Abstract
We propose a new method for smoothly interpolating probability measures using the geometry of optimal transport. To that end, we reduce this problem to the classical Euclidean setting, allowing us to directly leverage the extensive toolbox of spline interpolation. Unlike previous approaches to measure-valued splines, our interpolated curves (i) have a clear interpretation as governing particle flows, which is natural for applications, and (ii) come with the first approximation guarantees on Wasserstein space. Finally, we demonstrate the broad applicability of our interpolation methodology by fitting surfaces of measures using thin-plate splines.
Introduction
Interpolation is a fundamental task in data science, typically used to estimate values between discrete observations in Euclidean space. However, emerging fields like single-cell genomics and meteorology require interpolating between probability measures (distributions). The natural geometry for such distributions is the 2-Wasserstein space, which treats them as points on a curved manifold. Traditional methods for spline interpolation in this space, often referred to as P-splines, are computationally intensive and frequently produce non-deterministic "fuzzy" trajectories where mass splits across multiple paths—a behavior that lacks physical or biological realism.
Methodology: Transport Splines
The proposed "transport splines" framework addresses these challenges by decoupling the interpolation process into two manageable steps:
- Coupling via Monge Maps: The distributions at observed time points are linked using optimal transport (Monge) maps. This process establishes a one-to-one correspondence between points across the distributions.
- Euclidean Spline Fitting: Once coupled, the problem is reduced to the Euclidean domain. Classical interpolation techniques, such as cubic splines or thin-plate splines, are applied to the individual particle trajectories defined by the coupling.
This approach allows the method to inherit the efficiency and smoothness of classical splines. For Gaussian distributions, the process is particularly fast, involving only basic matrix operations.
Key Insights and Theoretical Results
A major contribution of this work is the shift from Kantorovich-style solutions (which allow mass splitting) to Monge solutions. This ensures that every "particle" follows a unique, smooth path, providing a clear physical interpretation of the interpolation as a particle flow.
Furthermore, this paper establishes the first quantitative approximation error bounds for Wasserstein interpolation. The authors prove an $O(\delta^2)$ convergence rate (where $\delta$ is the time step), matching the optimal theoretical rate for classical Euclidean splines.
Experimental Evaluation
The methodology was validated across three domains:
- Synthetic Gaussian Data: Demonstrating the transition from linear to cubic interpolation while maintaining the Gaussian property.
- N-Body Physics: Reconstructing empirical distributions of particles in motion, showing the method's ability to capture complex dynamics.
- Geospatial Temperature Modeling: Using thin-plate transport splines to interpolate temperature distributions across California, showcasing the framework's flexibility with spatial covariates.
As shown in the qualitative results, transport splines produce clean, non-crossing trajectories, whereas previous P-splines often result in a "fuzzy" superposition of paths that are difficult to interpret.
Conclusion and Future Directions
Transport splines offer a scalable, theoretically grounded, and physically interpretable solution for measure-valued interpolation. Future work will focus on extending the framework to handle noisy observations through "smoothing splines" and exploring higher-order accuracy by refining velocity field approximations.
Summary updated