Abstract
In this paper, we tackle the problem of recovering the phase of complex linear measurements when only magnitude information is available and we control the input. We are motivated by the recent development of dedicated optics-based hardware for rapid random projections which leverages the propagation of light in random media. A signal of interest $\mathbf{\xi} \in \mathbb{R}^N$ is mixed by a random scattering medium to compute the projection $\mathbf{y} = \mathbf{A} \mathbf{\xi}$, where $\mathbf{A} \in \mathbb{C}^{M \times N}$ is a realization of a standard complex Gaussian i.i.d. random matrix. While optics-based matrix multiplications are significantly faster and more energy-efficient than traditional CPU or GPU counterparts, they face two major challenges: only the intensity $|\mathbf{y}|^2$ is recorded by the camera, and the transmission matrix $\mathbf{A}$ is unknown. We demonstrate that even without knowledge of $\mathbf{A}$, we can recover the unknown phase of $\mathbf{y}$ for an equivalent transmission matrix with the same distribution as $\mathbf{A}$. Our method leverages the distributional invariance of $\mathbf{A}$ and the ability to control input signals to cast the phase retrieval problem as a Euclidean distance geometry problem. We demonstrate the algorithm's effectiveness in both numerical simulations and real hardware experiments, showing that it not only recovers the missing phase but also mitigates quantization effects and improves measured magnitudes.
Introduction
Random projections are fundamental to modern machine learning, enabling efficient dimensionality reduction and kernel approximations. However, as datasets grow, performing these operations on conventional hardware becomes computationally expensive and power-hungry. Optical Processing Units (OPUs) provide a promising alternative by utilizing the physical scattering of light to perform matrix-vector multiplications almost instantaneously with minimal energy consumption.
Despite their potential, OPUs are limited by two factors: the loss of phase information, as camera sensors only capture light intensity, and the difficulty of calibrating the complex, high-dimensional transmission matrix $\mathbf{A}$. Traditional phase retrieval methods typically require knowledge of $\mathbf{A}$, making them impractical for these "black box" optical systems. This work introduces a framework to recover phase information without needing to calibrate the operator.
The Measurement Phase Retrieval (MPR) Framework
The proposed Measurement Phase Retrieval (MPR) framework treats phase retrieval as a localization problem in the complex plane. By controlling the input signal $\mathbf{\xi}$, we can introduce "anchor" signals $\mathbf{r}_k$ and measure the intensity of their differences: $|\mathbf{A}(\mathbf{\xi} - \mathbf{r}_k)|^2$. These measurements correspond to squared Euclidean distances between the projected signal and the projected anchors in the complex plane.
The algorithm proceeds in three main steps:
- Multidimensional Scaling (MDS): Using pairwise distances between the signal and several anchors, the relative coordinates of the projections are determined in the 2D complex plane.
- Procrustes Analysis: To ensure consistency across multiple measurement frames, recovered anchor positions are aligned to a fixed reference.
- Refinement: A gradient descent approach is employed to minimize the "stress"—the discrepancy between measured and model-predicted distances—further refining the phase estimation.
A key theoretical insight is the distributional invariance of the transmission matrix. Since $\mathbf{A}$ is a realization of a standard complex Gaussian i.i.d. matrix, rotating or conjugating its rows does not change its statistical properties. Thus, recovering the phase up to a global rotation or reflection is sufficient for most randomized algorithms.
Experimental Evaluation
The MPR algorithm was evaluated using three distinct setups:
- Synthetic Data: Complex Gaussian random matrices and signals were used to analyze the impact of noise and quantization.
- MNIST Dataset: 500 binarized samples of handwritten digits were processed to test real-world applicability.
- Hardware Implementation: Experiments were conducted on a LightOn OPU, which utilizes a Digital Micro-mirror Device (DMD) for input and an 8-bit camera for output.
Results and Discussion
The experiments confirmed that the MPR method successfully approximates an end-to-end linear system. As the number of anchor signals increases, the linearity error decreases significantly, demonstrating the robustness of the distance-based approach.
Beyond phase recovery, the algorithm provides a "free" denoising effect. By integrating multiple views from different anchors, the accuracy of the magnitude measurements is improved, effectively gaining an additional bit of precision over the raw 8-bit sensor data. This redundancy also makes the system resilient to measurements falling below the camera's sensitivity threshold.
Finally, the team performed a Randomized Singular Value Decomposition (RSVD) on the MNIST dataset using the OPU. The resulting singular vectors were nearly identical to those computed via standard software, proving that the exact values of the random matrix are less important than its distribution for these classes of algorithms.
Conclusion
This work demonstrates that phase information can be accurately recovered from optical random projections without explicit knowledge of the transmission operator. By framing the problem through distance geometry, we enable the use of high-speed, energy-efficient optical hardware for complex-valued computations. Future work will explore optimizing the number of anchors to balance throughput and accuracy, as well as extending this framework to other physical computing platforms.
Summary updated