perturbation bounds for low-rank inverse approximations under noise
TL;DR¶
This paper provides the first non-asymptotic spectral norm perturbation bound for low-rank inverse approximations \(\|(\tilde{A}^{-1})_p - A_p^{-1}\|\) under additive noise. Using contour integration techniques, it derives sharp bounds depending on the eigenvalue gap, spectral decay, and noise alignment, improving upon classical full-inverse bounds by up to a factor of \(\sqrt{n}\).
Background & Motivation¶
Background: Low-rank pseudoinverses are widely used to accelerate inverse matrix approximations in ML, optimization, and scientific computing. Classical perturbation theory provides bounds only for the full inverse \(\|A^{-1} - \tilde{A}^{-1}\|\).
Key Challenge: Classical bounds ignore the structure introduced by truncation to rank \(p\), overlook the interaction between noise and eigenvalue gaps, and frequently yield overly pessimistic estimates.
Core Idea: Contour integration is applied to perform a localized resolvent expansion for the non-entire function \(f(z) = 1/z\), precisely controlling the perturbation of the Riesz projection near the smallest eigenvalue.
Method¶
Problem Formulation¶
Given an \(n \times n\) real symmetric positive definite matrix \(A\) with eigenvalues \(\lambda_1 \geq \cdots \geq \lambda_n > 0\) and a symmetric perturbation \(E\), let \(\tilde{A} = A + E\). The central quantity of interest is \(\|(\tilde{A}^{-1})_p - A_p^{-1}\|\), where \(A_p^{-1} = \sum_{i=n-p+1}^{n} \lambda_i^{-1} u_i u_i^\top\) is the best rank-\(p\) approximation to \(A^{-1}\).
Key Results¶
-
Main Theorem (Theorem 2.1): Under the condition \(4\|E\| \leq \min\{\lambda_n, \delta_{n-p}\}\): $\(\|(\tilde{A}^{-1})_p - A_p^{-1}\| \leq 5\left(\frac{\|E\|}{\lambda_n^2} + \frac{\|E\|}{\lambda_{n-p} \delta_{n-p}}\right)\)$ where \(\delta_{n-p} = \lambda_{n-p} - \lambda_{n-p+1}\) denotes the eigenvalue gap.
-
Comparison with Classical Bounds: The classical EYM-N bound takes the form \(\frac{8\|E\|}{3\lambda_n^2} + \frac{2}{\lambda_{n-p}}\), whose second term does not vanish as \(\|E\| \to 0\)—yielding nonzero error even in the noiseless case. The proposed bound correctly tends to zero as \(\|E\| \to 0\).
Technical Contributions¶
- Extension of Contour Integration: Extended from entire functions (e.g., the matrix exponential \(e^z\)) to the non-entire function \(f(z) = 1/z\).
- Localized Resolvent Expansion: A Cauchy contour is constructed around the smallest eigenvalue, exploiting local spectral structure.
- Riesz Projection Perturbation Control: Precise control over the effect of noise on eigenspace projections associated with low-curvature subspaces.
- Extension to General Symmetric Matrices: Covers rank-deficient cases via pseudoinverses.
Application: Improved PCG Convergence Rates¶
The proposed bounds yield tighter condition number estimates for PCG methods with low-rank-plus-regularization preconditioners, provably improving iteration count guarantees by a factor of \(n^{1/4}\).
Key Experimental Results¶
Main Results¶
| Matrix Type | Tracking Error (Ours) | Tracking Error (Classical) |
|---|---|---|
| Sample covariance | Small constant factor | 1–2 orders of magnitude overestimated |
| Discretized elliptic operator | Small constant factor | 1–2 orders of magnitude overestimated |
| BCSSTK09 stiffness | Small constant factor | 1–2 orders of magnitude overestimated |
Noise Condition Feasibility¶
Validated on the 1990 US Census covariance matrix and the BCSSTK09 stiffness matrix; safety margins comfortably exceed typical noise levels encountered in differential privacy and structural engineering applications.
Highlights & Insights¶
- Low-rank inverse approximations can be more robust than full inverses when the eigenvalue gap is sufficiently large—a counterintuitive finding, as truncation is typically assumed to introduce additional error.
- Spectrally aware perturbation bounds are far more informative in practice than worst-case bounds, as they exploit the spectral structure of the matrix.
- PCG convergence rates can be improved by a factor of \(n^{1/4}\) using the proposed bounds, yielding direct algorithmic benefits.
- Extending contour integration from entire to non-entire functions constitutes a non-trivial mathematical contribution.
Limitations & Future Work¶
- The noise condition \(4\|E\| < \min\{\lambda_n, \delta_{n-p}\}\) may be restrictive in certain settings.
- Only the spectral norm is considered; Frobenius norm bounds remain to be developed.
- Structured noise (e.g., low-rank perturbations) may admit tighter bounds.
Related Work & Insights¶
- vs. Classical Neumann Series Bounds: Classical bounds do not account for low-rank truncation, and the second term \(2/\lambda_{n-p}\) does not vanish as \(\|E\| \to 0\).
- vs. Schatten Norm Results: Existing work does not provide spectral norm guarantees.
- vs. Sherman–Morrison–Woodbury: Applicable only to structured low-rank updates, not to general additive noise.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ First non-asymptotic spectral norm bound for low-rank inverse perturbations, filling an important gap.
- Experimental Thoroughness: ⭐⭐⭐⭐ Validated on diverse real and synthetic matrices with high tracking accuracy.
- Writing Quality: ⭐⭐⭐⭐⭐ Rigorous and fluent exposition with clearly structured proofs.
- Value: ⭐⭐⭐⭐⭐ Foundational contribution to numerical computing and ML optimization.