Nonlinear filtering with fractional Brownian motion (Q5906317)

A special case of filtering is studied where the signal noise is a fractional Brownian motion denoted as \(B^h\) (Hurst coefficient \(h\in] {1\over 2}, 1[)\): the signal process is \(\theta_t= \int^t_0a(s,\xi) \theta_sds+ B^h_t\), and the observation process is \(\xi_t=\int^t_0 A(s,\xi)\theta_s ds+W_t\), where \(W\) is a standard Brownian motion independent of \(B^h\). If \(a\) and \(A\) are uniformly bounded, jointly measurable, Lipschitzian in the following sense: \(\forall t\in[0,T]\), \(\forall x\), \(y\in{\mathcal C}([0,1])\), \(\exists L_T\in R\) such that \[ \bigl| a(t,x)-a(t,y)\bigr| +\bigl| A(t,x)-A(t,y)\bigr|\leq L_T\left(| x_t-y_t|+ \int^t_0 | x_s-y_s | ds\right), \] then the filtering system admits a solution (the reference, theorem 3, does not appear in the paper). Then, the main result is that there exists a unique solution to the Stochastic Differential Equation of the filter: \(\widehat\theta_t= E[\theta_t/ {\mathcal F}^\xi_t]\), \(\gamma(t,\xi)= E[|\theta_t- \widehat\theta_t |^2/ {\mathcal F}_t^\xi]\), denoted as \(\gamma(t,\xi,t)\): \[ \widehat \theta_t=\int^t_0 a(r,\xi)\widehat \theta_rdr+\int^t_0 A(r,\xi)\gamma (t,\xi,r)d\overline W_r, \] \[ \begin{multlined} \gamma(s,\xi,t)= \int^t_0a(r,\xi) \bigl[\gamma(s,\xi,r) +\gamma(t, \xi,r)\bigr]dr -\int^t_0A^2(r,\xi)\gamma (s,\xi,r)\gamma (t,\xi,r)dr+\\ + \textstyle {1\over 2} \bigl[t^{2h}+ s^{2h}-| s-t|^{2h} \bigr]+C_h(t-s)^{2h}_+, \end{multlined} \] where \(\overline W\) is the innovation process \(\xi_t-\int^t_0A(r,\xi) \widehat\theta_rdr\). The main tool is the fact that the conditional law of \((\theta_{s_1}, \dots, \theta_{s_n}, s_i\leq t)\) given \({\mathcal F}_t^\xi\) is Gaussian. The boundary cases \((h=1/2,h=1)\) are studied and an extension to the multidimensional case is performed.