Overview
A Chinese tech forum roundtable (authored by "WB/Wenyuan", summer 2026) debates which of four mathematical "blades" is king for handling space and transformation: geometric algebra (GA), quaternions, matrices/linear algebra, and Riemannian geometry. The conclusion: there is no single winner—they are different blades forged from the same iron, each supreme in its own domain.
Key points
Family tree: all four are related
- Riemannian manifolds (M, g) are the most general stage (arbitrary curvature); Euclidean space ℝⁿ is the flat special case (K = 0).
- Quaternions ℍ ≅ Cl(3,0)⁺ are the even subalgebra of GA—a 3D-rotation special case (3D rotor = quaternion).
- By the Artin–Wedderburn theorem, every Clifford algebra Cl(p,q) is isomorphic to some matrix algebra—GA is the structured language, matrices its coordinate representation.
- The metric tensor g_ij is a "pointwise-varying matrix field"; GA meets Riemannian geometry in the Clifford bundle Cl(T*M, g), whose Dirac operator underlies relativistic quantum mechanics and the Atiyah–Singer index theorem.
- Matrices (5/10): 9 numbers for 3 DOF, gimbal lock issues with Euler angles, expensive SVD re-orthogonalization. Concedes.
- Quaternions (9/10): 4 numbers, no singularities, SLERP, cheap renormalization (~12 FLOPs vs. 30+), entrenched in Unity/Unreal/ROS/Eigen—engineering champion.
- GA rotors (9/10): algebraically equivalent to quaternions in 3D but geometrically clearer (rotation axis = bivector) and generalize to any dimension—theoretical champion.
- Riemannian (5/10): geodesics on SO(3) needed only for rotation statistics (pose averaging, medical registration).
- Matrices remain the fastest "output format" for bulk vertex transforms; quaternions/GA are the internal representation.
- GA wins *algebraic* unification: geometric product ab = a·b + a∧b unifies complex numbers, quaternions, exterior algebra, Pauli/Dirac matrices; Maxwell's four equations collapse to ∇F = J; conformal GA unifies lines, circles, planes, spheres as blades.
- Riemannian geometry wins *geometric* unification: Euclidean (K=0), spherical (K>0), hyperbolic (K<0); gravity, Fisher-information statistics, manifold optimization.
- They are complementary wings, fused in the Clifford bundle—GA the algebra wing, Riemannian the geometry wing.
- Matrices are the general but lossy "default language"; quaternions are a 3D special case, not a contender.
- Machine learning: Matrices 10/10 (attention, GEMM-optimized Tensor Cores, 90%+ of AI compute); GA 5/10 (GCAN, CliffordNet—emerging equivariant ML); Riemannian 4/10 (natural gradient, 2–5× speedups at medium scale only); quaternions 3/10 (QNNs).
- Robotics: Quaternions 9/10 (ROS tf2, IMU fusion, SLAM); matrices 8/10 (homogeneous transforms, Jacobians); GA 5/10 (gafro library, IEEE RAM 2025); Riemannian 5/10 (SO(3)/SE(3) planning).
- Graphics/XR: Matrices 10/10 (MVP pipeline); quaternions 9/10 (SLERP skeletal animation); GA 5/10 (CGA intersections in one line; arXiv:2511.15398 proposes GA as a unified XR interface); Riemannian 2/10.
- Overall engineering ranking: matrices > quaternions > GA ≈ Riemannian.
- Curved spaces / constrained optimization / information geometry → Riemannian tools (geomstats, Pymanopt)
- Large-scale linear compute / ML / render pipelines → matrices (NumPy, PyTorch, cuBLAS)
- 3D rotation interpolation / IMU fusion → quaternions (Eigen, GLM, ROS tf2); upgrade to GA rotors for high dimensions
- High-dimensional geometry / equivariant ML / physics modeling → GA (ganja.js, galgebra, gafro)
- Unsure → default to matrices
- Matrices: unshakeable short-term; long-term wildcards are quantum computing (HHL algorithm) and AI for Math (AlphaTensor).
- Quaternions: safe for a decade in games/IMU; growth in QNNs and differentiable physics; partial long-term replacement by GA rotors.
- GA: will not displace linear algebra within 10 years, but may enter the standard ML toolbox in 5–10 years if equivariant GNNs go mainstream.
- Riemannian: slowly becoming more practical (hyperbolic networks, manifold BatchNorm) but will stay an expert tool.
- Ultimate fusion: Clifford bundles on Riemannian manifolds—geometric deep learning on curved spaces—"a generational project."
- Hestenes, *Space-Time Algebra*, 1966
- Shoemake, *Animating Rotation with Quaternion Curves*, SIGGRAPH 1985
- Dorst, Fontijne, Mann, *Geometric Algebra for Computer Science*, 2007
- Ruhe et al., *Geometric Clifford Algebra Networks*, ICML 2023 (arXiv:2302.06594)
- Löw et al., *GAFRO: Geometric Algebra for Robotics*, IEEE RAM 2025
- Amari, *Information Geometry and Its Applications*, 2016
- do Carmo, *Riemannian Geometry*, 1992
- AlphaTensor, Nature 2022
- *One algebra for all: GA methods for CG and XR*, arXiv:2511.15398
- Projects: geomstats, Pymanopt, ganja.js, galgebra, Klein, gafro
Historical arc
Timeline highlights: Hamilton's quaternions (1843), Grassmann's exterior algebra (1844), Riemann's lecture on n-dimensional manifolds (1854), Clifford's geometric product (1878), the 1880s Gibbs/Heaviside "victory" of vector analysis that eclipsed both quaternions and GA for ~70 years (called "a miscarriage of justice in math history"), Hestenes' *Space-Time Algebra* (1966), Shoemake introducing quaternions to computer graphics (SIGGRAPH 1985), AlphaTensor's RL-discovered matrix multiplication (Nature 2022), and Geometric Clifford Algebra Networks (ICML 2023).
Lesson: the most elegant does not win—the most practical and teachable does.
PK Debate 1: King of 3D rotation
PK Debate 2: Unified framework
PK Debate 3: Engineering adoption
Decision tree (four-line mnemonic)
> Curved → Riemannian. Linear → matrices. Rotation → quaternions. Geometric reasoning → GA.
Final verdict
| Tool | Position | Analogy | |------|----------|---------| | Matrices | Universal language, king of engineering | Swiss army knife | | Quaternions | 3D rotation specialist | Precision boning knife | | GA | Geometrically unified language | Professional chef's knife set | | Riemannian | Master of curved spaces | Nuclear weapon—most powerful, highest barrier |