Optimal control of the Fokker-Planck equation under state constraints in the Wasserstein space (Q6105325)

Motivated by stochastic control problems, the author considers the problem \[ \min_{\alpha,m} \int_0^T \int_{\mathbb{R}^d} L(x,\alpha(t,x)dm(t)(x)dt + \int_0^T \mathcal{F(m(t))}dt + \mathcal{G}(m(T)) \] for \(\mathcal{F}\), \(\mathcal{G}\) bounded from below and continuously differentiable, subject to the Fokker-Planck equation \[ \partial_t m + \mathrm{div}(\alpha m) -\Delta m = 0,\quad m(0) = m_0, \] and pointwise state constraints \[ \Psi(m(t)) \leq 0 \] for \(\Psi\) convex and satisfying a Slater condition. Using a one-homogeneous penalization of the state constraints, the author derives necessary optimality conditions involving an Hamilton-Jacobi-Bellman equation for the adjoint state as well as a final-time transversality condition. Under the additional condition that \(\Psi\) is sufficiently smooth and that the transversality assumption \[ \int_{\mathbb{R}^d} |D_m \Psi(m,x)|^2 dm(x) \neq 0 \quad\text{whenever}\quad \Psi(m)=0 \] holds, it is shown that the penalization is exact, which implies that the optimal control \(\alpha\) is Lipschitz continuous in time and that the Lagrange multiplier for the state constraint is in \(L^\infty(0,T)\) instead of a measure.