Chaos hides in one line of algebra

Here is a claim to be suspicious of: you can manufacture genuine, unpredictable chaos with a single line of algebra. One multiplication, one subtraction, no randomness, no noise, nothing hidden up a sleeve. Feed a number in, get a number out, feed that number back in, and repeat. Do it a few hundred times and, depending on one dial, the output either settles to a steady value, or ticks between two values forever, or never repeats and behaves for all practical purposes like a coin you cannot call.

The line is the logistic map, and it started life as a toy model of a population. Think of $x_n$ as this year's headcount written as a fraction of the most the environment can hold, so $x_n$ lives between $0$ (extinct) and $1$ (packed to capacity). The dial $r$ is the growth rate. Small populations breed fast, which is the $r\,x_n$ part; crowded ones starve, which is the $(1 - x_n)$ part pulling the number back down. Turn $r$ up and you are asking what a faster-breeding species does to itself.

The whole model is one multiplication

Written out, the map is almost embarrassingly small. This is the entire dynamical system:

$$x_{n+1} = r\,x_n\,(1 - x_n)$$

Copy LaTeXwith $$ delimitersCopy mathexpression only

Where does it come to rest? A fixed point is a value the map returns unchanged, $x^{*} = r\,x^{*}(1 - x^{*})$. There is the boring root $x^{*} = 0$ (extinction stays extinct), and one that matters:

$$x^{*} = 1 - \frac{1}{r}$$

Copy LaTeXwith $$ delimitersCopy mathexpression only

For $r < 1$ this root is negative and meaningless, so the population dwindles to nothing. For $1 < r < 3$ it sits happily between $0$ and $1$ and the population settles there, which is exactly what you see on the left of the diagram. The question of whether it stays there is settled by the slope of the map at the fixed point. Differentiating equation (1) gives $f'(x) = r(1 - 2x)$, and plugging in equation (2) it collapses to a clean $f'(x^{*}) = 2 - r$. A fixed point holds firm while that slope has magnitude under $1$, so the steady state is stable for $1 < r < 3$ and loses its grip at exactly $r = 3$, where the slope hits $-1$. That is the first fork.

The cascade: doubling, then doubling faster

At $r = 3$ the steady state splits into a 2-cycle, a population that oscillates between a high year and a low year. Turn the dial further and that 2-cycle becomes unstable in the same way the fixed point did, and each branch splits again into a 4-cycle. Then an 8-cycle, a 16-cycle, and on up the powers of two. This is the period-doubling cascade, and the striking thing is not that it happens but how quickly it happens. The forks come at

$$r_1 = 3, \quad r_2 \approx 3.4495, \quad r_3 \approx 3.5441, \quad r_4 \approx 3.5644, \; \dots$$

Each gap is roughly four and a bit times shorter than the one before. Feigenbaum noticed that the ratio of successive gaps settles onto a fixed number:

$$\delta = \lim_{n\to\infty} \frac{r_n - r_{n-1}}{r_{n+1} - r_n} \approx 4.6692$$

Copy LaTeXwith $$ delimitersCopy mathexpression only

Because the gaps shrink geometrically, the whole infinite tower of doublings fits inside a finite stretch of $r$. Add up a geometric series with ratio $1/\delta$ and it converges. The doublings pile up at a limit point, the onset of chaos, at $r_\infty \approx 3.5699$. Past there the periodic order is spent, and the system is chaotic.

Feigenbaum δ

4.6692

the geometric rate the forks accelerate, the same for every smooth hump map

First bifurcation

r = 3

the steady state splits into a 2-cycle

Onset of chaos

r∞ ≈ 3.5699

infinitely many doublings, all squeezed below this

One number that does not care what equation you used

Here is where the logistic map stops being a cute biology model and starts being a fact about the world. That constant $\delta \approx 4.6692$ is not special to $r\,x(1-x)$. Take any smooth map with a single hump, a rounded maximum: the sine map $r\sin(\pi x)$, a cosine, a quartic, anything with one gentle peak. Every one of them period-doubles its way into chaos, and every one does it with the same $\delta$. The individual bifurcation values differ, the shape of the diagram differs, but the rate at which the forks accelerate is identical to the last digit you care to compute.

The route into chaos has a speed limit, and it is the same speed limit for a population model, a dripping tap, and a convecting fluid: a number near 4.6692 that nobody put there on purpose.

Feigenbaum found this in 1978 on a pocket calculator, watching the numbers from two completely different maps converge onto the same ratio and assuming at first that he had a bug. He did not. The constant is universal in the technical sense: it is a property of the whole class of one-hump maps, the way $\pi$ is a property of every circle rather than of any particular one. And the prediction is not just mathematical folklore. Experiments on convecting liquid helium, on electronic oscillators, on dripping taps, have all measured period-doubling cascades and read off a $\delta$ that agrees with 4.6692.

Windows of order inside the chaos

The chaotic band is not solid. Look again at the diagram between $r_\infty$ and $r = 4$ and you will spot vertical gaps, clean stripes where the scribble suddenly collapses back to a handful of points. These are windows of order, periodic cycles living in the middle of the chaos. The widest and most famous is the period-3 window, which opens abruptly at

$$r = 1 + \sqrt{8} \approx 3.8284$$

Drag the marker to about $3.84$ and watch the chaos snap into a clean 3-cycle, boom-boom-bust repeating cleanly. Then, as you nudge $r$ higher inside the window, that 3-cycle period-doubles into a 6-cycle, a 12-cycle, and cascades straight back into chaos, a perfect miniature of the whole diagram tucked inside itself. The structure is fractal: order and chaos interleaved at every scale.

The same fingerprint in a weather model

So far this is a one-dimensional map, a number chasing its own tail. Real chaos usually lives in continuous systems, differential equations flowing through time, and the surprise is that the logistic map's essential feature carries straight over. That feature is sensitive dependence on initial conditions: two starts a hair apart are pulled exponentially far apart as the system runs. The rate of that pulling-apart is the Lyapunov exponent, and for the logistic map it is an average of the log-slope along the trajectory:

$$\lambda = \lim_{n\to\infty} \frac{1}{n} \sum_{i=0}^{n-1} \ln\bigl| f'(x_i) \bigr|, \quad f'(x) = r(1 - 2x)$$

Copy LaTeXwith $$ delimitersCopy mathexpression only

The sign of $\lambda$ tells you everything. Where the map is periodic, $\lambda < 0$ and small errors heal. At every bifurcation and at $r_\infty$ it touches zero. Across the chaotic band $\lambda > 0$, and a positive Lyapunov exponent is the working definition of chaos in any system, discrete or continuous. If you plotted $\lambda$ against $r$ it would trace the diagram's mood exactly: negative in the ordered valleys, spiking positive in the chaos, dipping back below zero inside every window of order.

Edward Lorenz found the same positive exponent in three coupled differential equations for convection back in 1963. His trajectory never settles to a point or a loop; it winds forever around a two-winged surface, close but never repeating, and two starts differing in the fourth decimal place end up on opposite wings. That is the butterfly effect, and it is equation (4) with a positive answer, dressed in continuous clothes. The logistic map and the Lorenz system are not analogies for each other. They are two rooms in the same house, and the house is nonlinear dynamics.

Chaos you can feel

The last room is the one you can build on a desk. A double pendulum, one rod swinging off the end of another, obeys nothing but Newton's laws, exact and deterministic. It is also violently chaotic.

Release the arms from almost the same angle and they track together for a beat, then peel apart into completely different motions and never reconcile. That window of agreement, a second or two before divergence, is the prediction horizon set by the Lyapunov exponent, the same quantity as equation (4), the same idea as the two dots on the Lorenz butterfly. Turn up the gravity and the divergence comes faster, a bigger exponent, a shorter horizon.

Three rooms, then: a toy population model, a stripped-down weather model, and a swinging pendulum. A discrete map, a set of differential equations, and a lump of physical hardware. They share almost nothing on the surface, and underneath they share the one thing that matters, a route into chaos with a positive Lyapunov exponent and a structure that repeats itself as you zoom in. My own favourite consequence is the humbling one: the wall on weather forecasting near two weeks is not an engineering embarrassment waiting on a faster computer. It is the same mathematics as the fork at $r = 3$, and no supercomputer beats a logarithm. Something to sit with next time a forecast lets you down.

Next time you see a system doubling its period, whatever it is made of, you will know what is coming, and roughly how fast. Watch this space.

Reading further

• May, Simple mathematical models with very complicated dynamics (Nature 261, 1976). The paper that put the logistic map on everyone's desk, arguing that even a first course should teach how simple nonlinear models misbehave. Short, opinionated, and still the best entry point.
• Feigenbaum, Quantitative universality for a class of nonlinear transformations (J. Stat. Phys. 19, 1978). Where $\delta \approx 4.6692$ and its companion constant are pinned down, and where universality is argued for the whole class of one-hump maps.
• Strogatz, Nonlinear Dynamics and Chaos, chapter 10. The standard modern textbook treatment of the logistic map, the cobweb picture, Lyapunov exponents, and the renormalisation behind Feigenbaum's constants, drawn carefully.
• Lorenz, Deterministic Nonperiodic Flow (J. Atmos. Sci. 20, 1963). The continuous-time companion to this post: three equations, no randomness, and the first clear sighting of the butterfly effect.