Bridge Matching Sampler (BMS)

• Bridge Matching Sampler is a diffusion sampling framework that transports arbitrary priors to unnormalized target densities via iterative fixed-point matching in path space.
• It employs least-squares regression with damped updates to ensure stability and scalability in high-dimensional, multimodal scenarios like molecular dynamics and systems biology.
• Extensions of BMS include multi-marginal, underdamped, and interacting many-body systems, unifying diverse Schrödinger Bridge and score-matching techniques.

The Bridge Matching Sampler (BMS) is a family of scalable algorithms for learning diffusion processes that transport an arbitrary prior distribution to a target, typically specified via an unnormalized density, by iterative fixed-point matching in path space. BMS generalizes, unifies, and extends several classes of stochastic optimal control, diffusion sampling, and Schrödinger Bridge solvers. It achieves high scalability and stability through a tractable, single-objective optimization framework, enabling applications to high-dimensional generative modeling, unnormalized log-density sampling, molecular conformation, and trajectory inference in systems biology and physics.

1. Problem Setting and Theoretical Motivation

BMS addresses the problem of sampling from an unnormalized target density $\rho_{\rm target}(x)$ by learning a Markovian diffusion process: $$dX_t = \sigma(t)\,u(X_t, t)\,dt + \sigma(t)\,dB_t, \qquad X_0 \sim p_{\rm prior}$$ such that the terminal marginal $X_T \sim p_{\rm target}$ approximates the desired target distribution. The approach is grounded in the Schrödinger Bridge framework, which formulates the problem as minimizing kinetic energy (equivalently, KL divergence in path space) under endpoint constraints. BMS generalizes least-squares matching and adjoint Schrödinger Bridge methods within a fixed-point iteration over Markovian diffusions, underpinned by Nelson's relationship between forward/backward drifts and time-marginals of the path law (Blessing et al., 28 Feb 2026, Liu et al., 27 Jun 2025).

In the BMS formalism, sampling is recast as finding a fixed point of a "diffusion matching" operator. This operator alternates between constructing a reciprocal bridge process (matching prior and target endpoints) and projecting that (generally non-Markovian) bridge onto the set of Markov diffusions via least-squares fitting.

2. Algorithmic Structure: Fixed-Point and Least-Squares Matching

BMS defines an iterative fixed-point update: $u_{i+1} = \Phi(u_i)$ where the operator $\Phi$ is: $$\Phi(u)(x, t) = \mathbb{E}_{\Pi^u_{0,T}\,P_{|0,T}}\bigl[\xi(X, t) \mid X_t = x\bigr]$$ At each iteration, BMS simulates endpoint pairs $(X_0, X_T) \sim \Pi^i_{0,T}$ and, for each, draws intermediate states $X_t$ by evolving the conditional bridge law $P_{t|0,T}(X_t|X_0, X_T)$. The drift is fit by minimizing: $$u_{i+1} = \arg\min_{u \in \mathcal U}\, \mathbb{E}_{\Pi^i}\Biggl[\int_0^T \frac{1}{2} \|\xi(X, t) - u(X_t, t)\|^2\,dt\Biggr]$$ The explicit bridge drift $\xi(X, t)$ combines derivatives of $\log p_{\rm prior}$, $\log \rho_{\rm target}$, and the conditional path measure, with precise weights determined by the diffusion schedule (Blessing et al., 28 Feb 2026).

Crucially, BMS uses an independent coupling $\Pi^i_{0,T} = p_{\rm prior} \otimes p^{u_i}_T$, thereby removing the need for alternating optimization or explicit coupling estimation as in earlier Schrödinger Bridge samplers (Liu et al., 27 Jun 2025, Park et al., 18 Oct 2025).

3. Damped Updates and Regularization

Pure fixed-point updates can be unstable, especially in high-dimensional or multimodal settings. BMS introduces a damped iteration: $$u_{i+1} = \alpha \Phi(u_i) + (1 - \alpha) u_i, \qquad \alpha \in (0, 1]$$ Equivalently, this is a regularized regression: $$u_{i+1} = \arg\min_{u}\, \mathbb{E}_{\Pi^i}\Biggl[\int_0^T \frac{1}{2} \|\xi - u\|^2\,dt\Biggr] + \eta\,\mathbb{E}_{\Pi^i}\Biggl[\int_0^T \frac{1}{2} \|u_i - u\|^2\,dt\Biggr]$$ with $\eta = (1-\alpha)/\alpha$. The regularization forces stability and mitigates mode collapse, as empirically confirmed in GMM and molecular systems (Blessing et al., 28 Feb 2026).

4. Extensions: Multi-Marginal, Underdamped, and Entangled Particle Systems

Multi-Marginal BMS

For systems with multiple observed marginals at discrete timepoints (e.g., in single-cell genomics or climate data), the Multi-Marginal Schrödinger Bridge Matching (MSBM) algorithm generalizes BMS by splitting the temporal domain into subintervals. Each interval is paired with local SB fitting, and drifts are stitched globally. This construction preserves all intermediate marginals and parallelizes over time intervals, enabling tractable inference in high-dimensional, temporally structured data (Park et al., 18 Oct 2025, Theodoropoulos et al., 11 Jun 2025).

Underdamped Diffusion BMS

For dynamical systems with momentum (e.g., molecular dynamics), underdamped BMS employs second-order SDEs, operating on position-velocity phase space. The learning objective reduces to matching control and bridge drifts in phase space, and can be discretized via splitting schemes (e.g., OBABO). This extension provides tighter likelihood lower bounds and higher mixing efficiency for multimodal or stiff target distributions (Blessing et al., 2 Mar 2025).

EntangledSBM for Interacting Many-Body Systems

Entangled Schrödinger Bridge Matching (EntangledSBM) extends BMS to jointly-conditioned, interacting systems. The drift for each particle depends dynamically on all positions and velocities, and is parameterized by permutation-invariant architectures such as transformers. The cross-entropy loss is regularized globally over the entire trajectory, providing a solution to simulation and inference of rare transitions in biomolecular and cell-population systems that require non-factorized trajectory laws (Tang et al., 10 Nov 2025).

5. Practical Implementation and Empirical Results

BMS implementations use neural networks (ResNets or equivariant GNNs) to model the drift, with time embeddings for temporal conditioning. Training is performed by stochastic gradient descent over minibatches of bridge samples. Empirical evaluation demonstrates BMS's scalability—GMMs up to $d=2500$, molecular systems with $d\sim 126$, and efficient parallelization due to memory use $O(Kd)$. In all settings—high-dimensional GMMs, $n$-body systems, peptides, and trajectory inference—the BMS framework achieves or surpasses the state-of-the-art in stability, mode preservation, and sample fidelity (Blessing et al., 28 Feb 2026, Park et al., 18 Oct 2025, Tang et al., 10 Nov 2025).

Key empirical findings include:

• Stability and preservation of modes in multimodal targets, outperforming methods such as ASBS, DDS, and PIS in high dimensions.
• Lower error metrics (mode-TVD, Wasserstein-2, Ramachandran JS, free-energy RMSE) and reduced variance across training seeds.
• Tractable computational cost and memory footprint, enabling largescale molecular and high-dimensional synthetic experiments.

6. Relation to Other Bridge and Flow Matching Methods

BMS comprises and extends several methodological threads:

• It generalizes adjoint-based samplers (AS, ASBS) and the IPF path-space alternating minimization to arbitrary prior distributions and efficient one-step fitting (Shin et al., 17 Feb 2026, Liu et al., 27 Jun 2025).
• In the degenerate case (memoryless base SDE), BMS reduces to conditional score-matching in score-based diffusion models, explaining the increased instability and highly curved sampling trajectories in conventional diffusion generation (Shin et al., 17 Feb 2026).
• The Coupled Bridge Matching (BM$^2$) algorithm (Peluchetti, 2024) provides a non-iterative variant of BMS by cross-matching forward and backward drifts, showing competitive efficiency and convergence properties, while Augmented Bridge Matching (AugBM) preserves full endpoint coupling at the cost of non-Markovianity (Bortoli et al., 2023).

A summary of the relationships is provided in the table below:

Method	Coupling	Markovian Projection	Optimization
BMS	Independent	Yes	Least-squares
ASBS, AS	Arbitrary	Yes	Alternating AM/CM
BM$^2$	Coupled/learned	Yes	Joint cross-match
AugBM	Empirical pairings	No	Conditional drift

7. Convergence, Scalability, and Open Problems

BMS converges to a diffusion whose marginals match the reciprocal process at fixed points. For the damped variant, empirical experiments indicate high stability and robustness across large tasks. However, formal global convergence proofs—such as contractivity of the fixed-point or global monotonicity in path-space KL—remain open questions (Blessing et al., 28 Feb 2026).

In multi-marginal settings, monotonic decrease of KL to the global SB solution is established under Markovian/reciprocal projections (Park et al., 18 Oct 2025, Theodoropoulos et al., 11 Jun 2025). Empirically, the combination of regularized updates and efficient simulation enables scaling to thousands of dimensions with improved sample quality and reduced computational burden.

This article references specific developments in (Blessing et al., 28 Feb 2026, Liu et al., 27 Jun 2025, Park et al., 18 Oct 2025, Blessing et al., 2 Mar 2025, Tang et al., 10 Nov 2025, Shin et al., 17 Feb 2026, Peluchetti, 2024), and (Bortoli et al., 2023).