Eq. (6.26) — the -dimensional Hermite kernel as one exponential (as printed)
DisprovedNatureSLT.Kernels.hermite_kernel_multidimLet , , and let and be vectors of , with inner product and Euclidean norm . The second equality of (6.26) claims
The left side is the product of one-dimensional Hermite kernels (6.24), which by Theorem 6.1 is the kernel of the tensor-product Hermite basis; the right side is the form the book uses to motivate the semilocal kernels (6.27).
Formalization Note Only the second "=" of (6.26) is stated; the first is Theorem 6.1 applied to (6.24). The range is that of (6.24). This statement is believed to be false as printed for : the product of the prefactors is , and the factor is missing on the right (see the corrected variant). The power is a real power of a positive number.
import Mathlib
namespace NatureSLT.Kernels
/-- Eq. (6.26), second equality (Vapnik, *The Nature of Statistical Learning Theory*, 2nd ed.,
§6.3.2, p. 193), as printed: for `x, y ∈ ℝⁿ` and `0 ≤ q < 1`,
`Π_{i=1}^n (π(1 - q²))^{-1/2} exp{2x^i y^i q/(1 + q) - (x^i - y^i)² q²/(1 - q²)}
= (1 - q²)^{-n/2} exp{2(x * y) q/(1 + q) - |x - y|² q²/(1 - q²)}`,
where `x * y` is the inner product and `|x - y|` the Euclidean norm. -/
theorem hermite_kernel_multidim (n : ℕ) (q : ℝ) (hq0 : 0 ≤ q) (hq1 : q < 1)
(x y : EuclideanSpace ℝ (Fin n)) :
∏ i : Fin n, (1 / Real.sqrt (Real.pi * (1 - q ^ 2)) *
Real.exp (2 * x i * y i * q / (1 + q) - (x i - y i) ^ 2 * q ^ 2 / (1 - q ^ 2))) =
1 / (1 - q ^ 2) ^ ((n : ℝ) / 2) *
Real.exp (2 * inner ℝ x y * q / (1 + q) - ‖x - y‖ ^ 2 * q ^ 2 / (1 - q ^ 2)) := by sorry
end NatureSLT.Kernels
Read-back
What the Lean code literally says, in plain math · claude-opus-5-5
Data and hypotheses.
- is a natural number.
- is a real number with .
- and are points of Euclidean space .
- is the standard inner product, and is the Euclidean norm.
Statement.
The power is a real power with the positive base . No factor of appears on the right.
Equivalent form. The exponents on the left add up to the exponent on the right, because and . So the statement is equivalent to the equality of constant factors
Degenerate cases.
- Well-definedness. Because , both and , so every division and square root is well defined.
- . The left side is the empty product . The right side is . The statement holds.
- . The statement reads , that is, .
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.