Introduction
In scientific domains such as single-cell biology and meteorology, researchers frequently encounter "snapshot" data—observations of a system at discrete time intervals without the underlying continuous paths. Reconstructing these trajectories is a fundamental challenge in generative modeling, typically addressed through the Schrödinger Bridge (SB) problem, which seeks the most likely stochastic path between probability distributions.
While modern "matching" frameworks (e.g., Flow or Bridge Matching) have advanced the field, they often operate piecewise, interpolating between consecutive snapshots. This localized approach frequently fails to capture long-range temporal dependencies, resulting in "kinked" or physically inconsistent trajectories. Furthermore, existing momentum-based global solvers are often computationally prohibitive due to their reliance on numerical simulations during training.
Methodology
To overcome these limitations, the authors propose Momentum Multi-Marginal Schrödinger Bridge Matching (3MSBM). This framework introduces three primary innovations:
- Phase Space Lifting: By augmenting the state space to include both position ($x$) and velocity ($v$), 3MSBM ensures that inferred trajectories possess physical momentum and well-defined curvature.
- Multi-Marginal Momentum Brownian Bridge (3MBB): The authors derived a closed-form expression for a stochastic bridge conditioned on multiple future points. This allows the model's acceleration at any given time to be informed by all subsequent snapshots, ensuring global coherence.
- Simulation-Free Matching: The analytic nature of the 3MBB enables training via a regression-based matching objective. Unlike previous methods, 3MSBM does not require expensive SDE solvers during the training loop.
To handle datasets lacking explicit velocity information, the framework employs an Iterative Propagation scheme to refine velocity profiles across marginals over several training iterations.
Experimental Evaluation
The model was benchmarked against synthetic and real-world datasets:
- Lotka-Volterra: A predator-prey system with oscillatory dynamics.
- Gulf of Mexico (GoM): Ocean surface current velocity fields.
- Beijing Air Quality: PM2.5 pollutant tracking across multiple sensors.
- Embryoid Body (EB): High-dimensional (100D) single-cell sequencing data.
Results and Discussion
3MSBM consistently outperformed baselines in both accuracy and efficiency. It achieved the lowest Sliced Wasserstein Distance (SWD) and Maximum Mean Discrepancy (MMD) across all tasks. Notably, on the 100-dimensional EB dataset, 3MSBM was approximately 2.5x faster than the previous state-of-the-art momentum method (DMSB) while maintaining a significantly lower memory footprint.
Qualitatively, 3MSBM produces smooth, physically plausible curves that accurately pass through validation snapshots, avoiding the artifacts seen in vanilla SB implementations.
A key insight from the study is the exponential decay of influence: the impact of a future snapshot on current acceleration diminishes as the time gap increases. This allows for efficient "look-ahead" truncation without sacrificing global accuracy.
Conclusion
3MSBM represents a significant step forward in multi-marginal trajectory inference. By combining phase space lifting with a closed-form multi-marginal bridge, it achieves global coherence and physical smoothness without the computational burden of simulation-based training. Future work will explore the application of this framework to high-dimensional video generation and refining velocity estimation for non-Gaussian initial distributions.
Summary updated