---
title: Hegselmann-Krause Opinion Dynamics
url: https://www.emergentmind.com/topics/hegselmann-krause-opinion-dynamics
type: topic
---

# Hegselmann-Krause Opinion Dynamics

Hegselmann–Krause (HK) opinion dynamics is a bounded-confidence model for discrete-time opinion formation in which each agent replaces its current opinion by the average of the opinions lying within a prescribed confidence radius. In its classical form, the model is deterministic, synchronous, and state-dependent: the interaction graph is recomputed from the current opinion profile at every step, so the averaging operator changes with the state itself. This endogeneity makes the HK system a canonical object in the theory of nonlinear multiagent dynamics, where consensus, fragmentation, finite-time stabilization, stochastic perturbations, heterogeneity, and control can all be studied within a single formal framework [1401.4324].

## 1. Canonical bounded-confidence dynamics

In the standard one-dimensional synchronous HK model, a population of agents \(i=1,\dots,n\) holds scalar opinions \(x_i(t)\in[0,1]\) or \(x_i(t)\in\mathbb R\) at discrete time \(t\), and a confidence bound \(\epsilon>0\) or \(r>0\) specifies who influences whom. The confidence neighborhood of agent \(i\) is
\[
N_i(t)=\{\,j: |x_j(t)-x_i(t)|\le \epsilon\,\},
\]
and the update rule is
\[
x_i(t+1)=\frac{1}{|N_i(t)|}\sum_{j\in N_i(t)} x_j(t).
\]
The same structure extends to \(\mathbb R^d\) by replacing absolute value with a norm, typically \(\|\cdot\|_2\) or \(\|\cdot\|_\infty\), and by averaging vector opinions componentwise [1401.4324; 1412.6546; 2204.08515].

A useful equivalent description is graph-theoretic. At each time \(t\), the current opinion profile induces a confidence graph, influence graph, or influence network whose edges connect agents at distance at most \(\epsilon\). The HK update is then row-stochastic averaging over the current closed neighborhood. In social variants, the influence graph is the intersection of a fixed or time-varying social graph with the confidence graph, so that bounded confidence and exogenous network structure jointly determine interaction [2201.12923; 2112.13348].

For finite populations, the classical synchronous model reaches an absorbing configuration in which opinions partition into clusters: every agent in a cluster shares the same limiting value, and distinct cluster values are separated by more than the confidence bound. In the terminology of controlled HK systems, convergence occurs when every non-strategic agent is frozen, equivalently when for every pair \(i,j\) either \(x_i(T)=x_j(T)\) or \(|x_i(T)-x_j(T)|>r\) [1401.4324; 1411.4814].

## 2. Convergence, clustering, and phase structure

Finite-time stabilization is a basic feature of the synchronous classical model. For finite \(N\), the system reaches an absorbing configuration after finitely many steps, and the stationary state consists of one or more clusters separated by gaps larger than \(\epsilon\) [1401.4324]. In arbitrary finite dimension, Etesami and Başar established a dimension-free polynomial upper bound on the worst-case termination time: if \(T_n\) denotes the maximal number of synchronous HK steps until no agent moves, then
\[
T_n \le n^8+n,
\]
independently of the ambient dimension \(d\) [1412.6546].

The same literature shows that termination-time bounds remain far from sharp in several regimes. For the pure one-dimensional HK model, the controlled-dynamics survey records an \(O(n^3)\) upper bound and an \(\Omega(n^2)\) lower bound, while in dimension at least two it records an \(O(n^3)\) upper bound and a \(2^{\Omega(n^2)}\) lower bound [1411.4814]. This gap is one of the central unresolved quantitative questions in the theory.

The long-run number of clusters depends strongly on the confidence radius. Numerical work on dynamical phase transitions reports well-defined plateaus in the average number of final clusters \(\langle \nu\rangle\) as a function of \(\epsilon\), with sharp steps that steepen as \(N\to\infty\). The first transitions are estimated at approximately \(\epsilon_{c1}\approx 0.20\) for the transition from one to two clusters and \(\epsilon_{c2}\approx 0.14\) for the transition from two to three clusters [1401.4324].

A central mechanism behind the slowdown near these transitions is the appearance of mediator groups. Near \(\epsilon_{c1}\), sample trajectories often transiently form two large “wing” clusters and a very small central mediator cluster. The mediator cluster pulls the wings inward extremely slowly, so the convergence time \(\tau\) becomes large when the mediator population is small. Histograms of \(\tau\) show several peaks, and these peaks are interpreted as trajectories with mediator groups of size \(1,2,3,\dots\) [1401.4324]. This supports the interpretation of HK clustering transitions as dynamical phase transitions accompanied by critical slowing down.

## 3. Asynchronous, social, and mixed variants

The asynchronous HK model updates only one agent per step. In the social asynchronous model of Berenbrink et al., the agent chosen uniformly at random averages over its current influence neighborhood in a fixed undirected social network \(G=(V,E)\), where an edge is active only if its endpoints are within \(\epsilon\) in opinion space. Because exact freezing need not occur, the relevant stability notion is \(\delta\)-stability: every active edge has length at most \(\delta\). For arbitrary social graphs, the expected number of activations until a \(\delta\)-stable state is reached is
\[
O\!\Bigl(\frac{\Phi(x(0))\,n\,|E|}{\delta^2}\Bigr)\le O\!\Bigl(n|E|^2(\epsilon/\delta)^2\Bigr),
\]
and for the complete graph \(K_n\) the bound improves to
\[
O\!\Bigl(n^3(n^2+(\epsilon/\delta)^2)\Bigr),
\]
which improves the earlier complete-graph bound \(O(n^9(\epsilon/\delta)^2)\) [2201.12923; 1412.6546].

A complementary line of work studies friend-graph and non-deterministic variants. For the social HK system on a fixed undirected graph, the number of \(\varepsilon\)-non-trivial steps is \(O(n^5/\varepsilon^2)\), and the same asymptotic bound remains valid for friendly time-varying graphs. In a one-dimensional non-deterministic HK model with adversarial perturbations of amplitude \(\delta<1/(4n^2)\), the first time after which all agents remain forever in intervals of length at most \(\rho\) is
\[
O\!\bigl(n^4+\tfrac{\ln(1/\rho)}{\ln n}\bigr)
\]
for every initial condition and every admissible perturbation sequence [1503.01720].

The mixed HK model introduces time-varying stubbornness or openness parameters \(\alpha_i(t)\in[0,1]\), so that an agent blends its current opinion with the local average rather than replacing it outright. In this regime, finite-time convergence can fail: already for two agents at distance \(\epsilon\) with constant \(\alpha_1(t)=\alpha_2(t)=1/2\), the two opinions approach each other exponentially without ever coinciding in finite time [2010.03050]. Nevertheless, under bounded stubbornness one still obtains strong asymptotic structure. If \(\sup_{t,i}\alpha_i(t)<1\), then for every \(\delta>0\) there exists a finite time after which every connected component of the opinion graph has diameter at most \(\delta\), and each component then contracts asymptotically to a consensus point [2010.03050].

The higher-dimensional mixed model with time-varying social relationship extends this further by allowing random update sets and a profile graph \(\mathcal G(t)\) obtained by intersecting the social-update graph with the confidence graph. Under lower and upper bounds keeping all nontrivial \(\alpha_i(t)\) away from \(0\) and \(1\), every component of \(\mathcal G(t)\) becomes \(\delta\)-trivial in finite time almost surely for every \(\delta>0\), yielding asymptotic stability. By appropriate specialization of the update graph and the \(\alpha_i(t)\), this framework covers not only HK dynamics but also the Deffuant–Weisbuch pairwise model [2112.13348].

## 4. Heterogeneity, truth-seeking, and strategic control

A major generalization assigns agent-specific confidence bounds \(r_i\). In the heterogeneous synchronous model,
\[
\mathcal N_i(x(t))=\{\,j: |x_j(t)-x_i(t)|\le r_i\,\},\qquad
x_i(t+1)=\frac{1}{|\mathcal N_i(x(t))|}\sum_{j\in\mathcal N_i(x(t))}x_j(t).
\]
General convergence remains difficult, but a partial convergence theorem shows that there exist two nonempty disjoint sets of agents that freeze in finite time at extremal values \(c\le c'\). If \(r=\min_i r_i\) and \(c'-c\le 2r\), then the entire system merges at the next step; if \(c'-c>2r\), all remaining agents stay forever in the interval \((c+r,c'-r)\). Two immediate sufficient conditions for full convergence are \(x_{\max}(0)-x_{\min}(0)\le 2r\) and \(r\ge 1/2\) [1603.08099].

Topological classification provides a finer structural language for heterogeneous HK systems. Mirtabatabaei and Bullo distinguish closed-minded components, moderate-minded components, and open-minded weakly connected components in the directed proximity graph induced by heterogeneous radii. They conjecture that every trajectory eventually reaches constant topology. Under fixed topology, they define leader groups whose spectral radii determine the followers’ rate and direction of convergence, and they provide sufficient conditions guaranteeing convergence and monotonicity under invariant equi-topology neighborhoods [1010.2186].

Truth-seeking variants add an exogenous truth value \(h\in[0,1]\). In the symmetric bounded-confidence model with truth seekers, agent \(i\) places weight \(\alpha_i(t)\in\{0\}\cup[\alpha,1]\) on \(h\) and distributes the remaining weight across neighbors using coefficients \(\beta_{ij}(t)\in[\beta,1-\beta]\). Under finite \(n\), symmetric confidence intervals, and uniform lower bounds \(\alpha>0\) and \(\beta>0\), all truth seekers converge to the truth. Kurz and Rambau describe this as a proof of the generalized Hegselmann–Krause conjecture [1401.4383].

Control-theoretic extensions treat selected agents as strategic inputs. In the controlled discrete HK model, a set of strategic agents can place their opinions arbitrarily at each time step, while non-strategic agents still average over all neighbors, including the strategic ones. The objective is to minimize the first time \(T\) at which all non-strategic agents are frozen. If \(f(n,m)\) denotes the worst-case optimal convergence time with \(n\) non-strategic agents and \(m\) strategic agents, then with one strategic agent
\[
\Omega(n^{2/3})\le f(n,1)\le O(n^2),
\]
whereas for \(m=\lfloor n^a\rfloor\), \(a\in[0,1]\), one has the asymptotically matching tradeoff
\[
f(n,\lfloor n^a\rfloor)=\Theta(n^{2-2a})
\]
up to constant factors. At the high-control end, \(m=9n\) strategic agents suffice to guarantee convergence in at most two steps [1411.4814].

## 5. Noise, stochastic perturbations, and continuum limits

Several noisy HK models replace exact averaging by averaging plus random perturbation. In the bounded additive-noise model,
\[
x_i^*(t)=\frac{1}{|N_i(t)|}\sum_{j\in N_i(t)}x_j(t)+\xi_i(t+1),
\]
followed by clipping to \([0,1]\). The relevant notion is quasi-consensus: if \(d_V(t)=\max_{i,j}|x_i(t)-x_j(t)|\), the system reaches quasi-consensus when \(\limsup_{t\to\infty} d_V(t)\le \epsilon\). A sharp threshold holds at \(\sigma_c=\epsilon/2\): if \(|\xi_i(t)|\le \epsilon/2\) almost surely, then quasi-consensus is reached in finite time almost surely; if the noise has positive probability of exceeding \(+\epsilon/2\) and \(-\epsilon/2\), then almost surely the system does not reach quasi-consensus [1512.05058].

A different noisy formulation allows random opinion jumps with probability \(m\). With global jumps, cluster-forming instability is suppressed for \(m>m_c\approx 0.51\), independently of \(\epsilon\); for \(m<m_c\), the dominant wavenumber is \(q_{\max}\approx 2.8/\epsilon\), predicting approximately \(0.4/\epsilon\) clusters. With bounded local jumps of half-width \(\gamma\), the linear-stability analysis yields the critical line
\[
m_c(\epsilon,\gamma)=\frac{2\epsilon^2}{2\epsilon^2+\gamma^2}.
\]
This places HK clustering in an order-disorder framework analogous to pattern-formation models [1309.2858].

Noise also changes the robustness of fragmentation. One study shows that arbitrarily small nonzero noise destroys the fragmentation of the classical HK model and of homogeneous prejudiced or homogeneous stubborn-agent variants, leading instead to quasi-consensus. By contrast, a model with heterogeneous prejudices retains robust bipolar clustering under suitable separation of the prejudice values [1712.04277]. In heterogeneous HK systems with environment noise or communication noise, the minimal confidence threshold \(r_{\min}\) determines a phase transition: for environment noise, quasi-synchronization occurs if and only if the noise amplitude satisfies \(\eta\le r_{\min}/2\), and the hitting time to the \(2\eta\)-tube has an exponential-tail bound [1912.10753].

Continuous-time and mean-field formulations place HK dynamics within stochastic interacting-particle theory. A continuous-time noisy HK system with idiosyncratic and environmental Brownian terms converges, as \(N\to\infty\), to a McKean–Vlasov stochastic differential equation with common noise, and the associated nonlocal stochastic Fokker–Planck equation is well posed under the stated ellipticity and smoothness assumptions [2301.03955]. A separate continuous HK formulation derives an integro-differential conservation law, shows that it is equivalent to a finite weighted-cluster ODE system, proposes a midpoint particle method with second-order weak convergence, and proves monotonicity of a concentration functional via an integration-by-parts argument [2211.06265].

## 6. Multidimensional, invariant, and application-oriented extensions

Multi-topic HK dynamics can be defined in more than one way. In the average-based multidimensional model, each agent carries \(x_i(t)\in\mathbb R^d\), but neighborhoods are determined by the scalar average opinion
\[
\bar x_i(t)=\frac{1}{d}\sum_{k=1}^d x_i^k(t).
\]
The induced row-stochastic matrix \(A^{\mathrm{ave}}(t)\) updates the full opinion vector \(X(t)\), and the coordinate-wise ranges \(\nu_k(t)\) as well as the global \(\ell_\infty\)-range are nonincreasing. Moreover, the full vector system reaches consensus or clustering if and only if the scalar average-opinion dynamics does, and it reaches exactly the same clustering pattern as that scalar model [2204.08515].

In the uniform-affinity multidimensional model, two agents interact only when all coordinate-wise differences are at most \(\epsilon\), equivalently when \(\|x_i(t)-x_j(t)\|_\infty\le \epsilon\). Here the global range is again nonincreasing, but topic-wise order preservation requires additional conditions because \(\mathbb R^d\) lacks a total order. A one-step ordering theorem holds when every non-neighbor pair differs by more than \(\epsilon\) in every coordinate, and a full-time ordering theorem holds once all topics are simultaneously ordered \(x_1^k\le \cdots \le x_N^k\) for every coordinate \(k\) [2204.08515].

Recent work has also formalized structural invariances of the one-dimensional synchronous HK process. For a fixed initial profile \(X(0)\), the set of \(\epsilon\)-switches—critical confidence values at which the bounded-confidence process changes—is finite, so the enumeration algorithm terminates. If \(\epsilon^*\) is the next switch, then the trajectories for \(\epsilon\) and \(\epsilon^*\) coincide up to the switch time. The entire dynamics is invariant under positive-affine transformations \(x\mapsto \alpha x+\beta\), provided \(\epsilon\) is rescaled to \(\alpha\epsilon\); the \(\epsilon\)-switches scale in the same way [2508.15982].

Application-oriented extensions preserve the HK averaging core while adding global information sources or endogenous institutions. An AI-Oracle model lets each agent combine the average opinion of its bounded-confidence neighborhood, the Oracle’s opinion, and a true value \(T\). In simulations with universal Oracle access and no direct truth access, all opinions converge to a single consensus independent of \(\epsilon\); when only some agents access \(T\), universal Oracle use can guarantee convergence to \(T\) under the stated conditions, although universal Oracle use may also delay convergence when everyone already observes \(T\) [2502.19701]. A voter–party extension introduces mutual voter–voter, voter–party, party–voter, and party–party interactions; it exhibits cluster formation and a phase transition from disagreement to consensus, and in the deterministic case unanimous consensus is guaranteed when \(\mu_{pp}<\mu_{vp}\) and the interaction radii all exceed the relevant initial-hull diameter [2410.13378].

Taken together, these developments show that HK opinion dynamics is less a single model than a family of state-dependent averaging processes organized around bounded confidence. The classical synchronous system remains the reference point, but current research treats asynchronous activation, social constraints, heterogeneity, truth attraction, control, noise, multidimensionality, and invariant structure as first-class components of the theory. The resulting picture is one in which consensus and fragmentation are not separate phenomena but different asymptotic regimes of the same nonlinear averaging architecture [2112.13348; 1411.4814].

Source: https://www.emergentmind.com/topics/hegselmann-krause-opinion-dynamics