A class of stochastic optimal control problems in Hilbert spaces: BSDEs and optimal control laws, state constraints, conditioned processes. (Q2574611): Difference between revisions

Let \(H\) and \(U\) be real separable Hilbert spaces, \(W\) a standard cylindrical Wiener process on \(U\), \(A\) an infinitesimal generator of a \(C_ 0\)-semigroup on \(H\), \(F\) and \(G\) measurable functions mapping \([0,T]\times H\) into \(H\) and the space of bounded linear operators from \(U\) to \(H\), respectively. Suppose that the operators \(\text e^ {sA}G(t,x)\) are Hilbert-Schmidt whenever \(s>0\), \(t\in [0,T]\), \(x\in H\), and the functions \(F(s,\cdot )\), \(\text e^ {sA}G(t,\cdot )\) are Lipschitz continuous and of a sublinear growth, the former with constants independent of \(s\), the latter with constants behaving as \(Ls^ {-\gamma }\) for some \(s\in \mathopen [0,\frac 12\mathclose [\). Let \(r\) and \(q\) be measurable \(\mathopen [0,\infty \mathclose ]\)-valued functions on \(H\) and \([0,T] \times H\), respectively. A controlled diffusion process in \(H\) defined by the equation \[ dX^ {u}_ {t} = \bigl \{ AX^ {u}_ {t} + F(t,X^ {u}_ {t}) + G(t,X^ {u}_ {t})u_ {t}\bigr \}\,dt + G(t,X^ {u}_ {t}) \,dW_ {t}, \quad X^ {u}_ {0} = x \tag{1} \] is considered, the objective being to minimize the cost functional \[ J(u) = \mathbb E\int ^ {T}_ 0 \Bigl \{\frac 12\| u_ {t} \| ^ 2_ {U} + q(t,X^ {u}_ {t})\Bigr \}\,dt + \mathbb E r(X^ {u}_ {T}) \] over the set of all admissible controls \(u\), that is, over the set of all adapted square integrable \(U\)-valued processes. First, it is proved that there exists a unique mild solution to (1) for any fixed admissible control \(u\), notwithstanding that \(u\) need not be bounded. Secondly, by using backward stochastic differential equations, a measurable function \(\zeta : [0,T]\times H\to U\) is found such that \(\overline u_ {t} = \zeta (t,\overline X_ {t})\) is an optimal control, provided that \(\overline X\) solves the closed-loop equation, which is shown to admit a (weak) mild solution unique in law. Finally, the case of functions \(r\) and \(q\) taking only the values 0 and \(\infty \) (that corresponds to optimal control problems with state constraints) is paid particular attention.