Measurement Matrix Design for Phase Retrieval Based on Mutual Information

Nir Shlezinger, Ron Dabora, Yonina C. Eldar

TL;DR

This paper tackles the crucial problem of designing optimal measurement matrices for phase retrieval, a fundamental task in imaging and signal processing. Instead of relying on standard random matrices, the authors propose a novel approach that maximizes mutual information between the signal and its measurements, leading to significantly improved reconstruction accuracy and reduced measurement requirements.

What Problem Does This Paper Address?

Phase retrieval is the process of reconstructing a signal from the magnitudes of its linear measurements. This is a ubiquitous problem in fields like X-ray crystallography, diffraction imaging, and microscopy. A key component in this process is the "measurement matrix," which defines how the signal is transformed into observable data. Traditionally, researchers have assumed these matrices to be random Gaussian matrices, which work well in theory but are often impractical. In real-world applications, the measurement matrix is dictated by the physical setup and may have specific structures or constraints. Moreover, the presence of noise in measurements complicates recovery. This paper addresses the challenge of designing deterministic measurement matrices that are optimized for noisy phase retrieval, moving beyond the limitations of random matrices and offering practical design strategies.

Methods

The core of this paper's contribution lies in its novel design criterion: maximizing the mutual information (MI) between the signal of interest (SOI) and its noisy measurements. MI quantifies the statistical dependence between two random variables, and maximizing it is shown to be a robust approach for noisy phase retrieval, offering advantages over traditional Minimum Mean Squared Error (MMSE) criteria.

The authors achieve this by:

  • Reformulating Phase Retrieval as a MIMO Channel: They cleverly transform the non-linear phase retrieval problem into an equivalent linear complex MIMO (Multiple-Input Multiple-Output) channel. This allows them to leverage existing tools for analyzing MI in linear systems.
  • Deriving Necessary Conditions for Optimal Matrices: Using the MIMO channel interpretation and applying Karush-Kuhn-Tucker (KKT) conditions, they derive necessary conditions that an optimal measurement matrix must satisfy.
  • Low SNR Regime Analysis: In the low signal-to-noise ratio (SNR) regime, they derive a closed-form expression for the optimal measurement matrix, particularly for signals with "Kronecker symmetry" (a property satisfied by common distributions like zero-mean proper-complex Gaussian). This optimal matrix turns out to be a rank-one matrix.
  • Practical Design Algorithms: Recognizing that the optimal matrix might be difficult to obtain in general, they propose practical algorithms for designing both general (unconstrained) and structured (e.g., masked Fourier) measurement matrices. These algorithms involve approximations and iterative optimization steps.
  • Joint Optimization of Unitary Transformation and Measurement Matrix: They develop an alternating optimization algorithm to jointly find the optimal measurement matrix and an associated unitary transformation, further enhancing performance.

Datasets

Not discussed in the paper. The paper focuses on theoretical design and simulation-based evaluation using synthetic data generated according to specified signal distributions and noise levels.

Key Insights and Ideas

  • Mutual Information as a Superior Design Criterion: Maximizing mutual information proves to be a more effective and robust design criterion for noisy phase retrieval compared to MMSE, especially due to its invariance to linear transformations.
  • The Power of Deterministic Matrices: Designing deterministic matrices tailored to the problem can yield significant performance gains over generic random matrices.
  • Rank-One Matrices in Low SNR: The optimal measurement matrix in the low SNR regime for Kronecker symmetric signals is a rank-one matrix, which is counter-intuitive from an estimation perspective but explained by its ability to isolate the least noisy signal dimension.
  • Practical Algorithms for Real-World Constraints: The proposed algorithms provide a pathway to design effective measurement matrices even when dealing with structural constraints common in optical systems (e.g., masked Fourier matrices).
  • Joint Optimization Yields Further Gains: Optimizing both the measurement matrix and a related unitary transformation leads to additional performance improvements.

Key Results

  • Significant SNR Gains: Simulations show that the proposed deterministic unconstrained measurement matrix (AUC) achieves the best performance across various SOI distributions and recovery algorithms, outperforming random Gaussian matrices by 8-10 dB.
  • Reduced Measurement Requirements: The proposed designs require fewer measurements to achieve a similar estimation error compared to random matrices. For instance, AUC and AMF (masked Fourier) require fewer measurements than random coded diffraction patterns for comparable performance.
  • Masked Fourier Matrices Show Promise: The proposed masked Fourier measurement matrix (AMF) also demonstrates notable gains (0-2 dB) over random counterparts, highlighting its effectiveness for structured optical systems.
  • Low SNR Optimality Confirmed: The rank-one optimal matrix for low SNR (AOK) performs best in that regime, as designed.
  • Unitary Transformation Optimization Benefits: Optimizing the unitary transformation can further improve performance, yielding an SNR gain of up to 4 dB for AUC.

Limitations and Future Work

  • Non-Convexity: The optimization problem for designing the measurement matrix is generally non-convex, meaning the proposed algorithms may converge to local minima rather than the global optimum.
  • Approximations: The practical design algorithms rely on approximations, such as using the Linear MMSE (LMMSE) estimate instead of the true MMSE, which might introduce some sub-optimality.
  • Computational Complexity: While efficient algorithms are proposed, the computational complexity can still be significant, especially for large dimensions.
  • Future Work: The paper suggests exploring the design of measurement matrices for non-Gaussian noise scenarios beyond the scope of this work. Further investigation into the theoretical guarantees and practical implementation for more complex signal structures and measurement constraints would also be valuable.

Summary updated

Abstract

In phase retrieval problems, a signal of interest (SOI) is reconstructed based on the magnitude of a linear transformation of the SOI observed with additive noise. The linear transform is typically referred to as a measurement matrix. Many works on phase retrieval assume that the measurement matrix is a random Gaussian matrix, which, in the noiseless scenario with sufficiently many measurements, guarantees invertability of the transformation between the SOI and the observations, up to an inherent phase ambiguity. However, in many practical applications, the measurement matrix corresponds to an underlying physical setup, and is therefore deterministic, possibly with structural constraints. In this work we study the design of deterministic measurement matrices, based on maximizing the mutual information between the SOI and the observations. We characterize necessary conditions for the optimality of a measurement matrix, and analytically obtain the optimal matrix in the low signal-to-noise ratio regime. Practical methods for designing general measurement matrices and masked Fourier measurements are proposed. Simulation tests demonstrate the performance gain achieved by the proposed techniques compared to random Gaussian measurements for various phase recovery algorithms.
BibTeX
@article{shlezinger2018measurement,
  author = {Nir Shlezinger and Ron Dabora and Yonina C. Eldar},
  title = {{Measurement Matrix Design for Phase Retrieval Based on Mutual Information}},
  journal = {IEEE TSP},
  year = {2018},
  doi = {10.1109/tsp.2017.2759101},
}
Figure