Reproducing kernels and choices of associated feature spaces, in the form of \(L^2\)-spaces (Q2235966)

The Aronszajn method of reproducing kernels and their reproducing kernels Hilbert spaces has a lot of applications, not only in classical analysis but in various other domains of pure and applied mathematics. The key idea of Aronszajn is that a reproducing kernel Hilbert space (RKHS) is a Hilbert space \(\mathcal{H}(K)\) of functions \(f\) on a set \(S\), such that the values \(f(x)\) are ``reproduced'' from \(f\) and a vector \(K_s\) in \(\mathcal{H}(K)\), in such a way that the inner product \(K(s,t):= \left\langle K_s,K_t\right\rangle\) is a positive definite kernel. Every positive definite kernel \(K\) on a set \(S\) has an associated RKHS \(\mathcal{H}(K)\). This paper presents a new analysis of positive definite kernels, their reproducing kernels Hilbert spaces, and an associated family of feature spaces that may be chosen in the form \(L^2(\mu)\). The problem of choosing the best measure \(\mu\) for \(K\) and its associated RKHS is studied and it is seen that depends on the particular application at hand. An important role in obtaining the results is played by using the Ito calculus. The paper is organized in seven sections giving an unitary study on reproducing kernels and choices of associated feature spaces in the form of \(L^2\)-spaces. After an introduction into the topic and presenting the summary of the results in the first section, a sketch of the reproduction kernel theory is presented in the second section. Definitions of the feature space, Gaussian reproducing kernel Hilbert space, canonical centered Gaussian process are introduced. In the third section, using Ito calculus, explicit characterizations and computable formulas for the RKHS corresponding to the general class of Gaussian fields are obtained. The factorization problem is discussed for the purely atomic and non-atomic measures which are connected with the Parseval frames and diffusion problems, respectively. The fourth section is devoted to the study of Gaussian interpolation of Markov processes, and the fifth one to the Parseval forms in the measure category. In the sixth section, a unique isometry \(J:\mathcal{H}(K)\rightarrow L^2(\mu)\) is constructed and the property of the projection \(Q=JJ^*\) in \(L^2(\mu)\) is studied. In the last section of the paper, Hilbert spaces of signed measures \(\mathcal{L}(K)\) are constructed and an isometry of \(\mathcal{L}(K)\) onto \(\mathcal{H}(K)\) is obtained.