Reciprocity Can Erase Direction:tructural Boundaries and Controlled Utility of Magnetic Graph Networks

| Source: arXiv AI

Tags: Magnetic Laplacian, Graph Neural Networks, Directed Graphs, DeepDRiD, Top-k Graph Learning

A theory paper shows magnetic graph networks can lose their directional signal entirely when learned graphs become reciprocal, and gives exact conditions under which that collapse happens.

Details

Magnetic graph networks encode edge direction through complex phase. The authors show that support-based magnetic Laplacians can erase that direction before prediction, deriving identities linking imaginary energy to amplitude-weighted reciprocity. Top-k selection forces complete collapse when the observed-node count is at most k+1, and dense supports impose a lower bound on reciprocal arcs. Constructed examples give a provable chance-level limit for direction-erasing representations. In 72 neural fits, support phase scored AUROC 1.0 on a unilateral ring but 0.5 on reciprocal and complete supports, while continuous score phase stayed at 1.0. On the DeepDRiD clinical study the benefit of continuous phase was small (+0.0036). The authors frame this as a diagnostic framework, not evidence of clinical superiority.