A stochastic mirror descent algorithm for mean-field control problems

Source abstract

We develop a stochastic mirror descent method for continuous-time McKean-Vlasov control with measure-valued controls and general Bregman regularisation. The dependence on the state law leads to a coupled Wasserstein-Bregman stability problem and to a McKean-Vlasov adjoint equation, which have no counterpart in classical stochastic control. We establish a first-order representation of the cost and prove its relative smoothness in two regimes: nonlinear dynamics with bounded cost derivatives, and affine state-law dynamics with linearly growing cost derivatives, the latter covering linear-quadratic (LQ) and more general nonquadratic models. Our regularity conditions in the control measure are formulated through directional finite increments and do not require second-order flat derivatives. We further introduce relative semiconvexity along admissible forward-backward trajectories, allowing compensation between the dynamics, running cost, and terminal cost beyond pointwise Hamiltonian convexity. This yields energy dissipation, an O(1/n) convergence rate at the convexity threshold, and geometric convergence when regularisation dominates relative nonconvexity. Explicit LQ criteria and numerical experiments illustrate the scope of the results.