Published November 18, 2026 | Version v1

Xhaeyorin - The Four-Point Obstruction to Continuous 3D Rotation Canonicalization and Linear-Size Frames for Unordered Point Clouds

Description

A continuous rotation canonicalization assigns a single standard pose to an unordered point cloud, with the same answer for every rotated copy. For centered 3D clouds of a fixed size \(n\), this manuscript proves that such a rule exists on the full domain if and only if \(n\leq 3\). It gives explicit canonical poses for up to three points and proves impossibility from four points onward. This answers the proper-rotation version of the unordered-point-cloud question raised by Dym, Lawrence and Siegel at ICML 2024. For uncentered clouds rotated about a fixed origin, the corresponding threshold is \(n\leq 2\). raw.githubusercontent.com

The obstruction already appears in a family of asymmetric clouds whose points have distinct distances from the centroid. A full axial rotation traces a contractible loop in this family, but an equivariant orientation rule would send that loop to a noncontractible loop in \(SO(3)\). Every continuous single-orientation predictor therefore makes a 180° equivariance error somewhere along each axial rotation orbit of the family. For the explicit four-point example, the smallest achievable worst-case displacement after optimal point matching is exactly \(2(\sqrt{2}-1)/\sqrt{14+(\sqrt{2}-1)^2}\), or 22.0062% of the cloud’s Frobenius norm.

The manuscript constructs a continuous alternative by assigning a weighted distribution of orientations to each cloud. Each directed convex hull edge defines an orientation, weighted by its length times its exterior dihedral angle. The resulting distribution of canonicalized inputs remains continuous when the hull changes, points collide, or a cloud collapses to a plane or line. The construction uses at most \(6n-12\) orientations for \(n\geq4\) centered 3D points. The exact worst case is \(2n\) for planar clouds embedded in 3D with \(n\geq3\). Across all centered 3D clouds, the exact worst cases for 4, 6 and 12 points are 12, 24 and 60 orientations, respectively.

For geometric deep learning, the construction gives a finite averaging rule that turns continuous vector- and tensor-valued backbones into rotation-equivariant predictors while preserving their existing permutation symmetry. The manuscript also proves that, for fixed ambient dimension \(d\), the worst-case local orientation-support requirement grows as \(\Theta(n^{\lfloor d/2\rfloor})\) among clouds of rank at least \(d-1\). A concrete isotropic cloud made of two orthogonal regular 50-gons in 4D requires at least 10,000 physical orientations in the stated frame model. The exact global minimum for general \(n\) in 3D and a global Lipschitz bound for the hull construction remain open.

The deposit includes the 12-page manuscript, a Python reference implementation, a deterministic data generator, numerical results and 10,188 synthetic verification records. These comprise 50 three-dimensional equivariance cases, nine degeneracy perturbations, 128 four-point matching cases, one reference 4D cloud, and 10,000 proper rotation matrices with their induced point permutations.

Files

Xhaeyorin_HuggingFace_Dataset_With_Uploader.zip

Files (709.5 kB)

Name Size
md5:9f38f43b0620cf9e44d3bfdaa6ed14b8
487.6 kB Preview Download
md5:da15702278bddb009708aeb2d09370bf
222.0 kB Preview Download