Mathematics and physics applications
in sociodynamics simulation: the case
of opinion formation and diffusion
Giacomo Aletti1 , Ahmad K. Naimzada2 , and Giovanni Naldi3
1
2
3
Department of Mathematics and ADAMSS Center, Università degli studi di
Milano, via Saldini 50, 20133 Milano, Italy, giacomo.aletti@unimi.it
Department of Quantitative methods for Business Economics, Università degli
studi di Milano Bicocca, Piazza dell’Ateneo Nuovo 1, 20126 Milano, Italy,
ahmad.naimzada@unimib.it
Department of Mathematics and ADAMSS Center, Università degli studi di
Milano, via Saldini 50, 20133 Milano, Italy, giovanni.naldi@unimi.it
Summary. In this chapter, we briefly review some opinion dynamics models starting from the classical Schelling model and other agent-based modelling examples.
We consider both discrete and continuous models and we briefly describe different
approaches: discrete dynamical systems and agent-based models, partial differential
equations based models, kinetic framework. We also synthesized some comparisons
between different methods with the main references in order to further analysis and
remarks.
1 Opinion dynamics
Opinion dynamic models describe the process of opinion formation in groups
of individuals: as the opinion behavior emerges, evolves, spreads, erodes, or
disappears. Here we provide a brief overview of models and simulation tools for
the opinion dynamics. Most people hold and exchange opinions about a lot of
topics, from politics and sports to health, new products and the lives of others.
These opinions can be either the result of serious reflection or as is often the
case when information is hard to process or obtain, formed through interactions with others that hold views on given issues. The modelling of the opinion
dynamics try to understand when the opinion formation leads to consensus,
polarization or fragmentation within an interacting group. Opinion dynamics
is one of the most widespread topics of Sociophysics (application of methods
from physics to human relations) and Sociodynamics (the attempt to build
up a modelling strategy allowing in principle of an integrative quantitative
description of dynamic macro-phenomena in the society). Sociophysics covers numerous topics of social sciences and addresses many different problems
G. Naldi et al. (eds.), Mathematical Modeling of Collective Behavior
in Socio-Economic and Life Sciences, Modeling and Simulation in Science,
Engineering and Technology, DOI 10.1007/978-0-8176-4946-3 8,
c Springer Science+Business Media, LLC 2010
203
204
Giacomo Aletti, Ahmad K. Naimzada, and Giovanni Naldi
including social networks, language evolution, population dynamics, epidemic
spreading, terrorism, voting, and coalition formation. Among them the study
of opinion dynamics has become a main stream of research. In fact, the public opinion is nowadays a feature of central importance in modern societies
making the understanding of its underlining mechanisms a major challenge.
The dynamics of agreement/disagreement among individuals is complex, because the individuals are. Physicists working on opinion dynamics aim at
defining the opinion states of a population, and the elementary processes that
determine transitions between such states. The main question is whether this
is possible and whether this approach can shed new light on the process of
opinion formation.
Computer simulations play an important role in the study of social dynamics as they parallel more traditional approaches of theoretical physics and
mathematical modelling, where a system is described in terms of a set of
equations. One of the most successful methodologies used in social dynamics
is agent-based modelling [1, 2]. In agent-based modelling (ABM), a system is
modelled as a collection of autonomous decision-making entities called agents.
Each agent individually assesses its situation and makes decisions on the basis
of a set of rules. Agents may execute various behaviors appropriate for the
system they represent. Repetitive interactions between agents are a feature
of agent-based modelling, which relies on the power of computers to explore
dynamics out of the reach of pure mathematical methods. At the simplest
level, an agent-based model consists of a system of agents and the relationships between them. Even a simple agent-based model can exhibit complex
behavior patterns and provide valuable information about the dynamics of
the real-world system that it emulates. In addition, agents may be capable of evolving, allowing unanticipated behaviors to emerge. Sophisticated
ABM sometimes incorporates neural networks, evolutionary algorithms, or
other learning techniques to allow realistic learning and adaptation. ABM is
a mindset more than a technology: it consists of describing a system from the
perspective of its constituent units.
The description of emerging collective behaviors and self-organization in
multiagent interactions has gained increasing interest from various research
communities in biology, ecology, robotics, and control theory, as well as sociology and economics. In the biological context, the emergent behavior of bird
flocks, fish schools, or bacteria aggregations, among others, is a major research
topic in population and behavioral biology and ecology [3–11]. Likewise, the
coordination and cooperation among multiple mobile agents (robots or sensors) have been playing central roles in sensor networking, with broad applications in environmental control [12]. Emergent economic behaviors, such as
distribution of wealth in a modern society [13–17], or the formation of choices
and opinions [18–21], are also challenging problems studied in recent years in
which emergence of universal equilibria is shown. Also, the development of a
common language in primitive societies is yet another example of a coherent
collective behavior emerging within a complex system [22, 23].
Opinion dynamics and diffusion
205
The social world that we observe reflects a lot of interdependent processes,
with macro level structures of organizations, communities, and societies both
emerging from and constraining the micro level interactions of individuals.
Many social phenomena, such as the spread of epidemics, or the dissolution of
organizations, are inherently time varying and depend on interactions between
entities within a social system. Understanding the link between microlevel
interactions and macrolevel dynamics promises to have profound impact on
how human societies, organizations, and nations might be structured and
how related policy decisions should be made. As we mentioned earlier, an
increasing number of scientists are using mathematical and computational
models to elucidate theoretical problems in social dynamics, often by applying
general theories or methods that are well developed in the natural and physical
sciences with a view to gaining insight into the underlying generative processes
or the dynamic consequences of social relationships.
It may be surprising, but the application of concepts from the natural
sciences to social sciences is at least 25 centuries old. In fact, the Greek
philosopher Empedocle stated that humans are like liquids: some mix easily like wine and water, and others refuse to mix. The discovery of quantitative laws in the collective properties of a large number of people was one
of the pushing factors for the development of statistics. Many scientists and
philosophers called for some quantitative understanding on how such precise
regularities arise. Among many others, Hobbes, Laplace, Comte, Stuart Mill
shared this line of thought. More recently, Majorana [24, 25] in the 1940s
suggested to apply quantum-mechanical uncertainty to socio-economic questions. Weidlich [26] studied similar questions since 1971. In the same year,
Thomas Schelling (Nobel Prize for Economics in 2005) published [27] his
highly acclaimed model for for urban segregation in the first issue of Journal
of Mathematical Sociology. Moreover, in the same issue of the journal Sakoda
[28] presented a closely related work whose basic design was already present in
his unpublished dissertation of 1949. In [33] Galam gave a personal testimony
of Sociophysics going back to his 1982 publication. Other references may be
found in the books of Arnopoulos [34] and Schweitzer [35], while some review
articles can be found for example in [36–38].
1.1 Schelling model
We present here a significant and historical segregation model proposed by
Schelling in 1971 [27] for the study of ethnic segregation in the United States.
Schelling supposed that people had a threshold of tolerance of other ethnic
groups based on the neighbored people. If, for instance, the threshold of tolerance was 40%, people were content to stay in the place they live provided
that at least four in ten of their neighbors were from the same ethnic group.
If this were not so, they would try to move to another neighborhood in which
at least 40% were of their own group. The conventional assumption is that
ethnic segregation in the USA is at least partly due to the fact that whites
206
Giacomo Aletti, Ahmad K. Naimzada, and Giovanni Naldi
are prejudiced and have a tolerance threshold of over 50%. They therefore
moved out of urban neighborhoods that had a majority of blacks, leaving
the neighborhood with a still higher proportion of black people and, thus,
accelerating the tendency towards complete segregation. Schellings point was
that tolerance thresholds much lower than 50% could lead to the same result.
Even a threshold as low as 30% could result in almost complete segregation.
Thus, although people might be quite content with being in the minority in a
neighborhood, so long as they demanded that some small proportion of their
neighbors were of the same ethnic group as themselves, segregation could
emerge. The original Schelling’s model was very simple. Take a large chessboard, and place a certain number of black and white counters on the board,
leaving some free places. A counter prefers to be on a square where a certain
fixed percentage of the counters in his Moore neighborhood (his eight nearest
neighbors) are of its own color to the opposite situation. From the counters
who wish to migrate one is chosen at random and moves to a preferred location. This model, when simulated, yields complete segregation even though
people’s preferences for being with their own color are not strong. In Fig. 1 we
show some simulations by using different threshold of tolerance, each chessboard refers to the steady state when there are no more unhappy. Schelling’s
result has become famous precisely because the preferences of individuals for
Fig. 1. The role of preferences on the end pattern of segregation that emerges in
the Schelling’s model
Opinion dynamics and diffusion
207
segregation were not particularly strong. Note that the Schelling result is of
interest to economists because it illustrates the emergence of an aggregate
phenomenon that is not directly foreseen from the individual behavior.
The Schelling model is based on the idea that an individual agent makes
decisions based on his preferences or utility function. Then the agent’s satisfaction is equivalent to the energy stored in him. An increase in happiness
is a decrease in internal energy. An agent, therefore, wants to minimize his
energy, which is generated either by taking some action or through the interaction with his environment. The Schelling model assumes that the agent’s
utility depends on his local environment and that he moves if the utility falls
below a certain threshold. Such a process is easily simulated through the twodimensional Ising model [29–31]. In the simplest case, as black and white, we
have two Ising spin orientations. In this Ising model, each site i on a square
lattice carries a variable Si with value +1 or −1. For each pair (i, k) of nearest
neighbors produces an energy contribution Eik equal to Eik = −JSi Sk with
some proportionality constant J. The total energy E is proportional to the
total unhappiness is the sum of this pair energy over all neighbor pairs of the
lattice Eik . In statistical physics, different distributions of the spins Si are realized with a probability proportional to e(−E/KB T ) , where T is the absolute
temperature and KB the Boltzmann constant. We point out that the relevant
quantity is the ratio (KB T /J). The Metropolis kinetics of the system is simulated by flipping a spin if and only if a random number between 0 and 1 is
smaller than the probability e(−∆E/(KB T )) , where ∆E is the variation of the
total energy. When the Glauber kinetics [32] is used the flipping probability
is e(−∆E/(KB T )) /(1 + e(−∆E/(KB T )) ). For Glauber or Metropolis, after very
long times (measured by the number of sweeps through the lattice) one of the
two possibilities dominates at the end which depend on the temperature T . If
T is not larger than the critical temperature Tc , for Kawasaki dynamics the
fraction of black sites remains constant, and we get two large domains. For
higher temperatures above Tc only small clusters and no large domains are
formed. In this Ising model, two neighboring spins have due to their interaction JSi Sk a higher probability to belong to the same group than to belong
to the two different groups. Schelling’s model avoided probabilistic rules and
thus counted neighbors Si = ±1 at zero temperature. As Schelling moved only
one person at a time, and made no exchange of two people simultaneously as
in Kawasaki kinetics, he introduced a large fraction of empty residences. Thus
at each step, one unhappy person or family moves into the closest vacancy
where life would be happy. For a study of the model from this physical point
of view and a link between Schellings socio-economic model of segregation and
the physics of clustering we refer to [39] and [40].
There are several variants on Schelling’s original model modifying the form
of the utility function used by Schelling, the size of neighborhoods, the rules
for moving, and the amount of unoccupied space. Many of these fail to give
large domains: only small clusters are seen. In many real cities, there are huge
cluster which extends over many square kilometers. Thus, Schelling’s model
208
Giacomo Aletti, Ahmad K. Naimzada, and Giovanni Naldi
does not give always the desired results. In general, the critique social systems
models are too simplified is most of the times very well grounded. However, in
most situations qualitative (and even some quantitative) properties of large
scale phenomena can be simulated and reproduced. A discussion about the
application of self-organising segregation and migration models can be found
in [41–43].
2 Binary opinion dynamics and beyond
Scientists working on opinion dynamics aim at defining the opinion states of a
population, and the processes that determine transitions between such states.
In a mathematical model, opinion has to be a variable, or a set of variables.
This may appear too reductive, thinking about the complexity of a person and
of each individual position. Rightly, Humans do not like to be treated like a
number, and indeed the human brain is much more complex than a set of
variables. But one can observe that the value of an opinion or a decision could
be represented by a numerical vector. As example binary opinions and binary
choices are frequent in the everyday life: Windows/Linux, buying/selling, Coca
Cola/Pespi Cola, etc. Moreover, appear to act on the opinion some kind of
social forces even though they may appear of little relevance to any important
decisions. For example we can consider the following simple experiment, first
performed four decades ago and reproduced many times since. Stand on a
busy street corner and look up at the sky. The crowd will part around you,
indifferent to whatever it is you may be looking at. Now enlist the help of a
friend to stand beside you and also look skyward. Soon, others will stop and
gaze up as well. A similar thing would happen if you and your friend boarded
an empty elevator and faced the rear wall. As more passengers boarded many
would face the back wall too. Then a kind of peer pressure seems to appear.
2.1 The Sznajd model
There are several models used in modelling consensus formation and we tried
to name a few historical introduction listed earlier. Here we consider in more
detail a recent Sznajd model only as an example just to illustrate some issues
and problems in the simulation of the opinion formation. The general framework of the model is as follows: at any given time a distribution of opinions
exists in a given population; at each time step some group of individuals interact and as a result, one or several opinions are shifted (generally towards
some consensus among the group); as a result of these interactions, some form
of pattern of opinions is observed. An important aspect always present in social dynamics is topology which is the structure of the interaction network
describing who is interacting with whom, how frequently and with which intensity. In the Sznajd model [20, 46], as in the traditional statistical physics
and in the Ising model, the geometry is a one-dimensional regular lattice.
Opinion dynamics and diffusion
300
209
1
time
m(t)
200
0.5
100
0
0
0
50
X
0
100
100
200
300
time
500
1
300
m(t)
time
400
0.5
200
100
0
0
0
50
X
100
0
200
time
400
Fig. 2. Two examples of the dynamics of the Sznajd model, in the first row we have
the behavior for corresponding to the original rules of the model while in second
row the behavior for the modified rules in order to avoid the antiferromagnetic final
state. In the left column we reported the evolution of the social state in time and
space, in the right column we considered the evolution of the magnetization variable
m(t) with respect to the time
Agents are, thus, supposed to sit on vertices (nodes) of the lattice, and each
agent can interact with the two adjacent agents. More generally, it is possible to consider a d-dimensional lattice of N d sites carries opinion variables
Si , i = 1, 2, . . . , N d . The lattice may be square (four nearest neighbours), triangular (six nearest neighbours), many other choices are also possible. The
opinion variables could be scalar or vector real variables. Usually, they are
scalar variables and are integers between P1 and P2 or take values in a finite
set of opinions {P1 , P2 , . . . , Pk }. The individual opinions Si in the Sznajd
model are represented in our model by Ising spins (“yes” or “no,” −1 or 1),
that is to say the opinions are binary opinions. Such an approach is not new
but it had been used earlier (see for example [33] or [47]). At each time step,
each Si is calculated from a rule according the spin dynamics [20]:
• A pair of opinions (spins) Si , Si+1 is chosen to change their nearest neighbors, i.e., the opinions (spins) Si−1 and Si+1 .
210
Giacomo Aletti, Ahmad K. Naimzada, and Giovanni Naldi
• If Si Si+1 = 1 (the opinions are equal) then Si−1 = Si and Si + 2 = Si
(social validation).
• If Si Si+1 = −1 then Si−1 = Si+1 and Si+2 = Si .
The model was originally called USDF after the trade union maxim “United
we Stand, Divided we Fall,” in other words: If you see two or more people
sharing an opinion on certain issue, you are tempted to join. The crucial difference of the model compared to other Ising-type models is that information
flows outward. The dynamic rules of the model lead to two possible steady
states: the complete consensus (ferromagnetic state) in which all Si = 1 or
all Si = −1; the stalemate (antiferromagnetic state) in which 50% of opinions
Si = 1 and 50% are equal to −1. However, the last 50 − 50 state is realized in
a very special way: every member of the community disagrees with his nearest
neighbor. Even if the Ising model with only next nearest neighbor interaction
has such kind of fixed point (ferro and antiferromagnetic state) it is unrealistic
as a social state. Then, new dynamic rules were proposed [46], for example in
the case of the disagreements, Si Si+1 = −1, then a new update is computed:
Si−1 = Si and Si+2 = Si+1 . Also a reasonable rule is to assume that the
individual keeps its opinion in such a case, rather than opposing its neighbor.
That is if Si = −Si+1 , nothing happens at that time step, the system is not
altered. A global index which corresponds to the general opinion of the society
is the magnetization m(t) at the time t defined as:
1
(number of opinions 1) − (number of opinions − 1)
=
Si ,
m(t) =
N
N i
where N is the number of people in the community. In the Fig. 2 we show the
behavior of the opinions in a community and the evolution of the magnetization in the case of the original Sznajd model and for the modified model with
the different active update rule. In [20] Sznajd-Weron and Sznajd have investigated some statistical properties of the magnetization function m(t) and
some interesting behavior of the opinion changes of one particular individual.
In particular they observed that if an individual changes the opinion at time t,
he (she) will probably change it also at time t + 1. Moreover, if τ denotes the
time needed by an individual to change the opinion its distribution P (τ ) follows a power law with an exponent −3/2. It is well known that the changes
of opinion in a community are not only determined by the individual contacts
between neighbors but also the social impact has a role. It is possible to introduce a noise p, some kind of social temperature, which is the probability
that an individual, instead of following the dynamic rules, will make a random
decision. In [20] a computational study is performed to study the influence
on the distribution P (τ ) of the parameter p. Sznajd model has been modified
and applied in marketing, finance, and politics (see for example [38]).
In agreement with social theory assigning four nearest neighbors to an
individual rather than only two is a more realistic simplification of a real society. So in search for a more realistic model, two dimensional version of Sznajd
Opinion dynamics and diffusion
211
model was proposed by Stauffer et al. [48] for two-dimensional regular lattice.
In accordance with the dimensionality of the model, six different versions were
proposed. Here, we focus on two of the most realistic of the six:
• A 2 × 2 square of four neighbors are chosen at random. If all four opinions
of the square are the same then all eight neighbors to this square follow
the same value. Otherwise, the system is not altered, the neighbors are
unchanged.
• A bond is chosen at random. If the two opinions forming the bond are the
same, then all six neighbors to this bond follow the same value. Otherwise,
the system is not altered, the neighbors are unchanged.
In Fig. 3 we show one example of the behavior of the two-dimensional Sznajd
model following the two rules described earlier, simulations were performed
on a N × N square lattice, where individuals were placed on the lattice as
spins in the two dimensional Ising model. Periodic boundary conditions are
used in both directions. As for the original Sznajd model simulation, random sequential updating was used as in Monte-Carlo simulation. In [48] the
emergence of a phase transition phenomena with respect to the size of the
system was studied: this is the only important difference between the onedimensional model and the two-dimensional model. A natural extension of the
binary Sznajd model developed in order to fit the experimental observation
in some elections is to allow more than two opinions for each individual. So, if
there are M candidates running for the post then each individual can have M
possible opinions. This is an example of a many-opinion modification to the
Sznajd model, and also equivalent to extending the Ising model to M -state
Potts model [49]. Other steps toward more realism is to investigate the model
on a complex network with different topology, adding noise or diffusion (see
for examples [46]).
In a short note [50] Ochrombel suggested a drastic simplification of the
Sznajd model. In the Ochrombel version it is not necessary to have a cluster
of identical opinions to change the neighbors. Any individual is capable to
convince her neighbors to select the same opinion. We choose an agent i at
random. Then, choose j randomly among neighbors and set Sj = Si . In the
case of a fully connected network the Ochrombel simplification is equivalent to
voter model, whose dynamical properties were studied, e.g., in [51]. In this last
case for binary opinions the state is described by the magnetisation variable
m(t) only. It can be found in the thermodynamic limit N → ∞ that the
probability density Pm (m, t) of the magnetization at time t evolves according
to the partial differential equation [52]
∂
∂2
(1 − m2 )Pm (m, t) .
Pm (m, t) =
∂t
∂m2
(1)
For the original binary Sznajd model but in the case of a fullyconnected network a different time scaling is needed in order to get sensible thermodynamic
212
Giacomo Aletti, Ahmad K. Naimzada, and Giovanni Naldi
0
0
50
50
100
0
50
t=1
100
100
0
0
0
50
50
100
0
50
t=10
100
100
0
50
t=5
100
50
t=50
100
0.52
0.515
0.51
m(t)
0.505
0.5
0.495
0.49
0.485
0.48
0
10
20
30
40
50
t
Fig. 3. The time evolution for a two-dimensional Sznajd model as proposed by
Stauffer et al. [48] and the behavior of the corresponding magnetization function
m(t)
limit [52]. Then the equation for the probability density is the following
Fokker–Planck equation
∂
∂
Pm (m, t) = −
(1 − m2 )mPm (m, t) .
∂t
∂m
(2)
The dynamics is those of a pure deterministic drift and no diffusion ever
smears out the evolving probability packet. The possibility to have the Fokker–
Planck equation provides a great tool for the theoretical study of the opinion
dynamics model.
Opinion dynamics and diffusion
213
One issue of interest in the opinion dynamics concerns the importance of
the binary assumption: what would happen if opinion were a continuous variable such as the worthiness of a choice or some belief about the adjustment
of a control parameter? Continuous opinions invalidate some of the concepts
adopted in models with discrete choices so they require a different framework.
The initial state is usually a population of N agents with randomly assigned
opinions, represented by real numbers within some interval. The opinion clusters could be one (consensus), two (polarization), or more (fragmentation). In
principle, each agent can interact with every other agent. In practice, there
is a real discussion only if the opinions of the people involved are sufficiently
close to each other. This realistic aspect of human communications is called
bounded confidence. Usually, it is expressed by introducing a real number ε,
the uncertainty or tolerance, such that an agent, with opinion x, only interacts
with those of its peers whose opinion lies in the interval ]x − ε, x + ε[. As example consider the model introduced by Deffuant et al. [53]. There are, again,
N individuals with opinions Si . In each update step, two of them, say, i and
j, are chosen randomly. Then, we check to see whether their opinions differ
less than (or equally to) the confidence bound ε . In the positive case, their
opinions are slightly shifted towards each other,
Si (t + 1) = (1 − µ)Si (t) + µSj (t), Sj (t + 1) = (1 − µ)Sj (t) + µSi (t),
for |Si (t) − Sj (t)| ≤ ε, where ε is a parameter. For very large numbers of
individuals, N → ∞, the dynamics can be expressed in terms of the continuous distribution of opinions P (s, t) which satisfies an equation with a fairly
complex behavior [54].
3 Agents based model and discrete dynamical systems
In [45], the authors consider a micro-scale approach. The population is composed by n agents. At each discrete time t = 0, 1, . . . the opinion of the ith
agent is given by the quantity xi (t) and therefore the opinion profile at time t is
given by the vector x(t) = (x1 (t), . . . , xn (t)). The ith agent takes into account
the opinion of the jth agent by means of the weight aij (t, x(t)). Accordingly,
the opinion formation evolves in the following way:
xi (t + 1) = ai1 (t, x(t))x1 (t) + ai2 (t, x(t))x2 (t) + · · · + ain (t, x(t))xn (t)
or, in matrix form, the general model is given by:
x(t + 1) = A(t, x(t))x(t).
The only assumption on the matrix A is that it is a stochastic matrix (each
row sums to 1). Obviously, further assumptions on the weight matrix A lead
to different studied models. The review article [45] presents models which
can be analytically studied and others that are computationally faced. The
consensum property is given when all the opinion converge to the same opinion
214
Giacomo Aletti, Ahmad K. Naimzada, and Giovanni Naldi
that depends only on the initial opinion profile: given x(0), there exists c =
c(x(0)) such that xi (t) → c for any 1 ≤ i ≤ n. When x(t) converges to a
vector with different components, we face an opinion fragmentation. The first
part of [45] is devoted to a review of the linear models, while the second part
refers to some computational results on some nonlinear models.
3.1 Linear models
If A(t, x(t)) ≡ A, (3) reduces to the classical model of fixed weights
x(t + 1) = Ax(t).
A generalization of the previous model is given by Friedkin and Johnsen in [44]
x(t + 1) = Gx(0) + (I − G)Ax(t),
where G is a diagonal matrix containing the degree of adhesion to the initial
opinion: G = diag(g1 , g2 , . . . , gn ). Obviously, when G ≡ 0, (3.1) reduces to
(3.1). Note that in (3.1), x(t) = V (t)x(0), where
V (t) = M t +
t−1
i=0
Mi G
with M = (I − G)A.
Analytical conditions on the matrix A ensure limit properties as consensum or
fragmentation. As an example, if A is irreducible, limt x(t) = (I − M )−1 Gx(0)
when G = 0. Similar results, even if less sharp, may be found with the last
model presented in the review part of [45]. This model is still linear but timevariant (or, in the theory of Markov chains, inhomogeneous), that is
x(t + 1) = A(t)x(t),
where the entries of matrix A(t), i.e., the weights, are dependent on time
only. The time variant model (3.1) 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.
3.2 Nonlinear models
The models proposed in [45] depend on opinions itself and hence the models
are nonlinear. Analytical insights are not so easy to obtain. Therefore, computer simulations were used to investigate these models. An agent i takes only
those agents j into account whose opinions differ from his own not more than
a certain confidence level. More precisely, if
xi (t + 1) = |I(i, x(t))|−1
xk (t), where
k
I(i, x(t)) = #{1 ≤ k ≤ n : ǫl ≤ xk − xi ≤ ǫr },
Opinion dynamics and diffusion
215
the confidence level for each agent is the interval [ǫl , ǫr ]. We have the nonlinear
model
x(t + 1) = A(x(t))x(t),
where
aij (x) =
1
.
#{1 ≤ k ≤ n : ǫl ≤ xk − xi ≤ ǫr }
Asymptotic configurations are numerically studied both in the symmetric case
(i.e., ǫl = ǫr ), and in the asymmetric one.
4 The kinetic approach
The goal of the forthcoming kinetic model of opinion formation, is to describe
the evolution of the distribution of opinions in a society by means of microscopic interactions among individuals which exchange information. Opinion
is represented as a continuous variable w ∈ I, with, as example, I = [−1, 1].
The extremes w = ±1 represent extreme opinions. Toscani has recently studied [55], by means of kinetic collision-like models, the distribution of opinion
among individuals in a simple, homogeneous society. This model is based on
binary interactions. When two individuals with preinteraction opinion v and
w meet, then their posttrade opinions (postcollisional) v ∗ and w∗ are given by:
v ∗ = v − γP (|v|)(v − w) + ηv D(|v|)
w∗ = w − γP (|w|)(w − v) + ηw D(|w|)
where the coefficient γ ∈ (0, 1/2) is a given constant, while ηv and ηw are
random variables with the same distribution with variance σ 2 and zero mean.
The constant γ measures the compromise propensity while the variance σ 2
is related to the modification of opinion due to diffusion. The functions 0 ≤
P (·) ≤ 1 and 0 ≤ D(·) ≤ 1 describe the local relevance of the compromise
and diffusion for a given opinion. In absence of the diffusion contribution,
η = 0, if P (·) is not constant the total momentum is not conserved and it can
increase or decrease depending on the opinions before the interaction. While
if P is assumed constant, the interaction correspond to a granular gas like
interaction [56].
Let f (w, t) denote the distribution of opinion w ∈ I at time t ≥ 0. Standard
methods [57] of kinetic theory allow to describe the time evolution of f as a
balance between bilinear gain and loss of opinion terms
1
∂f
=
(3)
β ′ f (v ′ )f (w′ ) − βf (v)f (w) dwdηv dηw ,
∂t
J
B2 I
where (v ′ , w′ ) are the preinteraction opinions that generate the couple (v, w)
of opinions after the interaction, J is the Jacobian of the transformation of
(v, w) into (v ′ , w′ ), while the kernels β ′ and β are related to the details of the
binary interaction, finally B is the space of the random variables η. Following
216
Giacomo Aletti, Ahmad K. Naimzada, and Giovanni Naldi
Toscani [55] different Fokker–Planck equations can be obtained by considering
different limits and simplifications. In the following section we consider only
one example. For the case of a kinetic model for the the evolution of the
opinion in a closed group with respect to a choice between multiple options,
we refer to [58].
In some sense intermediate between the Kinetic approach and agent-based
model we can also consider an alternative approach based on active Brownian particles which interact via a communication field (see [59] and references
therein). This scalar field considers the spatial distribution of the individual
opinions; further, it has a certain lifetime, reflecting a collective memory effect,
and it can spread out in the community, modelling the transfer of information.
Also the active particle approach [60] leads to the derivation of evolution equations for a one-particle distribution function over the microscopic state. Two
types of equations, to be regarded as a general mathematical framework for
deriving the models, are derived corresponding to short- and long-range interactions. The active particle approach gives a general mathematical framework
in order to link microscopic and macroscopic descriptions of the dynamics.
4.1 An example of nonlinear continuous models
of opinion formation
In [61], the large-time behavior of solutions of the equation
∂
∂f
=γ
(1 − x2 )(x − m(t))f
∂t
∂x
(4)
is studied, where the unknown f (x, t) is a time-dependent probability density
which may represent the density of opinion in a community of agents. This
opinion varies between the two extremal opinions represented by ±1, so that
x ∈ [−1, 1]. The constant γ is linked to the spreading (γ = −1) or to the
concentration (γ = +1) of opinions. In (4) m(t) represents the mean value
of f (·, t),
x f (x, t) dx,
m(t) =
(5)
[−1,1]
and its presence introduces a nonlinear effect into its evolution.
Equation (4) describes the evolution of a probability density which represents the density of opinions in a community. For all values of the constant γ,
the time-evolution driven by this equations leads the density toward a equilibrium state that is described in terms of two Dirac masses (γ = −1) or to
a unique Dirac mass (γ = 1), see [61]. The convergence result toward equilibrium holds in weak-measure sense. A suitable way of treating convergence
results of (4) is based on a rewriting of this equation in terms of pseudo-inverse
functions. In fact, let F (x) denote the probability distribution induced by the
density f (x),
f (y) dy.
(6)
F (x) =
(−∞,x]
Opinion dynamics and diffusion
217
As F (·) is not decreasing, we can define its pseudo inverse function (also called
quantile function) by setting, for ρ ∈ (0, 1),
X F (ρ) = inf{x : F (x) ≥ ρ}.
Equation (4) for f (x, t) takes a simple form if written in terms of its pseudo
inverse X(ρ, t). [61, Theorem 3.1] shows in fact that the evolution equation
for X(ρ, t) reads
∂X(ρ, t)
= −γ (X(ρ, t) − m(t)) (1 − X 2 (ρ, t)),
∂t
(7)
1
where now ρ ∈ (0, 1) and m(t) = 0 X(ρ, t) dρ. Existence and uniqueness of
solutions is studied. Then results on their large-time behavior are derived.
First, it is noted that the initial masses in +1 (called p+1 ), and in −1 (called
p−1 ) remain unchanged in time. For what concerns concentration (γ = 1), the
steady state is characterized by the distribution p−1 δ−1 + (1 − p1 − p−1 )δa +
p1 δ+1 . Obviously, the mean value m∞ = p1 −p−1 +a(1−p1 −p−1 ) characterize
it. In [61], it is proved that if (1 − p1 )(1 − p−1 ) < 1 (i.e., if there are masses in
±1 at time t = 0) then, m∞ = p1 − p−1 . Otherwise, if log (1 + X(ρ, 0))/(1 −
X(ρ, 0)) ∈ L1 (0, 1) then
m∞ =
exp {T (0)} − 1
,
exp {T (0)} + 1
(8)
1
where T (0) = 0 log 1+X(ρ,0)
1−X(ρ,0) dρ.
The spreading case is difficult to handle analytically due to the nonlinearity
present through the term m(t). Therefore, in [61], numerical methods are
discussed to capture the behavior for large t. In Fig. 4 we show the evolution
of numerical and quantile solution for a benchmark case as reported in [61].
1
0.8
f(x;0)
concentration T = 2
spreading T = 2
0.6
0.4
f(x;·)
X(½;·)
0.2
exact
initial
numerical
0
−0.2
−0.4
−0.6
−0.8
−1
−0.5
0
x
0.5
1
−1
0
0.5
½
1
Fig. 4. A benchmark case from [61]: evolution of density function (left hand side
figure) and comparison between analytical and numerical solution for the quantile
function (right hand side figure)
218
Giacomo Aletti, Ahmad K. Naimzada, and Giovanni Naldi
5 New opportunities: an opinion
Different computational techniques are useful for developing and adapting
social dynamics models. Differential Equations Systems of both ordinary and
partial differential equations are useful tools for developing models at the
aggregate level, but are more limited when considering the interdependent
behavior of individuals in a population. Agent-based methods are particularly
useful for studying the microlevel behavior of individual objects in a system.
The interactions between agents may be stochastic according to defined interaction and choice probabilities, and may be subject to exogenous shocks.
Given the limited applicability of systems dynamics models using differential
equations to populations of heterogeneous and interdependent agents, agentbased models may be used in conjunction with such systemlevel models or
by coupling agents at different resolutions to serve as a basis for developing multilevel models. Markov chains and similar probabilistic models have a
wide spectrum of applicability in social and economic process even if the independence assumptions that they make limit their modeling power. General
theoretical frameworks provide the conceptual underpinnings for a variety of
modeling techniques. Examples include Interacting Particle Systems and game
theory. The fundamental complexity of the dynamics of social phenomena and
the complications associated with representing agents who engage in rational,
responsive, and realistic interactions present challenges. We point out only
few opportunities and the list is certainly not exhaustive.
•
•
•
•
In order to study the effectiveness of possible interventions in social and
economical problems, agent based methods must be simulated numerous
times to compare alternative intervention scenarios, which is a limitation
of method. Results of agent based methods are not optimal solutions, but
rather scenarios with various assumptions. Thus, one possible improvement
may affect the possibility of developing control-theoretic approaches for
agent-based methods, which could be applied in studying interventions.
More realistic topologies must be taken into account in the opinion formation models. Some models has already been developed for suitable networks
as scale-free networks but a more general stability analysis with respect to
topology changes is needed.
Kinetic approach for a nonhomogeneous society must be considered (for
example for hierarchical organization).
More effective validation methods for models in social sciences including
methodological goals, criteria for evaluating fit, and model interpretation
and use of error.
References
1. Epstein, J.M., Axtell, R.L.: Growing Artificial Societies: Social Science from the
Bottom Up. MIT Press, Cambridge, MA (1996)
2. Axelrod, R.: The Complexity of Cooperation: Agent-Based Models of Competition and Collaboration. Princeton University Press, Princeton, NJ (1997)
Opinion dynamics and diffusion
219
3. Couzin, I.D., Krause, J., Franks, N.R., Levin, S.: Effective leadership and decision making in animal groups on the move. Nature, 433, 513–516 (2005)
4. Cucker, F., Mordecki, E.: Flocking in noisy environments. J. Math. Pures Appl.,
89, 278–296 (2008)
5. Cucker, F., Smale, S.: Emergent behavior in flocks. IEEE Trans. Automat.
Control, 52, 852–862 (2007)
6. Flierl, G., Grnbaum, D., Levin, D., Olson, D.: From individuals to aggregations:
the interplay between behavior and physics. J. Theor. Biol., 196, 397–454 (1999)
7. Liu, Y., Passino, K.: Stable social foraging swarms in a noisy environment. IEEE
Trans. Automatic Contrl., 49, 30–44 (2004)
8. Topaz, C.M., Bertozzi, A.L.: Swarming patterns in a two-dimensional kinematic
model for biological groups. SIAM J. Appl. Math., 65, 152–174 (2004)
9. Topaz, C.M., Bertozzi, A.L., Lewis, M.A.: A nonlocal continuum model for
biological aggregation. Bull. Math. Biol., 68, 1601–1623 (2006)
10. Topaz, C.M., Bernoff, A.J., Logan, S., Toolson, W.: A model for rolling swarms
of locusts. Eur. Phys. J. Special Topics, 157, 93–109 (2008)
11. Shen, J.: Cucker-smale flocking under hierarchical leadership. SIAM J. Appl.
Math., 68, 694–719 (2008)
12. Jadbabaie, A., Lin, J., Morse, A.S.: Coordination of groups of mobile autonomous agents using nearest neighbor rules. IEEE Trans. Autom. Contrl.,
48, 988–1001 (2003)
13. Biham, O., Malcai, O., Richmond, P., Solomon S.: Theoretical analysis and
simulations of the generalized Lotka-Volterra model. Phys. Rev. E, 66, 031102
(2002)
14. Solomon, S., Richmond, P.: Stable power laws in variable economies; LotkaVolterra implies Pareto-Zipf. Eur. Phys. J. B, 27, 257–262 (2002)
15. Dragulescu, A., Yakovenko, V.M.: Statistical mechanics of money. Eur. Phys.
J. B, 17, 723–729 (2000)
16. Chakraborti, A., Chakrabarti, B.K.: Statistical Mechanics of money: effects of
saving propensity. Eur. Phys. J. B, 17, 167–170 (2000)
17. Chakraborti, A.: Distributions of money in models of market economy. Int. J.
Modern Phys. C, 13, 1315–1321 (2002)
18. Galam, S., Gefen, Y., Shapir, Y.: Sociophysics:a new approach of sociological
collective behavior. J. Math. Sociol., 9, 1–13 (1982)
19. Stauffer, D., de Oliveira, P.M.C.: Persistence of opinion in the Sznajd consensus
model: computer simulation. Eur. Phys. J. B, 30, 587–592 (2002)
20. Sznajd-Weron, K., Sznajd, J.: Opinion evolution in closed community. Int. J.
Mod. Phys. C, 11, 1157–1165 (2000)
21. Weisbuch, G., Deffuant, G., Amblard, F., Nadal, J.P.: Meet, discuss, and segregate. Complexity, 7, 55–63 (2002)
22. Cucker, F., Smale, S.: On the mathematics of emergence. Jpn. J. Math., 2,
197–227 (2007)
23. Cucker, F., Smale, S., Zhou, D.X.: Modeling language evolution. Found. Comput. Math., 4, 315–343 (2004)
24. Majorana, E., 1942, Scientia Feb–Mar, 58
25. Mantegna, R.N.: Presentation of the English translation of Ettore Majoranas
paper: “The value of statistical laws in physics and social science”. Quantitative
Finance, 5, 133 (2005)
26. Weidlich, W.: Sociodynamics: A Systematic Approach to Mathematical Modelling in the Social Sciences. Harwood Academic, Amsterdam (2000)
220
Giacomo Aletti, Ahmad K. Naimzada, and Giovanni Naldi
27. Schelling, T. C.: Dynamic models of segregation. J. Math. Sociol., 1, 143–186
(1971)
28. Sakoda, J.M.: The checkerboard model of social interaction. J. Math. Sociol.,
1, 119–132 (1971)
29. Ising, E.: Beitrag zur Theorie des Ferromagnetismus. Zeitschr f. Physik, 31,
253–258 (1925)
30. Brush, S.G. , History of the Lenz-Ising Model. Rev. Modern Phys., 39, 883-893
(1967)
31. McCoy, B.M., Wu, T.T.: The Two-Dimensional Ising Model. Harvard University
Press, Cambridge Massachusetts (1973)
32. Godreche, C., Luck, J.M.: Metastability in zero-temperature dynamics: Statistics of attractors. J. Phys. Cond. Matt., 17, S2573–S2590 (2005)
33. Galam, S.: Sociophysics: a personal testimony. Physica A, 336, 49–55 (2004)
34. Arnopoulos, P.: Sociophysics: Cosmos and Chaos in Nature and Culture. Nova
Science Publishers, New York (2005)
35. Schweitzer, F.: Browning Agents and Active Particles. On the emergence of
complex behavior in the natural and social sciences. Springer, Berlin (2003)
36. Schweitzer, F. (ed): Modeling Complexity in Economic and Social Systems,
World Scientific. River Edge, NJ (2002)
37. Chakrabarti, B.K., Chakraborti, A., Chatterjee, A. (eds): Econophysics and
Sociophysics. Wiley, Berlin (2006)
38. Meyers, R.A. (ed): Encyclopedia of Complexity and Systems Science. Springer,
New York (2009)
39. Vinkovic, D., Kirman, A.: A physical analogue of the Schelling model. PNAS,
103 (51), 19261–19265 (2006)
40. DallAsta, L., Castellano, C., Marsili, M.: Statistical physics of the Schelling
model of segregation. eprint arXiv:0707.1681 (2007)
41. Fossett, M.: Ethnic preferences, social distance dynamics, and residential segregation: theoretical explorations using simulation analysis. J. Math. Sociol., 30,
185–274 (2006)
42. Clarkab, W.A.V., Fossett, M.: Understanding the social context of the Schelling
segregation model. PNAS, 5(11), 4109–4114 (2008)
43. Stauffer, D., Solomon, S.: Ising, Schelling and self-organising segregation. Eur.
Phys. J. B, 57, 473–479 (2007)
44. Friedkin, N.E., Johnsen, E.C.: Social Influence Networks and Opinion Change.
Adv. Group Proces., 16, 1–29 (1999)
45. Hegselmann, R., Krause, U.: Opinion dynamics and bounded confidence: models, analysis and simulation. J. Artif. Soc Social Simul., 5 (2002)
46. Sznajd-Weron, K.: Sznajd model and its applications. Acta Phys. Polonica B,
36, 1001–1011 (2005)
47. Castellano, C., Fortunato, S., Loreto, V.: Statistical physics of social dynamics.
Rev. Modern Phys., 81, 591–646 (2009)
48. Stauffer, D., Sousa, A.O., De Oliviera, M.: Generalization to square lattice of
Sznajd sociophysics model. Int. J. Mod. Phys. C, 11, 1239–1245 (2000)
49. Wu, F.Y.: The Potts model. Rev. Mod. Phys., 54, 235–268 (1982)
50. Ochrombel, R.: Simulation of Sznajd sociophysics model with convincing single
opinions. Int. J. Mod. Phys. C, 12, 1091 (2001)
51. Ben-Naim, E., Frachebourg, L., Krapivsky, P.L.: Coarsening and persistence in
the voter model. Phys. Rev. E, 53, 3078 (1996)
Opinion dynamics and diffusion
221
52. Slanina, F., Lavicka, H.: Analytical results for the Sznajd model of opinion
formation. Eur. Phys. J. B, 35, 279–288 (2003)
53. Deffuant, G., Neau, D., Amblard, F., Weisbuch, G.: Mixing beliefs among
interacting agents. Adv. Compl. Syst., 3, 87–98 (2000)
54. Ben-Naim, E., Krapivsky, P.L., Redner, S.: Bifurcations and patterns in compromise processes. Physica D, 183, 190 (2003)
55. Toscani, G.: Kinetic models of opinion formation. Comm. Math. Sci., 4, 481–496
(2006)
56. McNamara, S., Young, W.R.: Kinetics of a onedimensional granular medium in
the quasielastic limit. Phys. Fluids A, 5, 34 (1993)
57. Cercignani, C., Illner, R., Pulvirenti, M.: The mathematical theory of dilute
gases. Springer Series in Applied Mathematical Sciences, Vol. 106 SpringerVerlag, New York (1994)
58. Boudin, L., Monaco, R., Salvarani, F.: A Kinetic Model for multidimensional
opinion formation. To appear in Phys. Rev. E (2009)
59. Holyst, J.A., Kacperski, K., Schweitzer, F.: Social impact models of opinion
dynamics. Ann. Rev. Comp. Phys., 9, 253–273 (2001)
60. Bellomo N.: Modeling Complex Living Systems. Birkhäuser, Boston (2008)
61. Aletti, G., Naldi, G., Toscani, G.: First-order continuous models of opinion
formation. SIAM J. Appl. Math., 67, 837–853 (2007)