Hegselmann–Krause Model

• Hegselmann–Krause Model is a bounded-confidence opinion dynamics system where agents update opinions by averaging with nearby agents within a specified threshold.
• The model preserves order in one-dimensional cases and leads to consensus or multiple clusters depending on the confidence range and initial opinion distribution.
• Extensions incorporate noise, asynchronous updates, and multidimensional variants, offering detailed insights into convergence rates and system robustness.

The Hegselmann–Krause (HK) model is a finite-agent opinion dynamics model in which each agent updates its opinion by averaging with agents whose opinions differ by at most a confidence threshold. In the standard synchronous formulation, there are $n$ agents, agent $i$ holds an opinion $x_i(t)\in\mathbb{R}^d$ at discrete time $t$, and all agents update simultaneously. The model is a canonical bounded-confidence system: interaction is endogenous, because the communication graph is determined by the current opinion configuration rather than fixed a priori. Across the literature, the HK framework has been studied in scalar and vector opinion spaces, on complete and sparse social graphs, under synchronous and asynchronous schedules, with stubbornness, noise, and mean-field limits, and with primary emphasis on consensus, clustering, and convergence time (Bhattacharyya et al., 2012, Li, 2024).

1. Canonical formulation

In the classical homogeneous HK model, the neighbor set of agent $i$ at time $t$ is

$$N_i(t)=\{j\in[n]: \|x_i(t)-x_j(t)\|\le \varepsilon\},$$

where $\varepsilon>0$ is the confidence threshold. The synchronous update rule is

$$x_i(t+1)=\frac{1}{|N_i(t)|}\sum_{j\in N_i(t)}x_j(t).$$

For scalar opinions on $[0,1]$, the same rule is commonly written with absolute values instead of a norm. In the homogeneous model the interaction graph is undirected, because if $i$0 is a neighbor of $i$1, then $i$2 is a neighbor of $i$3 (Bhattacharyya et al., 2012, Li, 2024).

A graph-theoretic representation is central. The opinion graph at time $i$4 has vertex set $i$5 and edge set

$i$6

Consensus means that all agents eventually share the same opinion, equivalently that the dynamics converge to a single common value. In one-dimensional formulations, ordered opinions $i$7 are especially useful, because connectivity can be reduced to checking whether all consecutive ordered agents are connected (Li, 2024).

The model also admits social-network variants in which influence requires both social adjacency and confidence compatibility. In that setting, the influence neighborhood is determined by a fixed graph $i$8 together with the bounded-confidence condition, so the active influence network is a time-varying subgraph of the social graph (Berenbrink et al., 2022, Bhattacharyya et al., 2015).

2. Consensus, clustering, and endogenous topology

The basic deterministic phenomenology is a competition between local averaging and topological fragmentation. For large enough confidence bounds, the whole population is connected through bounded-confidence averaging and collapses into one cluster. For smaller $i$9, the opinion space breaks into multiple disconnected interaction components, producing several clusters in the absorbing state (Slanina, 2014).

In one dimension, the order of opinions is preserved over time: if $x_i(t)\in\mathbb{R}^d$0, then $x_i(t)\in\mathbb{R}^d$1. This order-preserving structure yields a decomposability property: if two consecutive agents are separated by more than the confidence radius, they remain separated, and the system splits into independent subsystems. The same one-dimensional structure underlies the statement that if the initial opinion graph is disconnected, consensus cannot be achieved, and more generally that consensus can be achieved if and only if the opinion graph remains connected over time (Bhattacharyya et al., 2012, Li, 2024).

The stationary states of the scalar HK model are clustered absorbing configurations. A $x_i(t)\in\mathbb{R}^d$2-cluster absorbing state is described by values

$x_i(t)\in\mathbb{R}^d$3

with every agent belonging to exactly one cluster. The smallest time at which the system becomes absorbing is often called the consensus time or absorbing time, depending on whether the absorbing state has one cluster or several (Slanina, 2014).

Near the first transition between one-cluster and two-cluster regimes, a characteristic three-cluster mechanism appears: a left wing cluster, a right wing cluster, and a small central mediator cluster near $x_i(t)\in\mathbb{R}^d$4. The mediator group can be very small but still determine whether eventual consensus occurs, and the associated absorbing-time histograms develop multiple peaks corresponding to mediator groups of size $x_i(t)\in\mathbb{R}^d$5 (Slanina, 2014). This suggests that in the HK model consensus is not only a static connectivity property but also a dynamical process controlled by transient mesoscopic structure.

3. Convergence and freezing-time theory

A major line of research concerns how long HK dynamics can remain active before freezing. For the homogeneous HK system in $x_i(t)\in\mathbb{R}^d$6 with $x_i(t)\in\mathbb{R}^d$7 agents, the first polynomial-time convergence guarantee in arbitrary dimensions gave an explicit bound of $x_i(t)\in\mathbb{R}^d$8. In one dimension, the same work established an $x_i(t)\in\mathbb{R}^d$9 upper bound and a quadratic lower bound $t$0 in two dimensions (Bhattacharyya et al., 2012).

Subsequent work removed the ambient-dimension dependence from the synchronous termination bound. Using a Lyapunov function

$t$1

together with spectral estimates for the communication graph, a dimension-independent upper bound

$t$2

was derived for synchronous homogeneous HK dynamics (Etesami et al., 2014).

The currently strongest freezing-time estimate in the finite-dimensional homogeneous model is a dimension-uniform quartic upper bound. With confidence radius normalized to $t$3, the maximal freezing time $t$4 satisfies

$t$5

The proof introduces the active part of the energy,

$t$6

and an improved decrement inequality

$t$7

where $t$8 is the largest nontrivial eigenvalue of the transition matrix $t$9. Combined with graph-diameter bounds, this yields the quartic estimate and improves the previous $i$0 bound in dimensions $i$1 (Martinsson, 2015).

Lower bounds remain quadratic. Configurations on a circle and one-dimensional dumbbell constructions produce $i$2 freezing-time behavior, so the general theory still leaves a gap between $i$3 and $i$4 (Martinsson, 2015, Bhattacharyya et al., 2012).

4. Consensus probability and critical behavior

When initial opinions are random, the central question becomes the probability of consensus rather than deterministic convergence from a fixed profile. For synchronous HK dynamics in $i$5 with i.i.d. initial opinions supported on a convex set of positive Lebesgue measure, the probability of consensus is strictly positive. A general lower bound is

$i$6

and for $i$7 in one dimension,

$i$8

(Li, 2021).

A distinct graph-uniform lower bound holds for a multivariate spatial HK model on an arbitrary finite connected graph $i$9, where opinions lie in a bounded convex set $t$0. If $t$1 is the radius of $t$2 and $t$3, then

$t$4

where $t$5 is a center of $t$6. For the original one-dimensional model on $t$7, this yields

$t$8

(Lanchier et al., 2021).

Critical-threshold statements require careful qualification. Numerical and simulation-based work on the noiseless scalar model with i.i.d. uniform initial opinions on $t$9 identifies a first transition between one-cluster and two-cluster regimes near

$$N_i(t)=\{j\in[n]: \|x_i(t)-x_j(t)\|\le \varepsilon\},$$0

with a second transition near $$N_i(t)=\{j\in[n]: \|x_i(t)-x_j(t)\|\le \varepsilon\},$$1 (Slanina, 2014). By contrast, a recent asymptotic theorem for i.i.d. uniform initial opinions on $$N_i(t)=\{j\in[n]: \|x_i(t)-x_j(t)\|\le \varepsilon\},$$2 proves

$$N_i(t)=\{j\in[n]: \|x_i(t)-x_j(t)\|\le \varepsilon\},$$3

(Li, 2024). A common source of confusion is that these results address different critical objects: one concerns dynamical phase transitions between cluster-number regimes in finite-size simulations, while the other concerns an asymptotic consensus-probability statement for uniformly random initial data.

The proof of the $$N_i(t)=\{j\in[n]: \|x_i(t)-x_j(t)\|\le \varepsilon\},$$4 result is based on order preservation, persistence of graph disconnection, connectivity via consecutive ordered agents, and a neighborhood-overlap criterion ensuring edge persistence. For $$N_i(t)=\{j\in[n]: \|x_i(t)-x_j(t)\|\le \varepsilon\},$$5, the initial opinion graph is connected with probability tending to $$N_i(t)=\{j\in[n]: \|x_i(t)-x_j(t)\|\le \varepsilon\},$$6, and the overlap counts are controlled by a law-of-large-numbers argument (Li, 2024).

5. Noise, robustness, synchronization, and fragmentation

Noisy HK dynamics are model-dependent, and different noise mechanisms produce different macroscopic conclusions. In a model with random opinion jumps occurring with probability $$N_i(t)=\{j\in[n]: \|x_i(t)-x_j(t)\|\le \varepsilon\},$$7, two noise scenarios have been analyzed: unlimited jumps across the whole opinion space and bounded jumps inside a small interval around the current opinion. For unlimited jumps, the growth rate of perturbations is

$$N_i(t)=\{j\in[n]: \|x_i(t)-x_j(t)\|\le \varepsilon\},$$8

and the homogeneous state becomes stable for

$$N_i(t)=\{j\in[n]: \|x_i(t)-x_j(t)\|\le \varepsilon\},$$9

independent of $\varepsilon>0$0. For bounded jumps, an approximate critical line is

$\varepsilon>0$1

(Pineda et al., 2013).

A different conclusion arises in additive bounded-noise models on $\varepsilon>0$2 with truncation at the boundaries. In the standard noisy HK model, if

$\varepsilon>0$3

then almost surely the system reaches $\varepsilon>0$4-consensus in finite time. The same work shows that homogeneous prejudice and homogeneous stubbornness lose their fragmentation ability under arbitrarily small noise, whereas heterogeneous prejudices can preserve robust cleavage. In particular, if two prejudice values satisfy

$\varepsilon>0$5

then each prejudice group almost surely reaches $\varepsilon>0$6-consensus with its own prejudice value (Su et al., 2017).

On the full real line, noise-induced synchronization has a sharp amplitude criterion. For i.i.d. noise, if the noise support is contained in an interval of width $\varepsilon>0$7, then the noisy HK system reaches quasi-synchronization in finite time almost surely; if no interval of width $\varepsilon>0$8 contains the noise almost surely, then quasi-synchronization cannot be sustained (Su et al., 2017). In heterogeneous HK dynamics with environment noise, the critical amplitude for quasi-synchronization is

$\varepsilon>0$9

where $$x_i(t+1)=\frac{1}{|N_i(t)|}\sum_{j\in N_i(t)}x_j(t).$$0, and with communication noise the heterogeneous case can behave qualitatively differently from the homogeneous case: raising confidence thresholds of constituent agents may break quasi-synchronization (Chen et al., 2019).

These results do not contradict one another. They show that “noise” in HK theory is not a single perturbative regime but a family of mechanisms—random jumps, additive environment noise, communication noise, bounded versus full-space perturbations—whose effects on clustering and consensus are materially different.

6. Asynchronous, social, and mixed HK dynamics

The classical synchronous and asynchronous models can be embedded in a broader mixed HK framework with time-varying stubbornness. In the mixed model,

$$x_i(t+1)=\frac{1}{|N_i(t)|}\sum_{j\in N_i(t)}x_j(t).$$1

where $$x_i(t+1)=\frac{1}{|N_i(t)|}\sum_{j\in N_i(t)}x_j(t).$$2 means absolute stubbornness and $$x_i(t+1)=\frac{1}{|N_i(t)|}\sum_{j\in N_i(t)}x_j(t).$$3 recovers the standard HK update. Synchronous HK is the special case $$x_i(t+1)=\frac{1}{|N_i(t)|}\sum_{j\in N_i(t)}x_j(t).$$4, while asynchronous HK corresponds to a schedule in which all but one agent are absolutely stubborn at each step (Li, 2020).

Several structural properties of synchronous HK fail in the mixed model. Finite-time convergence can fail; two agents can merge at one time and later separate again; and $$x_i(t+1)=\frac{1}{|N_i(t)|}\sum_{j\in N_i(t)}x_j(t).$$5-equilibria may fail to exist for some trajectories (Li, 2020). Nonetheless, if

$$x_i(t+1)=\frac{1}{|N_i(t)|}\sum_{j\in N_i(t)}x_j(t).$$6

then every connected component becomes $$x_i(t+1)=\frac{1}{|N_i(t)|}\sum_{j\in N_i(t)}x_j(t).$$7-trivial in finite time for every $$x_i(t+1)=\frac{1}{|N_i(t)|}\sum_{j\in N_i(t)}x_j(t).$$8, and under the stronger uniform bound

$$x_i(t+1)=\frac{1}{|N_i(t)|}\sum_{j\in N_i(t)}x_j(t).$$9

the mixed dynamics are asymptotically stable (Li, 2020). A nondeterministic extension with random stubbornness and time-varying social graphs proves analogous almost-sure asymptotic stability results, and under additional connectivity assumptions yields consensus; the same framework contains both HK and Deffuant dynamics as special cases (Li, 2021).

For asynchronous social-network HK systems, where one uniformly random agent updates on a fixed graph $[0,1]$0, convergence can be quantified through a potential

$[0,1]$1

The expected number of updates until a $[0,1]$2-stable state is reached is

$[0,1]$3

and for the complete social network the bound improves to

$[0,1]$4

(Berenbrink et al., 2022). Related social-HK work proves that, for arbitrary initial positions and arbitrary social graph, the number of $[0,1]$5-non-trivial steps is

$[0,1]$6

and that a friendly time-varying social graph preserves the same polynomial bound (Bhattacharyya et al., 2015).

7. Multidimensional, mean-field, memory-driven, and domain-specific extensions

Multi-topic and multidimensional generalizations alter the definition of compatibility. In the average-based model, agent $[0,1]$7 compares only the average of its topic opinions with those of other agents, and the average vector evolves according to a scalar HK dynamics. Consensus or clustering of the full vector process occurs if and only if the average-opinion process does so. In the uniform affinity model, agents interact only if they are close on every topic, using an $[0,1]$8-type condition; the global opinion range is nonincreasing, and topic-wise order preservation holds under explicit sufficient conditions (Pasquale et al., 2022).

A continuous-time stochastic formulation with idiosyncratic and environmental noises leads, in the mean-field limit, to a McKean–Vlasov stochastic differential equation with common noise and to a nonlinear stochastic Fokker–Planck equation. The interaction kernel

$[0,1]$9

is non-Lipschitz because of the sharp confidence cutoff, so the analysis proceeds via regularized kernels $i$00, propagation of chaos for the regularized particle system, and then passage to the original limit (Chen et al., 2023).

Other extensions modify the update logic itself. A smart-agent model on networks distinguishes general agents, who follow HK averaging, from smart agents, who adopt the average opinion of the highest-scoring compatible neighbors. In the thermodynamic limit for homogeneous agent types, the critical threshold can take only one of two values depending on the behavior of the average degree $i$01: for the original HK model on $i$02, $i$03 if $i$04 stays finite and $i$05 if $i$06 (Zhu et al., 2021). A fractional-order extension replaces the memoryless update by a history-dependent rule with Grünwald–Letnikov-type coefficients $i$07, proving order preservation, asymptotic convergence, and consensus when the initial spread is within the confidence threshold (Jiang et al., 5 Jun 2025).

Domain-specific modifications include a voter–party model in which voters are influenced by nearby voters and parties, while parties are attracted by voters and repelled by other parties. In the deterministic case, a sufficient condition for unanimous consensus is

$i$08

and in the noisy mean-field analysis the critical voter-noise threshold in one dimension is approximated by

$i$09

(Cahill et al., 2024).

Taken together, these extensions indicate that the HK model is best understood not as a single dynamical system but as a bounded-confidence paradigm. Its invariant core is local averaging under endogenous compatibility; its research frontier lies in how that mechanism interacts with topology, stochasticity, heterogeneity, memory, and additional agent classes.