Opiniondynamics on
social networks
Giacomo Albi
The opiniondynamicsEmergence ofconsensus
Symmetricinteractions
Non-symmetricinteractions
Opinion dynamics on social networks
Giacomo Albiwww.giacomoalbi.com
Department of Computer Science,University of Verona, Italy
ECMI Modelling courseVerona, October, 2017
October 2017Giacomo Albi (University of Verona) Opinion dynamics on social networks October, 2017 1 / 20
Opiniondynamics on
social networks
Giacomo Albi
The opiniondynamicsEmergence ofconsensus
Symmetricinteractions
Non-symmetricinteractions
Outline
Problem posingWe want to model, simulate and control an interacting systemdescribing the evolution of individuals’ opinions over a social network
The objects of study areIndividual’s opinion, which evolves according to the exchange ofinformations with other individuals.The interaction network, the social system ruling the interactionsamong individuals (Facebook, Twitter, personal network,. . . )Influence by external factors: how advertisement, politicalpolicies affect the global opinion?
Giacomo Albi (University of Verona) Opinion dynamics on social networks October, 2017 2 / 20
Opiniondynamics on
social networks
Giacomo Albi
The opiniondynamicsEmergence ofconsensus
Symmetricinteractions
Non-symmetricinteractions
Modelling opinion dynamics
We propose the following model for opinion formation
ddt wi =
N∑j=1
Pij(t)(wj − wi), i = 1, . . . , N
where,wi ∈ [−1, +1] represents the opinion of the i-agent, where −1 and+1 represent two opposite opinions.Pij ≥ 0 is the communication function, quantifying the influencebetween agent i and j.
The underling process of such model represents a generic way to describealignment, where opinion of agent i aligns toward the opinion of agent j,
wi → wj
Giacomo Albi (University of Verona) Opinion dynamics on social networks October, 2017 3 / 20
Opiniondynamics on
social networks
Giacomo Albi
The opiniondynamicsEmergence ofconsensus
Symmetricinteractions
Non-symmetricinteractions
The communication function
In a general setting we assume the communication function Pij(t) to be anonlinear model of the agents’ positions, and the associated graph, G,
Pij(t) = Pij(w(t),G(t)), w(t) = {wk(t)}k
Note that Pij(t) can be seen as weights that naturally induce a directgraph structure on the set of agents.Hence we can define the graph induced by Pij , for any ε ≥ 0 andt ≥ 0 the graph Pε(t) as
Pε = {(i, j) ∈ {1, . . . , N}2|Pij(t) > ε}.
Thus P0(t) is the the set of edges (i, j) for which the communicationchannel from i to j is active at time t.
Giacomo Albi (University of Verona) Opinion dynamics on social networks October, 2017 4 / 20
Opiniondynamics on
social networks
Giacomo Albi
The opiniondynamicsEmergence ofconsensus
Symmetricinteractions
Non-symmetricinteractions
The communication function
In what follows, we will assume that interactions Pij(t) = Pij(w(t),G(t)) areruled by two main mechanisms
Network based interactions, where Pij is the adjacency matrix of anassociated graph, G, describing the set of direct connections among agents,
Pij(t) = Pij(G(t))
Metric based interactions, where the communication among agents Pij is afunction of the relative distance, dij = |wi − wj |.
Pij(|wi − wj |)
Giacomo Albi (University of Verona) Opinion dynamics on social networks October, 2017 5 / 20
Opiniondynamics on
social networks
Giacomo Albi
The opiniondynamicsEmergence ofconsensus
Symmetricinteractions
Non-symmetricinteractions
Network based interactions
We considered N = 30 agents with different adjacency matrix Pij
Giacomo Albi (University of Verona) Opinion dynamics on social networks October, 2017 6 / 20
Opiniondynamics on
social networks
Giacomo Albi
The opiniondynamicsEmergence ofconsensus
Symmetricinteractions
Non-symmetricinteractions
Metric based interactions
As an example we can consider Bounded Confidence type of interactions,
Pij =1Nχ(|wi − wj | ≤ C),
where agents interact only within a confidence level C .
Left: C = 0.25, Center: C = 0.45 Right: C = 0.65
Giacomo Albi (University of Verona) Opinion dynamics on social networks October, 2017 7 / 20
Opiniondynamics on
social networks
Giacomo Albi
The opiniondynamicsEmergence ofconsensus
Symmetricinteractions
Non-symmetricinteractions
Emergence of consensus
Definition: Global consensus
Let w : R+ → [−1, +1]N be a solution of the opinion model with initialdatum w(0) = w0. We say that w(·) converges to consensus if there existsw∞ ∈ [+1,−1] such that, for every i = 1, . . . , N it holds
limt→+∞
|wi(t)− w∞| = 0,
where | · | is the Euclidean norm. The value w∞ is called the consensusstate.
wi(t)→ w∞.
We will see that the w∞ is an emergent property of the global dynamics ofthe system.
Giacomo Albi (University of Verona) Opinion dynamics on social networks October, 2017 8 / 20
Opiniondynamics on
social networks
Giacomo Albi
The opiniondynamicsEmergence ofconsensus
Symmetricinteractions
Non-symmetricinteractions
Emergence of consensus
The average opinionIf the communication function is symmetric, i.e. Pij = Pji , theaverage opinion is an invariant of the system.
Prove that
ddt w̄(t) = 0, with w̄(t) = 1
N
N∑i=1
wi(t)
As a consequence, if the solution of the opinion model converges toconsensus, we have
w∞ = 1N
N∑i=1
w0i .
Giacomo Albi (University of Verona) Opinion dynamics on social networks October, 2017 9 / 20
Opiniondynamics on
social networks
Giacomo Albi
The opiniondynamicsEmergence ofconsensus
Symmetricinteractions
Non-symmetricinteractions
The Laplacian formulation
We consider the following reformulation of the opinion model
ddt
wi =N∑
j=1
Pij(t)wj − wi
N∑j=1
Pij(t) = − (LP(t)w)i , i = 1, . . . ,N
where LP(t) is defined as the Laplacian matrix associated to the communicationfunction P(t) as follows
LP =
∑
j 6=1 P1j −P12 −P13 . . . −P1N−P21
∑j 6=2 P2j −P23 . . . −P2N
−P31 −P32∑
j 6=3 P3j . . . −P3N...
......
. . ....
−PN1 −PN2 −PN3 . . .∑
j 6=N PNj
For symmetric interactions, Pij = Pji , the Laplacian matrix is symmetric.The laplacian formulation is connected to the heat equation
∂tu = ∂xxu
in particular if we consider a finite difference discretization of the lastequation.
Giacomo Albi (University of Verona) Opinion dynamics on social networks October, 2017 10 / 20
Opiniondynamics on
social networks
Giacomo Albi
The opiniondynamicsEmergence ofconsensus
Symmetricinteractions
Non-symmetricinteractions
The Laplacian matrix of a graph
The Laplacian matrix L associated to the (simple) graph G is defined as follows
L = D −A
where:A is the adjacency matrix of G,and D is the (diagonal) degree matrix of G, defined as Dii =
∑j 6=iAij , for
i = 1, . . . ,N .
L =
∗ ∗ ∗ ∗ ∗∗ ∗ ∗ ∗ ∗∗ ∗ ∗ ∗ ∗∗ ∗ ∗ ∗ ∗∗ ∗ ∗ ∗ ∗
Write the Laplacian matrix associated to the left graph.
Giacomo Albi (University of Verona) Opinion dynamics on social networks October, 2017 11 / 20
Opiniondynamics on
social networks
Giacomo Albi
The opiniondynamicsEmergence ofconsensus
Symmetricinteractions
Non-symmetricinteractions
The Laplacian matrix of an undirect graph
Symmetric interactionsFor undirect graphs, G, the Laplacian matrix is a symmetric semi-positve definitematrix.
The Laplacian matrix L has eigenvalues
0 = λ1 ≤ λ2 ≤ . . . ≤ λN
Indeed, every row sum and column sum of L is zero,
L1 = 0,
1 = (1, 1, . . . , 1)T , 0 = (0, 0, . . . , 0)T
thus we have that L is singular with eigenvalue 0.The number of connected components in the graph G is the geometricalmultiplicity of the 0 eigenvalue.The second smallest eigenvalue of L (could be zero), λ2(L) is the Fiedlernumber of G and it is associated to the algebraic connectivity.
Giacomo Albi (University of Verona) Opinion dynamics on social networks October, 2017 12 / 20
Opiniondynamics on
social networks
Giacomo Albi
The opiniondynamicsEmergence ofconsensus
Symmetricinteractions
Non-symmetricinteractions
On the algebraic connectivity
Algebraic connectivityThe algebraic connectivity, or Fiedler eigenvalue, of a graph G is thesecond-smallest eigenvalue of the Laplacian matrix of G.
This eigenvalue is greater than 0 if and only if G is a connected graph.Indeed the number of times 0 appears as an eigenvalue in the Laplacian isthe number of connected components of G.The magnitude of this value reflects how well connected the overall graph is.
Write the Laplacian matrix of both graphs and compute its eigenvalues.
Giacomo Albi (University of Verona) Opinion dynamics on social networks October, 2017 13 / 20
Opiniondynamics on
social networks
Giacomo Albi
The opiniondynamicsEmergence ofconsensus
Symmetricinteractions
Non-symmetricinteractions
Unconditional consensus under symmetry
Let us introduce the following new variable
vi(t) := wi(t)− w̄
where w̄ is the average of the system which is conserved over time thanks to symmetry.
We can write,
ddt
vi =N∑
j=1
Pij(vj − vi), vi(0) = w0i − w̄.
Thus we can compute the following chain of equalities
12
ddt
∑i
|vi(t)|2 =12
∑i
2〈ddt
vi , vi〉
=∑
i
∑j
Pij〈vj − vi , vi〉
=12
∑i
∑j
Pij〈vj − vi , vi〉 −12
∑i
∑j
Pij〈vj − vi , vj〉
= −12
∑i
∑j
Pij |vj − vi |2
Giacomo Albi (University of Verona) Opinion dynamics on social networks October, 2017 14 / 20
Opiniondynamics on
social networks
Giacomo Albi
The opiniondynamicsEmergence ofconsensus
Symmetricinteractions
Non-symmetricinteractions
Unconditional consensus under symmetry
Courant-Fischer-Weyl min-max principle
We can characterize the Fiedler number of the Laplacian matrix, λ2(LP) in termsof the vector v ∈ RN orthogonal to the first eigenvector 1 = (1, 1, . . . , 1)T , asfollows
λ2(LP) = min1T v=0
〈LPv, v〉〈v, v〉
≤(1/2)
∑i
∑j Pij |vi − vj |2∑
i |vi |2
Thanks to this result we can infer from the previous equalities the followingestimate
12
ddt
∑i
|vi(t)|2 = −12
∑i
∑j
Pij |vj − vi |2 ≤ −λ2(LP)∑
i
|vi(t)|2.
Thus we have an estimate for the following quantity
1N
∑i
|vi(t)|2 =1N
N∑j=1
|w(t)− w̄(0)|2 =: Vw(t).
Giacomo Albi (University of Verona) Opinion dynamics on social networks October, 2017 15 / 20
Opiniondynamics on
social networks
Giacomo Albi
The opiniondynamicsEmergence ofconsensus
Symmetricinteractions
Non-symmetricinteractions
Unconditional consensus under symmetry
In terms of Vw(t) we haveddt
Vw(t) ≤ −2λ2(LP)Vw(t)
then the following theorem holds
Theorem (S. Motsch- E. Tadmor ’14)
Given the opinion model
ddt
w = −LP(t)w, w(0) = w0
with symmetric communication P. Then the following concentration estimate holds
Vw(t) ≤ exp
(−2
∫ t
0
λ2(LP(s)) ds
)Vw(0).
In particular, if the interactions remain “strong enough” so that
limt→∞
∫ t
0
λ2(LP(s)) ds = +∞,
then there is convergence toward consensus, i.e. w(t)→ w∞ = w̄(0).Giacomo Albi (University of Verona) Opinion dynamics on social networks October, 2017 16 / 20
Opiniondynamics on
social networks
Giacomo Albi
The opiniondynamicsEmergence ofconsensus
Symmetricinteractions
Non-symmetricinteractions
Connectivity
A direct path in G from Vi to Vj is a sequence of edges
Γij =(
eii1 , ei1i2 , . . . , eik j)∈ G.
A graph G is connected when there is a path between every pair of vertices.Conversely a graph G is said to be disconnected if there exist two nodes inG such that no path in G has those nodes as endpoints.
Giacomo Albi (University of Verona) Opinion dynamics on social networks October, 2017 17 / 20
Opiniondynamics on
social networks
Giacomo Albi
The opiniondynamicsEmergence ofconsensus
Symmetricinteractions
Non-symmetricinteractions
Strong connectivity
The graph G is said to be strongly connected if for any pair of vertexes Vi ,Vjthere exists a direct path from Vi to Vj
Giacomo Albi (University of Verona) Opinion dynamics on social networks October, 2017 18 / 20
Opiniondynamics on
social networks
Giacomo Albi
The opiniondynamicsEmergence ofconsensus
Symmetricinteractions
Non-symmetricinteractions
Unconditional consensus under strong connectivity
Let Pε(t) be the graph induced by the communication functionPij(t), we recall that
Pε = {(i, j) ∈ {1, . . . , N}2|Pij(t) > ε}.
Theorem (Haskovec, ’15)
Given the opinion model
ddt
w = −LP(t)w, w(0) = w0
with associated graph Pε(t). If there exists an ε > 0 and a strongly connectedgraph G on the set of agents on which the system spends an infinite amount oftime, i.e.
T = {t ≥ 0|Pε(t) ≡ G}, `1(T ) = +∞.Then there is convergence toward consensus, i.e. w(t)→ w∞.
where `1 is the Lebesgue measure.
Giacomo Albi (University of Verona) Opinion dynamics on social networks October, 2017 19 / 20
Opiniondynamics on
social networks
Giacomo Albi
The opiniondynamicsEmergence ofconsensus
Symmetricinteractions
Non-symmetricinteractions
Bibliography
Bongini, M. (2016), Sparse Optimal Control of Multiagent Systems,PhD thesis, TUM.Watts, D. J., Dodds, P. S. (2007). Influentials, networks, and publicopinion formation. Journal of consumer research, 34(4), 441-458.Hegselmann, R., Krause, U. (2002). Opinion dynamics and boundedconfidence models, analysis, and simulation. Journal of artificialsocieties and social simulation, 5(3).Cucker, F., Smale, S. (2007). Emergent behavior in flocks. IEEETransactions on automatic control, 52(5), 852-862.Motsch, S., Tadmor, E. (2014). Heterophilious dynamics enhancesconsensus. SIAM review, 56(4), 577-621.Moreau, L. (2005). Stability of multiagent systems withtime-dependent communication links. IEEE Trans. Automat. Control, 50(2):169-182.Haskovec, J. (2015). A note on the consensus finding problem incommunication networks with switching topologies. Appl. Anal.,94(5):991-998.
Giacomo Albi (University of Verona) Opinion dynamics on social networks October, 2017 20 / 20