Liouville's Theorem

Liouville's theorem is not a museum piece — it is a working constraint in every field that pushes clouds of particles through phase space. Accelerator physicists live by it as the conservation of emittance: the phase-space area of a particle beam cannot be reduced by any conservative optics, so squeezing a beam tighter in position necessarily spreads it in momentum. Beating this limit — "beam cooling" at facilities like CERN and Fermilab — requires deliberately dissipative tricks (stochastic or electron cooling) that add exactly the non-Hamiltonian term Liouville forbids, and the theorem is what tells engineers precisely how much extra machinery they must build. The same law governs plasma simulations (the Vlasov equation is Liouville's theorem for a collisionless plasma), galactic dynamics (stars stream through a $6$-dimensional phase space whose fine-grained density is frozen, so galaxies "phase-mix" rather than truly relax), and the design of symplectic integrators — numerical schemes engineered to preserve phase-space volume exactly, so that long molecular-dynamics and planetary simulations don't spuriously gain or lose energy over billions of steps. Wherever a conserved probability rides a Hamiltonian flow, Liouville is quietly setting the rules.

Return to the ink. What Liouville's theorem forbids is not the stretching — the droplet is free to filament into a shape of unbounded complexity — but any change in the amount. Ride along with a single ink particle and the local density around you is fixed forever; that is the content of $\mathrm{d}\rho/\mathrm{d}t = 0$. The stirring is essential (it is what will eventually let coarse observers see equilibration), but the volume is sacred. Why is the flow incompressible? Because $q$ and $p$ are yoked by one Hamiltonian. The velocity along a position axis is $\dot q_i = \partial H/\partial p_i$; the velocity along the conjugate momentum axis is $\dot p_i = -\partial H/\partial q_i$. When you ask how fast the flow is spreading — its divergence — the position term brings in $\partial^2 H/\partial q_i\partial p_i$ and the momentum term brings in $-\partial^2 H/\partial p_i\partial q_i$, and for any smooth $H$ these mixed partials are equal. They cancel, pair by pair, coordinate by coordinate. Nothing about the form of $H$ matters; only its smoothness.

A quick contrast fixes the idea. Add friction — a damped particle with $\dot p = -\gamma p - \partial V/\partial q$ — and now $\partial \dot p/\partial p = -\gamma \neq 0$. The flow is compressible; volume shrinks; trajectories spiral inward onto an attractor. That is exactly what conservative Hamiltonian dynamics cannot do. Liouville's theorem is the precise statement that closed mechanical systems have no attractors: they can mix, but they can never collapse onto a smaller set. The second figure shows the two fates side by side.

Because volume is preserved, the "obvious" uniform probability on an energy shell is dynamically consistent: spread the probability evenly over accessible microstates and the flow keeps it even. This is the dynamical justification hiding underneath the postulate of equal a priori probabilities — the flow cannot pile probability up anywhere, so uniform is a genuine steady state.

An observer standing at a fixed point in phase space and an observer riding along with the flow disagree about how $\rho$ changes, and reconciling them is the whole story. The comoving observer's rate of change is the material (convective) derivative — the local rate plus the streaming brought by the velocity field. Liouville's theorem is the statement that, for a Hamiltonian flow, this comoving rate is identically zero:

where: $\rho(q,p,t)$ — phase-space probability density; $q_i,p_i$ — generalized coordinates and conjugate momenta; $\dot q_i,\dot p_i$ — velocities set by Hamilton's equations; $3N$ — configuration-space degrees of freedom. Read it as a conservation law for a label: the value of $\rho$ is painted onto each system and carried along unchanged, no matter how the cloud deforms. The engine underneath is the vanishing divergence of the phase-space velocity field. This is where "a single Hamiltonian steers both $q$ and $p$" becomes an equation:

where: $\mathbf{v}$ — the $6N$-dimensional velocity field on phase space. A divergence of zero means the flow neither compresses nor rarefies the cloud anywhere; it is the local face of global volume conservation. Combining a divergence-free velocity with the continuity equation for probability collapses the streaming into a single compact object, the Poisson bracket, giving the differential form of the theorem — the Liouville equation:

where: $\{\cdot,\cdot\}$ — Poisson bracket; $H$ — the Hamiltonian generating the flow. The local rate of change of $\rho$ at a fixed point is entirely the streaming of a frozen profile past that point; it is developed in full in the companion article on the Liouville equation. Finally, two consequences worth stating as equations. Volume conservation of any comoving region $R$, and the constancy of the fine-grained Gibbs entropy that follows from it:

where: $\mathrm{d}\Gamma = \prod_i \mathrm{d}q_i\,\mathrm{d}p_i$ — the invariant Liouville measure; $S_G$ — fine-grained Gibbs entropy; $k_B$ — Boltzmann's constant. The right-hand equation is the sharpest expression of what Liouville forbids: exact mechanics cannot raise fine-grained entropy, full stop.

1 · A phase-space patch in a harmonic trap (confidence builder)

Take a single oscillator, $H = p^2/2m + \tfrac12 m\omega^2 q^2$, with $m=1\ \mathrm{kg}$ and $\omega = 1.3\ \mathrm{s^{-1}}$ (natural units). Start with a small square cloud of initial conditions spanning $\Delta q = 0.2$ and $\Delta p = 0.2$. Show its phase-space area is unchanged after evolving to $t = 1.7\ \mathrm{s}$, even though the square becomes a tilted parallelogram.

Step 1. Initial area. The starting cloud is a square: $A_0 = \Delta q\,\Delta p = 0.2 \times 0.2 = \boxed{0.0400}$ in phase-space units.

Step 2. The flow. Each corner $(q_0,p_0)$ maps to $q(t) = q_0\cos\omega t + (p_0/m\omega)\sin\omega t$, $p(t) = -m\omega q_0\sin\omega t + p_0\cos\omega t$. This is a linear map with matrix $M(t) = \begin{pmatrix}\cos\omega t & \sin\omega t/m\omega\\ -m\omega\sin\omega t & \cos\omega t\end{pmatrix}$.

Step 3. Area scaling factor. A linear map multiplies areas by $|\det M|$. Here $\det M = \cos^2\omega t + \sin^2\omega t = 1$, for every $t$ — no $\omega$, $m$, or $t$ survives.

Step 4. Final area. $A(1.7) = |\det M|\,A_0 = 1 \times 0.0400 = \boxed{0.0400}$. The square has become a rotated, sheared parallelogram, but its area is untouched — a direct, hand-checkable instance of Liouville's theorem. Numerically evolving the four corners with a shoelace-area calculation confirms $A(t)=0.0400$ to one part in $10^{12}$ at $t = 0.5, 1.7, 3.9,$ and $10.0$.

2 · Compression forbidden: why a conservative gas has no attractor (exam level)

A dilute gas of $N$ molecules starts spread over a phase-space volume $\Gamma_0$. A student claims that after long times the gas "settles down" and its representative cloud shrinks to a much smaller volume $\Gamma_\infty \ll \Gamma_0$, the way a damped pendulum settles to rest. Use Liouville's theorem to show this is impossible, and estimate how the claim would violate volume conservation for $N = 100$ molecules.

Step 1. State what Liouville forbids. For an isolated (Hamiltonian) gas, the comoving volume is exactly conserved: $\Gamma_\infty = \Gamma_0$. Any genuine shrinkage $\Gamma_\infty < \Gamma_0$ requires $\nabla\cdot\mathbf{v} < 0$ somewhere, i.e. a dissipative term absent from a closed system.

Step 2. Quantify the divergence a fake "settling" would need. If the cloud shrank by even a modest factor $f = \Gamma_\infty/\Gamma_0 = 0.1$ over a relaxation time $\tau$, the average divergence would have to be $\langle\nabla\cdot\mathbf{v}\rangle = (\ln f)/\tau = \ln(0.1)/\tau \approx -2.3/\tau$ — strictly negative, and with no term in Hamilton's equations to supply it.

Step 3. The resolution. The gas does reach equilibrium in every practical sense, but its fine-grained volume does not shrink. Instead the cloud filaments through the accessible region until, to any coarse observer, it looks uniformly spread. The volume $\Gamma_0$ is conserved exactly; only the apparent (coarse-grained) volume grows toward the full accessible shell.

Step 4. Contrast with dissipation. A damped pendulum genuinely shrinks its phase volume because $\partial\dot p/\partial p = -\gamma$; its energy leaks into the environment. The isolated gas has no such leak — the settling is an illusion of coarse-graining, not a contraction of phase volume. This is the sharp dividing line between mechanics and thermodynamics, and it is why the second law lives in the coarse description.

Texts. Reif, Fundamentals of Statistical and Thermal Physics, §2.2 · Pathria & Beale, Statistical Mechanics, 3rd ed., §2.2–2.3 · Landau & Lifshitz, Statistical Physics Part 1, §3 · Goldstein, Poole & Safko, Classical Mechanics, 3rd ed., §9.9 (symplectic structure).

Provenance. Base article rigorous and retained; upgraded with intuition-first figures, worked examples, and history. A verifier should confirm: incompressibility $\nabla\cdot\mathbf{v}=0$ for a general smooth $H$ (mixed-partial cancellation); the SHO flow-matrix determinant $=1$ for all $t$; the free-particle shear Jacobian $=1$; the numerical area conservation of a $0.2\times0.2$ patch to $<10^{-12}$; and $\mathrm{d}(\rho\ln\rho)/\mathrm{d}t=0$ following from $\mathrm{d}\rho/\mathrm{d}t=0$. All checks pass in verify_aq-sm-liouville-theorem.py.

Prerequisites. Phase space · phase-space density and the distribution function · microstates and macrostates. Leads to. The Liouville equation (the PDE form, $\partial_t\rho = -\{\rho,H\}$) · statistical ensembles (stationary $\rho$ as equilibrium) · the ergodic hypothesis (does one trajectory sample the whole shell?) · Gibbs entropy (why the fine-grained value is frozen).