Community detection in networks asks: which nodes belong together? The standard approach is spectral clustering — compute eigenvalues of the adjacency or Laplacian matrix, and use the eigenvectors to partition nodes. The method works when the eigengap (the separation between the informative eigenvalue and the bulk) is large enough. When communities are weakly separated, the eigengap shrinks and spectral methods lose resolution.
Kotharkar (arXiv:2603.11060, 2026) amplifies the eigengap by reweighting edges using discrete Ricci curvature. In a two-community stochastic block model, edges within a community have positive Ricci curvature — the neighborhoods of their endpoints overlap heavily. Edges between communities have lower or negative curvature — the endpoints' neighborhoods point in different directions. Reweighting each edge by its curvature strengthens within-community connections and weakens between-community ones.
The reweighting creates a two-level weight structure: high weights inside communities, low weights between them. Spectral clustering on the reweighted graph achieves a larger eigengap than on the original graph, with non-asymptotic misclustering guarantees. A single curvature reweighting step is sufficient. Multiple iterations are analyzed as a deterministic recursion that the random iterates track uniformly.
The structural point: the curvature does not add information to the graph. It extracts geometric information that was already present in the local topology — the pattern of neighborhood overlap — and converts it into a form (edge weights) that spectral methods can exploit directly. The graph always contained the community structure in its curvature. The reweighting makes the implicit explicit. Geometry serves as a translator between local structure (which curvature measures) and global structure (which spectral methods find).