Sparse Signal Recovery From Quadratic Systems with Full-Rank Matrices

TL;DR

This paper introduces a novel two-stage algorithm for sparse signal recovery from quadratic measurements, providing theoretical guarantees and demonstrating superior efficiency and accuracy over existing methods.

Contribution

It develops a new Sparse Gauss-Newton algorithm with theoretical recovery guarantees and improved computational efficiency for high-dimensional sparse quadratic systems.

Findings

Theoretical guarantees for sparse quadratic system recovery with minimal measurements.

SGN algorithm achieves quadratic convergence and outperforms existing methods.

Numerical results show SGN requires fewer measurements and iterations for high sparsity levels.

Abstract

In signal processing and data recovery, reconstructing a signal from quadratic measurements poses a significant challenge, particularly in high-dimensional settings where measurements $m$ is far less than the signal dimension $n$ (i.e., $m \ll n$). This paper addresses this problem by exploiting signal sparsity. Using tools from algebraic geometry, we derive theoretical recovery guarantees for sparse quadratic systems, showing that $m\ge 2s$ (real case) and $m\ge 4s-2$ (complex case) generic measurements suffice to uniquely recover all $s$-sparse signals. Under a Gaussian measurement model, we propose a novel two-stage Sparse Gauss-Newton (SGN) algorithm. The first stage employs a support-restricted spectral initialization, yielding an accurate initial estimate with $m=O(s^2\log{n})$ measurements. The second stage refines this estimate via an iterative hard-thresholding Gauss-Newton method, achieving quadratic convergence to the true signal within finitely many iterations when $m\ge O(s\log{n})$. Compared to existing second-order methods, our algorithm achieves near-optimal sampling complexity for the refinement stage without requiring resampling. Numerical experiments indicate that SGN significantly outperforms state-of-the-art algorithms in both accuracy and computational efficiency. In particular, (1) when sparsity level $s$ is high, compared with existing algorithms, SGN can achieve the same success rate with fewer measurements. (2) SGN converges with only about $1/10$ iterations of the best existing algorithm and reach lower relative error.