Cooperation has a mathematics problem—and two powerful ways to solve it
A mathematical toolkit can explain when networks tip the odds toward cooperation—and show exactly where its answers stop scaling.
160,000 simulated people fit where exact theory has reached about 5,000.
A cooperator surrounded by defectors can sometimes help cooperation take over—not because the group suddenly becomes kinder, but because the arrangement of social ties changes who meets whom and whose success gets copied. For three decades, computer simulations have made this spatial reciprocity visible. The harder question has been why it happens, and under what conditions it should happen again.
Wang, Li and Szolnoki’s review traces how mathematicians have turned that question into a set of solvable problems. Its central message is both hopeful and usefully restrained: two approaches—pair approximation and coalescent theory—can explain when cooperation is favored in structured populations, but neither is a universal recipe. Their strongest answers depend on assumptions such as weak selection, simplified networks or a small enough population to calculate. Mathematics does not replace simulation; it tells us what simulations may be showing, and where our understanding still runs out.
That distinction matters beyond games played on lattices. Human cooperation travels through social networks, institutions and repeated interactions. A model cannot tell us, by itself, how a community will behave. But it can separate a plausible mechanism from an appealing story—and reveal which features of a network might make collective action easier or harder.
The Science
Evolutionary game theory asks how strategies spread when individuals receive different payoffs and learn from one another. In the classic prisoner’s dilemma, a person can gain by defecting even when mutual cooperation would benefit everyone. In a well-mixed population, that tension often makes defection hard to dislodge. On a network, however, people interact mainly with neighbors. Cooperators may form clusters, helping one another enough to persist while defectors exploit one another or remain isolated. This family of effects is called spatial reciprocity.
The paper is a review of theory, not a report of one new experiment. It follows research that began with simulations of spatial games and asks how mathematical tools have made their outcomes more interpretable. The models typically assign each person one of two strategies—cooperate or defect—and specify how people interact, earn payoffs and update their strategy. Update rules matter: pairwise comparison means a person compares with a neighbor; death-birth updating replaces a focal individual with a neighbor chosen in proportion to fitness; imitation means a person copies another’s strategy. The review focuses especially on the death-birth rule, which can produce nontrivial effects because all neighbors compete to replace the focal individual.
A parameter called selection strength, (\beta), controls how strongly payoffs shape learning. When (\beta) is near zero, payoff differences barely matter and strategy changes resemble random drift. This is the weak-selection limit. At stronger selection, payoff differences influence imitation more sharply, and the system can sustain changing patterns rather than quickly settling into all cooperation or all defection. The paper’s illustrative lattice makes this contrast visible: the same game can look nearly random under weak selection and remain patterned under stronger selection.
The paper’s two selection-strength examples
Selection-strength values used to illustrate weak and relatively strong selection in the paper’s 100 × 100 lattice example under death-birth updating.
| Label | Value |
|---|---|
| Weak selection, β = 10⁻² | 0.01 β |
| Relatively strong selection, β = 10¹ | 10 β |
The review explains two ways to move from these dynamics to predictions. Pair approximation tracks not only the overall fraction of cooperators, but also how often cooperators are neighbors of cooperators. That local information is a compact stand-in for a much larger network. In the weak-selection limit, neighboring-strategy patterns adjust faster than the overall population share. Researchers can therefore approximate the local pattern and reduce the dynamics to a simpler equation for the cooperator fraction.
Coalescent theory takes a different route. Instead of approximating the population with a few average quantities, it follows how ancestral lineages meet when strategy changes are neutral—when payoffs have no influence. The expected time for two lineages to meet, called a coalescent time, captures how network structure shapes shared ancestry and local correlations. Researchers then calculate how a small payoff effect perturbs that neutral baseline. This can yield exact conditions for whether selection increases the chance that cooperation takes over.
The two methods answer related questions at different levels. Pair approximation is an approximate, relatively compact framework that can generate replicator equations—equations describing how strategy frequencies change. Coalescent theory can be exact under its assumptions, but asks for detailed calculations tied to the update rule, payoffs and network. In finite populations, the outcome is often summarized through fixation probability: the chance that a strategy eventually becomes the only strategy left.
What They Found
The review’s most famous result is the “(b/c>k)” rule. In a donation game, a cooperator pays a cost (c) to provide a benefit (b) to another person. On an infinite, regular network with death-birth updating and no triangles, pair approximation predicts that cooperation is favored when the benefit-to-cost ratio exceeds the number of neighbors, (k). In plain language: on this particular kind of network, the benefit must be large enough relative to the cost to compensate for the number of neighbors among whom the effects are spread. This is a precise result, not a universal law. It relies on a regular network, weak selection and other stated assumptions.
A second key result comes from the neutral baseline. In a finite population of (N) individuals, with no payoff-driven selection, a single randomly placed cooperator has a (1/N) chance of eventually taking over. That is the baseline against which selection is judged. If weak selection raises the chance above (1/N), it favors cooperation; if it lowers it, it favors defection. The criterion turns an intuitive claim—“the network helps cooperation”—into a measurable comparison.
Coalescent theory broadens what can be tested. Early results applied to regular, symmetric networks such as lattices. A major advance extended the approach to arbitrary population structures and to more than one kind of starting configuration (Allen et al., 2017; Wang et al., 2026). The basic recipe is to derive a condition under neutral drift, identify the coalescent times required by the payoff function, solve for those times on the network, and substitute them back into the condition.
That recipe also makes a useful research map. If a cooperation condition already exists, researchers can calculate it on new networks. If the payoff function changes—for example, to a public-goods game in which people contribute to a shared benefit—they derive the new condition and determine which coalescent times it needs. If the strategy-update process itself changes, they must rebuild the derivation from the beginning. The authors describe these as three tiers, from network applications, through new games, to new learning dynamics. Progress at the most foundational tier can unlock many applications below it.
Exact analysis and simulation reach different scales
Population-size examples cited in the review: the approximate scale where laptop memory becomes insufficient for coalescent calculations, the largest exact-analysis population reported by the authors, and a lattice size the authors say laptops can readily simulate.
| Label | Value |
|---|---|
| Laptop memory obstacle | 1,000 individuals (approx.) |
| Largest exact analysis reported | 5,000 individuals (approx.) |
| Simulation example: 400 × 400 lattice | 160,000 individuals (approx.) |
The comparison of computational scales is sobering. The authors report that exact coalescent calculations can involve solving a matrix problem whose dimensions grow with population size, and that memory becomes a practical obstacle around populations of roughly (10^3) for a personal laptop. The largest population they identify in the literature is about (5\times10^3). By contrast, they say a laptop can readily simulate a (400\times400) lattice—(1.6\times10^5) individuals. These are examples cited by the review, not a controlled benchmark of all hardware or algorithms. They show why exactness and scale pull in opposite directions: a mathematically complete answer for a modest system may be more costly than a simulation of a much larger one.
The distinction between weak and strong selection is another central finding. In weak selection, the system begins close to neutral drift, allowing a perturbative calculation—one that measures how a small change shifts a known baseline. That strategy yields tractable analytical answers. At stronger selection, the population may settle into a fluctuating, non-equilibrium state, where the fraction cooperating does not quickly resolve into one of the two absorbing outcomes. General exact analysis there can become exponentially difficult, and solutions are known only for special structures, such as star networks.
Why This Changes Things
A simulation can show that cooperation flourishes in a particular model. It may not reveal whether the cause is clustering, degree differences, the update rule or a subtle interaction among them. Mathematical analysis can isolate which ingredients matter and state the assumptions under which a result follows. That is valuable wherever researchers use models to reason about collective action: in social science, ecology, economics and the study of institutions.
But the paper does not claim that human behavior reduces to a lattice game. Its contribution is more methodological. It gives researchers a way to ask disciplined “what if” questions: if the rules stay the same but the network changes, does cooperation become more likely? If a public-goods payoff replaces a pairwise donation game, what new correlations must be calculated? If people learn differently, does the old criterion survive? Such questions can help prevent a familiar mistake in complex-systems research: treating a striking simulation pattern as though its mechanism were already understood.
The exactness of coalescent theory is particularly important here. A condition derived for a specified model can clarify which network features enter the result, rather than relying on a visual intuition that “more clustering helps.” Yet exact does not mean universally realistic. These results still rest on a model’s assumptions, and the review emphasizes that many calculations are limited to weak selection, fixation and comparatively simple dynamics. A theorem can tell us what follows from those premises; it cannot establish that the premises describe a particular community.
The limitations of pair approximation make the same point from the other direction. It assumes, in its basic form, an infinite regular network without triangles. Because it tracks pairs rather than larger neighborhoods, it can miss the effect of clustered groups. In the public-goods example discussed by the authors, pair approximation does not capture a clustering coefficient that appears in the corresponding coalescent result. A compact model can be powerful precisely because it leaves details out; the danger is forgetting which details it left behind.
For practitioners, the likely value is not a ready-made policy lever. It is a better toolkit for testing claims about networks and cooperation. Before saying that a platform’s design, a community’s social ties or an institutional change will promote cooperation, one would need evidence about real incentives and behavior. The theory can sharpen the hypotheses and identify mechanisms worth measuring; it cannot substitute for that evidence.
What’s Next
The authors see a broad research frontier. Existing cooperation conditions can be applied to more network structures; new payoff functions can bring in social goods, nonlinear interactions or higher-order networks; and changes in learning—such as multilayer or time-varying networks—require new derivations. The tiers matter because a breakthrough in how strategy updates are represented can open a large family of subsequent game and network analyses.
There is also a computational frontier. Exact calculations can be valuable even when they do not scale as far as simulations, but the present cost constrains which populations can be studied. The review points to sparse-network methods as one way to ease that burden. Better algorithms and computing resources may extend the range, though larger populations alone will not solve the conceptual problem of choosing a useful model.
A more interpretive challenge remains. The mathematics often focuses on local statistics—who is connected to whom, and how often lineages meet. Simulations show large-scale patterns unfolding through time: clusters grow, fracture and move. The authors argue that connecting exact mathematical results to those physical pictures is an important next step. That could make theories more intelligible while also exposing mechanisms that a local summary misses.
The durable lesson is not that mathematics has finally solved cooperation. It is that the question has become more precisely divisible. We can now distinguish which parts of the puzzle admit exact answers, which require approximation, and which still depend on simulation. For a world that repeatedly asks people to cooperate across networks, that is meaningful progress—not a universal formula for goodwill, but a clearer way to find out when the structure around us makes cooperation more likely.
All theories are tools with certain limitations that complement our understanding of cooperation.
Sign in to join the conversation.
Comments (0)
No comments yet. Be the first to share your thoughts.