P M x + SigmaEps P ε x n)⁻¹ *ᵥ SigmaMCross P M x x₀) ⬝ᵥ
fun _ => 1)) := by sorry
end StochKriging.OptimalMSE
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.