A graph $G(V,E)$ has a vertex set $V$ and edge set $E$. Its matrix representations — adjacency $A$, degree $D$, and Laplacian $L = D - A$ — encode structure. The eigenvalues of $L$ reveal connectivity and clustering.
We use undirected graphs with no loops. The degree $d_i$ of vertex $v_i$ is the number of edges connected to it.
Properties: $A$ is real, symmetric ($a_{ij} = a_{ji}$), has zero diagonal ($a_{jj} = 0$), and real eigenvalues (symmetric). The row sum $[A]\vec{1}$ gives the degree vector.
$D$ is diagonal. Each entry $d_i$ is the degree of vertex $v_i$ — the sum of row $i$ of the adjacency matrix.
Reading entries of $L$ directly:
So to draw a graph from $L$: diagonal gives degrees, off-diagonal $-1$ entries indicate edges.
Property 4 is the most important: every row of $L$ sums to zero, because diagonal = sum of off-diagonal magnitudes. This guarantees $\lambda_1 = 0$ always.
A connected component is a maximal subgraph where every pair of vertices is connected by a path. If $G$ has $k$ components, then $L$ has exactly $k$ zero eigenvalues. For a connected graph: exactly one zero eigenvalue.
Named after Miroslav Fiedler (1989), who called this "algebraic connectivity". The sign pattern of the Fiedler vector separates vertices into two clusters based on graph structure — even for nearly-disconnected graphs.
This converts a data matrix into a graph. The threshold $\varepsilon$ controls how far apart two observations can be and still be considered "connected". Then spectral clustering on the Laplacian finds natural clusters.
Key difference from PCA: PCA uses the first (largest) $k$ eigenvectors. Spectral clustering uses the last (smallest) $k$ eigenvectors. This is because the Laplacian's smallest eigenvalues correspond to the least-connected structure — the natural clusters.
Since $w_{ij} \geq 0$ (non-negative weights) and $(u_i - u_j)^2 \geq 0$ always, the quadratic form is always non-negative. Therefore $L \succeq 0$.