On control of two-scale stochastic systems with linear dynamics in the fast variables (Q2563991)

The paper deals with a two-scale system of stochastic differential equations (SDE) \((x^\varepsilon_t, y^\varepsilon_t)\), \(t\geq 0\), where \(x^\varepsilon_t\) is the ``slow'' variable, \(y^\varepsilon_t\) is the ``fast'' variable and \(\varepsilon\) is a ``small'' positive parameter. The evolution of the system is described as follows: \[ dx^\varepsilon_t= A_1(t, x^\varepsilon_t, u_t)dt+ A_2(t) y^\varepsilon_t dt+ \sigma(t, x^\varepsilon_t) dw^x_t, \] \[ \varepsilon dy^\varepsilon_t= A_3(t) x^\varepsilon_t dt+ A_4(t) y^\varepsilon_t dt+ B(t)u_t dt+ \beta^\varepsilon dw^y_t \] with given initial conditions \(x^\varepsilon_0= x_0\), \(y^\varepsilon_0= 0\). Here \(W= (w^x, w^y)\) is a standard Wiener process, \(\beta^\varepsilon\) is suitably chosen small quantity depending on \(\varepsilon\), the control \(u= (u_t)\) is a \(W\)-predictable process with values in a compact subset \(U\subset \mathbb{R}^d\). These controls are called admissible and their set is denoted by \({\mathcal U}\). Under general conditions this system has a unique solution. The problem is to find the limit behavior, when \(\varepsilon\to 0\), of the minimal value \(J^\varepsilon_*:= \inf J^\varepsilon(u)\), where inf is taken in \(U\in{\mathcal U}\) with a given cost functional \[ J^\varepsilon:= E\int^T_0 [f(t, x^\varepsilon_t, u_t)+ b(t) y^\varepsilon_t]dt+ Eg(x^\varepsilon_T, y^\varepsilon_T). \] This problem is among those which are relatively easy to formulate but not so easy to solve. The authors explain clearly what kind of new difficulties arise. An essentially new moment here is that the cost functional depends also on the fast component in the terminal payoff. Previously used analytical methods (A. Bensoussan) and weak convergence methods (H. Kushner) are not applicable to the above quite general two-scale stochastic system. Thus, the authors follow a different approach of a probabilistic nature successfully developing new ideas and techniques. One of the main results is Theorem 2.1 establishing that the minimal values \(J^\varepsilon_*\) converge to \(\overline J(u)\) as \(\varepsilon\to 0\), where \(\overline J(u)\) is the minimal value of the cost functional of the reduced stochastic control system obtained from the given one. The authors present a detailed proof of this result (some of the intermediate statements are even of independent interest). There is no doubt that the method and the results in this paper are an important contribution in an actively developing area of contemporary stochastics and its applications.