Quaternions ($q = a + bi + cj + dk$) offer an elegant mathematical framework for processing multi-channel signals holistically. By mapping channels to the three imaginary components ($i, j, k$), operations can natively preserve cross-channel correlations, such as the color channels of an RGB image or the spatial dimensions of 3D rotations and EEG signals.
However, quaternions suffer from a fundamental mathematical hurdle: non-commutativity ($ij = k \neq ji$). Because the order of multiplication matters, standard linear algebra tools—such as the Discrete Fourier Transform (DFT), matrix inversion, and Singular Value Decomposition (SVD)—break down or become computationally prohibitive.
The traditional workaround has been to expand an $M \times N$ quaternion matrix into a $4M \times 4N$ real-valued matrix. While this allows researchers to use standard real-valued solvers, it destroys the underlying algebraic symmetry, quadruples the dimensionality, and incurs massive computational and memory overhead. A rising paradigm shifts away from this brute-force expansion, focusing instead on structure-preserving algorithms and block-diagonal spectral decompositions that operate directly within the quaternion domain.
The Core Mechanism: Simplex-Perplex Splitting
Why can we block-diagonalize quaternion matrices when full diagonalization is blocked by non-commutativity? The answer lies in exploiting how quaternions commute relative to a specific chosen axis.
Any quaternion can be split into two components based on a single imaginary axis (for instance, the $i$ axis). The simplex component contains the real part and the chosen axis ($a + bi$), which behaves exactly like a standard complex number. The perplex component contains the remaining orthogonal axes ($cj + dk$). Crucially, the simplex component commutes with the chosen axis, while the perplex component anti-commutes (or conjugate-commutes) with it.
In On the Matrix Form of the Quaternion Fourier Transform and Quaternion Convolution, Sfikas et al. leverage this splitting to construct a formal matrix representation of the Quaternion Fourier Transform (QFT). By isolating the commuting and conjugate-commuting parts, the authors demonstrate that complex four-dimensional quaternion operations can be reduced to smaller, independent Hermitian matrix blocks.
From Global Complexity to 2x2 Blocks
This structural reduction becomes incredibly powerful when applied to structured matrices, such as circulant matrices or convolution kernels. Because a quaternion circulant matrix cannot be fully diagonalized into scalar eigenvalues due to non-commutativity, the next best structural target is a block-diagonal form.
Standard Real Expansion:
[ 4M x 4N Real Matrix ] ---> High Memory & Compute Overhead
Structure-Preserving Block-Diagonalization:
[ 1x1 or 2x2 Blocks ] ---> Independent, Parallelizable Subproblems
In Block-Diagonalization of Quaternion Circulant Matrices with Applications, Pan et al. prove that a permuted discrete QFT can transform quaternion circulant matrices into a block-diagonal form consisting strictly of $1 \times 1$ and $2 \times 2$ blocks. This geometric reduction bypasses the need for massive matrix constructions. Instead of inverting a massive system, one can invert these tiny, independent blocks in parallel, reducing the computational complexity of computing a quaternion circulant matrix inverse to an optimal $\mathcal{O}(n \log n)$.
The practical utility of these $2 \times 2$ blocks is further realized in deep learning. To guarantee the stability and generalization of a neural network, researchers often bound its Lipschitz constant by computing the spectral norm (the largest singular value) of its layers. For convolutional layers, this requires an SVD over a massive transformation matrix.
In Unlocking the matrix form of the Quaternion Fourier Transform and Quaternion Convolution, Sfikas et al. show that the square norm of a quaternionic convolution simplifies directly into a block-diagonal structure of $2 \times 2$ matrices. Furthermore, the authors establish that the left spectrum (left eigenvectors) is the primary driver for these operations. By extracting singular values directly from these localized $2 \times 2$ blocks, they achieve orders-of-magnitude speedups in spectral norm computations for Quaternionic Neural Networks.
Structure Preservation in Solvers and Tensors
When dealing with general linear systems or higher-dimensional tensors, maintaining the quaternion structure directly yields significant algorithmic advantages over real-space conversions.
Iterative Subspace Solvers
For large-scale linear systems with multiple right-hand sides, Wang et al. develop two algorithms, Gl-QFOM and Gl-QGMRES, in Gl‐QFOM and Gl‐QGMRES: Two Efficient Algorithms for Quaternion Linear Systems With Multiple Right‐Hand Sides. Instead of expanding the problem, the authors project the system directly onto a quaternion Krylov subspace. Because the operations happen natively within the quaternion domain, the solver inherits the natural algebraic symmetries of the data, accelerating convergence and drastically reducing memory usage during tasks like image restoration.
Tensor Decompositions
The block-diagonalization framework scales naturally to multi-dimensional arrays (tensors). Pan et al. extend their circulant matrix block-diagonalization to define a Quaternion Tensor Singular Value Decomposition (QT-SVD) using a t-product framework along the "tubes" of a tensor.
Similarly, for color image inpainting, Miao et al. introduce the Quaternion Tensor Left Ring (QTLR) decomposition in Quaternion Tensor Left Ring Decomposition and Application for Color Image Inpainting. To learn low-rank structures efficiently, the authors design a non-recursive SVD adapted for quaternions. This algorithm avoids the heavy computational overhead of recursive tensor network updates while strictly preserving the algebraic boundaries of the quaternion space.
Low-Rank and Structural Approximations
While exact block-diagonalization minimizes complexity, some applications demand even faster execution via low-rank approximations or structural relaxations.
Quaternion CUR Decomposition
Even an optimized Quaternion SVD (QSVD) can be too slow for large-scale datasets. To circumvent this, Wu et al. introduce a quaternion CUR (QMCUR) method in Efficient quaternion CUR method for low-rank approximation to quaternion matrix. Instead of computing abstract, orthogonal singular vectors, QMCUR samples actual, highly informative rows and columns directly from the quaternion matrix. The matrix is then approximated as the product of these sampled slices and a small core quaternion matrix, providing a massive speedup for color image processing.
Statistical and Architectural Approximations
In widely linear signal processing, fully describing a quaternion signal requires tracking four distinct covariance matrices. In Simultaneous diagonalisation of the covariance and complementary covariance matrices in quaternion widely linear signal processing, Xiang et al. show that for highly correlated, "improper" signals, an approximate simultaneous diagonalization is possible. Because the signal components share tight structural similarities, diagonalizing a single primary covariance matrix naturally minimizes the off-diagonal energy in the remaining complementary matrices.
On the hardware-constrained side of deep learning, Grant proposes Quaternion Approximate Networks (QUAN) in Quaternion Approximate Networks for Enhanced Image Classification and Oriented Object Detection. Rather than executing full quaternion algebraic convolutions, QUAN decomposes them into separable, real-valued operations. This approximation retains the crucial rotation equivariance properties inherent to quaternions while slashing computational latency, making it highly effective for real-time robotic vision.
Key Foundational Work
On the Matrix Form of the Quaternion Fourier Transform and Quaternion Convolution (2023) — contributed 3 concept elements to this direction.
Simultaneous diagonalisation of the covariance and complementary covariance matrices in quaternion widely linear signal processing (2017) — contributed 1 concept elements to this direction.