Abstract
We introduce a novel framework for efficient sampling from complex,unnormalised target distributions by exploiting multiscale dynamics.Traditional score-based sampling methods either rely on learned approximationsof the score function or involve computationally expensive nested Markov chainMonte Carlo (MCMC) loops. In contrast, the proposed approach leveragesstochastic averaging within a slow-fast system of stochastic differentialequations (SDEs) to estimate intermediate scores along a diffusion path withouttraining or inner-loop MCMC. Two algorithms are developed under this framework:MultALMC, which uses multiscale annealed Langevin dynamics, and MultCDiff,based on multiscale controlled diffusions for the reverse-timeOrnstein-Uhlenbeck process. Both overdamped and underdamped variants areconsidered, with theoretical guarantees of convergence to the desired diffusionpath. The framework is extended to handle heavy-tailed target distributionsusing Student's t-based noise models and tailored fast-process dynamics.Empirical results across synthetic and real-world benchmarks, includingmultimodal and high-dimensional distributions, demonstrate that the proposedmethods are competitive with existing samplers in terms of accuracy andefficiency, without the need for learned models.