Abstract
We theoretically analyze the original version of the denoising diffusionprobabilistic models (DDPMs) presented in Ho, J., Jain, A., and Abbeel, P.,Advances in Neural Information Processing Systems, 33 (2020), pp. 6840-6851.Our main theorem states that the sequence constructed by the original DDPMsampling algorithm weakly converges to a given data distribution as the numberof time steps goes to infinity, under some asymptotic conditions on theparameters for the variance schedule, the $L^2$-based score estimation error,and the noise estimating function with respect to the number of time steps. Inproving the theorem, we reveal that the sampling sequence can be seen as anexponential integrator type approximation of a reverse time stochasticdifferential equation.