The Machine That Misplaces Tomorrow: A Double-Pendulum Atlas of Chaos

Two weights hang from two rods. The arrangement looks less like a portal into difficult mathematics than something a school cupboard might contain beside a cracked metre rule. Pull one weight aside, release it, and the machine begins with a few courteous swings. Then a rod turns over, the lower weight changes its mind in mid-air, and tomorrow becomes expensive to calculate.

Nothing random has been added. There is no gust generator, hidden dice cup, or mischievous number source. Given the same four starting numbers and the same physical parameters, the equations produce the same trajectory again. The trouble is that an exact starting state is a mathematical luxury. A real angle is measured to finite precision; a computer stores a finite approximation; two apparently identical releases are never perfectly identical.

The guided experiment below makes that small inconvenience visible. It copies one future, alters an angle by exactly 0.000001 degrees, and evolves both copies with the same fixed-step fourth-order Runge–Kutta calculation. At first they occupy the same pixels. Later they have separate appointments.

Guided observation · fixed-step RK4

Two futures, one invisible nudge

Five short observations. Scroll normally, or use the stage controls. Nothing here captures the page or changes the physics clock to suit the animation.

OBSERVATION 01 NO RANDOM INPUT

Simulated time

0.00 s

Lower-bob separation

0 m

Integrator

RK4 · Δt 1/240 s

The simulation uses equal one-kilogram point masses, equal one-metre massless rods, gravity of 9.81 metres per second squared, no friction, and angles measured from downward vertical. The shadow differs only by the displayed perturbation from observation three onward.

Primary · solid circle Shadow · dashed diamond

01 / 5

One machine

“No dice were thrown.”

One state waits at 120° and −10°. The equations, not a random source, decide what follows.

02 / 5

One future, copied

“Same masses. Same lengths. Same gravity. Same four starting numbers.”

The cool dashed shadow is an exact state copy. At reset, its lower bob occupies the same point.

03 / 5

The nudge

“One angle has moved by less than your cursor can sensibly point.”

Only θ₁ changes: exactly 0.000001°. The inset magnifies that offset ten million times; the calculation does not.

04 / 5

Agreement

“For a while, the futures agree.”

Both states advance with the same fixed 1/240 s RK4 step. Their physical separation remains in the readout.

05 / 5

Loss

“Then the machine misplaces tomorrow.”

The calculation has no random branch. The visible split is the evolved consequence of the original one-millionth-degree difference.

The full double-pendulum laboratory

The guided sequence keeps its hands off the controls so that one fact can arrive without an instrument panel falling on it. The laboratory below is the less polite part. Drag a bob, change masses and lengths, pause, single-step, compare nearby futures, build a prediction-horizon atlas, inspect phase space, and ask a deliberately bad integrator to incriminate itself.

The simulation clock is not the animation clock. Rendering follows the browser, but the physical state advances in fixed increments of 1/240 simulated second. A slow display may draw fewer intermediate pictures; it does not receive different equations. Named presets are starting points rather than claims about every double pendulum, and shared settings reproduce the same numerical experiment.

Calibrating rods, clocks, and numerical honesty…

Interactive · deterministic mechanics

The Pendulum Laboratory

Two rods, four state numbers, one lawful machine. Computed locally in your browser.

Hide controls Shortcuts

Pendulum LabLabShadow FuturesShadowsPrediction Horizon AtlasAtlasPhase-Space ObservatoryPhase space

Drag a mass · release from rest

Pause Reset Single step Clear trails Speed0.1×0.25×0.5×1×2×4×

Simulated time0.00 s

Wall-clock active0.0 s

Energy error0

IntegratorRK4 · 1/240 s

θ₁ / θ₂ now120.0° / -10.0°

EngineRunning at 1× simulated speed.

Numerical Honesty

A browser simulates another dynamical system laid over the physical one. A poor method or oversized step can invent energy, destroy it, or manufacture behaviour that belongs to the algorithm rather than the pendulum.

Open The Lying Integrator

Share setup Share this instant Download PNG Download state Copy data

The laboratory is initializing.

Deterministic is not predictable

Determinism is a rule about succession: if the state and equations are exact, the next state follows. Predictability is a practical claim about knowledge: can we specify the present accurately enough, calculate reliably enough, and remain usefully close for the period we care about? The first statement can be true while the second collapses.

Imagine two addresses written to many decimal places. The last digit differs. A tame system may keep the corresponding futures close, as two buses travelling along neighbouring lanes. A sensitive trajectory repeatedly stretches that small difference. The buses agree through several junctions, then one takes a flyover and the other discovers Behala at six in the evening.

That is sensitivity to initial conditions. It does not say that every small difference instantly becomes enormous, that every initial state is equally chaotic, or that equations have ceased to operate. Nearby trajectories often remain close for a while. Their later separation is produced by the same lawful dynamics that produced their early agreement.

A prediction horizon is therefore not the moment causality resigns. It is the time at which a stated uncertainty grows beyond a stated tolerance. A laboratory technician predicting the lower bob to within ten centimetres has a different horizon from someone asking only whether the apparatus remains attached to the wall. Horizon requires an initial uncertainty, a separation measure, and a threshold; without those, it is merely a dramatic noun.

Four numbers hiding behind two rods

The moving picture lives in a plane, but its instantaneous state needs four coordinates:

The two theta values are rod angles measured from the downward vertical. The two omega values are their angular velocities. Position alone is insufficient: a weight passing the same point upwards and downwards has two different immediate futures. Masses, rod lengths, and gravitational acceleration are parameters of the model rather than extra moving coordinates.

The rods are coupled. Moving the lower mass changes the torque felt by the upper assembly; moving the upper rod carries the lower pivot through space. Sines, cosines, products of velocities, and the angle difference all meet in the acceleration equations. This nonlinearity allows nearby states to be stretched, folded, brought close, and stretched again. The screen shows two rods. The calculation carries a path through four-dimensional phase space.

Angles wrap when a rod crosses a full turn, but the mechanism does not jump. A displayed angle can move from just below 180 degrees to just above −180 degrees while the physical direction changes only slightly. Comparisons must use the shortest angular difference rather than mistaking the label boundary for a two-pi catastrophe.

Shadow Futures

Shadow Futures performs a controlled duplication. Both pendulums receive the same masses, lengths, gravity, integrator, timestep, and angular velocities. One selected state coordinate receives a small perturbation. The guided experiment’s one-millionth of a degree is about 1.745 × 10⁻⁸ radians; the open Lab begins at 10⁻⁷ radians and lets you vary the nudge. Both are physically invisible at this scale, which is why a magnified inset shows the offset without pretending the mechanism itself has moved by that much on screen.

The useful readout is lower-bob separation in metres. The one-centimetre and ten-centimetre markers turn “they look different now” into reproducible questions. Reduce the perturbation by powers of ten and the initial agreement usually lasts longer along a sensitive trajectory. Usually is doing work there: finite-time behaviour can be uneven, and nearby starts can wander through regions with different local stretching.

The laboratory can also estimate a finite-time Lyapunov growth rate. In broad terms, it watches how the distance between nearby states changes, periodically rescales a shadow separation so that it remains numerically useful, and accumulates the logarithmic growth. A positive estimate over the chosen interval suggests exponential separation in that experiment. It is not labelled the Lyapunov exponent of all double pendulums, because the result depends on the trajectory, interval, perturbation norm, and renormalisation procedure.

No randomness is needed for the shadows to part. If the complete configuration is reset, the same calculated parting happens again.

The Prediction Horizon Atlas

One trajectory is an anecdote. The Prediction Horizon Atlas asks the same question across a grid of starting angles. For every atlas cell, it launches a base state and a nearby shadow under the displayed parameters, then records the first simulated time at which the selected physical-separation threshold is crossed. A dark long-horizon island means “these two nearby trajectories remained within this tolerance for this experiment and calculation window.” It does not mean “non-chaotic forever.” A bright short-horizon seam is evidence of rapid finite-time separation, not a moral judgement on that coordinate.

This distinction matters because the angle plane is not a universal chaos map. Change the masses, lengths, initial angular velocities, perturbation, threshold, integration time, or numerical resolution and you have asked a different question. Cells that reach the maximum duration are censored at that limit rather than promoted to proof of eternal regularity. The atlas legend keeps the experiment definition beside the colour scale for precisely this reason.

The atlas reveals structure that a single flailing mechanism conceals: broad quiet regions, tangled borders, and neighbouring cells with sharply different horizons. Select a cell to send its actual starting angles into the laboratory. Then choose a nearby cell across a seam. The comparison turns an image back into two inspectable trajectories rather than leaving the heatmap to cultivate mystique on its own.

Phase space: where the missing dimensions went

A phase portrait plots one state coordinate against another, such as angle against angular velocity. Time is no longer an axis; it is the order in which the trace is drawn. A regular oscillation may form a compact loop. A rotating rod crosses wrapped angle boundaries. A more complicated coupled motion can thread a densely folded region without being random.

Four dimensions will not fit honestly on a flat display, so every two-dimensional phase portrait is a projection. Switching among coordinate pairs changes which structure is visible and which has been laid on top of itself. This is not a defect peculiar to the interface. It is the ordinary price of looking at a higher-dimensional state through a two-dimensional window.

A Poincaré section becomes more selective. Instead of plotting every integration step, it takes attendance when the trajectory crosses a chosen surface in a chosen direction—for example, when the second angle crosses zero while its angular velocity is positive. The crossing time is interpolated between fixed integration samples rather than assigned to whichever sample landed nearest the surface. Regular motion can leave curves or small collections of points; chaotic motion can populate broader patterned regions. The section removes one continuous dimension and much visual clutter, but its appearance still depends on the selected surface and accumulated run time.

Numerical honesty and the lying integrator

The physical model is one dynamical system. The simulator is another machine laid over it, replacing continuous evolution with arithmetic steps. If that arithmetic is careless, the computer can manufacture energy, erase energy, or introduce motion that belongs to the algorithm.

The main laboratory uses fixed-step fourth-order Runge–Kutta, commonly shortened to RK4. Within each timestep it samples the derivative at four intermediate locations and combines them. RK4 is not exact, but at the laboratory’s default step it is a defensible general-purpose method for these trajectories. Mechanical energy is calculated from the same state, and relative drift from the starting energy remains visible as a numerical health check.

The Lying Integrator is an opt-in comparison. It evolves a second copy with explicit Euler using the same nominal timestep. Euler leans in the direction of the derivative measured only at the beginning of each step; on this problem it can accumulate conspicuous energy error. When its trajectory departs from RK4, that is not a second valid prophecy about the physical pendulum. It is an exhibit about numerical method.

This also prevents a common confusion. Physical chaos can amplify the tiny, unavoidable differences between two competent calculations, so long-term trajectories may disagree even when short-term accuracy and energy behaviour are good. Numerical incompetence can cause disagreement much earlier for entirely avoidable reasons. Chaos is real. Numerical incompetence is also real. They should not be allowed to borrow one another’s trousers.

What to try

1. Start with The Polite Oscillator, then increase one angle until the motion begins turning over.
2. Open Shadow Futures and reduce the angular perturbation by powers of ten. Compare the one-centimetre and ten-centimetre crossing times.
3. Find a long-horizon island in the atlas, then inspect a nearby short-horizon seam by launching both cells.
4. Change the lower mass and regenerate the atlas. The old colours described the old experiment.
5. Let a trajectory run, then inspect its phase portrait and a directional Poincaré section.
6. Compare RK4 with Euler in The Lying Integrator and watch both the mechanism and relative energy error.

Technical notes and assumptions

Internally, the model uses SI units, radians, and downward vertical as zero angle. The moving geometry converts the calculated angles into bob positions only when drawing. Trails are bounded, animation pauses outside the viewport or in a hidden tab, and the device-pixel ratio is capped so an unusually dense phone screen does not purchase four times the heat for no additional understanding.

Finite precision never grants permission to call the result random. It asks for more careful language: deterministic equations, a specified numerical method, a finite prediction horizon, and conclusions attached to the experiment that produced them. The machine has not misplaced causality. It has misplaced our useful timetable.