09.08.05 · classical-mech / chaos

Strange attractors: the Lorenz system, fractal dimension, and sensitive dependence

shipped3 tiersLean: none

Anchor (Master): Tucker, The Lorenz attractor exists (1999); Sparrow, The Lorenz Equations: Bifurcations, Chaos, and Strange Attractors (1982)

Intuition Beginner

A strange attractor is a geometric object in phase space that attracts all nearby trajectories but has a fractal (non-integer) structure. It is "strange" because it is neither a point, nor a curve, nor a surface — it has a dimension that is not a whole number.

Edward Lorenz discovered the first strange attractor in 1963 while modelling atmospheric convection. His equations are three coupled ODEs with only three variables and three parameters, yet they produce infinitely complex, non-repeating behaviour. Two simulations started with initial conditions differing by one part in a million diverge completely within a few "weather cycles." This is the butterfly effect: the flap of a butterfly's wing in Brazil could set off a tornado in Texas.

The Lorenz attractor looks like the wings of a butterfly. Trajectories spiral outward on one wing, then suddenly jump to the other wing and spiral there, then jump back. The jumps are irregular — no pattern repeats — yet the overall shape of the attractor is always the same: all trajectories are confined to a thin, fractal shell with a dimension of approximately 2.06.

Visual Beginner

Feature	Value (Lorenz system)	Meaning
Largest Lyapunov exponent $λ_{1}$	$\approx 0.906$	Nearby trajectories diverge exponentially
Kaplan-Yorke dimension $d_{K Y}$	$\approx 2.06$	Attractor is slightly thicker than a surface
Correlation dimension $d_{c}$	$\approx 2.05$	Confirms fractal structure
Dissipation rate	$λ_{1} + λ_{2} + λ_{3} = - 13.66$	Phase-space volume contracts
Predictability time	$\approx 1/ λ_{1} \approx 1$ time unit	Forecasts diverge in ~1 unit

Worked example Beginner

The Lorenz equations are three simple ODEs:

\overset{x}{˙} = σ (y - x), \overset{y}{˙} = x (ρ - z) - y, \overset{z}{˙} = x y - β z

with the classical parameter values $σ = 10$ , $ρ = 28$ , $β = 8/3$ .

The equilibrium points are: the origin $C^{+} = (0, 0, 0)$ and two symmetric points $C^{\pm} = (\pm β (ρ - 1), \pm β (ρ - 1), ρ - 1)$ . For $ρ = 28$ : $C^{\pm} = (\pm 8.485, \pm 8.485, 27)$ .

All three equilibria are unstable (saddle points with both stable and unstable directions). The trajectory cannot settle to any equilibrium, so it wanders forever on the attractor.

The rate of phase-space volume contraction is $div (f) = - σ - 1 - β = - 10 - 1 - 8/3 = - 41/3 \approx - 13.67$ . This means volumes shrink exponentially: $V (t) = V_{0} e^{- 41 t /3}$ . The attractor has zero volume but non-zero fractal dimension.

Check your understanding Beginner

Formal definition Intermediate+

The Lorenz system is the autonomous ODE system in $R^{3}$ :

\overset{x}{˙} = σ (y - x), \overset{y}{˙} = r x - y - x z, \overset{z}{˙} = x y - b z

with parameters $σ > 0$ (Prandtl number), $r > 0$ (normalised Rayleigh number), $b > 0$ (geometric factor). The system is symmetric under $(x, y) \to (- x, - y)$ .

An attractor for a flow $ϕ^{t}$ on $R^{n}$ is a compact invariant set $A$ that attracts all trajectories starting in a neighbourhood $U$ : $dist (ϕ^{t} (x), A) \to 0$ as $t \to \infty$ for all $x \in U$ .

A strange attractor is an attractor that (1) has sensitive dependence on initial conditions (positive largest Lyapunov exponent) and (2) has a fractal (non-integer) geometric structure.

Fractal dimensions. Several dimensions characterise the geometry of a strange attractor:

Hausdorff dimension $d_{H}$ : the unique value where the Hausdorff measure transitions from infinite to zero.
Box-counting dimension $d_{B} = lim_{ϵ \to 0} ln N (ϵ) / ln (1/ ϵ)$ where $N (ϵ)$ is the number of $ϵ$ -boxes needed to cover the set.
Information dimension $d_{I}$ : related to the rate at which information is needed to specify a point on the attractor.
Correlation dimension $d_{C}$ : defined via the correlation sum $C (ϵ) = lim_{N \to \infty} \frac{1}{N ^{2}} \sum_{i, j} Θ (ϵ - ∣ x_{i} - x_{j} ∣)$ , with $d_{C} = lim_{ϵ \to 0} d ln C (ϵ) / d ln ϵ$ .
Kaplan-Yorke dimension $d_{K Y} = j + \frac{\sum _{i = 1}^{j} λ _{i}}{∣ λ _{j + 1} ∣}$ where $j$ is the largest integer with $\sum_{i = 1}^{j} λ_{i} \geq 0$ .

For "nice" attractors, these dimensions satisfy $d_{H} \leq d_{B}$ and $d_{I} \leq d_{H}$ , and are often approximately equal.

Key derivation Intermediate+

Derivation (Sensitive dependence and fractal dimension of the Lorenz attractor).

Step 1: Dissipation and volume contraction. The divergence of the Lorenz vector field is $\frac{\partial x ˙}{\partial x} + \frac{\partial y ˙}{\partial y} + \frac{\partial z ˙}{\partial z} = - σ - 1 - b < 0$ . By Liouville's theorem for dissipative systems, any volume element $V (t)$ evolves as $V (t) = V_{0} e^{- (σ + 1 + b) t}$ . Volumes contract exponentially fast. Any attractor has zero volume.

Step 2: Stretching and folding. Near the origin (the saddle point $C^{0} = 0$ ), the linearised system has eigenvalues (for $σ = 10$ , $r = 28$ , $b = 8/3$ ): one real positive eigenvalue $λ_{u} \approx 11.83$ and two real negative eigenvalues $λ_{s 1} \approx - 22.83$ , $λ_{s 2} \approx - 2.67$ . Trajectories approaching the origin along the stable directions are stretched along the unstable direction and ejected toward one of the two symmetric equilibria $C^{\pm}$ . This is the stretch-and-fold mechanism.

Step 3: The Lorenz map. The dynamics can be reduced to a 1D map. Record the local maxima of $z (t)$ : $z_{1}, z_{2}, z_{3}, \dots$ . The map $z_{n} \mapsto z_{n + 1}$ is a tent-shaped map on an interval. This map is expanding (slope greater than 1 in absolute value), which gives positive Lyapunov exponent $λ_{1} > 0$ and proves sensitive dependence.

Step 4: Fractal structure. The Lorenz attractor is created by the repeated stretching (in the unstable direction) and folding (the jumps between the two wings). Each iteration of the stretch-fold process multiplies the number of layers. After infinitely many iterations, the cross-section of the attractor becomes a Cantor-like set. The attractor is locally the product of a Cantor set (in the cross-section) and a curve (along the flow), giving a dimension $d \approx 2 + ϵ$ where $ϵ$ is the fractal dimension of the Cantor cross-section.

Step 5: Kaplan-Yorke dimension. For $σ = 10$ , $r = 28$ , $b = 8/3$ , the Lyapunov spectrum is $λ_{1} \approx 0.906$ , $λ_{2} = 0$ , $λ_{3} \approx - 14.57$ . The Kaplan-Yorke dimension is:

d_{K Y} = 2 + \frac{λ _{1} + λ _{2}}{∣ λ _{3} ∣} = 2 + \frac{0.906}{14.57} \approx 2.062

The attractor is a fractal with dimension approximately 2.06 — a thin, layered surface in 3D phase space.

Bridge. The Lorenz attractor is the paradigmatic example of a strange attractor in a dissipative system. The foundational result is Tucker's 1999 proof (computer-assisted, using interval arithmetic) that the Lorenz attractor exists and satisfies the axioms of a geometric Lorenz model — a rigorous proof that the numerically observed attractor is indeed a strange attractor with a fractal structure. The central insight is that the combination of volume contraction (forcing the attractor to be "thin") and exponential stretching (forcing the attractor to be "layered") produces a fractal invariant set with non-integer dimension. This generalises the horseshoe picture 09.08.02 from Hamiltonian systems (where the invariant set has zero volume but positive measure in a codimension-1 section) to dissipative systems (where the attractor has zero volume but non-zero fractal dimension). The Kaplan-Yorke dimension formula provides the bridge between the Lyapunov spectrum 09.08.03 (a dynamical quantity) and the fractal dimension (a geometric quantity), connecting dynamics to geometry.

Exercises Intermediate+

Exercise 4 (medium, symbolic).

Define the correlation dimension $d_{C}$ of a set of points ${x_{i}}_{i = 1}^{N}$ on an attractor. Explain how it is computed from a time series.

Hint

The correlation sum $C (ϵ)$ counts the fraction of pairs of points closer than $ϵ$ . The correlation dimension is the slope of $ln C$ vs $ln ϵ$ .

Answer

The correlation sum is $C (ϵ) = \frac{2}{N ( N - 1 )} \sum_{i < j} Θ (ϵ - ∥ x_{i} - x_{j} ∥)$ where $Θ$ is the Heaviside step function. The correlation dimension is $d_{C} = lim_{ϵ \to 0} \frac{l n C ( ϵ )}{l n ϵ}$ , which is the slope of the log-log plot of $C$ vs $ϵ$ in the scaling region (the range of $ϵ$ where the slope is approximately constant).

For a time series $s (t_{k})$ , the attractor is first reconstructed by time-delay embedding: form vectors $x_{k} = (s (t_{k}), s (t_{k} + τ), \dots, s (t_{k} + (m - 1) τ))$ where $τ$ is a delay and $m$ is the embedding dimension. Then compute $C (ϵ)$ and extract $d_{C}$ from the scaling region. For the Lorenz attractor: $d_{C} \approx 2.05$ , consistent with $d_{K Y} \approx 2.06$ .

Exercise 5 (hard, symbolic).

Describe the geometric Lorenz model and explain how it captures the essential dynamics of the Lorenz system.

Hint

The geometric model replaces the ODE flow with a return map on a Poincare section. The map has two branches (corresponding to the two wings) and a discontinuity (corresponding to the stable manifold of the origin).

Answer

The geometric Lorenz model (Guckenheimer-Williams 1979, Afraimovich-Bykov-Shilnikov 1977) is an abstract model that captures the topological dynamics of the Lorenz attractor. It consists of:

A Poincare section $Σ$ (a rectangle transverse to the flow) divided into two halves $Σ_{L}$ and $Σ_{R}$ by the stable manifold of the origin $W^{s} (0)$ .
A return map $P : Σ \to Σ$ with the following properties: (a) $P$ maps $Σ_{L}$ and $Σ_{R}$ each to a thin "wedge" across $Σ$ ; (b) near the dividing line (the stable manifold), points are strongly stretched horizontally and compressed vertically (by a factor related to $λ_{u} / λ_{s 1}$ ); (c) the image of each half of $Σ$ crosses the dividing line, so orbits can switch wings.
The essential feature is the hyperbolic splitting: in the contracting direction, the map is uniformly contracting (factor $< 1$ ), and in the expanding direction, it is uniformly expanding (factor $> 1$ ). This makes the attractor a singular hyperbolic set.
The attractor contains the saddle equilibrium at the origin, which is a singular point (not part of any regular orbit). The attractor is not uniformly hyperbolic due to this singularity, but it satisfies the weaker condition of sectional hyperbolicity.

Tucker (1999) proved that the actual Lorenz system (for the classical parameters) is topologically conjugate to the geometric Lorenz model, confirming that the Lorenz attractor is a genuine strange attractor.

Exercise 6 (hard, symbolic).

Explain the Lorenz map (the 1D return map $z_{n} \mapsto z_{n + 1}$ of successive maxima of $z (t)$ ). Why is this map expanding, and what does this imply about the dynamics?

Hint

Record the local maxima of $z (t)$ and plot each maximum against the next. The resulting map looks like a tent map with slope magnitude greater than 1.

Answer

The Lorenz map is constructed by recording the local maxima $z_{1}, z_{2}, z_{3}, \dots$ of the $z$ -coordinate along a trajectory, and plotting the map $f : z_{n} \mapsto z_{n + 1}$ . Numerically, the map has the shape of a smooth tent: it increases, reaches a maximum near $z \approx z_{0}$ (the $z$ -value of the unstable equilibria $C^{\pm}$ ), then decreases, with a cusp-like peak.

The map is expanding: the slope $∣ f^{'} (z) ∣ > 1$ everywhere (except near the cusp, where it has a discontinuity). This follows from the stretching in the unstable direction near the origin: each passage near the origin stretches the $z$ -coordinate of the trajectory, and the folding (switching between wings) creates the tent shape. The expansion means the map has a positive Lyapunov exponent: $λ = lim_{N \to \infty} \frac{1}{N} \sum_{n = 0}^{N - 1} ln ∣ f^{'} (z_{n}) ∣ > 0$ . This provides a simple 1D proof of sensitive dependence for the Lorenz system.

The Lorenz map is not rigorous (it is an approximation that works well for the classical parameters but breaks down for other parameter values where the attractor may have a different structure). However, it provides excellent intuition for why the Lorenz attractor is chaotic.

Lean formalization Intermediate+

Mathlib has Hausdorff dimension theory and ODE existence theorems but not the Lorenz attractor, Tucker's proof, the geometric Lorenz model, or the Kaplan-Yorke conjecture. Tucker's proof requires verified numerical computation (interval arithmetic with rigorous error bounds), which is an active research area in formal mathematics but not yet in Mathlib. lean_status: none.

Advanced results Master

Tucker's theorem (1999). Warwick Tucker proved that the Lorenz equations (for the classical parameters $σ = 10$ , $r = 28$ , $b = 8/3$ ) possess a strange attractor. The proof is computer-assisted: it combines normal form theory (to rigorously bound the dynamics near the origin) with interval arithmetic (to rigorously propagate the flow and verify the geometric Lorenz conditions). The result confirms that the numerically observed attractor is not a long transient or a numerical artifact but a genuine mathematical object with sensitive dependence and fractal structure.

SRB measures. The Lorenz attractor supports a Sinai-Ruelle-Bowen (SRB) measure — an invariant measure that describes the statistical properties of typical trajectories. For dissipative systems, the SRB measure is the "physical measure": time averages along almost every trajectory converge to the SRB average. The existence of the SRB measure for the Lorenz attractor follows from Tucker's proof, which establishes the required hyperbolicity properties.

The relation to Rayleigh-Benard convection. The Lorenz equations are a severe truncation of the Boussinesq equations for Rayleigh-Benard convection (fluid heated from below). Saltzman (1962) derived the equations by expanding the velocity and temperature fields in Fourier modes and keeping only three modes. Lorenz then showed that even this drastically simplified model exhibits chaos, implying that the full weather system must also be chaotic.

The dimension spectrum. The generalised dimensions $D_{q}$ (for $q = 0, 1, 2, \dots$ ) form a spectrum: $D_{0}$ is the box-counting dimension, $D_{1}$ is the information dimension, and $D_{2}$ is the correlation dimension. For multifractal attractors, $D_{q}$ decreases with $q$ , reflecting the fact that the attractor has a range of local scaling behaviours. The Lorenz attractor has a relatively narrow dimension spectrum ( $D_{0} \approx 2.06$ , $D_{1} \approx 2.06$ , $D_{2} \approx 2.05$ ), indicating near-uniform scaling.

Synthesis. The strange attractor is the third pillar of chaos theory, completing the geometric picture (horseshoes 09.08.02), the quantitative picture (Lyapunov exponents 09.08.03), and the phenomenological picture (period-doubling 09.08.04). The foundational result is Tucker's proof that the Lorenz attractor is a genuine strange attractor, confirming six decades of numerical evidence. The central insight is that dissipative chaos produces fractal attractors: the combination of stretching (creating layers) and volume contraction (compressing layers into thin sheets) generates invariant sets with non-integer dimension. The Kaplan-Yorke formula provides the precise link between the dynamical quantities (Lyapunov exponents) and the geometric quantities (fractal dimensions), making the strange attractor a computable and characterisable object. Putting these together, the Lorenz attractor is the prototype for chaotic dynamics in nature: weather, turbulence, lasers, chemical reactions, and biological oscillators all exhibit strange attractors with the same qualitative features — exponential divergence, fractal geometry, and statistical predictability despite deterministic unpredictability.

Full proof set Master

Proposition (Volume contraction in the Lorenz system). Any region $U \subset R^{3}$ evolves under the Lorenz flow to a set of zero volume.

Proof. The Lorenz vector field $f = (σ (y - x), r x - y - x z, x y - b z)$ has divergence $\nabla \cdot f = - σ - 1 - b < 0$ . By the transport theorem, the volume of any evolving region $V (t) = ϕ^{t} (U)$ satisfies $d V / d t = \int_{V} (\nabla \cdot f) d V = - (σ + 1 + b) V$ . So $V (t) = V (0) e^{- (σ + 1 + b) t}$ . As $t \to \infty$ , $V (t) \to 0$ . The attractor has zero volume. $□$

Proposition (Existence of a trapping region). The Lorenz system has a globally attracting set: all trajectories eventually enter a bounded ellipsoid and never leave.

Proof. Define the Lyapunov function $V = x^{2} + y^{2} + (z - r - σ)^{2}$ . Its time derivative is $\dot{V} = 2 x \overset{x}{˙} + 2 y \overset{y}{˙} + 2 (z - r - σ) \overset{z}{˙}$ . Substituting the Lorenz equations: $\dot{V} = 2 σ x (y - x) + 2 y (r x - y - x z) + 2 (z - r - σ) (x y - b z)$ . Expanding and simplifying: $\dot{V} = - 2 σ x^{2} - 2 y^{2} - 2 b z^{2} + 2 b (r + σ) z$ . This is negative for large $∣ x ∣, ∣ y ∣, ∣ z ∣$ . Specifically, for $z$ large: $- 2 b z^{2} + 2 b (r + σ) z < 0$ when $z > r + σ$ . The set ${V \leq R}$ is a trapping region for sufficiently large $R$ . All trajectories are eventually trapped. $□$

Connections Master

Lyapunov exponents 09.08.03 determine the Kaplan-Yorke dimension of the attractor; $λ_{1} > 0$ is the definition of sensitive dependence.
Period-doubling 09.08.04 is the route by which strange attractors are born (the Lorenz system undergoes a sequence of bifurcations as $r$ increases from 1 to 28).
Horseshoes 09.08.02 are the Hamiltonian analogue; the Lorenz attractor is the dissipative analogue (with contraction instead of volume preservation).
Navier-Stokes 09.07.04 is the parent system from which the Lorenz equations are derived (via severe truncation of the Boussinesq equations).
Symplectic structure 09.04.05 is absent here: the Lorenz system is dissipative, not Hamiltonian, so attractors can exist (Hamiltonian systems cannot have attractors by Liouville's theorem 09.04.04).

Historical & philosophical context Master

Edward Lorenz (1917–2008) was a meteorologist at MIT. In 1961, he was running a weather simulation on a Royal McBee LGP-30 computer. To save time on a repeat run, he entered intermediate values from a previous printout, rounded to three decimal places (the computer used six). The new run diverged completely from the old one. This accidental discovery of sensitive dependence led to the 1963 paper "Deterministic Nonperiodic Flow," one of the most influential papers in 20th-century science.

Lorenz coined the term "butterfly effect" in a 1972 talk: "Does the flap of a butterfly's wings in Brazil set off a tornado in Texas?" The answer is: yes, in principle, but no individual butterfly causes any specific tornado. The atmosphere is a chaotic system with positive Lyapunov exponents, and the limit of predictability is approximately 2 weeks (determined by the inverse of the largest Lyapunov exponent of atmospheric dynamics).

Tucker's 1999 proof that the Lorenz attractor exists was a landmark in computer-assisted mathematics. It required rigorously verified numerical integration of the Lorenz equations using interval arithmetic, combined with analytical arguments near the origin. The proof fills a gap that had stood for 36 years: Lorenz's numerical observations were convincing but not rigorous, and several attempts to prove the existence of the attractor using only analytical methods had failed.

Bibliography Master

Lorenz, E. N., "Deterministic nonperiodic flow," J. Atmos. Sci. 20, 130-141 (1963).
Tucker, W., "The Lorenz attractor exists," C. R. Acad. Sci. Paris 328, 1197-1202 (1999).
Guckenheimer, J. and Williams, R. F., "Structural stability of Lorenz attractors," Publ. Math. IHES 50, 59-72 (1979).
Sparrow, C., The Lorenz Equations: Bifurcations, Chaos, and Strange Attractors (Springer, 1982).
Grassberger, P. and Procaccia, I., "Measuring the strangeness of strange attractors," Physica D 9, 189-208 (1983).
Strogatz, S. H., Nonlinear Dynamics and Chaos, 2nd ed. (CRC Press, 2015).

Prerequisites

09.08.03
09.08.04

Tier anchors

beginner: Strogatz, Nonlinear Dynamics and Chaos (2015), Ch. 9; Taylor, Classical Mechanics (2005), Ch. 13.5
intermediate: Strogatz, Nonlinear Dynamics and Chaos (2015), Ch. 9; Lorenz, Deterministic Nonperiodic Flow (1963)
master: Tucker, The Lorenz attractor exists (1999); Sparrow, The Lorenz Equations: Bifurcations, Chaos, and Strange Attractors (1982)

References

Lorenz — Deterministic nonperiodic flow, J. Atmos. Sci. 20, 130-141 (1963) · The original paper deriving the Lorenz equations from Rayleigh-Benard convection and demonstrating sensitive dependence
Tucker — The Lorenz attractor exists, C. R. Acad. Sci. Paris 328, 1197-1202 (1999) · Computer-assisted rigorous proof that the Lorenz attractor exists and is a strange attractor
Sparrow — The Lorenz Equations: Bifurcations, Chaos, and Strange Attractors (Springer, 1982) · Comprehensive analysis of the Lorenz system: bifurcations, periodic orbits, homoclinic explosions, strange attractor

Estimated time

beginner: 15m
intermediate: 35m
master: 55m