Total Least Squares Phase Retrieval

Sidharth Gupta, Ivan Dokmanic

TL;DR

This paper introduces a novel Total Least Squares (TLS) framework for phase retrieval, extending existing methods to account for errors not only in measurements but also in the sensing vectors themselves. This new approach, implemented via a gradient descent algorithm, demonstrates improved accuracy in simulations and real-world optical experiments, particularly when sensing vector errors are significant.

What Problem Does This Paper Address?

Phase retrieval is a fundamental problem in many scientific fields, aiming to reconstruct a signal (often an image or optical field) from its intensity measurements. A key challenge is that these measurements are typically quadratic, meaning the phase information is lost. While recent methods have successfully tackled this using least squares (LS) formulations, they often assume that only the measurements are noisy. In reality, the "sensing vectors" (which describe how the signal is measured) can also be inaccurate due to calibration errors or environmental factors. This paper addresses the critical gap of handling these simultaneous errors in both measurements and sensing vectors, which can degrade the performance of existing LS-based methods.

Methods

The authors propose a Total Least Squares (TLS) approach for phase retrieval. Unlike standard LS, which only corrects measurement errors to fit the model, TLS simultaneously corrects both measurement and sensing vector errors. They formulate this as an optimization problem and develop a gradient descent algorithm to solve it. A key innovation is leveraging the specific geometry of the phase retrieval problem to efficiently update the sensing vectors. Instead of solving a complex vector optimization problem for each sensing vector, they reduce it to a simpler scalar optimization problem, making the algorithm computationally feasible. They also derive analytical expressions for the reconstruction errors of both TLS and LS methods, providing theoretical insights into when each method is expected to perform better.

Datasets

Not discussed in the paper. The experiments use synthetic data generated with controlled noise levels and real optical hardware measurements.

Key Insights and Ideas

  • Sensing Vector Errors Matter: The paper strongly emphasizes that ignoring sensing vector errors can lead to suboptimal phase retrieval results, even when measurement errors are well-handled.
  • TLS Geometry is Key: The authors' clever use of the optimization problem's geometry is crucial for making the TLS approach computationally tractable. Reducing the sensing vector update to a scalar problem is a significant algorithmic contribution.
  • Error Analysis Provides Guidance: The derived analytical expressions for reconstruction errors offer a theoretical basis for understanding the trade-offs between TLS and LS under different error conditions. This allows for predicting when TLS will outperform LS.
  • Real-World Validation: The successful application of the TLS method on real optical hardware demonstrates its practical relevance beyond simulated environments.

Key Results

  • Improved Accuracy: Simulations show that the proposed TLS method achieves more accurate solutions than standard LS, especially when sensing vector errors are present.
  • Real Hardware Performance: Experiments on an Optical Processing Unit (OPU) confirm that TLS outperforms LS in reconstructing real-world optical signals, demonstrating higher Signal-to-Noise Ratios (SNRs) and better visual quality.
  • Theoretical Insights: The analytical error expressions reveal that TLS is generally preferred when sensing vector errors dominate, while LS might be better when measurement errors are significantly larger.
  • Efficiency: The geometric insight allows for an efficient implementation of the TLS algorithm, making it practical for use.

Limitations and Future Work

  • Computational Cost: While the proposed method is efficient, it still involves optimizing over both the signal and sensing vectors, which can be more computationally intensive than standard LS.
  • Multi-Signal TLS: A promising future direction is extending the TLS framework to multi-signal scenarios. This could allow for denoising sensing vectors collectively, potentially improving performance for subsequent signals that share the same sensing vectors, which is relevant for applications like optical neural network backpropagation.
  • Unknown Error Models: The current analysis often assumes Gaussian error models. Investigating performance with more complex or unknown error distributions could be valuable.

Summary updated

Abstract

We address the phase retrieval problem with errors in the sensing vectors. A number of recent methods for phase retrieval are based on least squares (LS) formulations which assume errors in the quadratic measurements. We extend this approach to handle errors in the sensing vectors by adopting the total least squares (TLS) framework that is used in linear inverse problems with operator errors. We show how gradient descent and the specific geometry of the phase retrieval problem can be used to obtain a simple and efficient TLS solution. Additionally, we derive the gradients of the TLS and LS solutions with respect to the sensing vectors and measurements which enables us to calculate the solution errors. By analyzing these error expressions we determine conditions under which each method should outperform the other. We run simulations to demonstrate that our method can lead to more accurate solutions. We further demonstrate the effectiveness of our approach by performing phase retrieval experiments on real optical hardware which naturally contains both sensing vector and measurement errors.
BibTeX
@article{gupta2022total,
  author = {Sidharth Gupta and Ivan Dokmanic},
  title = {{Total Least Squares Phase Retrieval}},
  journal = {IEEE Trans. Signal Process.},
  year = {2022},
  url = {https://openreview.net/forum?id=0vO5OIziYQw},
}
Figure