Function-measure kernels, self-integrability and uniquely-defined stochastic integrals

In this work we study the self-integral of a function-measure kernel and its importance on stochastic integration. A continuous-function measure kernel $K$ over $D \subset \mathbb{R}^{d}$ is a function of two variables which acts as a continuous function in the first variable and as a real Radon measure in the second. Some analytical properties of such kernels are studied, particularly in the case of cross-positive-definite type kernels. The self-integral of $K$ over a bounded set $D$ is the "integral of $K$ with respect to itself". It is defined using Riemann sums and denoted $\int_{D}K(x,dx)$. Some examples where such notion is well-defined are presented. This concept turns out to be crucial for unique-definiteness of stochastic integrals, that is, when the integral is independent of the way of approaching it. If $K$ is the cross-covariance kernel between a mean-square continuous stochastic process $Z$ and a random measure with measure covariance structure $M$, $\int_{D}K(x,dx)$ is the expectation of the stochastic integral $\int_{D} ZdM$ when both are uniquely-defined. It is also proven that when $Z$ and $M$ are jointly Gaussian, self-integrability properties on $K$ are necessary and sufficient to guarantee the unique-definiteness of $\int_{D}ZdM$. Results on integrations over subsets, as well as potential $\sigma$-additive structures are obtained. Three applications of these results are proposed, involving tensor products of Gaussian random measures, the study of a uniquely-defined stochastic integral with respect to fractional Brownian motion with Hurst index $H > \frac{1}{2}$, and the non-uniquely-defined stochastic integrals with respect to orthogonal random measures. The studied stochastic integrals are defined without use of martingale-type conditions, providing a potential filtration-free approach to stochastic calculus grounded on covariance structures.