Redescending Boltzmann–Lohe Rotation Synchronization and the Hub-Precision Phase Transition on SO(d)
Year: 2026
Type: preprint
Abstract: Group synchronization learns latent SO(d) node frames from noisy pairwise measurements on a graph. In multi-view three-dimensional registration and signed community detection, gross Haar outliers from wide-baseline loop closures occur alongside large variations in vertex degree. In such settings, quadratic methods, including spectral synchronization and Riemannian gradient descent, perform poorly. Static degree normalization reduces the influence of corrupted hubs, yet it also penalizes clean overview measurements. Here, we introduce redescending Boltzmann-Lohe synchronization by maximizing an annealed exponential chordal energy via closed-form Kabsch-Umeyama polar retractions in monotonic coordinate ascent. Under power-law noise and outlier contamination with effective exponent gamma_eff, we prove that asymptotic Riemannian Gauss-Markov error is minimized by weighting edges as (d_i d_j)^(-alpha) with alpha* = gamma_eff / 2. Since gamma_eff is unknown in practice, we derive Adaptive Residual-Precision Boltzmann-Lohe (ARP-Lohe), an Empirical-Bayes method that estimates node precisions from accepted chordal residuals. Across 24 scan graphs of the Stanford Bunny, mean geodesic error drops from 4.88 degrees under two-stage Lie-algebraic iteratively reweighted least squares to 4.42 degrees. Under hub corruption, ARP-Lohe achieves 1.65 degrees error on rotation graphs. | Author: Felipe Provezano Coutinho (Independent Researcher, Brazil; ORCID: 0000-0001-7206-5856) | Code & Replication: https://github.com/provezano/boltzmann-lohe-synchronization | Archived Software DOI: https://doi.org/10.5281/zenodo.24374037
Source: Figshare
Author: Felipe Coutinho
Language: English
Subfield: Applied Mathematics
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236710885
