5D Library › Scaling Laws, Allometric Growth and the Statistical Physics of Organisational Metabolism · 中文

Chapter 3 Urban Scaling: Superlinearity, Finite-Time Singularities and Accelerating Innovation

A city is not an enlarged animal

Companies often fail, but cities seldom disappear. In 2007 Bettencourt, Lobo, Helbing, Kühnert and West carried the question of Chapter 1 over to cities: do urban indicators also change with population according to power laws, and if so, are the exponents greater or smaller than 1 (Bettencourt et al. 2007)?

Evidence

Three classes of urban exponent. Using data for cities in the United States, the European Union, Germany and China, Bettencourt and colleagues fitted Y=Y0NβY = Y_0 N^{\beta} to each indicator. The results fall roughly into three classes: quantities tied to individual needs (housing, employment, household electricity and water use) have β≈1\beta \approx 1; infrastructure is sublinear, for example petrol stations 0.77, petrol sales 0.79, length of electrical cables 0.87 and road surface 0.83; socioeconomic quantities are superlinear, for example new patents 1.27, inventors 1.25, private research and development employment 1.34, total wages 1.12, gross domestic product 1.13 to 1.26 and serious crimes 1.16 (Bettencourt et al. 2007). Bettencourt (2013) derived β=1±δ\beta = 1 \pm \delta from a network model of interactions among people in a city, with δ≈1/6\delta \approx 1/6 under a set of simplifying conditions, that is, about 1.17 and 0.83 (Bettencourt 2013).

The framework’s “infrastructure about 0.85, socioeconomic output about 1.15” is a rounded summary of these numbers. Unlike organisms, cities have socioeconomic exponents above 1: the larger the city, the greater the output per person and the faster the pace.

Calculation

The first part computes how totals and per-person quantities change when the population doubles. The second part builds a synthetic urban system: 400 metropolitan areas whose core cities follow an approximately Zipf distribution; output is recorded in the core city where people work, population is counted where they live, and the larger the core, the larger the share of commuters coming in from outside its administrative boundary (illustrative parameters: the ratio of commuter-belt residents to core residents rises from 0.10 to about 0.67). The true exponent at the metropolitan level is set to 1.00 or 1.12, and the data are then fitted against metropolitan population and against core-city population. The last line makes the commuter share independent of city size, as a control.

# Urban scaling Y = Y0 N^beta. Part 1: what a doubling of population does. Part 2: a
# SYNTHETIC urban system showing how the city boundary changes the fitted exponent: output
# is produced where people WORK, population is counted where they LIVE, and larger cores
# draw proportionally more commuters from outside their administrative boundary. All
# parameters are illustrative (not estimates for any country).
import math, random
from statistics import mean

def ols(xs, ys):
    mx, my = mean(xs), mean(ys)
    return (sum((x - mx) * (y - my) for x, y in zip(xs, ys))
            / sum((x - mx) ** 2 for x in xs))

if __name__ == "__main__":
    print("beta   doubling N multiplies Y by   per-capita change")
    for b in (0.85, 1.00, 1.15):
        print(f"{b:.2f}            {2 ** b:.2f}                {2 ** (b - 1) - 1:+.1%}")
    rng = random.Random(2007)
    cores = [20000 / rng.random() for _ in range(3000)]      # Zipf-like core sizes
    cores = [c for c in cores if c < 5e6][:400]
    print(f"{len(cores)} metros, cores {min(cores):,.0f} to {max(cores):,.0f} residents")
    print("true beta (metro)   fit on metro residents   fit on core residents only")
    for beta in (1.00, 1.12):
        xm, xc, ys = [], [], []
        for c in cores:
            s = 0.10 + 0.25 * math.log10(c / 2e4)            # commuter belt / core
            metro = c * (1 + s)
            y = metro ** beta * math.exp(rng.gauss(0, 0.15))  # output recorded in core
            xm.append(math.log(metro)); xc.append(math.log(c)); ys.append(math.log(y))
        print(f"      {beta:.2f}                 {ols(xm, ys):.3f}"
              f"                     {ols(xc, ys):.3f}")
    xc, ys = [], []                     # same, but belt share NOT linked to core size
    for c in cores:
        s = rng.uniform(0.10, 0.70)
        xc.append(math.log(c)); ys.append(math.log((c * (1 + s)) ** 1.12)
                                          + rng.gauss(0, 0.15))
    print(f"core-only fit when the belt share does not grow with size (true 1.12):"
          f" {ols(xc, ys):.3f}")
# beta   doubling N multiplies Y by   per-capita change
# 0.85            1.80                -9.9%
# 1.00            2.00                +0.0%
# 1.15            2.22                +11.0%
# 400 metros, cores 20,195 to 3,947,016 residents
# true beta (metro)   fit on metro residents   fit on core residents only
#       1.00                 1.006                     1.091
#       1.12                 1.129                     1.225
# core-only fit when the belt share does not grow with size (true 1.12): 1.125

When the population doubles, a quantity with exponent 0.85 becomes 1.80 times larger and one with exponent 1.15 becomes 2.22 times larger, in line with the framework. Per person, however, this is only −9.9 % and +11.0 %. Popular accounts often say that “doubling a city’s population raises wages per person by 15 % and saves 15 % of infrastructure per person”; this takes the 0.15 in the exponent directly as a percentage, and the accurate figures are about 11 % and 10 %.

The second part shows the role of city boundaries. If the true relation is linear (1.00), fitting against metropolitan population gives 1.006, but fitting against core-city population gives 1.091; if the true exponent is 1.12, the two definitions give 1.129 and 1.225. The bias arises because the output of large cities includes a growing contribution from commuters who are not counted in the core city’s population. If the commuter share does not depend on city size, the core-city definition gives 1.125 and the bias disappears. The model is illustrative, but it points to a real problem: part of the superlinearity may come from how the boundary is drawn rather than from the city itself.

Evidence

Criticisms of urban scaling. Arcaute and colleagues (2015) used thresholds of commuting and population density to construct many different systems of cities for England and Wales, and found that most indicators were approximately linear under all definitions and that, where nonlinearity appeared, the exponent varied widely with the definition of the boundary; they concluded that population size alone is not enough to describe or predict the state of a city (Arcaute et al. 2015). Leitão, Miotto, Gerlach and Altmann (2016) compared five statistical models of fluctuations on 15 data sets from five databases and found that the conclusion “the exponent differs from 1” depends on how fluctuations are modelled and on the heavy-tailed distribution of city sizes: for cinema attendance in the European Union, for example, one model gives an exponent of 1.46 and another 1.00; their tests produced linear, superlinear, sublinear and undecidable cases, and they argued that the exponent and the fluctuation model should be estimated together (Leitão et al. 2016). Bettencourt and colleagues (2010), for their part, showed that each city’s deviation from the scaling law (a size-adjusted indicator) carries information of its own and can persist for decades (Bettencourt et al. 2010).

The growth equation and the finite-time singularity

Evidence

The growth equation of Bettencourt and colleagues. They assumed that a city’s resource supply Y=Y0NβY = Y_0 N^{\beta} is used to maintain the existing population (RR per person) and to support additional population (EE per person), giving dN/dt=(Y0/E)Nβ−(R/E)N\mathrm{d}N/\mathrm{d}t = (Y_0/E) N^{\beta} – (R/E) N. With β<1\beta < 1, the population follows an S-shaped curve to a ceiling, as in the biological growth of Chapter 2; with β>1\beta > 1, it tends to infinity in a finite time tct_c, a finite-time singularity. They noted that with finite resources this means growth stalls or even collapses before the singularity unless major innovations “reset” the parameters and start a new cycle, and that to sustain growth such innovations must arrive at an accelerating rate, the intervals between cycles shrinking as the population grows (Bettencourt et al. 2007).

Calculation

Write a=Y0/Ea = Y_0/E and b=R/Eb = R/E. For β>1\beta > 1, rewriting the equation as a linear equation in N1−βN^{1-\beta} gives tc=1(β−1)bln⁡11−(b/a)N01−β.t_c = \frac{1}{(\beta – 1) b} \ln \frac{1}{1 – (b/a) N_0^{1-\beta}}. The program takes the illustrative values a=0.03a = 0.03, b=0.01b = 0.01 (per year) and N0=1N_0 = 1, and compares the formula with numerical integration; it then divides growth into a series of “cycles”: in cycle jj resources cap the population at 2j2^j, and when the cap is reached an innovation doubles it. The program gives the number of years each doubling takes.

# Growth equation of Bettencourt et al. 2007: dN/dt = a N^beta - b N (N in units of the
# starting population; a, b per year; ILLUSTRATIVE values a = 0.03, b = 0.01, N0 = 1).
# beta < 1: growth stops at N = (a/b)^(1/(1-beta)). beta > 1: N reaches infinity at the
# finite time t_c = ln(1 / (1 - (b/a) N0^(1-beta))) / ((beta - 1) b). Part 2: a MODEL of
# "innovation cycles": in cycle j resources cap N at 2^j; an innovation doubles the cap.
import math
A, B, N0 = 0.03, 0.01, 1.0

def integrate(beta, t_end, n_stop=1e9, dt=0.001):
    n, t = N0, 0.0
    while t < t_end and n < n_stop:
        n += dt * (A * n ** beta - B * n); t += dt
    return t, n

def t_c(beta, n0=N0):
    return math.log(1 / (1 - (B / A) * n0 ** (1 - beta))) / ((beta - 1) * B)

def cycle_time(beta, lo, hi, steps=4000):
    """Time to grow from lo to hi: integral of dN / (a N^beta - b N), midpoint rule."""
    h = (hi - lo) / steps
    return sum(h / (A * (lo + (i + .5) * h) ** beta - B * (lo + (i + .5) * h))
               for i in range(steps))

if __name__ == "__main__":
    print(f"beta = 0.85: ceiling (a/b)^(1/0.15) = {(A / B) ** (1 / 0.15):.0f};"
          f" N after 2,000 years = {integrate(0.85, 2000)[1]:.0f}")
    t9, t15 = integrate(1.15, 1e4)[0], integrate(1.15, 1e4, 1e15)[0]
    print(f"beta = 1.15: formula t_c = {t_c(1.15):.1f} years; numerical N passes"
          f" 1e9 at {t9:.1f}, 1e15 at {t15:.1f}")
    print("cycle  N from -> to   duration (years)   ratio to previous   clock at end")
    clock, prev = 0.0, None
    for j in range(1, 13):
        d = cycle_time(1.15, 2 ** (j - 1), 2 ** j); clock += d
        r = f"{d / prev:.3f}" if prev else "  -  "
        if j <= 6 or j % 3 == 0:
            print(f"{j:4d}   {2 ** (j - 1):5d} -> {2 ** j:5d}      {d:7.1f}"
                  f"             {r}            {clock:7.1f}")
        prev = d
    q = 2 ** -0.15
    print(f"long-run ratio 2^-0.15 = {q:.3f}; all later cycles together"
          f" < {prev * q / (1 - q):.0f} years")
# beta = 0.85: ceiling (a/b)^(1/0.15) = 1516; N after 2,000 years = 1211
# beta = 1.15: formula t_c = 270.3 years; numerical N passes 1e9 at 260.3, 1e15 at 269.1
# cycle  N from -> to   duration (years)   ratio to previous   clock at end
#    1       1 ->     2         32.1               -                 32.1
#    2       2 ->     4         27.7             0.862               59.8
#    3       4 ->     8         24.0             0.867               83.8
#    4       8 ->    16         20.9             0.871              104.7
#    5      16 ->    32         18.3             0.875              123.0
#    6      32 ->    64         16.1             0.878              139.1
#    9     256 ->   512         11.1             0.885              176.9
#   12    2048 ->  4096          7.8             0.890              203.2
# long-run ratio 2^-0.15 = 0.901; all later cycles together < 71 years

With β=0.85\beta = 0.85 the population tends to a ceiling of 1,516 and after 2,000 years stands at 1,211, still creeping upwards. With β=1.15\beta = 1.15 the formula gives tc=270.3t_c = 270.3 years; in the numerical integration the population passes 10910^9 at 260.3 years and 101510^{15} at 269.1 years, consistent with the formula: the closer to the singularity, the faster the growth. The doubling time falls from 32.1 years for the first doubling to 16.1 for the sixth and 7.8 for the twelfth, and the ratio between successive doublings rises from 0.862 towards its limit 2−0.15=0.9012^{-0.15} = 0.901. This is a convergent geometric series: all doublings after the twelfth together take less than 71 years. In other words, “an innovation every time resources run out” does not remove the singularity; it merely cuts the same trajectory into ever shorter pieces. For growth to continue indefinitely, each innovation must come sooner than the last.

Left: population trajectories for \beta = 0.85, 1.00 and 1.15 (logarithmic vertical axis); the dotted line marks the singularity t_c for \beta = 1.15. Right: years needed for each doubling with \beta = 1.15, shrinking by about 10 % each time.

Note

The singularity is a consequence of the model, not an observation. No city’s population has gone to infinity in finite time. The finite-time singularity says that if the superlinear supply equation held for long, growth would approach an impossible end point ever faster, so the equation itself must fail somewhere. Bettencourt and colleagues read this as “innovation cycles must accelerate”; that reading is compatible with qualitative accounts of urban history, but no quantitative test has shown that innovation cycles are accelerating at any particular rate. Volume 28 discussed early-warning signals before critical transitions; this chapter describes only the mathematical structure of the superlinear growth equation and makes no forecast for any city.

Hypothesis · not established

The interaction-density hypothesis (not yet established). Bettencourt’s 2013 model attributes superlinearity to networks of face-to-face interaction among people (Bettencourt 2013). The author conjectures that if this mechanism holds, the superlinear exponents of wages, patents and similar indicators will fall in urban systems where the share of remote work rises sharply. The testable prediction is that, with the same metropolitan boundaries, the exponents after the spread of remote work are closer to 1 than before. If, in a pre-registered analysis, the 95 % confidence interval of the difference between the two periods’ exponents contains no negative values, the hypothesis is refuted.

Summary

Urban infrastructure grows sublinearly with population and socioeconomic output superlinearly; doubling the population multiplies the totals by 1.80 and 2.22, or about −10 % and +11 % per person. The exponents are sensitive to the definition of city boundaries and to the fluctuation model, and the synthetic model shows that commuting can “turn” a linear relation into a superlinear one of 1.09. The superlinear growth equation diverges in finite time and each doubling takes about 10 % less time than the last, so that “resetting by innovation” sustains growth only if innovation keeps accelerating; the singularity is a mathematical consequence, not an observation. The next chapter turns from cities back to firms: how long do companies live?

References

Arcaute, Elsa, Erez Hatna, Peter Ferguson, Hyejin Youn, Anders Johansson, and Michael Batty. 2015. “Constructing Cities, Deconstructing Scaling Laws.” Journal of the Royal Society Interface 12 (102): 20140745. https://doi.org/10.1098/rsif.2014.0745.
Bettencourt, Luı́s M. A. 2013. “The Origins of Scaling in Cities.” Science 340 (6139): 1438–41. https://doi.org/10.1126/science.1235823.
Bettencourt, Luı́s M. A., José Lobo, Dirk Helbing, Christian Kühnert, and Geoffrey B. West. 2007. “Growth, Innovation, Scaling, and the Pace of Life in Cities.” Proceedings of the National Academy of Sciences of the USA 104 (17): 7301–6. https://doi.org/10.1073/pnas.0610172104.
Bettencourt, Luı́s M. A., José Lobo, Deborah Strumsky, and Geoffrey B. West. 2010. “Urban Scaling and Its Deviations: Revealing the Structure of Wealth, Innovation and Crime Across Cities.” PLoS ONE 5 (11): e13541. https://doi.org/10.1371/journal.pone.0013541.
Leitão, Jorge C., José M. Miotto, Martin Gerlach, and Eduardo G. Altmann. 2016. “Is This Scaling Nonlinear?” Royal Society Open Science 3 (7): 150649. https://doi.org/10.1098/rsos.150649.

Fan, Y. (2026). Scaling Laws, Allometric Growth and the Statistical Physics of Organisational Metabolism: From Kleiber's Three-Quarter Power Law and Fractal Supply Networks to the Mortality of Firms, Read Beside Classical Chinese Organisational Thought [规模法则、异速生长与组织新陈代谢的统计物理] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23788553 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).