Double Pendulum Chaos

Definition

A double pendulum is one pendulum hung from the bottom of another. Two rigid arms, two pivots, two masses — and exactly two angles needed to describe everything: θ₁ (the upper arm from vertical) and θ₂ (the lower arm from vertical).

That sounds almost trivial. It is not. The double pendulum is the simplest mechanical system most people can build at home that is genuinely chaotic: its motion is generated by clean, deterministic equations with no randomness at all, yet it is effectively unpredictable. It has become the canonical desktop demonstration of deterministic chaos for exactly this reason.

The defining behavior: take two double pendulums and release them from almost the same position — say, a thousandth of a degree apart. For a few swings they move as one. Then, abruptly, they diverge and end up doing completely unrelated things. That runaway separation is sensitive dependence on initial conditions, and its rate is set by a positive Lyapunov exponent.

How it works

Start with what's special by comparison. A single pendulum is integrable: it has one degree of freedom and one conserved quantity (energy), so its state is pinned to a single closed curve in phase space. Its future is perfectly predictable; small errors stay small. Nothing chaotic can happen.

Now bolt a second arm to the bottom. You add a second angle, a second velocity — but you do not add a second conservation law. The double pendulum has two degrees of freedom and only one conserved quantity (energy, assuming no friction). It is non-integrable. With nothing forcing the trajectory onto a tidy surface, the path through phase space is free to stretch and fold: nearby points get pulled apart in one direction while the whole sheet is folded back to stay in a bounded region. Stretching is what makes errors grow; folding is what keeps the motion confined and re-mixes the trajectory so it never repeats.

The coupling is also nonlinear. The torque each arm exerts on the other depends on sin(θ₁ − θ₂), cos(θ₁ − θ₂), and the square of the angular velocities. Those nonlinear terms are not a detail — they are the engine of chaos. Linearize them (the small-angle approximation) and the double pendulum collapses into two ordinary coupled oscillators with two regular normal modes and zero Lyapunov exponent. The chaos lives entirely in the terms the small-angle approximation throws away.

Put those pieces together and you get the headline property. Define the gap between two trajectories as δ(t). In a chaotic regime it grows exponentially on average:

δ(t) ≈ δ₀ · e^(λt)        (λ > 0  ⟹  chaos)

λ is the largest Lyapunov exponent. Its sign is the rigorous, quantitative definition of chaos: λ > 0 means exponential separation; λ = 0 means regular; λ < 0 means trajectories converge. For an energetic double pendulum, λ is positive — typically a few inverse seconds.

A worked example — how fast does the error blow up?

Suppose two pendulums start a hair apart: an angular gap of δ₀ = 0.001 radian (about 0.057°). Take a representative largest Lyapunov exponent of λ = 3 s⁻¹ for an energetic, over-the-top swing. How long until the gap reaches order 1 radian — i.e., the two pendulums are doing visibly different things?

δ(t) = δ₀ · e^(λt)
1 rad = 0.001 · e^(3·t)
e^(3t) = 1000
3t = ln(1000) ≈ 6.908
t ≈ 2.30 seconds

So a gap a thousand times smaller than a degree explodes to full-scale disagreement in just over two seconds. That single number captures the entire personality of the system.

Now the brutal part. Suppose you measure the starting angle a thousand times more precisely: δ₀ = 0.000001 radian. How much longer can you predict?

e^(3t) = 1,000,000
3t = ln(10⁶) ≈ 13.816
t ≈ 4.61 seconds

A 1000× better measurement bought you 2.3 extra seconds — exactly (1/λ)·ln(1000), a constant, not a multiplier. This is the defining frustration of chaos: predictability grows only with the logarithm of your precision. To double your forecast horizon you must square your accuracy. Deterministic, yes — predictable, no.

Regimes — the same device is regular or chaotic depending on energy

The double pendulum is not chaotic by nature; it is chaotic by energy. The total energy you start it with selects the behavior:

Energy regime	Motion	Phase space	Largest Lyapunov exponent λ	Predictable?
Very low (tiny swings)	Two near-linear normal modes (in-phase & out-of-phase)	Smooth tori (KAM curves)	≈ 0	Yes — effectively integrable
Low–moderate	Mostly regular, thin chaotic seams near separatrix	Mixed: tori + thin chaotic layers	Small, near 0	Mostly
Moderate	Quasi-periodic with occasional erratic excursions	Shrinking tori, growing chaotic sea	Small positive	Short-term only
High (lower arm flips over)	Full chaos — flips, tumbles, no repetition	Mostly chaotic sea	Clearly positive (~ few s⁻¹)	No
Very high	Violent tumbling, frequent flips	Almost entirely chaotic	Large positive	No
Driven / damped variants	Limit cycles, strange attractors, intermittency	Attractor (fractal at chaos)	Tunable by drive	Depends on drive

This is the cleanest lesson the toy teaches: chaos is not a property of the equations alone but of where in their phase space you live. Turn the energy knob and you slide continuously from clockwork to chaos.

The equations and why there's no formula

Using the angles θ₁, θ₂ as generalized coordinates, you write the kinetic energy T and potential energy V of the two bobs, form the Lagrangian L = T − V, and apply the Euler–Lagrange equations. The output is a pair of coupled, second-order, nonlinear ODEs for the angular accelerations. Schematically (equal lengths ℓ, equal masses m, gravity g), the first reads:

θ₁'' =  [ −g(2m)·sinθ₁ − m·g·sin(θ₁−2θ₂)
          − 2·sin(θ₁−θ₂)·m·(θ₂'²·ℓ + θ₁'²·ℓ·cos(θ₁−θ₂)) ]
        / [ ℓ·(2m − m·cos(2θ₁−2θ₂)) ]

(θ₂'' is similar.) Two features matter. First, the sines and cosines of (θ₁ − θ₂) and the velocity-squared terms make the system nonlinear — drop them via small angles and chaos vanishes. Second, there is no closed-form solution. You cannot write θ₁(t) as elementary functions. The only way forward is numerical integration — and even that runs into the Lyapunov wall.

JavaScript — integrating and measuring the divergence

// State: [theta1, theta2, omega1, omega2]
const g = 9.81, m1 = 1, m2 = 1, L1 = 1, L2 = 1;

function deriv([t1, t2, w1, w2]) {
  const d = t1 - t2;
  const den1 = (m1 + m2) * L1 - m2 * L1 * Math.cos(d) * Math.cos(d);
  const den2 = (L2 / L1) * den1;
  const a1 = (m2 * L1 * w1 * w1 * Math.sin(d) * Math.cos(d)
            + m2 * g * Math.sin(t2) * Math.cos(d)
            + m2 * L2 * w2 * w2 * Math.sin(d)
            - (m1 + m2) * g * Math.sin(t1)) / den1;
  const a2 = (-m2 * L2 * w2 * w2 * Math.sin(d) * Math.cos(d)
            + (m1 + m2) * (g * Math.sin(t1) * Math.cos(d)
            - L1 * w1 * w1 * Math.sin(d) - g * Math.sin(t2))) / den2;
  return [w1, w2, a1, a2];
}

// 4th-order Runge–Kutta — accurate enough to expose chaos, not defeat it
function rk4(s, h) {
  const add = (a, b, k) => a.map((v, i) => v + b[i] * k);
  const k1 = deriv(s);
  const k2 = deriv(add(s, k1, h / 2));
  const k3 = deriv(add(s, k2, h / 2));
  const k4 = deriv(add(s, k3, h));
  return s.map((v, i) => v + (h / 6) * (k1[i] + 2 * k2[i] + 2 * k3[i] + k4[i]));
}

// Two pendulums, identical except for a 1e-3 rad nudge on theta1
let A = [2.0, 2.0, 0, 0];
let B = [2.0 + 1e-3, 2.0, 0, 0];
const h = 0.002;
for (let step = 0; step <= 2000; step++) {     // 4 seconds
  if (step % 250 === 0) {
    const gap = Math.hypot(A[0] - B[0], A[1] - B[1]); // angular separation
    console.log(`t=${(step * h).toFixed(2)}s  gap=${gap.toExponential(2)} rad`);
  }
  A = rk4(A, h);
  B = rk4(B, h);
}
// Gap starts ~1e-3, grows roughly e^(λt), and saturates near ~1 rad
// once the two pendulums are doing entirely different things.

Run it and the printed gap climbs almost geometrically, then plateaus once the trajectories are unrelated — the numerical fingerprint of a positive Lyapunov exponent. Fit a line to ln(gap) versus t over the growth window and the slope is an estimate of λ.

Where this shows up

• The butterfly effect, made tangible. Edward Lorenz's 1963 weather model is chaotic for the same reason: positive Lyapunov exponent. The double pendulum is the hand-held version — same mathematics, no atmosphere required.
• Weather and climate forecasting limits. The atmosphere's Lyapunov time is roughly a couple of weeks, which is precisely why forecasts decay past ~10 days no matter how good the model or the satellites. The double pendulum makes the logarithmic-precision wall obvious in seconds.
• Robotics and the acrobot. An underactuated two-link arm (the "acrobot") is a driven double pendulum. Controlling it to swing up and balance is a classic control-theory benchmark precisely because the uncontrolled dynamics are chaotic.
• Spacecraft and orbital mechanics. The three-body problem shares the double pendulum's non-integrability; chaotic transport along these tangled trajectories is what enables fuel-cheap "interplanetary superhighway" transfers.
• Validating chaos algorithms. Because its dynamics are simple to code yet genuinely chaotic, the double pendulum is a standard test case for Lyapunov-exponent estimators and symplectic integrators.
• Teaching nonlinear dynamics. Phase portraits, Poincaré sections, KAM tori, and the route to chaos can all be demonstrated on this one device — which is why it sits in nearly every dynamics classroom and science-museum lobby.

Common pitfalls and misconceptions

• "It's random." It is not. There is zero randomness in the equations. Same exact start ⟹ same exact future, every time. It only looks random because you can never reproduce the start exactly, and tiny differences explode. Deterministic, not random is the whole point.
• "Chaos means anything can happen." No. The motion is bounded by energy conservation — the bobs can't leave a fixed region, and the trajectory is confined to a constant-energy surface. Chaos is bounded unpredictability, not lawlessness.
• "A better computer would let us predict it forever." Predictability scales with the logarithm of precision. Going from 32-bit to 64-bit floats adds only a fixed few seconds of accurate forecast (≈ 29·ln2/λ). You can't out-compute a positive Lyapunov exponent.
• "Two simulations of the same start should match forever." Different machines, compilers, or even instruction orderings round differently in the last bit. That sub-bit difference is an initial-condition perturbation, so the two runs diverge — exactly as two physical pendulums do.
• "It's always chaotic." Only above a threshold energy. Small-swing motion is essentially two regular normal modes with λ ≈ 0. Energy is the on/off switch.
• "Energy isn't conserved — look how wild it is." Energy is conserved (frictionless ideal). Wildness in shape of motion is fully compatible with a fixed total energy. If your simulation's energy drifts, that's integrator error, not physics — use a symplectic method or a smaller step.

Predictability analysis — the math of the forecast horizon

Quantify how long you can trust a forecast. Let δ₀ be your initial uncertainty and δ_max the tolerance at which the prediction is useless (often order 1 in the relevant variable). The trajectories diverge until δ₀·e^(λT) = δ_max, giving the predictability horizon:

T_predict ≈ (1/λ) · ln(δ_max / δ₀)

Three consequences fall straight out:

• Logarithmic, not linear, payoff. Improving δ₀ by a factor of 10 adds only ln(10)/λ ≈ 2.3/λ seconds — the same fixed amount each time, no matter how good you already are.
• The Lyapunov time sets the scale. τ = 1/λ is the natural clock of the system. Useful prediction lasts a small multiple of τ; beyond a handful of Lyapunov times, all bets are off.
• Bits buy seconds. Floating-point precision is just δ₀ in disguise. Each extra bit halves δ₀, adding ln2/λ ≈ 0.69/λ seconds. With λ = 3 s⁻¹, every bit is worth ~0.23 s of forecast — and 64-bit doubles only beat 32-bit floats by ~29 bits, or about 6.7 extra seconds. That's the hard ceiling determinism cannot lift.

This is the deep payoff of the desktop toy: it turns an abstract inequality (λ > 0) into something you can watch unfold in your hand. The pendulum obeys the laws of physics perfectly — and is unpredictable anyway. Both statements are true at once, and the double pendulum is the simplest place to feel why.