Iterative Methods for Computing the Moore–Penrose Pseudoinverse of Quaternion Matrices, with Applications

Valentin Leplat, Salman Ahmadi-Asl, JunJun Pan, Ning Zheng · 2025

Read Iterative Methods for Computing the Moore–Penrose Pseudoinverse of Quaternion Matrices, with Applications by Valentin Leplat, Salman Ahmadi-Asl, JunJun Pan, Ning Zheng (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-16979.txt.

Introduction

Valentin Leplat, Salman Ahmadi-Asl, JunJun Pan, and Ning Zheng develop, in a 2025 math paper, step by step methods to compute a special inverse for matrices whose entries are quaternions. Quaternions extend complex numbers, and their multiplication does not commute, which makes many standard matrix tricks fail. The Moore-Penrose pseudoinverse is a general purpose inverse used across signal and data work.

Their methods work natively with quaternions, without translating the problem into real or complex form. One refines a classic Newton-Schulz loop, with care for left and right identities in non square cases. Higher order versions converge faster near the answer. They add a randomized sketch method, a hybrid of the two, and a gradient method on the normal equations. Each comes with convergence analysis and factorizations that cut the work per step.

Catalog record

CREATORValentin Leplat, Salman Ahmadi-Asl, JunJun Pan, Ning Zheng
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE67 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-16979

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Iterative Methods for Computing the Moore–Penrose Pseudoinverse of Quaternion Matrices, with Applications Valentin Leplat, Salman Ahmadi-Asl, JunJun Pan, Ning Zheng Abstract We develop quaternion–native iterative methods for computing the Moore–Penrose (MP) pseudoinverse of quaternion matrices and analyze their convergence. Our starting point is a damped Newton–Schulz (NS) iteration tailored to noncommutativity: we enforce the appropriate left/right identities for rectangular inputs and prove convergence directly in ℍ \mathbb{H} under a simple spectral scaling X 0 = α ​ A H X_{0}=\alpha A^{H} with α ∈ ( 0 , 2 / ‖ A ‖ 2 2 ) \alpha\in(0,2/\|A\|_{2}^{2}) . We then derive higher–order ( hyperpower ) NS schemes with exact residual recurrences that yield order- p p local convergence, together with factorizations that reduce the number of s × s s\times s quaternion products per iteration. Beyond NS, we introduce a randomized sketch–and–project method (RSP–Q), a hybrid RSP+NS scheme that interleaves inexpensive randomized projections with an exact hyperpower step, and a matrix–form conjugate gradient on the normal equations (CGNE–Q) that minimizes 1 2 ​ ‖ X ​ A − I ‖ F 2 \tfrac{1}{2}\|XA-I\|_{F}^{2} using only multiplies by A A and A H A^{H} . All algorithms operate directly in ℍ \mathbb{H} (no real or complex embeddings) and are matrix–free. Numerically, on controlled synthetic (tall random) matrices, NS with spectral scaling attains the best accuracy and runtime; higher–order NS matches the accuracy with a modest time increase. The hybrid RSP+NS is competitive, and CGNE–Q is promising. RSP–Q achieves acceptable accuracy but can be slower in practice because each step forms a sketched pseudoinverse via repeated thin-QR or small Hermitian SPD Gram solves—an identifiable bottleneck we outline how to mitigate. In three application case studies (CUR image/video completion, Lorenz filtering, FFT-based deblurring) we deploy only the NS family, which achieves state-of-the-art accuracy with the lowest wall time. These quaternion–native, matrix-free methods are suitable as drop-in solvers for large-scale quaternion inverse problems. 2020 MSC 65F20 , 65F10 , 15A09 , 15A33 1 Introduction Quaternions are a number system that extends complex numbers, consisting of one real part and three imaginary parts. They were introduced by William Rowan Hamilton and are typically represented as a combination of these components, where the imaginary units have specific multiplication rules that lead to non-commutative behavior. This means that the order in which quaternions are multiplied matters, which is a significant departure from the properties of real and complex numbers. Extending the classical matrix decompositions and concepts from complex matrices to quaternion matrices has been an active research topic started from [ 1 , 2 ] and the seminal paper by Zhang [ 3 ] was a relaunch for much research on the subject. For instance for the quaternion QR and quaternion SVD, see [ 4 ] and [ 5 ] , respectively. There are several books on quaternions each one focusing on an specific aspect of this extended number system, for theoretical aspects of quaternion linear algebra, we refer to [ 6 ] , for scientific aspects, the book [ 7 ] provides nice algorithmic materials, for signal processing applications see [ 8 ] , for computer graphic applications see [ 9 , 10 ] , for arithmetic aspects of quaternion see [ 11 ] From scientific computing perspective, the idea of so called structure preserving has been used to accelerate the quaternion computations by converting all computations in real or complex domain and benefit from efficient matrix operations in BLAS-3., we refer to [ 12 , 13 , 14 , 15 , 16 , 17 ] for such ideas. Moreover, solving linear systems of equations over quaternions and their special solutions, e.g. Hermitian, have found intensive interest in last decades, see [ 18 , 19 , 20 , 21 , 22 , 23 , 20 ] and the references therein. Recently, the randomization idea [ 24 ] was extended to quaternion matrices [ 25 , 26 , 27 , 28 , 29 , 30 , 31 ] . While quaternions define a number system with associativity property but dual quaternions [ 32 ] , split quaternions [ 33 ] and octonions [ 34 ] as its extensions do not have both commutativity and associativity properties. The split quaternions similar to quaternions are 4D but builds a non-division algebra (some non-zero elements lack inverses), dual quaternions and octonions both are 8D non-division algebras. The more general number system is Clifford algebra [ 35 ] which is 2^{n} D, and for example generates real and complex numbers for n=0 and n=1 , respectively. Quaternions find extensive applications in various fields, particularly in computer graphics [ 10 ] , robotics [ 36 , 37 ] , aerospace engineering [ 38 , 39 ] , physics simulations [ 40 ] , quantum computing [ 41 , 42 ] , and machine learning [ 43 , 44 ] . In computer graphics , they are used to represent rotations smoothly and avoid issues like gimbal lock that can occur with other rotation representations. In robotics, quaternions help describe the orientation of robotic arms and mobile robots, facilitating accurate control and movement. Similarly, in aerospace engineering, they are crucial for attitude control systems in aircraft and spacecraft. The advantages of using quaternions include their compact representation, which requires fewer components compared to rotation matrices, and their efficiency in computations. As technology advances and the demand for precise 3D modeling and simulations increases, the importance and relevance of quaternions continue to grow across various domains. Their unique properties and versatility make them an essential tool for anyone working with three-dimensional transformations and orientations. In this paper, we will study the Moore-Penrose pseudoinverse of quaternion matrices. It is well known that in both the real and complex domains, the Moore-Penrose (MP) pseudoinverse serves as a generalization of the matrix inverse, applicable to nonsquare or singular matrices where traditional inverses do not exist. By providing a unique solution that minimizes error in least-squares problems, the pseudoinverse enables robust handling of overdetermined or underdetermined systems of linear equations. This makes it a fundamental problem in linear algebra with significant applications in mathematics, engineering, and data science. To compute the MP pseudoinverse, the most robust and widely used method is the Singular Value Decomposition (SVD). Unlike closed-form (left/right pseudoinverse), which requires the matrix to have full column rank or row rank, SVD is stable almost for any matrix rank. Even so, its high computational cost and intensive memory requirements limit its application to large matrices. Other methods like QR decomposition and rank decomposition, though they are more efficient than the SVD, their performance degrades for large or sparse matrices and may be as costly as the SVD. Therefore, to improve computational efficiency, iterative methods such as Newton-style recursion (which has quadratic convergence) have been proposed. These iterative approaches can avoid full matrix decomposition and are more adaptable to parallel and distributed computing. While the MP pseudoinverse is classically defined over real or complex fields, its extension to quaternion matrices requires careful handling due to the non-commutativity in quaternion multiplication. Research on the MP pseudoinverse for quaternion matrices remains relatively limited. In [ 45 ] , the authors developed determinantal representations for generalized inverses over the quaternion skew field

Read the complete text