Abstract
Phase retrieval—the task of reconstructing a signal from the magnitude of its linear measurements—is a fundamental problem in imaging and signal processing. While non-convex optimization offers a direct approach, the standard quartic loss function is often plagued by spurious local minima, necessitating either high sampling complexity or sophisticated initialization strategies. This paper introduces three novel Quotient Intensity Models (QIMs) that reshape the optimization landscape. By incorporating a quotient structure into the loss function, we prove that the resulting landscape becomes "benign" under Gaussian measurements with $O(n)$ sampling complexity. This allows simple gradient descent to achieve global convergence from random initialization, reaching the information-theoretic limit for successful recovery.
1. Introduction
In fields ranging from X-ray crystallography to astronomical imaging, sensors often capture only the intensity of a signal, losing the phase information entirely. Mathematically, this is formulated as recovering a signal $x \in \mathbb{R}^n$ or $\mathbb{C}^n$ from measurements $y_j = |\langle a_j, x \rangle|^2$.
Traditionally, this problem is solved by minimizing a squared-error loss function. However, this "quartic" landscape is notoriously difficult to navigate. It typically requires a sampling complexity of $O(n \log n)$ to ensure a favorable geometry, or a "spectral initialization" to place the starting point within a local basin of attraction. This paper addresses a fundamental question: Can we design a loss function that is globally benign—containing no spurious local minima—even at the optimal sampling complexity of $O(n)$, thereby enabling success from a completely random starting point?
2. Quotient Intensity Models (QIMs)
The core innovation of this work is the introduction of Quotient Intensity Models (QIMs). The standard loss function suffers from "heavy tails" due to fourth-order powers of random variables, which create a rugged landscape. The QIMs mitigate this by using a quotient structure that acts as a damping mechanism, effectively smoothing the landscape.
Three specific models are proposed:
- QIM1: This model divides the error by the target intensity $(a_k \cdot x)^2$. While it appears singular if a measurement is zero, the authors demonstrate that for Gaussian vectors, this occurs with zero probability.
- QIM2: To enhance robustness, this model adds a regularization term $\beta|u|^2$ to the denominator, providing stability when target intensities are small.
- QIM3: This model employs a more sophisticated denominator involving both the current estimate and the target intensity, providing even stronger damping of the gradient's magnitude in high-error regions.
3. Theoretical Landscape Analysis
The primary theoretical contribution is the proof that these QIMs possess a benign geometric landscape. Using high-dimensional probability and geometric analysis, the authors establish that for Gaussian measurement matrices, if the number of measurements $m$ satisfies $m \ge Cn$ (where $C$ is a sufficiently large constant), the following properties hold with high probability:
- No Spurious Local Minima: Every local minimizer is a global minimizer (corresponding to the true signal or its phase-shifted equivalent).
- Strict Saddle Points: All saddle points are "strict," meaning they have at least one direction of negative curvature. This property is crucial because it allows gradient-based algorithms to escape saddles and continue descending toward the global optimum.
This $O(n)$ complexity represents the theoretical "gold standard," showing that QIMs can operate at the information-theoretic limit while remaining computationally simple.
4. Experimental Results
The theoretical findings were validated through extensive numerical simulations across three primary domains:
4.1 Synthetic 1D Signals
Using both real and complex Gaussian signals, the researchers found that QIM2 and QIM3 achieved a 100% success rate with approximately $6n$ measurements. When compared to state-of-the-art methods like Truncated Wirtinger Flow (TWF) and Reweighted Wirtinger Flow (RWF), the QIMs demonstrated superior or competitive convergence speeds, particularly when all algorithms were initialized randomly rather than using spectral methods.
4.2 Coded Diffraction Patterns (CDP)
To test applicability to optical systems, the models were applied to Coded Diffraction Patterns using random masks. The QIMs maintained high performance, requiring only about 7 octanary patterns for successful reconstruction, suggesting that the "benign" property extends beyond pure Gaussian measurements to more structured imaging models.
4.3 Natural Image Reconstruction
The practical utility of the QIMs was demonstrated by reconstructing a high-resolution ($1080 \times 1920$) RGB image of the Milky Way Galaxy. The models successfully recovered the image with high fidelity, proving that the theoretical advantages of the quotient structure translate effectively to large-scale, real-world data.
5. Conclusion and Future Work
The Quotient Intensity Models represent a significant step forward in taming the non-convexity of phase retrieval. By smoothing the optimization landscape, these models eliminate the need for specialized initialization and achieve optimal sampling complexity.
Future research directions include the automated selection of hyperparameters (such as $\beta$ in QIM2), formal proofs for non-Gaussian measurement structures (like CDP), and further investigation into the models' performance under extreme noise conditions, such as low-photon imaging environments. Despite these open questions, the QIM framework provides a robust and theoretically grounded approach to solving one of imaging science's most persistent challenges.
Summary updated