Canonical Perturbation Theory

Between 1846 and 1867, Charles-Eugène Delaunay computed the Moon's motion by hand: an expansion of the lunar Hamiltonian carried to seventh order in the small parameters, filling two enormous volumes with a single formula for the perturbed orbit. His engine was the machinery this article builds — canonical perturbation theory, the systematic art of solving a mechanics problem that is almost one you have already solved exactly. The solar system is the founding example: each planet's motion is a Kepler problem plus corrections of one part in a thousand from the other planets, and the entire predictive triumph of celestial mechanics — ephemerides, eclipse tables, the discovery of Neptune from Uranus's residuals — rests on turning "almost integrable" into a working expansion. The action-angle articles gave you the perfect starting point: an integrable Hamiltonian $H_0(J)$ whose motion is uniform winding on invariant tori. Now add $\varepsilon H_1(J, \theta)$ . The naive move — expand the trajectory in powers of $\varepsilon$ — fails spectacularly, producing secular terms that grow linearly in time and wreck the approximation just when you need it. The canonical cure is characteristically elegant: don't perturb the trajectory, perturb the coordinate system — find new action-angle variables in which the perturbed problem is again integrable, order by order. The reward is the physics of frequency renormalization: the pendulum clock that runs slow at large amplitude, computed in two lines. And then the sting. At second glance the method's denominators contain $k \cdot \omega$ — and near resonances these small divisors blow the series apart. Poincaré proved the divergence is real and profound. Understanding exactly how canonical perturbation theory fails is the road to the KAM theorem you have already met; this article is where that road begins.

Key takeaways

Setup: $H = H_0(J) + \varepsilon H_1(J, \theta)$ in the action-angle variables of the integrable $H_0$ , with unperturbed frequencies $\omega = \partial H_0/\partial J$ .
Naive time-series expansion generates secular terms $\propto t\sin\omega t$ : the true motion has a shifted frequency, and expanding a shifted cosine in powers of $\varepsilon$ manufactures fake unbounded growth.
The canonical fix: a near-identity canonical transformation $S = \theta \cdot J' + \varepsilon S_1(\theta, J')$ chosen so the new Hamiltonian depends only on the new actions: $K(J') = H_0(J') + \varepsilon\langle H_1\rangle + O(\varepsilon^2)$ .
First-order results: energy shift = the angle-average $\langle H_1\rangle$ ; frequency shift $\Delta\omega = \varepsilon\,\partial\langle H_1\rangle/\partial J$ ; the oscillating part of $H_1$ is absorbed into the transformation, not the dynamics.
Worked classics: quartic oscillator $\omega = \omega_0 + 3\varepsilon A^2/2m\omega_0$ (verified numerically to 4 digits here); pendulum $\omega = \omega_0(1 - \varphi_m^2/16)$ — a 30° grandfather clock loses ~24 min/day against its 5° rating.
The generating function's Fourier coefficients are $S_{1,k} = iH_{1,k}/(k \cdot \omega)$ : near resonances $k \cdot \omega \approx 0$ these small divisors diverge — resonant tori are dense, the series is generically divergent (Poincaré), and its rescue on Diophantine tori is precisely the KAM theorem.
Key idea: perturb the coordinates, not the trajectory — average the perturbation over the fast angles to get the new integrable motion, and pay for the elegance at every resonance, where the small divisors announce that some tori cannot be deformed, only destroyed.

The core insight

For a near-integrable Hamiltonian $H_0(J) + \varepsilon H_1(J,\theta)$ , a near-identity canonical transformation removes the angle dependence order by order, yielding renormalized actions with energy $K = H_0 + \varepsilon\langle H_1\rangle + O(\varepsilon^2)$ and shifted frequencies $\omega + \varepsilon\,\partial_J\langle H_1\rangle$ — an expansion whose coefficients carry the small divisors $1/(k\cdot\omega)$ that diverge on resonant tori and make the theory asymptotic rather than convergent, the precise pathology the KAM theorem was built to tame.

Where this lives today

Canonical perturbation theory is production software in three industries. Spaceflight: satellite orbit prediction runs on averaged equations — the workhorse SGP4 propagator behind every piece of space-situational-awareness infrastructure is a first-order canonical theory (Brouwer's 1959 solution, Lie-transformed descendants) of Earth's oblateness perturbation $J_2$ , with $\langle H_1 \rangle$ delivering the secular node regression and perigee precession that mission designers exploit deliberately (Sun-synchronous orbits are an application of equation (3)). Accelerator physics: a storage ring is a multi-torus Hamiltonian system in which the betatron tunes must be steered away from the small-divisor chasms — the "tune diagram" pinned in every control room is the resonance comb computed in this article, with beam lifetimes decided by exactly the $k\cdot\omega$ arithmetic of Derivation C. Plasma confinement: guiding-centre and gyrokinetic theories — the formulation in which tokamak turbulence is simulated — are Lie-series canonical perturbation expansions in the gyroradius, averaging out the fast Larmor angle to produce the slow drift Hamiltonian. The conceptual legacy is broader still. Modern implementations have replaced von Zeipel's mixed-variable generating functions with Lie transforms (Deprit, Hori) — same physics, explicit recursions that computer algebra executes to tenth order without complaint — and the same operator machinery reappears as normal-form theory across dynamics, as the Birkhoff normal form near equilibria, and (with commutators for Poisson brackets) as time-independent perturbation theory in quantum mechanics, whose energy shift $\langle\psi|\hat H_1|\psi\rangle$ is the direct heir of $\langle H_1\rangle$ , small denominators $E_n - E_m$ included. And the divergence story matured into physics rather than pathology: Nekhoroshev estimates bound chaotic diffusion for exponentially long times, explaining why the demonstrably non-integrable solar system has nonetheless behaved itself for $4.5$ billion years — the deepest answer yet to the question Delaunay's volumes first made quantitative.

The physical picture

Start with what integrability bought you. In action-angle variables, the motion of $H_0(J)$ is the simplest dynamics imaginable: every action frozen, every angle advancing uniformly at $\omega_i = \partial H_0/\partial J_i$ — trajectories wind around invariant tori like thread on a spool. Now switch on a weak coupling $\varepsilon H_1(J, \theta)$ : the actions begin to breathe ( ${\dot J = -\varepsilon\,\partial H_1/\partial\theta \ne 0}$ ), the winding wobbles, and the exact tori are gone. But for small $\varepsilon$ the motion still looks torus-like — slightly warped spools, slightly uneven winding. Perturbation theory's wager is that the warp can be transformed away: that there exist new, slightly distorted action-angle coordinates $(J', \theta')$ in which the perturbed motion is again frozen actions and uniform angles. Why coordinates rather than trajectories? Because the trajectory expansion breaks on a subtlety of oscillators that you can see with no formalism at all. The perturbed system oscillates at a slightly shifted frequency — $\cos[(\omega + \varepsilon\Delta)t]$ . Expand that in powers of $\varepsilon$ : the first correction is $-\varepsilon\Delta\,t\sin(\omega t)$ , growing without bound in time. Nothing in the physics grows — the orbit is a barely-distorted circle — but the expansion of a detuned cosine manufactures growth, and by $t \sim 1/\varepsilon$ the "small correction" is as large as the answer. These secular terms poisoned two centuries of lunar theory. The canonical method disposes of them by construction: it never expands oscillations in time; it renormalizes the frequency itself, putting the $\varepsilon\Delta$ inside the cosine where it belongs.

Figure

Perturb the coordinates, not the trajectory — the near-identity transformation straightens the warped winding

The transformed picture also announces its own limits. The coordinate change must wiggle in step with the perturbation's angle dependence — each Fourier mode $e^{ik\cdot\theta}$ of $H_1$ requires a coordinate wiggle of amplitude $H_{1,k}/(k\cdot\omega)$ . When the torus's frequencies are resonant ( $k\cdot\omega = 0$ for some integer vector $k$ ), the required wiggle is infinite: no smooth deformation of that torus can absorb a perturbation that pushes in sync with the winding. Resonance is not a technical nuisance — it is the mechanism by which perturbations do qualitatively new things (islands, separatrices, chaos), and the perturbation series registers this honestly, by diverging. The dense set of resonant tori, the survival of the "very irrational" ones, the destruction of the rest: you have seen the outcome in the KAM article. Here you see the origin.

The math: the expansion, its first order, and its price

The problem, in the variables integrability gave us:

H(J, \theta) = H_0(J) + \varepsilon H_1(J, \theta), \qquad \omega(J) \equiv \frac{\partial H_0}{\partial J} \qquad (1)

where: $(J, \theta)$ are $n$ pairs of action-angle variables of $H_0$ , $H_1$ is $2\pi$ -periodic in every angle — so it carries a Fourier expansion $H_1 = \sum_k H_{1,k}(J)\,e^{ik\cdot\theta}$ over integer vectors $k$ — and $\varepsilon \ll 1$ flags the perturbation's size (a bookkeeping parameter: set it to 1 at the end and let $H_1$ 's smallness speak for itself). The mean over the torus, $\langle H_1\rangle(J) = H_{1,0}(J)$ , will play the starring role. The method seeks a near-identity canonical transformation via a type-2 generating function,

S(\theta, J') = \theta \cdot J' + \varepsilon S_1(\theta, J') + \varepsilon^2 S_2(\theta, J') + \cdots \qquad (2)

where: the leading term $\theta \cdot J'$ generates the identity transformation, so at $\varepsilon = 0$ nothing happens — the new variables deform continuously away from the old ones: $J = J' + \varepsilon\,\partial S_1/\partial\theta$ , $\ \theta' = \theta + \varepsilon\,\partial S_1/\partial J'$ . The strategy: choose $S_1, S_2, \ldots$ to scrub the angle dependence out of the Hamiltonian order by order. To first order the result (derived below) is

K(J') = H_0(J') + \varepsilon\,\langle H_1\rangle(J') + O(\varepsilon^2), \qquad \omega'(J') = \omega(J') + \varepsilon\,\frac{\partial \langle H_1\rangle}{\partial J'} \qquad (3)

where: the physics is in what survived: only the average of the perturbation moves the energy and the clock. The oscillating part of $H_1$ — everything with $k \ne 0$ — is absorbed into the coordinate deformation and contributes to observables only at $O(\varepsilon^2)$ . This is the rigorous backbone of every "averaging" argument in mechanics, and it is worth internalizing as an instinct: fast wiggles don't shift frequencies at first order; means do. The absorption has a price tag. The generating function's Fourier coefficients come out as

S_{1,k}(J') = \frac{i\,H_{1,k}(J')}{k \cdot \omega(J')} \qquad (k \ne 0) \qquad (4)

where: the denominators $k\cdot\omega$ are the small divisors. For a single degree of freedom $k\omega$ never vanishes ( $k \ne 0$ ) and the theory is clean. For $n \ge 2$ , the resonant condition $k\cdot\omega = 0$ defines surfaces that are dense in frequency space (rationally related frequencies), and even off-resonance the divisors dip arbitrarily close to zero for large $|k|$ — convergence of the whole series then hinges on how fast the numerators $H_{1,k}$ decay versus how viciously the divisors shrink: the exact battlefield of the KAM theorem.

Derivation AFirst-order canonical perturbation theory — killing the angle dependence

We are deriving the first-order results (3) and (4): the transformed Hamiltonian, the frequency shift, and the generating function that does the work — together with the solvability condition that explains why the average survives. The strategy: substitute the near-identity transformation into the Hamiltonian, expand to first order in $\varepsilon$ , and impose that the result depend on $J'$ alone; the imposition turns out to be a linear PDE for $S_1$ solvable mode by mode. Watch the $k = 0$ mode — it is the one mode the PDE cannot remove, and that obstruction, not a choice, is what puts $\langle H_1\rangle$ into the answer.

Step 1. Write the transformation generated by (2): $J = \partial S/\partial\theta = J' + \varepsilon\,\partial S_1/\partial\theta$ . Substitute into $H$ and expand in $\varepsilon$ : $H_0(J) = H_0(J') + \varepsilon\,\omega(J')\cdot\frac{\partial S_1}{\partial\theta} + O(\varepsilon^2)$ , using $\partial H_0/\partial J = \omega$ ; and $\varepsilon H_1(J, \theta) = \varepsilon H_1(J', \theta) + O(\varepsilon^2)$ . Why $H_1$ 's argument can be downgraded: it already carries an explicit $\varepsilon$ , so the $O(\varepsilon)$ difference between $J$ and $J'$ inside it is a second-order affair — the standard bookkeeping gift of perturbation theory.

Step 2. Collect the new Hamiltonian through first order: $K = H_0(J') + \varepsilon\left[\omega\cdot\dfrac{\partial S_1}{\partial\theta} + H_1(J', \theta)\right]$ . Demand the bracket be $\theta$ -independent — call the unknown constant (in $\theta$ ) $K_1(J')$ : $\ \omega\cdot\dfrac{\partial S_1}{\partial\theta} = K_1(J') - H_1(J', \theta)$ . Why this is the whole method in one line: the left side is a derivative along the unperturbed winding; we are asking the coordinate wiggle $S_1$ to cancel the perturbation's variation along each unperturbed trajectory.

Step 3. Solve by Fourier. Expand $S_1 = \sum_{k\ne0} S_{1,k} e^{ik\cdot\theta}$ (no $k=0$ term — a $\theta$ -independent piece of $S_1$ generates nothing) and $H_1 = \sum_k H_{1,k} e^{ik\cdot\theta}$ . Mode $k \ne 0$ : $\ i(k\cdot\omega)\,S_{1,k} = -H_{1,k}$ , giving (4). Mode $k = 0$ : the left side has no $k=0$ component at all, so the equation forces $K_1 = \langle H_1\rangle$ . Why the average cannot be transformed away: the operator $\omega\cdot\partial_\theta$ annihilates constants — its range contains only zero-mean functions. The mean of $H_1$ is orthogonal to everything a periodic coordinate change can absorb; it is the irreducible physical content of the perturbation at this order. Solvability conditions of this kind — "the resonant part must be kept in the dynamics" — recur throughout physics, from Lindstedt–Poincaré to quantum-mechanical degenerate perturbation theory.

Step 4. Read off the observables. New Hamiltonian: $K(J') = H_0 + \varepsilon\langle H_1\rangle$ ; new frequencies $\omega' = \partial K/\partial J'$ — equation (3). The new actions $J'$ are conserved to this order, and the old variables oscillate around them: $J = J' + \varepsilon\,\partial S_1/\partial\theta$ wiggles with zero mean and amplitude set by (4). Why no secular terms can appear: the time dependence of every quantity now enters only through $\theta' = \omega' t + \theta'_0$ — uniformly winding angles inside periodic functions. The frequency shift lives inside the angle, exactly where the naive expansion refused to put it. *Verified: one degree of freedom sanity check — for $H_1$ independent of $\theta$ the theory must be exact at first order in the shift: indeed $S_1 = 0$ and $K = H_0 + \varepsilon H_1(J')$ ✓. Dimensional check: $[S_1] = [\text{action}]$ since $[H_1/\omega] = [E\,T]$ ✓. The quartic-oscillator prediction of this machinery is verified against direct numerical integration to four digits in worked example 2. Treatment as in Goldstein Ch. 12, Lichtenberg & Lieberman Ch. 2.*

Go deeper · Derivation B · 2 minThe quartic oscillator and the pendulum — frequency renormalization in full

Derivation BThe quartic oscillator and the pendulum — frequency renormalization in full

We are cranking the machinery on the two standard test bodies: the quartic (Duffing-type) oscillator $H = \dfrac{p^2}{2m} + \dfrac{1}{2}m\omega_0^2 q^2 + \varepsilon q^4$ , and — via a sign flip — the pendulum at finite amplitude, obtaining the classic $\omega = \omega_0(1 - \varphi_m^2/16)$ . The strategy: express the perturbation in action-angle variables of the harmonic oscillator, average over the angle, differentiate with respect to the action. Watch how three lines of algebra replace the elliptic integrals of the exact treatment — and where the elliptic answer eventually parts company with the perturbative one (the computed figure below shows the divorce).

Step 1. Recall the harmonic oscillator's action-angle dictionary: $q = \sqrt{\dfrac{2J}{m\omega_0}}\,\sin\theta$ , $\ p = \sqrt{2m\omega_0 J}\,\cos\theta$ , $\ H_0 = \omega_0 J$ . Why this is the right dictionary: it is the worked centerpiece of the action-angle article — the check is $H_0 = \frac{p^2}{2m} + \frac12 m\omega_0^2q^2 = \omega_0 J(\cos^2\theta + \sin^2\theta) = \omega_0 J$ ✓.

Step 2. Express and average the perturbation: $H_1 = q^4 = \left(\dfrac{2J}{m\omega_0}\right)^2 \sin^4\theta$ , and $\langle\sin^4\theta\rangle = \tfrac38$ , so $\langle H_1\rangle = \dfrac{3J^2}{2m^2\omega_0^2}$ . Why $\langle\sin^4\rangle = 3/8$ : write $\sin^4 = \left(\frac{1 - \cos2\theta}{2}\right)^2$ and keep the constant terms: $\frac14 + \frac18$ .

Step 3. Apply (3): $K = \omega_0 J' + \varepsilon\,\dfrac{3J'^2}{2m^2\omega_0^2}$ , hence

\omega' = \omega_0 + \varepsilon\,\frac{3J'}{m^2\omega_0^2} = \omega_0\left(1 + \frac{3\varepsilon A^2}{2m\omega_0^2}\right) \qquad (5)

using $J' \approx \tfrac12 m\omega_0 A^2$ for amplitude $A$ . Why the frequency grows with amplitude: $\varepsilon > 0$ stiffens the potential at large $|q|$ — bigger swings sample stiffer walls and oscillate faster. This matches the Landau–Lifshitz secular-perturbation result $\Delta\omega = 3\alpha A^2/8\omega_0$ for $\ddot q + \omega_0^2 q + \alpha q^3 = 0$ upon the dictionary $\alpha = 4\varepsilon/m$ ✓.

Step 4. The pendulum is the same calculation with a soft spring. Expand $H = \dfrac{p^2}{2m\ell^2} - mg\ell\cos\varphi$ about the bottom: the quartic term of the potential is $-\dfrac{mg\ell}{24}\varphi^4$ — a negative $\varepsilon$ (the pendulum's restoring force weakens at large angle). Run (5) with effective mass $m\ell^2$ , $\omega_0^2 = g/\ell$ , $\varepsilon = -mg\ell/24$ , amplitude $\varphi_m$ : $\ \dfrac{\Delta\omega}{\omega_0} = \dfrac{3\varepsilon\varphi_m^2}{2(m\ell^2)\omega_0^2} = -\dfrac{3\,mg\ell\,\varphi_m^2}{24 \cdot 2\,m\ell^2 (g/\ell)} = -\dfrac{\varphi_m^2}{16}$ :

\omega = \omega_0\left(1 - \frac{\varphi_m^2}{16}\right), \qquad T = T_0\left(1 + \frac{\varphi_m^2}{16} + \cdots\right) \qquad (6)

Why this number was civilization-relevant: pendulum clocks keep time only if amplitude is controlled — equation (6) is the "circular error" that Huygens fought with cycloidal cheeks and clockmakers finally beat with small-amplitude escapements; worked example 1 converts it into minutes per day. Against the exact elliptic-integral frequency, (6) errs by only $0.002\%$ at $30°$ and $0.17\%$ at $90°$ — the computed figure shows the two curves and their parting. *Verified: the quartic prediction (5) at $m = \omega_0 = A = 1$ , $\varepsilon = 0.01$ gives $\omega' = 1.0150$ ; direct RK4 integration of the exact equation of motion (this article, $dt = 10^{-5}$ ) gives a period $T = 6.1911$ , i.e. $\omega_{\rm num} = 1.01486$ — agreement to $1.4\times10^{-4}$ , consistent with the neglected $O(\varepsilon^2)$ ✓. Pendulum check at $\varphi_m = 30°$ : exact $\pi/2K(\sin15°) = 0.982889$ vs (6) $= 0.982865$ ✓. Treatment: Goldstein §12.2–12.3, Landau & Lifshitz §28.*

Go deeper · Derivation C · 3 minSmall divisors — where the series breaks, and the road to KAM

Derivation CSmall divisors — where the series breaks, and the road to KAM

We are confronting the convergence question for $n \ge 2$ degrees of freedom: the Fourier solution (4) divides by $k\cdot\omega$ , and we derive what that does to the theory — the density of resonances, Poincaré's nonexistence verdict, and the two structural insights (Diophantine tori + superconvergent iteration) that turn catastrophe into the KAM theorem. The strategy is to follow the arithmetic of $k\cdot\omega$ honestly. Watch the competition of two exponentials-versus-power-laws: numerators $H_{1,k}$ decay geometrically in $|k|$ for analytic perturbations, but the divisors can be beaten down polynomially — the outcome is decided by which frequencies you stand on.

Step 1. Locate the fatalities. On a torus with $k\cdot\omega(J') = 0$ for some integer vector $k \ne 0$ , the coefficient $S_{1,k}$ in (4) is infinite unless $H_{1,k}$ happens to vanish: the transformation does not exist. Such resonant frequency vectors are dense: for $n = 2$ , resonance means $\omega_1/\omega_2$ rational, and rationals are dense in the reals. Why density is the disaster: no open set of tori is safe — arbitrarily close to any initial condition sits a torus on which the expansion is meaningless. A series "valid except on a dense set" is not an analytic function of $J'$ ; this is the technical face of Poincaré's theorem that no analytic constants of motion beyond $H$ survive a generic perturbation.

Step 2. Quantify the near-misses. Even off resonance, divisors dip: for typical irrational ratios, $\min_{|k|\le N}|k\cdot\omega|$ decays like a power of $1/N$ (the computed resonance-comb figure shows the landscape — a spike of safety at every strongly irrational ratio, chasms at every rational). Each order of the expansion multiplies more divisors together; the naive bound on the $r$ -th order term grows like $(r!)^{\text{power}}$ — the series is asymptotic, not convergent: for fixed truncation it is superbly accurate (error $O(\varepsilon^{r+1})$ ), but at fixed $\varepsilon$ it eventually diverges as $r \to \infty$ . Why practitioners use it anyway: an asymptotic series truncated at optimal order achieves accuracy exponentially small in $1/\varepsilon$ — Delaunay's lunar theory and every modern ephemeris are optimally-truncated divergent series, and Nekhoroshev's theorem turns this into rigorous stability over times $\exp(1/\varepsilon^{a})$ .

Step 3. Identify the survivors. Restrict to tori whose frequencies are Diophantine: $|k\cdot\omega| \ge \gamma/|k|^\tau$ for all $k \ne 0$ , with $\tau > n - 1$ . Then every divisor is controlled by a power of $|k|$ , while analytic $H_1$ supplies numerators $|H_{1,k}| \le C e^{-\sigma|k|}$ ; each order of the series is finite, with bounds worsening only polynomially. Such frequency vectors are not rare: their complement has measure $O(\gamma)$ — almost every torus is Diophantine. Why this alone is not enough: finite orders do not make a convergent series; the factorial growth from accumulating divisors remains. A second idea is needed.

Step 4. Change the iteration, not the estimates. KAM's decisive move replaces the order-by-order expansion (error $\varepsilon^{r}$ after $r$ steps, divisors accumulating linearly) with a Newton-like scheme: solve the first-order problem, then treat the whole remainder $O(\varepsilon^2)$ as the new perturbation and repeat. Errors square at each step — $\varepsilon \to \varepsilon^2 \to \varepsilon^4 \to \cdots$ — and this superconvergence overwhelms the polynomially-worsening divisor losses on Diophantine tori. The limit exists: for $\varepsilon$ small enough, each Diophantine torus survives, smoothly deformed, carrying its original frequencies. Why this hands you the KAM article's headline: the perturbation series of THIS article diverges on a dense set (resonances) yet its Newton-resummed cousin converges on a fat Cantor set (Diophantine tori) — chaos and order interleaved at all scales, both verdicts delivered by the same small divisors read with different arithmetic. The resonant gaps, where nothing converges, are where the Chirikov-overlap article picks up the story. *Verified: the resonance comb below is computed (minimum of $|k_1 x + k_2|$ over $|k_1| + |k_2| \le 8$ , 521 frequency ratios $x$ ): chasms at $x = 0, \frac12, \frac13, \frac23, \ldots$ widths ordered by denominator, local maxima at noble ratios — the golden mean $x = 0.618\ldots$ sits atop the tallest interior spike, consistent with its Diophantine-optimality role in the KAM article ✓. Measure bookkeeping: excluded resonant zones $\sum_k \gamma/|k|^{\tau+1} = O(\gamma) \to 0$ ✓. Treatments: Poincaré, Méthodes Nouvelles I (1892); Arnold, Math. Methods, App. 8; Lichtenberg & Lieberman Ch. 3.*

Figure

The small-divisor landscape, computed — the worst denominator available to the series, as a function of the frequency ratio

Which equation answers which question

you want	use	notes
the setup	$H = H_0(J) + \varepsilon H_1(J,\theta)$ , eq. (1)	Fourier-expand $H_1$ in the angles
first-order energy	$K = H_0 + \varepsilon\langle H_1\rangle$ , eq. (3)	only the torus-average survives
first-order frequency shift	$\Delta\omega = \varepsilon\,\partial_J\langle H_1\rangle$ , eq. (3)	the anti-secular renormalization
the coordinate deformation	$S_{1,k} = iH_{1,k}/(k\cdot\omega)$ , eq. (4)	zero-mean wiggle of $J$ at $O(\varepsilon)$
stiffened oscillator	$\omega' = \omega_0 + 3\varepsilon A^2/2m\omega_0$ , eq. (5)	Duffing at first order; verified numerically
pendulum amplitude error	$\omega = \omega_0(1 - \varphi_m^2/16)$ , eq. (6)	the clockmaker's equation
convergence verdict	Derivation C	divergent on dense resonances; asymptotic off them; KAM on Diophantine tori

Limiting cases

Case	Result	Meaning
$H_1 = H_1(J)$ only	$S_1 = 0$ ; theory exact at first order	nothing to average away
one degree of freedom	$k\omega \ne 0$ always	no small divisors; series safe (and often convergent)
$\varepsilon \to 0$	$J' \to J$ , $K \to H_0$	near-identity honestly reduces to identity
$\langle H_1\rangle = 0$	no first-order shift; effects start at $O(\varepsilon^2)$	e.g. perturbations odd in $q$
resonant torus, $k\cdot\omega = 0$	$S_{1,k} = \infty$ unless $H_{1,k} = 0$	secular growth of actions; islands/chaos form — Chirikov's territory
Diophantine torus, analytic $H_1$	every order finite; Newton scheme converges	the KAM survivors
degenerate $H_0$ (harmonic: $\omega$ independent of $J$ )	frequencies can't detune with $J$	resonances are $J$ -independent: standard KAM hypothesis fails (proper degenerate theory needed)
pendulum $\varphi_m \to 0$	$\omega \to \omega_0$	isochronism recovered; $(6)$ 's error $< 0.002\%$ at $30°$

Common trapThe perturbation series converges — just add more terms for more accuracy

For the generic multi-frequency system, both halves are false, and knowing why is a professional competence. The series is divergent: Poincaré proved that the expansion in $\varepsilon$ has zero radius of convergence for generic $H_1$ — resonant tori, where the divisors $k\cdot\omega$ vanish, are dense, so no $\varepsilon$ -ball avoids them. And "more terms" eventually makes things worse: as an asymptotic series, it improves until an optimal truncation order $r_*(\varepsilon)$ and then deteriorates factorially — Delaunay's seventh order was a judgment call, not laziness. What the series legitimately delivers is spectacular but bounded: optimally truncated, the error is exponentially small in $1/\varepsilon$ (Nekhoroshev's theorem converts this into rigorous action-stability over times $\sim e^{1/\varepsilon^a}$ — "for practical eternities, integrable"), and on Diophantine tori a different resummation (KAM's Newton iteration) genuinely converges. The instinct to resist: treating perturbative output as an approximation to persisting tori everywhere. Inside every resonance zone, the true dynamics contains islands and stochastic layers that no power series in $\varepsilon$ describes — the series isn't inaccurate there, it is wrong in kind, and the Chirikov-overlap article is the map of that territory. Use perturbation theory off resonance, respect its exponentially-good-but-finite warranty, and hand the resonant zones to the geometric methods built for them.

Worked examples

1 · The clockmaker's equation (confidence builder)

A grandfather clock's pendulum is regulated to 2.000 s period at its design amplitude $\varphi_m = 5°$ . A worn escapement lets the amplitude grow to $30°$ . How much time does the clock lose per day?

Step 1. The pendulum's period at amplitude, from (6): $T(\varphi_m) = T_{\rm iso}(1 + \varphi_m^2/16)$ , where $T_{\rm iso}$ is the isochronous (zero-amplitude) period. At $5° = 0.08727$ rad: correction $\varphi_m^2/16 = 4.76\times10^{-4}$ . At $30° = 0.5236$ rad: correction $= 0.5236^2/16 = 1.714\times10^{-2}$ .

Step 2. The clock was regulated at $5°$ — its 2.000 s already includes the small correction. The fractional slowdown at $30°$ relative to regulation: $\Delta T/T = 1.714\times10^{-2} - 4.76\times10^{-4} = 1.666\times10^{-2}$ .

Step 3. Per day: $86400 \times 1.666\times10^{-2} = \boxed{1440\ \text{s} \approx 24\ \text{minutes lost per day}}$ — a catastrophic error from a "small" nonlinearity, which is why every precision escapement in horological history is a device for keeping $\varphi_m$ tiny and constant.

Step 4. Check the tool against truth: at $30°$ the exact (elliptic-integral) frequency ratio is $0.982889$ ; equation (6) gives $0.982865$ — the first-order theory is off by $2.4\times10^{-5}$ , i.e. our 24-minute answer is good to about 2 seconds. At $90°$ the PT error grows to $0.17\%$ (see the computed figure): the expansion parameter is $\varphi_m^2/16$ , and the series knows its own limits.

2 · Duffing, predicted and then measured (exam level)

For $H = \frac{p^2}{2} + \frac{q^2}{2} + \varepsilon q^4$ with $\varepsilon = 0.01$ , launched at amplitude $A = 1$ : predict the oscillation frequency to first order, then verify by integrating the exact equations, and state what order the residual is.

Step 1. Prediction from (5) with $m = \omega_0 = 1$ : $\omega' = 1 + \tfrac32\varepsilon A^2 = 1 + 0.015 = \boxed{1.0150}$ .

Step 2. Numerical truth (RK4 on $\ddot q = -q - 4\varepsilon q^3$ , $dt = 10^{-5}$ , launched at $q = 1$ , $\dot q = 0$ ): measured period $T = 6.1911$ , so $\omega_{\rm num} = 2\pi/T = \boxed{1.01486}$ .

Step 3. Residual: $\omega' - \omega_{\rm num} = 1.4\times10^{-4} \approx \varepsilon^2$ — exactly the advertised $O(\varepsilon^2)$ leftovers. (Fine print worth owning: part of the residual is the dictionary $J \approx \frac12 m\omega_0 A^2$ , itself an $O(\varepsilon)$ -corrected relation between the launch amplitude and the true action; a second-order treatment fixes both consistently.)

Step 4. Extract the working habit: first-order canonical PT is a two-line calculation — average the perturbation over the unperturbed angle, differentiate by the action — and it lands within $\varepsilon^2$ of a five-decimal numerical integration. For the multidimensional problems where numerics can't sweep parameter space (satellite lifetimes, accelerator tunes), those two lines, iterated by computer algebra to modest order, are still the working method — with Derivation C's warrant and warning attached.

Figure

Pendulum frequency vs amplitude, computed — exact elliptic result against the first-order perturbative line

Origins · 2 minHenri Poincaré and how this was discovered

Henri Poincaré

1854–1912 · Paris

Poincaré inherited celestial perturbation theory as a craft — Delaunay's decades of hand computation, Le Verrier's Neptune — and left it a mathematical science with its death certificate and its resurrection plan filed in the same work. His prize memoir on the three-body problem (1889, for King Oscar II of Sweden) and the three volumes of Les Méthodes Nouvelles de la Mécanique Céleste (1892–99) proved the theorems this article stands on: that the classical series diverge in general, that no new analytic constants of motion survive a generic perturbation, and that the divergence's mechanism — the small divisors — encodes real dynamics, not clumsy technique. Famously, the prize memoir's first printed version contained an error whose correction cost Poincaré more than the prize money and gave mathematics the homoclinic tangle: chasing the consequences of a resonance, he found intersecting stable and unstable manifolds weaving "a kind of trellis... of infinitely tight mesh" he declined to even attempt drawing — the first glimpse of deterministic chaos, sixty years before the word. He worked the asymptotic-series doctrine into respectability (divergent, yet "convergent in the sense of the astronomers"), and the modern edifice — KAM on the surviving tori, Chirikov and Smale in the resonant wreckage — is built exactly along the fault lines he mapped.

History

1846

— Le Verrier turns perturbation series backwards: from Uranus's residuals to a predicted planet; Neptune is found within a degree — the method's greatest advertisement.

1860

— Delaunay publishes the first volume of his lunar theory: canonical perturbation theory carried to seventh order by twenty years of hand algebra.

1892

— Poincaré's Méthodes Nouvelles proves the series generically divergent — small divisors are real dynamics — and finds the homoclinic tangle hiding in resonances.

1916

— von Zeipel systematizes the generating-function elimination of angles, the algorithm modern textbooks (and this article) teach.

1954

— Kolmogorov announces the superconvergent rescue: Diophantine tori survive; Arnold and Moser complete the proofs (1962–63), turning this article's failure analysis into KAM.

1969

— Deprit's Lie-transform perturbation theory replaces von Zeipel's mixed-variable functions with explicit, machine-friendly recursions — the form used in software ever since.

Going further

Texts. Goldstein, Poole & Safko, Classical Mechanics, Ch. 12 — the canonical treatment this article follows · Lichtenberg & Lieberman, Regular and Chaotic Dynamics, Ch. 2–3 — the professional's version, with Lie transforms and the KAM connection · Landau & Lifshitz, Mechanics, §28 — anharmonic frequency shifts with minimal apparatus · Arnold, Mathematical Methods of Classical Mechanics, Ch. 10 & App. 8 — averaging and small divisors with full rigor · Poincaré, Les Méthodes Nouvelles de la Mécanique Céleste (1892–99) — where the subject learned its limits.

Provenance. Authored fresh for the mechanics track (Phase-C gap add-back; no AllQFT base). A verifier should recompute: the averaging identities $\langle\sin^4\rangle = 3/8$ and $\langle H_1\rangle = 3J^2/2m^2\omega_0^2$ ; the Duffing shift (5) against the numerical integration quoted (RK4, $dt = 10^{-5}$ : $T = 6.1911$ , $\omega = 1.01486$ vs predicted $1.0150$ , residual $1.4\times10^{-4} \sim \varepsilon^2$ ); the pendulum reduction to $-\varphi_m^2/16$ (with effective mass $m\ell^2$ and $\varepsilon = -mg\ell/24$ ) and its check against $\pi/2K(\sin\varphi_m/2)$ ( $30°$ : $0.982865$ vs $0.982889$ ; $90°$ : $0.84579$ vs $0.84721$ ); the clock arithmetic ( $\Delta T/T = 1.67\%$ → 24 min/day); the computed resonance comb (min $|k_1x + k_2|$ , $|k| \le 8$ ) and its golden-mean maximum.

Builds on: Action-Angle Variables · Frequencies from Action-Angle Variables · Hamilton-Jacobi Equation · Degeneracy and Commensurate Frequencies. Leads to: The KAM Theorem · Resonant Tori, Resonance Overlap, and the Chirikov Criterion · Adiabatic Invariants.

Why does only the average ⟨H₁⟩ affect the motion at first order? The oscillating part is just as big.

Because of who is asked to respond. The oscillating part of $H_1$ pushes the system at frequencies $k\cdot\omega \ne 0$ — off-resonance forcing, which produces a bounded, zero-mean wiggle of the actions (amplitude $\sim \varepsilon H_{1,k}/k\cdot\omega$ , the content of equation 4): the system jiggles but doesn't drift. The average part pushes at frequency zero — on resonance with conservation itself — and a zero-frequency push, however small, accumulates: it permanently re-prices the energy landscape $K(J')$ and re-tunes the frequencies. The formal statement is Step 3's solvability condition: the operator $\omega\cdot\partial_\theta$ , which the coordinate change can deploy, only produces zero-mean functions, so the mean of $H_1$ is precisely the part no coordinate change can touch. The same logic, exported: it is why gravity-wave forces on a satellite average out but oblateness's mean torque precesses the orbit; why in quantum mechanics the first-order energy shift is the diagonal matrix element (the "average" in a stationary state); and why the drift approximation in plasma physics keeps the gyro-averaged force. First-order physics is the physics of what survives averaging.

Doesn't the pendulum have only one degree of freedom — why bother with canonical PT when the exact elliptic solution exists?

Three reasons, in ascending importance. Pedagogically, the pendulum is the calibration standard: because the exact answer exists, you can measure the method's accuracy (the computed figure: 0.002% at 30°) instead of taking it on faith — always test new machinery where the truth is known. Practically, the two-line averaging calculation is simply cheaper: for the clock question, $\varphi_m^2/16$ answers in seconds what elliptic integrals answer in minutes, and the perturbative form exposes the scaling (error $\propto$ amplitude squared) that the exact $K(k)$ conceals inside a special function. Structurally — the deep reason — one degree of freedom is where the method is safe ( $k\omega$ never vanishes; no small divisors), so it isolates the renormalization idea from the convergence pathology. The multi-frequency problems where canonical PT is genuinely irreplaceable (satellite + oblateness + drag + Moon; three-body problem; accelerator lattices) have no exact solutions at all, and there the pendulum-honed machinery is the only analytic access — carrying the small-divisor warning label that the one-dimensional cases never needed.

What exactly goes wrong at a resonance, physically? The formula dividing by zero is math, not physics.

The physics is a push that never lets up. Off resonance, each Fourier mode of the perturbation pushes the trajectory with a phase $k\cdot\theta(t)$ that drifts through all values — the pushes alternate sign and cancel, leaving only the bounded wiggle of equation (4). On a resonant torus, $k\cdot\omega = 0$ means the combination $k\cdot\theta$ is frozen: that mode pushes with a constant phase, always in the same direction along the torus, and the actions respond secularly — they drift without bound in the linearized theory. What actually happens (the divergence is the linear theory confessing it can't say) is that the drift feeds back: the actions move, the frequencies detune, the push reverses, and the trajectory executes a slow, large oscillation around the resonance — the pendulum-like island structure your Chirikov article dissects (its "pendulum reduction" is exactly the honest nonlinear completion of this divergence). So the small divisor marks a real qualitative change: locally, the dynamics reorganizes from deformed winding into libration around a resonant orbit, with a separatrix and — when perturbations of neighbouring resonances interfere — a stochastic layer. The series doesn't fail to converge to the truth; the truth stops being the kind of object (a deformed torus with the original frequencies) that the series is built to describe.

Is ⟨H₁⟩ the same thing as what the adiabatic-invariants article computes, or the time average of H₁?

Related but distinct, and the distinctions carry content. Versus time averaging: on a nonresonant torus, the flow visits the torus uniformly (equidistribution), so the infinite-time average of $H_1$ along a trajectory equals the phase average $\langle H_1\rangle$ — the two agree exactly where the theory works. On a resonant torus the trajectory is confined to a lower-dimensional winding, time averages depend on initial phase, and the disagreement is yet another face of the small-divisor breakdown. Versus adiabatic theory: both are averaging methods, but they average over different smallness. Here the Hamiltonian is time-independent and slightly non-integrable ( $\varepsilon$ = coupling strength); there the Hamiltonian is integrable at each instant but slowly time-dependent ( $\varepsilon$ = slowness), and the averaging is over the fast angle at frozen parameters, yielding conservation of $J$ itself rather than a shift of $K(J)$ . The formalisms converge in the modern Lie-transform view — both are one-parameter families of near-identity canonical maps, and textbooks (Lichtenberg & Lieberman) develop them as one subject — which is also the version to learn if you go further: mixed-variable generating functions like (2) become genuinely painful at second order, while Lie series iterate mechanically.

How do I actually carry this to second order — and is it worth it?

The recipe extends mechanically, with one new structural feature. Write $S = \theta J' + \varepsilon S_1 + \varepsilon^2 S_2$ ; expanding $H$ now produces at order $\varepsilon^2$ the terms $\omega\cdot\partial_\theta S_2 + \left[\tfrac12\,\partial_J\omega : (\partial_\theta S_1)^2 + \partial_J H_1\cdot\partial_\theta S_1\right]$ , where the bracket is built entirely from first-order output. Demand $\theta$ -independence: $K_2$ = the angle-average of the bracket, and $S_2$ absorbs its oscillating part with the same $1/(k\cdot\omega)$ inversion. The new feature: the perturbation now beats against its own first-order wiggle — second order is where mode couplings $H_{1,k}H_{1,k'}$ with $k + k' = 0$ contribute, which is why $\langle H_1\rangle = 0$ perturbations (odd ones, typically) first shift frequencies at $O(\varepsilon^2)$ , and why the divisors now appear squared. Worth it? For precision celestial mechanics and accelerator design, unavoidable — but nobody does it with mixed-variable generating functions anymore: the inversion $\theta(\theta', J')$ needed to express results in new variables becomes an algebraic swamp. The professional tool is the Lie transform (Deprit/Hori): the transformation is generated as a flow, corrections come from nested Poisson brackets in a clean recursion, and computer algebra grinds to order ten. Same physics, same divisors, industrial packaging.

Why does the harmonic oscillator — of all things — cause trouble for this theory as H₀?

Because its frequency doesn't care about its action, and detuning-with-amplitude is the theory's hidden immune system. For a nondegenerate $H_0$ (the standard assumption: $\det(\partial^2 H_0/\partial J^2) \ne 0$ ), frequencies vary across the tori, so each resonance $k\cdot\omega(J) = 0$ is a thin surface in action space: move slightly in $J$ and you are off it — resonances are localized, and their nonlinear islands are bounded by the detuning. The isotropic harmonic oscillator is maximally degenerate: $H_0 = \omega\cdot J$ with constant $\omega$ , so $\partial^2 H_0/\partial J^2 \equiv 0$ — if the frequencies are resonant anywhere they are resonant everywhere, no torus can escape by detuning, and both this article's ordering and KAM's nondegeneracy hypothesis fail. The physics is real, not technical: perturbed identical oscillators (coupled pendulums at equal length, particle accelerators sitting on a rational tune) exhibit full-system resonant response, and the correct treatment reorganizes around the resonance from the start — secular perturbation theory in the resonant combination angle, degenerate normal forms, or the isotropic oscillator's own hidden-symmetry structure. The pattern is worth carrying to quantum mechanics, where it is the same dichotomy: nondegenerate perturbation theory divides by $E_n - E_m$ , and degenerate levels force you to diagonalize within the degenerate block before any division — the quantum small divisor, handled by the quantum version of "reorganize around the resonance first."

Canonical Perturbation Theory

Where this lives today

The physical picture

The math: the expansion, its first order, and its price

Which equation answers which question

Limiting cases

Worked examples

1 · The clockmaker's equation (confidence builder)

2 · Duffing, predicted and then measured (exam level)

Going further

Before this

Where this leads