5D Library › Collective Dynamics, Coupled-Oscillator Synchrony and Distributed Consensus · 中文

Chapter 4 Leadership without Signals: Informed Minorities, Uninformed Individuals and Group Consensus

Who leads the way

In migrating flocks and foraging schools often only a few individuals know where the food or the destination is; they give no “follow me” signal, and the others may not even know who is informed. How is the group brought in the right direction? And when its members disagree, who decides? Conradt and Roper in 2003 used a simple model to compare two ways in which animal groups can decide: by a single “despotic” individual, or “democratically” by most members; in their model democratic decisions usually had a lower average cost, because they less often produced extreme choices (Conradt and Roper 2003). This chapter discusses two studies by Couzin and colleagues: “effective leadership” in 2005 and “uninformed individuals promote democratic consensus” in 2011.

Evidence

A few informed individuals can move a shoal. Reebs’s fish-shoal experiment of 2000 addressed whether a minority of individuals that know where food is can determine the foraging movements of the whole shoal (Reebs 2000). Experiments of this kind provide the behavioural background to the models below.

Effective leadership: the share needed falls as groups grow

Evidence

Couzin and colleagues, 2005. Couzin, Krause, Franks and Levin studied the question in Nature with a spatial model: individuals keep a minimum distance from one another and align with and move towards neighbours within a certain range; a proportion pp of them know a preferred direction gg and add it to their desired direction with weight ω\omega; nobody knows who is informed (Couzin et al. 2005). There are three main conclusions. First, the accuracy of the group’s direction rises with pp towards saturation, and the larger the group, the smaller the proportion of informed individuals needed to guide it; a very small proportion can give very high accuracy. Second, ω\omega is positively related to accuracy, but larger ω\omega makes the group more likely to split, a trade-off between accuracy and cohesion. Third, when two informed subgroups prefer different directions and are equal in number, the group takes the average direction if the preferences differ a little and chooses one of the two at random if they differ a lot; if the subgroups are unequal, the group tends to follow the majority even when it is larger by just one (for example 6 against 5). The paper mentions that scouts make up about 5 % of a honeybee swarm and suggests that swarms are already in the saturated region of accuracy.

The framework turns this into a “law of the informed proportion declining with group size”; in fact the paper gives only the qualitative trend and simulated curves, not a universal proportion, and “about 5 %” is a remark about honeybees, not a law of the model. Dyer and colleagues in 2009 tested similar questions with people: they report three experiments, and in walking crowds of 100 and 200 people a small minority of informed individuals was likewise enough to guide a large uninformed group (Dyer et al. 2009).

The program examines the first conclusion with a simplified model that has directions only and no positions: at each step every individual looks at 7 randomly drawn companions (a crude stand-in for the topological neighbours of Chapter 3) and turns to the sum of their unit directions and its own; an informed individual adds 0.5g0.5 g; the turning error has standard deviation 0.4 radians. The run lasts 400 steps; accuracy is the mean cosine of the angle between the group’s direction and gg over steps 200 to 400 (1 means on target, 0 means no information); each combination is simulated 10 times. Finally, two subgroups of 5 informed individuals each (50 in all) prefer directions DD degrees apart, and the program checks whether the group ends up taking the average direction or choosing one of them.

Calculation

# Leadership without signalling, in a heading-only caricature of Couzin et al. (2005).
# Each step every individual looks at k = 7 randomly drawn others (no positions are
# modelled) and turns to the direction of the sum of their unit headings plus its own; an
# informed individual adds w * g (g = preferred direction, w = 0.5). Gaussian turning
# error SD 0.4 rad. Nobody knows who is informed. 400 steps; accuracy = mean of
# cos(group heading - g) over steps 200..400 (1 = on target, 0 = no information).
import math, random

def run(N, n_inf, rng, w=0.5, k=7, sd=0.4, steps=400, goals=None):
    th = [rng.uniform(-math.pi, math.pi) for _ in range(N)]
    goals = goals or [0.0] * n_inf            # preferred directions of the informed
    acc, heads = 0.0, []
    for t in range(steps):
        c = [math.cos(a) for a in th]; s = [math.sin(a) for a in th]; new = []
        pick = rng.choices(range(N), k=k * N)          # k random others per individual
        for i in range(N):
            nb = pick[k * i:k * i + k]
            sx = c[i] + sum(c[j] for j in nb); sy = s[i] + sum(s[j] for j in nb)
            n = math.hypot(sx, sy) or 1.0; sx, sy = sx / n, sy / n
            if i < len(goals):
                sx += w * math.cos(goals[i]); sy += w * math.sin(goals[i])
            new.append(math.atan2(sy, sx) + rng.gauss(0, sd))
        th = new
        if t >= steps // 2:
            psi = math.atan2(sum(map(math.sin, th)), sum(map(math.cos, th)))
            acc += math.cos(psi); heads.append(psi)
    return acc / (steps - steps // 2), heads[-1]

if __name__ == "__main__":
    rng = random.Random(2005)
    print("accuracy (mean of 10 runs) by group size N and share of informed individuals")
    shares = (0.05, 0.10, 0.20, 0.40)
    print("    N    " + "".join(f"  p = {p:4.2f}" for p in shares))
    for N in (10, 30, 100, 300):
        row = []
        for p in shares:
            n = max(1, round(p * N))
            row.append(sum(run(N, n, rng)[0] for _ in range(10)) / 10)
        print(f"  {N:4d}   " + "".join(f"   {a:6.3f}  " for a in row))
    print("two informed subgroups of 5 (N = 50) whose targets differ by D degrees;")
    print("share of 10 runs ending within 15 deg of the average / of either target")
    for D in (60, 150):
        g = math.radians(D / 2); mid = side = 0
        for _ in range(10):
            psi = run(50, 10, rng, goals=[g] * 5 + [-g] * 5)[1]
            mid += abs(psi) < math.radians(15)
            side += min(abs(psi - g), abs(psi + g)) < math.radians(15)
        print(f"   D = {D:3d}:  average {mid / 10:.2f}   one target {side / 10:.2f}")
# accuracy (mean of 10 runs) by group size N and share of informed individuals
#     N      p = 0.05  p = 0.10  p = 0.20  p = 0.40
#     10       0.896      0.883      0.920      0.962  
#     30       0.944      0.943      0.976      0.987  
#    100       0.968      0.985      0.991      0.996  
#    300       0.994      0.995      0.998      0.999  
# two informed subgroups of 5 (N = 50) whose targets differ by D degrees;
# share of 10 runs ending within 15 deg of the average / of either target
#    D =  60:  average 0.80   one target 0.20
#    D = 150:  average 0.60   one target 0.00

The first part reproduces “the share needed falls with group size”: to reach an accuracy above 0.95, a group of 10 needs 40 % informed (0.962), a group of 30 needs 20 % (0.976), and groups of 100 and 300 already reach it at 5 % (0.968 and 0.994). A closer look shows that accuracy depends roughly on the number of informed individuals rather than their share: 4 informed among 10 give 0.962, 6 among 30 give 0.976, 5 among 100 give 0.968. In this mean-field-like simplified model both the random drift of the group’s direction and the pull of the informed shrink as 1/N1/N, so their ratio depends only on the number informed, and the share needed falls roughly as 1/N1/N. The spatial model of Couzin and colleagues also finds a falling share, but the exact form of the decline depends on spatial structure, and this simplified model cannot replace it.

The second part shows a limit of the simplified model: when the two targets are 60 degrees apart, 8 runs in 10 end near the average direction; when they are 150 degrees apart, no run chooses either target (6 end near the average and the rest wander in between). This model only compromises; it never “chooses one of two”. The analytical model of Leonard and colleagues in 2012 shows that the bifurcation between compromise and choice requires a nonlinear coupling mechanism (Leonard et al. 2012); a model without such a mechanism cannot be used to discuss how groups choose under conflict.

Uninformed individuals return control to the majority

Evidence

Couzin and colleagues, 2011. Couzin, Ioannou, Demirel, Gross, Torney, Hartnett, Conradt, Levin and Leonard reported theory and experiments in Science (Couzin et al. 2011). The experiments used golden shiners: 6 fish were trained to swim to a blue target (the majority, with a weaker preference), and 5 fish were trained to swim to a yellow target (the minority; golden shiners are naturally attracted to yellow and learn the yellow association faster and more reliably, so this subgroup had the strong preference). With no untrained individuals the shoal swam to yellow in more than 80 % of trials — the strongly opinionated minority controlled the group; with 5 untrained fish added, the shoal went to blue about half the time; with 10 added, about 70 % of the time. The model showed the same trend, and with too many uninformed individuals the group’s choice became nearly random. The authors’ explanation is that individuals tend to follow the perceived majority, and strongly opinionated individuals resist this social pressure better; uninformed individuals respond only to numerical differences and so curb the influence of the forceful few, bringing the group back to the “numerical majority” (these percentages follow the Princeton University news release of December 2011; the independent check could not obtain the Science paper itself, and the contemporary Science News report gives consistent figures — more than 80 % with no untrained fish, about half with 5, minority wins below 40 % with 10 — but neither is the paper itself, and the figures must be checked against it before being cited).

The framework renders this result as “how an undecided majority restrains hijacking by an extreme minority”, which needs two corrections. First, the untrained fish added numbered 0, 5 or 10 and were not necessarily a majority; their effect was to return control to the more numerous of the informed sides. Second, the paper’s “strongly opinionated minority” refers to the strength of preference and has no political meaning; applying it to “extremists” in human politics is only an analogy, and fish-shoal experiments can provide no evidence for any political claim.

The effect is sensitive to the model. Hartnett and colleagues found in 2016, in a spin model, that the direction of the effect of unbiased individuals depends on how nonlinear individuals’ responses to their neighbours are: with a relatively linear response they reinforce the majority preference, promoting a “democratic” outcome even against a strongly biased minority; with a highly nonlinear response the effect is reversed, and they promote the strongly biased minority and inhibit the majority (Hartnett et al. 2016). “The uninformed return control to the majority” is therefore not a general conclusion; it depends on the specific form of the interaction rules. The program below uses a simple model of the book’s own to show another case: under a rule in which individuals copy one random other (as in the voter model), the uninformed have no effect at all.

Calculation

The program uses the book’s own two-option model (not the model of Couzin and colleagues): 6 individuals weakly prefer A, 5 strongly prefer B, and nUn_U have no preference. At each update one random individual, with probability ss (sA=0.02s_A = 0.02, sB=0.30s_B = 0.30, 0 for the uninformed), returns to its own preference; otherwise it imitates one random other (linear rule) or adopts the majority choice of three random others (majority rule). After 200 sweeps the program records the share choosing A; each condition is simulated 60 times.

For the linear rule the mean-field solution can be written down directly: a weak A-type is in A with probability sA+(1−sA)xs_A + (1-s_A)x, a strong B-type with probability (1−sB)x(1-s_B)x, an uninformed individual with probability xx; summing over everyone, the nUxn_U x terms cancel on both sides, giving x=nAsAnAsA+nBsB,x = \frac{n_A s_A}{n_A s_A + n_B s_B}, independent of nUn_U.

Calculation

# Does a pool of uninformed individuals stop a strongly opinionated minority from winning?
# Two-option model (this book's, after the set-up of Couzin et al. 2011): 6 individuals
# weakly prefer A, 5 strongly prefer B, n_U have no preference. At each update one random
# individual: with probability s (s_A = 0.02, s_B = 0.30; 0 for the uninformed) it returns
# to its own preference; otherwise it copies (linear rule) one random other, or (nonlinear
# rule) the majority of three random others. Outcome after 200 sweeps: share choosing A.
import random

N_A, N_B, S_A, S_B = 6, 5, 0.02, 0.30

def run(n_u, rule, rng, sweeps=200):
    n = N_A + N_B + n_u
    st = [1] * N_A + [0] * N_B + [int(rng.random() < 0.5) for _ in range(n_u)]
    tot = sum(st)
    for _ in range(sweeps * n):
        i = rng.randrange(n)
        if i < N_A and rng.random() < S_A: new = 1
        elif N_A <= i < N_A + N_B and rng.random() < S_B: new = 0
        else:
            m = 1 if rule == "linear" else 3
            votes = 0
            for _ in range(m):
                j = rng.randrange(n - 1); votes += st[j + (j >= i)]
            new = int(2 * votes > m)
        tot += new - st[i]; st[i] = new
    return tot / n

if __name__ == "__main__":
    # linear rule, stationary mean field: an A-type is in A w.p. s_A + (1 - s_A) x,
    # a B-type w.p. (1 - s_B) x, an uninformed one w.p. x; summing over everyone gives
    # x = n_A s_A / (n_A s_A + n_B s_B), whatever n_U is
    x = N_A * S_A / (N_A * S_A + N_B * S_B)
    print(f"linear rule, mean-field share choosing A = {x:.3f} for every n_U")
    rng = random.Random(2011)
    print("rule        n_U   mean share A   P(>= 2/3 choose A)   P(>= 2/3 choose B)")
    for rule in ("linear", "majority"):
        for n_u in (0, 5, 10, 20, 40):
            xs = [run(n_u, rule, rng) for _ in range(60)]
            pa = sum(x >= 2 / 3 for x in xs) / 60; pb = sum(x <= 1 / 3 for x in xs) / 60
            print(f"{rule:9s}   {n_u:3d}      {sum(xs) / 60:.3f}           {pa:.2f}"
                  f"                 {pb:.2f}")
# linear rule, mean-field share choosing A = 0.074 for every n_U
# rule        n_U   mean share A   P(>= 2/3 choose A)   P(>= 2/3 choose B)
# linear        0      0.083           0.03                 0.90
# linear        5      0.049           0.03                 0.97
# linear       10      0.087           0.05                 0.92
# linear       20      0.082           0.03                 0.90
# linear       40      0.057           0.00                 0.98
# majority      0      0.011           0.00                 1.00
# majority      5      0.025           0.02                 0.98
# majority     10      0.002           0.00                 1.00
# majority     20      0.244           0.27                 0.73
# majority     40      0.396           0.42                 0.58

Under the linear rule the mean field predicts a share of 0.074 choosing A for every nUn_U; in the simulations the mean share lies between 0.049 and 0.087 for every nUn_U, and over 90 % of runs end with B (the strong minority) holding more than two thirds. Under this copy-one-other rule the uninformed merely amplify the existing proportions and have no effect at all. The majority rule is different: with nUn_U from 0 to 10, B still wins almost always; with nU=20n_U = 20 the share of runs won by A rises to 0.27, and with nU=40n_U = 40 to 0.42. The uninformed weaken the control of the strong minority, but in this simplified model they only push the outcome towards a coin toss; they do not, as in the fish experiment, let the majority win 70 % of the time — a 6-to-5 advantage is too small, and reproducing the experiment needs more mechanisms, such as spatial structure.

Hypothesis · not established · The role of the undecided depends on the nonlinearity of ‘following the majority’

From the analysis above, the author conjectures that in experiments on two-option decisions by small human groups, whether adding members who are told no preference weakens the influence of a “strongly preferring minority” depends on whether members update in the nonlinear “follow the visible majority” way; if the design lets members consult only one other person at a time (linear imitation), adding undecided members will not change the outcome. This can be tested by comparing two ways of presenting information (showing the choice of one random other, or showing the majority choice of several). If adding undecided members also raises the winning rate of the majority significantly under the linear presentation, the conjecture is wrong. (The model of Hartnett and colleagues in 2016 points the other way: with a relatively linear response the uninformed reinforce the majority, with a highly nonlinear one they favour the minority; the same experiment can distinguish the two predictions.) The conjecture concerns only small-group decisions in the laboratory and makes no political claim.

Summary

A few informed individuals can guide a group without signals; the conclusion of Couzin and colleagues in 2005 is that the larger the group, the smaller the proportion of informed individuals needed, but there is no universal proportion. The book’s simplified model shows that accuracy depends mainly on the number of informed individuals, and also exposes its limitation: it can compromise but not choose. The fish experiments of Couzin and colleagues in 2011 show that untrained individuals can weaken the control of a strongly opinionated minority and return the choice to the numerical majority; in the book’s model in which individuals copy one random other, the uninformed have no effect at all, and the effect appears only under a “majority of three” rule, while the model of Hartnett and colleagues in 2016 shows that with highly nonlinear responses the direction of the effect is reversed — whether the effect appears, and in which direction, depends on the specific form of the interaction rules. The next chapter turns to the averaging of opinions: the DeGroot model.

References

Conradt, Larissa, and Timothy J. Roper. 2003. “Group Decision-Making in Animals.” Nature 421 (6919): 155–58. https://doi.org/10.1038/nature01294.
Couzin, Iain D., Christos C. Ioannou, Güven Demirel, et al. 2011. “Uninformed Individuals Promote Democratic Consensus in Animal Groups.” Science 334 (6062): 1578–80. https://doi.org/10.1126/science.1210280.
Couzin, Iain D., Jens Krause, Nigel R. Franks, and Simon A. Levin. 2005. “Effective Leadership and Decision-Making in Animal Groups on the Move.” Nature 433 (7025): 513–16. https://doi.org/10.1038/nature03236.
Dyer, John R. G., Anders Johansson, Dirk Helbing, Iain D. Couzin, and Jens Krause. 2009. “Leadership, Consensus Decision Making and Collective Behaviour in Humans.” Philosophical Transactions of the Royal Society B 364 (1518): 781–89. https://doi.org/10.1098/rstb.2008.0233.
Hartnett, Andrew T., Emmanuel Schertzer, Simon A. Levin, and Iain D. Couzin. 2016. “Heterogeneous Preference and Local Nonlinearity in Consensus Decision Making.” Physical Review Letters 116 (3): 038701. https://doi.org/10.1103/PhysRevLett.116.038701.
Leonard, Naomi E., Tian Shen, Benjamin Nabet, Luca Scardovi, Iain D. Couzin, and Simon A. Levin. 2012. “Decision Versus Compromise for Animal Groups in Motion.” Proceedings of the National Academy of Sciences 109 (1): 227–32. https://doi.org/10.1073/pnas.1118318108.
Reebs, Stephan G. 2000. “Can a Minority of Informed Leaders Determine the Foraging Movements of a Fish Shoal?” Animal Behaviour 59 (2): 403–9.

Fan, Y. (2026). Collective Dynamics, Coupled-Oscillator Synchrony and Distributed Consensus: From the Kuramoto Phase Transition and Vicsek Flocking to DeGroot Spectral Convergence, Read Beside Classical Chinese Thought on Harmony and Conformity [集群动力学、耦合振子同步与分布式共识的数学模型] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23828063 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).