The Double Pendulum

The driven damped pendulum of the previous article needed three ingredients for chaos — nonlinearity, dissipation, forcing — and a hidden third dimension smuggled in through the drive phase. The double pendulum needs none of that apparatus. Hang one pendulum from the bob of another, lift the pair to horizontal, let go. No motor, no friction worth modelling, energy exactly conserved, equations derivable in ten lines of Lagrangian mechanics — and the motion is so sensitive to its starting point that no computer on Earth can forecast the third flip of the lower arm from any realistic measurement of the release. This article is the course's cleanest specimen of Hamiltonian chaos, and it earns that role by an exact count. Two rods give two degrees of freedom: a four-dimensional phase space $(\theta_1, \theta_2, p_1, p_2)$ . Time-independence gives one conserved quantity: the energy. The Liouville–Arnold theorem of the W4 articles says integrability — motion forever on invariant tori — requires as many independent commuting constants as degrees of freedom: two. The double pendulum has one. That missing second integral is not a technical footnote; it is the entire story. Trajectories are confined to a three-dimensional energy surface but not to two-dimensional tori within it, and in the room left over, chaos lives. The system is also the perfect stage for the KAM theorem in action. At low energy the coupling barely flexes and quasiperiodic tori dominate — the Poincaré section shows nested closed curves, a portrait of near-integrability. Feed in more energy and resonant tori shred into island chains, chaotic layers widen around broken separatrices, and above the energy at which the inner arm can flip over the top, a connected chaotic sea floods most of the surface. One tabletop toy walks you from Liouville's ordered world to Poincaré's tangled one, with the crossover controlled by a single dial: energy.

Key takeaways

The double pendulum — two point masses on two massless rods — has two degrees of freedom, a four-dimensional phase space, and a Lagrangian whose kinetic energy couples the angles through $\cos(\theta_1 - \theta_2)$ .
It is conservative and autonomous: the Hamiltonian $H = E$ is conserved, phase volume is preserved, and there are no attractors — chaos here is Hamiltonian, not dissipative.
Two degrees of freedom with only one conserved quantity makes the system non-integrable by the Liouville–Arnold count: no second integral, no global foliation by tori.
At small amplitude it linearizes to two normal modes; for equal masses and lengths, $\omega_\pm^2 = (g/\ell)(2 \pm \sqrt2)$ with mode shapes $\theta_2/\theta_1 = \mp\sqrt2$ .
As energy rises, KAM tori break up: Poincaré sections show closed curves dissolving into island chains and a chaotic sea, essentially complete above the lower-arm flip energy $2m_2g\ell_2$ .
The Lyapunov spectrum is symmetric, $\{+\lambda, 0, 0, -\lambda\}$ : stretching is exactly repaid by contraction, as Liouville demands.
Key idea: two degrees of freedom plus only one conserved quantity (energy) makes the double pendulum non-integrable; the configuration-dependent inertial coupling drives a transition from quasiperiodic tori at low energy to a chaotic sea at high energy.

Where this lives today

The double pendulum is the drosophila of modern dynamics-adjacent engineering. In robotics it is the two-link arm: the Acrobot (double pendulum actuated only at the joint) and Pendubot are the standard benchmarks of underactuated control, where the task — swing up to and balance at the up-up saddle — must exploit exactly the nonlinear energy-pumping this article describes, and where controllers are certified against the chaotic sensitivity they must overpower. Gymnasts, gibbons, and golf swings are routinely modelled as driven double pendula for the same reason. In numerical science it is the canonical stress test for symplectic integrators — the geometric methods that preserve phase volume and keep long-time energy errors bounded, without which the sections of Figure 2 would be artifacts — and, most recently, the standard adversarial example for machine-learned dynamics: neural surrogates that forecast it beautifully for two flips and then dissolve into nonsense are a teaching moment about positive Lyapunov exponents, not a bug report. In this chapter's architecture the double pendulum is the hinge between worlds. It carries the Hamiltonian inheritance of W3-W4 — Liouville, tori, action-angle order — into the chaos block, and it hands the next three articles their subject matter: sensitive dependence (the flip-time unpredictability quantified here by hand), Lyapunov exponents (the measurement industrialized), and Poincaré sections (the instrument, formalized). The chaotic sea it displays is also KAM's sea — the same island-and-torus geometry that governs asteroid transport and accelerator beam stability, met here in a toy you can build from two rulers.

The physical picture

Where exactly does the trouble enter? Not through the potential — gravity acts on each rod as innocently as on a single pendulum. It enters through the inertia. The lower bob rides on a moving pivot, so its velocity mixes both angular rates, and the kinetic energy acquires a cross term $m_2\ell_1\ell_2\,\dot\theta_1\dot\theta_2\cos(\theta_1 - \theta_2)$ . The effective inertia each rod feels depends on the angle between the rods: swing them parallel and they drag each other along; fold them apart and the coupling reverses. Mass itself, in the generalized-coordinate sense, is a function of configuration — and a configuration-dependent mass matrix is a nonlinearity that no small trick removes.

Figure

The double pendulum — two rods, two angles, and the tangled trace its lower bob actually draws

At low energy the geometry barely changes — $\cos(\theta_1 - \theta_2) \approx 1$ — and the system is the coupled-oscillator problem of the W2 articles: two normal modes, in-phase and out-of-phase, every gentle motion a beat pattern between them. This is the integrable skeleton. Crank the energy up and the coupling term sweeps its full range; energy sloshes violently between the arms; the lower arm gathers enough to flip over the top, and each flip is a decision point where microscopically different histories choose differently. Two releases differing by a hundredth of a degree track each other for a couple of swings — then one flips and the other does not, and all resemblance ends.

Figure

The two normal modes of the small-amplitude double pendulum — the integrable skeleton beneath the chaos

The right arena for seeing the transition is the Poincaré section — here a bona fide section of an autonomous flow, unlike the stroboscope of the driven pendulum. Fix the energy, record $(\theta_2, \dot\theta_2)$ each time $\theta_1$ crosses zero upward, and watch the geometry of the invariant sets. At $E = 0.5$ (units of $m g \ell$ ): nested closed curves — trajectories winding on surviving KAM tori, quasiperiodic and forever regular. At $E = 2.2$ : an archipelago — island chains around resonant orbits, tori surviving in patches, a connected chaotic band prowling between them. At $E = 4.0$ : the sea — one single trajectory scatters its dots across the whole accessible region. All three panels below are computed from the exact equations of motion; nothing is schematic.

Figure

Poincaré sections of the double pendulum at three energies — the KAM transition from order to a chaotic sea, computed

Hold on to one structural contrast with the previous article. The driven damped pendulum contracted phase volume and owned attractors — transients died into a fixed point, cycle, or strange set. Here Liouville's theorem (the W4 article Liouville's Theorem and the Phase Fluid) forbids all of that: volume is preserved, nothing attracts, and a chaotic trajectory never settles — it wanders its chaotic sea forever, returning arbitrarily close to its start (Poincaré recurrence) without ever repeating. Hamiltonian chaos is not decay onto a strange object; it is eternal, measure-preserving restlessness.

The mechanics, assembled

Coordinates first. With the pivot at the origin and $y$ measured upward, the bob positions are

x_1 = \ell_1\sin\theta_1, \quad y_1 = -\ell_1\cos\theta_1; \qquad x_2 = \ell_1\sin\theta_1 + \ell_2\sin\theta_2, \quad y_2 = -\ell_1\cos\theta_1 - \ell_2\cos\theta_2 \qquad (1)

where: $\theta_1, \theta_2$ — angles of the upper and lower rods from the downward vertical — the natural generalized coordinates, automatically encoding both constraints. The Lagrangian machinery of the W3 articles then produces (Derivation A builds it):

L = \tfrac12(m_1{+}m_2)\ell_1^2\dot\theta_1^2 + \tfrac12 m_2\ell_2^2\dot\theta_2^2 + m_2\ell_1\ell_2\,\dot\theta_1\dot\theta_2\cos(\theta_1{-}\theta_2) + (m_1{+}m_2)g\ell_1\cos\theta_1 + m_2g\ell_2\cos\theta_2 \qquad (2)

where: the third term is the villain of the piece — the configuration-dependent inertial coupling. Note what is absent: no term mixes $\theta_1$ and $\theta_2$ in the potential; gravity treats the rods separately. All the nonlinear coupling is kinetic, which is exactly why it cannot be tamed by any small-oscillation trick once the relative angle swings freely. The Euler–Lagrange equations that follow (Derivation A) are two coupled second-order ODEs,

(m_1{+}m_2)\ell_1\ddot\theta_1 + m_2\ell_2\ddot\theta_2\cos(\theta_1{-}\theta_2) + m_2\ell_2\dot\theta_2^2\sin(\theta_1{-}\theta_2) + (m_1{+}m_2)g\sin\theta_1 = 0 \qquad (3)

\ell_2\ddot\theta_2 + \ell_1\ddot\theta_1\cos(\theta_1{-}\theta_2) - \ell_1\dot\theta_1^2\sin(\theta_1{-}\theta_2) + g\sin\theta_2 = 0 \qquad (4)

where: the $\dot\theta^2\sin(\theta_1{-}\theta_2)$ terms are centripetal couplings — each rod feels the other's rotation as a flung force. These equations are exact, compact, and analytically unsolvable in general; every trajectory in this article's figures is their numerical integration. Passing to the Hamiltonian (Legendre transform, as in the W3-4 articles) gives canonical momenta $\mathbf p = M(\theta_1{-}\theta_2)\,\dot{\boldsymbol\theta}$ with a non-diagonal, angle-dependent mass matrix, and

H(\theta_1, \theta_2, p_1, p_2) = T + V = E \qquad (5)

where: $H$ equals the energy because the constraints are time-independent and $L$ has no explicit $t$ . This is the system's only general constant of motion. Angular momentum is not conserved (the pivot exerts torque), and no hidden integral exists: by the Liouville–Arnold criterion — $n$ independent commuting integrals for integrability with $n = 2$ here — the double pendulum is non-integrable. Motion is confined to the 3D energy surface $H = E$ but not, in general, to any 2D torus inside it. The room between those dimensions is where the chaotic sea lives, and the KAM theorem governs how much of the surface the surviving tori still occupy at each energy. At small amplitude, integrability is restored approximately: linearizing (3)–(4) about the hanging equilibrium yields the coupled-oscillator problem whose normal modes (Derivation B, for equal masses and lengths) are

\omega_\pm^2 = \frac{g}{\ell}\big(2 \pm \sqrt2\big), \qquad \frac{\theta_2}{\theta_1}\bigg|_{\omega_\mp} = \pm\sqrt2 \qquad (6)

where: $\omega_-$ — the in-phase mode (bobs swing together, lower swinging $\sqrt2$ wider); $\omega_+$ — the out-of-phase mode. Their ratio $\omega_+/\omega_- = 1 + \sqrt2 \approx 2.414$ is irrational: even the linearized motion is quasiperiodic, never exactly repeating — a foreshadowing of the tori to come. For general mass ratio $\mu = m_2/m_1$ (equal lengths) the frequencies are $\omega_\pm^2 = \tfrac{g}{\ell}\big(1 + \mu \pm \sqrt{\mu(1+\mu)}\big)$ . Finally the chaos diagnostics, stated here and developed in the next three articles. The maximal Lyapunov exponent

\lambda = \lim_{t\to\infty}\frac1t \ln\frac{\lVert\delta\mathbf z(t)\rVert}{\lVert\delta\mathbf z(0)\rVert} \qquad (7)

where: $\delta\mathbf z$ — an infinitesimal phase-space separation — is positive on chaotic trajectories and zero on tori. And because the flow is Hamiltonian, the full spectrum is rigidly symmetric:

\{\lambda_1, \lambda_2, \lambda_3, \lambda_4\} = \{+\lambda,\ 0,\ 0,\ -\lambda\}, \qquad \textstyle\sum_i \lambda_i = 0 \qquad (8)

where: the zero sum is Liouville's volume preservation; the two zeros are the directions along the flow and across neighbouring energy surfaces; and the exact $\pm$ pairing is the symplectic structure at work — every stretching direction is mirrored by an equally strong contracting one. Contrast the driven pendulum's $(+0.16, 0, -0.66)$ : dissipation broke the symmetry there; here it is exact.

Which tool answers which question

you want	use	notes
equations of motion	Euler–Lagrange on (2) → (3)–(4)	exact; integrate numerically
small-swing behaviour	normal modes (6)	W2 coupled-oscillator machinery
is the motion regular or chaotic?	Poincaré section at $\theta_1 = 0$ , $\dot\theta_1 > 0$	curves = tori, scatter = chaos
how chaotic?	maximal Lyapunov exponent (7)	Benettin's algorithm; spectrum (8) for free
where chaos must dominate	energy vs flip thresholds	lower-arm flip at $\Delta V = 2m_2g\ell_2$
why no attractor can exist	Liouville volume preservation	Hamiltonian: no transients, recurrence instead

Derivation AFrom geometry to the equations of motion

We are deriving the Lagrangian (2) and the equations of motion (3)–(4) — the exact dynamical content of the double pendulum. The strategy is the standard W3 pipeline: Cartesian positions from the two angles, velocities by differentiation, $L = T - V$ , then Euler–Lagrange for each coordinate. The step to watch is the time derivative of the coupling term: differentiating $\cos(\theta_1 - \theta_2)$ produces velocity-product terms, and a satisfying cancellation leaves each equation with a single clean centripetal coupling $\propto \dot\theta^2$ — miss the cancellation and the equations come out wrong in a way that numerical integration will happily conceal until energies drift.

Step 1. Differentiate the positions (1): $v_1^2 = \ell_1^2\dot\theta_1^2$ , and for the lower bob

v_2^2 = \ell_1^2\dot\theta_1^2 + \ell_2^2\dot\theta_2^2 + 2\ell_1\ell_2\,\dot\theta_1\dot\theta_2\cos(\theta_1 - \theta_2)

Why the cross term: $\dot{\mathbf r}_2$ is the vector sum of the pivot velocity (magnitude $\ell_1\dot\theta_1$ ) and the relative velocity (magnitude $\ell_2\dot\theta_2$ ); the angle between them is $\theta_1 - \theta_2$ , and the law of cosines does the rest — the identity $\cos\theta_1\cos\theta_2 + \sin\theta_1\sin\theta_2 = \cos(\theta_1 - \theta_2)$ in kinematic clothing.

Step 2. Assemble $L = \tfrac12 m_1v_1^2 + \tfrac12 m_2 v_2^2 - g(m_1y_1 + m_2y_2)$ , which is exactly (2). The potential is a sum of single-pendulum terms; every coupling is kinetic.

Step 3. Euler–Lagrange for $\theta_1$ . The momentum is $\partial L/\partial\dot\theta_1 = (m_1{+}m_2)\ell_1^2\dot\theta_1 + m_2\ell_1\ell_2\dot\theta_2\cos(\theta_1{-}\theta_2)$ ; its time derivative brings in $-m_2\ell_1\ell_2\dot\theta_2(\dot\theta_1 - \dot\theta_2)\sin(\theta_1{-}\theta_2)$ from the rotating coupling. The coordinate derivative is $\partial L/\partial\theta_1 = -m_2\ell_1\ell_2\dot\theta_1\dot\theta_2\sin(\theta_1{-}\theta_2) - (m_1{+}m_2)g\ell_1\sin\theta_1$ .

Step 4. Subtract and watch the cancellation: the $\dot\theta_1\dot\theta_2\sin(\cdot)$ terms from the two sources are equal and opposite, leaving

(m_1{+}m_2)\ell_1\ddot\theta_1 + m_2\ell_2\ddot\theta_2\cos(\theta_1{-}\theta_2) + m_2\ell_2\dot\theta_2^2\sin(\theta_1{-}\theta_2) + (m_1{+}m_2)g\sin\theta_1 = 0

Why the survivor is $\dot\theta_2^2$ : the upper rod feels the lower one's rotation only through the centrifugal pull of the lower bob's circular motion about the moving joint — velocity products of mixed origin cancel by the symmetry of the coupling in $\theta_1 - \theta_2$ .

Step 5. Repeat for $\theta_2$ (the common factor $m_2$ divides out):

\ell_2\ddot\theta_2 + \ell_1\ddot\theta_1\cos(\theta_1{-}\theta_2) - \ell_1\dot\theta_1^2\sin(\theta_1{-}\theta_2) + g\sin\theta_2 = 0

with the opposite sign on the centripetal term because the coupling angle enters as $\theta_1 - \theta_2$ . Together with $H = T + V$ conserved (no explicit time dependence), these two ODEs are the complete, exact dynamics. Verified: matches Morin ch. 6 and Goldstein ch. 1; in the small-angle limit the equations reduce to the linear system of Derivation B, and the numerical integrations behind Figures 1–2 conserve $E$ to better than $10^{-9}$ per swing with these forms.

Go deeper · Derivation B · 1 minNormal modes of the equal-mass, equal-length pendulum

Derivation BNormal modes of the equal-mass, equal-length pendulum

The goal is the small-oscillation skeleton (6): the two frequencies and mode shapes that govern every low-energy motion — and that define the unperturbed tori whose KAM fate the rest of the article tracks. The strategy is the generalized eigenvalue problem of the W2 normal-modes articles: linearize the equations of motion, read off constant mass and stiffness matrices, and solve $\det(K - \omega^2 M) = 0$ . The step to watch is that $M$ is non-diagonal — the inertial coupling survives linearization even though the geometric nonlinearity dies — so the modes diagonalize the pair $(M, K)$ , not either matrix alone.

Step 1. Linearize (3)–(4) about $\theta_1 = \theta_2 = 0$ with $m_1 = m_2 = m$ , $\ell_1 = \ell_2 = \ell$ : set $\cos(\theta_1{-}\theta_2) \to 1$ , $\sin\theta_i \to \theta_i$ , and drop all $\dot\theta^2$ terms (quadratic in small quantities):

2\ell\ddot\theta_1 + \ell\ddot\theta_2 + 2g\theta_1 = 0, \qquad \ell\ddot\theta_1 + \ell\ddot\theta_2 + g\theta_2 = 0

Step 2. Identify the matrices: with $\boldsymbol\theta = (\theta_1, \theta_2)^\top$ ,

M = \ell\begin{pmatrix} 2 & 1 \\ 1 & 1 \end{pmatrix}, \qquad K = g\begin{pmatrix} 2 & 0 \\ 0 & 1 \end{pmatrix}

Why $M$ carries the coupling: the cross term in the kinetic energy is quadratic in velocities and survives at leading order, while the potential was never coupled — the opposite arrangement to spring-coupled pendula, where $K$ couples and $M$ is diagonal.

Step 3. Impose the mode ansatz $\boldsymbol\theta = \mathbf a\,e^{i\omega t}$ and solve the characteristic equation:

\det(K - \omega^2 M) = 0 \;\Rightarrow\; \ell^2\omega^4 - 4g\ell\,\omega^2 + 2g^2 = 0

Step 4. Quadratic formula in $\omega^2$ :

\omega^2_\pm = \frac{4g\ell \pm \sqrt{16g^2\ell^2 - 8g^2\ell^2}}{2\ell^2} = \frac{g}{\ell}\big(2 \pm \sqrt2\big)

Both roots positive: the hanging equilibrium is linearly stable, a genuine center in each mode plane.

Step 5. Extract the shapes: substituting back, $a_2/a_1 = \mp\sqrt2$ for $\omega_\pm$ . The slow mode swings the arms in phase with the lower bob sweeping $\sqrt2$ times wider; the fast mode folds them against each other. Frequency ratio $\omega_+/\omega_- = \sqrt{(2+\sqrt2)/(2-\sqrt2)} = 1 + \sqrt2$ — irrational, so generic small motion is a quasiperiodic beat that never repeats: the innocent, integrable-looking ancestor of the tori in Figure 2's first panel. Verified: matches Taylor ch. 11 and Landau–Lifshitz §23; the ratio $(2+\sqrt2)/(2-\sqrt2) = 3 + 2\sqrt2$ and $\sqrt{3+2\sqrt2} = 1+\sqrt2$ are exact; general- $\mu$ formula reduces to these at $\mu = 1$ .

Go deeper · Derivation C · 2 minOne missing integral — non-integrability and the structure of the chaos diagnostics

Derivation COne missing integral — non-integrability and the structure of the chaos diagnostics

We now make precise why the double pendulum is chaotic at all, and how the diagnostics of Figures 1–2 are constructed. The strategy has three moves: count integrals against the Liouville–Arnold requirement; build the Poincaré section as an honest area-preserving return map; and derive the symmetric Lyapunov spectrum (8) from volume preservation. The step to watch is the counting argument — it shows non-integrability is structural (about what conserved quantities exist), not about equations being "too hard to solve."

Step 1. Count. Degrees of freedom: $n = 2$ ; phase space dimension 4. Liouville–Arnold (W4-4) requires two independent integrals in involution for integrability. Available: $H$ only. Angular momentum about the pivot fails — gravity exerts torque; momentum conjugate to nothing else is conserved because the kinetic coupling ties the angles together. Generic motion therefore need not lie on 2-tori. Why one cannot simply hope for a hidden integral: a second analytic integral would force every Poincaré section point to lie on a smooth curve (the section of its joint level set with the energy surface). The computed sections at $E = 2.2$ and $4.0$ show single trajectories filling two-dimensional regions — direct numerical falsification of any globally analytic second integral. (Rigorous non-integrability proofs for specific parameter families use Morales–Ramis differential Galois theory; the numerical verdict is the same.)

Step 2. Build the section. On the energy surface $H = E$ (3D), take the plane $\theta_1 = 0$ with orientation $\dot\theta_1 > 0$ ; record $(\theta_2, \dot\theta_2)$ at each crossing. Fixing $E$ and the crossing convention makes the remaining pair complete coordinates: $p_1$ is recovered from $H = E$ . The return map is area-preserving (a consequence of the symplectic structure restricted to the section) — no attractors, so dots never converge to anything; they either organize on curves (tori) or fill regions (chaos).

Step 3. Read the KAM story in the section. At low $E$ the system is a small perturbation of two uncoupled oscillators; KAM guarantees tori with sufficiently irrational frequency ratios survive — the nested curves at $E = 0.5$ . Resonant tori ( $\omega_1/\omega_2$ rational) break by the Poincaré–Birkhoff mechanism into alternating elliptic/hyperbolic point chains — the islands at $E = 2.2$ — with chaotic layers seeded at the hyperbolic points' tangled separatrices. As $E$ grows the layers merge; above the lower-arm flip threshold the connected sea dominates ( $E = 4.0$ ).

Step 4. Derive the spectrum (8). Liouville: the flow preserves 4-volume, so $\sum_i\lambda_i = 0$ . Two exponents vanish identically: along the flow direction (perturbations that merely slide the trajectory in time neither grow nor decay) and transverse to the energy surface (energy conservation pins that direction). Symplectic pairing forces the remaining two to be exact negatives: $\{+\lambda, 0, 0, -\lambda\}$ .

Step 5. Interpret. A positive $\lambda$ on a sea trajectory quantifies the flip-to-flip unpredictability of Figure 1's trace; $\lambda = 0$ on the tori of the first panel. Plotting $\lambda$ against $E$ maps the order-to-chaos transition as a function on one dial — the measurement carried out in Worked Example 3 and industrialized in the two articles that follow. Verified: the Liouville–Arnold count and torus/section correspondence follow the W4-4 article; area preservation of the return map and the $\{+\lambda,0,0,-\lambda\}$ structure match Lichtenberg & Lieberman; the section phenomenology is reproduced by the direct computations behind Figure 2.

Limiting cases

Case	Result	Meaning
small oscillations	two normal modes, eq. (6)	integrable skeleton; quasiperiodic beats; KAM tori intact
$m_2 \to 0$	upper rod = simple pendulum; lower rod slaved	the lower pendulum becomes a driven pendulum with quasiperiodic drive — chaos in the test-particle limit without back-reaction
$\ell_2 \to 0$ or rods locked	single pendulum of the combined mass	one degree of freedom: integrable, elliptic functions, no chaos possible
$E \ll 2m_2g\ell_2$	sections dominated by KAM curves	chaotic layers exist but are exponentially thin
$E \gtrsim 2m_2g\ell_2$ (lower-arm flip)	connected chaotic sea opens	flips act as decision points; $\lambda > 0$ over most of the surface
$E \to \infty$	gravity negligible; free double rotor	integrability returns (angular momentum conserved without gravity) — chaos lives at intermediate energy
up-up equilibrium	doubly unstable saddle	inverted stabilization needs feedback — the Acrobot control problem

Common trapChaotic is not random, and this chaos has no attractor

Two conflations to refuse. First, the double pendulum's motion is fully deterministic: identical initial conditions give identical futures, forever. Chaos means a positive Lyapunov exponent — exponential amplification of initial differences — not indeterminism; the correct statement is that prediction horizons grow only logarithmically with measurement precision. Second, do not import dissipative vocabulary: there is no strange attractor here, because nothing attracts. Liouville's theorem preserves phase volume; transients never die; a chaotic trajectory wanders its chaotic sea forever and, by Poincaré recurrence, revisits any neighbourhood of its start infinitely often. The chaotic sea is a positive-volume region of the energy surface, not a zero-volume fractal set. (And a third, practical trap: "small oscillations" analysis applies only about the down-down equilibrium — the other three equilibria are saddles, and linearizing there yields growth, not modes.)

Worked examples

1 · The two clocks inside the toy (confidence builder)

A demonstration double pendulum has equal point masses and equal lengths $\ell = 0.30$ m. Find the normal-mode frequencies, the mode shapes, and the beat period when only the upper arm is displaced slightly.

Step 1. From (6) with $g/\ell = 9.81/0.30 = 32.7\ \mathrm{s^{-2}}$ : $\omega_-^2 = 32.7(2 - 1.414) = 19.2$ , so $\omega_- = 4.38\ \mathrm{rad/s}$ ( $f_- = 0.70$ Hz); $\omega_+^2 = 32.7 \times 3.414 = 111.6$ , so $\omega_+ = \boxed{10.6\ \mathrm{rad/s}}$ ( $f_+ = 1.68$ Hz). Ratio $\omega_+/\omega_- = 2.414 = 1 + \sqrt2$ ✓.

Step 2. Shapes: slow mode $\theta_2 = +\sqrt2\,\theta_1$ (in phase, lower arm sweeping 41% wider); fast mode $\theta_2 = -\sqrt2\,\theta_1$ (folding). Displacing only the upper arm excites both modes equally-ish, and energy then beats between the arms.

Step 3. Beat frequency $= (\omega_+ - \omega_-)/2\pi = (10.6 - 4.38)/6.283 = 0.99$ Hz... but note the beat between arms is set by the mode-frequency difference: period $2\pi/(\omega_+ - \omega_-) = 6.283/6.19 = \boxed{1.01\ \mathrm{s}}$ — the familiar demo rhythm of the lower arm "stealing" the swing roughly once a second. Because $\omega_+/\omega_-$ is irrational, the full motion never exactly repeats: quasiperiodicity, visible already in the linear theory.

2 · Energy budget of the routes to chaos (exam level)

The same pendulum has $m_1 = m_2 = 0.10$ kg, $\ell = 0.30$ m. Compute the energy (above the hanging state) needed to (i) flip the lower arm, (ii) flip the upper arm, (iii) reach the fully inverted state — and the energy of the standard "release from horizontal" demonstration. Where does the demo sit on the order-chaos dial?

Step 1. Potential landmarks from $V = -(m_1{+}m_2)g\ell\cos\theta_1 - m_2g\ell\cos\theta_2$ . Hanging: $V_0 = -(0.2)(9.81)(0.3) - (0.1)(9.81)(0.3) = -0.883$ J.

Step 2. Lower-arm flip (down-up configuration): raising $m_2$ by $2\ell$ costs $\Delta E = 2m_2g\ell = 2(0.1)(9.81)(0.3) = \boxed{0.589\ \mathrm{J}}$ . This is the threshold above which the chaotic sea becomes connected — the flip channel opens.

Step 3. Upper-arm flip (up-down): raising both masses' support by $2\ell$ costs $2(m_1{+}m_2)g\ell = \boxed{1.18\ \mathrm{J}}$ . Fully inverted (up-up): $2(m_1{+}m_2)g\ell + 2m_2g\ell - 2m_2g\ell$ ... compute directly: $V_{uu} = +0.883$ J, so $\Delta E = \boxed{1.77\ \mathrm{J}}$ — the ceiling of the potential landscape, a doubly unstable saddle.

Step 4. Release from horizontal ( $\theta_1 = \theta_2 = 90°$ ): $V_h = 0$ , so $E = 0.883$ J above hanging — $1.5\times$ the lower-flip threshold, comfortably inside the chaotic sea but below the upper-flip energy. That is why the classic demo shows wild lower-arm flips within the first seconds while the upper arm merely swings hard: the demo is tuned by geometry to the most chaotic accessible band.

3 · How long can you trust a forecast? (applied)

For the horizontal release, numerical experiments give a maximal Lyapunov exponent $\lambda \approx 2\ \mathrm{s^{-1}}$ (order of magnitude, for tabletop dimensions). If the release angles are set with precision $10^{-3}$ rad, when does the forecast fail? What does $10\times$ better preparation buy?

Step 1. Error growth: $\delta(t) = 10^{-3}e^{2t}$ rad reaches $\sim 1$ rad when $t = \tfrac12\ln 10^{3} = \boxed{3.5\ \mathrm{s}}$ — three or four swings, matching the observed moment when two "identical" demos part ways at a flip.

Step 2. Tenfold better preparation ( $10^{-4}$ rad): $t = \tfrac12\ln 10^4 = 4.6$ s — a single extra second. Each factor of 10 in precision buys $\ln 10/\lambda = 1.15$ s. Micron-and-microradian heroics would still not carry a forecast past ten seconds: the logarithm is merciless.

Step 3. The full spectrum for this trajectory is $\{+2, 0, 0, -2\}\ \mathrm{s^{-1}}$ by (8) — the contraction direction is exactly as strong as the stretching one, and phase volume is conserved even as neighbouring trajectories flee each other. Chaos with a balanced budget.

Origins · 1 minGeorge David Birkhoff and how this was discovered

George David Birkhoff

1884–1944 · Harvard

Birkhoff was the first mathematician to make Poincaré's tangled geometry into a working American school of dynamics. In 1913 he proved "Poincaré's last geometric theorem" — the fixed-point theorem Poincaré had published, unproven, months before his death — which guarantees that area-preserving twist maps of an annulus carry infinitely many periodic orbits: exactly the mechanism by which the double pendulum's resonant tori shatter into the island chains of Figure 2 (now called Poincaré–Birkhoff chains). His 1931 ergodic theorem gave the other half of this article its meaning: for measure-preserving flows, time averages exist and equal space averages on ergodic components — the precise sense in which a sea trajectory "explores everything available to it." Birkhoff's Dynamical Systems (1927) organized the field's vocabulary — recurrence, transitivity, minimal sets — decades before computers could draw a single section. The double pendulum is a museum of his theorems: island chains by Poincaré–Birkhoff, sea statistics by the ergodic theorem, and the whole subject's insistence that where equations cannot be solved, geometry and measure still rule.

History

1890

— Poincaré's memoir on the three-body problem exposes homoclinic tangles: two degrees of freedom suffice for unsolvable dynamics.

1913

— Birkhoff proves Poincaré's last geometric theorem; resonant island chains in area-preserving maps become a theorem.

1954

— Kolmogorov announces the persistence of invariant tori under small perturbations; Arnold and Moser complete the KAM proof over the next decade.

1964

— Hénon and Heiles publish the first computed Poincaré sections of a two-degree-of-freedom system, showing tori and chaos coexisting exactly as the double pendulum's sections do.

1980

— Benettin, Galgani, Giorgilli and Strelcyn give the standard algorithm for computing full Lyapunov spectra, making the order-chaos dial quantitative.

Going further

References. Morin, Introduction to Classical Mechanics, Ch. 6 (Lagrangian mechanics) and normal-modes chapter · Goldstein, Poole & Safko, Classical Mechanics, 3rd ed., Ch. 1–2, Ch. 11 (chaos) · Landau & Lifshitz, Mechanics, Vol. 1, §5 (small oscillations), §23–24 · Taylor, Classical Mechanics, Ch. 11 (coupled oscillators), Ch. 12 (nonlinear mechanics and chaos) · Lichtenberg & Lieberman, Regular and Chaotic Dynamics, 2nd ed.

Provenance. Equations of motion and Lagrangian verified against Morin Ch. 6 and Goldstein Ch. 1; a verifier should re-derive the two Euler–Lagrange equations and confirm the cancellation of $\dot\theta_1\dot\theta_2$ cross terms. Normal-mode frequencies $\omega_\pm^2=(g/\ell)(2\pm\sqrt2)$ for equal masses/lengths should be recomputed from $\det(K-\omega^2M)=0$ ; the general $\mu$ formula matches Landau §23 and Taylor Ch. 11. The Lyapunov spectrum structure $\{+\lambda,0,0,-\lambda\}$ and $\sum\lambda_i=0$ follow from Liouville's theorem (Lichtenberg & Lieberman). Frequency ratio $1+\sqrt2$ and energy ratio $3+2\sqrt2$ are exact algebraic values.

Builds on: Conservative Systems and Hamiltonian Phase Flow · Fixed Points and Equilibria · Linear Stability Analysis and the Jacobian · Classification of Fixed Points: Nodes, Saddles, Spirals, Centers · Phase Portraits and the Phase Plane Leads to: Period-Doubling Cascade and the Route to Chaos · Bifurcations and the Loss of Stability · Poincaré–Bendixson Theorem

Has the double pendulum's non-integrability actually been proven, or is it just numerical folklore?

Both more and less than folklore. The numerical evidence is overwhelming and, in one precise sense, decisive: a single trajectory filling a two-dimensional region of the Poincaré section is incompatible with a globally analytic second integral, whose level sets would confine section points to curves. Rigorous proofs exist but are parameter-specific: using Morales–Ramis theory (differential Galois analysis of the variational equations along particular solutions), non-integrability has been established for wide ranges of mass and length ratios of the planar double pendulum. What no theorem does is certify chaos at your particular energy and parameters — KAM guarantees order in part of phase space at low energy, and positive Lyapunov exponents are measured, not derived. The honest summary: non-integrability is proven where checked, chaos is measured wherever sought, and no one has ever found — nor could find — the missing integral.

Why exactly two zero Lyapunov exponents?

Each zero has a distinct geometric origin. The first is universal for any bounded trajectory of an autonomous flow: perturb along the velocity direction and you get the same orbit shifted in time — the separation grows only linearly (from period mismatch), so the exponent is zero. The second is bought by energy conservation: perturbing across energy surfaces changes $E$ slightly, putting the neighbour on an adjacent invariant surface it can never leave; separations transverse to the surface stay bounded, another zero. In a dissipative system the second mechanism is absent (no conserved $E$ ), which is why the driven pendulum had spectrum $(+, 0, -)$ with a single zero. Count zeros and you are reading off the flow's conserved structure — one per continuous symmetry/integral plus one for time translation along the orbit.

If phase volume is conserved, what plays the role that the strange attractor played in the driven pendulum?

The chaotic sea itself — a positive-volume invariant region, not a zero-volume attracting set. In dissipative chaos, contraction squeezes all transients onto a fractal attractor and the interesting measure lives there. In Hamiltonian chaos nothing is squeezed: the sea is a fat subset of the energy surface on which the dynamics is (approximately) ergodic and mixing — time averages equal phase-space averages over the sea, and an initially small blob of uncertainty spreads filament-by-filament until it wraps the entire sea at constant volume, like dye stirred into batter. The fractal geometry has not vanished, though: it migrates to the boundaries — island shorelines, cantori (broken tori acting as leaky barriers), and the stickiness of trajectories near them, which produces the long memory and anomalous statistics characteristic of Hamiltonian seas.

Is the low-energy motion truly regular, or just chaotic too slowly to notice?

Strictly speaking, the second — but with "too slowly" so extreme it deserves the name order. KAM guarantees that a large-measure set of tori survives at small energy, and motion started on those tori is exactly quasiperiodic forever. Between the tori, arbitrarily thin chaotic layers exist at every energy (around every resonance's broken separatrix), so a section computed at infinite resolution would always show some scatter. But the layer widths shrink exponentially with the perturbation strength, and — crucially — in two degrees of freedom the surviving tori enclose the chaotic layers: a trajectory born in a thin layer is trapped between neighbouring tori and can never wander far. So low-energy chaos is real but confined and exponentially thin. Only as energy grows and tori break wholesale do the layers connect into a sea — and only in three or more degrees of freedom can chaotic trajectories sneak past tori entirely (Arnold diffusion), a subtlety beyond this toy.

Why is over-the-top flipping of the lower arm such a good marker for the onset of global chaos?

Because a flip requires crossing the separatrix of the lower arm's own pendulum motion, and separatrices are where chaos is born. Near the separatrix the unperturbed period diverges, so the motion becomes infinitely sensitive to the phase of the coupling "kicks" from the upper arm — the resonance-overlap mechanism that shreds tori operates maximally there. Below the flip energy $2m_2g\ell_2$ , libration and rotation are separated by intact barriers and chaotic layers stay local; once the flip channel opens energetically, the layers around the librating and rotating separatrices merge into a connected sea, and the qualitative character of typical trajectories changes from "beats with jitter" to "flip roulette." The much-shared plots of flip time versus release angles — with their intricate fractal boundaries between flip-now and flip-later regions — are pictures of exactly this: the sea's fractal shoreline rendered in initial-condition space.

Do real demonstration pendula, with distributed-mass rods and bearing friction, obey this article?

Qualitatively yes, with two honest amendments. Distributed mass changes coefficients, not structure: a compound double pendulum has the same configuration space, the same $\cos(\theta_1 - \theta_2)$ kinetic coupling (now with moments of inertia in the matrix), the same one-integral count, and hence the same KAM-to-sea storyline; normal-mode frequencies and flip thresholds shift numerically. Friction is the more principled amendment: real bearings make the system weakly dissipative, so strictly it has no invariant tori and every demo ends at the hanging fixed point. But on the timescale of interest — dozens of swings, with $Q$ in the hundreds — the motion shadows the conservative dynamics closely: sections computed from real high-speed-camera data show the KAM structures blurred, not erased. The right mental model is a Hamiltonian phase portrait slowly leaking downhill in energy: it crosses the whole order-chaos dial of this article on its way to rest, which is exactly what makes the demo so watchable.

What happens in the limit where the lower mass is tiny — and why is that limit interesting?

Take $m_2 \to 0$ with lengths fixed. The upper pendulum no longer feels the lower one (all coupling terms carry $m_2$ ), so it swings as an exact simple pendulum — integrable, periodic. But the lower rod still feels its pivot moving: it becomes a pendulum hanging from a prescribed oscillating support — a parametrically and directly driven pendulum with zero back-reaction. This test-particle limit is a conceptual bridge to the previous article: chaos survives at $m_2 = 0^+$ in the lower arm's dynamics (a periodically driven pendulum with effectively 1.5 degrees of freedom), even though the "system" now conserves the upper arm's energy separately. It illustrates a recurring theme of Hamiltonian chaos: what matters is the dimension and structure of the dynamics a trajectory experiences, not the bookkeeping of which parts carry mass. It is also, historically, how restricted problems (the restricted three-body problem above all) let Poincaré find chaos with pencil and paper resources.

The Double Pendulum

Where this lives today

The physical picture

The mechanics, assembled

Which tool answers which question

Limiting cases

Worked examples

1 · The two clocks inside the toy (confidence builder)

2 · Energy budget of the routes to chaos (exam level)

3 · How long can you trust a forecast? (applied)

Going further

Before this

Where this leads