Reproduce “On the Expressive Power of Spectral Invariant Graph Neural Networks” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Bohang Zhang, Lingxiao Zhao, Haggai Maron

This research provides a comprehensive theoretical framework to analyze the expressive power of spectral invariant Graph Neural Networks, demonstrating that most existing architectures are bounded by the proposed Eigenspace Projection GNN (EPNN) and are fundamentally strictly less expressive than the 3-Weisfeiler-Lehman test.

Key insights

  • EPNN serves as a unifying upper bound, demonstrating that most prior spectral invariant architectures are either less expressive or equivalent to it.
  • All existing spectral invariant GNNs, including EPNN, are strictly less expressive than the 3-Weisfeiler-Lehman (3-WL) test.
  • Commonly used 2-dimensional spectral information (e.g., projection matrices) does not provide additional expressive power when combined with already powerful higher-order GNNs like k-IGN (for k>2) or PPGN.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.