When and how an error yields a Dirichlet form (Q860788): Difference between revisions

As a general theoretical framework, consider a probability space \((\Omega,{\mathcal A},P)\) and random variables \(Y\) and \(Y_n\) with value in a measurable space \((E,{\mathcal F})\). For an algebra \({\mathcal D}\) of bounded functions from \(E\) to \(R\) which is dense in \(L^2(E,{\mathcal F},P_Y)\) and containing the constants, the following bias operators are introduced, where \(P_Y\) is the distribution of \(Y\) under \(P\): For any \(\varphi \in {\mathcal D}\), \(\overline{A}[\varphi]\), \(\underline{A}[\varphi]\) and \(\tilde{A}[\varphi]\) are defined by \[ \lim_{n\to \infty} \alpha_n E\left[(\varphi(Y_n)-\varphi(Y))\chi(Y)\right]=E_Y\left[ \overline{A}[\varphi]\chi\right],\tag{H1} \] \[ \lim_{n\to \infty} \alpha_n E\left[(\varphi(Y) -\varphi(Y_n))\chi(Y_n)\right]=E_Y\left[\underline{A}[\varphi]\chi\right],\tag{H2} \] \[ \lim_{n\to \infty} \alpha_n E\left[(\varphi(Y_n)-\varphi(Y))(\chi(Y_n)-\chi(Y))\right] =E_Y\left[\tilde{A}[\varphi]\chi\right]\tag{H3} \] for any \(\chi \in {\mathcal D}\), respectively. If (H3) exists, then the closure \(({\mathcal E},D)\) of \(\tilde{\mathcal E}[\varphi,\chi], \varphi,\chi \in {\mathcal D}\), gives a Dirichlet form with a square field operator \(\Gamma[\varphi]=\tilde{A}[\varphi^2]-2\varphi \tilde{A}[\varphi]\). Moreover, a characterization of the locality of \(({\mathcal E},D)\) is given. In that case, it is shown that the asymptotic error calculus of the form \[ \begin{multlined} \lim_{n\to \infty}\alpha_n E\left[(F(f_1(Y_n),\dots,f_p(Y_n))-F(f_1(Y),\dots,f_p(Y)))^2\right] \\=E_Y\left[\sum_{i,j=1}^p F'_i(f_1,\dots.f_p)F'_j(f_1,\dots,f_p)\Gamma[f_i,f_j]\right]\end{multlined} \] is valid. These general results are applied to various examples related to Brownian bridges, Wiener spaces and stochastic differential equations, and the explicit forms of the bias operators \(\overline{A}[\varphi]\), \(\underline{A}[\varphi]\) and \(\tilde{A}[\varphi]\) are calculated.