Recovering Stochastic Dynamics via Gaussian Schrödinger Bridges

Charlotte Bunne, Ya-Ping Hsieh, marco cuturi, Andreas KrauseOpenReview 2022 · OpenReview

Abstract

The static optimal transport (OT) problem between Gaussians seeks to recover an optimal map, or more generally a coupling, to morph a Gaussian into another. It has been well studied and applied to a wide variety of tasks. Here we focus on the dynamic formulation of OT, also known as the Schrödinger bridge (SB) problem, which has recently seen a surge of interest in machine learning due to its connections with diffusion-based generative models. In contrast to the static setting, much less is known about the dynamic setting, even for Gaussian distributions. In this paper, we provide closed-form expressions for SBs between Gaussian measures. In contrast to the static Gaussian OT problem, which can be simply reduced to studying convex programs, our framework for solving SBs requires significantly more involved tools such as Riemannian geometry and generator theory. Notably, we establish that the solutions of SBs between Gaussian measures are themselves Gaussian processes with explicit mean and covariance kernels, and thus are readily amenable for many downstream applications such as generative modeling or interpolation. To demonstrate the utility, we devise a new method for modeling the evolution of single-cell genomics data and report significantly improved numerical stability compared to existing SB-based approaches.

Introduction

In machine learning, understanding how one distribution of data evolves into another over time—such as a stem cell transforming into a specialized lung cell—is a fundamental challenge addressed by Optimal Transport (OT). While static OT identifies the final mapping of mass, the Schrödinger Bridge (SB) problem seeks the dynamic path: the stochastic process connecting two distributions while remaining close to a reference prior, such as Brownian motion.

Solving SB problems is notoriously difficult, often resulting in numerical instability and high sensitivity to hyperparameters. While closed-form solutions for static Gaussian OT have existed for decades, a general closed-form solution for the dynamic Gaussian SB remained elusive. This gap forced researchers to rely on "data-oblivious" priors that ignored the basic shape of the target data, leading to significant training challenges.

Methodology: The GSBFLOW Framework

This work bridges the theoretical gap by deriving explicit, closed-form expressions for the mean and covariance of the stochastic process connecting two Gaussians. The authors reframe the SB problem as an action minimization problem on the Bures-Wasserstein manifold (the space of Gaussian distributions), showing that the optimal path is a geodesic modified by a potential energy term related to Fisher information.

The GSBFLOW framework connects Gaussian space theory with data space dynamics through closed-form SDE solutions.

To implement these findings, the authors propose GSBFLOW, a two-stage algorithm:

  1. Gaussian Approximation: The complex data distributions are approximated as Gaussians, and the GSB is calculated in closed form.
  2. Refined Training: This GSB serves as a "warm start" or refined prior for training neural networks, which then capture the remaining non-Gaussian details of the true dynamics. By starting with a "smart" prior that matches the first two moments of the data, the subsequent training becomes significantly more robust.

Experimental Results

The authors evaluated GSBFLOW on synthetic dynamics and high-stakes biological datasets, specifically tracking human embryonic stem cell development and iPSC reprogramming. The evaluation focused on the principal components of gene expression space, representing over 1,500 highly variable genes.

Quantitative Performance

GSBFLOW demonstrated superior numerical stability compared to standard SB baselines like VESDE, which frequently failed to produce results due to scaling or initialization issues. In the iPSC reprogramming task, GSBFLOW achieved a significantly lower Wasserstein Loss ($W_\epsilon$) than existing methods.

Quantitative comparison of Wasserstein Loss across single-cell genomics tasks, showing GSBFLOW's superior performance.

Qualitative Insights

Beyond numerical metrics, GSBFLOW successfully recovered the branching structures of cell differentiation, accurately identifying neural crest versus mesoderm lineages. This capability suggests that the geometric interpretation of the Schrödinger Bridge as a shortest-path problem on a curved manifold provides a powerful lens for biological modeling.

Conclusion and Future Work

This paper provides the first general closed-form solution for the dynamic Gaussian Schrödinger Bridge, offering a principled way to handle matrix equations involving covariance via Lyapunov operators. While the current $O(d^3)$ computational complexity poses challenges for raw high-dimensional images, the method excels in dimensionality-reduced spaces common in genomics. Future work will focus on extending this framework to non-Euclidean manifolds and developing more scalable matrix approximations.

Summary updated

Abstract

We propose a new framework to reconstruct a stochastic process {ℙt:t∈[0,T]} using only samples from its marginal distributions, observed at start and end times 0 and T. This reconstruction is useful to infer population dynamics, a crucial challenge, e.g., when modeling the time-evolution of cell populations from single-cell sequencing data. Our general framework encompasses the more specific Schrödinger bridge (SB) problem, where ℙt represents the evolution of a thermodynamic system at almost equilibrium. Estimating such bridges is notoriously difficult, motivating our proposal for a novel adaptive scheme called the GSBflow. Our goal is to rely on Gaussian approximations of the data to provide the reference stochastic process needed to estimate SB. To that end, we solve the \acs{SB} problem with Gaussian marginals, for which we provide, as a central contribution, a closed-form solution and SDE-representation. We use these formulas to define the reference process used to estimate more complex SBs, and show that this does indeed help with its numerical solution. We obtain notable improvements when reconstructing both synthetic processes and single-cell genomics experiments.
BibTeX
@misc{bunne2022recovering,
  author = {Charlotte Bunne and Ya-Ping Hsieh and marco cuturi and Andreas Krause},
  title = {{Recovering Stochastic Dynamics via Gaussian Schrödinger Bridges}},
  year = {2022},
  url = {https://openreview.net/forum?id=FCRjXQIzwd},
  howpublished = {OpenReview},
}

Loading related content...

Figure