Skip to content

Spectral Analysis

Analysis of information propagation and network structure through the spectral properties (eigenvalues and eigenvectors) of graph matrices such as the adjacency matrix and Laplacian matrix. Spectral methods provide a unified, mathematically principled framework for understanding global structural properties of cascades without requiring explicit topological feature engineering.

Key insights

  • Spectral properties encode structure: Eigenvalues and eigenvectors of propagation graphs capture global structural information such as connectivity, balance, and diffusion capacity without manually enumerating cascade features.
  • Interpretable bounds: Spectral bounds (inequalities relating eigenvalues to specific structural properties like branching capacity, cascade scale, or diffusion dynamics) translate abstract spectral quantities into interpretable cascade characteristics.
  • Unified representation: Five categories of bounds—branching capacity, cascade scale, structural cohesion, propagation span, and diffusion dynamics—provide complementary perspectives on how information spreads.

Methodological foundations

  • Graph Laplacian: The Laplacian matrix \(L = D - A\) (where \(D\) is the degree matrix and \(A\) is the adjacency matrix) captures graph structure; eigenvalues and eigenvectors reveal bottlenecks (algebraic connectivity), clustering, and diffusion behavior.
  • Spectral radius: The largest eigenvalue \(\lambda_1\) of the adjacency matrix bounds key quantities like maximum degree and branching capacity.
  • Cheeger constant: Related to the second-smallest Laplacian eigenvalue \(\mu_{n-1}\), it measures how easily information can be separated into disconnected groups.
  • Random walk properties: Spectral gaps (differences between Laplacian eigenvalues) relate to mixing time and conductance, capturing how efficiently information spreads.

Applications to fake news detection

  • Cai & Zafarani (2026) derives 35 spectral bounds organized into five structural categories and shows they reliably distinguish fake from real news on Weibo22 and Twitter16 datasets, achieving classification accuracy comparable to handcrafted topological features while offering greater interpretability through explicit structural-property correspondence.

See also