Synthetic-Trained Radial Adaptation for Quaternion Graph Lifting

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

We propose a synthetic-trained adaptive radial prediction filter for reversible quaternion graph lifting. Controlled synthetic regimes vary seven geometry factors; regime-optimal admissible filters are compressed into three prototypes. A seven-component decoder-visible descriptor drives a lightweight mixture estimator, with both bank and estimator trained entirely on synthetic data and frozen for target deployment. A parameter-free reference-routed update preserves exact reconstruction. On TotalCapture, the method reduces geodesic prediction MSE by versus a nonadaptive cubic B-spline predictor that uses a single radial filter across all neighborhoods, and versus uniform Karcher averaging, without target-domain fitting or transmitted node-wise adaptation parameters.