Schrödinger–Bass Bridge Problem

• The Schrödinger–Bass Bridge problem is a stochastic optimal transport framework that interpolates between drift and volatility control using a quadratic cost functional.
• It leverages explicit PDE reductions and duality theory to derive closed-form feedback controls and monotone transport maps for achieving prescribed endpoint marginals.
• Numerical methods like LightSBB-M demonstrate its efficiency in generative modeling and image translation by significantly reducing 2-Wasserstein distances compared to classical approaches.

The Schrödinger--Bass Bridge (SBB) problem refers to a parametric family of stochastic optimal transport problems that interpolate between the classical Schrödinger Bridge (drift control, entropy-regularized) and the Bass martingale optimal transport (volatility control) regimes. The SBB formulation provides a quadratic cost over both drift and volatility, generating a stochastic process whose endpoint marginals match prescribed distributions and whose law is closest to a Brownian reference in relative entropy. The SBB has received recent rigorous characterizations, including explicit PDE reductions, duality theory, closed-form feedback controls, and high-fidelity numerical solvers for high-dimensional generative data applications.

1. Mathematical Formulation: Pathwise Control and Objective

The general SBB problem asks: Given endpoint probability laws $\mu_0,\,\mu_T$ on $\mathbb{R}^d$ and a time horizon $[0,T]$, find a continuous semimartingale $X_t$ of the form

$$dX_t = \alpha_t\,dt + \sigma_t\,dW_t, \quad X_0 \sim \mu_0,\ X_T \sim \mu_T$$

minimizing the quadratic path-energy functional

$$SBB(\mu_0,\mu_T) := \inf_{(\alpha,\sigma)} \mathbb{E} \left[ \frac{1}{2} \int_0^T \left(|\alpha_t|^2 + \beta|\sigma_t - 1|^2\right)dt \right]$$

where $\beta > 0$ interpolates between pure drift ($\beta \gg 1$) and pure volatility ($\beta \to 0$) penalty regimes. For $\beta \gg 1$, $\mathbb{R}^d$0 and the problem recovers the classical Schrödinger Bridge; for $\mathbb{R}^d$1, the drift vanishes and the transport is performed entirely via volatility ("Bass martingale transport") (Alouadi et al., 25 Jan 2026, Alouadi et al., 27 Jan 2026).

Strong duality holds: $\mathbb{R}^d$2, where the dual potentials $\mathbb{R}^d$3 solve a nonlinear backward Hamilton-Jacobi-Bellman PDE with $\mathbb{R}^d$4 everywhere.

2. PDE Derivation and Explicit Solution in One Dimension

In dimension one, the SBB admits an explicit reduction via Legendre transforms and heat equations. Introduction of a dual field $\mathbb{R}^d$5 leads to a nonlinear HJB PDE:

$\mathbb{R}^d$6

Changing variable $\mathbb{R}^d$7 and Legendre-transforming in $\mathbb{R}^d$8 yields a linear backward heat equation in the new unknown $\mathbb{R}^d$9, i.e.,

$[0,T]$0

The endpoint coupling is enforced by linking $[0,T]$1 and the forward Kolmogorov density via time-dependent inverse gradient maps. The process $[0,T]$2 is realized as a time-dependent monotone transport ("stretch") of a classical Schrödinger bridge $[0,T]$3 by $[0,T]$4, where

$[0,T]$5

Thus, the SBB solution is the composition of a strictly monotone map with an entropic bridge (Alouadi et al., 25 Jan 2026).

3. Limiting Behavior: Schrödinger and Bass Regimes

The SBB family interpolates between two fundamental stochastic transport regimes:

• Schrödinger Bridge Limit ($[0,T]$6): Volatility control becomes rigid, recovering the classical minimum-energy drift bridge with prescribed marginals (Alouadi et al., 25 Jan 2026, Alouadi et al., 27 Jan 2026).
• Bass Martingale Limit ($[0,T]$7): Drift penalty dominates, imposing $[0,T]$8 with $[0,T]$9 subject to end-marginal constraints, as in classical monotone martingale transport (Alouadi et al., 25 Jan 2026). The gradient map simplifies to the Bass/Brenier transport.
• Intermediate ($X_t$0): SBB yields a strictly stretched semimartingale that unifies both mechanisms.

4. Feedback Controls: Analytic Expressions and Transport Maps

Optimal feedbacks are rendered in closed form upon solving the dual PDEs. For the multidimensional setting, the feedbacks are given by (Alouadi et al., 27 Jan 2026):

$X_t$1

where $X_t$2 is the Bellman potential. Alternatively, expressing the solution as a stretched Schrödinger bridge in terms of a potential $X_t$3 and time-dependent map $X_t$4, yields

$X_t$5

with $X_t$6 governed by the entropic bridge SDE and $X_t$7 stretched by the gradient map. These controls allow bypassing black-box SDE solvers and yield robust simulation of the SBB process.

5. Duality, Existence, and Uniqueness

Strong duality is established: the primal infimum over controlled diffusions equals the dual supremum over smooth Bellman potentials $X_t$8 with $X_t$9 under regularity conditions $$dX_t = \alpha_t\,dt + \sigma_t\,dW_t, \quad X_0 \sim \mu_0,\ X_T \sim \mu_T$$0 and finite second moments of $$dX_t = \alpha_t\,dt + \sigma_t\,dW_t, \quad X_0 \sim \mu_0,\ X_T \sim \mu_T$$1 (Alouadi et al., 25 Jan 2026, Alouadi et al., 27 Jan 2026). The backward linear heat reduction ensures uniqueness. The stretching map $$dX_t = \alpha_t\,dt + \sigma_t\,dW_t, \quad X_0 \sim \mu_0,\ X_T \sim \mu_T$$2 is a Brenier map from the reference bridge's marginal to the target at each time, with the endpoint constraints enforced via coupling of pushforwards of $$dX_t = \alpha_t\,dt + \sigma_t\,dW_t, \quad X_0 \sim \mu_0,\ X_T \sim \mu_T$$3 and $$dX_t = \alpha_t\,dt + \sigma_t\,dW_t, \quad X_0 \sim \mu_0,\ X_T \sim \mu_T$$4.

6. Numerical Algorithms and Applications

LightSBB-M is a practical algorithm to solve the SBB by alternating bridge-matching (mean regression, Gaussian denoising loss) with transport-map learning, using neural networks to parameterize the score-model and map invertibility. Empirically, LightSBB-M attains the lowest 2-Wasserstein distances compared to SB and diffusion baselines ($$dX_t = \alpha_t\,dt + \sigma_t\,dW_t, \quad X_0 \sim \mu_0,\ X_T \sim \mu_T$$5–$$dX_t = \alpha_t\,dt + \sigma_t\,dW_t, \quad X_0 \sim \mu_0,\ X_T \sim \mu_T$$6 improvement), with superior generative fidelity in image translation (adult$$dX_t = \alpha_t\,dt + \sigma_t\,dW_t, \quad X_0 \sim \mu_0,\ X_T \sim \mu_T$$7child faces, FFHQ) (Alouadi et al., 27 Jan 2026). The solver converges in a few iterations—a direct consequence of the explicit analytic control formulae and heat-equation reductions.

Algorithm	2-Wasserstein Distance (Moons$$dX_t = \alpha_t\,dt + \sigma_t\,dW_t, \quad X_0 \sim \mu_0,\ X_T \sim \mu_T$$88-Gauss)	Image Translation Fidelity
LightSB-M	$$dX_t = \alpha_t\,dt + \sigma_t\,dW_t, \quad X_0 \sim \mu_0,\ X_T \sim \mu_T$$9	Moderate
SBB (Ours)	$$SBB(\mu_0,\mu_T) := \inf_{(\alpha,\sigma)} \mathbb{E} \left[ \frac{1}{2} \int_0^T \left(|\alpha_t|^2 + \beta|\sigma_t - 1|^2\right)dt \right]$$0	High / Diverse

Strengths of SBB/LightSBB-M include analytic controls, robust convergence, and flexible manipulation of transport geometry via $$SBB(\mu_0,\mu_T) := \inf_{(\alpha,\sigma)} \mathbb{E} \left[ \frac{1}{2} \int_0^T \left(|\alpha_t|^2 + \beta|\sigma_t - 1|^2\right)dt \right]$$1. Limitations involve the lack of a rigorous convergence proof for the alternating solver and overhead for high-dimensional neural map inversion (Alouadi et al., 27 Jan 2026).

7. Implications and Extensions

The SBB paradigm establishes a unified theoretical and computational framework connecting entropy-regularized (Schrödinger) and martingale (Bass) transport, with explicit monotone map structures at all intermediate points. The explicit semimartingale representation, PDE reductions, and existence/uniqueness theorems underpin scalable implementations for generative modeling, optimal stochastic control, and high-dimensional time-series synthesis. Future directions include rigorous complexity analysis, extension to time-series/factor models, and leveraging SBB in data-driven scientific and financial synthesis.

References: (Alouadi et al., 25 Jan 2026, Alouadi et al., 27 Jan 2026, Hagmann, 2023, Alouadi et al., 27 Jan 2026).