Stochastic Mappings and Random Distribution Fields. A Correlation Approach

Given an arbitrary set (a topological space) \Lambda and a (continuous) m.s.o. stochastic mapping \Phi: \Lambda\rightarrow {\mathscr H} the following assertions hold:

1. (i) the scalar covariance function \gamma_{\Phi}, the operator covariance function \Gamma_{\Phi} respectively, is a complex, a \mathscr{C}_{1}(H)-valued respectively, positive definite (and continuous) kernel on \Lambda;

2. (ii) the space {\mathscr H}_{(\Phi)} (the module {\mathscr H}_{\Phi} respectively), associated to \Phi is isomorph as Hilbert space, (as normal Hilbert {\mathcal B}(H)-module respectively) to the reproducing kernel Hilbert space K_{\gamma_{\Phi}}, (the reproducing kernel normal Hilbert {\mathcal B}(H)-module {\mathscr H}_{\Gamma_{\Phi}} respectively);

3. (iii) the measurements space G_{\Phi}, associated to \Phi is isomorph to the Hilbert space G_{\Gamma_{\Phi}}, having \Gamma_{\Phi} as operatorial reproducing kernel. In this way the two Hilbert spaces ensure respectively the description of the {\mathcal B}(H)-modules {\mathscr H}_{\Phi}, {\mathscr H}_{\Gamma_{\Phi}} by means of the operatorial models mentioned in (?), respectively (?).
   Moreover, G_\Phi can be described directly in terms of the m.s.o. stochastic mapping \Phi by (?).

If \Phi and \Psi are two m.s.o.s.m. having the operator covariance functions \Gamma_{\Phi} and \Gamma_{\Psi} respectively, and the operator cross covariance function \Gamma_{\Phi, \Psi}, then \Phi is subordinate to \Psi, iff the function K_\lambda(\cdot)=\Gamma_{\Phi, \Psi}(\lambda,\cdot) is in the {\mathcal B}(H)-module {\mathscr H}_{\Gamma_{\Psi}} with \Gamma_{\Psi} as reproducing kernel and

1. (i) If \Lambda is an arbitrary set (a topological space), then for any positive definite \mathscr{C}_{1}(H) - valued kernel \Gamma on \Lambda (continuous on \Lambda\times \Lambda), there exists a (continuous) multivariate second order stochastic mapping \Phi: \Lambda\rightarrow {\mathscr H}, which assures for \Gamma a Kolmogorov factorization:

   \Gamma(\lambda, \mu)=[\Phi(\lambda), \Phi(\mu)]_{{\mathscr H}}\ ;\ \lambda, \mu \in \Lambda,

   i.e. \Gamma coincides with the operator covariance function \Gamma_{\Phi} of \Phi.

2. (ii) If \Phi_{1} and \Phi_{2} are two (continuous) m.s.o. stochastic mappings on \Lambda for which \Gamma_{\Phi_{1}}=\Gamma_{\Phi_{2}}, then \Phi_{1} and \Phi_{2} are gramian unitary equivalent, i.e. there exists a gramian unitary operator \mathrm{W}:{\mathscr H}_{\Phi_{2}}\rightarrow {\mathscr H}_{\Phi_{1}} such that

Let \Gamma be a {\mathscr C}_1(H)-valued sesquilinear kernel on {\mathcal D}_d^m, which is continuous on the diagonal of {\mathcal D}_d^m\times{\mathcal D}_d^m. Then there is an m-order distribution kernel C_\Gamma on {\mathbb R}^d such that

Moreover, when \Gamma is positive on the diagonal, then C_\Gamma is a positive definite m-order distribution kernel.

For a given m.s.o. random distribution field U\in\big({\mathcal D}^{m}_{d}\big)'({\mathscr H}), the modular domain {\mathscr H}_{U}, respectively the measurements space G_{U} will be isomorph to the {\mathcal B}(H)-module {\mathscr H}_{\Gamma^{C_{U}}}, respectively the Hilbert space G_{\Gamma^{C_{U}}} both reproduced by the \mathscr{C}_{1}(H)-valued kernel \Gamma^{C_{U}}, while the vector domain {\mathscr H}_{(U)} of U is isomorphic to the Hilbert space reproduced by the complex valued kernel \gamma^{C_{U}}.

Given two m.s.o.r.d.f. U and V then U is subordinate to V, iff for each fixed \varphi \in {\mathcal D}_d, the function K_\varphi (\cdot) given by

lies in the {\mathcal B}(H)-module {\mathscr H}_{\Gamma^{C_V}} and \bigl[K_\varphi , K_\psi\bigr]_{{\mathscr H}_{\Gamma^{C_V}}} = C_U (\varphi \otimes \bar \psi).

For an m-order distribution kernel on {\mathbb R}^d, C\in\big({\mathcal D}^{m}_{2d}\big)'(\mathscr{C}_{1}(H)) the following statements are equivalent:

• (i) C\in pd\big({\mathcal D}^{m}_{2d}\big)'(\mathscr{C}_{1}(H)); i.e. the operatorial kernel \Gamma^{C} on {\mathcal D}^{m}_{d} defined by \Gamma^{C}(\varphi,\psi):=C(\varphi\otimes\overline{\psi})\ (\varphi,\psi\in {\mathcal D}^{m}_{d}) is positive definite.

• (ii) The operatorial kernel \Gamma^{C} is sesquilinear and positive (in the sense of positivity from {\mathcal B}(H)) on the diagonal of {\mathcal D}^{m}_{d}\times{\mathcal D}^{m}_{d}.

• (iii) There is a m.s.o. random distribution field U\in\big({\mathcal D}^{m}_{d}\big)'(\mathscr{H}) such that C=C_{U} (i.e. C has a Kolmogorov type factorization).

If the m.s.o. random distribution fields U^{1},\,U^{2}\in \big({\mathcal D}^{m}_{d}\big)'({\mathscr H}) have the same operator covariance distribution, i.e. C_{U^{1}}=C_{U^{2}}, then there exists a gramian unitary operator W:{\mathscr H}_{U^{1}}\rightarrow {\mathscr H}_{U^{2}} such that

For a m.s.o.r.d.f. \{U_{\varphi}\}_{\varphi\in {\mathcal D}_d} taking values in {\mathscr H}, there exists a unique decomposition

of U such that

1. (i) \{U_{\varphi}^{det}\}_{\varphi\in {\mathcal D}_d} is operator deterministic and \{U_{\varphi}^{p}\}_{\varphi\in {\mathcal D}_d} is operator purely nondeterministic,

2. (ii) U^{det} and U^{p} are operator subordinate to U,

3. (iii) U^{det} and U^{p} are operator uncorrelated, i.e. the operator cross covariance distribution C_{U^{det}, U^p} vanishes on {\mathcal D}_{2d}.

Moreover the gramian orthogonal decomposition

holds.

For the above mentioned classes of m.s.o. random distribution fields the following inclusions hold with continuous embedding

The valued domains of the restrictions of the operator covariance distribution mapping (?) to the spaces from (5.1) satisfy respectively the following inclusions

Statements analogous as for the operator covariance distribution mapping (?) in the corollaries above, hold for the corresponding mappings

Moreover, these operator covariance bimeasure mappings are natural extensions of the covariance function mapping

associated to classical m.s.o. random fields, as was defined in [Kaki, Section 4.1, pp. 148].

The positive definite operator valued bimeasures from (5.3) and the operator covariance function of (5.3) satisfy

respectively.

The m.s.o.r.d.f.’s F, \ \xi^F,\ U^F as well as \xi,\ U^\xi are simultaneously (operator) deterministic, (operator) nondeterministic or (operator) purely nondeterministic. Also the corresponding summands in the Wold decomposition are the same, i.e.

as well as