Dynamical Systems: Hamiltonian & KAM¶

Job: develop Hamiltonian dynamics at the frontier — symplectic manifolds and Hamilton's equations, the Arnold–Liouville complete integrability theorem with action-angle coordinates, the Kolmogorov–Arnold–Moser (KAM) theorem on the persistence of invariant tori under small perturbation, the small-divisor obstruction, Nekhoroshev exponential stability, and the open Arnold diffusion problem.

Core¶

Hamiltonian systems¶

A symplectic manifold is $(M^{2n},\omega)$ with $\omega$ a closed nondegenerate 2-form (see symplectic geometry). A Hamiltonian $H:M\to\mathbb{R}$ defines the vector field $X_H$ by $i_{X_H}\omega=-dH$; Hamilton's equations are $\dot x=X_H(x)$. Liouville's theorem: the flow $\varphi^t_H$ preserves the symplectic form and the Liouville volume $\omega^n/n!$ (phase-space volume is preserved). The Poisson bracket $\{F,G\}=\omega(X_F,X_G)$ makes $C^\infty(M)$ a Lie algebra; $H$ is conserved along the flow ($\{H,H\}=0$).

Theorem (Arnold–Liouville, complete integrability). If $H$ has $n$ conserved quantities $F_1=H,F_2,\dots,F_n$ in involution ($\{F_i,F_j\}=0$) that are independent on an open dense set, then on each regular level set $F^{-1}(c)\cong\mathbb{T}^n$ (a compact connected component), the dynamics is quasi-periodic: there exist action-angle coordinates $(I,\theta)\in\mathbb{R}^n\times\mathbb{T}^n$ in which $H=H_0(I)$ and the flow is $\dot I=0,\;\dot\theta=\omega(I):=\partial H_0/\partial I$ — a linear flow on the invariant torus $\mathbb{T}^n$ with frequency vector $\omega(I)$.

The integrable case is the unperturbed reference for KAM: the dynamics on the invariant tori $\mathbb{T}^n$ is quasi-periodic, and the frequency map $I\mapsto\omega(I)$ encodes the rotation numbers.

KAM theorem¶

Theorem (Kolmogorov 1954, Arnold 1963, Moser 1962). Let $H(I,\theta)=H_0(I)+\varepsilon H_1(I,\theta)+O(\varepsilon^2)$ be a near-integrable Hamiltonian on $\mathbb{R}^n\times\mathbb{T}^n$ (or a symplectic manifold), with $H_0$ non-degenerate (the frequency map $I\mapsto\omega(I)=\partial H_0/\partial I$ is a local diffeomorphism, $\det\partial^2 H_0/\partial I^2\ne 0$). For sufficiently small $\varepsilon$, a Cantor set of positive-measure invariant tori persists: for each Diophantine frequency $\omega$ (satisfying $|\langle k,\omega\rangle|\ge\gamma|k|^{-\tau}$ for all $k\in\mathbb{Z}^n\setminus\{0\}$, with $\gamma>0,\tau>n-1$), there is an invariant torus $\mathbb{T}^n_\omega$ on which the dynamics is conjugate to $\theta\mapsto\theta+\omega$. The surviving tori form a Cantor set $\mathcal{K}_\varepsilon\subset\mathbb{T}^{2n}$ of measure $|\mathcal{K}_\varepsilon|\to 1$ as $\varepsilon\to 0$. The tori are $C^\infty$ (or $C^r$ with $r\to\infty$ as $\varepsilon\to 0$ depending on the version).

Integrable structure ($\varepsilon=0$)	Perturbed structure ($\varepsilon>0$ small)
Full-measure family of invariant tori $\mathbb{T}^n_I$	Cantor set $\mathcal{K}_\varepsilon$ of invariant tori, $
Frequency map $I\mapsto\omega(I)$ diffeo	Diophantine frequencies survive; resonant tori generically break
Quasi-periodic dynamics $\theta\mapsto\theta+\omega$	Same on surviving tori (conjugate); gaps between tori host chaotic dynamics
Smooth foliation by tori	Cantor foliation (nowhere dense, positive measure)
All actions conserved	Actions drift in the gaps (Arnold diffusion — open)

The small-divisor problem¶

The KAM construction solves the cohomological equation $\omega\cdot\partial_\theta u=H_1-\bar H_1$ (the first-order correction to the normal form), where $\bar H_1(I)=\int H_1(I,\theta)\,d\theta$ is the average. Expanding $H_1=\sum_k H_{1,k}(I)e^{i\langle k,\theta\rangle}$, the solution is $u=\sum_k\frac{H_{1,k}}{i\langle k,\omega\rangle}e^{i\langle k,\theta\rangle}$ — but the denominators $\langle k,\omega\rangle$ can be arbitrarily small (the small divisors / resonances). The Diophantine condition $|\langle k,\omega\rangle|\ge\gamma|k|^{-\tau}$ bounds the small divisors polynomially and gives a convergent (for $\tau$ large enough) Fourier series for $u$. The Kolmogorov / Newton iteration (superconvergent: the error at step $j$ is $O(\varepsilon^{2^j})$) overcomes the small divisors by combining the Diophantine control with a quadratic iteration that gains convergence faster than the small divisors lose it.

Nekhoroshev stability¶

Theorem (Nekhoroshev 1977). Under the convexity / steepness of $H_0$ (a generic condition stronger than non-degeneracy, e.g. convex $H_0$), the actions are stable for exponentially long times: there exist $a,b,c>0$ (depending on $n$ and the steepness) such that for $|t|\le\exp(c/\varepsilon^a)$, $|I(t)-I(0)|\le\varepsilon^b$. The stability is everywhere (not just on the KAM Cantor set), for exponentially long (but finite) time.

The KAM vs Nekhoroshev dichotomy: KAM gives perpetual stability on a Cantor set (the surviving tori); Nekhoroshev gives long-time stability everywhere (exponential time, off the tori). Together they cover the phase space — the KAM tori trap most of the measure, and the Nekhoroshev theorem controls the gaps for exponentially long times — but neither is global perpetual stability.

Arnold diffusion¶

Open problem (Arnold diffusion). In the gaps between KAM tori, can the actions drift by $O(1)$ over long (polynomial / sub-exponential) times, for generic perturbations? KAM (perpetual stability on the Cantor set) and Nekhoroshev (exponential stability everywhere) do not rule out the slow drift in the gaps. Arnold's 1964 example gives a mechanism (a chain of whiskered tori with heteroclinic connections); the rigorous results (Mather, Bernard, Kaloshin–Zhang) construct diffusion in specific settings, but the generic / general case (the diffusion time scale, the genericity of Arnold diffusion) is open.

The Chirikov overlap criterion (heuristic): when the resonances overlap (the widths of the resonant islands sum to more than their separation), the KAM tori break and diffusion occurs. The rigorous theory replaces the heuristic with the variational / geometric mechanisms (Mather's variational method, the Aubry–Mather set, the Bernard / Kaloshin constructions), but the full picture is open.

See it¶

The Kolmogorov normal form on a single Diophantine torus, via the Newton iteration solving the cohomological equation. Fix a Diophantine frequency $\omega$ (satisfying $|\langle k,\omega\rangle|\ge\gamma|k|^{-\tau}$) and consider $H(I,\theta)=H_0(I)+\varepsilon H_1(I,\theta)+O(\varepsilon^2)$ near the torus $I=I_0$ with $\omega(I_0)=\omega$. The KAM step: find a near-identity canonical transformation $(I,\theta)=\Phi(J,\psi)$ that removes the $\theta$-dependence to first order, putting $H$ into the Kolmogorov normal form

\[ H\circ\Phi = h_\omega(J) + \varepsilon\bar H_1(J) + O(\varepsilon^2),\qquad h_\omega(J)=H_0(I_0)+\langle\omega,J-I_0\rangle+\tfrac12\langle A(J-I_0),(J-I_0)\rangle, \]

on a thin neighborhood of the torus $J=I_0$. The transformation $\Phi$ is constructed by solving the cohomological equation $\omega\cdot\partial_\psi u(\psi)=H_1(I_0,\psi)-\bar H_1(I_0)$ for the angle-dependent part $u$ (the generator of the canonical transformation). Expanding $H_1(I_0,\psi)=\sum_k H_{1,k}e^{i\langle k,\psi\rangle}$, the solution is

\[ u(\psi)=\sum_{k\ne 0}\frac{H_{1,k}}{i\langle k,\omega\rangle}e^{i\langle k,\psi\rangle}, \]

which converges absolutely by the Diophantine condition (the denominators are $\ge\gamma|k|^{-\tau}$, so $|u|\le C\sum_k|H_{1,k}||k|^\tau$, convergent if $H_1$ is analytic in a strip). The Newton iteration repeats the step: the error after the first step is $O(\varepsilon^2)$, and the next step solves a cohomological equation with $\varepsilon^2$-sized RHS, giving $O(\varepsilon^4)$ error, then $O(\varepsilon^8)$, etc. — superconvergence $\varepsilon^{2^j}$. The small divisors lose a factor $|k|^\tau$ per step, but the Newton quadratic gain dominates: the iteration converges for $\varepsilon<\varepsilon_0(\gamma,\tau,n)$, and the limit is a smooth (analytic for analytic $H$) canonical transformation putting $H$ into the Kolmogorov form on the surviving torus. This single computation — the cohomological equation with the Diophantine denominator control and the Newton superconvergence — is the operational content of KAM: the small divisors are controlled polynomially, and the quadratic iteration overcomes them.

Check¶

Explain why the Diophantine condition is the right hypothesis for KAM, and identify what fails for the resonant frequencies $\omega$ with $\langle k,\omega\rangle=0$.

The cohomological equation $\omega\cdot\partial_\theta u=H_1-\bar H_1$ has a solution only when the RHS is in the image of $\omega\cdot\partial_\theta$; for a Fourier mode $e^{i\langle k,\theta\rangle}$, this requires dividing by $i\langle k,\omega\rangle$. If $\omega$ is resonant ($\langle k,\omega\rangle=0$ for some $k\ne 0$), the denominator vanishes and the Fourier mode $H_{1,k}$ cannot be removed — the normal-form construction breaks down, and the corresponding torus generically splits / breaks into resonant islands. The Diophantine condition $|\langle k,\omega\rangle|\ge\gamma|k|^{-\tau}$ excludes all resonances with a polynomial lower bound, giving a convergent (for analytic $H_1$ and $\tau>n-1$) Fourier series for $u$ — the small divisors are small but polynomially controlled, and the Newton iteration overcomes them. The resonant tori (with $\langle k,\omega\rangle=0$) form a measure-zero set (the resonances are dense but of zero measure for non-degenerate $H_0$); the Diophantine tori have full measure, and KAM preserves them.

State the precise dichotomy between KAM (perpetual stability on a Cantor set) and Nekhoroshev (exponential stability everywhere), and identify the open problem in the gap.

KAM: the surviving tori form a Cantor set $\mathcal{K}_\varepsilon$ of measure $|\mathcal{K}_\varepsilon|\to 1$ as $\varepsilon\to 0$; on $\mathcal{K}_\varepsilon$ the dynamics is quasi-periodic and the actions are conserved perpetually (for all time). Nekhoroshev: off $\mathcal{K}_\varepsilon$ (in the gaps), the actions are stable for exponentially long time $|I(t)-I(0)|\le\varepsilon^b$ for $|t|\le\exp(c/\varepsilon^a)$ — finite but exponential time, everywhere. The gap (the open problem) is the behavior at times longer than the Nekhoroshev time: can the actions drift by $O(1)$ (Arnold diffusion) in the gaps, for generic perturbations? KAM + Nekhoroshev together bound the drift up to the Nekhoroshev time, but beyond it the dynamics in the gaps is open. The Arnold diffusion question is whether the drift occurs generically, on what time scale (the conjectured $\exp(c/\varepsilon^a)$ scaling and the generic / average behavior), and via what mechanism (Mather's variational method, the Aubry–Mather set, the Bernard / Kaloshin geometric constructions are the rigorous approaches, but the general case is open).

Pitfalls¶

KAM does not imply global stability: it leaves a measure-zero but topologically connected chaotic sea (the gaps between KAM tori); Arnold diffusion in the gaps is open. Treating KAM as "the perturbation is stable" overstates the result.
KAM tori form a Cantor set (closed, nowhere dense, positive measure): not a smooth foliation; the surviving tori are $C^\infty$ individually but the Cantor set is not a manifold. The phase space is a Cantor foliation, not a smooth one.
The Diophantine condition is the right hypothesis, but the resonant tori break: only the Diophantine (non-resonant) tori survive; the resonant ones generically break into island chains. The full-measure KAM Cantor set is the Diophantine part, not the whole phase space.
Nekhoroshev requires convexity / steepness of $H_0$ (a generic condition stronger than non-degeneracy): without steepness, the exponential stability may fail; the Nekhoroshev theorem is not automatic from KAM non-degeneracy.
Arnold diffusion is open, not a theorem in general: the rigorous results (Mather, Bernard, Kaloshin–Zhang) construct diffusion in specific settings; the general / generic case is open. Do not state "Arnold diffusion always occurs" — it is conjectural.
The Chirikov overlap criterion is a heuristic, not a theorem: it gives a qualitative prediction of when KAM tori break, but the rigorous theory uses different mechanisms (variational / geometric); do not cite Chirikov as a proof.
The KAM tori are $C^\infty$ (or $C^r$), but the Cantor foliation is not a smooth family: the regularity in $\varepsilon$ and in the frequency $\omega$ is subtle; the tori are smooth individually but the Cantor set structure limits the smoothness of the foliation.

Refs¶

Kolmogorov, On dynamical systems with an integral invariant on the torus (Dokl. 1954); Arnold, Small denominators and the problem of stability of motion in classical and celestial mechanics (Russian Math. Surveys 1963); Moser, On invariant curves of area-preserving mappings (Nachr. 1962)
Nekhoroshev, Exponential estimate of the time of stability of nearly integrable Hamiltonian systems (Russian Math. Surveys 1977)
Arnold, Instability of dynamical systems with many degrees of freedom (Dokl. 1964, Arnold diffusion)
Mather, Variational construction of connecting orbits; Bernard, Connecting orbits and time-dependent Hamiltonians; Kaloshin–Zhang (Arnold diffusion constructions)
Lochak, Canonical perturbation theory via simultaneous approximation; Chierchia, KAM theory

Next¶

Hyperbolic Dynamics · Bifurcations & Chaos · Ergodic & Physical Measures · DS index · Foundations · Symplectic geometry · Mathematical physics · PDE