5D Library › Collective Dynamics, Coupled-Oscillator Synchrony and Distributed Consensus · 中文
Chapter 2 The Wobbling Millennium Bridge and the Master Stability Function of Network Synchrony
A bridge that swayed on its opening day
Evidence
The engineering report. The report of Dallard and colleagues in The Structural Engineer in 2001 records that the London Millennium Footbridge opened on 10 June 2000 and that on that day the deck moved sideways unexpectedly as crowds crossed (Dallard et al. 2001). The Arup engineers who investigated concluded that the lateral force exerted by pedestrians depends on the motion of the deck itself: when the bridge sways, walkers adjust their gait to keep their balance, and that adjustment feeds energy back into the bridge. The report states that the phenomenon is not peculiar to this design and could occur on any bridge with a lateral frequency below about 1.3 Hz carrying enough people; the retrofit added fluid-viscous dampers and tuned mass dampers to raise the damping above 20 % of critical, and tests on a prototype matched the analytical predictions.
The engineering tests found that the swaying started abruptly once the number of people on the bridge exceeded a critical value. Later literature reports this critical number differently: Lenci and Marcheggiani, quoting the tests of Dallard and colleagues, say that the critical number for the north span was about 155 (Lenci and Marcheggiani 2011); Belykh and colleagues used about 165 as the observed value to calibrate their model (Belykh et al. 2017). The book adopts no single figure and says only “about 155 to 165 people”.
The synchronisation model of Strogatz and colleagues
Strogatz, Abrams, McRobie, Eckhardt and Ott proposed in Nature in 2005 that the sway of the Millennium Bridge can be seen as an instance of Kuramoto synchronisation: the bridge is a driven damped oscillator, each pedestrian is a phase oscillator, the bridge’s motion pulls walkers’ steps towards synchrony, and synchronised steps drive the bridge harder (Strogatz et al. 2005). The model reads where is the lateral displacement of the bridge; , and are the modal mass, damping and stiffness of that mode; is the amplitude of each pedestrian’s lateral force; is the gait phase of pedestrian (one left and one right step make a cycle); is that walker’s natural stepping frequency; and are the amplitude and phase of the bridge’s vibration; measures walkers’ sensitivity to the sway; and is a phase lag. They took the simplest case , and obtained by fitting engineering test data. Eckhardt and colleagues gave a fuller analysis in 2007 (Eckhardt et al. 2007).
Calculation
Deriving the critical crowd size (derived in this book from the model above). Put with and , and average over one cycle of the bridge (the damping is small, so and change slowly). This gives the “slow-flow” equations Walkers with lock at the stable point , where : they feed energy into the bridge; walkers who are not locked contribute zero to the time average of . For small the contribution of the locked walkers is about ( being the density of stepping frequencies), so Put differently, each walker contributes on average a negative damping of ; once the number exceeds the total damping turns negative and the sway begins to grow.
The program simulates these slow-flow equations. The modal mass of the north span kg, stiffness N/m and damping kg/s are the values Lenci and Marcheggiani quote from Strogatz and colleagues (Lenci and Marcheggiani 2011); is from Strogatz and colleagues; the force amplitude per walker N and the standard deviation of stepping frequencies 0.63 rad/s are illustrative values. Starting from 100 walkers, 10 are added every 150 seconds up to 200; five randomly drawn crowds are simulated, recording the bridge amplitude, the walkers’ order parameter and the number of locked walkers.
Calculation
# Millennium Bridge model of Strogatz et al. (2005), averaged ("slow-flow") form.
# M X'' + B X' + K X = G sum_i sin(Theta_i)
# Theta_i' = Omega_i + C A sin(Psi - Theta_i + a)
# With X = A sin(Psi), Psi = Omega0 t + psi and a = pi/2, averaging over one bridge cycle:
# A' = -(B / 2M) A + (G / 2 M Omega0) sum_i sin(phi_i), phi_i = Theta_i - Psi
# psi' = -(G / 2 A M Omega0) sum_i cos(phi_i)
# phi_i' = Omega_i - Omega0 - psi' + C A cos(phi_i)
# Locked walkers inject energy; linear threshold N_c = 2 B Omega0 / (pi G C P(Omega0)).
# M, K, B: north-span values quoted by Lenci & Marcheggiani (2011) from Strogatz et al.;
# C = 16 /(m s) from Strogatz et al.; G = 30 N, SD(Omega) = 0.63 rad/s are ILLUSTRATIVE.
import math, random
M, KS, B, C, G, SD = 113000.0, 4.73e6, 11000.0, 16.0, 30.0, 0.63
W0 = math.sqrt(KS / M)
def crowd(n_max, seed=2000, hold=150.0, dt=0.05, step=10, n0=100):
rng = random.Random(seed)
nu = [rng.gauss(0, SD) for _ in range(n_max)] # Omega_i - Omega0
phi = [rng.uniform(-math.pi, math.pi) for _ in range(n_max)]
A, rows = 0.001, [] # 1 mm of ambient motion
for n in range(n0, n_max + 1, step):
for _ in range(int(hold / dt)):
s = sum(math.sin(p) for p in phi[:n]); c = sum(math.cos(p) for p in phi[:n])
dpsi = -G * c / (2 * A * M * W0)
A += dt * (-B / (2 * M) * A + G * s / (2 * M * W0))
A = max(A, 1e-4)
phi[:n] = [p + dt * (v - dpsi + C * A * math.cos(p))
for p, v in zip(phi[:n], nu[:n])]
R = math.hypot(sum(math.cos(p) for p in phi[:n]), s) / n
locked = sum(abs(v - dpsi) < C * A for v in nu[:n])
rows.append((n, 1000 * A, R, locked))
return rows
if __name__ == "__main__":
P0 = 1 / (SD * math.sqrt(2 * math.pi))
zeta = B / (2 * M * W0)
print(f"Omega0 = {W0:.3f} rad/s ({W0 / (2 * math.pi):.3f} Hz), damping ratio"
f" {100 * zeta:.2f} %, P(Omega0) = {P0:.3f} s")
print(f"linear threshold N_c = 2 B Omega0 / (pi G C P) = "
f"{2 * B * W0 / (math.pi * G * C * P0):.1f} walkers")
runs = [crowd(200, seed=s) for s in range(5)] # 5 independent crowds
print("walkers amplitude (mm): mean of 5 crowds [min - max] walker order R locked")
for k, (n, *_rest) in enumerate(runs[0]):
a = [r[k][1] for r in runs]; R = sum(r[k][2] for r in runs) / 5
lk = sum(r[k][3] for r in runs) / 5
print(f" {n:3d} {sum(a) / 5:6.2f} [{min(a):6.2f} - {max(a):6.2f}]"
f" {R:.3f} {lk:5.1f}")
# Omega0 = 6.470 rad/s (1.030 Hz), damping ratio 0.75 %, P(Omega0) = 0.633 s
# linear threshold N_c = 2 B Omega0 / (pi G C P) = 149.1 walkers
# walkers amplitude (mm): mean of 5 crowds [min - max] walker order R locked
# 100 2.37 [ 1.04 - 4.47] 0.117 7.6
# 110 2.33 [ 0.49 - 5.68] 0.079 6.4
# 120 4.30 [ 0.98 - 11.23] 0.159 14.0
# 130 8.44 [ 1.89 - 16.82] 0.193 26.2
# 140 12.98 [ 1.08 - 27.22] 0.292 38.0
# 150 17.40 [ 2.49 - 33.35] 0.292 51.6
# 160 22.09 [ 2.17 - 38.78] 0.327 63.2
# 170 28.49 [ 3.29 - 47.94] 0.384 82.0
# 180 40.66 [ 5.75 - 56.68] 0.547 119.6
# 190 53.71 [ 25.16 - 63.28] 0.688 152.6
# 200 66.42 [ 62.03 - 69.13] 0.792 181.0With these parameters the lateral frequency of the bridge is 1.030 Hz, the damping ratio 0.75 %, and the critical crowd for is 149.1 walkers. This is of the same order as the 155 to 165 people observed in engineering practice, but it cannot be taken as independent confirmation of the model: was fitted to test data, and and the dispersion of stepping frequencies are illustrative; changing them changes in proportion.
The simulation shows two more noteworthy points. First, a finite crowd has no sharp threshold: with 100 walkers the mean amplitude over the five crowds is already 2.37 mm, with 150 it is 17.40 mm but varies widely between crowds (2.49 to 33.35 mm); with 200 it is 66.42 mm and all five crowds exceed 62 mm. Second, the amplitude and the walkers’ synchrony grow together: with 100 walkers , near the random level; with 150, 0.292; with 200, 0.792, while the number of locked walkers rises from about 8 to 181. In this model only a small fraction of walkers is locked when the instability starts; the negative damping they supply is enough to start the sway, and the sway then pulls more people into lock. So even in the synchronisation model of Strogatz and colleagues, “first everyone synchronises, then the bridge sways” is not the right picture.
Later criticism
Evidence
Is synchrony the cause or the result? Belykh, Jeter and Belykh in 2017 studied the question with biomechanically inspired inverted-pendulum models of gait and argued that phase-locking of pedestrians was necessary for the large-amplitude wobbling of the Millennium Bridge, but that the initial small wobble may come with many phase slips and unsynchronised adjustments, so that the popular account “synchrony triggers the wobble” may be an oversimplification; their model becomes unstable at about 165 people, and the wobble stops only when the crowd shrinks to about 135, a hysteresis (Belykh et al. 2017). Belykh, Bocian, Champneys, Daley, Jeter, Macdonald and McRobie went further in Nature Communications in 2021: pedestrians who are not synchronised can, on average, already give the bridge negative damping; the total damping can be written with critical crowd ; and synchrony of footsteps appears mainly after the instability, as a consequence rather than a cause (Belykh et al. 2021).
The two explanations do not conflict mathematically: this chapter’s derivation likewise writes the critical condition as “negative damping per walker times the number of walkers equals the bridge’s damping”. They differ on where the negative damping comes from — from the few who lock, or from the average effect of every walker adjusting foot placement to keep balance. The latter conclusion matters more for engineering: with enough people, a bridge can lose stability even if walkers do not coordinate at all.
Hypothesis · not established · The sway precedes synchrony
The author’s conjecture (following the view of Belykh and colleagues in 2021): when people are added gradually to a footbridge whose lateral frequency is below about 1.3 Hz, the moment at which the bridge’s amplitude starts to grow persistently comes before the moment at which the walkers’ gait order parameter rises clearly above the random level (for example above ). This can be tested in field crowd tests with motion sensors. If in most tests rises significantly first and the amplitude grows only afterwards, the conjecture is wrong.
Synchrony on networks: the master stability function
The bridge and its walkers are globally coupled; in many systems coupling runs along a network. Pecora and Carroll showed in 1990 that two chaotic systems can synchronise (Pecora and Carroll 1990); in 1998 they proposed a general method for deciding whether identical oscillators on a network can synchronise — the master stability function (Pecora and Carroll 1998).
Calculation
The master stability function. Let identical systems obey , where is the network Laplacian (rows summing to 0) with eigenvalues . Linearise about the synchronous state and decompose along the eigenvectors of ; each transverse mode satisfies These equations have the same form and differ only in the parameter . Write for the largest Lyapunov exponent of this equation; this is the master stability function, and the synchronous state is stable if and only if for every . If the stable region is an unbounded interval , only is needed, and strong enough coupling always synchronises; if it is a bounded interval , both and are needed, and a suitable exists if and only if
The ratio condition above follows directly from the master stability function of Pecora and Carroll in 1998; using the eigenratio as an index for comparing the “synchronisability” of different networks is the work of Barahona and Pecora in 2002 (Barahona and Pecora 2002): they found that adding a few random shortcuts to a regular ring markedly lowers the ratio, and that small-world networks synchronise about as well as random graphs. The framework’s attribution of this criterion to Pecora and Carroll 1998 is not wrong, but the 2002 source should be added; and it applies only when the stable region is bounded. A large ratio means two kinds of failure: too small — the network is like two blocks joined by a thin thread, and the slow mode does not converge; or too large — short-wavelength modes are pushed out of the stable region by too strong a coupling. Heagy, Pecora and Carroll identified this second “short-wavelength bifurcation” in 1995: for a regular ring, adding oscillators increases , so a ring that could synchronise loses synchrony when more oscillators are added (Heagy et al. 1995). Nishikawa and colleagues found in 2003 that the more heterogeneous the degree distribution, the harder a network is to synchronise, even when its average path is shorter — the “paradox of heterogeneity” (Nishikawa et al. 2003). The review of Arenas and colleagues in 2008 summarises these results (Arenas et al. 2008).
An exactly solvable bounded example. Coupled logistic maps with . The Lyapunov exponent of is , and the -th transverse mode is multiplied by at each step, where are the eigenvalues of the normalised Laplacian (Jost and Joy 2002). Stability requires for all , so a synchronising exists if and only if . The program finds the eigenvalues of several networks by the Jacobi method, computes the interval of synchronising (also requiring , which keeps inside ), takes the that minimises the growth of the worst mode, then perturbs the synchronous state by and iterates 3,000 steps to test the prediction.
Calculation
# Master stability function for coupled chaotic maps (an exactly solvable bounded case).
# x_i(t+1) = (1 - e) f(x_i) + (e / k_i) sum_{j ~ i} f(x_j), f(x) = 4 x (1 - x)
# Transverse mode k grows by |1 - e lambda_k| e^h, h = ln 2 (Lyapunov exponent of f), and
# lambda_k the eigenvalues of the normalised Laplacian. Stable iff e lambda_k lies in
# (1 - 1/2, 1 + 1/2) for all k >= 2: a window exists iff lambda_N / lambda_2 < 3.
import math, random
def jacobi_eigs(S, sweeps=60): # eigenvalues of a symmetric matrix
a = [row[:] for row in S]; n = len(a)
for _ in range(sweeps):
off = sum(a[i][j] ** 2 for i in range(n) for j in range(n) if i != j)
if off < 1e-22: break
for p in range(n - 1):
for q in range(p + 1, n):
if abs(a[p][q]) < 1e-15: continue
t = (a[q][q] - a[p][p]) / (2 * a[p][q])
t = (1 if t >= 0 else -1) / (abs(t) + math.sqrt(t * t + 1))
c = 1 / math.sqrt(t * t + 1); s = t * c
for k in range(n):
u, v = a[k][p], a[k][q]
a[k][p], a[k][q] = c * u - s * v, s * u + c * v
for k in range(n):
u, v = a[p][k], a[q][k]
a[p][k], a[q][k] = c * u - s * v, s * u + c * v
return sorted(a[i][i] for i in range(n))
def norm_lap(adj): # I - D^-1/2 A D^-1/2, same spectrum as I - D^-1 A
n = len(adj); d = [len(x) for x in adj]
return [[(i == j) - (1 / math.sqrt(d[i] * d[j]) if j in adj[i] else 0.0)
for j in range(n)] for i in range(n)]
def ring(n, k=1):
return [{(i + s) % n for s in range(-k, k + 1) if s} for i in range(n)]
def graphs():
rng = random.Random(7); sw, sw2 = ring(20, 2), ring(20, 2)
for m, g in ((6, sw), (40, sw2)): # random shortcuts
for _ in range(m):
i, j = rng.sample(range(20), 2); g[i].add(j); g[j].add(i)
two = [set() for _ in range(20)] # two complete 10-cliques joined by one edge
for i in range(20):
two[i] = {j for j in range(20) if j != i and (j < 10) == (i < 10)}
two[0].add(10); two[10].add(0)
star = [set(range(1, 20))] + [{0} for _ in range(19)]
full = [set(range(20)) - {i} for i in range(20)]
return [("ring N=5", ring(5)), ("ring N=6", ring(6)), ("ring N=20, k=2", ring(20, 2)),
("ring N=20, k=2 + 6 shortcuts", sw),
("ring N=20, k=2 + 40 shortcuts", sw2), ("two cliques + 1 bridge", two),
("star N=20", star), ("complete N=20", full)]
def sync_error(adj, e, steps=3000, seed=1):
rng = random.Random(seed); x0 = rng.random()
x = [x0 + 1e-7 * rng.uniform(-1, 1) for _ in adj]
for _ in range(steps):
fx = [4 * v * (1 - v) for v in x]
x = [(1 - e) * fx[i] + e * sum(fx[j] for j in nb) / len(nb)
for i, nb in enumerate(adj)]
return max(x) - min(x)
if __name__ == "__main__":
print("graph lam_2 lam_N ratio window for e best e"
" growth spread")
for name, adj in graphs():
lam = jacobi_eigs(norm_lap(adj)); l2, lN = lam[1], lam[-1]
lo, hi = 0.5 / l2, min(1.0, 1.5 / lN) # e <= 1 keeps x inside [0, 1]
e = min(1.0, 2 / (l2 + lN)) # minimises the worst mode
g = 2 * max(abs(1 - e * l2), abs(1 - e * lN)) # per-step growth, worst mode
win = f"({lo:.3f},{hi:.3f})" if lo < hi else " none "
print(f"{name:30s} {l2:5.3f} {lN:5.3f} {lN / l2:6.2f} {win} {e:5.3f}"
f" {g:5.3f} {sync_error(adj, e):.0e}")
print("spread = max - min of x after 3000 steps, from a 1e-7 perturbation of sync")
# graph lam_2 lam_N ratio window for e best e growth spread
# ring N=5 0.691 1.809 2.62 (0.724,0.829) 0.800 0.894 0e+00
# ring N=6 0.500 2.000 4.00 none 0.800 1.200 5e-01
# ring N=20, k=2 0.120 1.559 13.00 none 1.000 1.760 1e-01
# ring N=20, k=2 + 6 shortcuts 0.180 1.583 8.79 none 1.000 1.640 7e-01
# ring N=20, k=2 + 40 shortcuts 0.424 1.546 3.65 none 1.000 1.153 4e-01
# two cliques + 1 bridge 0.019 1.192 63.99 none 1.000 1.963 2e-01
# star N=20 1.000 2.000 2.00 (0.500,0.750) 0.667 0.667 0e+00
# complete N=20 1.053 1.053 1.00 (0.475,1.000) 0.950 0.000 0e+00
# spread = max - min of x after 3000 steps, from a 1e-7 perturbation of syncPrediction and simulation agree exactly: in networks whose growth factor is below 1 (the ring of 5, the star, the complete graph) the perturbation dies out and the final spread is 0; in networks whose growth factor exceeds 1 the spread grows beyond 0.1. Three results stand out. First, the “short-wavelength bifurcation”: the ring of 5 oscillators has a ratio of 2.62 and can synchronise; the ring of 6 has 4.00 and cannot. Second, shortcuts help but may not be enough: a ring of 20 in which each node links to two neighbours on each side has a ratio of 13.00; 6 random shortcuts bring it to 8.79 and 40 bring it to 3.65, still above 3. Two complete subgraphs joined by a single edge give a ratio of 63.99. Third, under normalised coupling (each node averages its neighbours) the star has ratio 2.00 and can synchronise; under unnormalised coupling the star’s Laplacian eigenvalues are and the ratio equals . Whether heterogeneity hinders synchrony depends on whether the coupling is normalised by degree — a modelling choice that must be stated whenever networks are compared.
Summary
The sway of the Millennium Bridge can be described by the synchronisation model of Strogatz and colleagues, whose critical crowd is equivalent to “negative damping per walker times the number of walkers equals the bridge’s damping”; the observed critical number is reported as 155 to 165 in different sources. Both the model and later research show that synchrony and sway grow together, and that the onset of the sway does not need everyone to be in step; the 2021 study argues that synchrony is a result of the instability. The master stability function separates dynamics from network structure; the criterion follows directly from the method of Pecora and Carroll in 1998 and was used by Barahona and Pecora in 2002 as an index of network synchronisability; it applies only when the stable region is bounded, and its verdict depends on whether coupling is normalised. The next chapter turns from oscillators to individuals that move.
References
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).