Rylan Schaeffer

Logo
Resume
Research
Learning
Blog
Teaching
Jokes
Kernel Papers


Reproducing Kernels

Aronszajn’s theorem tells us that there is an equivalence between kernel functions and feature maps + inner products. However, it says nothing about the uniqueness or importance of any particular Hilbert space or kernel. However, there is one kernel that is more important than all others: the reproducing kernel (RK).

Definitions

When thinking about the reproducing kernel (RK), we think of a Hilbert space of functions from \(X\) to \(\mathbb{R}\); that is, each \(x \in X\) will be represented by \(\phi(x) := K_x \in H\).

Definition 1: Let \(X\) be a set and \(H \subset \mathbb{R}^X\) be a Hilbert space of functions with inner product \(\langle \rangle_H\). The function \(K: X \times X \rightarrow \mathbb{R}\) is called the reproducing kernel of \(H\) if

  1. \(H\) contains all functions of the form: \(\forall x \in X, K_x: x' \rightarrow K(x, x')\)
  2. \(\forall x \in X and f \in H\), the reproducing property holds: $$f(x) = \langle f, K_x \rangle_H

If a RK exists, we say that H is its corresponding RKHS.

Definition 2: The HIlbert space \(H \subset \mathbb{R}^X\) is a RKHS if and only if \(\forall x \in X\), the linear mapping \(F: H \rightarrow \mathbb{R}\) is continuous. That is, if we fix some \(k\) and we move in the Hilbert space, we’ll be able to find a \(f\) such that \(f \rightarrow f(x)\).

Definition 3: let \(X\) be a non-empty set and \(H\) a Hilbert space of functions \(f: X \rightarrow \mathbb{R}\). H is called a RKHS if there exists a function \(k: X \times X \rightarrow \mathbb{R}\) (i.e. a kernel) with the following properties:

  1. \(k\) is reproducing i.e. \(\forall f(\cdot) \in H, \langle f(\cdot), k(\cdot, x) \rangle = f(x)\). Importantly, \(\langle k(x_i, \cdot), k(x_j, \cdot) \rangle = k(x_i, x_j)\).

  2. \(k\) spans \(H\) i.e. \(H = \overbar{\text{span} \{k(x, \cdot) | x \in X \}}\), where the overline denotes completion of the space i.e. adding all limits of all Cauchy sequences.

Corollary: if \((f_n)_n\) converges to \(f \in H\), then \((f_n(x))_n\) converges to \(f(x)\). In english, convergence in the RKHS implies pointwise converges of the function.

Properties

Proof: Suppose for purposes of contradiction, the RKHS \(H\) has two RKs \(K, K'\). Then \(\forall x \in X\)

\(\begin{align*} ||K_x - K_x'||_H^2 &= \langle K_x - K_x', K_x - K_x' \rangle\\ &= K_x(x) - K_x(x) - K_x^'(x) + K_x^'(x)\) &= 0 \end{align*} $$