Octonion Phase Retrieval

Roman Jacome, Kumar Vijay Mishra, Brian M. Sadler, Henry Arguello

TL;DR

This paper introduces Octonion Phase Retrieval (OPR), a novel framework designed to reconstruct 8-dimensional signals—such as 8-band multispectral images—from intensity-only measurements. By leveraging a pseudo-real-matrix representation to bypass the mathematical headaches of non-associative octonion algebra, the authors develop the Octonion Wirtinger Flow (OWF) algorithm, which significantly outperforms traditional channel-concatenation methods in both sample efficiency and reconstruction accuracy.

What Problem Does This Paper Address?

Phase retrieval (PR) is a classic signal processing challenge: how do you reconstruct a signal when you can only measure its magnitude or intensity, losing all phase information? This problem is foundational to fields like X-ray crystallography, astronomy, and optical imaging.

While standard PR deals with complex numbers, modern imaging often captures multi-dimensional data. For instance, color images use quaternions (4-dimensional hypercomplex numbers) to jointly process RGB channels and preserve their mutual correlations. However, when dealing with 8-band multispectral images, quaternions fall short.

Enter octonions—8-dimensional hypercomplex numbers. Octonions are perfect for modeling 8-band data, but they come with a massive mathematical catch: octonion algebra is non-associative ($ (x \cdot y) \cdot z \neq x \cdot (y \cdot z) $) and non-commutative. This lack of associativity means standard calculus breaks down; you cannot easily define derivatives or chain rules to run gradient-based optimization algorithms like Wirtinger Flow. Consequently, researchers previously had to flatten or concatenate the 8 channels into a standard complex vector, completely discarding the rich cross-channel interactions.

Methods

The authors bridge this gap by introducing Octonion Wirtinger Flow (OWF). To circumvent the lack of native octonion derivatives, they employ a mathematical clever trick: a pseudo-real-matrix representation.

  1. Real-Matrix Mapping: They map the octonion variables into an equivalent $8 \times 8$ real-valued matrix representation ($\beth(x)$) and an $8 \times 1$ real vector ($\aleph(x)$). This transforms the non-associative octonion multiplications into standard, associative real matrix-vector multiplications.
  2. Gradient Descent in the Real Domain: Using this real-matrix framework, they successfully derive the gradients of the phase retrieval cost function, allowing them to perform Wirtinger-like gradient descent updates.
  3. Spectral Initialization: Non-convex optimization algorithms are highly sensitive to initialization. The authors adapt the power method from quaternion algebra to octonions, allowing them to compute the leading right eigenvector of an octonion-valued matrix to find a high-quality starting point.

Mathematically, they prove that the signal can be recovered uniquely up to a global right-octonion phase factor (the trivial ambiguity inherent to this non-commutative space).

Datasets

The algorithm was evaluated using two primary data sources:

  • Synthetic Data: Random signals drawn from an octonion Gaussian distribution ($\mathcal{N}_\mathbb{O}^n$) to rigorously test sample complexity and noise robustness.
  • Real-World Data: The CAVE multispectral image dataset, from which the authors extracted 8 equispaced spectral bands (ranging from 400 to 700 nm) cropped to $32 \times 32$ pixels.

Key Insights and Ideas

  • Algebraic Cohesion Over Concatenation: Treating the 8 spectral bands as a single hypercomplex octonion entity inherently preserves cross-channel correlations. Flattening the data into a long complex vector forces the algorithm to learn these relationships from scratch, wasting data and compute.
  • The Matrix Workaround: The pseudo-real-matrix representation demonstrates that you don't need a fully realized hypercomplex calculus framework to perform hypercomplex optimization; you just need a clever isomorphic mapping to the real domain.

Key Results

  • Lower Sample Complexity: In synthetic testing, OWF achieved near-perfect signal recovery with a sampling complexity ratio ($m/n$) of just over 10. In contrast, traditional concatenated Wirtinger Flow required an $m/n > 20$ to succeed.
Figure 1. (a) Success rate of OWF and concatenated WF for different value of sampling complexity m/n with n = 100. (b) Convergence rate of OWF for m/n = 20 for 2000 iterations.
  • Noise Robustness: Under additive Gaussian noise, OWF maintained high recovery accuracy provided the sample complexity $m/n > 17$ and the Signal-to-Noise Ratio (SNR) was above 20 dB.
Figure 2. Success rate of OWF for different values of sampling complexity m/n with n = 30 with measurements under additive Gaussian noise
  • Superior Real-World Reconstruction: When tested on the CAVE multispectral dataset, OWF reconstructed the 8-band images with a Peak Signal-to-Noise Ratio (PSNR) of 39.01 dB, vastly outperforming standard real-valued gradient descent on concatenated channels, which only managed 24.16 dB. Qualitative results show that OWF accurately preserved both spatial structures and spectral signatures.
Figure 3. Reconstruction with real data. (a) Statistical performance for 64 images of the spectral image dataset. (b) RGB representation of the 8-channel spectral image and its individual 8 components on the right panel. (c) OWF- reconstructed image and its components; recovered image’s PSNR = 39.01 dB (d) GD-reconstructed image and its components; recovered image’s PSNR = 24.16 dB. (e) Recovered spectral signature.

Limitations and Future Work

The authors note that because they rely on a pseudo-real matrix representation rather than native octonion calculus, their performance plots lack a distinct, sharp phase transition typically seen in standard PR analysis.

Future work could focus on developing more direct hypercomplex optimization tools that do not rely on real-matrix expansion, or scaling this framework to even higher-dimensional hypercomplex algebras (like sedenions) for massively multispectral applications.

Summary updated

Abstract

Signal processing over hypercomplex numbers arises in many optical imaging applications. In particular, spectral image or color stereo data are often processed using octonion algebra. Recently, the eight-band multispectral image phase recovery has gained salience, wherein it is desired to recover the eight bands from the phaseless measurements. In this paper, we tackle this hitherto unaddressed hypercomplex variant of the popular phase retrieval (PR) problem. We propose octonion Wirtinger flow (OWF) to recover an octonion signal from its intensity-only observation. However, contrary to the complex-valued Wirtinger flow, the non-associative nature of octonion algebra and the consequent lack of octonion derivatives make the extension to OWF non-trivial. We resolve this using the pseudo-real-matrix representation of octonion to perform the derivatives in each OWF update. We demonstrate that our approach recovers the octonion signal up to a right-octonion phase factor. Numerical experiments validate OWF-based PR with high accuracy under both noiseless and noisy measurements.
BibTeX
@article{jacome2024octonion,
  author = {Roman Jacome and Kumar Vijay Mishra and Brian M. Sadler and Henry Arguello},
  title = {{Octonion Phase Retrieval}},
  journal = {IEEE Signal Processing Letters},
  year = {2024},
  doi = {10.1109/lsp.2024.3411934},
}

Loading related content...

Figure