Supra-Laplacian in Multilayer Networks

• Supra-Laplacian is a block-structured matrix that generalizes the classical Laplacian to capture both intra-layer and inter-layer connections in multiplex and temporal networks.
• It exhibits a Kronecker-sum spectral structure, revealing eigenvalue multiplicities and shifts that control diffusion dynamics, synchronizability, and structural phase transitions.
• The operator underpins spatio-temporal encoding in graph neural networks, with efficient iterative eigensolvers enabling scalable analysis of large multiplex systems.

The supra-Laplacian is a block-structured, symmetric matrix that generalizes the classical Laplacian operator from single-layer graphs to multiplex, multilayer, or temporal networks. It encodes intra-layer (within-snapshot) structure as well as inter-layer (cross-snapshot or cross-layer) connectivity, thereby capturing the full spectrum of spatial and temporal dependencies. The supra-Laplacian is central in the spectral theory of multilayer networks, providing the foundational tool for the analysis of diffusion, synchronization, structural phase transitions, and the construction of spatio-temporal positional encodings in graph neural architectures.

1. Formal Definition and Block Structure

Given a multiplex or multilayer network with $M$ layers, each with $N$ nodes, the supra-Laplacian $\mathcal{L}$ is constructed by arranging the node-layer pairs as a single set of $MN$ entities. For each layer $\alpha$, let $W^{(\alpha)}$ denote its $N \times N$ adjacency (or strength) matrix, and $L^{(\alpha)} = S^{(\alpha)} - W^{(\alpha)}$ its (combinatorial) Laplacian, where $S^{(\alpha)}$ is the diagonal strength matrix. Inter-layer connections are described by an $M \times M$ adjacency $N$0 and Laplacian $N$1.

The general supra-Laplacian is given by

$N$2

In block notation, each block $N$3 (of size $N$4) is: $N$5 This formalism encompasses models for node-aligned multiplex networks (Sole-Ribalta et al., 2013), temporal multilayer representations (Galron et al., 2 Jun 2025, Karmim et al., 2024), and constant block Jacobi models for periodic or chain-coupled temporal layers (Kuncheva et al., 2023).

2. Spectral Properties and Kronecker Structure

The supra-Laplacian exhibits a Kronecker-sum spectral structure, allowing partial decoupling of its eigenproblem. Specifically, if $N$6 is an eigenpair of $N$7, then $N$8 is an eigenpair of $N$9. Thus, the $\mathcal{L}$0 eigenvalues of $\mathcal{L}$1 each appear with multiplicity $\mathcal{L}$2 in $\mathcal{L}$3. Other eigenmodes correspond to perturbations of the intra-layer spectra, leading to $\mathcal{L}$4 additional eigenvalues that interpolate from the union of the intra-layer Laplacians (for vanishingly weak coupling) to the spectrum of the average or superposed network under strong inter-layer coupling (Sole-Ribalta et al., 2013, Gomez et al., 2012).

For the temporally ordinal case with constant nearest-neighbor coupling $\mathcal{L}$5, the supra-Laplacian is block tridiagonal, and discrete Fourier analysis in the layer index yields a spectrum as a union of $\mathcal{L}$6 shifted single-layer spectra (Kuncheva et al., 2023).

3. Structural Phases and Eigengaps

The eigenspectrum of the supra-Laplacian reveals distinct structural regimes as the ratio $\mathcal{L}$7 of inter-layer to intra-layer coupling is varied (Cozzo et al., 2016):

• In the layer-dominated phase ($\mathcal{L}$8), the second eigenvalue is $\mathcal{L}$9 and inter-layer coupling dominates the fastest relaxation processes.
• In the genuine multiplex phase ($MN$0), nontrivial spectral gaps appear, and neither intra- nor inter-layer structures singularly dominate.
• In the aggregate-dominated phase ($MN$1), exactly $MN$2 eigenvalues remain bounded (approaching the spectrum of the aggregate Laplacian), while the others diverge linearly in $MN$3.

For $MN$4 identical layers, $MN$5 and $MN$6, where $MN$7 are the second-smallest and largest eigenvalues of the individual layer Laplacian. Notably, the interlacing property ensures that quotient Laplacians (aggregate and layer-network reductions) control the appearance, location, and size of eigengaps and thus different dynamical time scales.

4. Analytical Asymptotics and Physical Consequences

Asymptotic analyses clarify how the spectrum controls physical processes:

• Diffusion: For $MN$8, the slowest decay mode is $MN$9, so the diffusion time scale is $\alpha$0. For weak coupling ($\alpha$1), $\alpha$2, yielding $\alpha$3. For strong coupling ($\alpha$4), $\alpha$5 approaches the second eigenvalue of the average network, resulting in super-diffusion: the full multiplex can equilibrate faster than any individual layer (Gomez et al., 2012, Sole-Ribalta et al., 2013).
• Synchronizability: According to the Master-Stability-Function formalism, the stability of the fully synchronized state is controlled by the eigenratio $\alpha$6. The scaling of $\alpha$7 in both weak and strong coupling regimes determines system synchronizability (Sole-Ribalta et al., 2013).
• Quantum Entropy and Commute Times: Von Neumann entropy of the supra-Laplacian peaks in the multiplex phase, and commute times between node replicas converge to those on the aggregate in the strong-coupling regime (Cozzo et al., 2016).

5. Supra-Laplacian in Spatio-Temporal Graph Embeddings

The supra-Laplacian provides the spectral foundation for spatio-temporal positional encodings in temporal and dynamic graph learning architectures. For a temporal graph with $\alpha$8 snapshots of $\alpha$9 nodes, the combinatorial or normalized supra-Laplacian $W^{(\alpha)}$0 is constructed from the block adjacency that connects each node across consecutive time layers.

A key result shows that the $W^{(\alpha)}$1 smallest eigenvectors of $W^{(\alpha)}$2 solve the constrained minimization: $W^{(\alpha)}$3 with $W^{(\alpha)}$4 the $W^{(\alpha)}$5 Laplacian eigenvectors in each time slice, and $W^{(\alpha)}$6 weighting temporal smoothness (Galron et al., 2 Jun 2025). This objective interpolates between purely local (per-slice) positional embeddings and temporally coherent encodings. Projecting node-time pairs onto the leading supra-Laplacian eigenvectors yields a geometric embedding that faithfully captures both structural and temporal regularities, supporting effective spatial-temporal learning in Transformer-based or message-passing neural networks (Karmim et al., 2024, Galron et al., 2 Jun 2025).

Empirical work demonstrates that such encodings yield consistent improvements in link-prediction and representation tasks, with fast sparse eigensolvers (LOBPCG, Lanczos) making them practical for large graphs (Galron et al., 2 Jun 2025).

6. Perturbative Analysis, Block Jacobi Models, and Generalized Fiedler Vectors

In periodic or chain-coupled temporal/multilayer networks with constant inter-layer weights, the supra-Laplacian assumes a block-circulant Jacobi structure. The spectral problem is then reducible to a set of shifted single-layer problems via discrete Fourier diagonalization (Kuncheva et al., 2023). For small $W^{(\alpha)}$7, the near-zero eigenvalues and eigenvectors of the supra-Laplacian are well-approximated by linear combinations of the zero-modes of each layer, modulated by Fourier coefficients in the layer index. The generalization of the Fiedler vector to the multi-layer case consists of these low-frequency modes, controlling the minimal variation/smoothness across both layers and time, directly impacting diffusion mixing times and modularity-based community detection in multilayer graphs.

7. Algorithmic Aspects and Practical Computation

The eigendecomposition of the supra-Laplacian, in its dense form, scales cubically with graph size $W^{(\alpha)}$8, quickly becoming computationally expensive in large-scale, high-frequency temporal or multilayer scenarios. Iterative methods such as LOBPCG and Lanczos are therefore preferred for extracting the leading $W^{(\alpha)}$9 spectral modes, with warm-restarts from per-slice solutions leveraged for additional efficiency (Galron et al., 2 Jun 2025). In practice, the approximation quality of these iterative methods is sufficient for downstream tasks, matching exact eigendecomposition within 0.3% median AUC for link-prediction, and showing 10–56$N \times N$0 improvements in runtime (for $N \times N$1 up to $N \times N$2 across multiple real-world benchmarks).

A step-by-step algorithm for generating block-structured supra-Laplacian encodings suitable for spatio-temporal Transformers includes: pre-processing for node alignment and layer connectivity, efficient extraction of the low-lying spectrum, padding for isolated nodes, and concatenation of static node features with spectral embeddings (Karmim et al., 2024, Galron et al., 2 Jun 2025).

---

References:

• Spectral properties of the Laplacian of multiplex networks (Sole-Ribalta et al., 2013)
• Diffusion dynamics on multiplex networks (Gomez et al., 2012)
• Characterization of multiple topological scales in multiplex networks through supra-Laplacian eigengaps (Cozzo et al., 2016)
• Supra-Laplacian Encoding for Transformer on Dynamic Graphs (Karmim et al., 2024)
• Understanding and Improving Laplacian Positional Encodings For Temporal GNNs (Galron et al., 2 Jun 2025)
• Spectral properties of the Laplacian of temporal networks following a constant block Jacobi model (Kuncheva et al., 2023)