Trajectory Inference via Acceleration Matching

This paper introduces Acceleration Matching (AM), an algorithm for inferring smooth trajectories from unpaired observations collected at discrete time points. The method lifts the interpolation problem into phase space and learns an explicit conditional acceleration field that generates random trajectories consistent with prescribed marginal distributions. According to the abstract, training uses only positional data and avoids trajectory simulation and expensive preprocessing. The authors report competitive or superior numerical results on several benchmark problems, although the supplied metadata does not include detailed datasets, baselines, metrics, or ablation results.