The Lorenz System

In the winter of 1961, Edward Lorenz reran a weather simulation on his Royal McBee computer, typing in the numbers from a mid-run printout as the starting point. The new run tracked the old one for a while — then wandered off entirely, until the two "identical" weathers had nothing to do with each other. The printout had rounded 0.506127 to 0.506: a part-in-a-thousand difference, and the atmosphere in the machine had amplified it into a different future. Lorenz distilled the phenomenon into the smallest system that could hold it — three coupled ordinary differential equations, stripped from the physics of a fluid layer heated from below — and in 1963 published the flow that now bears his name. Its solutions trace the most recognizable object in nonlinear dynamics: two spiral wings, a trajectory forever switching between them in an unpredictable sequence, never crossing itself, never repeating. The butterfly. You arrive with the full chaos toolkit from this chapter — sensitive dependence, Lyapunov exponents, strange attractors, Poincaré maps — and the Lorenz system is where the toolkit assembles into a single worked masterpiece. Every abstraction becomes computable here: the positive exponent is $\lambda_1 \approx 0.9$ , the attractor's fractal dimension is $2.06$ , the return map is a tent with a cusp, and the transition sequence in the parameter $\rho$ — pitchfork, subcritical Hopf, homoclinic explosion — is the standard tour of how a dissipative flow goes chaotic. This article derives the equations from convection (honestly, truncation and all), walks the fixed-point ladder as $\rho$ climbs, proves the flow dissipative and trapped, and reads the computed figures the way Lorenz read his printout: as the discovery that determinism and predictability are different things.

Key takeaways

The Lorenz system is $\dot x = \sigma(y - x)$ , $\dot y = x(\rho - z) - y$ , $\dot z = xy - \beta z$ — a three-mode truncation of Rayleigh–Bénard convection: $x$ = circulation strength, $y$ , $z$ = temperature-field amplitudes; $\sigma$ = Prandtl number, $\rho$ = normalized Rayleigh number, $\beta$ = roll geometry.
The flow is uniformly dissipative: $\nabla\!\cdot\!\vec F = -(\sigma + 1 + \beta) = -41/3$ at classic parameters — phase-space volumes shrink by $\sim10^{6}$ per time unit, so the attractor has zero volume.
Fixed-point ladder in $\rho$ : origin (no convection) stable until $\rho = 1$ ; pitchfork births steady rolls $C_\pm = (\pm\sqrt{\beta(\rho-1)}, \pm\sqrt{\beta(\rho-1)}, \rho - 1)$ ; rolls lose stability in a subcritical Hopf at $\rho_H = \sigma(\sigma{+}\beta{+}3)/(\sigma{-}\beta{-}1) \approx 24.74$ .
At $\sigma = 10$ , $\beta = 8/3$ , $\rho = 28$ the trajectory settles onto the butterfly: a strange attractor with Lyapunov spectrum $(\;\approx 0.906,\ 0,\ -14.57)$ and Kaplan–Yorke dimension $\approx 2.06$ .
Sensitive dependence is quantitative: a $10^{-8}$ perturbation reaches order-one separation in $t \approx 19$ time units; each tenfold improvement in initial data buys only $\ln 10/\lambda_1 \approx 2.5$ more.
The Lorenz map — successive maxima of $z(t)$ plotted $z_{n+1}$ vs $z_n$ — collapses the flow to a one-dimensional cusp map with $|f'| > 1$ everywhere, killing every periodic orbit's stability in one stroke.
Key idea: three innocuous quadratic ODEs, provably dissipative and bounded, generate motion that is deterministic yet unpredictable — the canonical demonstration that chaos needs neither noise nor complexity, only stretching and folding in three dimensions.

The core insight

The Lorenz equations — a three-mode Galerkin truncation of thermal convection — trap every trajectory in a bounded region while contracting phase-space volume at the fixed rate $\sigma + 1 + \beta$ , forcing solutions onto a zero-volume set on which one direction nevertheless stretches ( $\lambda_1 \approx 0.906 > 0$ ): a strange attractor, on which determinism coexists with a finite predictability horizon.

Where this lives today

Operational weather prediction is the Lorenz lesson institutionalized. Every major centre (ECMWF, NOAA) runs 30–100-member ensembles — the same model launched from perturbed initial conditions sampling the analysis uncertainty — because Worked Example 2's logarithm makes the single deterministic forecast a wasting asset; the ensemble spread is today's operational error bar, and "ensemble collapse versus spread" diagnostics are direct descendants of the two-trajectory overlay computed above. The Lorenz-63 system itself, along with its 1996 sibling, remains the universal test rig for data assimilation — the Kalman-filter variants that fuse satellite observations into model states are tuned on these three equations before graduating to the $10^9$ -variable atmosphere. Beyond meteorology, the system turned out to be less a model than a normal form that nature keeps rediscovering: Haken's 1975 mapping of the single-mode laser equations onto (1) made "Lorenz chaos" observable in optics; thermosyphon loops and the malkus water-wheel realize it mechanically; and chaotic-synchronization schemes (Pecora–Carroll 1990) — two Lorenz circuits slaved through one transmitted variable — seeded a communications-security literature. In dynamics proper, Tucker's 2002 computer-assisted proof made the Lorenz attractor the first physically-derived flow proven strange, and the geometric-Lorenz framework it validated is now the template for "robust chaos" classification. The three equations run in every nonlinear-dynamics course, every numerical-methods assignment, and — as the irreducible emblem of sensitive dependence — in the epistemology of every discipline that forecasts anything.

The physical picture

Heat a shallow layer of fluid from below. For gentle heating the fluid conducts: it sits still and the temperature interpolates linearly from hot bottom to cool top. Push harder and conduction becomes unstable — hot fluid is buoyant, and past a critical Rayleigh number the layer spontaneously organizes into counter-rotating convection rolls, the pattern you can see in miso soup. Lorenz asked the minimal question: keep just enough variables to describe one roll — how fast it turns ( $x$ ), how the temperature difference between the rising and falling currents evolves ( $y$ ), and how the vertical temperature profile is distorted from the linear one ( $z$ ) — and let those three amplitudes interact as the fluid equations dictate. The result is the system above. The quadratic terms $xz$ and $xy$ are the fingerprints of advection — the roll carrying its own temperature field around — and they are where all the trouble lives. The parameter $\rho$ is the knob: it measures the heating in units of the convection threshold. Below $\rho = 1$ , nothing moves and the origin — pure conduction — attracts everything. Past 1, two mirror-image steady rolls $C_\pm$ (clockwise, counterclockwise) take over: steady convection. And past $\rho_H \approx 24.74$ the steady rolls themselves go unstable — but nothing stable replaces them. The Hopf bifurcation here is subcritical: it ejects the trajectory rather than cradling it in a limit cycle. What remains is the butterfly — the trajectory spirals outward around one roll (one wing), gets flung across to the neighbourhood of the other, spirals there, and returns, the sequence of wing visits forming an aperiodic, effectively random itinerary.

Figure

The Lorenz butterfly, computed — 14,000 RK4 steps at sigma = 10, rho = 28, beta = 8/3, projected on the x–z plane

Two structural facts discipline the wildness, and their coexistence is the phenomenon. First, the system is dissipative: the divergence of the flow field is a negative constant, so any blob of initial conditions collapses in volume exponentially — after ten time units, a unit cube of uncertainty occupies $10^{-59}$ of its original volume. Everything ends up on a set of zero volume. Second, the system is globally bounded: an explicit ellipsoid traps every trajectory forever (derived below), so the collapse cannot be an escape to infinity, and — by the three-dimensional loophole in the Poincaré–Bendixson theorem you met earlier in this chapter — the zero-volume set it collapses onto need not be a point or a cycle. At $\rho = 28$ it is neither: it is a fractal sheaf of dimension $2.06$ , on which neighbouring trajectories separate at the exponential rate $\lambda_1 \approx 0.906$ even as the volume they jointly occupy shrinks. Stretch along the attractor, squeeze onto it: the same stretch-and-fold anatomy your strange-attractors article dissected, here in its original habitat.

The math: the equations, the ladder, the trap

The system, with the three classic parameters:

\dot x = \sigma(y - x), \qquad \dot y = x(\rho - z) - y, \qquad \dot z = xy - \beta z \qquad (1)

where: $x(t)$ — the convective circulation amplitude (sign = rotation sense); $y(t)$ — the temperature contrast between rising and falling currents; $z(t)$ — the distortion of the vertical temperature profile from linearity; $\sigma$ — the Prandtl number $\nu/\kappa$ (momentum vs thermal diffusivity, $\sigma = 10$ roughly water-like); $\rho = \mathrm{Ra}/\mathrm{Ra}_c$ — the Rayleigh number in units of the convection threshold; $\beta$ — a roll aspect-ratio factor, $8/3$ for the most unstable roll shape. Note the symmetry $(x, y, z) \to (-x, -y, z)$ : physics cannot prefer clockwise to counterclockwise rolls, and the attractor's two wings are this symmetry made visible. Only two terms are nonlinear — $-xz$ and $+xy$ , the advection terms — and setting them to zero leaves a damped linear system that could never do anything interesting. Chaos enters through the quadratics. Dissipation is a one-line computation with global force:

\nabla\!\cdot\!\vec F = \frac{\partial \dot x}{\partial x} + \frac{\partial \dot y}{\partial y} + \frac{\partial \dot z}{\partial z} = -(\sigma + 1 + \beta) = -\frac{41}{3} \approx -13.67 \qquad (2)

where: the divergence is constant — independent of position and of $\rho$ — so every phase-space volume $V(t) = V(0)\,e^{-(\sigma+1+\beta)t}$ exactly. This single number rules out two entire categories of behaviour: no volume-preserving Hamiltonian-style recurrence (this flow is as far from the KAM world of the previous articles as dynamics gets), and no attracting invariant torus carrying quasiperiodic motion — a 2-torus would bound a 3-volume that the flow is forbidden to preserve. Whatever attracts must have zero volume: point, cycle, or something stranger. The fixed-point ladder organizes the road to chaos. Solving $\dot{\vec x} = 0$ : the origin always; and for $\rho > 1$ the pair

C_\pm = \left(\pm\sqrt{\beta(\rho - 1)},\ \pm\sqrt{\beta(\rho - 1)},\ \rho - 1\right) \qquad (3)

where: $C_\pm$ are the steady convection rolls, born from the origin in a supercritical pitchfork at $\rho = 1$ (the conduction state hands over exactly at the convection threshold — the model remembers its fluid ancestry). At the classic $\rho = 28$ : $C_\pm = (\pm 8.49, \pm 8.49, 27)$ — the two red dots at the wings' centres in the figure, each an unstable spiral-saddle whose outward spiral is the wing. The stability analysis (Derivation B) locates the rolls' death at

\rho_H = \frac{\sigma(\sigma + \beta + 3)}{\sigma - \beta - 1} = \frac{470/3}{19/3} = \frac{470}{19} \approx 24.74 \qquad (4)

where: the formula requires $\sigma > \beta + 1$ (otherwise the rolls never destabilize — high-Prandtl convection stays steady, an honest physical prediction). The bifurcation at $\rho_H$ is a subcritical Hopf: the limit cycle that should catch the trajectory exists only below $\rho_H$ and is unstable — at the bifurcation the rolls shed their basin and the dynamics jumps, discontinuously and hysteretically, to the distant strange attractor (which already exists from $\rho \approx 24.06$ , giving a window of genuine bistability between steady rolls and chaos — flip a coin, or rather an initial condition, to choose). Your Hopf-bifurcation article's warning about the subcritical case — "the system leaves town" — was written with exactly this destination in mind.

Derivation AFrom convection to three ODEs — what Lorenz kept and what he threw away

We are sketching, honestly, how the partial differential equations of two-dimensional Rayleigh–Bénard convection collapse to system (1) — because knowing what was discarded is what licenses (and limits) every physical claim made with the model. The strategy is Galerkin truncation: expand the flow and temperature fields in spatial Fourier modes, keep exactly three amplitudes, and let the PDE's quadratic advection terms couple them. Watch where the closure happens — the discarded modes are not small, they are simply dropped, and that is why the Lorenz system is a caricature of convection (a faithful one near onset, a free-standing mathematical object beyond).

Step 1. Start from the Boussinesq equations for a fluid layer of depth $h$ heated from below: incompressible Navier–Stokes with buoyancy $\alpha g\,\delta T$ driving vertical motion, plus advection–diffusion for temperature. In 2-D, write the velocity through a streamfunction $\psi(x_*, z_*, t)$ and split the temperature into the linear conduction profile plus a deviation $\theta(x_*, z_*, t)$ . Two coupled PDEs remain, with two dimensionless groups: the Prandtl number $\sigma = \nu/\kappa$ and the Rayleigh number $\mathrm{Ra} = \alpha g h^3 \Delta T/\nu\kappa$ . Why these groups: $\sigma$ compares how fast momentum and heat diffuse; $\mathrm{Ra}$ compares buoyant driving to the two damping channels. Rayleigh's 1916 linear analysis gives instability at $\mathrm{Ra}_c = 27\pi^4/4 \approx 657.5$ for free boundaries.

Step 2. Expand in the modes of the linear instability. Saltzman (1962) kept many; Lorenz kept three: $\psi \propto X(t)\sin(\pi a x_*/h)\sin(\pi z_*/h)$ — one convection roll of aspect parameter $a$ ; $\theta \propto Y(t)\cos(\pi a x_*/h)\sin(\pi z_*/h)$ — the temperature contrast between up- and downdrafts, and $Z(t)\sin(2\pi z_*/h)$ — the horizontally uniform distortion of the mean profile. Why exactly these: the first two are the critical mode pair of Rayleigh's analysis (the roll and its thermal shadow); the third is the lowest mode the quadratic advection generates from them — the minimal closure that lets the convection modify the very gradient driving it.

Step 3. Insert the expansion, project each PDE onto the three kept modes, discard every other mode the quadratic terms generate. The projected advection integrals produce precisely the $-xz$ and $+xy$ couplings, and after rescaling time by the thermal diffusion rate and the amplitudes by their natural units, system (1) emerges with $\rho = \mathrm{Ra}/\mathrm{Ra}_c$ and $\beta = 4/(1 + a^2)$ . Why $\beta = 8/3$ : the most unstable roll at onset has $a^2 = 1/2$ , giving $\beta = 4/(3/2) = 8/3$ — the odd-looking classic constant is the geometry of the preferred roll, nothing more.

Step 4. State the fine print. The truncation is uncontrolled: the discarded modes are fed energy by the kept ones and would feed back at any finite amplitude, so system (1) is quantitatively faithful to real convection only near onset ( $\rho$ slightly above 1) — at $\rho = 28$ real convection is turbulent in many modes and looks nothing like the butterfly. Why the model survives this indictment: Lorenz's claim was never "this is convection at $\mathrm{Ra} = 28\,\mathrm{Ra}_c$ " but "here is the simplest deterministic system exhibiting nonperiodic flow" — the caricature turned out to contain the universal mechanism (stretch, fold, reinject) that infinitely many real systems share; and remarkably, the same three ODEs re-derive exactly for other physical systems — thermosyphon loops, single-mode lasers (Haken 1975), water-wheel dynamics — where the three-mode truncation is not an approximation but the whole physics. *Verified: dimensional bookkeeping — $[\alpha g h^3\Delta T/\nu\kappa]$ is dimensionless ✓; $\beta = 4/(1+a^2)$ at $a^2 = 1/2$ gives $8/3$ ✓; the $\rho \to 1^+$ steady-roll amplitude from (3), $x^2 = \beta(\rho-1)$ , reproduces the square-root amplitude law of a supercritical pitchfork, matching weakly nonlinear convection theory ✓. Full derivation: Lorenz, J. Atmos. Sci. 20, 130 (1963); Saltzman, J. Atmos. Sci. 19, 329 (1962); Strogatz Ch. 9.*

Go deeper · Derivation B · 3 minThe fixed-point ladder — pitchfork at rho = 1, subcritical Hopf at rho = 24.74

Derivation BThe fixed-point ladder — pitchfork at rho = 1, subcritical Hopf at rho = 24.74

We are deriving the stability biography of the fixed points as $\rho$ climbs at fixed $\sigma = 10$ , $\beta = 8/3$ : the origin's loss of stability at $\rho = 1$ , and the convection rolls' death at $\rho_H = 470/19$ — the two rungs that set up the strange attractor. The tool is the Jacobian eigenvalue analysis from the fixed-points and linearization articles, now in three dimensions; the new technology is the Routh–Hurwitz condition for a cubic, which locates the Hopf crossing without solving for the eigenvalues themselves. Watch the sign of one product — it is the entire bifurcation.

Step 1. Linearize at the origin. The Jacobian of (1) at $(0,0,0)$ block-diagonalizes: $z$ decouples with eigenvalue $-\beta$ , and the $(x, y)$ block is $\begin{pmatrix} -\sigma & \sigma \\ \rho & -1 \end{pmatrix}$ , with characteristic equation $\lambda^2 + (\sigma + 1)\lambda + \sigma(1 - \rho) = 0$ . For $\rho < 1$ both roots are negative (product $\sigma(1-\rho) > 0$ , sum negative): conduction is stable. At $\rho = 1$ the product changes sign — one eigenvalue crosses zero: a pitchfork (symmetric under the flow's $(x,y) \to (-x,-y)$ ), delivering the roll pair (3). At $\rho = 28$ : $\lambda_\pm = \tfrac{1}{2}\left(-11 \pm \sqrt{121 + 40 \times 27}\right) = \tfrac{1}{2}(-11 \pm \sqrt{1201})$ , i.e. $\lambda_+ = 11.83$ , $\lambda_- = -22.83$ , with $\lambda_z = -2.67$ : a saddle with a one-dimensional unstable manifold. Why this saddle matters at $\rho = 28$ : its unstable manifold is the trajectory's escape route from the region between the wings — the two symmetric branches of that manifold feed the two wings, and the origin's saddle geometry is what shuffles wing-to-wing transits.

Step 2. Linearize at $C_\pm$ . Substituting (3) into the Jacobian and expanding $\det(J - \lambda I) = 0$ gives, after algebra that uses $x^2 = \beta(\rho - 1)$ at the fixed point,

\lambda^3 + (\sigma + \beta + 1)\lambda^2 + \beta(\sigma + \rho)\lambda + 2\sigma\beta(\rho - 1) = 0 \qquad (5)

Why all three coefficients are positive for $\rho > 1$ : by Descartes' rule there is then no positive real root — the rolls can only die by a complex pair crossing the imaginary axis, i.e. a Hopf bifurcation, never by a real eigenvalue crossing zero.

Step 3. Apply Routh–Hurwitz. A cubic $\lambda^3 + a_2\lambda^2 + a_1\lambda + a_0$ with positive coefficients has all roots in the left half-plane iff $a_2 a_1 > a_0$ ; the Hopf crossing is the equality $a_2 a_1 = a_0$ (where the cubic factors as $(\lambda + a_2)(\lambda^2 + a_1)$ — one real negative root, one pure-imaginary pair). Set $(\sigma + \beta + 1)\,\beta(\sigma + \rho) = 2\sigma\beta(\rho - 1)$ and solve for $\rho$ :

\rho_H = \frac{\sigma(\sigma + \beta + 3)}{\sigma - \beta - 1} \qquad (6)

with the crossing frequency $\omega_H = \sqrt{a_1} = \sqrt{\beta(\sigma + \rho_H)}$ . At the classic parameters: $\rho_H = 470/19 = 24.737$ , $\omega_H = \sqrt{(8/3)(34.737)} = 9.62$ . Why $\sigma > \beta + 1$ is required: otherwise (6) is negative — no crossing exists and steady rolls persist to arbitrary $\rho$ within the model; the chaos of the Lorenz system is a low-Prandtl-number phenomenon.

Step 4. Determine the Hopf's character — the step that decides everything. The first-Lyapunov-coefficient computation (or the numerical continuation that standard software reproduces) shows the bifurcation is subcritical: an unstable limit cycle exists for $\rho < \rho_H$ , shrinking onto the roll as $\rho \to \rho_H^-$ . For $\rho$ just below $\rho_H$ , each roll sits inside a fragile basin fenced by that unstable cycle; at $\rho_H$ the fence collapses onto the fixed point and the basin evaporates. There is no small stable oscillation to inherit the dynamics — trajectories are expelled to whatever global structure exists, and since $\rho \approx 24.06$ that structure is the strange attractor. In $24.06 < \rho < 24.74$ chaos and steady convection coexist (hysteresis); above $24.74$ the butterfly is the only attractor. Why this matters beyond Lorenz: this is the canonical instance of a "dangerous" bifurcation — no warning amplitude growth, a hard jump to a distant attractor, and hysteresis on the way back — the failure mode your Hopf article flagged, realized in the founding example. *Verified: (6) recomputed — numerator $10 \times (10 + 8/3 + 3) = 470/3$ , denominator $10 - 8/3 - 1 = 19/3$ , ratio $470/19 = 24.7368$ ✓; origin eigenvalues at $\rho = 28$ : $\sqrt{1201} = 34.655$ , giving $11.83$ , $-22.83$ ✓, and their sum $-11 - 8/3 = -41/3$ matches the trace (2) ✓; the eigenvalue sum at $C_\pm$ , $-a_2 = -(\sigma + \beta + 1) = -41/3$ , matches the universal divergence ✓. Standard treatment: Strogatz Ch. 9, Sparrow (1982).*

Go deeper · Derivation C · 3 minDissipative and trapped — why a strange attractor is forced, and why no stable orbit survives

Derivation CDissipative and trapped — why a strange attractor is forced, and why no stable orbit survives

We are proving the two global facts that leave chaos as the only option at $\rho = 28$ : every trajectory eventually enters a bounded ellipsoid and stays (no escape), while every volume shrinks to zero (no fat attractor) — and then running Lorenz's own argument, via the cusp return map, that no stable periodic orbit is hiding in the wreckage. The strategy for trapping is a Lyapunov-function estimate: build a quadratic energy-like function whose time derivative is negative everywhere outside a bounded region. Watch the completing-of-squares — the choice of function looks unmotivated until the cross terms cancel, which is the entire craft.

Step 1. Define $V(x, y, z) = \rho x^2 + \sigma y^2 + \sigma(z - 2\rho)^2$ . Differentiate along the flow (1): $\dot V = 2\rho x \dot x + 2\sigma y \dot y + 2\sigma(z - 2\rho)\dot z$ . The dangerous cubic terms cancel in the combination — $2\rho x(\sigma y) + 2\sigma y(-xz \cdot 1) \ldots$ — collecting everything: $\dot V = -2\sigma\left(\rho x^2 + y^2 + \beta z^2 - 2\beta\rho z\right)$ . Why the weights $(\rho, \sigma, \sigma)$ and the shift $z - 2\rho$ : they are reverse-engineered precisely so the $xyz$ advection terms — which conserve the quadratic form — drop out, leaving only the linear damping visible. The advection terms in (1) redistribute "energy" between modes but create none; the function $V$ is built to see only the dissipation.

Step 2. Complete the square in $z$ : $\dot V = -2\sigma\left[\rho x^2 + y^2 + \beta(z - \rho)^2 - \beta\rho^2\right]$ . Thus $\dot V < 0$ everywhere outside the bounded ellipsoid $E:\ \rho x^2 + y^2 + \beta(z-\rho)^2 \le \beta\rho^2$ . Any trajectory outside a large $V$ -level set containing $E$ has strictly decreasing $V$ , so it must enter that level set in finite time and can never leave. Why this kills escape: combined with local existence theory it makes solutions global and uniformly bounded — the Lorenz flow is a closed ecology; whatever complexity occurs must occur in a box roughly $|x| \lesssim 20$ , $|y| \lesssim 27$ , $0 < z \lesssim 48$ — exactly the window of the computed figures.

Step 3. Combine with dissipation (2): inside the trapping region, every volume shrinks as $e^{-41t/3}$ , so the attractor $A = \bigcap_{t>0} \phi_t(\text{trap})$ has zero 3-volume; yet it must be "large enough" to carry bounded aperiodic motion with a positive Lyapunov exponent. The measured resolution of this tension: Lyapunov spectrum $(\lambda_1, \lambda_2, \lambda_3) \approx (0.906, 0, -14.57)$ — note $\lambda_1 + \lambda_2 + \lambda_3 = -13.67 = -(\sigma + 1 + \beta)$ , as the constant divergence demands — giving Kaplan–Yorke dimension $D_{KY} = 2 + \lambda_1/|\lambda_3| = 2.062$ . Why the dimension is barely above 2: the squeezing ( $14.57$ ) overwhelms the stretching ( $0.906$ ), so the attractor is locally almost a surface — two sheets that appear to merge but, since trajectories cannot cross, must actually be infinitely many sheets packed at Cantor spacing: the fractal transverse structure your strange-attractors article measured in the Hénon map, here forced by pure arithmetic.

Step 4. Run Lorenz's periodic-orbit execution. Record each successive maximum $z_n$ of $z(t)$ and plot $z_{n+1}$ against $z_n$ (the computed figure below): the points collapse — to visual precision — onto a single-valued curve with a cusp: a one-dimensional return map, legitimate because the attractor is locally two-sheeted, so one coordinate on a transversal almost suffices (this is the Poincaré-section machinery of the return-maps article, in its most famous application). The measured curve satisfies $|f'(z)| > 1$ everywhere. A periodic orbit of the flow is a periodic point of $f$ ; its stability multiplier is a product of factors $f'(z_i)$ , each exceeding 1 in magnitude — so every periodic orbit is unstable. Why this argument was historic: from one computed curve, Lorenz excluded the entire classical repertoire of long-time behaviour — no stable equilibrium (Step 2 of the previous derivation), no stable cycle (this step), no quasiperiodic torus (volume contraction) — leaving, by elimination, sustained aperiodic motion on a set of zero volume: the object later christened a strange attractor. The 1963 paper contains, in effect, the definition by exclusion. *Verified: exponent-sum identity $0.906 + 0 - 14.57 = -13.67 = -41/3$ ✓; $D_{KY} = 2 + 0.906/14.57 = 2.062$ , matching box-counting estimates $2.06 \pm 0.01$ ✓; the trapping-ellipsoid bound $z \le 2\rho \approx 56$ is consistent with the computed attractor's $z_{\max} \approx 45$ ✓; cusp-map slope: the computed map's measured slopes exceed 1 across the whole branch pair (figure) ✓. Treatments: Strogatz Ch. 9, Sparrow (1982), Guckenheimer & Holmes §2.3.*

Which equation answers which question

you want	use	notes
the equations & parameter meanings	eq. (1)	$\sigma$ Prandtl, $\rho$ = Ra/Ra $_c$ , $\beta$ roll geometry
volume budget / why no tori	$\nabla\!\cdot\!\vec F = -(\sigma{+}1{+}\beta)$ , eq. (2)	constant; exponents must sum to it
the steady states	origin & $C_\pm$ , eq. (3)	pitchfork at $\rho = 1$
where steady convection dies	$\rho_H = \sigma(\sigma{+}\beta{+}3)/(\sigma{-}\beta{-}1)$ , eq. (4)/(6)	subcritical Hopf; needs $\sigma > \beta + 1$
boundedness	trapping ellipsoid, Derivation C	$V = \rho x^2 + \sigma y^2 + \sigma(z-2\rho)^2$
predictability horizon	$t \approx \ln(\Delta/\delta_0)/\lambda_1$	$\lambda_1 \approx 0.906$ at classic parameters
attractor geometry	$D_{KY} = 2 + \lambda_1/\|\lambda_3\| \approx 2.06$	nearly-2D sheaf of sheets
periodic-orbit stability	Lorenz cusp map, $\|f'\| > 1$	all cycles unstable

Limiting cases

Case	Result	Meaning
$\rho < 1$	origin globally stable	conduction: heat crosses, fluid rests
$1 < \rho < 24.06$	$C_\pm$ attract (spirally for $\rho \gtrsim 1.35$ )	steady convection rolls; transient chaos appears $\rho \gtrsim 13.93$
$24.06 < \rho < 24.74$	rolls AND strange attractor coexist	hysteresis window of the subcritical Hopf
$\rho = 28$	the classic butterfly	$\lambda_1 \approx 0.906$ , $D \approx 2.06$
$\rho \to \infty$ (large- $r$ limit)	periodic windows return (e.g. $\rho \approx 99.5$ – $100.8$ , $145$ – $166$ )	chaos is not monotone in the driving
$\sigma \le \beta + 1$	no Hopf: rolls stable for all $\rho$	high-Prandtl fluids convect steadily in-model
nonlinearities deleted	damped linear flow to the origin	the quadratic advection terms carry all the chaos
$t \to \infty$ , any volume	$V(t) = V(0)e^{-41t/3} \to 0$	attractor has zero volume, always

Figure

Sensitive dependence, computed — two trajectories launched 10⁻⁸ apart in x, overlaid until they part company near t = 19

Common trapChaos means the Lorenz equations are unstable and blow up — or random

Neither, and the two computed global facts of Derivation C are the antidote. The flow is bounded: every trajectory, forever, lives inside an explicit ellipsoid — nothing blows up, and indeed all three Lyapunov exponents sum to $-13.67$ : the dynamics is overwhelmingly contracting. And it is perfectly deterministic: run the same initial condition twice and you get the identical trajectory to machine precision; the equations harbour no noise, and the wing-hopping itinerary is a function of the initial state alone. What chaos actually asserts is a divorce between determinism and predictability: the single positive exponent $\lambda_1 \approx 0.906$ means any error in specifying the state — and physical measurement guarantees one — grows by $e$ every $1.1$ time units, so the horizon of useful prediction scales only logarithmically with precision: $10^{-8} \to$ order-one in $t \approx 20$ ; a million-fold better measurement buys $\ln(10^6)/0.906 \approx 15$ units more. A second trap hides in the word "unstable": the individual trajectories separate exponentially, but the attractor is robustly stable as a set — perturb the state and you fall right back onto the butterfly (only where on it is scrambled), and perturbing the parameters deforms the attractor smoothly across wide ranges. That two-level structure — statistical robustness carrying pointwise unpredictability — is why weather ensembles work: no member is right, the cloud of members is.

Worked examples

1 · Anatomy of the classic attractor (structure audit)

At $\sigma = 10$ , $\beta = 8/3$ , $\rho = 28$ , locate all fixed points, classify them, and verify the eigenvalue bookkeeping against the universal volume-contraction rate.

Step 1. Fixed points: origin, plus $C_\pm$ from (3): $\sqrt{\beta(\rho-1)} = \sqrt{(8/3)(27)} = \sqrt{72} = 8.485$ , so $C_\pm = (\pm 8.485, \pm 8.485, 27)$ — confirmed as the wing centres in the computed butterfly.

Step 2. Origin: eigenvalues $\lambda_z = -\beta = -2.667$ and $\tfrac12\left(-11 \pm \sqrt{1201}\right) = \{+11.83, -22.83\}$ . One unstable direction: a saddle whose unstable manifold launches trajectories toward the wings. Sum: $11.83 - 22.83 - 2.67 = -13.67 = -(\sigma + 1 + \beta)$ ✓.

Step 3. $C_\pm$ : the cubic (5) has coefficients $a_2 = 41/3 = 13.67$ , $a_1 = \beta(\sigma + \rho) = (8/3)(38) = 101.3$ , $a_0 = 2\sigma\beta(\rho - 1) = 2 \times 10 \times (8/3) \times 27 = 1440$ . Routh–Hurwitz product $a_2 a_1 = 1385 < 1440 = a_0$ : the stability condition fails — as it must, since $\rho = 28 > \rho_H = 24.74$ . Numerically the roots are one real $\approx -13.85$ and a complex pair $\approx +0.094 \pm 10.19\,i$ : a weakly repelling spiral (checks: real parts sum to $0.094 \times 2 - 13.85 = -13.67$ ✓; the slow outward spiral rate $0.094$ against orbital frequency $10.2$ is why each wing shows many nearly-closed loops before ejection — the trajectory needs $\sim\!70$ turns to e-fold outward, but leaves sooner because ejection is finite-amplitude, not asymptotic).

Step 4. Read the mechanism off the numbers: wings = slow unstable spirals around $C_\pm$ (rate 0.09), transits = rides on the origin-saddle's manifolds (rates ±12–23), and the global contraction $-41/3$ squeezes it all onto the $D = 2.06$ sheaf. Three linearizations plus one divergence — the entire butterfly, accounted for.

2 · The predictability budget (the Lorenz lesson, quantified)

Using $\lambda_1 = 0.906$ : (a) the error-doubling time; (b) the horizon for initial error $10^{-8}$ reaching order one; (c) what a thousandfold precision improvement buys; (d) the horizon's form if these were weather models.

Step 1. Doubling time: $t_2 = \ln 2/\lambda_1 = 0.693/0.906 = \boxed{0.765}$ time units. (The computed figure's divergence is consistent: measured $\lambda \approx 0.93$ over the fit window — finite-time exponents fluctuate around the asymptotic $0.906$ .)

Step 2. Horizon: $t_* \approx \ln(\Delta/\delta_0)/\lambda_1 = \ln(10^8)/0.906 = 18.42/0.906 = \boxed{20.3}$ — the overlay figure shows visible separation at $t \approx 18.8$ , within the finite-time scatter ✓.

Step 3. Improve $\delta_0$ to $10^{-11}$ : gain $\ln(10^3)/0.906 = 6.9/0.906 = \boxed{7.6}$ units — a 37% extension for a 1000× measurement effort. The logarithm is merciless: precision is the wrong currency for buying forecast range.

Step 4. The atmospheric translation Lorenz himself drew: with error-doubling measured in a couple of days (modern estimates: ~1.5–2 days for synoptic scales), initial-condition improvements asymptote the useful forecast at roughly two weeks — not because models are poor, but because $\ln$ beats any budget. This single scaling argument reshaped meteorology: effort moved from chasing perfect initial data to ensemble forecasting — sampling the initial-error ball and propagating the whole cloud, converting the doomed pointwise question into the answerable statistical one.

Figure

The Lorenz map, computed — 320 successive maxima of z(t), plotted z(n+1) against z(n): a one-dimensional cusp with slope magnitude everywhere above 1

Origins · 2 minEdward Norton Lorenz and how this was discovered

Edward Norton Lorenz

1917–2008 · MIT

Lorenz came to meteorology through wartime weather forecasting, armed with a Harvard mathematics degree, and spent the 1950s in the MIT group testing whether statistical methods could match physics-based prediction. His answer — an accidental one — settled a deeper question. The 1961 rounding incident on his Royal McBee LGP-30 (0.506127 entered as 0.506) revealed that his twelve-variable weather model amplified initial differences exponentially; rather than filing it as a computer artifact, Lorenz spent two years distilling the effect to its minimal carrier, publishing "Deterministic Nonperiodic Flow" in the Journal of the Atmospheric Sciences in 1963 — three equations, the cusp return map, the elimination argument against stable orbits, and the predictability horizon, essentially the complete modern theory in twelve pages that mathematicians took a decade to find. His 1972 AAAS talk title — "Does the flap of a butterfly's wings in Brazil set off a tornado in Texas?" (coined by session chair Philip Merilees) — gave the phenomenon its immortal name; Lorenz's own point was characteristically precise: if the flap can seed the tornado, it can equally prevent one — sensitivity is about the inseparability of small causes, not their malevolence. He kept working — the 1996 "Lorenz-96" model is still the standard testbed for data assimilation — hiking and cross-country skiing into his eighties, a quiet man who had, in Ian Stewart's phrase, ended the Laplacian dream politely.

History

1962

— Saltzman's seven-mode convection truncation exhibits nonperiodic solutions; Lorenz, in the audience of the problem, prunes it to three modes.

1963

— "Deterministic Nonperiodic Flow" appears: equations, butterfly, cusp map, predictability horizon — cited a handful of times in its first decade, then tens of thousands.

1972

— The butterfly gets its name: Lorenz's AAAS talk on a Brazilian butterfly and a Texas tornado enters the language.

1979

— Guckenheimer and Williams build the geometric Lorenz model — an idealized flow provably carrying a robust strange attractor — leaving open whether the real equations match it.

2002

— Warwick Tucker closes the question with a computer-assisted proof (interval arithmetic + normal forms): the Lorenz attractor exists and is a strange attractor, resolving Smale's 14th problem.

Going further

Texts. Lorenz, "Deterministic Nonperiodic Flow," J. Atmos. Sci. 20, 130 (1963) — still the best first read · Strogatz, Nonlinear Dynamics and Chaos, Ch. 9 — the modern standard treatment this article's conventions follow · Sparrow, The Lorenz Equations: Bifurcations, Chaos, and Strange Attractors (1982) — the full parameter-space atlas · Guckenheimer & Holmes, Nonlinear Oscillations…, §2.3 & §5.7 · Tucker, "A rigorous ODE solver and Smale's 14th problem," Found. Comput. Math. 2, 53 (2002) · Gleick, Chaos (1987), for the history.

Provenance. Authored fresh for the mechanics track (Phase-C gap add-back; no AllQFT base). All figures computed for this article (RK4, $dt = 0.004$ , transients discarded): butterfly (3,500 plotted points), two-trajectory divergence ( $\delta_0 = 10^{-8}$ , separation reaching 1 at $t \approx 18.8$ ; fitted finite-time exponent 0.93 vs literature $\lambda_1 = 0.9056$ ), and the cusp map (320 successive $z$ -maxima, range 31.0–46.4). A verifier should recompute: $\rho_H = 470/19 = 24.737$ and $\omega_H = 9.62$ ; origin eigenvalues $\{11.83, -22.83, -8/3\}$ at $\rho = 28$ (sum $-41/3$ ✓); $C_\pm = (\pm\sqrt{72}, \pm\sqrt{72}, 27)$ ; Routh–Hurwitz numbers $a_2a_1 = 1385 < a_0 = 1440$ at $\rho = 28$ ; the $C_\pm$ spiral pair $0.094 \pm 10.19i$ ; $D_{KY} = 2 + 0.906/14.57 = 2.062$ ; the predictability arithmetic of worked example 2.

Builds on: Sensitive Dependence on Initial Conditions · Lyapunov Exponents · Strange Attractors and Fractal Dimension · Hopf Bifurcation and Birth of Limit Cycles · Poincaré Sections and Return Maps. Leads to: Discrete Maps and the Logistic Map · (beyond this track) turbulence, predictability theory, and data assimilation.

Is the butterfly effect really saying a butterfly can cause a tornado?

Read it as Lorenz did: in a flow with $\lambda_1 > 0$ , the future beyond the predictability horizon is inseparable from arbitrarily small details of the present — the flap-sized perturbation grows to synoptic scale in finite time, so the weather-with-flap and weather-without-flap genuinely differ, and the tornado may occur in one and not the other. That is not "the butterfly powers the tornado" (energy is irrelevant; the tornado's energy was always available in the atmosphere) and not "butterflies make weather worse" (Lorenz's own caveat: the flap is as likely to prevent the tornado as to seed it — you cannot even assign the sign). The honest content is epistemic: causes below your measurement resolution are dynamically promoted to consequences above it, so deterministic laws do not confer determinable futures. Two fine-print items keep the metaphor honest: the growth saturates at attractor size (a butterfly cannot make Venus's climate on Earth — sensitive dependence operates within the attractor, not across attractors), and real atmospheric error growth is scale-dependent (small scales double fastest), so the single- $\lambda$ picture of Lorenz-63 is the cartoon, refined but not overturned by modern predictability theory.

Why does chaos need three dimensions for a flow — the logistic map manages it in one?

The constraint is topological and specific to continuous, deterministic flows: trajectories cannot cross. In a two-dimensional phase plane, a trajectory is a curve that fences the plane — anything trapped and non-approaching-a-fixed-point must, by the Poincaré–Bendixson theorem from your earlier article, wind onto a limit cycle. Stretching (for sensitivity) plus folding (for boundedness) cannot be arranged in 2-D without a crossing; the third dimension is exactly the room to fold a sheet back over itself. Maps escape the theorem because iteration is not a continuous curve: the logistic map's $x_{n+1} = rx_n(1-x_n)$ jumps, so it can fold the interval in one dimension — and non-invertibility (two preimages) is its substitute for the third dimension. The Lorenz system sits precisely at the minimum: three variables, and its faithful one-dimensional shadow — the cusp map of successive maxima — is (like the logistic map at full chaos) non-invertible. The dimension count is also why the Lorenz attractor's $D \approx 2.06$ is no accident: a flow's strange attractor in 3-D must exceed dimension 2 (it carries a 2-D sheet's worth of non-crossing trajectories) but, being volume-starved by dissipation, barely.

Is the Lorenz attractor actually proven to be a strange attractor, or just numerically observed?

Proven — but the proof took 39 years and a computer. The subtlety: numerics alone cannot distinguish a true strange attractor from, say, an absurdly long stable periodic orbit (such orbits do exist at other parameter values, e.g. the $\rho \approx 99.5$ – $100.8$ window), and floating-point trajectories are shadowed by true ones only under hypotheses that chaos itself strains. Guckenheimer and Williams (1979) proved that an idealized "geometric Lorenz flow" — built from the observed sheet structure — carries a robust strange attractor; what remained open (listed by Smale in 1998 as his 14th problem) was whether the actual equations (1) realize that geometry. Warwick Tucker's 2002 proof closed it with rigorous interval-arithmetic integration (every numerical step carries certified error bounds) combined with normal-form analysis near the origin, where finite integration cannot cope with the saddle's slowdown. Conclusion: for the classic parameters, the Lorenz equations possess a transitive strange attractor with a positive Lyapunov exponent, robust under perturbation. It stands as the emblem of computer-assisted proof in dynamics — the machine not as illustrator but as certified witness.

Does the Lorenz system actually describe the convection experiment it came from?

Only in a thin sliver, and honesty about this is part of the article's contract. The three-mode truncation is quantitatively faithful near onset — $\rho$ slightly above 1, where steady rolls appear with the predicted square-root amplitude — but by $\rho = 28$ a real Rayleigh–Bénard cell has excited many spatial modes the truncation discarded: real convection at that driving shows oscillatory instabilities, then spatiotemporal complexity in which the pattern, not just three amplitudes, dances. The model's chaotic regime is not a prediction about that fluid. Yet the equations found their experiments elsewhere: systems whose physics genuinely possesses only a few modes. Haken showed in 1975 that the single-mode laser maps exactly onto (1) (with reinterpreted parameters — field, polarization, inversion), and "Lorenz chaos" was subsequently observed in far-infrared lasers; thermosyphons (fluid convecting in a closed loop) and the Malkus–Howard water-wheel realize the equations mechanically, wing-hopping as spontaneous reversals of rotation. The philosophical residue is the interesting part: a model can fail its parent system and still be exact for others, because what it captures is a mechanism — and the mechanism, not the fluid, is what proved universal.

What exactly do the wings correspond to physically, and what is a wing-switch?

In the convection reading: the two wings are the two rotation senses of the roll. On the right wing $x > 0$ — fluid circulating, say, clockwise, spiralling around the unstable steady-roll state $C_+$ ; on the left wing, counterclockwise around $C_-$ . A trajectory looping within one wing is convection whose amplitude oscillates — overshooting and undershooting the (unstable) steady value with growing swing, since $C_\pm$ are weakly repelling spirals (outward rate 0.094 against orbital frequency 10.2 from Worked Example 1). A wing-switch is a spontaneous reversal of the convection roll: the oscillation grows until the circulation stalls near $x \approx 0$ , the trajectory falls under the influence of the origin-saddle, and its unstable manifold flings it into the other rotation sense. In the Malkus water-wheel realization the reversals are visible and startling — the wheel spins, falters, and turns the other way at irregular intervals. The itinerary of switches (3 loops right, 1 left, 4 right...) is the system's symbolic dynamics, and its effective randomness — extractable as a coin-flip sequence from perfectly deterministic equations — is sensitive dependence rendered as narrative: which wing you occupy ten switches hence depends on digits of the initial condition you could never measure.

The Lyapunov exponents and the dimension are quoted numerically — can any of this be computed analytically?

Almost none of it, and the honest accounting is instructive. Analytically available: the fixed points and their eigenvalues (Derivations B), the Hopf threshold (6), the divergence (2), the trapping ellipsoid, and one exact global identity — the exponents must sum to $-(\sigma + 1 + \beta)$ , which provides a stringent check on any numerical spectrum (and, with the exact $\lambda_2 = 0$ along the flow, means only one independent number, $\lambda_1$ , is genuinely empirical: $\lambda_3$ follows). Not analytically available: $\lambda_1$ itself, the attractor's dimension, the invariant measure, and the diffusion statistics of wing-switches — these require the numerical machinery of your Lyapunov-exponents article (tangent-space integration with Gram–Schmidt renormalization yields $\lambda_1 = 0.9056$ at classic parameters; the figure's two-trajectory fit, $\approx 0.93$ over a finite window, illustrates the finite-time scatter around it). This division of labour is the normal condition of chaos theory: rigorous structure theorems fencing a quantitatively numerical interior. Even Tucker's existence proof is computer-assisted — the analytic-only toolkit provably cannot decide questions that depend on following trajectories through finite-amplitude folding.

Why is the wing-switching sequence "random" if the equations are deterministic — where does the randomness come from?

From the initial condition's infinite decimal tail, promoted digit by digit into observable events. The mechanism is cleanest in the Lorenz map: the cusp map, like the full-chaos logistic map from your discrete-maps article, is expanding ( $|f'| > 1$ ), so each iteration roughly doubles the scale at which the current state's fine structure matters — after $n$ switches, the wing choice is reading digits of the initial condition around the $n$ -th binary place. The "randomness" of the itinerary is therefore exactly the randomness of a typical real number's digit sequence: deterministic dynamics acts as a magnifying glass on the initial condition, and since no physical preparation controls infinitely many digits, the output is operationally a stochastic process (in fact the symbolic dynamics is well-approximated by a Markov chain, and this is how "deterministic chaos generates entropy" is made precise — the Kolmogorov–Sinai entropy, equal by Pesin's identity to the sum of positive exponents, here $\approx 0.906$ nats per unit time, is literally the rate at which the system converts unread initial-condition digits into new observable information). Determinism fixes the mapping; measurement finiteness supplies the coin flips.

The Lorenz System

Where this lives today

The physical picture

The math: the equations, the ladder, the trap

Which equation answers which question

Limiting cases

Worked examples

1 · Anatomy of the classic attractor (structure audit)

2 · The predictability budget (the Lorenz lesson, quantified)

Going further

Before this

Where this leads