A model of anonymous influence
with anti-conformist agents

We consider a society $N$ with $|N|=n$ agents who make a ‘yes’ or ‘no’ decision on some issue (binary choice). Initially every agent has an opinion on that issue, but by knowing the opinion of the others or by some social interaction, in each period the agent may modify his opinion due to mutual influence. In other words, there is an evolution in time of the opinion of the agents, which may or may not stop at some stable state of the society.

We define the state of the society at a given time as the set $S\subseteq N$ of agents whose opinion is ‘yes’. As usual, the cardinality of a set is denoted by the corresponding lower case, e.g., $s=|S|$. Our fundamental assumption is that the evolution of the state is ruled by a homogeneous Markov chain, that is, the state evolves at discrete time steps, the new opinions depend only on the opinions among the society in the last period, and the transition matrix giving the probability of all possible transitions from one state to another is constant over time.

We study the opinion formation process of the society, where the updating of opinion relies only on how many agents said ‘yes’ and how many said ‘no’ in the previous period, disregarding who said ‘yes’ and who said ‘no’. For this reason we call this opinion formation process anonymous. Both conformist and anti-conformist behaviors are allowed, i.e., agents can revise their opinions by following the trend as well as in a way contrary to the trend.

We now formalize the previous ideas. An (anonymous) aggregation rule describes for a given agent how the opinions of the other agents are aggregated in order to determine the updated opinion of this agent. Specifically, it is a mapping $p$ from $\{0,1,\ldots,n\}$ to $[0,1]$, assigning to any $0\leq s\leq n$, representing the number of agents saying ‘yes’, a quantity $p(s)$ which is the probability of saying ‘yes’ at next time step (and consequently, $1-p(s)$ is the probability of saying ‘no’). We focus on three classes of aggregation rules:

and mixed aggregation rules:

Note that $p\in A^{m}$ is not necessarily monotone, and that $p(0)>0$ and $p(n)<1$.

Let $p_{i}$ be agent $i$’s (anonymous) aggregation rule. Supposing that the update of opinion is done independently accross the agents, the probability of transition from a state $S$ to a state $T$ is

Observe that, as the aggregation rule is anonymous, the probability of transition to $T$ is the same for all states $S$ having the same size $s$.

We distinguish between three types of agents. We say that agent $i$ is purely conformist if $p_{i}\in A^{c}$, purely anti-conformist if $p_{i}\in A^{a}$, and is a mixed agent if $p_{i}\in A^{m}$. The society of agents is partitioned into

where $N^{c}$ is the set of purely conformist agents, $N^{a}$ the set of purely anti-conformist agents, and $N^{m}$ is the set of mixed agents. When $N^{m}=\emptyset$, we call it the pure case.

We make the following simplifying assumption: we suppose that all purely conformist agents, although having possibly different aggregation rules, have in common the same minimum number of ‘yes’ for them to say ‘yes’ with positive chance, and the same minimum number of ‘yes’ for them to say ‘yes’ with probability 1. Formally, we require that for each $i\neq j$ in $N^{c}$,

$l^{c}$ can be interpreted as the (maximum) number of ‘yes’ for which no effect on the probability of saying ‘yes’ arises, while $r^{c}$ is the (maximum) number of ‘no’ for which no effect on this probability is visible (see Figure 1). This assumption permits to conduct a precise analysis of convergence while remaining reasonable (it can be considered as a mean field approximation).

Similarly, we assume that all purely anti-conformist agents share the same minimum number of ‘yes’ for them to say ‘no’ with positive chance and the same minimum number of ‘yes’ for them to say ‘no’ with probability 1, i.e., for each $i\not=j$ in $N^{a}$,

As above, $l^{a}$ (respectively, $r^{a}$) is the maximum number of ‘yes’ (respectively, ‘no’) for which no effect on the probability of saying ‘yes’ is visible (see Figure 1).

These assumptions carry over mixed agents: any mixed agent $i$ has an aggregation rule $p_{i}$ which is a convex combination with coefficient $\alpha_{i}$ of a conformist aggregation rule (with fixed $l^{c},r^{c}$) and an anti-conformist aggregation rule (with fixed $l^{a},r^{a}$). We allow, however, that two mixed agents have different convex combination coefficients $\alpha_{i}$. Hence, a mixed agent $i$ can be seen as an agent who does not have a fixed behavior, but who is conformist with probability $\alpha_{i}$ and anti-conformist with probability $1-\alpha_{i}$. On Figure 1, we can see that a mixed agent with $\alpha_{i}=1/2$ has a very indecisive behavior.

Based on the above assumptions, in this paper we fully characterize all possible absorbing classes based on size comparison among $l^{c},r^{c},l^{a},r^{a}$ and the number of conformist and anti-conformist agents in the society.

We recall that (see, e.g., Kemeny and Snell 1976; Seneta 2006) for a Markov chain with set of states $E$ and transition matrix $\Lambda$ and its associated digraph $\Gamma$, a class is a subset $C$ of states such that for all states $e,f\in C$, there is a path in $\Gamma$ from $e$ to $f$, and $C$ is maximal w.r.t. inclusion for this property. A class is absorbing if for every $e\in C$ there is no arc in $\Gamma$ from $e$ to a state outside $C$. An absorbing class $C$ is periodic of period $k$ if it can be partitioned in blocks $C_{1},\ldots,C_{k}$ such that for $i=1,\ldots,k$, every outgoing arc of every state $e\in C_{i}$ goes to some state in $C_{i+1}$, with the convention $k+1=1$. When each $C_{1},\ldots,C_{k}$ reduces to a single state, one may speak of cycle of length $k$, by analogy with graph theory.

In our framework, states are subsets of agents and therefore classes are collections of sets, which we denote by calligraphic letters, like $\mathcal{C},\mathcal{B}$, etc. By definition, an absorbing class indicates the final state of opinion of the society. For instance, if an absorbing class reduces to a single state $S$, it means that in the long run, the society is dichotomous (unless $S=N$ or $S=\emptyset$, in which case consensus is reached): there is a set of agents $S$ who say ‘yes’ forever, while the other ones say ‘no’ forever. Otherwise, there are endless transitions with some probability from one set $S\in\mathcal{C}$ to another one $S^{\prime}\in\mathcal{C}$.

We now study the properties of the transition matrix $\Lambda$, with entries $\lambda_{S,T}$, $S,T\in 2^{N}$, where $\lambda_{S,T}$ is given by (4). Our aim is to find under which conditions one has a possible transition from $S$ to $T$, i.e., $\lambda_{S,T}>0$. From (4), we have:

We start with the pure case, i.e., $N^{m}=\emptyset$. We first observe that $p_{i}(0)=1$ if $i\in N^{a}$ and 0 otherwise, and $p_{i}(n)=1$ if $i\in N^{c}$ and 0 otherwise. Therefore, we have in any case the sure transitions

Moreover, we get immediately from (5) to (8) for any $s\neq 0,n$,

By combining these conditions and their negation in various ways, one can see that we can only have transitions to $\emptyset,N,N^{a},N^{c}$ and any of the subsets or supersets of $N^{a},N^{c}$. A convenient notation here is the interval notation: for two sets $S\subseteq T$, $[S,T]$ denotes the collection of sets $K$ such that $S\subseteq K\subseteq T$. For instance, $[\emptyset,N^{a}]$ and $[N^{a},N]$ denote respectively the collection of subsets of $N^{a}$ and the collection of supersets of $N^{a}$. Table 1 summarizes the possible transitions, adding also those from $S=\emptyset$ and $S=N$.

Let us introduce $Z=(l^{c},r^{c},l^{a},r^{a})$ and write $p_{i}^{Z}$ to emphasize the dependency of $p_{i}$ on these parameters (and similarly for $\lambda_{S,T}$). Equations (9) to (12) show striking symmetries, in particular, when interchanging conformists and anti-conformists. $Z$ being given, we introduce the reversal of $Z$, $Z^{\partial}:=(r^{c},l^{c},r^{a},l^{a})$, and the interchange of $Z$, $Z^{\prime}=(l^{a},r^{a},l^{c},r^{c})$ which amounts to interchanging conformists with anti-conformists. Considering these operations, we observe the following symmetries:

1. (i)

   Interchange:

   	$\displaystyle p_{i}^{Z}(s)>0\text{ for }i\in N^{c}$		$\displaystyle\Leftrightarrow$		$\displaystyle p_{i}^{Z^{\prime}}(s)<1\text{ for }i\in N^{a}$
   	$\displaystyle p_{i}^{Z}(s)<1\text{ for }i\in N^{c}$		$\displaystyle\Leftrightarrow$		$\displaystyle p_{i}^{Z^{\prime}}(s)>0\text{ for }i\in N^{a}$

   (idem with $N^{a}$, $N^{c}$ exchanged)

2. (ii)

   Reversal:

   	$\displaystyle p_{i}^{Z}(s)>0\text{ for }i\in N^{c}$		$\displaystyle\Leftrightarrow$		$\displaystyle p_{i}^{Z^{\partial}}(n-s)<1\text{ for }i\in N^{c}$
   	$\displaystyle p_{i}^{Z}(s)<1\text{ for }i\in N^{c}$		$\displaystyle\Leftrightarrow$		$\displaystyle p_{i}^{Z^{\partial}}(n-s)>0\text{ for }i\in N^{c}$

   (idem with $N^{a}$, $N^{c}$ exchanged)

3. (iii)

   Interchange and reversal:

   	$\displaystyle p_{i}^{Z}(s)>0\text{ for }i\in N^{c}$		$\displaystyle\Leftrightarrow$		$\displaystyle p_{i}^{(Z^{\partial})^{\prime}}(n-s)>0\text{ for }i\in N^{a}$
   	$\displaystyle p_{i}^{Z}(s)<1\text{ for }i\in N^{c}$		$\displaystyle\Leftrightarrow$		$\displaystyle p_{i}^{(Z^{\partial})^{\prime}}(n-s)<1\text{ for }i\in N^{a}$

   (idem with $N^{a}$, $N^{c}$ exchanged)

The second case is of particular interest and leads to the following lemma.

Next we consider the mixed case, assuming $p_{i}=\alpha_{i}q^{c}+(1-\alpha_{i})q^{a}$ with $q^{c}\in A^{c}$, $q^{a}\in A^{a}$ and fixed $l^{c},r^{c},l^{a},r^{a}$. We can easily derive the conditions for $p_{i}(s)$ to be 0 or 1, using (9) to (12). For any $0<s<n$, we obtain

and $0<p_{i}(s)<1$ for all other cases. Finally, if $S=\emptyset$, then $\lambda_{S,T}>0$ for every $T\in[N^{a},N^{a}\cup N^{m}]$, and if $S=N$, then $\lambda_{S,T}>0$ for every $T\in[N^{c},N^{c}\cup N^{m}]$. Table 2 presents possible transitions from $S\in 2^{N}$ in the mixed case.

We recall the anonymous model of conformism (Förster et al. 2013) and show that it is equivalent to our class of conformist aggregation rules. By doing this we show that the present model is a natural extension of Förster et al. 2013.

The assumption that agents modify their opinions in a Markovian way is basically that underlying Grabisch and Rusinowska 2013. As the number of states is $2^{n}$, the size of the transition matrix is $2^{n}\times 2^{n}$. In order to avoid this exponential complexity, like in the present paper the latter reference uses a simple mechanism to generate the transition matrix, based on aggregation functions, that is, nondecreasing mappings $A:[0,1]^{n}\rightarrow[0,1]$ satisfying $A(0,\ldots,0)=0$ and $A(1,\ldots,1)=1$. Specifically, supposing that agent $i$ aggregates opinions by the function $A_{i}$, the probability that agent says ‘yes’ at next time step, given that $S$ is the set of agents saying ‘yes’ at present, is given by

where $1_{S}$ is the indicator function of $S$, i.e., $1_{S}(i)=1$ iff $i\in S$ and 0 otherwise. This model is presented and studied in general in Grabisch and Rusinowska 2013.

The most common example of aggregation function, used, e.g., in DeGroot 1974, is the weighted arithmetic mean

where $x=(x_{1},\ldots,x_{n})$ and the $w^{i}_{j}$’s are weights on the entries, satisfying $w^{i}_{j}\geq 0$ and $\sum_{j=1}^{n}w^{i}_{j}=1$. The weight $w^{i}_{j}$ represents to which extent agent $i$ puts confidence on the opinion $x_{j}$ of agent $j$. It depicts a situation where every agent knows the identity of every other agent, and is able to assess to which extent he trusts or agrees with the opinion or personal tastes of others.

Förster et al. 2013 investigate the model of conformism with anonymous social influence, which depends only on the number of agents with a certain opinion, not on their identity. They use the ordered weighted averages (commonly called OWA operators, Yager 1988), which are the unique anonymous aggregation functions:

where the entries $x_{1},\ldots,x_{n}$ are rearranged in decreasing order: $x_{(1)}\geq x_{(2)}\geq\cdots\geq x_{(n)}$, and $w=(w_{1},\ldots,w_{n})$ is the weight vector defined as above. Hence, the weight $w_{j}$ is not assigned to agent $j$ but to rank $j$, and thus permits to model quantifiers. For example, taking $w_{1}=1$ and all other weights being 0 models the quantifier “there exists”. Indeed, it is enough to have one of the entries being equal to 1 to get 1 as output. In our context, it means that only one agent saying ‘yes’ is enough to make your opinion being ’yes’ for sure. Similarly, “for all” is modeled by $w_{n}=1$ and all other weights being 0, and means that you need that all agents (including you) say ‘yes’ to be sure to continue to say ‘yes’. Intermediate situations can be modeled as well.

For the anonymous model of conformism, there exist two types of absorbing classes (Förster et al. 2013):

1. (i)

   any single state $S\in 2^{N}$ (including the consensus states $\emptyset$ and $N$);

2. (ii)

   union of intervals of the type $[S,S\cup K]$, where $S,K\neq\emptyset,N$ (recall that $[S,S\cup K]=\{T\in 2^{N}\mid S\subseteq T\subseteq S\cup K\}$).

For the second case, when the absorbing class is reduced to a single interval $[S,S\cup K]$, it depicts a situation in the long run where agents in $S$ say ‘yes’, those outside $S\cup K$ say ‘no’, and those in $K$ oscillate between ‘yes’ and ‘no’ forever. Interestingly, no periodic class can occur, although in general for arbitrary aggregation functions cycles can occur (Grabisch and Rusinowska 2013).

We now establish the relation with our framework. Supposing that agent $i$ aggregates opinions anonymously by $\mathrm{OWA}_{w^{i}}$ with weight vector $w^{i}$, we have from (15) and (17), for any $S\subseteq N$:

We see that $p_{i}(s)$ is an aggregation rule in the sense of Section 2.1, and that, due to the properties of $w^{i}$, $p_{i}\in A^{c}$. Conversely, for any $p\in A^{c}$, we can define $w\in{\mathbb{R}}^{n}$ by $w_{j}=p(j)-p(j-1)$, $j=1,\ldots,n$, and by properties of $p\in A^{c}$, $w$ is a well-defined weight vector s.t. $p\equiv\mathrm{OWA}_{w}$. The introduction of the classes of aggregation rules $A^{a},A^{m}$ are therefore natural generalizations of the conformist model of Förster et al. 2013.

Note that the numbers $l^{c},r^{c},l^{a},r^{a}$ defined in (5) - (8) correspond to the numbers of left and right zeroes in the weight vectors of an OWA operator, for conformists and anti-conformists, respectively. Moreover, the assumptions (5) - (8) simply mean that while agents in $N^{c}$ ($N^{a}$, respectively) may have different weight vectors, the number of left and right zeroes is the same for all of them. The number of left/right zeroes indicates how many people the agent needs in order to start being influenced towards the yes/no opinion. In particular, a non symmetrical weight vector w.r.t. the number of left and right zeroes means that the agent is biased towards the ‘yes’ or ‘no’ answer, i.e., he needs a different number of people to start being convinced to say ‘yes’ or ‘no’.

We mentioned in Section 2.3 that for the anonymous model of conformism, it was found that any state could be absorbing and that other absorbing classes are intervals of sets. However, the case with $N^{a}=\emptyset$ is much more particular as every agent has the same $l,r$, and there are very few possibilities left, as shown below. To this end, it suffices to rewrite Table 1 which becomes:

It follows that only $\emptyset,N$ are absorbing states and neither $2^{N}$ nor any of its subcollections is an absorbing class, because a transition from any $S\neq\emptyset,N$ to $\emptyset$ or $N$ is possible.

Both $\emptyset,N$ are possible, regardless of the values of $r^{c},l^{c}$. This means that the society converges to a consensus, which depends on the initial state $S$. If $s\leq l^{c}$, there is convergence to ’no’ in one step, and if $s\geq n-r^{c}$, there is convergence to ’yes’ in one step.

The analysis of the extreme case with $N^{c}=\emptyset$ is also done by rewriting Table 1 which becomes now:

Then there is only one absorbing class which is the cycle $\emptyset\stackrel{{\scriptstyle 1}}{{\rightarrow}}N\stackrel{{\scriptstyle 1}}{{\rightarrow}}\emptyset$. We see that, without much surprise, a society of anti-conformist agents can never reach a stable state.

We observe the following basic facts:

• (F0)

  $\emptyset\stackrel{{\scriptstyle 1}}{{\rightarrow}}N^{a}$, $N\stackrel{{\scriptstyle 1}}{{\rightarrow}}N^{c}$ (as already observed).

• (F1)

  If $\mathcal{S}\stackrel{{\scriptstyle 1}}{{\rightarrow}}\mathcal{T}$, $\mathcal{S}^{\prime}\stackrel{{\scriptstyle 1}}{{\rightarrow}}\mathcal{T}^{\prime}$ and $\mathcal{S}\subset\mathcal{S}^{\prime}$, then $\mathcal{T}\subseteq\mathcal{T}^{\prime}$.

• (F2)

  Applying (F0) and (F1), we find that in a transition $\mathcal{S}\rightarrow\mathcal{T}$, $\emptyset\in\mathcal{S}$ implies $N^{a}\in\mathcal{T}$ and $N\in\mathcal{S}$ implies $N^{c}\in\mathcal{T}$.

• (F3)

  Consider $\mathcal{S}\stackrel{{\scriptstyle 1}}{{\rightarrow}}\mathcal{T}_{1}\stackrel{{\scriptstyle 1}}{{\rightarrow}}\cdots\stackrel{{\scriptstyle 1}}{{\rightarrow}}\mathcal{T}_{p}$, with $p\geq 2$. If $\mathcal{S}\subseteq\mathcal{T}_{1}$, then $\mathcal{S}\subseteq\mathcal{T}_{1}\subseteq\cdots\subseteq\mathcal{T}_{p}$.

• (F4)

  $2^{N}$ is a possible absorbing class. Indeed, take $l^{c}=r^{c}=l^{a}=r^{a}=0$. From Table 1 we immediately see that for any $S\neq\emptyset,N$ we have $S\stackrel{{\scriptstyle 1}}{{\rightarrow}}2^{N}$. Since the power set of the set of states is the “default” absorbing class when no other can exist, we exclude it from our study and do not consider transitions to $2^{N}$.

• (F5)

  From Table 1, we see that we have to deal only with the sets $\emptyset,N^{a},N^{c},N$ and the intervals $[\emptyset,N^{c}],[\emptyset,N^{a}],[N^{a},N],[N^{c},N]$ ($2^{N}$ being excluded by (F4)), i.e., only these can be constituents of an absorbing class. We put

  $${\mathbb{B}}=\{\{\emptyset\},\{N^{a}\},\{N^{c}\},\{N\},[\emptyset,N^{c}],[\emptyset,N^{a}],[N^{a},N],[N^{c},N]\}$$

  the set of collections relevant to our study. Intervals not reduced to a singleton are called nontrivial intervals.

• (F6)

  $\mathcal{S}\subseteq 2^{N}$ is an absorbing class if and only if $\mathcal{S}\stackrel{{\scriptstyle 1}}{{\rightarrow}}\mathcal{S}$ and $\mathcal{S}$ is connected (i.e., there is a path (chain of transitions) from $S$ to $T$ for any $S,T\in\mathcal{S}$).

(F6) will be our unique tool to find aperiodic absorbing classes, while periodic classes are of the form $\mathcal{S}_{1}\stackrel{{\scriptstyle 1}}{{\rightarrow}}\cdots\stackrel{{\scriptstyle 1}}{{\rightarrow}}\mathcal{S}_{p}$, with $\mathcal{S}_{1},\ldots,\mathcal{S}_{p}\subset 2^{N}$ and being pairwise disjoint (no common set between $\mathcal{S}_{i},\mathcal{S}_{j}$), and $\mathcal{S}_{1}\cup\cdots\cup\mathcal{S}_{p}$ must be connected.

Since $N^{m}=\emptyset$, we have $n^{a}=n-n^{c}$, where $n^{a}=|N^{a}|$ and $n^{c}=|N^{c}|$. Hence, the model is entirely determined by $l^{c},r^{c},l^{a},r^{a},n^{c},n$. We recall that these parameters must satisfy the following constraints:

Based on these facts, we can show the main result of this section.

The proof of Theorem 1 can be found in the appendix.

A complete analysis and interpretation of these results seems to be out of reach. Nevertheless, this becomes quite possible and instructive when considering special situations (see below). We recall that these results are given without any simplifying assumption and are valid whatever the value of the parameters describing the society. Once they are known, Theorem 1 immediately gives the possible absorbing classes. We stick in what follows to general comments, reserving more specific ones and a deeper analysis to the particular situations studied hereafter:

1. (i)

   Although there are many absorbing classes, they can be grouped in three categories. The first category ((1) and (2)) comprises absorbing states (singleton classes) and represent a stable state of the society in the long run, which happens to be a polarization: a part of the society says ’yes’ (the anti-conformists in case (1), the conformists in case (2)), while the other part says ’no’. Observe that consensus never occurs.

   The second category (from (11) to (20)) comprises aperiodic classes which are not reduced to singletons. (11) and (12) are simple cases where the class is an interval. For (11), it means that in the long run all the conformists say ‘no’, while the anti-conformists have a chaotic behavior, in the sense that, at each time step, a (different) fraction of them says ’yes’ while the remaining part says ’no’. We may call this a fuzzy polarization. Cases (13) through (18) can be interpreted in a similar way, although the behavior is yet more chaotic: taking (13) for example, in the long run, at each time step the set of agents saying ’yes’ can be any fraction of anti-conformists or any fraction of conformists (but not a mixed group). We may call this a chaotic polarization. The extreme case of chaotic behavior is given by (20): at any time step in the long run, the set of agents saying ’yes’ can be any subset of agents.

   The third category (from (3) to (10)) comprises cycles and periodic classes, which also have a natural interpretation. Cycles (from (3) to (7)) are succession in time of states of polarization and/or consensus, endlessly in the same order. Periodic classes ((8) to (10)) mix the cyclic behavior with intervals as in the case of fuzzy polarization, and for this reason may be called fuzzy cycles. For example, in case (8) at some step in the long run all anti-conformists say ‘yes’, in the following step they all say ‘no’ but a fraction of conformists says ‘yes’, then in the next step again all anti-conformists say ‘yes’, etc.

2. (ii)

   Cases (1) to (20) are not exclusive. This can immediately be seen by considering cases (1) and (2). Indeed, for a given society, which is represented by the set of parameters $n$, $n^{c}$, $l^{a}$, $r^{a}$, $l^{c}$ and $r^{c}$, under some conditions both cases (1) and (2) are possible, and therefore two different absorbing classes might occur, $N^{a}$ and $N^{c}$. However, the process will end up in only one of them, with some probability.

3. (iii)

   The analysis for conformists and anti-conformists is not symmetric. For example, $[\emptyset,N^{a}]$ is a possible absorbing class but not $[\emptyset,N^{c}]$. However, while there is no symmetry between “$a$” and “$c$” in this framework, there exists symmetry between $S$ and $N\setminus S$ as pointed out in Lemma 1.

Prior to the study of the mixed case, we consider the situation where the society is formed only by mixed agents: $N=N^{m}$. It is easily done by using Eqs. (13) and (14):

which permits to rewrite Table 1:

We can see that $\emptyset,N$ are not absorbing states, hence the only absorbing class is $2^{N}$.

Next, we consider a society in which pure conformists, pure anti-conformists, and mixed agents co-exist, i.e., $n=n^{a}+n^{c}+n^{m}$, with $n^{a}>0$, $n^{c}>0$ and $n^{m}>0$. We get the following result.