Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.36090

Setting Things Straight: On the Straightness and Non-Straightness of Two-Step Rectified Flow

Tuan Pham·Saptarshi Roy·Alessandro Rinaldo·Purnamrita Sarkar
View original source ↗Read source document ↗Export metadata ↓

Abstract

Rectified flow (RF) is an ODE-based generative modeling framework that has achieved remarkable success in practice over the past few years. By design, RF seeks to produce flows that have a high degree of straightness, in the sense that the resulting ODE trajectories can be obtained with relatively few (ideally, one) sampling steps. However, a rigorous analysis of straightness or near-straightness of RF flows remains missing in the literature. In this paper, we study settings in which the coupling generated by two-step RF is provably straight at the population level. We provide a complete characterization in two popular classes of data distributions: (i) smooth, strongly log-concave densities and (ii) Gaussian mixture models. Our main results yield sharp conditions for the coupling generated by two-step RF to be straight. Non-straight couplings are constructed when these conditions fail, which demonstrate the sharpness of these conditions. Using the theory developed on the straightness of two-step RF, we propose a notion of approximate straightness that measures the straightness of the learned couplings and can be estimated from the data. We describe a procedure for constructing asymptotic confidence intervals for our proposed notion of approximate straightness and propose a data-driven method for selecting the number of reflow rounds in multiple iterations of RF.

Explore related subjects

math.STstat.TH

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Tuan Pham, Saptarshi Roy, Alessandro Rinaldo, Purnamrita Sarkar. 2026-10-26. Setting Things Straight: On the Straightness and Non-Straightness of Two-Step Rectified Flow. https://arxiv.org/abs/2610.36090

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Wasserstein approximation of measures via orthogonal frames: statistical and computational complexities

Dimension-reduction methods such as PCA identify a subspace capturing most of the variation in the data, but not a distinguished basis within it. Selecting a basis in which the samples are sparsely represented is a recurring task in unsupervised learning, addressed by methods in factor analysis, dictionary learning, and subspace clustering. We pose this task as Wasserstein approximation of the data distribution by measures supported on the axes of an orthogonal frame, which, in turn, reduces to a nonsmooth optimization problem over the orthogonal group $\mathrm{O}(d)$. We show that, under mild assumptions, both empirical risk minimizers and a locally initialized stochastic gradient method recover the population-optimal frame at the near-parametric rate $\widetilde{O}_d(n^{-1/2})$, where $n$ is the sample size. The main difficulty is that, although the population objective is smooth, the empirical objective is not-this mismatch precludes standard covering arguments. We overcome this issue by replacing Lipschitz continuity of the empirical gradient with a coarsely Lipschitz bound, a tool that may be of independent interest. Numerical experiments support our theoretical findings.

math.ST↗
arXiv2026

Compatibility-based Bayesian inference: asymptotic limits, variational principles, and Wasserstein geometry

Approximate Bayesian Computation (ABC) replaces intractable likelihood evaluations with comparisons between observed and simulated data. We formulate discrepancy-based inference through nonnegative compatibility weights. Prior-integrated convergence of rescaled weights gives total variation convergence as the tolerance vanishes. For a fixed positive tolerance and increasing sample size, we give a general representation-based formulation of the conditional-prior limit known from ABC and coarsened Bayesian inference, and establish convergence in expected total variation under stabilization in probability and a prior-null boundary condition. We also derive acceptance-persistence bounds under partial stabilization. The Gibbs variational identity yields convergence of minimum values and stability of approximate minimizers; the fixed-tolerance limit is characterized separately by constrained relative-entropy minimization. For additional penalty-based functionals of the limiting Wasserstein discrepancy, we prove an exact reduction to distributions over model laws and, under a geodesic assumption, a representation by distances from a Wasserstein ball. These penalty-based updates are distinguished from the original fixed-tolerance ABC limit and illustrated through a Gaussian comparison.

math.ST↗
arXiv2026

A theory of stratification learning

Given i.i.d. sample from a stratified mixture of immersed manifolds of different dimensions, we study the minimax estimation of the underlying stratified structure. We provide a constructive algorithm allowing to estimate each mixture component at its optimal dimension-specific rate adaptively. The method is based on an ascending hierarchical co-detection of points belonging to different layers, which also identifies the number of layers and their dimensions, assigns each data point to a layer accurately, and estimates tangent spaces optimally. These results hold regardless of any ambient assumption on the manifolds or on their intersection configurations. They open the way to a broad clustering framework, where each mixture component models a cluster emanating from a specific nonlinear correlation phenomenon.

math.ST↗