mathematics//dynamical systems//nonlinear dynamics//chaos
Chaos is the behaviour of a deterministic dynamical system whose trajectories stay bounded yet separate exponentially from nearly identical starts, so that exact rules give no long-term prediction. It matters to engineers as a hard limit: it caps how far ahead a model can forecast the weather, a turbulent flow or a double pendulum, whatever the computer, and it tells an estimator that the model alone will not hold. There is no randomness in a chaotic system; started from exactly the same state it would do exactly the same thing. The trouble is that the state is never known exactly, and the error grows as
Chaos is the behaviour of a deterministic dynamical system whose trajectories stay bounded yet separate exponentially from nearly identical starts, so that exact rules give no long-term prediction. It matters to engineers as a hard limit: it caps how far ahead a model can forecast the weather, a turbulent flow or a double pendulum, whatever the computer, and it tells an estimator that the model alone will not hold. There is no randomness in a chaotic system; started from exactly the same state it would do exactly the same thing. The trouble is that the state is never known exactly, and the error grows as
∥δ(t)∥≈∥δ0∥ eλ1t,Tp≈1λ1lnΔ∥δ0∥.\|\delta(t)\|\approx\|\delta_0\|\,e^{\lambda_1 t},\qquad T_p\approx\frac{1}{\lambda_1}\ln\frac{\Delta}{\|\delta_0\|}.∥δ(t)∥≈∥δ0∥eλ1t,Tp≈λ11ln∥δ0∥Δ.
δ0\delta_0δ0 is the initial difference, λ1\lambda_1λ1 the largest Lyapunov exponent (the average rate at which nearby trajectories separate) and TpT_pTp the prediction horizon, the time until the separation reaches a tolerance Δ\DeltaΔ.
The logarithm is the bad news. Measuring the initial state a thousand times better extends the horizon by only ln(1000)/λ1≈6.9/λ1\ln(1000)/\lambda_1\approx6.9/\lambda_1ln(1000)/λ1≈6.9/λ1: every factor of ten in precision buys the same handful of seconds. That is why forecasts have a useful horizon of days and not months, and more computers do not change it.
initial separation10⁻⁶ rad horizon, 0.1 rad7.5 s divergence rate λ1.4 1/s Two double pendulums started at 2.00 and 2.40 rad. Released 10⁻⁶ rad apart, they are 0.1 rad apart after 7.5 s; on the log scale the separation climbs as a straight line, at about 1.4 per second. Starting a thousand times closer (10⁻⁹ rad) only stretches that to 9.7 s.
Drag the initial separation from 10−310^{-3}10−3 down to 10−910^{-9}10−9 rad, a million times more precise, and watch the horizon grow by adding seconds rather than multiplying; switch on small angles and the separation stops growing.
Noise, chaos and instability are three different things. Noise is random and has no rule to predict it, only statistics. Chaos is deterministic, predictable in the short term and only statistically in the long term. A linearly unstable system diverges to infinity, and predictably so, while a chaotic one stays inside a bounded region (equilibrium and stability).
It needs nonlinearity and room. A continuous autonomous system needs at least three state variables to be chaotic, so a pendulum is regular and a double pendulum is the classic chaotic machine; a discrete map can be chaotic with a single state. Most nonlinear machines are not chaotic at all (nonlinear dynamics).
A Lyapunov exponent belongs to a trajectory, averaged over a long time; it is a different object from the eigenvalue of a Jacobian at one point, and from a Lyapunov function, which proves stability.
What engineering does about it is to predict statistics and keep correcting. A fleet planner simulates many scenarios rather than one trajectory (Monte Carlo method), and an estimator feeds the model measurements often enough that the error never has time to grow (complex system).