Introduction
In many advanced imaging applications, such as X-ray crystallography, diffraction imaging, and astronomy, physical sensors are only capable of recording the intensity (magnitude) of a light wave's Fourier transform. The crucial phase information is lost during the measurement process. Reconstructing the original signal from these magnitude-only measurements is known as the phase retrieval problem.
While traditional phase retrieval is well-studied, modern applications often deal with signals that are sparse in a specific basis. Compressive Phase Retrieval (CPR) leverages this sparsity to reconstruct signals using significantly fewer measurements than the Nyquist-Shannon sampling theorem would otherwise require. However, many existing CPR algorithms suffer from either extreme computational complexity (e.g., semidefinite programming) or a lack of theoretical guarantees when applied to realistic Fourier-based optical setups. This paper introduces a fast, two-stage algorithm designed to bridge the gap between theoretical efficiency and practical Fourier optics.
Methodology: A Two-Stage Decoupled Approach
The core innovation of this work is a "divide and conquer" strategy that decouples the non-linear phase retrieval problem from the sparse recovery problem. The authors propose a framework that transforms a complex non-convex optimization into a sequence of two manageable steps:
1. Algebraic Phase Retrieval
Instead of solving for the signal and phases simultaneously, the algorithm first recovers the missing phases of the Fourier measurements. This is achieved through a specific physical setup using four structured optical masks (transmittance functions). By modulating the signal with these masks before it passes through a lens, the phases can be recovered up to a single global phase factor using a low-complexity algebraic method. The authors provide explicit constructions for these masks—defined as $p_s[n] = a_s \delta[n] + b_s$—ensuring the mapping is injective and physically realizable.
2. Compressive Sensing Reconstruction
Once the phases are recovered, the problem is reduced to a standard linear inverse problem. The algorithm then employs Basis Pursuit ($\ell_1$-minimization), a cornerstone of compressive sensing, to reconstruct the $k$-sparse signal from the now "phase-complete" measurements. By using established solvers like SPGL1, this stage remains computationally efficient even for high-dimensional signals.
Experimental Validation
The authors validated the proposed algorithm using numerical simulations of complex synthetic signals ($N=512$) with varying sparsity levels ($k$).
Measurement Efficiency
The results demonstrate that the algorithm achieves successful recovery with high probability when the number of measurements $M$ scales as $M \approx 4k \log_2(N/k)$. This logarithmic scaling with respect to the signal dimension $N$ is the "gold standard" for efficiency in compressive sensing, allowing for significant data compression without sacrificing reconstruction quality.
Speed and Stability
By avoiding the heavy computational overhead of semidefinite programming methods like PhaseLift, the two-stage approach offers significantly faster execution times. Furthermore, the algorithm demonstrates robust stability under noise. In simulations involving additive white Gaussian noise (SNR from 0 to 60 dB), the reconstruction error was shown to decrease linearly as the SNR increased, confirming the method's viability for real-world optical environments where measurement noise is inevitable.
Discussion and Limitations
While the algorithm represents a significant step forward in fast CPR, certain constraints remain. A technical requirement for the algebraic recovery step is that the first entry of the signal, $x[1]$, must be non-zero. While this is a mild restriction for many physical signals, it is a necessary condition for the current proof of uniqueness. Additionally, like most phase retrieval techniques, the signal is recovered up to a global phase ambiguity ($e^{i\phi}$), which is generally acceptable in optical imaging.
Future work may involve deriving formal theoretical error bounds for the integrated two-stage process and investigating whether the number of required masks can be reduced from four to three without compromising the stability of the reconstruction.
Summary updated