Multivectors: The LEGO Bricks of Geometry — A Unified Mathematical Universe

Forum topic · 小凯 · 2026-03-29

Summary

This article introduces the multivector, the core object of geometric (Clifford) algebra, originally proposed by William Kingdon Clifford in 1878. Unlike traditional vector algebra, multivectors organize geometric entities by grade: scalars (grade 0), vectors (grade 1), bivectors (grade 2), and trivectors (grade 3), giving an n-dimensional multivector space of dimension 2^n. The geometric product ab = a·b + a∧b unifies inner and outer products into one operation, capturing both similarity and difference. The article explains key structures — blades (subspaces), versors (transformation operators), and rotors (elegant rotation encodings with the half-angle that explains spinor behavior) — and shows applications: Maxwell's four equations collapse into ∇F = J in spacetime algebra; conformal geometric algebra unifies translation and rotation for robotics and graphics; and RotorQuant (2026) reportedly achieved 44× parameter reduction and 10–19× speedups by replacing matrix rotations with Clifford rotors in LLM KV-cache quantization. The article also traces family relations to complex numbers, quaternions, spinors, and differential forms, and discusses challenges like non-commutativity, dimension explosion, and immature toolchains.

Multivectors: The LEGO Bricks of Geometry

This is an English adaptation of a Chinese forum article introducing multivectors — the central objects of geometric (Clifford) algebra.

Key points

Motivation: For centuries, mathematics used separate languages for different geometric objects: scalars (magnitude), vectors (arrows), cross products (areas), quaternions (rotations), matrices (Lorentz transformations). In 1878, William Kingdon Clifford proposed a unified algebra where all of these live together.
Graded structure: Multivectors are classified by grade — grade 0 (scalar), grade 1 (vector), grade 2 (bivector, oriented area), grade 3 (trivector, oriented volume), etc. In an n-dimensional space with orthonormal basis {e₁,…,eₙ}, the multivector space has 2ⁿ basis elements (e.g., 1 + 3 + 3 + 1 = 8 in 3D).
The geometric product: The heart of the framework is

ab = a \cdot b + a \wedge b

One product simultaneously yields the inner product (a scalar measuring "similarity") and the wedge product (a bivector measuring "difference"/area). The product is associative but non-commutative — and this non-commutativity correctly captures the order-dependence of rotations.

Duality: Multiplication by the unit pseudoscalar I = e₁∧…∧eₙ maps k-vectors to (n−k)-vectors. In 3D, this explains why the cross product can represent areas: the resulting vector is the dual of the bivector. The wedge product, not the cross product, is the natural measure of area in any dimension.
Special inhabitants:
Blades (v₁∧…∧v_k) represent k-dimensional subspaces. Not every multivector is a blade (e.g., e₁₂ + e₃₄ in 4D).
Versors are products of blades and generate orthogonal transformations via the sandwich product VxV⁻¹. A reflection is x' = −nxn; two reflections compose into a rotation R = mn.
Rotors are even versors: R = exp(θ/2 · B̂) = cos(θ/2) + sin(θ/2)B̂, applied as v' = RvR̃. The half-angle explains why spin-1/2 particles need 720° to return to their original state.

Applications

Electromagnetism: In the Spacetime Algebra, Maxwell's four equations collapse into a single equation ∇F = J, where F = E + IB is one bivector field — electric and magnetic fields are two parts of a single geometric object, matching special relativity's insight.
Rigid-body motion: Conformal Geometric Algebra (CGA) uses a 5D space where points, lines, planes, and spheres are all multivectors, and both translations T = 1 + ½te∞ and rotations are versors, so any rigid motion is a single versor V = TR applied via the sandwich product.
AI — RotorQuant (2026): Replaces random orthogonal matrix rotations in KV-cache quantization with Clifford rotors over Cl(3,0): reported 10–19× faster on NVIDIA GPUs, 9–31× on Apple Silicon, 44× fewer parameters (372 vs 16,399 at d=128), with ~99.0% attention fidelity.

Family relations

Complex numbers: the even subalgebra of Cl(0,1); complex multiplication = scaling + rotation.
Quaternions: the even subalgebra of Cl(3,0), with i = e₂₃, j = e₃₁, k = e₁₂; quaternion multiplication is the geometric product of bivectors, and qvq⁻¹ matches the rotor sandwich product.
Spinors: minimal left ideals of Clifford algebras; the ½ factor in the rotor exponential explains spinor sign flips under 360° rotation.
Differential forms: elements of the exterior algebra, a metric-free subset of Clifford algebra (wedge only, no inner product).

Challenges and open questions

Non-commutativity requires rebuilding intuition; multivector spaces grow as 2ⁿ (dimension explosion), requiring sparse representations.
The software ecosystem (GAViewer, GAlgebra, clifford, Versor, Kingdon) lags behind NumPy/Eigen in maturity.
Frontier directions: Clifford neural networks, Geometric Algebra Transformers (GATr), multivector parameterization of neural network weights, and multivector optimizers.