Exact Solutions to the Quantum Schrödinger Bridge Problem
Mykola Bordyuh, Djork-Arné Clevert, Marco Bertolini · 2025
Read Exact Solutions to the Quantum Schrödinger Bridge Problem by Mykola Bordyuh, Djork-Arné Clevert, Marco Bertolini (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-2509-25980.txt.
Introduction
Mykola Bordyuh, Djork-Arne Clevert, and Marco Bertolini solve a quantum version of the bridge problem exactly in a 2025 study. The classical bridge asks for the most likely stochastic path between two distributions; the quantum version lets the dynamics follow the Schrödinger equation instead of a real wave equation. The problem was known to mathematicians, but the authors recast it from a Lagrangian angle aimed at generative modeling.
Their derivation shows the evolution equation takes a tractable form with exact solutions in the cases they treat. Exact bridges could give quantum-inspired generative models the same firm footing diffusion bridges gave classical ones.
Catalog record
| CREATOR | Mykola Bordyuh, Djork-Arné Clevert, Marco Bertolini |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 78 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-25980
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Exact Solutions to the Quantum Schrödinger Bridge Problem Mykola Bordyuh, Djork-Arné Clevert, Marco Bertolini Abstract The Quantum Schrödinger Bridge Problem (QSBP) describes the evolution of a stochastic process between two arbitrary probability distributions, where the dynamics are governed by the Schrödinger equation rather than by the traditional real-valued wave equation. Although the QSBP is known in the mathematical literature, we formulate it here from a Lagrangian perspective and derive its main features in a way that is particularly suited to generative modeling. We show that the resulting evolution equations involve the so-called Bohm (quantum) potential, representing a notion of non-locality in the stochastic process. This distinguishes the QSBP from classical stochastic dynamics and reflects a key characteristic typical of quantum mechanical systems. In this work, we derive exact closed-form solutions for the QSBP between Gaussian distributions. Our derivation is based on solving the Fokker-Planck Equation (FPE) and the Hamilton-Jacobi Equation (HJE) arising from the Lagrangian formulation of dynamical Optimal Transport. We find that, similar to the classical Schrödinger Bridge Problem, the solution to the QSBP between Gaussians is again a Gaussian process; however, the evolution of the covariance differs due to quantum effects. Leveraging these explicit solutions, we present a modified algorithm based on a Gaussian Mixture Model framework, and demonstrate its effectiveness across several experimental settings, including single-cell evolution data, image generation, molecular translation and applications in Mean-Field Games. 1 Introduction The Schrödinger Bridge Problem (SBP), in its dynamical formulation via entropy-regularized optimal transport cuturi2013sinkhorn , seeks the most likely stochastic evolution that transports mass between two arbitrary distributions, \pi_{0} and \pi_{1} , while remaining close (in a relative entropy sense) to a reference diffusion process such as a Wiener process leonard2013survey ; chen2021stochastic . In machine learning, diffusion generative models can be viewed as a special case of the SBP, where the goal is to learn an optimal transformation from a simple reference distribution (typically Gaussian) to a complex target distribution song2020score ; ho2020denoising ; de2021diffusion ; vincent2011connection ; hyvarinen2005estimation ; sohl2015deep ; huang2021variational . Under the transition density of Brownian motion, paths that transport mass from arbitrary \pi_{0} to \pi_{1} are highly unlikely; the SBP addresses this by selecting, among all such low-probability trajectories, the one that is most likely. This optimal stochastic process has time-marginal density p(\mathbf{x},t) that admits the factorization p(\mathbf{x},t)=\phi(\mathbf{x},t)\,\widehat{\phi}(\mathbf{x},t) , where the potentials \phi and \widehat{\phi} satisfy heat equations \displaystyle\partial_{t}\phi \displaystyle=-\beta\Delta\phi~, \displaystyle\partial_{t}\widehat{\phi} \displaystyle=\beta\Delta\widehat{\phi}~, (1) with \beta denoting the diffusion (Wiener) coefficient. This structure closely mirrors quantum mechanics, where the probability density is given by p(\boldsymbol{x},t)=\psi(\boldsymbol{x},t)\overline{\psi}(\boldsymbol{x},t) , with \psi satisfying the (imaginary-time) Schrödinger equation i\frac{\partial\psi}{\partial t}=-\beta\Delta\psi , highlighting a deep mathematical connection between stochastic control and quantum dynamics griffiths2019introduction ; schrodinger1926undulatory . This analogy was rigorously formalized by Guerra and Morato guerra1983quantization , and extended in subsequent works carlen1984conservative ; nagasawa2012schrodinger ; nelson2020quantum , which demonstrated that the SBP admits an alternative formulation as a stochastic control problem, governed not by classical optimal transport dynamics monge1781memoire ; kantorovich1942translocation , but by the kinematics of the Schrödinger equation. In this work, we aim to bridge the conceptual and methodological gap between the “classical” SBP and the Guerra-Morato (GM) formulation, referred to as the Quantum Schrödinger Bridge Problem (QSBP), following the terminology of pavon2002quantum . Our first contribution is to reformulate the GM Lagrangian in terms of a forward-backward stochastic differential equation (SDE) system and derive the necessary conditions that the associated quantities must satisfy. This reformulation enables the adaptation of various techniques originally developed for solving the classical SBP de2021diffusion ; shi2023diffusion ; chen2021likelihood ; vargas2021solving ; tong2023simulation ; neklyudov2022action to the quantum setting. These methods typically involve learning forward and backward stochastic processes between the marginal distributions \pi_{0}(\mathbf{x}) and \pi_{1}(\mathbf{x}) using two neural networks, with a Iterative Proportional Fitting Procedure (IPFP) ruschendorf1995convergence ; ireland1968contingency ; essid2019traversing optimization procedure. However, such approaches are often computationally intensive due to the necessity of sampling the whole trajectory, which can hinder their usability in practice shi2023diffusion . To address these limitations, a complementary line of research has focused on the analytical tractability of Gaussian distributions to develop more efficient SBP solvers chewi2020gradient ; altschuler2021averaging ; korotin2023light . Exact solutions to the SBP are known in only a few cases, with the Gaussian setting being one of the most prominent mallasto2022entropy ; bunne2023schrodinger . In this work, we extend the set of exact solutions by deriving the closed-form expression of the QSBP for Gaussian marginals. Our approach differs significantly from that of bunne2023schrodinger , which employ tools from Riemannian geometry. In contrast, we adopt a Lagrangian framework and solve both the Fokker–Planck equation and the quantum Hamilton–Jacobi equation explicitly, offering a distinct theoretical perspective for addressing bridging problems with tractable distributions. In short, this work places the quantum Schrödinger Bridge on equal footing with the classical SBP in both theoretical understanding and practical usability . Specifically, we make the following contributions: • In Section 2 , we introduce the optimality conditions for the QSBP and show that they lead to the Quantum Hamilton-Jacobi Equation, whose solution is governed by the Schrödinger equation. While these results are not novel per se, we reformulate them in a manner that is suitable for generative score-matching models. • In Section 3 , we derive a closed-form solution to the QSBP for Gaussian measures. • In the subsequent sections, we extend the Gaussian Mixture Model (GMM) algorithm to our exact solution and apply it to a variety of bridging tasks, including image generation, single-cell data modeling and latent molecular properties translation. We also demonstrate an application to Mean Field Games via a variational formulation of the quantum Lagrangian. 2 From Classical to Quantum Schrödinger Bridge The goal of this section is to set up notation and motivate and formally introduce our problem setting. Let us denote by path measure any positive measure \mathbb{Q}\in M_{+}(\Sigma) , where \Sigma=C\left([0,1],X\right) is the space of all continuous paths with time-coordinate t\in[0,1] . 1 1 1 We are restricting our attention to the continuous case here. If we want to extend the definition to the discrete case as well, we need to replace \Sigma=D\left([0,1],X\right) with the space of càdlàg (right-continuous left-limited) paths. Let P(\Sigma) be the space of probability measures on \Sigma , that is, \mathbb{P}\in M_{+}(\Sigma) and \int_{\Sigma}d\mathbb{P}=1 . Given two (potentially unknown, but from which we can sample) distributions \pi_{0} and \pi_{1} on X and a reference path measure \mathbb{R}\in M_{+}(\Sigma) , the Schrödinger Bridge Problem (SBP) aims at selecting a path measure \mathbb{Q}\in M_{+}(\Sigma) such that it generates a mapping between the distributions, \mathbb{Q}_{0,1}=\pi_{0,1} , and it satisfies some notation of optimality, for instance that the relative entropy H(\mathbb{Q}|\mathbb{P})=\int_{\Sigma}\log\left(\frac{d\mathbb{Q}}{d\mathbb{P}}\right)d\mathbb{P} is minimized. It is often the case that \mathbb{P} is a reversible Markov process, for instance a linear stochastic processes bunne2023schrodinger or the reversible Brownian motion on X=\mathbb{R}^{n} leonard2013survey . A useful construction of such a path measure \mathbb{Q} that descends from diffusion models is represented in terms of a solution to a stochastic differential equation (SDE) \displaystyle\mathrm{d}\boldsymbol{x}(t) \displaystyle=\mathbf{b}_{+}(\boldsymbol{x}(t),t)\,\mathrm{d}t+\sqrt{2\beta(t)}\,\mathrm{d}\mathbf{W}(t)~, \displaystyle\boldsymbol{x}(0) \displaystyle\sim\pi_{0}~,\ \ \boldsymbol{x}(1)\sim\pi_{1} (2) where \mathbf{b}_{+}:\Sigma\rightarrow TX is the vector-valued drift coefficient. \beta(t) is the diffusion coefficient, which we assume to be \boldsymbol{x} -independent for simplicity, and \mathrm{d}\mathbf{W} is the reversible Brownian motion on X . As