Liouville's Theorem

Ask an accelerator physicist about Liouville's theorem and you will get a number: emittance, the phase-space area of a particle bunch, the quantity their whole machine is designed around. Under Hamiltonian beam optics — quadrupoles, dipoles, RF cavities — emittance cannot decrease: focus the beam to a smaller spot and the angular spread grows in exact compensation. Genuine cooling therefore requires deliberately non-Hamiltonian physics. Simon van der Meer's stochastic cooling — measure the bunch's fluctuations, feed back a correcting kick — shrank the antiproton beams that discovered the W and Z bosons and earned the 1984 Nobel Prize. Its descendants set the agenda still: the muon-collider design effort rests on ionization cooling (energy loss in absorbers, the only trick fast enough for particles that live 2.2 microseconds), demonstrated by the MICE collaboration in 2020, and CERN's LHC luminosity upgrade juggles emittance budgets through every stage of the injector chain. The theorem's statistical face is just as current. The Gibbs entropy $-k_B\!\int\!\rho\ln\rho\,\mathrm{d}\Gamma$ is exactly constant under Liouville flow — so the second law lives in coarse-graining, in the filament of the hero figure becoming indistinguishable from a uniform smear at any finite resolution. The quantum version of that tension is a live research industry: unitary evolution preserves von Neumann entropy just as Liouville preserves Gibbs's, and the eigenstate thermalization hypothesis — tested in today's cold-atom quantum simulators — is the working answer to how closed quantum systems nevertheless appear to thermalize.

Drop a spot of dye into the phase fluid — a small disk of nearby initial conditions — and let Hamilton's equations carry it. Points near the bottom of the disk sit at lower energy and circulate at a different rate than points near the top, so the disk shears. Keep going and it draws out into a long, curved filament, folded along the flow lines. It looks like the blob is being destroyed. It is not. Every intermediate snapshot below has exactly the same area, computed from the actual integrated flow:

That is the behavior of dye in water, not smoke in air: an incompressible fluid can be stirred into arbitrarily intricate shapes but never concentrated or diluted. The microscopic reason is a term-by-term cancellation: wherever the $q$-flow stretches the blob, the $p$-flow squeezes it at precisely the same rate, because both rates are the same mixed second derivative of the one function $H$ running the show. The cleanest place to see it is the hyperbolic flow of $H = qp$, which stretches $q$ by $e^{t}$ while squeezing $p$ by $e^{-t}$ — aggressive deformation, identically zero net flux out of any box:

Now contrast a system that is not Hamiltonian. Add a drag force $-b\dot x$ to an oscillator and the phase velocity field acquires a genuine sink: $\nabla\cdot\mathbf{v} = -b/m < 0$ everywhere. Blobs shrink exponentially and every trajectory spirals into the origin — the geometric face of dissipation, drawn in full in the damped-harmonic-oscillator article. Liouville's theorem is precisely the statement that conservative dynamics has no such drain:

Because the comoving density never changes, you can think of phase points as indestructible fluid particles that neither merge nor split: the flow is a continuous, volume-preserving relabeling of phase space. Stretching is always paid for by compression in the conjugate direction — and that bookkeeping, enforced by the symplectic pairing, is the entire content of the theorem.

Everything in this article is a consequence of one geometric statement. Carry a region $V(t)$ of phase space along the flow — every point in it evolving by Hamilton's equations — and its $2n$-dimensional volume is a constant of the motion. In symbols,

where: $V(t)$ — a region transported by the flow; $\mathrm{d}\Gamma$ — the phase-space volume element; $n$ — the number of degrees of freedom. Read it as the incompressibility of the phase fluid: shapes deform freely, measure never changes. Nothing about the shape is protected — the hero figure's filament is the generic fate. The engine underneath (1) is a property of the velocity field itself. Assemble the $2n$ rates given by Hamilton's equations into $\mathbf{v} = (\dot q_1,\dots,\dot q_n,\dot p_1,\dots,\dot p_n)$ and compute its divergence:

where: $\dot q_i = \partial H/\partial p_i$ and $\dot p_i = -\partial H/\partial q_i$ have been inserted before differentiating. Equality of mixed partials (Clairaut–Schwarz, valid for any twice-differentiable $H$) kills every term in the sum. Notice what the calculation does not use: no conservation of energy, no time-independence of $H$, no special form of the potential. The cancellation is a structural identity of the canonical equations. Now switch from geometry to statistics — the viewpoint this article owns. Describe an ensemble by a density $\rho(q,p,t)$, so that $\rho\,\mathrm{d}\Gamma$ counts the systems in each cell. Representative points are never created or destroyed, so $\rho$ obeys a continuity equation in $2n$ dimensions, exactly like charge or mass density in ordinary fluid dynamics:

where: $\rho\mathbf{v}$ — the current of phase points. This is bookkeeping, true for any flow, dissipative or not; the Hamiltonian structure has not yet been used. Combine (2) with (3): the divergence expands as $\nabla\cdot(\rho\mathbf{v}) = \mathbf{v}\cdot\nabla\rho + \rho\,(\nabla\cdot\mathbf{v})$, the compressive term drops, and the advective term $\mathbf{v}\cdot\nabla\rho$ is — substituting Hamilton's equations — nothing but a Poisson bracket. The result is the Liouville equation, the master equation of classical statistical mechanics:

where: $\mathrm{d}\rho/\mathrm{d}t$ — the total (comoving) derivative, the rate of change seen riding with the flow. An observer surfing a trajectory measures constant density forever; the density at a fixed location generally does change, at the rate $\partial\rho/\partial t = \{H,\rho\}$, as the filament sweeps past. Equation (4) is the evolution law $\dot F = \{F, H\}$ of any phase function — see the time-evolution article — applied to the density, with the sign flip inherited from the comoving picture. The geometric and kinetic statements meet in the Jacobian of the flow map. Let $\phi_t$ send initial conditions $(q_0, p_0)$ to $(q(t), p(t))$; a standard identity of flows says the Jacobian determinant grows at the rate of the divergence:

where: $J(t)$ — the volume-stretch factor of the time-$t$ map. With zero divergence, $J$ stays pinned at 1: the flow map is a volume-preserving change of coordinates. For an infinitesimal step this is visible by hand: the Jacobian matrix of $\phi_\varepsilon$ is $I + \varepsilon\,\mathbb{J}\,\mathrm{Hess}\,H$, whose determinant is $1 + \varepsilon\,\mathrm{tr}(\mathbb{J}\,\mathrm{Hess}\,H) + O(\varepsilon^2) = 1 + O(\varepsilon^2)$ — the trace vanishes by the same mixed-partial symmetry. Time evolution is in fact a canonical transformation, and every canonical transformation has unit Jacobian; Liouville is the time-flow special case of that symplectic fact. The payoff for statistical mechanics is immediate. Ask which densities are stationary — unchanged at every fixed point of phase space. By (4) that requires $\{\rho, H\} = 0$, and any smooth function of the Hamiltonian qualifies:

where: $f$ — any smooth function. Set $f(H) \propto \delta(H - E)$ and you have the microcanonical ensemble; $f(H) \propto e^{-\beta H}$, the canonical ensemble. Liouville's theorem is what makes these distributions self-consistent equilibria: uniform weight on equal phase volumes is preserved by the dynamics, so "equal a priori probability" is not eroded by time evolution. (It does not by itself force relaxation toward them — that is ergodicity's job, as the FAQ spells out.)

Which statement of the theorem answers which question

1 · How fast dissipation drains phase area (confidence builder)

A damped oscillator has mass $m = 0.50\ \mathrm{kg}$, spring constant $k = 2.0\ \mathrm{N/m}$, and linear drag coefficient $b = 0.10\ \mathrm{kg/s}$: $\ \dot x = v$, $\ \dot v = -(k/m)x - (b/m)v$. Find the divergence of its phase flow, the law by which a blob's area decays, and the area half-life.

Step 1. Compute the divergence in the $(x,v)$ plane. $\partial\dot x/\partial x = 0$ and $\partial\dot v/\partial v = -b/m$, so $\nabla\cdot\mathbf{v} = -b/m = -0.10/0.50 = \boxed{-0.20\ \mathrm{s^{-1}}}$ — uniform over the whole plane, and independent of the spring constant. (The same value results in the canonical $(x,p)$ plane with $p = mv$: the drag term $-\,(b/m)p$ in $\dot p$ has $p$-derivative $-b/m$.)

Step 2. Integrate the transport law. With constant divergence, $\mathrm{d}A/\mathrm{d}t = (\nabla\cdot\mathbf{v})A$ gives exponential decay, $A(t) = A_0\,e^{-bt/m} = A_0\,e^{-0.20\,t}$ — every blob, anywhere in the plane, regardless of shape.

Step 3. Solve for the half-life: $e^{-bt_{1/2}/m} = \tfrac12$ gives $t_{1/2} = \dfrac{m\ln 2}{b} = \dfrac{0.50 \times 0.6931}{0.10} = \boxed{3.47\ \mathrm{s}}$. After $10\ \mathrm{s}$ the area is $e^{-2} = 13.5\%$ of the original. The contrast figure above is this exact system: computed areas $0.363 \to 0.182 \to 0.091$ at $t = 0,\ 3.47,\ 6.93\ \mathrm{s}$ — one half-life each step.

Step 4. Read the moral. The number $-b/m$ is the failure of Liouville's theorem for this system, and its sign is why the damped oscillator can have an attractor at the origin while its undamped twin cannot.

2 · Free-streaming shear of a rectangle (exam level)

An ensemble of free particles of mass $m = 0.60\ \mathrm{kg}$ initially fills the rectangle $0 \le q \le 0.20\ \mathrm{m}$, $\ 0 \le p \le 0.30\ \mathrm{kg\,m/s}$ in the phase plane. Evolve for $t = 2.0\ \mathrm{s}$ and show explicitly — corner by corner — that the occupied area is unchanged, even though the region is violently sheared.

Step 1. Solve the dynamics exactly. For $H = p^2/2m$: $\ p(t) = p_0$ and $q(t) = q_0 + p_0 t/m$. The map is a pure shear along $q$ with shear coefficient $t/m = 2.0/0.60 = 3.33\ \mathrm{m/(kg\,m/s)}$.

Step 2. Push the corners through. $(0,0) \to (0,0)$; $\ (0.20, 0) \to (0.20, 0)$; $\ (0.20, 0.30) \to (0.20 + 3.33\times0.30,\ 0.30) = (1.20,\ 0.30)$; $\ (0, 0.30) \to (1.00,\ 0.30)$. The bottom edge stays put; the top edge slides right by a full metre.

Step 3. Compute the sheared area by the shoelace formula on the parallelogram $(0,0),(0.20,0),(1.20,0.30),(1.00,0.30)$: the cross terms give $2A = 0.20\times0.30 - 0 + 1.20\times0.30 - 1.00\times0.30 + 0 = 0.12$, so $A = \boxed{0.060\ \mathrm{J\,s}}$ — exactly the initial $\Delta q\,\Delta p = 0.20 \times 0.30 = 0.060\ \mathrm{J\,s}$. Equivalently: a shear matrix has determinant 1.

Step 4. Measure the distortion that hides inside "area conserved." The $q$-extent grew from $0.20\ \mathrm{m}$ to $1.20\ \mathrm{m}$ — a factor of 6 — while the momentum spread is untouched. Position uncertainty exploded, yet in phase-space measure nothing happened. This is the beam-physics statement that a drifting bunch debunches without any loss of phase-space quality.

Step 5. Check the gravity variant. Add uniform gravity, $H = p^2/2m + mgz$: now $p(t) = p_0 - mgt$ shifts every point down by the same amount while free-flight shear acts as before. A rigid translation composed with a unit-determinant shear still has Jacobian 1 — the parallelogram moves and tilts, its area stays $0.060\ \mathrm{J\,s}$. Area is conserved in both cases.

References. Goldstein, Poole & Safko, Classical Mechanics, 3rd ed., §9.9 · Landau & Lifshitz, Mechanics, 3rd ed., §46 footnote and statistical-physics companion · Morin, Introduction to Classical Mechanics, Ch. 15 (Hamiltonian formalism) · Taylor, Classical Mechanics, Ch. 13 · Arnold, Mathematical Methods of Classical Mechanics, §16

Provenance. Equations and derivations standard; verifier should recompute $\nabla\cdot\mathbf{v}=0$ from Hamilton's equations via equality of mixed partials, confirm $\mathrm{d}J/\mathrm{d}t=(\nabla\cdot\mathbf{v})J$ with $J(0)=1$, and check the Liouville equation derivation from the $2n$-dimensional continuity equation. Cross-references: Goldstein, Poole & Safko §9.9; Landau & Lifshitz Mechanics and Statistical Physics; Arnold §16; Taylor Ch. 13. No numerical constants asserted.

Builds on: Hamilton's Canonical Equations · Phase Space · Hamiltonian Flow and the Phase Fluid · The Hamiltonian Function. Leads to: Poincaré Recurrence Theorem · Microcanonical Ensemble · Fixed Points and Stability in Phase Space · Symplectic Geometry and Canonical Invariants.