Sparse Early Warnings of Cooperation Collapse on Networks

When cooperation breaks down in a networked population, the usual health measure, the fraction of agents currently cooperating, is often the last thing to move. Defectors can take over a few well-connected positions while the population-wide cooperation rate stays flat, and then everything collapses at once. Medicine has a name for this kind of measure: a lagging indicator. Serum creatinine, for example, rises 48–72 hours after a kidney has already been injured. Clinicians respond by searching for upstream markers, and they often use LASSO, a regression that adds a penalty on the absolute size of coefficients so that most of them become exactly zero. This leaves a short, readable panel of predictors. We propose to do the same for cooperation. We would build a large library of features that describe where defectors sit in the network and how payoffs are distributed around them, then use sparse, stability-checked selection to find a small panel of structural signals. The test is whether that panel warns of collapse early enough that a cheap intervention can still reverse it.

The claim is not only that a sparse model predicts collapse. It is that the selected panel changes with selection strength in the way prior work predicts, which gives a data-driven account of when hubs protect cooperators and when they spread defection. The claim also has to hold up under the test that matters for lagging indicators: an alarm raised by the panel must leave enough time for rescue, compared with alarms raised by the cooperation rate, by general early-warning statistics, or by simply watching the largest hubs.

Core Claim

In evolutionary games on networks, collapse after an invasion is driven by strategy conditioned on position: which structural niches defectors occupy and whether they are clustered. Topology alone is static, and the global cooperation rate lags. So a sparse model over time-lagged, position-conditioned strategy features should give longer lead time at a fixed false-alarm rate than any global statistic. That extra lead time should turn into higher rescue rates when a fixed-budget intervention is triggered by the alarm.

A second, falsifiable claim follows. Ranking the invasions of cheaters in structured populations compared six hand-picked ways of placing invaders and found that degree-based placement is most damaging under weak selection, while clustered "P-weighted" placement is most damaging under strong selection. If our selection step is finding real mechanism, it should recover that switch without being told. We expect defector occupancy of high-degree nodes to be selected under weak selection, and the cohesion of defector clusters to be selected under strong selection. If the panel does not shift this way, we would treat it as fitting noise.

Research Program

Simulated populations. The main data come from simulated death–birth evolutionary dynamics with N = 500 nodes, following the setup of Ranking the invasions of cheaters in structured populations and Strategically positioning cooperators can facilitate the contagion of cooperation. Networks include Barabási–Albert scale-free graphs, Erdős–Rényi random graphs, lattices and small-world graphs. Games are the Prisoner's Dilemma, Stag Hunt and Snowdrift. We sweep the temptation payoff $b$, the selection strength $w$ and the invader seeding rule (random, degree, k-shell, P-weighted). Those papers measured an endpoint: the critical seeding density needed for takeover. We instead record the full trajectory of every run and label each time step by whether cooperation falls below half its pre-invasion level within the next $H$ steps.

Feature library. At each step we compute several hundred features, deliberately redundant. They include:

  • the defector fraction within each degree decile and each k-shell;
  • defector–cooperator edge density, broken down by the degree of the endpoints;
  • payoff gaps across the cooperator–defector boundary at hubs;
  • the size and cohesion of connected defector clusters;
  • local clustering around defectors.

Each feature is included at several lags. The global cooperation rate and its rolling variance and autocorrelation are in the library too, so the selection step is free to pick them if they are best.

Sparse model. We fit an $L_1$-penalized logistic model:

$$\hat\beta = \arg\min_\beta \; -\frac{1}{T}\sum_{t} \log p_\beta(y_t \mid x_t) + \lambda \sum_{j} |\beta_j| ,$$

where $x_t$ is the feature vector at step $t$, $y_t$ indicates collapse within $H$ steps, $T$ is the number of training steps, and $\lambda$ sets how many coefficients survive. Many features are near-duplicates, and LASSO is known to pick one of a correlated set arbitrarily. We therefore report stability selection: a feature counts as part of the panel only if it survives in a large share of refits on bootstrapped runs.

Runs, not time steps, are the unit of splitting, because steps within a run are strongly correlated. We fit panels separately for each selection-strength regime to test the hub claim. We also fit one pooled panel to ask whether a single "universal" panel exists.

Out-of-family check. As a secondary test, we move the panel to LLM-agent societies. We use the multilingual repeated Prisoner's Dilemma setup of Evolution of Cooperation in LLM Societies, in which 50 GPT-4o agents play over 9 topologies and 3 languages. That study reported sudden collapses on star and complete networks and sustained cooperation of 0.6–0.7 on scale-free networks, but analyzed them only with after-the-fact Pearson correlations. We would take the panel trained on synthetic dynamics, freeze its coefficients, and ask whether it forecasts those collapses. This is a transfer test, not a retraining. LLM runs are expensive, so this will be a small number of runs with an open-weight model, and we will report it as suggestive.

The idea from medicine. The lagging-indicator direction supplies the evaluation, not the method. The acute kidney injury work, Novel Plasma Biomarker-Based Model for Predicting Acute Kidney Injury After Cardiac Surgery, found upstream plasma markers (GDF15 alone reached an AUC of 0.80) that move before creatinine. The pneumonia study, Characterization of diffuse lung function in children with Mycoplasma pneumoniae pneumonia, found that oxygen saturation stays normal while lung diffusion capacity collapses. Both show that an indicator is only useful if it fires inside the window when damage can still be reversed.

We turn that into a measurable quantity: the rescue probability. When an alarm fires, a fixed intervention flips $k$ agents to cooperation at the highest-degree nodes, the placement rule shown to lower takeover thresholds in Strategically positioning cooperators. Rescue probability is the fraction of collapses this averts.

The medical sources also constrain the work. The kidney study chose its candidates with co-expression network clustering and then fit plain logistic regression. The clinical sparse scores in Nomogram for Predicting Cardiovascular Mortality in Incident Peritoneal Dialysis Patients (7 predictors, C-statistic 0.83 at one year) and Early COVID-19 respiratory risk stratification using machine learning (6 variables, AUC 0.789) treat patients as flat, unstructured vectors. We carry over their emphasis on a short, readable panel, but our features are defined relative to the interaction graph and change over time.

Baselines

Global cooperation rate (naive). The reference point is an alarm that fires when the rolling global cooperation rate drops by a set fraction. We tune that fraction on the training runs to match the sparse model's false-alarm rate, then report lead time and rescue probability. The sparse panel beats it only if it adds lead time at equal false alarms and that added time turns into more rescues. No reported result was found on Lacuna for forecasting cooperation collapse by any method. The nearest related figure is the endpoint-only 53.3% survival of GPT-4o reported for GovSim in Cooperate or Collapse, which is not a forecasting number.

Generic early-warning statistics. This is the strongest competitor. The literature on tipping points watches for rising variance and rising lag-one autocorrelation in a single aggregate time series. The idea is that a system approaching a tipping point recovers more slowly from small shocks. These statistics need no knowledge of the network at all. We compute them on the global cooperation rate with the same window lengths and the same tuning budget. We also give them to a logistic model without a penalty, so that any gap comes from the features themselves, not from the fitting procedure.

Hub watching. The simplest structural heuristic tracks defector occupancy among the top-5 or top-decile nodes, ranked by each of the six heuristics in Ranking the invasions of cheaters. We give it an oracle advantage: the best of the six is chosen separately for each regime.

Non-linear model on the same features. Gradient-boosted trees receive the identical feature library and the same run-level splits and tuning budget. In the COVID cohort this family beat LASSO on AUC, 0.86 versus 0.789, in Early COVID-19 respiratory risk stratification. That gap is the kind of shortfall we expect and will report.

Ablations isolating the key ingredient. Two ablations test whether position conditioning is what matters. The first removes all position-conditioned features and keeps only topology-free aggregates, including cooperation, variance, autocorrelation and the global defector-cluster size. The second removes the strategy features and keeps only static topology.

What counts as success. The proposal succeeds if the sparse panel beats the second, third and fourth baselines on lead time at a matched false-alarm rate, stays within a small AUC margin of the trees while using roughly ten or fewer features, and the first ablation loses most of the gain.

Why not just…?

Why not just watch the cooperation rate and its variance? This is the real threat. If the approach to collapse shows the classic slowing-down pattern in the aggregate series, generic statistics would match our panel and the structural story would be unnecessary. We suspect they will fail for hub invasions. There, collapse looks more like a cascade spreading from a few taken-over nodes than a slow drift toward a tipping point, and Evolution of Cooperation in LLM Societies observed exactly such abrupt collapses on centralized networks. That is only a hypothesis. The comparison against the second baseline, split by topology and seeding rule, settles it. If generic statistics match us on lattices but not on scale-free graphs, that is itself a finding about which collapses are cascades. If they match us everywhere, we will report the negative result and drop the claim.

Why not just monitor the biggest hubs? Because which hub-based signal matters depends on the regime. Ranking the invasions of cheaters shows that degree-based placement is most damaging only under weak selection, and that clustered placement is most damaging under strong selection. Our hub baseline is even allowed to choose the best heuristic per regime after the fact. Beating that oracle version is the bar. If the sparse panel does not beat it, the honest conclusion is that hand-picked hub signals are enough.

Why not just use gradient-boosted trees or a graph neural network? They will probably score somewhat higher on AUC, as in the COVID comparison. Our argument is not raw accuracy. A readable ten-feature panel can be checked against mechanism, for instance through the predicted switch between weak and strong selection, and a black box cannot be checked this way. We will report the AUC gap openly. If the trees also deliver much longer lead time and much higher rescue rates, the argument for sparsity rests only on readability, and we will say so.

Why not just use stepwise selection? With hundreds of near-collinear lagged features, stepwise selection gives panels that change from one resample to the next. We will show this directly by reporting how stable stepwise selection is across bootstraps next to our stability-selection frequencies.

Decisive Checks

Three results decide the proposal.

The first is the rescue curve. We plot rescue probability against alarm time for every method under the same intervention budget $k$. The sparse panel must shift this curve by a clear margin over generic early-warning statistics on scale-free graphs.

The second is recovery of known mechanism. Stability selection must pick high-degree defector occupancy under weak selection and defector-cluster cohesion under strong selection, as Ranking the invasions of cheaters would predict. We will also sweep selection strength, graph density and $b$ continuously to locate where the panel switches. This directly maps the unresolved boundary between dispersed and clustered seeding raised by Strategically positioning cooperators.

The third is transfer. With coefficients frozen, the panel should give earlier warning than the cooperation rate in the LLM-agent runs, where language-dependent drift toward defection was reported. If it holds only on synthetic death–birth dynamics, we will describe the panel as specific to that dynamics, not as a general early-warning signal for cooperation.