Chaos and Strange Attractors

A forecast that shouldn't have changed#

In the winter of 1961, Edward Lorenz was running a toy weather model on a Royal McBee LGP-30 — a vacuum-tube machine that printed one line of numbers per simulated day. He wanted to re-examine a run, so rather than start from the beginning he typed in a state from the middle of the printout and set it going again.

The new run tracked the old one for a while. Then it drifted. Then it diverged completely: a different weather, with different storms, from what was supposed to be the same simulation.

Lorenz first assumed a vacuum tube had failed. It hadn't. The machine held six decimal digits internally but printed only three. He had typed 0.506 where the computer had been carrying 0.506127. A discrepancy of one part in ten thousand — far smaller than any real thermometer's error — had grown until it dominated the forecast.

Nothing in the model was random. Given the same numbers, it produced the same answer every time. And yet it was, in a deep sense, unpredictable.

Chaos is not randomness#

This is the misconception worth killing immediately: chaotic does not mean random. Lorenz's equations contain no random number generator, no noise term, no probability distribution. They are as deterministic as $F = ma$. Run them twice from bit-identical initial conditions and you get bit-identical output.

What chaos means is sensitive dependence on initial conditions. Two starting states that differ by $\varepsilon$ do not stay $\varepsilon$ apart — their separation grows exponentially, roughly

$$d(t) \approx \varepsilon\, e^{\lambda t}$$

until it saturates at the size of the system. Because every real measurement has finite precision, $\varepsilon$ is never zero, and the exponential does the rest.

It is worth contrasting this with Brownian motion. A pollen grain jitters because it is genuinely being kicked by an unknowable number of molecules; the randomness is in the process itself, and re-running the same experiment gives a different path no matter how well you measure the starting point. Chaos is the opposite situation with the same practical consequence: the rule is exact, the future is uniquely determined, and you still cannot forecast it. Determinism guarantees predictability only if you have infinite precision.

The other place this shows up in a form you can hold in your hand is the double pendulum — a second rod hung off the bob of the first. One pendulum is the textbook example of clockwork regularity. Two, coupled, is a chaotic system: release it twice from angles you cannot distinguish by eye and after a few seconds the two motions have nothing to do with each other.

The simplest chaotic system there is#

You do not need fluid dynamics to get chaos. You need one nonlinear equation. The logistic map models a population that grows when small and is suppressed when crowded:

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

Here $x_n \in [0,1]$ is the population as a fraction of the maximum the environment can hold, and $r$ is the growth rate. That is the whole system: one variable, one parameter, one multiplication and one subtraction.

Its behaviour as $r$ increases is astonishing. For $r < 1$ the population dies out. For $1 < r < 3$ it settles to a single steady value. At $r = 3$ that fixed point loses stability and the orbit begins alternating between two values — a period-2 cycle. At $r \approx 3.449$ each of those splits again, giving period 4. Then 8, then 16, each cascade arriving sooner than the last, until at $r_\infty \approx 3.5699$ the period becomes infinite and the orbit never repeats at all.

The diagram paints itself as $r$ sweeps left to right: for each $r$, it plots the values the orbit eventually visits. Watch the single line fork into two, the two into four, and the forks pile up ever faster until the picture smears into a cloud. Then scrub the $r$ slider and read the orbit panel below. A few things to try:

• Park at $r = 2.9$: the orbit panel is a flat line. One value, forever.
• Step through $r = 3.2$, $3.5$, $3.55$: two values, four, eight. Count the dots.
• Cross $r = 3.5699$ and the orbit panel never repeats — but notice it is still confined to a band, not wandering anywhere it likes.
• Now find the clear vertical gap near $r = 3.83$. Inside the chaos, order returns: a clean period-3 window. Widen the slider a hair either side and it dissolves back into chaos. Those windows are everywhere in the diagram, at every scale.

Feigenbaum's number#

Measure where the bifurcations happen — $r_1 = 3$, $r_2 \approx 3.4495$, $r_3 \approx 3.5441$, $r_4 \approx 3.5644$ — and take the ratio of successive gaps. Mitchell Feigenbaum did this on a pocket calculator in 1975 and found it converging:

$$\delta = \lim_{n\to\infty}\frac{r_n - r_{n-1}}{r_{n+1} - r_n} = 4.669201\ldots$$

The remarkable part is not that a constant exists but that it is universal. The same $\delta$ governs the period-doubling route of any smooth map with a single quadratic maximum — and, as experiments through the late 1970s and 1980s confirmed, of dripping faucets, convecting liquid helium, oscillating chemical reactions, and driven electronic circuits. Systems with nothing physically in common share a number, because they share a geometry.

Measuring the divergence rate#

The parameter $\lambda$ in $d(t) \approx \varepsilon e^{\lambda t}$ is the Lyapunov exponent, and it is the quantitative definition of chaos. Formally,

$$\lambda = \lim_{t\to\infty}\frac{1}{t}\ln\frac{|\delta(t)|}{|\delta(0)|}$$

for the separation $\delta$ of two infinitesimally close trajectories. Its sign is the whole story: $\lambda < 0$ means perturbations decay and the system is stable; $\lambda = 0$ is the marginal, periodic case; $\lambda > 0$ means chaos.

For the logistic map there is a clean discrete version, averaging the log of the slope the orbit actually experiences:

$$\lambda = \lim_{N\to\infty}\frac{1}{N}\sum_{n=0}^{N-1}\ln\bigl|r(1 - 2x_n)\bigr|$$

At $r = 4$ this evaluates exactly to $\ln 2$: one bit of information about the initial condition is destroyed per iteration. Whatever precision you started with is gone after that many steps.

Lorenz's own system — his stripped-down model of convection rolls in a heated fluid layer — is the continuous counterpart:

$$\frac{dx}{dt} = \sigma(y - x),\qquad \frac{dy}{dt} = x(\rho - z) - y,\qquad \frac{dz}{dt} = xy - \beta z$$

with the classic values $\sigma = 10$, $\rho = 28$, $\beta = 8/3$. Three variables, three equations, two nonlinear terms ($xz$ and $xy$), and a largest Lyapunov exponent of $\lambda \approx 0.906$.

Two trajectories start on the Lorenz attractor separated by $\varepsilon$ in a single coordinate. The gold and pink dots are indistinguishable at first — they trace the same loops, cross to the same wing at the same moment. The lower panel plots their separation on a log scale, so exponential growth appears as a straight line, and the green dashed reference has slope $\lambda / \ln 10$.

What to look for:

• The measured curve tracks the reference slope closely — that straight-line stretch is the Lyapunov exponent, read directly off the plot.
• Growth stops at the "attractor size" line. Separation cannot exceed the diameter of the attractor, so the curve flattens: at that point the two forecasts are as different as two randomly chosen states, and all skill is gone.
• Drag $\varepsilon$ from $10^{-2}$ down to $10^{-10}$. Eight orders of magnitude better precision — and the curve simply starts lower and takes a bit longer to climb. The horizon readout moves from about 5 to about 25 time units. Improving your measurements by a factor of a billion did not buy a factor of a billion in forecast time; it bought roughly a factor of five.

That last point is the practical heart of chaos. Inverting $\varepsilon e^{\lambda T} = D$ for the horizon gives

$$T \approx \frac{1}{\lambda}\ln\frac{D}{\varepsilon}$$

Precision enters through a logarithm. Each extra decimal digit of accuracy buys the same fixed increment $\ln(10)/\lambda$ of extra forecast — never more.

Bounded, but never repeating#

Watch the top panel long enough and a paradox surfaces. The trajectory never repeats, never settles, never crosses itself — yet it never leaves a compact region either, and it never fills that region. It is drawn to a specific two-winged shape and stays on it forever.

That shape is a strange attractor. "Attractor" because nearby states are pulled onto it: start the Lorenz system anywhere and it converges to that surface. "Strange" because the set is fractal — the Lorenz attractor has a dimension of about 2.06, more than a surface, less than a solid — and because motion on it is chaotic.

The two properties are what make chaos different from both order and noise. Volumes contract, so the long-run behaviour is confined and structured; but lengths along the attractor stretch, so points on it separate. Stretching and folding, over and over, like kneading dough: two grains of flour pressed together end up arbitrarily far apart, yet the dough stays in the bowl. This is why chaotic systems have excellent statistics and terrible forecasts. You can say with confidence what the climate of the attractor is — which regions get visited, how often, with what average — while being unable to say where the point will be at any specific future time.

The two-week wall#

Modern weather models are extraordinary. They ingest millions of satellite, radiosonde, buoy, and aircraft observations, run on some of the largest computers on Earth, and resolve the atmosphere down to a few kilometres. Their skill has improved by about a day per decade for forty years: today's six-day forecast is as good as the five-day forecast was ten years ago.

And yet the ceiling holds at roughly two weeks. The atmosphere's largest Lyapunov exponent corresponds to error-doubling times of a day or two at synoptic scales. An initial uncertainty of one part in a million — far beyond current observation networks, which leave gaps over oceans and between soundings — grows to forecast-destroying size in about two weeks. Recent work suggests the true intrinsic limit may sit near 14–15 days even with a perfect model and near-perfect data.

This is not an engineering problem. Better satellites raise the wall by days, not decades, because of that logarithm. It is a property of the equations of the atmosphere, and no amount of computing power removes it.

The response was to stop asking for a single answer. Operational centres now run ensemble forecasts: fifty-odd copies of the model, each started from a slightly perturbed state, all equally consistent with the observations. If the ensemble members stay together, the atmosphere is in a predictable regime and the forecast is confident. If they scatter, it isn't. That scatter is the Lyapunov exponent doing its work in production, and the "70% chance of rain" in a forecast app is, quite literally, a count of how many ensemble members got wet.

The same logic reaches well beyond weather. Chaos sets the limits on long-term orbital prediction in the solar system (the inner planets' positions become unpredictable on roughly 10-million-year timescales), on cardiac and neural dynamics, on population ecology, and on the behaviour of turbulent flows. Wherever the equations are nonlinear and the Lyapunov exponent is positive, prediction has a horizon — and knowing the size of that horizon is more useful than pretending it isn't there.