Radial kernels and their reproducing kernel Hilbert spaces (Q609742): Difference between revisions

Let \(R\) be a continuous convex function on a Hilbert space \(H\). In learning theory, \[ A(\lambda):= \inf_{h\in H} \{\lambda\| h\|^2+ R(h)\}- \inf_{h\in H} R(h) \] is called an approximation error function. Here, \(H\) is a reproducing kernel Hilbert space (RKHS) of functions on \(\mathbb{R}^d\), i.e., such that the evaluations \(\delta_x: h\mapsto h(x)\), \(x\in\mathbb{R}^d\), \(h\in H\), are continuous (belong to \(H^*\)). Every RKHS gives rise to a kernel \(k:\mathbb{R}^d\times \mathbb{R}^d\to\mathbb{R}\), a function satisfying \(k(\cdot,x')\in H\) and \(h(x')=\langle h,k(\cdot,x')\rangle\) for all \(x'\in\mathbb{R}^d\) and \(h\in H\); this \(H\) is denoted by \(H_k\). The authors consider the case of radial kernels (i.e., \(k(x,x')= k(\| x-x'\|)\) for some \(k: \mathbb{R}_+\to \mathbb{R}\)). Due to Schoenberg's theorem, such a kernel can be represented as \[ k(x,x')= \int_{\mathbb{R}^d} k_t(x,x')\,d\mu(t), \] where \(k_t(x,x'):= e^{-t^2\| x- x'\|^2}\) and \(\mu\) is a Borel measure. The main result of the present paper evaluates \(A_{H_k}(\lambda)\) for such kernel as follows: \[ A_{H_k}(\lambda)\leq \int_{\mathbb{R}_+} A_{H_{k_t}}(\lambda)\,d\mu(t), \] where \(k_t\) and \(\mu\) are taken from the above representation.