An Eulerian Perspective on Straight-Line Sampling

We study dynamic measure transport for generative modeling: specifically, flows induced by stochastic processes that bridge a specified source and target distribution. The conditional expectation of the process’ velocity defines an ODE whose flow map achieves the desired transport. We ask \emph{which processes produce straight-line flows} – i.e., flows whose pointwise acceleration vanishes and thus are exactly integrable with a first-order method? We provide a concise PDE characterization of straightness as a balance between conditional acceleration and the divergence of a weighted covariance (Reynolds) tensor. Using this lens, we fully characterize affine-in-time interpolants and show that straightness occurs exactly under deterministic endpoint couplings. We also derive necessary conditions that constrain flow geometry for general processes, offering broad guidance for designing transports that are easier to integrate.

💡 Research Summary

The paper “An Eulerian Perspective on Straight-Line Sampling” addresses a fundamental challenge in generative modeling: how to design stochastic processes that induce the most efficient transport trajectories between a source and a target distribution. In the context of continuous-time generative models, such as Flow Matching and Diffusion models, the computational cost of sampling is heavily dictated by the curvature of the learned trajectories. If the trajectories are straight, simple first-order numerical integrators like the Euler method can achieve high accuracy with minimal steps. However, curved trajectories necessitate complex, high-order ODE solvers, significantly increasing inference latency.

The authors approach this problem through an Eulerian lens, focusing on the evolution of the velocity field rather than individual particle trajectories. The core contribution of the paper is the derivation of a precise PDE characterization of “straightness.” The researchers demonstrate that for a flow to be straight (i.e., to have vanishing pointwise acceleration), there must be a delicate mathematical balance between the conditional acceleration of the process and the divergence of a weighted covariance tensor, which is analogous to the Reynolds stress tensor in fluid dynamics. This implies that the geometric curvature of the flow is intrinsically linked to the spatial evolution of the process’s variance.

Furthermore, the paper provides a rigorous analysis of affine-in-time interpolants, a class of widely used paths in modern generative modeling. The authors prove that straightness is achieved exactly under deterministic endpoint couplings, providing a theoretical justification for the efficiency of deterministic flow matching. Beyond specific cases, the paper also derives necessary conditions that constrain the geometry of flows for more general stochastic processes.

In summary, this work moves beyond empirical heuristic-based path design and provides a rigorous mathematical framework for understanding and engineering straight-line flows. By characterizing the relationship between stochasticity and flow curvature, the paper offers essential guidance for the development of next-generation generative models that prioritize both sampling fidelity and computational efficiency. This theoretical advancement is crucial for scaling generative AI to real-time applications where rapid and accurate sampling is paramount.