Lyapunov exponents and the horizon of prediction

For a differentiable map, a small displacement evolves approximately by the derivative evaluated at the current orbit point. After several steps, the relevant factor is a product of derivatives, not a derivative evaluated once at the starting point. Taking logarithms turns this product into an average of local logarithmic stretching. When a suitable long-time limit exists, it defines an exponent for that orbit. Different invariant sets or exceptional orbits of the same system can have different exponents.

For a multidimensional flow, evolve the variational equation along the trajectory and track independent perturbation directions. An exponent has units of inverse time when time is dimensional, whereas a map exponent is commonly quoted per iteration. Smooth coordinate changes whose derivatives and inverse derivatives stay uniformly bounded along the orbit preserve asymptotic exponents, but finite-time values and their numerical conditioning can change. Always state the clock and norm when reporting a finite-time sensitivity estimate or comparing two calculations.

While a perturbation remains in the linear regime, a positive growth rate suggests an uncertainty envelope proportional to $e^{\lambda t}$. Set a tolerable error and solve for the time at which the envelope reaches it. Reducing initial uncertainty buys time only logarithmically. Improving precision by a factor of ten adds a fixed interval determined by the growth rate, rather than multiplying the forecast horizon by ten. This is a statement about state prediction, not about the inability to predict any statistic.

A chaotic system can still have reproducible distributions, average fluxes, or other coarse observables. Weather-like trajectory uncertainty and climate-like statistical regularity are conceptually compatible, although proving either for a particular model requires additional work. Model error and noisy forcing also enter a forecast; an initial-condition exponent alone does not capture them. Once the perturbation reaches the attractor's size, the exponential separation law saturates and cannot be extrapolated to physically impossible distances.

A positive finite-time exponent can arise during a transient near an unstable state before a trajectory settles to a stable attractor. An unbounded equation such as a scalar exponentially growing mode also separates trajectories without a bounded recurrent chaotic attractor. A persuasive chaos claim therefore combines sustained positive stretching with the relevant bounded invariant dynamics and checks appropriate to the system. A single number computed over a short interval is a diagnostic, not a complete dynamical classification.

Numerically, two ordinary trajectories eventually separate too far for their difference to represent a tangent vector, while making the initial difference too small invites roundoff. Evolving the variational equation and periodically renormalizing tangent vectors avoids that particular saturation problem. Multiple directions need orthogonalization to recover a spectrum. Repeat the calculation with longer observation, discarded transients, and numerical refinement. Convergence evidence does not prove every mathematical limit, but it makes the claimed finite-time result reproducible and exposes obvious numerical artifacts.