Download:
Abstract:
How can we optimally denoise from noisy measurements to recover the signal distribution without knowing the noise distribution, only the noise level? We revisit this challenging denoising problem in multi-dimensions: independent general noise $Z$ is added to corrupt the signal $X$, resulting in the noisy measurement $Y = X + \sigma Z$, where $\sigma $ is a small known constant. Our goal is to recover the underlying signal distribution $P_X$ based on the distribution of noisy measurements $P_Y$. We propose and analyze new universal denoisers that are agnostic to a range of signal and noise distributions. These denoisers achieve optimal distributional matching with order-of-magnitude improvements over the Bayes-optimal denoiser derived from Tweedie’s formula. These denoisers shrink the distribution $P_Y$ towards $P_X$ with accuracy of $O(\sigma^4)$ and $O(\sigma^6)$, respectively, both in terms of matching generalized moments and densities. Our denoisers are inspired by optimal transport theory, are optimal in the sense that they approximate the Monge-Ampère equation with higher-order accuracy, and can be efficiently implemented using score matching techniques.
Citation
Tengyuan Liang. 2025. “Distributional Shrinkage I: Universal Denoisers in Multi-Dimensions.” arXiv:2511.xxxxx.