Hypercomplex Widely Linear Processing: Fundamentals for Quaternion Machine Learning

Sayed Pouria Talebi; Clive Cheong Took

stat.MLMar 12, 2026

Hypercomplex Widely Linear Processing: Fundamentals for Quaternion Machine Learning

Authors: Sayed Pouria Talebi, Clive Cheong Took

Organizations: Computer Science Department, University of Roehampton, London WC2R-2LS, U.K. · Electronic Engineering Department, Royal Holloway University of London, TW20 0EX, U.K.

Abstract

Numerous attempts have been made to replicate the success of complex-valued algebra in engineering and science to other hypercomplex domains such as quaternions, tessarines, biquaternions, and octonions. Perhaps, none have matched the success of quaternions. The most useful feature of quaternions lies in their ability to model three-dimensional rotations which, in turn, have found various industrial applications such as in aeronautics and computergraphics. Recently, we have witnessed a renaissance of quaternions due to the rise of machine learning. To equip the reader to contribute to this emerging research area, this chapter lays down the foundation for: - augmented statistics for modelling quaternion-valued random processes, - widely linear models to exploit such advanced statistics, - quaternion calculus and algebra for algorithmic derivations, - mean square estimation for practical considerations. For ease of exposure, several examples are offered to facilitate the learning, understanding, and(hopefully) the adoption of this multidimensional domain.

Figures & tables

$\mathbb{N}$	set of natural numbers
$\mathbb{R}$	set of real-valued numbers
$\mathbb{R}^{+}$	set of positive real-valued numbers
$\mathbb{C}$	set of complex-valued numbers
$\mathbb{H}$	set of quaternion-valued numbers
$\Re\{\cdot\}$	operator returning the real component
$\Im\{\cdot\}$	operator returning the imaginary component

Table 1

Figure 1: Schematic of a rotation around

\eta

by an angle of

\theta

, with

q_{\text{pre}}

and

q_{\text{post}}

pointing to the pre- and post-rotation orientation of the object in question.

Figure 2: Schematic demonstrating

q\in\mathbb{H}

along side its involution

q^{\kappa}

. The real axis has been omitted for simplicity as

\Re\{q\}

does not change under quaternion involutions. In the top left graph,

q

(in black), its projection onto the

\imath-\jmath

plane (in blue), and its projection onto the

\kappa

axis (in red) is shown. In the top right graph, projection of

q

onto the

\imath-\jmath

plane is rotated around the centre by

\pi

. In the bottom left graph,

q^{\kappa}

is constructed from the projection of

q

onto the

\kappa

axis and the rotated projection of

q

onto the

\imath-\jmath

plane. In the bottom right graph, all elements of

q

and

q^{\kappa}

are shown together.

Autocorrelation matrices in $\mathbb{H}$ (of size $3\times 3$ )
$\mathbf{R}_{c}(n)=\left[\begin{array}[]{ccc}{\color[rgb]{0.25,0.25,0.25}73}&15.33-13.33\imath&{\color[rgb]{0.5,0.5,0.5}18.67-10.33\imath}\\ &+0.33\jmath+9\kappa&{\color[rgb]{0.5,0.5,0.5}+5.33\jmath-10\kappa}\\ 15.33+13.33\imath&{\color[rgb]{0.5,0.5,0.5}73}&15.33-13.33\imath\\ -0.33\jmath-9\kappa&&+0.33\jmath+9\kappa\\ {\color[rgb]{0.5,0.5,0.5}18.67+10.33\imath}&15.33+13.33\imath&{\color[rgb]{0.5,0.5,0.5}73}\\ {\color[rgb]{0.5,0.5,0.5}-5.33\jmath+10\kappa}&-0.33\jmath-9\kappa&\end{array}\right]$
$\mathbf{R}_{\imath}(n)=\left[\begin{array}[]{ccc}{\color[rgb]{0.5,0.5,0.5}35+4\jmath-6.67\kappa}&31.33-1.33\imath&{\color[rgb]{0.5,0.5,0.5}17.33-13\imath}\\ &+22.33\jmath+19.67\kappa&{\color[rgb]{0.5,0.5,0.5}-0.67\jmath-10.67\kappa}\\ 31.33+1.33\imath&{\color[rgb]{0.5,0.5,0.5}35+4\jmath-6.67\kappa}&31.33-1.33\imath\\ +22.33\jmath-19.67\kappa&&+22.33\jmath+19.67\kappa\\ {\color[rgb]{0.5,0.5,0.5}17.33+13\imath}&31.33+1.33\imath&{\color[rgb]{0.5,0.5,0.5}35+4\jmath-6.67\kappa}\\ {\color[rgb]{0.5,0.5,0.5}-0.67\jmath-10.67\kappa}&+22.33\jmath+19.67\kappa&\end{array}\right]$
$\mathbf{R}_{\jmath}(n)=\left[\begin{array}[]{ccc}{\color[rgb]{0.5,0.5,0.5}-28.33+14.67\imath-5.33\kappa}&-24.67+20.67\imath&{\color[rgb]{0.5,0.5,0.5}-14+14.33\imath}\\ &-15\jmath-29.67\kappa&{\color[rgb]{0.5,0.5,0.5}-4.67\jmath+12.67\kappa}\\ -24.67+20.67\imath&{\color[rgb]{0.5,0.5,0.5}-28.33+14.67\imath-5.33\kappa}&-24.67+20.67\imath\\ +15\jmath-29.67\kappa&&-15\jmath-29.67\kappa\\ {\color[rgb]{0.5,0.5,0.5}-14+14.33\imath}&-24.67+20.67\imath&{\color[rgb]{0.5,0.5,0.5}-28.33+14.67\imath-5.33\kappa}\\ {\color[rgb]{0.5,0.5,0.5}+4.67\jmath+12.67\kappa}&+15\jmath-29.67\kappa&\end{array}\right]$
$\mathbf{R}_{\kappa}(n)=\left[\begin{array}[]{ccc}{\color[rgb]{0.5,0.5,0.5}-51.67+36\imath-5.33\jmath}&-8.67+12.67\imath&{\color[rgb]{0.5,0.5,0.5}-16.67+15.67\imath}\\ &-7.67\jmath-5.67\kappa&{\color[rgb]{0.5,0.5,0.5}-4\jmath+6.67\kappa}\\ -8.67+12.67\imath&{\color[rgb]{0.5,0.5,0.5}-51.67+36\imath-5.33\jmath}&-8.67+12.67\imath\\ -7.67\jmath+5.67\kappa&&-7.67\jmath-5.67\kappa\\ {\color[rgb]{0.5,0.5,0.5}-16.67+15.67\imath}&-8.67+12.67\imath&{\color[rgb]{0.5,0.5,0.5}-51.67+36\imath-5.33\jmath}\\ {\color[rgb]{0.5,0.5,0.5}-4\jmath-6.67\kappa}&-7.67\jmath+5.67\kappa&\end{array}\right]$

Table 2: Quaternion-valued autocorrelation matrices considered in Example 3.2 . Observe the Hermitian structure of

{R}_{c}(n)

as shown in ( 49 ) and the

\eta-

Hermitian property of

\mathbf{R}_{\imath}(n)

\mathbf{R}_{\jmath}(n)

, and

\mathbf{R}_{\kappa}(n)

as illustrated in ( 46 ).

Auto-correlation	$\mathbf{R}_{c}(n)$	$\mathbf{R}_{\imath}(n)$
matrix
$\Re\{\cdot\}$	$\mathbf{R}_{rr}(n)+\mathbf{R}_{\imath\imath}(n)+\mathbf{R}_{\jmath\jmath}(n)+\mathbf{R}_{\kappa\kappa}(n)$	$\mathbf{R}_{rr}(n)+\mathbf{R}_{\imath\imath}(n)-\mathbf{R}_{\jmath\jmath}(n)-\mathbf{R}_{\kappa\kappa}(n)$
$\Im_{\imath}\{\cdot\}$	$\mathbf{R}_{\imath r}(n)-\mathbf{R}_{r\imath}(n)+\mathbf{R}_{\kappa\jmath}(n)-\mathbf{R}_{\jmath\kappa}(n)$	$\mathbf{R}_{\imath r}(n)-\mathbf{R}_{r\imath}(n)+\mathbf{R}_{\jmath\kappa}(n)-\mathbf{R}_{\kappa\jmath}(n)$
$\Im_{\jmath}\{\cdot\}$	$\mathbf{R}_{\jmath r}(n)-\mathbf{R}_{r\jmath}(n)+\mathbf{R}_{\imath\kappa}(n)-\mathbf{R}_{\kappa\imath}(n)$	$\mathbf{R}_{r\jmath}(n)+\mathbf{R}_{\jmath r}(n)-\mathbf{R}_{\kappa\imath}(n)-\mathbf{R}_{\imath\kappa}(n)$
$\Im_{\kappa}\{\cdot\}$	$\mathbf{R}_{\kappa r}(n)-\mathbf{R}_{r\kappa}(n)+\mathbf{R}_{\jmath\imath}(n)-\mathbf{R}_{\imath\jmath}(n)$	$\mathbf{R}_{\kappa r}(n)+\mathbf{R}_{r\kappa}(n)+\mathbf{R}_{\imath\jmath}(n)+\mathbf{R}_{\jmath\imath}(n)$

Table 3: Quaternion-valued auto-correlation matrices in terms of their real-valued counterparts

Auto-correlation	$\mathbf{R}_{\jmath}(n)$	$\mathbf{R}_{\kappa}(n)$
matrix
$\Re\{\cdot\}$	$\mathbf{R}_{rr}(n)-\mathbf{R}_{\imath\imath}(n)+\mathbf{R}_{\jmath\jmath}(n)-\mathbf{R}_{\kappa\kappa}(n)$	$\mathbf{R}_{rr}(n)-\mathbf{R}_{\imath\imath}(n)-\mathbf{R}_{\jmath\jmath}(n)+\mathbf{R}_{\kappa\kappa}(n)$
$\Im_{\imath}\{\cdot\}$	$\mathbf{R}_{\imath r}(n)+\mathbf{R}_{r\imath}(n)+\mathbf{R}_{\kappa\jmath}(n)+\mathbf{R}_{\jmath\kappa}(n)$	$\mathbf{R}_{\imath r}(n)+\mathbf{R}_{r\imath}(n)-\mathbf{R}_{\jmath\kappa}(n)-\mathbf{R}_{\kappa\jmath}(n)$
$\Im_{\jmath}\{\cdot\}$	$\mathbf{R}_{\jmath r}(n)-\mathbf{R}_{r\jmath}(n)+\mathbf{R}_{\kappa\imath}(n)-\mathbf{R}_{\imath\kappa}(n)$	$\mathbf{R}_{r\jmath}(n)+\mathbf{R}_{\jmath r}(n)+\mathbf{R}_{\kappa\imath}(n)+\mathbf{R}_{\imath\kappa}(n)$
$\Im_{\kappa}\{\cdot\}$	$\mathbf{R}_{\kappa r}(n)+\mathbf{R}_{r\kappa}(n)-\mathbf{R}_{\imath\jmath}(n)-\mathbf{R}_{\jmath\imath(n)}$	$\mathbf{R}_{\kappa r}(n)-\mathbf{R}_{r\kappa}(n)+\mathbf{R}_{\imath\jmath}(n)-\mathbf{R}_{\jmath\imath}(n)$

Table 4: Quaternion-valued auto-correlation matrices in terms of their real-valued counterparts

Autocorrelation matrices in $\mathbb{R}$ (of size $3\times 3$ )
$\mathbf{R}_{rr}(n)=\left[\begin{array}[]{ccc}7&3.33&1.33\\ 3.33&7&3.33\\ 1.33&3.33&7\end{array}\right]$	$\mathbf{R}_{\imath\imath}(n)=\left[\begin{array}[]{ccc}47&20&16.67\\ 20&47&20\\ 16.67&20&47\end{array}\right]$
$\mathbf{R}_{\jmath\jmath}(n)=\left[\begin{array}[]{ccc}15.33&-8&1\\ -8&15.33&-8\\ 1&-8&15.33\end{array}\right]$	$\mathbf{R}_{\kappa\kappa}(n)=\left[\begin{array}[]{ccc}3.67&0&-0.33\\ 0&3.67&0\\ -0.33&0&3.67\end{array}\right]$
$\mathbf{R}_{\imath r}(n)=\left[\begin{array}[]{ccc}12.67&4.67&1.67\\ 12&12.67&4.67\\ 13.33&12&12.67\end{array}\right]$	$\mathbf{R}_{\jmath r}(n)=\left[\begin{array}[]{ccc}-0.33&0&-1\\ 7.33&-0.33&0\\ -1.33&7.33&-0.33\end{array}\right]$
$\mathbf{R}_{\kappa r}(n)=\left[\begin{array}[]{ccc}-3&-1.67&-0.33\\ -3.33&-3&-1.67\\ 1.33&-3.33&-3\end{array}\right]$	$\mathbf{R}_{\jmath\imath}(n)=\left[\begin{array}[]{ccc}-0.33&16&-10\\ 8.67&-0.33&16\\ -1.67&8.67&-0.33\end{array}\right]$
$\mathbf{R}_{\kappa\imath}(n)=\left[\begin{array}[]{ccc}-2.33&-11.33&-3.33\\ -3.67&-2.33&-11.33\\ 1.67&-3.67&-2.33\end{array}\right]$	$\mathbf{R}_{\kappa\jmath}(n)=\left[\begin{array}[]{ccc}-5.33&-1&0.33\\ 5&-5.33&-1\\ -1&5&-5.33\end{array}\right]$

Table 5: Real-valued autocorrelation matrices corresponding to their quaternion-valued counterparts in Table 2 . The double subscripts indicate the correlation between the two relevant quaternion components. For example,

R_{\imath r}(n)

corresponds to the cross correlation between the

\imath-

imaginary and real component.

Jun 15, 2026

On the Entropy Formula for Real, Complex, and Quaternionic Deep Linear Networks

We extend the entropy formula of Menon and Yu for the real Deep Linear Network (DLN) to its complex and quaternionic analogues, obtaining a unified formula for DLNs over

\mathbb{R}

\mathbb{C}

, and

\mathbb{H}

Luis Contreras, Marco Nahas, Tejas Kotwal

Complex-Valued Neural Networks Deep Linear Networks

Ask PDF

Aug 7, 2026

Classical $\mathrm{SU}(2)$ Models Match or Exceed Shallow Variational Quantum Circuits on Vision Benchmarks

Quaternion-valued neural networks and variational quantum circuits (VQCs) both derive local transformations from

\mathrm{SU}(2)

geometry, yet their performance on classical supervised learning remains poorly understood. We compare real-valued, quaternion-valued, and quantum classification heads on identical frozen features across MNIST, FashionMNIST, and CIFAR-10. CIFAR-10 uses a learned 16-dimensional bottleneck and frozen ImageNet-pretrained ResNet18 features to separate architecture from representation quality. Quaternion classifiers match or approach real-valued baselines while outperforming shallow VQCs. On MNIST and FashionMNIST, quaternion networks nearly equal real-valued MLPs, whereas product-state VQCs show lower accuracy and higher cost. On CIFAR-10, quaternion networks retain 94--97% of real-valued performance and remain stable under a 32-fold increase in dimensionality. Product-state circuits underperform quaternion classifiers, while entanglement gives modest grayscale gains but reverses under pretrained CNN features (9.25 pp degradation vs.\ product-state). Fubini--Study/QFI natural gradients improve geometric alignment but not short-horizon loss reduction vs.\ Adam. A Friedman test on five-seed MNIST detects model differences (

χ^2=12.796

p=0.0051

n=5

), with Wilcoxon tests yielding large effect sizes (

d>5

) for QuatNet vs.\ quantum comparisons. For FashionMNIST and CIFAR-10, large effects (

d>2.0

) are the primary statistic given

n=3

. These results indicate that quaternion networks provide efficient, stable

\mathrm{SU}(2)

alternatives to shallow VQCs on tasks lacking intrinsic quantum structure. Shared local

\mathrm{SU}(2)

geometry and shallow entanglement are insufficient, within the regime studied, to confer practical quantum advantage. Conclusions are limited to shallow, measurement-limited circuits on such tasks.

Christopher Fulton, Irene Tsapara, Lawrence Fulton

United States Air Force Test Pilot School, Edwards Air Force Base, California, 93524, USA · Department of Engineering, Data, and Computer Science; National University, San Diego, California, 92123, USA · Applied Analytics, Boston College, 140 Commonwealth Ave, Chestnut Hill, MA, 02467, USA

Variational Quantum Circuits Image Classification

Ask PDF

May 12, 2026

A Family of Quaternion-Valued Differential Evolution Algorithms for Numerical Function Optimization

The numerical optimization of continuous functions is a fundamental task in many scientific and engineering domains, ranging from mechanical design to training of artificial intelligence models. Among the most effective and widely used algorithms for this purpose is Differential Evolution (DE), known for its simplicity and strong performance. Recent research has shown that adapting AI models to operate over alternative number systems-such as complex numbers, quaternions, and geometric algebras-can improve model compactness and accuracy. However, such extensions remain underexplored in bio-inspired optimization algorithms. In particular, the use of quaternion algebra represents an emerging area in computational intelligence. This paper introduces a family of novel Quaternion-Valued Differential Evolution (QDE) algorithms that operate directly in the quaternion space. We propose several mutation strategies specifically designed to exploit the algebraic and geometric properties of quaternions. Results show that our QDE variants achieve faster convergence and superior performance on several function classes in the BBOB benchmark compared to the traditional real-valued DE algorithm.

Gerardo Altamirano-Gomez, Álvaro Gallardo, Carlos Ignacio Hernández Castellanos

Instituto de Investigaciones en Matem´aticas Aplicadas y Sistemas, Universidad Nacional Aut´onoma de M´exico, M´exico · Universidad Iberoamericana, M´exico

Differential Evolution Evolutionary Optimization Black-Box Optimization

Ask PDF

Abstract

Explore similar work

On the Entropy Formula for Real, Complex, and Quaternionic Deep Linear Networks

Classical SU(2)\mathrm{SU}(2)SU(2) Models Match or Exceed Shallow Variational Quantum Circuits on Vision Benchmarks

A Family of Quaternion-Valued Differential Evolution Algorithms for Numerical Function Optimization

Classical $\mathrm{SU}(2)$ Models Match or Exceed Shallow Variational Quantum Circuits on Vision Benchmarks