Projective Geometric Algebra Models

• Projective Geometric Algebra Models are Clifford algebras built over projective spaces, unifying the representation of points, lines, and planes.
• They implement a coordinate-free framework using exterior algebra and pseudoscalars to encode and compute geometric primitives across various metric signatures.
• They underpin efficient computational methods for isometries and kinematic motions, with applications in robotics, computer graphics, and geometric deep learning.

A @@@@1@@@@ (PGA) model is a Clifford algebra constructed over a projective (homogeneous) vector space with a metric of chosen signature, providing a unified coordinate-free algebraic framework for encoding, computing, and transforming geometric primitives—points, lines, planes, etc.—and their relationships in homogeneous metric spaces. PGA subsumes classical vector, exterior, quaternion, and dual-quaternion algebras, and serves as a structure-preserving foundation for Euclidean, non-Euclidean (elliptic, hyperbolic), and certain kinematic spaces. It enables direct and uniform application of geometric products, meets, joins, and versor (sandwich) operators for isometries and geometric transformations, distinguishing itself from coordinate-based or matrix-centric computational models (Sokolov, 2013, Gunn, 2014, Gunn, 2011, Sokolov, 2014).

1. Algebraic Foundations and Construction

A PGA is constructed on the exterior algebra $\Lambda(V^*)$ of the $(n+1)$-dimensional real vector space $V^*$ with an (optionally degenerate) symmetric bilinear form of signature $(p,q,r)$; typically, $r=1$ for the homogeneous direction. Basis vectors $\{e_0, e_1, \ldots, e_n\}$ satisfy $e_0^2=\varepsilon_0,\,e_i^2=\varepsilon_i$ with $\varepsilon_i\in\{+1,-1,0\}$; for Euclidean spaces, the canonical signature is $(n,0,1)$ with $e_0^2=0$ and $e_i^2=1$ ($i=1,\ldots,n$). The full algebra has $2^{n+1}$ dimensions and is graded by multivector degree from $0$ (scalars) up to $n+1$ (pseudoscalars) (Sokolov, 2013, Gunn, 2014).

The fundamental geometric product for vectors $a, b$ in $V^*$ is

$ab = a\cdot b + a\wedge b$

where $a\cdot b$ is the symmetric inner product determined by the metric, and $a\wedge b$ is the antisymmetric exterior (wedge) product. This product extends by associativity and linearity to the whole algebra.

Duality is implemented via the pseudoscalar $I = e_0 \wedge \cdots \wedge e_n$: for any $k$-blade $A$, its dual is $A^* = A I^{-1}$ (or a grade-reversal isomorphism for degenerate/metric-neutral cases).

2. Geometric Primitive Representations

In the dual PGA model (preferred in Euclidean applications), geometric elements correspond to specific grades:

• Points: In dimension $n$, a point $(x^1, \ldots, x^n) \in \mathbb{R}^n$ is encoded as a grade-$n$ normalized blade

$$P = e_0 \wedge e_1 \wedge \cdots \wedge e_n + x^1 e_1 \wedge \cdots \wedge e_n + \cdots + x^n e_0 \wedge \cdots \wedge e_{n-1}$$

• Lines: Simple $k$-blades. In 3D, lines are grade-2 blades (intersection of two planes).
• Planes: Grade-1 elements in $V^*$, i.e., $a = a_1 e_1 + a_2 e_2 + a_3 e_3 + d e_0$, encode planes of normal $\mathbf{n} = (a_1, a_2, a_3)$ offset $d$.
• Higher codimension subspaces: General $k$-blades.
• Pseudoscalar: The top-grade blade $I$, representing orientation and dualization (Sokolov, 2013, Gunn, 2014, Gunn, 2019).

Meet (intersection) is the exterior product $\wedge$ in the dual algebra, while the join (span) is implemented either via duals and wedge in the primal algebra or directly (regressive product): $A\vee B = (A^*\wedge B^*)^*$.

3. Homogeneous Metric Geometries via Signature

By tuning the signature $(p,q,1)$ in the underlying metric space, PGA realizes a spectrum of homogeneous (Cayley–Klein) geometries:

• Euclidean: $(n,0,1)$, with degenerate homogeneous direction, enabling standard $\mathbb{E}^n$ geometry (Gunn, 2011, Sokolov, 2013, Gunn, 2014).
• Minkowski (pseudo-Euclidean): $(n-1,1,1)$, for signature $(-+++)$, relevant to kinematic and spacetime models.
• Hyperbolic/Elliptic: $(n,0,0)$ or $(n-1,1,0)$, for curved geometries (no null direction). The metric-neutral algebraic construction yields coordinate-free, uniform formulas for all distances, angles, and trigonometric measurements (e.g., rational trigonometry) across Euclidean and non-Euclidean settings (Gunn, 2014, Sokolov, 2014, Sokolov, 2013, Sokolov, 2016).

4. Geometric Transformations and Versor Operations

PGA encapsulates isometries as versor (spinor/pinor) operations, leveraging sandwich products:

• Reflections: For $a^2 \neq 0$, reflection of $M$ in the hyperplane $a$ is $M' = -a M a^{-1}$.
• Rotations: Composed of two reflections, yielding a rotor $R = ba$ with $a^2 = b^2 = 1$. Rotation by angle $\theta$ about an axis/bivector $B$:

$R = e^{-\frac{1}{2}\theta B},\quad M' = R M R^{-1}$

• Translations: In Euclidean $(n,0,1)$, use a null vector $e_0$. For normalized direction $a$:

$$T(\tau) = \exp(-\frac{1}{2} \tau a e_0) = 1 - \frac{1}{2}\tau a e_0$$

with $X' = T X T^{-1}$ implementing a translation (Sokolov, 2013, Gunn, 2014, Gunn, 2011).

• Screw Motions: In 3D, general Euclidean motions compose into motors—products of translators and rotors—modeled directly as even-grade versors; the even subalgebra is isomorphic to dual quaternions (Gunn, 2020, Gunn, 2014).

The spin group $\mathrm{Spin}(n,0,1)$, consisting of normalized even versors, double-covers the connected Euclidean group $SE(n)$, with all rigid-body motions realized by sandwiching. The logarithm of a motor recovers twist/screw parameters (Gunn, 2011).

5. Computational and Practical Characteristics

PGA provides a reductionist, polymorphic syntax, encapsulating a range of subalgebras:

• Vector algebra: Realized in the subspace of ideal elements.
• Exterior algebra: Retained in native wedge products for meet/join operations.
• Quaternions and dual quaternions: Isomorphic to even subalgebras, facilitating efficient rotor-based and screw-theoretic computation in 3D (Gunn, 2019).
• Automatic differentiation: Dual basis and degenerate pseudoscalar $I$ with $I^2=0$ enable seamless embedding of dual numbers for forward-mode differentiation. Practical implementations exploit data sparsity, vectorization, and specialized kernels to execute meet, join, and sandwich operations with computational complexity on par with or less than dense matrix methods of classical geometric computation. PGA minimizes overparameterization and eliminates affine-bias in chained transformations (Gunn, 2019, Gunn, 2020).

6. Comparison with Other Geometric Algebra Models

PGA sits naturally between lower-level vector/Grassmann methods and the conformal geometric algebra (CGA) "2-up" model:

• Advantages over vector/matrix methods: Uniform, coordinate-free representation across all primitive dimensions; compact closure of all isometries in product form; robust numerical behavior.
• Comparison with CGA: PGA is $(n+1)$-dimensional while CGA is $(n+2)$-dimensional, resulting in lower computational cost and simpler normalization for flat (non-conformal) geometry. However, direct sphere and conformal object manipulation is less tractable in PGA due to non-blade representation of spheres (Gunn, 2014).
• Relationship to line/Klein geometry: For general projective transformations, the larger Clifford algebra $Cl_{3,3}$ and corresponding Pin/Spin representations are employed, underpinning the algebra of 3D projective transformations, their factorizations, and their embeddings via Klein's quadric (Klawitter, 2013, Klawitter, 2014, Li et al., 2015).

Model	Dim.	Encoding	Flat Primitives	Spheres/Conformal	Cost
Vector	$n$	points, lines	all as tuples	indirect	minimal
PGA	$n+1$	blades	native	indirect	moderate
CGA	$n+2$	all as blades	native	direct	higher

7. Applications and Extensions

PGA is foundational in multiple contemporary computational domains:

• Robotic Kinematics and Dynamics: Efficient models of pose, screw motion, inertia computation, and base parameter identification (e.g., the "tetrahedral-point" model for base parameter analysis) (Sun et al., 2 Sep 2025).
• Geometric Deep Learning: Expressive, symmetry-preserving architectures such as the Geometric Algebra Transformer (GATr) and extensions (iP-GATr) utilize the Cl(3,0,1) PGA model for equivariant operations on 3D data (Brehmer et al., 2023, Haan et al., 2023).
• Physics and Computer Graphics: Encapsulation of rigid-body motion, rotations, translations, scene descriptions, and integration with dual quaternion representations (Gunn, 2020).
• Protein Design: Emerging generative models (e.g., Clifford Frame Attention) leverage the higher-order geometric messaging afforded by PGA to improve physical expressivity of 3D structures (Wagner et al., 2024).
• Generalized Homogeneous Spaces: Via appropriate signature selection, PGA enables rational, coordinate-free trigonometry and measurements in Euclidean, elliptic, hyperbolic, and Minkowski spaces (Gunn, 2014, Sokolov, 2016, Sokolov, 2013, Sokolov, 2014).

The model's computational efficiency, unification of geometric primitives, and robust isometry representation have led to its adoption across domains requiring algebraically-grounded, symmetry-respecting geometric computation and learning.

---

References:

(Sokolov, 2013, Gunn, 2011, Gunn, 2014, Gunn, 2019, Gunn, 2020, Sokolov, 2014, Gunn, 2014, Sokolov, 2013, Sokolov, 2016, Klawitter, 2013, Klawitter, 2014, Li et al., 2015, Sun et al., 2 Sep 2025, Brehmer et al., 2023, Haan et al., 2023, Wagner et al., 2024)