In classical linear algebra, we operate within the comfortable boundaries of real and complex numbers. These fields are commutative, associative, and free of non-zero elements that multiply to zero. However, modern applications in robotics, computer graphics, and signal processing frequently demand the representation of complex spatial geometries and dynamic transformations.
To model both rotation and translation simultaneously, or to track a value alongside its infinitesimal change, researchers turn to hypercomplex number systems: quaternions, dual numbers, and dual quaternions.
The challenge is that these algebraic structures break the foundational assumptions of standard linear algebra. Quaternions are non-commutative ($ij \neq ji$), and dual numbers introduce a nilpotent element $\epsilon$ (where $\epsilon \neq 0$ but $\epsilon^2 = 0$), creating zero divisors that make matrix inversion and factorization notoriously difficult. A growing body of research focuses on generalizing fundamental matrix operations, decompositions, and equations to these challenging algebraic domains by preserving their intrinsic geometric structures.
The Core Obstacle: Non-Commutativity and Zero Divisors
Why can't we just apply standard algorithms to dual or quaternion matrices? Consider the simple matrix equation $AXB = C$. In real linear algebra, we can use Kronecker products to flatten the matrices into vectors and solve the resulting linear system. But if the elements belong to a dual quaternion algebra, we cannot freely rearrange the order of multipliers.
Furthermore, the presence of the dual unit $\epsilon$ means that we cannot blindly divide or perform row reduction; any element with a zero "real part" has no multiplicative inverse. If an algorithm encounters a dual component during a pivot step, standard Gaussian elimination collapses.
To circumvent this, researchers must redesign algorithms from the ground up to respect the algebraic boundaries of the underlying system rather than forcing them into real-valued representations, which drastically increases computational and memory overhead.
Generalizing Matrix Decompositions
Matrix decompositions like QR, LU, and Singular Value Decomposition (SVD) are the workhorses of numerical computation. Extending them to dual and quaternion spaces requires tracking how the hypercomplex components propagate through orthogonal and triangular transformations.
QR and SVD Decompositions
In real spaces, QR decomposition relies on orthogonal matrices to decompose data into an upper triangular form. When dealing with dual matrices—which represent a standard value alongside its infinitesimal change—maintaining true orthogonality is difficult.
Renjie Xu et al. addressed this in QR decomposition of dual matrices and its application by developing a proper orthogonal decomposition tailored for dual matrices. This allows researchers to isolate dynamic wave patterns in complex datasets, such as brain activity, without losing track of the infinitesimal variations represented by the dual component. This structural preservation mirrors historical breakthroughs in classical matrix products, such as the efficient QR-like algorithms for computing the SVD of matrix sequences proposed by Gene H. Golub et al. in Computing the SVD of a General Matrix Product/Quotient.
LU and Takagi Factorizations
For structural decompositions like LU, the primary hurdle is managing the interaction between the real and dual parts during elimination. In LU decomposition and generalized Autoone-Takagi decomposition of dual matrices and their applications, Yimin Wei et al. demonstrated that the Sylvester equation serves as the mathematical bridge needed to systematically handle dual components. By leveraging the Sylvester equation, the authors successfully extended the Autoone-Takagi decomposition (originally designed for complex symmetric matrices) to dual complex symmetric matrices and dual quaternion $n$-Hermitian matrices, establishing a firm theoretical framework for numerical linear algebra in robotics.
Classical Space Dual / Quaternion Space
[ Commutative Field ] ---------> [ Non-Commutative / Zero Divisors ]
Standard Elimination Sylvester Equation / Structure Preservation
| |
v v
Standard LU/QR Dual LU / Dual QR
Solving Hypercomplex Equations and Inversions
Moving beyond factorizations, solving explicit matrix equations and computing inverses in hypercomplex spaces requires redefining classical identities.
Frobenius Inversion
Frobenius inversion provides a block-wise method to invert matrices. However, its derivation assumes that the underlying blocks commute. In Correction to: Generalizing Frobenius inversion to quaternion matrices, Qiyuan Chen et al. focused on overcoming these theoretical hurdles, proving that Frobenius inversion can indeed be generalized to quaternion matrices despite their non-commutative nature.
The AXB = C Equation
The matrix equation $AXB = C$ is foundational for control theory and graphics. In Dual Quaternion Matrix Equation AXB = C with Applications, Qing‐Wen Wang et al. derived the necessary and sufficient conditions for this equation to hold over the dual quaternion algebra. By generalizing the concept of $\phi$-Hermiticity to dual quaternions, the authors provided a flexible framework to find specialized, structure-preserving solutions directly applicable to rigid body transformations.
Iterative Solvers for Large Systems
When dealing with large-scale problems like color image restoration, direct inversion is computationally impossible. In Gl‐QFOM and Gl‐QGMRES: Two Efficient Algorithms for Quaternion Linear Systems With Multiple Right‐Hand Sides, Qing‐Wen Wang et al. designed structure-preserving Krylov subspace methods (Gl-QFOM and Gl-QGMRES). Instead of expanding a quaternion system into a real matrix four times its size, these algorithms operate directly on the quaternion structure, yielding faster convergence and significantly lower memory consumption.
The Geometry of Zeros
The ultimate utility of these algebraic extensions often lands on solving polynomials that govern physical motion. In kinematics, rigid body trajectories are modeled as "motion polynomials" over dual quaternions. Factorizing these polynomials or finding their roots is equivalent to breaking down a complex spatial maneuver into a sequence of simpler, elemental rotations and translations.
In Zeros of slice functions and polynomials over dual quaternions, Graziano Gentili et al. established the geometric properties of "slice functions" over dual quaternions. Because dual quaternions do not form a simple field, their polynomials do not behave like complex polynomials; a single dual quaternion polynomial can possess entire spheres of roots. By characterizing the precise nature of these zeros, the authors provided the theoretical foundation necessary to factorize motion polynomials reliably, closing the gap between abstract hypercomplex algebra and practical robotic engineering.
Key Foundational Work
Zeros of slice functions and polynomials over dual quaternions (2021) — contributed 3 concept elements to this direction.
Computing the SVD of a General Matrix Product/Quotient (2000) — contributed 1 concept elements to this direction.