Kolmogorov-Arnold-Moser theorem

Intuition [Beginner]

Picture a frictionless solar system. Each planet traces an ellipse around the Sun and the orbits never overlap or drift apart. In the right coordinates the entire motion is a clock: each orbit is a circle, and the system is a collection of independent clocks ticking at fixed rates. Mathematicians call such systems integrable, and the geometric picture is that phase space is foliated by tori — donuts of higher dimension — with the trajectory winding around each torus at a constant rate.

Now switch on a small perturbation. Maybe the planets begin to pull on each other. The clocks no longer tick independently. A skeptical question follows: do any of the donuts survive, or does everything collapse into chaos?

The KAM theorem answers: most of the donuts survive. If a donut's frequencies are sufficiently irrational — far from being rationally related — then a slightly deformed copy of it persists in the perturbed system. Donuts whose frequencies are rationally related, or close to it, do break apart and produce chaotic regions. The surviving donuts have full measure when the perturbation is small enough.

This is why solar systems can be stable for billions of years even though the equations are not exactly solvable.

Visual [Beginner]

Two phase portraits side by side. On the left the unperturbed system: clean nested donuts, each a steady-rotation orbit. On the right the perturbed system: most donuts persist, slightly bent; a few thin chaotic bands appear where rational-frequency donuts used to be.

The picture captures the headline result: most invariant tori survive a small perturbation, with a Cantor-like set of surviving frequencies whose complement has small measure.

Worked example [Beginner]

Take a pendulum coupled weakly to a rotor. The pendulum swings at frequency $1.0$, the rotor turns at frequency $1.7320508$ — close to $\sqrt{3}$, an irrational number that is hard to approximate by rationals. With no coupling each motion is independent: phase space is a product of circles, and trajectories live on two-dimensional donuts.

Switch on a coupling of strength $0.01$. The donut at the pendulum-rotor pair $(1.0,1.7320508)$ has frequency vector $(1.0,1.732\dots )$ which satisfies a strong irrationality condition. KAM guarantees a slightly deformed donut survives, with the same frequency vector, and the trajectory continues to wind around it densely.

Compare to a different starting condition with frequencies $(1.0,2.0)$ — exactly $2:1$ resonant. Here KAM offers no guarantee: the resonant donut is destroyed and a thin chaotic layer appears where regular motion used to be.

Takeaway: irrational frequency ratios protect orbits from a small perturbation; rational ratios do not.

Check your understanding [Beginner]

Formal definition [Intermediate+]

Let $(M^{2n},\omega )$ be a symplectic manifold and let $H_0:M\to \mathbb{R}$ be a completely integrable Hamiltonian with $n$ Poisson-commuting integrals $F_1,\dots ,F_n$ (with $F_1=H_0$). The Liouville-Arnold theorem 05.02.03 gives, on a neighbourhood of a compact regular common level set, action-angle coordinates 05.02.04 $$ (I, \theta) \in U \times \mathbb{T}^n \subset \mathbb{R}^n \times \mathbb{T}^n, $$ with $\omega ={\sum }_{i=1}^{n}d{I}_i\wedge d{\theta }_i$ and $H_0=H_0(I)$ depending only on the actions. The Hamiltonian flow on each torus $\{I=I^{\ast }\}$ is the linear flow $$ \dot\theta = \omega(I^*), \qquad \omega(I) := \frac{\partial H_0}{\partial I}(I) \in \mathbb{R}^n. $$ The map $I\mapsto \omega (I)$ is the frequency map of the integrable system.

A perturbation in this setup is a Hamiltonian $$ H_\epsilon(I, \theta) = H_0(I) + \epsilon H_1(I, \theta), $$ with $H_1$ a smooth (or real-analytic) function on $U\times {\mathbb{T}}^n$ and $ϵ\in \mathbb{R}$ a small parameter.

Diophantine condition. A frequency vector ${\omega }^{\ast }\in {\mathbb{R}}^n$ is $(\gamma ,\tau )$-Diophantine if $$ |\langle k, \omega^* \rangle| \geq \frac{\gamma}{|k|^\tau} \quad \text{for all } k \in \mathbb{Z}^n \setminus {0}, $$ where $|k|:=|k_1|+\cdots +|k_n|$. Write $D_{\gamma ,\tau }\subset {\mathbb{R}}^n$ for the set of such frequencies. For $\tau >n-1$ the complement of $D_{\gamma ,\tau }$ in any bounded set $\Omega$ has Lebesgue measure $\le C(\Omega ,\tau )\gamma$, so ${\bigcup }_{\gamma }{D}_{\gamma ,\tau }$ has full Lebesgue measure in ${\mathbb{R}}^n$.

Kolmogorov non-degeneracy. $H_0$ is non-degenerate at $I^{\ast }\in U$ if $$ \det \frac{\partial^2 H_0}{\partial I , \partial I}(I^) \neq 0, $$ equivalently the frequency map $I\mapsto \omega (I)$ is a local diffeomorphism near $I^$. This condition lets one parametrise tori by their frequencies rather than by their actions, which is what the KAM iteration requires.

The Hamiltonian $H_{ϵ}$ defines a flow ${\Phi }_{ϵ}^t$ on $U\times {\mathbb{T}}^n$. A subset $T\subset U\times {\mathbb{T}}^n$ is an invariant torus with frequency ${\omega }^{\ast }$ if $T$ is diffeomorphic to ${\mathbb{T}}^n$, ${\Phi }_{ϵ}^t$-invariant, and the flow restricted to $T$ is smoothly conjugate to the linear flow $\theta \mapsto \theta +t{\omega }^{\ast }$ on ${\mathbb{T}}^n$.

Key theorem with proof [Intermediate+]

Theorem (KAM, Kolmogorov 1954 / Arnold 1963 / Moser 1962). Let $H_{ϵ}=H_0(I)+ϵH_1(I,\theta )$ be a real-analytic Hamiltonian on $U\times {\mathbb{T}}^n$. Suppose:

1. (Non-degeneracy.) $H_0$ is non-degenerate at $I^ \in U$,i.e.\ast$\det \partial^2_I H_0(I^*) \neq 0$.
2. (Diophantine frequency.) $\omega^ := \partial_I H_0(I^)$is$(\gamma, \tau)$-Diophantineforsome$\gamma > 0$,$\tau > n - 1$.
3. (Real-analyticity.) $H_0,H_1$ extend holomorphically to a strip $|\mathrm{Im}\theta |<r$, $|\operatorname{Re} I - I^| < s$, with bounded norms.*

Then there exists ${ϵ}_0={ϵ}_0(\gamma ,\tau ,r,s,\Vert H_0\Vert ,\Vert H_1\Vert )>0$ such that for $|ϵ|<{ϵ}_0$ the Hamiltonian $H_{ϵ}$ has an invariant torus $T_{\omega^, \epsilon}$near${I = I^}$onwhichtheflowisreal-analyticallyconjugateto$\theta \mapsto \theta + t \omega^$on$\mathbb{T}^n$* ^{[Kolmogorov 1954; ref: TODO_REF Arnold 1963]}.

Measure-theoretic conclusion. The union $\bigcup_{\omega^ \in D_{\gamma, \tau}} T_{\omega^, \epsilon}$hasLebesguemeasure$\geq (1 - C\sqrt{\epsilon}) \cdot \operatorname{vol}(U \times \mathbb{T}^n)$for$\epsilon$ small.

Proof (Newton-iteration / KAM scheme — outline). The construction is a quadratically-convergent Newton scheme that produces, at each step, a symplectic change of variables straightening the perturbation closer to the integrable form. Set $H^{(0)}:=H_{ϵ}$ with perturbation size ${ϵ}^{(0)}:=ϵ$.

Step 1: linearised conjugacy equation. At iteration $n$, write $$ H^{(n)}(I, \theta) = H^{(n)}0(I) + R^{(n)}(I, \theta), $$ with $R^{(n)}$ a remainder of analytic norm ${ϵ}^{(n)}$ on a strip $|\mathrm{Im}\theta |<r_n$. Seek a symplectomorphism ${\varphi }^{(n)}$ generated, in a generating-function sense 05.05.03, by an auxiliary function $F^{(n)}(I,\theta )$ — that is, ${\varphi }^{(n)}$ is the time-one map of the Hamiltonian vector field $X{F^{(n)}}$.Toleadingorderin$F^{(n)}$, the pulled-back Hamiltonian equals $$ H^{(n)} \circ \phi^{(n)} = H^{(n)}_0(I) + {H^{(n)}_0, F^{(n)}}(I, \theta) + R^{(n)}(I, \theta) + O(|F^{(n)}|^2). $$ Choose $F^{(n)}$ so that the first-order correction cancels the angle-dependent part of $R^{(n)}$: $$ {H^{(n)}0, F^{(n)}} = \langle R^{(n)} \rangle - R^{(n)}, $$ where $\langle R^{(n)} \rangle(I) := (2\pi)^{-n} \int{\mathbb{T}^n} R^{(n)}(I, \theta), d\theta$istheangle-average.Theangle-averageisabsorbedinto$H^{(n+1)}_0 := H^{(n)}_0 + \langle R^{(n)} \rangle$.

Step 2: solving the cohomological equation. Fix $I^{\ast }$ and freeze the frequency at ${\omega }^{\ast }={\partial }_I{H}_0^{(n)}(I^{\ast })$. The Poisson bracket $\{H_0^{(n)},F^{(n)}\}{|}_{I=I^{\ast }}=-⟨{\omega }^{\ast },{\partial }_{\theta }{F}^{(n)}⟩$, so the equation becomes $$ \langle \omega^, \partial_\theta F^{(n)}(I^, \theta) \rangle = R^{(n)}(I^, \theta) - \langle R^{(n)} \rangle(I^). $$ Expand both sides in Fourier series in $\theta$: $$ F^{(n)}(I^, \theta) = \sum_{k \neq 0} F^{(n)}_k e^{i \langle k, \theta \rangle}, \qquad R^{(n)}(I^, \theta) = \sum_{k} R^{(n)}_k e^{i\langle k, \theta\rangle}. $$ Equating coefficients gives the explicit solution $$ F^{(n)}_k = \frac{R^{(n)}k}{i \langle k, \omega^* \rangle} \quad (k \neq 0), $$ provided every denominator $⟨k,{\omega }^{\ast }⟩$ is non-zero. The Diophantine bound $|⟨k,{\omega }^{\ast }⟩|\ge \gamma /|k{|}^{\tau }$ controls the divergence of these denominators: a Fourier mode at wavenumber $k$ is amplified by at most $|k{|}^{\tau }/\gamma$. For real-analytic $R^{(n)}$ on a strip of width $r_n$, $|R_k^{(n)}|\le \Vert R^{(n)}{\Vert }_{r_n}{e}^{-r_n|k|}$, so the small-denominator amplification is dominated by the exponential decay and $F^{(n)}$ is real-analytic on a strictly smaller strip. Quantitatively (Pöschel-style estimate on a strip $r_n-{\delta }_n$ with $0<{\delta }_n<r_n$): $$ |F^{(n)}|{r_n - \delta_n} \leq C \frac{\epsilon^{(n)}}{\gamma , \delta_n^{n + \tau}}, $$ where $C$ depends only on $n$ and $\tau$.

Step 3: quadratic convergence. The next-step remainder $R^{(n+1)}:=H^{(n)}\circ {\varphi }^{(n)}-H_0^{(n+1)}$ contains only second-order terms in $F^{(n)}$ (the leading correction having been arranged to cancel) plus the variation of $H_0^{(n)}$ across the level $\{I=I^{\ast }\}$. A Cauchy-estimate computation gives $$ \epsilon^{(n+1)} = |R^{(n+1)}|{r{n+1}} \leq C \frac{(\epsilon^{(n)})^2}{\gamma^2 , \delta_n^{2(n + \tau) + 2}}, $$ on a strip of width $r_{n+1}:=r_n-{\delta }_n$. Choosing ${\delta }_n:=r_0/2^{n+2}$, the strips $r_n$ remain bounded below by $r_0/2$. With this choice, induction on $n$ shows $$ \epsilon^{(n)} \leq K^{2^n - 1} (\epsilon^{(0)})^{2^n} $$ for a constant $K=K(\gamma ,\tau ,r_0)$. Provided ${ϵ}^{(0)}$ is small enough that $K{ϵ}^{(0)}<1$, the sequence ${ϵ}^{(n)}$ tends to zero super-exponentially. The composed transformations ${\varphi }^{(0)}\circ {\varphi }^{(1)}\circ \cdots$ converge in the analytic norm on the strip $r_{\infty }=r_0/2$ to a real-analytic symplectomorphism ${\Phi }_{\infty }$ defined near $I=I^{\ast }$ on ${\mathbb{T}}^n$.

Step 4: the surviving torus. In the limit coordinates, $H_{ϵ}\circ {\Phi }_{\infty }=H_{\infty }(I)$ depends only on $I$, and the level set $\{I=I^{\ast }\}$ in the new coordinates is invariant with linear flow at frequency ${\omega }^{\ast }$. Pulling back through ${\Phi }_{\infty }$ gives the invariant torus $T_{{\omega }^{\ast },ϵ}\subset U\times {\mathbb{T}}^n$ for the original Hamiltonian, with the conjugacy a real-analytic diffeomorphism. $\square$

Bridge. The KAM scheme builds toward the modern theory of Floer homology 05.08.02 and the Arnold conjecture 05.08.01, where lower bounds on the number of fixed points of Hamiltonian symplectomorphisms appear again as a torus-counting principle: KAM tori are non-degenerate fixed points of the Poincaré time-one map on a transversal section, and the Conley-Zehnder index 05.08.04 keeps track of their stability. The cohomological equation $\{H_0,F\}=R-⟨R⟩$ that drives KAM also appears again in normal-form theory and in the derivation of the Birkhoff normal form near elliptic fixed points. Putting these together, the foundational reason most tori survive is exactly the Diophantine bound: it provides the quantitative input that converts a divergent Fourier sum into a convergent one, and the Newton iteration's quadratic convergence is what tolerates the step-by-step loss of analyticity.

Exercises [Intermediate+]

Lean formalization [Intermediate+]

lean_status: none — Mathlib lacks the differential-geometric, function-space, and measure-theoretic infrastructure needed for KAM. A formal statement would look like the following pseudocode, with each axiom replaced by a real definition once the prerequisites are in Mathlib.

[object Promise]

A formal route would assemble: smooth and real-analytic function spaces with weighted norms, the symplectic form on ${\mathbb{R}}^n\times {\mathbb{T}}^n$, the Liouville-Arnold theorem, the Diophantine condition with its measure estimate, the cohomological equation with Fourier-division solution, and a Newton-iteration convergence lemma combining the Cauchy-estimate inequality with super-exponential bookkeeping.

Advanced results [Master]

The KAM theorem is the headline statement of a much larger structural circle of ideas concerning persistence of invariant manifolds under Hamiltonian perturbation. Six refinements deepen, generalise, or sit beside the basic theorem.

Moser's twist theorem (1962). For an area-preserving diffeomorphism of the annulus $\mathbb{T}\times [a,b]$ of the form $(\theta ,I)\mapsto (\theta +\alpha (I),I)+O(ϵ)$ with ${\alpha }^{\prime }(I)\ne 0$, KAM circles with Diophantine rotation number persist for $ϵ$ small. Moser's original 1962 paper handled the $C^k$ case via a Nash-Moser smoothing scheme; this was the version that opened the technique to dynamical-systems applications outside the analytic category ^{[Moser 1962]}. The two-dimensional case has the additional feature that surviving KAM circles confine the dynamics — there is no Arnold-diffusion analogue in dimension two — so KAM alone gives full perpetual stability in the planar twist setting.

Pöschel's $C^k$-smooth theorem (1982). For a $C^k$ Hamiltonian with $k>2(n+\tau )$, the KAM theorem holds with the conjugacy a $C^{k-2(n+\tau )}$ map, and the dependence on the frequency parameter ${\omega }^{\ast }$ is Whitney-smooth on the Cantor set $D_{\gamma ,\tau }$ ^{[Pöschel 1982]}. Pöschel's iteration scheme separates the cohomological equation from the smoothing operator: at each step one truncates Fourier modes above a wavenumber $N_n$ growing geometrically, applies the Diophantine bound only to retained modes, and absorbs the high-mode tail into the next remainder. The threshold $k>2(n+\tau )$ is sharp up to a logarithmic factor. Salamon's 2004 reformulation ^{[Salamon 2004]} sharpens this further with explicit constants and a streamlined inverse-function-theorem viewpoint that emphasises the Newton structure over the Nash-Moser smoothing.

Lower-dimensional and partial KAM tori. The standard theorem yields full-dimensional invariant tori (dimension $n$ in a $2n$-dimensional phase space). Eliasson, Kuksin, and others established analogous persistence for lower-dimensional invariant tori — typically tori carrying linear flow at a frequency ${\omega }^{\ast }\in {\mathbb{R}}^d$ with $d<n$, with the remaining $n-d$ directions hyperbolic or elliptic. Persistence requires a Diophantine-like condition on ${\omega }^{\ast }$ together with Melnikov non-degeneracy conditions controlling the normal frequencies. Lower-dimensional KAM tori are the framework underlying the modern theory of Hamiltonian PDE: invariant tori for nonlinear Schrödinger and KdV equations are infinite-dimensional analogues of finite lower-dimensional KAM tori.

Nekhoroshev's theorem (1977). Under a steepness or quasi-convexity hypothesis on $H_0$ — strictly stronger than Kolmogorov non-degeneracy — every solution of $H_{ϵ}$, regardless of initial condition, has action variables that drift by at most ${ϵ}^a$ over a time of length $\exp (c{ϵ}^{-b})$, for explicit Nekhoroshev exponents $a,b>0$ depending only on $n$ and the steepness data. KAM and Nekhoroshev are complementary: KAM controls measure-typical orbits perpetually, while Nekhoroshev controls every orbit for exponentially long times. In the resonance gaps between KAM tori, Nekhoroshev provides the only stability statement available.

Aubry-Mather theory. When KAM tori break under increasing perturbation strength, Aubry and Mather (independently, 1982-1984) showed that they leave behind invariant Cantor sets — cantori — that retain the rotation-number structure of the destroyed torus but have measure zero. Cantori are the minimisers of an action functional on the space of invariant measures with prescribed rotation vector, and they exist for every irrational rotation number (not just Diophantine ones). They form the dynamical skeleton replacing destroyed tori in the symplectic-twist-map setting and are the central object of modern weak KAM theory.

Arnold diffusion. In dimension $2n\ge 6$, KAM tori do not separate phase space — their complement is connected. Arnold conjectured (1964) and proved in a model example that there exist orbits with action drift growing without bound over arbitrarily long times, threading through the resonance gaps between KAM tori. The mechanism — heteroclinic chains of whiskered tori — is now called Arnold diffusion. Quantitative bounds compatible with Nekhoroshev are an active area.

Synthesis. The Newton-iteration / Nash-Moser scheme that drives KAM is one prototype of how to invert a non-linear functional equation when the linearised inverse loses regularity. The cohomological equation $\{H_0,F\}=R-⟨R⟩$ has solutions that lose $\tau +n$ derivatives (in $C^k$) or shrink the strip of analyticity (in the analytic category) by an arbitrary $\delta$; the Newton iteration's quadratic convergence is fast enough to absorb this loss at every step, provided the Diophantine exponent $\tau$ is finite. The same scheme appears again in normal-form theory for vector fields near elliptic fixed points (the Birkhoff normal form), in the construction of foliations by invariant manifolds for hyperbolic systems with resonant eigenvalues, in the Hamilton-Jacobi inversion that produces action-angle coordinates 05.05.03, and in nonlinear PDE persistence theorems for $\mathrm{KdV}$ and nonlinear Schrödinger. The foundational reason these scattered results all use the same iteration is that the linearised problem is in each case a small-divisor problem in disguise: the inverse has poor regularity but predictable dependence on a Diophantine parameter, and the iteration's quadratic gain dominates the linear loss. Putting these together, the bridge between the analytic input — the cohomological equation — and the geometric output — invariant tori — is the foundational reason KAM unifies celestial mechanics, normal-form theory, and Hamiltonian PDE into a single technical apparatus.

Full proof set [Master]

Lemma (Cauchy bound for analytic Fourier coefficients). Let $f$ be holomorphic on the strip $|\mathrm{Im}\theta |<r$ with $\sup |f|=:\Vert f{\Vert }_r<\infty$, periodic in $\theta$. Then $|f_k|\le \Vert f{\Vert }_r{e}^{-r|k|}$ for every $k\in {\mathbb{Z}}^n$.

Proof. Translate the contour of integration in $f_k=(2\pi {)}^{-n}{\int }_{{\mathbb{T}}^n}f(\theta )e^{-i⟨k,\theta ⟩}d\theta$ into the strip by $\mathrm{Im}\theta =-\mathrm{sgn}(k)\cdot r^{\prime }$, $r^{\prime }<r$. The integrand picks up $e^{-r^{\prime }\sum |k_i|}=e^{-r^{\prime }|k|}$ in absolute value while the boundary of integration stays in the analyticity domain by Cauchy's theorem. Take $r^{\prime }\to r$. $\square$

Lemma (small-divisor estimate). Let $\omega^ \in \mathbb{R}^n$be$(\gamma, \tau)$-Diophantineandlet$R(\theta) = \sum_k R_k e^{i\langle k, \theta \rangle}$beareal-analyticfunctiononthestrip$|\operatorname{Im}\theta| < r$with$\langle R \rangle = R_0 = 0$.Thentheuniquezero-meansolution$F$of$\langle \omega^, \partial_\theta F \rangle = R$ is real-analytic on every strict sub-strip and $$ |F|_{r - \delta} \leq C(n, \tau) \frac{|R|_r}{\gamma , \delta^{n + \tau}} \qquad (0 < \delta < r). $$

Proof. The Fourier expansion of $F$ is $F={\sum }_{k\ne 0}{F}_k{e}^{i⟨k,\theta ⟩}$ with $F_k=R_k/(i⟨k,{\omega }^{\ast }⟩)$. The Diophantine bound gives $|F_k|\le |k{|}^{\tau }|R_k|/\gamma$. The previous lemma gives $|R_k|\le \Vert R{\Vert }_r{e}^{-r|k|}$. On the smaller strip $r-\delta$, $$ |F|{r - \delta} \leq \sum{k \neq 0} |F_k| e^{(r - \delta)|k|} \leq \frac{|R|r}{\gamma} \sum{k \neq 0} |k|^\tau e^{-\delta|k|}. $$ Group the sum by shells $\{|k|=m\}$: there are $\le C{n}^n{m}^{n-1}$ lattice points in shell $m$, so the sum is bounded by $C{\int }_0^{\infty }{m}^{n+\tau -1}{e}^{-\delta m}dm=C\Gamma (n+\tau )/{\delta }^{n+\tau }$. Absorb $\Gamma (n+\tau )$ into the constant $C(n,\tau )$. $\square$

Lemma (one Newton step). In the setup of the KAM theorem, suppose at step $n$ the Hamiltonian is $H^{(n)}=H_0^{(n)}+R^{(n)}$ on the strip $|\mathrm{Im}\theta |<r_n$, $|\operatorname{Re} I - I^n| < s_n$,with$|R^{(n)}|{r_n, s_n} \leq \epsilon^{(n)}$,$\partial_I H^{(n)}0(I^_n) = \omega^$,$\det \partial^2_I H^{(n)}0 \neq 0$,andtheinverseHessianisboundedby$M$.Choose$0 < \delta_n < r_n / 2$and$0 < \sigma_n < s_n / 2$.Thenthereexistsareal-analyticsymplectomorphism$\phi^{(n)}$onthestrip$|\operatorname{Im}\theta| < r_n - 2\delta_n$,$|\operatorname{Re} I - I^*{n+1}| < s_n - 2\sigma_n$,and$H^{(n+1)} := H^{(n)} \circ \phi^{(n)}$hastheform$H^{(n+1)}0 + R^{(n+1)}$ with* $$ |R^{(n+1)}|{r_{n+1}, s{n+1}} \leq C(n, \tau, M) \frac{(\epsilon^{(n)})^2}{\gamma^2 \delta_n^{2(n+\tau)} \sigma_n^2}, $$ where $r_{n+1}=r_n-2{\delta }_n$, $s_{n+1}=s_n-2{\sigma }_n$.

Proof. Decompose $R^{(n)}(I,\theta )=R^{(n)}(I_n^{\ast },\theta )+(I-I_n^{\ast })\cdot {\partial }_I{R}^{(n)}(I_n^{\ast },\theta )+R_2(I,\theta )$ where $R_2$ is quadratic in $I-I_n^{\ast }$. Apply the small-divisor lemma to the angle-dependent part of $R^{(n)}(I_n^{\ast },\cdot )$ to produce an auxiliary $F_1^{(n)}(\theta )$. Apply it again to the angle-dependent part of ${\partial }_I{R}^{(n)}(I_n^{\ast },\cdot )$ to produce $F_2^{(n)}(\theta )$. Set $F^{(n)}(I,\theta ):=F_1^{(n)}(\theta )+(I-I_n^{\ast })\cdot F_2^{(n)}(\theta )$ and let ${\varphi }^{(n)}$ be the time-one map of $X_{F^{(n)}}$. By construction ${\varphi }^{(n)}$ cancels $R^{(n)}$ to first order in $F$ and to first order in $I-I_n^{\ast }$. The new integrable part is $$ H^{(n+1)}_0(I) := H^{(n)}_0(I) + \langle R^{(n)} \rangle(I) + \text{constant correction from } F^{(n)}2. $$ The new action $I^*{n+1}$istheuniquesolutionof$\partial_I H^{(n+1)}0(I^*{n+1}) = \omega^$,givenbytheinversefunctiontheoremonthefrequencymap(using$M$).Theremainder$R^{(n+1)}$contains:thesecond-orderLie-seriestermsfrom$\phi^{(n)}$(controlledby$|F^{(n)}|^2$viaCauchyestimates),the$R_2$term(alreadyquadraticin$I - I^_n$,hencesmallafterrestrictingtothesmallerstrip$\sigma_n$),andthedifferencebetween$\partial_I H^{(n)}0$at$I^_n$and$I^_{n+1}$(controlledby$|I^*{n+1} - I^*_n| \leq M |\langle R^{(n)} \rangle|$).Eachpieceisboundedby$C(\epsilon^{(n)})^2 / (\gamma^2 \delta_n^{2(n+\tau)} \sigma_n^2)$.$\square$

Theorem (KAM convergence). Choose ${\delta }_n:=r_0/2^{n+2}$, ${\sigma }_n:=s_0/2^{n+2}$. There exists ${ϵ}_0={ϵ}_0(\gamma ,\tau ,n,r_0,s_0,M)$ such that for ${ϵ}^{(0)}<{ϵ}_0$ the iterated Hamiltonians $H^{(n)}$ converge in the analytic norm on the strip $r_{\infty }=r_0/2$, $s_{\infty }=s_0/2$ to a Hamiltonian $H_{\infty }$ depending only on $I$, with $\partial_I H_\infty(I^\infty) = \omega^*$.Theiteratedsymplectomorphism$\Phi^{(n)} := \phi^{(0)} \circ \cdots \circ \phi^{(n)}$convergestoareal-analyticsymplectomorphism$\Phi\infty$,and${I = I^\infty}$pullsbacktoaninvarianttorusof$H\epsilon$withlinearflowatfrequency$\omega^$.*

Proof. From the one-step lemma with the chosen ${\delta }_n,{\sigma }_n$, the constant $A_n:=C/({\gamma }^2{\delta }_n^{2(n+\tau )}{\sigma }_n^2)=C{\gamma }^{-2}(4/r_0{)}^{2(n+\tau )}(4/s_0{)}^2\cdot 2^{2(n+2)(n+\tau +1)}$ grows like $2^{2(n+\tau +1)n}$ in $n$. The recursion ${ϵ}^{(n+1)}\le A_n({ϵ}^{(n)}{)}^2$ iterates, after taking logarithms, to $\log {ϵ}^{(n)}\le 2^n\log {ϵ}^{(0)}+{\sum }_{j=0}^{n-1}{2}^{n-1-j}\log A_j$. The sum $\sum 2^{-j-1}\log A_j$ converges since $\log A_j$ grows polynomially in $j$ while $2^{-j}$ decays geometrically. Set $K:=\exp ({\sum }_{j=0}^{\infty }{2}^{-j-1}\log A_j)$; then ${ϵ}^{(n)}\le K^{2^n}({ϵ}^{(0)}{)}^{2^n}$. Provided $K{ϵ}^{(0)}<1/2$, ${ϵ}^{(n)}\to 0$ super-exponentially. Each ${\varphi }^{(n)}$ is close to the identity in the analytic norm (its generator $F^{(n)}$ is bounded by $C{ϵ}^{(n)}/(\gamma {\delta }_n^{n+\tau })$, which is summable in $n$), so ${\Phi }^{(n)}={\varphi }^{(0)}\circ \cdots \circ {\varphi }^{(n)}$ converges in the analytic norm on the strip $r_{\infty },s_{\infty }$ to a real-analytic symplectomorphism ${\Phi }_{\infty }$. The composed sequence $H^{(n)}$ converges to $H_{\infty }(I)$ depending only on actions (since each step removes the angle-dependence of the leading remainder), and the level $\{I=I_{\infty }^{\ast }\}$ is invariant with linear flow at frequency ${\omega }^{\ast }$. $\square$

Theorem (measure of surviving tori). For $ϵ$ small, the set of actions $I\in U$ at which $H_{ϵ}$ has an invariant torus of frequency $\omega (I)$ has Lebesgue measure $\ge (1-C\sqrt{ϵ})\mathrm{vol}(U)$.

Proof. Apply the previous theorem at each $\gamma >0$ for which $ϵ<{ϵ}_0(\gamma )$. The dependence is ${ϵ}_0(\gamma )\asymp {\gamma }^2$ (from the Cauchy/small-divisor estimates compounded in the convergence), so for given $ϵ$ the largest usable $\gamma$ is of order $\sqrt{ϵ}$. The Lebesgue measure of the bad frequency set is $\mathrm{vol}(\omega (U)\setminus D_{\gamma ,\tau })\le C(\tau )\gamma$ by the standard count of resonance strips: each strip $\{|⟨k,\omega ⟩|<\gamma /|k{|}^{\tau }\}$ has width $2\gamma /|k{|}^{\tau +1}$, and summing ${\sum }_{k\ne 0}|k{|}^{-(\tau +1)}$ converges for $\tau >n-1$. Pull this estimate back to action space via the frequency map $\omega :U\to {\mathbb{R}}^n$: by Kolmogorov non-degeneracy, $\omega$ is a local diffeomorphism with bounded Jacobian, so the bad action set has measure $\le C\gamma =C\sqrt{ϵ}$. Multiplying by $\mathrm{vol}({\mathbb{T}}^n)$ for the angular component gives the stated estimate in phase space. $\square$

Connections [Master]

• Integrable system 05.02.03 — the unperturbed half of the KAM setup; Liouville-Arnold produces the foliation by tori that KAM tries to preserve.

• Action-angle coordinates 05.02.04 — the canonical chart in which the KAM scheme is set up; the iteration is intrinsically a sequence of corrections to the action-angle structure, and Whitney-smooth dependence on ${\omega }^{\ast }$ extends action-angle coordinates to the Cantor set of surviving frequencies.

• Symplectic manifold 05.01.02 — the ambient category in which the perturbation problem lives; symplecticity of the iterated transformations ${\varphi }^{(n)}$ is what makes the iteration well-posed, since at each step the generator $F^{(n)}$ produces a Hamiltonian symplectomorphism by construction.

• Generating functions 05.05.03 — the technical mechanism producing each ${\varphi }^{(n)}$; in the original Kolmogorov-Arnold formulation the symplectomorphisms are described through type-II generating functions, while in the modern Lie-series approach they arise as time-one maps of Hamiltonian flows.

• Moser's trick 05.01.05 — the path-method counterpart to the KAM iteration; both produce symplectomorphisms by integrating a time-dependent vector field constructed from a primitive, but Moser's trick works in finite-dimensional cohomology while KAM works in an infinite-dimensional Diophantine-controlled function space.

• Hamiltonian vector field 05.02.01 — the iterated transformations are flows of Hamiltonian vector fields with generators built from the cohomological equation; the Poisson bracket $\{H_0,F\}$ that appears in the linearised equation is the Lie derivative of $H_0$ along $X_F$.

• Poisson bracket 05.02.02 — the cohomological equation $\{H_0,F\}=R-⟨R⟩$ is the fundamental linear operator of the iteration; its Fourier representation as multiplication by $i⟨k,{\omega }^{\ast }⟩$ is the small-divisor problem.

• Arnold conjecture / Floer homology [05.08.01, 05.08.02] — KAM tori are stable invariant submanifolds whose Floer-theoretic count contributes to lower bounds for fixed points of Hamiltonian symplectomorphisms; the Conley-Zehnder index 05.08.04 of a KAM torus encodes its stability type.

• Pseudoholomorphic curve 05.06.02 — the Gromov non-squeezing theorem 05.07.01 is a hard symplectic-rigidity statement complementary to KAM's soft persistence statement; both express the rigidity of symplectic structures, but along orthogonal axes.

The bridge from the analytic input (the cohomological equation with Diophantine bound) to the geometric output (invariant tori for the perturbed Hamiltonian) is the foundational reason KAM unifies a wide circle of results in symplectic dynamics. Putting the connections together, the same Newton-iteration scheme drives normal-form theory near elliptic fixed points, persistence of lower-dimensional invariant tori, and infinite-dimensional analogues for Hamiltonian PDE.

Historical & philosophical context [Master]

Andrey Kolmogorov announced the theorem in a four-page note in 1954: On preservation of conditionally periodic motions under small perturbations of the Hamiltonian, Doklady Akad. Nauk SSSR 98 (1954), 527-530 ^{[Kolmogorov 1954]}. The note states the result and sketches the iteration scheme without giving the analytic estimates needed for a complete proof. Kolmogorov delivered an elaboration at the 1954 International Congress of Mathematicians in Amsterdam; the iteration scheme he described is now recognised as the prototype of Newton's method on a function space and as a precursor to the Nash-Moser implicit function theorem. The 1954 announcement framed the result as an answer to Poincaré's question on the stability of the solar system, formulated in Les Méthodes Nouvelles de la Mécanique Céleste (1892-1899).

Vladimir Arnold's 1963 paper Proof of A. N. Kolmogorov's theorem on preservation of conditionally periodic motions under small perturbation of the Hamiltonian, Russian Math. Surveys 18 (1963), 9-36 ^{[Arnold 1963]}, supplied the analytic estimates and the rigorous convergence argument for the real-analytic case. Arnold also extended the technique in subsequent papers to handle properly degenerate cases (where Kolmogorov's non-degeneracy fails for the unperturbed Hamiltonian but holds for the perturbed time-averaged Hamiltonian) and to apply the theorem to celestial mechanics. The 1963 application to the planar three-body problem established that, on a positive-measure set of initial conditions, the planar restricted three-body problem has perpetual stability — the first rigorous theorem of its kind for a problem of celestial mechanics.

Jürgen Moser's 1962 paper On invariant curves of area-preserving mappings of an annulus, Nachr. Akad. Wiss. Göttingen II (1962), 1-20 ^{[Moser 1962]}, handled the smooth (rather than analytic) case via a Nash-style smoothing scheme. Moser's paper proved persistence for $C^{333}$ twist maps; later refinements (Rüssmann, Pöschel, Salamon) brought the smoothness threshold down to a sharp $C^{2n+\tau }$-type bound. Moser's 1962 introduction of the smoothing-truncation technique into KAM-style iterations was a significant analytic advance and one of the origins of the Nash-Moser implicit function theorem in its modern form.

Subsequent decades have refined every aspect: Pöschel's 1982 paper Integrability of Hamiltonian systems on Cantor sets, Comm. Pure Appl. Math. 35 ^{[Pöschel 1982]}, gave Whitney-smooth dependence on the Diophantine parameter and sharp $C^k$ thresholds; Eliasson, Kuksin, Wayne, and others extended KAM to lower-dimensional tori and to infinite-dimensional Hamiltonian PDE in the 1980s and 1990s; Arnold-Kozlov-Neishtadt's Mathematical Aspects of Classical and Celestial Mechanics ^{[Arnold-Kozlov-Neishtadt]} is the canonical encyclopaedic treatment. The theorem was the central technical input to two of the major successes of twentieth-century dynamical systems: the Arnold-Avez programme of perpetual stability for celestial mechanics, and the Kuksin-Bourgain programme of persistent invariant tori for nonlinear partial differential equations.

Bibliography [Master]

[object Promise]