Symmetries and Cluster Synchronization in Multilayer Networks

Introduction

• Heterogeneous components can interact through multiple connection types; power grids, for example, combine transmission and communication networks that both matter for understanding system failures.A cited study reports that increased coupling between those layers can reduce system vulnerability.Source: Introduction · p. 1
• Prior multilayer synchronization studies addressed complete synchronization with diffusive coupling, while many cluster-synchronization studies considered networks with one node type and one connection type.Source: Introduction · p. 1
• Cluster synchronization requires equitable clusters whose nodes receive equal total input from each cluster, while multilayer cases had received little study beyond a special experiment with homogeneous nodes.Source: Introduction · p. 2
• The paper defines and computes multilayer-network symmetries, distinguishing independent layer symmetries from dependent symmetries involving simultaneous node swaps across layers.It studies cluster-synchronization stability analytically and validates its framework experimentally.Source: Introduction · p. 2

Formulation and Dynamical Equations

• The network formulation allows different node and connection types, assumes undirected couplings, and uses symmetry permutations to identify invariant synchronous motions.Two nodes are synchronized when their states agree for all time; such synchronous motions define invariant manifolds.Source: Formulation and Dynamical Equations · pp. 2–3

Symmetries of Multilayer Networks

• Symmetry permutations are block diagonal across layers, and interlayer coupling determines which layer permutations can be paired to preserve the whole network.The compatibility conditions restrict the possible pairings to subgroups of each layer’s symmetry group.Source: Symmetries of Multilayer Networks · p. 5
• Independent-layer symmetries leave interlayer coupling unchanged, whereas dependent-layer symmetries involve simultaneous node swaps across layers.Breaking an independent-layer symmetry does not change possible clusters in other layers unless symmetries in other equivalence classes also break.Source: Symmetries of Multilayer Networks · pp. 6–7
• Several real multilayer networks possess very large symmetry groups, based on structural analysis that treats nodes as homogeneous within each layer and distinct across layers.The analysis uses network structure rather than individual node attributes.Source: Symmetries of Multilayer Networks · p. 7, Table 1 · p. 7
• Generated multiplex networks with scale-free layers typically lack symmetries, except when the layers’ power-law degree exponents are low; some analyzed real networks have more symmetries.This observation motivated an algorithm for generating multilayer networks with prescribed symmetry counts.Source: Symmetries of Multilayer Networks · p. 7

Stability Analysis

• The stability analysis constructs intralayer and interlayer quotient matrices for orbital clusters, then linearizes the dynamics about the quotient-network evolution.This part introduces the stability procedure but supplies no stability results.Source: Stability Analysis · p. 8
• The IRR transformation separates synchronization-manifold motion from transverse blocks; ILS breakings involve one layer, whereas DLS breakings involve different layers and depend on both layers’ dynamics.Transverse blocks correspond to an individual cluster or a subset of intertwined clusters, reducing stability analysis to smaller independent problems.Source: Stability Analysis · pp. 9–10
• DLS perturbations are harder to stabilize than ILS perturbations: the analytic model’s best stability occurs at δc = 0, and increasing layer differences can spread eigenvalues.The eigenvalue argument supports a general conjecture that DLS clusters are more difficult to stabilize when systems in different layers differ; the paper motivates this using Gershgorin discs.Source: Stability Analysis · pp. 11–12

An experimental testbed for cluster synchronization

• The experimental testbed combines one jumper resonator, two FitzHugh–Nagumo oscillators, and four Colpitts oscillators, with inductive and resistive couplings between layers.The Colpitts layer has inductive intralayer coupling; coupling between Colpitts and FHN oscillators is resistive, while FHN-to-jumper coupling is inductive.Source: An experimental testbed for cluster… · pp. 12, 14

Stability Analysis

• For discrete-time maps, synchronization is simultaneously lost in both layers for δc ⪆0.35, with ρ+ exceeding 1; in another DLS case, within-layer synchronizations are simultaneously lost for δc ⪆0.4.Source: Stability Analysis · p. 12, Figure 2 · p. 13

An experimental testbed for cluster synchronization

• Three clustered patterns are possible: β and γ fully synchronized; β fully synchronized while γ has either pairing [1,3]/[2,4] or [1,4]/[2,3]; or β is unsynchronized while γ has clusters [1,2]/[3,4].Pattern 1a is stable when both transverse MLE curves are negative. Figure 5 summarizes the stability analyses of all three patterns.Source: An experimental testbed for cluster… · p. 15
• As kC decreases, the system progresses from sole stable pattern 1a through patterns 2-(π/2) and 2-(π), to pattern 1b or its slight modification.Figure 6 identifies the clustered behaviors and their parameter regions.Source: Figure 6 · p. 17, An experimental testbed for cluster… · pp. 17–18
• At kC = 0.2, a supercritical pitchfork creates pattern 2; at kC = 0.14, a saddle-node creates an anti-phase pattern 2 solution that remains stable until kC = −0.28.The anti-phase state has a phase separation of π between the FHN oscillators and a slightly higher oscillation frequency.Source: An experimental testbed for cluster… · p. 17
• At kC = 0.07, pattern 2-(π/2) loses stability to an unsynchronized solution, which disappears at kC = 0; for kC ∈ [−0.27, 0], two solutions coexist.The coexisting states are pattern 2-(π) and pattern 1b or a slightly nonsynchronized modification of 1b, depending on whether the red curve is negative or positive.Source: An experimental testbed for cluster… · p. 18
• The modified pattern 1b persists because antiphase Colpitts pairs give the FHN a very small net signal, allowing the Colpitts oscillators to nearly synchronize despite slightly different inputs.The FHN oscillators remain locked in this modified state, while the Colpitts oscillators are not perfectly synchronized.Source: An experimental testbed for cluster… · p. 18
• At kC = −0.28, pattern 2-(π) undergoes catastrophic symmetry breaking, leaving pattern 1b or its slight modification as the only attractor; experiments qualitatively agree with the numerical results.The slight modification occurs when the red curve is positive.Source: An experimental testbed for cluster… · p. 18

Discussion

• The paper defines and computes multilayer-network symmetries, analyzes cluster synchronization analytically and numerically, and reports many symmetries in real multilayer datasets.Source: Discussion · p. 18
• Independent layer symmetries do not alter possible clusters in other layers, whereas dependent layer symmetries involve interlayer connections and can affect those dynamics.Independent-layer symmetry-breaking bifurcations can break a cluster in one layer without changing possible clusters in another.Source: Discussion · p. 18
• An IRR change of coordinates separates stability into lower-dimensional blocks, with intertwined clusters either all stable or all unstable.Dependent-layer-symmetry blocks differ from single-layer blocks and affect the stability of clusters involving those symmetries.Source: Discussion · p. 18
• Analytical results for a specific network class and simulations with Van der Pol, Lorenz, and Roessler oscillators indicate that dependent-layer-symmetry clusters are harder to synchronize when layer systems differ.Source: Discussion · p. 18
• The seven-node circuit exhibits all predicted cluster states in the expected part of parameter space, broadly matching theoretical predictions.The circuit contains slightly heterogeneous nodes and noise, although the model assumes identical nodes and connections and neglects noise and other nonidealities.Source: Discussion · pp. 18–19
• Stability analysis scales to typically smaller networks than symmetry finding, and the framework assumes identical systems within each layer.The study cannot quantitatively specify when noise or heterogeneity changes predictions, though its circuit experiment demonstrates some tolerance.Source: Discussion · p. 19
• The framework leaves directed intra- and interlayer connections and cluster-synchronization patterns unrelated to symmetries as possible extensions.Source: Discussion · p. 19

Formal Proofs

• The conjugate sets Hα and Hβ are subgroups of their respective layer symmetry groups.The proof verifies identity, inverse existence, and closure for Hα, with the proof for Hβ stated to be identical.Source: Formal Proofs · pp. 19–20
• To compute Hα, check which permutations in Gα satisfy the conjugacy relation for permutations in Gβ; Hβ can be computed similarly.In the undirected case, finding Hα automatically determines Hβ.Source: Formal Proofs · p. 20
• For undirected networks, either one of the two stated conjugacy relations suffices to establish that Hα and Hβ are subgroups.Symmetry of the overall coupling matrix relates the off-diagonal coupling blocks as Aβα = (Aαβ)^T.Source: Formal Proofs · p. 20
• The kernel Kα₁ is a normal subgroup, and the subgroups Kαᵢ have equal numbers of elements.The proof establishes normality by conjugation and relates the subgroup sizes through the uniqueness of group products.Source: Formal Proofs · p. 21

Properties of the Equivalence Classes and structure of the matrix T

• The full-system T matrix has block-diagonal form and can be assembled from the independently computed T matrices of each layer.Source: Properties of the Equivalence Classes… · p. 21
• In the projection-operator sums, repeated appearances of each layer’s group elements are canceled by the full group order in the denominator.The full group order is K nα nβ, leaving the correct dimension divisor for each layer’s sum.Source: Properties of the Equivalence Classes… · p. 21

Computing the symmetry group of a simple multilayer network

• The example computes Hα by checking which α-layer permutations satisfy the conjugacy relation with β-layer permutations.Source: Computing the symmetry group of a… · p. 22
• The example’s independent-layer symmetry permits α-layer clusters [1] and without changing possible β-layer clusters.The α-layer bifurcations associated with these symmetries do not cause bifurcations in possible β dynamics or synchronous clusters; stability is a separate issue.Source: Computing the symmetry group of a… · p. 22
• In this network, β-layer nodes 1 and 3 depend respectively on the sums of α-layer nodes 1 and 2, and nodes 3 and 4.Thus, when β nodes 1 and 3 are not synchronized, they are separate trivial clusters, which does not prevent either of the stated α-layer cluster patterns.Source: Computing the symmetry group of a… · p. 23

• The symmetry problems for multiplex and multidimensional networks are closely related, so the section develops the multiplex computation through that mapping.Source: Symmetries of multiplex and… · p. 23
• A multiplex network can be rewritten as a multidimensional network with one layer and separate adjacency matrices for different interaction types.This equivalence provides the basis for computing multiplex symmetries using the multidimensional-network formulation.Source: Symmetries of multiplex and… · p. 23
• The multidimensional network's symmetry group is the largest common subgroup of the symmetry groups for its interaction types.A permutation in each group commutes with the corresponding interaction-type adjacency matrix.Source: Symmetries of multiplex and… · pp. 23–24
• Computationally, GAP or SAGE can find the symmetry group after the network is remapped as a labeled graph.The labeled adjacency representation distinguishes absent edges, single interactions, and multiple interactions by their types and weights.Source: Symmetries of multiplex and… · p. 24

• The three-layer extension applies the symmetry procedure beyond the earlier cases with at most two layers and one interlayer coupling.Source: A network with three layers · p. 24
• For directed three-layer graphs, six interlayer conjugacy relationships must hold; for undirected graphs, half are redundant.The relationships match each layer permutation across the two other layers' coupling matrices.Source: A network with three layers · p. 24
• Compatible permutations define subgroups within each layer, with equivalence requiring the same outcomes across the defining conjugacy relationships.The construction forms disjoint equivalence sets and uses them to assemble the final network symmetry group.Source: A network with three layers · pp. 24–25
• The three-layer construction identifies independent layer symmetries within the layers' equivalence-class subgroups.Source: A network with three layers · p. 25

• The extension is formulated for a network with two interlayer coupling types and their corresponding adjacency matrices.Source: Two interlayer coupling types · p. 25
• With two coupling types between two layers, each layer permutation must satisfy conjugacy relationships for both types.For directed graphs this gives four relationships, two of which are redundant for undirected graphs.Source: Two interlayer coupling types · p. 25
• The layer subgroups and equivalence classes are formed using the multiple conjugacy constraints, then used to construct the network symmetry group.Equivalence requires the same left-hand-side outcomes across all defining relationships.Source: Two interlayer coupling types · p. 25

• The construction covers multilayer networks with any number of layers and interlayer coupling types.Source: Any number of layers and inter-layer… · p. 25
• Each interlayer adjacency matrix contributes a conjugacy relationship, so the number of relationships grows with the number of coupling types.Source: Any number of layers and inter-layer… · p. 25
• Computational effort grows as O(A), where A is the number of interlayer adjacency matrices.Source: Any number of layers and inter-layer… · p. 25

Computation of the symmetry group

• A symmetry may exchange only interchangeable edge types and nodes of the same type.The authors use multipartite labeled-graph software to handle both constraints.Source: Computation of the symmetry group · p. 26
• The labeled supra-adjacency matrix records multilayer connections, with zero entries for unconnected node pairs and tuples for connected pairs.Source: Computation of the symmetry group · p. 26
• A predefined partition groups renumbered nodes by layer, limiting candidate permutations to nodes within the same subset.Source: Computation of the symmetry group · p. 26
• The symmetry group is found by solving ΠA_lab = A_labΠ subject to the predefined partition of nodes allowed to swap.Source: Computation of the symmetry group · p. 26

Mathematical model of the experimental system

• Kirchhoff’s laws provide the differential equations for the circuit, including the jumper and Colpitts oscillator equations.Source: Mathematical model of the experimental… · p. 27
• The experiment varies Colpitts intralayer coupling kC by changing inductor separation while holding the other stated couplings constant.C1 and C3 are immobilized while C2 and C4 are moved; fixed resistors and plastic clips maintain the other couplings.Source: Mathematical model of the experimental… · p. 28
• The noise-free model omits circuit noise and stray resistance, inductance, and capacitance, while component mismatch makes uncoupled oscillators slightly heterogeneous.The uncoupled nodes have slightly different frequencies and amplitudes.Source: Mathematical model of the experimental… · p. 28
• The model assumes identical oscillators within each layer before expressing the circuit in the general form (1).The equations distinguish intra-layer from inter-layer coupling.Source: Mathematical model of the experimental… · p. 28
• Inductive coupling through the jumper induces direct diffusive coupling between FHN-layer currents because it acts on the derivative of current.The coupling is represented by a dashed line in the FHN layer in Figure 4.Source: Mathematical model of the experimental… · p. 29