Continuous Opinion Dynamics under Bounded 8 0 Confidence: A Survey 0 2 n Jan Lorenz∗ a J July 12, 2007 8 1 ] h Abstract p Models of continuous opinion dynamics under bounded confidence - c have been presented independently by Krause and Hegselmann and by o Deffuant et al in 2000. They have raised a fair amount of attention in s the communities of social simulation, sociophysics and complexity sci- . s ence. The researchers working on it come from disciplines as physics, c mathematics, computer science, social psychology and philosophy. i s Inthesemodelsagentsholdcontinuousopinionswhichtheycangrad- y ually adjust if they hear the opinions of others. The idea of bounded h confidenceis thatagentsonlyinteract iftheyareclose in opiniontoeach p other. Usually, the models are analyzed with agent-based simulations in [ a Monte-Carlo style, but they can also be reformulated on the agent’s 2 density in the opinion space in a master-equation style. The contribu- v tion of this survey is fourfold. First, it will present the agent-based and 2 density-based modeling frameworks including the cases of multidimen- 6 sional opinions and heterogeneous bounds of confidence. Second, it will 7 givethebifurcationdiagramsofclusterconfigurationinthehomogeneous 1 modelwithuniformlydistributedinitialopinions. Third,itwillreviewthe . 7 several extensions and the evolving phenomena which have been studied 0 so far, and fourth it will state some open questions. 7 0 opinion dynamics, continuous opinions, cluster formation, bifurcation patterns : PACS Nos.: 89.20.-a;89.65.-s v i X 1 Introduction r a The term “opinion dynamics” nowadays summarizes a wide class of different models differing in heuristics, formalization as well as in the phenomena of interest. The latter range from emergence of fads, minority opinion spreading, collective decision making, finding and not finding of consensus, emergence of political parties, minority opinion survival, emergence of extremism and so on. ∗Department of Mathematics and Computer Science, University of Bremen, Bibliothek- straße, 28329 Bremen, Germany, [email protected]; Present adress: ETH Zu¨rich, Chairof Sys- temsDesign,Keuzplatz5,8032Zu¨rich,Switzerland,[email protected] 1 This paper deals with these phenomena in models of continuous opinion dynamics under bounded confidence. ‘Continuous’ refers to the opinion issue and not to the time. Thus opinions in continuous opinion dynamics should be expressable in real numbers where compromising in the middle is always possi- ble. Example issues are prices, tax rates or predictions about macroeconomic variables. The politicalspectrumis alsooftenmappedtoacontinuumfromleft to right wing. Inopiniondynamicsoneconsidersasetofagentswhereeachholdsanopinion from a certainopinion space. She may change her opinion when she gets aware ofthe opinionsof others. In the physics literature discrete opinionspaces (clas- sically binary opinions) (1; 2; 3) have dominated research due to their striking analogy with spin systems. Sometimes they have been extended to more than two spin values, which are ordered and thus get closer to continuous opinion dynamics (4; 5). In recent years two models of genuinely continuous opinion dynamics under boundedconfidencehaveraisedthe interestofthe sociophysicscommunity: the models of Hegselmann and Krause(6; 7; 8) and Deffuant, Weisbuch and others (9;10). Duetophysicist’sresearchsomeprogressinunderstandingthedynamics of these models has been made, especially by introducing the master equation ontheagentsdensityintheopinionspaceforthesetypeofmodels. Forarecent review in this broader context of opinion dynamics including discrete opinions from a physicist’s perspective see (11). Let us consider a population of agents which hold diverse opinions about certain issues expressible in real numbers. Each agent is willing to change her opinion if she hears the opinions of others by adjusting towardsthose opinions. Everyadjustmentintermsofaveragingispossibleduetothecontinuousnature of the opinions. Further on, consider agents to have bounded confidence. That means an agent is only willing to take those opinions into account, which differ less than a certain bound of confidence ε from her own opinion. TheDeffuant-Weisbuch(DW)modelandtheHegselmann-Krause(HK)model both rely on the idea of repeated averaging under bounded confidence. They differ in their communicationregime. Inthe DW model agentsmeetin random pairwiseencountersafterwhichtheycompromiseornot. IntheHKmodel,each agent moves to the average opinion of all agents which lie in her area of confi- dence (including herself). Actually, the DW model contains another parameter which controlls how close an agent moves to the opinion of the other. But it has turned out that this parameter has only an effect on convergence time in the basic model. Therefore we neglect it in the basic analysis and discuss its impact afterwards. The DW modelwaspartly inspiredby the famous Axelrodmodelaboutthe disseminationofculture(12)wheresomethingsimilartotheboundedconfidence assumptionisimplemented. Ithasbeendevelopedinaprojectaboutimproving agri-environmentalpolicies in the European union. The HK model has been presented by Krause (6; 7) in a mathematical context as a nonlinear version of older consensus models (13; 14; 15). It was analyzed through computer simulations by Hegselmann and Krause (8) in the 2 context of social simulation and has gained a lot of attention since. The next section presents the two models in their original agent-based ver- sionsandintheirdensity-basedformulationinspiredbystatisticalphysics. Sec- tion 3 shows and explains the bifurcation diagrams of both models which serve as a reference for the review of several extensions in Section 4. 2 The models From the heuristic description of the models one can either define agent-based dynamics for a finite population of n agents or density-based dynamics for a density function whichdetermines the agentsdensity in the opinionspace. The latter approach can then be interpreted as taking the infinite limit n → ∞ of agent-based models. It is a classical tool of statistical physics often called the derivationofamasterequationorrateequationwhichwasfirstappliedtosocial systems bei Weidlich (16; 17). Nevertheless,thebasicingredientofbothapproachesisacontinuousopinion space S ⊂ Rd. The d real numbers represent opinions on d different subjects. Usually, only compact and convex sets are regarded as appropriate opinion spaces. The simplest case that we will discuss is the one-dimensional interval [0,1]. But definitions extend naturally to more dimensions. In an agent-based model the state variable is an opinion profile x(t) ∈ Sn, which is a vector of vectors. For agent i, it contains the opinion vector xi ∈S. Foraninitialopinionprofilex(0)dynamicsaredefinedrecursivelyasx(t+1)= f(t,x(t)). The pair (S,f) is then a discrete dynamical system. In a density based model, the state variable is a density function on the opinion space P(t,·) : S → R with P(t,x)dx = 1 for all t. Dynam- ≥0 S ics are then defined for a given initial dRensity function P(0,·) as the evolu- tion of the density function in time. Time is sometimes regarded as discrete and sometimes as continuous. In the continuous-time case we consider the differential equation ∂ P(t,x) = g(P(t,·)) where g operates on the space of ∂t density functions. Then we try to find solutions of the differential equation analytically or with numerical solvers. In the discrete-time case we replace the differential operator by the difference operator and write ∆P(t+1,x) = P(t+1,x)−P(t,x) = g(P(t,·)). We can then directly compute the solution recursivelyP(t+1,x)=P(t,x)+∆P(t+1). Actually,bothdiscrete(18;19)and continuous(20; 21) time wereused to compute solutions numerically. It turned out that discrete and continuous solutions of the DW model are quite similar butdiscreteandcontinuoussolutionsoftheHKmodeldifferqualitatively,which we will point out later. Inthe followingwewill define the agent-basedDW andHK modelindetail, as wellas their density-basedcounterparts. The definitions will alreadyinclude the extensions of multidimensional opinions and heterogeneous bounds of con- fidence. In a first step it is enough to think of the opinion space as the interval [0,1] and to regard all bounds of confidence as equal. 3 2.1 The Deffuant-Weisbuch model We proceed with the definition of the agent-based version of the DW model. Definition 1 (Agent-based DW model) Let there be n ∈ N agents and an appropriate opinion space S ⊂ Rd. Given an initial profile x(0) ∈ Sn, bounds of confidence ε ,...,ε > 0, and a norm k·k we define the agent-based DW 1 n process as the random process (x(t))t∈N that chooses in each time step t ∈ N two random agents i,j which perform the action (xj(t)+xi(t))/2 if xi(t)−xj(t) ≤ε xi(t+1) = i (cid:26) xi(t) oth(cid:13)erwise. (cid:13) (cid:13) (cid:13) The same for xj(t+1) with i and j interchanged. If ε =···=ε we call the model homogeneous, otherwise heterogeneous. 1 n It has been shown (22) that the homogeneous process always converges to a limit opinion profile (usually not in finite time). The same is observed in simulations of the heterogeneous case, but a proof is lacking. A limit profile in the homogeneouscase is anopinionprofile where for eachtwo opinions x ,x it i j holds that they are either equal (belong to the same cluster) or have a distance larger than ε. Thus further changes are not possible regardless of the choice of i and j. Further on, it is easy to see that the average opinion over all agents is conserved during dynamics (23), but only in the homogeneous case. For example trajectories of the process see (9; 19). For the density-based DW model we extend the model defined in (20) to populations of agents with heterogeneous bounds of confidence. Definition 2 (Density-based DW model) Let S ⊂ Rd be an appropriate opinion space, [ε ,ε ] be an interval of possible bounds of confidence and the 1 2 initial density function on the opinion space times the interval of bounds of confidence be P(0,·,·) : S×[ε ,ε ] → [0,∞]. with ε2dxdεP(0,x,ε) = 1. 1 1 2 S ε1 For abbreviation we define the aggregated density aRs PR(t,x) = ε2dεP(t,x,ε). ε1 With this we define the differential equation R ε2 ∂ P(t,x,ε)= dx1 dε¯ dx2 P(t,x1,ε¯)P(t,x2) + ∂t Z Z (cid:16) Z h S ε1 kx1−x2k≤ε¯ x −x 1 2 P(t,x )P(t,x ,ε¯) δ(x− ) 1 2 i 2 (cid:19) − dx (P(t,x ,ε)P(t,x )δ(x−x )+P(t,x )P(t,x ,ε)δ(x−x )) 2 1 2 1 1 2 2 Z kx1−x2k≤ε  (1) 1P(0,·,·) is a density function at time zero over the opinion space and the interval of bounds of confidence, P(0,x,ε)dxdεrepresents the proportionof agents whichholdopinions in[x,x+dx]andbounds ofconfidence in[ε,ε+dε]. 4 The continuous-time density-based DW process is then defined as the solution of the initial value problem. The discrete-time density-based DW process is the sequence (P(t,·,·))t∈N recursively defined by P(t+1,x,ε)=P(t,x,ε)+∆P(t,x,ε) with ∆P(t,x,ε)= ∂ P(t,x,ε). ∂t Notice that it is important to distinguish the aggregated density P(t,x) defined above and the density for a given bound of confidence P(t,x,ε) to un- derstandhowanagentlooksatallagentsregardlessoftheirboundofconfidence and adjusts to the ones within her bound of confidence. For ahomogeneousboundofconfidenceε=[ε ,ε ]the dynamicalEquation 1 2 (1) reduces to ∂ x +x 1 2 P(t,x)= dx dx P(t,x )P(t,x ) 2δ(x− ) 1 2 1 2 ∂t Z Z (cid:16) (cid:16) 2 S kx1−x2k≤ε fractionjoiningstatex | {z } −(δ(x−x )+δ(x−x )) 1 2 (cid:17) (cid:17) fractionleavingstatex | {z }(2) which has been reportedfirst in (20) and gives a good starting point for under- standingpopulationswithheterogeneousboundsofconfidence. Theδ-functions represent the position where the mass of the two opinions x and x jumps to. 1 2 Equation (2) can be transformed to a form free of δ-distributions, ∂ P(t,x)= dy4P(t,y)P(t,2x−y) − dy2P(t,y). (3) ∂t Z Z kx−yk≤2ε kx−yk≤ε One insight is that the gain term for opinion x is a ‘bounded convolution’ of P(t,·) with itself at the point 2x. It is pointedoutin(20) thatEquation(2)conservesthe massandthe mean opinion of the density. For Equation (1) this holds only for the mass, which is conserved for each value of ε and of course for the aggregated density, too. Thisgivesrisetothefactthatunderheterogeneousboundsofconfidenceoverall drifts (in terms of a drift of the mean opinion) to more extremal opinions may happen. This has been simulated by (24) in a very stylized model. In (20) it is shown for S = [0,1] and ε ≥ 0.5 that P(t,·) in Equation (2) goes with t → ∞ to a limit distribution P(∞,x) = δ(x−x ) where x is the 0 0 mean opinion. For lower epsilon it is postulated (and verified by simulation) r that P(∞,x) = m δ(x−x ) where r is the number of evolving opinion i=1 i i clusters, xi ∈ [0,P1] is the position of cluster i and mi is its mass. All clusters r r must fulfill the conservation laws m = 1 and x m is equal to the i=1 i i=1 i i conserved mean opinion. FurtherPon, all different cPlusters i 6= j must fulfill |x −x |>ε. This conicides with the result for the agent-based model. i j 5 Ingeneralthe righthandside ofEquation(2)does notdependcontinuously onP(t,·). 2 ButthishappensonlyinthecasewhereP(t,·)containsδ-functions and is thus not a ‘normal’ function in x but a distribution. Noncontinuity of therighthandsideof (2)implies thatthere isnooratleastnounique solution. Fortunately,itturnedoutthatδ-functions donotevolveduringtheprocessbut onlyinthe limitt→∞. IfwestartwithP(t,·)withoutδ-functionstrajectories can be computed as it is done in (20) with a fourth order Adams-Bashforth algorithm. Moreover, it turned out in comparing results of (20) and (19) that the discrete-time density-based process leads to nearly the same limit densities as the continuous-time density-based process. In (18; 19), additional to discrete time, the space P(t,·) has been discretized to p(t) ∈ Rn . Dynamics can then ≥0 be defined as an interactive Markov chain (a state and time discrete Markov chain where transition probabilities depend on the actual state). Essentially, the interactive Markov chain is the same what evolves by discretization of the continuous opinion space for numerical computation in (20). Inanutshell,therearedifferentapproachesbutresultsforthedensity-based DW model lie all close together. This does not hold for the HK model as we will see. Examples for trajectories of the process can be found in (19; 20). 2.2 The Hegselmann-Krause model Here we will give the HK model, first in its original agent-based version. Definition 3 Agent-based HK model Let there be n ∈ N agents and an appropriate opinion space S ⊂Rd. Given an initial profile x(0) ∈ Sn, bounds of confidence ε ,...,ε > 0 and 1 n a norm k·k we define the HK process (x(t))t∈N recursively through x(t+1)=A(x(t),ε ,...,ε )x(t), (4) 1 n with A(x,ε ,...,ε ) being the confidence matrix defined 1 n 1 if j ∈I (i,x) Aij(x,ε1,...,εn):= #Iεi(i,x) εi (cid:26) 0 otherwise, with I (i,x):={j ∈n| xi−xj ≤ε } beingthe confidencesetof agent i with εi i respect to opinion profile(cid:13)x. In (cid:13)other words, the new opinion xi(t+1) is the (cid:13) (cid:13) arithmetic average over all opinions in x(t) differing from it by not more than ε . i If ε =···=ε we call the model homogeneous, otherwise heterogeneous. 1 n 2Consider P(t,·) = 1δ(x−1)+ 1δ(x+1+ξ) then we have a fixed point for ε = 2 2 2 and any ξ > 0. But for ξ = 0 applying the right hand side of Equation (2) would give 1δ(x−1)+ 1δ(x)+ 1δ(x+1). 4 2 2 6 It has been shown (25; 26) that the homogeneous process always converges to a limit opinion profile in finite time. The same convergence is observed in simulationsfortheheterogeneouscasebutitdoesnotoccurinfinite time anda proofis lacking. A limit profile in the homogeneouscase is alwaysa fixed point x∗ = A(x∗,ε)x∗. In x∗ it holds for each pair of two opinions xi,xj that they are either equal (belong to the same cluster) or have a distance larger than ε. The proof for this is not difficult but also not trivial and can be found in (19). In contrast to the DW model the mean opinion is not conserved. The mean opinion is only conservedwhen the initial profile is symmetric around its mean opinion. For example trajectories of the process see (8; 19). We nowturnourattentiontothe density-basedversionoftheHKmodelfor which we extend models independently developed in (21) and (18). For the definition of the density-based HK model we need the definition of the ε-local mean x+ε yP(t,y)dy M (x,P(t,·),ε)= x−ε . 1 R x+ε P(t,y)dy x−ε R ItgivestheexpectedvalueofP(t,·)intheinterval[x−ε,x+ε]. Thenominator gives the first moment in the ε-interval around x while the denominator gives the necessary renormalizationby the probability mass in that interval. Definition 4 (Density-based HK model) Let S ⊂ Rd be an appropriate opinion space, [ε ,ε ] be an interval of possible bounds of confidence and the 1 2 initial density function on the opinion space times the interval of bounds of confidence be P(0,·,·) : S ×[ε ,ε ] → [0,∞] with ε2dxdεP(0,x,ε) = 1. 3 1 2 S ε1 For abbreviation we define the aggregated density aRs PR(t,x) = ε2dεP(t,x,ε). ε1 With this we define the differential equation R ε2 ∂ P(t,x,ε)= dy dε¯P(t,y,ε¯)δ(x−M1(y,P(t,·),ε¯)) ∂t Z (cid:16) Z (cid:17) S  ε1 − P(t,x,ε)δ(y−M (x,P(t,·),ε)) (5) 1 i The continuous-time density-based HK process is then defined as the solution of the initial value problem. The discrete-time density-based HK process is the sequence (P(t,·,·))t∈N recursively defined by P(t+1,x,ε)=P(t,x,ε)+∆P(t,x,ε) with ∆P(t,x,ε)= ∂ P(t,x,ε). ∂t 3P(0,·,·) is a density function at time zero over the opinion space and the interval of bounds of confidence, P(0,x,ε)dxdεrepresents the proportionof agents whichholdopinions in[x,x+dx]andbounds ofconfidence in[ε,ε+dε]. 7 As for the density-based DW model it is important to distinguish the ag- gregated density P(t,x) defined above and the density for a given bound of confidence P(t,x,ε) to understand how an agent looks at all agents regardless of their bound of confidence and takes into account all within her bound of confidence. For ahomogeneousboundofconfidenceε=[ε ,ε ]the dynamicalEquation 1 2 (5) reduces to ∂ P(t,x)= δ(M (y,P(t,·),ε)−x)P(t,y)−δ(M (x,P(t,·),ε)−y)P(t,x)dy. 1 1 ∂t Z S fractionjoiningstatex fractionleavingstatex | {z } | {z }(6) This can be transformed to the form ∂ P(t,y) P(t,x)= dyδ(y) −P(t,x) ∂t Z |M′(y,P(t,·),ε)| 1 {yisrootofx−M1(y,P(t,·),ε)}  (7) withtheM′ beingthederivativewithrespecttoy. ThefunctionM (y,P(t,·),ε) 1 1 is monotone in y for all P. Therefore x−M (y,P(t,·),ε) has usually only one 1 or no root. In the case of one root the integral in Equation (7) reduces to P(y0) with y being that root. But there might be a continuum of |M1′(y0,P(t,·),ε)| 0 roots due to a plateau of x−M (y,P(t,·),ε) at zero, because M (y,P(t,·),ε) 1 1 is not strongly monotone. It is shown in (21) that Equation (6) conserves the mass. Further on it is claimed that the mean opinion is also conserved. Although the derivation in (21) relies on the symmetry of the initial opinion density around its mean the authors forgot to mention that the mean opinion is not generally conserved for other initial conditions. Ingeneralthe righthandside ofEquation(6)does notdependcontinuously on P(t,·). (See example above.) But unfortunately a δ-peak in the density may evolve in this model by computing the right hand side of Equation (6) even with a given P(t,·) which does not contain a δ-peak. 4 Therefore the existence of unique solutions to (6) is not generally assured for every initial condition. Neverthelessthey werecomputedin(21)witha fourthorderRunge- Kutta algorithm and smooth looking dynamics were reached. But the picture changes drastically when we switch to the discrete-time density-basedHKmodel,whichhasbeenstudiedin(18;27;19)asaninteractive Markov chain with 1000 and 1001 opinion classes. This coincides with the discretization of the interval [0,1] into 1000 bins in (21). In the discrete-time casetwo phenomena occurwhichwere notreportedin(21): Consensusstriking 4Consider P(t,·) to be uniformly distributed on [0,1]∪[2,1] and ε = 0.5. Then all the 3 3 ’agents’ withopinions inthe interval [0,1]have ε-localmeanequal to 1 atthat position we 6 6 willhaveaδ-peak. 8 back for lower ε after a phase of polarization (27) and the importance of an odd versus an even number of opinion classes (18). Under an odd number, consensus is much easier to reach because under an even number, the central mass contracts into two bins which might be divided more easily compared to theoddcase. Inthe oddcasethemassisstoredinonebin,andmassinonebin canonly move jointly by definition ofthe process. We discussthe differences of continuous-time and discrete-time results in the next section. A numericalissue whichis notclearly defined in(21) is how movingmass is assigned to bins when computing the right hand side of Equation (6). Usually, the ε-local mean of a given bin lies not directly in a bin but between two bins. So, there has to be a rule how to distribute it between these two bins. In (18; 27; 19) it is done proportionally to the distances from each bin. If the ε-local mean lies closer to one bin, this bin receives more of the moving mass. In this difference of continuous versus discrete time, the HK model distin- guishes from the DW model. Nevertheless, the possible fixed points of the dynamics are the same as in the DW model. It is again every distribution like r r P(∞,x)= m δ(x−x ) with |x −x |>ε for all i6=j and m =1. i=1 i i i j i=1 i This coincidPes with the result for the agent-based model. But forPboth models it is not easy to determine the limit density out of the initial density. Nearly every study until now has studied only uniformly distributed initial densities. Example trajectories of the process can be found in (21; 19). 3 Bifurcation Diagrams In this section we give the basic bifurcation diagrams for both homogeneous density-based models. A bifurcation diagram shows the location of clusters in the limit density versus the continuum of values of the bound of confidence ε. So, one can determine the attractive cluster patterns for each bound of confidence and observe transitions of attractive patterns at critical values of ε. The bifurcation diagrams serve as references for discussions on robustness to extensions of the homogeneous models and the impact of these extensions, e.g. on the critical bounds of confidence. ThetopplotofFigure1showsthebifurcationdiagramforthehomogeneous density-basedDWmodelwithuniforminitialdensityintheopinionspace[0,1]. A bifurcation diagram is made by computing the trajectory of the corre- sponding interactiveMarkovchainwith 1001opinionclassesuntilclusters have evolvedwhichspreadnotmorethanεandarefurtherawayfromeachotherthan ε. Clusters are determined by collecting intervals in the opinion space where the density is positive. Therefore, one has to decide on a threshold for which mass is regarded as zero. This threshold was 10−9 here. (For computational details see (19).) Thebifurcationdiagramshownhereresemblesthebifurcationdiagramshown in (20) but with a transformation of variables as ∆ = 1 . In (20) ∆ is the 2ε changing variable and determines the opinion space [−∆,∆] and the bound of confidence is fixed to one. Here we have a fixed opinion space and a changing 9 bound of confidence because it is always done that way in agent-based simula- tionsandthisframeworkcanbeextendedtomodelswithheterogeneousbounds of confidence. The fat lines in the top plot display the locations of clusters which have a major mass, the thin lines the location of clusters with a minor mass and the medium line the central cluster. Dotted lines are just for orientation, they show the interval [0.5−ε,0.5+ε]. The two lower plots show the masses of the related clusters. The most striking feature is the existence of minor clusters at the extremes and between major clusters. In agent-based simulations these minor clusters often do not evolve because of low population sizes, nevertheless outliers are frequently reported. These outliers make it difficult to count the number of major clusters correctly. In (9; 10) authors therefore decided to not count clusters with only one agent. The bifurcation diagram in (20) clarified that minor clusters exist for structural reasons. For ε≥0.5 only one big central cluster evolves. As ε decreases bifurcations and nucleations of clusters occur. First, the nucleation of two minor extremal clusters,then the bifurcation ofthe centralcluster into two major clusters,and thirdtherebirthofthecentralcluster. Fordetailssee(19;20). Thisbifurcation pattern is then repeated in shorter ε-intervals. The length of these intervals seems to scale with 1. This can be better seen in the bifurcation diagram ε of (20), where the bifurcation pattern seems to repeat itself on intervals that converge towards a length of about 2.155. But this constant has only been derived numerically and it is not clear that the bifurcation pattern will repeat that regularly, although it looks very much like that. This result resembles the rough 1 -rule reported in (9; 10) for agent-based simulation, which says that 2ε the number of major clusters after cluster formation is roughly determined as the integer part of 1. 2ε Acertainε-phaseofinterestisε∈[0.27,0.35]. Asεdecreasesthemassofthe minorclusterdecreases,too. Thusthemassofthecentralclustermustincrease slightlyuntilthecentralclusterbifurcates. Thus,surprisinglythecentralcluster has a larger mass very close to the critical ε. For further minor clusters the same intermediate low-mass region exist. It even looks as if the minor cluster disappears for a short intermediate ε-range. The question if these gaps exist has also been posed in (20). ThetopplotofFigure2showsthebifurcationdiagramforthehomogeneous density-basedHK modelwith uniforminitialdensity inthe opinionspace[0,1]. The bifurcation diagram was computed the same way as for the DW model. Butthe detectionofthe finaldensitywasmucheasierbecausethetime-discrete process converges in finite time. (For computational details see (19).) The lines in the top plot display the locations of clusters. Unlike the DW modelthe HKmodelhasno minorclusters. The lowerplotshowsthe massesof the related clusters. The convergence time (which is finite here) is the subject of the lowest plot. There are certain ε-values were convergence time is really long. The plot shows a zoom, which does not display the longest times. The longest time was 21431. For about ε≥0.19 consensus is reached, but for values only slightly above, 10