arXiv Science⌕ Search

arXiv · 2611.17625

Finite speed Ornstein-Uhlenbeck-Kac processes and their applications in generative modeling

Abstract

We study the Ornstein-Uhlenbeck-Kac (OU-K) process, where instead of a Brownian motion, a finite-speed Kac process defines the driving process. The OU-K process can be analyzed through a rescaled, time-inhomogeneous Kac process, characterized by a switching rate from the Kac part, and a convergence rate from the Ornstein-Uhlenbeck (OU) part. The law of this inhomogeneous Kac process describes the solution of a wave equation, the Euler-Poisson-Darboux (EPD) equation. For Kac switching rates being multiples of the OU convergence rate, we can explicitly derive this EPD solution via the sum of (the derivative of) two Legendre polynomials, restricted between two Dirac points representing the wave fronts. Using the framework of piecewise-deterministic Markov processes (PDMP), we describe the conditional velocity fields of the OU-K process via our polynomial-law-formula. This allows us to apply the OU-K process in a generative flow model. Since the OU-K velocity fields prove to be bounded by the finite Kac speed, the usual time truncations of diffusion models become unnecessary. We demonstrate advantages over diffusion and flow matching models in terms of generation sharpness on singular data sets, while achieving competitive results on CIFAR-10 image generation.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Richard Duong, Janis Auffenberg, Gabriele Steidl. 2026-11-13. Finite speed Ornstein-Uhlenbeck-Kac processes and their applications in generative modeling. https://arxiv.science/abs/2611.17625

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite arXiv Science by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

Approximation of the Lévy-driven stochastic heat equation on the sphere

The stochastic heat equation on the sphere driven by an additive square-integrable Lévy process is approximated by a spectral method in space and forward and backward Euler-Maruyama schemes in time. New sharp regularity results are proven for its solution. The spectral approximation is based on a truncation of the series expansion with respect to the spherical harmonic functions. For a given regularity of the initial condition and two different settings of regularity for the driving noise, strong convergence rates for the spectral approximation and for the Euler-Maruyama methods are proven. Moreover, weak rates of up to twice the strong rates are shown for Wiener noise and for Lévy noise with independent and up to a deterministic decay rate factor identically distributed components, while for general square-integrable Lévy noise the weak rate depends explicitly on the polynomial order of the test function. Numerical simulations confirm the theoretical results.

math.PR↗

A Multivariate Generalized Cox Model for Dependent and Simultaneous Default Times

We propose a framework for dependent default times where each name defaults when an increasing driver crosses an independent unit exponential threshold. As in Gueye and Jeanblanc (2022), the drivers may jump, so defaults may occur at stopping times of the reference filtration, including predictable ones, and a common jump may cause no, one or several simultaneous defaults. For general drivers, we derive the conditional joint survival law by a backward recursion; it is an explicit product of survival functions if and only if, for every set of names, the predictable reduction of the compensator of their first default time is deterministic. If, moreover, defaults avoid predictable stopping times, the law is a fatal shock law with hazards identified from the drivers; predictable default atoms can rule this out. Examples include Lévy subordinators and, with random reductions but explicit laws, Liu's model and shot-noise drivers with memory.

math.PR↗

Concentration Inequalities for Branching Random Walks

Motivated by phase transition problems in CSPs, we prove a more general concentration inequality that retains classical sub-Gaussian tails under a mild global linear-growth condition $|S_n| < Cn$, relaxing the bounded-increment assumption to finite exponential moments and requiring neither independence nor the martingale property. We further extend it to branching random walks (BRWs), obtaining the first concentration inequality for BRWs.

math.PR↗