Opinion Dynamics and Bounded Confidence Models, Analysis and Simulation
Abstract
When does opinion formation within an interacting group lead to consensus, polarization or fragmentation? The article investigates various models for the dynamics of continuous opinions by analytical methods as well as by computer simulations. Section 2 develops within a unified framework the classical model of consensus formation, the variant of this model due to Friedkin and Johnsen, a time-dependent version and a nonlinear version with bounded confidence of the agents. Section 3 presents for all these models major analytical results. Section 4 gives an extensive exploration of the nonlinear model with bounded confidence by a series of computer simulations. An appendix supplies needed mathematical definitions, tools, and theorems.
2 Figures
OPINION DYNAMICS AND BOUNDED CONFIDENCE
MODELS, ANALYSIS, AND SIMULATION
*
Rainer Hegselmann
Department of Philosophy, University Bayreuth
Ulrich Krause
Department of Mathematics, University Bremen
Abstract When does opinion formation within an interacting group lead to consensus,
polarization or fragmentation? The article investigates various models for the dynamics
of continuous opinions by analytical methods as well as by computer simulations. Sec-
tion 2 develops within a unified framework the classical model of consensus formation,
the variant of this model due to Friedkin and Johnsen, a time-dependent version and a
nonlinear version with bounded confidence of the agents. Section 3 presents for all
these models major analytical results. Section 4 gives an extensive exploration of the
nonlinear model with bounded confidence by a series of computer simulations. An ap-
pendix supplies needed mathematical definitions, tools, and theorems.
Keywords: opinion dynamics, consensus/dissent, bounded confidence, nonlinear dy-
namical systems.
1. INTRODUCTION
Consider a group of interacting agents among whom some process of opinion formation takes place.
The group may be a small one, e.g., a group of experts asked by the UN to merge their different
assessments on, say, the magnitude of the world population in the year 2020, into one single judge-
ment. On the other extreme, the group may be an entire society in which the individuals ruled by
various networks of social influences develop a wide spectrum of opinions. In each case there is a
process of opinion dynamics that may lead to a consensus among the agents or to a polarization
between the agents or, more general, that results in a certain fragmentation of the patterns of opin-
ions. An understanding if not an analysis of the involved process of opinion formation is hardly
possible without an explicit formulation of a mathematical model.
An early formulation of such a model was given by J.R.P. French in 1956 in order to understand
complex phenomena found empirically about groups (French 1956). See also the related mathe-
matical analysis by F. Harary in 1959 (Harary 1959). Another source of opinion dynamics is the
work by M.H. De Groot in 1974 (De Groot 1974) and by K. Lehrer in 1975 (Lehrer 1975) and sub-
sequent work by C. G. Wagner, S. Chatterjee, E. Seneta, J. Cohen among others. (See (Wagner
1978), (Chatterjee 1975), (Chatterjee and Seneta 1977), (Cohen et al. 1986) and the references
given therein. There is an interesting connection to the Delphi technique for pooling opinions of
experts, (cf. (De Groot 1974)). The focus of these contributions is on consensus and how to reach it.
This is true in particular for the encompassing study of Lehrer and Wagner (Lehrer and Wagner
*
The major part of the simulation results in ch. 4 were produced during the academic year 1999 - 2000 when
Rainer Hegselmann was a member of the research group "Making Choices" at the Centre for Interdiscipli-
nary Research (ZiF) of the University of Bielefeld. He would like to thank ZiF for hospitality and the stimu-
lating intellectual environment. Many thanks as well to the participants of the three international confer-
ences (1998, 1999, and 2000, all organized by this research group) where R. Hegselmann could present the
basic approach and first results.
J
ournal of Artifica
l
Societies and Social Simulation (JASSS) vol.5, no. 3, 2002
http://jasss.soc.surrey.ac.uk/5/3/2.html
2
1991) which emphasizes rational consensus as a fundamental feature from justice to epistemology.
The importance of modelling disagreement beside consensus was pointed out especially by R. P.
Abelson (Abelson 1964) and by N.E. Friedkin and E.C. Johnsen (Friedkin and Johnsen 1999). For
example, there is the following remark in (Abelson 1964, p. 153): “Since universal ultimate agree-
ment is an ubiquitous outcome of a very broad class of mathematical models, we are naturally led to
inquire what on earth one must assume in order to generate the bimodal outcome of community
cleavage studies. Offhand, there are at least three potential ways of generating such a bimodal out-
come.”
From French to Friedkin/Johnsen, as difficult as these models might be in their details, they are all
comparatively simple in the sense that they are all linear models. This means in particular that the
analysis needed can be carried out by powerful linear techniques such as matrix theory, Markov
chains and graph theory. The first nonlinear model within this line of thinking was formulated and
analyzed in (Krause 1997) and (Krause 2000). For this model see also (Beckmann 1997), (Dittmer
2000), (Dittmer 2001) and (Hegselmann and Flache 1998). A model similar in spirit has recently
been investigated in (Weisbuch et al. 2001) and (Deffuant et al. 2000). Of course, having an ex-
plicit mathematical model does not mean at all that one has explicit mathematical answers. Already
R. Abelson observed that a “More drastic revision of the model can be introduced by making the
system nonlinear”. (Abelson 1964, p. 150). Though elementary, the model is nonlinear in that the
structure of the model changes with the states of the model given by the opinions of the agents (see
Section 2). Not only that helpful mathematical tools like Markov chains are no longer applicable, it
turns out, moreover, that rigorous analytical results are difficult to obtain. For that reason we carry
out the analysis of the above nonlinear model to a large extent by simulations on the computer. In-
deed, though we are proud to present analytical results, too, we want to emphasize the importance
of careful computer simulations for social dynamics in general, and opinion dynamics in particular,
whenever nonlinearities are involved. (See (Hegselmann and Flache 1998). For computer simula-
tions on binary opinion dynamics see (Holyst et al 2001), (Latané and Nowak 1997), (Stauffer
2001), (Stocker et al. 2001).
In this paper we address the classical question of reaching a consensus as well as that of disagree-
ment leading to polarization and, more general, the dynamics of opinion fragmentation of which
consensus and polarization are just two interesting special cases. Section 2 of the paper develops
and reviews within a unified framework four models of continuous dynamics: The classical model
of consensus formation, a variation of this model due to Friedkin and Johnsen, a model with time-
dependent (inhomogeneous) structure and a nonlinear model dealing with bounded confidence. Sec-
tion 3 then presents for all these models major analytical results of which some are well-known and
some are new. For the convenience of the reader we state the results with a minimum of mathemat-
ics. Precise mathematical definitions and theorems together with additional hints are delegated to an
Appendix. Section 4 contains an extensive exploration of the nonlinear model with bounded confi-
dence by a series of computer simulations. In the presentation we took care to explain step by step
the computer experiments made.
2. MODELLING OPINION DYNAMICS
Consider a group of agents (or experts or individuals of some kind) among whom some process of
opinion formation takes place. In general, an agent will neither simply share nor strictly disregard
the opinion of any other agent, but will take into account the opinions of others to a certain extent
in forming his own opinion. This can be modelled by different weights which any of the agents puts
on the opinions of all the other agents. This process of forming the actual opinion by taking an av-
erage over opinions can be repeated again and leads, therefore, to a dynamical process in discrete
time. Intuitively, one may expect that this process of repeatedly averaging opinions will bring newly
formed opinions of different agents closer to each other until they flow into a consensus among all
agents. In the next two sections we will show that the dynamics of opinion formation can be much
3
more complex than one would intuitively expect, even in the above simple model. The crucial point
here is that the weights put on the opinions of others may change for reasons explained below. Only
for the classical case of constant weights and enough confidence among agents the phenomenon of
consensus is a typical one (cf., e.g., (De Groot 1974), (Lehrer 1975), and the pioneering but less
formal (French 1956)).
Let n be the number of agents in the group under consideration. To model the repeated process of
opinion formation we think of time as a number or rounds or of periods, that is a as discrete time
{
}
0,1,2,T = L . It will be assumed that the opinion of an agent can be expressed by a real number
as, e.g., in the case of an expert who has to assess a certain magnitude. This assumption is made for
simplicity, because already in that case the opinion dynamics considered can be quite intricate. This
case is sometimes referred to as “continuous opinion dynamics” in contrast to the, even more re-
stricted, case of “binary opinion dynamics” (cf. (Weisbuch et al. 2001)).
Later on we will also consider “higher dimensional” opinions (see the time-variant model in the
next section). For a fixed agent, say
i where 1 in≤≤ , we denote the agents opinion at time t (in
round t) by ( )
i
x
t . Thus ( )
i
x
t is a real number and the vector
1
() ( (), , ())
n
x
txt xt= L in n-
dimensional space represents the opinion profile at time t. Fixing an agent
i , the weight given to
any other agent, say j , we denote by
ij
a . To keep things simple we introduce
ii
a such that
12
1
ii in
aa a+++=L . Furthermore, let 0
ij
a ≥ for all ,ij. Having these notations, opinion forma-
tion of agent
i can be described as averaging in the following way
11 22
(1) () () ().
iii inn
x
taxtaxt axt+= + ++L (2.1)
That is, agent
i adjusts his opinion in period 1t + by taking a weighted average with weight
ij
a for
the opinion of agent j at time t. Of course, weights can be zero. For example, if agent
i disregards
all other opinions, this means
1
ii
a = and 0
ij
a = for
j
i≠ ; or, if i follows the opinion of j then
1
ij
a = and 0
ik
a = for kj≠ . It is important to note that the weights may change with time or with
the opinion, that is
(, ())
ij ij
aatxt= can be a function of t and/or of the whole profile ()
x
t . By
collecting the weights into a matrix,
(, ()) ( (, ()))
ij
Atxt a txt= , with n rows and n columns, we
obtain a stochastic matrix, i.e., a nonnegative matrix with all its rows summing up to 1. Thus, using
matrix notation, the general form of our model (GM) can be compactly written as
(1) (,())()for .(GM)xt Atxt xt t T+= ∈
The main problem we are dealing with in this paper is the following one: Given an initial profile
(start profile)
(0)
x
and the dynamics specified by the weights – what can be said about the final
behavior of the opinion profile, i.e., about
()
x
t for t approaching infinity? In particular, when does
the group of agents approach a consensus c, i.e., it holds
lim ( )
i
t
x
tc
→∞
= for all agents 1, ,in= L ?
In this generality, however, one cannot hope to get an answer, neither by mathematical analysis nor
by computer simulations. Therefore, we will treat various interesting specializations of the above
general model (GM). First, we begin with the classical model of fixed weights, i.e.,
(1) ()for , (CM)xt Axt t T+= ∈
where A is a fixed stochastic matrix and
()
x
t the column vector of opinions at time t. This model
has been proposed and employed to assess opinion pooling by a dialogue among experts (De Groot
1974), (Lehrer 1975).
4
There is an interesting variation of this model developed by (Friedkin and Johnsen 1990, 1999).
This model addresses opinion formation under social influence and assumes that agent
i adheres to
his initial opinion to a certain degree
i
g
and by a susceptibility of 1
i
g
− the agent is socially influ-
enced by the other agents according to a classical model. By this variation the classical model be-
comes
(
)
11
(1) (0)(1 ) () ()
iiiii inn
x
tgx gaxtaxt+= +− +L (2.2)
or, in matrix notation,
(1) (0)( ) ()for .(FJ)xt Gx I G Axt t T+= +− ∈
Here G is the diagonal matrix with the
i
g
, 0 1
i
g≤≤, in the diagonal and I is the identity ma-
trix. Obviously (CM) is a special case of (FJ), namely for
0, 1
i
g
in=≤≤. This model has been
used to estimate from experiments the susceptibility of agents to interpersonal influence (French
1956).
A model similar to (CM) but in continuous time, that is a system of differential equations instead of
difference equations, has been explored and applied rather early by (Abelson 1964). Though there
are similarities in the search for consensus among agents, in the present article we will stick to the
discrete time framework.
The models (CM) and (FJ), as well as Abelson’s model, are both linear which makes them system-
atically tractable by analytical methods. The next type of model is still linear but time-variant (or, in
the theory of Markov chains, inhomogeneous), that is
(1) ()()for , (TV)xt At xt t T+= ∈
where the entries of matrix
()At , i.e., the weights, are dependent on time only. The time variant
model (TV) portrays, e.g., the so called “hardening of positions” where agents put in the course of
time more and more weight on their own opinion and less weight on the opinion of others. In the
next section it will be shown that analytical results are still available, even for “higher dimensional”
opinions, but the results are less sharp than in the time-invariant classical model.
The most difficult type of model occurs if the weights depend on opinions itself because then the
model turns from a linear one to a nonlinear one. Thus, the model is of type (GM) where
(())Axt
does not explicitly depend on time. It is still completely hopeless to analyse the model in this gener-
ality. There is, however, a particular kind of nonlinearity which captures an important aspect of
reality and seems at the same time tractable. But, compared with the other models, analytical in-
sights are not so easy to obtain. For that reason we will investigate this model detailed and exten-
sively by computer simulations in Section 4. The model we are going to exhibit portrays bounded
confidence among the agents in the following sense. An agent
i takes only those agents j into ac-
count whose opinions differ from his own not more than a certain confidence level
i
ε . Fixing an
agent
i and an opinion profile
1
(, , )
n
x
xx= L this set of agents is given by
{
}
(, ) 1
ij i
Iix j n x x ε=≤≤ − ≤ (2.3)
where
⋅ denotes the absolute value of a real number. To make things not too complicated we as-
sume that agent
i puts an equal weight on all (,jIix∈ ). That is to say, in the light of the general
model (GM) we let the weights given by
() 0
ij
ax= for (, )jIix∉ and
1
() (,)
ij
ax Iix
−
= for
(, )jIix∈ ( ⋅ for a finite set denotes the number of elements). Thus the model with bounded con-
fidence is given by
5
(
)
1
(, ())
(1) ,() ()for .(BC)
ij
jIixt
xt Iixt x t t T
−
∈
+= ∈
∑
This model has been developed by (Krause 1997, 2000); see also (Beckmann 1997) and (Dittmer
2000, 2001). For another attempt to model a lack of confidence see (Deffuant et al. 2000), (Weis-
buch et al. 2001). There, in a pairwise comparison between agents, “opinion adjustments only pro-
ceed when opinion difference is below a given threshold” ((Deffuant et al. 2000) , p. 2).
3. ANALYTICAL RESULTS FOR THE VARIOUS MODELS OF OPINION DYNAMICS
In the following we present several major results for the various models. Some are well known,
others are less well known and some are new results.
A. The classical model
The classical model (CM) was given by
(1) ()
x
tAxt+= for
{
}
0,1,2,tT∈= L ,
where A is a stochastic matrix. Obviously, ( ) (0)
t
xt Ax= for all tT∈ and, hence, the analysis
amounts to analyze the powers of a given matrix.
Result 1: On consensus
· If any two agents put jointly a positive weight on a third one then for every initial opin-
ion profile a consensus will be approached (which consensus, of course, depends on the
initial profile).
· More general, a consensus will be approached for every initial opinion profile if and only
if finally for t big enough any two agents put jointly a positive weight on a third one.
For a formal statement of this result together with additional hints see Theorem 1 in the Appendix.
A special case of Result 1 is given if every agent puts a positive weight on any other agent. This
corresponds to the intuition that the group should approach a consensus if every agent takes the
opinion of any other agent into account. On the other extreme is the case where every agent sticks
to his opinion without taking care of any other opinion. In this case A is the identity matrix and, of
course, there is no consensus, except the special case of a consensus already in the initial opinion
profile. Result 1 describes the confidence pattern between these extremes that allows for a consen-
sus. An example for three agents is given by the stochastic matrix
11
22
21
33
13
44
0
0
0
A
=
.
For this matrix A the power
2
A is strictly positive (does not contain any zero). A matrix for which
some power is strictly positive is also called primitive. By Result 1 a primitive matrix A is suffi-
cient for consensus. (The second part of Result 1 provides a condition that is not only sufficient but
necessary, too.)
Moreover, by employing what is called the normal form of Gantmacher, a more refined result is
possible (see Theorem 2 in the Appendix). Intuitively, the group of agents can be split up into sub-
groups such that the first g isolated subgroups consist only of essential agents in the sense that an
agent puts weight only on agents in his group. Such a subgroup has primitive structure if its subma-
trix of weights is primitive.
6
Result 2: On opinion fragmentation
· For any given initial profile the opinion dynamics approaches a stable opinion pattern if and
only if the subgroups of essential agents are all primitive. For the final stable opinion pattern
only the initial opinions of essential agents play a role.
· In particular, the stable opinion pattern reduces to a consensus if and only if there exists just
one subgroup of essential agents
(1)g = .
(For more information see theorem 2 in the Appendix.)
An example of four agents is given by
1
1
2
2
1
2
3
3
13
44
11
22
00
00
.
00
00
A
=
There are two subgroups of essential agents formed by agents 1 and 2 and agents 3 and 4, respec-
tively. The final stable opinion pattern consists of a partial consensus between agents 1 and 2 and a
partial consensus between agents 3 and 4. In general one obtains no consensus between these sub-
groups.
B. The Friedkin-Johnsen model
This model (FJ) was given by
(1) (0)( ) ()
x
tGxIGAxt+= + − fortT∈ .
If
0G = then (FJ) specializes to the classical model and it suffices to discuss 0G ≠ .
Result 3: On opinion fragmentation with positive degrees
· If there is at least one agent with positive degree and there are “enough positive weights”
then for any given initial profile the opinion dynamics approaches a stable opinion pattern
that can be computed from the weights, the degrees, and the initial profile.
· The above stable pattern represents a consensus if and only if there prevails already a con-
sensus among all agents with positive degree.
(For more information see Theorem 3 in the Appendix.)
Thus, in the case of positive degrees one can expect a consensus only for very special initial pro-
files. Assuming a primitive structure for the whole group this is very different from what holds for
the classical model.
An example for four agents is given by
.220 .120 .359 .300
.147 .215 .344 .294
, diagonalof
0010
.089 .178 .446 .286
AGA
==
In this case all agents have positive degree and a consensus is possible only if there is a consensus
at the beginning. For
(0) (25,25,75,85)x = , e.g., one obtains as final opinion pattern (60, 60, 75,
75) (Friedkin and Johnsen 1999, p. 6). Since A has a strictly positive column, in the classical model
one would obtain for
(0)
x
as above a consensus according to Result 1.
7
C. The time-variant model
The model with time-variance (TV) was given by
(1) ()()
x
tAtxt+= for tT∈ ,
where the weights collected in matrix
()At depend on time t. As one might expect from the classi-
cal model, the time variance will be no obstruction to a consensus as long as the weights remain
sufficiently positive. If, however, the weights tend to zero very quickly then consensus disappears.
This interesting phenomenon we illustrate by a simple example which works already for a group of
two agents. Suppose agent 1 does never care about the opinion of agent 2. In a first scenario let
agent 2 put at time t, for
2t ≥ , the weight
1
t
−
to agent 1 and the remaining weight
1
1 t
−
− to him-
self. Thus, the position of agent 2 is hardening by putting less and less weight to agent 1. It is not
difficult to verify that
11
212
() (1 ) (2) (2)xt t x t x
−−
=− + and, because of
11
() (0)xt x= for all t, that
2
()
x
t tends to
1
(0)x . Therefore, the two agents approach always a consensus in spite of hardening
the positions. In a second scenario let agent 2 hardening his position faster, by putting at time t, for
2t ≥ , the weight
2
t
−
to agent 1 and
2
1 t
−
− to himself. A little computation shows that
2
()
x
t tends
to
(
)
1
12
2
(2) (2)xx+ and, of course,
11
() (0)xt x= for all t. Thus, for many initial profiles a consen-
sus will not be approached. This little example shows that approaching a consensus does depend on
the speed by which the weights are changing (cf. (Cohen 1986)). This vague insight is made more
precise by the following result.
Result 4: One time-variance consensus
Consensus will be approached for every initial opinion profile, provided there exists a sequence of
time points with the following property:
For the accumulated weights
ij
b between time points it holds that for any two agents i and j there
exists a third one k such that
ik
b and
jk
b are positive and the minima of both sum up to infinity for
summing over all time points. Roughly speaking, the result says that for a consensus the weights
cannot tend too fast to zero because their sum should be infinite. For a precise statement see Theo-
rem 4 in the Appendix. The above result can be formulated and proved also for the case of multidi-
mensional opinions, that is, ( )
i
x
t is not just a number but a vector in higher dimensions. The proof
for this result is telling because it shows that the set in higher dimensions which is spanned by the
agents’ opinions must shrink to the point of consensus if time develops. (See the Lemma before
Theorem 4 in the Appendix.)
D. Opinion dynamics with bounded confidence (BC)
The (BC) model was given by
(
)
1
(, ())
( 1) , () ()for ,
ij
jIixt
x
tIixt xttT
−
∈
+= ∈
∑
where
{
}
(, ) 1
ij i
Iix j n x x ε=≤≤ −≤
and 0
i
ε > is the given confidence level of agent i . Though some properties hold still for different
levels of confidence, in the following we will assume throughout a uniform level of confidence, i.e.,
i
εε= for all agents i . Later on, in the next section when dealing with simulations, we will consider
not only symmetric confidence intervals, i.e.,
[]
,εε−+ , but also asymmetric confidence intervals,
8
i.e.,
[]
,
lr
εε−+ where ,
l left r right
εε ε ε==are positive but can be different from each other. The set
(, )
I
ix in the asymmetric case is then given by
{
}
(, ) 1
ljir
Iix j n x xεε=≤≤−≤ −≤ .
In the asymmetric case
lr
εε≠ we can have a one sided split between two agents i and j, namely
ljir
xxεε<−≤ if
lr
εε< and
rjil
xxεε<−≤ if
rl
εε< . In a one sided split one agent (i if
lr
εε<
and j if
rl
εε< ) takes the other agent (j and i , respectively) into account, but not vice versa.
The confidence levels ,
lr
εε serve as parameters of the model and the set of possible values of
(, )
lr
εε is called the parameter space.
A particular feature of this model, compared with the other models discussed, is that a consensus
will be reached in finite time, if there is a consensus at all. For an opinion profile
1
(, , )
n
x
xx= L we
say that there is a split (or crack) between agents
i and j if
ij
xx ε−>.
An opinion profile
1
(, , )
n
x
xx= L we call an ε -profile if there exists an ordering
12 n
ii i
x
xx≤≤≤L
of the opinions such that two adjacent opinions are within confidence, i.e.,
1
for all 1 1
kk
ii
x
xknε
+
−≤ ≤≤−
First we collect some fundamental properties of this model for the case
lr
εε= (see ( Krause 2000)).
Properties
I. The dynamics does not change the order of opinions, i.e., () ()
ij
x
txt≤ for all ij≤ im-
plies that
(1) (1)
ij
xt x t+≤ + for all ij≤ .
II. If a split between two agents occurs at some time it will remain a split forever.
III. If for an initial profile a consensus is approached then the opinion profile must be an
ε -
profile for all times.
IV. For
2,3, 4n= a consensus is approached if and only if the initial profile is an ε -profile.
The last property (IV) is no longer true for
5n= . That is, for five agents it can happen that they
don’t approach a consensus though their initial opinions are close in the sense of an
ε -profile. If,
however, the initial profile is an equidistant
ε -profile, i.e., the distance between adjacent opinions is
exactly
ε , then for a number of five agents, too, a consensus will be approached. A careful consid-
eration shows that for a number of six agents even in the case of an equidistant
ε -profile a consen-
sus cannot be reached. These remarks indicate that the number of agents involved can matter for the
dynamics. For the simplest case of just two agents it is not difficult to give a complete analysis of
the dynamics. For an arbitrary n, however, the mathematical analysis of the dynamics turns out to
be rather difficult and up to now there are only a few general results available. For that reason we
will present in the next section an extensive analysis of this model for higher values of n by com-
puter simulation. Of the few general results we present two fundamental ones.
Result 5: On consensus for bounded confidence
Consensus will be approached for a given initial opinion profile, provided for an equidistant se-
quence of time points the following property holds:
For any two agents
i and j there exists a third one k such that a chain of confidence leads from i
to k as well as from j to k. In this case, the consensus is reached in finite time.
9
1
0
1
e
left
e
right
(See Theorem 5 in the Appendix for a precise formal statement.)
Result 6: On opinion fragmentation for bounded confidence
For any given initial profile there exists a finite
*
tT∈ and a division of all agents into maximal
subgroups such that within each of the subgroups there holds a consensus for all
*
tt≥ . (Of course,
those partial consenses will be different from each other in general.)
(See Theorem 6 in the Appendix for a precise formulation.)
4. SIMULATIONS
In this section we will explore the model with bounded confidence by means of simulations. We
want to do that, firstly, in a systematic way, and , secondly, following the KISS-principle: Keep it
simple, stupid! Under that principle it is fairly natural to start in a first step with homogenous and
symmetric
e-intervals. Homogeneity means that the size and shape of the confidence interval is the
same for all agents. Symmetry means, that the interval has the same size to the left and to the right,
i.e.
lr
ee= . In a second step we will analyse different types of asymmetry.
4.1 SYMMETRIC CONFIDENCE
With continuous, one–dimensional, and normalized opinions x taken from the interval
[
]
0,1x Î
only confidence intervals 0 , 1
lr
ee££ make sense. Thus, the parameter space is given by the unit
square. The diagonal represents symmetric confidence. A systematic approach could be 'walking
along the diagonal', starting the trip at the point
0,0 .
Figure 1: The parameter space, walking along the diagonal.
We will randomly generate a start distribution of 625 opinions. Updating is simultaneous. Figure 2
shows three stops on the tour along the diagonal. The stops are single runs. They all get going with
the same start distribution. The ordinate indicates the opinions. Since it is useful for the following
analysis the opinions are additionally encoded by colours ranging from red (
0x = ) to magenta
(
1x = ). The abscissa represents time and shows the first 15 periods. Obviously it takes less than 15
periods to get a stable pattern: With the quite small confidence interval of 0.01
lr
ee== exactly 38
different opinions survive in the end. Under the much bigger confidence of 0.15
lr
ee== the agents
end up in two camps, and with 0.25
lr
ee== the result of the dynamics is consensus. As it looks,
the size of the confidence interval really matters.
10
Figure 2 shows only single runs. To get a better feeling of what is going on we run systematically
simulations walking along the diagonal. The simulations start with 0.01
lr
ee== , 0.02
lr
ee== ,
…, 0.4
lr
ee== . (For reasons that become obvious below, there is nothing new and interesting in
the parameter space for , 0.4
lr
ee> .) For each of these 40 steps we repeat the simulation 50 times,
always starting with a different random start distribution. Each run is continued until the dynamics
becomes stable. Figure 3 gives an overview.
(a)
0.01
lr
ee==
(b)
0.15
lr
ee==
(c)
0.25
lr
ee==
Figure 2: Stops while walking along the diagonal
11
20
40
60
80
100
10
20
30
40
0.05
0.1
0.15
0.2
20
40
60
80
100
Figure 3: Walking along the diagonal – simulation results.
The x-axis in Figure 3 represents the opinion space
[
]
0,1 divided into 100 intervals. Our 40 step
walk along the diagonal is represented by the y-axis. (Note: the steps are not time steps!) The z-axis
represents the average (!) relative frequencies of opinions in the 100 opinion intervals of the opinion
space after the dynamics has stabilised.
Figure 3 deserves careful inspection: At the beginning of our walk along the diagonal of Figure 1,
i.e. the y-axis of Figure 3, there is only little confidence. For example, step 2 or 3 means that
0.01
lr
ee== or 0.02
lr
ee== . In terms of single runs we are speaking about opinion dynamics like
that in Figure 2a. The z-values show that as an average we find under that little confidence a small
fraction of the opinions in all intervals of the opinion space. (Note: That does not imply that in a
single run all intervals are occupied after stabilisation; see Figure 2a and Figures 12a, 12b.) No part
of the opinion space seems to have especially high or especially low frequencies. That changes as
we step further along the diagonal. Look, for instance, at step 15, i.e. 0.15
lr
ee== . A single run
example for a dynamics based on confidence intervals of this size is given by Figure 2b, where we
end with two camps holding different opinions. Figure 3 makes it clear that this is a fairly typical
result: As we step forward along the diagonal the average distribution of stabilised relative frequen-
cies becomes less and less uniform. To the left and to the right of the centre mountains of increasing
height emerge. As we continue our walk on the diagonal the landscape changes dramatically again:
At about step 25, i.e. 0.25
lr
ee== the 'Lefty' and the 'Righty Mountains' come to a sudden and
steep end. At the same time a new and steep centre mountain emerges. An example for a single run
in this region of the parameter space is given by Figure 2c. Figure 3 shows that this single run with
a total confidence interval including half of the opinion space is a typical run, leading to a consen-
sus including all or almost all agents. Except in the centre almost all opinion intervals are empty and
the corresponding opinions are eliminated.
Summing up and generalising what we see in Figure 3 one might say: As the homogeneous and
symmetric confidence interval increases we transit from phase to phase. More exactly, we step from
fragmentation (plurality) over polarisation (polarity) to consensus (conformity) .
x
z
y
12
The e-profile splits in t
6
. From now on
the split sub–profiles belong to different
'opinion worlds' or communities which
do no lon
g
er inte
r
ac
t
.
E
xtreme opinions are under a one sided influence and move
direction centre. The ran
g
e of the
p
rofile
s
hrinks.
At the extremes opinions condense.
Condensed regions attract opinions from less populated
areas within their
e–reach. In the centre opinions > 0.5
move upwards, opinions < 0.5 move downwards.
That a dynamics, governed by averaging among those opinions which are within a certain fairly
small confidence interval, leads to an evenly distributed variety of opinions (plurality) will probably
not surprise that much (see Section 3, part D, Result 6). That fairly large confidence intervals drive
our dynamics towards conformity is probably even less surprising (see Section 3, part D, Result 5).
Big confidence intervals should drive all opinions direction centre and there they converge. (For all
values , 0.4
lr
ee> the result is always conformity, and that is the reason why we stop the walk on
the diagonal at the point 0.4
lr
ee== .) Thus, the real task of understanding is the phase in between:
What mechanisms drive our dynamics under ‘middle sized’ confidence into polarisation?
Figure 4: Regular start profile, 0.2
lr
ee== .
Figure 4 gives decisive hints to understand polarity. The dynamics starts with a regular profile, i.e.
a profile for which the distance between any two neighbouring opinions is the same. A grey area
between two neighbouring opinions indicates that the distance between the two is not greater than
,
lr
ee. Careful inspection shows:
· Extreme opinions are under a one sided influence and move direction centre. As a conse-
quence the range of the profile shrinks.
· At the extremes opinions condense.
· Condensed regions attract opinions from less populated areas within their e–reach.
· In the centre opinions > 0.5 move upwards, opinions < 0.5 move downwards.
· The e–profile splits in period t
6
. From now on the split sub–profiles belong to different
'opinion worlds' or communities which do no longer interact.
Figure 5 shows one aspect of the dynamics in Figure 4 more in detail. The x–axis gives the periods,
the y–axis of Figure 5 shows the change from one period to the next, i.e.
(1) () (1) 1,2,
iii
txtxtt∆−= − − = L (4.1)
for all opinions from Figure 4. The opinions are indicated by their colour.
13
Figure 5: Opinion changes from period to period (50 opinions, regular start profile,
0.2
lr
ee== ).
Figure 5 makes clear that opinion changes start at the extremes and are there (at the start!) most
extreme for the most extreme opinions. Opinions next to 1 (magenta) move downwards (negative
i
D ), opinions next to 0 (red) move upwards (positive
i
D ). From period 0 to 1 nothing changes more
in the middle of the opinion space. But as times goes by changes work through the opinion space
direction centre. The opinions directly in the centre are the last to be affected. They make the big-
gest ‘jumps’. From period 7 to 8 onwards nothing changes anymore, i.e. 0
i
D= for all opinions
i
x
.
The effects described above become in some respect more obvious with more opinions. Figure 6
shows for 500 opinions what Figure 5 showed for 50. The opinions of the regular profile are again
indicated by their colours. Since colouring of the profile is done sequentially, an earlier coloured
opinion may later become hidden by other opinions. Figure 6, top shows a colouring of the profile
in an ascending order (0 to 1) while Figure 6 follows a descending order (1 to 0). By visual inspec-
tion it becomes immediately clear that changes start at the extremes and reach the centre only with
some delay.
The decisive key for an understanding of polarity are obviously the splits in the
e-profile, induced
by a one sided–influence at the extremes, a shrinking range of opinions combined with an increas-
ing frequency of opinions in certain areas of the opinion space which attract opinions in less ‘popu-
lated’ areas within their reach. Under simultaneous updating cracked profiles can never get con-
nected again. Agents/opinions outside ones own sub profile are ‘out of range/reach’ in a quite se-
vere sense: They are not only outside ones own confidence interval, but also outside the confidence
intervals of all others one takes seriously, outside the confidence intervals of all others which those
others take seriously etc.
14
Figure 6: Opinion changes from period to period (500 opinions, regular start profile,
0.2
lr
ee== ). Top: Colouring in an ascending order. Bottom: Colouring in an descending
order.
The splits do not only explain polarity, they explain the stabilisation of our dynamics in general. For
fairly small confidence intervals the stabilisation leads to a fairly high number of surviving opin-
ions, i.e. plurality. Figure 7 shows a regular start profile with 100 opinions and 0.05
lr
ee== . The
profile splits 8 times. Again a grey area between two neighbouring opinions indicates that the dis-
tance between the two is not greater than ,
lr
ee. For ‘middle sized’ confidence intervals we get only
a small number of surviving opinions, i.e. polarisation; Figure 4 gives an example. Under large
confidence intervals the profile splits never or the splits leave alone extreme and quite small minori-
ties while an overwhelming majority converges in the centre of the opinion space. Figure 8 gives an
15
example of total consensus. Note that even the four opinions in the centre of the opinion space
move for a short while out of the centre. But they merge there again much earlier than the upper and
lower part of the profile arrives in the centre as well.
Figure 7: 100 opinions, 0.05
lr
ee== , 8 splits.
Figure 8: 100 opinions 0.25
lr
ee== , no split, total consensus.
The simulation results from Figure 3 are based on randomly generated start profiles (uniform distri-
bution), not on regular start profiles. But the effects described so far do not essentially depend on
the regularity of the start profile. What irregularity adds are density fluctuations in the initial distri-
bution of opinions. They are additional causes for splits and induce opinion changes deep inside the
start–profile without any delay right at the beginning of the process.
16
1
0
1
e
left
e
right
4.2 ASYMMETRIC CONFIDENCE
Up till now we considered walks along the diagonal of the parameter space (Figure 1), i.e. symme-
try in the sense that
lr
ee= . But confidence may be asymmetric. One can think of several types of
asymmetry. In the following we will analyse two cases. In the first case (4.2.1) the asymmetry is
independent of the opinion an agent holds. Whatever their opinion might be, all agents have the
same confidence intervals ,
lr
ee with
lr
ee¹ . In the second case (4.2.2) the asymmetry is dependent
upon the opinion the agent holds: An agent with an opinion more to the right [left] has more confi-
dence in the right [left] direction.
4.2.1 OPINION INDEPENDENT ASYMMETRY
How to get an overview about what is going on under opinion independent asymmetries? Again our
approach is a systematic walk through the parameter space. But instead of taking the route along the
diagonal we now walk on straight lines below (or above) the diagonal as indicated by Figure 9. For
this type of asymmetry it does not matter whether confidence is biased to the right or biased to the
left. Therefore it is only one of the triangles, either the one below or the one above the diagonal, that
has to be analysed. We confine ourselves to the triangle below, i.e. a bias to the right. For all effects
we find in that area of the parameter space there exist corresponding effects in the triangle above
(bias to the left).
Figure 9: Research strategy for opinion independent asymmetries.
To get a first feeling we look at the three single run examples of asymmetric confidence in Figure
10. We find phenomena we are already familiar with, for instance fast stabilisation, plurality, polar-
ity, and conformity. But obviously the asymmetric confidence drives the dynamics somehow into
the direction favoured by the asymmetry, i.e. to the right.
17
(a)
0.02
0.04
l
r
e
e
=
=
(b)
0.03
0.15
l
r
e
e
=
=
(c)
0.10
0.25
l
r
e
e
=
=
Figure 10: Single runs, 625 opinions, random start profile.
To get a more systematic overview we present the results of four stepwise walks below the diago-
nal. In the first walk we follow the straight line 0.9
lr
ee= . We start with 0.01
r
e = and make 40
18
steps until we get to the point 0.36, 0.4
lr
ee==. For each value of these 40 steps we repeat the
simulation 50 times, always starting with a different random (uniform) start distribution. Each run is
continued until the dynamics becomes stable. The other three walks follow in the same way the
lines 0.75
lr
ee= , 0.5
lr
ee= , and 0.1
lr
ee= . We always stop when 0.4
r
e = is reached. Figure 11
gives an overview.
20
40
60
80
100
10
20
30
40
0.05
0.1
0.15
0.2
20
40
60
80
100
(a) 0.9
lr
ee=
20
40
60
80
100
10
20
30
40
0.05
0.1
0.15
0.2
20
40
60
80
100
(b) 0.75
lr
ee=
19
20
40
60
80
100
10
20
30
40
0.05
0.1
0.15
0.2
20
40
60
80
100
(c) 0.5
lr
ee=
20
40
60
80
100
10
20
30
40
0.05
0.1
0.15
0.2
20
40
60
80
100
(d) 0.25
lr
ee=
20
20
40
60
80
100
10
20
30
40
0.05
0.1
0.15
0.2
20
40
60
80
100
(e) 0.1
lr
ee=
Figure 11: Walking below the diagonal – simulation results.
Inspection of
Figure 11a–e supports the following observations:
· As
r
e (and thereby
l
e ) increases all four walks finally lead again into a region of the parame-
ter space where
consensus prevails. But as
l
e compared to
r
e becomes smaller and smaller
the resulting consensus moves into the favoured, i.e. here into the right direction.
· For a very small
r
e and
l
e the dynamics stabilises with a lot of surviving opinions. Thus
again we have a phase one might coin
plurality.
· As
r
e and
l
e increases polarisation emerges. As long as
l
e is only a little bit smaller than
r
e (cf. Figure 11a,0.9
lr
ee= ) it is the type of polarisation we know from the symmetric
case: in the left and in the right of the of the opinion space extreme opinion camps emerge,
grow, and get closer to each other with an increasing
r
e (and thereby increasing
l
e ). It is a
somehow a
'symmetric' polarisation: The camps have about the same size and the same dis-
tance from the centre (or the borders, respectively) of the opinion space.
· As
r
e becomes significantly greater than
l
e we observe a new type of asymmetric polarisa-
tion. The most obvious effect is that a big opinion camp emerges at the right border of the
opinion space. This effect is extreme if
l
e is only a small fraction of
r
e (cf. Figure 11e). To
the left of this main camp – in a certain distance, but still to the right of the centre of the
opinion space– we observe smaller but nevertheless outstanding frequencies. This reflects
the fact, that asymmetric confidence tends to produce in a certain region of the parameter
space two or few opinion camps of different size: The bigger one normally more to the bor-
der of the opinion space.
Figures 11a – 11e (and similarly Figure 3) show how many opinions on an average over 50 simu-
lation runs end up (after stabilisation) in each of the 100 intervals in which the opinion space was
divided. Thus, one
does not generally see, how many opinions on the average survive at all. This
information is given by
Figures 12a and 12b. Both figures show the average number of different
21
opinions that survive after the dynamics has stabilised. The results are based on 25 simulation runs
for each pair
, , , 0, 0.02, 0.04, 0.06, ..., 0.4
lr lr
e eee= . All simulations start with 625 randomly
generated opinions.
Figure 12a shows the number n of surviving opinions with 10n £ , while Fig-
ure 12b
is a detail of Figure 12a with 2n £ as the upper limit of the z–axis. Both figures clarify a
bit more our speaking of plurality, polarization, and consensus as three different phases: For very
small confidence symmetric or asymmetric intervals lots of opinions survive (
plurality). But as
either
l
e or
r
e or both increase we observe a sharp decline of surviving opinions. Soon one gets to
the green–yellow base where for the most part only 2 opinions survive (
polarization). Figure 12b–
in principle a magnification of
Figure 12a– shows that with an further increase in
l
e or
r
e this po-
larization turns into
consensus.
Figure 12c shows the final average opinion in the whole population for all points
,
lr
ee with , 0.4
lr
ee£ , again based on 25 simulation runs with 625 random opinions at the begin-
ning. It is no surprise that for symmetric confidence this final average opinion is about 0.5. With
asymmetric confidence the mean opinion moves into the direction favoured by the asymmetry. This
effect is extreme if there is only little confidence in the non favoured direction.
Figure 12d shows
that the effect becomes milder as the confidence in the non favoured direction increases as well.
A decisive step for an explanation of the phenomena stated above is an understanding of the
new
type of
splits that occur under asymmetric confidence. Since
lr
ee¹ it may be the case that
1
() ()
iir
xt xt e
+
-£, while
1
() ()
iil
xt xt e
+
->, where
i
x
and
1i
x
+
are two neighbouring opinions in our
e –profile. In such a situation the opinion
1
()
i
x
t
+
affects the opinion ( )
i
x
t since
1
()
i
x
t
+
is within
the
r
e –reach of ( )
i
x
t . But at the same
1
()
i
x
t
+
is not affected (any longer) by ( )
i
x
t since ( )
i
x
t is out-
side the
l
e –reach of
1
()
i
x
t
+
. We thus have a one–sided split (see definition in Section 3, part D). In
the opinion dynamics given by
Figure 13 several one–sided splits occur: Where ever a grey area
between two neighbouring opinions has
only white lines in the direction top right, there we have a
one–sided split. In contrast to that a pattern resulting from both, white lines direction bottom right
and white lines direction top right, indicates that the two neighbouring opinions are mutually within
their relevant
e–reach: ( )
i
x
t in the
l
e –reach of
1
()
i
x
t
+
, and
1
()
i
x
t
+
in the
r
e –reach of ( )
i
x
t .
0
2
4
6
8
10
0
2
4
6
8
10
0
0.5
1
1.5
2
0
0.5
1
1.5
2
Figure 12a: Number of remaining opinions af-
ter stabilisation.
Figure 12b: Number of remaining opinions
after stabilisation, magnification of Figure 12a.
22
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
Figure 12c: Final average opinions u after
stabilisation
Figure 12d: Detail of Figure 12c: final average
opinions
0.3 0.7u££ after stabilisation.
Legend for 12a–12d
Figure 12: Simulation results: Number of surviving opinions and final average.
Even without any split in the profile bare asymmetry of confidence drives the opinion dynamics in
the favored direction. But one–sided splits amplify this effect dramatically. How the mechanism
works can be seen in
Figure 13: Opinions right above a split become the new extreme left opinions
of a remaining
e–sub–profile that continues to converge. At least for a while this sub–profile is no
longer influenced by more left opinions
. The split off e–sub–profile below the one–sided split con-
verges and moves upwards. It is still under the influence of opinions above the one sided split. By
that the whole dynamics is driven to the right. Note that, contrary to two sided–splits one–sided
splits can close again.
0.4
e
left
e
right
0
0.4
23
One–sided split.
One sided splits can close. Two sided splits
never do that.
1
() ()
iir
xt xt e
+
-£
1
() ()
iil
xt xt e
+
-£
The split off e-sub–profile right below the one-sided split con-
verges and moves upwards. It is still under the influence of
opinions above the one sided split.
Opinions right above a split become the
new extreme left opinions of a remaining
e-sub-profile that continues to converge.
The sub–profile is
at least for while not
influenced by more left opinions.
lr
ee<
lr
ee=
lr
ee>
0
1
0.5
Figure 13: One sided splits (100 opinions,
0.8, 0.24
lr
ee==).
4.2.2 OPINION DEPENDENT ASYMMETRY
The confidence intervals analysed in 4.2.1 are asymmetric. But the asymmetry is independent of the
opinion
itself. However, it is a quite common phenomenon that that those holding a more left (right)
opinion often listen more to other left (right) opinions while being sceptical upon more right (left)
views. So it seems natural to model an
opinion dependent asymmetry in the following way: The
more left (right) an opinion is, the more the confidence interval is biased direction left (right). For
the special case of a 'centre' opinion, i.e.
0.5x = , the confidence interval should be symmetric.
Figure 14: Opinion dependent confidence intervals.
Figure 14 illustrates how –following this idea– a confidence interval of a given total (!) size e is
shifted to the left (right) for an opinion to the left (right) of the centre. For the centre opinion the left
and the right confidence interval are of equal size.
How to model opinion dependent asymmetries of confidence intervals? One possible strategy is to
introduce an opinion dependent bias
l
b
to the left and a bias
r
b
to the right such that , 0
lr
bb³ and
1
lr
bb+=. Having done that we use
l
b
and
r
b
to divide any given confidence interval e into a left
24
0.5
0.5
1
0 1
opinion x
bias
b
r
b
l
b
and a right part. Following our intuition stated above, the values
r
b
should be generated by a
monotonically
increasing function f of the opinion x with
[
]
0,1x Î . By 1()
f
x- we get the mono-
tonically
decreasing function that we need to generate the values
l
b
–again following the intuition
stated above. Different slopes for the function
f would then allow to model the strength of the bias.
Figure 15: Parameter space for opinion dependent asymmetry.
Figure
15 illustrates an easy way to elaborate in detail this type of asymmetry. The x-axis indicates
the opinion. The
y-axis or, respectively, the blue lines indicate the opinion dependent bias
(
)
r
x
b to
the right and
()
l
x
b to the left. The blue graphs are generated by rotations around the blue point
0.5,0.5 , i.e. according to the function
1
()
2
m
fx mx
-
=+
. It is
()
1
2
r
m
xmxb
-
=+
and
(
)
(
)
1
lr
x
xbb=- . These biases are used to determine how a confidence interval of any given size
e is partitioned into an
l
e and
r
e . We define
() ()
rr
uuebe= and
(
)
(
)
ll
uuebe= . Then it always
holds that
(
)
(
)
lr
xxee e+=. It is also guaranteed that
(
)
(
)
0.5 0.5
lr
ee= . In this setting it is the
slope
m in
1
()
2
m
fx mx
-
=+
that controls the strength of the bias. For 0m = we do not have any
bias. Both parts of the confidence interval have, whatever the opinion might be, the same size. As
m
increases (by rotating the blue graph anti clockwise around
0.5,0.5 ) the bias becomes stronger
and stronger: People with a more left (right) opinion listen less and less to the right (left) side of the
opinion space. We will confine ourselves to slopes within the range
01m££. To give an example:
For an opinion
0.6x = , a total confidence interval of 0.4e = , and 0.5m = we get 0.18
l
e = and
0.22
r
e = . For any positive m it holds that the more one's opinion is located to the left (right), the
more one's confidence interval is shifted to the left (right).
25
1
0
m
e
This approach offers an easy way to analyse the effects of opinion dependent asymmetries of confi-
dence intervals (
Figure 16): For different absolute sizes of confidence intervals we start with sym-
metry. In each case we let
m stepwise increase and study the resulting dynamics by means of simu-
lation. The analysed area of the parameter space will be
01m££.
Figure 16: Opinion dependent asymmetries: Analysing the parameter space.
Figure 17 gives an overview. The graphics are of the same type as in Figures 3 and 11. The x-axis
represents the opinion space
[
]
0,1 divided into 100 intervals. The z-axis represents the average (!)
relative frequencies of opinions in the 100 opinion intervals of the opinion space after the dynamics
has stabilised.
In contrast to the former figures the y-axis does not represent changes in e ; it now
represents
changes of the parameter m which controls the strength of the opinion dependent bias of
e . Along the y-axis m is increasing from 0 to 1 in 26 steps of size 0.04 (while in the former graph-
ics we saw 40 steps of an increasing
l
e or
r
e , respectively). Thus, each graphics represents a walk
along one of the blue horizontal lines in
Figure 16. Figures 17a to c show the simulation results
based on
0.2, 0.4, 0.6e = .
20
40
60
80
100
10
20
0.05
0.1
0.15
0.2
20
40
60
80
100
(a) 0.2e =
26
20
40
60
80
100
10
20
0.05
0.1
0.15
0.2
20
40
60
80
100
(b) 0.4e =
20
40
60
80
100
10
20
0.05
0.1
0.15
0.2
20
40
60
80
100
(c) 0.6e =
Figure 17: Increasing bias m (26 steps, 0,...,1m = ) for three different confidence intervals.
The simulations in
Figure 17a are based on 0.2e = . Along the y-axis we start (step 1) with 0m = ,
what implies 0.1
lr
ee== . Thus, we now get going where we got by step 10 when walking along
the diagonal in our first experiments with
symmetric confidence intervals (Figure 2 and Figure 3).
For a symmetric confidence interval of that size a very mild polarization starts to emerge: The
z-
values show that as an average we find (on the average!) small fractions of remaining opinions in
all intervals of the slightly shrunk opinion space. At the extremes the frequencies are a bit higher, a
27
consequence of the one-sided influence which drives the opinions direction centre. But we are far
away from a full fledged polarization as we will get for 0.25
lr
ee== (step 25 in Figure 3). Figure
17a, firstly, shows that a sufficiently high opinion dependent bias will produce a blatant polariza-
tion even under the condition of a comparatively small confidence interval. That polarization is,
secondly, more severe in the sense, that the opinion distances of the two major camps are greater
than the distances we observe in the symmetric cases. As
m increases the distance between the two
major camps becomes greater and greater. For an
m=1 the two camps occupy the most extreme po-
sitions 0 and 1.
Figure 17b shows the simulation results for an increasing m based on 0.4e = . 0m = corresponds
step 20 of our walk along the diagonal in the first experiments with
symmetric confidence intervals
(
Figure 2 and Figure 3). For a symmetric confidence interval of that size we got a clear polariza-
tion. With an increasing opinion dependent bias the polarization becomes even more dramatic under
both perspectives, size of the camps and the distance between them.
The simulation results for a total confidence interval of
0.6e = are shown in Figure 17c.
0m = corresponds step 30 of our walk along the diagonal in the first experiments with symmetric
confidence intervals (
Figure 2 and Figure 3). An all including consensus is the result and that re-
mains true for a mild opinion dependent bias
m. But a certain point (about step 11, i.e. 0.4m » ) that
consensus breaks down. A sharper and sharper polarisation is the final result.
(a) 0m = (b) 0.25m =
(c) 0.5m = (d) 0.75m =
28
1,
() ()
iiri
xt xt e
+
-£
1,1
() ()
iili
xt xt e
++
-£
(e) 0.99m =
Figure 18: Five different biases for 0.6e =
For a better understanding of the effects described so far it is helpful to look at single runs. Figure
18 shows a sequence of single runs in which the opinion dependent bias gets stronger and stronger.
All runs are based on the confidence interval
0.6e = and an increasing m. To keep things simple
we use a
regular profile of 50 opinions. The runs show: As m increases it takes longer to get to a
consensus. In
Figure 18c the bias is sufficiently strong to cause a break down of the former consen-
sus. In period 4 the profile splits finally and two polarised opinion camps remain. As
m increases
further, the distance between the two camps becomes greater. The principle cause for all these ef-
fects is that with an increasing
m those at the extremes become less and less affected by opinions
more in the centre of the opinion space. The decisive point becomes obvious by a comparison of
Figures 18a and 18e: Under symmetric confidence those at the extremes are under a one–sided in-
fluence that drives them direction centre, thereby causing a shrinking of the whole opinion space
(
Figure 18a). With an increasing opinion based bias the drive direction centre disappears or –
depending upon the size of the confidence interval– is significantly weakened
at the extremes. In
Figure 18e – based on an heavy bias of 0.99m = – the opinions at the extremes stay where they are.
At the same time they attract step by step opinions in whose (at least)
one– sided e –reach they are.
Thus, instead of a drive direction centre the opinion based bias generates a
drift to the extremes.
4.3 MAIN RESULTS
We can summarise our results, firstly, for the case of symmetric and opinion independent asymme-
tries:
· With an increasing symmetric or asymmetric confidence we step from plurality to polarisation
and then to consensus.
· Under (a)symmetric confidence polarisation is (a)symmetric as well.
· In the symmetric case the major causes for polarisation are splits in the opinion profiles. They
are caused by shrinking and condensing at the extremes, condensing induced by condensing,
and condensing induced by density fluctuations in randomly generated start profiles.
· In the case of asymmetric confidence an asymmetric polarisation is caused and amplified espe-
cially by one–sided splits. At least temporarily one of the two resulting sub–profiles influences
the other one, while not longer being influenced by the other one itself. This favours conver-
gence into the direction favoured by the asymmetric confidence.
29
· With asymmetric confidence mean and median move into the favoured direction. This effect is
extreme if there is only little confidence in the non favoured direction. The effect becomes
milder as the confidence in the non favoured direction increases.
As to the effects of
opinion dependent asymmetries we can, secondly, conclude:
· With an increasing opinion depending bias the drift direction centre is significantly weakened at
the extremes and –depending upon the size of both, the bias and the confidence interval– may
even totally disappear.
· For small confidence intervals which produce plurality in the symmetric case it holds: with an
increasing bias at least a moderate polarisation starts earlier.
· With an increasing bias polarization is amplified: The two opinion camps at the extremes be-
come bigger and their final position is more to the extremes.
· With an increasing bias reaching a consensus becomes more and more difficult. Polarisation is
the result instead. If consensus is still feasible, it takes more time to get there.
One might ask whether the results depend upon simultaneous updating. The answer is: no. Random
serial updating gives extreme opinions a slightly better chance to survive. But none of the results
stated above depends crucially on simultaneous updating.
Future directions of research will, firstly, include the analysis of opinion spaces of
higher dimen-
sions. (For a first analytical result see theorem 4 in the Appendix). Secondly, we will analyse the
effects of network structures in which interactions are restricted to
neighbouring others, i.e. indi-
viduals living, for instance, within ones v. Neumann or Moore neighbourhood. First simulations
show that this type of locality matters dramatically: If the neighbourhoods in which the agents in-
teract are fairly small (though overlapping!), then the phase in between plurality and consensus, i.e.
polarization, disappears. At least under bounded confidence locality of interactions may prevent
societies from sharp polarisation.
APPENDIX: Theorems and Hints
A. The classical model (GM)
The model (1) (),
x
tAxttT+= ∈ has the consensus property if for every (0)x ∈ R
n
there exists a
c∈R such that lim ( )
i
t
x
tc
→∞
= for all
{
}
1, ,in∈ L , all tT∈ .
Theorem 1
·
If for any two
{
}
,1,,ij n∈ L there exists some
{
}
1, ,kn∈ L such that 0
ik
a > and 0
jk
a >
then the consensus property holds.
· The consensus property holds if and only if there exists some
0
tT∈ such that the matrix
power
0
t
A contains at least one strictly positive column.
For the first part of Theorem 1 see (De Groot 1974), for the second part see (Berger 1981).
Theorem 2 Let A be in Gantmacher normal form with diagonal blocks , 1 ,
i
Aisgs≤≤ ≤.
· lim ( )
t
x
t
→∞
exists for every (0)x ∈Rn
n
if and only if the
i
A are all primitive for 1 ig≤≤ or,
equivalently, if 1 is the only root of
A of modulus 1.
· The consensus property holds if an only if 1g = and
1
A is primitive or, equivalently, if 1
is the only root of
A of modulus 1 and 1 is a simple root.
30
For a proof see (Gantmacher 1959).
B. The Friedkin-Johnsen model (FJ)
From the model (FJ), i.e.,
(1) (0)( ) ()
x
tGxIGAxt+= + − for tT∈
one obtains by induction
() () (0)
x
tVtx= for tT∈ , where
1
0
()
t
tk
k
Vt M M G
−
=
=+
∑
with ()
M
IGA=− .
For
0G = , model (FJ) specializes to (CM) and , ( )
t
M
AV t A==.
Theorem 3
· Let 0G ≠ and suppose A is an irreducible matrix. For every (0)x ∈R
n
there exists
()lim()
x
x
xt
→∞
∞= and one has the formula
1
()( ) (0)
x
IM Gx
−
∞= − .
· Consensus ()(,,)
x
cc∞= L holds if and only if (0)
i
x
c= for all i with 0
i
g > .
See (Friedkin and Johnsen 1990, Appendix).
C. Time-variant model (TV)
Consider multidimensional opinions, i.e., the opinion of agent i at time t is given by ( )
i
xt∈ R
m
,
where
1m≥ is the number of opinions considered. The (TV) model then reads
1
(1) ()()
n
ij
ij
j
x
tatxt
=
+=
∑
for 1,intT≤≤ ∈ .
This shows that ( 1)
i
xt+ is a convex combination of
1
(), , ()
n
x
txtL in R
m
. For points
1
,,
p
zz∈L
m
R
m
the set of all convex combinations of these points is denoted by conv
{
}
1
,,
p
z
zL .
For a subset
M ⊂R
m
the diameter of M is
{
}
()sup ' ,'dM a a aa M=− ∈, where ⋅ is the Euclidean norm on R
m
(but it could be any norm
on
R
m
).
Lemma For
1
,,
n
xx∈L R
m
and
1
,1
n
ik
ik
k
yaxin
=
=≤≤
∑
the following estimate holds
{}
(
)
{}
{}
(
)
11
1,
1
conv ,, 1min min , ,,
n
nn
ik jk
ij n
k
dyy aadxx
≤≤
=
≤−
∑
LL.
This estimate shows that the diameter of the set spanned by the agents’ opinions shrinks by a cer-
tain factor from one period to the next one.
This Lemma is the crucial step in proving the following theorem.
Let for
,
s
tT∈ with
s
t<
the matrix
(
)
(, ) (, )
ij
B
ts b ts= denote the matrix product
(1)(2) ()At At As−−L which models the accumulated weights between periods s and t.
31
Theorem 4 Suppose there exist a sequence
012
0 ttt=<<<L in T and a sequence
12
,,δδL in
[
]
0,1 with
1
m
m
δ
∞
=
=∞
∑
such that
{}
11
1
min ( , ), ( , )
n
ik m m jk m m m
k
btt b tt δ
−−
=
≥
∑
for all
1m≥ , all 1,ij n≤≤.
Then for any
1
(0), , (0)
n
xxL in R
m
there exists a consensus
{
}
*1
conv x (0), , (0)
n
xx∈ L , i.e.,
*
lim ( )
i
t
x
tx
→∞
= for all 1 in≤≤ .
Furthermore, for
1
(0), , (0)
n
yyL in R
m
with corresponding consensus
*
y one has the sensitivity
property that
**
1,
max (0) (0)
ij
ij n
xy x y
≤≤
−≤ −
.
For Theorem 4 in the special case of m = 1 see (Chatterjee 1975), (Chatterjee and Seneta 1977).
For Theorem 4 in case of more general average procedures see (Krause 2000). Theorem 4 as above
for multidimensional opinions is new.
D. Opinion dynamics with bounded confidence (BC)
For two agents
{
}
,1,,ik n∈ L and
s
t<
, s and t in T, a sequence
01
(,, , )
ts
ii i
−
L of agents is called
a confidence chain from
i to k for (s, t) if it holds that
0
,
ts
iii k
−
== and
(
)
(
)
1
, ( ) (0) given for 1, 2, , .
jj
iIi xtjx j ts
−
∈− =−L
Theorem 5 Consensus will be approached (in finite time) for a given initial profile, provided there
exist
1h≥ such that for all m the following property holds: For any two agents i and j there exists
a third one k such that a confidence chain exists from
i to k and from j to k for
(
)
(1),mhmh− .
The proof of Theorem 5 employs Theorem 4.
Theorem 6 For any given initial profile there exist
*
tT∈ , natural numbers
1
1
k
nnn<< < <L and
nonnegative numbers
j
c for 0 jk≤≤ such that for every j ()
ij
x
tc= for all
(
)
10 1
1,
jj k
nin n n n
++
≤< = = for all
*
tt≥ .
Theorem 6 can be derived by applying Theorem 5 to certain subgroups.
In a different manner, Theorem 6 was first proved by J.C. Dittmer (Dittmer 2000), (Dittmer 2001, p.
4618). There, instead of Theorem 5, the following result is used (Dittmer 2001, p. 4617):
If the opinion profile is an
ε -chain for every time point then a consensus will be reached in finite
time.
(This can be obtained also as a special consequence of Theorem 5.)
REFERENCES
ABELSON, R P (1964), “Mathematical models of the distribution of attitudes under controversy”.
In Frederiksen N and Gulliksen H (Eds.), Contributions to Mathematical Psychology, New York,
NY: Holt, Rinehart, and Winston.
32
BECKMANN T (1997) Starke und schwache Ergodizität in nichtlinearen Konsensmodellen. Di-
ploma thesis Universität Bremen.
BERGER R L (1981) A necessary and sufficient condition for reaching a consensus using De
Groot’s method.
J. Amer. Statist. Assoc. 76. pp. 415 – 419.
CHATTERJEE S (1975) Reaching a consensus: Some limit theorems.
Proc. Int. Statist. Inst. pp.
159 – 164.
CHATTERJEE S and Seneta E (1977) Toward consensus: some convergence theorems on repeated
averaging. J. Appl. Prob. 14. pp. 89 – 97.
COHEN J, Hajnal J , and Newman C M (1986) Approaching consensus can be delicate when posi-
tions harden. Stochastic Proc. and Appl. 22. pp. 315 – 322.
DEFFUANT G, Neau D, Amblard F, and Weisbuch G (2000) Mixing beliefs among interacting
agents. Advances in Complex Systems 3. pp. 87 – 98.
DE GROOT M H (1974) Reaching a consensus. J. Amer. Statist. Assoc. 69. pp. 118 – 121.
DITTMER J C (2000) Diskrete nichtlineare Modelle der Konsensbildung. Diploma thesis Univer-
sität Bremen.
DITTMER J C (2001) Consensus formation under bounded confidence. Nonlinear Analysis 47.
pp. 4615 – 4621.
FRIEDKIN N E and Johnsen E C (1990) Social influence and opinions. J. Math. Soc. 15. pp. 193
– 206.
FRIEDKIN N E and Johnsen E C (1999) Social influence networks and opinion change. Advances
in Group Processes 16. pp. 1 – 29.
FRENCH J R P (1956) A formal theory of social power. Psychological Review 63. pp. 181 –194.
FUJIMOTO T (1999) A simple model of consensus formation. Okayama Economic Review 31. pp.
95 – 100.
GANTMACHER F R (1959) Applications of the Theory of Matrices. Interscience, New York.
HARARY F (1959) “A criterion for unanimity in French’s theory of social power”. In Cartwright
D (Ed.), Studies in Social Power. Institute for Social Research, Ann Arbor.
HEGSELMANN R and Flache A (1998) Understanding complex social dynamics – a plea for cellu-
lar automata based modelling. Journal of Artificial Societies and Social Simulation, vol. 1 no. 3.
<http://www.soc.surrey.ac.uk/JASSS/1/3/1.html>
HEGSELMANN R, Flache A and Möller V, “Cellular automata models of solidarity and opinion
formation: sensitivity analysis”. In Suleiman R, Troitzsch K G, Gilbert N and Müller U (Eds.), So-
cial Science Microsimulation: Tools for Modeling, Parameter Optimization, and Sensitivity Analy-
sis, Heidelberg: Springer. pp. 151 – 178.
HOYLST J A, Kacperski K and Schweitzer F (2001) Social impact models of opinion dynamics.
Ann. Rev. Comp. Physics IX. pp.253 – 273.
KRAUSE U (1997), „Soziale Dynamiken mit vielen Interakteuren. Eine Problemskizze“. In Krause
U and Stöckler M (Eds.) Modellierung und Simulation von Dynamiken mit vielen interagierenden
Akteuren, Universität Bremen. pp. 37 – 51.
KRAUSE U (2000), “A discrete nonlinear and non—autonomous model of consensus formation”.
In Elaydi S, Ladas G, Popenda J and Rakowski J (Eds.), Communications in Difference Equations,
Amsterdam: Gordon and Breach Publ. pp. 227 – 236.
33
LATANÉ B and Nowak A (1997), “Self-organizing social systems, necessary and sufficient condi-
tions for the emergence of clustering, consolidation, and continuing diversity. In Barnet G and Bos-
ter F (Eds), Progress in Communication Science: Persuasion. Norwood, NJ: Ablex. pp. 43 – 74.
LEHRER K (1975) Social consensus and rational agnoiology. Synthese 31. pp. 141 – 160.
LEHRER K and Wagner C G (1981) Rational Consensus in Science and Society. Dordrecht: D.
Reidel Publ. Co.
STAUFFER D (2001) Monte Carlo simulations of Sznajd models. Journal of Artificial Societies
and Social Simulation, vol. 5, no. 1. <http://www.soc.surrey.ac.uk/JASSS/5/1/4.html>
STOCKER R, Green D G and Newth D (2001) Consensus and cohesion in simulated social net-
works. Journal of Artificial Societies and Social Simulation, vol. 4, no. 4.
<http://www.soc.surrey.ac.uk/JASSS/4/4/5.html>
WAGNER C G (1978) Consensus through respect: a model of rational group decision-making.
Philosophical Studies 34. pp. 335 – 349.
WEISBUCH G, Deffuant G, Amblard F and Nadal J P (2001), Interacting agents and continuous
opinion dynamics. <http://arXiv.org/pdf/cond-mat/0111494>
- CitationsCitations710
- ReferencesReferences26
- Given a matrix M , its vectorization vec M is a column vector obtained by stacking the columns of M , one on top of one another [40], e.g. vec ( 1 0 2 1) = [1, 2, 0, 1] . Lemma 1: [40] For any three matrices A, B, C such that the product ABC is defined, one has vec ABC = (C ⊗ A) vec B.
[Show abstract] [Hide abstract] ABSTRACT: Unlike many complex networks studied in the literature, social networks rarely exhibit regular unanimous behavior, or consensus of opinions. This requires a development of mathematical models that are sufficiently simple to be examined and capture, at the same time, the complex behavior of real social groups, where opinions and the actions related to them may form clusters of different sizes. One such model, proposed in [1], deals with scalar opinions and extends the idea in [2] of iterative pooling to take into account the actors' prejudices, caused by some exogenous factors and leading to disagreement in the final opinions. In this paper, we offer a novel multidimensional extension, which represents the dynamics of agents' opinions on several topics, and those topic-specific opinions are interdependent. As soon as opinions on several topics are affected simultaneously by the same influence networks, they automatically become related. However, we introduce an additional relation, interdependent topics, by which the opinions being formed on one topic are functions of the opinions held on other topics. We examine rigorous convergence properties of the proposed model and find explicitly the steady opinions of the agents. Although our model assumes synchronous communication among the agents, we show that the same final opinion may be reached "on average" via asynchronous gossip-based protocols.- Clearly, a cellular automaton can be regarded as a SPS in the sense that the rules governing the value change of cells can be straightforwardly converted into a production rule, while the rest, including the grid itself, can be captured by a syntactic structure. No matter the rules are applied synchronously or asynchronously, this can be captured in structured production systems with different rule application strategies discussed in Section 4. It can be seen that many other approaches, e.g., opinion dynamics[3,7]and, can be reformulated as structured production system as well.
[Show abstract] [Hide abstract] ABSTRACT: In this extended abstract, we propose Structured Production Systems (SPS), which extend traditional production systems with well-formed syntactic structures. Due to the richness of structures, structured production systems significantly enhance the expressive power as well as the flexibility of production systems, for instance, to handle uncertainty. We show that different rule application strategies can be reduced into the basic one by utilizing structures. Also, many fundamental approaches in computer science, including automata, grammar and logic, can be captured by structured production systems.- And finally updating rule is one of the important aspects to be considered in this algorithm. For updating the opinions, various strategies are adopted as stated in the literatures [10][11][12][13][14]. But, here Durkheimian opinion dynamics has been used for updating rule.
- Consensus of multi-agent systems has drawn considerable attention from various fields over the past two decades. For example, physicists investigate the synchronization phenomena of coupled oscillators, flashing fireflies, and chirping crickets [1][2][3]; biologists, physicists, and computer scientists try to understand and model the flocking phenomenon of animals behavior [4][5][6]; sociologists simulate the emergence and spread of the public opinions [7, 8]. Because of the importance, effort has been devoted to the mathematical analysis of consensus of flocks [9][10][11][12].
[Show abstract] [Hide abstract] ABSTRACT: Consensus conditions and convergence speeds are crucial for distributed consensus algorithms of networked systems. Based on a basic first-order average-consensus protocol with time-varying topologies and additive noises, this paper first investigates its critical consensus condition on network topology by stochastic approximation frameworks. A new joint-connectivity condition called extensible joint-connectivity that contains a parameter $\delta$ (termed the extensible exponent) is proposed. With this and a balanced topology condition, we show that a critical value of $\delta$ for consensus is $1/2$. Optimization on convergence rate of this protocol is further investigated. It is proved that the fastest convergence rate, which is the theoretic optimal rate among all controls, is of the order $1/t$ for the best topologies, and is of the order $1/t^{1-2\delta}$ for the worst topologies which are balanced and satisfy the extensible joint-connectivity condition. For practical implementation, certain open-loop control strategies are introduced to achieve consensus with a convergence rate of the same order as the fastest convergence rate. Furthermore, a consensus condition is derived for non stationary and strongly correlated random topologies. The algorithms and consensus conditions are applied to distributed consensus computation of mobile ad-hoc networks; and their related critical exponents are derived from relative velocities of mobile agents for guaranteeing consensus.- This section briefly introduces the basic knowledge regarding the HK model [21], and the consensus in opinion dynamics, which will provide a foundation for this study.
[Show abstract] [Hide abstract] ABSTRACT: Opinion dynamics provides a modeling tool for the public opinion management. The existing studies mainly focused on building the evolution model of opinions. However, the control of public opinions has been a key problem in practical opinion dynamics. The objective of this paper is to propose an opinion control rule to support the consensus reaching. Based on the bounded confidence model, the consensus model with the minimum adjustment is proposed. Next, based on the proposed consensus model, we propose the opinion control rule to support the consensus reaching. Furthermore, a numerical example is given to illustrate the feasibility of the proposed opinion control rule. Through simulation experiments, we investigate the effects of adjustment thresholds and bounded confidences on the opinion control rule.- Many applications, however, lead to multi-agent systems with time-varying interaction topologies. An example of such systems is given by the bounded confidence models [10], [11], [31], [59] of opinion dynamics, where each person interacts only to " like-minded " individuals, whose opinions are close to his/her own opinion. Mathematically this implies that the gain w ij in (2) depends on the deviation |x i − x j | and vanishes when this deviation becomes large.
[Show abstract] [Hide abstract] ABSTRACT: Many cooperative behaviors of multi-agent teams emerge from local interactions among the agents, where an agent interacts with a few “adjacent” teammates, but has no information about the remaining agents. For instance, the self-organization of many biological populations –including swarms of insects, flocks of birds, and schools of fish –are based on such local interaction rules: the motion and decisions of an individual agent are determined by the behavior of its nearest neighbors in the population. A special case of multi-agent coordination is consensus, that is, the agreement of agents on some quantity of interest or, more generally, the full or partial synchronization of their state trajectories. Establishing consensus is a “benchmark” problem in multi-agent systems study, which allows to reveal the main principles of multi-agent coordination and, in particular, the role of the system's interaction graph (or topology). Consensus lies in the heart of many natural phenomena (e.g., synchronous oscillation of neural cells, which maintains a stable heart rhythm) and engineering designs (e.g., attitude synchronization of satellites). In this article, we give a brief review of distributed consensus algorithms, focusing on the basic ideas and relevant mathematical theory, in particular, graph-theoretic methods.
People who read this publication also read
Data provided are for informational purposes only. Although carefully collected, accuracy cannot be guaranteed. Publisher conditions are provided by RoMEO. Differing provisions from the publisher's actual policy or licence agreement may be applicable.
This publication is from a journal that may support self archiving.
Learn more