Spectral Embedding Showdown: New Theory Reveals When Graph Methods Agree

Two cornerstone methods for analyzing graph data—Adjacency Spectral Embedding (ASE) and Laplacian Spectral Embedding (LSE)—frequently yield different latent subspaces when applied to the same network, leaving practitioners unsure which to trust. In a new paper on arXiv (arXiv:2605.22346), authors Minh Triet Pham and Ian Gallagher provide a rigorous theoretical explanation for this disagreement. They show that regularity (every node having identical degree) is a sufficient condition for perfect alignment: if the graph is regular, ASE and LSE produce identical subspaces. Any deviation from regularity introduces disagreement.

The authors derive an explicit bound showing that disagreement is driven by two structural forces: degree heterogeneity pushes the embeddings apart, while community structure strength (measured by the eigengap) pulls them back together. Their mathematical results are backed by extensive simulations on thousands of synthetic networks, confirming that the ratio of heterogeneity to community strength is a strong predictor of when the two embeddings can be treated as interchangeable. This work gives network scientists a principled way to decide which embedding method to use based on the graph's structural properties, with practical implications for social network analysis, bioinformatics, and recommendation systems.