5D Library › Scaling Laws, Allometric Growth and the Statistical Physics of Organisational Metabolism · 中文
Chapter 6 Span of Control, Managerial Hierarchy and Dunbar’s Number
The accounts of hierarchy
When a team grows, the usual response is to add levels: each manager leads a number of people, and above the managers there are further managers. The number of people a manager leads directly is the span of control. With span and levels there are people at the bottom and in the whole organisation. Hierarchy reduces the communication channels from the of everyone connected to everyone to about ; the price is that information passes through several levels, and the managers at each level must be paid.
Evidence
Simon on executive pay. In 1957 Simon asked why executives in large companies are paid more. His model was that an organisation is a pyramid with a roughly fixed span, in which pay at each level is a fixed multiple of pay at the level below; a larger company then has more levels, and top pay rises as a power of company size, with an exponent set jointly by the span and the pay multiple between levels, with no need to assume that the top executives of large companies have a higher “marginal product” (Simon 1957).
Evidence
Williamson’s control loss. Williamson (1967) added control loss to the same pyramid: each level carries out only a fraction () of its superior’s intentions. He showed that the optimal number of levels is then finite, so that the optimal size of the firm is bounded; as control loss disappears (), the bound disappears too (Williamson 1967).
Calculation
A simplified version of Williamson’s set-up: span ; level has people paid (Simon’s pay multiple between levels); each person at the bottom produces ; net revenue is minus the total wage bill and a non-wage cost for each production worker. Spreading the pay of the levels above over the production workers, the total cost per production worker is about . Net revenue is then about . Treating as continuous and setting the derivative to zero gives the first-order condition , that is without the second term this is the level at which revenue and cost break even at the bottom, , an upper bound on size rather than the optimum. The program takes the illustrative values , , , and and compares net revenue level by level; finally it computes the elasticity of top pay with respect to size in Simon’s model, .
# Hierarchy, control loss and firm size, in the set-up of Williamson (1967): span of
# control s, n levels (level 1 = top), s^(i-1) people at level i, wage w0 * b^(n-i) at
# level i (Simon 1957: pay rises by a factor b per level). Each level passes on only a
# fraction alpha of its superior's intentions, so output is theta (alpha s)^(n-1).
# Net revenue = P theta (alpha s)^(n-1) - wages - r s^(n-1). All numbers ILLUSTRATIVE:
# s = 6, b = 1.5, w0 = 1, r = 0.5, P theta = 3. Net revenue is about s^(n-1) times
# (P theta alpha^(n-1) - c); setting its derivative in n to 0 gives the first-order
# condition alpha^(n-1) = (c / P theta) ln s / ln(alpha s).
import math
S, B, W0, R, PT = 6, 1.5, 1.0, 0.5, 3.0
COST = W0 + R + W0 * (B / S) / (1 - B / S) # cost per production worker, incl. bosses
def net(n, alpha):
wages = sum(W0 * B ** (n - i) * S ** (i - 1) for i in range(1, n + 1))
return PT * (alpha * S) ** (n - 1) - wages - R * S ** (n - 1)
def staff(n):
return sum(S ** (i - 1) for i in range(1, n + 1))
if __name__ == "__main__":
print(f"cost per production worker including the hierarchy above: {COST:.3f}")
print("alpha best n first-order n* break-even n staff at optimum fidelity")
for alpha in (0.80, 0.85, 0.90, 0.93, 0.95):
n = max(range(1, 40), key=lambda k: net(k, alpha))
foc = 1 + (math.log(COST / PT) + math.log(math.log(S) / math.log(alpha * S))
) / math.log(alpha)
even = 1 + math.log(COST / PT) / math.log(alpha)
print(f" {alpha:.2f} {n:2d} {foc:5.1f} {even:5.1f}"
f" {staff(n):12,d} {alpha ** (n - 1):.2f}")
print(f" 1.00 no limit: every production worker adds {PT - COST:.2f} net")
e = math.log(B) / math.log(S)
print(f"Simon: top pay ~ (firm size)^(ln b / ln s) = size^{e:.3f};")
print(f" a firm 10 times larger pays its top level {10 ** e:.2f} times more")
# cost per production worker including the hierarchy above: 1.833
# alpha best n first-order n* break-even n staff at optimum fidelity
# 0.80 2 2.6 3.2 7 0.80
# 0.85 3 3.4 4.0 43 0.72
# 0.90 5 5.1 5.7 1,555 0.66
# 0.93 7 7.2 7.8 55,987 0.65
# 0.95 10 10.0 10.6 12,093,235 0.63
# 1.00 no limit: every production worker adds 1.17 net
# Simon: top pay ~ (firm size)^(ln b / ln s) = size^0.226;
# a firm 10 times larger pays its top level 1.68 times moreThe cost per production worker, including the levels above, is 1.833. With the optimum is 5 levels and 1,555 people; with , 10 levels and about 12.09 million people; with , only 2 levels and 7 people. The first-order condition gives of 5.1 and 10.0, in line with the level-by-level comparison; the break-even level (5.7 and 10.6) is about half a level to one level higher, because the last level adds the most people and must pay off as a whole, not merely break even at the margin. At the optimum, the fidelity of information from top to bottom, , lies between 0.63 and 0.80. Without control loss, each production worker nets 1.17 and there is no limit to size.
What matters here is not the particular numbers but the sensitivity: raising from 0.90 to 0.95 takes the optimal size from about fifteen hundred people to over twelve million. Anything that reduces distortion from level to level (written standards, information systems, modules that can run independently) greatly raises the feasible size of an organisation. In Simon’s model, with a span of 6 and a pay multiple of 1.5, the elasticity of top pay with respect to size is 0.226: a company ten times larger pays its top level about 1.68 times more.
Evidence
Modularity and “mirroring”. Simon (1962) proposed that complex systems tend to be nearly decomposable hierarchies, with dense connections within subsystems and sparse connections between them, and that such structures evolve and are maintained more easily (Simon 1962). Parnas (1972) argued for dividing software into modules by “information hiding”, so that each module exposes only its interface (Parnas 1972). Conway (1968) observed that organisations that design systems produce designs that copy their own communication structures, an observation later called Conway’s law (Conway 1968). MacCormack, Baldwin and Rusnak (2012) tested this “mirroring” hypothesis with matched pairs: in their sample of paired software products, those developed by loosely coupled open-source communities were in every case more modular than comparable products developed by tightly coupled firms, with the potential for changes to propagate differing by up to about eightfold (MacCormack et al. 2012).
Dunbar’s number and its disputes
Evidence
The social brain hypothesis and 150. Dunbar (1992) found that the neocortex ratio of primates (the volume of the neocortex relative to the rest of the brain) predicts group size (Dunbar 1992). In 1993 he substituted the human neocortex ratio of 4.1 into the regression equation and obtained a predicted group size of 147.8, with a 95 % confidence interval of 100.2 to 231.1; he noted that three sizes recur in hunter-gatherer societies: residential bands of 30 to 50, clans or villages of 100 to 200, and tribes of 500 to 2,500 (Dunbar 1993). Hill and Dunbar (2003) estimated social network size from Christmas-card lists and found an average of 153.5 people (124.9 counting only those contacted directly) (Hill and Dunbar 2003). Zhou, Sornette, Hill and Dunbar (2005) found that social group sizes form a layered structure in which successive layers differ by a factor of about 3, the familiar 5, 15, 50, 150 (Zhou et al. 2005).
Evidence
Dunbar’s number deconstructed. Lindenfors, Wartel and Lind (2021) re-estimated the figure with updated primate data and phylogenetic methods; the Bayesian method gave estimates of 69 to 109 for average human group size and generalised least squares 16 to 42, but with 95 % intervals as wide as 4 to 520 and 2 to 336 respectively; they concluded that specifying any one number in this way is futile (Lindenfors et al. 2021). In a reply in 2021, Dunbar argued that when the horizontal axis (the brain measure) itself contains error, ordinary least squares systematically underestimates the slope and reduced major axis regression should be used; on the same data the former gives 71 and the latter 158. He also noted that the relation between group size and brain size falls into several “grades”, each needing its own equation, and listed evidence from telephone networks, military units, historical villages and elsewhere (Dunbar 2021).
Calculation
The first part substitutes into Dunbar’s equation and looks at how small changes in the neocortex ratio or the slope change the prediction. The second part builds a synthetic “primate” data set: 36 species with neocortex ratios between 1.2 and 3.2, generated from Dunbar’s line with scatter and with measurement error in the neocortex ratio; it fits the data by OLS and by RMA and extrapolates both to 4.1, outside the range of the data. The third part computes the ratios between layer sizes.
# Dunbar's number as an extrapolation. Part 1: Dunbar's (1993) equation
# log10 N = 0.093 + 3.389 log10 CR with the human neocortex ratio CR = 4.1. Part 2: a
# SYNTHETIC "primate" data set (36 species, CR 1.2 to 3.2, generated from Dunbar's line
# with scatter and with measurement error in CR) fitted by OLS and by RMA, then
# extrapolated to CR = 4.1, outside the range of the data. Part 3: the layer sizes.
import math, random
from statistics import mean, stdev, NormalDist
rng = random.Random(150)
def dunbar(cr, a=0.093, b=3.389):
return 10 ** (a + b * math.log10(cr))
if __name__ == "__main__":
print(f"CR = 4.1 -> N = {dunbar(4.1):.1f}; CR 3.9 / 4.3 -> {dunbar(3.9):.0f} /"
f" {dunbar(4.3):.0f}; slope 3.389 -/+ 10 % -> {dunbar(4.1, b=3.05):.0f} /"
f" {dunbar(4.1, b=3.728):.0f}")
xs_true = [math.log10(rng.uniform(1.2, 3.2)) for _ in range(36)]
xs = [x + rng.gauss(0, 0.04) for x in xs_true] # CR measured with error
ys = [0.093 + 3.389 * x + rng.gauss(0, 0.25) for x in xs_true]
mx, my = mean(xs), mean(ys)
sxy = sum((x - mx) * (y - my) for x, y in zip(xs, ys))
sxx = sum((x - mx) ** 2 for x in xs); syy = sum((y - my) ** 2 for y in ys)
b_ols = sxy / sxx; r = sxy / math.sqrt(sxx * syy); b_rma = b_ols / r
x0 = math.log10(4.1)
print(f"synthetic data: r = {r:.2f}; slope OLS {b_ols:.2f}, RMA {b_rma:.2f}"
f" (true 3.389)")
for lab, b in (("OLS", b_ols), ("RMA", b_rma)):
print(f" {lab} prediction at CR = 4.1: {10 ** (my + b * (x0 - mx)):.0f}")
res = [y - my - b_ols * (x - mx) for x, y in zip(xs, ys)]
s = stdev(res) * math.sqrt(35 / 34)
se = s * math.sqrt(1 + 1 / 36 + (x0 - mx) ** 2 / sxx)
z = NormalDist().inv_cdf(0.975); c = my + b_ols * (x0 - mx)
print(f" 95 % prediction interval for ONE species at CR = 4.1 (OLS):"
f" {10 ** (c - z * se):.0f} to {10 ** (c + z * se):.0f}")
layers = [5, 15, 50, 150, 500, 1500]
print("layers " + ", ".join(map(str, layers)) + "; ratios "
+ ", ".join(f"{b / a:.1f}" for a, b in zip(layers, layers[1:])))
# CR = 4.1 -> N = 147.8; CR 3.9 / 4.3 -> 125 / 174; slope 3.389 -/+ 10 % -> 92 / 238
# synthetic data: r = 0.77; slope OLS 2.83, RMA 3.68 (true 3.389)
# OLS prediction at CR = 4.1: 100
# RMA prediction at CR = 4.1: 172
# 95 % prediction interval for ONE species at CR = 4.1 (OLS): 24 to 419
# layers 5, 15, 50, 150, 500, 1500; ratios 3.0, 3.3, 3.0, 3.3, 3.0The equation itself is sensitive to its inputs: a neocortex ratio of 3.9 or 4.3 gives 125 or 174; a slope changed by ±10 % gives 92 or 238. In the synthetic data, measurement error on the horizontal axis lowers the OLS slope from the true 3.389 to 2.83, while the RMA slope is 3.68; extrapolated to 4.1 they predict 100 and 172, a difference in the same direction as Dunbar’s 71 against 158. More important, even with OLS the 95 % prediction interval for a single species at 4.1 runs from 24 to 419. Part of the disagreement is therefore about method (OLS underestimates a slope measured with error, while RMA relies on an assumption about the ratio of errors on the two axes); the more fundamental problem is that the human neocortex ratio lies outside the range of the primate data, and any extrapolation carries very wide uncertainty.
Note
When applied to organisations. “150 people” is often treated as a design limit for company departments or communities. Even if Dunbar’s number holds, it describes how many stable social relationships one person can maintain, not how many people a team can coordinate effectively, nor the optimal size of a department. This chapter’s conclusion is that control loss and communication costs do set limits on organisational size, but those limits depend on , the span and how far the task can be decomposed; they must be estimated from an organisation’s own data and cannot be extrapolated from primate brain size.
Summary
Span and hierarchy reduce communication channels from order to order , at the price of distortion from level to level and managerial pay. Simon’s model explains why executive pay rises as a power of size; Williamson’s control loss gives a finite optimal number of levels, and the optimal size is extremely sensitive to the fidelity of information. Nearly decomposable modular structures are the classic way to reduce coordination costs. Dunbar’s 147.8 is an extrapolation with a wide interval; the reanalysis by Lindenfors and colleagues gives an interval of 4 to 520, and Dunbar replies that they used the wrong regression method; the observation that layers differ by a factor of about 3 is a separate matter from the specific number 150. The next chapter reads this modern research beside classical Chinese texts on “governing the many”.
References
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).