Learning non-equilibrium diffusions with Schrödinger bridges: from exactly solvable to simulation-free
Stephen Y. Zhang, Michael P. H. Stumpf · 2025
Read Learning non-equilibrium diffusions with Schrödinger bridges: from exactly solvable to simulation-free by Stephen Y. Zhang, Michael P. H. Stumpf (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-16644.txt.
Introduction
Stephen Zhang and Michael Stumpf tackle the Schrödinger bridge problem for systems that live out of equilibrium, in a 2025 paper. Given snapshots of a system's start and end states, the task is to reconstruct its most likely path between them.
Most methods assume a simple random-walk reference, which fits only systems pulled by a potential. The authors allow a richer reference motion with a general drift matrix, covering the non-gradient forces common in biology. For Gaussian snapshots they derive the exact bridge, static and dynamic. For general cases they propose a simulation-free learning method based on flow and score matching. Tests on synthetic and real single-cell data beat rival methods on accuracy and training speed.
Catalog record
| CREATOR | Stephen Y. Zhang, Michael P. H. Stumpf |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 117 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-16644
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Learning non-equilibrium diffusions with Schrödinger bridges: from exactly solvable to simulation-free Stephen Y. Zhang, Michael P. H. Stumpf Abstract We consider the Schrödinger bridge problem which, given ensemble measurements of the initial and final configurations of a stochastic dynamical system and some prior knowledge on the dynamics, aims to reconstruct the “most likely” evolution of the system compatible with the data. Most existing literature assume Brownian reference dynamics, and are implicitly limited to modelling systems driven by the gradient of a potential energy. We depart from this regime and consider reference processes described by a multivariate Ornstein-Uhlenbeck process with generic drift matrix 𝑨 ∈ ℝ d × d \bm{A}\in\mathbb{R}^{d\times d} . When 𝑨 \bm{A} is asymmetric, this corresponds to a non-equilibrium system in which non-gradient forces are at play: this is important for applications to biological systems, which naturally exist out-of-equilibrium. In the case of Gaussian marginals, we derive explicit expressions that characterise exactly the solution of both the static and dynamic Schrödinger bridge. For general marginals, we propose mvOU-OTFM , a simulation-free algorithm based on flow and score matching for learning an approximation to the Schrödinger bridge. In application to a range of problems based on synthetic and real single cell data, we demonstrate that mvOU-OTFM achieves higher accuracy compared to competing methods, whilst being significantly faster to train. 1 Introduction We are interested in reconstruction of stochastic dynamics of individuals from static population snapshots. This is a central problem with applications arising across the natural and social sciences, whenever longitudinal tracking of individuals over time is either impossible or impractical Weinreb et al. (2018) ; Muzellec et al. (2017) ; Stuart and Wolfram (2020) ; Schiebinger et al. (2019) ; Miles et al. (2024) ; Abel and Sander (2014) . In simple terms, consider a system of indistinguishable particles \bm{x}_{t}\in\mathbb{R}^{d} undergoing some unobserved temporal dynamics. The practitioner observes the system to have distribution \bm{x}_{0}\sim\rho_{0} at an initial time t=0 and later to be \bm{x}_{1}\sim\rho_{1} at a final time t=1 . The question is then: can we, under some suitable assumptions, reconstruct the continuous-time behaviour of the system for the unobserved time interval 0<t<1 ? The Schrödinger bridge problem (SBP), by now a centrepiece of the theoretical literature on this topic, places this task on a theoretical footing in terms of a mathematical formulation of this problem as a large deviations principle on the path space Léonard (2013) and a (stochastic) least action principle intimately linked to optimal transportation theory Chen et al. (2021) . Given a reference process that encodes prior knowledge on the dynamics, the SBP can be understood as identifying the “most likely” evolution of the system that is compatible with the snapshot observations. The SBP and related topics have enjoyed a great deal of recent interest from both applied and theoretical perspectives. Many applications arise from biological modelling of cell dynamics Bunne et al. (2024) ; Schiebinger (2021) . The majority of existing work assumes, either explicitly or implicitly, that the system of interest is potential driven , that is, driven by the gradient of a potential energy Lavenant et al. (2021) ; Chizat et al. (2022) ; Tong et al. (2023b) ; Bunne et al. (2023) ; Tong et al. (2020) ; Yeo et al. (2021) ; Bunne et al. (2022) . In fact, most studies do not consider any prior drift, corresponding to the setting where the reference process is purely an isotropic Brownian motion. A few studies Bunne et al. (2023) ; Shi et al. (2023) ; De Bortoli et al. (2021) allow for scalar Ornstein-Uhlenbeck (OU) processes, motivated by applications in the generative modelling domain. Since scalar OU dynamics have unidimensional drift which can always be written as the gradient of a scalar function, all of these are also potential driven systems. Changing the choice for the reference process provides a route to dealing with systems which are not potential driven, i.e. driven by a drift that is a non-conservative vector field. Abundant motivations for modelling these dynamics arises from biological systems and other kinds of active matter; these exist naturally far from equilibrium Zhu et al. (2024) ; Parrondo et al. (2009) ; Gnesotto et al. (2018) ; Fang et al. (2019) , exhibiting irreversible dynamics at a non-equilibrium steady state Lax (1960) . While a few studies Vargas et al. (2021) ; Shen et al. (2024) ; Koshizuka and Sato (2022) allow for reference processes described by a SDE with a generic drift term, they rely on computationally expensive simulation procedures such as numerical integration and suffer from accuracy issues in high dimensions. Here we explore a complementary approach and concentrate our attention on a family of linear reference dynamics as a middle ground between physical relevance and analytical tractability. Specifically, we consider reference processes arising in a class of linear drift-diffusion dynamics described by the SDE \mathrm{d}\bm{X}_{t}=\bm{A}(\bm{X}_{t}-\bm{m})\>\mathrm{d}t+\bm{\sigma}\>\mathrm{d}\bm{B}_{t}. (1) These processes are also known as multivariate Ornstein-Uhlenbeck (mvOU) processes, often used as a model of non-equilibrium systems in the physics literature Lax (1960) ; Godrèche and Luck (2018) ; Gilson et al. (2023) . Indeed, for asymmetric drift matrix \bm{A} the system ( 1 ) has a drift that is no longer the gradient of a potential. As an example in the case of isotropic diffusion \bm{\sigma}=\sigma\mathbf{I} , when \bm{A} is asymmetric and has all eigenvalues with negative real part, a non-equilibrium steady state exists where the dynamics are irreversible and exhibit nonzero probability currents while the population density is unchanged Gilson et al. (2023) ; Godrèche and Luck (2018) . While allowing a broader range of reference dynamics, the statistics of ( 1 ) remain analytically tractable – in particular it is a Markovian Gaussian process with explicit formulae for its mean, covariance, and transition kernel. This provides avenues to efficient solution of the SBP that sidesteps the need for numerical integration, and as we will see, even analytical expressions in the Gaussian case. In particular, the family of SBP problems that we model is strictly larger than the standard (Brownian) SBP, which we recover as a special case when \bm{A},\bm{m}=\bm{0} . The contributions of our paper are twofold. Leveraging results on mvOU processes, we characterise for Gaussian endpoints the Gaussian Schrödinger Bridge (GSB) with reference dynamics described by ( 1 ). To handle general marginals, we develop a simulation free training procedure based on score and flow matching Tong et al. (2023b) to solve the SBP. We conclude by demonstrating our results in a range of synthetic and real data examples. We find that our approach solves the generalised SBP faster and more accurately than comparable simulation-based algorithms, at the cost of assuming the form ( 1 ) of the reference process. Our findings highlight the tradeoff between (i) analytically tractable but more restrictive models and (ii) more expressive non-parametric models, which are much less tractable to learn, both in terms of computational cost and statistical complexity. Figure 1 : The Schrödinger bridge problem with a multivariate Ornstein-Uhlenbeck reference process ( 1 ) can be solved via a generalised entropic transport problem and characterisation of the \mathbb{Q} -bridges. For non-Gaussian endpoints, score and flow matching provide a route to building neural approximations without simulation. 2 Background and