Abstract:The problem of recovering a signal $\boldsymbol x\in \mathbb{R}^n$ from a quadratic system $\{y_i=\boldsymbol x^\top\boldsymbol A_i\boldsymbol x,\ i=1,\ldots,m\}$ with full-rank matrices $\boldsymbol A_i$ frequently arises in applications such as unassigned distance geometry and sub-wavelength imaging. With i.i.d. standard Gaussian matrices $\boldsymbol A_i$, this paper addresses the high-dimensional case where $m\ll n$ by incorporating prior knowledge of $\boldsymbol x$. First, we consider a $k$-sparse $\boldsymbol x$ and introduce the thresholded Wirtinger flow (TWF) algorithm that does not require the sparsity level $k$. TWF comprises two steps: the spectral initialization that identifies a point sufficiently close to $\boldsymbol x$ (up to a sign flip) when $m=O(k^2\log n)$, and the thresholded gradient descent which, when provided a good initialization, produces a sequence linearly converging to $\boldsymbol x$ with $m=O(k\log n)$ measurements. Second, we explore the generative prior, assuming that $x$ lies in the range of an $L$-Lipschitz continuous generative model with $k$-dimensional inputs in an $\ell_2$-ball of radius $r$. With an estimate correlated with the signal, we develop the projected gradient descent (PGD) algorithm that also comprises two steps: the projected power method that provides an initial vector with $O\big(\sqrt{\frac{k \log L}{m}}\big)$ $\ell_2$-error given $m=O(k\log(Lnr))$ measurements, and the projected gradient descent that refines the $\ell_2$-error to $O(\delta)$ at a geometric rate when $m=O(k\log\frac{Lrn}{\delta^2})$. 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.

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is: recovering the signal $x\in\mathbb{R}^n$ from the quadratic system $\{y_i = x^\top A_i x, i = 1,\ldots,m\}$, where $A_i$ are full - rank matrices. Specifically, the paper focuses on how to effectively recover the signal $x$ by using prior knowledge (such as sparsity or generative - model priors) in the high - dimensional case (i.e., $m\ll n$). ### Specific Problem Description 1. **Sparse Quadratic System**: - The author considered the $k$-sparse signal $x$ and introduced an algorithm named "Thresholded Wirtinger Flow" (TWF). - The TWF algorithm is divided into two steps: - **Spectral Initialization**: When $m = O(k^2\log n)$, it can find an initial point that is close enough to $x$ (up to a sign flip). - **Thresholded Gradient Descent**: Given a good initialization condition, when $m=O(k\log n)$, it can linearly converge to $x$. 2. **Quadratic System with Generative - Model Priors**: - Assume that the signal $x$ is within the range of an $L$-Lipschitz continuous generative model, with an input of $k$ - dimensional and within an $\ell_2$ - ball with a radius of $r$. - The author developed an algorithm named "Projected Gradient Descent" (PGD), which is also divided into two steps: - **Projected Power Method**: When $m = O(k\log(Lnr))$, it provides an initial vector with an $\ell_2$ error of $O\left(\sqrt{\frac{k\log L}{m}}\right)$. - **Projected Gradient Descent**: When $m = O\left(\frac{k\log Lrn}{\delta^2}\right)$, it refines the $\ell_2$ error to $O(\delta)$ at a geometric rate. ### Application Background These quadratic systems appear in many practical applications, such as: - **Unassigned Distance Geometry Problem**: Reconstruct the relative positions of a set of points on a line or a ring. - **Sub - wavelength Imaging**: Recover the sub - wavelength features of an object using partially spatially incoherent light. - **Cryo - EM (Cryo - Electron Microscopy)**: Reconstruct the three - dimensional density map of a protein from noisy two - dimensional projections with unknown viewing directions. ### Main Contributions 1. **Sparse Quadratic System**: - Proposed the TWF algorithm, which does not need to know the sparsity level $k$ in advance and significantly outperforms existing provable algorithms (such as Sparse Power Factorization, SPF) in experiments. 2. **Quadratic System with Generative - Model Priors**: - Proposed the PGD algorithm, which solves this specific problem for the first time and shows the ability to accurately recover images from a small number of quadratic measurements on the MNIST dataset. Through these methods, the paper provides new solutions for signal recovery in high - dimensional quadratic systems and demonstrates their effectiveness in practical applications.

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

Convergence of Projected Subgradient Method with Sparse or Low-Rank Constraints

Sparse Bayesian Learning Approach for Discrete Signal Reconstruction

Projected Gradient Descent Algorithms for Solving Nonlinear Inverse Problems with Generative Priors

Sparse Cholesky Factorization for Solving Nonlinear PDEs via Gaussian Processes

In the recovery of sparse vectors from quadratic measurements, the presence of linear terms breaks the square root bottleneck

An Oracle Gradient Regularized Newton Method for Quadratic Measurements Regression

Generative Principal Component Analysis

Uncertainty quantification for sparse Fourier recovery

Solving Dense Linear Systems Faster Than via Preconditioning

Non-convex matrix sensing: Breaking the quadratic rank barrier in the sample complexity

Solving Random Quadratic Systems of Equations Is Nearly as Easy as Solving Linear Systems

High-precision randomized iterative methods for the random feature method

Cross Low-Dimension Pursuit for Sparse Signal Recovery from Incomplete Measurements Based on Permuted Block Diagonal Matrix.

A Computational Study of Using Black-box QR Solvers for Large-scale Sparse-dense Linear Least Squares Problems

Low-rank matrix recovery with non-quadratic loss: projected gradient method and regularity projection oracle

Faster Linear Systems and Matrix Norm Approximation via Multi-level Sketched Preconditioning

A Fast Algorithm for Sparse Reconstruction Based on Shrinkage, Subspace Optimization, and Continuation

On A Class of Greedy Sparse Recovery Algorithms -- A High Dimensional Approach

Recovery of damped exponentials using structured low rank matrix completion