Abstract
Score-based diffusion models have become a foundational paradigm for moderngenerative modeling, demonstrating exceptional capability in generating samplesfrom complex high-dimensional distributions. Despite the dominant adoption ofprobability flow ODE-based samplers in practice due to their superior samplingefficiency and precision, rigorous statistical guarantees for these methodshave remained elusive in the literature. This work develops the firstend-to-end theoretical framework for deterministic ODE-based samplers thatestablishes near-minimax optimal guarantees under mild assumptions on targetdata distributions. Specifically, focusing on subgaussian distributions with$\beta$-H\"older smooth densities for $\beta\leq 2$, we propose a smoothregularized score estimator that simultaneously controls both the $L^2$ scoreerror and the associated mean Jacobian error. Leveraging this estimator withina refined convergence analysis of the ODE-based sampling process, wedemonstrate that the resulting sampler achieves the minimax rate in totalvariation distance, modulo logarithmic factors. Notably, our theorycomprehensively accounts for all sources of error in the sampling process anddoes not require strong structural conditions such as density lower bounds orLipschitz/smooth scores on target distributions, thereby covering a broad rangeof practical data distributions.