arXiv:2608.04608 2026 Architecture 2 ideas extracted · analyzed Aug 31, 2026
The paper gives a constructive method for designing finite-alphabet local observations that uniquely identify translated locations in a group or Cayley graph. This can transfer to graph neural networks as a discrete local positional encoding: assign node labels so that ordered labels in a connected neighborhood are injective across nodes, avoiding globally unique IDs and potentially reducing positional-encoding size. Its orientation-doubling lemma also provides a principled way to distinguish forward and reversed local views with one additional binary channel.
Replace globally unique node IDs with a small discrete alphabet labeling whose ordered local neighborhood words uniquely identify nodes. Feed these separating words to a GNN or graph transformer as positional features, preserving node distinguishability with bounded local receptive fields.
For local structures with a forward/reverse ambiguity, expose both ordered directions and add one explicit orientation bit. This creates a shared bidirectional positional encoder that can distinguish reflected neighborhoods without maintaining two completely independent directional encoders.