Perturbed Amplitude Flow for Phase Retrieval

Bing Gao, Xinwei Sun, Yang Wang, Zhiqiang XuIEEE TSP 2020 · DOI

Introduction

Phase retrieval is a classic inverse problem central to fields such as X-ray crystallography, astronomy, and coherent diffractive imaging. In these applications, sensors typically record only the intensity or magnitude of a signal, losing the phase information. Mathematically, the goal is to recover a vector $x \in \mathbb{C}^n$ (or $\mathbb{R}^n$) from $m$ measurements of the form $b_j = |\langle a_j, x \rangle|$, where $a_j$ are known measurement vectors.

Because the phase is missing, the problem is inherently non-convex. Traditional approaches like convex relaxation are computationally expensive, while early iterative methods often lacked theoretical guarantees. Recent "flow" algorithms, such as Wirtinger Flow (WF), utilize gradient descent on non-convex objectives. However, these methods can be sensitive to "bad" gradients—measurements where the signal magnitude is small—leading to instability. To mitigate this, previous state-of-the-art methods like Truncated Amplitude Flow (TAF) relied on complex data-filtering tricks (truncation or re-weighting), which complicate both implementation and theoretical analysis, especially for complex-valued signals.

The Perturbed Amplitude Flow (PAF) Method

This paper introduces Perturbed Amplitude Flow (PAF), a simple yet powerful algorithm that addresses gradient instability through a novel objective function. Instead of the standard distance-to-amplitude loss, the authors propose a perturbed amplitude-based model:

$$\min_z f_\epsilon(z) := \frac{1}{m} \sum_{j=1}^m \left( \sqrt{|\langle a_j, z \rangle|^2 + \epsilon_j^2} - \sqrt{b_j^2 + \epsilon_j^2} \right)^2$$

The "secret sauce" of PAF lies in the perturbation parameter $\epsilon_j$. By setting $\epsilon_j = \sqrt{\alpha}b_j$, the loss function remains smooth and the gradients stay well-behaved even when measurements are near zero. This eliminates the need for the "if-then" logic found in truncation-based methods, providing a unified framework for both real and complex fields.

The PAF workflow consists of two main stages:

  1. Initialization: A spectral method is used to find a starting point within the basin of attraction of the global optimum.
  2. Gradient Descent: Vanilla gradient descent is performed on the perturbed loss function, ensuring stable and efficient convergence.

Experimental Results and Discussion

The authors evaluated PAF against several benchmarks, including Wirtinger Flow (WF), Truncated Wirtinger Flow (TWF), and Truncated Amplitude Flow (TAF), using both synthetic Gaussian data and natural images.

Convergence and Noise Robustness

A key advantage of PAF is its linear convergence rate. As shown in Figure 3, the relative error of PAF drops exponentially with the iteration count. In both noiseless and noisy environments, PAF demonstrates superior or comparable convergence speed to existing methods. Notably, in the presence of noise, PAF maintains a stable error floor, whereas other methods may exhibit higher sensitivity.

Relative error vs. iteration count for noiseless (a) and noisy (b) cases. PAF shows rapid linear convergence and robustness to noise.

Success Rates and Sample Complexity

The efficiency of a phase retrieval algorithm is often measured by its sample complexity—the number of measurements $m$ required to recover a signal of dimension $n$. PAF achieves exact recovery at the information-theoretic limit of $O(n)$ measurements.

As illustrated in Figure 6, PAF consistently outperforms original Amplitude Flow (AF) and Wirtinger Flow (WF) in terms of success rate, particularly in the "low-data" regime where $m/n$ is small (between 2 and 4). This performance holds for both real-valued and complex-valued signals, validating the universal applicability of the perturbation approach.

Recovery probability vs. measurement ratio $m/n$ for real (a) and complex (b) signals. PAF achieves high success rates with fewer measurements than standard flow methods.

Practical Application: Image Recovery

Beyond synthetic tests, PAF was applied to reconstruct high-resolution images (e.g., the Milky Way Galaxy) using Coded Diffraction Patterns (CDP). The algorithm successfully recovered the images with high precision, proving significantly faster than Truncated Wirtinger Flow (TWF) while maintaining competitive accuracy.

Conclusion

Perturbed Amplitude Flow (PAF) represents a significant step forward in non-convex phase retrieval. By replacing complex truncation heuristics with a simple perturbation term, the authors have developed an algorithm that is both theoretically robust and practically efficient. PAF achieves optimal sample complexity, linear convergence, and universal applicability to real and complex signals, making it a versatile tool for modern imaging applications. Future work may focus on extending these theoretical guarantees to more structured, non-Gaussian measurement matrices common in optical hardware.

Summary updated

BibTeX
@article{gao2020perturbed,
  author = {Bing Gao and Xinwei Sun and Yang Wang and Zhiqiang Xu},
  title = {{Perturbed Amplitude Flow for Phase Retrieval}},
  journal = {IEEE Trans. Signal Process.},
  year = {2020},
  doi = {10.1109/tsp.2020.3022817},
}
Figure