How the Leopard Got Its Spots

Here is a claim that should sound wrong. Take two chemicals. Let one of them catalyse its own production, let the other one feed it, and let both simply spread out the way ink spreads in water. Start them mixed almost evenly across a dish, with only a whisper of noise to break the symmetry. Wait. Spots appear. Not because anything tells them where to go, not because a mould stamps them in, but because the maths of "react and spread" has no stable flat answer. The blank sheet is the one thing the system cannot keep.

That is the idea Alan Turing published in 1952, in what turned out to be his last major paper, The Chemical Basis of Morphogenesis. He was asking how a ball of identical cells in an embryo decides to become a striped tail or a spotted flank when every cell starts with the same instructions. His answer was that diffusion, which we all learn is the great smoother, can be the very thing that creates structure. Let's watch it happen first, then take it apart.

The two-chemical recipe

The model in the lab is the Gray-Scott system, a particularly clean activator-substrate pair. Call the two concentrations $u$ and $v$. The reaction is autocatalysis: one unit of $v$ recruits more $v$ by consuming $u$, written as a chemical scheme it looks like this.

$$U + 2V \longrightarrow 3V, \qquad V \longrightarrow P$$

Copy LaTeXwith $$ delimitersCopy mathexpression only

So $v$ is the activator: the more of it there is, the faster it makes more of itself (that $uv^2$ term, cubic in the reactants). And $u$ is the substrate it burns through. Now add diffusion, write down the rate of change of each concentration at every point, and you get the two coupled partial differential equations the lab is actually integrating.

$$\begin{aligned} \frac{\partial u}{\partial t} &= D_u\,\nabla^2 u - u v^2 + F(1 - u) \\[4pt] \frac{\partial v}{\partial t} &= D_v\,\nabla^2 v + u v^2 - (F + k)\,v \end{aligned}$$

Copy LaTeXwith $$ delimitersCopy mathexpression only

Read it left to right. The $\nabla^2$ term is diffusion, the bit that spreads each chemical toward its neighbours. The $uv^2$ term is the reaction that moves $u$ into $v$. The $F(1-u)$ term tops $u$ back up (the feed, replacing what gets used), and $(F+k)v$ removes $v$ (the kill). Four moving parts, two of them just spreading and reacting. There is genuinely nothing else in here, no hidden template, which is what makes the result feel like a magic trick.

The one number that matters more than any other is the ratio of the two diffusion rates. In the lab, $D_u = 0.16$ and $D_v = 0.08$: the substrate spreads twice as fast as the activator. Hold that thought, because it turns out to be the whole secret.

Why the great smoother makes structure

The intuition Turing gave is short-range activation with long-range inhibition. The activator $v$ is slow, so wherever it gets a local lead it builds a sharp peak right where it is. The substrate it depends on is fast, so a growing peak drains $u$ from a wide area around itself. Nearby, $u$ has been eaten and can't be replenished quickly, so no second peak can start there. Far away, the substrate is still plentiful and another peak can rise. The peaks therefore sit at a preferred distance from each other. That spacing is the pattern.

The phrase "long-range inhibition" is a useful lie worth flagging. The substrate $u$ does not actively suppress $v$ the way a true inhibitor would. It just gets eaten, and because it refills slowly and spreads quickly, a local feast of $v$ leaves a wide ring of starvation around it. The inhibition is effective, a consequence of depletion plus fast diffusion, rather than a chemical that hunts $v$ down. The behaviour is the same; the mechanism is gentler than the name suggests.

The maths of the instability

We can make "a uniform mixture cannot survive" precise, and this is where the post connects to every other dynamical-system story on this site. Freeze the diffusion for a second and treat the reaction terms as an ordinary two-variable system. It has a steady mixed state where production balances loss, and you check its stability with the Jacobian $J$ of the reaction kinetics, exactly the linearisation we use to classify fixed points in phase portraits. For Gray-Scott in the patterning regime, that mixed state is stable on its own: leave the chemicals to react with no spreading and they settle.

Now switch diffusion back on and ask what a small ripple of wavenumber $q$ does. A bumpy perturbation grows or shrinks at a rate $\lambda(q)$, the larger of two eigenvalues, and its sign is controlled by a single function.

$$h(q^2) = D_u D_v\,q^4 - (D_u\, g_v + D_v\, f_u)\,q^2 + \det J$$

Copy LaTeXwith $$ delimitersCopy mathexpression only

Here $f_u$ and $g_v$ are entries of that reaction Jacobian. When $h(q^2) < 0$ for some band of $q$, ripples at those wavelengths grow, even though the no-diffusion state was stable. That is a diffusion-driven instability, and it is the engine of the whole thing. The fastest-growing wavelength wins and sets the spot spacing you measure on the screen.

A footnote for the connoisseur: classical linear activator-inhibitor models often need the inhibitor to diffuse roughly ten times faster than the activator before patterns appear, which is awkward biologically, because real morphogens of similar size diffuse at similar rates. Gray-Scott gets away with a ratio of two because its cubic autocatalysis does extra work the linear models can't. It is one reason this particular toy is so beloved.

Reading the parameter map

Everything above sets the scale of the pattern. What kind of pattern you get, dots versus mazes versus something that splits and crawls, lives in the two dials $F$ and $k$. Pearson's 1993 survey mapped this plane and found a small continent of wildly different behaviours packed into a narrow strip of parameter space.

The preset above is the strange one. The dots grow, and when a dot gets too fat for the wavelength the dispersion relation allows, it becomes unstable to splitting and pinches into two, which then push apart and repeat. It looks exactly like cells dividing. There are no cells. There is no instruction to divide. It is the same two equations, parked at a different $(F, k)$.

Two failure modes are worth a try while you have the controls. Push the kill rate $k$ too high and $v$ can't sustain itself; the whole field collapses to the dead, uniform "all substrate" state and stays there. Pull the diffusion rates together (you can't in this lab, but it's the thought experiment from the callout) and the pattern never starts. Both are reminders that the pattern lives in a window, and outside it the blank sheet wins after all.

From two chemicals to a leopard

Turing wrote about chemistry, but the title of this post is stolen, gently, from James Murray, who in 1988 used exactly this reasoning to argue about animal coats. The argument has a lovely, testable consequence about geometry. A reaction-diffusion pattern fitting into a narrow, tapering domain, like a tail, is forced toward stripes, because a single wavelength can only pack across the width one way. A broad domain, like a flank, has room for spots. So you can have a spotted body running out into a striped tail, but a striped body cannot end in a spotted tail. Look at the big cats and the rule holds: plenty of spotted cats with banded tails, no banded cats with spotted tails.

A spotted animal can have a striped tail, but a striped animal cannot have a spotted tail. The geometry of the canvas decides, and the chemistry just fills it in.

This is not only a story about ink in a dish. In 2020 a team led by Stephan Getzin showed that the "fairy circles" in the Australian outback near Newman, regular gaps in spinifex grass, fit a Turing model where the grass biomass is the slow activator and soil water is the fast-spreading inhibitor. The vegetation engineers its own water supply, the water redistributes faster than the plants do, and the landscape settles into a honeycomb you can see from a plane. Same maths, a different pair of "chemicals".

The same instability, elsewhere

The reason this clicks, if you've read around the site, is that it is the same move as everywhere else. A stable fixed point, a perturbation, a parameter that flips stability and lets the perturbation grow. In the Lorenz post a tiny difference in initial conditions blows up and you get chaos; here a tiny ripple blows up and you get order. Both are linear instabilities of an equilibrium, classified with the same Jacobian we use in phase portraits. The difference is only what the growing mode looks like: a trajectory peeling away from a point, or a wavelength peeling away from a flat field.

We tend to think of "smoothing" and "structure" as opposites. Turing's quiet, radical point was that with two ingredients spreading at different speeds, the smoothing is the structure. Drag the Feed and Kill sliders for a while and you'll start to feel where the window is. The blank sheet, it turns out, was never an option.

Reading further

• Turing, The Chemical Basis of Morphogenesis (1952). The original, Phil. Trans. R. Soc. B 237, 37-72. Surprisingly readable, and the source of the whole field; his "morphogen" instability is equation (3) in disguise.
• Pearson, Complex Patterns in a Simple System (1993). Science 261, 189-192. The map of the Gray-Scott $(F, k)$ plane, including the self-replicating-spot regime in the lab.
• Murray, How the Leopard Gets Its Spots (1988). Scientific American 258(3), 80-87. The animal-coat argument and the spotted-body-striped-tail rule, from the master of mathematical biology.
• Getzin et al., Fairy circle Turing patterns (2021). J. Ecology 109, 399-416. Reaction-diffusion confirmed in the field, in Australian spinifex grassland.
• Karl Sims, Reaction-Diffusion Tutorial. The clearest hands-on explanation of the Gray-Scott numerics, and the source of the exact Laplacian stencil this lab uses.