Diffusion Schrödinger Bridge Matching

• DSBM is a framework for computing stochastic processes that transport one probability distribution to another while minimizing kinetic-action costs and enforcing endpoint constraints.
• It employs variational approximations, iterative drift matching, and path update techniques to achieve scalable and accurate solutions in high-dimensional settings.
• DSBM unifies score-based diffusion models, flow matching, and entropic optimal transport, yielding robust convergence with practical applications in image translation and molecular optimization.

Diffusion Schrödinger Bridge Matching (DSBM) defines a family of scalable algorithms for computing stochastic processes that transport one probability distribution to another while minimizing a dynamical entropy-regularized cost. Rooted in stochastic optimal control, entropic optimal transport, and variational inference, DSBM is a central methodology for high-dimensional generative modeling, scientific data assimilation, and distribution-matching tasks where kinetic costs or additional task-specific regularization must be imposed. DSBM generalizes both score-based diffusion models and flow matching approaches by enforcing endpoint marginal constraints characteristic of the Schrödinger Bridge, while leveraging regression-style variational objectives and scalable neural parameterizations.

1. Mathematical Formulation of the Diffusion Schrödinger Bridge

Given two boundary distributions $\mu_0, \mu_1$ on $\mathbb{R}^d$, the classical diffusion Schrödinger Bridge (SB) problem seeks a stochastic process $(X_t)_{t\in[0,1]}$ (typically an Itô SDE)

$$dX_t = u_t(X_t) \, dt + \sigma\, dW_t, \quad X_0 \sim \mu_0,\,X_1\sim\mu_1$$

whose marginal laws $(p_t)_{t\in[0,1]}$ satisfy the Fokker–Planck equation

$$\partial_t p_t(x) = -\nabla\cdot\left[u_t(x) p_t(x)\right] + \frac{\sigma^2}{2}\Delta p_t(x), \quad p_0 = \mu_0,\, p_1 = \mu_1,$$

and among all such $(u_t)$, minimizes the kinetic-action

$$J_{\mathrm{SB}}[u] = \int_0^1 \mathbb{E}_{p_t} \left[ \frac{1}{2}\|u_t\|^2 \right] dt.$$

This is a dynamic, path-space entropy-regularized optimal transport (EOT) problem and can also be characterized as a stochastic control problem or variational problem. The SB solution interpolates between deterministic OT and classic diffusion (score-based, memoryless) models as $\sigma\to 0$ or as the endpoint constraint degenerates (Shi et al., 2023).

Generalized SB Matching (GSBM) introduces a nonnegative state cost $V_t(x)$, yielding

$\mathbb{R}^d$0

subject to the same Fokker–Planck and endpoint constraints. The optimal drift $\mathbb{R}^d$1 is related to the logarithmic gradient of a value potential $\mathbb{R}^d$2, solution to a backward Kolmogorov (Hamilton–Jacobi–Bellman) equation: $\mathbb{R}^d$3 This expresses the SB as a linearly-solvable conditional SOC (Liu et al., 2023).

2. Variational Algorithms and Iterative Fitting

Direct computation of the SB law or optimal drift is tractable only in limited cases. DSBM algorithms circumvent this by exploiting variational approximations and alternating projections—known as Iterative Markovian Fitting (IMF) or IPF (Iterative Proportional Fitting). The essential elements are:

• Path law and marginal update: Given a drift $\mathbb{R}^d$4, sample $\mathbb{R}^d$5 pairs and solve conditional optimal control $\mathbb{R}^d$6 (SOC step). Update the pathwise law to $\mathbb{R}^d$7, then average to a new marginal $\mathbb{R}^d$8.
• Drift matching: For fixed $\mathbb{R}^d$9 (possibly represented as mixtures over bridges), update the neural drift $(X_t)_{t\in[0,1]}$0 to minimize squared deviations from the optimal conditional drift $(X_t)_{t\in[0,1]}$1 under $(X_t)_{t\in[0,1]}$2:

$(X_t)_{t\in[0,1]}$3

• Alternation: Iterate between path update (SOC) and drift matching, ensuring at each step that the endpoint marginals remain feasible—that is, $(X_t)_{t\in[0,1]}$4 (Liu et al., 2023, Shi et al., 2023).

For computational stability and scalability:

• Spline or Brownian bridge parameterizations are used for conditional processes.
• Simulation-free updates and efficient importance-weighted (path integral) corrections can debias Gaussian or spline path approximations.
• At each iteration, the variational objective $(X_t)_{t\in[0,1]}$5 is monotonically decreased (Liu et al., 2023).

3. Connections to Related Methods and Unified Perspectives

DSBM generalizes and connects the following methodologies:

• Score-based diffusion models/SGMs: The first IPF step in SB reduces to score-matching on the terminal distribution; as the number of DSBM iterations increases, the process more strictly interpolates between prescribed endpoints (Bortoli et al., 2021, Shi et al., 2023).
• Flow Matching Models (FMMs): In the deterministic (zero-diffusion) limit with linear interpolation, DSBM recovers conditional flow matching. More generally, it yields entropic regularized flows or ODE bridges (Kim, 27 Mar 2025).
• DSBM as IMF: Alternating Markovian and reciprocal projections, DSBM finds the SB as the unique Markov-reciprocal law with given endpoints (Shi et al., 2023).
• Optimal Bridge Matching: For known couplings, recent theory shows that the KL projection of any reciprocal process to the SB manifold recovers the bridge in one step, connecting DSBM losses to energy-based modeling objectives for entropic OT (Gushchin et al., 2024).
• Adjoint Schrödinger Bridge Matching (ASBM): ASBM further accelerates convergence in high-dimensions by first constructing the optimal coupling $(X_t)_{t\in[0,1]}$6 (via a data-to-energy SOC) and then learning the adjoint (backward) drift via a single regression, achieving low-variance, efficient trajectories (Shin et al., 17 Feb 2026).

These developments suggest a unified variational and control-theoretic framework bridging score-based SDEs, entropic OT, and stochastic control, with DSBM/GSBM algorithms at the center.

4. Algorithmic Flow and Practical Realizations

A schematic DSBM iteration proceeds as follows (Liu et al., 2023, Shi et al., 2023):

1. Initialization: Parameterize $(X_t)_{t\in[0,1]}$7 (e.g., as a neural network), initialize endpoint coupling $(X_t)_{t\in[0,1]}$8.
2. Conditional SOC (Path Update): For each sampled endpoint pair $(X_t)_{t\in[0,1]}$9, solve or approximate the conditional optimal control to obtain $$dX_t = u_t(X_t) \, dt + \sigma\, dW_t, \quad X_0 \sim \mu_0,\,X_1\sim\mu_1$$0.
3. Importance Resampling (optional): Resort to path-integral correction to debias drift based on state cost $$dX_t = u_t(X_t) \, dt + \sigma\, dW_t, \quad X_0 \sim \mu_0,\,X_1\sim\mu_1$$1.
4. Drift Matching: Fit $$dX_t = u_t(X_t) \, dt + \sigma\, dW_t, \quad X_0 \sim \mu_0,\,X_1\sim\mu_1$$2 by minimizing the conditional squared error over simulated bridge samples.
5. Averaging: Aggregate conditional marginals to form new $$dX_t = u_t(X_t) \, dt + \sigma\, dW_t, \quad X_0 \sim \mu_0,\,X_1\sim\mu_1$$3.
6. Alternate: Repeat until convergence to the stationary point of the general action $$dX_t = u_t(X_t) \, dt + \sigma\, dW_t, \quad X_0 \sim \mu_0,\,X_1\sim\mu_1$$4.

For the generalized formulation, action-matching (functional) losses, explicit minimization over conditionals, and SOC steps are flexible to encompass a broad range of costs beyond kinetic energy.

Pseudocode

$$dX_t = u_t(X_t) \, dt + \sigma\, dW_t, \quad X_0 \sim \mu_0,\,X_1\sim\mu_1$$8 (Liu et al., 2023)

5. Theoretical Properties and Convergence

DSBM/GSBM algorithms possess global convergence guarantees under the generic monotonicity of the alternating minimizations. Each alternation ensures $$dX_t = u_t(X_t) \, dt + \sigma\, dW_t, \quad X_0 \sim \mu_0,\,X_1\sim\mu_1$$5, and at the optimum the drift and path law are a fixed point of the scheme. Crucially, DSBM maintains exact feasibility at each iteration (endpoint marginals match) and works entirely by sampling from marginal distributions—a property enabling high-dimensional scalability.

Compared to prior iterative SB/sinkhorn or OT solvers, DSBM:

• Does not require full density evaluation of $$dX_t = u_t(X_t) \, dt + \sigma\, dW_t, \quad X_0 \sim \mu_0,\,X_1\sim\mu_1$$6,
• Avoids evaluating high-rank kernel couplings,
• Scales to high dimensions and complex state costs due to spline-conditioned Gaussian-bridge approximations,
• Is simulation-free in the conditional SOC step (no necessity for backpropagation through SDE solves) (Liu et al., 2023, Shi et al., 2023).

6. Empirical Applications and Performance

GSBM and DSBM frameworks have been validated on a wide array of scientific and machine learning problems, including:

• Crowd navigation: Optimal path planning with obstacles encoded as $$dX_t = u_t(X_t) \, dt + \sigma\, dW_t, \quad X_0 \sim \mu_0,\,X_1\sim\mu_1$$7; GSBM yields low kinetic cost and precise terminal fitting.
• Opinion depolarization: High-dimensional, nonlinear mean-field costs; DSBM accurately matches marginals and transports with interaction-aware paths.
• Manifold-structured data: LiDAR, image domain transfer, unpaired domain translation.
• Molecular optimization: Discrete GSBM extension with CTMCs enables minimal graph edit transformations for molecular property optimization (Kim et al., 2024).
• Image-to-image translation: Latent domain GSBM achieves competitive FID scores and faster convergence than standard IPF-based SB solvers.

GSBM’s convergence is robust—maintaining marginal constraints and feasible transports at each step. Empirically, it requires fewer alternations and supports both continuous and discrete domains (Liu et al., 2023, Kim et al., 2024).

7. Impact, Extensions, and Future Directions

DSBM and its generalized variants have established an effective, scalable, and theoretically principled approach for high-dimensional distribution matching under both classical (unregularized) and strictly regularized OT objectives. Extensions include:

• One-shot optimal SB matching, which shows that, with the right projection, standard bridge-matching can recover the SB in a single step (Gushchin et al., 2024).
• Semi-supervised and partially-aligned settings, where small sets of key-point alignments are leveraged to accelerate convergence and generalize transport maps (e.g., Feedback SB Matching) (Theodoropoulos et al., 2024).
• Discrete-state and graph-domain generalizations, leveraging continuous-time Markov chains for combinatorial structure (Kim et al., 2024).
• Hybridizations with pre-trained diffusion score models and neural ODE surrogates for accelerated sample efficiency and efficient inference (Khilchuk et al., 14 Dec 2025, Tang et al., 25 Aug 2025).
• Direct application to large-scale inverse problems, Bayesian inference, and scientific simulation, where task-specific costs and constraints must be respected throughout (Liu et al., 2023, Shi et al., 2023).

The DSBM methodology thus provides a unifying, extensible toolkit for tractable and robust learning of transport processes under general cost structures, bridging the gap between stochastic control, variational inference, and modern generative modeling (Liu et al., 2023).