Gap-Free Differentially Private PCA for Gaussian Data

Forum topic · 小凯 · 2026-09-29

Summary

Researchers Alina Ene and Huy L. Nguyen present a gap-free differentially private algorithm for the principal component analysis (PCA) problem on Gaussian data. The paper, posted to arXiv as 2609.31614, addresses a core challenge in privacy-preserving machine learning: computing principal components while protecting individual data records under differential privacy. The gap-free property means the algorithm's accuracy guarantee does not depend on a preexisting spectral gap assumption—previously, many private PCA results required the underlying covariance matrix to have a well-separated top eigenvalue to guarantee utility. By removing this assumption, the proposed method extends reliable differentially private PCA to settings where the data covariance matrix has nearly degenerate leading eigenvalues. The work is relevant to researchers in differential privacy, unsupervised learning, and statistical estimation, as PCA is a foundational building block in data analysis pipelines. The arXiv listing is available at https://arxiv.org/abs/2609.31614, and the post was collected automatically by zhichai.net on 2026-09-29.

Paper Overview

Field: Machine Learning Authors: Alina Ene, Huy L. Nguyen Posted: 2026-09-25 arXiv: 2609.31614

Abstract

We give a gap-free differentially private algorithm for the principal component analysis (PCA) problem with Gaussian data.

Context

The key contribution is the gap-free property: unlike earlier differentially private PCA results that require a spectral gap (a well-separated top eigenvalue of the covariance matrix) to achieve good utility, this algorithm works without any such separation assumption. This makes private PCA applicable to Gaussian datasets whose leading eigenvalues may be nearly degenerate.

*Collected automatically on 2026-09-29.*