Theory

Measures & invariants of chaos

Numerical invariants distinguish chaos from periodicity from noise and quantify it. The following are the most useful, with brief definitions and pointers to the tools that compute them on this site.

Lyapunov exponents

Quantify the exponential rate at which infinitesimally close trajectories diverge. For an iterated map x_{n+1} = f(x_n),

λ = lim_{N→∞}  (1/N) · Σ_{i=0}^{N−1} log |f'(x_i)|

A positive value is the defining quantitative signature of chaos. For continuous flows, the leading Lyapunov exponent λ₁ is computed by integrating the linearised dynamics along the trajectory, re-normalising at every step (Benettin et al. 1980, Wolf et al. 1985). The full Lyapunov spectrum λ₁ ≥ λ₂ ≥ … ≥ λ_n contains as many exponents as the phase-space dimension.

Try /tools/lyapunov for a live λ(parameter) chart on 1D maps.

λ(r) = lim_{N→∞} (1/N) Σ_{i=0}^{N−1} log |f'(x_i)|

orange segments: λ > 0  →  chaotic
grey segments:   λ ≤ 0  →  stable periodic

Kaplan-Yorke (Lyapunov) dimension

D_KY = k + (λ_1 + ... + λ_k) / |λ_{k+1}|

where k = largest integer with λ_1 + ... + λ_k ≥ 0

Estimates the Hausdorff dimension of the attractor from the Lyapunov spectrum. For Lorenz at classical parameters: λ ≈ (0.906, 0, −14.572), giving D_KY ≈ 2.062. Cheap to compute, often accurate.

Box-counting dimension

D_0 = lim_{ε → 0}  log N(ε) / log(1/ε)

Cover the attractor with axis-aligned boxes of side ε and count how many are non-empty. Equivalently the slope on a log-log plot. The easiest dimension to estimate from numerical data.

Correlation dimension (Grassberger-Procaccia)

D_2 = lim_{ε → 0}  log C(ε) / log ε

C(ε) = (1/N²) · |{(i, j) : ||x_i − x_j|| < ε}|

Standard estimator from time-series data after Takens-embedding the scalar series. Grassberger and Procaccia (1983).

Topological & Kolmogorov-Sinai entropy

Information-theoretic measures of how fast trajectories diverge in a partition-coded representation. For 1D maps with absolutely continuous invariant measure, K-S entropy = λ⁺(Pesin’s formula).

Sample, approximate, permutation entropy

Robust to noise and short series. Bandt-Pompe permutation entropy (2002) became a standard quick chaos detector: it counts the relative frequency of ordinal patterns of length m in the time series.

0-1 test for chaos (Gottwald-Melbourne)

Maps a scalar deterministic time series x_n to a 2D random-walk-like process p, q via p_{n+1} = p_n + φ_n cos(c n). The asymptotic mean-square displacement scales linearly with time for chaos (output ≈ 1) and remains bounded for regular dynamics (output ≈ 0). Single scalar verdict, no embedding needed.

Recurrence plots & RQA

Plot 1 at (i, j) when the embedded trajectory at time i is within ε of the one at time j; 0 otherwise. The visual texture encodes determinism. Recurrence quantification analysis extracts numerical measures (recurrence rate, determinism, laminarity, divergence) from this binary matrix (Marwan et al. review, Phys. Rep. 2007).

R[i,j] = 1  iff  ‖x_i − x_j‖ < ε  (embedded coordinates, m = 3, τ = 4)

texture readings:
  diagonal lines          deterministic, recurrent structure
  isolated short lines    chaotic regime
  vertical clusters       laminar / intermittent phases
  uniform noise           stochastic process

Power spectrum

The classic frequency-domain signature. Periodic dynamics → discrete harmonic spikes. Quasi-periodic → spikes at incommensurate frequencies and combinations. Chaos → broadband, often with broad bumps; noise → flat (white) or 1/f^α (coloured).

periodic       → spikes at harmonic frequencies, dynamic range > 60 dB
quasi-periodic → spikes at incommensurate frequencies and their integer combinations
chaotic        → broadband (continuous) spectrum, may include broad peaks
noise          → flat spectrum (white) or 1/f^α (coloured)

Embedding theory

Takens’ theorem (1981): for a generic smooth observable on a finite-dimensional attractor of box-counting dimension D, the time-delay embedding (x(t), x(t − τ), …, x(t − (m−1)τ)) with m > 2D is a diffeomorphic image of the attractor. Practical algorithms: average-mutual-information to choose τ (Fraser-Swinney 1986), false-nearest-neighbours to choose m (Kennel et al. 1992).

Routes to chaos

• Period-doubling cascade (Feigenbaum)
• Quasi-periodic (Ruelle-Takens-Newhouse)
• Intermittency (Pomeau-Manneville types I, II, III)
• Crisis-induced transitions
• Chaotic transients

For more, see /bifurcations and /learn.