Theorem 6.1 — the kernel of a tensor-product basis is the product of one-dimensional kernels
ProvedNatureSLT.Kernels.tensor_product_kernelLet every coordinate of range over the same set , and let , , be a finite or infinite family of real coordinatewise basis functions on , not necessarily orthonormal, whose one-dimensional kernel
converges for all . The -dimensional basis (6.25) consists of the tensor products over all multi-indices . Then the kernel defining the inner product in this basis converges and is the product of one-dimensional kernels:
This is how the book builds every multidimensional kernel of Chapter 6 (Hermite, spline and Fourier) from a one-dimensional one.
Formalization Note The left side is the sum over multi-indices of products of basis functions, not a definition of the multidimensional kernel as a product. "Converges" is unconditional summability in ; for a finite family it is automatic. For both sides equal .
import Mathlib
namespace NatureSLT.Kernels
/-- Theorem 6.1 (Vapnik, *The Nature of Statistical Learning Theory*, 2nd ed., §6.3.2, p. 193),
with the basis (6.25). Every coordinate `x^k` of `x = (x^1, …, x^n)` ranges over the same set `X`
(the book's interval `I`), and `b_i`, `i ∈ ι`, is a finite or infinite set of coordinatewise basis
functions on `X`, not necessarily orthonormal. The one-dimensional kernel is
`K(s, t) = Σ_i b_i(s) b_i(t)`, assumed convergent for all `s, t`. The `n`-dimensional basis
consists of the tensor products `b_{i₁ … iₙ}(x) = b_{i₁}(x^1) ⋯ b_{iₙ}(x^n)` over all multi-indices
`(i₁, …, iₙ) ∈ ιⁿ`, and the kernel defining the inner product in this basis,
`Σ_{i₁, …, iₙ} b_{i₁ … iₙ}(x) b_{i₁ … iₙ}(y)`, converges and equals the product of
one-dimensional kernels `Π_{k=1}^n K(x^k, y^k)`. -/
theorem tensor_product_kernel {X ι : Type*} (b : ι → X → ℝ)
(hb : ∀ s t : X, Summable (fun i : ι => b i s * b i t)) (n : ℕ) (x y : Fin n → X) :
Summable (fun I : Fin n → ι => (∏ k, b (I k) (x k)) * (∏ k, b (I k) (y k))) ∧
∑' I : Fin n → ι, (∏ k, b (I k) (x k)) * (∏ k, b (I k) (y k)) =
∏ k, ∑' i : ι, b i (x k) * b i (y k) := by sorry
end NatureSLT.Kernels
Read-back
What the Lean code literally says, in plain math · claude-opus-5-5
Data.
- and are arbitrary types. Either may be empty, finite, countable or uncountable.
- is a family of real-valued functions on .
- is a natural number.
- and are -tuples of elements of .
Hypothesis. For all , the family is summable. For real families this means unconditional summability, equivalently absolute summability. It implies that only countably many terms are nonzero.
Conclusion. The conclusion is about multi-indices , that is, functions from to . Both of the following hold.
- The family indexed by these multi-indices
is summable over .
- Its sum equals a product of one-dimensional sums:
The basis functions are not assumed orthonormal or otherwise constrained.
Degenerate cases.
- . There is exactly one multi-index (the empty tuple). Its term is the empty product squared, which is , and the right side is the empty product . This holds even when is empty.
- empty and . There are no multi-indices, so the left sum is . Each one-dimensional sum on the right is an empty sum, also , so the right side is .
- empty and . No tuples exist, so the statement is vacuous.
- Default values of sums. The sum notation would return for a non-summable family. Here the right-hand sums are summable by hypothesis, and the left-hand family is asserted summable, so no default value is used.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.