Nonlocality as the Price of Discovery: A Symmetry Between Transport Operators in Neurodegeneration and Attention in Language Models

Abstract

Short framing note. It observes that two independent modeling corrections are the same statement: in neurodegeneration, a streamline-weighted graph Laplacian is a nonlocal transport operator while voxelwise anisotropic diffusion is local; in language models, all-to-all attention is a nonlocal interaction while topographic architectures impose locality. It fixes the claim that locality buys interpretability and validability while nonlocality buys discovery, as a trade-off that holds across substrates. It makes no empirical claim beyond the two cited corrections. Date of record 2026-10-11 (JST); OpenTimestamps proof included.