Lemma C.5 — One-step suffix contraction
OpenGinibrePermanent977.Anticoncentration.Supplement.lemma_C_5Use the independent-row setting of Appendix C. Fix and . Suppose the quaternionic row obeys
where is a standard quaternionic Gaussian row. Let the whole vector have an arbitrary Borel probability law, independent of . For one fresh standard Gaussian vector , independent of all original data, define simultaneously
The random vector is independent of the future rows. Conditional on those rows and , almost surely and simultaneously for every ,
If has the law of , equality holds for every such . No moments of are assumed, and the coordinates of need not be independent. This is the common quaternionic replacement input for Theorems 4.1 and C.1.
Formalization Note Independent full rows carry the future suffix; rows up to the current row are unused dummy coordinates. The actual current row is integrated separately. Independence from the entire dummy-extended expresses independence from the future rows. Source is Lean suffixEval (k-1).
Differences from the source:
- encoding: 1-based row and column indices → zero-based Fin indices, with source row k+1 represented by k; reason: finite reindexing preserves all rows and increasing multiplication order
- encoding: independent random inputs → coordinate functions under their probability product laws; reason: these are exactly the joint laws of independent Borel random variables; no moment or density restriction is added
- encoding: conditional expectation given the future rows and a → almost-everywhere inequality for the explicit current-row and fresh-Gaussian integrals, simultaneously for every s ≥ 0; reason: product independence identifies these integrals with conditional expectation versions; unused earlier rows do not enter either suffix
- encoding: fresh coefficients exist, with equality for Gaussian R_k → the explicit contractCoeff from the proof; equality when the current-row law is the pushforward of gaussVec by σ • ·; reason: this is exactly the one-vector construction on p.21, including its right multiplication order
import Mathlib import Definitions.Def_GinibrePermanent977_Anticoncentration_Supplement_Model open MeasureTheory ProbabilityTheory open scoped Quaternion BigOperators
namespace GinibrePermanent977.Anticoncentration.Supplement
/-- Lemma C.5, p.21. The product sample space carries random a, independent rows B and
one fresh Gaussian vector h. The row at k is integrated separately under μ. The conditional
comparison holds almost surely simultaneously for every s≥0. The explicit coefficient map
is independent of all the remaining rows and gives equality when μ is the scaled Gaussian law. -/
theorem lemma_C_5 (n k : ℕ) (hk : k < n) (σ : ℝ) (hσ : 0 < σ)
(μ : Measure (Fin n → Scalar.quaternion.K)) [IsProbabilityMeasure μ]
(hproj : ∀ (c : Fin n → Scalar.quaternion.K) (s : ℝ), 0 ≤ s →
(∫ r, Real.exp (-s * ‖rowAction r c‖ ^ 2) ∂μ) ≤
∫ g, Real.exp (-s * σ ^ 2 * ‖rowAction g c‖ ^ 2) ∂gaussVec .quaternion n)
(α : Measure (Coeff n k)) [IsProbabilityMeasure α]
(ρ : Fin n → Measure (Fin n → Scalar.quaternion.K))
[∀ i, IsProbabilityMeasure (ρ i)] :
IndepFun
(fun x : Coeff n k × ((Fin n → Fin n → Scalar.quaternion.K) ×
(Fin n → Scalar.quaternion.K)) => contractCoeff k hk σ x.1 x.2.2)
(fun x => x.2.1)
(α.prod ((Measure.pi ρ).prod (gaussVec .quaternion n))) ∧
∀ᵐ t ∂(α.prod (Measure.pi ρ)), ∀ s : ℝ, 0 ≤ s →
((∫ r, Real.exp (-s * ‖suffixEval k (by omega) (replaceRow k t.2 r) t.1‖ ^ 2) ∂μ) ≤
∫ h, Real.exp (-s * ‖suffixEval (k+1) (by omega) t.2
(contractCoeff k hk σ t.1 h)‖ ^ 2) ∂gaussVec .quaternion n) ∧
(μ = (gaussVec .quaternion n).map (fun g => σ • g) →
(∫ r, Real.exp (-s * ‖suffixEval k (by omega) (replaceRow k t.2 r) t.1‖ ^ 2) ∂μ) =
∫ h, Real.exp (-s * ‖suffixEval (k+1) (by omega) t.2
(contractCoeff k hk σ t.1 h)‖ ^ 2) ∂gaussVec .quaternion n) := by sorry
end GinibrePermanent977.Anticoncentration.Supplement