A simple proof of the discrete time geometric Pontryagin maximum principle on smooth manifolds (Q2173993)

The systems evolve in an \(n\)-dimensional manifold \(\mathbb M\) according to the equation \[ x_{t+1} = f_t(x_t, u_t), \quad t = 0,\dots, T - 1, \] where \(x_t \in \mathbb{M}\), \(u_t \in \mathbb{R}^m\) and \( \{f_t; t = 0, \dots, T-1 \}\) is a family of maps from \(\mathbb{M} \times \mathbb{R}^m\) into \(\mathbb{M}\), so that (discrete) trajectories that start in \(\mathbb M\) stay in \(\mathbb M\). The sequence \( \{u_t; t = 0, \dots, T-1 \}\) in \(\mathbb R^m\) is the control sequence, and the problem is that of minimizing the functional \[ \sum_{t=0}^{T-1} c_t(x_t, u_t) + c_T(x_T) \] over all control sequences. There are three types of constraints: state, control and frequency. The state constraints are \(x_0 = q_0\) and \(g_t(x_t) \le 0\) \((t = 1, \dots, T)\), where \(g_t\) maps \(\mathbb M\) into \(\mathbb R^{n(t)}\). The control constraint is the usual \(u_t \in U_t\) \((t = 0, \dots, T-1)\). Finally, the frequency constraints exclude certain components of the discrete Fourier transform of the control sequence, and can be reduced to \(\sum F_tu_t = 0\), where the \(F_t\) are appropriate matrices. The authors introduce the adjoint trajectory and prove Pontryagin's maximum principle in the usual Hamiltonian minimization form \(\langle H_t, u_t\rangle \le \langle H_t, \bar u_t \rangle = 0\) for controls \(\{u_t\}\) near the optimal control \(\{\bar u_t\}\), although \(H_t\) is not the true Hamiltonian due to the constraints. The principle includes a complementary slackness condition and a transversality condition. The method used is imbedding \(\mathbb M\) in an Euclidean space, proving the maximum principle there and then restrict it to \(\mathbb M\). The authors state that constraints are easier to include in discrete rather in continuous time problems, which suggests that discretization should be the initial step of the treatment of the latter.