Documentation

LeanPool.OSforGFF.General.GaussianRBF

Gaussian RBF Kernel is Positive Definite #

Proves that the Gaussian radial basis function kernel h ↦ exp(-‖h‖² / 2) is a positive definite function on any real inner product space. The proof uses three steps: (1) the inner product kernel ⟨x, y⟩ is positive definite; (2) exponentiation preserves positive definiteness via the Hadamard series in HadamardExp.lean; (3) the Gaussian kernel factors as exp(-‖x - y‖² / 2) = exp(-‖x‖² / 2) · exp(-‖y‖² / 2) · exp(⟨x, y⟩), reducing from the RBF to the inner product kernel.

def IsPositiveDefiniteKernel {α : Type u_2} (K : αα) :

Definition of a positive definite kernel K : α × α → ℂ.

Equations
Instances For

    Step 2: The inner product kernel is positive definite.

    theorem real_valued_PD_kernel_gives_PSD_matrix {α : Type u_2} (K : αα) (h_real : ∀ (x y : α), (K x y).im = 0) (h_symm : ∀ (x y : α), K x y = K y x) (hK : IsPositiveDefiniteKernel K) (m : ) (x : Fin mα) :
    (Matrix.of fun (i j : Fin m) => (K (x i) (x j)).re).PosSemidef

    Helper: For a real-valued PD kernel, the kernel matrix is PSD. Forward direction of the bridge between complex kernels and real matrices.

    theorem exp_is_pd_kernel {α : Type u_2} (K : αα) (hK : IsPositiveDefiniteKernel K) (h_real : ∀ (x y : α), (K x y).im = 0) (h_symm : ∀ (x y : α), K x y = K y x) :
    IsPositiveDefiniteKernel fun (x y : α) => Complex.exp (K x y)

    Step 3b: Exponential of a symmetric real-valued PD kernel is PD. Uses the Hadamard series machinery from HadamardExp.lean (same as OS3 proof).

    The Gaussian RBF kernel is positive definite on any inner product space.