The Graph Laplacian and Spectral Clustering

Grouping stocks into clusters by how similarly they behave — sector-like groupings discovered from the data itself, not from an official sector label — is hard for standard clustering methods when the natural groups aren't neat round blobs but instead form chains, rings, or oddly shaped clusters in similarity space. K-means and similar methods struggle here because they assume clusters look like circles around a center. The graph Laplacian sidesteps that assumption entirely, turning clustering into an eigenvalue problem on a matrix built from pairwise similarities.

An analogy: a spring network settling into its natural groups

Imagine every pair of assets connected by a spring, whose stiffness is proportional to how correlated they are — tightly co-moving assets get a stiff spring, unrelated ones get a weak or absent spring. If you let this whole spring network vibrate, its natural, lowest-energy vibration patterns will tend to move tightly-connected clusters together as a block, with looser connections between blocks bending more easily. The graph Laplacian is the matrix that governs exactly this spring network's vibration, and its lowest eigenvectors are precisely those natural, cluster-revealing vibration patterns.

The math, one symbol at a time

Given a similarity (weighted adjacency) matrix $W$ where $W_{ij}$ measures how strongly nodes $i$ and $j$ are connected, define the degree matrix $D$ (diagonal, with $D_{ii} = \sum_j W_{ij}$, the total connection strength of node $i$), and the graph Laplacian

$$L = D - W.$$

In words: $L$'s diagonal holds each node's total connectivity, and its off-diagonal entries subtract off the specific pairwise strengths. The Laplacian's key property, via the quadratic form (see Quadratic Forms): $x^\top L x = \frac{1}{2}\sum_{i,j} W_{ij}(x_i - x_j)^2$ — a sum that's small exactly when strongly-connected nodes $i,j$ are assigned similar values $x_i \approx x_j$. Spectral clustering finds the eigenvectors of $L$ with the smallest eigenvalues (excluding the trivial all-ones eigenvector with eigenvalue 0) and uses their entries as coordinates to cluster nodes — nodes that the low-energy eigenvectors treat similarly end up grouped together.

Worked example 1: two tightly connected pairs, weakly linked

Four assets: 1–2 strongly correlated (weight 5), 3–4 strongly correlated (weight 5), and only a weak link 2–3 (weight 0.1). $D = \text{diag}(5, 5.1, 5.1, 5)$, and $L = D - W$ has this near-block structure. Its eigenvalues come out approximately $0, 0.1, 5, 10$ — the second-smallest eigenvalue (0.1, not counting the trivial 0) is tiny precisely because the graph is almost two disconnected pieces joined by one weak link; the corresponding eigenvector takes one sign on nodes {1,2} and the opposite sign on {3,4}, directly revealing the two clusters by the sign pattern alone.

Worked example 2: reading off clusters from eigenvector signs

Continuing the example, the second-smallest eigenvector might come out approximately $(0.5, 0.49, -0.49, -0.5)$. Thresholding at zero: nodes 1 and 2 (positive) form one cluster, nodes 3 and 4 (negative) form the other — recovering exactly the intended grouping, using nothing but an eigenvector, with no assumption that clusters are round or evenly sized. If a third, genuinely separate cluster existed, the third-smallest eigenvector would typically reveal it, which is why spectral clustering with $k$ clusters generally uses the $k$ smallest nontrivial eigenvectors together as coordinates, then runs a simple clustering step (like k-means) on those coordinates instead of on the raw data.

What this means in practice

Spectral clustering via the graph Laplacian is used to discover data-driven groupings of assets from a correlation or similarity network — clusters that don't need to be predefined by industry classification and can capture co-movement patterns that cut across official sectors. It's related to hierarchical clustering methods used in portfolio construction (see Nested Clustered Optimization) and, more broadly, to any method that treats a correlation matrix as a graph rather than just a table of numbers.