Spectral Modularity Optimization

• Spectral modularity optimization is a technique that reformulates the modularity objective into a quadratic or trace form, then relaxes it to an eigenvalue problem for graph partitioning.
• It employs eigenvector computations and methods like k-means clustering to efficiently extract community structures in undirected, weighted, and multilayer networks.
• Extensions include adjustments via a resolution parameter, detectability thresholds, and differentiable formulations, enhancing the robustness of network analysis.

Spectral modularity optimization is a family of methods for graph partitioning and community detection in which a modularity objective is rewritten as a quadratic or trace form and then relaxed to an eigenvalue problem. For an undirected graph $G=(V,E)$ with adjacency matrix $A$, degrees $k_i=\sum_j A_{ij}$, and total number of edges $m=|E|$, the modularity matrix is $B_{ij}=A_{ij}-\frac{k_i k_j}{2m}$; with a resolution parameter $\gamma$, it becomes $B_{ij}=A_{ij}-\gamma \frac{k_i k_j}{2m}$. In this framework, maximizing modularity over discrete community assignments is approximated by leading-eigenvector or leading-eigensubspace computations, followed by sign-thresholding, $k$-means, or vector partitioning (Kawamoto et al., 2015, Zhang et al., 2015).

1. Objective functions and matrix formulations

The standard Newman–Girvan modularity for a partition $\{c_i\}$ is

$$Q \;=\; \frac{1}{2m}\sum_{i,j=1}^n\Bigl[A_{ij}-\frac{d_i\,d_j}{2m}\Bigr]\delta(c_i,c_j),$$

with $A$0. Equivalently, defining the modularity matrix

$A$1

and a one-hot indicator matrix $A$2, one has

$A$3

For a two-way partition encoded by spins $A$4, the same objective becomes a quadratic form in $A$5 (Zhuang et al., 25 Mar 2025).

A more general statement is that modularity maximization with the resolution parameter offers a unifying framework of graph partitioning. In that setting,

$A$6

where $A$7. This introduces an explicit scale parameter into the objective while preserving the same spectral kernel form (Kawamoto et al., 2015).

The same trace-maximization viewpoint extends to weighted graphs and to related objectives. For an edge-weighted graph $A$8, Bolla writes modularity as

$A$9

and introduces the normalized modularity matrix

$k_i=\sum_j A_{ij}$0

A further generalization is the “modularity Laplacian” formulation, which places both the modularity $k_i=\sum_j A_{ij}$1-measure and the modularity density $k_i=\sum_j A_{ij}$2-measure into a common trace optimization with orthogonality and nonnegativity constraints (Bolla, 2013, Jiang et al., 2011).

2. Spectral relaxation and extraction of partitions

Maximizing modularity over discrete assignments is NP-hard in general, so the standard spectral relaxation replaces discrete variables by real ones. In the two-way case one approximately maximizes $k_i=\sum_j A_{ij}$3 under a spherical constraint, choosing $k_i=\sum_j A_{ij}$4 to be the eigenvector associated with the largest eigenvalue of $k_i=\sum_j A_{ij}$5. A bisection is then obtained by the sign rule: assign $k_i=\sum_j A_{ij}$6 to community $k_i=\sum_j A_{ij}$7 if $k_i=\sum_j A_{ij}$8, and to $k_i=\sum_j A_{ij}$9 if $m=|E|$0. If all components have the same sign, the method returns the trivial unpartitioned solution (Kawamoto et al., 2015).

For $m=|E|$1-way clustering, the relaxation is usually written with a real matrix $m=|E|$2 or $m=|E|$3 subject to orthonormality constraints. By the variational Rayleigh–Ritz principle, the maximum is attained when the columns span the top-$m=|E|$4 eigenspace of the modularity operator. In the weighted-graph setting, the row-vectors

$m=|E|$5

derived from $m=|E|$6 are used as vertex representatives, and a weighted $m=|E|$7-means step produces a partition with low variance when a spectral gap is present (Bolla, 2013).

Direct multiway spectral modularity optimization avoids repeated bisection by mapping modularity maximization to a vector partitioning problem. Retaining the top $m=|E|$8 positive eigenvalues and defining

$m=|E|$9

one obtains the relaxed approximation

$B_{ij}=A_{ij}-\frac{k_i k_j}{2m}$0

This yields a max-sum vector partitioning problem in $B_{ij}=A_{ij}-\frac{k_i k_j}{2m}$1. The reported comparison shows that vector partitioning directly approximates modularity and yields higher normalized mutual information particularly as community-size imbalance increases (Zhang et al., 2015).

3. Resolution parameter, detectability, and phase transitions

In sparse stochastic block models, spectral modularity optimization exhibits a detectability transition. For the planted 2-block model with $B_{ij}=A_{ij}-\frac{k_i k_j}{2m}$2 and $B_{ij}=A_{ij}-\frac{k_i k_j}{2m}$3, define

$B_{ij}=A_{ij}-\frac{k_i k_j}{2m}$4

The leading eigenvector has nonzero overlap with the planted partition if and only if

$B_{ij}=A_{ij}-\frac{k_i k_j}{2m}$5

Equivalently, the detectability threshold is

$B_{ij}=A_{ij}-\frac{k_i k_j}{2m}$6

or, in terms of $B_{ij}=A_{ij}-\frac{k_i k_j}{2m}$7,

$B_{ij}=A_{ij}-\frac{k_i k_j}{2m}$8

Below this threshold the leading eigenvector carries no information, whereas above it one recovers communities with a positive fraction of correctly classified nodes (Kawamoto et al., 2015).

A central result is that the condition for the emergence of a community-informed eigenvector does not depend on the resolution parameter $B_{ij}=A_{ij}-\frac{k_i k_j}{2m}$9. In the replica-symmetric calculation, the boundary between detectable and undetectable phases is determined by a saddle-point equation in which neither $\gamma$0 nor $\gamma$1 depends on $\gamma$2. Hence the detectability threshold is universal in $\gamma$3. This means that, as long as one stays in the detectable regime, varying $\gamma$4 will not destroy detectability, although it will shift the numerical values of the eigenvalues (Kawamoto et al., 2015).

The same analysis identifies a distinct first-order transition to an unpartitioned phase when $\gamma$5 is sufficiently small. In the random $\gamma$6-regular graph,

$\gamma$7

and

$\gamma$8

implies that no nontrivial partition exists even if the planted structure is above the information-theoretic threshold. At $\gamma$9, the eigenvalue of the uniform vector $B_{ij}=A_{ij}-\gamma \frac{k_i k_j}{2m}$0 crosses the bulk edge $B_{ij}=A_{ij}-\gamma \frac{k_i k_j}{2m}$1, and the system jumps from the detectable eigenvector to the trivial one. A common practical prescription is therefore to pick $B_{ij}=A_{ij}-\gamma \frac{k_i k_j}{2m}$2; a common choice is $B_{ij}=A_{ij}-\gamma \frac{k_i k_j}{2m}$3 (Kawamoto et al., 2015).

4. Normalized, parametrized, and signed spectral structures

The normalized modularity matrix provides a degree-balanced alternative to the unnormalized modularity operator. Its spectrum lies in $B_{ij}=A_{ij}-\gamma \frac{k_i k_j}{2m}$4, with $B_{ij}=A_{ij}-\gamma \frac{k_i k_j}{2m}$5 as a trivial eigenvalue, and a spectral gap between the $B_{ij}=A_{ij}-\gamma \frac{k_i k_j}{2m}$6 largest absolute-value eigenvalues and the remainder implies testability of the structural eigenvalues, of the corresponding eigensubspace, and of the sum of the inner variances of the $B_{ij}=A_{ij}-\gamma \frac{k_i k_j}{2m}$7 clusters obtained by weighted $B_{ij}=A_{ij}-\gamma \frac{k_i k_j}{2m}$8-means on the vertex representatives. In Bolla’s formulation, both cohesive clusters and anti-communities can be detected by including eigenvectors from both ends of the spectrum (Bolla, 2013).

A further generalization is the $B_{ij}=A_{ij}-\gamma \frac{k_i k_j}{2m}$9-parametrized normalized modularity matrix

$k$0

In heterogeneous graph models, this operator admits a spiked-random-matrix approximation, and there exists an optimal value $k$1 for which the detection of communities is best ensured. The same analysis shows that raw eigenvectors should be regularized before clustering by premultiplication with $k$2, because the informative low-rank spike is aligned with $k$3 rather than with $k$4 itself (Ali et al., 2016).

The resolution-parametrized modularity matrix also has a direct connection to the normalized Laplacian. For

$k$5

the corresponding normalized operator has spectrum

$k$6

where $k$7 are Laplacian eigenvalues. In this analysis, the maxima of the quadratic function with $k$8 as the kernel matrix always reside in the Fiedler space of the normalized graph Laplacian $k$9 or the null space of $\{c_i\}$0, or their combination, and the Fiedler value $\{c_i\}$1 marks the critical $\{c_i\}$2 value in the transition of candidate community configuration states between graph division and aggregation. The introduced Fiedler pseudo-set is identified as the de facto critical region for the state transition (Floros et al., 2023).

Spectral modularity methods can also be formulated to detect communities and anti-communities simultaneously. In that setting, positive extreme eigenvalues indicate communities, negative extreme eigenvalues indicate anti-communities, and the relevant invariant subspaces are localized by matrix angles based on Frobenius inner products. This extends the usual modularity narrative from exclusively positive structure to both large positive and large negative normalized modularity (Fasino et al., 2017).

5. Optimization schemes, refinements, and differentiable formulations

Several algorithmic lines refine the basic leading-eigenvector relaxation. One fast spectral algorithm searches for the partition with the maximum value of the modularity via the interplay of several refinement steps that include both agglomeration and division. Its workflow is spectral bisection, Fine Tuning, Final Tuning, and Agglomeration, repeated until the total $\{c_i\}$3 is no longer positive. The same work derives Erdős–Rényi null-model predictions for the mean and variance of $\{c_i\}$4 and defines the modularity effect size by the $\{c_i\}$5-score

$\{c_i\}$6

This places spectral modularity optimization in both an optimization and a statistical-effect-size setting (III et al., 2014).

A different tightening of the relaxation is successive spectral relaxation. Instead of thresholding the leading eigenvector in one shot, it iteratively fixes high-confidence vertices and re-solves a residual constrained problem for the remainder by a constrained power method. The residual objective

$\{c_i\}$7

is optimized under a norm constraint, and the method is reported as highly suitable for parallel execution, with nearly linear improvement in running speed when increasing the number of computing nodes (Li, 2018).

Nonnegative relaxation replaces the traditional unconstrained spectral relaxation by

$\{c_i\}$8

The explicit nonnegative constraint makes the solutions very close to the ideal community indicator matrix and allows direct node assignment. Because the near-orthogonal columns can be reformulated as posterior probability, the method can also be exploited to identify fuzzy or overlapping communities. The associated multiplicative update is a modularity-optimization analogue of Lee–Seung-style NMF updates (Jiang et al., 2011).

Recent work has made spectral modularity optimization fully differentiable. In the GyralNet setting, the cluster-assignment matrix is parameterized by a two-layer GCN with softmax output,

$\{c_i\}$9

and the loss is

$$Q \;=\; \frac{1}{2m}\sum_{i,j=1}^n\Bigl[A_{ij}-\frac{d_i\,d_j}{2m}\Bigr]\delta(c_i,c_j),$$0

Because every operation is differentiable, one can back-propagate through $$Q \;=\; \frac{1}{2m}\sum_{i,j=1}^n\Bigl[A_{ij}-\frac{d_i\,d_j}{2m}\Bigr]\delta(c_i,c_j),$$1 with respect to $$Q \;=\; \frac{1}{2m}\sum_{i,j=1}^n\Bigl[A_{ij}-\frac{d_i\,d_j}{2m}\Bigr]\delta(c_i,c_j),$$2 and thereby perform “spectral” modularity optimization entirely via gradient descent. The paper summarizes this as a fully differentiable “GCN$$Q \;=\; \frac{1}{2m}\sum_{i,j=1}^n\Bigl[A_{ij}-\frac{d_i\,d_j}{2m}\Bigr]\delta(c_i,c_j),$$3softmax$$Q \;=\; \frac{1}{2m}\sum_{i,j=1}^n\Bigl[A_{ij}-\frac{d_i\,d_j}{2m}\Bigr]\delta(c_i,c_j),$$4trace-objective+regularizer” pipeline (Zhuang et al., 25 Mar 2025).

6. Generalizations and application domains

Spectral modularity optimization is not restricted to edge-based, single-layer community detection. One generalization replaces edges by triangles as the building blocks of modules. The triangle modularity

$$Q \;=\; \frac{1}{2m}\sum_{i,j=1}^n\Bigl[A_{ij}-\frac{d_i\,d_j}{2m}\Bigr]\delta(c_i,c_j),$$5

reduces, for a binary split, to a quadratic form

$$Q \;=\; \frac{1}{2m}\sum_{i,j=1}^n\Bigl[A_{ij}-\frac{d_i\,d_j}{2m}\Bigr]\delta(c_i,c_j),$$6

where $$Q \;=\; \frac{1}{2m}\sum_{i,j=1}^n\Bigl[A_{ij}-\frac{d_i\,d_j}{2m}\Bigr]\delta(c_i,c_j),$$7 is the triangle-modularity matrix. The reported comparison states that the information reported by the analysis of modules of triangles complements the information of the classical modularity analysis (Serrour et al., 2010).

In multilayer networks with multiple aspects, the modularity function can be flattened into a supra-adjacency representation and optimized spectrally. The mSpec method builds a supra-modularity matrix $$Q \;=\; \frac{1}{2m}\sum_{i,j=1}^n\Bigl[A_{ij}-\frac{d_i\,d_j}{2m}\Bigr]\delta(c_i,c_j),$$8, extracts its leading eigenvector, and uses recursive spectral bisection when the modularity gain is positive. The experiments on several empirical multilayer networks and an EEG testbed are reported to demonstrate the feasibility and reliable performance of the proposed method (Zhang et al., 2016).

Spectral modularity ideas also appear in graph design rather than only graph analysis. Lubin, Shore, and Ishakian formulate communication-network design as a non-convex program that maximizes the Laplacian spectral gap $$Q \;=\; \frac{1}{2m}\sum_{i,j=1}^n\Bigl[A_{ij}-\frac{d_i\,d_j}{2m}\Bigr]\delta(c_i,c_j),$$9 subject to a lower bound on $A$00, thereby balancing modularity and mixing time. Their optimization combines SDP representations with the Concave-Convex Procedure, and the resulting graphs are described as favoring “liaisons” rather than single brokers (Lubin et al., 2013).

Application-specific pipelines continue to combine modularity optimization with spectral machinery. DynMSA uses Random Matrix Theory, modularity optimisation, and spectral clustering to identify clusters of stocks with high intra-cluster correlations and low inter-cluster correlations, while GyralNet subnetwork partitioning combines topological structural similarity and DTI-derived connectivity patterns with differentiable spectral modularity optimization. The reported outcomes include stable clusters, regime-change detection, and biologically meaningful cortical organization, which suggests that spectral modularity optimization now functions as a general optimization template across sparse graphs, weighted graphs, heterogeneous graphs, multilayer networks, dynamic correlation networks, and differentiable graph representation pipelines (Wirth et al., 2024, Zhuang et al., 25 Mar 2025).