Welcome to Scribd. Sign in or start your free trial to enjoy unlimited e-books, audiobooks & documents.Find out more
Download
Standard view
Full view
of .
Look up keyword
Like this
1Activity
0 of .
Results for:
No results containing your search query
P. 1
Social Consensus Through the Influence of Committed Minorities

Social Consensus Through the Influence of Committed Minorities

Ratings: (0)|Views: 65|Likes:
Published by D. Manchester
US Army Research Laboratory funded research from Rensallier on Social Consensus Through the Influence of Committed Minorities shows that once 10 percent or more of a population holds an opinion, eventually the entire population will adopt that opinion much more quickly.
US Army Research Laboratory funded research from Rensallier on Social Consensus Through the Influence of Committed Minorities shows that once 10 percent or more of a population holds an opinion, eventually the entire population will adopt that opinion much more quickly.

More info:

Published by: D. Manchester on Jul 27, 2011
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

07/27/2011

pdf

text

original

 
Social consensus through the influence of committed minorities
J. Xie,
1
S. Sreenivasan
,
1,2
G. Korniss,
2
W. Zhang,
3
C. Lim,
3
and B. K. Szymanski
1
1
Dept. of Computer Science, Rensselaer Polytechnic Institute, 110 8th Street, Troy, NY 1218
Dept. of Physics, Rensselaer Polytechnic Institute, 110 8th Street, Troy, NY 12180 
Dept. of Mathematics, Rensselaer Polytechnic Institute, 110 8th Street, Troy, NY 1218
We show how the prevailing majority opinion in a population can be rapidly reversed by a smallfraction
p
of randomly distributed
committed 
agents who consistently proselytize the opposing opin-ion and are immune to influence. Specifically, we show that when the committed fraction growsbeyond a critical value
p
c
10%, there is a dramatic decrease in the time,
c
, taken for the entirepopulation to adopt the committed opinion. In particular, for complete graphs we show that when
 p < p
c
,
c
exp(
α
(
 p
)
), while for
p > p
c
,
c
ln
. We conclude with simulation results forErd˝os-R´enyi random graphs and scale-free networks which show qualitatively similar behavior.
PACS numbers: 87.23.Ge, 89.75.Fb
I. INTRODUCTION
Human behavior is profoundly affected by the influ-enceability of individuals and the social networks thatlink them together. Well before the proliferation of onlinesocial networking, offline or interpersonal social networkshave been acknowledged as a major factor in determininghow societies move towards consensus in the adoption of ideologies, traditions and attitudes [1,2]. As a result, the dynamics of social influence has been heavily studiedin sociological, physics and computer science literature[37]. In the sociological context, work on
diffusion of innovations
has emphasized how individuals adopt newstates in behavior, opinion or consumption through theinfluence of their neighbors. Commonly used models forthis process include the threshold model [8] and the Bassmodel[9]. A key feature in both these models is that oncean individual adopts the new state, his state remains un-changed at all subsequent times. While appropriate formodeling the diffusion of innovation where investment ina new idea comes at a cost, these models are less suitedto studying the dynamics of competing opinions whereswitching one’s state has little overhead.Here we address the latter case. From among the vastrepertoire of models in statistical physics and mathemati-cal sociology, we focus on one which is a 2-opinion variant[10] of the Naming Game (NG) [1115] and that we re- fer to as the
binary agreement model 
. The evolution of the system in this model takes place through the usualNG dynamics, wherein at each simulation time step, arandomly chosen speaker voices a random opinion fromhis list to a randomly chosen neighbor, designated thelistener. If the listener has the spoken opinion in his list,both speaker and listener retain only that opinion, elsethe listener adds the spoken opinion to his list (see Ta-ble 1). The order of selecting speakers and listeners isknown to influence the dynamics, and we stick to choos-
Corresponding author: sreens@rpi.edu
ing the speaker first, followed by the listener.It serves to point out that an important differencebetween the binary agreement model and the predomi-nantly used opinion dynamics models [4,6,1618] is that an agent is allowed to possess both opinions simultane-ously in the former, and this significantly alters the timerequired to attain consensus starting from uniform initialconditions. Numerical studies in [10] have shown that forthe binary agreement model on a complete graph, start-ing from an initial condition where each agent randomlyadopts one of the two opinions with equal probability,the system reaches consensus in time
c
ln
(in con-trast, for example, to
c
for the voter model). Here
is the number of nodes in the network, and unit timeconsists of 
speaker-listener interactions. The binaryagreement model is well suited to understanding howopinions, perceptions or behaviors of individuals are al-tered through social interactions specifically in situationswhere the cost associated with changing one’s opinion islow [19], or where changes in state are not deliberateor calculated, but unconscious [20]. Furthermore, by itsvery definition, the binary agreement model may be ap-plicable to situations where agents while trying to influ-ence others, simultaneously also have a desire to reachglobal consensus[21].Another merit of the binary agreement dynamics inmodeling social opinion change seems worth mentioning.Two state epidemic-like models of social “contagion” (ex-amples in [22]) suffer from the drawback that the rulesgoverning the conversion of a node from a given stateto the other are not symmetric for the two states. Incontrast, in the binary agreement model, both singularopinion states are treated symmetrically in their suscep-tibility to change.Here we study the evolution of opinions in the binaryagreement model starting from an initial state where allagents adopt a given opinion
B
, except for a finite frac-tion
p
of the total number of agents who are
committed agents
and have state
A
. Committed agents, introducedpreviously in [23], are defined as nodes that can influ-ence other nodes to alter their state through the usualprescribed rules, but which themselves are immune to
 
2influence. In the presence of committed agents adoptingstate
A
, the only absorbing fixed point of the system isthe consensus state where all influenceable nodes adoptopinion
A
- the opinion of the committed agents. Thequestion that we specifically ask is: how does the con-sensus time vary with the size of the committed fraction?More generally, our work addresses the conditions underwhich an inflexible set of minority opinion holders canwin over the rest of the population.The effect of having un-influencable agents has beenconsidered to some extent in prior studies. Biswas et al.[24] considered for two state opinion dynamics modelsin one dimension, the case where some individuals are“rigid” in both segments of the population, and studiedthe time evolution of the magnetization and the fractionof domain walls in the system. Mobilia et. al.[25] con-sidered the case of the voter model with some fractionof spins representing “zealots” who never change theirstate, and studied the magnetization distribution of thesystem on the complete graph, and in one and two di-mensions. Our study differs from these not only due tothe particular model of opinion dynamics, but also in itsexplicit consideration of different network topologies andof finite size networks, specifically deriving how consen-sus times scale with network size for the particular caseof the complete graph. Furthermore, the above men-tioned studies do not explicitly consider the initial statethat we care about - one where the entire minority setis un-influencable. A notable exception to the latter isthe study by Galam and Jacobs [26] in which the authorsconsidered the case of “inflexibles” in a two state model of opinion dynamics with opinion updates obeying a major-ity rule. While this study provides several useful insightsand is certainly the seminal quantitive attempt at under-standing the effect of committed minorities, its analysisis restricted to the mean-field case, and has no explicitconsideration of consensus times for finite systems.
II. COMPLETE GRAPHSA. Infinite network size limit
We start along similar lines as[26] by considering thecase where the social network connecting agents is a com-plete graph with the size of the network
→ ∞
. Wedesignate the densities of uncommitted nodes in states
A,B
as
n
A
,n
B
respectively. Consequently, the density of nodes in the mixed state
AB
is
n
AB
= 1
 p
n
A
n
B
,where
p
is the fraction of the total number of nodes, thatare committed. Neglecting correlations between nodes,and fluctuations, one can write the following rate equa-tions for the evolution of densities:
dn
A
dt
=
n
A
n
B
+
n
2
AB
+
n
AB
n
A
+32
 pn
AB
dn
B
dt
=
n
A
n
B
+
n
2
AB
+
n
AB
n
B
 pn
B
(1)
Before interaction After interactionA
A
A A - AA
A
B A - ABA
A
AB A - AB
B
A B - ABB
B
B B - BB
B
AB B - BAB
A
A A - AAB
A
B AB - ABAB
A
AB A - AAB
B
A AB - ABAB
B
B B - BAB
B
AB B - BTABLE I: Shown here are the possible interactions in the bi-nary agreement model. Nodes can possess opinion
A
,
B
or
AB
, and opinion updates occur through repeated selectionof speaker-listener pairs. Shown in the left column are theopinions of the speaker (first)and listener (second) before theinteraction, and the opinion voiced by the speaker during theinteraction is shown above the arrow. The column on rightshows the states of the speaker-listener pair after the interac-tion.
The terms in these equations are obtained by consider-ing all interactions which increase (decrease) the densityof agents in a particular state and computing the prob-ability of that interaction occurring. Table 1 lists allpossible interactions. As an example, the probability of the interaction listed in row eight is equal to the proba-bility that a node in state AB is chosen as speaker and anode in state B is chosen as listener (
n
AB
n
A
) times theprobability that the speaker voices opinion
A
(
12
).The fixed-point and stability analyses (see Appendix)of these
mean-field 
equations show that for any value of 
p
,the consensus state in the committed opinion (
n
A
= 1
 p
,
n
B
= 0) is a stable fixed point of the mean-field dynam-ics. However, below
p
=
p
c
=
52
32
3
 
5 +
24
1
2
32
3
 
5
24
1
2
, two additional fixed points appear:one of these is an unstable fixed point (saddle point),whereas the second is stable and represents an
active
steady state where
n
A
,n
B
and
n
AB
are all non-zero (ex-cept in the trivial case where
p
= 0) . Fig.1(a) shows(asterisks) the steady state density of nodes in state
B
obtained by numerically integrating the mean-field equa-tions at different values of the committed fraction
p
andwith initial condition
n
A
= 0
,n
B
= 1
 p
. As
p
is in-creased, the stable density of 
B
nodes
n
B
abruptly jumpsfrom
0
.
6504 to zero at the critical committed fraction
 p
c
. A similar abrupt jump also occurs for the stable den-sity of 
A
nodes from a value very close to zero below
 p
c
, to a value of 1, indicating consensus in the
A
state(not shown). In the study of phase transitions, an “orderparameter”is a suitable quantity changing (either contin-
 
3uously or discontinuously) from zero to a non zero-valueat the critical point. Following this convention, we use
n
B
- the density of uncommitted nodes in state
B
- asthe order parameter appropriate for our case, character-izing the transition from an active steady state to theabsorbing consensus state.In practice, for a complete graph of any finite size,consensus is always reached. However, we can still probehow the system evolves, conditioned on the system nothaving reached consensus. Fig.1(a) shows the resultsof simulating the binary agreement model on a completegraph for different system sizes (solid lines). For
p < p
c
,in each realization of agreement dynamics, neglecting theinitial transient, the density of nodes in state
B
,
n
B
, fluc-tuates around a non-zero steady state value, until a largefluctuation causes the system to escape from this activesteady state to the consensus state. Fig.1(a) shows thesesteady state values of 
n
B
conditioned on survival, for sev-eral values of 
p
. As expected, agreement of simulationresults with the mean-field curve improves with increas-ing system size, since Eq. (1) represents the true evolu-tion of the system in the asymptotic network size limit.Accordingly the critical value of the committed fractionobtained from the mean-field equations is designated as
 p
c
(
), although, for brevity, we refer to it simply as
p
c
throughout this paper.The existence of the transition as
p
is varied and whenthe initial condition for densities is (
n
A
= 0
,n
B
= 1
 p
)can be further understood by observing the motion of thefixed points in phase space. Fig.1(b) shows how the sta-ble fixed point and the unstable fixed point move in phasespace as
p
is varied from 0 to
p
c
. The active steady statemoves downward and right while the saddle point movesupwards and left. At the critical value
p
c
the two meetand the only remaining stable fixed point is the consen-sus fixed point. A similar observation was made in themodel studied in[26]. The fact that the value of 
n
B
con-verges to
0
.
65 and does not smoothly approach zeroas the stable fixed point and the saddle point approacheach other, explains the origin of the
first-order 
natureof the phase transition. Fig2shows the representativetrajectories obtained by integrating the mean-field equa-tions for the cases where
p
= 0
.
05 (
< p
c
) and
p
= 0
.
1(
> p
c
).
B. Finite network size: Scaling results forconsensus times
Even though consensus is always reached for finite
,limits on computation time prohibit the investigation of the consensus time,
c
, for values of 
p
below or very closeto
p
c
. We therefore adopt a semi-analytical approachpre-scribed in[27] that allows us to estimate the consensustimes for different
for an appreciable range of 
p
includ-ing values below
p
c
. We start with the master equationwhich describes the evolution of the probability that thenetwork of size
has
n
(
m
) uncommited nodes in state
A
(
B
). We denote by
c
, the number of committed nodes,and by
l
(=
n
m
c
), the number of uncommittednodes in state
AB
.
dp
nm
dt
1
=1
2
 p
nm
(2
ln
+32
lc
+ 2
nm
+
l
(
l
1)+ 2
lm
+
mc
) +
p
n
1
,m
3(
l
+ 1)(
n
1 +
c
)2+
p
n
+1
,m
(
n
+ 1)(2
m
+
l
1)2+
p
n
2
,m
(
l
+ 2)(
l
+ 1)2+
p
n,m
1
3(
l
+ 1)(
m
1)2+
p
n,m
+1
(
m
+ 1)(2
n
+ 2
c
+
l
1)2+
p
n,m
2
(
l
+ 2)(
l
+ 1)2
(2)The factor of 1
/N 
in the LHS comes from the fact thata transition between states takes place in an interval of time 1
/N 
. The transition rates in each term are the prod-uct of two densities, which is responsible for the overallfactor of 1
/N 
2
in the RHS. The probabilities are definedover all allowed states of the system ( i.e. 0
n
c
,and 0
m
c
n
for given
n
) and the al-lowed transitions from any point
{
n,m
}
in the interiorof this state-space are
{
n,m
} → {
n,m
±
1
}
,
{
n,m
} →{
n
±
1
,m
}
,
{
n,m
}{
n,m
+ 2
}
,
{
n,m
}{
n
+ 2
,m
}
.We know from the mean field equations that in theasymptotic limit, and below a critical fraction of com-mitted agents, there exists a stable fixed point. For fi-nite stochastic systems, escape from this fixed point isalways possible, and therefore it is termed
metastable
.For a finite system, the probability of having escapedto the metastable fixed point as a function of time is
e
(
t
) = 1
s
(
t
) where
s
(
t
) is the survival probabil-ity. The surviving fraction is constrained to be in theallowed region of 
n,m
quadrant excluding the true fixedpoint
{
c,
0
}
. If the number of committed agents isfar lower than
p
c
we expect that this surviving frac-tion will occupy configurations around the metastablefixed point, and the occupation probabilities
p
n,m
willbe peaked around the metastable fixed point. In sys-tems which exhibit such long lived metastable states inaddition to an absorbing fixed point, applying a quasista-tionary (QS) approximation has been found to be usefulin computing quantities of interest[2729]. This approx- imation assumes that after a short transient, the occu-pation probability, conditioned on survival, of allowedstates excluding the consensus state, is stationary. Fol-lowing this approximation, the distribution of occupationprobabilities conditioned on survival can be written as,˜
 p
nm
=
p
nm
(
t
)
/P 
s
(
t
)[27] and using this form in the mas-

You're Reading a Free Preview

Download
scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->