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

Chapter 2 Ontogenetic Growth: Why Growth Stops

An energy budget

An animal grows from a fertilised egg to an adult and then stops growing. Why does it stop? One explanation is that the genes set an end point; another comes from the energy accounts: growth needs surplus energy, and as the body gets larger the energy needed to maintain existing tissue grows faster than the supply, until it absorbs all of it. In 2001 West, Brown and Enquist wrote the metabolic scaling of Chapter 1 as such a budget (West et al. 2001).

Evidence

A general model of ontogenetic growth. West and colleagues assumed that the whole organism’s metabolic rate B=B0m3/4B = B_0 m^{3/4} is spent partly on maintaining existing cells, in proportion to the number of cells and so to body mass mm, and the rest on making new cells. This gives dm/dt=am3/4−bm\mathrm{d}m/\mathrm{d}t = a m^{3/4} – b m, where aa and bb are fixed by the metabolic rate of a cell, the mass of a cell and the energy needed to make one. Growth stops when dm/dt=0\mathrm{d}m/\mathrm{d}t = 0, at the asymptotic mass M=(a/b)4M = (a/b)^4. Rewriting the variables as r=(m/M)1/4r = (m/M)^{1/4} and the dimensionless time τ=at/(4M1/4)−ln⁡[1−(m0/M)1/4]\tau = a t / (4M^{1/4}) – \ln[1 – (m_0/M)^{1/4}] gives r=1−e−τr = 1 – e^{-\tau}, and they showed that growth data for mammals, birds, fish and crustaceans fall close to this curve (West et al. 2001).

The form of the equation is not new. The growth equation that von Bertalanffy proposed in the mid-twentieth century has exactly this structure, synthesis proportional to a power of mass and breakdown proportional to mass, usually with the exponent 2/3 (Bertalanffy 1957). WBE’s contribution was to link the exponent to the network theory of Chapter 1 and to provide the non-dimensionalisation that lays different species on top of one another.

Calculation

A fourth-order Runge–Kutta method solves the growth equation for four “species” (illustrative parameters, not taken from the paper) and checks whether all of them collapse onto r=1−e−τr = 1 – e^{-\tau}. The second part uses the 3/4 model of the “ungulate” (the fourth species) to generate observations with 3 % or 10 % multiplicative error, fits them with the 3/4 model and with a 2/3 model (both with free aa and MM), and compares the root-mean-square log errors.

# Ontogenetic growth (West, Brown & Enquist 2001): dm/dt = a m^(3/4) - b m, so growth
# stops at M = (a/b)^4. Universal form: r = (m/M)^(1/4), tau = a t / (4 M^(1/4))
# - ln(1 - (m0/M)^(1/4)) gives r = 1 - exp(-tau) for every species. The four "species"
# below have ILLUSTRATIVE parameters (not taken from the paper). Part 2 asks whether a
# 2/3-power model (von Bertalanffy) can fit the same curve.
import math, random

def grow(a, b, m0, t_end, dt=0.01, p=0.75):
    """RK4 solution of dm/dt = a m^p - b m; returns a list of (t, m) every 1 time unit."""
    f = lambda m: a * m ** p - b * m
    m, t, out, step = m0, 0.0, [(0.0, m0)], round(1 / dt)
    for i in range(1, round(t_end / dt) + 1):
        k1 = f(m); k2 = f(m + dt * k1 / 2); k3 = f(m + dt * k2 / 2); k4 = f(m + dt * k3)
        m += dt * (k1 + 2 * k2 + 2 * k3 + k4) / 6; t = i * dt
        if i % step == 0: out.append((t, m))
    return out

SPECIES = {"fish": (0.20, 0.10, 0.5, 150), "bird": (1.0, 0.25, 1.0, 60),
           "rodent": (0.6, 0.20, 0.8, 80), "ungulate": (2.0, 0.20, 20.0, 150)}

def universal(a, M, m0, t):
    return 1 - math.exp(-(a * t / (4 * M ** 0.25) - math.log(1 - (m0 / M) ** 0.25)))

if __name__ == "__main__":
    print("species    a     b    M = (a/b)^4   m at end   max |r - (1 - e^-tau)|")
    worst = 0.0
    for name, (a, b, m0, T) in SPECIES.items():
        M = (a / b) ** 4; traj = grow(a, b, m0, T)
        dev = max(abs((m / M) ** 0.25 - universal(a, M, m0, t)) for t, m in traj)
        worst = max(worst, dev)
        print(f"{name:9s} {a:4.1f}  {b:4.2f}  {M:10.1f}  {traj[-1][1]:9.1f}"
              f"      {dev:.1e}")
    print(f"all four collapse onto r = 1 - exp(-tau); worst deviation {worst:.1e}")
    # Part 2: noisy 'measurements' (3 % or 10 % error) from the 3/4 model of the ungulate,
    # fitted by least squares with the 3/4 model and with a 2/3 model (free a and M).
    rng = random.Random(2001)
    a, b, m0, T = SPECIES["ungulate"]
    true = grow(a, b, m0, T)[::5]
    for noise in (0.03, 0.10):
        data = [(t, m * math.exp(rng.gauss(0, noise))) for t, m in true]

        def sse(p, aa, MM):
            pred = grow(aa, aa / MM ** (1 - p), m0, T, dt=0.05, p=p)[::5]
            return sum((math.log(y) - math.log(q[1])) ** 2
                       for (t, y), q in zip(data, pred))

        res = []
        for p in (0.75, 2 / 3):
            best = min((sse(p, aa, MM), aa, MM)
                       for aa in [0.4 + 0.1 * i for i in range(40)]
                       for MM in [8000 + 400 * j for j in range(41)])
            res.append((math.sqrt(best[0] / len(data)), best[1], best[2]))
        print(f"noise {noise:.2f}: 3/4 model rms {res[0][0]:.4f} (M {res[0][2]:.0f});"
              f" 2/3 model rms {res[1][0]:.4f} (M {res[1][2]:.0f})")
# species    a     b    M = (a/b)^4   m at end   max |r - (1 - e^-tau)|
# fish       0.2  0.10        16.0       15.1      8.9e-16
# bird       1.0  0.25       256.0      238.4      4.0e-13
# rodent     0.6  0.20        81.0       77.0      4.4e-14
# ungulate   2.0  0.20     10000.0     9982.6      3.8e-13
# all four collapse onto r = 1 - exp(-tau); worst deviation 4.0e-13
# noise 0.03: 3/4 model rms 0.0268 (M 10000); 2/3 model rms 0.0418 (M 10400)
# noise 0.10: 3/4 model rms 0.1201 (M 10000); 2/3 model rms 0.1245 (M 10400)

The asymptotic masses of the four species range from 16 to 10,000, a factor of more than six hundred, yet after non-dimensionalisation the largest deviation from 1−e−τ1 – e^{-\tau} is of order 10−1310^{-13}, merely numerical error: for this equation the collapse is a mathematical identity, not an empirical discovery. What needs empirical testing is whether data obey the equation. The second part shows that with 3 % measurement error the 3/4 model’s error (0.0268) is clearly smaller than the 2/3 model’s (0.0418), so the two can be told apart; with 10 % error the values are 0.1201 and 0.1245, almost indistinguishable, because the 2/3 model needs only a small adjustment of aa and MM to fit the same curve. This is the basis of the point made by Banavar and colleagues in 2002: the universal growth curve follows from more general principles, and existing growth data cannot distinguish 3/4 from 2/3 (Banavar et al. 2002).

Left: growth curves of four illustrative species (mass divided by each species’ asymptotic mass). Right: after non-dimensionalisation all four curves lie on 1 - e^{-\tau} (grey line); for this equation the collapse is a mathematical identity.

Maintenance swallows supply

Calculation

In dm/dt=amp−bm\mathrm{d}m/\mathrm{d}t = a m^{p} – b m the supply is ampa m^{p} and maintenance is bmb m; their ratio bm/(amp)=(m/M)1−pb m / (a m^p) = (m/M)^{1-p}, and the ceiling is M=(a/b)1/(1−p)M = (a/b)^{1/(1-p)}. With p=3/4p = 3/4, the share of supply taken by maintenance is (m/M)1/4(m/M)^{1/4}. The second part of the program is an organisational analogy: treat “supply” as the gross margin a team can reinvest and “maintenance” as a cost proportional to headcount, and compare, for different exponents pp, the height of the ceiling, its sensitivity to a 10 % gain in efficiency, and the share of supply already taken by maintenance at half the ceiling. The exponents pp are assumptions, not estimates for any real firm.

# Where the energy goes in dm/dt = a m^p - b m. Supply a m^p, maintenance b m, ceiling
# M = (a/b)^(1/(1-p)). Share of supply spent on maintenance = (m/M)^(1-p). Part 1: the
# biological model (p = 3/4). Part 2: an ILLUSTRATIVE organisational analogy in which
# "supply" is the gross margin a team can reinvest and "maintenance" is a cost per person;
# the exponents p are assumptions, not estimates for any real firm.
import math

def ceiling(a, b, p):
    return (a / b) ** (1 / (1 - p))

def time_to(frac, a, b, p, m0_frac=0.01, dt=0.001):
    """Time for m to grow from m0_frac*M to frac*M (Euler steps, m in units of M)."""
    M = ceiling(a, b, p); m, t = m0_frac * M, 0.0
    while m < frac * M:
        m += dt * (a * m ** p - b * m); t += dt
    return t

if __name__ == "__main__":
    print("p = 3/4: share of supply spent on maintenance at m/M =")
    fr = (1 / 16, 0.25, 0.5, 0.9)
    print("   " + "   ".join(f"{f:.3f}: {f ** 0.25:.2f}" for f in fr))
    a, b = 1.0, 0.5
    t50, t90, t99 = (time_to(f, a, b, 0.75) for f in (0.5, 0.9, 0.99))
    print(f"time from 1 % to 50 %, 90 %, 99 % of M (a = 1, b = 0.5):"
          f" {t50:.1f}, {t90:.1f}, {t99:.1f}")
    print("analogy: ceiling N = (a/b)^(1/(1-p)) and its sensitivity to efficiency")
    print("   p    N for a/b = 1.5   N if a up 10 %   ratio   maintenance share at N/2")
    for p in (0.75, 0.85, 0.90, 0.95):
        n0, n1 = ceiling(1.5, 1, p), ceiling(1.65, 1, p)
        print(f"  {p:.2f}  {n0:12.1f}    {n1:12.1f}      {n1 / n0:5.2f}"
              f"          {0.5 ** (1 - p):.2f}")
    print("  1.00  no ceiling: exponential growth if a > b, decline if a < b")
# p = 3/4: share of supply spent on maintenance at m/M =
#    0.062: 0.50   0.250: 0.71   0.500: 0.84   0.900: 0.97
# time from 1 % to 50 %, 90 %, 99 % of M (a = 1, b = 0.5): 11.7, 26.2, 44.9
# analogy: ceiling N = (a/b)^(1/(1-p)) and its sensitivity to efficiency
#    p    N for a/b = 1.5   N if a up 10 %   ratio   maintenance share at N/2
#   0.75           5.1             7.4       1.46          0.84
#   0.85          14.9            28.2       1.89          0.90
#   0.90          57.7           149.6       2.59          0.93
#   0.95        3325.3         22370.7       6.73          0.97
#   1.00  no ceiling: exponential growth if a > b, decline if a < b

The biological part shows why growth slows in its later stages: at only 1/16 of the ceiling, half of metabolism already goes to maintenance; at half, 84 %; at 90 %, 97 %. Growing from 1 % to 50 % of the ceiling takes 11.7 time units, from 50 % to 90 % another 14.5, and from 90 % to 99 % a further 18.7. Growth does not “stop suddenly”; it approaches a ceiling set by a/ba/b ever more slowly.

The analogy shows two things. First, the closer the supply exponent is to 1, the higher the ceiling and the more sensitive it is to efficiency: raising aa by 10 % lifts the ceiling by only 46 % when p=0.75p = 0.75, but multiplies it by 2.59 when p=0.90p = 0.90 and by 6.73 when p=0.95p = 0.95; with p=1p = 1 there is no ceiling, and the system grows exponentially if a>ba > b and shrinks if a<ba < b. Second, the closer the exponent is to 1, the larger the share of supply taken by maintenance half-way: with p=0.90p = 0.90, a system at half its ceiling already spends 93 % of its supply on maintenance, leaving 7 % for growth. A nearly linear organisation can therefore feel that “growth is getting harder and harder” even when it is still far from its ceiling.

Note

Limits of the analogy. An organism’s aa and bb are fixed by cell-level quantities and change little over its life; a firm’s “supply exponent” and “maintenance cost” can be changed by organisational design, technology and markets, and a firm can shrink, split or merge, which an animal cannot. Chapter 4 will show that most exits of listed companies are acquisitions, not “deaths of old age”. This section’s program shows only the mathematical consequence of the structure “sublinear supply plus linear maintenance equals a ceiling”; it says nothing about the value of pp for any firm.

Hypothesis · not established

The maintenance-swallowing hypothesis (not yet established). The author conjectures that in software and digital-service teams, the share of working hours spent maintaining features already in service is a leading indicator of slowing revenue growth: teams in which this share rises steadily show lower revenue growth over the following four quarters than comparable teams in which it is stable. If, in pre-registered longitudinal data on many teams, the change in this share is unrelated to subsequent growth (the 95 % confidence interval of the correlation contains 0 and its upper limit is below 0.1), the hypothesis is refuted. Step 4 of the Chapter 8 checklist gives a fictional example.

Summary

The WBE model of ontogenetic growth writes growth as sublinear supply minus linear maintenance, with an asymptotic mass of (a/b)4(a/b)^4; the universal curve after non-dimensionalisation is an identity for this equation, and existing data can hardly distinguish 3/4 from 2/3. The share of supply taken by maintenance rises with mass and reaches 84 % at half the ceiling. Applied to organisations, the structure implies that the closer supply is to linear, the higher and more efficiency-sensitive the ceiling, but the more of the supply maintenance absorbs along the way. The next chapter turns to a system that does not stop growing: the city.

References

Banavar, Jayanth R., John Damuth, Amos Maritan, and Andrea Rinaldo. 2002. “Ontogenetic Growth: Modelling Universality and Scaling.” Nature 420 (6916): 626. https://doi.org/10.1038/420626a.
Bertalanffy, Ludwig von. 1957. “Quantitative Laws in Metabolism and Growth.” The Quarterly Review of Biology 32 (3): 217–31. https://doi.org/10.1086/401873.
West, Geoffrey B., James H. Brown, and Brian J. Enquist. 2001. “A General Model for Ontogenetic Growth.” Nature 413 (6856): 628–31. https://doi.org/10.1038/35098076.

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).