5D Library › Scaling Laws, Allometric Growth and the Statistical Physics of Organisational Metabolism · 中文
Chapter 5 The Ringelmann Effect and Brooks’s Law: More People, Less Each?
Pulling on a rope
In the 1880s the French agricultural engineer Maximilien Ringelmann studied how efficiently people worked when pushing or pulling a load horizontally in different ways, and he published the results in 1913 (Ringelmann 1913). One of his tables was later cited again and again: taking the output of one person as 1, each member of a group of 2 to 8 produced on average only 0.93, 0.85, 0.77, 0.70, 0.63, 0.56 and 0.49; the total output of a group of eight was only 3.92 times that of one person. The table entered social psychology through a summary by the German psychologist Walther Moede in 1927 and is usually described as the result of rope pulling; the original was not presented in English until Kravitz and Martin did so in 1986, and they pointed out that Ringelmann said only that the table summarised “a great number of findings”, without naming the task, so that its attribution to rope pulling comes from Moede and later writers (Kravitz and Martin 1986).
Evidence
Ringelmann’s original article and his interpretation. Kravitz and Martin pointed out that Ringelmann’s main interest was in methods of work, with group comparisons secondary, and that the declining table summarised a great many observations. He attributed the group loss mainly to coordination loss, a lack of simultaneity in effort; he was aware of motivational factors (he mentioned prisoners who relied on their neighbours to do the work) but thought them unlikely in his tests. In one preliminary series, a person alone pulled 85.3 kg on average, each member of a 7-person group 65.0 kg and each member of a 14-person group 61.4 kg, a decline that levelled off as numbers grew (Kravitz and Martin 1986).
Evidence
Coordination loss or motivation loss? Ingham, Levinger, Graves and Peckham (1974) repeated the rope-pulling experiment and found that the force per person in real groups fell from one person to three but did not fall noticeably beyond that. Their second experiment used “pseudo-groups”: blindfolded participants believed that the people behind them were pulling too, but the others were only pretending. The participants’ effort still fell, although there was no coordination problem at all, so a substantial part of the loss is motivational (Ingham et al. 1974). Latané, Williams and Harkins (1979) replicated this conclusion with clapping and shouting and named it social loafing (Latané et al. 1979). Karau and Williams’s 1993 meta-analysis of 78 studies showed that social loafing is robust, and that whether individual contributions can be evaluated, expectations about co-workers’ performance, the meaningfulness of the task and cultural background all markedly change its size (Karau and Williams 1993).
“More people, less effort” therefore has at least two sources: coordination loss (unsynchronised effort, time spent communicating) and motivation loss (individual contributions cannot be seen). The Chapter 8 trial compares team structures and is aimed mainly at the first; but a squad structure also makes each person’s contribution more visible and may reduce the second as well.
Channels and the optimal number
Calculation
Among people there are pairwise communication channels. If each person alone produces and each channel costs the team of output, then Equivalently, output per person falls linearly with group size, exactly the shape of Ringelmann’s table. The program fits that table by least squares to obtain and ; compares it with an illustrative model in which the loss stops growing after three people (output per person settling at 0.80); and finally counts the channels among 12 people and the number of “relationships” between a manager and subordinates calculated by Graicunas in 1937 (Graicunas 1937).
# Ringelmann's (1913) summary table, translated by Kravitz & Martin 1986: relative output
# per worker for groups of 1 to 8 (often cited as rope pulling; Ringelmann did not name
# the task). Fit per-capita output y = 1 - c (N - 1); total output
# Y = N y = alpha N - (beta/2) N (N - 1) with alpha = 1, beta = 2c, i.e. a fixed loss per
# communication channel; peak at N* = alpha/beta + 1/2. Then: what if the per-capita loss
# levels off after three people (the curvilinear shape reported by Ingham et al. 1974;
# the plateau value 0.80 below is ILLUSTRATIVE)? Finally: channel counts for 12 people.
import math
PER_WORKER = [1.00, 0.93, 0.85, 0.77, 0.70, 0.63, 0.56, 0.49]
if __name__ == "__main__":
ns = range(1, 9)
c = (sum((n - 1) * (1 - y) for n, y in zip(ns, PER_WORKER))
/ sum((n - 1) ** 2 for n in ns)) # least squares through (1, 1)
beta = 2 * c
print(f"fitted loss per extra worker c = {c:.4f}; beta = 2c = {beta:.4f};"
f" N* = 1/beta + 1/2 = {1 / beta + 0.5:.2f}")
print("N per worker (table) total (table) total (fit) channels N(N-1)/2"
" plateau model")
for n, y in zip(ns, PER_WORKER):
plateau = n * (1 - 0.20 * (1 - math.exp(-(n - 1) / 1.0)))
print(f"{n:2d} {y:.2f} {n * y:4.2f} "
f"{n - beta / 2 * n * (n - 1):4.2f} {n * (n - 1) // 2:3d}"
f" {plateau:5.2f}")
print("12 people: all-to-all channels", 12 * 11 // 2,
"| three teams of four plus three lead-to-lead links", 3 * 6 + 3)
k = 4
print(f"Graicunas (1937) relationships for a manager with {k} direct reports:"
f" {k * (2 ** (k - 1) + k - 1)}; with 12: {12 * (2 ** 11 + 11)}")
# fitted loss per extra worker c = 0.0737; beta = 2c = 0.1474; N* = 1/beta + 1/2 = 7.28
# N per worker (table) total (table) total (fit) channels N(N-1)/2 plateau model
# 1 1.00 1.00 1.00 0 1.00
# 2 0.93 1.86 1.85 1 1.75
# 3 0.85 2.55 2.56 3 2.48
# 4 0.77 3.08 3.12 6 3.24
# 5 0.70 3.50 3.53 10 4.02
# 6 0.63 3.78 3.79 15 4.81
# 7 0.56 3.92 3.90 21 5.60
# 8 0.49 3.92 3.87 28 6.40
# 12 people: all-to-all channels 66 | three teams of four plus three lead-to-lead links 21
# Graicunas (1937) relationships for a manager with 4 direct reports: 44; with 12: 24708Ringelmann’s table fits the “fixed loss per channel” model almost perfectly: the fitted loss per extra person is 0.0737, so and ; in the table total output is 3.92 for both 7 and 8 people, with the peak exactly between them. But if the loss stops growing after three people, as Ingham and colleagues observed, total output keeps rising with group size (6.40 with eight people) and there is no optimal number. Which model is closer to reality depends on the task: in pulling a rope, each additional person adds no new “interface” to coordinate, whereas in software or assembly tasks each person’s work may depend on everyone else’s.
Twelve people communicating freely have 66 channels; split into three squads of four, with only the three squad leads in contact with one another, they have . Graicunas, counting direct relationships between a superior and subordinates, relationships among subordinates and cross relationships, obtained 44 relationships for a manager with 4 direct reports and 24,708 for one with 12. The figure is often used to argue that spans of control should not be too wide, but it assumes that every possible relationship needs equal managing, an assumption with no empirical basis.
Brooks’s law
Evidence
A rule of thumb. Drawing on his experience of managing the development of the operating system for IBM’s System/360, Brooks wrote in The Mythical Man-Month (1975): “Adding manpower to a late software project makes it later.” His reasons were that newcomers need training, which takes time from existing members; that the work of communication grows as as numbers grow; and that many tasks cannot be divided at will (Brooks 1975). In the 1995 anniversary edition Brooks called the law an “outrageous oversimplification” while maintaining that it captures the general rule (Brooks 1995). McConnell (1999) argued that projects are often late because the original estimates were too optimistic, that adding people early is reasonable, and that the law holds roughly only near the end of a project (McConnell 1999).
Evidence
The team scaling fallacy. Staats, Milkman and Fox (2012) had 267 executive MBA students build a LEGO figure in teams of two or four. Two-person teams took 36.33 person-minutes on average and four-person teams 52.64: the four-person teams used about 45 % more labour in total. The paper reports only person-minutes; dividing by team size gives about 13 minutes of clock time for the four-person teams and 18 for the two-person teams (a further 20 teams did not finish within the 30-minute limit). Another group of participants, asked to predict the labour needed by each kind of team, underestimated both, and underestimated the four-person teams more. The authors called this the “team scaling fallacy” (Staats et al. 2012).
The experiment reminds us to distinguish two outcomes: completion time and total labour. Adding people can shorten completion time while lowering each person’s efficiency; Brooks’s law says that under certain conditions even completion time gets longer.
Calculation
A day-by-day illustrative model: six people start a project of 1,500 person-days; each person loses 3 % of working time for every additional colleague (the “fixed loss per channel” of the previous section); newcomers ramp up linearly from zero to full speed over 40 days, and each newcomer still ramping up takes half a day of an existing member’s time for training. The program compares the effect on the finishing day of adding people on different days. All parameters are illustrative.
# A day-by-day MODEL of Brooks's law (all parameters ILLUSTRATIVE). Six people start a
# project of 1,500 person-days. Each person loses c = 3 % of the day for every other team
# member (a fixed cost per communication channel, as in Chapter 5). Newcomers ramp up
# linearly from 0 to full speed over R days, and each one still ramping takes half of a
# veteran's day for training. Question: when does adding k people shorten the project?
WORK, N0, C, MENTOR = 1500.0, 6, 0.03, 0.5
def finish(k, day_add, ramp=40, work=WORK):
done, day = 0.0, 0
while done < work:
day += 1
new = [min(1.0, (day - day_add) / ramp)] * k if day > day_add else []
n = N0 + len(new)
learning = sum(1 for x in new if x < 1.0)
done += (1 - C * (n - 1)) * (N0 + sum(new) - MENTOR * learning)
return day
if __name__ == "__main__":
base = finish(0, 0)
print(f"no one added: {base} days"
f" (output {N0 * (1 - C * (N0 - 1)):.1f} person-days per day)")
ks = (2, 4, 6, 12)
print("change in finishing day, people added k: 2 4 6 12")
for d in (1, 100, 200, 260):
print(f" added on day {d:3d}: "
+ " ".join(f"{finish(k, d) - base:+5d}" for k in ks))
print(" day 260, ramp-up 10 days, not 40: "
+ " ".join(f"{finish(k, 260, 10) - base:+5d}" for k in ks))
print(f"steady-state output n(1 - c(n - 1)) peaks at n = {1 / (2 * C) + 0.5:.1f}:"
+ "".join(f" n={n}: {n * (1 - C * (n - 1)):.1f}" for n in (6, 12, 17)))
# no one added: 295 days (output 5.1 person-days per day)
# change in finishing day, people added k: 2 4 6 12
# added on day 1: -47 -73 -88 -98
# added on day 100: -28 -43 -52 -56
# added on day 200: -9 -13 -15 -14
# added on day 260: +3 +5 +7 +11
# day 260, ramp-up 10 days, not 40: -5 -7 -8 -9
# steady-state output n(1 - c(n - 1)) peaks at n = 17.2: n=6: 5.1 n=12: 8.0 n=17: 8.8Without additions the project takes 295 days. Adding 6 people on the first day brings completion forward by 88 days; adding them on day 100, by 52 days; on day 200, by only 15 days, and adding 12 (14 days earlier) is no better than adding 6. Adding people on day 260, with about 35 days to go, delays completion by 3, 5, 7 and 11 days for 2, 4, 6 and 12 people: this is Brooks’s law. If newcomers need only 10 days to ramp up, adding people on day 260 instead brings completion forward by 5 to 9 days. In steady state, team output is highest at 17.2 people. The model shows that Brooks’s law requires three conditions together: little work left, slow ramp-up and high communication costs; it is a conditional rule of thumb, not a law.
Hypothesis · not established
The interface-document hypothesis (not yet established). The author conjectures that written interface specifications lower the coordination cost of each cross-squad channel, so that the total coordination cost of a squad structure is lower than that of free communication among everyone. The Chapter 8 trial compares the combination of squads, liaisons and interface documents with flat, free communication, so it tests the hypothesis only indirectly: if the trial reaches the verdict “no difference worth acting on” or “the squads are slower”, the combined form of the hypothesis loses its support; a positive result cannot be credited to the interface documents alone, and testing them on their own would need a third arm with squads but no documents (see “What this trial cannot show” in Chapter 8).
Summary
Ringelmann’s table shows output per person falling linearly with group size, equivalent to a fixed loss per communication channel, with an optimal size of about 7; but Ingham and colleagues found that the loss levels off after three people and that much of it is motivational rather than a matter of coordination. Brooks’s law is a conditional rule of thumb: near the end of a project, with slow ramp-up and high communication costs, adding people delays completion; adding them early usually helps. The next chapter widens the view from a single team to a whole hierarchical organisation.
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).