The Liouville-Arnold Theorem: Integrability and the Existence of Action-Angle Variables

Intuition Beginner

Some mechanical systems are special: they have enough conserved quantities to be solved exactly. The Liouville-Arnold theorem tells you precisely what happens when a system with $n$ degrees of freedom has $n$ independent conserved quantities that all "get along" (a technical condition called Poisson commuting). The answer is geometric: the motion is trapped on surfaces shaped like $n$-dimensional tori.

Think of a spinning top. A symmetric top has two natural conserved quantities: the total energy and the component of angular momentum along the symmetry axis. These two quantities pin the motion to a surface shaped like a 2-torus (a doughnut). The top traces a path on this torus, winding around it in two independent directions at fixed rates. If you know the two conserved quantities, you know which torus the motion lives on. If you know where on the torus the motion starts, you know the entire future.

The Liouville-Arnold theorem generalises this picture to any number of degrees of freedom. When you have $n$ commuting conserved quantities for an $n$-degree-of-freedom system, the level sets (the places where all $n$ quantities take fixed values) are $n$-dimensional tori. On each torus, the motion is a straight-line winding: every coordinate increases at a constant rate, like the hands of $n$ independent clocks ticking simultaneously.

Each torus has its own set of frequencies. A torus for a planetary orbit has one frequency for radial oscillation and another for angular sweep. A torus for a 3D rigid body has three frequencies, one for each independent rotation. The frequencies are determined by which torus you are on (i.e., by the values of the conserved quantities). Different tori generally have different frequencies.

Why tori? A torus is the product of circles: $T^n=S^1\times S^1\times \cdots \times S^1$. Each conserved quantity generates a circular flow, like the motion of a clock hand. When these flows commute (which is what "Poisson commuting" means physically), they combine to give a flow on the product of $n$ circles, which is an $n$-dimensional torus. Compactness of the level set (bounded and closed) is what forces each factor to be a circle rather than a line.

An integrable system, then, is one that is exactly solvable because its phase space is layered into invariant tori, and the motion on each torus is as simple as it can possibly be: $n$ independent clocks ticking at constant rates.

Visual Beginner

Figure 1. A schematic view of the Liouville-Arnold foliation. The $2n$-dimensional phase space is sliced into a family of $n$-dimensional invariant tori, one for each set of values $(F_1,F_2,\dots ,F_n)$ of the commuting integrals. The motion of the system is confined to a single torus, winding around it in $n$ independent directions at frequencies ${\omega }_1,{\omega }_2,\dots ,{\omega }_n$ determined by the values of the integrals.

Imagine phase space as a block of Swiss cheese where each hole is a torus. The system sits on exactly one torus (determined by its initial conditions) and traces a path on that torus forever. If the frequencies are rationally related (e.g., ${\omega }_1/{\omega }_2=p/q$ for integers $p$ and $q$), the path eventually closes and repeats. If not, the path densely fills the entire torus without ever exactly repeating -- this is called quasi-periodic motion.

Worked example Beginner

The planar Kepler problem as a Liouville-integrable system.

A planet of mass $m$ orbiting a star of mass $M$ under Newtonian gravity has two degrees of freedom (the radial coordinate $r$ and the angle $\varphi$). Two independent, commuting integrals suffice for integrability.

The Hamiltonian is $H=p_r^2/(2m)+L^2/(2m{r}^2)-GMm/r$ where $L$ is the angular momentum. The two commuting integrals are:

1. $F_1=H$ (the energy, conserved by time-translation symmetry).
2. $F_2=L$ (the angular momentum, conserved by rotational symmetry).

These commute because the Hamiltonian is rotation-invariant: the angular momentum generates rotations, and rotating a rotation-invariant function changes nothing.

By the Liouville-Arnold theorem, the joint level sets $H=E$ and $L=\ell$ are 2-tori for bound orbits ($E<0$). The planet's motion winds around each torus with two frequencies:

• The radial frequency ${\omega }_r$: how fast the planet oscillates between perihelion (closest approach) and aphelion (farthest point).
• The angular frequency ${\omega }_{\varphi }$: how fast the planet sweeps out angle.

For the Kepler problem, the frequencies are equal: ${\omega }_r={\omega }_{\varphi }$. This means every bound orbit closes exactly -- the planet returns to its starting point after one revolution. This is a special property of the $1/r$ potential and the harmonic oscillator, related to the Bertrand theorem ^{[source pending]}.

A second example: the 2D harmonic oscillator. The Hamiltonian $H=(p_x^2+p_y^2)/(2m)+\frac{1}{2}m({\omega }_x^2{x}^2+{\omega }_y^2{y}^2)$ splits as $H=H_x+H_y$ where $H_x$ is the $x$-oscillator energy and $H_y$ is the $y$-oscillator energy. The two commuting integrals are $F_1=H$ and $F_2=H_x$. Since $H_x$ depends only on $(x,p_x)$ and $H_y$ depends only on $(y,p_y)$, they commute automatically. The level sets $H=E$ and $H_x=E_x$ are 2-tori. The orbit closes when ${\omega }_x/{\omega }_y$ is a ratio of integers (a rational number); otherwise the trajectory fills the torus densely.

Check your understanding Beginner

Formal definition Intermediate+

Let $(M^{2n},\omega )$ be a symplectic manifold with Hamiltonian $H\in C^{\infty }(M)$.

Definition (Liouville integrability). The system is completely integrable in the sense of Liouville if there exist $n$ smooth functions $F_1=H,F_2,\dots ,F_n$ on $M$ such that:

1. Functional independence. The differentials $d{F}_1,d{F}_2,\dots ,d{F}_n$ are linearly independent on a dense open subset of $M$ (equivalently, the $n\times 2n$ matrix $(\partial F_i/\partial x^j)$ has rank $n$ on a dense set).
2. Involution. $\{F_i,F_j\}=0$ for all $1\le i,j\le n$ (the integrals pairwise Poisson-commute).

The level set (or invariant manifold) of the integrals at the value $f=(f_1,\dots ,f_n)$ is

$$M_f=\{x\in M:F_i(x)=f_i,i=1,\dots ,n\}.$$

The map $F=(F_1,\dots ,F_n):M\to {\mathbb{R}}^n$ is called the momentum map. By the involution condition, each $F_i$ is constant along the Hamiltonian flow of every $F_j$ (since ${\dot{F}}_i=\{F_i,F_j\}=0$ along the flow of $F_j$), so each flow preserves every level set $M_f$.

Definition (Liouville torus). If $M_f$ is compact and connected, the Liouville-Arnold theorem (proved below) guarantees that $M_f$ is diffeomorphic to the $n$-torus $T^n$. In this case $M_f$ is called a Liouville torus.

Definition (Action variables). Let ${\gamma }_1,{\gamma }_2,\dots ,{\gamma }_n$ be a basis of independent non-contractible loops on the Liouville torus $M_f$ (a basis of the first homology group $H_1(T^n,\mathbb{Z})\cong {\mathbb{Z}}^n$). The action variables are

$$\boxed{I_i=\frac{1}{2\pi }{\oint }_{{\gamma }_i}\sum _{k=1}^n{p}_{k}d{q}^k=\frac{1}{2\pi }{\oint }_{{\gamma }_i}{p}_{k}d{q}^k.}$$

These depend only on the values $f=(f_1,\dots ,f_n)$ (i.e., on which torus we are on), so they are constants of the motion. The second equality uses the summation convention.

Definition (Angle variables). The conjugate angle variables ${\theta }_1,\dots ,{\theta }_n$ are coordinates on the torus $M_f$ chosen so that $(I_1,\dots ,I_n,{\theta }_1,\dots ,{\theta }_n)$ form a Darboux chart: $\omega =d{\theta }^i\wedge d{I}_i$. They are defined up to additive constants (choice of origin on each circle factor of the torus). The angles evolve linearly: ${\dot{\theta }}_i={\omega }_i(I)$.

Definition (Frequency vector). In action-angle coordinates the Hamiltonian depends only on the actions: $H=H(I)$. The equations of motion become ${\dot{I}}_i=0$ and

$${\omega }_i(I)=\frac{\partial H}{\partial I_i}.$$

This is the frequency vector. Each torus has its own frequency, determined by the action values that label it.

Counterexamples to common slips

• Functional independence is not the same as algebraic independence. The functions $F_1=q^2$, $F_2=q^4$ on ${\mathbb{R}}^2$ are algebraically dependent ($F_2=F_1^2$) but also fail functional independence at $q=0$ because $d{F}_1=2qdq$ and $d{F}_2=4q^{3}dq$ are linearly dependent there. Functional independence requires the differentials to be linearly independent as covectors.

• Having $n$ conserved quantities is not enough; they must commute. The Toda lattice is integrable because its integrals commute. But a generic Hamiltonian system on a $2n$-dimensional phase space can have many conserved quantities that do not commute pairwise, and the Liouville-Arnold theorem does not apply. The commutation condition $\{F_i,F_j\}=0$ is the restrictive hypothesis.

• The level set need not be compact. If $M_f$ is non-compact, it is diffeomorphic to ${\mathbb{R}}^k\times T^{n-k}$ rather than $T^n$, and action-angle variables exist only for the compact directions. The spherical pendulum at high energy has non-compact level sets in the total-energy--angular-momentum plane.

• Action variables are not unique. Different choices of homology basis $\{{\gamma }_i\}$ are related by $\mathrm{GL}(n,\mathbb{Z})$ transformations (integer matrices with determinant $\pm 1$), so the action variables are unique up to unimodular linear combinations.

Key theorem with proof Intermediate+

Theorem (Liouville 1855; Arnold 1963; Mineur 1936; Jost 1959). Let $(M^{2n},\omega )$ be a symplectic manifold with Hamiltonian $H$, and let $F_1=H,F_2,\dots ,F_n$ be $n$ smooth functions in involution that are functionally independent on a dense open set. Let $M_f$ be a compact connected component of the level set $\{F_i=f_i\}$. Then:

1. $M_f$ is diffeomorphic to the $n$-torus $T^n$.
2. The Hamiltonian flows of $F_1,\dots ,F_n$ are complete on $M_f$ and define a transitive action of ${\mathbb{R}}^n$ by translations, with a discrete stabiliser isomorphic to ${\mathbb{Z}}^n$.
3. In a neighbourhood of $M_f$, there exist action-angle variables $(I,\theta )$ in which $\omega =d{\theta }^i\wedge d{I}_i$ and the equations of motion are $\dot{I}=0$, $\dot{\theta }=\omega (I)$.
4. $M_f$ is a Lagrangian submanifold: the restriction $\omega {|}_{M_f}=0$.

Proof. The argument proceeds in four stages.

Stage 1: The involution condition implies the flows commute. Since $\{F_i,F_j\}=0$, the Hamiltonian vector fields $X_{F_i}$ and $X_{F_j}$ satisfy $[X_{F_i},X_{F_j}]=X_{\{F_i,F_j\}}=0$. The commuting vector fields $X_{F_1},\dots ,X_{F_n}$ define an action of ${\mathbb{R}}^n$ on $M$ by translations: ${\Phi }_t(x)=g_{F_1}^{t_1}\circ \cdots \circ g_{F_n}^{t_n}(x)$, independent of the order of composition.

Stage 2: Compact level sets are tori. Each flow $g_{F_i}^{t_i}$ preserves $M_f$ (because $F_j$ is constant along the flow of $F_i$). Restricted to the compact $M_f$, each flow generates a closed orbit -- compactness prevents escape to infinity. The stabiliser of any point $x\in M_f$ is the set of "period vectors" $t=(t_1,\dots ,t_n)$ that return $x$ to itself. This stabiliser is a lattice $\Lambda \cong {\mathbb{Z}}^n\subset {\mathbb{R}}^n$. Therefore $M_f\cong {\mathbb{R}}^n/\Lambda \cong T^n$. The linear independence of $d{F}_i$ ensures the ${\mathbb{R}}^n$-action is free, making $\Lambda$ discrete of rank exactly $n$ ^{[source pending]}.

Stage 3: Construction of action variables. Choose a basis of cycles ${\gamma }_1,\dots ,{\gamma }_n$ for $H_1(M_f,\mathbb{Z})\cong {\mathbb{Z}}^n$. Define

$$I_i=\frac{1}{2\pi }{\oint }_{{\gamma }_i}{p}_{k}d{q}^k.$$

Because $\omega =d{p}_k\wedge d{q}^k$ is closed and $M_f$ is Lagrangian (proved below), the integral depends only on the homology class of ${\gamma }_i$ and on the values $f$ -- not on the specific representative curve. The map $f\mapsto I$ is locally invertible by the implicit function theorem (since $d{F}_i$ are linearly independent), so the $I_i$ are well-defined functions of the integrals and hence constants of the motion.

Stage 4: Angle variables and the Darboux chart. Fix a reference point $x_0\in M_f$. Define angle variables by integrating the canonical 1-form $\alpha =p_{k}d{q}^k$ along paths on $M_f$: the angle ${\theta }_i(x)$ measures the displacement along the $i$-th flow from $x_0$ to $x$. The $2\pi$-periodicity follows from the lattice $\Lambda$: traversing a full period adds $2\pi I_i$ to the integral, and dividing by $I_i$ gives ${\theta }_i\mapsto {\theta }_i+2\pi$. The symplectic form in the new coordinates is $\omega =d{\theta }^i\wedge d{I}_i$, verified by computing the Poisson brackets $\{I_i,{\theta }_j\}={\delta }_{ij}$, $\{I_i,I_j\}=0$, $\{{\theta }_i,{\theta }_j\}=0$. This confirms $(I,\theta )$ is a Darboux chart.

That $M_f$ is Lagrangian: the tangent space $T_x{M}_f=\mathrm{span}\{X_{F_1},\dots ,X_{F_n}\}$ (the annihilator of $d{F}_1,\dots ,d{F}_n$). For any tangent vectors $v=\sum a_i{X}_{F_i}$ and $w=\sum b_j{X}_{F_j}$:

$$\omega (v,w)=\sum _{i,j}{a}_i{b}_j\omega (X_{F_i},X_{F_j})=\sum _{i,j}{a}_i{b}_j\{F_j,F_i\}=0$$

by the involution hypothesis. So $\omega {|}_{M_f}=0$ and $M_f$ is isotropic. Since $\dim M_f=n=\frac{1}{2}\dim M$, it is Lagrangian. $\square$

The existence of action-angle coordinates is local: it holds in a tubular neighbourhood of $M_f$. Global existence on all of $M$ requires further conditions (see the Master section on Duistermaat's theorem).

Worked example: the 1D harmonic oscillator

For the 1D oscillator $H=p^2/(2m)+\frac{1}{2}m{\omega }^2{q}^2$ (one degree of freedom, $n=1$), the Liouville-Arnold theorem says the compact level sets $H=E$ are 1-tori (circles). There is one action variable:

$$I=\frac{1}{2\pi }\oint pdq.$$

On the ellipse $H=E$ in the $(q,p)$ plane, the enclosed area is $\pi \cdot \sqrt{2mE}\cdot \sqrt{2E/(m{\omega }^2)}=2\pi E/\omega$, so $I=E/\omega$. Inverting: $H=\omega I$. The frequency is $\dot{\theta }=dH/dI=\omega$, constant and independent of $I$. This is a special property of the harmonic oscillator: all tori have the same frequency.

In angle variables, $\theta =\omega t+{\theta }_0$, and the original coordinates are recovered by $q=\sqrt{2I/(m\omega )}\cos \theta$, $p=-\sqrt{2m\omega I}\sin \theta$.

Exercises Intermediate+

Lean formalization Intermediate+

lean_status: none. Mathlib has symplectic linear algebra (LinearAlgebra.SymplecticGroup, bilinear-form-skew-symmetric machinery) and smooth manifold theory (Geometry.Manifold, tangent and cotangent bundles as smooth bundles). What is missing is:

1. Hamiltonian vector fields on manifolds. The correspondence $F\mapsto X_F$ defined by $\omega (X_F,\cdot )=dF$ requires contraction of a vector field with a 2-form on a smooth manifold, not yet available in usable form.

2. Commuting flows and the period lattice. The proof that compact level sets of commuting flows are tori requires the orbit-stabiliser theorem for the ${\mathbb{R}}^n$-action and identification of the stabiliser with a lattice $\Lambda \cong {\mathbb{Z}}^n$.

3. Homology and the action integral. The definition of action variables as integrals over basis cycles requires $H_1(T^n,\mathbb{Z})$ and Stokes' theorem on the Lagrangian level set to show the integral depends only on the homology class.

4. Darboux verification. Proving $(I,\theta )$ are Darboux coordinates requires computing Poisson brackets in the new coordinates and verifying the canonical relations.

The single most impactful formalization target would be the Caratheodory-Jacobi-Lie theorem (Exercise 7), which is a local result requiring only Darboux machinery closer to Mathlib's current capabilities. This unit ships without a Lean module and is reviewer-attested.

Advanced results Master

Arnold's theorem on the full phase space

The Liouville-Arnold theorem as stated above guarantees action-angle variables only in a tubular neighbourhood of a single torus. Arnold's original formulation addresses the global picture: if the momentum map $F:M\to {\mathbb{R}}^n$ is a proper submersion (so all level sets are compact and the map has no critical values), then $M$ is diffeomorphic to $T^n\times B$ where $B\subset {\mathbb{R}}^n$ is an open set (the image of the momentum map), and global action-angle coordinates exist on all of $M$ ^{[source pending]}.

When the momentum map has critical values (degenerate tori), the global picture is richer. The singular fibres where the rank of $dF$ drops are not tori but more general manifolds (pinched tori, cylindrical surfaces, etc.). The topology of the singular fibres and their relationship to the regular fibres is the subject of singularity theory for integrable systems, developed by Fomenko and collaborators ^{[source pending]}.

Topological obstructions and monodromy

Duistermaat (1980) identified three obstructions to the global existence of action-angle coordinates ^{[source pending]}:

1. Monodromy. As one loops around a singular fibre in the base $B$ (the space of integral values), the basis cycles ${\gamma }_1,\dots ,{\gamma }_n$ may undergo a linear transformation ${\gamma }_i\mapsto {\sum }_j{M}_{ij}{\gamma }_j$ with $M\in \mathrm{GL}(n,\mathbb{Z})$. This monodromy matrix measures the twisting of the torus fibration. If $M\ne 1$, global action-angle coordinates do not exist. The spherical pendulum has monodromy $M=\left(\begin{array}{cc}1& 1\\ 0& 1\end{array}\right)$, the simplest non-identity case ^{[source pending]}.

2. Nonzero Chern class. The Lagrangian torus bundle may have a non-vanishing Chern class in $H^2(B,\mathbb{Z})$, obstructing the existence of global angle variables even when the action variables are globally defined.

3. Caustics. The singular fibres (caustics) where the rank of the momentum map drops form a discriminant locus in the base. The topology of this locus constrains the global properties of the fibration. Near a focus-focus singularity (the generic type for two degrees of freedom), the monodromy is never the identity.

Semi-local normal form

Near a non-degenerate singular fibre, the integrable system admits a Williamson normal form that classifies the local geometry. For an equilibrium point (a torus of dimension zero), the Williamson theorem (1936) classifies linear Hamiltonian systems into elliptic, hyperbolic, and focus-focus components. Each component contributes a model normal form:

• Elliptic: $H=\omega (I_1+I_2)/2$ (a pair of harmonic oscillators).
• Hyperbolic: $H=\lambda I$ (a saddle in the action variable).
• Focus-focus: $H=\delta (I_1\log r-I_2\theta )$ in appropriate polar coordinates.

The semi-local normal form theorem (Eliasson, 1984; Miranda and Vu Ngoc, 2007) extends Williamson's linear classification to a nonlinear normal form near singular fibres, providing action-angle coordinates in a punctured neighbourhood of the singularity ^{[source pending]}. This is the bridge between the local Darboux theorem and the global torus fibration.

Connections to KAM theory

The Kolmogorov-Arnold-Moser (KAM) theorem 09.08.01 asks: when an integrable Hamiltonian $H_0(I)$ is perturbed to $H_0(I)+ϵH_1(I,\theta )$, do the Liouville tori survive? The answer depends on the frequencies. Tori with "sufficiently irrational" frequencies -- satisfying a Diophantine condition $|k\cdot \omega |>c|k{|}^{-\tau }$ for all $k\in {\mathbb{Z}}^n\setminus \{0\}$ and suitable constants $c,\tau >0$ -- persist as invariant tori of the perturbed system, slightly deformed but still carrying quasi-periodic motion ^{[source pending]}.

The non-degeneracy conditions defined in this unit determine whether enough tori satisfy the Diophantine condition:

• Kolmogorov condition ($\det ({\partial }^{2}H/\partial I_i\partial I_j)\ne 0$): ensures the frequency map $I\mapsto \omega$ is a local diffeomorphism, so a positive measure set of tori has Diophantine frequencies.
• Arnold (isoenergetic) condition: ensures persistence at fixed energy, a weaker requirement that applies to systems like Kepler where the Hessian degenerates.
• Russmann condition: if the image of the frequency map does not lie in any hyperplane through the origin, then a positive measure of tori persists. This is the weakest of the standard conditions ^{[source pending]}.

The Mishchenko-Fomenko theorem (non-compact level sets)

When $M_f$ is non-compact, it is diffeomorphic to ${\mathbb{R}}^k\times T^{n-k}$. The Mishchenko-Fomenko theorem extends the Liouville-Arnold framework: the commuting flows still define a translational action, and action-angle variables exist locally, but the angle variables corresponding to the non-compact directions are unbounded. The theorem also covers the case of non-commutative integrability, where the integrals form a Lie algebra rather than all commuting, and the level sets are orbits of a Lie group action ^{[source pending]}.

Nekhoroshev's theorem on exponential stability

Even when the tori break up under perturbation, Nekhoroshev (1977) proved that the action variables remain nearly constant for exponentially long times: $|I(t)-I(0)|<C{ϵ}^a$ for $|t|<T_0\exp (({ϵ}_0/ϵ{)}^b)$, where $ϵ$ is the perturbation size and $a,b$ depend on the number of degrees of freedom and a steepness condition on $H(I)$. This is stronger than KAM (which only guarantees that some tori persist) but weaker in that it allows slow drift rather than exact conservation ^{[source pending]}.

The Atiyah-Guillemin-Sternberg convexity theorem

For Hamiltonian group actions, the Liouville-Arnold framework has a convexity refinement. The Atiyah (1982) and Guillemin-Sternberg (1982) theorem states: if a torus $T^k$ acts on a compact symplectic manifold $(M^{2n},\omega )$ in a Hamiltonian fashion with momentum map $\mu :M\to {\mathbb{R}}^k$, then $\mu (M)$ is a convex polytope in ${\mathbb{R}}^k$. This is the symplectic counterpart of the Schur-Horn theorem in linear algebra and connects integrable systems to toric geometry: the polytope determines the symplectic manifold up to equivariant symplectomorphism (Delzant's theorem, 1988) ^{[source pending]}.

Synthesis. The Liouville-Arnold theorem is the structural centrepiece of integrable Hamiltonian mechanics: it converts $n$ algebraic conditions (commuting integrals) into a geometric picture (torus foliation with action-angle coordinates). The non-degeneracy conditions bridge to perturbation theory via KAM; the global obstructions (monodromy, Chern class) show that local integrability does not guarantee global simplicity; and the convexity theorems for group actions connect the integrable picture to toric geometry and representation theory. The entire framework rests on three pillars: the symplectic structure 09.04.01, the Poisson bracket 09.04.02, and the Darboux theorem that guarantees the existence of canonical coordinates. Cross-reference the dynamics spine [38.07] for the rigorous global proof infrastructure.

Full proof set Master

Proposition 1 (The period lattice is discrete of rank $n$). Let $X_{F_1},\dots ,X_{F_n}$ be $n$ commuting Hamiltonian vector fields, linearly independent on the compact level set $M_f$. Then the stabiliser $\Lambda =\{t\in {\mathbb{R}}^n:{\Phi }_t(x)=x\}$ is a lattice of rank $n$ in ${\mathbb{R}}^n$.

Proof. The map $\Phi :{\mathbb{R}}^n\times M_f\to M_f$ given by $\Phi (t,x)=g_{F_1}^{t_1}\circ \cdots \circ g_{F_n}^{t_n}(x)$ is a smooth group action because the flows commute. The linear independence of $X_{F_i}$ means the derivative $d{\Phi }_0:{\mathbb{R}}^n\to T_x{M}_f$ is an isomorphism, so the action is locally free. For any $x\in M_f$, the stabiliser ${\Lambda }_x=\{t:{\Phi }_t(x)=x\}$ is a closed subgroup of ${\mathbb{R}}^n$ (it is the preimage of the closed set $\{x\}$ under the continuous map $t\mapsto {\Phi }_t(x)$). By the classification of closed subgroups of ${\mathbb{R}}^n$, ${\Lambda }_x$ is isomorphic to ${\mathbb{R}}^j\times {\mathbb{Z}}^k$ for some $j+k\le n$. Local freeness (the derivative is an isomorphism) excludes the ${\mathbb{R}}^j$ component: the exponential map is injective in a neighbourhood of $0$, so no one-parameter family of non-identity elements can fix $x$. Hence $j=0$ and ${\Lambda }_x\cong {\mathbb{Z}}^k$ with $k\le n$.

Compactness forces $k=n$. If $k<n$, the quotient ${\mathbb{R}}^n/{\Lambda }_x\cong {\mathbb{R}}^{n-k}\times T^k$ is non-compact, but the orbit map ${\mathbb{R}}^n/{\Lambda }_x\to M_f$ is a continuous bijection with continuous inverse (by compactness of $M_f$ and local freeness), contradicting non-compactness of the quotient. Therefore $k=n$, ${\Lambda }_x\cong {\mathbb{Z}}^n$, and $M_f\cong {\mathbb{R}}^n/{\mathbb{Z}}^n\cong T^n$. $\square$

Proposition 2 (The action integrals are well-defined). The integral $I_i=(2\pi {)}^{-1}{\oint }_{{\gamma }_i}{p}_{k}d{q}^k$ depends only on the homology class $[{\gamma }_i]\in H_1(M_f,\mathbb{Z})$ and the values $f=(f_1,\dots ,f_n)$, not on the specific representative curve.

Proof. Let $\gamma$ and ${\gamma }^{\prime }$ be two loops in the same homology class on $M_f$. Then $\gamma -{\gamma }^{\prime }=\partial S$ for some 2-chain $S$ in $M_f$. By Stokes' theorem:

$${\oint }_{\gamma }{p}_{k}d{q}^k-{\oint }_{{\gamma }^{\prime }}{p}_{k}d{q}^k={\int }_{S}d{p}_k\wedge d{q}^k={\int }_S\omega .$$

Since $M_f$ is Lagrangian ($\omega {|}_{M_f}=0$), the right-hand side vanishes. Hence the integral depends only on the homology class. Dependence on $f$ alone follows because the cycles ${\gamma }_i$ are determined by the topology of the torus $M_f$, and the 1-form $p_{k}d{q}^k$ varies smoothly with $f$. $\square$

Proposition 3 (The action-angle map is a symplectomorphism). Let $(I,\theta )$ be the action-angle coordinates constructed in the Liouville-Arnold theorem. Then $\omega =d{\theta }^i\wedge d{I}_i$ in these coordinates.

Proof. The canonical 1-form $\alpha =p_{k}d{q}^k$ satisfies $d\alpha =-\omega$. In the $(I,\theta )$ coordinates, the 1-form $\alpha$ restricts to $p_{k}d{q}^k=I_{i}d{\theta }^i$ on each torus (by construction of $\theta$ as the displacement coordinate). In the tubular neighbourhood, the full 1-form is $\alpha =I_{i}d{\theta }^i$ (the $d{I}_i$ components vanish because $\alpha$ evaluates on a vector tangent to the torus as $I_i$ times the angular displacement, and on a vector transverse to the torus as zero by the definition of $\theta$ as the fibre coordinate). Therefore $\omega =-d\alpha =d{\theta }^i\wedge d{I}_i$. Alternatively, verify the Poisson brackets: $\{I_i,I_j\}=0$ (proven in Exercise 4), $\{I_i,{\theta }_j\}={\delta }_{ij}$ (by construction, since ${\theta }_j$ is the coordinate conjugate to $I_j$), and $\{{\theta }_i,{\theta }_j\}=0$ (because the angle coordinates are coordinates on a single torus and the symplectic form vanishes there). These three relations are equivalent to $\omega =d{\theta }^i\wedge d{I}_i$. $\square$

Proposition 4 (The frequency vector is smooth). If $H$ is smooth and the Kolmogorov condition holds at a torus, then the frequency map $I\mapsto \omega (I)={\nabla }_{I}H(I)$ is a local diffeomorphism.

Proof. The Jacobian of the frequency map is the Hessian ${\partial }^{2}H/\partial I_i\partial I_j$. The Kolmogorov condition is the statement that this Hessian is non-degenerate. By the inverse function theorem, the map $I\mapsto \omega$ is a local diffeomorphism. This means nearby tori have genuinely different frequencies: the frequency map is locally injective, so the set of tori with Diophantine frequencies has positive measure in action space. $\square$

Connections Master

• Hamiltonian mechanics [09.04.01, 09.04.02] -- the Legendre transform produces the Hamiltonian $H$ and Hamilton's equations that the Liouville-Arnold theorem takes as its starting point. The symplectic structure $\omega =d{q}^i\wedge d{p}_i$ is the geometric arena in which the invariant tori live.

• Action-angle variables 09.06.01 -- this unit provides the general existence theorem (Liouville-Arnold) that justifies the action-angle construction introduced in 09.06.01. The earlier unit develops the computational machinery; this unit proves it works.

• KAM theorem 09.08.01 -- the Kolmogorov-Arnold-Moser theorem describes the persistence of Liouville tori under small perturbations. The non-degeneracy conditions (Kolmogorov, Arnold, isoenergetic, Russmann) defined here are precisely the hypotheses that KAM requires.

• Dynamics spine [38.07] -- the dynamics spine provides the rigorous proof infrastructure (stable manifold theorem, KAM persistence theorems, twist-map machinery) that underpins the Liouville-Arnold framework. The semi-local normal form proved here connects directly to the global dynamical-systems perspective.

• Completely integrable systems 09.06.03 -- the Toda lattice, Kepler problem, and other concrete integrable systems provide explicit realisations of the abstract Liouville-Arnold framework. Each system's independent integrals define Liouville tori, and the action-angle construction can be carried out explicitly.

• Quantum mechanics: Bohr-Sommerfeld quantisation -- Einstein (1911) observed that the action variables provide the natural quantisation conditions $I_i=n_i\hslash$ for the old quantum theory. The EBK (Einstein-Brillouin-Keller) quantisation uses Maslov-index-corrected conditions. This was one of the earliest hints that integrability is a fundamental structural property, not merely a convenience.

• Integrable systems in infinite dimensions -- the Liouville-Arnold framework extends formally to PDEs with a Hamiltonian structure (KdV, nonlinear Schrodinger equation, Toda lattice). The analogue of the Liouville torus is the space of scattering data, and the inverse scattering transform plays the role of the action-angle map.

• Toric geometry -- the Atiyah-Guillemin-Sternberg convexity theorem and Delzant's theorem connect the Liouville-Arnold picture to toric geometry: the image of the momentum map is a convex polytope that determines the symplectic manifold. This is the geometric foundation of mirror symmetry and toric degenerations in algebraic geometry.

Historical and philosophical context Master

Joseph Liouville (1855) first observed that $n$ commuting integrals imply the existence of invariant tori, in the context of the Jacobi method for integrating Hamilton's equations ^{[source pending]}. His result was stated in local coordinates and did not emphasise the global torus structure. Liouville's contribution was the recognition that the involution condition produces a family of invariant manifolds on which the dynamics simplify.

Henri Mineur (1936) independently proved a version of the theorem that explicitly constructs the angle variables and identifies the torus structure, in a paper that was largely overlooked until recently ^{[source pending]}. Mineur's construction of the period lattice and the action integrals is essentially the one used in modern presentations.

Res Jost (1959) gave a clean formulation that highlights the geometric content: the commuting flows define a transitive ${\mathbb{R}}^n$-action on the compact level set, and the quotient by the period lattice is the torus ^{[source pending]}. Jost's version is sometimes called the Liouville-Arnold-Jost theorem in recognition of the three independent discoveries.

Vladimir Arnold (1963) gave the modern formulation in his landmark paper on the stability of Hamiltonian systems, proving the global torus diffeomorphism and constructing the action-angle variables on a tubular neighbourhood ^{[source pending]}. Arnold's proof is the one presented in the Key theorem section above, using the period lattice of the commuting flows. Arnold also identified the connection to perturbation theory that led to the KAM theorem.

Albert Einstein (1911) recognised the significance of the action variables for quantum theory ^{[source pending]}. His observation that the Bohr-Sommerfeld conditions only make sense for integrable systems was prescient: it identified integrability as a fundamental physical property, not merely a mathematical convenience. This insight lay dormant until the development of KAM theory and the study of classical chaos brought the distinction between integrable and non-integrable systems to the forefront of physics.

Johannes Duistermaat (1980) opened the study of global obstructions to action-angle coordinates, introducing the monodromy of Lagrangian fibrations and showing that the spherical pendulum exhibits non-identity monodromy ^{[source pending]}. This result revealed that the local picture guaranteed by Liouville-Arnold can fail globally even in very simple systems.

Nikolai Nekhoroshev (1972, 1977) developed the theory of action-angle variables for non-compact level sets and proved his exponential-stability theorem for perturbed integrable systems, bridging the gap between exact conservation in integrable systems and the breakdown of tori in chaotic systems ^{[source pending]}.

A philosophical point: the Liouville-Arnold theorem reveals that integrability is a geometric property, not merely an algebraic one. Having $n$ commuting integrals is an algebraic condition, but the consequence -- torus foliation and action-angle coordinates -- is geometric. This mirrors the philosophy of symplectic geometry: the algebraic structure (Poisson bracket, involution) determines the geometric structure (torus foliation, Lagrangian fibration), and vice versa. The theorem is a paradigmatic example of the algebra-geometry correspondence that pervades modern mathematical physics.

Bibliography Master

• Arnold, V. I. "Proof of a theorem of A. N. Kolmogorov on the preservation of conditionally periodic motions under a small change of the Hamiltonian." Uspekhi Mat. Nauk 18 (1963), 13-40. ^{[source pending]}

• Arnold, V. I. Mathematical Methods of Classical Mechanics, 2nd ed. Springer GTM 60, 1989. §49-50, Appendices 6-7. ^{[source pending]}

• Goldstein, H., Poole, C., and Safko, J. Classical Mechanics, 3rd ed. Pearson, 2002. Ch. 10. ^{[source pending]}

• Landau, L. D. and Lifshitz, E. M. Mechanics, 3rd ed. Pergamon, 1976. §49. ^{[source pending]}

• Liouville, J. "Sur l'integration des equations differentielles de la dynamique." J. Math. Pures Appl. 20 (1855), 137-138. ^{[source pending]}

• Mineur, H. "Sur les systemes conservatifs dans lesquels le mouvement est rapporte a des variables fixes par des relations homogenes." J. Math. Pures Appl. 25 (1936), 93-126. ^{[source pending]}

• Jost, R. "Eine Bemerkung zum Liouvilleschen Satz." Helvetica Phys. Acta 32 (1959), 117-121. ^{[source pending]}

• Einstein, A. "Zum Quantensatz von Sommerfeld und Epstein." Ann. Phys. 339 (1911), 91-107. ^{[source pending]}

• Eliasson, L. H. "Normal forms for Hamiltonian systems with Poisson commuting integrals." Preprint, 1984. ^{[source pending]}

• Duistermaat, J. J. "On global action-angle coordinates." Comm. Pure Appl. Math. 33 (1980), 131-168. ^{[source pending]}

• Nekhoroshev, N. N. "Action-angle variables and their generalizations." Russian Math. Surveys 27 (1972), 65-121. ^{[source pending]}

• Nekhoroshev, N. N. "An exponential estimate of the time of stability of nearly integrable Hamiltonian systems." Uspekhi Mat. Nauk 32 (1977), 5-66. ^{[source pending]}

• Fomenko, A. T. Integrability and Nonintegrability in Geometry and Mechanics. Kluwer, 1988. ^{[source pending]}

• Miranda, E. and Vu Ngoc, S. "A singular Poincare lemma." Int. Math. Res. Not. 2007 (2007), rnm015. ^{[source pending]}

• Weinstein, A. "Symplectic manifolds and their Lagrangian submanifolds." Adv. Math. 6 (1971), 329-346. ^{[source pending]}

• Williamson, J. "On the algebraic problem concerning the normal forms of linear dynamical systems." Amer. J. Math. 58 (1936), 141-163. ^{[source pending]}

• Atiyah, M. F. "Convexity and commuting Hamiltonians." Bull. London Math. Soc. 14 (1982), 1-15. ^{[source pending]}

• Guillemin, V. and Sternberg, S. "Convexity properties of the moment mapping." Invent. Math. 67 (1982), 491-513. ^{[source pending]}