Optimal Differentially Private Sampling of Unbounded Gaussians
Abstract
We provide the first $\widetilde{\mathcal{O}}(d)$-sample algorithm for sampling from unbounded Gaussian distributions under the constraint of $(\varepsilon, \delta)$-differential privacy. This is a quadratic improvement over previous results for the same problem, settling an open question of Ghazi, Hu, Kumar, and Manurangsi.
Cite
Text
Iverson et al. "Optimal Differentially Private Sampling of Unbounded Gaussians." Proceedings of Thirty Eighth Conference on Learning Theory, 2025.Markdown
[Iverson et al. "Optimal Differentially Private Sampling of Unbounded Gaussians." Proceedings of Thirty Eighth Conference on Learning Theory, 2025.](https://mlanthology.org/colt/2025/iverson2025colt-optimal/)BibTeX
@inproceedings{iverson2025colt-optimal,
title = {{Optimal Differentially Private Sampling of Unbounded Gaussians}},
author = {Iverson, Valentio and Kamath, Gautam and Mouzakis, Argyris},
booktitle = {Proceedings of Thirty Eighth Conference on Learning Theory},
year = {2025},
pages = {2893-2941},
volume = {291},
url = {https://mlanthology.org/colt/2025/iverson2025colt-optimal/}
}