A Trust Region Approach for Learning Schrödinger Bridges

Takeshi Koshizuka, Denis Blessing, Max Zimmer, Sebastian Pokutta, Lorenz RichterarXiv 2026 · arXiv

TL;DR

This paper introduces the Trust Region Schrödinger Bridge Sampler (TR-SBS), a robust algorithm for sampling from unnormalized densities. By constraining successive updates within a Kullback-Leibler (KL) trust region, the authors stabilize the notoriously brittle Iterative Proportional Fitting (IPF) procedure, preventing mode collapse and enabling efficient sampling in high-dimensional, multi-modal landscapes.

What Problem Does This Paper Address?

Sampling from a target distribution where you only know the unnormalized density (the "energy function") is a fundamental challenge in machine learning and physics. Schrödinger Bridges (SBs) offer a powerful framework for this by finding a stochastic process that transports a simple prior (like a Gaussian) to the complex target with minimal "control cost."

The standard way to solve for SBs is Iterative Proportional Fitting (IPF). However, when target samples are unavailable—the "data-free" regime—IPF becomes highly unstable. Because the optimal solution is often far from the initialization, the algorithm's "projections" can be too aggressive, leading to numerical explosions or mode collapse, where the sampler ignores large parts of the distribution.

Methods

The authors propose Trust Region IPF (TR-IPF). Instead of taking a full step toward the target projection in each iteration, they constrain the update to stay within a fixed KL-divergence radius $\epsilon$ from the current iterate.

Key Innovations:

  • Adaptive Step Sizes: They prove that this trust-region constraint is equivalent to a mirror descent scheme on path space with an adaptively chosen step size $\gamma_i$. This step size is derived from a Lagrange multiplier $\lambda_i$ that is optimized in each iteration.
  • TR-SBS Algorithm: They implement this theory using a "matching" objective. Unlike previous methods like the Adjoint Schrödinger Bridge Sampler (ASBS), TR-SBS incorporates importance weighting.
  • Variance Control: Crucially, the trust region keeps the variance of these importance weights under control, making the training numerically stable even in high dimensions.
Conceptual comparison of standard IPF vs. the proposed Trust Region approach.

As shown in the figure above, standard IPF (left) can take large, unstable jumps, while the Trust Region approach (right) ensures equidistant, stable updates in terms of KL divergence.

Datasets

The authors evaluated TR-SBS on several challenging benchmarks:

  1. Many-Well (MW) Potential: A high-dimensional ($d=64$) landscape with up to $2^{32}$ symmetric modes, testing global mass discovery.
  2. Gaussian Mixture Model (GMM): Used to test scalability with dimension (up to $d=100$) and the number of components.
  3. Alanine Dipeptide: A molecular dynamics benchmark ($d=60$) where the goal is to capture a rare metastable conformation (C7ax) that accounts for only ~0.17% of the population.
  4. Ginzburg-Landau ($\phi^4$ scalar field theory): A physics model characterized by strong spatial correlations and phase transitions.
  5. Lennard-Jones cluster (LJ-13): A molecular cluster benchmark.

Key Insights and Ideas

  • The Stability-Speed Tradeoff: The trust region radius $\epsilon$ acts as a safety valve. A smaller $\epsilon$ provides rock-solid stability but requires more iterations to converge.
  • Solving "Mode Blindness": Purely gradient-based samplers often miss well-separated modes because they don't "see" the relative weights between them. By safely re-introducing importance weighting via the trust region, TR-SBS can balance mass between distant modes correctly.
  • Mirror Descent Interpretation: The paper provides a rigorous link between Schrödinger Bridges and mirror descent, showing that the trust region update is essentially a "geometric average" of the current path measure and the target.
  • Self-Correcting Dynamics: In early training, the algorithm automatically chooses small step sizes to avoid collapse. As the model improves and the "moving target" becomes clearer, the step size naturally increases.
Ablation study showing how the adaptive step size $\gamma$ starts small for stability and increases to accelerate convergence.

Key Results

  • Robustness to Complexity: On the Many-Well task, TR-SBS maintained low error even as the number of modes ($m$) and the barrier height ($\delta$) increased, whereas baselines like ASBS failed significantly.
Figure 3: Results on Many-Well potential. We evaluate robustness to increasing modes and robustness to mode distance. (a, b) Performance vs. number of modes. (c, d) Performance vs. mode distance. Our proposed method (TR-SBS, red) achieves the lowest estimation error and marginal distance for large m and δ, handling both combinatorial complexity and high energy barriers.
  • Capturing Rare Events: In the Alanine Dipeptide experiment, TR-SBS was the only method to consistently and stably capture the rare C7ax state. Standard ASBS either missed the state entirely or diverged numerically.
  • High-Dimensional Scaling: On GMMs, TR-SBS achieved the lowest Sinkhorn divergence and Total Variation Distance across all tested dimensions, proving it can handle the "curse of dimensionality" better than existing diffusion-based samplers like PIS or DIS.

Limitations and Future Work

  • Computational Overhead: Because TR-SBS takes more conservative steps, it can require significantly more iterations (and thus more target gradient evaluations) than standard IPF variants.
  • Hyperparameter Sensitivity: While more stable, the method still depends on the choice of the radius $\epsilon$ and the diffusion scale $\sigma_{max}$.
  • Idealized Assumptions: The theoretical convergence guarantees assume exact updates, which are never perfectly achieved in practice due to neural network approximation and finite sampling.

The authors suggest that future work could involve automatically adapting the trust region radius during training to further optimize the balance between speed and stability.

Summary updated

Abstract

We study the numerical approximation of Schrödinger Bridges in the setting where samples from the target distribution are unavailable and only an unnormalized target density is given. In this regime, the widely used iterative proportional fitting (IPF) procedure (and its variants) often exhibits severe numerical instabilities, stemming from ill-conditioned underlying control problems and from the fact that the optimal solution may lie far from the initialization. To address these issues, we propose a trust region variant of IPF that constrains successive updates to remain at a fixed Kullback-Leibler divergence from the current iterate. This restriction significantly improves numerical stability. Moreover, we show that the resulting method admits an interpretation as a mirror descent scheme with an adaptively chosen step size. We demonstrate the effectiveness of our approach on several challenging sampling tasks.
BibTeX
@misc{koshizuka2026trust,
  author = {Takeshi Koshizuka and Denis Blessing and Max Zimmer and Sebastian Pokutta and Lorenz Richter},
  title = {{A Trust Region Approach for Learning Schrödinger Bridges}},
  year = {2026},
  eprint = {2611.11649},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.11649},
}
Figure