1: \begin{abstract}

2:     \acp{dm} as generative priors have recently shown great potential for denoising tasks but lack theoretical understanding with respect to their \ac{mse} optimality. This paper proposes a novel denoising strategy inspired by the structure of the \ac{mse}-optimal \ac{cme}. The resulting \ac{dm}-based denoiser can be conveniently employed using a pre-trained \ac{dm}, being particularly fast by truncating reverse diffusion steps and not requiring stochastic re-sampling. 

3:     We present a comprehensive (non-)asymptotic optimality analysis of the proposed diffusion-based denoiser, demonstrating polynomial-time convergence to the \ac{cme} under mild conditions.

4:     Our analysis also derives a novel Lipschitz constant that depends solely on the \ac{dm}'s hyperparameters. 

5:     Further, we offer a new perspective on \acp{dm}, showing that they inherently combine an asymptotically optimal denoiser with a powerful generator, modifiable by switching re-sampling in the reverse process on or off.

6:     The theoretical findings are thoroughly validated with experiments based on various benchmark datasets.

7: \end{abstract}

8: