no. 16 · Tuesday 17 November 2026
Soft-constrained Schrödinger bridge for generative AI: numerical algorithms and approximation guarantees
Abstract
We study a soft-constrained formulation of the Schrödinger bridge problem, in which the terminal distribution constraint is relaxed into a penalty on the discrepancy between the terminal law of the controlled process and a target distribution. This leads to an unconstrained McKean–Vlasov stochastic control problem, where the interaction enters only through the terminal cost. We characterize the optimizer via a McKean–Vlasov forward-backward stochastic differential equation and establish well-posedness under very mild assumptions. We also analyze two approximation regimes relevant to data-driven computation: approximation of the target law by empirical samples, and finite-particle approximation of the McKean–Vlasov control problem. For both, we derive convergence results and error estimates. Finally, building on the well-posedness and approximation results, we investigate two neural numerical methods: one based on the forward-backward system and the other on direct control optimization. Both methods are tested on state-of-the-art generative tasks. Our results provide both an algorithmic foundation and its theoretical guarentee for soft-constrained Schrödinger bridge problems in sample-based generative modeling.
- fields
- Optimization and Control, Analysis of PDEs