Epidemics Speak the Same Mathematical Language — Even When They Look Different

At the moment an outbreak tips from possibility to inevitability, its fate is already sealed — not by the virus, not by the network, but by a single number: the slope of a curve at zero prevalence. That slope, according to a new analysis, determines whether an epidemic will rise gradually or explode suddenly into a full-blown crisis. And astonishingly, that same mathematical rule governs outbreaks as different as a biological pandemic spreading through physical contact, a meme going viral on social media, or two diseases reinforcing each other in a population. The finding, from a trio of network scientists, reveals that beneath wildly different mechanisms — from adaptive social distancing to higher-order group interactions — lies a universal skeleton of contagion. What looks like complexity collapses into simplicity when you realize that some processes move much faster than others.

The Science

Laurent Hébert-Dufresne (University of Vermont, Santa Fe Institute, Complexity Science Hub), Péter L. Simon (Eötvös Loránd University), and István Z. Kiss (Northeastern University London) set out to solve a long-standing puzzle in mathematical epidemiology: why do so many different models of disease spread — each with its own assumptions about human behavior, network structure, or biological interaction — produce such similar outcomes? Some models include detailed contact networks; others allow people to cut risky ties when disease spreads; still others simulate how diseases interact when two are circulating at once. Each adds layers of realism, but also layers of mathematical complexity, often resulting in systems of equations too unwieldy to solve.

Their insight, detailed in a 2026 arXiv preprint, is that in all these models, certain variables — like the density of risky connections or the proportion of infected neighbors — adjust rapidly compared to the slow creep of overall infection prevalence. These are the "fast variables." By assuming these fast variables equilibrate instantly at any given level of prevalence, the researchers show that a vast array of mechanistic models can be reduced to a single differential equation:

$$\dot{i}=i\left[{\beta }_{\mathrm{eff}}(i)-\gamma \right]$$

Here, $i$ is the fraction of the population infected, $\gamma$ is the recovery rate, and $\beta_{\rm eff}(i)$ is an effective force of infection that depends on $i$ itself. This form is not assumed; it emerges naturally from the underlying dynamics. The function $\beta_{\rm eff}(i)$ encodes all the complexity — network structure, behavioral adaptation, co-infection effects — but only through its dependence on prevalence. The result is a unification: models that appear distinct at the microscopic level collapse onto the same macroscopic skeleton.

The authors tested their framework on five canonical models:

• The classical pairwise SIS model on regular networks
• A higher-order SIS model with simplicial (triangle-based) interactions
• An adaptive network model where individuals rewire away from infected contacts
• A model of two interacting SIS contagions (e.g., diseases or ideas)
• The pairwise SIR model

Each was reduced using the same four-step recipe: identify the slow variable (prevalence), isolate the fast variables (e.g., $\langle SI \rangle / \langle I \rangle$), impose quasi-static equilibrium ($\dot{\mathbf{p}} = 0$), and substitute into the prevalence equation. The method works whether the fast-variable system is one-dimensional or two-dimensional, solvable analytically or not. Even when closed-form solutions are unavailable, the effective force of infection $\beta_{\rm eff}(i)$ can be measured numerically from simulations, making the approach broadly applicable.

What They Found

The most striking result is that the epidemic threshold — the point at which an outbreak becomes self-sustaining — is universally determined by the condition:

$${\beta }_{\mathrm{eff}}(0)=\gamma$$

This means that regardless of whether people are avoiding infection, diseases are interacting, or transmission happens in groups, the critical transmission rate $\tau_c$ is set by the effective infection pressure at the very beginning of an outbreak. This single condition replaces a zoo of ad hoc threshold calculations.

Even more powerful is the prediction of how the transition happens. The nature of the epidemic onset — continuous (gradual) or discontinuous (sudden, with hysteresis and bistability) — depends on the derivative of $\beta_{\rm eff}(i)$ at zero:

• If $\partial_i \beta_{\rm eff}|_{i=0} < 0$, the transition is continuous: infections rise smoothly from zero as transmission exceeds the threshold.
• If $\partial_i \beta_{\rm eff}|_{i=0} > 0$, the transition is discontinuous: a tiny increase in transmission can trigger a large-scale outbreak, and reducing transmission back below threshold may not stop it.

This criterion, derived in Appendix 6.1, transforms how we classify epidemic models. It’s not the presence of network structure or behavioral feedback that matters — it’s the sign of this slope.

The authors illustrate this collapse in

, which plots $\beta_{\rm eff}(i)/\gamma$ against prevalence $i$ for four different models, all evaluated at their respective epidemic thresholds. Every curve passes through the point $(0,1)$, reflecting the universal threshold condition. But their slopes differ:

• The classical pairwise SIS model (solid blue) has a negative slope, predicting a continuous transition.
• The simplicial model (dashed orange), when the higher-order transmission rate $\beta$ is strong enough, has a positive slope, leading to a discontinuous jump.
• The adaptive network model (dashed green) can switch from negative to positive slope depending on the rewiring rate $w$.
• The interacting contagions model (dashed red) behaves similarly, with the transition type controlled by the synergy parameter $\alpha$.

Epidemic thresholds and transition types across models

Label	Value
Classical pairwise SIS	1.2 γ/(n−1)
Higher-order (simplicial)	1.1 γ/(n−1)
Adaptive network	1.4 γ/(n−1)
Interacting contagions	1 γ/(n−1)
Pairwise SIR	1.3 γ/(n−1)

Thresholds vary by model, but all satisfy $\beta_{\rm eff}(0) = \gamma$. Transition type depends on slope of $\beta_{\rm eff}(i)$ at i=0.

summarizes the thresholds and transition types across models. Notably, the classical SIS and SIR models are always continuous, while the others can flip between continuous and discontinuous depending on a single parameter.

The numerical reduction method works remarkably well.

shows the fast variable $p = \langle SI \rangle / \langle I \rangle$ plotted against prevalence $i$ from simulations of the full model. A simple linear fit (degree $K=1$) captures the quasi-static relationship, and when used to construct $\hat{\beta}_{\rm eff}(i)$, it reproduces the full prevalence trajectory almost exactly (see inset). This means that even for models too complex to solve analytically, we can extract their essential dynamics from simulation data alone.

Why This Changes Things

For decades, epidemiologists have treated different contagion mechanisms as fundamentally distinct. Network models required different tools than behavioral models, which in turn differed from models of interacting pathogens. This paper shows that many of these differences are illusory — at least near the epidemic threshold, they are macroscopically indistinguishable if they produce the same $\beta_{\rm eff}(i)$.

This has profound implications. First, it means that observing the early growth of an outbreak — say, the first 0.1% of cases — could in principle reveal not just whether an epidemic will take off, but how it will take off. A concave-down effective transmission rate suggests a manageable, gradual increase; a concave-up rate warns of a potential explosion.

Second, it reframes public health interventions. Instead of asking "How do we model social distancing?" we can ask "How does social distancing change the slope of $\beta_{\rm eff}(i)$ at zero?" The answer determines whether a policy merely raises the threshold or fundamentally alters the nature of the transition. For instance, adaptive rewiring in networks (people avoiding infected contacts) can, beyond a critical rate, turn a gradual epidemic into a sudden one — a counterintuitive result that aligns with earlier findings (Gross et al., 2006) but is now placed within a universal framework.

Third, the work bridges biological and social contagions. The same mathematics governs both a virus spreading through physical contact and a belief spreading through social reinforcement. This is not just metaphorical; the interacting contagions model directly applies to phenomena like anti-vaccine sentiment and disease spread, where one "infection" lowers resistance to the other. The finding that synergy $\alpha > 2$ leads to discontinuous transitions suggests that when misinformation and disease feed each other, outbreaks can be abrupt and hard to reverse — a warning for pandemic preparedness.

The collapse onto a one-dimensional skeleton also has practical value. Instead of simulating millions of agents with complex rules, policymakers could use simple, interpretable models calibrated to real-world data. The numerical method in Section 4 shows how to do this: run a detailed simulation, measure how fast variables respond to prevalence, and build a reduced model that captures the essence. This could accelerate scenario planning for emerging threats.

What's Next

The framework has limits. It assumes a clear separation of timescales — that fast variables equilibrate much quicker than prevalence changes. In real populations, this may not always hold, especially during rapid behavioral shifts or when multiple pathogens evolve at different rates. The authors note that the SIR model resists full reduction because both susceptibility and infection are slow variables, requiring a two-dimensional approximation. This suggests the method works best for endemic or recurring outbreaks, not one-time events.

Another open question is how to extend the approach to heterogeneous populations. The models analyzed assume regular networks or mean-field approximations. Real-world contact patterns are far more varied, with superspreaders and tightly-knit communities. While the authors suggest the method generalizes, explicit treatment of degree heterogeneity or clustering would strengthen the case.

Perhaps the most exciting direction is empirical validation. Can we measure $\beta_{\rm eff}(i)$ from real outbreak data? For digital contagions — memes, misinformation, cryptocurrency adoption — the answer may already be yes, given high-resolution interaction logs. For biological diseases, it would require fine-grained surveillance at very low prevalence, a challenge but not an impossibility with wastewater monitoring or pooled testing.

The authors also hint at deeper theoretical implications. If diverse mechanisms collapse onto the same skeleton, then multiple models may be equifinal — different paths to the same outcome. This challenges the idea that mechanistic detail is always necessary for prediction. In some cases, the "why" may matter less than the "how much."

Ultimately, this work offers a new lens: not just for modeling epidemics, but for understanding how complexity gives rise to simplicity in nature. The next pandemic may look unlike any we’ve seen — but if it spreads through interactions that adjust quickly to risk, its mathematical heart will be familiar.