WorksMathematics
A Derivative-Free Bellman Equation for Dynamic Time-to-Collision: Two-Sided Certificates and Explicit ReLU Approximation
Abstract
Dynamic time-to-collision (TTC) is the first contact time of a prediction flow and defines a feedback map implicitly. We study its capped value function \VV for Carathéodory dynamics with bounded, measurable forcing. A self-contained argument establishes local Lipschitz continuity on the transversal collision domain. A derivative-free stopped-step Bellman equation then characterizes \VV uniquely: it advances the prediction for a fixed duration or until contact and adds the elapsed time. A two-sided error bound uses only the candidate's residual along complete stopped trajectories and its error at their endpoints. Exact continuum iteration recovers \VV after finitely many sweeps, including one additional sweep for initializations that violate the boundary condition. We also characterize an optional-duration infimum operator: its fixed points are the dynamic-programming subsolutions, so its residual alone does not identify \VV. Finally, on compact subsets of the transversal domain, an optimization-free ReLU construction has explicit size, controlled local residual, and robustness to errors in its nodal TTC values. The optimal velocity model provides the running application.