Optimal control under reduced regularity (Q1030777)

This paper is aimed at studying optimal control under reduced regularity. The authors deal with the numerical solution of the following control-constrained optimal control problem \[ J(\overline{u}) = \min_{u \in U_{ad} } J(u)\text{ with }J(u) = \frac{1}{2}\|Su - y_d\|^2_{L^2(\Omega)}+\frac{\nu}{2}\|u\|^2_{L^2(\Omega)}\tag{1} \] where the operator \(S\) associates the state \(y=Su\) to the control \(u\) as the weak solution of equential quadratic programming methods \[ \begin{gathered} Ly=u\text{ in }\Omega,\quad y = 0\text{ on }\Gamma=\partial \Omega, \quad U=L^\infty (\Omega)\\ \text{ and }[a,b]\subset \mathbb R, U_{ad}= \{u \in U: a \leq u(x) \leq b \text{ a.e. in } \Omega\}\text{(the set of admissible controls.)}\tag{2}\end{gathered} \] The second order elliptic differential operator \(Ly: = \nabla .(A \nabla y) + a. \nabla y + a_0y\) (with smooth coefficient functions \(A(x) \in \mathbb R^{3 \times 3}, a(x) \in \mathbb R^3\) and \(a_0(x) \in \mathbb R\) such that: \(\exists m_0 >0: m_0|\xi|^2 \leq \xi^T A\xi\) \(\forall x \in \Omega, \forall \xi \in \mathbb R^3, a_0- \frac{1}{2} \nabla.a \geq 0, \forall x \in \Omega\), \(A\) is symmetric) is investigated. The numerical solution of the optimal control problem (1), (2) is based on a linear or bilinear discretization of the state variable leading to the discrete solution operator \(S_h (J_h(u)= \frac{1}{2}\|S_h u - y_d\|^2_{L^2(\Omega)}+\frac{\nu}{2}\|u\|^2_{L^2(\Omega)})\). Main result: The authors investigate the three-dimensional case in the presence of corner and edge singularities and show for both methods that the approximation order \(\alpha = 2\) can be retained although the state and adjoint state do not have full regularity. The paper does neither target on a comparison of the two approaches nor on new concepts for the numerical solution of optimal control problems. The reduced regularity of the solution leads to a lower convergence order if standard methods on quasi-uniform meshes are used. Estimates of the finite element error: \([Su - S_h u] \) are given in the \(H^1 (\Omega)\)- and \(L^2(\Omega)\)-norms. It is proved that the optimal control can be calculated by the finite element method at a rate of \(O(h^2)\) provided that the mesh is sufficiently graded. The presented numerical results confirm the expected convergence rates.