The Pendulum That Stands on Its Head
A pendulum has exactly one place it will rest: hanging straight down. The upside-down position is the textbook example of unstable — the faintest breath topples it. Yet a small, fast vertical shake can make a second stable resting place appear, pointing straight up. Place the bob near upright and small disturbances can stay bounded. Below you can build the effect and compare the motion with the leading-order averaged threshold.
Averaged potential U(θ)
The leading-order approximation to the slow landscape. Down (θ=0) is always a valley. The fast shake adds the green bump that, past threshold, digs a second valley at the top (θ=180°). The dot is the bob now.
Motion is paused because your system requests reduced motion. Press Stand it up or ▶ Play to run the simulation when you choose to.
At the default settings, press Stand it up to place the bob near upright. Pressing Knock it gives a strong impulse and can topple it. Try a lower drive, then stand it up again. The line and badge use the leading-order averaged prediction for a small, fast shake; they do not measure the trajectory's stability across the full control range.
What you are looking at
The setup is the simplest in mechanics: a rigid rod, a weight on the end, a pivot at the top. Leave it alone and it hangs. Lift it to the side and it swings. There are only two places it can balance without moving — hanging down, which is stable, and pointing up, which is not. The upward balance is the standard picture of instability: it exists in principle, but any real disturbance grows, and over it goes.
Now grab the pivot and shake it straight up and down, fast, with a small amplitude. Place the bob near upright first. In the stable regime it can stay inverted, oscillating about vertical after a small disturbance. Shaking does not lift a bob that is exactly at rest hanging down in this model, and a strong knock can topple an upright bob. This is Kapitza's pendulum, described by A. Stephenson in 1908 and explained by the physicist Pyotr Kapitza in 1951.
Why shaking helps — the effective potential
Nothing is balancing on the fast jiggle in the way a broom balances on a fingertip. The trick is
subtler and deeper. When one motion is much faster than another, you can average over the fast one
and ask what slow landscape is left behind. Do that here and the rapid shaking contributes a real,
time-averaged restoring force — an effective potential. For a pivot driven as
s(t) = a·cos(ωt), the leading-order approximation to the slow motion has the potential
where θ is measured from straight down. The first term is ordinary gravity: a single valley at the
bottom (θ=0). The second term — the gift of the shaking — is largest where sin²θ is
largest, at the horizontal, and it pushes the bob away from sideways and toward the
vertical, either vertical. That is the green bump in the panel above. Crank the shaking up
and the bump grows until the top, once a hilltop, caves into a genuine valley. The bob can sit in
it. This is exactly the morphing you can watch in the U(θ) panel as you raise the drive.
The averaged threshold
The top (θ=180°) flips from hilltop to valley at a precise condition. Differentiate the effective potential, demand that θ=180° be a minimum, and out drops a clean inequality:
In words: the peak speed of the pivot, a·ω, must beat √(2gL).
Make the shake bigger or faster, or the pendulum shorter, and you cross the line. Written without
units, with ω₀ = √(g/L) the natural swing rate, the same averaged threshold is just
K = (a/L)(ω/ω₀) > √2 ≈ 1.414 — the number in the readout. The drive-frequency
slider's "threshold" mark is this formula solved for ω, recomputed whenever you change the amplitude
or length.
That derivation assumes a small shake (a≪L) and a fast drive (ω≫ω₀). The line is an approximate stability prediction for the full pendulum. The bob's trajectory is computed separately, by integrating the unaveraged equation of motion step by step:
A separate script sweeps the drive and records where a bob released near upright stops falling
during its observation window, then compares that finite-time result with √2.
The check — exact ODE vs. the formula
The verifier (research/kapitza-pendulum/verify.py, pure NumPy, fixed-step RK4) integrates the true equation of motion — no averaging anywhere in the simulation. First it confirms the integrator is faithful: with the shaking off, total energy holds to 3.1×10⁻⁸ (relative) over 40 small-angle periods. This checks the unforced integrator at that resolution. Then, for each amplitude, it bisects on drive frequency to find where the inverted bob (released with a 0.10 rad tilt) stays within 90° of upright during the observation window:
| amplitude a/L | finite-window ω/ω₀ | K = (a/L)(ω/ω₀) | K ⁄ √2 |
|---|---|---|---|
| 0.04 | 34.27 | 1.371 | 0.969 |
| 0.06 | 23.37 | 1.402 | 0.992 |
| 0.08 | 17.74 | 1.419 | 1.003 |
| 0.12 | 11.84 | 1.421 | 1.005 |
| 0.16 | 8.91 | 1.425 | 1.008 |
These recorded finite-window results range from 3.07% below to 0.77% above
√2 = 1.414. They do not demonstrate convergence as the drive speeds up: the fastest
row is 3.07% below √2. Below the averaged threshold the verifier watches only 60 drive cycles.
At a=0.01 m and K=1.38, that window ends at 1.7444 s; extending the same trajectory to
240 cycles finds a 90° excursion at about 1.9077 s. A short no-fall observation does not
establish lasting stability.
The two fastest rows are below √2. Higher-order corrections to averaging exist, but these rows do not isolate such a correction.
A second, independent check: above threshold the bob does a slow oscillation about the top, and the
effective potential predicts its small-oscillation period 2π / √(a²ω²/2L² − g/L). Measured from the full
simulation (stroboscopically, one sample per shake), the period matches the prediction to
0.4%–2.6% across three settings. The error grows from 0.37% at 120 rad/s to 2.57% at
180 rad/s; amplitude and K also change, so these rows do not establish frequency convergence. Re-run with
python3 verify.py; the table above converts its frequencies to ω/ω₀.
The same trick, four places you'd never expect
The reason this is worth more than a curiosity: the mechanism — a saddle or a hilltop that a fast oscillation turns into a trap — is one of the most reused ideas in physics. It even has a name when it acts on charges: the ponderomotive force.
Trapping a single atom
You cannot hold a charged particle still with static electric fields alone — Earnshaw's theorem forbids a stable stationary equilibrium in charge-free space; a static electrostatic trap has an escape direction. Wolfgang Paul's solution (Nobel Prize, 1989) was to make the saddle oscillate. Exactly as with the pendulum, the fast switching turns the unstable saddle into a stable effective well, and a lone ion sits in vacuum for hours. The mechanical cartoon of this is literally a ball on a saddle that spins — same averaging, same math. Trapped ions are now a leading platform for atomic clocks and quantum computers.
Plasma, accelerators, and lasers
The same time-averaged force pushes charges out of regions of strong oscillating field — it structures the wake behind an intense laser pulse. RF ponderomotive plugging is being explored to improve confinement in mirror-fusion designs. And strong focusing, the alternating-gradient principle that lets a particle accelerator's magnets focus a beam that each magnet alone would defocus, is the Kapitza idea applied in space rather than time: alternate the "wrong" force fast enough and the average becomes a restoring one. Every high-energy collider runs on it.
Corrections
Corrected 2026-10-14: The page called (aω)² > 2gL an exact stability threshold and labelled the simulated up state stable above it. It is a leading-order averaged prediction for small drive amplitude and high frequency; the slider and badge now identify that scope; found by the assay line, 2026-10-14. Butikov, equation (5).
Corrected 2026-10-14: The page promised recovery after Knock it. In a default run starting at simulation time zero (2 cm amplitude, 25 cm length, 20 Hz), Stand it up followed immediately by Knock it topples the bob; found by the assay line, 2026-10-14. Try the controls.
Corrected 2026-10-14: The page said shaking makes the bob rise. This simulation leaves a bob at rest straight down hanging; Stand it up places it near upright, where the drive can stabilize it; found by the assay line, 2026-10-14. Try the controls.
Corrected 2026-10-14: The page said the sweep was within about 1% of √2 and converged as the drive sped up. Its fastest row is 3.07% below √2, and the finite observation window can miss a later fall; these rows do not demonstrate convergence; found by the assay line, 2026-10-14. Sweep and window check.
Corrected 2026-10-14: The page called the fastest-drive results an excess above √2 and identified a higher-order correction. The two fastest rows are below √2, and these observations do not isolate a higher-order correction; found by the assay line, 2026-10-14. Recorded sweep.
Corrected 2026-10-14: The page said the measured slow periods converged toward the prediction as the drive sped up. Across these settings the error grows from 0.37% at 120 rad/s to 2.57% at 180 rad/s; amplitude and K also change; found by the assay line, 2026-10-14. Period comparison.
Corrected 2026-10-14: The dek said strong focusing serves every particle accelerator. Brookhaven's 2020 account describes its use in most present-day accelerators and all high-energy colliders; found by the assay line, 2026-10-14. Brookhaven account.
Corrected 2026-10-14: The page said Earnshaw's theorem forbids any static field from holding an ion. It forbids a stable stationary equilibrium using electrostatic fields alone in charge-free space; a Penning trap confines moving ions with static electric and magnetic fields; found by the assay line, 2026-10-14. Paul, pp. 613-614.