11.4 Lyapunov exponents and the horizon of prediction

Every lesson so far has leaned on the phrase “nearby trajectories separate exponentially.” It is time to attach a number to it. That number is the Lyapunov exponent, the average exponential rate at which infinitesimally close states pull apart. Its sign is the cleanest definition of chaos there is — positive means chaotic — and its value sets a hard limit on how far ahead any deterministic system can be forecast. The limit is far more pessimistic than intuition suggests, because it tightens only logarithmically as measurements improve.

Defining the exponent

Take two trajectories separated initially by a tiny vector $\boldsymbol\delta_0$ . In a chaotic system the separation $|\boldsymbol\delta(t)|$ grows, on average, like an exponential:

\boxed{\; |\boldsymbol\delta(t)| \;\approx\; |\boldsymbol\delta_0|\, e^{\lambda t}. \;}

where

$\boldsymbol\delta(t)$: separation between two nearby trajectories at time t
$\boldsymbol\delta_0$: initial (infinitesimal) separation
$\lambda$: largest Lyapunov exponent — the average exponential separation rate 1/time

Solving for the rate and taking the limits that make it intrinsic:

\lambda \;=\; \lim_{t\to\infty} \lim_{|\boldsymbol\delta_0|\to 0} \frac{1}{t} \ln \frac{|\boldsymbol\delta(t)|}{|\boldsymbol\delta_0|}.

The order of limits matters: $|\boldsymbol\delta_0|\to 0$ first keeps the separation infinitesimal so the growth stays in the linear regime, and $t\to\infty$ then averages the rate over the whole attractor. An $n$ -dimensional system has $n$ Lyapunov exponents, one per direction of stretching or squeezing (the Lyapunov spectrum); “the” Lyapunov exponent means the largest, since it dominates a generic separation after a short time.

The sign of $\lambda$ classifies the dynamics completely:

$\lambda < 0$ — neighbours converge; the system forgets perturbations and is predictable (fixed points, stable cycles).
$\lambda = 0$ — neighbours hold their separation; marginal (the neutral direction along any flow’s trajectory always contributes a zero exponent).
$\lambda > 0$ — neighbours diverge exponentially; chaos.

For the Lorenz attractor of the previous lesson, the three exponents are approximately $(+0.906,\; 0,\; -14.57)$ per unit time: one positive (the chaos), one zero (the flow direction), one strongly negative (the volume-contracting squeeze). Their sum, $-13.67$ , is exactly the divergence $-(\sigma+1+\beta)$ we computed — the Lyapunov spectrum must sum to the average volume-contraction rate, a useful consistency check.

Watching the separation grow

prediction horizon ≈ 18.6 time units

initial separation δ₀ = 1e-6

Make δ₀ a thousand times smaller and the horizon barely moves — it grows only as log(1/δ₀). That is why long-range forecasting is hopeless even with near-perfect data.

Two Lorenz trajectories start a distance $\delta_0$ apart. The main panel plots $\log_{10}|\boldsymbol\delta(t)|$ , and the lower panel overlays the two $x(t)$ traces. Three things to read off:

The straight line on the log plot is the exponent. Exponential growth $e^{\lambda t}$ is a straight line in a log plot, with slope $\lambda/\ln 10$ . The curve hugs the dashed reference line of slope $0.906/\ln 10$ until it saturates.
Saturation is geometric, not dynamical. The separation cannot exceed the diameter of the attractor — once the two trajectories are on opposite wings, they are as far apart as they can get, and the log curve flattens. Exponential growth is only the early story.
The traces tell the human version. $x_A(t)$ and $x_B(t)$ sit exactly on top of each other, then peel apart at one specific moment and thereafter share nothing. That moment is the prediction horizon.

The prediction horizon, and why precision barely helps

Here is the practical payload of the whole chapter. Suppose you know the present state to within $\delta_0$ and you are willing to tolerate a forecast error up to some fixed size $L$ — the scale at which “the forecast is now useless,” roughly the size of the attractor. How far ahead can you predict?

▶ The horizon grows only as the logarithm of precision Derivation

The forecast stays useful until the separation, growing as $\delta_0 e^{\lambda t}$ , reaches the tolerance $L$ :

\delta_0\, e^{\lambda t_{\text{horizon}}} \;=\; L.

Solve for the time:

\boxed{\; t_{\text{horizon}} \;=\; \frac{1}{\lambda} \ln\!\frac{L}{\delta_0}. \;}

where

$t_{\text{horizon}}$: time until the forecast error reaches the tolerance
$L$: tolerable error / attractor size
$\delta_0$: uncertainty in the initial state
$\lambda$: largest Lyapunov exponent

Now read what the logarithm does. To double the horizon you must push $\ln(L/\delta_0)$ to twice its value, which means making $\delta_0$ smaller by the square root — an enormous improvement in measurement for a merely linear gain in lead time. Improving $\delta_0$ by a factor of 1000 (three more decimal digits on every sensor) buys only an extra $\ln(1000)/\lambda = 6.9/\lambda$ of time — a fixed additive amount, not a multiplicative one. Each additional good measurement digit extends the forecast by the same small constant. The exponential in the dynamics becomes a logarithm in what you can do about it, and the logarithm is merciless.

Put numbers to it. The atmosphere’s largest Lyapunov exponent corresponds to error roughly doubling every $\sim 1.5$ days. Halving today’s observational uncertainty extends useful forecasts by only that same $\sim 1.5$ days; a thousandfold better global observing system would add only about $\ln(1000)/\!\ln 2 \approx 10$ doubling times, i.e. a couple of weeks. This is the rigorous content of the butterfly: not that weather is random — it is fully deterministic — but that its positive Lyapunov exponent caps deterministic weather prediction at around two weeks, and no conceivable improvement in data or computing power moves that wall by much. The slider above makes the point directly: drive $\delta_0$ down by orders of magnitude and the horizon creeps up by almost nothing.

This is the sharp distinction between chaos and the randomness of Brownian motion. A random walk is unpredictable because it is stochastic — there is no rule to know. A chaotic system is unpredictable despite being deterministic — the rule is exact and known, but reading the future from it requires impossibly precise knowledge of the present. Both defeat prediction; only one contains any actual randomness.

⏳ The history — Lyapunov's stability, repurposed for instability

Aleksandr Lyapunov introduced his exponents in his 1892 doctoral thesis The General Problem of the Stability of Motion, written in Kharkiv and concerned with the opposite of chaos: he wanted rigorous conditions under which a mechanical equilibrium is stable, so that perturbations decay. His “first method” linearised the dynamics and read stability from the exponential rates — negative rates meaning a perturbation dies away. The same exponents, taken positive, became the central diagnostic of chaos eighty years later, once Lorenz and others realised that the interesting systems were the ones where Lyapunov’s quantity comes out positive. The numerical recipe for extracting the full spectrum from data or simulation — periodically renormalising the separation vector to keep it infinitesimal — was worked out by Benettin and others around 1980 and is still the standard method.

What we use this for

These ideas reach into the other books wherever a system is nonlinear enough to misbehave:

The onset of turbulence. The Lorenz model is a stripped-down convecting fluid; real turbulence is its high-dimensional descendant. The transition from smooth (laminar) to turbulent flow in the fluid-mechanics chapter, past a critical Reynolds number, is a route into chaos with infinitely many active modes.
Nonlinear acoustics. When sound amplitude grows large the small-perturbation theory breaks and the governing equations turn genuinely nonlinear; strongly driven nonlinear oscillators period-double and can become chaotic, the same cascade seen here.
Bubble dynamics. A bubble driven hard by an acoustic field obeys the nonlinear Rayleigh–Plesset equation; under strong driving its radial oscillations period-double and go chaotic, and the resulting “chaotic” bubble response is observed in cavitation experiments.
The limits of simulation. The numerical ODE solvers we use to integrate any of these flows accumulate roundoff that, in a chaotic system, is amplified exactly like any other perturbation — so a long simulated chaotic trajectory is never the “true” one, only a representative orbit on the same attractor. Knowing $\lambda$ tells you how long a computed trajectory can be trusted pointwise.

That is the chapter. Determinism fixes the future; a positive Lyapunov exponent hides it. Chaos is the precise statement that those two facts are compatible — and the bifurcation diagram, the strange attractor, and the prediction horizon are its three faces.