5D Library › Evolutionary Game Theory, Direct and Indirect Reciprocity and the Emergence of Cooperation · 中文

Chapter 6 Punishment: Altruistic Punishment, Antisocial Punishment and Governing the Commons

Public goods and free riding

The donation games of the earlier chapters are matters between two people. Many cooperation problems involve many people: an irrigation channel, an open-source project, a community’s shared rules. In experiments such problems are studied with the public goods game: each person has an endowment and may put part of it into a common pool; the pool is multiplied by a factor greater than 1 but smaller than the number of players and divided equally among everyone. Each unit put in returns only part of itself to the contributor, so each person does better by contributing less; but if everyone contributes, everyone is better off. Chapter 1 of Volume 27 discussed the experimental regularities of this game: people usually start by contributing about 40 to 60 % of their endowment, and contributions decline when the game is repeated. This chapter continues from there: what happens if participants are allowed to pay to punish free riders?

Evidence

Altruistic punishment. Fehr and Gächter (2002) had 240 university students play a public goods game in groups of four, regrouped every period so that the same people never met again. In the punishment condition, after seeing the others’ contributions, participants could spend their own money to reduce others’ income: every unit spent took 3 units from the target. Because they would not meet again, a punisher could not benefit from any later change in the person punished, and punishment was in this sense “altruistic”. The result: cooperation flourished when punishment was possible and broke down when it was ruled out; most participants used punishment, and it was directed mainly at those who contributed less than others. The authors suggested that negative emotions towards free riders are the proximate cause of punishment (Fehr and Gächter 2002).

Evidence

Antisocial punishment. Herrmann, Thöni and Gächter (2008) ran the same kind of experiment with 1,120 university students in 16 cities in 15 countries (Boston, Melbourne, Nottingham, St. Gallen, Zurich, Copenhagen, Bonn, Dnipropetrovs’k, Seoul, Istanbul, Chengdu, Minsk, Samara, Athens, Riyadh and Muscat): fixed groups of four, an endowment of 20 each, a marginal per-capita return of 0.4, 10 periods each with and without punishment, and the same 1:3 punishment technology. They found a great deal of antisocial punishment: punishment of people who had contributed as much as, or more than, the punisher. Its extent varied widely; with punishment, average contributions ranged from 5.7 in Athens and 6.9 in Riyadh to 17.7 in Copenhagen and 18.0 in Boston; across the 16 participant pools, the more antisocial punishment, the lower cooperation (Spearman correlation −0.90). The cooperation-enhancing effect of punishment was at least weakly significant in 11 of the pools and not significant in the other five; the stronger the antisocial punishment, the smaller the increase. Weaker national norms of civic cooperation and a weaker rule of law went with more antisocial punishment. The authors concluded that punishment opportunities are socially beneficial only when complemented by strong social norms of cooperation (Herrmann et al. 2008).

Evidence

The efficiency cost of punishment. Punishment itself destroys income: both the punisher and the punished pay. Gächter, Renner and Sefton (2008) had participants in groups of three play the public goods game for 10 or 50 periods. With punishment, contributions were higher at both lengths and rose more with 50 periods; but only with 50 periods were net earnings, after the costs of punishment, higher than without punishment; with 10 periods they were lower (Gächter et al. 2008).

A program: public goods with punishment and antisocial punishment

How punishment can evolve is a theoretical puzzle: punishers pay a cost, while cooperators who do not punish enjoy the benefits (the “second-order free rider” problem). Antisocial punishment complicates matters further: in a model of the public goods game with optional participation, Rand and Nowak (2011) found that once punishment may be aimed at cooperators, the co-evolution of punishment and cooperation can be blocked (Rand and Nowak 2011).

Calculation

# Public goods with prosocial and antisocial punishment. Population N = 100; groups of
# n = 5 are drawn at random. Contributors pay 1; the pot is multiplied by r = 3 and shared
# equally among the participants. Then punishers act: each punishing act costs the
# punisher 1 and the target 3 (the 1:3 technology of Fehr & Gachter 2002 and Herrmann et
# al. 2008).
#   C  contributes, never punishes      D  defects, never punishes
#   P  contributes, punishes every D
#   A  defects, punishes every contributor (antisocial punishment)
#   L  loner: stays out and gets 1 (optional participation; groups with fewer than two
#      participants also give 1). Punishment acts only among participants.
# Expected payoffs in a population with i mutants follow from hypergeometric group draws.
# Imitation by pairwise comparison with rare mutations, as in p07: long-run share of time
# in each strategy, for weak (beta = 0.05) and strong (beta = 1) selection. All parameters
# are illustrative.
import math

R_MULT, NG, FINE, COST, SIGMA = 3.0, 5, 3.0, 1.0, 1.0
CONTRIB = {"C": 1, "D": 0, "P": 1, "A": 0, "L": None}

def payoff(me, others):
    """Payoff of 'me' in a group whose other members are 'others'."""
    part = [s for s in [me] + others if s != "L"]
    if me == "L" or len(part) < 2: return SIGMA
    pay = R_MULT * sum(CONTRIB[s] for s in part) / len(part) - CONTRIB[me]
    po = [s for s in others if s != "L"]
    if me == "P": pay -= COST * sum(CONTRIB[s] == 0 for s in po)
    if me == "A": pay -= COST * sum(CONTRIB[s] == 1 for s in po)
    if CONTRIB[me] == 0: pay -= FINE * sum(s == "P" for s in po)
    else: pay -= FINE * sum(s == "A" for s in po)
    return pay

def hyper(k, K, N, n):       # P(k of the n others are mutants), K mutants in N others
    return math.comb(K, k) * math.comb(N - K, n - k) / math.comb(N, n)

def expected(me, mut, res, i, N):
    """Expected payoff of a 'me'-player when the population has i mutants (mut) and N - i
    residents (res); the other NG - 1 group members are drawn from the remaining N - 1."""
    K = i - (me == mut)
    return sum(hyper(k, K, N - 1, NG - 1) * payoff(me, [mut] * k + [res] * (NG - 1 - k))
               for k in range(NG))

def fixation(mut, res, N, beta):
    tot, prod = 1.0, 1.0
    for i in range(1, N):
        d = expected(mut, mut, res, i, N) - expected(res, mut, res, i, N)
        prod *= math.exp(-beta * d); tot += prod
        if prod > 1e300: return 0.0
    return 1 / tot

def abundance(names, N=100, beta=1.0):
    n = len(names)
    T = [[0.0] * n for _ in range(n)]
    for j, res in enumerate(names):
        for i, mut in enumerate(names):
            if i != j: T[j][i] = fixation(mut, res, N, beta) / (n - 1)
        T[j][j] = 1 - sum(T[j])
    v = [1 / n] * n
    for _ in range(100000):
        w = [sum(v[j] * T[j][i] for j in range(n)) for i in range(n)]
        if max(abs(x - y) for x, y in zip(v, w)) < 1e-14: break
        v = w
    return v

for beta in (0.05, 1.0):
    print(f"selection strength beta = {beta}: long-run share of time; share contributing")
    for names in (["C", "D"], ["C", "D", "P"], ["C", "D", "P", "A"], ["C", "D", "L", "P"],
                  ["C", "D", "L", "P", "A"]):
        v = abundance(names, beta=beta)
        coop = sum(x for s, x in zip(names, v) if CONTRIB[s] == 1)
        sh = ", ".join(f"{s} {x:.3f}" for s, x in zip(names, v))
        print(f"  {'+'.join(names):10s} {sh};  {coop:.3f}")
print("group of 5, payoff per player (contributing 1 is the cost of cooperating):")
for grp in (["C"] * 5, ["D"] * 5, ["P"] * 5, ["P"] * 4 + ["D"], ["P"] * 3 + ["A"] * 2,
            ["C"] * 4 + ["A"]):
    pays = [payoff(grp[k], grp[:k] + grp[k + 1:]) for k in range(5)]
    print(f"  {''.join(grp)}: " + " ".join(f"{p:+.1f}" for p in pays)
          + f"   group total {sum(pays):+.1f}")
# selection strength beta = 0.05: long-run share of time; share contributing
#   C+D        C 0.109, D 0.891;  0.109
#   C+D+P      C 0.126, D 0.631, P 0.243;  0.369
#   C+D+P+A    C 0.063, D 0.385, P 0.134, A 0.418;  0.197
#   C+D+L+P    C 0.249, D 0.114, L 0.104, P 0.533;  0.782
#   C+D+L+P+A  C 0.222, D 0.092, L 0.115, P 0.533, A 0.038;  0.755
# selection strength beta = 1.0: long-run share of time; share contributing
#   C+D        C 0.000, D 1.000;  0.000
#   C+D+P      C 0.000, D 1.000, P 0.000;  0.000
#   C+D+P+A    C 0.000, D 0.500, P 0.000, A 0.500;  0.000
#   C+D+L+P    C 0.047, D 0.026, L 0.061, P 0.866;  0.913
#   C+D+L+P+A  C 0.047, D 0.025, L 0.061, P 0.866, A 0.000;  0.913
# group of 5, payoff per player (contributing 1 is the cost of cooperating):
#   CCCCC: +2.0 +2.0 +2.0 +2.0 +2.0   group total +10.0
#   DDDDD: +0.0 +0.0 +0.0 +0.0 +0.0   group total +0.0
#   PPPPP: +2.0 +2.0 +2.0 +2.0 +2.0   group total +10.0
#   PPPPD: +0.4 +0.4 +0.4 +0.4 -9.6   group total -8.0
#   PPPAA: -7.2 -7.2 -7.2 -10.2 -10.2   group total -42.0
#   CCCCA: -1.6 -1.6 -1.6 -1.6 -1.6   group total -8.0

The program’s population has 100 individuals; groups of 5 are drawn at random; each contribution of 1 adds 3 to the common pool, which is shared equally among the participants; each act of punishment costs the punisher 1 and the target 3. There are five strategies: cooperators (C), defectors (D), cooperators who punish defectors (P), defectors who punish cooperators (A, antisocial punishers), and loners (L) who do not take part and receive a fixed 1. Individuals imitate those who earn more, and mutations are rare.

Under weak selection (β=0.05\beta = 0.05), with only cooperators and defectors the population spends 11 % of the time contributing; adding punishers raises this to 37 %; adding antisocial punishers as well lowers it to 20 %, with antisocial punishers present 42 % of the time. This matches the direction of the experiment by Herrmann and colleagues: antisocial punishment cancels part of the effect of punishment. But the conclusion depends on the details of the model: when opting out (loners) is allowed, punishers are present more than half of the time and contributions are 78 %, and adding antisocial punishers lowers this only to 76 %, because antisocial punishers are themselves punished most in a population dominated by punishers. This differs from the conclusion of Rand and Nowak (2011): in their model cooperators, defectors and loners could each choose to punish any of the three types, giving 24 strategies; when only cooperators could punish defectors, average cooperation was about 87 %, and when all punishment strategies were allowed it fell to about 34 %, the same as without punishment, with loners doing most of the punishing of cooperators (Rand and Nowak 2011). In the present program loners punish nobody and antisocial punishers do not punish loners, so this result does not appear. The difference itself shows how sensitive the conclusions are to the model’s specification. Under strong selection (β=1\beta = 1), without loners punishment cannot get started at all: a single punisher in a population of defectors must punish everyone in its group at great cost, so the population stays with defection (and with antisocial punishers, who likewise do not contribute); with loners, punishers are present 87 % of the time.

The last part of the program lists each person’s payoff in a group of 5 to show the efficiency cost of punishment. When all cooperate each gets 2.0, the same as when all are punishers, because nobody needs to be punished. With 4 punishers and 1 defector the group’s total payoff is −8.0, lower not only than the 10.0 of full cooperation but also than the 0 of full defection: each punisher gets only 0.4 and the defector is driven down to −9.6. When 3 punishers and 2 antisocial punishers punish each other, the group total is −42.0. Punishment when used always destroys income; its value lies before use: if its threat is enough to make people cooperate, it rarely needs to be used. This is the logic of the 50-period experiment of Gächter and colleagues.

Note

Limits of the model. The program is illustrative: payoffs, group size, the 1:3 punishment technology, the updating rule and the selection strength are all chosen values, and with other values the shares of the strategies would change. What it shows is how the mechanisms relate to one another (punishment can sustain contributions; antisocial punishment can weaken it; whether it can get started depends on whether opting out is possible), not the proportions of any kind of person in any real group. “Punishment” in the laboratory is an anonymous one-off deduction of money; sanctions in real communities are usually public, open to appeal and liable to provoke retaliation, and the two differ greatly.

Ostrom’s eight design principles

Outside the laboratory, many communities have managed shared resources over long periods without coercion by the state: pastures, fisheries, forests, irrigation systems. In 1990 Ostrom compared such long-enduring and failed “common-pool resource” institutions and induced eight features that long-enduring institutions usually share (Ostrom 1990):

  1. clearly defined boundaries: who has the right to use the resource, and where the resource ends;
  2. rules for use and provision that fit local conditions;
  3. collective-choice arrangements: those affected by the rules can take part in changing them;
  4. monitoring: monitors are accountable to the users, or are users themselves;
  5. graduated sanctions: sanctions on rule-breakers rise with the seriousness and frequency of the offence;
  6. conflict-resolution mechanisms: low-cost, accessible arenas for settling disputes;
  7. minimal recognition of rights to organise: external governments do not deny the community’s right to make its own rules;
  8. nested enterprises: for parts of larger systems, governance activities are organised in multiple layers.

Evidence

Empirical support for the design principles. Cox, Arnold and Villamayor Tomás (2010) reviewed 91 studies that tested these principles and concluded that they are well supported empirically, while raising several theoretical issues and offering a revised formulation (Cox et al. 2010).

The eight principles map clearly onto the models above: boundaries and collective choice determine who is in the “population” and who sets the rules; monitoring corresponds to gathering the reputation information of Chapter 4; graduated sanctions and conflict resolution correspond to the questions of this chapter: how to punish without destroying cooperation, and how to tell deliberate rule-breaking from accidental slips (the execution errors of Chapter 3). The framework’s “graduated monitoring” is two principles: monitoring (principle 4) and graduated sanctions (principle 5). The principles were induced from cases, not derived from models; institutions with these features are more likely to endure, but the features are not sufficient conditions and cannot be transplanted to any community as they stand.

Hypothesis · not established

Graduated-and-forgetting hypothesis (not yet established). The author conjectures that in a community based on mutual help among members, a “graduated and forgetting” rule, under which records of violations are cleared step by step after a period without violations, will keep the same deterrence against deliberate free riders as a “three violations and you are out” rule while expelling fewer members in good standing for accidental slips, and that members will judge it fairer. The program in Chapter 8 shows how such a comparison is calculated, with a fictional example. A testable prediction comes from randomised comparison of community rules: the frequency of deliberate violations is similar under the two rules, while the loss of members in good standing is lower under the “graduated and forgetting” rule. If deliberate violations increase markedly under that rule, or the loss of members in good standing does not fall, the hypothesis is refuted.

Chapter summary

Allowing people to pay to punish free riders sustains contributions to public goods in experiments, even when punishers will never meet the punished again. But experiments across societies found a great deal of antisocial punishment, and where norms of cooperation and the rule of law are weak, punishment raises cooperation little. Punishment always destroys income when used, and only when interaction lasts long enough does the cooperation it brings make up for that cost. Models show that whether punishment can get started and how antisocial punishment affects it both depend on details such as whether opting out is allowed. Ostrom’s design principles, induced from long-enduring institutions for common resources, include monitoring, graduated sanctions and conflict resolution; they echo the models of this book but guarantee nothing. The next chapter reads these studies beside Confucian and Mohist writings on love and return.

References

Cox, Michael, Gwen Arnold, and Sergio Villamayor Tomás. 2010. “A Review of Design Principles for Community-Based Natural Resource Management.” Ecology and Society 15 (4): 38. https://doi.org/10.5751/ES-03704-150438.
Fehr, Ernst, and Simon Gächter. 2002. “Altruistic Punishment in Humans.” Nature 415 (6868): 137–40. https://doi.org/10.1038/415137a.
Gächter, Simon, Elke Renner, and Martin Sefton. 2008. “The Long-Run Benefits of Punishment.” Science 322 (5907): 1510. https://doi.org/10.1126/science.1164744.
Herrmann, Benedikt, Christian Thöni, and Simon Gächter. 2008. “Antisocial Punishment Across Societies.” Science 319 (5868): 1362–67. https://doi.org/10.1126/science.1153808.
Ostrom, Elinor. 1990. Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press.
Rand, David G., and Martin A. Nowak. 2011. “The Evolution of Antisocial Punishment in Optional Public Goods Games.” Nature Communications 2: 434. https://doi.org/10.1038/ncomms1442.

Fan, Y. (2026). Evolutionary Game Theory, Direct and Indirect Reciprocity and the Emergence of Cooperation: From Hamilton's Rule and Axelrod's Tournaments to Reputation and Costly Punishment, Read Beside Classical Confucian and Mohist Ethics [演化博弈、直接与间接互惠及合作涌现的数学模型] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23788551 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).