A Clifford (geometric) algebra is a number system in which one object, a multivector, carries scalars, vectors, oriented areas, volumes and higher pieces at once. These objects are increasingly used as the internal features of equivariant neural networks and physics surrogates. Shirokov has proved that every multivector has a singular value decomposition and a polar decomposition, the latter factoring it into a "rotation-like" unitary part times a "stretch-like" positive part. Both are stated in the algebra's own operations. However, the proofs still pass through matrix representations. There is no algorithm that computes these factors natively, and no one has measured the numerics of any native decomposition.
We propose the first such algorithm: an inverse-free Newton–Schulz iteration for the polar decomposition that uses only multivector products, one conjugation, and scalar extraction. Two older results make it cheaper and better behaved than a naive transcription. The positive factor and every Gram product $W^\dagger W$ that the iteration forms live in a fixed "Hermitian" subspace. In definite signatures that subspace is exactly a union of Shirokov's grade-mod-4 "quaternion types", which cuts the work on those products roughly in half. It also lets us re-project each iterate after every step, which suppresses roundoff drift. The trace–scalar identity supplies exact native norms, which we use for scaling and for deciding when to stop. The outcome would be a signature-uniform, differentiable routine, together with the first numerical benchmark this literature has lacked.
Core Claim
Let $M$ be a multivector in $\mathcal{G}_{p,q}$ or its complexification. Let $\dagger$ be the internal Hermitian conjugation built from reversion and grade involution in On SVD and Polar Decomposition in Real and Complexified Clifford Algebras. The iteration is
$$W_0 = M/\alpha,\qquad W_{k+1} = \tfrac12\, W_k\,\big(3 - \Pi_H(W_k^\dagger W_k)\big),$$
Here $\alpha$ is a scaling constant and $\Pi_H$ is the projection onto $\dagger$-self-adjoint multivectors. The iteration converges quadratically to the unitary factor $W$ of $M = WP$ whenever every singular value of $W_0$ lies in $(0,\sqrt3)$, the classical Newton–Schulz condition. The positive factor is then recovered as $P = \Pi_H(W^\dagger M)$.
Two structural facts make the iteration native and cheap:
- Where the Hermitian elements live. In Euclidean signature, $e_A^\dagger = \widetilde{e_A}$, so the Hermitian subspace is spanned by grades congruent to 0 or 1 mod 4. These are quaternion types $\bar0 \oplus \bar1$ in the sense of Quaternion Typification of Clifford Algebra Elements. In anti-Euclidean signature the Hermitian subspace is types $\bar0 \oplus \bar3$. In mixed signature the sign of $e_A^\dagger$ also depends on how many negative-square vectors $e_A$ contains, so the Hermitian subspace is a refinement: grade mod 4 crossed with that count mod 2.
- How the scaling is set. The trace identity from On computing the determinant, other characteristic polynomial coefficients, and inverse in Clifford algebras of arbitrary dimension gives $\langle M^\dagger M\rangle_0$ as the mean squared singular value. Setting $\alpha^2 = d\,\langle M^\dagger M\rangle_0$, with $d$ the representation dimension, therefore guarantees that the largest singular value of $W_0$ is at most 1. The same quantity applied to $W_k^\dagger W_k - 1$ gives an exact orthogonality residual for stopping.
The novelty has three parts: a matrix-free polar algorithm, the use of quaternion-type structure to decide which coefficients to compute and to project, and the extension of typification to the refined grading needed in mixed signature.
Research Program
Theory. We first extend quaternion typification to the refined grading needed for mixed signature. The question is whether the product of two Hermitian elements splits cleanly, with the anticommutator staying Hermitian and the commutator becoming anti-Hermitian, in the same way the four-type table works for commutators and anticommutators in A Classification of Lie Algebras of Pseudo-Unitary Groups in the Techniques of Clifford Algebras. Where it does, we derive product kernels that compute only the Hermitian output coefficients of $W^\dagger W$. We also derive product kernels for $W_k(3-H)$ that exploit the Hermitian structure of $H$.
Convergence and error analysis. We prove the convergence and backward-error statements natively. The spectral facts come from the trace identity and from the fact that $\dagger$ corresponds to a matrix adjoint (On Some Lie Groups Containing Spin Group in Clifford Algebra). Any matrix argument used here serves only to certify properties; the computation itself never forms a matrix.
Extension to SVD. Once $P$ is available, a Jacobi-style sweep can diagonalize it. Each rotation would be an exponential of an anti-Hermitian element from types $\bar2\oplus\bar3$, computed with the quaternion-arithmetic function formulas of Development of the Method of Quaternion Typification of Clifford Algebra Elements. This sweep is a stretch goal; the polar routine is the deliverable.
Implementation. Everything is written as batched GPU kernels over coefficient arrays. The same code runs for every $(p,q)$ up to $n=10$ and for complexified algebras.
Baselines
Naive reference: exact-arithmetic ground truth. The reference point is the Cartan–Bott matrix representation of each algebra, with the polar factor computed in exact or extended precision and mapped back. Every method is scored as error against this ground truth.
Strongest existing competitor: matrix-map plus library polar. This means mapping $M$ to its matrix representation, calling a batched matrix SVD or polar routine in double precision, and mapping back. It is the strongest competitor because it is what practitioners would actually do, and the libraries behind it are highly optimized. We run it on the same GPU, with the same batches and the same precision, and time it end to end including both mapping steps. For quaternionic cases, such as the $2\times2$ quaternionic matrices of $\mathcal{G}_{1,3}$, it runs on the standard complex embedding, which doubles the matrix size, because common libraries lack quaternionic routines.
No accuracy or runtime number for any native multivector decomposition exists. On Singular Value Decomposition and Polar Decomposition in Geometric Algebras states explicitly that such evaluation has not been done, so there is no reported result on Lacuna for native multivector SVD or polar numerics. Beating this baseline means matching its backward error to within a small constant factor while being faster on batches of many small multivectors ($n \le 6$), or else reaching the same accuracy with exactly structured factors where the matrix route leaks error across grades. If the matrix route is both faster and equally accurate at all $n$ we test, we will report that.
Ablations isolating the key ingredient. An untyped native Newton–Schulz, with full products and no projection, isolates the value of typification. A $\mathbb{Z}_2$ (even/odd) variant isolates the value of mod-4 over parity. A Newton polar iteration with native Faddeev–LeVerrier inversion, following On computing the determinant, other characteristic polynomial coefficients, and inverse in Clifford algebras of arbitrary dimension, represents the only native inverse currently available. All variants share the same scaling, the same stopping rule, and the same iteration cap.
Why not just…?
Why not just map to matrices and call LAPACK? This is the real threat, and we treat it as the headline comparison rather than dismissing it. The case for native computation rests on three things.
- Mapping costs gather/scatter work and, for real algebras, may force non-minimal or quaternionic representations that complicate the trace factor (On Rank of Multivectors in Geometric Algebras).
- A single native code path covers every signature, whereas the target matrix algebra changes with $(p,q)$ (On SVD and Polar Decomposition in Real and Complexified Clifford Algebras).
- Native factors can be exactly structured by construction.
The only speed evidence in the file is symbolic: roughly one day native versus over four days via matrices at $n=6$ (On computing the determinant, other characteristic polynomial coefficients, and inverse in Clifford algebras of arbitrary dimension On computing the determinant, other characteristic polynomial coefficients, and inverse in Clifford algebras of arbitrary dimension). That does not transfer to floating point, which is why the head-to-head timing experiment is decisive.
Why not just use even/odd parity instead of quaternion types? Parity tells us that an even $M$ keeps every iterate even, and we use that. But parity cannot identify the Hermitian subspace, because Hermiticity depends on grade mod 4 (Quaternion Typification of Clifford Algebra Elements). A parity-only projection leaves half the non-Hermitian drift in place. The $\mathbb{Z}_2$ ablation measures exactly what is lost.
Why not just lift Faddeev–LeVerrier and form $P^{-1}$ or $\sqrt{M^\dagger M}$ from the characteristic polynomial? Faddeev–LeVerrier is exact symbolically but prone to catastrophic cancellation in floating point. Its explicit formulas also grow combinatorially, reaching 231 summands for the determinant at $n=8$ (On computing the determinant, other characteristic polynomial coefficients, and inverse in Clifford algebras of arbitrary dimension On computing the determinant, other characteristic polynomial coefficients, and inverse in Clifford algebras of arbitrary dimension). The inverse-based ablation will show whether this concern is real at the condition numbers we test. Our iteration avoids inverses altogether.
Proposed Experiments
Synthetic accuracy and conditioning. We build synthetic suites for $\mathcal{G}_{3,0}$, $\mathcal{G}_{1,3}$, $\mathcal{G}_{3,1}$, $\mathcal{G}_{4,1}$, $\mathcal{G}_{3,3}$, $\mathcal{G}_{6,0}$, $\mathcal{G}_{8,0}$ and complexified $\mathcal{G}_{5}$. Multivectors are drawn with prescribed condition numbers from $10^0$ to $10^{12}$ by composing random unitary elements with known positive factors. We report four metrics:
- orthogonality residual $\langle (W^\dagger W-1)^\dagger(W^\dagger W-1)\rangle_0$;
- relative reconstruction error of $M = WP$;
- the non-Hermitian mass remaining in $P$;
- iterations to converge.
Speed. We measure wall-clock time and FLOPs at batch sizes from 1 to $10^6$.
Downstream use. We insert the routine as a unitary-projection step on multivector weights in Clifford Group Equivariant Neural Networks on the 5-body charged-particle task. That paper reports test MSE 0.0039, and Multivector Neurons: Better and Faster O(n)-Equivariant Clifford GNNs reports 0.0035. We also test the O(1,3) top-tagging task, where CGENN reports accuracy 0.942. In both cases we use the published training recipe and three seeds. The downstream claim is modest: the operation should cost little time and not hurt accuracy.
Decisive checks. Three results would settle the proposal:
- Accuracy. Typed Newton–Schulz reaches backward error within 10× of double-precision matrix polar up to condition number $10^8$.
- Speed. It is faster than map-plus-library at batched $n\le6$.
- Structure. Removing the typed projection, or replacing it with parity, measurably increases Hermiticity drift and iteration count.
If typing gives no measurable gain over the untyped native iteration, the contribution shrinks to "a native polar algorithm with benchmarks," and we would report it that way.
References
- On SVD and Polar Decomposition in Real and Complexified Clifford Algebras
- Quaternion Typification of Clifford Algebra Elements
- On computing the determinant, other characteristic polynomial coefficients, and inverse in Clifford algebras of arbitrary dimension
- A Classification of Lie Algebras of Pseudo-Unitary Groups in the Techniques of Clifford Algebras
- On Some Lie Groups Containing Spin Group in Clifford Algebra
- Development of the Method of Quaternion Typification of Clifford Algebra Elements
- On Singular Value Decomposition and Polar Decomposition in Geometric Algebras
- On Rank of Multivectors in Geometric Algebras
- On computing the determinant, other characteristic polynomial coefficients, and inverse in Clifford algebras of arbitrary dimension
- Clifford Group Equivariant Neural Networks
- Multivector Neurons: Better and Faster O(n)-Equivariant Clifford GNNs