Multilayer Modularity Function

• Multilayer modularity function is a quality metric that generalizes classical modularity to networks with multiple layers, measuring both intra-layer connectivity and inter-layer coupling.
• The function uses a resolution parameter (γ) to control community granularity and an inter-layer coupling (ω) to enforce label consistency, enabling analyses over dynamic or multiplex networks.
• Various null models and heuristic algorithms, like the generalized Louvain method, are employed to optimize this NP-hard problem and robustly detect statistically significant communities.

A multilayer modularity function is a quality metric that generalizes classical modularity to networks with multiple layers, each potentially corresponding to a different type of relationship, time point, modality, or experimental condition. This function measures the statistical significance of observed community structure with respect to a null model, extended to account for both intra-layer connectivity and inter-layer correspondences. The multilayer modularity framework canonically incorporates explicit parameters controlling the resolution within each layer and the degree of coupling across layers, supporting a broad range of applications in network science, neuroscience, and temporal or multiplex complex systems.

1. Formal Definition and Key Parameters

The multilayer modularity quality function $Q$ is most generally expressed, following Mucha et al. (2010), as

$$Q = \frac{1}{2\mu} \sum_{i,j=1}^N \sum_{s,r=1}^T \left[ (A_{ij}^{s} - \gamma_s P_{ij}^{s}) \delta_{s r} + \delta_{i j} \omega_j^{sr} \right] \delta(g_i^s, g_j^r)$$

where:

• $N$ = number of nodes per layer
• $T$ = number of layers
• $A_{ij}^s$ = observed edge weight between $i$ and $j$ in layer $s$
• $P_{ij}^s$ = expected edge weight from a layer-specific null model
• $\gamma_s$ = resolution parameter for layer $$Q = \frac{1}{2\mu} \sum_{i,j=1}^N \sum_{s,r=1}^T \left[ (A_{ij}^{s} - \gamma_s P_{ij}^{s}) \delta_{s r} + \delta_{i j} \omega_j^{sr} \right] \delta(g_i^s, g_j^r)$$0
• $$Q = \frac{1}{2\mu} \sum_{i,j=1}^N \sum_{s,r=1}^T \left[ (A_{ij}^{s} - \gamma_s P_{ij}^{s}) \delta_{s r} + \delta_{i j} \omega_j^{sr} \right] \delta(g_i^s, g_j^r)$$1 = inter-layer coupling weight connecting node $$Q = \frac{1}{2\mu} \sum_{i,j=1}^N \sum_{s,r=1}^T \left[ (A_{ij}^{s} - \gamma_s P_{ij}^{s}) \delta_{s r} + \delta_{i j} \omega_j^{sr} \right] \delta(g_i^s, g_j^r)$$2 in layer $$Q = \frac{1}{2\mu} \sum_{i,j=1}^N \sum_{s,r=1}^T \left[ (A_{ij}^{s} - \gamma_s P_{ij}^{s}) \delta_{s r} + \delta_{i j} \omega_j^{sr} \right] \delta(g_i^s, g_j^r)$$3 to itself in layer $$Q = \frac{1}{2\mu} \sum_{i,j=1}^N \sum_{s,r=1}^T \left[ (A_{ij}^{s} - \gamma_s P_{ij}^{s}) \delta_{s r} + \delta_{i j} \omega_j^{sr} \right] \delta(g_i^s, g_j^r)$$4
• $$Q = \frac{1}{2\mu} \sum_{i,j=1}^N \sum_{s,r=1}^T \left[ (A_{ij}^{s} - \gamma_s P_{ij}^{s}) \delta_{s r} + \delta_{i j} \omega_j^{sr} \right] \delta(g_i^s, g_j^r)$$5 = community label of node $$Q = \frac{1}{2\mu} \sum_{i,j=1}^N \sum_{s,r=1}^T \left[ (A_{ij}^{s} - \gamma_s P_{ij}^{s}) \delta_{s r} + \delta_{i j} \omega_j^{sr} \right] \delta(g_i^s, g_j^r)$$6 in layer $$Q = \frac{1}{2\mu} \sum_{i,j=1}^N \sum_{s,r=1}^T \left[ (A_{ij}^{s} - \gamma_s P_{ij}^{s}) \delta_{s r} + \delta_{i j} \omega_j^{sr} \right] \delta(g_i^s, g_j^r)$$7
• $$Q = \frac{1}{2\mu} \sum_{i,j=1}^N \sum_{s,r=1}^T \left[ (A_{ij}^{s} - \gamma_s P_{ij}^{s}) \delta_{s r} + \delta_{i j} \omega_j^{sr} \right] \delta(g_i^s, g_j^r)$$8 = normalizing constant, $$Q = \frac{1}{2\mu} \sum_{i,j=1}^N \sum_{s,r=1}^T \left[ (A_{ij}^{s} - \gamma_s P_{ij}^{s}) \delta_{s r} + \delta_{i j} \omega_j^{sr} \right] \delta(g_i^s, g_j^r)$$9
• $N$0, $N$1 = Kronecker deltas

The double sum includes all pairs of node-layer indices and rewards both within-layer modular structure and persistence of community labels across layers. The effect of $N$2 is to control the granularity of communities detected in each layer, while $N$3 enforces the persistence of labels (community consistency) for each node across layers (Esfahlani et al., 2021, Bazzi et al., 2014).

2. Roles and Effects of Resolution and Coupling Parameters

Resolution Parameter ($N$4)

• Scales the null-model term $N$5; higher $N$6 favors detection of smaller, finer communities, mitigating the classical resolution limit.
• $N$7: under-penalizes null, resulting in larger/coarser modules; $N$8: over-penalizes null, yielding smaller/finer modules (Esfahlani et al., 2021, Bazzi et al., 2014). Scanning $N$9 reveals a multiscale hierarchy of community structure.

Inter-layer Coupling ($T$0)

• Couples the same node across different layers, rewarding consistency of community labels.
• $T$1: decouples layers, each independently partitioned.
• Large $T$2: enforces identical partitions across layers ("locking").
• Intermediate $T$3: allows partial alignment with "wiggling" boundaries (Esfahlani et al., 2021, Bazzi et al., 2014). In time-ordered/dynamic scenarios, $T$4 is typically nonzero only for adjacent layers; in other contexts, it can couple all node-layer pairs.
• Crucially, there exists a resolution limit on the maximum detectable $T$5; beyond an upper bound $T$6, true community changes across layers are "invisible" to modularity maximization (Vaiana et al., 2018).

Parameter	Typical Role	Effect When Increased
$T$7	Layer resolution	Smaller, finer communities
$T$8	Inter-layer coupling	Labels more persistent across layers

3. Null Models and Structural Variants

The choice of null model $T$9 determines the expected edge structure against which observed modularity is compared. The classical choices include:

• Newman–Girvan (degree-corrected configuration): $A_{ij}^s$0
• Uniform ("surprise" model): $A_{ij}^s$1
• Signed variants: for handling positive and negative weights (Bazzi et al., 2014, Zhang et al., 2016)
• Spatial or statistical blockmodel-based extensions, including SBMs and degree-corrected Poisson models for likelihood-based modularity (Pamfil et al., 2018, Paul et al., 2016)

The null model influences both the modularity matrix and the range of structural patterns captured.

4. Optimization and Algorithmic Implementation

Maximizing multilayer modularity is a combinatorial, NP-hard problem with a high degree of degeneracy. The standard approach is a greedy heuristic based on the generalized Louvain or GenLouvain method, often adapted to the supra-adjacency matrix representation. Key algorithmic strategies include:

• Multiple independent runs with random initializations to explore near-degenerate maxima.
• Construction of co-assignment tensors/matrices by aggregating results from multiple runs.
• Extraction of consensus partitions post hoc using thresholding and re-clustering (Esfahlani et al., 2021, Bazzi et al., 2014).
• Use of belief-propagation (BP) algorithms for modularity maximization, providing both hard and soft community assignments, and robustly identifying non-random structure via convergence diagnostics (Weir et al., 2019).

Recent contributions include spectral and variational methods based on gradient flows and total variation regularization, capable of leveraging fast eigendecomposition to accelerate community detection and facilitating energy-minimization perspectives (Bergermann et al., 2024).

5. Resolution Limits and Theoretical Insights

A central theoretical result is the existence of coupled resolution bounds: for each node group whose community assignment changes across layers, there is an upper bound $A_{ij}^s$2 on the coupling parameter $A_{ij}^s$3 beyond which modularity cannot detect this change (Vaiana et al., 2018). Explicitly,

$A_{ij}^s$4

where $A_{ij}^s$5 is the set of nodes changing community, and $A_{ij}^s$6 depends on the number and sizes of communities. This generalizes the single-layer Fortunato–Barthélemy limit and establishes a tradeoff between within-layer resolution and inter-layer smoothing.

Further, the linear dependence of $A_{ij}^s$7 on $A_{ij}^s$8 reveals that increasing in-layer resolution allows proportionally stronger coupling without loss of change detectability—up to a point.

A practical implication is that for layers with low structural connectivity, even small $A_{ij}^s$9 can induce spurious label persistence, and adjusting $i$0 or $i$1 does not always recover meaningful cross-layer change (Vaiana et al., 2018).

6. Extensions, Specializations, and Parameter-Free Alternatives

The multilayer modularity framework has been extended in several directions:

• Redundancy-based and projection-based modularity: Parameter-free approaches have been proposed where the resolution and coupling terms are estimated directly from redundancy (multi-layer support) and projection (community overlap across layers) (Amelio et al., 2019, Amelio et al., 2017). Here, the resolution $i$2 is data-dependent, reducing penalization for communities with rich multilayer support.
• Multiobjective and Pareto-filtered maximization: Instead of aggregating all objectives, variance-aware multiobjective Louvain heuristics maximize both the mean and variance of layerwise modularities, preserving information on consistency and heterogeneity across multiplex layers (Venturini et al., 2021).
• Supra-adjacency and aspect-aware modularity: Unified frameworks have been proposed using block-matrix representations for networks with multiple aspects, such as time and modality, and encompassing temporal, multiplex, and spatial layers (Zhang et al., 2016).
• Likelihood-based modularities: Direct equivalence with degree-corrected multilayer SBMs can be established, leading to modularity functions that optimize the profile log-likelihood under various null models (independent-layers, shared-degree, etc.) (Pamfil et al., 2018, Paul et al., 2016).

7. Practical Usage and Empirical Considerations

Application-dependent strategies are critical in selecting the null model, tuning $i$3 and $i$4, and interpreting modular partitions. In neuroscientific contexts, multilayer modularity maximization enables detection and tracking of dynamic brain modules across time, tasks, subjects, or modalities (Esfahlani et al., 2021). Systematic parameter sweeps (grid search) over $i$5 and $i$6, combined with consensus clustering and post-processing for partition persistence, are standard practice.

Empirical studies have revealed:

• $i$7 controls the spatial or topological scale of detected communities, with larger values revealing finer modular structure.
• $i$8 mediates the degree of temporal or cross-modal label persistence, with strong coupling suppressing genuine dynamic or functional heterogeneity (Esfahlani et al., 2021, Bazzi et al., 2014).
• The use of parameter-free or data-driven resolution and coupling terms yields modularity measures that adapt to real structural redundancies and cross-layer community support, avoiding arbitrary choices and improving interpretability (Amelio et al., 2019, Amelio et al., 2017).
• Optimization landscapes are degenerate; consensus/frequentist aggregations and rigorous evaluation across runs are essential for reproducibility.

• Zamani Esfahlani et al., "Modularity maximization as a flexible and generic framework for brain network exploratory analysis" (Esfahlani et al., 2021)
• Bazzi et al., "Community detection in temporal multilayer networks, with an application to correlation networks" (Bazzi et al., 2014)
• Paul et al., "Resolution Limits for Detecting Community Changes in Multilayer Networks" (Vaiana et al., 2018)
• Zhang et al., "Community Detection Using Multilayer Edge Mixture Model" (Zhang et al., 2016)
• Weir et al., "Multilayer Modularity Belief Propagation To Assess Detectability Of Community Structure" (Weir et al., 2019)
• Bergermann and Stoll, "Gradient flow-based modularity maximization for community detection in multiplex networks" (Bergermann et al., 2024)
• Pamfil et al., "Relating modularity maximization and stochastic block models in multilayer networks" (Pamfil et al., 2018)
• Amelio and Tagarelli, "Modularity in Multilayer Networks using Redundancy-based Resolution and Projection-based Inter-Layer Coupling" (Amelio et al., 2019)
• Amelio and Tagarelli, "Revisiting Resolution and Inter-Layer Coupling Factors in Modularity for Multilayer Networks" (Amelio et al., 2017)
• Liu et al., "Modularity in Complex Multilayer Networks with Multiple Aspects: A Static Perspective" (Zhang et al., 2016)
• Paul and Chen, "Null Models and Community Detection in Multi-Layer Networks" (Paul et al., 2016)
• Aldecoa et al., "A Variance-aware Multiobjective Louvain-like Method for Community Detection in Multiplex Networks" (Venturini et al., 2021)