🔢 mathematics

Contact Wasserstein Geodesics for Non-Conservative Schrödinger Bridges

This paper introduces the Non-Conservative Generalized Schrödinger Bridge (NCGSB), a novel framework based on contact Hamiltonian mechanics that overcomes energy-conservation limitations to model varying-energy stochastic processes via a computationally efficient, non-iterative Contact Wasserstein Geodesic (CWG) method implemented with ResNet architectures.

Original authors: Andrea Testa, Søren Hauberg, Tamim Asfour, Leonel Rozo

Published 2026-05-08

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Andrea Testa, Søren Hauberg, Tamim Asfour, Leonel Rozo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Predicting the Journey, Not Just the Destination

Imagine you are trying to predict how a crowd of people moves from a starting point (like a train station) to an ending point (like a stadium) over the course of an hour. You have photos of the crowd at the start and the finish, but you don't know what happened in between.

Traditional methods for solving this problem act like a perfectly efficient, frictionless robot. They assume that the "energy" of the crowd stays exactly the same the whole time. If the crowd starts moving fast, they must keep moving fast. If they slow down, it's only because they are taking a longer, more winding path, not because they are actually tired or stopping to rest.

The Problem: Real life isn't like that. Crowds get tired, storms lose intensity, and cells change energy as they grow. The old methods can't model these "energy-changing" events because they are stuck in a world where energy is never lost or gained.

The Solution: This paper introduces a new framework called NCGSB (Non-Conservative Generalized Schrödinger Bridge). Think of it as a smart, tired, and energetic crowd simulator. It allows the "energy" of the system to change over time, just like in the real world.

The Core Concepts, Simplified

1. The "Bridge" (Schrödinger Bridge)

In math, a "bridge" connects two points. Here, it connects two distributions (snapshots of data).

Old Way: Imagine a tightrope walker who must maintain a constant speed and balance perfectly. If they stumble, the math breaks.
New Way (NCGSB): Imagine a hiker on a mountain trail. They might sprint up a hill (gaining potential energy), then coast down a slope (losing energy), or stop to rest (dissipating energy). The new method can model this hiker perfectly, even if their energy levels fluctuate wildly.

2. The "Contact" (Contact Hamiltonian Mechanics)

The paper uses a fancy math concept called "Contact Hamiltonian Mechanics."

The Analogy: Think of a standard map (Symplectic geometry) where you can only move in directions that keep your total energy constant. It's like a video game where you can't gain or lose health points; you just move around.
The New Map (Contact Geometry): This is a 3D map where the "height" represents energy. You can move up (gain energy) or down (lose energy). The paper uses this 3D map to create a path that naturally accounts for things like friction, heat loss, or biological growth.

3. The "Geodesic" (The Shortest Path)

In geometry, a "geodesic" is the shortest path between two points on a curved surface (like a plane flying over the Earth).

The Innovation: The authors realized that finding the most likely path for this energy-changing crowd is the same as finding the shortest path on a specific, curved mathematical surface.
The Shortcut: Instead of trying to calculate this path by guessing and checking millions of times (which is slow and expensive), they turned the problem into a ResNet (a type of AI neural network).
- Analogy: Imagine trying to find the best route through a maze. The old way is to try every single path one by one. The new way is to build a robot (the ResNet) that learns the shape of the maze instantly and walks the perfect path in one go.

4. "Guided Generation" (Steering the Crowd)

Sometimes you don't just want any path; you want a specific kind of path.

The Analogy: You want the crowd to move from the station to the stadium, but you also want them to avoid a construction zone or pass through a specific park.
How it works: The new method allows you to "tweak" the mathematical map (the metric) to make certain paths cheaper or more expensive. If you want the crowd to go through the park, the math makes that route look like a smooth highway, while the construction zone becomes a muddy swamp. The AI naturally steers the crowd toward the park.

What Did They Actually Do? (The Results)

The authors tested their new "Energy-Changing Bridge" on several real-world tasks to prove it works better and faster than the old methods:

Navigating a Mountain (LiDAR Data): They simulated a path moving across a 3D map of Mount Rainier. The old methods tried to fly straight through the mountain (impossible) or took huge detours. The new method stayed on the ground (the mountain surface) and found a smooth, low-energy path.
Watching Cells Grow (Single Cell Sequencing): They tracked how stem cells turn into different types of cells. Since cells consume and release energy as they grow, the old "constant energy" models failed to predict the middle steps accurately. The new model captured the biological reality of energy changing over time.
Predicting Weather (Sea Surface Temperature): They tried to predict the temperature of the ocean in March and July, given data from January, May, and September. The new method predicted the "missing" months much more accurately and in a fraction of the time it took other methods.
Robotics: They reconstructed videos of robots moving objects. The new method filled in the missing frames of the video so smoothly that the robot's movement looked natural, whereas other methods made the robot jitter or move unrealistically.
Face Aging: They took pictures of adults and tried to generate what they looked like as children. The new method successfully transformed the faces while keeping them looking like the same person, whereas other methods often produced blurry or unrecognizable results.

Why Does This Matter?

It's Faster: The new method doesn't need to run thousands of slow, repetitive loops. It solves the problem almost instantly (near-linear speed).
It's More Realistic: By allowing energy to change, it can model real-world physics (like friction, heat, and biology) that previous math tools simply couldn't handle.
It's Flexible: It can be "guided" to follow specific rules or goals, making it useful for tasks like image editing or planning robot movements.

In short: The authors built a new mathematical "GPS" that understands that the journey isn't always about maintaining a perfect speed. It understands that sometimes you speed up, sometimes you slow down, and sometimes you stop to rest—and it can calculate the most likely path for all of those scenarios.

Technical Summary: Contact Wasserstein Geodesics for Non-Conservative Schrödinger Bridges

Problem Statement

The Schrödinger Bridge (SB) problem seeks the most likely stochastic path connecting two marginal distributions, typically relative to a reference process like Brownian motion. While SBs offer a principled framework for modeling stochastic processes in fields ranging from cellular dynamics to meteorology, existing methods face two primary limitations:

Energy Conservation Assumption: Standard SB formulations assume a constant energy profile throughout the stochastic path. This restricts the model to conservative systems, failing to capture real-world phenomena involving energy dissipation or external interactions (e.g., storms losing intensity or stem cell differentiation).
Computational Intractability: Current solvers often rely on iterative forward-backward stochastic simulations or progressive refinement of reference dynamics. These approaches operate in infinite-dimensional probability spaces, leading to high computational costs, slow convergence, and sensitivity to initialization.

Methodology

1. Non-Conservative Generalized Schrödinger Bridge (NCGSB)

The authors propose a reformulation of the SB problem based on contact Hamiltonian mechanics, a generalization of classical Hamiltonian mechanics that allows for energy variation.

State Augmentation: The system's state space is augmented with a scalar variable $z_t$ representing the Lagrangian action. The dynamics of $z_t$ depend recursively on the system's current state and a damping factor $\gamma$ .
Energy Variation: Unlike the Generalized SB (GSB), which minimizes a static cost functional, the NCGSB minimizes a time-varying state. The cost functional is defined as the integral of $\partial_t z_t$ , where the evolution of $z_t$ includes a damping term $-\gamma z_t$ . This structure allows the total energy of the system to vary over time, naturally modeling dissipative behaviors.
Guided Generation: The framework supports guided generation by incorporating a task-specific loss term $\|y - f(x_{t_s})\|^2$ into the Lagrangian action. This steers the generative process toward specific conditions at arbitrary time steps without breaking the underlying data manifold.

2. Contact Wasserstein Geodesics (CWG)

To make the NCGSB problem tractable, the authors lift the problem from the infinite-dimensional Wasserstein manifold to a finite-dimensional space using a geometric perspective.

Geometric Reformulation: The optimality conditions of the NCGSB are shown to correspond to a geodesic flow on an extended manifold $P_+(M) \times \mathbb{R}$ , governed by a Contact Jacobi Metric $\tilde{g}_J$ . This metric incorporates kinetic energy, potential energy, and an entropy term derived from the contact Hamiltonian.
ResNet Parameterization: The continuous geodesic is discretized into a sequence of probability transitions. The authors map this discretization to a Residual Network (ResNet) architecture. Each ResNet block corresponds to a segment of the geodesic, parameterizing the pushforward map between consecutive time steps.
Non-Iterative Solver: By framing the problem as a geodesic computation in a finite-dimensional parameter space $\Theta$ $Θ$ , the method avoids the expensive outer iteration loops typical of Iterative Proportional Fitting (IPF) or forward-backward SDE solvers. The training objective minimizes a loss function comprising:
- Terminal and intermediate marginal matching (Wasserstein-2 distance).
- Energy minimization along the path, weighted by the scaling factor $\Phi_{t_k}$ derived from the Contact Jacobi metric.

3. Complexity and Scalability

The proposed solver achieves near-linear complexity with respect to both the data dimensionality $D$ and the batch size $N$ . This is a significant improvement over existing methods that scale quadratically or exponentially with dimension or require iterative convergence.

Key Contributions

NCGSB Formulation: A novel non-conservative formulation of the Schrödinger Bridge that models energy-varying systems by extending the Hamiltonian framework to contact Hamiltonian dynamics.
CWG Framework: The introduction of the Contact Wasserstein Geodesic (CWG), a general geometric solver compatible with all SB variants. It transforms the problem into a tractable geodesic computation on a finite-dimensional manifold, enabling fast, non-iterative training via ResNet architectures.
Guided Generation: A method for steering the learned bridge using task-specific losses by modulating the Riemannian metric, allowing for conditional generation without compromising the stochastic nature of the bridge.

Experimental Results

The framework was validated across diverse tasks, demonstrating superior performance in accuracy, feasibility, and computational efficiency compared to baselines (DSBM, SB-Flow, GSBM, SBIRR, DM-SB).

LiDAR Manifold Navigation: CWG successfully computed bridges on a complex 3D manifold (Mount Rainier) while respecting low-altitude constraints. It achieved lower optimality costs and higher feasibility (boundary satisfaction) than baselines, with training times roughly 5x faster than GSBM.
Single-Cell Sequencing: In reconstructing stem cell differentiation (a non-conservative biological process), CWG outperformed mmSB baselines in Wasserstein error on validation sets. The energy-varying formulation captured the rapid early-stage dynamics of differentiation more accurately than energy-conserving models.
Image Generation:
- Sea Surface Temperature: CWG predicted intermediate monthly temperature maps with significantly lower FID scores than baselines and achieved training speeds orders of magnitude faster (e.g., 17s vs. 1227s for a specific year).
- Robotic Manipulation: The model generated realistic intermediate frames for robot pick-and-place tasks, outperforming baselines in image quality (FID) and training time.
- Unpaired Image Transfer (FFHQ): In mapping adult faces to children, CWG successfully satisfied boundary constraints (Feasibility), whereas baselines failed to match the target distribution despite lower transport costs.
Guided Generation: The method successfully steered generation in LiDAR navigation (targeting specific mountain sides) and robotic tasks (placing items at specific coordinates) with minimal degradation in image quality.

Significance and Claims

The paper claims that by moving from conservative to non-conservative formulations via contact geometry, the Schrödinger Bridge framework can model a broader class of real-world stochastic processes that involve energy dissipation. The primary significance lies in the computational efficiency of the CWG solver; by leveraging the geometric structure of the problem and ResNet parameterization, it eliminates the need for costly iterative loops. This enables the application of SBs to high-dimensional, complex tasks (like image generation and molecular dynamics) that were previously computationally prohibitive. The authors emphasize that their approach provides a "principled framework" that is both more expressive (handling energy variation) and significantly faster than existing state-of-the-art methods.