Pointwise evaluation bound on the span of kernel sections (Cauchy-Schwarz)
DisprovedHighDimStat.Rkhs.reproducing_boundfunctional-analysishighdimstatreproducing-kernelrkhssource-faithful-child
Decomposition of HighDimStat.Rkhs.thm12_11_moore_aronszajn (Wainwright, Theorem 12.11), step 2. Pointwise evaluation bound on the span of kernel sections: for f = ∑ᵢ cᵢK(·,xᵢ), |f z| ≤ √(K z z) · √(Gram sum), the Cauchy-Schwarz corollary of the pre-inner product. Hence point evaluation z ↦ f z is continuous for the pre-inner-product norm, so it extends to the completion — the reproducing property (12.3) of Theorem 12.11.
Preamble
import Mathlib
namespace HighDimStat.Rkhs
/-- A kernel `K` on `X` is positive semidefinite: every finite Gram matrix is PSD. -/
def IsPSDKernel {X : Type} (K : X → X → ℝ) : Prop :=
∀ (n : ℕ) (x : Fin n → X) (c : Fin n → ℝ),
0 ≤ Finset.sum Finset.univ (fun i =>
Finset.sum Finset.univ (fun j => c i * c j * K (x i) (x j)))
end HighDimStat.Rkhs
Formal statement
namespace HighDimStat.Rkhs
/-- Step 2 of the Moore-Aronszajn construction: the pointwise evaluation bound on
the span of kernel sections. For `f` a finite linear combination of the
kernel sections `K · (x i)` one has `|f z| ≤ sqrt (K z z) * sqrt (Gram sum)`,
the Cauchy-Schwarz corollary that
makes point evaluation continuous for the pre-inner-product norm. -/
theorem reproducing_bound {X : Type} (K : X → X → ℝ) (hK : IsPSDKernel K) :
∀ (n : ℕ) (x : Fin n → X) (c : Fin n → ℝ) (z : X),
abs (Finset.sum Finset.univ (fun i => c i * K z (x i))) ≤
Real.sqrt (K z z) *
Real.sqrt (Finset.sum Finset.univ (fun i =>
Finset.sum Finset.univ (fun j => c i * c j * K (x i) (x j)))) := by
sorry
end HighDimStat.Rkhs
Source
Wainwright, High-Dimensional Statistics, CUP 2019, p. 389 (PDF p. 409), Theorem 12.11, Eq. (12.3)