Criticality

Critical Mass, Length, Temperature

Many systems and processes, whether in nature or society, can appear unchanging until they reach a certain transition point, after which their behavior shifts dramatically. Think of disease outbreaks, the spread of information in a network, or battery failures in electric vehicles. Let’s go through some examples.

Every symbol in the formulas is underlined with dots. Hover over it, or tap it on a phone, to see what it stands for.

Disease Outbreaks

Consider a disease outbreak such as COVID-19 in 2020. The goal of many governments was not just to reduce infections slightly, but to bring the average number of new infections per case, known as the reproduction number, or \(\tip{R_0}{reproduction number: people one infected person infects on average}\) , below the critical threshold of 1. When \(\tip{R_0}{reproduction number: people one infected person infects on average} < 1\) , each infected person, on average, transmits the disease to fewer than one other person, and the outbreak dies out. However, if \(\tip{R_0}{reproduction number: people one infected person infects on average} > 1\) even by a small margin, the number of infections can grow exponentially.

Social Networks

On platforms like Facebook, the algorithm often recommends “friends of friends.” This means that when a piece of information is shared, its likelihood of reaching someone outside your immediate circle can increase sharply once the network crosses a connectivity threshold (the formation of a giant component). This threshold behavior is key to understanding how information, trends, or even misinformation can spread rapidly through a network.

Battery Explosions

In electric vehicles, safety depends on thermal stability. A critical point is reached when damage in a cell propagates faster than the battery can dissipate the excess energy. This triggers a runaway process, where one failing cell causes its neighbors to fail as well, leading to a sudden and dramatic failure.

---

These examples demonstrate how systems can exhibit qualitatively different behavior depending on one or more parameters. This behavior is known as criticality. Although these examples might seem similar, they can be classified into two broad categories of critical behavior. After some necessary terminology, we will examine the differences between them and explore interactive simulations.

Terminology

To compare these examples, we only need two words:

• Order Parameter (\(\tip{X}{order parameter: the quantity that tells us the state of the system}\) ): the quantity that tells us in which state the system is. For the outbreak this is the number of infected people, for a magnet its magnetization, for a network the size of its largest connected group.
• Control Parameter (\(\tip{p}{control parameter: the knob we turn}\) ): the knob we turn. For the outbreak this is \(\tip{R_0}{reproduction number: people one infected person infects on average}\) , for the magnet the temperature, for the network how many connections there are.

Criticality means that there is one specific setting of the knob, the critical value \(\tip{p_c}{critical value of the control parameter}\) , where the state of the system changes qualitatively. Below it the outbreak dies out, above it the outbreak spreads. We are not interested in what happens in the next few minutes, but in where the system ends up after a long time.

The formal definition

We look at the long-time or equilibrium behavior of the order parameter:

\[\tip{X_{\infty}}{long-term state of the system}(\tip{p}{control parameter: the knob we turn}) := \lim_{\tip{t}{time} \to \infty} \tip{X}{order parameter: the quantity that tells us the state of the system}(\tip{t}{time}; \tip{p}{control parameter: the knob we turn})\]

In statistical models without explicit time, we interpret \(\tip{X_\infty}{long-term state of the system}\) as the thermodynamic limit (as system size \(\tip{N}{system size: number of parts of the system} \to \infty\) ). A system exhibits criticality if there exists a value \(\tip{p_c}{critical value of the control parameter}\) at which \(\tip{X_\infty}{long-term state of the system}(\tip{p}{control parameter: the knob we turn})\) changes qualitatively.

Two Classes of Systems Exhibiting Criticality

The key question is: what happens to the end state if we turn the knob slowly across \(\tip{p_c}{critical value of the control parameter}\) ? Either it jumps, or it changes smoothly. This gives two classes of critical behavior.

Tipping Point Criticality (Discontinuous)

Think of a chair that you tilt backwards. For a long time nothing happens. Then, at one specific angle, it falls. A tiny change of the control parameter causes a sudden jump of the end state.

The simplest example is exponential growth, \(\tip{N}{population size}(\tip{t}{time, counted in generations}) = \tip{N_0}{population at the start} \tip{k}{growth factor per generation}^\tip{t}{time, counted in generations}\) , where every generation is \(\tip{k}{growth factor per generation}\) times as large as the one before. As long as \(\tip{k}{growth factor per generation} < 1\) , every generation is smaller than the previous one and the population dies out. As soon as \(\tip{k}{growth factor per generation} > 1\) , it grows without limit. At \(\tip{k}{growth factor per generation} = 1\) the long-term population jumps from zero to infinity.

Often the way back is not the same as the way there. Once the chair lies on the floor, tilting it back by the same small angle does not bring it up again. This is called hysteresis: the path taken to “tip” the system is different from the path required to return it to its original state.

The formal version

At a tipping point the order parameter is discontinuous at \(\tip{p_c}{critical value of the control parameter}\) :

\[\lim_{\tip{p}{control parameter: the knob we turn} \to \tip{p_c}{critical value of the control parameter}^+} \tip{X_{\infty}}{long-term state of the system}(\tip{p}{control parameter: the knob we turn}) \neq \lim_{\tip{p}{control parameter: the knob we turn} \to \tip{p_c}{critical value of the control parameter}^-} \tip{X_{\infty}}{long-term state of the system}(\tip{p}{control parameter: the knob we turn})\]

In the language of dynamical systems, the fixed point at \(\tip{N}{population size}=0\) of the growth \(\tip{N}{population size}(\tip{t}{time, counted in generations}) = \tip{N_0}{population at the start} \tip{k}{growth factor per generation}^\tip{t}{time, counted in generations}\) is stable for \(\tip{k}{growth factor per generation} < 1\) and becomes unstable for \(\tip{k}{growth factor per generation} > 1\) . The continuous-time analogue \(\tip{\dot{N}}{rate of change of the population over time} = \tip{r}{growth rate per unit of time}\tip{N}{population size}\) behaves the same way: the fixed point at \(\tip{N}{population size}=0\) loses stability as \(\tip{r}{growth rate per unit of time}\) crosses 0.

Continuous Critical Transition

Now imagine a group of people, where we add random acquaintances one by one. At first only small groups know each other. Beyond a certain density, a giant group appears that connects a large part of everyone. But at the critical point this giant group is still tiny, and it grows smoothly as more connections are added. Nothing jumps, but something new is born at \(\tip{p_c}{critical value of the control parameter}\) .

The classic example is percolation: as connectivity increases, a “spanning cluster” appears. Its size starts at zero and grows continuously as the network becomes denser. In statistical physics, these transitions are known as second-order phase transitions.

The formal version

In a continuous critical transition, the order parameter is continuous at \(\tip{p_c}{critical value of the control parameter}\) :

\[\lim_{\tip{p}{control parameter: the knob we turn} \to \tip{p_c}{critical value of the control parameter}^+} \tip{X_\infty}{long-term state of the system}(\tip{p}{control parameter: the knob we turn}) = \lim_{\tip{p}{control parameter: the knob we turn} \to \tip{p_c}{critical value of the control parameter}^-} \tip{X_\infty}{long-term state of the system}(\tip{p}{control parameter: the knob we turn})\]

What changes at \(\tip{p_c}{critical value of the control parameter}\) is not the value but its behavior: below \(\tip{p_c}{critical value of the control parameter}\) it is identically zero, above it it grows, typically like a power of \(\tip{p}{control parameter: the knob we turn} - \tip{p_c}{critical value of the control parameter}\) .

Criticality and Power Laws

Power laws (\(\tip{y}{a quantity that follows a power law} \sim \tip{x}{size of an event}^\tip{\alpha}{exponent: how fast the probability falls with size}\) ) show up in two very different roles: as the mechanism that creates a threshold, and as the signature a system shows when it sits exactly at the threshold.

1. Competing Power Laws (The Mechanism)

Why does a big pot of soup stay hot so much longer than a small cup? The heat stored in a body grows with its volume (\(\sim \tip{x}{size of an event}^3\) ), but it can only escape through its surface (\(\sim \tip{x}{size of an event}^2\) ). Double the size, and the volume grows eightfold while the surface only grows fourfold. Whatever grows with the higher power wins in the end. So if a system depends on both surface and volume, varying its size causes a crossover: beyond a certain scale, the volume term dominates and can drive the system past a threshold.

2. Power-Law Distributions & Scale Invariance (The Signature)

Most things in daily life have a typical size. Body heights cluster around an average, and nobody is ten metres tall. Near a continuous transition this is no longer true. Event sizes span all scales: very many tiny ones, a few huge ones, and no typical size in between. Mathematically, this is a power-law distribution, where the probability of an event of size \(\tip{x}{size of an event}\) falls off like

\[\tip{f}{probability density: how likely an event of size x is}(\tip{x}{size of an event}) \propto \tip{x}{size of an event}^{-\tip{\alpha}{exponent: how fast the probability falls with size}}.\]

Conditions on the exponent

For a random variable \(\tip{X}{random size of an event}\) with a power-law tail, the probability density function is

\[\tip{f}{probability density: how likely an event of size x is}(\tip{x}{size of an event}) \propto \tip{x}{size of an event}^{-\tip{\alpha}{exponent: how fast the probability falls with size}}, \quad \tip{x}{size of an event} \geq \tip{x_{\min}}{smallest size from which the power law holds}.\]

Normalization requires \(\tip{\alpha}{exponent: how fast the probability falls with size} > 1\) , and a finite mean requires \(\tip{\alpha}{exponent: how fast the probability falls with size} > 2\) . For \(1 < \tip{\alpha}{exponent: how fast the probability falls with size} \leq 2\) the average is dominated by the rare huge events and does not settle down, no matter how many samples you take.

Additionally, the system becomes scale-invariant: it looks statistically similar regardless of the magnification level, like a coastline that looks equally ragged from a plane and from a cliff. The distance over which one part of the system “feels” another, the correlation length, grows without limit as we approach the critical point.

The correlation length

The correlation length \(\tip{\xi}{correlation length: how far one part of the system feels another}\) diverges as we approach criticality, again following a power law:

\[\tip{\xi}{correlation length: how far one part of the system feels another} \sim |\tip{p}{control parameter: the knob we turn} - \tip{p_c}{critical value of the control parameter}|^{-\tip{\nu}{critical exponent of the correlation length}}.\]

The exponent \(\tip{\nu}{critical exponent of the correlation length}\) is one of the critical exponents. Remarkably, it depends only on broad features of the system, such as its dimension, and not on the microscopic details. This is called universality.

Real World Examples

Thermal Runaway & Nuclear Chain Reaction

Both thermal runaway and nuclear chain reactions can be modeled by an exponential growth function where the exponent is determined by a balance of production and loss:

\[\tip{f}{net growth: production minus loss}(\tip{s}{size of the system}) \propto \text{Production} - \text{Loss} \sim \tip{a_0}{production per unit of volume} \tip{s}{size of the system}^3 - \tip{a_1}{loss per unit of surface} \tip{s}{size of the system}^2\]

The production of heat or neutrons scales with Volume (\(\tip{s}{size of the system}^3\) ), while the loss (dissipation or leakage) scales with Surface Area (\(\tip{s}{size of the system}^2\) ). The system becomes critical when the ratio of Volume to Surface Area reaches a specific threshold.

In reactor physics, the control parameter is the effective multiplication factor \(\tip{k_{\text{eff}}}{effective multiplication factor: new neutrons caused by one neutron}\) , with criticality at \(\tip{k_{\text{eff}}}{effective multiplication factor: new neutrons caused by one neutron} = 1\) . Geometry (via neutron leakage) pulls \(\tip{k_{\text{eff}}}{effective multiplication factor: new neutrons caused by one neutron}\) down, while material properties and moderation push it up. The \(\tip{V}{volume}/\tip{A}{surface area}\) scaling argument captures this intuition directly.

The Geometry of Criticality: Cube vs. Sphere

The critical size depends on the volume-to-surface ratio (\(\tip{V}{volume}/\tip{A}{surface area}\) ):

• Sphere: \(\tip{V}{volume}/\tip{A}{surface area} = \tip{r}{radius of the sphere}/3\)
• Cube: \(\tip{V}{volume}/\tip{A}{surface area} = \tip{L}{side length of the cube}/6\)

Setting these equal (\(\tip{r}{radius of the sphere}/3 = \tip{L}{side length of the cube}/6\) ) reveals that \(\tip{L_c}{critical side length of the cube} = 2\tip{r_c}{critical radius of the sphere}\) . This means a cube-shaped battery must have a side length exactly twice a sphere’s radius to reach the same level of thermal stability.

Griffith Crack Length

Imagine a sail under tension or an air-filled balloon. Small imperfections or micro-cracks may appear in the material. Under stress, these tiny cracks can suddenly start to grow, potentially leading to catastrophic failure. This behavior was first analyzed by A.A. Griffith; the following logic follows the intuitive framework in J.E. Gordon’s Structures.

In physics, systems naturally evolve toward lower energy states. When a crack forms, two competing energy terms are at play:

• Surface Energy Cost (\(\sim \tip{L}{length of the crack}\) ): Creating new surfaces (tearing the material) requires energy. This increases the total energy of the system.
• Strain Energy Release (\(\sim \tip{L}{length of the crack}^2\) ): As the crack opens, the material “relaxes,” releasing stored elastic energy. This lowers the total energy of the system.

The net change in energy for a crack of length \(\tip{L}{length of the crack}\) can be modeled as:

\[\tip{\Delta E}{change in total energy caused by the crack}(\tip{L}{length of the crack}) = \tip{\beta}{energy cost of new crack surface per unit of length} \tip{L}{length of the crack} - \frac{\tip{\alpha}{how strongly stored elastic energy is released as the crack opens}}{2} \tip{L}{length of the crack}^2\]

A crack grows spontaneously if doing so reduces the total energy, which happens once the energy barrier peaks. That gives the critical crack length

\[\tip{L_c}{critical crack length: beyond it the crack grows by itself} = \frac{\tip{\beta}{energy cost of new crack surface per unit of length}}{\tip{\alpha}{how strongly stored elastic energy is released as the crack opens}}\]

• For \(\tip{L}{length of the crack} < \tip{L_c}{critical crack length: beyond it the crack grows by itself}\) : The energy cost of creating a new surface is higher than the energy released. The crack is stable.
• For \(\tip{L}{length of the crack} > \tip{L_c}{critical crack length: beyond it the crack grows by itself}\) : The strain energy release dominates. Every incremental increase in length releases more energy than it consumes, leading to a runaway effect.

Where the critical length comes from

The crack grows when the derivative \(\frac{d\tip{\Delta E}{change in total energy caused by the crack}}{d\tip{L}{length of the crack}}\) turns negative, so the threshold is the stationary point of \(\tip{\Delta E}{change in total energy caused by the crack}\) :

\[\frac{d\tip{\Delta E}{change in total energy caused by the crack}}{d\tip{L}{length of the crack}} = \tip{\beta}{energy cost of new crack surface per unit of length} - \tip{\alpha}{how strongly stored elastic energy is released as the crack opens} \tip{L}{length of the crack} = 0 \implies \tip{L_c}{critical crack length: beyond it the crack grows by itself} = \frac{\tip{\beta}{energy cost of new crack surface per unit of length}}{\tip{\alpha}{how strongly stored elastic energy is released as the crack opens}}\]

In a thin plate under tensile stress \(\tip{\sigma}{tensile stress pulling on the plate}\) , the classical Griffith criterion gives \(\tip{\sigma_c}{critical stress at which the crack starts to grow} \approx \sqrt{2\tip{E}{Young's modulus: stiffness of the material}\tip{\gamma}{surface energy per unit of area} / (\pi \tip{a}{half of the crack length})}\) , where \(\tip{E}{Young's modulus: stiffness of the material}\) is Young’s modulus, \(\tip{\gamma}{surface energy per unit of area}\) is the surface energy per unit area, and \(\tip{a}{half of the crack length}\) is the half-crack length. This links the toy-model parameters \(\tip{\alpha}{how strongly stored elastic energy is released as the crack opens}, \tip{\beta}{energy cost of new crack surface per unit of length}\) to measurable material constants.

Once a crack is longer than \(\tip{L_c}{critical crack length: beyond it the crack grows by itself}\) , it feeds itself: the longer it gets, the more energy it releases, and the faster it grows. This explains why structural failures often appear so instantaneous and violent.

How fast does the crack grow?

To capture the growth dynamics, we can use a toy differential equation:

\[\frac{d\tip{L}{length of the crack}}{d\tip{t}{time}} = \tip{\lambda}{constant that sets the time scale of the growth} \left( \tip{\alpha}{how strongly stored elastic energy is released as the crack opens} \tip{L}{length of the crack}^2 - \tip{\beta}{energy cost of new crack surface per unit of length} \tip{L}{length of the crack} \right) = \tip{\lambda}{constant that sets the time scale of the growth} \tip{L}{length of the crack} (\tip{\alpha}{how strongly stored elastic energy is released as the crack opens} \tip{L}{length of the crack} - \tip{\beta}{energy cost of new crack surface per unit of length})\]

where \(\tip{\lambda}{constant that sets the time scale of the growth}\) is a constant setting the time scale. Because the driving force (\(\tip{\alpha}{how strongly stored elastic energy is released as the crack opens} \tip{L}{length of the crack}^2\) ) is proportional to the square of the state variable, the model predicts faster-than-exponential growth once the threshold is crossed: the crack length reaches infinity in finite time (finite-time blow-up).

Percolation

Percolation refers to the study of connectivity in random media, modeled as water “seeping through” a porous rock or a message “hopping” through a network. Unlike the SIR or Ising models, which evolve over time toward an equilibrium, percolation is a static probabilistic model. We aren’t watching the system change; we are “sampling” a snapshot of a random world to see if it is connected.

The simplest version is a grid of squares. Each square is “open” with probability \(\tip{p}{probability that a square (or connection) is open}\) , and water can only flow between neighbouring open squares. The question is: can water get from the top of the grid to the bottom?

• For small \(\tip{p}{probability that a square (or connection) is open}\) : The open squares form small, isolated “islands.” Water does not get far.
• For large \(\tip{p}{probability that a square (or connection) is open}\) : One giant connected cluster spans the whole grid, and water flows through easily.

In a small grid, a path might appear by luck even at low \(\tip{p}{probability that a square (or connection) is open}\) . But the larger the grid, the sharper the switch becomes, until it is a clean “yes or no” at a critical value. For the square grid it is \(\tip{p_c}{critical probability: where a spanning path appears} \approx 0.5927\) . Below it, practically never a path; above it, practically always.

Variants, the giant component and the spanning probability

We typically study percolation using a lattice (a grid) or a graph:

• Site Percolation: Each square on a grid is “occupied” (open) with probability \(\tip{p}{probability that a square (or connection) is open}\) . Water can only flow between adjacent occupied squares.
• Bond Percolation: All sites are open, but the connections between them only exist with probability \(\tip{p}{probability that a square (or connection) is open}\) .
• Erdős–Rényi Graph: This is the mean-field version of percolation. It ignores spatial geometry entirely; the giant component emerges when the mean degree \(\tip{c}{mean degree: average number of connections per node} = \tip{p}{probability that a square (or connection) is open}\tip{n}{number of nodes in the graph}\) crosses 1.

The transition is sharp only in the thermodynamic limit (\(\tip{N}{number of sites in the system} \to \infty\) ). We can analyze it through two related lenses:

1. The Giant Component (The Theory): As \(\tip{p}{probability that a square (or connection) is open}\) increases, small clusters of nodes begin to merge. At \(\tip{p_c}{critical probability: where a spanning path appears}\) , a Giant Component emerges: a single connected cluster that contains a finite fraction of all nodes in the infinite system. The theoretical order parameter is the fraction of sites in this infinite cluster, \(\tip{P_\infty}{fraction of sites in the infinite cluster}\) , which grows continuously from 0 for \(\tip{p}{probability that a square (or connection) is open} > \tip{p_c}{critical probability: where a spanning path appears}\) .

2. The Spanning Path (The Observation): In our simulation, we observe the spanning probability \(\tip{\Pi}{spanning probability: chance that a path crosses the box}(\tip{p}{probability that a square (or connection) is open}; \tip{L}{side length of the grid})\) in a finite \(\tip{L}{side length of the grid} \times \tip{L}{side length of the grid}\) box. As \(\tip{L}{side length of the grid} \to \infty\) , \(\tip{\Pi}{spanning probability: chance that a path crosses the box}(\tip{p}{probability that a square (or connection) is open}; \tip{L}{side length of the grid})\) sharpens into a step function centered at \(\tip{p_c}{critical probability: where a spanning path appears}\) .

A spanning path exists because the Giant Component has finally grown large enough to “bridge” the system boundaries.

Criticality and Fractals

Exactly at \(\tip{p}{probability that a square (or connection) is open} = \tip{p_c}{critical probability: where a spanning path appears}\) , the system is in a state of “perfect” complexity. The clusters are neither isolated dots nor a solid block; they are fractals. If you were to zoom in on a cluster at the critical point, it would look statistically identical to the whole. This is the scale invariance that defines criticality.

At this threshold, most clusters are tiny, a few are huge, and there is no typical cluster size: the power-law signature from above. While the specific \(\tip{p_c}{critical probability: where a spanning path appears}\) varies between models, this signature remains the universal hallmark of the transition.

The cluster size distribution

At \(\tip{p}{probability that a square (or connection) is open} = \tip{p_c}{critical probability: where a spanning path appears}\) the distribution of cluster sizes follows a power law:

\[\tip{P}{probability}(\text{cluster size} = \tip{s}{size of a cluster}) \sim \tip{s}{size of a cluster}^{-\tip{\tau}{exponent of the cluster size distribution}}\]

This tells us that while most clusters are tiny, the “spanning cluster” is beginning to touch every scale of the system simultaneously.

The Ising Model of Magnetization

The Ising Model explains how local interactions between individual atoms lead to macroscopic ferromagnetism. Imagine a grid of tiny compass needles, where each one can only point up or down. Every needle would like to point in the same direction as its neighbours. Heat works against this: the hotter it is, the more often needles flip randomly.

• Control Parameter: the temperature \(\tip{T}{temperature}\) .
• Order Parameter: the magnetization \(|\tip{M}{magnetization: how strongly the spins agree on one direction}|\) , how strongly the needles agree on one direction overall.

When it is hot, the random flips win, the needles point every which way, and the magnet has no magnetization. When we cool it down below the Curie Temperature \(\tip{T_c}{Curie temperature: critical temperature of the magnet}\) , the needles “choose” a direction and align. Which direction is pure chance, but once chosen, the whole magnet follows. Below \(\tip{T_c}{Curie temperature: critical temperature of the magnet}\) , \(|\tip{M}{magnetization: how strongly the spins agree on one direction}|\) grows continuously from zero, a continuous critical transition. Exactly at \(\tip{T_c}{Curie temperature: critical temperature of the magnet}\) , aligned patches of every size coexist, mirroring the scale invariance seen in percolation.

The model in formulas

Each site \(\tip{i}{a site of the lattice}\) of the lattice has a “spin” \(\tip{s_i}{spin at site i: +1 (up) or -1 (down)}\) that can be either up (\(+1\) ) or down (\(-1\) ). The energy of the system is

\[\tip{H}{total energy of the spin configuration} = -\tip{J}{coupling strength: how much neighbours want to align} \sum_{\langle \tip{i}{a site of the lattice},\tip{j}{a neighbouring site of the lattice} \rangle} \tip{s_i}{spin at site i: +1 (up) or -1 (down)} \tip{s_j}{spin at site j: +1 (up) or -1 (down)}\]

where \(\tip{J}{coupling strength: how much neighbours want to align} > 0\) favors neighboring spins pointing in the same direction and the sum runs over neighbouring pairs. We look at the limit of time to infinity (\(\tip{t}{time} \to \infty\) ) to find the steady-state equilibrium. The order parameter is the net magnetization \(\tip{M}{magnetization: how strongly the spins agree on one direction} = \frac{1}{\tip{N}{number of spins}}\sum_\tip{i}{a site of the lattice} \tip{s_i}{spin at site i: +1 (up) or -1 (down)}\) , reported as \(|\tip{M}{magnetization: how strongly the spins agree on one direction}|\) because the direction the spins pick is arbitrary (spontaneous symmetry breaking).

For the 2D nearest-neighbor Ising model (zero external field), \(\tip{k_B}{Boltzmann constant: converts temperature to energy} \tip{T_c}{Curie temperature: critical temperature of the magnet} / \tip{J}{coupling strength: how much neighbours want to align} = 2 / \ln(1 + \sqrt{2}) \approx 2.269\) . The critical exponents include \(\tip{\beta}{critical exponent of the magnetization} = 1/8\) (magnetization) and \(\tip{\nu}{critical exponent of the correlation length} = 1\) (correlation length).

Disease Outbreak Revisited: SIR Model

Let’s come back to the first example. The classical SIR model divides the population into three groups: susceptible (\(\tip{S}{susceptible: people who can still get infected}\) ), infected (\(\tip{I}{infected: people who are currently ill}\) ), and recovered (\(\tip{R}{recovered: people who are immune}\) ). Susceptible people get infected when they meet infected people, and infected people recover after a while. Everything is governed by the basic reproduction number \(\tip{R_0}{reproduction number: people one infected person infects on average}\) , the number of people one infected person infects on average.

• For \(\tip{R_0}{reproduction number: people one infected person infects on average} < 1\) : Every infected person passes the disease on to less than one other person. The epidemic dies out.
• For \(\tip{R_0}{reproduction number: people one infected person infects on average} > 1\) : The epidemic takes off, and a finite fraction of the population gets infected in the end.

Here is the surprise: although the start of an outbreak feels like a tipping point, because of the exponential growth, the end state is continuous. Just above \(\tip{R_0}{reproduction number: people one infected person infects on average} = 1\) only very few people get infected in total, and this fraction grows smoothly with \(\tip{R_0}{reproduction number: people one infected person infects on average}\) . Furthermore, close to \(\tip{R_0}{reproduction number: people one infected person infects on average} = 1\) the epidemic grows very slowly, so the system takes a very long time to “decide” whether the outbreak will explode or vanish. This is called critical slowing down.

The SIR equations and the final size

The movement between the groups is governed by nonlinear differential equations:

\[\frac{d\tip{S}{susceptible: people who can still get infected}}{d\tip{t}{time}} = -\tip{\beta}{infection rate: contacts that lead to an infection per unit of time} \frac{\tip{S}{susceptible: people who can still get infected}\tip{I}{infected: people who are currently ill}}{\tip{N}{total population}}, \quad \frac{d\tip{I}{infected: people who are currently ill}}{d\tip{t}{time}} = \tip{\beta}{infection rate: contacts that lead to an infection per unit of time} \frac{\tip{S}{susceptible: people who can still get infected}\tip{I}{infected: people who are currently ill}}{\tip{N}{total population}} - \tip{\gamma}{recovery rate: inverse of the average illness duration} \tip{I}{infected: people who are currently ill}, \quad \frac{d\tip{R}{recovered: people who are immune}}{d\tip{t}{time}} = \tip{\gamma}{recovery rate: inverse of the average illness duration} \tip{I}{infected: people who are currently ill}\]

where \(\tip{\beta}{infection rate: contacts that lead to an infection per unit of time}\) is the infection rate and \(\tip{\gamma}{recovery rate: inverse of the average illness duration}\) the recovery rate, so \(\tip{R_0}{reproduction number: people one infected person infects on average} = \tip{\beta}{infection rate: contacts that lead to an infection per unit of time}/\tip{\gamma}{recovery rate: inverse of the average illness duration}\) . This model assumes everyone mixes perfectly with everyone (mean-field theory).

In the language of dynamical systems, the SIR model exhibits a transcritical bifurcation at \(\tip{R_0}{reproduction number: people one infected person infects on average} = 1\) : below it the only stable state is \(\tip{I}{infected: people who are currently ill}=0\) , above it \(\tip{I}{infected: people who are currently ill}=0\) becomes unstable.

The final infected fraction \(\tip{z}{final infected fraction of the population} = \tip{R}{recovered: people who are immune}(\infty)/\tip{N}{total population}\) satisfies \(1 - \tip{z}{final infected fraction of the population} = e^{-\tip{R_0}{reproduction number: people one infected person infects on average} \tip{z}{final infected fraction of the population}}\) , so \(\tip{z}{final infected fraction of the population}\) increases smoothly from zero as \(\tip{R_0}{reproduction number: people one infected person infects on average}\) passes 1. The early growth rate is \(\tip{\beta}{infection rate: contacts that lead to an infection per unit of time} - \tip{\gamma}{recovery rate: inverse of the average illness duration}\) , which vanishes at the threshold: this is the critical slowing down.

The first simulation is the well-mixed model above: drag \(\tip{R_0}{reproduction number: people one infected person infects on average}\) across 1 and watch the final size lift off continuously, with no jump. The second is a spatial stochastic variant, shown for contrast.

Remarks

All the examples above are different faces of the same mathematical idea: a small, smooth change of a parameter that changes the long-term behavior of a system qualitatively. Mathematicians call this a bifurcation.

Going deeper: bifurcations and further reading

Critical phenomena are the physical manifestation of bifurcations. While equilibrium statistical critical points are most cleanly characterized by universal scaling, many can be understood via coarse-grained dynamics:

• Discontinuous transitions align with saddle-node or subcritical pitchfork bifurcations, often exhibiting hysteresis.
• Continuous transitions (supercritical pitchforks) are characterized by scale invariance and diverging correlation lengths.

For further study, explore Renormalization Group Theory (the mathematical “zoom lens”) and Mean-Field Theory (the approximation of average behavior).

Conclusion

• Thresholds Matter: Systems often appear stable until a parameter crosses a critical value, triggering a regime shift.
• Tipping vs. Continuous: Tipping points involve violent, often irreversible jumps; continuous transitions involve the smooth unfolding of a new state.
• Power Laws: They act as both the mechanism (competing scales) and the signature (scale invariance) of criticality.

Sources

• Nonlinear Dynamics and Chaos by Steven H. Strogatz
• The Los Alamos Primer by Robert Serber
• Structures by J.E. Gordon

Simulation codes are inline JavaScript snippets that can be inspected via browser developer tools. The charts share one colour convention: blue for measured values, grey for theory, and a dashed red line for a critical value.