dlp_conditional_lemma2_sigma_fiber_matrix_chaos_inl
ProvedConditional Lemma 2 (equation (6) of Section 4) of de la Peña–Montgomery-Smith, Decoupling Inequalities for the Tail Probabilities of Multivariate U-Statistics (Ann. Probab. 23 (1995) 806–816, arXiv:math/9309211), instantiated at order for the concrete matrix-valued off-diagonal -sign chaos on the symmetric Rademacher fiber. Let be the tetrahedral bilinear sign chaos with matrix coefficients (inlined as the guarded double sum), and let be a conditioning matrix. Following de la Peña's Proposition 1, fix a dual unit pair () norming , i.e. (spectral norm). Then, provided the scalar dual image is mean-zero with strictly positive variance (the standing Proposition 1 non-degeneracy hypotheses), the -mass of the survival event is bounded below: . The constant comes from Bonami's degree-2 hypercontractivity () feeding the Paley–Zygmund positivity , with the norming-functional containment . This is the genuine conditional-positivity content (eq 6) of the de la Peña pair-decoupling forward bound bernoulli_pair_decoupling_spectral_tail_bound_offdiag. (Suffix _inl: matrix chaos fully inlined so the declaration is a single theorem.)
import Definitions.Def_matrix_completion_tangent import Definitions.Def_matrix_completion_rademacher import Mathlib.Analysis.InnerProductSpace.Basic open MatrixCompletion open scoped BigOperators Classical InnerProductSpace
theorem dlp_conditional_lemma2_sigma_fiber_matrix_chaos_inl
{n1 n2 : Nat}
(a : (Fin n1 × Fin n2) → (Fin n1 × Fin n2) → RealMatrix n1 n2)
(T : RealMatrix n1 n2)
(xv : EuclideanSpace ℝ (Fin n2)) (yv : EuclideanSpace ℝ (Fin n1))
(hxv : ‖xv‖ ≤ 1) (hyv : ‖yv‖ ≤ 1)
(hnorm : ⟪Matrix.toEuclideanLin T xv, yv⟫_ℝ = spectralNorm T)
(hmean :
rademacherExpectation
(fun eps => ⟪Matrix.toEuclideanLin
(∑ w1 : Fin n1 × Fin n2, ∑ w2 : Fin n1 × Fin n2,
(if w1 = w2 then (0 : RealMatrix n1 n2)
else (rademacherSign eps w1.1 w1.2 * rademacherSign eps w2.1 w2.2) • a w1 w2)) xv, yv⟫_ℝ) = 0)
(hvar :
0 < rademacherExpectation
(fun eps => (⟪Matrix.toEuclideanLin
(∑ w1 : Fin n1 × Fin n2, ∑ w2 : Fin n1 × Fin n2,
(if w1 = w2 then (0 : RealMatrix n1 n2)
else (rademacherSign eps w1.1 w1.2 * rademacherSign eps w2.1 w2.2) • a w1 w2)) xv, yv⟫_ℝ) ^ 2)) :
rademacherExpectation
(fun eps =>
if spectralNorm T ≤ spectralNorm (T +
(∑ w1 : Fin n1 × Fin n2, ∑ w2 : Fin n1 × Fin n2,
(if w1 = w2 then (0 : RealMatrix n1 n2)
else (rademacherSign eps w1.1 w1.2 * rademacherSign eps w2.1 w2.2) • a w1 w2)))
then (1 : ℝ) else 0) ≥ 1 / 324 := by sorry