Liouville Integrable Systems

Newton solved the two-body problem in 1687. Three and a half centuries later, the three-body problem remains unsolved — not because we lack cleverness, but because, as Poincaré proved, no solution of the kind we want exists. What separates the systems that yield exact solutions — Kepler orbits, spinning tops, chains of springs — from the ones that tumble into chaos? The answer is not "enough conserved quantities." It is a sharper, more geometric condition: enough conserved quantities that commute with each other. A Hamiltonian system with $n$ degrees of freedom is Liouville integrable when it admits $n$ functionally independent constants of motion $F_1,\dots,F_n$ in involution — every pairwise Poisson bracket vanishes. That involution clause looks like a technicality; it is actually the entire story. Conservation alone pins a trajectory to a level set. Involution makes the $n$ conserved quantities generate $n$ commuting flows, and commuting flows weave the level set into an $n$ -dimensional torus on which the motion is nothing but uniform winding. Every exactly solvable system in mechanics — from the harmonic oscillator to the Toda lattice — is solvable for exactly this reason. This article defines integrability precisely, proves why involution converts algebra into torus geometry, and establishes why $n$ is the maximum possible number of commuting integrals. It is the dividing line of Hamiltonian mechanics: on one side, action-angle variables and exact solutions; on the other, KAM theory, chaos, and the three-body problem. Mastering the definition here is what makes the Liouville-Arnold Theorem — the structure theorem of the whole subject — a short walk rather than a leap.

Key takeaways

A Hamiltonian system with $n$ degrees of freedom is Liouville integrable if it has $n$ functionally independent conserved quantities $F_1,\dots,F_n$ with all pairwise Poisson brackets zero: $\{F_i,F_j\}=0$ (involution).
Involution is strictly stronger than conservation: it makes the $n$ Hamiltonian flows commute, so the common level set $M_f=\{F_i=f_i\}$ is woven by commuting flows — a torus $\mathbb{T}^n$ when compact and connected.
On each invariant torus the motion is conditionally periodic: $\theta_i(t)=\theta_i(0)+\omega_i t$ with $\omega_i=\partial H/\partial I_i$ — all complexity hides in the change of variables, none in the dynamics.
At most $n$ independent functions can be in involution on a $2n$ -dimensional phase space — the commuting flow directions span an isotropic subspace, capped at half the dimension. Integrability saturates the cap.
Separability of the Hamilton-Jacobi equation automatically manufactures $n$ integrals in involution — every separable system is integrable.
Integrability is exceptional and fragile: generic Hamiltonians have no integrals beyond energy, and small perturbations of integrable systems destroy most tori (KAM theory).
Key idea: Liouville integrability = $n$ independent integrals in involution. The involution (commuting) condition, not just conservation, is what forces phase space to foliate into invariant tori carrying linear conditionally-periodic motion.

Where this lives today

Integrability graduated from a classical-mechanics chapter to a working tool across physics. The Toda lattice and Calogero-Moser systems are the fruit-fly models of the trade, and the discovery that the Korteweg-de Vries equation is an infinite-dimensional Liouville integrable system — with the inverse scattering transform playing the role of action-angle variables and soliton amplitudes as the actions — created a whole discipline of integrable PDEs. On the quantum side, involution becomes commutativity of operators: the Bethe ansatz solves 1D quantum magnets and cold-atom gases exactly because a tower of commuting conserved charges survives quantization, and experiments like the quantum Newton's cradle (Kinoshita et al., 2006) show 1D Bose gases refusing to thermalize — their many commuting integrals force relaxation to a generalized Gibbs ensemble that remembers every action, not just the energy. The classical-quantum bridge runs through this page in the other direction too. Semiclassical quantization — Bohr-Sommerfeld Quantization and its EBK refinement — quantizes precisely the invariant tori built here ( $I_i=(n_i+\mu_i/4)\hbar$ ), which is why it works only for integrable systems; the Berry-Tabor conjecture (1977) predicts that quantum spectra of classically integrable systems show Poisson-distributed level spacings, while chaotic ones show random-matrix repulsion — a dichotomy tested daily in Rydberg atoms, microwave billiards, and quantum dots. Integrability is the baseline against which quantum chaos is defined.

The physical picture

Picture the $2n$ -dimensional phase space, and think of each integral $F_i$ as a knob frozen at the value $f_i$ . Each freeze slices away one dimension: $n$ frozen knobs cut the $2n$ -dimensional space down to an $n$ -dimensional sheet, the common level set $M_f$ . So far this is bookkeeping — any $n$ conserved quantities would do the same. The involution condition is what disciplines the sheet's shape. Each $F_i$ generates its own Hamiltonian flow, a way of sliding points around phase space; vanishing brackets mean these $n$ flows commute — you can slide first along flow 1 then flow 2, or in the opposite order, and land on the same point. A compact $n$ -dimensional surface carrying $n$ independent, everywhere-nonzero, mutually commuting flow directions has only one possible topology: the $n$ -torus. Not a sphere — a sphere cannot even carry one nonvanishing vector field (the hairy ball theorem) — a torus.

Figure

Freezing n commuting integrals slices phase space down to an invariant torus carrying linear winding motion

Once the torus is recognized, the dynamics collapses. The natural coordinates on a torus are angles, one per winding direction; the actual trajectory advances each angle at a constant rate, $\theta_i \to \theta_i + \omega_i t$ . All the apparent complexity of the motion in the original variables — the looping of a Kepler ellipse, the nutation of a top — lives entirely in the nonlinear change of variables from $(\theta, I)$ back to $(q,p)$ . The dynamics itself is as simple as dynamics can be: straight-line motion at constant speed, wrapped around a torus. That is what "exactly solvable" means, made geometric. The picture also shows what is at stake when integrability fails. Energy conservation alone confines an $n$ -degree-of-freedom trajectory to a $(2n-1)$ -dimensional energy shell; for $n\ge 2$ that leaves room for the orbit to wander ergodically and chaotically. Each additional commuting integral collapses another dimension, and the full set of $n$ crushes the available space down to tori too thin and too orderly for chaos. Integrable and chaotic systems live on opposite sides of exactly this line.

Figure

Poincare section view: an integrable system shows nested invariant curves while a perturbed system shows surviving tori embedded in chaotic scatter

What integrability demands

Three separate conditions go into the definition, and each earns its keep. Start with the one that does the geometric heavy lifting. For two functions on phase space, the Poisson bracket measures how each changes along the flow the other generates — the full algebraic structure is developed in Algebra of Poisson Brackets. Involution demands this mutual influence vanish for every pair of our $n$ integrals:

\{F_i,F_j\}=\sum_{k=1}^{n}\left(\frac{\partial F_i}{\partial q_k}\frac{\partial F_j}{\partial p_k}-\frac{\partial F_i}{\partial p_k}\frac{\partial F_j}{\partial q_k}\right)=0 \qquad (1)

where: $F_i,F_j$ — any two of the $n$ integrals; $q_k,p_k$ — canonical coordinates and momenta. Physically: each integral is constant not only along the time flow but along the flow generated by every other integral — the $n$ quantities form a mutually non-interfering family, an abelian subalgebra of the Poisson algebra. In the Kepler problem, $\{L^2, L_z\}=0$ is exactly such a statement: the total angular momentum does not change as you rotate the orbit about the $z$ -axis. Second, the family must actually be conserved. Taking one of the $F_i$ to be $H$ itself (always allowed), conservation of the others is the statement that their bracket with the Hamiltonian vanishes:

\frac{\mathrm{d}F_i}{\mathrm{d}t}=\{F_i,H\}+\frac{\partial F_i}{\partial t}=0 \qquad (2)

where: $H$ — the Hamiltonian, itself a member of the family; $\partial F_i/\partial t$ — explicit time dependence, zero for the autonomous integrals we consider. Note that once $H\in\{F_1,\dots,F_n\}$ , conservation is contained in involution: $\{F_i,H\}=0$ is just one row of condition (1). This is why the definition is usually stated as " $n$ functions in involution, one of which is $H$ ." Third, the integrals must be genuinely different functions, not disguised copies of each other ( $H$ and $H^2$ are both conserved but carry the same information). Independence is the demand that their gradients span an $n$ -dimensional space at almost every point:

\mathrm{d}F_1\wedge\mathrm{d}F_2\wedge\cdots\wedge\mathrm{d}F_n\neq 0 \qquad (3)

where: $\mathrm{d}F_i$ — the differential (gradient one-form) of $F_i$ ; $\wedge$ — the wedge product, which vanishes exactly when the gradients are linearly dependent. Geometrically, (3) is what guarantees each frozen knob slices away a genuine dimension, so the common level set

M_f=\{(q,p):F_i(q,p)=f_i,\ i=1,\dots,n\} \qquad (4)

is a smooth $n$ -dimensional submanifold. where: $f_i$ — the frozen values; $M_f$ — the invariant manifold on which the trajectory lives forever. When $M_f$ is compact and connected, the Liouville-Arnold Theorem upgrades it to a torus $\mathbb{T}^n$ and delivers the action-angle chart; this article proves the two structural facts that theorem rests on.

Why involution means commuting flows

Everything hinges on translating the algebraic statement $\{F_i,F_j\}=0$ into geometry. The dictionary entry is the Hamiltonian vector field: every phase-space function $F$ generates a flow, and brackets of functions correspond to commutators of flows.

Derivation ACommuting integrals generate commuting flows

We prove that involution of two integrals is equivalent to the commutativity of the flows they generate — the single fact that forces the level set to be a torus and the foundation on which the Liouville-Arnold Theorem builds. The strategy is to express the Lie bracket of two Hamiltonian vector fields as the Hamiltonian vector field of the Poisson bracket, using the Jacobi identity; involution then kills the bracket instantly. Watch how the Jacobi identity — the one non-trivial axiom of the Poisson algebra — is precisely what makes the dictionary between functions and flows respect brackets.

Step 1. Attach to each function $F$ its Hamiltonian vector field, the differential operator that measures rate of change along $F$ 's flow: $X_F=\sum_k\left(\tfrac{\partial F}{\partial p_k}\tfrac{\partial}{\partial q_k}-\tfrac{\partial F}{\partial q_k}\tfrac{\partial}{\partial p_k}\right)$ , so that for any observable $G$ , $X_F G=\{G,F\}$ . Taking $F=H$ recovers ordinary time evolution, $\dot G=\{G,H\}$ .

Step 2. Compute the commutator of two such fields acting on an arbitrary test function $G$ : $[X_F,X_{G'}]G = X_F(\{G,G'\})-X_{G'}(\{G,F\}) = \{\{G,G'\},F\}-\{\{G,F\},G'\}$ . Why a test function: two vector fields are equal exactly when they act identically on all smooth functions, so proving the operator identity on arbitrary $G$ proves it as geometry.

Step 3. Apply the Jacobi identity $\{\{G,G'\},F\}+\{\{G',F\},G\}+\{\{F,G\},G'\}=0$ to rewrite the right-hand side: $\{\{G,G'\},F\}-\{\{G,F\},G'\} = \{G,\{G',F\}\}$ . Hence the operator identity

[X_F,X_{G'}] = X_{\{G',F\}}

— the commutator of two Hamiltonian flows is itself a Hamiltonian flow, generated by the Poisson bracket of the generators.

Step 4. Set $F=F_i$ , $G'=F_j$ with $\{F_i,F_j\}=0$ . Then $[X_{F_i},X_{F_j}]=X_0=0$ : the flows commute exactly. Conversely, if the flows commute then $X_{\{F_j,F_i\}}=0$ , so $\{F_i,F_j\}$ is a constant on each connected component — and equals zero whenever the bracket vanishes anywhere on it. Commuting integrals and commuting flows are the same condition in two languages.

Step 5. Because each $X_{F_i}$ is tangent to $M_f$ (each $F_j$ is constant along it), the $n$ independent commuting flows define an action of the abelian group $\mathbb{R}^n$ on the $n$ -dimensional manifold $M_f$ : the point $x$ is carried to $\Phi^{t_1}_1\circ\cdots\circ\Phi^{t_n}_n(x)$ , and commutativity makes the order irrelevant. When $M_f$ is compact and connected, the subgroup of $(t_1,\dots,t_n)$ that returns $x$ to itself is a full-rank lattice $\Lambda\cong\mathbb{Z}^n$ , so $M_f\cong\mathbb{R}^n/\Lambda=\mathbb{T}^n$ . The lattice construction is carried out in detail in the Liouville-Arnold Theorem. Why the sign convention is consistent: expanding both sides of $[X_F,X_G]=X_{\{G,F\}}$ on the test functions $G=q_k$ and $G=p_k$ reproduces the identity component by component, confirming that the convention $X_FG=\{G,F\}$ is used throughout.

The torus, once found, comes with canonical coordinates. Around each of its $n$ independent cycles $\gamma_i$ one defines the action

I_i=\frac{1}{2\pi}\oint_{\gamma_i} \sum_k p_k\,\mathrm{d}q_k \qquad (5)

where: $\gamma_i$ — the $i$ -th irreducible loop on the torus; the normalization makes the conjugate angle advance by $2\pi$ per circuit. The construction, and the proof that the $I_i$ are themselves in involution, is the business of Action Integrals over Invariant Tori and Action-Angle Variables — here we need only the outcome: $(I,\theta)$ is a canonical chart in which the Hamiltonian depends on the actions alone,

H=H(I_1,\dots,I_n),\qquad \dot I_i=-\frac{\partial H}{\partial \theta_i}=0,\qquad \dot\theta_i=\frac{\partial H}{\partial I_i}\equiv\omega_i(I) \qquad (6)

where: $\omega_i$ — the torus frequencies, constant on each torus because the $I_i$ are. Integrating trivially,

\theta_i(t)=\theta_i(0)+\omega_i t \pmod{2\pi} \qquad (7)

where: the mod $2\pi$ records the winding. Every observable, being a periodic function on the torus, becomes a Fourier sum over frequencies $m\cdot\omega$ with integer vectors $m$ — conditionally periodic motion. Whether the orbit closes (commensurate $\omega_i$ ) or fills the torus densely (incommensurate) is the subject of Degeneracy and Commensurate Frequencies, and the frequencies themselves are extracted systematically in Frequencies from Action-Angle Variables.

Why n is the maximum

The definition asks for exactly $n$ integrals in involution. Fewer, and the system is only partially integrable. But why not demand more — say $2n-1$ , enough to pin the orbit completely? Because more is impossible: the symplectic geometry of phase space caps a commuting family at $n$ members.

Go deeper · Derivation B · 2 minAt most n independent functions can be in involution

Derivation BAt most n independent functions can be in involution

We prove the completeness bound: on a $2n$ -dimensional phase space, no more than $n$ functionally independent functions can have all pairwise Poisson brackets zero. The strategy is pure linear algebra at a single point — involution says the Hamiltonian vector fields of the family span a subspace on which the symplectic form vanishes identically, and such isotropic subspaces cannot exceed half the total dimension. The bound explains the word complete in "completely integrable": $n$ is not a modest requirement but the saturation of a hard geometric ceiling.

Step 1. Recall that the Poisson bracket is the symplectic form evaluated on Hamiltonian vector fields: $\{F,G\}=\omega(X_F,X_G)$ , with $\omega=\sum_k \mathrm{d}q_k\wedge\mathrm{d}p_k$ nondegenerate — for every nonzero vector $v$ there exists $w$ with $\omega(v,w)\neq 0$ .

Step 2. Let $F_1,\dots,F_m$ be independent and in involution, and fix a regular point where the differentials $\mathrm{d}F_i$ are linearly independent. Nondegeneracy of $\omega$ makes the map $F \mapsto X_F$ injective on gradients, so the vector fields $X_{F_1},\dots,X_{F_m}$ span an $m$ -dimensional subspace $V$ of the tangent space.

Step 3. Involution translates to: $\omega(X_{F_i},X_{F_j})=\{F_i,F_j\}=0$ for all pairs — the form $\omega$ restricted to $V$ is identically zero. Such a subspace is called isotropic.

Step 4. An isotropic subspace of a $2n$ -dimensional symplectic vector space has dimension at most $n$ . Consider the $\omega$ -orthogonal complement $V^{\omega}=\{w:\omega(v,w)=0\ \forall v\in V\}$ ; nondegeneracy gives $\dim V^{\omega}=2n-\dim V$ , and isotropy means $V\subseteq V^{\omega}$ , so $\dim V\le 2n-\dim V$ , i.e. $m=\dim V\le n$ . Why this cannot be evaded by cleverness: the bound is pointwise linear algebra, independent of the dynamics — no choice of coordinates, no hidden symmetry, no non-smooth trick manufactures an $(n+1)$ -th commuting integral.

Step 5. When the bound is saturated ( $m=n$ ), $V=V^{\omega}$ is a maximal isotropic — Lagrangian — subspace, and it is precisely the tangent space to the level set $M_f$ . Each invariant torus of an integrable system is therefore a Lagrangian submanifold: the symplectic form vanishes along it entirely. This is the geometric signature that makes the generating-function machinery of Hamilton-Jacobi Equation theory work globally. Why the dimension count works: $\dim V^{\omega}=2n-\dim V$ is simply the rank-nullity theorem applied to the nondegenerate symplectic pairing, and the Lagrangian-submanifold conclusion follows directly from saturating the bound.

There is a beautiful shortcut to producing such maximal families. Whenever the time-independent Hamilton-Jacobi equation separates — the setting of Separation of Variables in Hamilton-Jacobi Theory — the characteristic function splits as

W(q,\alpha)=\sum_{k=1}^{n} W_k(q_k;\alpha_1,\dots,\alpha_n) \qquad (8)

where: $W_k$ — a function of the single coordinate $q_k$ and the $n$ separation constants $\alpha_i$ . Each separated momentum $p_k=\partial W_k/\partial q_k$ depends only on its own coordinate, and solving the separated equations for the constants expresses each $\alpha_i$ as a function $\alpha_i(q_k,p_k)$ of disjoint variable pairs — whence $\{\alpha_i,\alpha_j\}=0$ automatically. Separability therefore manufactures integrability: the separation constants are the commuting integrals. (The converse fails: the Toda lattice is integrable but not separable in any obvious coordinates — involution is the deeper notion, and the Staeckel Conditions for Separability characterize exactly when the shortcut is available.)

Limiting cases

Case	Result	Meaning
One degree of freedom	$n=1$ : energy alone suffices	Every autonomous 1-DOF Hamiltonian is integrable; the level curve $H=E$ is the torus $\mathbb{T}^1$ (a closed orbit) for bounded motion. Chaos needs $n\ge 2$ .
Maximal degeneracy	all $\omega_i$ commensurate	Every orbit closes: genuinely periodic motion on a circle inside the torus. Kepler and the isotropic oscillator live here — see Degeneracy and Commensurate Frequencies.
Non-compact level set	$M_f\cong \mathbb{T}^k\times\mathbb{R}^{n-k}$	Unbounded (scattering) motion replaces some winding directions by straight lines; a cylinder, not a torus. Action variables exist only for the compact directions.
Generic perturbation	$H=H_0(I)+\varepsilon H_1(\theta,I)$	Integrability generically destroyed. KAM: tori with Diophantine frequency vectors survive small $\varepsilon$ ; resonant tori shatter first.
Partial involution	$m<n$ commuting integrals	Motion confined to $(2n-m)$ -dimensional sets, not tori; the system can still be chaotic within them (e.g. Hénon-Heiles has only $H$ ).

Common trapHaving n conserved quantities makes a system integrable

It does not — this is the single most common misreading of the definition. Conservation ( $\{F_i,H\}=0$ ) only pins the orbit to a level set; involution ( $\{F_i,F_j\}=0$ for every pair) is what makes the flows commute and the level set a torus. The point-mass in a generic potential with $H$ , $L_x$ , $L_y$ , $L_z$ all conserved is the cautionary tale: $\{L_x,L_y\}=L_z\neq 0$ , so despite four conserved quantities only two of them ( $H$ and, say, $L_z$ … plus $L^2$ ) can join a commuting family. Count integrals in involution, never just integrals. And remember the topological fine print: even a commuting family gives tori only where the level set is compact and connected.

Worked examples

Example 1 — the 2D isotropic oscillator, integrability audited by hand (confidence builder). A particle of mass $m=1.0\ \mathrm{kg}$ moves in the plane under $V=\tfrac12 m\omega^2(x^2+y^2)$ with $\omega=2.0\ \mathrm{rad/s}$ . Take as candidate integrals the two partial energies $F_1=E_x=\tfrac{p_x^2}{2m}+\tfrac12 m\omega^2 x^2$ and $F_2=E_y=\tfrac{p_y^2}{2m}+\tfrac12 m\omega^2 y^2$ . Suppose the motion has $E_x=0.40\ \mathrm{J}$ and $E_y=0.10\ \mathrm{J}$ .

Step 1. Involution check. $E_x$ depends only on $(x,p_x)$ , $E_y$ only on $(y,p_y)$ : every term of $\{E_x,E_y\}=\tfrac{\partial E_x}{\partial x}\tfrac{\partial E_y}{\partial p_x}-\tfrac{\partial E_x}{\partial p_x}\tfrac{\partial E_y}{\partial x}+\tfrac{\partial E_x}{\partial y}\tfrac{\partial E_y}{\partial p_y}-\tfrac{\partial E_x}{\partial p_y}\tfrac{\partial E_y}{\partial y}$ contains a derivative that vanishes identically. So $\{E_x,E_y\}=0$ , and since $H=E_x+E_y$ , both bracket to zero with $H$ as well. Independence: $\mathrm{d}E_x\wedge \mathrm{d}E_y\neq0$ wherever neither oscillator sits at rest. Two independent integrals in involution, $n=2$ : integrable.

Step 2. Actions. For a 1-DOF oscillator the loop integral gives $I=E/\omega$ (computed in Action-Angle Variables), so $I_x=0.40/2.0=\mathbf{0.20\ J\,s}$ and $I_y=0.10/2.0=\mathbf{0.050\ J\,s}$ .

Step 3. Hamiltonian in action form and frequencies. $H=\omega(I_x+I_y)=2.0\times(0.20+0.050)=\mathbf{0.50\ J}$ — matching $E_x+E_y=0.40+0.10=0.50\ \mathrm{J}$ . Frequencies: $\omega_x=\partial H/\partial I_x=2.0\ \mathrm{rad/s}$ and $\omega_y=2.0\ \mathrm{rad/s}$ .

Step 4. Read the torus. The frequency ratio is $\omega_x/\omega_y=1$ — perfectly commensurate — so the orbit on the 2-torus closes after one circuit: in real space, a closed Lissajous figure (an ellipse), with amplitudes $A_x=\sqrt{2E_x/(m\omega^2)}=\sqrt{0.80/4.0}=\mathbf{0.45\ m}$ and $A_y=\sqrt{0.20/4.0}=\mathbf{0.22\ m}$ . Detune the trap to $\omega_y=1.9\ \mathrm{rad/s}$ and the ratio $20{:}19$ closes only after 19 $x$ -cycles; make it irrational and the orbit never closes, densely filling the $0.45\times0.22$ m rectangle.

Example 2 — Kepler actions for a real satellite (exam level). A satellite of mass $m=1000\ \mathrm{kg}$ orbits Earth ( $GM=3.986\times10^{14}\ \mathrm{m^3/s^2}$ ) on an ellipse of semi-major axis $a=7.00\times10^{6}\ \mathrm{m}$ (a low-Earth orbit, ~620 km altitude). The Kepler problem is integrable with commuting family $\{H, L^2, L_z\}$ ; separation in spherical coordinates yields three actions with the famous degenerate combination $E=-mk^2/[2(I_r+I_\theta+I_\varphi)^2]$ , where $k=GMm$ .

Step 1. Energy from the orbit: $E=-\dfrac{GMm}{2a}=-\dfrac{(3.986\times10^{14})(1000)}{2(7.00\times10^{6})}=\mathbf{-2.85\times10^{10}\ J}$ .

Step 2. Total action from the energy-action relation. Inverting $E=-mk^2/(2J^2)$ with $J\equiv I_r+I_\theta+I_\varphi$ and $k=GMm=3.986\times10^{17}\ \mathrm{N\,m^2}$ : $J=k\sqrt{\dfrac{m}{2|E|}}=3.986\times10^{17}\sqrt{\dfrac{1000}{2(2.85\times10^{10})}}=3.986\times10^{17}\times(1.325\times10^{-4})=\mathbf{5.28\times10^{13}\ J\,s}$ . Cross-check by the independent formula $J=m\sqrt{GMa}=1000\sqrt{(3.986\times10^{14})(7.00\times10^{6})}=1000\times5.28\times10^{10}$ — same value. ✓

Step 3. One frequency, three times. $\omega=\partial H/\partial J = mk^2/J^3 = \dfrac{(1000)(3.986\times10^{17})^2}{(5.28\times10^{13})^3}=\dfrac{1.589\times10^{38}}{1.474\times10^{41}}=\mathbf{1.078\times10^{-3}\ rad/s}$ . Because $H$ depends only on the sum of the actions, all three frequencies are equal — complete degeneracy.

Step 4. Sanity check against Kepler's third law: $T=2\pi\sqrt{a^3/GM}=2\pi\sqrt{(3.43\times10^{20})/(3.986\times10^{14})}=2\pi(928\ \mathrm{s})=5.83\times10^{3}\ \mathrm{s}\approx 97\ \mathrm{min}$ , giving $\omega=2\pi/T=1.078\times10^{-3}\ \mathrm{rad/s}$ . ✓ The equality of all three torus frequencies is why the orbit is a closed ellipse rather than a space-filling rosette — the geometric fingerprint of the hidden Runge-Lenz symmetry, unpacked in Degeneracy and Commensurate Frequencies.

The history band

Origins · 1 minJoseph Liouville and how this was discovered

Joseph Liouville

1809–1882 · École Polytechnique analyst who defined integrability, founded a great journal, and proved phase volume incompressible

Liouville was the great connector of nineteenth-century French mathematics: founder and forty-year editor of the Journal de Mathématiques Pures et Appliquées ("Liouville's Journal"), discoverer of transcendental numbers, and rescuer of Galois's manuscripts from oblivion. Mechanics owes him twice. His 1838 theorem on the conservation of phase-space volume underlies statistical mechanics; his 1855 observation — that $n$ integrals in involution suffice to solve a $2n$ -dimensional Hamiltonian system by quadratures (integrations alone) — defined what "solvable" means in mechanics ever since. Characteristically, he proved it as a statement about reducing differential equations, a century before anyone saw the tori: the geometric reading had to wait for Arnold in 1963.

History

1687

— Newton solves the two-body problem; the three-body problem immediately resists all assaults

1838

— Liouville proves phase-space volume is conserved under Hamiltonian flow

1855

— Liouville shows n integrals in involution solve a 2n-dimensional system by quadratures

1890

— Poincaré proves the three-body problem has no additional analytic integrals — genericity of non-integrability

1963

— Arnold completes the geometric statement: compact level sets of involutive integrals are tori with action-angle coordinates

1965

— Zabusky and Kruskal discover solitons in KdV, opening infinite-dimensional integrability and inverse scattering

1967

— Toda finds his integrable lattice; Flaschka's Lax pair later explains its hidden commuting integrals

Going further

References. Arnold, Mathematical Methods of Classical Mechanics, 2nd ed., §49 (Liouville's theorem on integrable systems) and §39–41 · Goldstein, Poole & Safko, Classical Mechanics, 3rd ed., §10.6–10.8 · Landau & Lifshitz, Mechanics, 3rd ed., §49–50 · José & Saletan, Classical Dynamics, Ch. 6 · Babelon, Bernard & Talon, Introduction to Classical Integrable Systems, Ch. 1–2 The Lie-bracket identity $[X_F,X_G]=X_{\{G,F\}}$ follows from the Jacobi identity; the isotropic-subspace dimension bound follows from symplectic linear algebra; and the Kepler energy-action relation is $E=-mk^2/[2(I_r+I_\theta+I_\varphi)^2]$ . The numeric examples are: 2D oscillator with $E_x=0.40$ J, $E_y=0.10$ J, $\omega=2.0$ rad/s giving $I_x=0.20$ J·s, $I_y=0.050$ J·s; LEO satellite with $a=7.00\times10^6$ m giving $J=5.28\times10^{13}$ J·s, $\omega=1.078\times10^{-3}$ rad/s $=2\pi/T$ with $T\approx97$ min.

Builds on: Action-Angle Variables · Hamilton-Jacobi Equation · Separation of Variables in Hamilton-Jacobi Theory · Time-Independent Hamilton-Jacobi Equation

Leads to: Liouville-Arnold Theorem · Action Integrals over Invariant Tori · Frequencies from Action-Angle Variables · Degeneracy and Commensurate Frequencies · Staeckel Conditions for Separability

Is the Hamiltonian itself always one of the n integrals?

By convention, yes — and it costs nothing. Any autonomous Hamiltonian Poisson-commutes with itself, so $H$ can always be enrolled as $F_1$ . The real content of the definition is finding the other $n-1$ integrals that commute both with $H$ (conservation) and with each other (involution). Occasionally it is convenient to use a different member of the family as "the Hamiltonian" — any $F_i$ generates a legitimate flow on the same tori, just with different frequencies. This freedom is exploited in the Liouville-Arnold Theorem proof, where all $n$ flows are treated on an exactly equal footing.

Where do the extra commuting integrals actually come from in practice?

Almost always from symmetry, visible or hidden. Visible: every cyclic coordinate donates its conserved momentum ( $L_z$ from axial symmetry, $p_\varphi$ and $p_\psi$ for the Lagrange top), and Noether's theorem industrializes the harvest. Hidden: the Kepler problem's Runge-Lenz vector and the isotropic oscillator's quadrupole-like integrals come from symmetries of phase space that do not act on configuration space alone — their fingerprint is degeneracy of the frequencies. And sometimes the source is algebraic: for the Toda lattice the integrals are eigenvalues of a Lax matrix $L$ obeying $\dot L=[M,L]$ , with no geometric symmetry in sight. That is why integrability, not symmetry, is the primitive notion.

Why does involution make the level set a torus rather than a sphere or something else?

Because the $n$ commuting flows provide $n$ everywhere-independent, everywhere-nonvanishing tangent vector fields on the $n$ -dimensional level set. Most manifolds cannot support even one nonvanishing field — the 2-sphere famously cannot (hairy ball theorem) — and demanding $n$ independent commuting ones is enormously restrictive: it makes the manifold a homogeneous space of $\mathbb{R}^n$ , i.e. $\mathbb{R}^n/\Lambda$ for some discrete stabilizer lattice $\Lambda$ . Compactness forces $\Lambda$ to have full rank $n$ , and $\mathbb{R}^n$ modulo a rank- $n$ lattice is the $n$ -torus. Nothing about the specific Hamiltonian enters — only the count and commutativity of the flows.

Is every separable system integrable, and every integrable system separable?

The first implication is a theorem: separation of the Hamilton-Jacobi equation produces $n$ separation constants which are automatically independent integrals in involution — the separated momenta pair disjoint variables, so all brackets vanish termwise. The converse is false and importantly so. The Toda lattice is Liouville integrable (Lax-pair integrals), yet no coordinate system separates its Hamilton-Jacobi equation in the classical Staeckel sense; the same holds for the Calogero-Moser system and most "modern" integrable models. Separability is a coordinate-dependent accident with a precise characterization (the Staeckel Conditions for Separability); integrability is the coordinate-free, structural property.

The theorem needs compact level sets. What actually goes wrong when they are not compact?

Nothing pathological — you just get cylinders instead of tori. The lattice $\Lambda\subset\mathbb{R}^n$ of flow-times returning a point to itself can have rank $k<n$ , and then $M_f\cong\mathbb{T}^k\times\mathbb{R}^{n-k}$ : $k$ winding directions and $n-k$ runaway directions. Physically the $\mathbb{R}$ factors are unbounded motions — a scattering orbit in the Kepler problem ( $E>0$ ), a particle sliding off to infinity. Action variables exist only for the torus directions; the non-compact directions keep honest momentum-like coordinates. The hyperbolic Kepler orbit is the standard picture: one angle (orientation on the branch) and one unbounded coordinate.

How can motion on a torus be dense but never repeat — does it not run out of room?

It does not, because the torus is a continuum and the orbit is a one-dimensional thread of measure zero. With incommensurate frequencies ( $m\cdot\omega\neq0$ for every nonzero integer vector $m$ ), the trajectory returns arbitrarily close to its start — Poincaré recurrence — without ever exactly closing; over time it visits every neighbourhood of every point, uniformly (Weyl equidistribution). This "quasi-periodic" motion is the fingerprint of integrability: complicated-looking, never-repeating, yet utterly non-chaotic — nearby orbits separate only linearly in time, not exponentially. The Lissajous figures of a detuned 2D oscillator are the visual: an ever-denser weave that never quite overlaps itself.

If integrable systems are a measure-zero exception, why do they dominate every mechanics course?

Three defensible reasons. Pedagogically, they are the systems that can be finished — exact solutions teach structure that numerics cannot. Physically, many flagship systems really are integrable or nearly so: the two-body problem, rigid bodies with symmetry, small oscillations (always integrable — normal modes are commuting oscillators), the hydrogen atom. Strategically, perturbation theory needs a solvable zeroth order: the entire KAM edifice, and most of celestial mechanics, is built as $H_0(I)+\varepsilon H_1$ around an integrable skeleton. The solar system is "a perturbed integrable system" to remarkable accuracy — which is why the exception governs the generic case.

What is the quantum counterpart of Liouville integrability?

Replace functions by operators and Poisson brackets by commutators: a quantum system is (loosely) integrable when it has $n$ independent commuting conserved operators — a complete set of commuting observables that includes $\hat H$ . The consequences mirror the classical ones. Semiclassically, eigenstates condense onto quantized tori ( $I_i=(n_i+\mu_i/4)\hbar$ , the EBK rule of Bohr-Sommerfeld Quantization); spectrally, integrable systems show Poisson level statistics while chaotic ones show random-matrix level repulsion (Berry-Tabor vs Bohigas-Giannoni-Schmit). In many-body physics the notion sharpens: Bethe-ansatz-solvable chains carry infinitely many commuting charges, which block thermalization and force generalized Gibbs ensembles — classical involution echoing in cold-atom experiments.