On the mathematical foundations of stable RKHSs (Q2188280)

The theory of reproducing kernel Hilbert spaces is essential tool for functions estimation from sparse and noisy data. It is widely employed in machine learning, computer vision and in dynamical systems identification. The contribution of this manuscript is threefold: first, the novel stability tests on some classes of kernel are proposed; second, it is shown that any stable kernel admits a spectral (Mercer) feature map \(\phi: \, {\mathcal X}\rightarrow l_2 \) with \(\phi(x) = \{ \sqrt{\lambda_i} \rho_i(x) \}_{i=1}^{\infty},\) where \(\lambda_i\) and \(\rho_i\) are, respectively, the kernel eigenvalues and eigen functions forming an orthonormal basis in \({\mathcal l}_2\), and finally, the necessary and sufficient condition for RKHS stability in the Mercer feature space are derived.