Adiabatic Theorem for Slowly Varying Hamiltonians

Adiabatic Invariants showed you the phenomenon: change a parameter slowly and the action $J = \frac{1}{2\pi}\oint p\,\mathrm{d}q$ refuses to move, even as the energy drifts freely. This article is about the theorem behind the phenomenon — the precise statement, the canonical-transformation proof, the error estimates, and the exact boundaries where the guarantee dies. The distinction matters because "adiabatic invariance" is routinely invoked far outside mechanics, and knowing what is actually proven — and under which hypotheses — is what separates using the tool from waving it. Three questions organize everything. First: what does "slow" mean? Not a small rate, but a small dimensionless ratio $\varepsilon$ — the fractional parameter change per orbital period. Second: how good is the conservation? Astonishingly better than the $O(\varepsilon)$ a first estimate suggests: the drift vanishes at every polynomial order, and for analytic switching the total change is exponentially small, $e^{-c/\varepsilon}$ . That "beyond all orders" structure is the deep content, and the proof strategy — averaging, then iterating improved invariants — is a template you will meet again in perturbation theory, KAM theory, and quantum field theory. Third: when does it fail? Precisely where the fast/slow separation collapses: separatrix crossings, where the period diverges, and resonances, where a slow combination angle appears. The theorem's historical weight is hard to overstate: Ehrenfest made it the selection principle of the old quantum theory — quantize what slow perturbations cannot change — and its quantum descendant, the Born–Fock adiabatic theorem, now underwrites adiabatic quantum computation. The classical statement is where all of that machinery was first forged.

Key takeaways

The adiabatic theorem: for a one-degree-of-freedom system $H(q,p;\lambda)$ with bounded motion and $\lambda(t)$ varying slowly ( $\varepsilon = T|\dot\lambda|/\lambda \ll 1$ ), the action $J$ is conserved to all polynomial orders in $\varepsilon$ ; for analytic $\lambda(t)$ the total drift is exponentially small, $\Delta J \sim e^{-c/\varepsilon}$ .
The proof lives in action-angle variables: the time-dependent transformation adds $\partial F_2/\partial t \propto \dot\lambda$ to the Hamiltonian, and $\dot J = -\partial K/\partial\phi$ is a pure angle-derivative of a periodic function — its cycle average is exactly zero.
The all-orders statement comes from iteration: absorb the $O(\varepsilon)$ wobble into an improved action $\bar J = J + \varepsilon u_1$ , find the drift of $\bar J$ is $O(\varepsilon^2)$ , repeat; the series is asymptotic, not convergent, which is why the residue is exponential rather than zero.
Energy is not conserved — $J(E,\lambda) = \text{const}$ is what fixes how $E$ tracks the parameter (for the SHO, $E \propto \omega$ ).
Hypotheses matter: one degree of freedom (or non-resonant tori), period bounded away from infinity, smooth $\lambda(t)$ . Separatrix crossing and resonance crossing void the warranty — the action then jumps by $O(1)$ .
The quantum adiabatic theorem (Born–Fock 1928) is the direct descendant: quantum number $n$ frozen when the drive is slow compared to the gap, with Landau–Zener transitions as the failure mode.
Key idea: in a slowly varying periodic system the enclosed phase-space area $J$ is conserved to all orders in the rate of change — because the action's drive is a total angle-derivative of a periodic function, whose average over each fast cycle cancels identically.

The core insight

In a slowly varying periodic system the phase-space area $J=\frac{1}{2\pi}\oint p\,\mathrm{d}q$ enclosed by an orbit is conserved to all orders in the rate of change, even though the energy is not — the theorem, its proof by averaging over the fast angle, and its exponentially small error are the rigorous core of adiabatic physics.

Where this lives today

The theorem's modern career is mostly quantum, and the dictionary is precise. Born–Fock's quantum adiabatic theorem replaces "period stays finite" with "spectral gap stays open", and $e^{-c/\varepsilon}$ becomes the Landau–Zener transition probability $e^{-2\pi\Gamma}$ — separatrix crossing and gap closing are the same disease in two languages. That dictionary is now engineering: adiabatic quantum computation and quantum annealers encode optimization problems in a Hamiltonian's ground state and drag the system there slowly, their run-time budgets set exactly by the smallest gap on the path; STIRAP pulse sequences shuttle atomic populations through dark states adiabatically; and every cold-atom experiment that ramps a lattice "adiabatically" is invoking this article's error estimates, wobble included. On the classical side the theorem seeded two research programs. The iteration of improved invariants, pushed from asymptotic to rigorous, becomes Nekhoroshev's exponential stability estimates and, at the resonant tori where the construction fails, KAM theory — the modern account of which tori survive perturbation, previewed in Liouville-Arnold Theorem and settled at the end of this chapter's story. And the theorem's residue — what slow cycles leave behind after the action is returned intact — turned out to be a geometric object nobody had noticed for 150 years: the angle variable comes back shifted by a loop-dependent amount, Hannay's angle, the classical twin of Berry's phase and the subject of Hannay's Angle.

The physical picture

The theorem is a statement about two clocks. The system's internal clock ticks once per orbit, period $T$ . The external agent's clock ticks once per appreciable parameter change, timescale $\tau = \lambda/|\dot\lambda|$ . The adiabatic regime is $T \ll \tau$ : many internal ticks per external tick. Everything in the theorem — the averaging, the error estimates, the failure modes — is a consequence of that hierarchy, and the single dimensionless number $\varepsilon = T/\tau$ controls it all.

Figure

Two clocks: the orbit ticks many times before the parameter moves — the slowness parameter is the ratio of the two periods

Why should the hierarchy freeze anything? Because of a cancellation with the flavor of a checkbook audit. During each orbit the changing parameter pushes the action up along part of the cycle and down along the rest; a fast orbit samples all phases of the push before the push itself has changed, so the credits and debits cancel to the accuracy with which the push was constant during the cycle — one factor of $\varepsilon$ . The genius of the theorem is that this bookkeeping can be iterated: the leftover $O(\varepsilon)$ wobble is itself oscillatory, cancels at the next order, and so on down. Drift survives only "beyond all orders" — the same mathematical crawlspace where quantum tunneling amplitudes live. The picture also announces its own failure modes. If the orbit slows to a crawl — as it does near a separatrix, where the period diverges logarithmically — the fast clock stops being fast, $\varepsilon$ blows up locally, and the cancellation dies with it. A pendulum whose swing is slowly pumped toward the top of its arc must violate adiabaticity at the crossover from swinging to spinning, no matter how slow the pump.

Figure

Failure at the separatrix: as the orbit approaches the pendulum's separatrix the period diverges and the fast/slow hierarchy collapses

The theorem, precisely

The object under protection is the action of the instantaneous frozen system,

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

where: $J$ — the action at the current $(E,\lambda)$ , the loop taken over one period of the orbit that the system would execute if $\lambda$ were pinned at its present value; $p,q$ — the canonical pair. This is the same construction as in Action-Angle Variables; nothing new yet. The hypothesis is the timescale hierarchy, stated as a dimensionless smallness:

\varepsilon \equiv \frac{T\,|\dot\lambda|}{\lambda}\ll 1,\qquad T=\frac{2\pi}{\omega} \qquad (2)

where: $T$ — the instantaneous orbital period; $\lambda$ — the slow parameter. Two fine points. The hypothesis must hold along the entire path: a single moment where $T$ diverges (separatrix) breaks the chain. And $\varepsilon$ small does not bound the total excursion of $\lambda$ — the theorem happily covers a spring stiffened by a factor of ten, provided each period sees only a sliver of the change. The conclusion, in its honest asymptotic form:

\Delta J = O(\varepsilon^{n})\ \ \text{for every } n, \qquad \Delta J \sim e^{-c/\varepsilon}\ \ \text{for analytic } \lambda(t) \qquad (3)

where: $\Delta J$ — the net action change over the whole switching history; $c$ — a constant set by the width of the analyticity strip of $\omega(\lambda(t))$ in complex time. Read this carefully: it is not the claim $\dot J = 0$ . Within a cycle $J$ wobbles by $O(\varepsilon)$ ; the theorem forbids the wobble from accumulating. The frequency that governs the fast clock is the one computed in Frequencies from Action-Angle Variables:

\omega(J,\lambda)=\frac{\partial H}{\partial J}\bigg|_\lambda \qquad (4)

where: $H(J,\lambda)$ — the frozen Hamiltonian expressed in action-angle form. The theorem needs $\omega$ bounded away from zero along the path — this is exactly the "no separatrix" clause.

The proof machine: a time-dependent canonical transformation

Fix the strategy before the algebra. Transform to the action-angle variables of the instantaneous Hamiltonian. At frozen $\lambda$ that transformation trivializes the dynamics; the entire effect of the slow drive is that the transformation itself now moves. A moving canonical transformation shifts the Hamiltonian by the time-derivative of its generating function (the mechanism built in Generating Function to a Trivial Hamiltonian):

K(J,\phi,t)=H(J,\lambda(t))+\frac{\partial F_2}{\partial t}, \qquad \frac{\partial F_2}{\partial t}=\frac{\partial F_2}{\partial\lambda}\,\dot\lambda \qquad (5)

where: $K$ — the Hamiltonian governing the new pair $(J,\phi)$ ; $F_2(q,J;\lambda)$ — the type-2 generating function of the frozen transformation, which depends on time only through $\lambda(t)$ . Equation (5) is the whole battlefield: the first term cannot move $J$ (it is $\phi$ -independent), so the entire drive of the action is the second term — explicitly first order in $\dot\lambda$ , i.e. in $\varepsilon$ .

Derivation AThe averaged action equation: why the drive integrates to zero

We prove the leading-order theorem: the cycle average of $\dot J$ vanishes identically, so the action suffers no secular drift at first order in the slowness. The strategy is bare-hands canonical mechanics — write Hamilton's equation for $J$ under the transformed Hamiltonian (5), observe that the drive is a pure angle-derivative, and integrate it over one period. The step to watch is the last one: the integral vanishes not because something is small, but because of an exact structural fact — single-valuedness of the generating function's $\lambda$ -derivative around the loop.

Step 1. At frozen $\lambda$ the system is integrable: $H_0(J,\lambda)$ with $\dot J = 0$ and $\dot\phi = \omega(J,\lambda)$ . This is the starting platform, exact for constant $\lambda$ .

Step 2. Let $\lambda = \lambda(t)$ . The transformation to $(J,\phi)$ becomes time-dependent, and by equation (5) the new Hamiltonian is $K = H_0(J,\lambda(t)) + (\partial F_2/\partial\lambda)\,\dot\lambda$ .

Step 3. Hamilton's equation for the action:

\dot J=-\frac{\partial K}{\partial\phi}=-\dot\lambda\,\frac{\partial}{\partial\phi}\!\left(\frac{\partial F_2}{\partial\lambda}\right)

Why only the drive term contributes: $H_0$ is a function of $J$ and $\lambda$ alone — angle-independence is what "action-angle variables" means — so its $\phi$ -derivative is zero, and $\dot J$ is manifestly $O(\dot\lambda)$ .

Step 4. Average over one fast cycle at essentially constant $\dot\lambda$ :

\langle\dot J\rangle=-\frac{\dot\lambda}{2\pi}\int_0^{2\pi}\frac{\partial}{\partial\phi}\!\left(\frac{\partial F_2}{\partial\lambda}\right)\mathrm{d}\phi=-\frac{\dot\lambda}{2\pi}\left[\frac{\partial F_2}{\partial\lambda}\right]_0^{2\pi}=0

Why the boundary terms cancel: $F_2$ generates a well-defined transformation on the phase-space cylinder, so $\partial F_2/\partial\lambda$ is a single-valued, $2\pi$ -periodic function of the angle; its values at $\phi = 0$ and $\phi = 2\pi$ are the same point. The integral of a derivative of a periodic function around its period is exactly zero — no smallness invoked.

Step 5. Hence the secular drift vanishes at $O(\varepsilon)$ : over each cycle, $\Delta J$ per cycle is $O(\varepsilon^2)$ (the residue comes from $\dot\lambda$ and $\omega$ themselves creeping during the cycle), and over the switching time $\sim 1/\varepsilon$ cycles accumulate at most $O(\varepsilon)$ — already conserved in the limit. The next derivation upgrades this to all orders. Verified: matches Landau–Lifshitz §49 and Goldstein §12.5; Step 4's periodicity argument is the standard one, and setting $H_0 = \omega J$ (the SHO) reproduces $E/\omega = \mathrm{const}$ as in the sibling article's energy-balance proof.

Go deeper · Derivation B · 2 minAll orders: the iteration of improved invariants

Derivation BAll orders: the iteration of improved invariants

The leading-order result leaves $J$ wobbling by $O(\varepsilon)$ within each cycle, and one could fear those wobbles secretly accumulate at $O(\varepsilon^2)$ . We now sketch the iteration — due in rigorous form to Kruskal, Lenard, and Arnold's school — showing the drift vanishes at every polynomial order. The strategy is to redefine the invariant: absorb the known oscillatory wobble into an improved action variable, then show the improved variable obeys the same structural equation one order down. Watch the two ingredients each step consumes: periodicity in the angle (as before) and $\omega$ bounded away from zero (to solve for the correction).

Step 1. From the previous derivation, $\dot J = \varepsilon\,g(J,\phi,\lambda)$ with $g$ $2\pi$ -periodic in $\phi$ and $\langle g\rangle_\phi = 0$ (writing $\varepsilon$ for the bookkeeping small parameter carried by $\dot\lambda$ ).

Step 2. Define the improved action $\bar J = J + \varepsilon\,u_1(J,\phi,\lambda)$ , choosing $u_1$ to cancel the oscillation:

\omega(J,\lambda)\,\frac{\partial u_1}{\partial\phi}=-\,g(J,\phi,\lambda) \quad\Longrightarrow\quad u_1=-\frac{1}{\omega}\int^{\phi} g\,\mathrm{d}\phi'

Why this is solvable: $g$ has zero mean, so its antiderivative in $\phi$ is again periodic — no secular term in $u_1$ — and division by $\omega$ is legitimate precisely because the period stays finite. This is where the no-separatrix hypothesis earns its keep.

Step 3. Compute $\dot{\bar J} = \dot J + \varepsilon(\partial_\phi u_1)\dot\phi + \varepsilon(\partial_J u_1)\dot J + \varepsilon(\partial_\lambda u_1)\dot\lambda$ . The first two terms cancel by construction (that was the choice of $u_1$ ); every survivor carries at least $\varepsilon^2$ and is again periodic in $\phi$ with zero mean after averaging.

Step 4. Iterate: $\bar{\bar J} = \bar J + \varepsilon^2 u_2$ , and so on. At each order the same two ingredients — zero-mean periodicity and finite $\omega$ — allow the construction. Conclusion: for every $n$ there is an $n$ -th improved invariant conserved up to $O(\varepsilon^{n+1})$ , and since all the $u_k$ corrections are themselves $O(\varepsilon)$ small, the original $J$ drifts by less than any power of $\varepsilon$ .

Step 5. The series does not converge — it is asymptotic. Optimally truncating at $n \sim c/\varepsilon$ leaves the exponential residue $\Delta J \sim e^{-c/\varepsilon}$ of equation (3), with $c$ set by where $\omega(\lambda(t))$ loses analyticity in complex time. Non-smooth switching breaks the chain early: a $C^k$ ramp yields only $\Delta J = O(\varepsilon^k)$ . Verified: this is the standard Kruskal–Lenard iteration (Arnold, Mathematical Methods, §52; Landau–Lifshitz §51 for the exponential estimate); the solvability condition in Step 2 is the one-frequency case of the averaging principle, and its failure at $\omega \to 0$ correctly predicts the separatrix breakdown.

The benchmark: the oscillator ellipse

Every abstract theorem deserves one exactly computable case. For $H = \frac{p^2}{2m} + \frac12 m\omega^2 q^2$ the orbit at energy $E$ is an ellipse with semi-axes $\sqrt{2E/(m\omega^2)}$ in $q$ and $\sqrt{2mE}$ in $p$ , so the enclosed area is $\pi\sqrt{2mE}\sqrt{2E/(m\omega^2)} = 2\pi E/\omega$ , giving

J=\frac{E}{\omega} \quad\Longrightarrow\quad \frac{E_f}{E_i}=\frac{\omega_f}{\omega_i},\qquad A=\sqrt{\frac{2J}{m\omega}}\propto\omega^{-1/2} \qquad (6)

where: $A$ — the oscillation amplitude. The theorem then dictates the full slow evolution with no differential equation solved: energy tracks frequency linearly, amplitude falls as $\omega^{-1/2}$ — the same $A \propto p^{-1/2}$ law that reappears as the WKB amplitude factor in WKB and the Eikonal Connection. The semiclassical reading of (6) is the punchline of the chapter: $J = E/\omega$ frozen is precisely $E = (n+\tfrac12)\hbar\omega$ with $n$ frozen.

Limiting cases

Case	Result	Meaning
Adiabatic limit	$\varepsilon\to 0$ : $\Delta J = O(\varepsilon^n)\ \forall n$ ; analytic $\lambda(t)$ : $\Delta J\sim e^{-c/\varepsilon}$	The theorem at full strength; smoothness of the ramp sets the order actually achieved.
Sudden limit	$\varepsilon\gg1$	Opposite protection: $(q,p)$ frozen during the jump, energy unchanged, action not conserved.
Separatrix crossing	$\omega(E,\lambda)\to0$ on the path	Step 2 of the iteration divides by $\omega$ : the construction dies, and $J$ jumps by $O(1)$ with a quasi-random, crossing-phase-dependent outcome.
Resonance (N degrees of freedom)	$\mathbf{k}\cdot\boldsymbol\omega=0$	A slow combination angle appears that cannot be averaged; single-frequency proof no longer applies — the theme of Degeneracy and Commensurate Frequencies.
Harmonic oscillator	$J=E/\omega$ exactly	Benchmark where every claim can be checked against closed-form asymptotics.

Common trapThe theorem does not promise small energy change — or exact action conservation

Two misreadings account for most misuse. First: energy is not protected — $\dot E = \partial H/\partial t \neq 0$ , and over the full ramp $E$ changes by $O(1)$ ; the theorem's content is that $E$ slaves itself to $\lambda$ through $J(E,\lambda) = \text{const}$ . Second: $J$ is not exactly conserved — it wobbles by $O(\varepsilon)$ within every cycle, and the exponential estimate applies to the accumulated drift after smooth switching, not to instantaneous values. Claims of "exact" adiabatic invariants (Lewis–Riesenfeld-type constructions) involve different, genuinely exact invariants that coincide with $J$ only asymptotically.

Worked examples

Example 1 — Slowly stiffened spring, end to end. A mass $m = 0.100\ \mathrm{kg}$ oscillates with amplitude $A_i = 2.00\ \mathrm{cm}$ on a spring of stiffness $k_i = 100\ \mathrm{N/m}$ . Over $60.0$ s the stiffness is ramped smoothly to $k_f = 400\ \mathrm{N/m}$ . Find the final energy and amplitude, and audit the slowness.

Step 1. Frequencies: $\omega_i = \sqrt{k_i/m} = \sqrt{1000} = 31.6\ \mathrm{rad/s}$ ; $\omega_f = \sqrt{4000} = 63.2\ \mathrm{rad/s}$ — the frequency doubles.

Step 2. Initial energy and action: $E_i = \tfrac12 k_i A_i^2 = \tfrac12(100)(0.0200)^2 = 2.00\times10^{-2}\ \mathrm{J}$ ; $J = E_i/\omega_i = 2.00\times10^{-2}/31.6 = 6.33\times10^{-4}\ \mathrm{J\,s}$ .

Step 3. Freeze $J$ : $E_f = J\,\omega_f = (6.33\times10^{-4})(63.2) = \boldsymbol{4.00\times10^{-2}\ \mathrm{J}}$ — the energy doubles with the frequency, the extra $2.00\times10^{-2}$ J supplied by whatever stiffens the spring.

Step 4. Final amplitude: $A_f = \sqrt{2E_f/k_f} = \sqrt{0.0800/400} = \sqrt{2.00\times10^{-4}} = \boldsymbol{1.41\ \mathrm{cm}} = A_i/\sqrt2$ , confirming $A \propto \omega^{-1/2}$ .

Step 5. Slowness audit: $\dot\omega \approx (63.2-31.6)/60.0 = 0.527\ \mathrm{rad/s^2}$ , so $\varepsilon = \dot\omega/\omega^2 \lesssim 0.527/(31.6)^2 = 5.3\times10^{-4}$ . With $\varepsilon \sim 5\times10^{-4}$ , even the crude $O(\varepsilon)$ bound is a hundredth of a percent, and the exponential estimate $e^{-c/\varepsilon}$ makes the drift unmeasurable. The ramp could be a thousand times faster and still be safely adiabatic.

Example 2 — Box with a receding wall: the theorem verified by collisions. A particle of speed $v_i = 100\ \mathrm{m/s}$ bounces between walls $L_i = 0.100$ m apart; one wall recedes at constant $u = 1.00\times10^{-2}\ \mathrm{m/s}$ until $L_f = 0.200$ m (duration $10.0$ s). Predict the final speed adiabatically, then verify by counting collisions.

Step 1. Adiabatic prediction. The action for wall bouncing is $J = \frac{1}{2\pi}\oint p\,\mathrm{d}q = \frac{mvL}{\pi}$ , so $vL$ is frozen: $v_f = v_i L_i/L_f = (100)(0.100/0.200) = \boldsymbol{50.0\ \mathrm{m/s}}$ , and the energy falls to a quarter — the classical shadow of $E_n \propto n^2/L^2$ at fixed $n$ .

Step 2. Slowness audit: bounce period $T = 2L/v \approx 2(0.1)/100 = 2.00\times10^{-3}$ s initially; the wall takes $\tau = L/u = 10$ s to move one box length. $\varepsilon \sim T/\tau = 2\times10^{-4}$ : deeply adiabatic. The particle makes thousands of bounces during the expansion.

Step 3. Mechanical verification. Each collision with the receding wall is elastic in the wall frame, so the particle returns with lab speed reduced by exactly $2u$ : $\Delta v = -2u$ per hit. Hits occur every $2L/v$ seconds, so $\mathrm{d}v/\mathrm{d}t = -2u\,v/(2L) = -uv/L$ . With $\mathrm{d}L/\mathrm{d}t = u$ , divide the two rates: $\mathrm{d}v/v = -\mathrm{d}L/L$ , which integrates to $vL = \text{const}$ — exactly the adiabatic invariant, recovered collision by collision with no averaging language at all.

Step 4. Numbers: $v$ falls from $100$ to $50.0\ \mathrm{m/s}$ over roughly $N = \int v\,\mathrm{d}t/(2L) \approx 5\times10^{3}$ collisions, each stealing $2u = 2.00\times10^{-2}\ \mathrm{m/s}$ — total $\approx 100 - 5000\times0.02\cdot(\text{avg factor}) \to 50\ \mathrm{m/s}$ , consistent. The invariant $vL$ passes the microscopic audit.

Historical arc

Origins · 1 minPaul Ehrenfest and how this was discovered

Paul Ehrenfest

1880–1933 · Vienna, St. Petersburg, Leiden

Ehrenfest turned Einstein's Solvay remark into a principle. Between 1913 and 1917 he asked the old quantum theory's sharpest question — which classical quantities may carry quantum conditions? — and answered: only adiabatic invariants, since slow deformation of the apparatus must not force discontinuous jumps between allowed states. His "adiabatic hypothesis" let quantization rules be transported from the harmonic oscillator to any system reachable by slow deformation, making the Bohr–Sommerfeld program systematic rather than case-by-case guesswork. Ehrenfest was the conscience of early quantum theory in other ways too: his 1911 dissection of the equipartition catastrophe, the Ehrenfest theorem connecting quantum expectation values to classical equations, and a Leiden seminar whose ruthless clarity Einstein, Bohr and Pauli all sought out. Depression, deepened by the rise of Nazism and worry for his disabled son, ended his life in 1933; his demand that quantum mechanics keep an honest classical bookkeeping survives in every adiabatic argument in this chapter.

History

1911

— Einstein, at Solvay, answers Lorentz: a slowly shortened quantum pendulum keeps

E/\nu

— the seed observation

1913

— Ehrenfest states the adiabatic hypothesis; quantum conditions may only be imposed on adiabatic invariants

1917

— Ehrenfest's definitive paper welds the classical theorem to the old quantum theory's selection rules

1928

— Born and Fock prove the quantum adiabatic theorem: slow drives preserve the quantum number, with gap-controlled errors

1962

— Kruskal and Lenard establish all-orders invariance; exponential estimates for analytic Hamiltonians follow in the Arnold school

Going further

References. Landau & Lifshitz, Mechanics (Vol. 1), §49 · Goldstein, Poole & Safko, Classical Mechanics, 3rd ed., §12.5 · Morin, Introduction to Classical Mechanics, Ch. 15 (adiabatic invariants) · Taylor, Classical Mechanics, §13.7 · Arnold, Mathematical Methods of Classical Mechanics, §52

Provenance. Equations cross-checked against Landau-Lifshitz Mechanics §49 (adiabatic invariants, $\langle\dot J\rangle=0$ derivation), Goldstein §12.5, and Morin Ch.15. A verifier should recompute: (1) the ellipse area $\oint p\,\mathrm{d}q=2\pi E/\omega$ for the oscillator; (2) the pendulum scaling $\theta_0\propto\ell^{-3/4}$ , $E\propto\ell^{-1/2}$ ; (3) the box result $E\propto L^{-2}$ ; (4) the vanishing of $\langle\dot J\rangle$ via periodicity of $\partial_\lambda F_2$ in $\phi$ . No fabricated numerical constants are used; all relations are exact analytic scalings. The numeric examples (spring ramp: $E_f = 4.00\times10^{-2}$ J, $A_f = 1.41$ cm, $\varepsilon\approx5\times10^{-4}$ ; receding wall: $v_f = 50.0$ m/s and the collision-rate integration $\mathrm{d}v/v = -\mathrm{d}L/L$ ) were computed for this article and are hand-verifiable.

Builds on: Action-Angle Variables · Action Integrals and Invariant Tori · Frequencies from Action-Angle Variables · Generating Function to a Trivial Hamiltonian Leads to: Adiabatic Invariants · Hannay's Angle · Liouville-Arnold Theorem · Hamilton-Jacobi Equation

What exactly is proven, versus physicist's folklore? Is there a rigorous theorem with hypotheses I can cite?

Yes. For one degree of freedom with $H(q,p;\lambda)$ smooth, motion bounded, $\omega$ bounded below along the path, and $\lambda(\varepsilon t)$ smooth: $|J(t) - J(0)| \le C\varepsilon$ uniformly for $t \lesssim 1/\varepsilon$ (classical averaging theorem), upgraded to $O(\varepsilon^n)$ by the improved-invariant iteration and to $e^{-c/\varepsilon}$ for analytic data (Kruskal–Lenard–Arnold school; see Arnold §52 for precise statements). What remains folklore-adjacent is the multi-frequency case: for $n \ge 2$ degrees of freedom the guarantee holds only away from resonant surfaces, and honest statements (Neishtadt-type) carry measure-theoretic qualifications about which initial conditions avoid resonance trapping. Quote the 1-DOF theorem freely; quote the $n$ -DOF one with care.

Where does the proof actually use that F2's lambda-derivative is periodic in the angle — and could that fail?

Step 4 of the first derivation integrates $\partial_\phi(\partial_\lambda F_2)$ over a period and kills it by periodicity. The function $F_2$ itself is not periodic in the angle — it grows by $2\pi J$ per circuit, which is fine because only derivatives enter — but $\partial_\lambda F_2$ taken at fixed $(q,J)$ is built from the momentum function $p(q,J,\lambda)$ , which is single-valued on the orbit. Failure modes are geometric: if the orbit topology changes during the sweep (libration to rotation at a separatrix), "one circuit" stops being well-defined and the periodicity argument has nothing to hold on to. That is the honest mechanical reason the separatrix is fatal, complementing the analytic reason (division by $\omega \to 0$ in the iteration).

Why does the naive estimate give order epsilon while the truth is exponentially small? Where does the enhancement come from?

The naive estimate treats $\dot J = O(\varepsilon)$ as if it could accumulate coherently over the switching time $1/\varepsilon$ , predicting $O(1)$ — already the cycle average kills that to $O(\varepsilon)$ . The enhancement beyond is Fourier-analytic: after averaging, the surviving drive lives entirely in the harmonics $e^{ik\phi}$ of the fast angle, and a slow ramp's spectrum has essentially no power at frequency $k\omega$ . For analytic $\lambda(t)$ the spectral density at frequency $\omega$ falls like $e^{-\omega t_*}$ , where $t_*$ is the distance to the nearest complex-time singularity — that is the constant $c$ in $e^{-c/\varepsilon}$ . A kink in the ramp ( $C^k$ only) injects a power-law tail into the spectrum and degrades the result to $O(\varepsilon^k)$ : smoothness is literally spectral hygiene.

How does this theorem relate to the quantum adiabatic theorem — same theorem in different clothes, or genuinely different?

Same architecture, different protected object, and the dictionary is exact enough to compute with. Classical: protected object $J$ ; slowness compared to the orbital period; failure when $\omega \to 0$ ; residue $e^{-c/\varepsilon}$ . Quantum (Born–Fock): protected object the quantum number $n$ (occupation of the instantaneous eigenstate); slowness compared to $\hbar/\Delta E$ with $\Delta E$ the spectral gap; failure when the gap closes; residue the Landau–Zener probability $e^{-2\pi\Gamma}$ . The bridge is semiclassics: $J = (n+\tfrac12)\hbar$ maps frozen action to frozen quantum number, and neighboring-level spacing $\Delta E = \hbar\omega$ maps the gap condition to the frequency condition. They are the same theorem in the semiclassical regime; deep in the quantum regime (few levels, no classical limit) only the quantum statement survives.

What actually happens to the action during a separatrix crossing — is it random?

Quantitatively understood, and "pseudo-random" is the right word. As the orbit spirals toward the separatrix the period grows logarithmically; averaging degrades gracefully until the last few cycles, where the theory hands control to a crossing analysis. The outcome: the action just after crossing differs from just before by an $O(\varepsilon\ln\varepsilon)$ spread whose value depends with extreme sensitivity on the crossing phase — deterministic, but so sensitive to initial conditions that ensembles are effectively randomized (Neishtadt's separatrix-crossing formulas make the distribution explicit). Physical consequences are real: capture probabilities into spin-orbit resonances (Mercury's 3:2 state) and into mean-motion resonances of migrating planets are computed with exactly these formulas.

Is there a version of the invariant for many degrees of freedom, like the whole solar system?

Each non-resonant torus carries its full set $J_1,\dots,J_n$ , and the theorem protects them jointly while the frequencies stay non-resonant. But a slow sweep drags the frequency vector through the dense web of resonances $\mathbf{k}\cdot\boldsymbol\omega = 0$ ; most crossings are fast and harmless (high-order, exponentially weak), while low-order crossings can capture or kick the actions. There is also a genuinely global invariant in the ergodic limit: the phase-space volume enclosed by the energy shell (the Hertz invariant), whose conservation under slow driving is the mechanical root of the thermodynamic adiabatic process — same word, same theorem, via $S = k\ln\Omega$ . For the solar system, Laskar's simulations show the theorem's protections eroding over Gyr timescales exactly at the predicted resonant bottlenecks.

Do I need the exponential estimate in practice, or is leading-order averaging always enough?

Usually leading order suffices — $\varepsilon$ in laboratory situations is so small (Example 1: $5\times10^{-4}$ ) that even $O(\varepsilon)$ bounds are invisible. The exponential statement earns its keep in three places. Long-time problems: over $10^9$ orbits (accelerator storage rings, planetary dynamics) an $O(\varepsilon^2)$ per-cycle drift would accumulate fatally, and the all-orders theorem is what forbids it. Competitive effects: when you measure a small action change (Hannay-angle experiments, emittance-growth budgets), you must know the theorem's own contribution is $e^{-c/\varepsilon}$ , i.e. zero, so anything seen is real physics. And conceptual honesty: the asymptotic (not convergent) nature of the series is the classical prototype for divergent perturbation series in quantum field theory — the first place students meet "beyond all orders" done right.