acceptodds
Under review as a conference paper at ICLR 2027

Solving the Schrödinger Bridge via Bayesian Inference in Parameter Space

Abstract

Solving transport problems, i.e., finding a map that transports one given distribution to another, has numerous applications in machine learning. Currently, the Schrödinger Bridge (SB) is commonly used to describe the mapping between two arbitrary distributions. However, most existing SB solutions either suffer from slow convergence and sampling due to the difficulty of directly optimizing the likelihood or are highly sensitive to initialization and hyperparameters. To address these, we propose a novel Bayesian Schrödinger Bridge (BSB) formalism to address the SB problem by explicit likelihood factorization leveraging Bayesian inference. By explicitly introducing Bayesian inference in the continuous parameter space of a Gaussian mixture model, BSB allows tractable and explicit likelihood-based optimization of the likelihood instead of the variational lower bound, while enabling more accurate modeling of complex probability density transitions. BSB analytically derives closed-form Bayesian inference at each time step, enabling the design of SDE and higher-order ODE solvers for few-step high-quality generation. Extensive experiments validate our effectiveness, achieving faster convergence, robust training, and high-quality few-step SB solution.

Primary area:
generative models
Keywords:
Schrödinger Bridge, Bayesian inference

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_aQtHOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes a Bayesian formulation of the Schrödinger Bridge (BSB). BSB introduces a composite projection (Bayesian and posterior projections), which replaces the Markovian projection in Iterative Markovian Fitting. The method uses a Gaussian mixture model for the transport kernel, for which it provides a closed-form update. BSB contributes a method that directly optimizes the KL divergence to the reference bridge without requiring a variational lower bound, and it empirically provides high-quality few-step Schrödinger Bridge solutions. BSB is validated with component ablations and comparisons against baselines on FID, precision and LPIPS, on both image and synthetic datasets. The underlying algorithm is proven to converge.

Critical strengths

The experimental results are promising; SBS compares favorable against baseline and the ablation study show its components all contribute positively.

Critical weaknesses

The paper is hard to follow, mainly because it is not clear which variable the distribution p(y∣x)p(y|x) is defined over. Amortizing p(y∣x)p(y|x) by predicting its parameters with a network is the key to Theorem 1, so the explanation of p(y∣x)p(y|x) on L152 is not sufficient. The claim that SBS analytically performs closed-form Bayesian inference at each time step is too strong. The closed form is only available because of the amortization network and the Gaussian assumption on the data distribution. The theorem then reduces to a case of conjugacy, because the hard parts of inference are hidden in the network and the data assumption. Beyond this main objection, a further weakness of the paper is its notation (see the questions).

Questions for authors
  1. In what sense does SBS converge faster? And what is SBS claimed to converge faster than? What evidence do you provide in the paper of faster convergence?
  2. L168: “SB problem suffers from challenges in likelihood optimization and numerical stability”. Which challenges exactly are you referring to here, and how does SBS address them?
  3. What exactly is the random variable yy? q(y∣x)q(y|x) is called a posterior, but it is not clear what it is a posterior over.
  4. In Eqs. 3, 5 and 7, the =def\stackrel{\text{def}}{=} symbol seems to have two different meanings. Take Eq. 7: the first =def\stackrel{\text{def}}{=} is a definition (the projection is a kernel), and the second is a constraint (time homogeneity). Is this the right reading?
  5. In Eq. 17, is the second equality supposed to be there? If so, please provide the derivation.

Minor

NN in Eq. 3 is not defined. The RHS of Eq. 8 does not depend on q3k+2q^{3k+2}.

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

This paper introduces Bayesian Schrödinger Bridge (BSB), a simulation-free approach to solving the Schrödinger Bridge problem. Based on the Iterative Markovian Fitting (IMF) framework, BSB divides the Markovian projection into two stages: posterior projection and Bayesian inference projection. Instead of directly modeling the full transition kernel, the method trains a neural network to estimate only the posterior distribution of the endpoint given the current state. The transition distribution is then obtained through closed-form Bayesian inference with Gaussian Mixture Models (GMMs). Building on this formulation, the authors also propose SDE and ODE solvers to support efficient generation with few steps. Experiments show strong results across various image translation tasks and promising transport accuracy on entropic optimal transport benchmarks with known solutions.

Critical strengths
  • The proposed method is simulation-free during training, allowing efficient learning of Schrödinger Bridges without trajectory simulation.

  • The authors first compute the transition kernel conditioned on the endpoint, and then marginalize over the possible endpoints using the endpoint posterior to obtain the actual transition kernel. This structure allows the method to learn only the posterior distribution of the endpoint given the current state, enabling efficient training.

  • The proposed method outperforms all the baselines considered in the paper across multiple image translation tasks. Moreover, it achieves strong generation quality with fewer than 20 neural function evaluations (NFEs), demonstrating its efficiency in few-step generation.

Critical weaknesses
  • Major Concern: In each Markovian projection, the endpoint posterior should depend on the current coupling between the source and target distributions. However, according to Algorithm 1, the authors appear to train the endpoint posterior using independently sampled x0x_0 and x1x_1 from the two datasets. Does this mean that the proposed algorithm effectively constructs an independent coupling π(x0,x1)=p0(x0)p1(x1)\pi(x_0,x_1)=p_0(x_0)p_1(x_1) and connects the sampled endpoints using Brownian bridges? If so, this seems inconsistent with the objective of solving the Schrödinger Bridge problem, since the endpoint coupling is not updated toward the optimal coupling.

  • Minor Concern: The authors use GMMs to parameterize the distributions at each time step. However, the expressive capacity of GMMs may become limited when dealing with large-scale or highly complex datasets.

Questions for authors
  • Could the authors clarify the first concern raised in the Weaknesses section? Since the endpoint posterior appears to be trained using an independent coupling, how is the endpoint coupling updated throughout the algorithm? Does the method use the same independent coupling throughout training? This is my primary concern and the main factor affecting my evaluation. A satisfactory clarification or resolution of this issue would lead me to increase my score.

  • How well does the GMM parameterization scale to large and complex datasets? Could its limited expressive capacity become a bottleneck as the data dimensionality or complexity increases?

  • Simulation-free methods for solving Schrödinger Bridges already exist, such as [SF]²M [1]. Could the authors provide a more detailed comparison between the proposed method and existing simulation-free Schrödinger Bridge solvers?

[1] Simulation-Free Schrödinger Bridges via Score and Flow Matching.

Reviewer_s4AwOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes viewing the computation of the Schrödinger bridge through the lens of Bayesian inference.

Critical strengths

I find the results of the numerical experiments very interesting and promising. However, they are difficult to appreciate without a full understanding of the theory underpinning the method. Unfortunately, I find the presentation of the theory rather unclear, as I discuss in detail below. Should the presentation be improved, I would be happy to revise my assessment of the paper.

Critical weaknesses

Here, I list my concerns about the theory presented in the paper. My main questions relate to Sections 2 and 3.

  1. In my opinion, Figure 1 is quite difficult to understand. I do not think that panels a) and b) clearly convey the intuition behind IPF and IMF, respectively. Moreover, I struggle to understand the meaning of panel c). Specifically, what is the posterior predictive, and what is the posterior manifold? The method is only introduced on page 4, yet Figure 1 already contains many method-specific terms that have not been defined.

  2. I find the naming of the distributions ppriorp_\text{prior} and pdatap_\text{data} rather inconvenient. Firstly, in Schrödinger bridge problems these distributions are often of comparable difficulty, whereas in sampling and Bayesian inference the term prior refers to an initial, a priori belief that is refined through observations. Secondly, this naming conflicts with the standard prior/posterior terminology of Bayesian inference: the proposed algorithm does not seem to "update" ppriorp_\text{prior} in any sense. I propose using p0p_0 and p1p_1 instead.

  3. What is the reference transition p(xt∣x0,x1)p(x_t \mid x_0, x_1) in line 86?

  4. I am at a complete loss as to how to understand the paragraph Bayesian Flow Networks: yy, ρ\rho and θ\theta are not properly defined, the Bayesian inference problem is not fully specified, it is hard to tell which distribution is the prior and which is the posterior, and it is unclear what problem is being solved.

  5. Equation 11 appears to contain typos, and I am not sure how to interpret it.

  6. Although yy appears in Equations 6, 7, 8 and 10 and in Theorem 1, it is never formally defined.

  7. Theorems 1 and 2 are given without proofs. There is no appendix, although the paper refers to one several times (lines 222, 255 and 269).

  8. Algorithm 1 seems incomplete. How does the loss depend on θ\theta? Are fFf^{\rm F} and fBf^{\rm B} updated simultaneously?

  9. The distribution qq is redefined several times. Initially, it denotes a transition probability; however, line 237 defines it as a distribution over xx that depends on θ\theta (again, what is xx: x0x_0, xtx_t or x1x_1? And what is θ\theta?). Later, qq is redefined once more as q(y∣θ)q(y \mid \theta).

Questions for authors

See above

Reviewer_yS9FOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

Schrödinger Bridges provide a principled framework for generative modeling, their application to high-dimensional problems remains computationally challenging, particularly due to the iterative optimization required by many existing methods.The paper introduces Bayesian Schrodinger Bridge, a method for solving SB problems by modeling the transport process in the parameter space of probability distributions. The proposed method introduces Iterative Proportional Bayesian Fitting (IPBF). It breaks down the Markovian projection into posterior projection and Bayesian inference projection. The posterior projection uses a NN to predict the parameters of a conditional distribution over the transport endpoint. The Bayesian inference projection analytically updates the current distribution using the predicted posterior. The method uses Gaussian Mixture Models for the intermediate distributions; this allows for closed-form updates. An interesting feature of the formulation is that it allows for the construction of SDE and higher-order ODE solvers for few-step sampling

On the theory side, the authors establish the equivalence between the ideal IPBF and IMF, obtaining convergence to the SB under exact posterior projections ad regularity assumptions.

The numerical experiments consider 2d problems, image-to-image translation, and entropic OT benchmarks. The reported results show competitive performance in transport accuracy and generation quality.

Critical strengths
  1. The paper proposes a novel approach to solve SB problems by combining a neural posterior prediction with analytical Bayesian updates of GMM. The closed-form updates provide an alternative to directly learning the transition distribution and allow the construction of efficient sampling schemes.
  2. The authors show that IPBF is equivalent to IMF under exact posterior projections.
  3. The experimental section is very comprehensive. The authors compare IPBF against SOTA methods in synthetic, image-to-image translation, and entropic OT benchmarks. The ablation study on the number GMM shows a clear advantage of increasing the number of modes in the mixture.
  4. In the flow-based generative model community, the NFE is of high interest. The method can generate competitive quality with fewer sampling steps.
Critical weaknesses
  1. Although the theory establishes the connection of IPBF and IMF, it does not account for the neural approximation of the posterior projection or the GMM parametrization used in practice. Can the authors comment further on how these approximations might affect the convergence guarantees?
  2. Although the paper reports improvements in sampling efficiency, the current experiments do not report the computational cost of training. Can the authors provide a comparison of wall-clock training times for image-to-image translation experiments?
Questions for authors
  1. Do the authors have any theoretical insights on the behavior of BSB as the noise of the reference Wiener process approaches zero? Does the method recover quadratic OT?
  2. Have the authors considered extending the proposed framework to generalized SB problems [1]? In this context, the cost of transportation includes a potential term that models obstacles, interaction of the samples, and density-dependent effects. It would be interesting to understand whether the proposed Bayesian inference formulation could accommodate these more general transport costs.

[1] Liu et al., Generalized Schrödinger Bridge Matching ICLR 2024.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.