68
1

Solving Quadratic Systems with Full-Rank Matrices Using Sparse or Generative Priors

Abstract

The problem of recovering a signal xRn\boldsymbol{x} \in \mathbb{R}^n from a quadratic system {yi=xAix, i=1,,m}\{y_i=\boldsymbol{x}^\top\boldsymbol{A}_i\boldsymbol{x},\ i=1,\ldots,m\} with full-rank matrices Ai\boldsymbol{A}_i frequently arises in applications such as unassigned distance geometry and sub-wavelength imaging. With i.i.d. standard Gaussian matrices Ai\boldsymbol{A}_i, this paper addresses the high-dimensional case where mnm\ll n by incorporating prior knowledge of x\boldsymbol{x}. First, we consider a kk-sparse x\boldsymbol{x} and introduce the thresholded Wirtinger flow (TWF) algorithm that does not require the sparsity level kk. TWF comprises two steps: the spectral initialization that identifies a point sufficiently close to x\boldsymbol{x} (up to a sign flip) when m=O(k2logn)m=O(k^2\log n), and the thresholded gradient descent (with a good initialization) that produces a sequence linearly converging to x\boldsymbol{x} with m=O(klogn)m=O(k\log n) measurements. Second, we explore the generative prior, assuming that x\boldsymbol{x} lies in the range of an LL-Lipschitz continuous generative model with kk-dimensional inputs in an 2\ell_2-ball of radius rr. We develop the projected gradient descent (PGD) algorithm that also comprises two steps: the projected power method that provides an initial vector with O(klogLm)O\big(\sqrt{\frac{k \log L}{m}}\big) 2\ell_2-error given m=O(klog(Lnr))m=O(k\log(Lnr)) measurements, and the projected gradient descent that refines the 2\ell_2-error to O(δ)O(\delta) at a geometric rate when m=O(klogLrnδ2)m=O(k\log\frac{Lrn}{\delta^2}). Experimental results corroborate our theoretical findings and show that: (i) our approach for the sparse case notably outperforms the existing provable algorithm sparse power factorization; (ii) leveraging the generative prior allows for precise image recovery in the MNIST dataset from a small number of quadratic measurements.

View on arXiv
Comments on this paper