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

Chapter 1 The Kuramoto Model: the Synchronisation Transition of Coupled Oscillators, and Applause

From fireflies to a solvable model

When a population of oscillators, each with its own pace — flashing fireflies, cardiac pacemaker cells, an audience clapping — influence one another, when do they fall into step? Winfree in 1967 was the first to pose this as a problem about a large population of interacting limit-cycle oscillators, and saw in simulations that when the spread of frequencies fell below a certain level part of the population suddenly “froze” into a synchronised cluster, like a phase transition in time (Winfree 1967). Kuramoto (蔵本由紀, Kuramoto Yoshiki) in 1975 simplified Winfree’s idea into a model that can be solved exactly; it first appeared as a three-page note in volume 39 of Lecture Notes in Physics (Kuramoto 1975) and was developed systematically in his monograph (Kuramoto 1984). Strogatz’s review of 2000 traces the model from its proposal to its rigorous analysis (Strogatz 2000), and the review by Acebrón and colleagues in 2005 covers its many extensions (Acebrón et al. 2005).

In the Kuramoto model oscillator ii is described by a phase θi\theta_i alone; its natural frequency ωi\omega_i is drawn from a distribution g(ω)g(\omega), and every pair of oscillators pulls on each other through the sine of their phase difference: θ̇i=ωi+KN∑j=1Nsin⁡(θj−θi),i=1,…,N.\dot\theta_i = \omega_i + \frac{K}{N}\sum_{j=1}^{N}\sin(\theta_j – \theta_i), \qquad i = 1, \dots, N. K≥0K \ge 0 is the coupling strength. Three things are assumed: each oscillator is coupled equally to every other (global coupling), the coupling depends only on phase differences, and the amplitudes of the oscillators do not change. These assumptions make the model solvable, and they are also exactly where it departs from real systems.

The complex order parameter and the self-consistency equation

Kuramoto’s key step was to introduce the complex order parameter reiψ=1N∑j=1Neiθj.r e^{i\psi} = \frac{1}{N}\sum_{j=1}^{N} e^{i\theta_j}. r∈[0,1]r \in [0, 1] measures overall coherence: r=1r = 1 when all phases coincide and r≈0r \approx 0 when they are spread uniformly; ψ\psi is the mean phase. Substituted into the model, the equation of each oscillator becomes θ̇i=ωi+Krsin⁡(ψ−θi).\dot\theta_i = \omega_i + K r \sin(\psi – \theta_i). Each oscillator no longer interacts with the other N−1N-1 separately but feels only a “mean field” of strength KrKr. This makes the model solvable self-consistently: rr is determined by the phases, and the phases are determined by rr.

Calculation

Deriving the self-consistency equation. Let N→∞N \to \infty, let gg be symmetric about 0 (work in a rotating frame in which the mean frequency is 0), and look for a steady state in which rr and ψ\psi are constant, taking ψ=0\psi = 0.

Oscillators with |ωi|≤Kr|\omega_i| \le Kr lock at sin⁡θi=ωi/(Kr)\sin\theta_i = \omega_i/(Kr) (the stable solution with |θi|≤π/2|\theta_i| \le \pi/2); oscillators with |ωi|>Kr|\omega_i| > Kr drift, their stationary density on the circle is inversely proportional to their speed and symmetric under θ→θ+π\theta \to \theta + \pi, and they contribute nothing to ⟨eiθ⟩\langle e^{i\theta}\rangle. Hence r=∫−KrKrcos⁡θ(ω)g(ω)dω=Kr∫−π/2π/2cos⁡2θg(Krsin⁡θ)dθ,r = \int_{-Kr}^{Kr} \cos\theta(\omega)\, g(\omega)\, d\omega = Kr\int_{-\pi/2}^{\pi/2}\cos^2\theta\; g(Kr\sin\theta)\, d\theta, using the substitution ω=Krsin⁡θ\omega = Kr\sin\theta. r=0r = 0 (the incoherent state) is always a solution; solutions with r>0r > 0 satisfy 1=K∫−π/2π/2cos⁡2θg(Krsin⁡θ)dθ.1 = K\int_{-\pi/2}^{\pi/2}\cos^2\theta\; g(Kr\sin\theta)\, d\theta. As r→0+r \to 0^+ the right-hand side tends to Kg(0)π/2K g(0)\pi/2, so the threshold at which a partially locked state appears is Kc=2πg(0).K_c = \frac{2}{\pi g(0)}.

The derivation rests on several conditions; the framework gives only the formula, so they are listed here (Strogatz 2000). First, N→∞N \to \infty: at finite size rr keeps fluctuating and there is no sharp transition point. Second, gg is symmetric and unimodal (non-increasing on [0,∞)[0,\infty)); otherwise several locked clusters or an oscillating order parameter can appear. Third, global coupling: on sparse networks the threshold also depends on the network structure (Chapter 2). Fourth, the argument shows only that a partially locked steady state exists for K>KcK > K_c; its stability must be shown separately. Strogatz and Mirollo proved that the incoherent state is neutrally stable for K<KcK < K_c and unstable for K>KcK > K_c.

The order of the transition depends on the frequency distribution

How rr grows above threshold depends on the shape of gg. For the Lorentzian g(ω)=γ/[π(γ2+ω2)]g(\omega) = \gamma/[\pi(\gamma^2+\omega^2)], Kc=2γK_c = 2\gamma and r=1−Kc/K,K≥Kc,r = \sqrt{1 – K_c/K}, \qquad K \ge K_c, holds exactly for every KK (Kuramoto 1984). Ott and Antonsen showed in 2008 that for the Lorentzian the macroscopic dynamics of the whole population reduce to an ordinary differential equation for rr, whose steady state is this result (Ott and Antonsen 2008). A general smooth unimodal distribution gives only a square-root law near threshold, r≈16πKc3⋅μ−g″(0),μ=K−KcKc,r \approx \sqrt{\frac{16}{\pi K_c^3}\cdot\frac{\mu}{-g''(0)}}, \qquad \mu = \frac{K – K_c}{K_c}, a continuous supercritical bifurcation when g″(0)<0g''(0) < 0 — a second-order transition in the language of statistical physics (Strogatz 2000). The uniform distribution has g″(0)=0g''(0) = 0 and jumps at its edges, and Pazó proved in 2005 that the transition is then first order: rr jumps from 0 to π/4\pi/4 at KcK_c, and r−π/4r – \pi/4 grows as (K−Kc)2/3(K-K_c)^{2/3} thereafter (Pazó 2005). At KcK_c, KrKr equals exactly the half-width of the frequency distribution and all oscillators lock at once, with no intermediate partially locked stage; for the uniform distribution, therefore, KcK_c is the threshold at which incoherence loses stability and a fully locked state appears, and the phrase “a partially locked state appears” in the derivation above applies only to smooth unimodal distributions.

The program solves the self-consistency equation numerically for three distributions of scale 1 (Lorentzian with γ=1\gamma = 1, Gaussian with σ=1\sigma = 1, uniform on [−1,1][-1, 1]): for a given KK it first checks whether the right-hand side exceeds 1 as r→0+r \to 0^+, then finds rr by bisection. The Lorentzian result is compared with the exact solution, the Gaussian with the square-root law, and the uniform case is checked for the jump to π/4\pi/4 and the exponent 2/32/3.

Calculation

# Kuramoto model, N -> infinity: partially locked state from the self-consistency
# condition 1 = K * integral_{-pi/2}^{pi/2} cos^2(t) g(K r sin t) dt  (r > 0).
# Three symmetric unimodal frequency densities, each of scale 1:
# Lorentz (gamma = 1), Gauss (sigma = 1), uniform on [-1, 1].
import math

DENS = {"Lorentz": lambda w: 1 / (math.pi * (1 + w * w)),
        "Gauss": lambda w: math.exp(-w * w / 2) / math.sqrt(2 * math.pi),
        "uniform": lambda w: 0.5 if abs(w) <= 1 else 0.0}

def F(K, r, g, m=4000):              # midpoint rule (handles the uniform cut-off)
    h = math.pi / m
    return K * h * sum(math.cos(t) ** 2 * g(K * r * math.sin(t))
                       for t in (-math.pi / 2 + (j + 0.5) * h for j in range(m)))

def r_of_K(K, g):
    if F(K, 1e-9, g) <= 1:           # incoherence is the only state (K <= K_c)
        return 0.0
    lo, hi = 1e-9, 1.0               # F decreases with r for these densities
    for _ in range(50):
        mid = (lo + hi) / 2
        lo, hi = (mid, hi) if F(K, mid, g) > 1 else (lo, mid)
    return (lo + hi) / 2

if __name__ == "__main__":
    Kc = {k: 2 / (math.pi * g(0)) for k, g in DENS.items()}
    print("K_c = 2/(pi g(0)):  " + "   ".join(f"{k} {v:.4f}" for k, v in Kc.items()))
    print("  K/K_c   Lorentz  sqrt(1-Kc/K)    Gauss  sqrt-law    uniform")
    for x in (0.95, 1.0001, 1.01, 1.05, 1.10, 1.25, 1.50, 2.00, 3.00):
        rl, rg, ru = (r_of_K(x * Kc[k], DENS[k]) for k in DENS)
        exact = math.sqrt(max(0.0, 1 - 1 / x))
        # small-mu law: r ~ sqrt(16 mu / (pi Kc^3 (-g''(0)))), g''(0) = -1/sqrt(2 pi)
        mu = x - 1
        law = math.sqrt(16 * max(mu, 0) / (math.pi * Kc["Gauss"] ** 3
                                           / math.sqrt(2 * math.pi)))
        print(f"  {x:6.4f}   {rl:.4f}     {exact:.4f}      {rg:.4f}   {law:.4f}"
              f"     {ru:.4f}")
    print(f"uniform: r jumps from 0 to pi/4 = {math.pi / 4:.4f} at K_c (first order)")
    for x in (1.0001, 1.001, 1.01):     # Pazo 2005: r - pi/4 ~ (K - K_c)^(2/3)
        ru = r_of_K(x * Kc["uniform"], DENS["uniform"])
        q = (ru - math.pi / 4) / (x - 1) ** (2 / 3)
        print(f"   K/K_c = {x}: r - pi/4 = {ru - math.pi / 4:.5f},"
              f"  ratio to (K/K_c - 1)^(2/3) = {q:.3f}")
# K_c = 2/(pi g(0)):  Lorentz 2.0000   Gauss 1.5958   uniform 1.2732
#   K/K_c   Lorentz  sqrt(1-Kc/K)    Gauss  sqrt-law    uniform
#   0.9500   0.0000     0.0000      0.0000   0.0000     0.0000
#   1.0001   0.0100     0.0100      0.0177   0.0177     0.7868
#   1.0100   0.0995     0.0995      0.1755   0.1772     0.8108
#   1.0500   0.2182     0.2182      0.3775   0.3963     0.8480
#   1.1000   0.3015     0.3015      0.5099   0.5605     0.8734
#   1.2500   0.4472     0.4472      0.7123   0.8862     0.9144
#   1.5000   0.5774     0.5774      0.8497   1.2533     0.9463
#   2.0000   0.7071     0.7071      0.9370   1.7725     0.9721
#   3.0000   0.8165     0.8165      0.9762   2.5066     0.9879
# uniform: r jumps from 0 to pi/4 = 0.7854 at K_c (first order)
#    K/K_c = 1.0001: r - pi/4 = 0.00142,  ratio to (K/K_c - 1)^(2/3) = 0.657
#    K/K_c = 1.001: r - pi/4 = 0.00622,  ratio to (K/K_c - 1)^(2/3) = 0.622
#    K/K_c = 1.01: r - pi/4 = 0.02540,  ratio to (K/K_c - 1)^(2/3) = 0.547

Three points stand out. First, the thresholds of the three distributions are 2, 1.5958 and 1.2732, each equal to 2/(πg(0))2/(\pi g(0)); at K/Kc=0.95K/K_c = 0.95 all three have only the incoherent solution. Second, the numerical Lorentzian solution agrees with 1−Kc/K\sqrt{1 – K_c/K} to four decimals. For the Gaussian at K/Kc=1.01K/K_c = 1.01, r=0.1755r = 0.1755 against 0.1772 from the square-root law, a difference of about 1 %; but at K/Kc=1.10K/K_c = 1.10 the square-root law (0.5605) already exceeds the true value (0.5099) by about 10 %, so the approximation is useful only close to threshold. Third, for the uniform distribution rr is already 0.7868 at K/Kc=1.0001K/K_c = 1.0001, just above π/4=0.7854\pi/4 = 0.7854; the ratio of r−π/4r – \pi/4 to (K/Kc−1)2/3(K/K_c – 1)^{2/3} tends to a constant as the threshold is approached (0.547, 0.622, 0.657), in line with Pazó’s exponent. So “the Kuramoto transition is second order” holds only for one class of frequency distributions.

Finite size: no sharp transition

Real crowds and cell populations are finite. At finite NN, even without coupling, the average of NN random phases is not 0 but a random number of order N−1/2N^{-1/2}; coupling amplifies these fluctuations.

The program simulates the Kuramoto model in mean-field form for N=50N = 50 and N=400N = 400, with natural frequencies drawn at random from the standard normal distribution (three draws for each NN, fixed seeds), integrates with Heun’s method to t=120t = 120, averages rr over t=60t = 60 to 120, and compares the result with the self-consistent solution for N→∞N \to \infty.

Calculation

# Finite-N Kuramoto simulation in mean-field form:
#   d theta_i/dt = omega_i + K r sin(psi - theta_i),  r e^{i psi} = mean_j e^{i theta_j}
# omega_i ~ Gauss(0, 1), random draws (fixed seeds); Heun steps, dt = 0.05; r is averaged
# over t = 60..120 and over 3 frequency draws. Theory: N -> infinity self-consistency.
import math, random

def theory_r(K):              # Gauss sigma = 1: bisection on the self-consistency
    def F(r, m=2000):
        h = math.pi / m
        return K * h * sum(math.cos(t) ** 2 * math.exp(-(K * r * math.sin(t)) ** 2 / 2)
                           for t in (-math.pi / 2 + (j + .5) * h for j in range(m))) \
            / math.sqrt(2 * math.pi)
    if F(1e-9) <= 1:
        return 0.0
    lo, hi = 1e-9, 1.0
    for _ in range(40):
        mid = (lo + hi) / 2
        lo, hi = (mid, hi) if F(mid) > 1 else (lo, mid)
    return lo

def order(th):
    c = sum(math.cos(t) for t in th) / len(th); s = sum(math.sin(t) for t in th) / len(th)
    return math.hypot(c, s), math.atan2(s, c)

def simulate(N, K, seed, dt=0.05, t_end=120.0, t_avg=60.0):
    rng = random.Random(seed)
    om = [rng.gauss(0, 1) for _ in range(N)]
    th = [rng.uniform(-math.pi, math.pi) for _ in range(N)]
    acc, cnt = 0.0, 0
    for k in range(int(t_end / dt)):
        r, p = order(th)
        k1 = [o + K * r * math.sin(p - t) for o, t in zip(om, th)]
        pr = [t + dt * d for t, d in zip(th, k1)]
        r2, p2 = order(pr)
        th = [t + dt / 2 * (d + o + K * r2 * math.sin(p2 - q))
              for t, d, o, q in zip(th, k1, om, pr)]
        if k * dt >= t_avg:
            acc += r2; cnt += 1
    return acc / cnt

if __name__ == "__main__":
    Kc = 2 / (math.pi / math.sqrt(2 * math.pi))
    print(f"Gauss(0,1) frequencies: K_c = {Kc:.4f};  1/sqrt(N): N=50 {50 ** -.5:.3f},"
          f" N=400 {400 ** -.5:.3f}")
    print("  K/K_c   theory r   N = 50   N = 400")
    for x in (0.50, 0.90, 1.00, 1.10, 1.25, 1.50, 2.00):
        K = x * Kc
        sims = [sum(simulate(N, K, 100 * N + s) for s in range(3)) / 3 for N in (50, 400)]
        print(f"  {x:5.2f}    {theory_r(K):.3f}     {sims[0]:.3f}    {sims[1]:.3f}")
# Gauss(0,1) frequencies: K_c = 1.5958;  1/sqrt(N): N=50 0.141, N=400 0.050
#   K/K_c   theory r   N = 50   N = 400
#    0.50    0.000     0.210    0.067
#    0.90    0.000     0.391    0.148
#    1.00    0.000     0.552    0.336
#    1.10    0.510     0.632    0.538
#    1.25    0.712     0.719    0.730
#    1.50    0.850     0.878    0.862
#    2.00    0.937     0.944    0.940

Well above threshold (K/Kc≥1.25K/K_c \ge 1.25) both sizes agree with theory to within 0.03. Near threshold the differences are large: at K/Kc=1K/K_c = 1 the theoretical value is 0, but the simulation gives 0.336 for N=400N = 400 and 0.552 for N=50N = 50; even at K/Kc=0.5K/K_c = 0.5 the mean rr for N=50N = 50 is 0.210, larger than the typical value without coupling (of order N−1/2=0.141N^{-1/2} = 0.141). In a group of a few dozen, the “transition point” is a blurred band, and KcK_c cannot be treated as a precise switch. Volume 8 has already checked Kc=2γK_c = 2\gamma for the Lorentzian by simulation, and that is not repeated here.

Left: the order parameter r against K/K_c for N \to \infty under three frequency distributions; the Lorentzian and Gaussian give continuous transitions, the uniform distribution jumps to \pi/4 at K_c. Right: the Gaussian theory curve and finite simulations with N = 50 and N = 400 (one draw each); the smaller the group, the blurrier the threshold.

Why applause comes together and falls apart

Evidence

The applause studies of Néda and colleagues. Néda, Ravasz, Brechet, Vicsek and Barabási recorded audiences at theatre and opera performances in Romania and Hungary and found that applause turns from noise into a regular beat, lasts for a while, falls apart, and repeats (Néda, Ravasz, Brechet, et al. 2000). They had 73 secondary-school students clap separately and measured each one’s “fast” and “slow” clapping frequencies: the clapping period during synchrony was about twice that outside it, and the distribution of the slow mode was much narrower than that of the fast mode. The longer paper of the same year gives a ratio of dispersions DI/DII≈2.5D_I/D_{II} \approx 2.5 and a ratio of relative dispersions of about 1.3, and explains the effect with Kuramoto’s result Kc∝DK_c \propto D: in fast clapping the dispersion is large, K<KcK < K_c, and synchrony is impossible; once the audience doubles its period the dispersion shrinks, KcK_c falls below KK, and partial synchrony appears (Néda, Ravasz, Vicsek, et al. 2000). They also noted that during synchrony the number of claps per unit time halves and the average sound intensity in the room falls; wishing to be louder, the audience speeds up, the dispersion returns and synchrony breaks down — maximal synchrony and maximal noise cannot be had together.

The program uses the simulation parameters of the longer paper of Néda and colleagues: 70 clappers, coupling K=0.8s−1K = 0.8\ \mathrm{s^{-1}}; in the fast mode the mean angular frequency is 2πs−12\pi\ \mathrm{s^{-1}} with dispersion D=2π/6.9s−1D = 2\pi/6.9\ \mathrm{s^{-1}}; in the slow mode the period doubles and both the mean and the dispersion halve (a simplification: the measured ratio of dispersions is about 2.5, not 2). Each mode runs for 40 seconds; the program records the mean rr over the last 20 seconds, the claps per second, and the ratio of the peak to the mean of the number of claps in 0.05-second bins (the larger, the more like a regular beat). Ten audiences are simulated.

Calculation

# Rhythmic applause after Neda et al. (2000): N = 70 clappers, coupling K = 0.8 s^-1.
# Mode I (fast): mean 2 pi s^-1, dispersion D = 2 pi/6.9 s^-1 (their simulation values).
# Mode II: period doubled, so mean and D halve. Gauss: K_c = 2/(pi g(0)) = D sqrt(8/pi).
# Each phase lasts 40 s (Euler, dt = 0.01 s); r averaged over its last 20 s; 10 audiences.
# A clap = a phase crossing a multiple of 2 pi; "noise" = claps per second; "signal" =
# peak of the clap count in 0.05 s bins over its mean (1 for perfectly even clapping).
import math, random

N, K, DT = 70, 0.8, 0.01

def run(seed):
    rng = random.Random(seed)
    z = [rng.gauss(0, 1) for _ in range(N)]          # each person's place in the spread
    th = [rng.uniform(0, 2 * math.pi) for _ in range(N)]
    out = []
    for mean, D in ((2 * math.pi, 2 * math.pi / 6.9), (math.pi, math.pi / 6.9)):
        om = [mean + D * x for x in z]
        rs, bins, claps, steps = [], [], 0, int(40 / DT)
        for k in range(steps):
            c = sum(math.cos(t) for t in th) / N; s = sum(math.sin(t) for t in th) / N
            r, p = math.hypot(c, s), math.atan2(s, c)
            new = [t + DT * (o + K * r * math.sin(p - t)) for t, o in zip(th, om)]
            n = sum(math.floor(b / (2 * math.pi)) > math.floor(a / (2 * math.pi))
                    for a, b in zip(th, new))
            th = new
            if k >= steps // 2:
                rs.append(r); claps += n
                if k % 5 == 0: bins.append(0)
                bins[-1] += n
        mean_bin = sum(bins) / len(bins)
        out.append((sum(rs) / len(rs), claps / 20, max(bins) / mean_bin))
    return out

if __name__ == "__main__":
    for name, D in (("I ", 2 * math.pi / 6.9), ("II", math.pi / 6.9)):
        kc = D * math.sqrt(8 / math.pi)
        print(f"mode {name}: D = {D:.3f} s^-1, K_c = {kc:.3f} s^-1"
              f"  (K = {K})")
    res = [run(s) for s in range(10)]
    for m, name in enumerate(("I ", "II")):
        r = sum(x[m][0] for x in res) / 10; n = sum(x[m][1] for x in res) / 10
        pk = sum(x[m][2] for x in res) / 10
        print(f"mode {name}: mean r = {r:.3f}   claps per second = {n:5.1f}"
              f"   peak/mean of 0.05 s bins = {pk:.2f}")
    print(f"incoherent reference 1/sqrt(N) = {N ** -0.5:.3f}")
# mode I : D = 0.911 s^-1, K_c = 1.453 s^-1  (K = 0.8)
# mode II: D = 0.455 s^-1, K_c = 0.727 s^-1  (K = 0.8)
# mode I : mean r = 0.179   claps per second =  69.7   peak/mean of 0.05 s bins = 3.13
# mode II: mean r = 0.555   claps per second =  34.7   peak/mean of 0.05 s bins = 5.36
# incoherent reference 1/sqrt(N) = 0.120

In the fast mode Kc=1.453>K=0.8K_c = 1.453 > K = 0.8 and the mean rr is only 0.179, close to the level of 70 random phases (of order N−1/2=0.120N^{-1/2} = 0.120); in the slow mode Kc=0.727<0.8K_c = 0.727 < 0.8, the mean rr rises to 0.555 and the peak-to-mean ratio of clapping rises from 3.13 to 5.36, but the claps per second fall from 69.7 to 34.7. The model reproduces the trade-off Néda and colleagues described: synchrony needs slowing down, and slowing down makes the applause quieter. KK here is an illustrative value; the program does not estimate the real coupling between members of an audience.

Hypothesis · not established · Slowing the rhythm synchronises more easily than strengthening coupling

The Kuramoto threshold Kc=2/(πg(0))K_c = 2/(\pi g(0)) depends on both coupling and dispersion. The author’s conjecture is that in group activities that need people to move in step — choirs, reading aloud together, group drills — lowering the tempo (and with it the absolute spread of people’s natural rhythms) brings the whole group into step more easily than asking people to “pay more attention to one another” (increasing coupling). This can be tested: in randomly assigned clapping or tapping tasks, compare the phase coherence rr under “beat slowed by half” and under “extra visual cues to strengthen coupling”. If, at equal dispersion, strengthening coupling raises coherence clearly more than slowing the tempo, the conjecture is wrong. This is only an analogy for group activities and implies no advice about managing organisations.

Summary

The Kuramoto model uses the complex order parameter to turn an NN-body problem into a mean-field problem and, under global coupling, N→∞N \to \infty and a symmetric unimodal frequency distribution, gives the critical coupling Kc=2/(πg(0))K_c = 2/(\pi g(0)). For the Lorentzian, r=1−Kc/Kr = \sqrt{1 – K_c/K} holds exactly; general smooth distributions give a continuous transition with a square-root law near threshold; the uniform distribution gives a first-order transition in which rr jumps to π/4\pi/4. At finite size the threshold is smeared by fluctuations. The key to synchronised applause is that doubling the period shrinks the frequency dispersion, while synchrony lowers the average loudness, the audience speeds up, and synchrony collapses. The next chapter turns to a risk that synchrony can bring: the London Millennium Bridge.

References

Acebrón, Juan A., Luis L. Bonilla, Conrad J. Pérez Vicente, Félix Ritort, and Renato Spigler. 2005. “The Kuramoto Model: A Simple Paradigm for Synchronization Phenomena.” Reviews of Modern Physics 77 (1): 137–85. https://doi.org/10.1103/RevModPhys.77.137.
Kuramoto, Yoshiki. 1975. “Self-Entrainment of a Population of Coupled Non-Linear Oscillators.” In International Symposium on Mathematical Problems in Theoretical Physics, edited by Huzihiro Araki, vol. 39. Lecture Notes in Physics. Springer. https://doi.org/10.1007/BFb0013365.
Kuramoto, Yoshiki. 1984. Chemical Oscillations, Waves, and Turbulence. Vol. 19. Springer Series in Synergetics. Springer.
Néda, Zoltán, Erzsébet Ravasz, Yves Brechet, Tamás Vicsek, and Albert-László Barabási. 2000. “The Sound of Many Hands Clapping.” Nature 403 (6772): 849–50. https://doi.org/10.1038/35002660.
Néda, Zoltán, Erzsébet Ravasz, Tamás Vicsek, Yves Brechet, and Albert-László Barabási. 2000. “Physics of the Rhythmic Applause.” Physical Review E 61 (6): 6987–92. https://doi.org/10.1103/PhysRevE.61.6987.
Ott, Edward, and Thomas M. Antonsen. 2008. “Low Dimensional Behavior of Large Systems of Globally Coupled Oscillators.” Chaos 18 (3): 037113.
Pazó, Diego. 2005. “Thermodynamic Limit of the First-Order Phase Transition in the Kuramoto Model.” Physical Review E 72: 046211. https://doi.org/10.1103/PhysRevE.72.046211.
Strogatz, Steven H. 2000. “From Kuramoto to Crawford: Exploring the Onset of Synchronization in Populations of Coupled Oscillators.” Physica D 143: 1–20. https://doi.org/10.1016/S0167-2789(00)00094-4.
Winfree, Arthur T. 1967. “Biological Rhythms and the Behavior of Populations of Coupled Oscillators.” Journal of Theoretical Biology 16 (1): 15–42.

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).