Adiabatic Invariants

At the first Solvay Congress in 1911, Hendrik Lorentz stood up with a deceptively innocent puzzle: take a pendulum that carries exactly one quantum of energy, $E = h\nu$ , and slowly shorten its string. The frequency $\nu$ rises, the energy changes — so does the pendulum still hold a whole quantum, or has slow, gentle pulling destroyed the quantization? Einstein's answer from the floor was immediate: for slow shortening, $E/\nu$ stays constant. Energy and frequency drift together, in lockstep, and the quantum condition survives. That ratio — the action $J = \frac{1}{2\pi}\oint p\,\mathrm{d}q$ — is the star of this article. For fixed parameters the actions are constants of motion, as built in Action-Angle Variables. The theory of adiabatic invariants asks the sharper question: what survives when a parameter $\lambda$ — a string length, a spring constant, a magnetic field — is not fixed but varies in time? If the variation is slow on the scale of the orbital period, the action remains frozen even though the energy drifts substantially. Nothing in Hamilton's equations conserves it exactly; it is conserved because the fast oscillation averages the disturbance away. The payoff spans a century of physics. Adiabatic invariance is why Bohr–Sommerfeld quantization was self-consistent, why the Van Allen belts trap particles for years, why accelerator beams keep their emittance while being ramped, and why a plasma in a magnetic mirror bounces back instead of leaking out the ends. Master the averaging argument once and you own all of it.

Key takeaways

An adiabatic invariant is a quantity — almost always an action variable $J=\frac{1}{2\pi}\oint p\,\mathrm{d}q$ — that stays nearly constant when a system's parameters change slowly compared to the oscillation period.
"Slow" is a ratio, not a rate: the fractional parameter change per orbital period must be tiny, $\varepsilon = |\dot\lambda|/(\omega\lambda) \ll 1$ .
The conservation is far better than the naive $O(\varepsilon)$ : for smooth (analytic) switching the total drift is exponentially small, $\Delta J \sim e^{-c/\varepsilon}$ .
Canonical example: a pendulum with a slowly shortened string keeps $J = E/\omega$ constant, so the energy grows as $E \propto \ell^{-1/2}$ and the angular amplitude as $\theta_{\max}\propto \ell^{-3/4}$ .
The magnetic moment $\mu = m v_\perp^2/2B$ is the gyration action in disguise — its invariance is the working principle of magnetic mirrors and the Van Allen radiation belts.
Invariance fails where the period diverges (separatrix crossing) or at resonances between internal frequencies — there the action can jump by $O(1)$ .
Key idea: under slow variation of parameters the angle sweeps many cycles before the parameter moves, so $\dot J = -\partial H/\partial\theta$ averages to zero over each cycle — the enclosed phase-space area is preserved, with only exponentially small residual drift.

The core insight

Under slow variation of parameters, the angle variable advances much faster than the parameter changes, so the equation $\dot J=-\partial H/\partial\theta$ averages to zero over a cycle: the action $J=\frac{1}{2\pi}\oint p\,\mathrm{d}q$ — the phase-space area divided by $2\pi$ — is conserved to all orders in the slowness, with only exponentially small residual drift.

Where this lives today

Adiabatic invariance is quietly load-bearing across modern physics. In space physics, the three invariants of trapped-particle motion — the gyration moment $\mu$ , the bounce action, and the drift-shell flux — organize the entire theory of the Van Allen belts; violations of the third invariant by magnetic storms are how the belts fill and drain, a live issue for satellite electronics. In accelerator physics, the betatron action of each particle is adiabatically invariant while fields ramp over millions of turns, which is why beam emittance is preserved — and why "adiabatic damping" of the normalized emittance during acceleration can be engineered rather than hoped for. Fermi's 1949 model of cosmic-ray acceleration is adiabatic invariance run in reverse: magnetic clouds acting as slow-moving mirrors that stochastically pump particle energy. The deepest legacy is the classical–quantum bridge this chapter is building. Ehrenfest's principle — quantize the adiabatic invariants — explains why $\oint p\,\mathrm{d}q = nh$ was the right guess: quantum numbers must label quantities that slow perturbations cannot change, and those are exactly the actions (the full story is in Bohr-Sommerfeld Quantization). The quantum adiabatic theorem is the direct descendant, underwriting adiabatic quantum computation and STIRAP control of atoms. And the residual, geometric part of slow cyclic evolution — the piece that survives after the dynamical phase is subtracted — is Hannay's angle classically and Berry's phase quantum mechanically, the subject of Hannay's Angle.

The physical picture

Picture the orbit as a closed loop in the $(q,p)$ plane. The action is $1/2\pi$ times the area of that loop. Change a parameter slowly and the loop morphs — taller here, narrower there — but like an incompressible fluid blob being gently kneaded, the area it encloses stays fixed. Change the parameter fast and the picture shatters: the orbit is caught mid-cycle at some particular phase, the loop tears away to a new energy contour, and the enclosed area jumps. Slowness is what makes the deformation phase-blind.

Figure

A slow parameter change morphs the phase-space orbit but preserves the area it encloses — the action is the frozen quantity

Think of a child on a swing. Pumping at the right phase changes the amplitude a lot — that is resonant driving. But if you raise and lower the seat very slowly compared to the swing period, the swing "doesn't notice" any particular phase: the perturbation delivered on the upswing is undone on the downswing, cycle after cycle, and only the phase-averaged effect survives. For the action that average is exactly zero. Slowness randomizes the phase of the disturbance relative to the motion. The flagship application makes the picture concrete. A charged particle gyrating in a magnetic field is a little oscillator whose "parameter" is the local field strength $B$ . Let the particle drift along a field line into a region of stronger $B$ — a magnetic mirror. Its gyration action, the magnetic moment $\mu = mv_\perp^2/2B$ , is frozen, so $v_\perp$ must grow with $B$ ; since the magnetic force does no work, $v_\parallel$ must shrink to compensate — until it hits zero and the particle is reflected back. No wall touched it. That bounce, repeated between two mirror points, is what holds Van Allen-belt protons for years and fusion plasma in mirror machines for as long as collisions allow.

Figure

Magnetic mirror: the frozen gyration action converts parallel motion into perpendicular gyration until the particle reflects

Where does the picture break? Wherever the motion stops being fast. Near a separatrix — the pendulum poised at the top of its swing — the period diverges, so no parameter change is slow enough, and the action jumps as the orbit crosses over. That failure is not a technicality: separatrix crossing and capture into resonance is real dynamics in planetary spin-orbit history and particle traps, and it is where the adiabatic chapter hands off to chaos theory.

From action variable to adiabatic invariant

Everything rests on one definition. For a one-dimensional bounded motion at fixed parameters, the loop integral over one period,

J=\frac{1}{2\pi}\oint p\,\mathrm{d}q \qquad (1)

where: $J$ — the action variable, $1/2\pi$ times the phase-space area enclosed by the orbit; $p, q$ — canonical momentum and coordinate; $\oint$ — one full circuit of the libration or rotation orbit. At fixed $\lambda$ this is the same $J$ built in Action-Angle Variables — constant because $H = H(J)$ there. The new question is its fate when $\lambda = \lambda(t)$ . The claim needs a precise notion of "slow". Compare the fractional change of the parameter during one orbit to unity:

\varepsilon \equiv \frac{1}{\omega}\left|\frac{\dot\lambda}{\lambda}\right| \ll 1 \qquad (2)

where: $\varepsilon$ — the slowness (adiabaticity) parameter; $\omega$ — the instantaneous orbital frequency; $\lambda$ — the slowly varying external parameter. Note what is not required: $\dot\lambda$ need not be small in absolute terms, and the total change of $\lambda$ can be enormous — the string can be shortened to half its length, the field doubled — as long as each single period sees only a sliver of it. Adiabatic invariance is a statement about a ratio of timescales, not about the size of the excursion. With $\lambda$ time-dependent, transform to the action-angle pair $(J,\theta)$ of the instantaneous Hamiltonian. The transformation is now time-dependent, so $H$ picks up angle dependence of order $\dot\lambda$ , and Hamilton's equation for the action reads

\frac{\mathrm{d}J}{\mathrm{d}t}=-\frac{\partial H}{\partial\theta}\bigg|_{J,\lambda}, \qquad \left\langle\frac{\mathrm{d}J}{\mathrm{d}t}\right\rangle_\theta = O(\varepsilon^2) \qquad (3)

where: $\theta$ — the fast angle, advancing at rate $\omega$ ; $\langle\cdot\rangle_\theta$ — average over one cycle of $\theta$ . The right-hand side is a pure $\theta$ -derivative, so its Fourier expansion $\sum_{k\neq0}(-ik)H_k e^{ik\theta}$ has no constant term: the angle-average vanishes identically at leading order, and the residual drift enters only because $\lambda$ itself creeps during the averaging window. This single structural fact — a derivative has no mean — is the entire mechanism. $J$ wobbles by $O(\varepsilon)$ within a cycle but accumulates no secular change. The two derivations below make this bulletproof: first the harmonic oscillator by direct energy bookkeeping, then the general one-dimensional case.

Derivation AHarmonic oscillator: E over omega is frozen

We prove adiabatic invariance in the cleanest laboratory: $H = \tfrac{p^2}{2m} + \tfrac12 m\omega^2(t)\,q^2$ with $\omega(t)$ varying slowly. The strategy is energy bookkeeping plus the virial theorem: compute how fast the energy drifts because $\omega$ changes, average that drift over one fast cycle, and watch the result organize itself into $\dot E/E = \dot\omega/\omega$ — which says exactly that $E/\omega$ is constant. The one step to watch is the cycle average: it treats $\omega$ and $\dot\omega$ as constant during a single period, which is precisely the adiabatic assumption $\dot\omega/\omega^2 \ll 1$ .

Step 1. The energy changes only through the explicit time dependence of $H$ . Along a trajectory, $\mathrm{d}E/\mathrm{d}t = \partial H/\partial t$ (the $\dot q, \dot p$ terms cancel by Hamilton's equations), so

\frac{\mathrm{d}E}{\mathrm{d}t}=\frac{\partial H}{\partial t}=m\,\omega\dot\omega\, q^2

Why only the explicit term survives: $\dot E = \frac{\partial H}{\partial q}\dot q + \frac{\partial H}{\partial p}\dot p + \frac{\partial H}{\partial t}$ , and the first two terms are $\frac{\partial H}{\partial q}\frac{\partial H}{\partial p} - \frac{\partial H}{\partial p}\frac{\partial H}{\partial q} = 0$ .

Step 2. Average over one period, holding $\omega,\dot\omega$ fixed during it (legitimate to leading order in $\varepsilon$ ):

\left\langle\frac{\mathrm{d}E}{\mathrm{d}t}\right\rangle=m\,\omega\dot\omega\,\langle q^2\rangle

Step 3. For a harmonic oscillator the time-averaged potential energy is half the total energy: $\tfrac12 m\omega^2\langle q^2\rangle=\tfrac12 E$ , hence $\langle q^2\rangle = E/(m\omega^2)$ . Why: with $q = a\cos(\omega t + \phi)$ , $\langle\cos^2\rangle = 1/2$ gives $\langle q^2\rangle = a^2/2$ and $E = \tfrac12 m\omega^2 a^2$ — the equipartition of the SHO between kinetic and potential energy on average.

Step 4. Combine Steps 2 and 3:

\left\langle\frac{\mathrm{d}E}{\mathrm{d}t}\right\rangle=m\omega\dot\omega\cdot\frac{E}{m\omega^2}=\frac{\dot\omega}{\omega}\,E

so the fractional energy drift matches the fractional frequency drift, $\dot E/E = \dot\omega/\omega$ .

Step 5. Therefore

\frac{\mathrm{d}}{\mathrm{d}t}\ln\frac{E}{\omega}=\frac{\dot E}{E}-\frac{\dot\omega}{\omega}=0 \quad\Longrightarrow\quad J=\frac{E}{\omega}=\text{const}

to leading order; the neglected terms are $O(\dot\omega^2)$ and average away over the full switching for smooth $\omega(t)$ . Verified: matches Landau–Lifshitz §49 and Goldstein §12.5; the SHO action $J = E/\omega$ is the standard result, and equation (5) below follows by substituting $E = \tfrac12 m\omega^2 a^2$ .

The oscillator result generalizes with no more machinery than differentiating the loop integral carefully.

Go deeper · Derivation B · 1 minGeneral one-dimensional systems: the averaged action does not drift

Derivation BGeneral one-dimensional systems: the averaged action does not drift

We now show $\langle\dot J\rangle = 0$ for any one-dimensional bounded motion $H(q,p;\lambda)$ with slow $\lambda(t)$ , not just the SHO. The strategy: regard $J$ as a function of the instantaneous energy and parameter, $J(E,\lambda)$ , compute its total time derivative through both arguments, and convert both resulting loop integrals to time averages — where they turn out to be the same integral with opposite signs. The delicate point is keeping track of what is held fixed in each partial derivative.

Step 1. The action defined on the instantaneous orbit is $J(E,\lambda)=\frac{1}{2\pi}\oint p(q,E,\lambda)\,\mathrm{d}q$ , so its drift has two channels:

\frac{\mathrm{d}J}{\mathrm{d}t}=\frac{\partial J}{\partial E}\,\dot E+\frac{\partial J}{\partial\lambda}\,\dot\lambda

Step 2. The energy drifts only through $\lambda$ : $\dot E = \partial H/\partial t = (\partial H/\partial\lambda)\,\dot\lambda$ , evaluated along the orbit.

Step 3. Differentiate the loop integral under the integral sign, at fixed $\lambda$ and at fixed $E$ respectively:

\frac{\partial J}{\partial E}=\frac{1}{2\pi}\oint\frac{\partial p}{\partial E}\,\mathrm{d}q=\frac{T}{2\pi}, \qquad \frac{\partial J}{\partial\lambda}=\frac{1}{2\pi}\oint\frac{\partial p}{\partial\lambda}\,\mathrm{d}q

Why the first integral is the period: differentiating $H(q,p,\lambda)=E$ at fixed $q,\lambda$ gives $\partial p/\partial E = 1/(\partial H/\partial p) = 1/\dot q$ , and $\oint \mathrm{d}q/\dot q = T$ . This is the same identity $T = \partial(2\pi J)/\partial E$ used in Frequencies from Action-Angle Variables.

Step 4. Similarly, differentiating $H(q,p,\lambda)=E$ at fixed $q, E$ gives $\partial p/\partial\lambda = -(\partial H/\partial\lambda)/\dot q$ .

Step 5. Assemble, converting $\mathrm{d}q = \dot q\,\mathrm{d}t$ so each loop integral becomes an integral over one period:

\frac{\mathrm{d}J}{\mathrm{d}t}=\frac{\dot\lambda}{2\pi}\left[\int_0^T\frac{\partial H}{\partial\lambda}\bigg|_{\text{orbit}}\mathrm{d}t-\int_0^T\frac{\partial H}{\partial\lambda}\bigg|_{\text{orbit}}\mathrm{d}t\right]=0

Why the two terms coincide: the first comes from $\frac{\partial J}{\partial E}\dot E = \frac{T}{2\pi}\langle\partial H/\partial\lambda\rangle\dot\lambda$ written as a time integral, the second from $\frac{\partial J}{\partial\lambda}\dot\lambda$ after Step 4 — identical integrands, opposite signs. The cancellation is exact at fixed $(E,\lambda)$ ; the residual $O(\dot\lambda^2)$ comes from their drift during the cycle. Verified: reproduces Landau–Lifshitz §49 and Arnold §52; the cancellation structure is the classical averaging theorem specialized to one frequency, and Step 3 recovers the standard identity $\omega = \partial H/\partial J$ .

For smooth switching the story is even better than "no drift on average". If $\lambda(t)$ is analytic and its derivatives vanish at the endpoints, every polynomial order in $\varepsilon$ cancels, and the total change over the entire switching history is exponentially small,

\Delta J \sim \exp(-c/\varepsilon) \qquad (4)

where: $c$ — a constant set by the analyticity strip of the frequency $\omega(\lambda(t))$ in complex time. This is the same "beyond all orders" structure as quantum tunneling amplitudes — a first hint that adiabatic invariance and semiclassical quantum mechanics share machinery. For merely $C^k$ switching, $\Delta J = O(\varepsilon^{k})$ : smoothness literally buys orders of conservation.

The two headline invariants

Specializing (1) to the harmonic oscillator, where $H = \omega J$ exactly, gives the form Einstein quoted at Solvay:

E = \omega(\lambda)\,J \quad\Longrightarrow\quad \frac{E}{\omega}=\text{const} \qquad (5)

where: $E$ — oscillator energy, which does change; $\omega(\lambda)$ — the instantaneous frequency. Energy is pumped in or extracted by the agent moving the parameter, but always in the fixed ratio $\mathrm{d}E = J\,\mathrm{d}\omega$ . Substituting $E = \tfrac12 m\omega^2 a^2$ gives the amplitude and momentum scalings $a_{\max}\propto\omega^{-1/2}$ , $p_{\max}\propto\omega^{1/2}$ : the phase-space ellipse narrows in $q$ and stretches in $p$ at fixed area. The second headline invariant governs magnetized plasma. A particle gyrating about a field line is an oscillator whose action is, up to constants, the magnetic moment:

\mu = \frac{m v_\perp^2}{2B} = \text{const} \qquad (6)

where: $\mu$ — the magnetic moment (first adiabatic invariant of plasma physics); $v_\perp$ — the speed perpendicular to the field line; $B$ — the field magnitude sampled by the guiding center. To see that $\mu$ is a gyration action, compute the loop integral for one Larmor circle of radius $r_L = mv_\perp/(qB)$ : the gyration action comes out proportional to $\tfrac{m v_\perp^2}{2B} = \mu$ , so slow variation of $B$ — in space along the guiding-center path, or in time — freezes $\mu$ . "Slow" here means the field changes little per gyration: satisfied by kilometers-scale geomagnetic gradients against millimeter-to-meter Larmor radii, which is why the invariance is so spectacularly good in the magnetosphere.

Limiting cases

The validity of everything above hinges on the single ratio $\varepsilon$ of equation (2) — and on the period staying finite.

Case	Result	Meaning
Ideal adiabatic limit	$\varepsilon\to 0$ , $\Delta J\sim e^{-c/\varepsilon}$	Smooth, infinitely slow variation: the action is conserved to all orders; only exponentially small drift remains.
Sudden (anti-adiabatic) limit	$\varepsilon\gg 1$	The parameter jumps in a time short compared to the period: $(q,p)$ are momentarily frozen, and the new action is whatever loop the frozen point lands on — generally different.
Separatrix crossing	$\omega\to 0$ , $T\to\infty$	Near a separatrix the period diverges, so adiabaticity fails at any rate; the action changes by $O(1)$ , with a quasi-random outcome set by the crossing phase.
Resonance (several frequencies)	$\mathbf{k}\cdot\boldsymbol\omega=0$	Low-order commensurabilities among the unperturbed frequencies spoil the averaging; actions can drift secularly — the theme of Degeneracy and Commensurate Frequencies.

Worked examples

Example 1 — Einstein's pendulum, with numbers. A pendulum of mass $m = 0.250\ \mathrm{kg}$ swings with angular amplitude $\theta_{\max} = 0.100\ \mathrm{rad}$ on a string of length $\ell_1 = 1.00\ \mathrm{m}$ . The string is drawn up through a hole in the pivot, slowly (over minutes), to $\ell_2 = 0.500\ \mathrm{m}$ . Find the new amplitude and energy ( $g = 9.81\ \mathrm{m/s^2}$ ).

Step 1. Initial frequency and energy. $\omega_1 = \sqrt{g/\ell_1} = \sqrt{9.81/1.00} = 3.13\ \mathrm{rad/s}$ . For small oscillations $E = \tfrac12 m g \ell\,\theta_{\max}^2$ , so $E_1 = \tfrac12(0.250)(9.81)(1.00)(0.100)^2 = 1.23\times10^{-2}\ \mathrm{J}$ .

Step 2. The frozen action. $J = E_1/\omega_1 = 1.23\times10^{-2}/3.13 = 3.92\times10^{-3}\ \mathrm{J\,s}$ .

Step 3. New frequency: $\omega_2 = \sqrt{9.81/0.500} = 4.43\ \mathrm{rad/s}$ . Adiabatic invariance fixes the new energy: $E_2 = J\,\omega_2 = (3.92\times10^{-3})(4.43) = 1.73\times10^{-2}\ \mathrm{J}$ — the energy grew by the factor $\omega_2/\omega_1 = \sqrt{2} = 1.41$ , pumped in by the hand pulling the string against the average tension.

Step 4. New amplitude from $E_2 = \tfrac12 m g \ell_2\,\theta_{\max,2}^2$ : $\theta_{\max,2}^2 = 2E_2/(mg\ell_2) = 2(1.73\times10^{-2})/[(0.250)(9.81)(0.500)] = 2.83\times10^{-2}$ , so $\boldsymbol{\theta_{\max,2} = 0.168\ \mathrm{rad}}$ . Cross-check against the scaling law $\theta_{\max}\propto\ell^{-3/4}$ : $0.100\times(0.500)^{-3/4} = 0.100\times1.682 = 0.168\ \mathrm{rad}$ ✓. Halving the string raises the amplitude by 68% while the linear swing $\ell\,\theta_{\max}$ shrinks from $10.0$ to $8.4\ \mathrm{cm}$ — both energy and frequency rose, in the locked ratio $E/\omega$ .

Example 2 — Is this proton trapped? (magnetic mirror, exam level). A proton ( $m = 1.67\times10^{-27}\ \mathrm{kg}$ , $q = 1.60\times10^{-19}\ \mathrm{C}$ ) sits at the midplane of a mirror machine where $B_0 = 0.200\ \mathrm{T}$ , moving at $v = 1.00\times10^{6}\ \mathrm{m/s}$ with pitch angle $\alpha = 30.0°$ (so $v_\perp = v\sin\alpha$ ). The field rises to $B_{\max} = 1.00\ \mathrm{T}$ at the throats. Does the proton reflect, and if so at what field?

Step 1. Magnetic moment. $v_\perp = (1.00\times10^{6})(0.500) = 5.00\times10^{5}\ \mathrm{m/s}$ , so $\mu = m v_\perp^2/(2B_0) = (1.67\times10^{-27})(2.50\times10^{11})/(0.400) = 1.04\times10^{-15}\ \mathrm{J/T}$ .

Step 2. Reflection condition. The magnetic force does no work, so $v^2$ is constant; the particle reflects where all of it is perpendicular: $v_\perp^2(B_m) = v^2$ . With $\mu$ frozen, $v_\perp^2 = v_\perp^2(0)\,B/B_0$ , so $B_m = B_0\,v^2/v_{\perp0}^2 = B_0/\sin^2\alpha = 0.200/0.250 = \boldsymbol{0.800\ \mathrm{T}}$ .

Step 3. Trapping check. $B_m = 0.800\ \mathrm{T} < B_{\max} = 1.00\ \mathrm{T}$ , so the proton reflects before the throat: trapped. The general criterion is the loss cone: particles escape if $\sin^2\alpha < B_0/B_{\max} = 0.200$ , i.e. $\alpha < 26.6°$ . Our $30.0°$ clears it by only $3.4°$ — a collision that scatters the pitch angle below $26.6°$ ejects the proton, which is exactly why mirror machines leak through collisional scattering into the loss cone.

Step 4. Adiabaticity audit. The gyrofrequency at the midplane is $\omega_c = qB_0/m = (1.60\times10^{-19})(0.200)/(1.67\times10^{-27}) = 1.92\times10^{7}\ \mathrm{rad/s}$ — a gyroperiod of $0.33\ \mathrm{\mu s}$ , while the bounce between mirror points in a meters-long machine takes tens of microseconds. The field seen by the particle changes by a few percent per gyration at most: $\varepsilon \sim 10^{-2} \ll 1$ , comfortably adiabatic.

Historical arc

Origins · 1 minAlbert Einstein and how this was discovered

Albert Einstein

1879–1955 · Bern, Zurich, Berlin, Princeton

At the 1911 Solvay Congress — the first summit of the quantum crisis — Einstein answered Lorentz's pendulum puzzle in a sentence: shorten the string slowly and $E/\nu$ stays constant, so the quantum condition survives the change. The remark handed the old quantum theory its selection principle — quantize the adiabatic invariants — and Ehrenfest spent the following decade building it into a systematic program, calling it the adiabatic hypothesis. It was a characteristic Einstein move: he had already used the same instinct in 1905, asking which quantities stay meaningful when a light wave is slowly transformed, and he would use it again in 1917 when his radiation theory needed to know which classical structures survive contact with quanta. The question behind this whole article — what is robust under slow deformation? — is one he answered from the floor, apparently without needing to compute.

History

1902

— Rayleigh analyzes a pendulum with slowly varying length and finds

E/\nu

constant — adiabatic invariance before the quantum needed it

1911

— Solvay Congress: Lorentz asks what slow shortening does to a quantized pendulum; Einstein answers that

E/\nu

is invariant

1917

— Ehrenfest publishes the adiabatic hypothesis: the quantities to quantize are precisely the adiabatic invariants, welding old quantum theory to classical mechanics

1940

— Alfvén develops guiding-center theory; the magnetic moment

\mu

emerges as the first adiabatic invariant of plasma physics, foundation of mirror confinement

1962

— Kruskal, Lenard and others prove all-orders and exponential invariance theorems, turning a physicist's rule of thumb into rigorous asymptotics

Going further

References. Goldstein, Poole & Safko, Classical Mechanics, 3rd ed., §12.5 (Adiabatic Invariants) · Landau & Lifshitz, Mechanics, 3rd ed., §49–51 · Morin, Introduction to Classical Mechanics, Ch. 15 (Appendix on adiabatic invariants) · Taylor, Classical Mechanics, Ch. 14 · Arnold, Mathematical Methods of Classical Mechanics, §52

Provenance. A verifier should recompute: (1) the SHO action $J=E/\omega$ and the energy-balance derivation $\dot E/E=\dot\omega/\omega$ (Landau–Lifshitz §49); (2) the general averaging proof $\langle\dot J\rangle=0$ using $T=\oint\partial p/\partial E\,\mathrm{d}q$ (LL §49, Goldstein §12.5); (3) the magnetic-moment action $J\propto \mu=mv_\perp^2/2B$ with $r_L=mv_\perp/qB$ ; (4) the pendulum scalings $\theta_{\max}\propto\ell^{-3/4}$ , $E\propto\ell^{-1/2}$ from $E/\omega$ const with $\omega=\sqrt{g/\ell}$ ; (5) Bohr–Sommerfeld $\oint p\,\mathrm{d}q=nh$ and WKB $E_n=\hbar\omega(n+\tfrac12)$ . All are standard textbook results (Goldstein, Landau–Lifshitz, Morin, Taylor). The numeric examples (pendulum $\ell$ halved: $E_2 = 1.73\times10^{-2}$ J, $\theta_{\max,2}=0.168$ rad; mirror proton: $\mu = 1.04\times10^{-15}$ J/T, $B_m = 0.800$ T, loss cone $26.6°$ ) were computed for this article and are hand-verifiable.

Builds on: Action-Angle Variables · Action Integrals and Invariant Tori · Frequencies from Action-Angle Variables · Hamilton-Jacobi Equation Leads to: Liouville Integrable Systems · Liouville-Arnold Theorem · Adiabatic Theorem for Slowly Varying Hamiltonians · Hannay's Angle

If the energy changes, who is doing work on the oscillator — and why in exactly the ratio dE equals J d-omega?

The external agent moving the parameter does the work. Pulling the pendulum string against its time-varying tension, or ramping the current that makes $B$ , injects energy at a rate $\dot E = (\partial H/\partial\lambda)\dot\lambda$ — this is Step 2 of the general derivation. What the theorem adds is the bookkeeping: averaged over cycles, the injected energy is exactly what is needed to keep $E/\omega$ (more generally $J$ ) fixed, no more and no less. The agent cannot choose to put energy into the action — angle-averaged, the action channel is closed, and only the $E$ -per- $\omega$ ratio channel is open. A fast, phase-locked agent can pump the action: that is parametric resonance, the anti-adiabatic regime.

Doesn't Liouville's theorem already guarantee phase-space area is conserved? Why do we need a separate theorem?

Different areas. Liouville's theorem conserves the area of a material blob of initial conditions transported by the flow — exactly, for any time dependence, fast or slow. The adiabatic theorem concerns the area enclosed by one orbit's instantaneous energy contour, which is not a material curve: points on it evolve onto a slightly different contour as $\lambda$ changes. For fast driving, a material blob initially spread along an energy contour ends up smeared across many contours — Liouville holds, but each trajectory's action changed wildly. Slowness is what keeps the energy contour approximately material, and that approximate statement, with its $e^{-c/\varepsilon}$ error, is genuinely new content.

Is J conserved better or worse than energy here? I'm used to E being the sacred quantity.

For time-dependent Hamiltonians the roles invert. Energy is not conserved at all — $\dot E = \partial H/\partial t$ is first order in $\dot\lambda$ , and over the full switching $E$ changes by $O(1)$ (Einstein's pendulum gains 41% energy). The action drifts by $e^{-c/\varepsilon}$ : overwhelmingly better. This role reversal is the practical point of the whole subject — when parameters vary slowly, stop tracking energy and start tracking action. Energy is then derived: $E(t) = H(J,\lambda(t))$ with $J$ frozen at its initial value.

What happens with several degrees of freedom — is each J_i separately invariant?

Yes, provided the frequencies stay away from low-order resonances along the whole parameter path. The averaging argument runs on each angle separately, but a resonance $\mathbf{k}\cdot\boldsymbol\omega = 0$ makes some combination of angles slow, and a slow angle cannot be averaged away — the corresponding action combination can then drift ( $O(1)$ effects at strong resonances). Since $\omega_i(J,\lambda)$ moves as $\lambda$ varies, a long slow sweep generically walks through resonances; whether the invariant survives depends on crossing speed and resonance strength. This is the doorway to Degeneracy and Commensurate Frequencies and, ultimately, KAM theory.

How is the sudden approximation the opposite limit, concretely?

Take the particle in a box of length $L$ , wall moved to $2L$ . Adiabatic: $J \propto pL$ frozen, so $p \to p/2$ , $E \to E/4$ — the particle cools by doing work on the receding wall over many bounces. Sudden: the wall leaves before the particle next reaches it, so $(q,p)$ are untouched; the energy is unchanged, and the new action is $J' \propto p\,(2L) = 2J$ — the action doubled. Same initial state, same final box, opposite conserved quantity. Between the limits ( $\varepsilon \sim 1$ ) nothing clean is conserved, and you must integrate the dynamics.

Why is the invariance exponentially good rather than just order epsilon? That seems too strong to be true.

Within one cycle $J$ really does oscillate by $O(\varepsilon)$ — but those wobbles are reversible: each Fourier component $e^{ik\theta}$ of the perturbation integrates to something oscillatory, not secular. Push the cancellation to second order, then third: at every polynomial order the would-be drift is again a total derivative of an oscillatory quantity. The only obstruction is non-analyticity — kinks in $\lambda(t)$ inject a spectrum of fast Fourier components that resonate with the motion. For analytic $\lambda(t)$ the nearest singularity in complex time sits a distance $\sim 1/\varepsilon$ away in phase units, giving the $e^{-c/\varepsilon}$ estimate, exactly parallel to how WKB tunneling amplitudes arise beyond all orders in $\hbar$ .

Do adiabatic invariants survive dissipation, like a slowly damped oscillator?

No — and the accelerator-physics usage can confuse you here. Dissipation breaks the Hamiltonian structure: a weakly damped SHO loses energy at fixed $\omega$ , so $J = E/\omega$ decays with the energy; nothing is invariant. What accelerator physicists call "adiabatic damping" is Hamiltonian: as the beam's longitudinal momentum $p_0$ is ramped up, the transverse action in the scaled variables shrinks like $1/p_0$ , but the properly canonical (normalized) emittance is the frozen adiabatic invariant. Rule of thumb: first identify the honest canonical pair, then apply the theorem — and never to genuinely dissipative dynamics.

Lorentz's pendulum carried one quantum. Does classical adiabatic invariance actually prove the quantum state survives?

It made it consistent, which is what the old quantum theory needed in 1911: if states are selected by $\oint p\,\mathrm{d}q = nh$ and the left side cannot change under slow perturbation, then slow perturbations cannot dequantize a state or force transitions — quantum numbers are adiabatically rigid labels. The honest quantum statement came later (Born–Fock, 1928): a system in the $n$ -th discrete eigenstate stays in the $n$ -th instantaneous eigenstate under slow change, with transition amplitudes exponentially small — the quantum adiabatic theorem, whose classical shadow is exactly this article. The correspondence is quantitative: $E_n = \hbar\omega(n+\tfrac12)$ with $n$ fixed is $E/\omega = \text{const}$ .