Graph Theory - Cheeger's Inequality



Cheeger's Inequality

Cheeger's Inequality relates the spectral gap (the difference between the largest and second-largest eigenvalues of the graph's Laplacian matrix) to the isoperimetric constant (or Cheeger constant) of the graph, which measures the "bottleneck" or "edge expansion" of a graph when divided into two parts.

Cheeger's Inequality provides a relationship between the second-smallest eigenvalue of the graph's Laplacian matrix, denoted as 2, and the conductance of the graph, denoted as h(G). The inequality is given by −

h(G)  2 / 2  2 * h(G)

Where,

  • h(G) is the conductance of the graph, which measures how well connected the graph is and how easily it can be divided into two disconnected parts.
  • 2 is the second-smallest eigenvalue of the graph's Laplacian matrix. This eigenvalue gives information about the graph's connectivity and can be used to detect bottlenecks in the graph's structure.

Interpretation of Cheeger's Inequality

The inequality provides a bound on the conductance in terms of the second-smallest eigenvalue of the Laplacian. A lower value of 2 indicates that the graph has weak connectivity, and the graph is likely to have a small cut (a way of partitioning the graph into two parts with few edges between them).

On the other hand, a high value of 2 indicates that the graph is well-connected and difficult to partition into disconnected components.

Cheeger's Constant and Conductance

The conductance of a graph G is defined as the minimum ratio of the size of the edge cut between two sets of vertices and the total degree of the vertices in the smaller set. Formally, the conductance is given by −

h(G) = min{S ⊆ V} (cut(S, V \ S) / min(vol(S), vol(V \ S)))

Where,

  • cut(S, V \ S) is the number of edges between the set of vertices S and its complement V \ S.
  • vol(S) is the total degree of the vertices in the set S, i.e., the sum of the degrees of the vertices in S.

The goal is to find a small cut with a small volume in order to measure how easily the graph can be split into two parts. Cheeger's Inequality states that the second-smallest eigenvalue of the Laplacian matrix gives a bound on this conductance.

Example: Conductance of a Graph

Consider a simple graph where vertices 1 and 2 are connected by an edge, vertices 2 and 3 by another edge, and vertices 3 and 1 form a cycle. The degree matrix, adjacency matrix, and Laplacian matrix for this graph are as follows −

Adjacency matrix A:
[[0, 1, 1],
 [1, 0, 1],
 [1, 1, 0]]

Degree matrix D:
[[2, 0, 0],
 [0, 2, 0],
 [0, 0, 2]]

Laplacian matrix L:
[[ 2, -1, -1],
 [-1,  2, -1],
 [-1, -1,  2]]

Eigenvalues of the Laplacian: [0, 4, 4]

Conductance h(G) = 0.5

In this case, the conductance h(G) is 0.5, indicating that the graph can be split into two parts with relatively few edges between them. This result is consistent with the eigenvalue analysis, as the second-smallest eigenvalue 2 = 4 suggests moderate connectivity.

Applications of Cheeger's Inequality

Cheeger's Inequality has various applications across different fields, such as −

  • Graph Partitioning: Cheeger's Inequality can be used to identify bottlenecks or weakly connected components in a graph. It provides a bound on how easy it is to partition the graph into two disconnected parts.
  • Spectral Clustering: In spectral clustering, Cheeger's Inequality helps to determine the number of clusters in a graph. A small second-smallest eigenvalue indicates the presence of natural partitions in the graph.
  • Community Detection: By analyzing the Laplacian matrix and applying Cheeger's Inequality, it is possible to identify communities or clusters within large social or biological networks.
  • Network Robustness: Cheeger's Inequality can be used to measure the strongness of a network against failures or attacks. A higher 2 suggests that the network is well-connected and can tolerate more failures without becoming disconnected.

Example: Spectral Clustering and Graph Partitioning

Given a graph, one can compute the second-smallest eigenvalue 2 and use this value to identify how well the graph can be partitioned into two disconnected components.

If the value of 2 is small, it suggests that the graph can be easily split into two parts. A large value of 2 indicates that the graph is highly connected and resistant to splitting.

Laplacian matrix L:
[[ 4, -1,  0, -1],
 [-1,  4, -1,  0],
 [ 0, -1,  4, -1],
 [-1,  0, -1,  4]]

Eigenvalues of the Laplacian: [0, 2, 4, 4]

Conductance h(G) = 0.25

In this case, the conductance is 0.25, indicating that the graph is not easily split into two disconnected parts. The second-smallest eigenvalue 2 = 4 confirms the graph's high connectivity.

Relation to Cheeger's Constant

Cheeger's constant represents the infimum of the conductance across all subsets of vertices in the graph. Formally, it is defined as the minimum value of the conductance function −

h(G) = inf{S ⊆ V} (cut(S, V \ S) / min(vol(S), vol(V \ S)))
The infimum (often abbreviated as "inf") refers to the greatest value that is less than or equal to every element in a given set, essentially the "largest lower bound".

The Cheeger constant provides an important measure of the bottleneck of the graph, and Cheeger's Inequality relates this constant to the second-smallest eigenvalue of the Laplacian matrix. Thus, the inequality provides a spectral characterization of the graph's connectivity.

Advertisements