Pontryagin's principle for some probabilistic control problems (Q6589680)

Let \(\mathbb{H}\) and \(\mathbb{U}\) be two Hilbert spaces, \(T\) a given finite final time, and \(U\) a closed metric space. For every parameter \(z\in \mathbb{ R}^{n}\), the authors consider the differential equation: \(\overset{.}{X} _{z}(t)=AX_{z}(t)+B(z)u(t)+E(t,z)\), with the initial condition \( X_{z}(0)=x_{0}\in \mathbb{H}\), where \(A:\mathcal{D}(A)\subset \mathbb{H} \rightarrow \mathbb{H}\) is a linear operator generating a strongly continuous analytic semi-group on the Hilbert space \(\mathbb{H}\), the mapping \(B(z):U\rightarrow \lbrack \mathcal{D}(A)]^{\prime }\) is a linear operator, \(E:[0,T]\times \mathbb{R}^{n}\rightarrow \mathbb{H}\) is a given source term, and the control input \(u\) belongs to \(\mathcal{U}=\{u\in L^{2}([0,T];U)\) and \(u(s)\in U\) a.e. on \([0,T]\}\). \((\Omega ,\mathcal{F}, \mathbb{P})\) being a probability space, where \(\Omega \) represents the sample space, \(\mathcal{F}\) is the \(\sigma \)-algebra of events, and \(\mathbb{ P}:\Omega \rightarrow \lbrack 0,1]\) is a probability measure, the authors consider the controlled system: \(\overset{.}{x}_{\omega }(t)=Ax_{\omega }(t)+B(\xi (\omega ))u(t)+E(t,\xi (\omega ))\), \(x_{\omega }(0)=x_{0}\), where \(\xi \) is a given \(n\)-dimensional random variable in the probability space \( (\Omega ,\mathcal{F},\mathbb{P})\). The optimal control problem is written as: \(\mathrm{Maximize}\{\mathbb{E}[\int_{0}^{T}\ell (t,x^{u}(t),u(t),\cdot )dt]+ \mathbb{P}[\Psi (x^{u}(T))\leq 0]\), \(u(t)\in \mathcal{U}\}\), where \(\ell :[0,T]\times \mathbb{H}\times \mathbb{U}\times \Omega \rightarrow \mathbb{R}\) is a Carathéodory function which satisfies some measurability, continuity, Fréchet differentiability, and growth properties, and \(\Psi : \mathbb{H}\rightarrow \mathbb{R}\) is a convex and continuously Fréchet differentiable function which satisfies growth conditions together with its gradient. This problem may be written as: \(\mathrm{Maximize}\{\mathbb{E} [\int_{0}^{T}\ell (t,\mathcal{G}(u,\xi )(t),u(t),\cdot )dt]+\mathbb{P}[\Psi ( \mathcal{G}(u,\xi )(T))\leq 0]\), \(u(t)\in \mathcal{U}\}\), where \(\mathcal{G} (u,z)\) is the unique mild solution in \(L^{2}([0,T];\mathbb{H})\cap C([0,T]; \mathbb{H})\) to the preceding controlled system, which may be expressed in terms of the semigroup associated with \(A\). The main result of the paper presents optimality conditions of the optimal control problem in the form of Pontryagin's principle. For the proof, the authors first recall the definition of an elliptical distribution. They prove the continuity of the probability function associated with an elliptical distribution and its representation as a spherical integral, then its differentiability and the representation of its derivative as a spherical integral. They prove properties of the mappings \(\mathcal{G}\). The paper ends with two examples of such situations: the heat equation with Neumann boundary control and the 1D heat equation with pointwise control.