Deterministic Decomposition of Stochastic Generative Dynamics
Xingyu Song, Yuan Mei, Naoya Takeishi · 2026
Read Deterministic Decomposition of Stochastic Generative Dynamics by Xingyu Song, Yuan Mei, Naoya Takeishi (2026) 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-2605-08794.txt.
Introduction
Xingyu Song, Yuan Mei, and Naoya Takeishi split the driving field of stochastic generative models into two interpretable parts. Generative systems transport probability from a simple start to real data; deterministic versions use one velocity field, while stochastic versions mix drift and diffusion into that field until their roles blur. The authors prove the field decomposes naturally into a transport part that moves the distribution and an osmotic part created by diffusion and set by the distribution's score.
Their Bridge Matching framework learns both parts through marginal and conditional training. Recombining them with an adjustable osmotic weight gives controllable, interpretable sampling. Generative modeling gains a dial for how much randomness shapes the path from noise to data.
Catalog record
| CREATOR | Xingyu Song, Yuan Mei, Naoya Takeishi |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 93 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.08794
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Deterministic Decomposition of Stochastic Generative Dynamics Xingyu Song, Yuan Mei, Naoya Takeishi Abstract Modern generative models can be understood as probability transport from a simple base distribution to a target data distribution. Deterministic transport models offer tractable velocity-field parameterizations, whereas stochastic generative models capture richer density evolution through drift and diffusion. Yet when stochastic dynamics are described through deterministic velocity fields, the effects of drift and diffusion are often compressed into a single effective field, obscuring the distinct roles of deterministic evolution and stochastic fluctuation. In this work, we show that the deterministic field b t b_{t} of a stochastic generative process admits a natural transport–osmotic decomposition that separates deterministic transport from stochastic, diffusion-induced effects: b t = u t + d t b_{t}=u_{t}+d_{t} , where u t u_{t} governs marginal probability transport and d t d_{t} captures an osmotic effect induced by diffusion and determined by the marginal score. Based on this decomposition, we propose Bridge Matching, a flow-based framework for learning decomposed generative dynamics through both marginal and conditional formulations. In generative modeling experiments, we recombine the learned components as b t = u t + λ d d t b_{t}=u_{t}+\lambda_{d}d_{t} , showing that the proposed decomposition enables interpretable and controllable sampling by adjusting the osmotic contribution in probability transport. The implementation is publicly available here . 1 Introduction Generative modeling has become a central paradigm in modern machine learning, enabling the synthesis of high-dimensional data across domains such as vision ( 41 ; 10 ) , audio ( 28 ; 11 ) , and scientific simulation ( 13 ; 42 ) . A unifying view of modern generative modeling is probability transport ( 24 ; 25 ; 1 ; 29 ) : samples are generated by evolving a simple base distribution \pi_{0} into a target data distribution \pi_{1} . A key distinction among these approaches lies in the dynamics governing the evolution of marginal distributions. In deterministic formulations, samples evolve according to an ordinary differential equation (ODE) defined by a velocity field ( 3 ; 16 ) , so that the associated marginal density evolves according to the continuity equation ( 47 ) . This makes the learned dynamics relatively tractable: the velocity field directly characterizes the transport of probability mass over time, and can be learned by regression-based objectives ( 25 ; 46 ) . In stochastic formulations, by contrast, samples evolve according to stochastic differential Equations (SDEs), while their marginal densities are governed by the Fokker–Planck equation ( 2 ; 35 ; 43 ) . Here, density evolution is determined jointly by drift and diffusion, making stochastic dynamics expressive but also more difficult to interpret and control. Nevertheless, deterministic and stochastic models should not be viewed as separate families. Instead, several existing perspectives connect stochastic generative dynamics to deterministic dynamics or deterministic representations. For example, probability flow ODEs define deterministic dynamics whose marginal distributions match those of the corresponding stochastic diffusion process ( 43 ; 4 ) . Stochastic interpolants and Schrödinger-bridge-based methods further emphasize pathwise and bidirectional structures, connecting endpoint distributions through stochastic or deterministic interpolating dynamics ( 1 ; 5 ; 24 ; 7 ) . Related deterministic transport formulations construct velocity fields that realize prescribed probability paths through the continuity equation ( 25 ; 29 ) . Other works improve or regularize learned generative trajectories through path straightening or curvature-based criteria ( 22 ; 23 ) . Together, these connections show that stochastic and deterministic generative models are closely related at the level of marginal evolution. However, existing deterministic reformulations of stochastic dynamics typically represent marginal evolution with a single effective velocity field. Although this enables deterministic sampling and efficient training, it collapses the mechanisms underlying stochastic dynamics into one representation. In particular, probability-mass transport and diffusion-induced density reshaping are absorbed into the same marginal velocity, making it difficult to analyze, control, or recombine these effects independently. Consequently, even when deterministic dynamics reproduce the same marginal distributions, the contribution of stochasticity remains hidden. In this work, we show that stochastic generative dynamics admit a more structured deterministic representation. Our key observation is that stochastic processes exhibit an intrinsic bidirectional structure: the forward and backward drift fields associated with the same marginal path are coupled through the marginal score. This coupling naturally induces a Transport–Osmotic decomposition into two complementary deterministic vector fields as shown in Fig. 1 . The transport field governs the movement of probability mass and determines marginal evolution through a continuity equation, whereas the osmotic field captures diffusion-induced, density-dependent effects. In this view, stochasticity is not treated merely as an external noise source, but is represented as a structured component of deterministic marginal dynamics. Building on this decomposition, we propose Bridge Matching , a framework for learning decomposed generative dynamics Rather than learning a single effective velocity field directly, Bridge Matching learns separate transport and osmotic (density-shaping) fields that can be recombined to construct forward or backward generative dynamics. We develop both marginal and conditional formulations: the marginal formulation follows the distribution-level definition of the decomposition, while the conditional formulation provides a tractable implementation based on conditional probability paths with analytically available scores. This enables the decomposition to be learned with regression-style objectives similar to Flow Matching, while preserving the structural distinction between transport and diffusion-induced effects. Figure 1 : Overview of the proposed Transport–Osmotic decomposition. Dynamics from source sample x_{0} to target sample x_{1} are decomposed into transport field u_{t} and osmotic field d_{t} over the marginal distribution \pi_{t} . Smooth and noisy curves denote deterministic and stochastic trajectories. Contributions. 1. We derive a Transport–Osmotic decomposition of stochastic generative dynamics, showing that deterministic and stochastic models share a common marginal transport backbone while differing in how diffusion-induced effects enter the sample dynamics. 2. We propose Bridge Matching , a training framework that learns the transport and osmotic fields through marginal and conditional formulations, connecting score-based information with flow-style regression objectives. 3. We empirically show that recombining the learned fields improves generation performance and enables controllable, interpretable sampling by adjusting the relative strength of transport and density-dependent effects. 2 Generative modeling via probability transport In the data space \mathcal{X}\subseteq\mathbb{R}^{d} , a generative model defines a parametric probability distribution p_{\theta}(x) with learnable parameters \theta\in\Theta . The goal is to approximate an unknown data distribution p_{\text{data}}(x) such that p_{\theta}(x)\approx p_{\text{data}}(x) . A unifying perspective formulates generative modeling as a probability transport problem ( 47 ; 24 ) , where samples from a simple base distribution \pi_{0} are progressively transformed into samples from a target data distribution \pi_{1} ( 34 ; 40 ; 25 ) . This transformation is described by a continuous family of intermediate distributions \{\pi_{t}\}_{t\in[0,1]} , with boundary conditions \pi_{0} and \pi_{1} at t=0 and t=1 . From this perspective, generative modeling paradigms differ in how they specify the probability path and in the sample-level dynamics used to realize it. More specifically, these dynamics may be deterministic, as in continuous normalizing flows and flow matching, or stochastic, as in diffusion and score-based models ( 3 ; 43 ; 16 ) . 2.1 Deterministic transport and the continuity equation Following the probability transport view, deterministic transport models realize a distribution-level path \{\pi_{t}\}_{t\in[0,1]} through sample-level deterministic dynamics. Here, \pi_{t} denotes the