Riemann GeoResolver: Non-Euclidean Attention Framework from arXiv

A new theoretical paper on arXiv (2608.10416) introduces the Riemann GeoResolver, a non-Euclidean attention framework that extends the Euclidean Resolver to hyperbolic and spherical geometries. The paper establishes three core theorems for inverse-distance attention (IDA) in Euclidean space: circuit separation, showing IDA achieves exact retrieval with O(1) resources while softmax requires Ω((log n)^2) width; a Polyak–Lojasiewicz inequality with a stronger constant than softmax, implying linear convergence, O(log n) Lipschitz scaling under low-rank/clustering assumptions, Θ(1) Hessian spread, and absence of spurious local minima; and a width-independent effective rank bound that limits noise memorization, with softmax memorizing arbitrary labels when d_h ≥ n, while IDA limits test error to O(η^2). The non-Euclidean extension replaces Euclidean distance with hyperbolic geodesic distance, building on the Euclidean prototype. The paper is categorized as a cross-type announcement and is available at https://arxiv.org/abs/2608.10416.