arXiv:2607.07939 2026 Architecture 1 ideas extracted · analyzed Aug 30, 2026
The paper gives a constructive way to represent finite graph vertices as elements of a small product of cyclic groups while preserving all-pairs shortest-path distances through an isometric Cayley-graph embedding. The transferable asset is the combination of injective discrete coordinates, exact distance preservation, and compact non-binary factors such as \(\mathbb{Z}_m\), which can replace large one-hot or high-dimensional positional encodings in graph neural networks. A practical adaptation is to search for a compact abelian host on small training graphs, use its cyclic coordinates or group characters as deterministic node positional features, and compare against Laplacian, random-walk, and one-hot encodings. The paper's lower bounds also provide certificates for when a proposed binary code dimension is impossible, making representation search falsifiable rather than purely heuristic.
Replace one-hot node IDs or large positional encodings in a GNN with coordinates from a compact abelian Cayley graph. The coordinates preserve graph-shortest-path geometry exactly, while Fourier characters of cyclic factors provide smooth neural features with fewer channels.