The Hidden Rules of How Things Spread

In 2014, an Ebola outbreak began spreading through Guinea, then Liberia, then Sierra Leone. Epidemiologists could map exactly who infected whom—they had the network of transmission. What they couldn't see was how the infection spread: Was each infection an independent roll of the dice, independent of how many people around you were sick? Was it more likely to spread if your neighbors were infected? Was there a threshold effect—a critical mass of sick people that suddenly made infection inevitable?

This is the puzzle at the heart of a new paper by researchers at Tampere University, Goethe University Frankfurt, Central European University, and the National Autonomous University of Mexico. The question isn't just academic. When you can't identify the mechanism driving a spreading process, you can't predict how it will behave under new conditions. You can't know whether closing schools will slow an outbreak, or whether a new advertising campaign will tip public opinion, because you don't understand the underlying dynamics.

The paper's central result is both reassuring and sobering. The authors—Javier Ureña-Carrión, Tiago P. Peixoto, and Gerardo Iñiguez—show that we can statistically infer which spreading mechanism is operating, even from incomplete observations of a complex system. But they also show that this inference often fails in real-world conditions, and that knowing when inference will work is just as important as knowing how to do it.

"We find that accuracy generally increases in sparse networks, which are prevalent in real-world systems, and near phase transitions," the researchers write. In other words, the very conditions that characterize most social, biological, and technological networks—where everyone isn't connected to everyone, and where processes often hover near tipping points—happen to be the conditions where our inference tools work best.

The Science

To understand what Ureña-Carrión and colleagues accomplished, it helps to first understand what they were trying to solve. In network science, we've gotten good at two things: measuring network structure (who's connected to whom) and simulating dynamics on networks (what happens when something spreads). What's harder is the inverse problem: given a network and observations of a spreading process, can you figure out the rules that generated those observations?

This is harder than it sounds. A network is a snapshot of potential connections; a spreading process unfolds over time, often stochastically, and the rules governing it might depend on the number of infected neighbors a node has, or the node's own degree, or more subtle factors. The data you collect is always incomplete—you can't observe every single interaction—and the same observable pattern could in principle be generated by many different underlying mechanisms.

The researchers focused on binary-state dynamics: situations where each node can be in one of two states—susceptible or infected, off or on, disagree or agree. This framework encompasses a remarkable range of real-world phenomena. Disease epidemics follow this pattern. So do social contagion processes like viral memes, opinion dynamics like voting behavior, and even some models of technological adoption. Each of these can be described as a stochastic process where nodes switch states based on rules that depend on their neighbors' states.

The key insight is that these rules—the "mechanisms"—can be characterized by transition probabilities. For a node with degree k (meaning it has k neighbors) and m infected neighbors, there's a probability F${k,m}$ that it will become infected, and a probability R${k,m}$ that an infected node will recover. Different mechanistic assumptions translate into different functional forms for these probabilities.

Consider the classic susceptible-infected-susceptible (SIS) model of disease spread. Here, the infection probability depends on the number of infected neighbors according to a simple exponential formula: F${k,m}$ = 1 - (1 - θ${F}$)$^{m}$. This means each infected neighbor contributes independently to the probability of infection. In contrast, a threshold model assumes infection only happens once a certain fraction of neighbors are infected—the probability jumps from near-zero to near-one at a critical threshold.

The authors identified six archetypal mechanisms that appear repeatedly in the literature: Independent, Simple, Voter, Ising, Threshold, and Majority. Each represents a different hypothesis about how spreading depends on neighborhood context. The Independent model assumes infection depends only on intrinsic node properties, not neighbors. The Simple model (like SIS) assumes each neighbor contributes independently. The Voter model assumes infection probability scales with the fraction of infected neighbors. The Ising model—borrowed from statistical physics—assumes a smooth, temperature-dependent transition. The Threshold model assumes a critical mass effect. And the Majority model assumes a node switches to match most of its neighbors.

Figure 1 from the paper illustrates this framework. Part (a) shows diagrams of each mechanism, highlighting how they use varying information about a node's neighborhood to drive state changes. Part (b) shows how infection and recovery mechanisms combine to determine dynamics. The authors use a visualization where recovery matrices are "mirrored"—displaying nodes with the opposite state—so the same graphical format can show both infection and recovery transitions.

The mathematical framework centers on likelihood: given a particular trajectory of states over time, what is the probability of observing that trajectory under each candidate mechanism? By comparing these probabilities via log-likelihood ratios, the researchers could determine which mechanism was most consistent with observed data.

The crucial innovation is showing that the expected likelihood ratio between a true model and a competing model can be expressed in terms of the Kullback-Leibler divergence—a measure of how different two probability distributions are—weighted by the amount of observed data in different parts of the system. This gives a principled way to predict how detectable a mechanism will be before conducting the full analysis.

To test their framework, the researchers simulated dynamics on synthetic networks drawn from the configuration model—a standard approach that generates random networks with specified degree distributions. They explored a parameter space encompassing different combinations of infection and recovery mechanisms, different network structures (varying average degree and degree variance), and different initial conditions. For each point in this space, they could simulate data, apply their inference method, and measure how often it correctly identified the true underlying mechanism.

They also developed analytical approximations using what are called Approximate Master Equations (AMEs)—coupled differential equations that track the average fractions of susceptible and infected nodes in each network class. In the thermodynamic limit of infinite system size, these equations give exact predictions. The authors showed that despite being derived for infinite systems, these approximations accurately predicted inference outcomes in finite systems, a result that validates a major assumption underlying much of network science.

What They Found

The first major finding concerns how detectability scales with system size. As networks get larger, the absolute amount of data grows, and with it the signal-to-noise ratio in the likelihood comparisons. But the researchers made a more subtle observation: while raw likelihood ratios diverge as system size increases, the relative detectability—scaled by the largest system considered—stabilizes. This means that a researcher working with 250 nodes can use approximations derived from systems of 100,000 nodes to predict how well their inference will work.

Figure 2 shows this elegantly. Panel (c) displays how relative detectability evolves as system size increases, for 250 random points in the parameter space. All curves approach a plateau, and the variance across different parameter configurations shrinks. Panel (d) validates this approach: when the authors partitioned the parameter space and computed asymptotic detectability, they found that lower asymptotic detectability values predicted lower true positive rates in finite systems, confirming that the theoretical measure captures practical inference quality.

The second finding concerns the role of network density. The researchers systematically varied the average degree of their synthetic networks and found a consistent pattern: as networks become denser, detectability decreases. This finding has profound implications because real-world networks tend to be sparse. Social networks, for example, don't connect everyone to everyone—most people have dozens or hundreds of connections, not thousands. The same is true for most biological and technological networks. "Accuracy generally increases in sparse networks, which are prevalent in real-world systems," the authors note—a finding that reframes the sparse-network case as actually favorable for inference rather than a limitation.

Figure 3 illustrates this density dependence. Panel (a) shows heatmaps of asymptotic detectability across parameter space for epidemic and threshold dynamics on fixed networks. Panel (b) reveals potential confusion patterns: the same Simple infection mechanism could be most closely approximated by different alternative mechanisms depending on the parameter region. Panel (c) shows how increasing network density systematically decreases detectability for most models. The exception—Ising High temperature dynamics—maintains broader detectability across degree distributions.

The third finding is perhaps the most counterintuitive. The researchers found that detectability peaks after the epidemic threshold—that is, in the parameter regime where the infection can sustain itself but has not yet overwhelmed the system. At the threshold itself, where the dynamics hover in a critical state, detectability is actually lower. This is surprising because critical points are often assumed to maximize information about underlying mechanisms. The authors show that while more data is generated near critical points, the differences between mechanisms become harder to distinguish precisely at the threshold.

Panel (d) of Figure 3 makes this clear. The expected likelihood ratio peaks after the critical point κ$_C$, driven by larger amounts of data in that regime, but the normalized detectability λ* is systematically lower before and at κ$_C$—exactly where intuition might suggest inference would work best.

To test their framework on real data, the researchers applied their inference method to seven diverse empirical datasets: cooperation experiments (where individuals decide whether to cooperate based on neighbors' behavior), social media posts (tracking which content spreads), urban traffic patterns, supply chain dynamics, and animal interactions. Each required preprocessing to convert continuous data into binary states, and the researchers explicitly investigated how this preprocessing affected results.

They found that temporal aggregation—the size of the time window used to discretize continuous data—fundamentally altered which mechanism was selected. For the social media dataset they analyzed, very small time windows suggested noisy simple infections. But aggregating to daily windows shifted the inferred mechanism toward what looked more like independent dynamics. For urban traffic data, low traffic volumes behaved like noisy simple or threshold infections, while high traffic was consistently simple. The preprocessing choices that researchers make—whoever is cleaning the data—can alter the scientific conclusions.

Why This Changes Things

The paper's framework represents a significant advance in the toolkit available to researchers studying spreading processes. Historically, choosing a model for disease transmission, opinion dynamics, or information diffusion has relied heavily on researcher intuition and conventions within specific fields. Epidemiologists default to SIS or SEIR models because those are what they've always used. Social scientists gravitate toward threshold models because they match intuitive notions of tipping points. But these choices are made with little rigor around whether the chosen model actually matches the data.

The framework Ureña-Carrión and colleagues provide allows researchers to systematically compare candidate mechanisms using a principled statistical procedure. Rather than assuming a model and checking its fit, researchers can ask: given this dataset, what evidence supports each of several mechanistic hypotheses? This shifts the question from "does our model fit?" to "what does the data actually suggest?"

The finding that sparse networks are favorable for inference is particularly significant. Network scientists have long characterized sparse networks as challenging because many analytical tools assume dense connectivity. This paper reframes sparsity as an advantage: when each node has fewer connections, the "signature" of the underlying mechanism is clearer, because nodes sample less of the network and are more strongly constrained by local structure. A dense network, where everyone is connected to everyone, essentially averages away the local dynamics that reveal mechanism.

The critical-point finding challenges intuitions that have guided research for decades. Criticality—the idea that spreading processes naturally evolve toward states where small perturbations can cascade globally—has been invoked to explain everything from brain dynamics to market crashes to viral media. If detection is actually harder at critical points, this raises questions about how well we can identify critical dynamics in real systems. It also suggests that the information available at critical points is subtly different from what researchers have assumed.

The preprocessing results should give any empirical researcher pause. Data never comes pre-cleaned and binary; converting continuous observations into states suitable for this analysis requires choices that the paper shows can alter scientific conclusions. When does someone become "infected" by a meme—is it when they first see it, when they share it, or when they comment on it? When does a driver join a traffic jam—when they slow down, when they stop, or when they finally merge? The framework can't resolve these ambiguities, but it can quantify how much they matter, giving researchers a clearer picture of how much their preprocessing choices drive their conclusions.

What's Next

The framework as developed has limitations that suggest important research directions. The analysis focused on static networks, but real-world systems often feature dynamic or temporal networks—relationships that form and dissolve over time. The extension to temporal structure would be significant but challenging, as the state space grows dramatically when networks can change.

The authors analyzed one-versus-all comparisons: given one true mechanism, how well can we distinguish it from each alternative? A natural next step would be full Bayesian model comparison with prior probabilities over mechanisms, which would better reflect how researchers actually work when they have prior knowledge or theoretical commitments.

The empirical applications revealed a troubling pattern: in many real-world datasets, model selection fails to converge on any single mechanism—the likelihood ratios don't clearly favor one candidate over others. This could reflect genuine model uncertainty (multiple mechanisms contributing), misspecification (none of the candidate mechanisms actually generated the data), or data limitations (the signal is too weak). Disentangling these possibilities is an important direction for future work.

The finding that temporal aggregation affects inference opens questions about the natural timescales of spreading processes. If model selection changes when you aggregate from hourly to daily windows, which resolution captures the "true" mechanism? The answer likely depends on the system: for rapidly evolving social media dynamics, hourly windows might be more appropriate, while for slower opinion changes, daily or weekly aggregation might better capture the decision-making process.

Perhaps most importantly, the framework suggests that inference quality is highly heterogeneous across parameter space. In some conditions—certain mechanisms, certain network structures, certain observation periods—inference works well. In others, it fails reliably. Rather than treating inference as a black box that either works or doesn't, researchers can now ask where their specific system falls in this landscape and calibrate their confidence accordingly.

This matters for practical applications. If a public health official is trying to understand how a disease spreads to design interventions, they should know whether the available data can support confident mechanistic inference. If it can't—if the conditions are in a regime where multiple mechanisms are nearly indistinguishable—then policy might need to hedge against multiple scenarios rather than betting on a single inferred mechanism.

The paper closes by noting that their results "show that statistical model selection can fail under common conditions and suggest new directions for overcoming these limitations." This is honest science: acknowledging the limits of what we've built rather than overclaiming its capabilities. The framework is powerful, but it's also a map of its own blind spots.

What the researchers have built is a lens for seeing the invisible rules that govern how things spread. In a world increasingly shaped by contagion—whether of disease, information, or behavior—understanding those rules isn't just academically interesting. It's essential for navigating the cascading failures and viral cascades that characterize modern complex systems. This paper shows us both how far we've come in developing that understanding and how far we still have to go.