Optimization of source parameters in multipoint nonseparated conditions for linear dynamical systems (Q2030113)

The authors handle the problem of optimizing the right-hand sides of linear nonlocal multipoint conditions for a linear system of ordinary differential equations. It is pointed out that such a problem can arise in the form of an inverse problem of recovering parameters of boundary conditions or even in the optimization of the dynamics of an object by means of sources describing the boundary conditions. Convexity, respectively, strong convexity assumption on the functions defining the objective functional assure that the same property holds for it. Further, differentiability of the objective functional and an explicity formula for its gradient are obtained under corresponding conditions. First-order necessary optimality conditions are proven for the optimization parameters in the boundary conditions, which allows the use of first-order methods for a numerical solution of the problem. More specifically, keeping the authors notation, one considers the linear system of differential equations: \[ \dot{x}(t) = A(t)x(t)+B(t), \quad t \in [t_1,t_f] \tag{1} \] with constraints \[ \sum^\ell_{i=1} \alpha_i\,x(\bar{t}_i) = \mathcal{V}, \tag{2} \] where \(x(t) \in \mathbb{R}^n\) is a \(C^1\) function, \(A(t)\) and \(\alpha_i\) \, \((i=1,2,\dots,L)\) are \(n\)-dimensional square matrices, \(A(t)\) is continuous and non-constant for \(T \in [t_1,t_f]\), \(B(t) \in \mathbb{R}^n\) is a \(C^1\) function, \(L \ge 2\) is an integer and \(\bar{t}_i \in [t_1,t_f]\) for \(i=1,2,\dots,L\). Here \(\mathcal{V} \in V \subset \mathbb{R}^n\) is the parameter of the problem. Given the feasible convex compact set \(V\), one searches \(\mathcal{V}\) in \(V\) that minimizes the functional \[ J(\mathcal{V}) = \int_{t_1}^{t_f} f^0(x,(t),\mathcal{V},t)dt + \Phi(\bar{x},\mathcal{V}), \tag{3} \] where \(f^0(x,\mathcal{V},t)\) and \(\Phi(\bar{x},\mathcal{V})\) are continuously differentiable functions of their first two variables. The first result of the paper states that by considering problem (1)--(3): If \(f^0(x,\mathcal{V},t)\) and \(\Phi(\bar{x},\mathcal{V})\) are convex (resp strictly convex) functions of their first two variables then the functional \(J(\mathcal{V})\) is also convex (resp strictly convex). The second result assures, under the conditions on the functions and parameters in (1)--(3), the differentiability of \(J\). Further, it is proven that \[ \mathrm{grad}\, J\mathcal(V) = \int_{t_1}^{t_f} \frac{\partial f^0(x,\mathcal{V},t)}{\partial\mathcal{V}}\,dt + \frac{\partial\Phi(\bar{x},\mathcal{V})}{\partial\mathcal{V}} - \lambda, \tag{4} \] where \(\lambda \in \mathbb{R}^n\). The authors also give necessary optimality conditions for (1)--(3) in the form: Suppose \(\big(\mathcal{V}^*, x^*(t)\big)\) be a solution of (1)--(3) then it is necessary and sufficient for an arbitrary admissible \(\mathcal{V} \in V\), \[ (\mathrm{grad}\, J(\mathcal{V}^*), \,\mathcal{V}-\mathcal{V}^*) \ge 0, \] where \(\mathrm{grad} J(\mathcal{V}^*)\) is given by (4), among other possibilities. In the final part of the paper it is presented a detailed scheme for the numerical solution, by applying iterative first order methods, in two concrete instances for two-dimensional cases.