A proximal gradient method for control problems with non-smooth and non-convex control cost (Q2231051)

In the paper the authors study a possibly non-smooth optimal control problem of type \begin{align*} \min(f(u)+\int_\Omega g(u(x))\,dx:\,u\in L^2(\Omega))\,. \tag{P}\end{align*} Here \(\Omega\) is a Lebesgue measurable subset of \(\mathbb{R}^n\), \(f:L^2(\Omega)\to\mathbb{R}\) is smooth and \(g:\mathbb{R}\to\mathbb{R}\cup\{+\infty\}\), which can be non-smooth and non-convex, is chosen to promote sparsity, that is, local solutions of (P) are zero on a significant part of \(\Omega\). Since \(\int_\Omega g(u(x))\,dx\) is not weakly lower semicontinuous in \(L^2(\Omega)\), due to the lack of convexity of \(g\), the direct method of calculus of variations cannot be used to show existence of solutions of (P). This problem could be overcome by adding a regularizing terms of the type \(\frac{\alpha}{2}\|u\|_{H^1}^2\) but, even if solutions to the regularized problem would converge to the solution of (P) as \(\alpha\to 0\), it is not clear how this can be exploited algorithmically. The method proposed by the authors is the so-called ``proximal gradient method'' (or forward-backward algorithm) to compute candidates for solutions. For any iterate \(u_k\), the next iterate \(u_{k+1}\) is computed as \begin{align*} u_{k+1}\in\mathrm{arg}\min_{u\in L^2(\Omega)}(f(u_k)+\nabla f(u_k)\cdot(u-u_k)+\frac{L}{2}\|u-u_k\|_{L^2(\Omega)}^2+j(u))\,,\tag{1} \end{align*} where \(L>0\) is a proximal parameter and \(j(u):=\int_\Omega g(u(x))\,dx\). Problem (1) can be equivalently written as \begin{align*} u_{k+1}\in\mathrm{prox}_{L^{-1}j}(u_k-\frac{1}{L}\nabla f(u_k))\,, \end{align*} where, for any \(\gamma>0\), the prox-map is defined as \[ \mathrm{prox}_{\gamma j}(z):=\mathrm{arg}\min_{u\in L^2(\Omega)}(\frac{1}{2}\|u-z\|_{L^2(\Omega)}^2+\gamma j(u))\,. \] The aim of the authors is to show that weak limit points in \(L^2(\Omega)\) of the proximal gradient method are \(L\)-stationary and, to prevent convexification, they employ ideas of the second author [SIAM J. Control Optim. 57, No. 2, 854--879 (2019; Zbl 1410.49011)]. The properties they show are weaker than those provided by Pontryagin's maximum principle and weaker than L-stationarity.