Sensitive Dependence

In 1961 Edward Lorenz, rerunning a numerical weather model, typed in a number from a printout—0.506 instead of the stored 0.506127—and went for coffee. The truncated run, started from a state differing by less than one part in a thousand, tracked the original for a while and then peeled away completely, until the two simulated weathers shared nothing but their statistics. The machine was deterministic. The equations had no noise. Yet a rounding error in the fourth decimal place had erased the forecast.

This is sensitive dependence on initial conditions, the defining signature of deterministic chaos: a system whose rule of evolution is fixed and noiseless, but whose trajectories from nearby starts separate so fast that finite-precision knowledge of the present buys vanishing knowledge of the far future. The same phenomenon governs the swing of a double pendulum and, in the relevant regime, the tumble of a thrown die. It belongs squarely to the Atlas of structures that are simple to state and unbounded in consequence, and it is rendered in Plate IV — the Lorenz attractor.

What “deterministic chaos” means

A dynamical system is a state $\mathbf{x}(t)$ evolving under a fixed rule, $\dot{\mathbf{x}} = \mathbf{f}(\mathbf{x})$, with no random term anywhere. Determinism guarantees that the same initial state always produces the same future. Chaos is the claim that this guarantee is, in practice, useless: you can never specify the same initial state, only a nearby one, and “nearby” decays.

Make this precise by following a tiny perturbation. Let two trajectories start a distance $\delta_0$ apart. In a chaotic system the separation grows, on average, exponentially,

$\|\delta(t)\| \approx \delta_0\, e^{\lambda t},$

with $\lambda > 0$ the largest Lyapunov exponent. The exponent is an average stretching rate along the attractor, the long-time analogue of an eigenvalue of the linearized flow.Locally the perturbation obeys $\dot\delta = J(\mathbf{x})\,\delta$, where $J$ is the Jacobian of $\mathbf{f}$. The Lyapunov exponent is the time-averaged growth rate of $\delta$ along the trajectory—an eigenvalue smeared over the whole orbit. A negative $\lambda$ means errors shrink and the system is predictable; $\lambda = 0$ is the marginal case of regular oscillation; $\lambda > 0$ is chaos.

The consequence is brutal because exponentials are brutal. Suppose you know the present state to one part in $10^6$ and you need it to one part in $10$ to forecast usefully. You can tolerate growth by a factor of $10^5$, which takes time $t = \ln(10^5)/\lambda \approx 11.5/\lambda$. Now demand a hundredfold better measurement, $10^8$ precision. You buy only $\ln(10^3)/\lambda \approx 6.9/\lambda$ more time. Halving the error again and again extends the forecast by a constant increment each time. Prediction horizon grows like the logarithm of effort, which is the mathematician’s way of saying it does not grow.

The Lorenz system

Lorenz distilled atmospheric convection—warm fluid rising, cool fluid sinking—into three ordinary differential equations:

$\dot{x} = \sigma(y - x), \qquad \dot{y} = x(\rho - z) - y, \qquad \dot{z} = xy - \beta z.$

Here $x$ measures the convection roll’s intensity, $y$ and $z$ the temperature structure; $\sigma, \rho, \beta$ are physical parameters. At the classic values $\sigma = 10$, $\rho = 28$, $\beta = 8/3$ the system never settles to a fixed point and never repeats. The two nonlinear terms, $xz$ and $xy$, are the entire source of the trouble: a linear system can only decay, grow, or oscillate, never generate chaos.

Plotted in three dimensions the trajectory traces the famous butterfly: two lobes, with the orbit winding around one, then crossing to the other after an unpredictable number of loops. The sequence of lobe-switches is where the sensitivity lives. Two trajectories agreeing to fifteen decimal places will eventually disagree about which wing they are on, and from that moment they are uncorrelated.

A pendulum you can build

The Lorenz attractor is an abstraction of a fluid; the double pendulum is a length of metal you can hold in your hand—one rigid arm hinged to a second, both swinging under gravity. Its equations of motion follow from undergraduate Lagrangian mechanics and are entirely deterministic. Released from a steep angle it does not swing; it flails, and two such pendulums released from visually identical positions diverge into different motions within seconds. There is no friction term needed and no randomness invoked. The nonlinearity of the coupling between the two arms is enough.

This is the right antidote to a common misreading. Chaos is not the absence of law. It is law that amplifies ignorance. The double pendulum obeys Newton exactly; what it lacks is the courtesy of being stable under the imprecision with which we know its starting angle.

Chaos is not randomness

A fair die is the bridge case. Its faces are equiprobable not because physics is random—a rigid body bouncing on a table is governed by deterministic mechanics—but because the map from launch conditions to final face is so finely interleaved that any honest spread in how you release it smears across all six outcomes. Chaos manufactures effective randomness from deterministic dynamics, which is why the language of probability—of expectation and variance—is the only useful description left once the trajectory itself is unknowable.The deep result is that a chaotic system, though deterministic, admits an invariant probability measure on its attractor: long-time averages of an observable converge even though the orbit never does.

But chaos and randomness are not the same. A truly random sequence has no generating rule; a chaotic one is produced by a short, exact recipe. The distinction is operational. Knowing the Lorenz equations exactly tells you everything about the attractor—its shape, its statistics, the long-run fraction of time spent on each wing—while telling you nothing about where on it the system will be an hour hence. The structure is knowable; the trajectory is not. This is also why a chaotic system is not memoryless in the Markov sense at the level of its full state—the next state is a deterministic function of the current one—yet behaves, when coarsely observed, like a stochastic process.

Strange attractors and the forecast wall

Two facts about the butterfly seem to be in tension. The trajectory diverges from its neighbors exponentially, yet it stays inside a bounded region forever—it never flies off to infinity and never settles down. The reconciliation is geometric: the flow stretches nearby points apart and then folds the stretched sheet back onto itself, like a baker kneading dough. Repeated stretching-and-folding produces a set that is bounded, never self-intersecting, and infinitely layered—a fractal of non-integer dimension. This is a strange attractor: attracting, because nearby states are drawn onto it; strange, because motion on it is chaotic.

The practical upshot has a name in meteorology: the predictability horizon. In the real atmosphere the error-doubling time is on the order of a day or two, which is precisely why detailed weather forecasts degrade past roughly two weeks no matter how many sensors and how much computing power are thrown at them. The limit is not engineering; it is mathematical. Modern forecasting answers it not by chasing the unreachable exact trajectory but by running ensembles—many slightly perturbed initial states—and reading off the spread as a probability. We have stopped predicting the weather and started predicting the distribution of weathers, which is the only honest thing a positive Lyapunov exponent allows.

Further reading

• Edward N. Lorenz, “Deterministic Nonperiodic Flow,” Journal of the Atmospheric Sciences (1963) — the founding paper.
• The site’s Atlas — neighboring structures in the same family.