Source-linked AI summary
Structural balance and opinion separation in trust-mistrust social networks
Weiguo Xia, Ming Cao, Karl Henrik Johansson
TL;DR
The paper asks how structural balance governs opinion separation beyond fixed, strongly connected networks. It analyzes discrete- and continuous-time DeGroot-type dynamics under weaker fixed connectivity and time-varying joint connectivity, finding neutralization for influential unbalanced subnetworks, polarization for balanced ones, and intermediate opinions elsewhere.
Problem
Existing results connected structural balance with polarization mainly for fixed, strongly connected networks, leaving opinion evolution under weaker or time-varying connectivity open.
Method
The paper studies discrete-time and continuous-time DeGroot-type models on fixed networks with influential strongly connected subnetworks and on time-varying networks using joint graphs.
Results
Unbalanced influential subnetworks neutralize all opinions, while balanced subnetworks polarize internally and drive other agents’ opinions between the polarized values; analogous criteria hold for joint graphs.
Takeaways & Limitations
DeGroot-type dynamics can produce opinion clustering under weaker connectivity, in addition to consensus and polarization.
Abstract
from arXiv · showhide
Structural balance theory has been developed in sociology and psychology to explain how interacting agents, e.g., countries, political parties, opinionated individuals, with mixed trust and mistrust relationships evolve into polarized camps. Recent results have shown that structural balance is necessary for polarization in networks with fixed, strongly connected neighbor relationships when the opinion dynamics are described by DeGroot-type averaging rules. We develop this line of research in this paper in two steps. First, we consider fixed, not necessarily strongly connected, neighbor relationships. It is shown that if the network includes a strongly connected subnetwork containing mistrust, which influences the rest of the network, then no opinion clustering is possible when that subnetwork is not structurally balanced; all the opinions become neutralized in the end. In contrast, it is shown that when that subnetwork is indeed structurally balanced, the agents of the subnetwork evolve into two polarized camps and the opinions of all other agents in the network spread between these two polarized opinions. Second, we consider time-varying neighbor relationships. We show that the opinion separation criteria carry over if the conditions for fixed graphs are extended to joint graphs. The results are developed for both discrete-time and continuous-time models.
I. INTRODUCTION
The paper extends research on DeGroot-type opinion dynamics by examining how structural balance relates to opinion separation under weaker connectivity and time-varying network topologies.
- Structural balance specifies trust–mistrust arrangements associated with stable polarized opinions but does not specify how opinions update.DeGroot models address this gap by repeatedly averaging neighbors’ opinions.
- Prior results show that strongly connected, structurally balanced networks can produce opposite opinions across two camps, whereas structurally unbalanced networks converge all opinions to zero.
- The paper studies fixed networks containing directed spanning trees and dynamically changing networks with joint connectivity, using both discrete-time and continuous-time models.
- For fixed networks, an unbalanced strongly connected influential subgraph neutralizes all opinions, while a balanced one polarizes internally and bounds other agents between the polarized values.
- For time-varying networks, analogous opinion-separation conclusions hold when the fixed-graph conditions are applied to joint graphs.
- The results identify opinion clustering under weaker connectivity as an additional behavior beyond polarization and consensus.
II. MOTIVATING EXAMPLE
The motivating examples show that trust–mistrust DeGroot dynamics can exhibit clustering and switching-related behavior beyond the polarization expected in strongly connected balanced networks.
- Each agent updates its scalar opinion by a weighted average of its neighbors’ opinions and its own opinion, with signed edge weights representing trust or mistrust.A negative edge contributes a negative neighbor term to the update.
- Fig. 2 compares state evolution for fixed topology G1, fixed topology G2, and switching between G1–G3 or G1–G4.Initial opinions lie in [−1, 1].
- G1 is structurally balanced and strongly connected, so its agents form two camps with opposite agreed values.
- Although G1 and G4 are each structurally balanced, their different bipartitions make agreement under switching unclear.
- G2 has a directed spanning tree without strong connectivity, producing opinion clustering rather than polarization or consensus.The paper develops theoretical explanations for this behavior in Theorems 2 and 5.
III. PROBLEM FORMULATION
The paper formulates signed DeGroot dynamics on directed networks, defines structural balance and polarization, and studies their relationship under fixed spanning-tree and time-varying connectivity conditions.
- Each agent has a scalar opinion, and a directed signed edge records which neighbor influences it and whether that relationship is trusting or mistrusting.
- Discrete-time model: The discrete-time model is written as x(t + 1) = P(t)x(t), where P(t) has positive diagonal entries.
- Continuous-time model: The continuous-time model uses signed interaction weights and a signed Laplacian matrix to represent opinion evolution.
- Both models make trusted opinions pull agents closer and mistrusted opinions push them apart, linking edge-sign distributions to opinion evolution.
- A directed signed graph is structurally balanced when vertices can be partitioned so within-group edges are positive and between-group edges are negative.All-positive graphs remain structurally balanced when one group is empty.
- Polarization requires nonzero equal limiting absolute opinion magnitudes and opposite limiting values for at least some agents.
- The central problem is whether structural balance predicts opinion separation when fixed graphs only contain a spanning tree or when connectivity is time varying.
IV. DISCRETE-TIME MODEL
The discrete-time analysis enlarges signed graphs into nonnegative consensus graphs, converting structural balance into connectivity properties that characterize opinion polarization or neutralization.
- Consensus reformulation: The signed update system is transformed into a classical nonnegative consensus system, allowing graphical conditions on the signed graph to be analyzed through the enlarged graph.The associated graphs of the signed and transformed systems are isomorphic, and the transformed matrix is stochastic.
- Graph transformation: The enlarged graph duplicates each agent into positive and negative copies and replaces signed edges with positive directed edges.Positive original edges preserve copy signs, while negative edges cross between positive and negative copies.
- Graph transformation: Structural balance holds exactly when the enlarged graph is disconnected into two strongly connected components.The two components correspond to the two camps induced by a structural-balance bipartition.
- Graph transformation: Structural unbalance holds exactly when the enlarged graph is strongly connected.A negative cycle combined with strong connectivity yields paths between positive and negative copies.
A. G(P(t)) is fixed
For fixed networks containing a directed spanning tree, the structural balance of the strongly connected root subgraph determines whether opinions polarize or converge to zero, while downstream agents remain bounded between polarized values.
- Root subgraph cases: If the strongly connected root subgraph is structurally unbalanced, every agent’s opinion converges to zero for every initial value.The transformed consensus system has a single limiting value, which must be zero because it contains the positive and negative copies of each state.
- Root subgraph cases: If the root subgraph is structurally balanced with at least one negative edge, its agents polarize to equal-magnitude opposite values.The two corresponding components of the transformed graph independently reach consensus at opposite values.
- Root subgraph cases: For a structurally balanced root subgraph, every downstream agent converges to a value in the interval [−|C|, |C|], where |C| is the root subgraph’s polarized magnitude.The bound follows from the nonnegative consensus representation and the two root-component limits.
- Root subgraph cases: When the entire fixed graph is structurally balanced with at least one negative edge, the full system polarizes.The theorem distinguishes this full-network conclusion from the weaker boundedness result for downstream agents.
- Spectral analysis: The analysis gives a complete characterization of the fixed-network final state by examining the eigenvalues and eigenvectors of the transformed system matrix.The transformed matrix has two unit eigenvalues in the balanced case, with all other eigenvalues strictly inside the unit circle.
B. G(P(t)) is time-varying
For time-varying discrete-time networks, joint-graph connectivity and structural balance determine whether opinions polarize or become neutralized. With spanning-tree connectivity, polarized root agents bound the remaining agents’ states between the polarized values.
- Joint strong connectivity: A common bipartition across strongly connected joint graphs makes the system polarize into two opposite values.The transformed joint graphs have two strongly connected components whose states converge to opposite values, except for a zero-measure set of initial conditions.
- Joint strong connectivity: Strongly connected joint graphs without a common structural-balance bipartition make x(t) converge asymptotically to zero.The proof constructs a strongly connected signed union graph containing a negative cycle and applies consensus results to an auxiliary system.
- Structural unbalance: If structural unbalance occurs frequently enough in strongly connected intervals, the agents’ opinions become neutralized at zero instead of polarizing.The stated condition is that every interval contains some time at which the graph is strongly connected and structurally unbalanced.
- Spanning-tree connectivity: When joint graphs contain only a spanning tree, structurally balanced root agents polarize while all other agents’ states eventually lie between the polarized values.The root set must share a bipartition across the graphs and remain the root set of each interval’s union graph.
- Spanning-tree connectivity: In time-varying spanning-tree networks, non-root states may fail to converge even though they eventually remain between the polarized values.This differs from the fixed-topology case, where the other agents’ states converge and lie between the polarized values.
V. CONTINUOUS-TIME MODEL
The continuous-time analysis decomposes signed interactions into positive and negative parts and studies an associated transformed system to characterize the dynamics.
- Model transformation: The continuous-time model defines nonnegative matrices A+(t) and A−(t) from the signed adjacency matrix A(t).The signed dynamics are rewritten in terms of a transformed state y(t).
- Scope: The continuous-time results are developed for both fixed and dynamically changing interaction topologies.The paper separately analyzes fixed signed Laplacians and piecewise-constant topology or weight changes.
- Model transformation: The transformed dynamics are governed by a Laplacian matrix with nonpositive off-diagonal elements.Studying system (17) reveals the dynamical behavior of the original continuous-time system.
A. G(L(t)) is fixed
For fixed continuous-time topologies containing a spanning tree, the structurally balanced or unbalanced nature of the irreducible root subgraph determines polarization versus neutralization.
- Fixed topology: A spanning-tree signed Laplacian can be permuted into blocks with an irreducible root block L11 and downstream blocks.The block structure isolates the strongly connected root subgraph from the remaining agents.
- Unbalanced root subgraph: If the root subgraph G(L11) is structurally unbalanced, every initial state converges to zero.This conclusion applies to the continuous-time system under the stated spanning-tree structure.
- Balanced root subgraph: If G(L11) is structurally balanced and contains a negative edge, its agents polarize while other agents converge between the polarized values.If the entire graph is also structurally balanced with a negative edge, the whole system polarizes.
B. G(L(t)) is time-varying
For time-varying continuous-time networks, the paper extends the discrete-time analysis to piecewise-constant signed interactions and interval-based joint-graph conditions.
- Model: The continuous-time topology and edge weights are modeled as piecewise constant, with changes at interaction time instants and dwell times between changes.Nonzero adjacency weights lie between fixed positive bounds γ1 and γ2.
- Analysis: The continuous-time dynamics are rewritten as a transformed y-system and analyzed using ideas from the discrete-time results.The paper invokes an established convergence theorem to obtain the time-varying continuous-time conclusions.
- Balanced joint graphs: A common bipartition of the root vertex set together with spanning-tree connectivity across uniformly bounded intervals yields continuous-time polarization of the root agents.The theorem is stated for an infinite sequence of nonempty, uniformly bounded time intervals.
- Unbalanced joint graphs: If each interval’s union contains a spanning tree but the root set lacks the required bipartition, the continuous-time state converges asymptotically to zero.This is the continuous-time counterpart of the discrete-time structural-unbalance result.
VI. ILLUSTRATIVE EXAMPLES
Simulations compare fixed and switching graph topologies containing spanning trees. Structurally balanced influential subgraphs produce opposite limiting values, while some peripheral agents may fail to converge yet remain between them.
- The simulations use two graphs whose graph topologies contain spanning trees and an initial state x(0) = [0.9, 0.7, −0.9, −1, 0.2, 0.9]T.
- G(P1) is structurally unbalanced, but its subgraph G((P1)11) is structurally balanced.
- Under G(P1), agents 1, 2, and 3 achieve opposite values, while agents 4, 5, and 6 converge between them.
- When graphs switch between P1 at even times and P2 at odd times, agents 4 and 6 do not converge but remain between the opposite values of agents 1, 2, and 3.
VII. CONCLUSION
The conclusion links structural balance to opinion separation in trust–mistrust networks under DeGroot-type dynamics. It also identifies reduced reliance on DeGroot averaging and biased assimilation as future directions.
- The paper studies how structural balance relates to opinion separation in social networks containing trust and mistrust relationships.
- Under conditions related to structural balance, opinions either separate, potentially forming two polarized camps, or become neutralized.
- The authors propose developing opinion-separation models that rely less on DeGroot averaging rules.
- Biased assimilation is identified as a promising direction, with its inherent nonlinearity described as a main challenge.
APPENDIX A
The appendix analyzes powers of a block matrix using eigenvectors and spectral properties. It establishes convergence by isolating a simple eigenvalue at 1 and contracting the remaining components.
- Q1 is stochastic, so 1 is an eigenvalue; the lemma assumes it is simple and all other Q1 eigenvalues have magnitude below 1.
- The proof introduces a nonnegative left eigenvector ξ satisfying ξT Q1 = ξT and ξT 1 = 1.
- Two independent left eigenvectors and two independent right eigenvectors of Q corresponding to 1 are constructed from ξ and vectors involving η1 and η2.
- I − Q33 is invertible because ρ(Q33) < 1, enabling the relations that define η1 and η2.
- The eigenvector normalization and Jordan canonical form are used to show that Qk converges as k approaches infinity.