Held up by a force that averages to nothing

A pendulum has two equilibria and everybody knows which one is which. Hanging down is stable and standing up is not, because the potential has a minimum at one and a maximum at the other, and that is the end of it.

An ordinary pendulum makes the difficulty plain. Resolve the weight along the rod and across it: the across-component is the restoring force, and hanging down it points back towards the equilibrium. Turn the pendulum through 180° and that same component points away from the upright. That is the whole of what makes standing up unstable, and it is not a subtle effect — the force actively drives the rod away from the position that is about to be made stable.

It is not the end of it. Shake the pivot up and down fast enough and the pendulum stands upside down and stays there. Kapitza described it in 1951, and the demonstration is a jigsaw with a rod bolted to the blade.

What the shaking does to the equations #

In the frame of the moving pivot the only change is that gravity acquires a time-dependent part. If the pivot moves as $a\cos\Omega t$, the effective downward acceleration is $g + a\Omega^2\cos\Omega t$, and the equation of motion is

$$\ddot{\theta} + \frac{g + a\Omega^2\cos\Omega t}{L}\sin\theta = 0$$

which is a pendulum with a wobbling gravity. Nothing has been approximated to get there, and nothing about it suggests that the upright should become stable: the extra term averages to exactly zero over a cycle.

Put it as a potential and the difficulty looks insuperable. A maximum is unstable precisely because the force points away from it, and adding something that averages to nothing cannot change the average force. The argument seems complete: the mean of the added force is zero by construction, so the mean force is what it was, so the maximum stays a maximum. That is the intuition this essay exists to break, and it is worth holding it clearly before it breaks, because the flaw in it is not where a reader expects.

Why zero average is not zero effect #

The step everybody skips is that the average of a product is not the product of the averages, and the two things being multiplied here are correlated.

Split the motion into a slow part and a fast one, $\theta = \Theta + \xi$, where $\xi$ is a small ripple at the shaking frequency. The fast forcing $-(a\Omega^2/L)\cos(\Omega t)\sin\theta$ drives the ripple, so $\xi$ is proportional to $\cos\Omega t$ and to $\sin\Theta$. Now put that ripple back into the same force: the term $\cos(\Omega t)\cdot\xi\cos\Theta$ appears, containing $\cos^2\Omega t$, whose average is one half rather than zero.

What survives the averaging is an extra term in the potential,

$$\frac{U_{\text{eff}}}{MgL} = -\cos\theta + \frac{1}{4}\left(\frac{a}{L}\right)^2\left(\frac{\Omega}{\omega_0}\right)^2\sin^2\theta$$

which is largest sideways and zero both hanging and inverted — and which digs a well at 180°.

The criterion falls out by comparing the two terms’ curvature at $\theta = \pi$: the inverted position is stable when

$$(a\Omega)^2 > 2gL$$

which says that the speed of the shaking must exceed the speed something acquires falling through the pendulum’s own length. Neither the amplitude nor the frequency matters on its own, only the product — shaking twice as fast needs half the throw.

Does the exact equation agree? #

The derivation above averaged, and averaging is exactly the sort of step that produces a plausible criterion with the wrong constant in it. The equation can be integrated instead, with nothing separated and nothing averaged.

Three per cent, over rates from eight to thirty times the pendulum’s natural frequency, and the disagreement is largest at the slow end — which is exactly where it should be, because that is where the separation between “fast” and “slow” is weakest. The averaged theory is not a heuristic that happens to work; it is the leading term of a controlled expansion, and its error behaves accordingly.

The upper boundary is the more interesting output. Shaken hard enough, the pendulum falls over again — because the ripple $\xi$ is no longer small, the expansion the averaging rests on stops converging, and the fast motion itself becomes the instability. Stability is a band, not a threshold, and only its lower edge is what the effective potential describes.

The phase portrait of the unshaken pendulum says the same thing in a language that survives the repair. The crossing point at the top is the inverted equilibrium, and it is a saddle: one direction leads in, one leads out, and unstable means exactly that an arbitrarily small displacement finds the outgoing one. What the shaking does, within the band, is convert that saddle into a centre — and the trajectories near it stop leaving and start closing.

The neighbour on the same chart #

Modulating a pendulum’s parameters is also how a swing is pumped, and that is the opposite effect from the same equation: there, the modulation feeds energy in and the hanging equilibrium goes unstable.

The two phenomena live in different regions of that chart. Parametric resonance is the tongue near $\Omega = 2\omega_0$: a modulation at a resonance, taking energy in a little each cycle. Kapitza’s regime is far to the right of every tongue, at $\Omega \gg \omega_0$, where the modulation resonates with nothing and its only effect is the average one.

Distinguishing them matters because the words are nearly the same and the physics is opposite. One is a resonance and the other is explicitly not one.

The same trick holding a charge #

The mechanism generalises past pendulums, and its most consequential use is in a place with no gravity in it at all.

No arrangement of static charges can trap another charge: a potential obeying Laplace’s equation has no interior minimum, so every static electric trap has at least one direction out. The nearest thing available is a saddle — trapping along one axis and expelling along another.

Reverse the saddle fast enough and the particle is confined in both directions. The reason is the reason above: while the field pushes the particle outward it also displaces it into a stronger part of the field, so the inward half of the cycle is applied at a larger displacement than the outward half, and the average of the product does not vanish. The effective potential is a well, the residual fast wobble is called micromotion, and the whole apparatus is Paul’s ion trap, which shared the 1989 Nobel Prize.

The measured stability edge here is $q = 0.907$, against the 0.908 at which Mathieu’s first unstable region begins. That is the same equation as the pendulum’s, with the constant and alternating terms exchanged.

Every ion clock, every trapped-ion quantum computer and every mass spectrometer of the quadrupole kind is a saddle being switched.

The rotating saddle, which is the same statement #

There is a mechanical version of the ion trap that makes the averaging visible without any electricity, and it is worth describing because it is the clearest demonstration of the whole idea.

Machine a saddle-shaped surface — rising along one horizontal axis, falling along the other — and put a ball at the centre. It rolls off, along the falling direction. Now rotate the saddle about its vertical axis. If the rotation is fast enough the ball stays, oscillating gently about the centre, and it will return there if it is nudged.

A rotating saddle is confining for the same reason a switched one is, and it is the easier picture to hold. Put a ball on a saddle-shaped surface and spin the surface: by the time the ball has begun to roll down a falling direction, that direction has rotated and become a rising one. The average over a turn of a force that depends on where the ball has got to does not vanish, even though the average of the force at a fixed point does. That distinction is the entire mechanism, and it is why the naive argument above fails.

The rotating saddle and the switched saddle are not quite the same equation — a rotation adds a Coriolis term that a switch does not have, so the rotating version is stable over a narrower range — but the mechanism is identical, and a rotating saddle trap was built to demonstrate exactly that. It also makes the failure mode obvious: rotate too slowly and the ball leaves during a single half-turn, which is the lower edge, and rotate too fast and it is flung out, which is the upper one.

Who found it, and when #

The result has been discovered more than once, which is a common fate for phenomena that look like tricks.

Andrew Stephenson published the inverted pendulum’s stability in 1908, derived from the Mathieu equation, and it attracted almost no attention. Kapitza rederived it in 1951 with the averaging argument used here — he was under house arrest at the time, working in a converted shed — and it is his treatment that made it a piece of physics rather than a curiosity, because the effective potential is a general method and the Mathieu chart is a specific answer.

Paul’s trap came in the mid-1950s from the mass-spectrometry side, and the connection to the pendulum was not the route by which it was found. The two literatures took some time to notice each other, which is one reason the same idea carries at least four names: effective potential, ponderomotive potential, averaged dynamics, and Kapitza stabilisation.

What sets the slow frequency #

Inside the band the inverted pendulum is a genuine oscillator, with a period that can be computed and measured.

What sets the slow frequency is a curvature, as it always is. The frequency of small oscillations about a minimum is fixed by the curvature of the potential there and by nothing else about its shape. For the shaken pendulum the relevant curvature at 180° is a difference of two terms — the shaking’s contribution minus gravity’s — so the effective frequency falls to zero as the shaking is weakened to the edge of the stable band, which is what the edge of a stability region always looks like from inside it.

Differentiating the effective potential twice at $\theta = \pi$ gives

$$\omega_{\text{eff}}^2 = \frac{a^2\Omega^2}{2L^2} - \frac{g}{L}$$

so the inverted pendulum’s swing is slow just above threshold and speeds up as the shaking is increased. That is a testable prediction with no free constants, and measuring it is the standard way of demonstrating that the effective potential is real rather than a way of speaking.

The zero at threshold is worth pausing on. A frequency that goes to zero means a restoring force that goes to zero, so at the edge of the band the pendulum is neutrally stable — it will sit anywhere near the top and take arbitrarily long to come back. That is the generic signature of a stability boundary, and it is the same signature a system approaching a critical point shows: the response to a disturbance slows down before it changes sign.

The energy trade is the ordinary one, made in the effective potential rather than the real one. An ordinary pendulum swaps kinetic for potential energy twice a cycle; the inverted pendulum does the same swap, but the energy being traded is an average over the fast cycle rather than an instantaneous quantity. That is why the exchange is visible only in the slow motion, and why a stroboscopic view of a stabilised inverted pendulum shows something that looks exactly like an ordinary oscillator.

Where else an average creates a force #

In laser cooling and optical tweezers. A dipole in a rapidly oscillating light field is pushed toward high intensity or away from it depending on the detuning, by exactly this averaging. The trap that holds a bacterium in a microscope is the same calculation with a different oscillation.

In plasma physics, where it has a name. The ponderomotive force pushes charged particles out of regions of strong oscillating field, and it is what expels electrons from the focus of an intense laser. Its expression is the gradient of the mean-square field — which is the $\sin^2\theta$ term above, written for a field instead of a pendulum.

In everything with two well-separated timescales. A fast, small motion averaged over produces a slow effective dynamics with terms that the fast motion is invisible in. What a system actually minimises is a question that has to be asked again once the fast degrees of freedom are averaged away, because what is minimised is not what would have been guessed from the instantaneous forces.

The sign of the effect comes from a phase lag, which is where an average creates a force elsewhere too. Far above resonance a damped oscillator responds in antiphase with its drive, and the fast ripple of the shaken pendulum lives in exactly that regime. Antiphase is what makes the correlation between the rod’s position and the force on it come out with the sign that stabilises rather than the one that does not — reverse it and the same shaking would throw the rod down faster than gravity alone.

The same averaging that made accelerators possible #

The ion trap is the best-known application. The most consequential one is in a machine nobody associates with pendulums, and it arrived from an entirely separate direction.

A charged particle beam has to be kept near the axis of a machine that may be kilometres round, and the tool for bending a beam toward an axis is a quadrupole magnet. A quadrupole has the same defect the electrostatic trap does, and for the same reason: the field must satisfy Laplace’s equation, so a magnet that focuses in the horizontal plane necessarily defocuses in the vertical one. There is no such thing as a magnetic lens that converges in both.

Early accelerators lived with it. They used weak focusing — a small gradient providing a feeble restoring force in both planes at once — and the price was that the beam wandered over a large area, so the vacuum chamber had to be wide and the magnets surrounding it enormous. The Cosmotron’s magnet weighed two thousand tonnes for a beam a few centimetres across.

The escape is the trick of this essay. Alternate the quadrupoles: focusing, then defocusing, then focusing, along the beamline. Naively the two cancel. They do not, and the reason is a one-line calculation about thin lenses: two lenses of focal length $+f$ and $-f$ separated by a distance $d$ combine to give

$$\frac{1}{F} = \frac{1}{f} - \frac{1}{f} + \frac{d}{f^2} = \frac{d}{f^2},$$

which is positive whatever the order of the two. The pair converges.

The physical reading is exactly the pendulum’s. A particle passing through the focusing element is bent toward the axis, so it arrives at the defocusing element closer in, where the field is weaker — and a quadrupole’s field is proportional to the distance from the axis. The focusing kick is delivered at large displacement and the defocusing kick at small displacement, so the average of the product does not vanish, and it does not vanish in the helpful direction.

Courant, Livingston and Snyder published this in 1952; Christofilos had reached it independently two years earlier and had not managed to get it noticed. The consequence was immediate and enormous. Restoring forces rose by orders of magnitude, beam sizes fell from tens of centimetres to millimetres, and magnets fell from thousands of tonnes to tens. Every synchrotron built since is strong-focusing, and none of the machines that found the particles of the last seventy years would have been affordable without it.

That two communities — one shaking a pendulum, one steering protons — arrived at the same mechanism within a year of each other, neither aware of the other, is worth noticing. What they had in common was the constraint: Laplace’s equation forbids a static minimum, and alternating in time or in space is the only way round it.

A chain of them, and the rope trick #

The derivation above treats the pendulum as one rigid rod, and it is fair to ask how much of the result depends on that. The answer is that it depends on it hardly at all, and the demonstrations are among the most startling in mechanics.

Link two rods end to end, pivot the bottom one on a shaker, and the doubly-inverted configuration — both rods pointing up, the top one balanced on the end of the bottom one — can be stabilised by the same vertical vibration. So can three. Acheson and Mullin published photographs of a three-link chain standing upside down on a vibrating support in 1993, and the accompanying theory shows that a chain of any number of links has a stabilising regime, with the required shaking speed rising as links are added.

The multi-link case is not a curiosity added on. It is a check on the mechanism, because a chain of $N$ links has $N$ unstable modes rather than one, and the effective-potential argument has to turn all of them stable simultaneously. That it does — and that the required amplitude and frequency come out where the theory says — is much stronger evidence than any single-rod demonstration could be.

Taking the number of links to infinity gives a flexible chain, and that has been done too: a length of ordinary curtain wire, held at the bottom by a vibrating clamp, stands upright and sways. It has been called the Indian rope trick, and unlike the original it works.

What the limit shows is which quantity actually matters. A stiff rod resists bending on its own; a floppy wire does not, and stands anyway, so the stabilisation is not doing anything with the material’s stiffness. The wire is being held up by an effective potential built entirely out of the correlation between its own displacement and the shaking — which is to say by exactly the term derived at the top of this essay, applied to every point along its length at once.

Why the intuition fails in the first place #

It is worth naming the general error, because it recurs and it is not specific to pendulums.

The wrong argument is: the extra force averages to zero, so its effect averages to zero. That step is a swap of two operations — averaging and evaluating the force at the position — and the swap is legitimate only when the position does not vary over the averaging window. Here it does vary, by exactly the amount the same force produced, so the correlation between the two is not a small correction but the entire effect.

The general form of the statement is that $\langle F(x(t))\rangle \neq F(\langle x\rangle)$ whenever $F$ is nonlinear and $x$ wobbles. Expanding $F$ about the mean position, the first surviving term is $\tfrac{1}{2}F''\langle \xi^2\rangle$ — a force proportional to the curvature of the field and to the mean square of the wobble, and pointing toward weaker field or stronger field according to the sign. That is the ponderomotive force, and it is why fast oscillation is a way of building a potential rather than a way of adding noise.

The same swap fails in the same way in several other places on this site: the average of a fluctuating rate is not the rate at the average when the rate is exponential in the fluctuating quantity, which is why a reaction’s rate is dominated by the rare hot moments rather than by the typical ones.

What the pictures cannot show #

The averaging is an expansion, and only its first term is drawn. The effective potential is the leading term in $\omega_0/\Omega$, and the corrections are what the three per cent disagreement measures. At a shaking rate of three or four times the natural frequency, rather than thirty, the picture is qualitatively right and quantitatively poor.

There is no damping in the integration. Real damping widens the band at the bottom, because it removes the energy a slightly-off start would otherwise keep, and a real demonstration is more forgiving than this calculation. It also eventually stops the shaking from being free.

The pivot’s motion is sinusoidal. A jigsaw is not, and a square-wave drive gives a different constant in the criterion — the mean square of the velocity is what appears, so the shape of the drive matters but only through one number.

And the pendulum is rigid and planar. A real rod shaken hard flexes, and a real bob swings out of the plane; both add degrees of freedom that the two-dimensional equation has no room for, and the second is what usually ends a demonstration.

The ladder from here #

Later rungs on this anchor: the Mathieu equation proper, with its stability chart read as one object rather than as two regimes; the second-order averaging that produces the upper stability boundary; the ponderomotive force in a field rather than for a pendulum; and the Paul trap’s design, where the same criterion is used to choose an operating point rather than to explain a demonstration.

The neighbouring ladders are the pumped swing, which is the resonant corner of the same chart, the pendulum’s small lie, which is the approximation this essay makes at a different point, and the three ways of coming to rest, which is what happens to the fast motion when damping is added.