Resonant Tori, Resonance Overlap, and the Chirikov Criterion

The KAM theorem is a treaty with a hole in it. It saves the strongly irrational tori and abandons every resonant torus — along with a finite-width zone around each — to an unspecified fate. This article goes into the abandoned zones and maps what actually happens there, because "unspecified fate" turns out to be the most structured demolition in physics: each broken resonant torus is replaced by a chain of islands, pearls of stability strung around the old torus, built around alternating stable and unstable periodic points whose existence is guaranteed by the Poincaré–Birkhoff theorem. Zoom into any single island and you find, universally, a pendulum: the same Hamiltonian that governs a swing governs every resonance in every nonlinear system, which is why one derivation serves all of them. The pendulum reduction delivers the number that controls everything downstream: the island's width in action grows as $\sqrt{\varepsilon}$ — the square root of the perturbation, enormously larger than the $O(\varepsilon)$ naive estimate. Resonances are not small features; they are amplified ones. And that amplification sets up the decisive confrontation. Islands sit at every rational winding number, each with its $\sqrt{\varepsilon}$ girth, separated by the surviving KAM tori that wall chaos into thin layers along each separatrix. Fatten the perturbation and the chains grow toward each other; the walls between them thin, then break. Boris Chirikov's 1959 criterion says the transition to global chaos happens, roughly, when neighbouring islands touch: half-width plus half-width equals spacing. It is a heuristic — it ignores the secondary structure that accelerates the collapse, and overestimates the true threshold by a factor of about $2.5$ in the standard map — but it is the heuristic that organizes real machines: accelerator apertures, tokamak field errors, and asteroid-belt gaps are all engineered or explained by counting island widths against island spacings. This article derives the chain, the pendulum, the width, and the criterion, and computes the pictures of islands growing until the sea connects.

Key takeaways

KAM abandons each resonant torus $\mathbf{k}\cdot\boldsymbol{\omega} = 0$ ; the Poincaré–Birkhoff theorem says what replaces it: $2kq$ alternating elliptic/hyperbolic periodic points — an island chain around the old torus.
Near any single resonance, averaging over fast angles reduces the Hamiltonian universally to a pendulum $H_{\text{res}} = \tfrac12 G\,\Delta J^2 + \varepsilon F \cos\psi$ in the slow resonance phase $\psi$ .
The island half-width in action is $\Delta J_{\max} = 2\sqrt{\varepsilon |F|/|G|}$ — an $O(\sqrt{\varepsilon})$ structure from an $O(\varepsilon)$ perturbation; the libration frequency is $\omega_{\text{lib}} = \sqrt{\varepsilon |FG|}$ .
Each island's separatrix is dressed with a thin chaotic layer (the homoclinic tangle of its hyperbolic points); while KAM tori survive between chains, that chaos stays locally confined.
Chirikov's criterion: global chaos begins when neighbouring resonances overlap — sum of half-widths $\ge$ spacing. For the standard map it predicts $K \approx \pi^2/4 \approx 2.47$ ; the true golden-torus breakup is $K_c = 0.9716$ , so overlap with a rule-of-thumb factor $\sim 2/3$ on widths is the practical estimate.
The criterion overestimates because it ignores secondary resonances and layer growth — but its scaling is right, and it is the working design rule in accelerators, plasmas, and celestial mechanics.
Key idea: every broken resonance is a pendulum of width $\propto \sqrt{\varepsilon}$ ; chaos goes global when neighbouring pendulums touch — half-width plus half-width exceeding the spacing between resonance centres.

The core insight

Each resonant torus breaks into a pendulum-like island chain of action half-width $2\sqrt{\varepsilon|F|/|G|}$ ; isolated chains mean local chaos only, and the transition to global chaos arrives when neighbouring chains grow wide enough to overlap — Chirikov's criterion.

Where this lives today

Chirikov wrote the criterion to keep plasma confined, and that is still its day job: tokamak and stellarator designers compute magnetic-island widths from field-error harmonics ( $\Delta \propto \sqrt{b_{mn}}$ , the same square root) against the spacing of rational flux surfaces, and specify coil tolerances so the islands stay below overlap — the "stochastic field-line" regions that appear when they fail are direct heat leaks. Accelerator physics runs the identical calculation in tune space: nonlinear magnet errors make each resonance line $\mathbf{k}\cdot\boldsymbol{\nu} \in \mathbb{Z}$ a channel of finite width, and the working point plus beam tune-spread must fit in the gaps — dynamic-aperture engineering is applied resonance packing. In the solar system, overlap of mean-motion resonances explains the Kirkwood gaps and — via Wisdom's work on the 3:1 — delivered meteorites their transport route to Earth; overlap of first-order resonances sets the general chaotic-zone width around any planet, $\Delta a/a \simeq 1.3\,(m_p/M_\odot)^{2/7}$ , the standard formula for exoplanet-system stability screening. Chemists use the same threshold for intramolecular energy flow between vibrational modes (IVR): below overlap a molecule's energy stays mode-locked for long times, above it the energy thermalizes — reaction-rate theories switch regimes exactly at the Chirikov boundary. Wherever someone asks "when do nonlinear resonances stop being decoration and start being transport," this criterion is still the first — and usually the deciding — estimate. Within the chapter, the two bookends are adjacent articles: the standard map (next) is the laboratory where every number here — $4\sqrt K$ , $2\pi$ , $0.9716$ , $K^2/4$ — can be checked by iteration, and the homoclinic-tangle machinery that makes the "thin chaotic layer" rigorous is the horseshoe article that closes the block.

The physical picture

Start on a Poincaré section of a near-integrable system and walk outward through the nested invariant circles. Most circles are intact KAM curves. But at every radius where the winding number passes through a rational $p/q$ , the smooth circle is gone; in its place sits a necklace of $q$ islands. At the centre of each island bead is a stable (elliptic) periodic point — the orbit that repeats after exactly $q$ laps — and between consecutive beads sits an unstable (hyperbolic) point, a saddle whose invariant manifolds weave the necklace together as a chain of separatrix loops. This is not an occasional decoration: it happens at every rational, so the section is an infinite hierarchy of necklaces between surviving curves, with smaller necklaces between those, all the way down.

Figure

Anatomy of an island chain, computed — elliptic centre, hyperbolic points, librating interior, chaotic separatrix layer

Why islands, and why a pendulum inside each? Sit in a rotating frame that keeps pace with the resonance — measure only the slow phase $\psi$ , the combination of angles that would be frozen exactly on resonance. Relative to this frame the perturbation is no longer an oscillating nuisance; its resonant Fourier component becomes a static potential $\varepsilon F\cos\psi$ , a fixed corrugation in which the slow phase can be trapped. The action mismatch $\Delta J$ from exact resonance supplies inertia through the twist ( $G = \partial\omega/\partial J$ converts action offset into phase drift). Trapped oscillation of $\psi$ in the corrugation is libration — the interior of the island; free running of $\psi$ over the corrugation crests is circulation — the KAM-like curves outside. The dividing orbit is a separatrix, and the whole structure is precisely a pendulum: trapped swings, over-the-top rotation, and the infinite-period boundary between them. The pendulum's separatrix is where the trouble concentrates. An exactly integrable pendulum has a clean separatrix, but the resonance pendulum is still weakly shaken by all the other, off-resonant Fourier modes we averaged away. A shaken separatrix does not survive as a curve — its stable and unstable manifolds split and cross transversally, weaving the homoclinic tangle (the horseshoe article makes this rigorous), so each island chain is wrapped in a thin stochastic layer of genuine chaos. At small $\varepsilon$ the layers are exponentially thin and quarantined between KAM curves: chaos exists but cannot travel. The endgame is territorial: island widths grow like $\sqrt{\varepsilon}$ , spacings stay fixed, and when two chains' layers meet, the quarantine fails — an orbit can hop from one resonance to the next, then the next, and diffuse across all of action space. Chirikov's criterion is simply the geometry of that first contact.

The mathematics of a broken resonance

The stage is the perturbed Hamiltonian in action-angle form, with the perturbation resolved into the Fourier modes that drive individual resonances — each mode $\mathbf{k}$ speaks to the tori where $\mathbf{k}\cdot\boldsymbol{\omega}$ is small, as established in the two previous articles:

H(\mathbf{J}, \boldsymbol{\theta}) = H_0(\mathbf{J}) + \varepsilon \sum_{\mathbf{k}\neq 0} H_{\mathbf{k}}(\mathbf{J})\, e^{i\mathbf{k}\cdot\boldsymbol{\theta}}, \qquad \omega_i = \frac{\partial H_0}{\partial J_i} \qquad (1)

where: $H_{\mathbf{k}}$ — Fourier amplitude of the resonance driven by integer vector $\mathbf{k}$ (for smooth perturbations these decay rapidly with $|\mathbf{k}|$ , which is why low-order resonances dominate); $\boldsymbol{\omega}$ — the frequency map of the integrable part. The small-denominator analysis of the integrable-baseline article showed perturbation theory failing near each surface $\mathbf{k}\cdot\boldsymbol{\omega}(\mathbf{J}) = 0$ ; everything in this article happens inside those failure zones. What replaces a resonant torus is fixed by topology before any approximation is made. On the Poincaré section the unperturbed dynamics near winding number $p/q$ is a twist map — circles rotating at radius-dependent rates — and Poincaré's "last geometric theorem", proved by Birkhoff, counts the survivors of the resonant circle's destruction:

Derivation APoincaré–Birkhoff — what a resonant torus leaves behind

We establish the skeleton of every island chain: when a twist map's rational circle is destroyed by a perturbation, an even number $2kq$ of period- $q$ points survive, alternating elliptic and hyperbolic around the ring. The route is a fixed-point argument on the $q$ -th iterate of the map, using the twist to trap zeros of the radial displacement and area preservation to force their signs. The count and the alternation are exactly what the computed sections show: $q$ island centres, $q$ saddles stitched between them.

Step 1. Set up the unperturbed twist map on an annulus of the section, $T_0: (J, \phi) \mapsto (J, \phi + 2\pi\Omega(J))$ with $\Omega'(J) \neq 0$ (the twist). On the circle $J = J_r$ where $\Omega(J_r) = p/q$ , every point is a fixed point of $T_0^q$ — a full circle of degenerate fixed points, an accident of exact integrability that cannot survive.

Step 2. Perturb: $T_\varepsilon = T_0 + O(\varepsilon)$ , area-preserving. Consider the $q$ -th iterate $T_\varepsilon^q$ and ask where a point returns to its own angular position. By the twist, just inside $J_r$ the iterate over-rotates and just outside it under-rotates (or vice versa); by continuity, on each radial line there is a point with zero net rotation. These points form a closed curve $\mathcal{C}$ near the old circle. Why the twist is indispensable: without $\Omega'(J) \neq 0$ there is no over/under-rotation sandwich, no trapped zero, and the argument — like the KAM proof before it — collapses.

Step 3. On $\mathcal{C}$ , the map $T_\varepsilon^q$ moves points purely radially. If it moved every point outward, the region enclosed by $\mathcal{C}$ would map strictly inside a larger region — impossible for an area-preserving map ( $\mathcal{C}$ and its image enclose equal areas). So the radial displacement changes sign an even number of times: generically $2kq$ zeros ( $k \ge 1$ ), which are the fixed points of $T_\varepsilon^q$ , i.e. period- $q$ orbits of $T_\varepsilon$ in groups of $q$ .

Step 4. Classify by index: walking along $\mathcal{C}$ , the displacement field alternates between pointing out and pointing in, so consecutive fixed points carry opposite Poincaré index — alternately elliptic (index $+1$ , centres of islands) and hyperbolic (index $-1$ , saddles). The elliptic points seed stability (small KAM structure around each); the hyperbolic points' manifolds form the chain's separatrix loops and, under further perturbation, its chaotic layer. Verified: statement and proof strategy per Lichtenberg & Lieberman §3.2b and Arnold–Avez; the fixed-point count $2kq$ with elliptic/hyperbolic alternation is the classical Poincaré–Birkhoff conclusion.

Topology fixed the skeleton; dynamics dresses it. To see the flesh — how wide the island is, how fast it swings — reduce the full Hamiltonian near one resonance to its universal normal form:

Go deeper · Derivation B · 2 minReduction of a single resonance to a pendulum

Derivation BReduction of a single resonance to a pendulum

We derive the pendulum that lives inside every island: near a resonance of a two-degree-of-freedom system, a resonant canonical transformation plus averaging turns $H$ into $\tfrac12 G\,\Delta J^2 + \varepsilon F\cos\psi$ , from which the island's width and internal frequency read off by pendulum energetics. The three moves — pass to slow-fast variables, Taylor-expand the integrable part, average the perturbation over the fast angle — are the standard secular machinery of celestial mechanics; the thing to watch is where the twist coefficient $G$ enters, because it converts perturbation strength into geometric width.

Step 1. Pick the resonance $k_1\omega_1 + k_2\omega_2 = 0$ at $\mathbf{J} = \mathbf{J}_0$ . Change to the slow phase and its conjugate action by the generating function $F_2 = (k_1\theta_1 + k_2\theta_2)P + \theta_2 Q$ :

\psi = k_1\theta_1 + k_2\theta_2, \qquad \phi = \theta_2, \qquad J_1 = k_1 P, \qquad J_2 = k_2 P + Q

On resonance $\dot\psi = k_1\omega_1 + k_2\omega_2 = 0$ : $\psi$ is the stroboscopically frozen combination — the swing angle of the coming pendulum — while $\phi$ still rotates fast.

Step 2. Expand $H_0$ about $\mathbf{J}_0$ to second order in the action offset $\Delta J$ conjugate to $\psi$ :

H_0 \approx H_0(\mathbf{J}_0) + \underbrace{(k_1\omega_1 + k_2\omega_2)}_{=\,0}\Delta J + \tfrac12 G\, \Delta J^2, \qquad G = \sum_{ij} k_i k_j \frac{\partial^2 H_0}{\partial J_i \partial J_j}

Why the second-order term is the leading one: the linear term vanishes BY the resonance condition — this is why the local dynamics is soft and easily reshaped by even a tiny perturbation. $G$ is the twist projected onto the resonance direction; it plays the role of inverse mass.

Step 3. Average over the fast angle $\phi$ . Every Fourier mode of the perturbation that depends on $\phi$ oscillates rapidly and contributes only at higher order (the averaging theorem); the sole survivor is the resonant harmonic, $\varepsilon F \cos\psi$ with $F = 2|H_{\mathbf{k}}(\mathbf{J}_0)|$ and a phase choice absorbing its argument:

\bar{H} = \tfrac12 G\, \Delta J^2 + \varepsilon F \cos\psi \qquad (2)

where: $\Delta J$ — action distance from exact resonance; $\psi$ — slow resonance phase; $G$ — twist coefficient (inverse inertia); $\varepsilon F$ — depth of the resonant corrugation. This is the pendulum Hamiltonian, exactly as promised — mass $1/G$ , gravity torque $\varepsilon F$ .

Step 4. Read the equations of motion and confirm: $\dot\psi = G\,\Delta J$ , $\dot{\Delta J} = \varepsilon F \sin\psi$ , hence $\ddot\psi = \varepsilon F G \sin\psi$ — libration inside the island at small amplitude with frequency

\omega_{\text{lib}} = \sqrt{\varepsilon |F G|}

Step 5. Width from separatrix energetics. The separatrix passes over the corrugation crest with zero slow speed: $\bar{H}_{\text{sep}} = \varepsilon F$ (taking $G, F > 0$ ). At the trough phase $\psi = \pi$ the action offset is maximal: $\tfrac12 G \Delta J_{\max}^2 - \varepsilon F = \varepsilon F$ , so

\Delta J_{\max} = 2\sqrt{\frac{\varepsilon |F|}{|G|}} \qquad (3)

where: $\Delta J_{\max}$ — the island half-width in action. The square root is the headline: halving your perturbation shrinks the resonance zones only by $1/\sqrt2$ . A $10^{-4}$ perturbation makes $10^{-2}$ -scale islands. Verified: reduction and width follow Lichtenberg & Lieberman §2.4 and Chirikov (1979) §2; the pendulum energetics $\Delta J_{\max} = 2\sqrt{\varepsilon F/G}$ reproduces Landau–Lifshitz §49-adjacent secular theory. Check of limits: $F \to 0$ (no resonant harmonic) or $G \to 0$ (no twist) both collapse the island, as they must.

With one resonance fully understood, the global question is a packing problem: how do the chains fit side by side, and when do they collide? Chirikov's criterion is the geometry of first contact between neighbouring pendulums:

\frac{\tfrac12\Delta J^{(1)} + \tfrac12\Delta J^{(2)}}{|J^{(1)}_{\text{res}} - J^{(2)}_{\text{res}}|} \;\ge\; 1 \quad \Longrightarrow \quad \text{global chaos across the pair} \qquad (4)

where: $\Delta J^{(i)}$ — the full widths $= 4\sqrt{\varepsilon|F_i|/|G_i|}$ of the two chains, so the numerator is the sum of half-widths; the denominator is the action distance between resonance centres. Below threshold, a surviving KAM torus between the chains guarantees (for two degrees of freedom) that no orbit crosses; at threshold the last such torus dies and the two stochastic layers fuse into one connected sea. The criterion's genius is its bluntness — every hard effect (layer thickness, secondary chains, higher harmonics) is swept into " $\approx$ ", and its predictions land within factors of 2–3 universally.

Go deeper · Derivation C · 2 minTwo resonances of the standard map and the overlap estimate

Derivation CTwo resonances of the standard map and the overlap estimate

We apply the full machinery to the cleanest testbed — the kicked rotor, whose stroboscopic dynamics is the standard map — and derive Chirikov's classic estimate for its global-chaos threshold. The kicked rotor's periodic impulses supply an infinite ladder of resonances at every integer momentum multiple of $2\pi$ ; each is a pendulum by Derivation B, all with the SAME width; overlap of adjacent rungs then gives $K \gtrsim \pi^2/4 \approx 2.47$ . Comparing with the true $K_c = 0.9716$ calibrates how conservative the criterion is — and why.

Step 1. The kicked rotor, $H = \tfrac12 p^2 + K\cos\theta \sum_n \delta(t - n)$ , has kicks whose Fourier expansion in time is a comb of harmonics:

K\cos\theta \sum_n \delta(t-n) = K\cos\theta \sum_{m=-\infty}^{\infty} e^{i 2\pi m t} = K \sum_m \cos(\theta - 2\pi m t)

Each term $\cos(\theta - 2\pi m t)$ is a travelling wave — one resonance, resonant where the rotor keeps pace with it.

Step 2. Resonance centres: $\dot\theta = p$ , so the $m$ -th wave is stationary in the rotor's frame when $p = 2\pi m$ . The ladder of resonances sits at integer multiples of $2\pi$ with uniform spacing $\delta p = 2\pi$ — the denominators of criterion (4).

Step 3. Width of each rung: near $p = 2\pi m$ , keeping only the resonant wave gives the pendulum $\bar H = \tfrac12 (\Delta p)^2 + K\cos\psi$ — here $G = 1$ (the rotor's bare twist) and $\varepsilon F = K$ . By equation (3), half-width $\Delta p_{\max} = 2\sqrt{K}$ , full width $4\sqrt{K}$ , identical for every rung.

Step 4. Overlap: half-width plus half-width against spacing,

\frac{2\sqrt{K} + 2\sqrt{K}}{2\pi} \ge 1 \quad\Longrightarrow\quad K \ge \frac{\pi^2}{4} \approx 2.47

Step 5. Compare with truth. The numerically exact threshold — the death of the last (golden) rotational KAM curve, Greene's computation — is $K_c = 0.971635$ : the plain criterion overshoots by a factor $\approx 2.5$ . The physics of the overshoot is visible in the computed portraits: as $K$ grows, secondary island chains (notably the half-integer resonance from second-order effects at $p = \pi$ ) bloom between the primary rungs, and each separatrix's chaotic layer fattens — the gap is invaded from within and from both sides well before the primaries themselves touch. Chirikov's practical patch: apply the criterion with widths inflated by including the layer and the strongest secondary — the celebrated "two-thirds rule", $\tfrac{2}{3} \times$ spacing instead of full spacing — which lands at $K \approx 1.1$ , adjacent to the truth. Verified: Fourier comb, ladder spacing $2\pi$ , width $4\sqrt K$ , and $K_{\text{overlap}} = \pi^2/4$ follow Chirikov (1979) and Lichtenberg & Lieberman ch. 4; $K_c = 0.971635$ per Greene (1979). Numbers recomputed: $4\sqrt{2.47} = 6.28 = 2\pi$ ✓.

Figure

Resonance overlap in action, computed — island chains growing with K until their chaotic layers merge into a connected sea

Limiting cases

Case	Result	Meaning
$\varepsilon \to 0$	widths $\propto \sqrt{\varepsilon} \to 0$ , chains isolated	KAM regime: chaos quarantined in exponentially thin layers
exactly on resonance, $\varepsilon$ small	pendulum libration at $\omega_{\text{lib}} = \sqrt{\varepsilon FG}$	motion trapped in the island, stable and regular
overlap ratio $= 1$	last dividing torus dies	onset of global transport across the two chains
$K \gg K_c$ (standard map)	diffusion $D \to K^2/4$	kicks decorrelate: momentum random-walks, quasilinear scaling
high-order resonance $\lvert\mathbf{k}\rvert$ large	$F \sim e^{-\rho\lvert\mathbf{k}\rvert}$ : negligible width	analytic perturbations make high-order chains invisible
twist $G \to 0$	width formula diverges — reduction fails	degenerate resonances need higher-order normal forms
$n \ge 3$ degrees of freedom	chains form a connected web even below overlap	Arnold diffusion along resonances; overlap still governs FAST transport

Common trapThe criterion is a scaling law wearing the costume of an equation

Chirikov's criterion looks like a sharp inequality; treating it as one produces factor-of-2.5 embarrassments. It systematically overestimates the global-chaos threshold because three destroyers act before the primary islands physically touch: each separatrix already carries a chaotic layer whose width grows with $\varepsilon$ ; secondary resonances (island chains between island chains, from second-order perturbation terms) erect stepping stones across the gap; and the dividing KAM torus does not need to be squeezed to zero width — it dies of its own Diophantine marginality first, per the KAM article. The honest usage: compute widths and spacings, form the overlap ratio $s$ , and read $s \ll 2/3$ as safely ordered, $s \gtrsim 2/3$ as endangered, $s \ge 1$ as certainly chaotic — the "two-thirds rule". Also mind the opposite error: below overlap is not "no chaos" but "no global chaos" — the layers are chaotic at any $\varepsilon > 0$ , and in three or more degrees of freedom Arnold diffusion creeps along the resonance web even when every pairwise ratio is far below threshold. The criterion answers one question only: when does transport across resonances become fast.

Worked examples

1 · Overlap arithmetic on the standard map (confidence builder)

For the standard map, evaluate the overlap ratio $s(K) = 4\sqrt{K}/2\pi$ (sum of half-widths over spacing) at $K = 0.5$ , at Greene's $K_c = 0.9716$ , and find the $K$ the bare criterion certifies. Interpret.

Step 1. At $K = 0.5$ : half-width $2\sqrt{0.5} = 1.414$ ; sum of half-widths $= 2.83$ ; spacing $2\pi = 6.283$ . Ratio $s = 2.83/6.283 = \mathbf{0.45}$ . Well below unity — and indeed the computed portrait at $K = 0.5$ shows fat KAM bands between the integer chains.

Step 2. At $K_c = 0.9716$ : $s = 4\sqrt{0.9716}/2\pi = 3.943/6.283 = \mathbf{0.63}$ . Global chaos actually begins when the primaries fill only $63\%$ of the gap — the missing $37\%$ is eaten by the $p = \pi$ secondary chain and the two stochastic layers. This measured $0.63$ is the two-thirds rule.

Step 3. Bare criterion: $s = 1$ requires $4\sqrt{K} = 2\pi$ , i.e. $K = \pi^2/4 = \mathbf{2.47}$ — overestimating $K_c$ by the factor $2.47/0.9716 = 2.5$ . Moral: trust the $\sqrt{K}$ scaling and the geometry; apply the $2/3$ calibration for numbers.

2 · Designing a rotor to stay ordered (exam level)

A rotor with $H_0 = J^2/2$ (so $G = 1$ ) is driven by two waves: $\varepsilon F_1 \cos(\theta - t)$ and $\varepsilon F_2 \cos(\theta - 2t)$ , with $F_1 = 0.8$ , $F_2 = 0.45$ . (i) Locate the resonances. (ii) Find the overlap ratio at $\varepsilon = 0.01$ . (iii) What maximum $\varepsilon$ does the two-thirds rule certify as globally ordered?

Step 1. Resonances where the rotor surfs each wave: $\dot\theta = J$ equal to the wave speed. Wave 1: $J_1 = 1$ ; wave 2: $J_2 = 2$ . Spacing $\delta J = 1$ .

Step 2. Half-widths at $\varepsilon = 0.01$ from equation (3): $\Delta J^{(1)}_{\max} = 2\sqrt{0.01 \times 0.8/1} = 2\sqrt{0.008} = \mathbf{0.179}$ ; $\Delta J^{(2)}_{\max} = 2\sqrt{0.01\times0.45} = 2\sqrt{0.0045} = \mathbf{0.134}$ . Overlap ratio $s = (0.179 + 0.134)/1 = \mathbf{0.31}$ — comfortably ordered; each wave's island chain is quarantined by KAM curves near $J = 1.5$ .

Step 3. Threshold by the two-thirds rule: need $s(\varepsilon) = 2\sqrt{0.8\varepsilon} + 2\sqrt{0.45\varepsilon} = (2\times0.894 + 2\times0.671)\sqrt{\varepsilon} = 3.13\sqrt{\varepsilon} \le 2/3$ . So $\sqrt{\varepsilon} \le 0.213$ , $\varepsilon \le \mathbf{0.045}$ . (The bare criterion would allow $\varepsilon = 0.102$ — the difference between an aggressive estimate and a safe design margin.) Note the design lever: thresholds scale as the square of allowed drive, so a $2\times$ stronger wave pair demands $4\times$ wider resonance spacing.

3 · The 3:1 Kirkwood gap, order-of-magnitude (applied)

An asteroid near the 3:1 resonance with Jupiter has orbital frequency $\omega_a = 3\omega_J$ , with $\omega_J = 2\pi/11.86$ yr $^{-1}$ . Jupiter's perturbation strength is $\varepsilon \sim M_J/M_\odot \approx 10^{-3}$ . Estimate the resonance's fractional width in semi-major axis and compare with the observed gap.

Step 1. For Kepler motion $H_0 \propto -1/(2J^2)$ (per the action-angle article), frequencies scale as $\omega \propto J^{-3}$ , so fractional widths translate: $\Delta a/a = \tfrac{2}{3}\,\Delta \omega/\omega \sim \Delta J/J$ up to $O(1)$ factors. The pendulum width gives $\Delta J/J \sim 2\sqrt{\varepsilon f}$ with $f$ an $O(0.1\text{–}1)$ resonance-strength factor (eccentricity-dependent for 3:1).

Step 2. Numbers: $2\sqrt{10^{-3} \times 0.3} \approx 2 \times 0.017 = \mathbf{0.035}$ — a few-percent action width, i.e. $\Delta a/a \sim 2\text{–}3\%$ at moderate eccentricity. At the 3:1 location $a = 2.50$ AU that is $\Delta a \sim \mathbf{0.06\text{–}0.08}$ AU.

Step 3. The observed 3:1 Kirkwood gap is $\sim 0.1$ AU wide — the estimate lands within a factor of order one. The gap exists because inside the resonance's chaotic layer eccentricities random-walk upward until the asteroid crosses Mars's or Earth's orbit and is removed: resonance-overlap chaos, acting as a $10^8$ -year vacuum cleaner. (The full story — overlapping secular resonances inside the mean-motion island — is exactly a nested application of this article's machinery.)

Origins · 1 minGeorge David Birkhoff and how this was discovered

George David Birkhoff

1884–1944 · Harvard

Birkhoff became, overnight, the most famous mathematician in America by proving Poincaré's "last geometric theorem" — the fixed-point count of Derivation A — in 1913, a year after Poincaré had published it unproven, with an apology, months before his death. Poincaré knew exactly what hung on it: the theorem is the guarantee that a destroyed resonant torus leaves behind alternating stable and unstable periodic orbits, the skeleton of every island chain. Birkhoff went on to build much of modern dynamics around such structures: his ergodic theorem (1931) founded the statistical description of motion; his normal forms organize behaviour near elliptic points; his "remarkable curves" anticipated the cantori that throttle transport after KAM curves die. He was also the first to see clearly that near a homoclinic tangle the periodic orbits are infinite in number and intricately interleaved — the bridge from this article's island chains to the horseshoe that closes the chapter. Working in an era when American mathematics looked to Europe, Birkhoff reversed the traffic; Wiener recalled that European visitors asked first for Birkhoff. His son Garrett became a distinguished algebraist — the Birkhoffs are one of mathematics' rare father-son dynasties.

History

1912

— Poincaré publishes his "last geometric theorem" without proof: an area-preserving annulus twist map has at least two fixed points.

1913

— Birkhoff proves it, securing the elliptic/hyperbolic skeleton of every broken resonance.

1959

— Chirikov, analyzing plasma confinement in mirror machines, formulates the resonance-overlap criterion — the first quantitative prediction of a chaos threshold.

1968

— Rosenbluth, Sagdeev, Taylor and Zaslavsky compute the exponentially thin stochastic layer of a modulated pendulum's separatrix.

1979

— Chirikov's Physics Reports survey codifies the standard map as the universal testbed; Greene pins the exact threshold

K_c = 0.9716

via the golden curve.

Going further

References. Goldstein, Poole & Safko, Classical Mechanics, 3rd ed., §12 (canonical perturbation theory, action-angle, KAM) · Landau & Lifshitz, Mechanics (Vol 1), §49–52 (adiabatic invariants, action variables) · Morin, Introduction to Classical Mechanics, ch. on oscillations and the pendulum · Taylor, Classical Mechanics, ch. 12 (nonlinear mechanics and chaos) · Chirikov, 'A universal instability of many-dimensional oscillator systems', Physics Reports 52, 263 (1979) · Lichtenberg & Lieberman, Regular and Chaotic Dynamics, 2nd ed., ch. 2–4 (definitive treatment of resonance overlap and the standard map) · Greene, 'A method for determining a stochastic transition', J. Math. Phys. 20, 1183 (1979) — K_c = 0.971635

Provenance. Pendulum reduction and island widths follow Landau-Lifshitz Mechanics §49 and the canonical perturbation theory of Goldstein §12; small-denominator/KAM discussion per Goldstein and Arnold. Standard-map overlap estimate $K_{\text{overlap}}=\pi^2/4\approx2.47$ and exact threshold $K_c=0.971635$ are standard (Chirikov 1979; Greene 1979; Lichtenberg & Lieberman ch.4). A verifier should: (1) confirm $\Delta J_{\max}=2\sqrt{\epsilon F/G}$ from pendulum energy conservation; (2) recompute standard-map island full width $=4\sqrt K$ , spacing $2\pi$ , hence $K_{\text{overlap}}=\pi^2/4$ ; (3) verify $K_c\approx0.9716$ against Greene's renormalization / Mackay's value. No fabricated numbers used. Additionally for this upgrade: Poincaré–Birkhoff count $2kq$ with alternating indices per Lichtenberg & Lieberman §3.2b; worked-example numbers recomputed directly (overlap ratios $0.45$ and $0.63$ ; rotor thresholds $\varepsilon \le 0.045$ two-thirds vs $0.102$ bare; Kirkwood estimate $\Delta a/a \sim 2\sqrt{\varepsilon f} \approx 3.5\%$ at $f=0.3$ ); the chaotic-zone exponent $2/7$ (Wisdom 1980) quoted for context. Figures are computed standard-map iterations at the stated $K$ values.

Builds on: Integrable Systems and Action-Angle Variables · The KAM Theorem · Conservative Systems and Hamiltonian Phase Flow · Classification of Fixed Points: Nodes, Saddles, Spirals, Centers Leads to: The Standard Map and Chaotic Hamiltonian Dynamics · Homoclinic and Heteroclinic Orbits and the Smale Horseshoe

Why does an O(ε) perturbation produce an O(√ε) island — where does the amplification come from?

From the flatness of the unperturbed Hamiltonian along the resonance direction. Off resonance, a perturbing mode fights the full frequency $\mathbf{k}\cdot\boldsymbol{\omega}$ and produces only $O(\varepsilon)$ wiggles. On resonance that frequency is zero: the linear term in the local expansion of $H_0$ vanishes identically (Step 2 of the pendulum reduction), leaving only the quadratic $\tfrac12 G \Delta J^2$ to resist the perturbation's $\varepsilon F\cos\psi$ . Balancing a quadratic restoring term against a fixed-amplitude forcing gives displacement $\propto \sqrt{\varepsilon F/G}$ — the same reason a ball in a nearly flat trough rolls far under a tiny tilt. This is the single most practical scaling in the subject: it is why "small" perturbations dominate long-time transport, why the destroyed measure in the KAM census is $\sqrt{\varepsilon}$ rather than $\varepsilon$ (each of the surviving-order resonances contributes its width), and why doubling machine tolerances buys only a $\sqrt2$ improvement in resonance sizes.

Is the chaotic layer around a separatrix really unavoidable — the pendulum we derived is integrable, after all?

The pendulum is integrable, but it is an averaged model: the terms discarded in Step 3 of the reduction — the off-resonant Fourier modes — come back as a fast, weak, periodic shaking of that pendulum. A shaken pendulum is precisely the system whose separatrix splits: its stable and unstable manifolds, which coincided in the integrable limit, now intersect transversally (Melnikov's method computes the splitting angle, exponentially small in the frequency ratio, $\sim e^{-c/\sqrt{\varepsilon}}$ ), and one transverse intersection forces the full homoclinic tangle and embedded horseshoe of the next-but-one article. So every island chain is born already wearing a chaotic film. At small $\varepsilon$ the film is exponentially thin — invisible on coarse plots, which is why low- $K$ standard-map portraits look perfectly ordered — but it thickens with $\varepsilon$ and supplies the connective tissue once chains approach overlap. The layer is also why the criterion overestimates: the "gap" between islands must only be bridged by the layers, not by the islands proper.

Do the islands themselves ever become chaotic inside?

The island interior is a miniature of the whole problem — that is the deepest structural fact in the subject. Around the elliptic centre, motion is again near-integrable (librations with frequency $\omega_{\text{lib}}$ replacing rotations), so it has its own actions, its own rational/irrational winding numbers, its own KAM curves — and its own resonances: when $\omega_{\text{lib}}$ is commensurate with the fast frequency, secondary island chains appear inside the island, each reducible to its own pendulum, wrapped in its own micro-layer, containing tertiary islands... The hierarchy is self-similar and infinite, and renormalization treatments of the critical KAM curve exploit exactly this nesting. Practically, islands hollow out from the edge inward as $\varepsilon$ grows: the outermost librational curves die first (they are closest to the separatrix layer), the chaotic film invades, and near the global threshold only a small stable core survives around the elliptic point — which itself can finally destabilize through a period-doubling cascade, connecting this chapter back to the dissipative one. "Inside every island is a smaller ocean" is the accurate slogan.

How do I actually compute island widths for a real system, where F and G are not handed to me?

Recipe with the three quantities you must dig out. (1) The resonance: find where $\mathbf{k}\cdot\boldsymbol{\omega}(\mathbf{J}) = 0$ for the low-order $\mathbf{k}$ of interest — this needs the frequency map of your integrable approximation (analytically, or numerically by frequency analysis of orbits). (2) G, the projected twist: $G = \sum k_i k_j \,\partial^2 H_0/\partial J_i\partial J_j$ evaluated at the resonance — for Kepler-like or rotor-like $H_0$ this is a one-line derivative; for numerical models, difference the measured frequencies. (3) F, the resonant amplitude: the Fourier coefficient of the perturbation along $\mathbf{k}$ at the resonant torus, $F = 2|H_{\mathbf{k}}(\mathbf{J}_0)|$ — obtained by an angle-space FFT of the perturbation evaluated on the unperturbed orbit. Then $\Delta J_{\max} = 2\sqrt{\varepsilon|F/G|}$ , and the sanity checks: the predicted libration frequency $\sqrt{\varepsilon|FG|}$ should match the small oscillation you see around the island centre in simulation, and the measured width on a Poincaré section (scan initial conditions through the chain at the island's phase) should match $\Delta J_{\max}$ within tens of percent. If it does not, the usual culprits are a non-negligible second harmonic in $\psi$ or proximity to a degenerate ( $G \approx 0$ ) point.

Does resonance overlap explain chaos in systems with three or more degrees of freedom too?

It explains the fast chaos, but the geography changes qualitatively. For $n \ge 3$ the resonance surfaces $\mathbf{k}\cdot\boldsymbol{\omega} = 0$ are codimension-one sheets in an $n$ -dimensional frequency space, and they intersect: the resonance web is connected even at tiny $\varepsilon$ , when every individual pairwise overlap ratio is far below one. Along this web Arnold diffusion trickles — motion along resonance lines and through their junctions — at rates exponentially small in $1/\sqrt{\varepsilon}$ (Nekhoroshev bounds), utterly unlike the fast cross-resonance transport of overlap. Chirikov's criterion then answers a different, still central question: on which parts of the web is transport fast? Where local overlap holds (near multiple-resonance junctions especially), diffusion is quick and quasilinear; where it fails, orbits are stuck with the exponentially slow web-crawling. Practical stability analyses of, e.g., planetary systems or acculmulator rings therefore run a two-tier audit: Chirikov overlap for the fast channels, Nekhoroshev/frequency-map arguments for the residual slow leaks.

Why is everything reduced to a pendulum and not, say, to an oscillator?

Because a resonance is intrinsically a phase-trapping phenomenon, and trapping-versus-running is the pendulum's dichotomy, not the oscillator's. The slow phase $\psi$ lives on a circle; its effective potential $\varepsilon F\cos\psi$ is necessarily periodic in $\psi$ ; and a particle in a periodic potential either librates in a well or circulates over the crests — that phase portrait, with its separatrix joining the two regimes through a saddle, is the pendulum's and cannot be anything else. The harmonic oscillator, by contrast, has an unbounded quadratic potential, no separatrix, no circulation, and — as the integrable-baseline article stressed — no twist; it cannot represent the coexistence of trapped and passing orbits that defines a resonance zone. This is also why the pendulum earns the title "the hydrogen atom of nonlinear dynamics": it is the universal local normal form at any generic resonance of any twist system, so everything learned about one pendulum (width, libration frequency, separatrix layer, response to shaking) is a theorem about every resonance in every machine, molecule, and planetary system. The only escapes from pendulum universality are non-generic: vanishing twist $G$ , vanishing resonant harmonic $F$ , or symmetry-forced higher harmonics — each with its own (well-catalogued) alternative normal form.