Schrödinger Bridge & Bass Formulation

• The SBB formulation is a framework that integrates entropic optimal transport, stochastic control, and martingale transport by jointly controlling drift and volatility.
• It generalizes classical Schrödinger bridges and Bass couplings by introducing a parameter to balance penalization between drift and volatility, enabling smooth interpolation between transport regimes.
• Advanced numerical methods and PDE formulations yield scalable solvers, with applications spanning generative modeling, optimal transport, and stochastic control.

The Schrödinger Bridge and Bass (SBB) formulation generalizes the classical Schrödinger bridge problem by incorporating joint control of drift and volatility in stochastic dynamic couplings. SBB bridges entropic optimal transport, stochastic control, and martingale transport, providing a one-parameter family of interpolation problems between probability measures that subsume both Schrödinger bridges and Bass martingale couplings. The SBB formalism is now pivotal in mathematical physics, statistics, optimization, generative modeling, and control theory.

1. Classical Schrödinger Bridge and Bass Coupling

The Schrödinger bridge problem originated as the challenge of describing the most likely path evolution of a system of Brownian particles, conditioned to start and end at prescribed distributions. The formulation replaces the quadratic transportation cost in the Monge-Kantorovich problem by relative entropy (Kullback-Leibler divergence) with respect to a reference stochastic process. For probability measures $\mu,\nu\in\mathcal{P}_2(M)$ on a manifold $M$ and a reference law $P$ (typically Brownian motion or Langevin diffusion), the dynamic Schrödinger problem is

$$\min_{Q\in\mathcal{P}(\Omega)} H(Q\mid P) \quad\text{s.t.}\quad (X_0)_\#Q = \mu,\; (X_1)_\#Q=\nu$$

where $H(Q\mid P)$ is the relative entropy and $(X_t)$ are canonical coordinate maps on path space $\Omega=C([0,1],M)$ (Conforti, 2017).

Bass's formulation constructs Markov couplings via SDEs whose drift is expressed as

$$b_t(x)=\frac{1}{2}\nabla\left[\log\Psi_t(x)-\log\Phi_t(x)\right]$$

where $\Psi_t$ and $\Phi_t$ are suitable marginal potentials. This overlap with the Schrödinger bridge formalism allows one to view Bass's coupling as the zero-noise, martingale-constrained instance of the entropic optimal transport problem (Conforti, 2017).

2. The SBB Stochastic Optimal Coupling and Interpolation

The Schrödinger–Bass Bridge (SBB) extends the classical SB by introducing separate penalization for both drift and volatility in the pathwise cost. For marginals $\mu_0,\mu_T$ on $\mathbb{R}^d$, and parameter $\beta>0$, the SBB optimal coupling is defined by

$$\min_{(b,\sigma)} \mathbb{E}\left[\int_0^T \left(\|b_t(X_t)\|^2 + \beta\|\sigma_t(X_t) - \sqrt{\epsilon}I_d\|_F^2\right) dt\right]$$

subject to $X_0\sim\mu_0$, $X_T\sim\mu_T$ and the SDE

$dX_t = b_t(X_t)\,dt + \sigma_t(X_t)\,dW_t$

(Alouadi et al., 27 Jan 2026, Alouadi et al., 25 Jan 2026). The interpolating parameter $\beta$ regulates the relative penalization between drift and volatility:

• $\beta\to\infty$: $\sigma_t\to\sqrt{\epsilon}I_d$, recovering classical Schrödinger bridge (drift-only control).
• $\beta\to0$: drift is suppressed, yielding pure volatility (Bass martingale transport).

The optimal law is characterized as the minimizer of KL divergence to a reference process with zero drift and fixed volatility $\sqrt{\epsilon}I_d$ (Alouadi et al., 27 Jan 2026).

3. PDE-Derivation, Potentials, and Stretched Bridge Representation

The SBB problem admits a dual variational formulation wherein the value function $v(t,x)$ solves a Hamilton–Jacobi–Bellman PDE subject to a convexity constraint on its Hessian. In one dimension,

$$\partial_t v + \frac{1}{2}|\partial_x v|^2 + \frac{1}{2}\frac{\partial_{xx}v}{1-\partial_{xx}v/\beta}=0$$

(Alouadi et al., 25 Jan 2026). Employing Legendre transforms and representing dual variables via the heat equation enables explicit construction of monotone transport maps $\mathcal{X}, \mathcal{Y}$. The “Stretched Schrödinger Bridge” formalism expresses the SBB optimal process $X_t$ as

$$X_t = \mathcal{X}(t, Y_t) = Y_t + \frac{1}{\beta}\,\partial_y\log h(t, Y_t)$$

where $Y_t$ is the classical Schrödinger bridge, and $h(t,y)$ solves the backward heat equation; $\mathcal{X}$ stretches the base bridge according to the SBB interpolation (Alouadi et al., 25 Jan 2026).

As $\beta\to\infty$, the stretching vanishes and $X_t\to Y_t$, reproducing Sinkhorn and entropic OT. As $\beta\to0$, the transform approaches the optimal transport map for the Bass martingale coupling (Alouadi et al., 25 Jan 2026).

4. Geometric and Functional Properties

Schrödinger bridges (and by extension SBB) admit a Riemannian geometric interpretation on the Wasserstein manifold $\mathcal{P}_2(M)$ (Conforti, 2017). Here, path-marginals $(\mu_t)$ evolve with velocity field $v_t$ satisfying

$\partial_t\mu_t + \nabla\cdot(\mu_t v_t)=0$

and their acceleration obeys

$$\frac{D}{dt}v_t = \frac{\sigma^2}{8}\operatorname{grad}I(\mu_t)$$

where $I(\mu_t)$ is the Fisher information (Conforti, 2017). This viewpoint yields second-order geodesic equations, functional inequalities for entropic transport cost, and convexity results for Fisher information along the bridge. Generalized Talagrand inequalities and Ricci/Bakry-Émery curvature bounds apply in this setting (Conforti, 2017).

5. Numerical Algorithms and Scalable Solvers

Recent advances have produced highly efficient algorithms for computing SBB plans. The LightSBB-M algorithm (Alouadi et al., 27 Jan 2026) alternates between transport map regression and score-matching on a sampled bridge, leveraging analytic formulas for drift and volatility from the SBB dual, and updating a transport map via regression. A typical step proceeds by sampling endpoint pairs, solving for analytic score drifts, applying regression updates, and matching the coupling in Y-space. Empirically, only a few iterations ($K\approx5$) suffice for convergence.

Closed-form controls for optimal drift and volatility are derived from the dual variational problem (using Fenchel–Legendre transforms), and efficient bridge-matching loss functions are employed. SBB-based solvers demonstrate superior performance on 2-Wasserstein distance metrics and generative tasks relative to prior SB and diffusion algorithms (Alouadi et al., 27 Jan 2026).

The SBB PDE formulation also admits solution via alternating Sinkhorn-like updates on boundary data and forward/backward heat equations, with extensions for Riccati sweeps in discrete-time (LQR-SBB) (Alouadi et al., 25 Jan 2026, Lambert, 12 Jun 2025).

6. Specializations: Reflected, LQR, and State-Cost SBB

Reflected SBB formulations govern optimal path planning under uncertainty with reflecting boundary conditions, yielding coupled Fokker-Planck and HJB systems. Efficient finite-element discretizations naturally enforce Neumann reflection and are contractive in the Hilbert metric (Kalise et al., 18 Nov 2025).

With quadratic state-costs, exactly solvable SBB kernels are represented by the Mehler kernel, derived via Weyl calculus. Here, Schrödinger potentials recover optimal drift and transition densities in closed form, and the Bass formulation is interpretationally equivalent to weighted Brownian bridges (Feynman-Kac weights) (Teter et al., 2024). Discrete-time formulations with LQR cost yield closed-form Riccati updates for Kantorovich potentials and transition kernels, with Bures-Wasserstein geodesics as a limiting case (Lambert, 12 Jun 2025).

7. Applications and Computational Modality

SBB bridges are widely applied in generative modeling, stochastic control, and optimal transport. Empirical advances include fast unpaired image-to-image translation tasks, generative diffusion with minimal neural function evaluations, and the design of scalable continuous normalizing flow algorithms for entropy-regularized optimal transport (Jing et al., 22 Mar 2025, Kim et al., 2024, Alouadi et al., 27 Jan 2026). Score-matching methods, probability-flow ODE reformulations, and duality-based neural parameterizations have shown both theoretical convergence and practical efficiency, enabling their deployment in high-dimensional, non-Euclidean, or boundary-constrained domains.

In summary, the SBB formulation unifies entropic transport, martingale constraints, and stochastic control by offering an analytically tractable, geometrically grounded, and computationally scalable framework for optimal dynamic coupling and interpolation between probability measures. It encompasses classical Schrödinger bridges, Bass martingales, and their interpolants, and extends to advanced problem settings, such as reflected domains, quadratic state costs, and optimal trajectory planning in uncertainty (Conforti, 2017, Alouadi et al., 27 Jan 2026, Alouadi et al., 25 Jan 2026, Kalise et al., 18 Nov 2025, Teter et al., 2024, Lambert, 12 Jun 2025, Jing et al., 22 Mar 2025, Kim et al., 2024).