acceptodds
Withdrawn from ICLR 2027

Distributionally Robust Schrödinger Bridge

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Schrödinger bridge (SB) learns stochastic transport between prescribed initial and target distributions. When the initial distribution shifts at test time, the learned dynamics can fail to recover the target distribution. We introduce the Distributionally Robust Schrödinger Bridge (DRSB), which learns a single controller that accounts for uncertainty in the initial distribution. The DRSB objective consists of control energy and a KL penalty between the resulting terminal distribution and the target distribution. DRSB seeks a single controller that minimizes the worst-case value of this objective as the initial distribution varies within an ambiguity set around the nominal distribution. We derive an exact variational formulation of this objective and connect its fixed-terminal-cost subproblem to stochastic optimal control and distributionally robust optimization. This formulation motivates an alternating algorithm that updates the adversarial initial distribution, estimates the terminal log-density ratio, and trains the controller. We develop Wasserstein and Sinkhorn variants using stochastic control optimality conditions to approximate the gradients required for adversarial updates. Experiments on two-dimensional transport tasks and image-to-image translation show improved robustness to input perturbations relative to standard SB, with a tradeoff in nominal performance. On Gaussian mixture transport, Sinkhorn DRSB also achieves lower mean sliced Wasserstein distance than fixed-level noise augmentation at both tested unseen noise levels.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Schrödinger bridge, Distributionally robust optimization, Stochastic optimal control

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_NUbMOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Distributional robust Schrödinger bridge (DRSB), which trains a controller U_t that minimizes the worst-case sum of energy and a KL term for terminal states, over an ambiguity set of initial distributions around the empirical distribution of training examples. The terminal KL is represented in its Donsker-Varadhan variational form with the maximizing function g estimated by a binary classifier. For fixed g, the objective is linear in the initialization distribution, so the standard DRO duality applies to the inner maximization. Training is done in alternative manner. Experiments on 2D Gaussian and GMM transport and on FFHQ latent image translation show the effectiveness of DRSB.

Critical strengths
  • This paper addresses the robustness of a learned transport map to test-time shifts in the initial distribution, without retraining, is important and not well-studied for SB models.
  • the variational form of the KL makes the subproblem linear in the initial distribution. This is the key insight: without it, the KL is a nonlinear function of the initial distribution and standandard DRO duality cannot be applied.
  • The experiments could show the effectiveness of DRSB.
Critical weaknesses
  • Theoretical contributions are limited. Theorem 3.4 (main contribution of this paper?) is a direct substitution of the Donsker-Varadhan formula. It should be better presented as a proposition. The convergence guarantees are not provided. The derived optimization problem is a min-max-sup problem, which is complex and not guaranteed to converge.
  • There is no sensitivity analysis over eps and \rho?
  • Eq 12 should use argmin/argmax
  • Problem 3.1 minimizers over u while the integrand uses u^\theta
  • proposition (3.3) should read Proposition 3.3
Questions for authors

See weaknesses.

Reviewer_LvwuOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The consider the Schröndiger bridge problem when there is uncertainty in the initial distribution. The proposed method replace the hard constraint at terminal time by a Kullback-Leibler penalty and takes a DRO perspective by minizing the worst case of the control energy over an ambiguity set of initial distributions. The ambiguity set is defined using either Wasserstein distance or a Sinkhorn distance.
Convex duality for the Kullback-Leibler divergence yields a min-max-sup form which can be solved iteratively by using DRO techniques, a likelihood ratio estimate, and an adjoint matching update for the control variable. Experiments are two dimensional toy model and image translation.

Critical strengths
  1. The problem is very clearly posed and the dual problem looks very neat and natural.

  2. The Sinkhorn problem is practical because of the explicit form of the dual variables.

  3. The model perform well on Gaussian mixture models.

Critical weaknesses
  1. The theoretical aspects of the papers are very thin and amount to a standard application of Donsker-Varadhan and the duality for Wassertein DRO. This is pretty straightforward.

  2. There is no analysis, or even a discussion of the iterative procedure for the minimax problem. I am not an expert in such analysis but the complete lack of discussion is troubling. Without this the paper cannot be accepted in my view.

  3. The experiments look a little shaky. For example it looks like table 1 selects the best performing case of table 6 and 7, so the statement of the papers about improvement across all three pertubation types look shaky. The image experiments also look selected to show the best case.

Questions for authors
  1. Which delta and which shift direction do the "Noisy" and "Shifted" rows of Table 1 report, and why these given Tables 6a and 7a?

  2. The mumu update holds g fixed, so the adversary maximizes a Donsker-Varadhan lower bound. Can you show, even empirically, that this increases the true objective in Equation 5a? More generally can you provide some analysis of the iteration scheme?

Reviewer_RsS1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the Schrödinger Bridge problem under uncertainty in the initial distribution. The authors propose learning a single controller that minimizes the worst-case sum of control energy and a KL penalty between the terminal and target distributions. Uncertainty in the initial distribution is modeled using Wasserstein and Sinkhorn ambiguity sets, with corresponding training algorithms derived from distributionally robust optimization and stochastic optimal control. The method is evaluated on synthetic tasks and image-to-image translation.

Critical strengths
  • The paper addresses a well-motivated problem of constructing a Schrödinger Bridge that is robust to changes in the initial distribution.
  • The authors derive a variational representation of the proposed objective and present training procedures for both Wasserstein- and Sinkhorn-based formulations. Their experimental evaluation includes low-dimensional transport problems and image-to-image translation tasks.
Critical weaknesses

(W1) The ambiguity radius ϵ\epsilon is an important parameter that determines which distribution shifts are considered during optimization. Although Appendix F.2 provides a sensitivity analysis, the paper does not explain how ϵ\epsilon should be selected from data or related to plausible test-time shifts. In particular, it is unclear whether the evaluated perturbations fall within the prescribed ambiguity sets. This makes the practical scope of the robustness claims difficult to assess.

(W2) Proposition 3.3 assumes that the terminal and target distributions have positive densities and finite KL divergence. In the image-to-image translation experiment, the method operates in a 512-dimensional latent space, but it is unclear how these assumptions are justified for the target distribution. In particular, the empirical distribution of finitely many latent representations is discrete and does not have a Lebesgue density. The paper should clarify whether ν\nu denotes an underlying population distribution or its empirical approximation, and how the theoretical assumptions relate to the implementation.

(W3) Although the paper includes experiments in both 2D and 512-dimensional latent spaces, it does not systematically investigate how performance and training stability change with dimensionality. The appendix discusses stability issues in high-dimensional settings, but a controlled comparison across several dimensions would provide a clearer picture of the scalability of the proposed approach.

(W4) A more detailed analysis of the statistical reliability of the improvements would help clarify how consistently the proposed method benefits from distributionally robust training.

(W5) To the best of my understanding, the paper does not provide a link to the source code. Since the proposed method involves several optimization steps and training stabilization techniques, the lack of a publicly available implementation makes it more difficult to evaluate the method independently.

Questions for authors

(Q1) How should the ambiguity radius ϵ\epsilon be selected in practice? Could the authors report the distances between the nominal and evaluated test distributions and compare them with the chosen ambiguity radii?

(Q2) In Proposition 3.3, does ν\nu represent the underlying target distribution in latent space or its empirical approximation? What justifies the density and finite-KL assumptions in the image-to-image translation experiment?

(Q3) Could the authors evaluate performance and training stability across several data dimensions while keeping the experimental setup otherwise comparable?

(Q4) Could the authors provide additional evidence for the reliability of the reported robustness improvements, for example through confidence intervals or statistical significance analysis?

(Q5) Could the authors provide the code?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.