Imagine you have two identical jigsaw puzzles dumped on a table, but one has been rotated and the pieces are unnumbered. You need to figure out which piece matches which AND how much the whole thing was rotated — simultaneously. The standard approach (alternating minimization, a.k.a. Procrustes-Wasserstein) picks a rotation guess, solves the matching, re-estimates the rotation, and repeats. It works, but it gets stuck in local optima the way a hiker descending a foggy ridge can walk into the wrong valley. Rubix doesn't guess. It proves which valley is deepest. The core insight is beautiful in its economy. For two centered planar point sets of size n, every possible matching σ defines a single complex number zσ — a correlation between the two sets under that matching. Collect all n! of these complex numbers and take their convex hull: you get what the authors call the permutation polygon. The farthest point from the origin on this polygon IS the global optimum. No grid search, no random restarts, no annealing. The geometry of the assignment space encodes the answer. The theoretical contribution is sharp. The authors prove a tight bound of n(n−1) vertices for n ≥ 2, resolving an open problem posed by Günter Rote on rotation-assignment structures. This is not an asymptotic sketch — the paper is 67 pages with full proofs and experimental appendices, the kind of manuscript that reads like a monograph. In exact arithmetic, assignment queries recover the full polygon in O(n⁵) operations. Practically, the 2D result is already useful. On timed MPEG-7 shape pairs, Rubix hits every numerical reference value in 12 ms on average — 50× faster than a rotation grid at the same accuracy. That's the kind of speedup that changes whether you can use a method interactively versus as an overnight batch job. For 3D, the story is more nuanced: the paper extends the approach via assignment-based bounds inside a branch-and-bound framework, which is no longer globally optimal in the same clean sense but still demonstrably outperforms alternating minimization on real 3D scans, shape retrieval, and noisy crystal classification. The ladder here is clearly articulated. The enemy is alternating minimization (ICP variants, Procrustes-Wasserstein cycling), which is fast but fragile. Classical ICP is the workhorse of every 3D registration pipeline from SLAM to medical imaging. Rubix doesn't claim to replace ICP at massive scale — it claims to solve the problem exactly where ICP approximates, and to do it fast enough to matter. The 3D extension via branch-and-bound is less clean but still advances over the status quo for moderate point set sizes. The integrity profile is strong for a theory paper. The proofs are complete and self-contained. The experiments use standard benchmarks (MPEG-7 for 2D, real 3D scan datasets) rather than synthetic scenarios chosen to flatter the method. The authors are honest about where the 3D branch-and-bound scaling becomes expensive. What's missing is code availability — the paper doesn't mention a public release, which limits immediate reproducibility. The obvious next experiment is scaling the 3D branch-and-bound to point sets of hundreds or thousands of points and benchmarking against modern learned registration methods (DCP, RPM-Net, GeoTransformer). The authors stopped at moderate-scale 3D — my read is this is partly compute budget (branch-and-bound is exponential in the worst case) and partly that the paper's identity is as a theory contribution that also works in practice, not a systems paper competing on scale. The next paper from this group will likely be the engineering one.
