Epidemics on Networks

Every epidemic starts with a contact: one infected person stands near one susceptible person and the pathogen leaps across. We have been building mathematical models of that leap since Daniel Bernoulli quantified smallpox inoculation in 1760 — but the most influential framework, the SIR model, only crystallised in 1927 when Kermack and McKendrick published their differential equations.

SIR divides a closed population into three compartments. Susceptible (S) individuals have not yet been infected. Infected (I) individuals carry the pathogen and can transmit it. Recovered (R) individuals are immune and play no further role. Two rates govern the flow: $\beta$ (the transmission rate) moves people from S to I, and $\gamma$ (the recovery rate) moves them from I to R.

The textbook version assumes homogeneous mixing — every person contacts every other with equal probability, like molecules in a well-stirred flask. That simplification yields a tidy $R_0 = \beta / \gamma$ (the basic reproduction number): when $R_0 > 1$ the epidemic grows; when $R_0 < 1$ it dies. Clean, predictive, and often wrong.

Real populations are not well-stirred. They are networks: airports connect cities, households cluster families, schools concentrate children. On a contact graph the same $\beta$ and $\gamma$ can produce radically different outcomes depending entirely on who is connected to whom. Understanding that dependence is the heart of network epidemiology.

The network below has 20 nodes connected by contact edges. One node starts infected (red). At each step every infected node tries to infect each susceptible neighbour with probability $\beta$; then infected nodes recover with probability $\gamma$.

<!-- {{c_html_intro}} -->
<div class="controls">
  <label>{{lbl_beta}} <input id="betaSlider" type="range" min="0.05" max="0.95" step="0.05" value="0.35">
    <span id="betaVal">0.35</span></label>
  <label>{{lbl_gamma}} <input id="gammaSlider" type="range" min="0.05" max="0.95" step="0.05" value="0.20">
    <span id="gammaVal">0.20</span></label>
</div>
<div class="canvas-wrap">
  <canvas id="net" width="420" height="280"></canvas>
</div>
<div class="stats" id="stats"></div>
<div class="btns">
  <button id="btnStep" type="button">{{btn_step}}</button>
  <button id="btnRun" type="button">{{btn_run}}</button>
  <button id="btnReset" type="button" class="ghost">{{btn_reset}}</button>
</div>
<div class="status" id="status"></div>

/* {{c_css_intro}} */
* { box-sizing: border-box; }
body { font-family: system-ui, sans-serif; color: #222; margin: 0; padding: .5rem; }
.controls { display: flex; flex-wrap: wrap; gap: .6rem 1.2rem; margin-bottom: .5rem; font-size: .88rem; }
.controls label { display: flex; align-items: center; gap: .35rem; }
.controls input[type=range] { width: 110px; }
.canvas-wrap { border: 1px solid #cdd9e3; border-radius: 8px; overflow: hidden; background: #f7f9fb; }
canvas { display: block; }
.stats { font-size: .82rem; color: #555; margin: .35rem 0; min-height: 1.2em; }
.btns { display: flex; gap: .5rem; flex-wrap: wrap; margin-bottom: .35rem; }
button { font: 600 13px system-ui; padding: .4rem .85rem; border: 1px solid #1d3557;
         background: #1d3557; color: #fff; border-radius: 8px; cursor: pointer; }
button.ghost { background: #fff; color: #1d3557; }
.status { font-size: .95rem; font-weight: 600; min-height: 1.4em; }
.status.ok { color: #0a7d33; }
.status.info { color: #1d3557; }

// Code not found

Watch how the epidemic percolates — or fails to. The same parameters can produce a small local outbreak when the seed is isolated, or a near-total sweep when the seed is a hub. The network structure is doing the work, not the averages.

Once you move from a well-mixed population to a contact graph, two facts change almost everything.

Degree heterogeneity. In a random Erdős–Rényi graph every node has roughly the same number of contacts. Real networks — airline routes, sexual contact networks, the web — follow a heavy-tailed degree distribution: most nodes have few links, but a small number of hubs have vastly more. On such scale-free graphs the effective reproduction number is no longer just $\beta / \gamma$. A 1999 result by Pastor-Satorras and Vespignani showed that scale-free networks have no epidemic threshold: no matter how small $\beta$ is, an epidemic can persist indefinitely because hubs sustain transmission even when the rest of the network would extinguish it.

The heterogeneous mean-field correction. For a network with degree distribution $p_k$, the epidemic threshold becomes:

$$R_0^{\text{net}} = \frac{\beta}{\gamma} \cdot \frac{\langle k^2 \rangle}{\langle k \rangle}$$

where $\langle k \rangle$ is the mean degree and $\langle k^2 \rangle$ is the mean squared degree. When the degree distribution has a heavy tail, $\langle k^2 \rangle / \langle k \rangle$ grows without bound as the network grows — pushing the effective $R_0$ far above its well-mixed estimate.

Herd immunity. Vaccinating a random fraction $1 - 1/R_0$ of the population stops a well-mixed epidemic. On a heterogeneous network you can do far better by targeting hubs first — but the threshold itself shifts. For a scale-free network the fraction required for herd immunity depends on the second moment of the degree distribution, not just the mean. In practice this means that hub-targeted vaccination is disproportionately powerful, and that disease can persist in a network even when the majority of nodes are immune.

Clustering. Dense local triangles (think household contacts) slow the initial exponential growth because infected and susceptible nodes share many of the same neighbours, reducing the pool of new targets. Long-range ties — the "weak links" studied by network science — are what allow an epidemic to bridge clusters and go global.

The interplay of these factors means that predicting epidemic outcomes requires actual contact data, not just population averages. The famous failure to anticipate the geographic pattern of SARS in 2003 was partly a failure to account for the air-travel network — a lesson that shaped pandemic planning ever since.

The network-SIR framework is not confined to biology. Anything that spreads through contact — viruses, information, computer malware, social norms — follows the same mathematics:

• Pandemic preparedness: COVID-19 modelling combined individual contact networks (household, workplace, school) with airline mobility data to forecast geographic spread and evaluate non-pharmaceutical interventions. The topology of the contact graph determined which interventions worked.
• Targeted vaccination: if you can identify and vaccinate the highest-degree nodes first, you collapse the effective $R_0$ far more efficiently than random vaccination. This insight drives ring-vaccination strategies for smallpox and Ebola.
• Social media and misinformation: rumours and viral content follow a susceptible–infected–recovered-like cascade on follower graphs. The same hub-driven amplification that makes influenza hard to contain makes misinformation hard to debunk.
• Computer network security: worms and ransomware propagate across IP networks whose topology is well-studied. Quarantine strategies (analogous to $\gamma$) and patch deployment (analogous to vaccination) can be tuned using the same threshold analysis. See P vs NP for why finding the optimal set of nodes to immunize is itself a computationally hard problem.
• Ecology: the spread of invasive species and plant pathogens through habitat networks obeys the same percolation dynamics. Conservation planning now routinely uses network-SIR to identify critical habitat corridors to protect.

In every case the lesson is the same: ignore the contact structure and you will misforecast both the peak and the tail of the spreading process.

The SIR model is elegant in its simplicity: three compartments, two rates, one number $R_0$ that governs whether an epidemic grows or dies. But place that model on a real contact network and the elegance gives way to something richer — and harder.

Hubs with disproportionate connectivity become superspreaders. Tight clusters slow initial spread but act as reservoirs. Long-range ties bridge communities and turn local outbreaks into pandemics. The effective reproduction number is no longer $\beta / \gamma$ but depends on the full degree distribution of the contact graph. And the fraction of the population you need to vaccinate — or to remove — to stop the epidemic is not a universal constant but a property of the network.

This shift from averages to topology is one of the deepest ideas in mathematical epidemiology, and it connects directly to the broader theme of P vs NP: optimally targeting interventions on a network is computationally hard, even when the right target is obvious in hindsight. The epidemic does not wait for a tractable solution — but understanding the network structure at least tells you where to look first.