First-order hold lossless convexification: theoretical guarantees for discrete-time optimal control problems

Abstract: Lossless Convexification (LCvx) is a clever trick that transforms a class of nonconvex optimal control problems (where the nonconvexity arises from a lower bound on the control norm) into equivalent convex problems via convex relaxations, the goal being to solve these problems efficiently via polynomial-time numerical solvers. However, to solve these infinite-dimensional problems in practice, they must first be converted into finite-dimensional problems, and it remains an open area of research to ensure the theoretical guarantees of LCvx are maintained across this discretization step. Prior work has proven guarantees for zero-order hold control parameterization. In this work, we extend these results to the more general, and practically useful, first-order hold control parameterization. We first show that under mild assumptions, we are guaranteed a solution that violates our nonconvex constraint at no more than $n_x + 1$ vertices in our discretized trajectory (where $n_x$ is the dimension of our state-space). Then, we discuss an algorithm that, for a specific case of problems, finds a solution where our nonconvex constraint is violated along no more than $2n_x + 2$ edges in at most $\lceil \log_2 ((\rho_{\max} - \rho_{\min}) / \varepsilon_\rho) \rceil + 1$ calls to our solver (where $[\rho_{\min}, \rho_{\max}]$ represent the bounds on our control norm and $\varepsilon_\rho$ is some desired suboptimality tolerance). Finally, we provide numerical results demonstrating the effectiveness of our proposed method.