Deterministic chaos and Lyapunov exponents
Chaos is what a fully deterministic system does when nearby states pull apart exponentially fast: no randomness anywhere in the equations, yet two starting conditions that differ in the tenth decimal place end up on entirely different trajectories. The Lyapunov exponent measures that rate — a small separation grows roughly as an exponential of the exponent times elapsed time, so its reciprocal sets the time scale on which prediction decays — and a positive exponent together with motion that stays bounded is what marks genuine chaos rather than a system simply running away, which is why weather has a forecast horizon of days no matter how good the instruments get.
Practise this conceptRun a chaotic system twice from exactly the same starting state and you get exactly the same trajectory: nothing in the equations is random. Run it from two states that differ by a millionth of a millimetre and the two agree for a while, then drift apart until they have nothing to do with each other. That is the whole of chaos — deterministic rules, plus a separation between neighbours that grows exponentially rather than steadily.
The Lyapunov exponent is the growth rate in that exponential: a small initial separation is multiplied by e raised to the exponent times the elapsed time, so the reciprocal of the exponent is roughly how long a small error takes to triple. The sign is what matters most. Negative means neighbouring states converge and the motion is predictable; zero means they drift apart only in proportion to time; positive means the system forgets its initial condition at a definite rate. Two details keep the definition honest. The exponent is an average over the attractor, because the stretching rate differs from point to point on it, and the exponential law describes only the early growth — once the separation is as large as the attractor itself, the two trajectories are just two unrelated points on the same object and the gap cannot widen further.
That bound is exactly what separates chaos from ordinary instability. A ball balanced on a hilltop also has a positive exponent, but it rolls away and never returns, while a chaotic system keeps folding back on itself, stretching along one direction and squeezing along another — which is what builds the layered structure of a strange attractor. The practical consequence is the forecast horizon. Errors compound multiplicatively while precision is bought additively: measuring the initial state a hundred times better extends the useful forecast by the natural logarithm of a hundred divided by the exponent, a fixed number of days rather than a hundredfold gain. That is why a weather model whose exponent is about half a day inverse gains a week or so of skill from a vastly better observing network, and why no instrument will ever deliver a useful three-month forecast.
Key ideas
Requires: Bifurcations and the loss of stability
Formulas
| Symbol | Name | Unit |
|---|---|---|
| distance | m (metre) | |
| distance | m (metre) | |
| Lyapunov exponent | 1/s (per second) | |
| time | s (second) |
Common misconceptions
- Equating chaos with randomness. A chaotic system is deterministic: start it twice from exactly the same state and it repeats itself exactly, and there is no noise term anywhere in the equations. What is unpredictable is the consequence of the state never being known exactly — the sensitivity is to the initial condition, not to hidden dice.
- Taking a positive Lyapunov exponent to mean the motion itself grows without bound. What separates exponentially is the gap between two nearby trajectories, not the trajectories themselves: chaotic motion stays on a bounded attractor, and the exponential law describes the early growth of a small separation, stopping once that gap is comparable to the attractor's own size. A ball rolling off a hilltop also has a positive exponent and is not chaotic, precisely because it never comes back.
- Assuming better initial data buys proportionally more prediction. Because the error grows exponentially, measuring the starting state a hundred times more precisely does not multiply the useful forecast by a hundred — it adds a fixed amount of time, the natural logarithm of a hundred divided by the Lyapunov exponent. Precision buys time additively while the error compounds multiplicatively, which is why forecast horizons creep forward by days over decades of instrument improvement.
Resources
Part of these UT courses
This concept is mapped as covering part of the following course(s) in the University of Tartu curriculum.
- Nonlinear dynamicsLOFY.04.070