Foundations of Schrödinger Bridges for Generative Modeling

Abstract: At the core of modern generative modeling frameworks, including diffusion models, score-based models, and flow matching, is the task of transforming a simple prior distribution into a complex target distribution through stochastic paths in probability space. Schrödinger bridges provide a unifying principle underlying these approaches, framing the problem as determining an optimal stochastic bridge between marginal distribution constraints with minimal-entropy deviations from a pre-defined reference process. This guide develops the mathematical foundations of the Schrödinger bridge problem, drawing on optimal transport, stochastic control, and path-space optimization, and focuses on its dynamic formulation with direct connections to modern generative modeling. We build a comprehensive toolkit for constructing Schrödinger bridges from first principles, and show how these constructions give rise to generalized and task-specific computational methods.

• The paper establishes a mathematically rigorous formulation using Schrödinger bridges to design generative models via minimal-entropy deviations of reference processes.
• It employs a dynamic programming approach and Hopf–Cole transform to linearize nonlinear stochastic control and Fokker–Planck systems for efficient computation.
• It unifies stochastic optimal control, entropic optimal transport, and diffusion-based methods to yield scalable algorithms with provable convergence.

Foundations of Schrödinger Bridges for Generative Modeling

---

Introduction and Scope

The manuscript "Foundations of Schrödinger Bridges for Generative Modeling" (2603.18992) presents a mathematically rigorous, comprehensive treatment of Schrödinger bridge (SB) theory, unifying disparate research in stochastic control, entropic optimal transport (EOT), and generative modeling. By framing generative transport problems as minimal-entropy deviations of path-space measures from reference stochastic processes, the paper provides foundational tools to analyze and design modern generative frameworks including diffusion models, score-based approaches, and flow-matching architectures.

The exposition is structured to guide the reader from classical optimal mass transport (OMT) theory, through entropy regularization and the static SB problem, to its dynamic path-space form and its role as a generalization of regularized dynamical transport. Methodologically, the work prioritizes explicit connections between analytic and algorithmic constructions, including duality, contractive projections, stochastic optimal control theory, and computational schemes such as Sinkhorn and forward-backward SDEs.

---

Static Schrödinger Bridge, Entropic OT, and Sinkhorn Iteration

The static Schrödinger Bridge problem is formulated as a convex entropy-regularized relaxation of the classical Kantorovich OMT, casting the search for an optimal coupling $\pi_{0,T}$ as a KL-projection of a reference coupling $q$ onto the set of measures matching prescribed marginals. This yields, for arbitrary reference $q$ (including those arising from exponential tilts of cost functions), the variational principle

$$\pi_{0,T}^* = \underset{\pi_{0,T} \in \Pi(\pi_0, \pi_T)}{\arg\min} \text{KL}(\pi_{0,T} \| q)$$

as detailed in Section 2. This convex program admits a unique minimizer, ensuring strict regularity and avoiding the degeneracies of deterministic mappings.

A critical insight is the dual characterization in terms of Schrödinger potentials $(\varphi, \hat\varphi)$, leading to a pair of coupled log-integral equations (the Schrödinger system). These dual variables enforce marginal constraints and are the objects optimized in the classical IPF/Sinkhorn iteration, which alternately updates $\varphi$ and $\hat\varphi$ to enforce normalization over each marginal: $$\varphi(x) \leftarrow -\log\int e^{\hat\varphi(y) - c(x,y)} \pi_T(dy), \quad \hat\varphi(y) \leftarrow -\log\int e^{\varphi(x) - c(x,y)} \pi_0(dx)$$ The manuscript provides a meticulous proof of convergence, grounded in the strict convexity and monotonicity of the KL objective, and clarifies that the resulting coupling is always smooth and offers a stochastic interpretation of transport that is robust to initialization and noise. The chain rule for KL, data-processing inequalities, and advantages of entropy regularization (uniqueness, smoothing, and scalability) are formalized.

---

Dynamic Schrödinger Bridge: Path-Space Lifting and Controlled SDEs

Extending to the dynamic SB, the paper shifts from couplings of endpoints to couplings of full path measures, characterizing the time-evolution of distributions between fixed marginals over a time interval $[0,T]$. This lift is both conceptually and technically central, as it underlies the construction of continuous-time generative models.

The dynamic SB problem is described as the minimization of path-space KL divergence $\text{KL}(\mathbb{P} \| \mathbb{Q})$ between a controlled process $\mathbb{P}$ and a reference SDE law $\mathbb{Q}$, subject to marginal constraints at $t=0$ and $t=T$. This is made concrete by the explicit computation of the Radon-Nikodym derivative via Girsanov's theorem, yielding an objective of the form

$$\text{KL}(\mathbb{P}^u \| \mathbb{Q}) = \mathbb{E}_{\mathbb{P}^u}\left[ \frac{1}{2}\int_0^T \|\boldsymbol{u}(X_t, t)\|^2 dt \right]$$

where $\boldsymbol{u}$ is the control drift superimposed upon the reference process.

The Fokker-Planck equation is derived as the forward time PDE for marginals, while backward Kolmogorov/Feynman-Kac PDEs propagate terminal data backward—together constituting the forward-backward structure fundamental to SB analysis. Crucially, the path-space SB lifts OMT from static mappings to probability flows compatible with stochastic process simulation, supporting dynamical generative modeling.

---

Nonlinear and Hopf–Cole Linearization

The optimality system for the SB is cast as a nonlinear system coupling the value function (Lagrange multiplier enforcing the FP constraints) with the controlled evolution of the marginals. The authors develop the dynamic programming principle, deriving Hamilton–Jacobi–Bellman (HJB) equations for the value functional, and establish that the optimal control field is characterized by the spatial gradient of the backward potential: $$\boldsymbol{u}^*(x, t) = \sigma_t \nabla \log \varphi_t(x)$$ The key technical advance is the application of the Hopf–Cole transform, linearizing the otherwise nonlinear HJB–FP PDE system into a pair of forward-backward linear PDEs for the SB potentials $(\varphi_t, \hat\varphi_t)$. This enables analytically and numerically tractable computation, and underpins SB algorithms by facilitating alternated or iterative solutions in continuous or discretized time.

---

Schrödinger Bridge as Stochastic Optimal Control and Generative Modeling

The SB is shown to be a prototypical problem in stochastic optimal control (SOC): the construction of a control that interpolates between two marginals with minimum energy (entropy-regularized kinetic action), subject to diffusion. SOC tools—cost functionals, Bellman equations, and value function/dual potential formulations—are deployed to derive closed-form or Feynman–Kac expectation representations for potentials, clarify regularity (optimal control bounded by reference process), and connect with variational inference in generative models.

By developing explicit Radon–Nikodym and endpoint law expressions for the SB path measure, the analysis enables both theoretical interpretation and practical computation of the optimal generative path law. Importantly, the SB formulation naturally encompasses score-based diffusion modeling, flow matching, and recent path-regularized generative frameworks.

---

Extensions: Algorithmic, Variational, and Generalized SBs

The paper systematically treats several significant generalizations:

• Generalized SB: Incorporating mean-field and multi-marginal constraints, enabling, e.g., modeling of interacting particle systems or multi-stage transport.
• Unbalanced and branched SBs: SBs with mass creation/destruction or multiple terminal states.
• Fractional and discrete SBs: SBs with long-memory (fractional Brownian motion) or defined over finite state spaces (Markov chains, CTMCs).

Algorithmic frameworks include forward-backward SDEs (FBSDEs), time-reversal, Sinkhorn iteration in dynamic/path-space form, Doob-$h$ transformations, Markovian/reciprocal projections, and stochastic interpolant construction.

---

Numerical and Theoretical Implications

Explicit focus is given to computational aspects and algorithmic design. Sinkhorn and related iterative projection schemes are shown, both formally and empirically, to strongly converge under integrability conditions. The SB formalism provides the mathematical baseline upon which scalable, stable, and flexible generative algorithms are developed—crucially unifying the construction and training of neural generative models as entropy-regularized optimal control, and supporting recent advances in diffusion model training, simulation-free score/flow matching, and adjoint-based optimization.

Theoretical implications include: (1) strict convexity and regularization yields unique, robust couplings; (2) forward-backward potential structure provides analytic and computational handles for high-dimensional and/or non-Gaussian transport; (3) dynamic formulation and path-space tools enable analysis and generalization to time-inhomogeneous, mean-field, and non-equilibrium generative processes.

---

Conclusion

This monograph delivers a technically complete, coherent theory of Schrödinger bridges for generative modeling. It synthesizes classical and modern results on entropic optimal transport, path-space KL minimization, and stochastic optimal control, providing both foundational insight and a practical mathematical toolkit for the analysis and engineering of generative systems.

The implication for AI and generative modeling is profound: the SB framework offers the natural generalization and unification of diffusion, score-based, and flow-matching models, supports principled extensions to complex and structured data, and establishes a flexible platform for future developments in entropy-regularized, control-theoretic, and sample-efficient generative architecture design. As the field evolves, SB theory will remain the canonical reference point for understanding and constructing transport-based generative models in both the continuous and discrete settings.