Core Concepts
Biased spatial directions enhance spectral method effectiveness in high-dimensional phase retrieval.
Abstract
The article explores a spectral initialization method's performance in high-dimensional phase retrieval scenarios. Analyzing two models of the covariance matrix, it reveals improvements due to biased spatial directions and a phase transition phenomenon. The study extends to orthogonal and projector matrices, illustrating results with numerical simulations.
Introduction
Optical systems can only measure power spectral density, leading to the challenge of phase retrieval.
Generalized Linear Estimation
PhaseLift approach probabilistically recovers signals under mild conditions with Gaussian sensing vectors.
Unique Solution Investigation
Challenging question of unique solution existence in noiseless phase-retrieval problem is explored.
Reconstruction Algorithms
Iterative algorithms like Alternating Projections and Wirtinger Flow are discussed for phase retrieval.
Polynomial Time Algorithm
Approximate Message-Passing algorithm is highlighted as the best-known polynomial time algorithm for phase retrieval.
Spectral Method Enhancement
Spectral methods are employed for weak recovery problems by maximizing overlap between signals and estimates.
Wishart Matrix Analysis
Analytical extension of formulae for Wishart matrices is provided, showing a phase transition phenomenon.
Threshold Values
Threshold values for different covariance matrices are derived, showcasing universal results across models.
Stats
T = 2N −1 is necessary and sufficient to have a unique solution.
M(N) = 4N −4 when N = 2k + 1.