Title:Convergence of the denoising diffusion probabilistic models for general noise schedules

Abstract:This paper presents a theoretical convergence analysis of a denoising diffusion probabilistic model (DDPM) in its original discrete-time formulation introduced by Ho, Jain, and Abbeel (Advances in Neural Information Processing Systems}, 33 (2020), 6840-6851). We derive an explicit upper bound for the total variation distance between the sampling distribution of the discrete-time DDPM algorithm and a given target data distribution, under general noise schedule parameters. Our analysis requires only mild regularity assumptions on the data distribution and a linear growth condition on the estimated score function. The sampling scheme is interpreted as an exponential-integrator-type approximation of a reverse-time stochastic differential equation (SDE) over a finite time horizon. Tools from the Schrödinger problem are employed to control the distributional error in reverse time and connect it to its forward-time counterpart. Moreover, the score function in DDPMs naturally appears as an adapted solution of a forward-backward SDE, providing a basis for analyzing the time-discretization error in reverse-time SDE sampling.