Population dynamics are often recorded as separate snapshots at different times, without linking individual observations across time. As a result, the path the population took between measurements is not uniquely determined. A new paper on arXiv addresses this by introducing a regularization strategy based on second-order properties, specifically penalizing acceleration along the inferred trajectory.
The authors propose 'neural Wasserstein splines,' which combine optimal transport with neural networks to produce smooth, plausible trajectories. Because the penalty targets acceleration, the method favors paths without abrupt changes in velocity. The approach is also representation-invariant, meaning the result does not depend on how the population state is parameterized or represented.
This summary is based on a single source, so there are no differing viewpoints to note. The provided abstract is truncated, and the full paper would be needed to assess experimental validation or comparisons with other trajectory inference methods.