In this article I discuss matrices associated to graphs. As we will see, a graph can be represented as a matrix without any information loss. Hence, the properties of these matrices describe properties of the underlying graph.

Matrices associated to graphs

A graph $G = (V, E)$ s.t. $V = {v_{1}, \dots, v_{n}}$ and $E = {e_{1}, \dots, e_{m}}$ has several important associated matrices. For convenience, I often refer to vertex $v_{i}$ simply by its index ( $i$ ), and to an edge by the vertices it links (e.g., $i j$ ).

I will show examples on the following graph, named $G_{1}$ :

---
config:
  layout: elk
  look: handDrawn
---
graph LR
    vertex_1((1))
    vertex_2((2))
    vertex_3((3))
    vertex_4((4))

    vertex_1 === vertex_2
    vertex_1 === vertex_3
    vertex_1 === vertex_4
    vertex_2 === vertex_3

Degree matrix

Vertex degree is ised to define the degree matrix $D$ is a diagonal $n \times n$ matrix such that $D_{i i} = \deg i$ , and 0 elsewhere. For instance, for $G_{1}$ :

D (G_{1}) = [\begin{matrix} 3 & 0 & 0 & 0 \\ 0 & 2 & 0 & 0 \\ 0 & 0 & 2 & 0 \\ 0 & 0 & 0 & 1 \end{matrix}]

Incidence matrix

Incidence is used to define the incidence matrix $Q$ , a $n \times m$ matrix such that $Q_{i j}$ equals:

If $G$ is directed:
- $0$ if vertex $i$ and edge $e_{j}$ are not incident
- $1$ if edge $e_{j}$ originates at vertex $i$
- $- 1$ if edge $e_{j}$ terminates at vertex $i$
If $G$ is undirected:
- If $Q$ is unoriented:
  - $0$ if vertex $i$ and edge $e_{j}$ are not incident
  - $1$ otherwise
- If $Q$ is oriented: we pick an orientation of the graph, and use the incidence matrix of the resulting directed graph.

Adjacency matrix

Adjacency is used to define the adjacency matrix $A$ , a matrix $n \times n$ such that the $A_{i j}$ equals:

$0$ if vertices $i$ and $j$ are not adjacent (note that in simple graphs vertices are not self-adjacent)
$1$ otherwise

For $G_{1}$ :

A = [\begin{matrix} 0 & 1 & 1 & 1 \\ 1 & 0 & 1 & 0 \\ 1 & 1 & 0 & 0 \\ 1 & 0 & 0 & 0 \end{matrix}]

The adjacency matrix relates to the concept of paths in an unweighted graph: $(A^{k})_{i j}$ represents the number of paths of length $k$ from vertex $i$ to vertex $j$ . In a weighted graph, it represents the sum of products of weights. For instance, if edge weights represent transition probabilities, $(A^{k})_{i j}$ represents the probability of starting a walk at node $i$ and ending at node $j$ after $k$ steps.

Laplacian matrix

The Laplacian matrix $L$ is a $n \times n$ matrix such that the $L_{i j}$ equals::

For $i \neq j$ :
- $0$ if vertices $i$ and $j$ are not adjacent
- $- 1$ otherwise
For $i = j$ , the degree of $i$ .

More concisely, $L = D - A$ . Or, given any oriented incidence matrix $Q (G)$ , $L = Q Q^{T}$ .

For $G_{1}$ :

L = D - A = [\begin{matrix} 3 & - 1 & - 1 & - 1 \\ - 1 & 2 & - 1 & 0 \\ - 1 & - 1 & 2 & 0 \\ - 1 & 0 & 0 & 1 \end{matrix}]

The Laplacian relates to the connectedness of a graph, giving rise to spectral graph theory. It also is connected to flows. The diagonal entries represent the total outflow capacity from a vertex, while off-diagonal entries encode pairwise connection strengths.

Normalized Laplacian matrices

The presence of hubs results in large diagonal entries in the Laplacian. There are normalized versions of the Laplacian that downweigh such vertices by dividing the entries by the vertex degree.

The symmetrically normalized Laplacian $L_{sym}$ is a symmetric matrix derived as follows:

L_{sym} = D^{- 1 / 2} L D^{- 1 / 2}

The random walk normalized Laplacian $L_{rw}$ is a matrix closely related to random walks that is derived as follows:

L_{rw} = D^{- 1} L

Spectral graph theory

Spectral graph theory studies how the eigenvalues and eigenvectors of a graph’s associated matrices relate to its properties. Looking more closely at two of the matrices described above, we can see they have interesting properties:

If $G$ is undirected, $A$ is both real and symmetric. Hence, it is diagonalizable and has only real values.
Since for an undirected graph both $D$ and $A$ are symmetric, $L$ is also real and symmetric. In fact, $L$ is positive semi-definite. This implies that $L$ ’s eigenvalues are not only real, but also non-negative.

Spectral graph theory often focuses on studying the eigenvalues of the Laplacian.

Connectivity of the graph

The eigenvectors of $L$ are closely related to the connectivity of its associated graph.

A simple, but ultimately insightful property of $L$ is that, for an undirected graph, the sum over the rows or the columns equals 0. In other words, multiplying $L$ by an all-ones vector $1$ results in the zero vector. This tells us that $L$ has an eigenvalue of 0, corresponding to the eigenvector $1$ . Separately, linear algebra tells us that since $L$ is real and symmetric, it has real eigenvalues and orthogonal eigenvectors. And since $L$ is positive semi-definite, its eigenvalues are non-negative. As we have just seen, the first eigenvalue, $λ_{1}$ , of $L$ is 0, corresponding to the $1$ eigenvector. If a vector has multiple components, $L$ is block diagonal. This makes it easy to see that the indicator vectors, representing the membership of each vertex to one of the components, are eigenvectors with an eigenvalue of 0. This highlights another important property of the Laplacian: given an undirected graph, the multiplicity of the eigenvalue 0 of $L$ equals the number of components. Conversely, for a connected graph, $λ_{2} > 0$ . (The second smallest eigenvalue is sometimes called the Fiedler eigenvalue.)

More generally, less smooth eigenvectors (i.e., those in which consecutive elements change sharply) indicate a less connected. Equivalently, smaller eigenvalues correspond to smoother eigenvectors, and hence to better connected graphs.

Spectral clustering

The goal of spectral clustering is finding a partition of the graph into $k$ groups such that the are densely/strongly connected with each other, and sparsely/weakly connected to the others. (If we consider random walks, spectral clustering seeks a partition of the graph such that a random walker tends to stay within each partition, rarely shifting between disjoint sets.) An spectral clustering algorithm, in which seek to find k clusters, looks as follows:

\begin{algorithm}
\caption{Spectral Clustering}
\begin{algorithmic}[1]
\PROCEDURE{GraphSpectralClustering}{$$A, k$$}
    \STATE $$n \gets \text{number of nodes (rows in A)}$$

    \STATE Compute degree matrix $$D$$ where $$D[i,i] = \sum_{j=1}^n A[i,j]$$
    \STATE $$D_{\text{sqrt-inv}} \gets \text{diag}(1/\sqrt{D[i,i]})$$
    \STATE $$L_{\text{sym}} \gets I - D_{\text{sqrt-inv}} A D_{\text{sqrt-inv}}$$ \COMMENT{Symmetric normalized Laplacian}

    \STATE Compute first $$k$$ eigenvectors $$u_1, \ldots, u_k$$ of $$L_{\text{sym}}$$
    \STATE Form matrix $$U \in \mathbb{R}^{n \times k}$$ with columns $$u_1, \ldots, u_k$$

    \FOR{$$i = 1$$ \TO $$n$$}
        \STATE $$U[i] \gets U[i] / \|U[i]\|$$ \COMMENT{Row normalization}
    \ENDFOR

    \STATE $$\text{labels} \gets \text{KMeans}(U, k)$$ \COMMENT{Cluster embedded nodes}
    \RETURN $$\text{labels}$$
\ENDPROCEDURE
\end{algorithmic}
\end{algorithm}

Note: spectral clustering is often applied as a clustering technique on datasets. The aim is to divide the observation into $k$ groups based on their pairwise similarities. In that case, the first step consists on obtaining the graph. It will be a complete weighted graph in which the vertices are the different observations and the edges are weighted according to the similarity between each pair of vertices, as measured by an arbitrary function.

Graphs and Linear Algebra