Schrödinger Bridge Problems Overview

• Schrödinger Bridge Problems are entropy-minimizing interpolation tasks that find the path measure closest to a reference process while satisfying prescribed endpoint marginals.
• They integrate concepts from stochastic control, optimal transport, and large deviations, employing forward/backward iterative schemes like Sinkhorn iterations.
• Recent advances extend the framework to Gaussian–LQR bridges, geometric formulations on manifolds, quantum dynamics, and hybrid jump-diffusion systems.

Searching arXiv for recent and foundational Schrödinger bridge papers to ground the article. arXiv search query: "Schrödinger bridge problems recent theory control optimal transport"

Schrödinger bridge problems are entropy-minimizing interpolation problems for probability laws on path space: given a reference stochastic evolution and prescribed endpoint marginals, the objective is to find the path measure that matches the endpoints while remaining closest to the reference in relative entropy. In modern terms, they sit at the intersection of stochastic control, large deviations, and entropic optimal transport, and admit formulations on Euclidean spaces, discrete-time Markov chains, manifolds and Lie groups, as well as in data-driven, quantum, and hybrid jump-diffusion settings (Lambert, 12 Jun 2025, Mahmood et al., 14 Mar 2026, Miangolarra et al., 7 Mar 2025).

1. Classical formulation and the Schrödinger system

The classical problem originates in Schrödinger’s 1931 question: given two “snapshots” of a cloud of particles at times $0$ and $T$ with observed marginal densities, what is the most likely stochastic evolution of the cloud in between, assuming a prior dynamics such as Brownian motion (Lambert, 12 Jun 2025). In its modern dynamic form, if $R$ is a reference path measure and $P$ is any path measure with prescribed initial and terminal marginals, the bridge solves

$\min_{P:\,P_0=p_0,\;P_T=p_T} \mathrm{KL}(P\|R).$

For diffusions, Girsanov’s theorem yields an equivalent stochastic control problem with quadratic control energy, and in the zero-temperature limit this connects to classical optimal transport (Mahmood et al., 21 Jun 2026).

A discrete-time formulation makes the entropic structure especially explicit. For a path loss $\ell(x_0,\dots,x_K)$ and temperature $\varepsilon>0$, the problem becomes

$$\min_{p\in\mathcal{P}(p_0,p_K)} E_p[\ell(x_0,\dots,x_K)] + \varepsilon H(p) = \varepsilon \min_{p\in\mathcal{P}(p_0,p_K)} \mathrm{KL}\!\left(p \Big\| \exp(-\ell/\varepsilon)\right),$$

and when $\ell$ depends only on $(x_0,x_K)$ this is exactly entropy-regularized optimal transport (Lambert, 12 Jun 2025). The optimal path law has the multiplicative form

$T$0

where $T$1 are Gibbs factors built from dual Kantorovich potentials. These potentials satisfy the Schrödinger system, a pair of coupled forward/backward equations enforcing the endpoint marginals (Lambert, 12 Jun 2025).

This forward/backward structure underlies both the analytic theory and the standard computational viewpoint. In continuous settings it appears as coupled heat equations or adjoint Kolmogorov equations; in discrete settings it becomes iterative proportional fitting or Sinkhorn-type scaling (Mahmood et al., 21 Jun 2026). A closely related control-theoretic viewpoint represents the optimal drift as a logarithmic gradient of a Schrödinger potential, while time reversal relates forward and backward drifts through the marginal density, a relation that is central in likelihood-based numerical methods (Vargas et al., 2021).

2. Control-theoretic structure and Gaussian–LQR bridges

A prominent modern specialization is the discrete-time linear-quadratic-regulator Schrödinger bridge. There the path loss is pairwise and quadratic,

$T$2

with

$T$3

The first term is a potential term that attracts the process toward prescribed waypoints $T$4, while the second is a kinetic term penalizing large jumps and controlling diffusion or smoothness (Lambert, 12 Jun 2025).

Under Gaussian boundary marginals and positive definite $T$5, the problem is exactly solvable. The backward and forward Gibbs potentials remain Gaussian, the Schrödinger system closes in finite dimensions, and the Kantorovich potentials become quadratic cost-to-go functions propagated by dual Riccati recursions (Lambert, 12 Jun 2025). In this regime, the backward recursion is structurally identical to the discrete-time LQR Riccati equation, while the forward recursion is dual to it in the Kalman sense. The optimal bridge is then a non-homogeneous Gaussian Markov chain with closed-form transition kernels and Gaussian marginals.

This formulation sharpens the control interpretation of Schrödinger bridges. The backward potentials act as LQR-like value functions, the forward potentials as dual information-to-go quantities, and the resulting dynamics can generate curved, looping, or obstacle-avoiding Gaussian trajectories when attractive potentials are activated at selected times (Lambert, 12 Jun 2025). When $T$6, the construction reduces to Gaussian Schrödinger bridges with Brownian-like kinetic cost and recovers Bures transport between Gaussian measures in the small-$T$7 limit; when $T$8, Bures transport is extended to a richer, potential-driven geometry (Lambert, 12 Jun 2025).

3. Computational methods and data-driven formulations

Classical computation proceeds through the Schrödinger system by forward/backward iterations, often described as IPF or Sinkhorn in discrete settings (Lambert, 12 Jun 2025). In practice, several lines of work replace direct solution of the boundary-coupled equations by statistical estimation or bridge-matching procedures tailored to continuous and high-dimensional problems.

A data-driven approach considers the case where only samples from the endpoint marginals are available. The proposed methodology replaces the nonlinear boundary couplings by constrained maximum likelihood estimation and uses importance sampling to propagate the functions solving the Schrödinger system, specifically to avoid grid-based discretizations in high dimension (Pavon et al., 2018). This viewpoint is closely aligned with a later maximum-likelihood formulation of the dynamic bridge, where the bridge drift is recovered by reverse-time regression on discretized trajectories. In that setting, the dynamic Schrödinger problem with general diffusion prior is turned into an iterative proportional maximum likelihood procedure on drifts, and time reversal converts terminal constraints into initial-value simulation problems (Vargas et al., 2021).

A distinct kernel-based formulation characterizes Schrödinger bridges by a family of McKean–Vlasov stochastic control problems without terminal distribution constraints. The terminal constraint is replaced by an MMD penalty built from a characteristic kernel, and the penalized control problem is solved numerically with neural SDEs (Nakano, 2023). Under the stated assumptions, the laws generated by $T$9-optimal controls converge to the unique solution of the original Schrödinger problem, while the kernel machinery supplies a metric on probability laws compatible with weak convergence for translation-invariant kernels with positive Fourier transform (Nakano, 2023).

A unifying perspective places several modern bridge algorithms inside one framework. The proposed Unified Bridge Algorithm subsumes Flow Matching, mini-batch OT Flow Matching, mini-batch Schrödinger bridge Flow Matching, and deep Schrödinger bridge matching by specifying a pinned path family, a coupling on endpoints, a diffusion level, and a regression target for the drift (Kim, 27 Mar 2025). Within that taxonomy, Schrödinger bridge matching methods arise when the pinned bridge is the reference conditioned on endpoints and the learned drift is the Markovian projection of the reciprocal process.

4. Relaxed, empirical, and partially observed bridges

Exact endpoint constraints are not always appropriate in empirical settings. When the prescribed marginals are discrete empirical measures and the reference distribution is absolutely continuous, the classical static bridge may be ill-posed because no feasible coupling has finite entropy relative to the reference. A transport-relaxed Schrödinger bridge resolves this by replacing hard marginal constraints with optimal-transport penalties: $R$0 or, with penalty parameter $R$1,

$R$2

In the semi-discrete quadratic-cost setting, the dual problem reduces to a finite-dimensional concave optimization over max-affine potentials, and both primal and dual solutions are unique under the stated conditions (Jiang et al., 8 Feb 2026). As the penalty blows up, the solutions converge to a discrete Schrödinger bridge with a leading-order logarithmic divergence in the optimal value (Jiang et al., 8 Feb 2026).

A different departure from the classical setting arises with incomplete information. In a discrete Schrödinger bridge with partial marginal observations, only selected components of the marginal vectors are prescribed at each time, while the total mass and unobserved components remain unknown. The resulting optimization is convex but not strictly convex, and standard Sinkhorn-type methods cannot be directly applied (Mascherpa et al., 7 Apr 2026). An entropic proximal scheme is introduced for this setting, together with duality results, a characterization of optimal solutions, and an observability condition that determines when the optimizer is unique. The same framework is applied to contamination tracking in a water distribution network, where partial marginals correspond to measured pollutant concentrations at sensor locations (Mascherpa et al., 7 Apr 2026).

These two directions modify the usual interpretation of endpoint constraints in complementary ways. Transport relaxation softens exact matching to accommodate empirical data; partial-observation bridges retain exact constraints where measurements exist but infer the unobserved mass through entropy minimization and dynamics (Jiang et al., 8 Feb 2026, Mascherpa et al., 7 Apr 2026).

5. Geometric formulations on manifolds and Lie groups

Schrödinger bridges are not restricted to Euclidean spaces. On $R$3, an isotropic bridge for the kinematic equation is formulated intrinsically, with the control given by the angular velocity and the uncontrolled generator equal to $R$4, the Laplace–Beltrami operator (Mahmood et al., 21 Jun 2026). The associated Fokker–Planck equation is written entirely in geometric terms, and a Hopf–Cole transform converts the nonlinear HJB–Fokker–Planck system into a pair of forward/backward heat equations on the group. Existence and uniqueness of the corresponding Schrödinger system are established by proving that a fixed-point recursion is contractive in Hilbert’s projective metric on the cone of positive continuous functions (Mahmood et al., 21 Jun 2026).

The same geometric program extends from $R$5 to compact connected Lie groups. For a compact connected Lie group $R$6, the controlled diffusion is written in left-trivialized form,

$R$7

with Haar measure, Lie algebra inner product, and Laplace–Beltrami operator providing the intrinsic geometric data (Mahmood et al., 14 Mar 2026). The Schrödinger system again takes the form of coupled heat semigroup equations, now on $R$8, and the solution is unique up to reciprocal scaling. The optimal controller is expressed through the Riemannian gradient of $R$9, left-trivialized to the Lie algebra, so the entire construction remains coordinate-free and avoids Euclidean embeddings or local parameterizations (Mahmood et al., 14 Mar 2026).

These geometric results show that the classical forward/backward factorization survives on nonlinear configuration spaces. Heat-kernel positivity, Hilbert-metric contraction, and the multiplicative coupling of forward and backward potentials remain the central analytic ingredients, while the controller becomes an intrinsic gradient field on the group (Mahmood et al., 21 Jun 2026, Mahmood et al., 14 Mar 2026).

6. Nonclassical extensions: non-Markov, quantum, and hybrid regimes

The Markovian reference-process paradigm can break in averaged or uncertain systems. For an ensemble of linear diffusions subject to parameter perturbations, the averaged process is generally a non-Markov Gaussian process, and the corresponding bridge over the averaged system yields an optimal control that is not Markovian but a stochastic feedforward law depending on past and present noise (Adu et al., 2024). A path-integral construction expresses the global bridge as an integral over pinned endpoint controls weighted by a posterior on final states, showing that robust distribution steering under parameter uncertainty may require genuinely non-Markovian strategies (Adu et al., 2024).

Quantum extensions preserve the bridge logic while altering the objects being bridged. One formulation studies time-symmetric ensembles generated by pre- and post-selected measurements around a Markovian quantum experiment. The corresponding quantum Schrödinger bridge is the most likely joint distribution of initial and final outcomes consistent with observed endpoint density matrices, and the solution retains the classical product structure in the sense that endpoint and intermediate density matrices can be written as products of forward-evolving and backward-evolving matrices (Miangolarra et al., 7 Mar 2025). A related formulation derives exact Gaussian solutions to a quantum Schrödinger bridge problem from a Lagrangian treatment of dynamical optimal transport. There the evolution is governed by a Schrödinger equation, the effective Hamilton–Jacobi equation contains the Bohm potential, and Gaussian endpoints produce a Gaussian process whose covariance evolution differs from the classical bridge because of quantum effects (Bordyuh et al., 30 Sep 2025).

Hybrid continuous–discrete systems admit yet another extension. When the reference law is a regime-switching jump diffusion on $P$0, the bridge is again a regime-switching jump diffusion, with drift, jump intensity, and switching rates transformed by a Schrödinger potential through a Doob $P$1-transform (Zlotchevski et al., 8 Nov 2025). Under suitable assumptions, the paper derives both stochastic-calculus and analytic characterizations of the bridge and develops a regime-switching perspective on an unbalanced Schrödinger bridge with killing (Zlotchevski et al., 8 Nov 2025).

These developments show that several properties often associated with the classical bridge are conditional rather than universal. The optimal bridge is Markovian in many standard settings, including pairwise discrete-time models and Gaussian–LQR bridges, but not for averaged systems with parameter perturbations (Lambert, 12 Jun 2025, Adu et al., 2024). Likewise, Brownian motion on $P$2 is only one reference among many; Lie-group diffusions, quantum channels, and regime-switching jump diffusions all admit Schrödinger-type bridge constructions (Mahmood et al., 14 Mar 2026, Miangolarra et al., 7 Mar 2025, Zlotchevski et al., 8 Nov 2025).

Schrödinger bridge problems therefore form a broad variational framework rather than a single model class. Across these settings, the recurring elements are endpoint constraints, an entropy or KL objective relative to a reference evolution, and a pair of forward/backward objects—potentials, scaling functions, or operator analogues—that encode the optimal interpolation. Modern work extends this template toward geometric state spaces, empirical and partially observed marginals, structured control models, and generative modeling applications ranging from trajectory planning and contamination tracking to single-cell dynamics, image generation, molecular translation, and mean-field games (Lambert, 12 Jun 2025, Mascherpa et al., 7 Apr 2026, Bordyuh et al., 30 Sep 2025).