Cooperation is, on the face of it, irrational. In any single interaction where you can choose to cooperate at cost c to provide benefit b to another, or defect and pay nothing, defection dominates. The Prisoner's Dilemma formalizes this: mutual cooperation yields b − c for both, mutual defection yields 0 for both, but a defector facing a cooperator gets b while the cooperator gets −c. Since b > b − c > 0 > −c, defection is the strictly dominant strategy. Every rational agent should defect. Every population of rational agents should converge to universal defection.
And yet cooperation is everywhere. Cells cooperate in organisms. Organisms cooperate in ecosystems. Humans cooperate in societies built on trust. Something is wrong with the analysis—or rather, something is missing. The missing ingredient is structure.
In a well-mixed population—where every agent interacts with every other with equal probability—defection truly does dominate. Natural selection in well-mixed populations drives cooperators to extinction. This was the standard result for decades, and it led to a puzzle Martin Nowak called the fundamental problem of evolution: how does cooperation persist?
The answer, discovered by Nowak and May in 1992 and refined over the next two decades, is that spatial structure changes the game. When interactions happen on a graph—where each agent only plays with its neighbors—cooperators can form clusters. Within a cluster, cooperators interact mainly with other cooperators, receiving b − c per interaction instead of being exploited. Defectors on the cluster boundary exploit some cooperators, but they also interact with other defectors and receive 0. If the cluster is dense enough, the interior cooperators outcompete the boundary defectors.
This is spatial reciprocity: the graph itself creates the conditions for cooperation, without requiring memory, punishment, or reputation. The topology does the work.
The mathematical framework is the replicator equation. In a well-mixed population with strategy frequencies xi, the continuous-time dynamics are:
where fi is the fitness of strategy i and φ = ∑ xjfj is the average population fitness. Strategies that exceed average fitness grow; those that fall below it shrink. This is the mathematical distillation of natural selection.
On graphs, the dynamics are more subtle. A node’s fitness depends only on its neighbors, not the global population. The update rule matters: under Death-Birth updating, a random node is selected to die (vacate its position), and its neighbors compete to place a copy of their strategy, with probability proportional to fitness. Under Birth-Death, a node is selected to reproduce proportional to fitness, and its offspring replaces a random neighbor. These are not equivalent—Death-Birth favors cooperation more than Birth-Death.
The simulation below implements Death-Birth dynamics. A random node vacates. Its neighbors compete: each neighbor’s probability of filling the vacancy is proportional to 1 + δ·payoff, where δ is selection intensity. The winner’s strategy replaces the dead node’s strategy. Watch how cooperator clusters form and dissolve.
In 2006, Ohtsuki, Hauert, Lieberman, and Nowak proved a remarkably clean result: on regular graphs (where every node has exactly k neighbors), cooperation evolves under Death-Birth updating if and only if:
That is: the benefit-to-cost ratio must exceed the average degree. This is the Ohtsuki-Nowak rule, sometimes called the “graph selection criterion.” It is simple, exact (in the limit of weak selection and large population), and deeply counterintuitive.
The counterintuition: lower connectivity helps cooperation. A sparsely connected network is better for cooperators than a dense one. On a ring (k = 2), cooperation evolves whenever b/c > 2. On a complete graph (k = N − 1), the condition is nearly impossible to satisfy. Connectivity, which we usually associate with trust and community, actually enables exploitation. A defector in a dense network can exploit many cooperators simultaneously. A defector in a sparse network can only exploit a few—and is surrounded by other defectors who provide no payoff.
Try it in the simulation above: set the topology to regular lattice, lower k to 2, and watch cooperation thrive. Then increase k to 8 or 10 and watch it struggle. The phase transition at b/c = k is visible in real time.
In finite populations, evolution is stochastic. A mutant strategy—say, a single defector in a population of cooperators—has a fixation probability ρ: the probability that its lineage eventually takes over the entire population. Under neutral drift (no selection), ρ = 1/N. Selection can amplify or suppress this.
Graph topology affects fixation profoundly. Lieberman, Hauert, and Nowak showed that certain graph structures are amplifiers of selection—they increase the fixation probability of advantageous mutants beyond 1/N—while others are suppressors. Scale-free networks, where a few hubs have many connections, are strong amplifiers: a fit mutant that reaches a hub can spread rapidly. Regular graphs are neither amplifiers nor suppressors; they are “isothermal” and fixation probability equals the Moran process result.
This has direct implications for trust networks. A trust graph with hub structure (a few highly-trusted authorities) is more susceptible to invasion: if a malicious actor captures a hub, its influence spreads faster than on a flat, regular graph. Decentralization is not just an ideology; it is a structural defense against evolutionary invasion.
A Sybil attack in a reputation network is, in evolutionary terms, a coordinated mutant invasion. The attacker introduces a cluster of nodes running a malicious strategy—nodes that cooperate with each other (to build internal reputation) while defecting against the rest of the network (to extract value). This is not a single random mutant; it is a pre-formed cluster injected at a chosen location in the graph.
The fixation probability of such a cluster depends on two factors: the cluster’s internal connectivity (how well the Sybils support each other) and its attachment point (where in the graph the cluster connects). Hub-targeting—connecting Sybil nodes to high-degree nodes—exploits the amplification effect of scale-free topology. The hub’s high fitness (due to many cooperating neighbors) means it influences many updates; once it is “converted,” its strategy change propagates through the network. Periphery-seeding, by contrast, starts at low-degree nodes where propagation is slower but detection is also harder.
Click “Inject Sybil Cluster” in the simulation to see this in action. The injected cluster targets the highest-degree nodes, connecting coordinated defectors to the most influential positions. On a scale-free network, the effect is dramatic. On a regular lattice, the invasion is slower and more easily contained.
The Tier 2 DMI (Decentralized Metric of Identity) in NIP-XX’s kind 30085 events includes a graph diversity metric: a measure of how structurally diverse a node’s trust attestations are. This metric—based on the entropy of neighbor degree distribution and the proportion of non-redundant paths—is, in evolutionary game-theoretic terms, measuring resistance to fixation of malicious strategies.
A node whose trust graph is diverse (attestations from nodes of varying degree, in different communities, with non-overlapping neighborhoods) is structurally equivalent to a node in a graph that suppresses selection. A mutant strategy (Sybil invasion) has low fixation probability because the diverse topology prevents the cascade dynamics that amplifiers enable. Conversely, a node whose trust comes from a single cluster of similar-degree nodes is structurally in an amplifier: one compromised attestor can flip the whole group.
The evolutionary game theory here is not a metaphor. The replicator dynamics on graphs are the literal mathematical framework underlying why certain trust topologies are robust and others are fragile. The Ohtsuki-Nowak condition b/c > k is a theorem about when cooperation—honest behavior in a reputation system—is evolutionarily stable. The DMI graph diversity metric is an empirical proxy for the conditions that make cooperation a Nash equilibrium on the trust graph.