One Extra Site Can Make Cooperation 7× Easier to Spread — If You Pay the Hidden Price
There is a number that keeps showing up in the study of cooperation, and it is deceptively simple: three. It is the threshold at which a single defector, dropped alone into a group of cooperators, stops being able to take over. Below that ratio of benefit to cost, selfishness wins; above it, the collective sometimes survives. But Abhishek Chowdhury, a physicist at the Indian Institute of Technology Bhubaneswar, has found that how you wire the population matters just as much as that number — and that the geometry of who-talks-to-whom can make cooperation easier or dramatically harder to establish.
The paper's most striking result is almost counterintuitive: a bigger, more connected population is not always better for cooperation. On a simple path of four sites, no arrangement of contact weights ever favors cooperation at weak selection. Stretch that path to five sites and cooperation can win — but only if the benefit-to-cost ratio exceeds seven. Stretch it to six sites, and suddenly you can push that threshold all the way down to one, the theoretical floor where helping pays for itself. The difference between a ratio of seven and a ratio of one is not a change in the payoff structure. It is a change in how the individuals are connected (Chowdhury, 2026).
Threshold infima on small promoter networks
Sharp weak-selection threshold infima (minimum benefit-to-cost ratio for cooperation to be favored) for the three smallest promoter networks studied in the paper. Lower means cooperation is easier to establish.
| Label | Value |
|---|---|
| Four-cycle C4 | 3 |
| Five-site path P5 | 7 |
| Six-site path P6 | 1 |
This is a paper about protected domains — clusters of cooperators that shield one another from defection long enough to spread — and about the hidden price of building them. To make cooperation nearly free to invade, you need contact weights that differ from each other by a factor that grows like as you approach the ideal threshold . That is a staggering demand. At a threshold of 1.1, it means the strongest contact in the population must be roughly 65,600 times stronger than the weakest — and the more you demand, the slower the takeover becomes correction, and the harder it is to observe statistically. Improving the invasion threshold, Chowdhury shows, can require increasing contact heterogeneity, precision, and observation time all at once.
The Science
The setting is a donation game, the cleanest possible model of a public-good dilemma. A cooperator pays a cost to confer a benefit on a partner; a defector pays nothing and collects the benefit anyway. Played once, defection strictly dominates: whatever your partner does, you do better by not paying. But a population of cooperating microbes — or bacteria secreting a public enzyme, or cells contributing to a shared tissue — can all gain if everyone cooperates. The question is how a rare cooperative mutant ever gets a foothold when it starts surrounded by defectors.
The dynamics here are those of social copying. At each step, one individual is chosen as the "source" in proportion to its payoff — its fecundity — and copies its trait to a randomly chosen neighbor. A cooperator's payoff depends on its local environment: the fraction of cooperators among the neighbors it actually interacts with. The paper uses a reciprocal kernel: the same weights that determine who influences whom also determine who benefits from whom. This is source-first Birth–death updating with averaged payoffs, a convention that matters enormously, because different update rules are known to produce opposite conclusions about when cooperation spreads (Chowdhury, 2026).
The key comparison is a paired experiment. Introduce a single cooperator uniformly into an all-defector population and measure its fixation probability — the chance it takes over before being wiped out. Then run the mirror image: a single defector introduced into an all-cooperator population. Cooperation is favored when the cooperator's takeover probability exceeds the defector's. This symmetric comparison, the paper stresses, is fair in a way that comparing cooperation against neutral drift alone is not.
To analyze this without approximation, Chowdhury uses an exact weak-selection expansion. The genius of the method lies in what it tracks. The mean density of cooperators stays perfectly flat at neutrality — it never moves — so a crude "common density" approximation would conclude that nothing happens. But the correlations between sites evolve, and it is precisely those correlations that carry the response to selection. The fixation response reduces to an integral over the neutral history of pairwise disagreements between sites, objects with a clean ancestral interpretation: two lineages tracing backward through copying until they meet at a common ancestor. This reduces a problem with transient configurations to a set of ancestral pairs that can be solved exactly (Chowdhury, 2026).
What They Found
The results resolve into a crisp classification of small networks, each one a test case for a distinct mechanism.
The four-cycle — a square, four sites each linked to two others — is the minimal promoter: the smallest network of sites and edges that can favor cooperation at all. Its threshold infimum is . But here is the catch: this threshold is approached only in a singular limit where some contact weights become enormous and others vanish. No finite set of positive weights actually attains it)Skip the setup and you miss the point. The number 3, in this context, is not a universal constant of cooperation. It is the best possible for that particular four-site wiring, and even that requires extreme contact asymmetry.
The five-site path tells a richer story. Its threshold infimum jumps to seven — more than double the four-cycle's. The reason, Chowdhury shows, is that on five sites the population must form a "protected domain": a tight cluster of cooperators around a central, slowly-overwritten site that pins the cluster together while it competes across a weak bridge with the defector side. The establishment of that domain carries an adverse response that the four-cycle avoids. Seven is the price.
Then comes the surprise. Every path with at least six sites can approach a threshold of — the absolute floor at which cooperation barely pays for itself. Six sites change the mechanism qualitatively. Instead of one protected cluster fighting the world, the population forms two coherent triples, each anchored around its own slowly-overwritten central site. Each triple aligns its internal followers before the two cohorts compete across their weak central bridge (Chowdhury, 2026).
The six-site construction extends to every longer path.
Contact contrast required to approach threshold one
Minimum contact contrast (largest weight divided by smallest) required on the six-site path to reach a threshold of 1+epsilon. Computed as 656 * epsilon^-2 from the paper's asymptotic formula. Higher epsilon (further from threshold one) requires far less contrast.
| Label | Value |
|---|---|
| Contrast to reach 1.05 | 262,400 |
| Contrast to reach 1.10 | 65,600 |
| Contrast to reach 1.20 | 16,400 |
| Contrast to reach 1.50 | 2,624 |
None of these infima is attained at any finite positive weights; they are limits, approached as the contact hierarchy becomes increasingly extreme. Yet on six sites, a global mathematical bound forces an important conclusion: every near-barrier design that succeeds must follow the protected-domain hierarchy. There is no clever alternative wiring that sneaks around it. The paper proves this with a positive polynomial obstruction that rules out all other families of designs.
The cost of this idealized threshold appears when you ask how much contrast in contact strengths is required. To reach a threshold , the largest contact weight must exceed the smallest by a factor asymptotic to
Near the floor, precision becomes punishing. And the absorption time — how long the invasion takes on average — grows like , so the closer you push toward threshold one, the longer you must wait and the more observation time you need to see anything decisive. Crucially, these two costs trade off against each other: extra contrast can shorten the invasion, but hovering near the minimum leaves almost no room for error in the contact design.
Why This Changes Things
The deepest lesson is a reframing of what "cooperation is favored" means. In the standard reading, a benefit-to-cost ratio above some threshold guarantees that cooperation can spread. Chowdhury's analysis shows the threshold is not a property of payoffs alone — it is jointly determined by payoffs, wiring, time, and statistical resolvability. On the five-site path, you need ; on the six-site path, you can get to but only by building a contact structure whose precision requirements grow without bound as you approach the goal.
This connects to a long-standing question in the field about whether weighted networks with unequal site strengths can overcome the obstruction that defeats isothermal graphs — graphs where every site has equal influence. Ref. (Allen et al., 2017, Discussion) explicitly asked whether nonisothermal weighted graphs could favor cooperation under source-first averaged-payoff dynamics. Chowdhury's answer is a definitive, quantified yes — but with a sharp price tag attached, and the price grows as the threshold improves.
The science of cooperation has generally treated "does this network promote cooperation?" as a binary question. This paper replaces that with a continuum of achievable thresholds, each with its own cost in heterogeneity, precision, and time. The five-to-six-site transition is the cleanest illustration ever given of how a single added site can qualitatively transform what a population can accomplish — not by giving it more individuals, but by enabling a different arrangement of protected domains.
There is also an important negative result buried in the analysis. In the finite-selection window where the dynamics are actually observable, the maximal fixation advantage — the boost in takeover probability cooperation gets over defection — is only of order when the contrast exceeds its exact minimum by order . In plainer terms: you can make cooperation almost free to invade, but the measurable advantage shrinks even as the design becomes more extreme. An improving invasion threshold can coexist with a shrinking observable signal. This is the kind of result that disciplines the field's optimism about amplifier networks — the capacity to make cooperation win can come at the cost of making it hard to see that it wins.
What's Next
The sharp bounds in this paper are for paths and cycles, the simplest topologies. The natural next step is to ask whether the protected-domain mechanism survives on richer geometries — trees, grids, random graphs, or networks with community structure, where the two-triple construction of the six-site path might generalize to multi-cluster hierarchies. The toric and convex-dual mathematics the paper develops, which show that the optimal weight design is eventually exactly reflection-symmetric, seem likely to extend to larger symmetric supports, but nobody has checked yet.
The time cost is one of the most practically relevant findings. If a real microbial community must wait a time of order for a cooperative mutant to fix, then there is a natural clock against which the whole experiment runs. The tradeoff between contrast and observation time gives experimentalists a concrete target: given a maximum feasible experiment duration, what is the tightest threshold a contact design can realistically promise? The paper's constrained quadratic geometry provides exactly the framework to answer that.
There is also the question of what all this means for real public-goods systems, from bacteria secreting siderophores to tumor cells cooperating to resist therapy. The donation game is an abstraction, but the structural insight — that locally protected clusters beat globally diluted cooperation, and that building them costs heterogeneity, precision, and time — is exactly the kind of principle that survives translation to messier biology. A quantitative application would need a specific model of transport and uptake, as the paper notes, but the design principles are now on the table.
What makes this work genuinely exciting is that it turns a question that seemed closed — "can weighted graphs favor cooperation?" — back into an open optimization problem with crisp, provable answers and honest costs. Cooperation, it turns out, is not just something that happens at the right ratio of benefit to cost. It is something you have to engineer the neighborhood for, and every improvement in the threshold comes with a bill. Knowing the price is the first step to deciding whether it is worth paying.