Graph partitioning into isolated, high conductance clusters: theory, computation and applications to preconditioning

  • Authors:

  • Ioannis Koutis;Gary L. Miller

  • Affiliations:

  • Carnegie Mellon University, Pittsburgh, PA, USA;Carnegie Mellon University, Pittsburgh, PA, USA

  • Venue:
  • Proceedings of the twentieth annual symposium on Parallelism in algorithms and architectures
  • Year:
  • 2008

Quantified Score

Hi-index 0.02

Visualization

Abstract

We consider the problem of decomposing a weighted graph with n vertices into a collection P of vertex disjoint clusters such that, for all clusters C ε P, the graph induced by the vertices in C and the edges leaving C, has conductance bounded below by φ. We show that for planar graphs we can compute a decomposition P such that |P| n/ρ, where ρ is a constant, in O(log n) parallel time with O(n) work. Slightly worse guarantees can be obtained in nearly linear time for graphs that have fixed size minors or bounded genus. We show how these decompositions can be used in the first known linear work parallel construction of provably good preconditioners for the important class of fixed degree graph Laplacians. On a more theoretical note, we present upper bounds on the Euclidean distance of eigenvectors of the normalized Laplacian from the space of vectors which consists of the cluster-wise constant vectors scaled by the square roots of the total incident weights of the vertices.