Phase retrieval with one or two diffraction patterns by alternating projections with the null initialization (Q1645274): Difference between revisions

In this interesting paper, the authors consider the phase retrieval, i.e., the solution of the equation \(b=|A^{\ast} x_0|\), where \(x_0 \in X \subseteq {\mathbb C}^n\) represents the unknown signal, \(A^{\ast} \in {\mathbb C}^{N\times n}\) is the diffraction process with \(N \gg n\), and \(b \in {\mathbb R}_{+}^N\) are one or two diffraction patterns. The subset \(X\) represents all prior constraints on the signal. The diffraction process \(A^{\ast}\) is modeled by a multidimensional discrete Fourier transform. Using the uniqueness framework of the second author [Inverse Probl. 28, No. 7, Article ID 075008, 20 p. (2012; Zbl 1250.78024)], for any isometric \(A^{\ast}\) the authors show local geometric convergence for various alternating projections (AP), such as parallel AP, real-constrained AP, and serial AP, for the solution of phase retrieval. In practice, numerical stagnation of AP disappears under the uniqueness measurement schemes. Sharp bounds on the convergence rates of AP are given in terms of a spectral gap condition. It is shown that null initialization produces good-quality guess for fixed point iteration. Many numerical experiments point out that the AP with null initialization is a globally convergent algorithm for phase retrieval with one or two coded diffraction patterns. Further, AP converges faster to the exact signal than other iterative schemes.