TL;DR
This paper introduces a new family of barycenter models based on the Schrödinger-Bass Bridge (SBB), a framework that unifies deterministic transport and stochastic diffusion. By leveraging the inherent asymmetry of SBB, the authors develop four distinct "oriented" barycenter formulations that significantly outperform traditional Wasserstein-based methods in recovering latent structures from noisy, measure-valued data like blurred images and single-cell RNA sequences.
What Problem Does This Paper Address?
In many fields, data points aren't just vectors; they are probability distributions (measures). For example, a population of cells in a biological sample or a blurred image can be viewed as a distribution of mass. A "barycenter" is essentially a geometric average of these distributions.
Standard tools like Wasserstein barycenters focus on "optimal displacement"—finding the most efficient way to move mass from one shape to another. However, real-world data often involves both systematic movement (drift) and random spreading (diffusion or blur). Existing models often struggle to distinguish between these two. Furthermore, most transport metrics are symmetric, but many processes (like time-series development) are inherently directed. The authors address these gaps by applying the recently proposed Schrödinger-Bass Bridge to the problem of distributional averaging.
Methods
The core of the paper is the application of the Schrödinger-Bass Bridge (SBB). Unlike the classical Schrödinger Bridge (which has fixed volatility) or Bass Martingales (which have zero drift), SBB allows both the drift and the volatility of a stochastic process to be controlled and optimized.
The authors propose four specific barycenter models based on the directionality of the SBB cost:
- SBB-Encoder: Finds a common terminal state reached from all input domains.
- SBB-Decoder: Learns a common source (latent state) from which all inputs are generated.
- SBB-Symmetric: Averages both directions to avoid directional bias.
- SBB-Markovian: Specifically designed for interpolation between two time points.
Theoretical and Algorithmic Contributions: * Existence Proofs: They prove that these barycenters actually exist under mild assumptions. * Gaussian Closure: They demonstrate that for Gaussian input measures, the decoder barycenter is also Gaussian. This is a huge win for efficiency, as it reduces a complex infinite-dimensional problem to simple mean and covariance updates. * Hybrid Particle Descent: For non-Gaussian data, they derive a "Hybrid Wasserstein Gradient Particle Descent" scheme. This uses dual potentials (from the Hamilton-Jacobi-Bellman equation) to update a set of particles representing the barycenter.
Datasets
The models were evaluated on two primary tasks:
- Spatial MNIST: A latent-source recovery task where a "clean" digit image is reconstructed from 32 observations that have undergone random geometric transformations (scaling, translation, rotation) and heterogeneous anisotropic blur.
- Single-cell RNA Sequencing (scRNA-seq): Using the Waddington-OT dataset, which tracks the 18-day reprogramming of mouse embryonic fibroblasts. The models were used to reconstruct "ancestor" populations and interpolate intermediate cell states between recorded time points.
Key Insights and Ideas
- Asymmetry is a Feature, Not a Bug: Because the SBB cost is directed, the position of the barycenter relative to the inputs matters. The "Decoder" formulation is particularly powerful for latent recovery because it mirrors the generative process of "source $\rightarrow$ noisy observations."
- Unifying Drift and Diffusion: SBB’s ability to jointly control drift and volatility allows it to "de-blur" measures more effectively than Wasserstein methods, which try to explain all variance through displacement alone.
- The Power of Dual Potentials: By utilizing the dual representation of the SBB problem, the authors turned a difficult optimization over the space of measures into a practical particle-update rule that can be implemented on modern GPUs.
Key Results
The SBB models consistently outperformed baselines like standard Wasserstein barycenters, Sinkhorn-based methods, and TrajectoryNet.
In the Spatial MNIST experiment, the Decoder SBB achieved the lowest error across all metrics, successfully recovering sharper and more accurate latent digits from heavily blurred inputs.
In the scRNA-seq tasks, the SBB-Markovian and Decoder models were superior at reconstructing the MET-like ancestry of cell lineages and interpolating held-out time points, showing better agreement with biological ground truth (measured by $W_2$ distance and Maximum Mean Discrepancy).
Limitations and Future Work
While the results are promising, the authors note a few areas for further exploration:
- Uniqueness: While they proved the existence of these barycenters, the conditions for uniqueness remain an open theoretical question.
- Encoder Gaussian Closure: The paper proves Gaussian closure for the decoder barycenter, but numerical evidence suggests this property might not hold for the encoder version.
- Convergence: Future work is needed to establish formal convergence guarantees for the proposed particle descent scheme.
Summary updated