Liouville-Arnold Theorem

Liouville Integrable Systems set up a promise: find $n$ independent conserved quantities in involution and the dynamics will surrender completely. The Liouville-Arnold theorem is the payment of that promise, and it pays in a currency Liouville himself never imagined. Liouville's 1855 result was analytic — the equations of motion can be solved by quadratures, by evaluating integrals alone. Arnold's 1963 contribution was to see what the solutions are: every compact connected level set of the commuting integrals is an $n$ -dimensional torus, phase space is foliated by these tori like nested doughnuts, and on each one the motion is uniform linear winding in suitable angle coordinates. Algebra in, geometry out. The theorem earns its keep three times over. It is the existence theorem for action-angle variables — the canonical chart $(I,\phi)$ with $H=H(I)$ that every treatment of adiabatic invariants and canonical perturbation theory takes for granted. It is the launching pad for KAM theory, which asks what becomes of the tori under perturbation and answers with the deepest theorem of classical mechanics. And it is the skeleton of semiclassical quantization: the tori it constructs are precisely the objects on which Bohr-Sommerfeld and EBK quantization impose $\oint p\,\mathrm{d}q = 2\pi\hbar(n+\mu/4)$ . This article states the theorem precisely, walks the proof's architecture — commuting flows, the $\mathbb{R}^n$ action, the lattice quotient — and shows why the level sets are Lagrangian, which is the fact that makes the action integrals well-defined. The constructions it licenses are detailed in Action Integrals over Invariant Tori and Frequencies from Action-Angle Variables; here we prove the license itself.

Key takeaways

Theorem (Liouville-Arnold): given $n$ functionally independent integrals $F_1,\dots,F_n$ in involution on a $2n$ -dimensional phase space, every compact connected common level set $M_f$ is diffeomorphic to the torus $\mathbb{T}^n$ , and a neighbourhood of it admits canonical action-angle coordinates $(I,\phi)$ with $H=H(I)$ .
Proof architecture: involution makes the $n$ Hamiltonian flows commute → they define an $\mathbb{R}^n$ action on $M_f$ → transitivity plus compactness force $M_f\cong\mathbb{R}^n/\Lambda$ with $\Lambda$ a full-rank lattice — a torus.
The level sets are Lagrangian submanifolds ( $\omega$ restricts to zero), so $p\,\mathrm{d}q$ is closed on them and the actions $I_i=\frac{1}{2\pi}\oint_{\gamma_i}p\,\mathrm{d}q$ depend only on the homology class of the cycle — well-defined torus labels.
On each torus the flow is linear: $\dot\phi_i=\omega_i(I)=\partial H/\partial I_i$ , $\dot I_i=0$ — conditionally periodic motion, closed when frequencies are commensurate, dense when not.
The conclusion is local-in-actions but not automatically global: separatrices pinch the tori, non-compact level sets give cylinders, and Hamiltonian monodromy can obstruct one global action-angle chart.
The theorem is the unperturbed scaffold of KAM theory and the geometric object quantized by Bohr-Sommerfeld/EBK rules.
Key idea: $n$ independent integrals in involution force compact level sets to be $n$ -tori carrying linear (conditionally periodic) flow; the period integrals of $p\,\mathrm{d}q$ supply canonical actions $I$ with $H=H(I)$ and frequencies $\omega_i=\partial H/\partial I_i$ .

Where this lives today

The theorem's tori are the reference objects of modern dynamics. KAM theory — Kolmogorov, Arnold, Moser — answers the question the theorem begs: perturb $H(I)$ to $H(I)+\varepsilon H_1(I,\phi)$ and the tori with Diophantine frequency vectors survive as slightly deformed invariant tori, while resonant ones shatter into island chains and stochastic layers. That single result underwrites the 100-million-year stability estimates of planetary orbits, the confinement analysis of particle accelerators, and the design of storage rings, where tune ratios (frequency ratios of the machine's effective torus) are deliberately parked far from low-order resonances. The failure modes are just as productive. Monodromy, found by Duistermaat as a topological footnote in 1980, turned out to be observable: the quantum joint spectrum of integrable molecules (water's bending mode, CO₂'s Fermi dyad) shows a lattice defect exactly where classical monodromy sits, a discovery that made "quantum monodromy" a working tool of molecular spectroscopy in the 2000s. And the theorem remains the gateway to semiclassics: EBK quantization selects the tori with $I_i=\hbar(n_i+\mu_i/4)$ , which is meaningful only because the Liouville-Arnold theorem supplies tori to select — for chaotic systems, with no tori, quantization must fall back on Gutzwiller's periodic orbits. The dividing line drawn by this theorem thus runs straight through the middle of quantum mechanics: integrable spectra obey Poisson statistics, chaotic ones random-matrix statistics, and the Liouville-Arnold tori are what the word "integrable" means in that sentence.

The physical picture

The theorem is best seen as an assembly line with three stations. Station one converts algebra to motion: each integral $F_i$ generates a flow, and involution — proven equivalent to commuting flows in Liouville Integrable Systems — means the $n$ flows can be run in any order. Station two converts motion to topology: run flow 1 for time $t_1$ , flow 2 for $t_2$ , and so on; because order is irrelevant, the recipe " $(t_1,\dots,t_n)$ " is a well-defined displacement, and the level set becomes a canvas painted completely by $\mathbb{R}^n$ . Station three converts topology to a torus: on a compact canvas the painting must overlap itself, and the set of recipes that return you to the start forms a lattice $\Lambda$ — the level set is $\mathbb{R}^n$ rolled up by $\Lambda$ , which is precisely $\mathbb{T}^n$ .

Figure

The proof pipeline: involution becomes commuting flows, an abelian group action, and finally a lattice quotient — the torus

Why must the answer be a torus and nothing else? Because "compact $n$ -manifold on which $\mathbb{R}^n$ acts transitively by commuting flows" is an extraordinarily tight specification. The sphere fails at the first hurdle — it cannot carry even one nonvanishing vector field. Higher-genus surfaces carry fields but not commuting, everywhere-independent families. Only the quotient of $\mathbb{R}^n$ by a lattice survives, and that quotient is the torus in every dimension. Zoom back out to full phase space and turn the knob values $f=(f_1,\dots,f_n)$ : the tori stack into an $n$ -parameter family filling a $2n$ -dimensional region — a foliation, like the layers of an onion but $n$ -dimensional in a $2n$ -dimensional space. The actions $I_i$ are the natural coordinates across the layers (which torus), the angles $\phi_i$ the coordinates along them (where on the torus). Half the coordinates label, the other half wind; the Hamiltonian sees only the labels. That split is the entire dynamical content of integrability.

Figure

Phase space foliated by invariant tori: actions select the torus, angles locate the point, the trajectory winds at constant rates

The theorem, stated precisely

Fix a $2n$ -dimensional symplectic phase space $M$ and suppose given $n$ smooth functions $F_1,\dots,F_n$ (with $F_1=H$ , say) satisfying the involution hypothesis — the defining condition analysed at length in Liouville Integrable Systems:

\{F_i,F_j\}=0 \quad\text{for all } i,j=1,\dots,n \qquad (1)

where: $\{\cdot,\cdot\}$ — the Poisson bracket. Together with functional independence ( $\mathrm{d}F_1\wedge\cdots\wedge\mathrm{d}F_n\neq0$ on the level set), this is the complete hypothesis — nothing about separability, symmetry, or explicit solvability is assumed. The object of study is the common level set

M_f=\{x\in M:\ F_i(x)=f_i,\ i=1,\dots,n\} \qquad (2)

where: $f=(f_1,\dots,f_n)$ — the frozen values of the integrals; by independence, $M_f$ is a smooth $n$ -dimensional submanifold, invariant under all $n$ flows. The theorem then asserts three things in escalating strength. Topology: if $M_f$ is compact and connected,

M_f\ \cong\ \mathbb{T}^n=\mathbb{R}^n/(2\pi\mathbb{Z})^n \qquad (3)

where: $\mathbb{T}^n$ — the $n$ -torus. Dynamics: on $M_f$ the flow of $H$ is conjugate to a linear flow, $\phi_i(t)=\phi_i(0)+\omega_i t$ with constant $\omega_i$ . Canonical structure: in a whole neighbourhood of $M_f$ (a union of nearby tori) there exist canonical coordinates $(I,\phi)$ — actions and angles — with

I_i=\frac{1}{2\pi}\oint_{\gamma_i}\sum_k p_k\,\mathrm{d}q_k,\qquad H=H(I_1,\dots,I_n)\ \text{only} \qquad (4)

where: $\gamma_i$ — a basis of the $n$ independent cycles of the torus; the actions label the tori and the conjugate angles wind around them. The explicit construction of these integrals, cycle by cycle for real systems, is the subject of Action Integrals over Invariant Tori; the generating-function route from $(q,p)$ to $(I,\phi)$ via Hamilton's characteristic function is built in Action-Angle Variables and Hamilton's Characteristic Function. What those articles assume — and this one proves — is that the torus and the well-defined loop integrals exist at all.

The proof architecture

Derivation AThe torus theorem from commuting flows

We prove the topological heart of the theorem: a compact connected common level set of $n$ independent involutive integrals is diffeomorphic to $\mathbb{T}^n$ . The strategy is to let the $n$ commuting Hamiltonian flows define an action of the abelian group $\mathbb{R}^n$ on $M_f$ , show the action is transitive with discrete stabilizer, and identify $M_f$ with the quotient $\mathbb{R}^n/\Lambda$ . Compactness enters exactly once — to force the stabilizer lattice to have full rank — and it pays to watch where.

Step 1. Each field is tangent to the level set: $X_{F_j}F_i=\{F_i,F_j\}=0$ , so every flow $\Phi^{t}_{j}$ (the time- $t$ map of $X_{F_j}$ ) preserves each $F_i$ and therefore maps $M_f$ to itself. On the compact $M_f$ the flows are automatically complete — defined for all times.

Step 2. The flows commute: $[X_{F_i},X_{F_j}]=X_{\{F_j,F_i\}}=0$ (the bracket dictionary proven in Liouville Integrable Systems), hence $\Phi^{s}_{i}\circ\Phi^{t}_{j}=\Phi^{t}_{j}\circ\Phi^{s}_{i}$ . Define the joint map $\Phi^{\mathbf t}=\Phi^{t_1}_{1}\circ\cdots\circ\Phi^{t_n}_{n}$ for $\mathbf t=(t_1,\dots,t_n)$ . Commutativity makes $\mathbf t\mapsto\Phi^{\mathbf t}$ a group homomorphism: $\Phi^{\mathbf s+\mathbf t}=\Phi^{\mathbf s}\circ\Phi^{\mathbf t}$ — an action of the abelian group $\mathbb{R}^n$ on $M_f$ .

Step 3. Fix $x_0\in M_f$ and consider the orbit map $\Psi:\mathbb{R}^n\to M_f$ , $\Psi(\mathbf t)=\Phi^{\mathbf t}(x_0)$ . Its differential at $\mathbf 0$ sends the standard basis to the vectors $X_{F_1},\dots,X_{F_n}$ , which are linearly independent (because the $\mathrm{d}F_i$ are, and the symplectic form is nondegenerate). So $\Psi$ is a local diffeomorphism — the orbit is open in $M_f$ . Why the orbit is also closed: an orbit's complement is a union of other orbits, each open by the same argument; so each orbit is open and closed, and connectedness of $M_f$ leaves room for exactly one — the action is transitive. $M_f$ is completely painted by $\Psi$ .

Step 4. Let $\Lambda=\{\mathbf t\in\mathbb{R}^n:\Phi^{\mathbf t}(x_0)=x_0\}$ — the stabilizer, the set of "return recipes". It is a subgroup, and it is discrete: since $\Psi$ is a local diffeomorphism near $\mathbf 0$ , no small nonzero $\mathbf t$ can return $x_0$ . A discrete subgroup of $\mathbb{R}^n$ is a lattice of some rank $k\le n$ , generated by $k$ linearly independent vectors, so $\Psi$ descends to a diffeomorphism $M_f\cong\mathbb{R}^n/\Lambda\cong\mathbb{T}^k\times\mathbb{R}^{n-k}$ .

Step 5. Now compactness strikes: $\mathbb{T}^k\times\mathbb{R}^{n-k}$ is compact only if $k=n$ . Hence $\Lambda$ has full rank and $M_f\cong\mathbb{R}^n/\Lambda\cong\mathbb{T}^n$ . Rescaling the lattice generators to $2\pi$ defines the angle coordinates $\phi_i$ , in which each original flow — in particular the physical time evolution generated by $H$ — is a straight line wrapped on the torus. Verified: the argument is Arnold's (Mathematical Methods §49, Theorem and Lemmas 1-3); the rank-deficient case $k<n$ correctly reproduces the cylinders $\mathbb{T}^k\times\mathbb{R}^{n-k}$ of unbounded motion quoted in the limiting-cases table.

The second pillar makes the actions in (4) legitimate. A loop integral is a useful label only if it does not depend on which loop you draw — and that hinges on a geometric property of the tori that involution hands us for free.

Go deeper · Derivation B · 2 minThe tori are Lagrangian and the actions are well-defined

Derivation BThe tori are Lagrangian and the actions are well-defined

We prove that the symplectic form vanishes identically on each invariant torus — the tori are Lagrangian submanifolds — and deduce that $\oint p\,\mathrm{d}q$ around a cycle depends only on the cycle's homology class. This is the fact that turns the action integrals from loop-dependent curiosities into canonical coordinates, and it is where the "at most $n$ in involution" bound of Liouville Integrable Systems reappears as exactly- $n$ saturation.

Step 1. At each point of $M_f$ the tangent space is spanned by the $n$ commuting fields $X_{F_1},\dots,X_{F_n}$ (Step 3 of the previous derivation). Evaluate the symplectic form on any two spanning vectors: $\omega(X_{F_i},X_{F_j})=\{F_i,F_j\}=0$ by involution.

Step 2. So $\omega$ restricted to $TM_f$ is identically zero — $M_f$ is isotropic — and since $\dim M_f=n$ is the maximal possible dimension for an isotropic submanifold of a $2n$ -dimensional symplectic space, $M_f$ is Lagrangian.

Step 3. The canonical (Liouville) one-form $\vartheta=\sum_k p_k\,\mathrm{d}q_k$ satisfies $\mathrm{d}\vartheta=\sum_k \mathrm{d}p_k\wedge\mathrm{d}q_k=-\omega$ . Restricting to $M_f$ : $\mathrm{d}(\vartheta|_{M_f})=-\omega|_{M_f}=0$ . The one-form we integrate is closed on the torus. Why closedness is the whole game: by Stokes' theorem, the integrals of a closed one-form over two loops that can be deformed into one another (bounding a strip on the torus) are equal. The integral is a function of the homology class alone.

Step 4. The torus has exactly $n$ independent one-dimensional cycles — $H_1(\mathbb{T}^n)=\mathbb{Z}^n$ — so choosing a basis $\gamma_1,\dots,\gamma_n$ produces exactly $n$ well-defined numbers $I_i=\frac{1}{2\pi}\oint_{\gamma_i}\vartheta$ per torus: as many actions as degrees of freedom, no more, no fewer. Since nearby tori vary smoothly, the $I_i$ are smooth functions of the level values $f_j$ , invertible wherever the period matrix $\partial I_i/\partial f_j$ is nonsingular — and then $(I_1,\dots,I_n)$ replaces $(f_1,\dots,f_n)$ as the torus label.

Step 5. Conjugate angles come from a generating function: on a neighbourhood sliced by the tori, $S(q,I)=\int_{q_0}^{q}\vartheta|_{M_f}$ is well-defined up to the periods, and the type-2 transformation $\phi_i=\partial S/\partial I_i$ delivers canonical pairs with $\{\phi_i,I_j\}=\delta_{ij}$ and each $\phi_i$ jumping by exactly $2\pi$ around $\gamma_i$ — the normalization built into (4). The hands-on version of this construction is carried out in Action-Angle Variables. Verified: Arnold §50; the multivaluedness of $S$ by the periods $2\pi I_i$ is exactly why $\phi_i$ is an angle. For the harmonic oscillator the recipe gives $I=E/\omega$ , $\phi$ the phase of the ellipse — matching the pilot construction.

With $(I,\phi)$ canonical and $H=H(I)$ , the dynamics reads off in one line. Hamilton's equations give

\dot I_i=-\frac{\partial H}{\partial\phi_i}=0,\qquad \dot\phi_i=\frac{\partial H}{\partial I_i}\equiv\omega_i(I) \qquad (5)

where: $\omega_i(I)$ — the torus frequencies, constant on each torus because the actions are. Integrated: $\phi_i(t)=\phi_i(0)+\omega_i t$ , and every observable — a periodic function on the torus — becomes a Fourier series over the frequency comb $m\cdot\omega$ , $m\in\mathbb{Z}^n$ . This conditionally periodic motion closes into a periodic orbit when the frequencies satisfy integer relations and fills the torus densely when they do not; the arithmetic of that dichotomy, and its spectacular consequences for Kepler and the oscillator, belong to Degeneracy and Commensurate Frequencies, with the frequency-extraction toolkit in Frequencies from Action-Angle Variables.

Limiting cases

Case	Result	Meaning
Non-compact level set	$M_f\cong\mathbb{T}^k\times\mathbb{R}^{n-k}$	The stabilizer lattice has rank $k<n$ : cylinders, with unbounded (scattering) motion along the $\mathbb{R}$ factors. Actions exist only for the $k$ compact cycles.
Critical level set	$\mathrm{rank}(\mathrm{d}F_1,\dots,\mathrm{d}F_n)<n$	Separatrices and unstable equilibria: the level set is not a smooth manifold, tori pinch, periods diverge, action-angle chart breaks down.
Commensurate frequencies	$m\cdot\omega=0$ for some $m\in\mathbb{Z}^n\setminus\{0\}$	Resonant torus: orbits close or fill sub-tori; the action-angle chart is not unique; these tori are the first casualties of perturbation (KAM).
One degree of freedom	$n=1$	Every bounded orbit is already a 1-torus; $I=$ (enclosed area) $/2\pi$ , $\omega=\partial H/\partial I$ — the familiar oscillator results with no topology needed.
Global obstruction	nontrivial monodromy	Even with all tori regular, circling a singular fibre (e.g. the spherical pendulum's pinched torus) can return a different cycle basis: no single global action-angle chart exists.

Common trapIntegrable means action-angle variables exist globally

The theorem is more careful than the slogan. Its conclusions hold on a neighbourhood of one regular compact torus. Three distinct failures block globalization: (1) separatrices — on critical level sets the torus pinches and the chart is singular (the pendulum's separatrix, where the period diverges logarithmically); (2) non-compactness — scattering directions give cylinders with no action; (3) monodromy — in systems like the spherical pendulum, transporting the cycle basis $(\gamma_1,\gamma_2)$ around a singular value of the integrals brings it back changed (by an integer matrix), so no consistent global choice of actions exists, even though every individual torus is perfectly regular. Local integrability is a theorem; global action-angle coordinates are a further, sometimes false, hope.

Worked examples

Example 1 — anisotropic 2D oscillator: the theorem end to end (confidence builder). A unit mass moves in $V=\tfrac12\omega_x^2x^2+\tfrac12\omega_y^2y^2$ with $\omega_x=3.0\ \mathrm{rad/s}$ , $\omega_y=2.0\ \mathrm{rad/s}$ , and partial energies $E_x=0.90\ \mathrm{J}$ , $E_y=0.40\ \mathrm{J}$ . The commuting pair is $(E_x,E_y)$ ; the level set is the product of two phase-plane ellipses — a 2-torus, as the theorem demands.

Step 1. Actions from the ellipse areas (the loop integral $\oint p\,\mathrm{d}q$ equals the enclosed area $2\pi E/\omega$ for each oscillator): $I_x=E_x/\omega_x=0.90/3.0=\mathbf{0.30\ J\,s}$ , $I_y=E_y/\omega_y=0.40/2.0=\mathbf{0.20\ J\,s}$ .

Step 2. Hamiltonian in actions: $H=\omega_xI_x+\omega_yI_y=3.0(0.30)+2.0(0.20)=\mathbf{1.30\ J}$ — equal to $E_x+E_y=0.90+0.40$ . ✓

Step 3. Frequencies from the theorem's formula: $\omega_x=\partial H/\partial I_x=3.0\ \mathrm{rad/s}$ , $\omega_y=\partial H/\partial I_y=2.0\ \mathrm{rad/s}$ — the construction returns the input frequencies, as it must for this linear case.

Step 4. Read the winding. The ratio $\omega_x/\omega_y=3/2$ is rational, so the resonance relation $2\omega_x-3\omega_y=0$ holds: the torus is resonant and the orbit closes. Closure time: the smallest $T$ with $\omega_xT$ and $\omega_yT$ both multiples of $2\pi$ is $T=2\pi\cdot\mathrm{lcm}(1/3,1/2)\ \mathrm{s}=2\pi\ \mathrm{s}\approx\mathbf{6.28\ s}$ , in which $x$ completes exactly 3 cycles and $y$ exactly 2 — a closed 3:2 Lissajous figure. Detune $\omega_y$ to $\sqrt{2}\approx1.414$ and no integer relation survives: the same construction goes through, but the orbit now fills its torus densely.

Example 2 — particle in a box: actions deep in the classical regime (exam level). A nitrogen molecule ( $m=4.65\times10^{-26}\ \mathrm{kg}$ ) bounces between the walls of a 1D container of length $L=0.10\ \mathrm{m}$ at speed $v=500\ \mathrm{m/s}$ (a typical thermal speed). The bounded orbit is a 1-torus; find its action, verify the frequency formula against the bounce time, and locate the motion on the quantum ladder.

Step 1. Energy and momentum: $E=\tfrac12mv^2=\tfrac12(4.65\times10^{-26})(500)^2=5.81\times10^{-21}\ \mathrm{J}$ ; $|p|=mv=2.33\times10^{-23}\ \mathrm{kg\,m/s}$ .

Step 2. Action. One period traverses the box right at $+|p|$ and back at $-|p|$ : $\oint p\,\mathrm{d}q=2L|p|$ , so $I=\dfrac{2L|p|}{2\pi}=\dfrac{L\sqrt{2mE}}{\pi}=\dfrac{(0.10)(2.33\times10^{-23})}{\pi}=\mathbf{7.40\times10^{-25}\ J\,s}$ .

Step 3. Hamiltonian in the action and the frequency. Inverting Step 2, $H(I)=\dfrac{\pi^2I^2}{2mL^2}$ , so $\omega=\dfrac{\partial H}{\partial I}=\dfrac{\pi^2 I}{mL^2}=\dfrac{(9.870)(7.40\times10^{-25})}{(4.65\times10^{-26})(0.010)}=\mathbf{1.57\times10^{4}\ rad/s}$ .

Step 4. Independent check: the round-trip time is $T=2L/v=2(0.10)/500=4.0\times10^{-4}\ \mathrm{s}$ , giving $\omega=2\pi/T=1.57\times10^{4}\ \mathrm{rad/s}$ . ✓ The abstract derivative $\partial H/\partial I$ knows the bounce time.

Step 5. The quantum ladder. $I/\hbar=7.40\times10^{-25}/1.055\times10^{-34}=\mathbf{7.0\times10^{9}}$ : the molecule sits on roughly the seven-billionth quantized torus of Bohr-Sommerfeld Quantization — the depth of the classical regime, and a preview of why semiclassical methods are so accurate for macroscopic actions.

The history band

Origins · 1 minVladimir Arnold and how this was discovered

Vladimir Arnold

1937–2010 · Moscow geometer who turned mechanics into geometry — tori, KAM, and the catastrophe of small denominators

Arnold solved Hilbert's 13th problem at nineteen, and at twenty-six published the two results that carry his name here: the proof half of KAM theory (his advisor Kolmogorov had announced the idea in 1954) and the global geometric form of Liouville's integrability theorem. His textbook Mathematical Methods of Classical Mechanics re-founded the subject for two generations — mechanics as symplectic geometry, with the Liouville-Arnold tori as its central characters. The same year, 1964, he discovered that perturbed tori can leak — Arnold diffusion — showing his own theorem's tori are the exception that proves chaos. Famously combative about rigour and priority ("the Arnold principle: if a notion bears a personal name, then this name is not the name of the discoverer"), he insisted the theorem be credited to Liouville; the community, applying the Arnold principle to itself, kept both names.

History

1855

— Liouville proves n involutive integrals allow solution by quadratures

1890

— Poincaré exposes the generic failure of integrability in the three-body problem

1954

— Kolmogorov announces the persistence of strongly irrational tori under perturbation

1963

— Arnold publishes the global torus theorem and completes the KAM proof for analytic systems

1964

— Arnold discovers diffusion along resonances — the escape route his tori cannot block

1980

— Duistermaat identifies monodromy as the obstruction to global action-angle variables

Going further

References. Arnold, Mathematical Methods of Classical Mechanics, 2nd ed., §49–50 · Goldstein, Poole & Safko, Classical Mechanics, 3rd ed., §10.6–10.8 · Landau & Lifshitz, Mechanics, 3rd ed., §49–50 · José & Saletan, Classical Dynamics, §6.2 · Cushman & Bates, Global Aspects of Classical Integrable Systems (monodromy) · Duistermaat, Comm. Pure Appl. Math. 33, 687 (1980)

Provenance. Statement and proof structure follow Arnold, Mathematical Methods of Classical Mechanics §49–50 (involution, torus theorem, action-angle construction). Action-integral and frequency formulas verifiable against Goldstein §10.6–10.8 and Landau–Lifshitz §49–50. A verifier should recompute: (1) the orbit-map/lattice argument of the torus derivation against Arnold §49, Lemmas 1–3; (2) the Lagrangian property $\omega(X_{F_i},X_{F_j})=\{F_i,F_j\}=0$ and closedness of $p\,\mathrm{d}q$ ; (3) the worked numbers — anisotropic oscillator $I_x=0.30$ , $I_y=0.20$ J·s, $H=1.30$ J, closure at $T=2\pi$ s for the 3:2 resonance; box molecule $I=L\sqrt{2mE}/\pi=7.40\times10^{-25}$ J·s and $\omega=\pi^2I/(mL^2)=1.57\times10^4$ rad/s $=2\pi v/2L$ . All are hand-verifiable; no fabricated constants appear.

Builds on: Liouville Integrable Systems · Action-Angle Variables · Hamilton-Jacobi Equation · Hamilton's Characteristic Function

Leads to: Action Integrals over Invariant Tori · Frequencies from Action-Angle Variables · Degeneracy and Commensurate Frequencies · Separation of Variables in Hamilton-Jacobi Theory

Why must the stabilizer of the R^n action be discrete, and why does discreteness give a lattice?

Discreteness comes from the orbit map being a local diffeomorphism: near $\mathbf t=0$ the map $\mathbf t\mapsto\Phi^{\mathbf t}(x_0)$ is injective, so no nonzero recipe close to zero can return the point — the stabilizer has an isolated identity, and a subgroup with one isolated point has all points isolated. The classification of discrete subgroups of $\mathbb{R}^n$ is then a classical lemma: pick a shortest nonzero element, quotient it out, induct — every discrete subgroup is generated by $k\le n$ linearly independent vectors, i.e. is a lattice of rank $k$ . Full rank ( $k=n$ ) is not automatic; it is exactly what compactness buys, since $\mathbb{R}^n/\Lambda$ with a rank-deficient $\Lambda$ still has unrolled, non-compact directions.

Is the theorem constructive — how do I actually find the cycles and compute the actions for a real system?

The proof is constructive in principle and the practice is easier than the proof. For separable systems — the common case — each pair $(q_k,p_k)$ executes its own libration or rotation, and the natural cycles are "one full excursion of $q_k$ with the others frozen": the action reduces to the one-dimensional integral $I_k=\frac{1}{2\pi}\oint p_k(q_k;f)\,\mathrm{d}q_k$ familiar from Action-Angle Variables. The worked recipes — turning points, quadratures, the Kepler and oscillator catalogues — are the content of Action Integrals over Invariant Tori. For non-separable integrable systems (Toda, Calogero) the cycles must be located on the Liouville tori themselves, typically via the spectral curve of a Lax pair — constructive still, but with algebraic-geometry machinery.

What is Hamiltonian monodromy, concretely, in the spherical pendulum?

Take the pendulum free to swing on a sphere, with integrals $(H,L_z)$ . The image of phase space under $(H,L_z)$ has one special interior point, corresponding to the unstable upright equilibrium; its fibre is a "pinched torus". Circle a loop in the $(H,L_z)$ plane around that point, transporting a basis of cycles $(\gamma_1,\gamma_2)$ continuously along the regular tori. On return, the basis comes back as $(\gamma_1,\gamma_2+\gamma_1)$ — shifted by an integer matrix $\begin{pmatrix}1&0\\1&1\end{pmatrix}$ . Since actions are integrals over these cycles, no consistent global labelling of tori by $(I_1,I_2)$ exists: the action lattice has a screw dislocation. Duistermaat identified this in 1980; its quantum image — a defect in the joint spectrum lattice — is measured in molecular spectra.

How is Liouville-Arnold related to superintegrability, where there are more than n integrals?

More than $n$ independent integrals cannot all be in involution (the isotropic bound of Liouville Integrable Systems), but extra non-commuting integrals can coexist with a commuting core of $n$ . Such systems — Kepler with its Runge-Lenz vector, the isotropic oscillator with its quadrupole tensor — are called superintegrable. The extra integrals slice the $n$ -tori into lower-dimensional invariant sets: with the maximal $2n-1$ independent integrals, every bounded orbit is a closed curve. Dynamically this shows up as degenerate frequencies ( $\omega_r=\omega_\theta=\omega_\varphi$ for Kepler), which is why superintegrability, closed orbits, and frequency commensurability are one phenomenon — dissected in Degeneracy and Commensurate Frequencies.

The pendulum is integrable with n = 1. Why does its phase portrait still have a separatrix where action-angle variables fail?

Integrability guarantees tori only on regular, compact, connected level sets, and the pendulum's phase portrait is a map of where those adjectives hold: libration ovals inside the separatrix (tori of one family), rotation curves above and below (tori of another), and the separatrix itself — the level set of the unstable equilibrium — where $\mathrm{d}H=0$ at a point and regularity fails. There the period diverges logarithmically, the enclosed-area action is continuous but its derivative $\partial I/\partial E$ blows up, and no single action-angle chart covers both families. The theorem is not violated; its hypotheses simply exclude that one level set — and perturbation theory duly fails first in its neighbourhood, which is where chaotic layers are born.

Different books write the flow-bracket identity with opposite signs — does the sign matter for the theorem?

No, and it is worth seeing why. Whether one's conventions give $[X_F,X_G]=X_{\{F,G\}}$ or $[X_F,X_G]=-X_{\{F,G\}}$ (both appear: Arnold and Goldstein differ in the bracket sign and in $X_F$ 's definition), the theorem needs only the vanishing of the right-hand side when $\{F_i,F_j\}=0$ . Commuting is sign-blind. The same robustness holds for the Liouville form: $\vartheta=p\,\mathrm{d}q$ or $-q\,\mathrm{d}p$ differ by an exact form $\mathrm{d}(pq)$ , which integrates to zero around any cycle, so the actions are convention-independent. When cross-checking derivations across textbooks, fix conventions once from the fundamental bracket $\{q,p\}=1$ and translate; every physical conclusion — tori, frequencies, actions — survives untouched.

Does the theorem say anything about how the motion looks in the original coordinates?

It says the motion is a fixed quasi-periodic signal filtered through a static nonlinear map. The theorem produces $\phi(t)=\phi_0+\omega t$ on the torus; returning to $(q,p)$ applies the (generally complicated) change of variables, so every original coordinate is a multiply-periodic Fourier series $q_k(t)=\sum_{m\in\mathbb{Z}^n}a^{(k)}_m e^{i(m\cdot\omega)t}$ with fixed frequencies and fixed amplitudes. That is a sharp structural prediction: the power spectrum of any observable of an integrable system consists of sharp lines at integer combinations $m\cdot\omega$ — no broadband component. Numericists use exactly this as a chaos detector: broadened spectra or drifting frequencies in a simulation certify that the trajectory has left the Liouville-Arnold world.