Existence and uniqueness of the RKHS associated with a PSD kernel (Theorem 12.11)
DisprovedHighDimStat.Rkhs.thm12_11_moore_aronszajnTheorem 12.11 (the Moore-Aronszajn theorem). The chapter's central existence-and- uniqueness result: every positive semidefinite kernel function determines one Hilbert space of functions, with no further data needed.
Given any positive semidefinite kernel function , there is a Hilbert space of functions on (embedded via an injective linear map into ) in which satisfies the reproducing property (12.3) via a feature map — and this Hilbert space is unique: any two Hilbert spaces with the reproducing property for the same are linearly isometric, via an isometry that intertwines their embeddings into functions on (so they are, in the sense that matters, literally the same space of functions). It is known as the reproducing kernel Hilbert space (RKHS) associated with .
Formalization Note Existence and uniqueness are both formalized, per this chapter's
named pitfall that the uniqueness clause is real content. Existence is ∃ H together with
its Hilbert-space instances and the embedding/feature data. Uniqueness is stated as: for
any two witnesses and of the reproducing property for the
same , there is a linear isometric equivalence with
— the book's own proof concludes literally
(as sets of functions with the same inner product), which this captures as
"isometric via an equivalence commuting with the embeddings into functions," the natural
formal rendering of literal set equality once is presented abstractly rather than as
a concrete subtype of . X carries no topology or measure-theoretic
structure, matching the chapter's own remark that Theorem 12.11 holds for any PSD
kernel on any set — not conflated with Mercer's theorem's additional compactness and
continuity hypotheses.
import Mathlib import Definitions.Def_HighDimStat_Rkhs_Core
namespace HighDimStat.Rkhs
open scoped RealInnerProductSpace
/-- Theorem 12.11 (the Moore-Aronszajn theorem, p. 389): given any positive semidefinite
kernel function `K` on `X`, there is a Hilbert space `H` (embedded into functions on `X` via
`toFun`) in which the kernel satisfies the reproducing property (12.3) via a feature map
`feature`, and this Hilbert space is unique: any two such Hilbert spaces for the same `K` are
linearly isometric via an equivalence that intertwines their embeddings into functions on
`X`. -/
theorem thm12_11_moore_aronszajn {X : Type*} (K : X → X → ℝ) (hK : IsPSDKernel K) :
(∃ (H : Type) (_ : NormedAddCommGroup H) (_ : InnerProductSpace ℝ H) (_ : CompleteSpace H)
(toFun : H →ₗ[ℝ] (X → ℝ)) (feature : X → H), IsRKHS K toFun feature) ∧
(∀ (H₁ : Type) (_ : NormedAddCommGroup H₁) (_ : InnerProductSpace ℝ H₁)
(_ : CompleteSpace H₁) (toFun₁ : H₁ →ₗ[ℝ] (X → ℝ)) (feature₁ : X → H₁),
IsRKHS K toFun₁ feature₁ →
∀ (H₂ : Type) (_ : NormedAddCommGroup H₂) (_ : InnerProductSpace ℝ H₂)
(_ : CompleteSpace H₂) (toFun₂ : H₂ →ₗ[ℝ] (X → ℝ)) (feature₂ : X → H₂),
IsRKHS K toFun₂ feature₂ →
∃ e : H₁ ≃ₗᵢ[ℝ] H₂, ∀ f : H₁, toFun₂ (e f) = toFun₁ f) := by sorry
end HighDimStat.Rkhs
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.