betaFeedback

WorksEngineering

Robust Action Clusters and Cascades in Opinion Dynamics with Heterogeneous Action Thresholds

Abstract

We study the preservation and propagation of public actions in networks whose agents hold continuous private opinions. Agents update their opinions from their neighbors' observed actions and express them through action rules whose thresholds differ across agents. We characterize groups of agents that preserve a given common action regardless of the outsiders' actions. We call such a group, paired with that action, a robust action cluster. We show that a group is a robust action cluster for that action if and only if each member places enough of its influence weight on the group, where the required influence depends on the member's own action rule. We further give a sufficient condition under which an agent outside the cluster permanently adopts the cluster's action, and we iterate it to obtain a cascade of agents that are guaranteed to adopt that action. Finally, numerical examples show that replacing heterogeneous thresholds by their mean can change both the propagation of an action and the presence of robust action clusters.

Introduction

People often hold a nuanced private opinion but express it through a coarse public choice: a vote, a rating, or one of a few stated positions. Classical averaging models describe social influence through the exchange of opinions [1]. Continuous-opinion discrete-action (CODA) models distinguish these private opinions from the public actions through which agents influence one another [2], [3]. Two people can express the same position despite holding different opinions or using different rules to translate opinions into actions. A group defined by its public behavior can therefore be heterogeneous in both respects.

This distinction raises two questions. When can a group preserve a common action despite influence from the rest of the network? When does that action spread to others? Threshold models of collective behavior show that aggregate outcomes can depend on how individual response thresholds are distributed, not only on their mean [4, 5, 6]. In our setting, by contrast, each action is determined by a private opinion that retains memory of past interactions. No agent is stubborn: a group can preserve an action only through the opinion dynamics that all agents follow.

Related work. The binary CODA model of [3] studies preservation and propagation through majority conditions. Its state-dependent gain and history-dependent tie rule differ from the dynamics considered here. Related stochastic models couple multilevel opinions with binary actions [7]. Quantized behaviors are also analyzed in continuous time [8], [9] and coupled with external dynamics [10]. Complementary models study discrepancies between private and expressed opinions caused by social conformity [11] and the coevolution of binary actions and continuous opinions on two-layer networks [12]. Strategic communication on influence networks can also generate heterogeneous quantized communication rules [13]. Here, the action rules are prescribed and share a common finite action set.

Beyond binary actions, finite verbal scales [14] and a middle option [15] have been studied by simulation. Multistage contagion models further show that additional states can change cascade dynamics [16]. In threshold dynamics, cohesive groups can obstruct contagion [17], while robust coordination under asynchronous binary updates is studied in [18]. These works motivate examining how network cohesion and individual action rules jointly determine action preservation and propagation.

We study this interplay on weighted directed networks with inertial private opinions and heterogeneous thresholds on a common finite action set. For intermediate actions, robustness requires a two-sided cohesion condition that protects against influence toward both lower and higher actions. Actions can switch indefinitely, even with two agents (Example 1), so we study permanent adoption without assuming convergence of the entire network. Our contributions are threefold. First, we give a necessary and sufficient condition for a group to be a robust action cluster. It compares the influence weight each member places on the group with an agent-dependent influence threshold set by that member’s action rule. Thus, it separates the group’s cohesion from the individual rules. Second, a strict version of this condition guarantees that an agent placing enough influence weight on a robust action cluster adopts its action permanently, from any initial opinion. Applying it repeatedly yields a finite cascade of guaranteed adopters and a sufficient condition for permanent consensus on that action. The cascade identifies all guaranteed adopters for two actions, but provides only a sufficient condition for larger action sets. Third, two numerical examples illustrate that mean thresholds alone determine neither the propagation of an intermediate action from a robust action cluster nor the presence of robust action clusters. On a lattice, heterogeneous rules produce robust action clusters. The corresponding homogeneous rule, which gives every agent the mean thresholds, yields none in its final action profile, although the action profile remains fixed in both runs.

The model and its general properties are introduced in Sections II and III; Sections IV and V establish preservation and adoption conditions. Section VI illustrates permanent adoption from a robust seed and distinguishes action profiles that remain fixed from robust action clusters.

Notation. Vectors and matrices are bold (e.g., v∈Rn\mathbf{v} \in\mathbb{R}^{n}, W∈Rn×m\mathbf{W} \in\mathbb{R}^{n \times m}), with viv_i and WijW_{ij} denoting their respective entries. Vector inequalities are componentwise: x≤y\mathbf{x} \le\mathbf{y} means xi≤yix_i \le y_i for every ii. The vector 1\mathbf{1} has all entries equal to one, and II is the identity matrix. We write ∣S∣|S| for set cardinality, and R+\mathbb{R}_{+} for the set of nonnegative reals. We denote the interior and closure of a set CC by C∘C^{\circ} and C‾\overline{C}, respectively.

Model

Consider NN agents indexed by V:={1,…,N}\mathcal{V} := \{1,\ldots,N\}. Their interactions are described by a weighted directed graph G=(V,E,W)\mathcal{G} = (\mathcal{V},\mathcal{E},\mathbf{W}), where W=[Wij]∈R+N×N\mathbf{W} = [W_{ij}] \in\mathbb{R}_{+}^{N \times N}. The entry WijW_{ij} is the influence weight of agent jj on agent ii. A link (j,i)∈E(j,i) \in\mathcal{E} is present if and only if Wij>0W_{ij} > 0. Thus Ni:={j∈V:(j,i)∈E}\mathcal{N}_i := \{j \in\mathcal{V} : (j,i) \in\mathcal{E}\} is the set of agents observed by ii. We normalize each agent’s influence weights and assume the following properties for the interaction graph.

Standing Assumption (SA1): The matrix W\mathbf{W} is time-invariant, nonnegative, and row-stochastic (W1=1\mathbf{W}\mathbf{1} = \mathbf{1}), with zero diagonal.

For a group R⊆V\mathcal{R} \subseteq\mathcal{V}, agent ii places influence weight wi(R):=∑j∈RWijw_i(\mathcal{R}) := \sum_{j \in\mathcal{R}} W_{ij} on R\mathcal{R}. By (SA1), it is the fraction of ii’s total influence weight, so wi(V∖R)=1−wi(R)w_i(\mathcal{V} \setminus\mathcal{R}) = 1 - w_i(\mathcal{R}).

Private Opinions and Public Actions

At each time t∈{0,1,2,…}t \in\{0,1,2,\ldots\}, agent ii holds a private opinion xi(t)∈[−1,1]x_i(t) \in[-1,1] and broadcasts a public action. Agents observe only their neighbors’ actions.

All agents choose from a common ordered action set A:={a1<⋯<aL}⊂[−1,1]\mathcal{A} := \{a_1 < \cdots< a_L\} \subset[-1,1], with L≥2L \ge2. The actions strictly between a1a_1 and aLa_L are the intermediate actions. At time tt, agent ii broadcasts the action qi(xi(t))q_i(x_i(t)), a quantized version of its opinion. The map qiq_i, called the action rule of agent ii, assigns an action to each opinion through agent-dependent thresholds. It is defined as follows.

Definition 1 (Action rule): For each i∈Vi \in\mathcal{V}, let −1<θi,1<⋯<θi,L−1<1-1 < \theta_{i,1} < \cdots< \theta_{i,L-1} < 1 be the thresholds of agent ii, with the conventions θi,0:=−∞\theta_{i,0} := -\infty and θi,L:=+∞\theta_{i,L} := +\infty. The action rule qi:R→Aq_i: \mathbb{R} \to\mathcal{A} of agent ii is

qi(x):=aℓif x∈[θi,ℓ−1,θi,ℓ),ℓ=1,…,L.(1)q_i(x) := a_\ell\quad\text{if } x \in[\theta_{i,\ell-1},\theta_{i,\ell}), \quad\ell= 1,\ldots,L. \tag*{(1)}

By construction, each action rule qiq_i is nondecreasing and right-continuous: at a threshold, the agent takes the higher of the two adjacent actions (Fig. 1). The opinions at which an agent i∈Vi \in\mathcal{V} takes action a∈Aa \in\mathcal{A} form the action interval

Ii(a):={x∈R:qi(x)=a}=[θ‾i(a),θ‾i(a)),(2)\mathcal{I}_i(a) := \{x \in\mathbb{R} : q_i(x) = a\} = [\underline{\theta}_i(a),\overline{\theta}_i(a)), \tag*{(2)}

with lower and upper thresholds θ‾i(aℓ):=θi,ℓ−1\underline{\theta}_i(a_\ell) := \theta_{i,\ell-1} and θ‾i(aℓ):=θi,ℓ\overline{\theta}_i(a_\ell) := \theta_{i,\ell}. Only θi(a1)\theta_i(a_1) and θ‾i(aL)\overline{\theta}_i(a_L) are infinite.

We require self-consistency: each action is a fixed point of every action rule [8].

Standing Assumption (SA2): For all i∈Vi \in\mathcal{V} and a∈Aa \in\mathcal{A}, qi(a)=aq_i(a) = a; equivalently, a∈Ii(a)a \in\mathcal{I}_i(a).

In terms of the thresholds, this means that aℓ<θi,ℓ≤aℓ+1a_\ell< \theta_{i,\ell} \le a_{\ell+1} for every i∈Vi \in\mathcal{V} and ℓ=1,…,L−1\ell= 1,\ldots,L-1.

Opinion Dynamics

The observed actions in turn drive the evolution of private opinions. The social input to agent ii is the weighted average of its neighbors’ actions, ui(t):=∑j∈NiWijqj(xj(t))u_i(t) := \sum_{j \in\mathcal{N}_i} W_{ij}q_j(x_j(t)). It depends on the neighbors’ opinions only through their actions. The opinion of agent ii evolves according to the discrete-time dynamics

xi(t+1)=αixi(t)+(1−αi)ui(t),(3)x_i(t+1) = \alpha_i x_i(t) + (1-\alpha_i)u_i(t), \tag*{(3)}

where αi∈(0,1)\alpha_i \in(0,1) is the inertia (self-persistence) parameter of agent ii, determining the weight assigned to its current opinion. Since Wii=0W_{ii}=0, the social input excludes the agent’s own action, and its own opinion enters only through inertia.

We write x(t):=(x1(t),…,xN(t))T\mathbf{x}(t) := (x_1(t),\ldots,x_N(t))^{\mathsf{T}} and q(x):=(q1(x1),…,qN(xN))T\mathbf{q}(\mathbf{x}) := (q_1(x_1),\ldots,q_N(x_N))^{\mathsf{T}}. With Dα:=diag⁡(α1,…,αN)D_\alpha:= \operatorname{diag}(\alpha_1,\ldots,\alpha_N), the dynamics take the vector form

x(t+1)=Dαx(t)+(I−Dα)Wq(x(t))=:f(x(t)).(4)\mathbf{x}(t+1) = D_\alpha\mathbf{x}(t) + (I-D_\alpha)\mathbf{W}\mathbf{q}(\mathbf{x}(t)) =: \mathbf{f}(\mathbf{x}(t)). \tag*{(4)}

The public action profile at time tt is thus q(x(t))\mathbf{q}(\mathbf{x}(t)), and the state space is X:=[−1,1]N\mathcal{X} := [-1,1]^N. The rest of the paper analyzes this model and characterizes when a group of agents preserves or adopts a common action permanently.

General Properties of the Opinion Dynamics

We first establish basic structural properties of the map f\mathbf{f} on the state space X\mathcal{X}.

Lemma 1 (Well-posedness and monotonicity): We have: (i) f(X)⊆X\mathbf{f}(\mathcal{X}) \subseteq\mathcal{X}. (ii) f\mathbf{f} is monotone: x≤y\mathbf{x} \le\mathbf{y} implies f(x)≤f(y)\mathbf{f}(\mathbf{x}) \le\mathbf{f}(\mathbf{y}).

Proof. (i) Since A⊆[−1,1]\mathcal{A} \subseteq[-1,1] and W\mathbf{W} is row-stochastic, ui∈[−1,1]u_i \in[-1,1], and each update (3) is a convex combination of xix_i and uiu_i. (ii) The matrices DαD_\alpha and (I−Dα)W(I-D_\alpha)\mathbf{W} are nonnegative, and each qiq_i is nondecreasing.

Action rules of two agents

Figure 1. Action rules of two agents with A={−1,0,1}\mathcal{A}=\{-1,0,1\}, Ii(a2)=[−0.4,0.4)\mathcal{I}_i(a_2)=[-0.4,0.4), and Ij(a2)=[−0.1,0.7)\mathcal{I}_j(a_2)=[-0.1,0.7), on the opinion range [−1,1][-1,1]. Filled and open dots show right-continuity. Each step meets the identity (dashed) at its own action level, as required by (SA2).

Next, we show that if all agents share the same action, they preserve it for all time. For any group R⊆V\mathcal{R} \subseteq\mathcal{V} and any action a∈Aa \in\mathcal{A}, let Xa(R):={x∈X:qi(xi)=a for all i∈R}\mathcal{X}_a(\mathcal{R}) := \{x \in\mathcal{X} : q_i(x_i) = a \text{ for all } i \in\mathcal{R}\} be the states in which every member of R\mathcal{R} holds action aa, with the other agents unconstrained.

Lemma 2 (Global action consensus). For every a∈Aa \in\mathcal{A}, the consensus state a1a\mathbf{1} is a fixed point of ff, and x∈Xa(V)x \in\mathcal{X}_a(\mathcal{V}) implies f(x)∈Xa(V)f(x) \in\mathcal{X}_a(\mathcal{V}).

Proof. By (SA2), a∈Ii(a)a \in\mathcal{I}_i(a) for all i∈Vi \in\mathcal{V}, so q(a1)=a1\mathbf{q}(a\mathbf{1}) = a\mathbf{1} and f(a1)=a1f(a\mathbf{1}) = a\mathbf{1} by row-stochasticity of WW. More generally, if x∈Xa(V)x \in\mathcal{X}_a(\mathcal{V}), then ui=au_i = a for every i∈Vi \in\mathcal{V}, so (3) gives fi(x)=αixi+(1−αi)af_i(x) = \alpha_i x_i + (1-\alpha_i)a. Since xi,a∈Ii(a)x_i,a \in\mathcal{I}_i(a), their convex combination lies in Ii(a)\mathcal{I}_i(a), and Lemma 1(i) ensures f(x)∈Xf(x) \in\mathcal{X}, so f(x)∈Xa(V)f(x) \in\mathcal{X}_a(\mathcal{V}).

By Lemma 1(ii), ordered initial states give ordered trajectories. Monotonicity, however, does not imply convergence, since it still allows persistent switching. As in synchronous threshold dynamics [19], agents can switch actions periodically, even on a connected graph.

The following example shows that two agents can enter a period-two orbit, preventing opinion convergence.

Example 1 (Persistent switching): Consider two agents with action set A={−a,a}\mathcal{A} = \{-a,a\}, where 0<a≤10 < a \le1, and dynamics specified by

W=[0110],qi(x)={−a,x<0,a,x≥0.W = \begin{bmatrix} 0 & 1 \\ 1 & 0 \end{bmatrix}, \qquad q_i(x) = \begin{cases} -a, & x < 0, \\ a, & x \ge0. \end{cases}

For arbitrary α1,α2∈(0,1)\alpha_1,\alpha_2 \in(0,1), define ri:=a1−αi1+αi>0r_i := a\frac{1-\alpha_i}{1+\alpha_i} > 0. The opinion update maps (r1,−r2)(r_1,-r_2) to (−r1,r2)(-r_1,r_2) and back. Both states avoid the threshold because ri>0r_i > 0, so the agents’ actions switch indefinitely.

Since the action profile need not converge, we study which agents hold a given action permanently, rather than convergence of the whole network. Section IV provides a necessary and sufficient condition under which a group of agents keeps a common action independently of what the rest of the network does. This generalizes Lemma 2 from V\mathcal{V} to any R⊆V\mathcal{R} \subseteq\mathcal{V}. Section V then provides sufficient conditions under which that action propagates permanently to agents outside the group.

Robust Action Clusters

Agents holding a common action reinforce it through mutual influence, while the other agents may hold different actions. In this section, we investigate when such a group preserves its action regardless of the rest of the network. We start with the central definition of the paper, which concerns a group of agents paired with a public action.

Definition 2 (Robust action cluster): Let a∈Aa \in\mathcal{A} and ∅≠R⊆V\varnothing\ne\mathcal{R} \subseteq\mathcal{V}. R\mathcal{R} is a robust action cluster for action aa if

x∈Xa(R)⇒f(x)∈Xa(R),(5)x \in\mathcal{X}_a(\mathcal{R}) \Rightarrow f(x) \in\mathcal{X}_a(\mathcal{R}), \tag*{(5)}

i.e., Xa(R)\mathcal{X}_a(\mathcal{R}) is forward invariant under the dynamics (4). For shorthand, we write “(R,a)(\mathcal{R},a) is robust”. The agents in V∖R\mathcal{V}\setminus\mathcal{R} are the outsiders of R\mathcal{R}.

By Lemma 2, (V,a)(\mathcal{V},a) is always robust. In general, a robust action cluster preserves its action from every initial state in Xa(R)\mathcal{X}_a(\mathcal{R}), whatever the actions of its outsiders. Our first objective is to characterize the robust action clusters for each action a∈Aa \in\mathcal{A}. The second, addressed in Section V, is to determine when their action spreads to outsiders.

From the dynamics in (3), each updated opinion is a convex combination of the current opinion and the social input. Hence, an agent whose opinion and input both lie in Ii(a)\mathcal{I}_i(a) keeps action aa. We therefore characterize robustness of (R,a)(\mathcal{R},a) by bounding the social inputs that an agent can receive.

If x(t)∈Xa(R)x(t) \in\mathcal{X}_a(\mathcal{R}), splitting the social input between R\mathcal{R} and its outsiders gives ui(t)=a wi(R)+∑j∉RWijqj(xj(t))u_i(t)=a\,w_i(\mathcal{R})+\sum_{j\notin\mathcal{R}}W_{ij}q_j(x_j(t)). Since every action lies in [a1,aL][a_1,a_L] and the weights WijW_{ij} are nonnegative, each agent’s social input lies in the worst-case input interval Ui(R,a):=[u‾i(R,a),u‾i(R,a)]\mathcal{U}_i(\mathcal{R},a):=[\underline{u}_i(\mathcal{R},a),\overline{u}_i(\mathcal{R},a)], where

u‾i(R,a):=a wi(R)+(1−wi(R))a1,\underline{u}_i(\mathcal{R},a):=a\,w_i(\mathcal{R})+(1-w_i(\mathcal{R}))a_1,
u‾i(R,a):=a wi(R)+(1−wi(R))aL.(6)\overline{u}_i(\mathcal{R},a):=a\,w_i(\mathcal{R})+(1-w_i(\mathcal{R}))a_L. \tag*{(6)}

Both endpoints are attained when all outsiders hold a1a_1 or aLa_L, respectively. The interval narrows toward aa as wi(R)w_i(\mathcal{R}) increases (Figure 2). To preserve aa, a member must resist changes to both lower and higher actions.

Influence threshold of agent $i \in \mathcal{R}$ for action $a$

Figure 2. Influence threshold of agent i∈Ri \in\mathcal{R} for action aa. (a) Agent ii places influence weight wi(R)w_i(\mathcal{R}) on its own group, which holds aa, and 1−wi(R)1-w_i(\mathcal{R}) on the other agents, which may hold any action. (b) Worst-case input bounds (6) versus wi(R)w_i(\mathcal{R}). They cross the thresholds of Ii(a)\mathcal{I}_i(a) (gray) at κi+(a)\kappa_i^+(a) and κi−(a)\kappa_i^-(a), and both stay between them exactly when wi(R)≥κi(a)w_i(\mathcal{R}) \ge\kappa_i(a) (hatched); here, κi(a)=κi−(a)\kappa_i(a)=\kappa_i^-(a).

Requiring u‾i(R,a)≥θ‾i(a)\underline{u}_i(\mathcal{R},a) \ge\underline{\theta}_i(a) and u‾i(R,a)≤θ‾i(a)\overline{u}_i(\mathcal{R},a) \le\overline{\theta}_i(a) and solving for wi(R)w_i(\mathcal{R}) gives two lower bounds on wi(R)w_i(\mathcal{R}), which motivate the following definition.

Definition 3 (Influence threshold): For an agent ii and an action aa, let κi(a):=max⁡{κi−(a),κi+(a)}\kappa_i(a):=\max\{\kappa_i^-(a),\kappa_i^+(a)\} be the influence threshold, where

κi−(a):=θ‾i(a)−a1a−a1,κi+(a):=aL−θ‾i(a)aL−a,(7)\kappa_i^-(a):=\frac{\underline{\theta}_i(a)-a_1}{a-a_1}, \qquad\kappa_i^+(a):=\frac{a_L-\overline{\theta}_i(a)}{a_L-a}, \tag*{(7)}

with the convention κi(a1)=κi+(aL)=0\kappa_i(a_1)=\kappa_i^+(a_L)=0.

By (SA2), κi(a)∈[0,1]\kappa_i(a) \in[0,1]. The following theorem gives an exact condition for action preservation.

Theorem 1 (Action preservation): Let a∈Aa \in\mathcal{A} and ∅≠R⊆V\varnothing\ne\mathcal{R} \subseteq\mathcal{V}. Then (R,a)(\mathcal{R},a) is robust if and only if

wi(R)≥κi(a)for every i∈R.(8)w_i(\mathcal{R}) \ge\kappa_i(a) \quad\text{for every } i \in\mathcal{R}. \tag*{(8)}

Equivalently, Ui(R,a)⊆Ii(a)U_i(\mathcal{R},a) \subseteq\mathcal{I}_i(a) for every i∈Ri \in\mathcal{R}.

Proof: For each finite endpoint of Ii(a)\mathcal{I}_i(a), (6) gives the corresponding identity

u‾i(R,a)−θ‾i(a)=(a−a1)(wi(R)−κi−(a)),\underline{u}_i(\mathcal{R},a)-\underline{\theta}_i(a)=(a-a_1)(w_i(\mathcal{R})-\kappa_i^-(a)),
θ‾i(a)−u‾i(R,a)=(aL−a)(wi(R)−κi+(a)).\overline{\theta}_i(a)-\overline{u}_i(\mathcal{R},a)=(a_L-a)(w_i(\mathcal{R})-\kappa_i^+(a)).

The inclusion Ui(R,a)⊆Ii(a)U_i(\mathcal{R},a) \subseteq\mathcal{I}_i(a) holds if and only if wi(R)≥κi(a)w_i(\mathcal{R}) \ge\kappa_i(a), since infinite endpoints impose no restriction.

If xi∈Ii(a)x_i \in\mathcal{I}_i(a) and ui∈Ii(a)u_i \in\mathcal{I}_i(a), their convex combination in (3) remains in Ii(a)\mathcal{I}_i(a): since αi>0\alpha_i > 0, the next opinion stays strictly below θ‾i(a)\overline{\theta}_i(a) whenever that endpoint is finite.

Conversely, suppose that an input endpoint lies below θ‾i(a)\underline{\theta}_i(a). This input is attained when all outsiders hold a1a_1, and with xi=θ‾i(a)x_i = \underline{\theta}_i(a) the next opinion lies below Ii(a)\mathcal{I}_i(a). Similarly, if an input endpoint lies above θ‾i(a)\overline{\theta}_i(a), it is attained when all outsiders hold aLa_L, and with xi=θ‾i(a)−εx_i = \overline{\theta}_i(a)-\varepsilon for sufficiently small ε>0\varepsilon> 0 the next opinion lies above Ii(a)\mathcal{I}_i(a). In both cases, the other members can hold opinion aa, so x∈Xa(R)x \in\mathcal{X}_a(\mathcal{R}).

We call (8) the cohesion condition. As in [17], it requires members to place enough influence weight on the group, but the required weight depends on each member’s action rule.

Remark 1 (Binary CODA): With actions {−1,1}\{-1,1\} and a common threshold 00, κi(−1)=κi(1)=1/2\kappa_i(-1)=\kappa_i(1)=1/2. With equal weights on neighbors, the cohesion condition becomes ∣Ni∩R∣≥∣Ni∖R∣|\mathcal{N}_i \cap\mathcal{R}| \ge|\mathcal{N}_i \setminus\mathcal{R}| for every i∈Ri \in\mathcal{R}. This coincides with the criterion used for robust polarized clusters in [3], although the two models use different opinion updates and threshold tie rules.

Remark 2 (Robust action cluster within a candidate set): For a fixed action aa, unions of robust action clusters are robust by (8). Given a candidate set S⊆V\mathcal{S} \subseteq\mathcal{V}, repeatedly remove agents whose influence weight on the remaining set is below κi(a)\kappa_i(a). No member of a robust subset of S\mathcal{S} can be removed. At termination, the remaining set is therefore either empty or the unique largest robust action cluster contained in S\mathcal{S}. If none remain, S\mathcal{S} has no robust action cluster for aa.

Robust action clusters can also make disagreement persist. If disjoint groups R\mathcal{R} and R′\mathcal{R}' are robust action clusters for actions aa and bb, respectively, with a≠ba \ne b, every trajectory starting in Xa(R)∩Xb(R′)\mathcal{X}_a(\mathcal{R}) \cap\mathcal{X}_b(\mathcal{R}') preserves both actions, and such trajectories never reach action consensus.

We next ask when a robust action cluster makes other agents adopt its action permanently.

Guaranteed Adoption

Fix a∈Aa \in\mathcal{A} and let R0\mathcal{R}_0 be a robust action cluster for action aa, which we call a robust seed. A trajectory is admissible for this seed and target action aa if x(0)∈Xa(R0)x(0) \in\mathcal{X}_a(\mathcal{R}_0).

An agent ii adopts aa permanently if qi(xi(t))=aq_i(x_i(t))=a for all sufficiently large tt. A guaranteed adopter adopts aa permanently along every admissible trajectory. This guarantee must hold for all seed opinions consistent with action aa and all outsider opinions. An agent may therefore adopt aa from some initial states without being a guaranteed adopter. A trajectory reaches permanent aa-consensus if all agents hold aa from some time on.

We first give a sufficient condition for an outsider to be a guaranteed adopter. Iterating it yields a cascade of guaranteed adopters; when the cascade reaches every agent, every admissible trajectory reaches permanent aa-consensus.

An Adoption Condition

The cohesion condition of Theorem 1 concerns agents already holding action aa. To make an outsider adopt from any initial opinion, it suffices to keep its social input inside Ii(a)\mathcal{I}_i(a) with a positive margin from each finite endpoint. The following lemma expresses this requirement as the adoption condition (9), a strict version of the cohesion condition.

Lemma 3 (Adoption condition): Suppose (R,a)(\mathcal{R},a) is robust. If an outsider i∉Ri \notin\mathcal{R} satisfies

wi(R)>κi(a),(9)w_i(\mathcal{R}) > \kappa_i(a), \tag*{(9)}

or, equivalently, Ui(R,a)⊂Ii(a)U_i(\mathcal{R},a) \subset\mathcal{I}_i(a), then there is a finite TiT_i, independent of the initial state, such that every trajectory from Xa(R)\mathcal{X}_a(\mathcal{R}) satisfies qi(xi(t))=aq_i(x_i(t))=a for all t≥Tit \ge T_i.

Proof: The equivalence follows from (6) and (7). Let δi:=min⁡{u‾i(R,a)−θ‾i(a),θ‾i(a)−u‾i(R,a)}>0\delta_i:=\min\{\underline{u}_i(\mathcal{R},a)-\underline{\theta}_i(a),\overline{\theta}_i(a)-\overline{u}_i(\mathcal{R},a)\}>0, omitting the term of an infinite endpoint. By robustness, ui(t)∈Ui(R,a)u_i(t)\in U_i(\mathcal{R},a) for all tt; iterating (3) and using xi(0)∈[−1,1]x_i(0)\in[-1,1] gives ui(R,a)−2αit≤xi(t)≤u‾i(R,a)+2αitu_i(\mathcal{R},a)-2\alpha_i^t\le x_i(t)\le\overline{u}_i(\mathcal{R},a)+2\alpha_i^t. Since αi∈(0,1)\alpha_i\in(0,1), we may take the finite bound

Ti:=min⁡{t∈Z≥0:2αit<δi}.T_i:=\min\{t\in\mathbb{Z}_{\ge0}:2\alpha_i^t<\delta_i\}.

For t≥Tit\ge T_i, xi(t)x_i(t) lies strictly inside Ii(a)\mathcal{I}_i(a).

The strict inequality cannot, in general, be relaxed to equality: with a constant input equal to a finite endpoint of Ii(a)\mathcal{I}_i(a), an opinion starting beyond that endpoint converges to it without ever entering Ii(a).\mathcal{I}_i(a). The preservation and adoption conditions are independent of αi∈(0,1)\alpha_i\in(0,1), whereas the adoption-time bound depends on αi\alpha_i.

Cascade

An outsider that adopts an action permanently can help other outsiders meet the adoption condition. Iterating this condition yields a cascade of guaranteed adopters.

Definition 4 (Cascade): Starting from a robust seed R0\mathcal{R}_0, let Rk+1=Φa(Rk)\mathcal{R}_{k+1}=\Phi_a(\mathcal{R}_k) for k=0,1,…k=0,1,\ldots, where

Φa(R):=R∪{i∈V:wi(R)>κi(a)}.(10)\Phi_a(\mathcal{R}):=\mathcal{R}\cup\{i\in\mathcal{V}:w_i(\mathcal{R})>\kappa_i(a)\}. \tag*{(10)}

This increasing sequence stabilizes after at most N−∣R0∣N-|\mathcal{R}_0| strict enlargements; its limit is the cascade closure cl⁡a(R0)\operatorname{cl}_a(\mathcal{R}_0).

Each step of the cascade preserves robustness and guarantees permanent adoption for new members.

Proposition 1 (Cascade propagation). Let R0\mathcal{R}_0 be a robust seed for action aa. Every set Rk\mathcal{R}_k in the cascade sequence of Definition 4 is a robust action cluster for action aa, and every member of cl⁡a(R0)\operatorname{cl}_a(\mathcal{R}_0) is a guaranteed adopter of action aa.

Proof. We show by induction on kk that Rk\mathcal{R}_k is robust and that there is a finite time τk\tau_k such that every admissible trajectory satisfies x(t)∈Xa(Rk)x(t) \in X_a(\mathcal{R}_k) for all t≥τkt \ge\tau_k. For k=0k=0, both claims hold with τ0=0\tau_0=0. Assume they hold for kk. Each i∈Rk+1∖Rki \in\mathcal{R}_{k+1} \setminus\mathcal{R}_k satisfies wi(Rk)>κi(a)w_i(\mathcal{R}_k) > \kappa_i(a). Since the dynamics are time-invariant, Lemma 3 applied from time τk\tau_k gives qi(xi(t))=aq_i(x_i(t))=a for all t≥τk+Tit \ge\tau_k+T_i. The second claim therefore holds for k+1k+1, with τk+1:=τk+max⁡i∈Rk+1∖RkTi\tau_{k+1}:=\tau_k+\max_{i\in\mathcal{R}_{k+1}\setminus\mathcal{R}_k}T_i (or τk+1:=τk\tau_{k+1}:=\tau_k if no agent joins). Moreover, wi(Rk+1)≥wi(Rk)w_i(\mathcal{R}_{k+1}) \ge w_i(\mathcal{R}_k) for every ii, so every member of Rk+1\mathcal{R}_{k+1} satisfies (8): new members because wi(Rk)>κi(a)w_i(\mathcal{R}_k) > \kappa_i(a), and members of Rk\mathcal{R}_k by Theorem 1. Hence Rk+1\mathcal{R}_{k+1} is robust by Theorem 1. Applying the second claim to the limit cl⁡a(R0)\operatorname{cl}_a(\mathcal{R}_0) shows that each of its members is a guaranteed adopter. Thus cl⁡a(R0)=V\operatorname{cl}_a(\mathcal{R}_0)=\mathcal{V} is sufficient for permanent aa-consensus along every admissible trajectory.

The iteration in Definition 4 is a progressive linear-threshold cascade with thresholds κi(a)\kappa_i(a) and the strict adoption rule (9). As in [17], cohesive groups can block it. A nonempty group B\mathcal{B} is an obstruction if wi(B)≥1−κi(a)w_i(\mathcal{B}) \ge1-\kappa_i(a), or equivalently wi(V∖B)≤κi(a)w_i(\mathcal{V}\setminus\mathcal{B}) \le\kappa_i(a), for all i∈Bi \in\mathcal{B}. Then no member of B\mathcal{B} satisfies (9) for R=V∖B\mathcal{R}=\mathcal{V}\setminus\mathcal{B}, even when every other agent has adopted aa. The cascade reaches every agent if and only if every obstruction intersects the seed. An obstruction blocks only the cascade, not adoption itself: agents outside cl⁡a(R0)\operatorname{cl}_a(\mathcal{R}_0) may still adopt aa along particular trajectories. For L=2L=2, however, the cascade is exact. If bb denotes the other action, then κi(b)=1−κi(a)\kappa_i(b)=1-\kappa_i(a), so every obstruction B\mathcal{B} disjoint from the seed is a robust action cluster for bb, and its members keep bb along every admissible trajectory starting in Xb(B)X_b(\mathcal{B}). Since V∖cl⁡a(R0)\mathcal{V}\setminus\operatorname{cl}_a(\mathcal{R}_0) is such an obstruction whenever it is nonempty, cl⁡a(R0)\operatorname{cl}_a(\mathcal{R}_0) is then exactly the set of guaranteed adopters. For L>2L>2, the cascade gives only a sufficient condition for guaranteed adoption, since the worst-case inputs of different members of an obstruction need not be simultaneously realizable.

Numerical Illustrations

We illustrate Theorem 1 and the cascade of Definition 4 on two networks. Each example compares two runs: one with heterogeneous action rules, and one in which all agents use the homogeneous rule. Each threshold of this common rule is the mean, over all agents, of the corresponding heterogeneous thresholds. Within each example, the graph and initial opinions are identical in both runs. Table I lists the parameters of both examples, including the thresholds of every action rule. We say that the action profile remains fixed after time t0t_0 if no action changes after t0t_0. We verify this with the following criterion: every social input at time t0t_0 lies in the closure of the agent’s current action interval. Since αi>0\alpha_i>0, the update (3) then keeps every opinion in its action interval, so neither the actions nor the inputs change.

Rule (agents)Action thresholdsaaκi(a)\kappa_i(a)
Section VI-A: A={−1,−1/2,1/2,1}\mathcal{A}=\{-1,-1/2,1/2,1\}
A (8)(−0.70,−0.30,0.80)(-0.70,-0.30,0.80)1/21/20.47
B (3)(−0.70,0,0.80)(-0.70,0,0.80)1/21/20.67
Homogeneous (11)(−0.70,−0.218,0.80)(-0.70,-0.218,0.80)1/21/20.52
Section VI-B: A={−1,0,1}\mathcal{A}=\{-1,0,1\}
C (1250)(−0.60,0.60)(-0.60,0.60)0/±10/\pm10.40/0.800.40/0.80
D (1250)(−0.10,0.10)(-0.10,0.10)0/±10/\pm10.90/0.550.90/0.55
Homogeneous (2500)(−0.35,0.35)(-0.35,0.35)0/±10/\pm10.65/0.6750.65/0.675

Table I. Parameters of Section VI. Both examples use undirected graphs with uniform weights Wij=1/∣Ni∣W_{ij}=1/|{\cal N}_i| for j∈Nij\in{\cal N}_i and inertia αi=0.8\alpha_i=0.8. Each action rule is given by its thresholds (θi,1,…,θi,L−1)(\theta_{i,1},\ldots,\theta_{i,L-1}) and the influence threshold κi(a)\kappa_i(a) of (7) for target action aa.

Propagation of an Intermediate Action

The first example shows that mean thresholds alone do not determine whether an intermediate action propagates from a robust seed. The action set is A={−1,−1/2,1/2,1}\mathcal{A}=\{-1,-1/2,1/2,1\}, and the network of Figure 3a has eleven agents. Agents 1 to 4 form a complete graph and constitute the seed R0\mathcal{R}_0. The seed starts with action 1/21/2 and the seven outsiders with action −1-1. In the heterogeneous run, rules A and B are assigned as shown in Figure 3a, with both rules represented in the seed.

Initial networkAgents holding action 1/2

Figure 3. Propagation of action $1/2$ from a robust minority of four among eleven agents. (a) Initial network: shapes denote rules A and B in the heterogeneous run, colors give initial actions, and thick borders mark the seed R0={1,2,3,4}\mathcal{R}_0=\{1,2,3,4\}, which contains both rules. In the other run, every agent uses the homogeneous rule. (b) Number of agents currently holding action 1/21/2, from identical initial opinions. Under heterogeneous rules all eleven adopt permanently by t=30t=30; under the homogeneous rule only the four seed members hold action $1/2. Only 0≤t≤350\le t\le35 is displayed; both runs continue to T=180T=180.

Every seed member places influence weight 3/43/4 on R0\mathcal{R}_0, so its worst-case input interval Ui(R0,1/2)=[1/8,5/8]\mathcal{U}_i(\mathcal{R}_0,1/2)=[1/8,5/8], given by (6), lies strictly inside its action interval for action 1/21/2 under every rule. Hence R0\mathcal{R}_0 is a robust seed for action 1/21/2 in both runs, by Theorem 1. Under heterogeneous rules, outsider 5 places weight 1/21/2 on the seed, above rule A’s influence threshold of approximately 0.47, and joins the cascade first. Each later outsider places at least 1/21/2 on the growing group when it joins, and outsider 11, the only outsider using rule B, joins last with weight 1>0.671>0.67. Thus cl⁡1/2(R0)=V\operatorname{cl}_{1/2}(\mathcal{R}_0)=\mathcal{V}, and every agent is a guaranteed adopter of action 1/21/2. Under the homogeneous rule, the only outsiders with seed neighbors, agents 5 and 8, place weights $1/2 and 2/52/5 on the seed, both below κi(1/2)≈0.52\kappa_i(1/2) \approx0.52, so the cascade stops at R0\mathcal{R}_0.

Figure 3(b) shows the two runs. Under heterogeneous rules, all outsiders permanently adopt action 1/21/2. Agent 5 starts this propagation: its initial input −1/4-1/4 lies above rule A’s middle threshold −0.30-0.30 but below the mean middle threshold of approximately −0.218-0.218. Under the homogeneous rule, agents 5 and 8 move from −1-1 to −1/2-1/2 and keep inputs −1/4-1/4 and −2/5-2/5, below the mean middle threshold, while all other outsiders remain at −1-1. The resulting action profile remains fixed, with only the seed holding action 1/21/2. Indeed, every social input lies strictly inside its agent’s current action interval, so the criterion stated at the beginning of this section applies. The stalled cascade alone would show only that the cascade criterion does not certify adoption. The seed preserves its action in both runs, but replacing the individual rules by their mean stops the propagation.

Robust Action Clusters on a Square Lattice

The second example asks which groups in a fixed action profile are robust action clusters. A fixed action profile does not establish robustness. Theorem 1 identifies, from the final action profile alone, the groups that would keep their actions whatever the other agents do. With two actions, every agent of a fixed action profile belongs to such a group. Indeed, let R\mathcal{R} be the set of agents holding aLa_L. Every other agent holds a1a_1, so each member of R\mathcal{R} receives the constant input ui(R,aL)u_i(\mathcal{R},a_L) of (6) and keeps aLa_L as its opinion approaches that input. This input therefore lies in the closure of the action interval, and (R,aL)(\mathcal{R},a_L) is robust by Theorem 1. The same holds for a1a_1. This is the situation of the binary lattice in [3]. We therefore use three actions, the smallest action set in which a fixed action profile need not consist of robust action clusters.

Following [3], we use an open 50×5050 \times50 square lattice with horizontal and vertical edges and actions A={−1,0,1}\mathcal{A} = \{-1,0,1\}. Initial opinions are independent uniform draws from [−1,1][-1,1]. In one run, every agent uses the homogeneous rule; in the other, rules C and D are assigned at random to equal

numbers of agents (Table I). On the lattice, an interior agent, away from the boundary, has four neighbors, so an influence threshold translates into a number of neighbors. Under the homogeneous rule, κi(0)=0.65\kappa_i(0) = 0.65 and κi(±1)=0.675\kappa_i(\pm1) = 0.675, so an interior agent needs three of its four neighbors in its group, whatever action it holds. Under rule C, κi(0)=0.4\kappa_i(0) = 0.4, so an interior agent holding action 0 needs only two; under rule D, κi(0)=0.9\kappa_i(0) = 0.9, so it needs all four.

In any group disjoint from the lattice boundary, the left-most member of the top row has its upper and left neighbors outside the group, so it places influence weight at most 1/21/2 on the group. In Table I, all influence thresholds for actions ±1\pm1 exceed 1/21/2, and only rule C has κi(0)≤1/2\kappa_i(0) \leq1/2. By Theorem 1, no group disjoint from the boundary is a robust action cluster under the homogeneous rule. Under heterogeneous rules, among groups disjoint from the boundary, only those holding action 0 can be robust. Figure 4 shows both consequences. Under the homogeneous rule, the final action profile remains fixed but contains no robust action cluster, including at the boundary (Figure 4(a)). The absence of robust clusters is also observed in 15 additional initializations.

Figure 4. Robust action clusters on an open (non-periodic) 50 × 50 lattice with actions {−1,0,1}\{-1,0,1\}, from the same uniform random initial opinions; both final action profiles remain fixed. (a) Final actions qi(xi(T))q_i(x_i(T)), at T=1200T = 1200, when every agent uses the homogeneous rule: the final action profile contains no robust action cluster. (b) Final actions under heterogeneous rules: half of the agents, placed at random, use rule C and the others rule D (Table I); crosses mark the agents in robust action clusters, all holding action 0. (c) Final private opinions xi(T)x_i(T) under heterogeneous rules.

Under heterogeneous rules, the final action profile also remains fixed. Applying the procedure of Remark 2 to each final action set leaves no agents for actions ±1\pm1 and 701 agents for action 0 (Figure 4(b)); we split them into connected robust action clusters.

The other remaining agents also keep their actions, but none belongs to a robust action cluster. Their actions remain fixed only because of the actions currently held around them. For each connected group of these agents holding a common action, we start from the final state and fix all outsider actions to −1-1 in one run and to 11 in another. At least one run causes an agent in the group to change its action. Every robust action cluster keeps action 0 in both runs, as Theorem 1 predicts. Private opinions do not reveal this distinction. Agents holding action 0 have opinions ranging over [−0.5,0.5][-0.5,0.5] whether or not they belong to a robust action cluster (Figure 4(b)–(c)). As in Section VI-A, the homogeneous rule does not reproduce the behavior under heterogeneous rules. Here, it excludes the robust action clusters that rule C makes possible.

Conclusion

We characterized robust action clusters in networks with continuous private opinions and heterogeneous action rules. A group preserves a common action regardless of outsiders’ actions if and only if each member satisfies an action-dependent cohesion condition. A strict version yields a cascade of guaranteed adopters, which identifies all guaranteed adopters in the binary case. Numerical examples show that replacing heterogeneous thresholds by their means can alter both action propagation and the presence of robust clusters. Future work includes addressing seed formation, adoption through intermediate actions beyond the cascade, and asynchronous updates.

References

  1. [1]M. H. DeGroot, “Reaching a consensus,” Journal of the American Statistical Association, vol. 69, no. 345, pp. 118–121, 1974.DOI
  2. [2]A. C. R. Martins, “Continuous opinions and discrete actions in opinion dynamics problems,” International Journal of Modern Physics C, vol. 19, no. 4, pp. 617–624, 2008.
  3. [3]N. R. Chowdhury, I.-C. Morărescu, S. Martin, and S. Srikant, “Continuous opinions and discrete actions in social networks: A multi-agent system approach,” in Proc. 55th IEEE Conference on Decision and Control (CDC), 2016, pp. 1739–1744.
  4. [4]M. Granovetter, “Threshold models of collective behavior,” American Journal of Sociology, vol. 83, no. 6, pp. 1420–1443, 1978.DOI
  5. [5]D. J. Watts, “A simple model of global cascades on random networks,” Proceedings of the National Academy of Sciences, vol. 99, no. 9, pp. 5766–5771, 2002.
  6. [6]W. S. Rossi, G. Como, and F. Fagnani, “Threshold models of cascades in large-scale networks,” IEEE Transactions on Network Science and Engineering, vol. 6, no. 2, pp. 158–172, 2019.
  7. [7]V. S. Varma and I.-C. Morărescu, “Modeling stochastic dynamics of agents with multi-leveled opinions and binary actions,” in Proc. 56th IEEE Conference on Decision and Control (CDC), 2017, pp. 1064–1069.DOI
  8. [8]F. Ceragioli and P. Frasca, “Consensus and disagreement: The role of quantized behaviors in opinion dynamics,” SIAM Journal on Control and Optimization, vol. 56, no. 2, pp. 1058–1080, 2018.DOI
  9. [9]F. Ceragioli, P. Frasca, and R. Prisant, “A continuous opinions discrete actions model on line graphs,” Journal of Optimization Theory and Applications, vol. 206, no. 2, 2025, art. no. 47.DOI
  10. [10]A. Couthures, T. Mongaillard, V. S. Varma, S. Lasaule, and I.-C. Morărescu, “Analysis of a continuous opinion and discrete action model coupled with an external dynamics,” in Proc. European Control Conference (ECC), 2024, pp. 3392–3397.DOI
  11. [11]M. Ye, Y. Qin, A. Govaert, B. D. O. Anderson, and M. Cao, “An influence network model to study discrepancies in expressed and private opinions,” Automatica, vol. 107, pp. 371–381, 2019.DOI
  12. [12]H. D. Aghbolagh, M. Ye, L. Zino, Z. Chen, and M. Cao, “Coevolutionary dynamics of actions and opinions in social networks,” IEEE Transactions on Automatic Control, vol. 68, no. 12, pp. 7708–7723, 2023.
  13. [13]H. Munif, A. Couthures, V. S. Varma, S. Lasaule, and T. Başar, “Network-induced strategic communication in opinion dynamics,” arXiv preprint arXiv:2607.16036, 2026.arxiv.org/abs/2607.16036
  14. [14]D. Urbig, “Attitude dynamics with limited verbalisation capabilities,” Journal of Artificial Societies and Social Simulation, vol. 6, no. 1, 2003. [Online]. Available: https://www.jasss.org/6/1/2.html
  15. [15]A. C. R. Martins, “A middle option for choices in the continuous opinions and discrete actions model,” Advances and Applications in Statistical Sciences, vol. 2, no. 2, pp. 333–345, 2010, preprint arXiv:0904.1882.arxiv.org/abs/0904.1882
  16. [16]S. Melnik, J. A. Ward, J. P. Gleeson, and M. A. Porter, “Multistage complex contagions,” Chaos: An Interdisciplinary Journal of Nonlinear Science, vol. 23, no. 1, 2013, art. no. 013124.
  17. [17]S. Morris, “Contagion,” The Review of Economic Studies, vol. 67, no. 1, pp. 57–78, 2000.DOI
  18. [18]L. Arditti, G. Como, F. Fagnani, and M. Vanelli, “Robust coordination of linear threshold dynamics on directed weighted networks,” IEEE Transactions on Automatic Control, vol. 69, no. 10, pp. 6515–6529, 2024.arxiv.org/abs/2109.12685
  19. [19]E. Goles and J. Olivos, “Periodic behaviour of generalized threshold functions,” Discrete Mathematics, vol. 30, no. 2, pp. 187–189, 1980.

Paper details

vlnc:W_apedujxh
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents