Hermitian Dual Quaternion Matrix

• Hermitian dual quaternion measurement matrices are operators over dual quaternions that encode both rotational and translational constraints for pose and spatial estimation.
• They exhibit a rich spectral structure with dual number eigenvalues, enabling unitary diagonalization and minimax principles essential for identifiability and robust estimation.
• Applied in SLAM, SE(3) synchronization, and formation control, these matrices improve computation of pose initialization, sensitivity analysis, and overall system conditioning.

A Hermitian dual quaternion measurement matrix is a matrix-valued operator over the algebra of dual quaternions, playing a central role in contemporary pose-graph optimization, multi-agent formation control, simultaneous localization and mapping (SLAM), and SE(3) synchronization. Such matrices encode both rotational and translational measurement constraints and inherit a rich spectral structure from their Hermitian dual-quaternion algebra. Their spectral properties underpin identifiability, estimation quality, and robustness in geometric estimation and control pipelines (Qi et al., 2021, Cui et al., 2024, Zhao et al., 30 Jan 2026).

1. Algebraic Foundations: Dual Quaternions and Hermitian Structure

A dual quaternion is an element of $$\mathbb{DH} = \{q = q_{st} + q_{\mathcal{I}}\varepsilon : q_{st}, q_{\mathcal{I}} \in \mathbb{H},\, \varepsilon^2 = 0\}$$, where $\mathbb{H}$ denotes Hamilton's quaternion algebra. Addition and multiplication distribute accordingly: $$(a + b\varepsilon) + (c + d\varepsilon) = (a + c) + (b + d)\varepsilon;\qquad (a + b\varepsilon)(c + d\varepsilon) = ac + (ad + bc)\varepsilon,$$ and conjugation acts componentwise $$(q_{st} + q_{\mathcal{I}}\varepsilon)^* = q_{st}^* + q_{\mathcal{I}}^*\varepsilon$$. The standard part ($q_{st}$) encodes rigid body rotations, while the infinitesimal part ($q_{\mathcal{I}}$) encodes translations.

Given $$A = A_{st} + A_{\mathcal{I}}\,\varepsilon \in \mathbb{DH}^{n\times n}$$, Hermitian symmetry is enforced by $$A^* = A \implies A_{st} = A_{st}^*,\; A_{\mathcal{I}} = A_{\mathcal{I}}^*$$. Thus, both the standard and infinitesimal parts are quaternion Hermitian. This guarantees all eigenvalues of $A$ are dual numbers (Qi et al., 2021).

2. Spectral Theory and Eigenvalue Structure

For $A \in \mathbb{DH}^{n\times n}$ Hermitian, every right eigenvalue $\lambda$ (i.e., $A x = x \lambda$ for some appreciable $x \in \mathbb{DH}^n$) is of the dual number form $$\lambda = \lambda_{st} + \lambda_{\mathcal{I}}\,\varepsilon$$ (Qi et al., 2021, Qi et al., 2024). The standard part $\lambda_{st}$ is a real eigenvalue of $A_{st}$, and the infinitesimal part is given by

$$\lambda_{\mathcal{I}} = \frac{x_{st}^* A_{\mathcal{I}} x_{st}}{x_{st}^* x_{st}},$$

where $x = x_{st} + x_{\mathcal{I}}\varepsilon$. If the standard part has multiplicity $k$, the infinitesimal parts are determined by a $k\times k$ "supplement matrix" $S = W^* A_{\mathcal{I}} W$, where $W$ is a partial isometry with columns spanning the standard eigenspace (Qi et al., 2024).

Unitary diagonalization holds: every Hermitian dual quaternion matrix admits $$A = U \operatorname{diag}(\lambda_1, ..., \lambda_n) U^*$$, with $U^*U = I$ and $$\lambda_i \in \mathbb{R} \oplus \mathbb{R}\varepsilon$$ (Qi et al., 2021, Qi et al., 2022).

A minimax (Courant–Fischer-type) principle for the eigenvalues is valid: for $k=1,...,n$,

$$\lambda_k = \min_{\dim \mathcal{M}=k}\, \max_{x\in \mathcal{M}\backslash\{0\}} \frac{x^*A x}{x^*x}$$

where the order on dual numbers is: $a+\alpha\varepsilon \succeq b+\beta\varepsilon$ if $a > b$ or $a=b$ and $\alpha \ge \beta$ (Ling et al., 2022).

3. Determinants, Characteristic Polynomials, and Matrix Conditioning

For Hermitian $A\in \mathbb{DH}^{n\times n}$, the Moore and Chen determinants coincide and are defined so that

$\operatorname{det}(A) = \prod_{i=1}^n \lambda_i,$

with $\lambda_i$ running over all dual number eigenvalues. The characteristic polynomial $$p_A(\lambda) = \operatorname{det}(\lambda I - A) = \prod_{i=1}^n (\lambda - \lambda_i)$$ (Cui et al., 2024).

Invertibility is characterized spectrally: $A$ is invertible iff no $\lambda_i$ is zero. Positive definiteness and semidefiniteness correspond to all $\operatorname{Re}(\lambda_i) > 0$ or $\ge 0$, respectively. Conditioning is controlled via

$\kappa(A) = \max_i |\lambda_i|/\min_i |\lambda_i|,$

where $|\lambda| = |a| + \varepsilon |\alpha|$ for $\lambda = a + \alpha\varepsilon$. Well-conditioned measurement matrices have $$0 < a_{\min} \leq \operatorname{Re}(\lambda_i) \leq a_{\max}$$ and $|\alpha|$ small (Cui et al., 2024).

4. Measurement Matrix Construction in Pose Estimation and Control

In SLAM and SE(3) synchronization, measurement matrices $H \in \mathbb{DH}^{n\times n}$ are assembled from relative pose constraints: $$H_{ij} = \begin{cases} w_{ij}\tilde{Q}_{ij}, & i \neq j \ \sum_{k \neq i} w_{ik}, & i = j \end{cases}$$ where $\tilde{Q}_{ij}$ is a (possibly noisy) measurement of the relative pose from $i$ to $j$, and weights $w_{ij}=w_{ji}$ encode measurement confidence or incidence. Hermitian symmetry $H_{ij}^* = H_{ji}$ holds if $\tilde{Q}_{ji} = \tilde{Q}_{ij}^*$ (Zhao et al., 30 Jan 2026). In formation control, similar measurement matrices appear as Laplacians or adjacency matrices: $$\hat{L} = D - A, \quad A_{ij} = \hat{q}_i^* \hat{q}_j\;\; \text{if}\;\; (i, j) \in E$$ with $D$ the degree matrix (Qi et al., 2022, Chen et al., 21 May 2025).

In fully connected and noise-free settings, such measurement matrices are rank-one and positive semidefinite, while in practical scenarios, noise and partial observability introduce higher rank and diminish positivity. The spectrum reveals both global observability (via $\lambda_{i,st}$) and infinitesimal identifiability (via $\lambda_{i,\varepsilon}$) (Qi et al., 2021).

5. Algorithms for Spectral Decomposition

For $H \in \mathbb{DH}^{n\times n}$, power methods are used to compute dominant eigenpairs. The standard dual quaternion power method iterates $v^{(k)} = H v^{(k-1)}/\|H v^{(k-1)}\|$ and converges linearly to the dominant eigenvector if the largest standard part of $\lambda$ is simple (Cui et al., 2023). When $\lambda_1 = \lambda_2$ (standard parts), methods like DCAM-PM (dual-complex adjoint) and EDDCAM-EA leverage adjoint embeddings and Aitken-extrapolated iterations for faster convergence and for resolving defective cases (Chen et al., 21 May 2025).

For full spectral decomposition, the supplement-matrix method decomposes $H = H_{st} + H_{\mathcal{I}}\,\varepsilon$: the spectrum of $H_{st}$ yields all standard parts; the supplement matrices $S$ in each standard eigenspace yield the corresponding dual parts. Classical quaternion Hermitian solvers are leveraged in both steps (Qi et al., 2024).

6. Applications in Estimation, Synchronization, and Formation

Hermitian dual quaternion measurement matrices underpin various geometric estimation techniques:

• In SLAM, $H$ encodes the information matrix for pose estimation over a measurement graph, and its leading eigenvector provides a globally consistent pose initialization; refinement is performed with dual-quaternion projected generalized power methods for feasibility (Cui et al., 2023, Zhao et al., 30 Jan 2026).
• In SE(3) synchronization, spectral initializers and gradient refinements over the unit dual quaternion constraints achieve provable recovery bounds under bounded noise. Error contraction is linear up to a noise floor determined by the operator norm of the measurement perturbation (Zhao et al., 30 Jan 2026).
• In multi-agent formation control, Laplacian-type Hermitian dual quaternion matrices determine rigidity and identifiability. Zero standard-part eigenvalues correspond to gauge freedoms (global rigid-body motions), and positive standard parts encode locked configuration modes. Infinitesimal parts act as perturbative sensitivities to measurement errors (Qi et al., 2022, Qi et al., 2024, Chen et al., 21 May 2025).

Szabo's Gershgorin theorem yields spectral inclusion regions, and explicit examples detail small-scale construction and inversion (Qi et al., 2022, Cui et al., 2024).

7. Practical Design and Implementation Considerations

Well-conditioned Hermitian dual quaternion measurement matrices are constructed by assigning target eigenvalue spectra $\{\lambda_i\}$ with $$0 < a_{\min} \le \operatorname{Re}(\lambda_i) \le a_{\max}$$ and $|\operatorname{Im}(\lambda_i)|$ (i.e., dual part) small, then building $A = U \operatorname{diag}(\lambda_i) U^*$ for unitary $U$ (Cui et al., 2024). For numerical implementation, dual quaternion matrices are often mapped into $2n \times 2n$ complex (or $8n \times 8n$ real) block embeddings to leverage mature Hermitian algorithms (Chen et al., 21 May 2025).

Inversion and regularization utilize the dual quaternion SVD, with generalized inverses available if all singular/value standard parts are appreciable. Singular and near-singular modes are regularized via per-eigenvalue dual number thresholding (Ling et al., 2022). Block-sparsity and the symmetry pattern reflect measurement graph topology, and all computations are compatible with contemporary numerical linear algebra packages after lifting to quaternionic or complex representations.

---

References:

(Qi et al., 2021, Ling et al., 2022, Cui et al., 2023, Qi et al., 2024, Cui et al., 2024, Chen et al., 21 May 2025, Zhao et al., 30 Jan 2026, Qi et al., 2022)