Spectral Perturbation Bounds for Low-Rank Approximation with Applications to Privacy
By: Phuc Tran, Nisheeth K. Vishnoi, Van H. Vu
Potential Business Impact:
Protects private data while finding important patterns.
A central challenge in machine learning is to understand how noise or measurement errors affect low-rank approximations, particularly in the spectral norm. This question is especially important in differentially private low-rank approximation, where one aims to preserve the top-$p$ structure of a data-derived matrix while ensuring privacy. Prior work often analyzes Frobenius norm error or changes in reconstruction quality, but these metrics can over- or under-estimate true subspace distortion. The spectral norm, by contrast, captures worst-case directional error and provides the strongest utility guarantees. We establish new high-probability spectral-norm perturbation bounds for symmetric matrices that refine the classical Eckart--Young--Mirsky theorem and explicitly capture interactions between a matrix $A \in \mathbb{R}^{n \times n}$ and an arbitrary symmetric perturbation $E$. Under mild eigengap and norm conditions, our bounds yield sharp estimates for $\|(A + E)_p - A_p\|$, where $A_p$ is the best rank-$p$ approximation of $A$, with improvements of up to a factor of $\sqrt{n}$. As an application, we derive improved utility guarantees for differentially private PCA, resolving an open problem in the literature. Our analysis relies on a novel contour bootstrapping method from complex analysis and extends it to a broad class of spectral functionals, including polynomials and matrix exponentials. Empirical results on real-world datasets confirm that our bounds closely track the actual spectral error under diverse perturbation regimes.
Similar Papers
Perturbation Bounds for Low-Rank Inverse Approximations under Noise
Machine Learning (CS)
Makes computer math work better with messy data.
On Purely Private Covariance Estimation
Machine Learning (CS)
Keeps data private while still learning from it.
Structured Approximation of Toeplitz Matrices and Subspaces
Information Theory
Fixes broken data using math tricks.