Liouville's Theorem and Phase-Space Volume

There are two ways to discover that Hamiltonian flow preserves phase-space volume. The first — the route of Liouville's Theorem in the phase-space chapter — treats the ensemble as a fluid, computes the divergence of the velocity field $(\dot q,\dot p)$ , finds it identically zero, and reads off incompressibility. It is a beautiful argument, and that companion article develops everything it touches: the Liouville equation for $\rho$ , equilibrium ensembles, filamentation, entropy, attractors. This page tells the other story, the one that belongs to the canonical-transformations chapter you have just climbed. Seen from here, volume conservation is not a fact about fluids at all. It is a one-line corollary of the symplectic condition. Time evolution over any interval is a canonical transformation — the flow $H$ generates, assembled from the infinitesimal steps of this chapter — and every canonical transformation has Jacobian determinant exactly $+1$ , because $M^{\mathsf T}JM=J$ survives taking a determinant. No divergence, no continuity equation: the theorem falls out of the same algebra that certified your generating functions and Poisson brackets. The symplectic route buys more than elegance. It reveals that total volume is only the top rung of a whole ladder of conserved quantities — the Poincaré integral invariants $I_1,\dots,I_n$ , one for each even dimension — and that "volume-preserving" is a strictly weaker property than "canonical" the moment you have two or more degrees of freedom. That distinction is invisible from the fluid picture, yet it is the reason symplectic integrators outperform generic volume-preserving schemes, and it culminates in Gromov's astonishing non-squeezing theorem. One theorem, two proofs, and the second proof knows more.

Key takeaways

Time evolution $\Phi_t$ is a canonical transformation: its Jacobian $M$ obeys the symplectic condition $M^{\mathsf T}JM=J$ at every instant.
Taking determinants gives $(\det M)^2=1$ , and continuity from $M(0)=I$ (or the Pfaffian) sharpens it to $\det M=+1$ : phase-space volume is exactly conserved.
The invariant volume element is the top wedge power of the symplectic form, $\omega^n/n!$ with $\omega=\sum_i\mathrm{d}q_i\wedge\mathrm{d}p_i$ — volume conservation is area conservation compounded $n$ times.
Every intermediate power $\omega^k$ is conserved too: the Poincaré integral invariants $I_1,\dots,I_n$ , with $I_n$ = Liouville's theorem and $I_1$ = the conjugate-plane area sum.
For $n\ge 2$ , volume-preserving is strictly weaker than canonical: $\det M=1$ is one equation, $M^{\mathsf T}JM=J$ is $n(2n-1)$ — maps exist that keep the volume while wrecking every lower rung.
Key idea: Hamiltonian flow is a continuous canonical transformation, and every canonical transformation has unit Jacobian because $M^{\mathsf T}JM=J$ forces $\det M=+1$ — so the $2n$ -dimensional phase volume is invariant, along with the entire ladder of wedge-power invariants beneath it.

The core insight

Hamiltonian flow is a continuous canonical transformation, and every canonical transformation has Jacobian determinant $+1$ because $M^{\mathsf T}JM=J$ implies $(\det M)^2=1$ ; the $2n$ -dimensional volume element $\omega^n/n!$ is therefore invariant — and so is every lower wedge power of $\omega$ , which is more than plain volume conservation can say.

Where this lives today

The canonical-transformation reading of Liouville's theorem is embedded in nearly every long-time simulation running right now. Symplectic integrators — the leapfrog family, the splitting schemes of Poisson-Bracket Formulation of Dynamics, and their high-order descendants — are, by construction, discrete canonical maps: each step satisfies $M^{\mathsf T}JM=J$ exactly, so each step conserves the entire Poincaré ladder exactly, and the million-orbit stability of planetary ephemerides, molecular-dynamics runs, and accelerator tracking codes is Derivation A operating at machine precision. The design logic runs through this page's warning in reverse: because volume alone is too weak a certificate, integrator authors verify symplecticity of the update map — the bracket test — rather than watching energy or volume drift. The frontier is the gap between the rungs. Gromov's 1985 non-squeezing theorem says a canonical map can never squeeze a ball of radius $R$ through a hole of radius $r<R$ in any single conjugate plane, no matter how enormously it spreads the other coordinates — even though volume bookkeeping would happily permit it. This "symplectic camel" rigidity is the classical shadow of the uncertainty principle (a minimal $\Delta q_i\,\Delta p_i$ that no canonical trickery can beat), it constrains how much beam conditioning any conservative lattice can achieve beyond plain emittance conservation, and it seeded the field of symplectic capacities in mathematics. The fluid picture sees none of this: it is a theorem about the ladder, not the top rung.

The physical picture

Forget the fluid for a moment and think in the vocabulary of this chapter: charts and maps. The time- $t$ flow $\Phi_t$ takes each initial state $(q_0,p_0)$ to the evolved state $(q(t),p(t))$ . That is a map of phase space to itself — formally identical to the canonical transformations you have been testing with brackets and building with generating functions, except that nature composes this one for you. The whole content of Liouville's theorem is that $\Phi_t$ passes the canonicity test, and that passing it leaves the determinant of the Jacobian nowhere to go but $1$ .

Figure

The flow map through the canonical-transformation lens — a region is sheared by time evolution, but its Jacobian satisfies the symplectic condition, so the determinant and hence the volume cannot change

The picture to carry is a stack of accountants. The fluid proof audits one number — total volume — and certifies it unchanged. The symplectic proof audits a whole ledger: the area of the region's shadow on each conjugate $(q_i,p_i)$ plane, summed (that is $I_1$ , the quantity your canonicity test already guards); the four-dimensional shadows taken in conjugate pairs ( $I_2$ ); and so on up to the full $2n$ -volume ( $I_n$ ). A canonical map — and therefore Hamiltonian time evolution — balances every line of that ledger. A merely volume-preserving map balances only the last line, and for $n\ge2$ there are plenty of maps that do exactly that and no more.

Figure

The Poincaré ladder of invariants — canonical maps preserve every rung from the conjugate-plane area sum up to the full phase volume, while a merely volume-preserving map is only guaranteed the top rung

One caution the companion article states and we repeat, because it bites in this chapter too: conservation of volume says nothing about conservation of shape. The flow may stretch a region enormously along $q$ while crushing it along $p$ ; the symplectic condition polices the products, not the factors. What the canonical viewpoint adds is a sharp lower limit on the crushing — Gromov's non-squeezing theorem, at the end of this page — which no amount of volume bookkeeping could detect.

The theorem in one determinant

Recall the algebraic fingerprint of this chapter. A map $(q,p)\mapsto(Q,P)$ with Jacobian matrix $M=\partial(Q,P)/\partial(q,p)$ is canonical exactly when it satisfies the condition established in Symplectic Condition — the matrix form of the fundamental brackets of Fundamental Poisson Brackets:

M^{\mathsf T} J M = J,\qquad J=\begin{pmatrix}0&I_n\\-I_n&0\end{pmatrix} \qquad (1)

where: $M$ — the $2n\times2n$ Jacobian of the transformation; $J$ — the standard symplectic matrix; $I_n$ — the $n\times n$ identity. Equation (1) packs all $n(2n-1)$ independent fundamental-bracket conditions into one matrix identity; it is the same test Poisson Brackets as the Test for Canonicity runs bracket by bracket. Now do to (1) the crudest thing linear algebra allows: take the determinant of both sides. Determinants multiply, $\det M^{\mathsf T}=\det M$ , and $\det J=1$ , so the condition collapses to a single scalar sentence:

(\det M)^2=1\ \Longrightarrow\ \det M=+1 \qquad (2)

where: the sign is fixed at $+1$ , not merely $\pm1$ — Derivation A gives two independent reasons. Equation (2) is Liouville's theorem in embryo: a determinant is precisely the local volume-stretch factor of a map, and a map whose stretch factor is $1$ everywhere moves regions around without changing their content. Note what the argument did not use: no Hamiltonian, no equations of motion, no time. Unit Jacobian is a property of every canonical transformation — a coordinate change to action-angle variables, a generating-function map, anything that passes test (1). The change-of-variables formula turns (2) into the integral statement. For any region $V$ mapped to $V'=\Phi(V)$ ,

\int_{V'}\mathrm{d}^n Q\,\mathrm{d}^n P=\int_{V}|\det M|\;\mathrm{d}^n q\,\mathrm{d}^n p=\int_{V}\mathrm{d}^n q\,\mathrm{d}^n p \qquad (3)

where: $\mathrm{d}^n q\,\mathrm{d}^n p$ — the phase-space volume element. Reading (3) for the special canonical map "evolve for time $t$ " is Liouville's theorem; reading it for a coordinate change says the phase-space measure is chart-independent — the license behind every ensemble integral in statistical mechanics, and the reason the density of states does not care which canonical coordinates you compute it in.

Derivation AUnit Jacobian from the symplectic condition

We are proving that every matrix satisfying $M^{\mathsf T}JM=J$ has determinant exactly $+1$ . The determinant-squared step is two lines; the real content is pinning the sign, because $(\det M)^2=1$ alone would permit orientation-reversing "anticanonical" maps, and the difference matters — it is why Hamiltonian flow can never mirror-flip phase space. The strategy: take determinants for the magnitude, then use the Pfaffian — the polynomial square root of the determinant that every antisymmetric matrix possesses — for the sign. Watch that nothing here mentions dynamics: the result is pure linear algebra about the canonical group.

Step 1. Take the determinant of both sides of $M^{\mathsf T}JM=J$ . Multiplicativity gives

\det(M^{\mathsf T})\,\det J\,\det M=\det J

Step 2. Use $\det M^{\mathsf T}=\det M$ and cancel $\det J$ , which equals $1$ : the blocks of $J$ commute, so the block-determinant formula gives $\det J=\det\big(0\cdot0-(I_n)(-I_n)\big)=\det I_n=1$ . Then

(\det M)^2=1

Step 3. Fix the sign with the Pfaffian. Every real antisymmetric matrix $A$ has a polynomial invariant $\operatorname{Pf}(A)$ with $\operatorname{Pf}(A)^2=\det A$ and the transformation law $\operatorname{Pf}(B^{\mathsf T}AB)=\det(B)\operatorname{Pf}(A)$ for any $B$ . Apply it to the symplectic condition with $A=J$ , $B=M$ :

\operatorname{Pf}(M^{\mathsf T}JM)=\det(M)\operatorname{Pf}(J)\quad\text{and}\quad M^{\mathsf T}JM=J\ \Rightarrow\ \operatorname{Pf}(J)=\det(M)\operatorname{Pf}(J)

Since $\operatorname{Pf}(J)\neq0$ (it equals $\pm1$ depending on the coordinate ordering convention), divide: $\det M=+1$ . No symplectic matrix reverses orientation. Why a second argument is worth having: for the flow map specifically, continuity does the same job — $\Phi_0$ is the identity with $\det M=+1$ , and a continuously evolving determinant confined to the two-point set $\{+1,-1\}$ can never jump. The Pfaffian argument is stronger: it covers every canonical map, including discrete ones (a symplectic integrator step, a generating-function chart change) that are not connected to the identity through the family you happen to be using.

Step 4. Conclude with the change of variables: $|\det M|=1$ everywhere makes every region's volume invariant under the map, equation (3). For $n=1$ the symplectic condition reads $\{Q,P\}=\det M=1$ directly — condition and conclusion coincide, which is why one bracket decides everything in one degree of freedom.

Time evolution passes the test

Equation (2) applies to time evolution only if the flow map is actually canonical. This chapter has already assembled the pieces: an infinitesimal time step is an infinitesimal canonical transformation with generator $H$ — that is the content of Hamiltonian as Generator of Time Evolution, built on the F2 machinery of Infinitesimal Canonical Transformations — and compositions of canonical maps are canonical. What remains is to see the symplectic condition hold exactly along the flow, not just to first order in the step. The clean statement is differential. Let $M(t)=\partial(q(t),p(t))/\partial(q_0,p_0)$ be the Jacobian of the flow map at time $t$ ; differentiating Hamilton's equations with respect to the initial conditions gives its equation of motion,

\frac{\mathrm{d}M}{\mathrm{d}t}=J\,S(t)\,M,\qquad S(t)=\nabla^2 H\big(q(t),p(t)\big) \qquad (4)

where: $S(t)$ — the (symmetric) Hessian matrix of second derivatives of $H$ , evaluated along the trajectory; $J S$ — the linearization of the Hamiltonian vector field. Equation (4) is called the variational or tangent-map equation; numerically integrating it beside the trajectory is how one computes stability exponents in practice. Its structure — antisymmetric $J$ times symmetric $S$ — is exactly what makes the following derivation close.

Go deeper · Derivation B · 2 minThe flow map is canonical at every instant

Derivation BThe flow map is canonical at every instant

We are proving that $M(t)^{\mathsf T}JM(t)=J$ for all $t$ , which certifies the time- $t$ flow as a canonical transformation and (through Derivation A) delivers Liouville's theorem. Rather than compose infinitely many infinitesimal steps and track error terms, we use the sharpest version of the argument: show that the combination $M^{\mathsf T}JM$ is a constant of the motion of the variational equation — its time derivative vanishes identically because a symmetric Hessian sandwiched between $J$ 's cancels against its transpose. The one thing to watch is where symmetry of $S$ enters; that is the only property of $H$ the proof consumes, so the result holds for any $C^2$ Hamiltonian, driven or autonomous.

Step 1. Differentiate the candidate invariant along the flow, using the product rule and equation (4) with its transpose $\dot M^{\mathsf T}=M^{\mathsf T}S J^{\mathsf T}$ (since $S^{\mathsf T}=S$ ):

\frac{\mathrm{d}}{\mathrm{d}t}\left(M^{\mathsf T}JM\right)=\dot M^{\mathsf T}JM+M^{\mathsf T}J\dot M=M^{\mathsf T}S\,J^{\mathsf T}J\,M+M^{\mathsf T}JJ\,S\,M

Step 2. Evaluate the two matrix products of $J$ with itself. From $J^{\mathsf T}=-J$ and $J^2=-I$ : $\ J^{\mathsf T}J=+I$ and $JJ=-I$ . Substitute:

\frac{\mathrm{d}}{\mathrm{d}t}\left(M^{\mathsf T}JM\right)=M^{\mathsf T}S\,M-M^{\mathsf T}S\,M=0

Why the cancellation is not luck: the two terms are transposes of one another up to the antisymmetry of $J$ and the symmetry of $S$ — the same interplay that makes $\partial^2 H/\partial q\,\partial p$ cancel in the fluid proof's divergence. The two derivations are spending the identical mathematical coin ( $S=S^{\mathsf T}$ , Clairaut's theorem) in different currencies.

Step 3. Fix the constant by the initial condition. At $t=0$ the flow map is the identity, $M(0)=I$ , so $M^{\mathsf T}JM\big|_0=J$ . A quantity with zero time derivative keeps its initial value:

M(t)^{\mathsf T}\,J\,M(t)=J\qquad\text{for all }t

Step 4. Chain the conclusions. $\Phi_t$ satisfies the symplectic condition → Derivation A gives $\det M(t)=+1$ → equation (3) gives $V(t)=V(0)$ for every comoving region. Liouville's theorem, proved without ever mentioning a density. Why time-dependence of H costs nothing: the derivation used $S(t)$ as an arbitrary symmetric matrix at each instant — nothing required $H$ to be conserved or even autonomous. A driven system's flow is every bit as canonical, and its phase volume every bit as rigid; what fails for driven systems is energy conservation, a logically independent statement. This is the cleanest way to see that Liouville's theorem and energy conservation are different theorems with different hypotheses.

The wedge ladder — what volume conservation is made of

The determinant proof is efficient but slightly wasteful: it crushes the full symplectic condition — $n(2n-1)$ scalar equations — into the single scalar $\det M=1$ . The wedge-product language shows what the crushing discards. The symplectic form

\omega=\sum_{i=1}^{n}\mathrm{d}q_i\wedge\mathrm{d}p_i \qquad (5)

where: $\wedge$ — the antisymmetric (exterior) product of differentials, so $\mathrm{d}q\wedge\mathrm{d}p=-\mathrm{d}p\wedge\mathrm{d}q$ — is the coordinate-free packaging of the fundamental brackets: preserving $\omega$ is the symplectic condition, one form instead of a matrix identity. Geometrically, $\omega$ eats a small parallelogram in phase space and returns the sum of the signed areas of its shadows on the $n$ conjugate planes. That single sentence is worth internalizing — it is the invariant your canonicity test has been protecting all along. Because canonical maps preserve $\omega$ , they automatically preserve every wedge power $\omega^k=\omega\wedge\cdots\wedge\omega$ , and the top power is — up to a constant — the volume element itself:

\frac{\omega^{n}}{n!}=\mathrm{d}q_1\wedge\mathrm{d}p_1\wedge\cdots\wedge\mathrm{d}q_n\wedge\mathrm{d}p_n \qquad (6)

where: the $n!$ cancels the overcounting from the $n!$ orderings of the identical factors in the wedge power. Equation (6) is Liouville's theorem stated as geometry: the volume form is built out of the symplectic form, so no map can preserve $\omega$ and fail to preserve volume. Integrated over $2k$ -dimensional surfaces, the intermediate powers give the full ladder,

I_k=\int_{\sigma_{2k}}\omega^{k},\qquad k=1,\dots,n\quad\text{all conserved by the flow} \qquad (7)

where: $\sigma_{2k}$ — any $2k$ -dimensional comoving surface (or region boundary) in phase space; $I_1$ — the conjugate-plane area sum, the invariant beam physicists call emittance; $I_n$ — total volume, Liouville's theorem. These are the Poincaré integral invariants. The ladder is strictly ordered in strength: preserving $\omega$ (all rungs) implies preserving volume (top rung), but a map can preserve the top rung alone — Worked Example 2 constructs one with three lines of arithmetic and the canonicity test of this chapter catches it immediately.

Go deeper · Derivation C · 2 minVolume as the top rung — the ladder from the symplectic form

Derivation CVolume as the top rung — the ladder from the symplectic form

We are deriving equation (6) — that the $n$ -th wedge power of $\omega$ is $n!$ times the volume form — and with it the conservation of the whole ladder (7). The computation is a bookkeeping exercise in wedge algebra, and doing it once honestly demystifies both the $n!$ and the claim that "volume conservation is area conservation compounded $n$ times". The strategy: expand the power multinomially, kill the repeated factors by antisymmetry, and count the survivors. Keep track of one subtlety — the two-forms $\mathrm{d}q_i\wedge\mathrm{d}p_i$ commute with each other (even total degree), so no signs appear, only counting.

Step 1. Expand $\omega^n=(\sum_i \mathrm{d}q_i\wedge\mathrm{d}p_i)^n$ by the multinomial theorem. A generic term picks the factor $\mathrm{d}q_{i_1}\wedge\mathrm{d}p_{i_1}$ from the first copy, $\mathrm{d}q_{i_2}\wedge\mathrm{d}p_{i_2}$ from the second, and so on — a word $(i_1,\dots,i_n)$ with letters from $\{1,\dots,n\}$ .

Step 2. Kill the repeats. If any index appears twice, the term contains $\mathrm{d}q_i\wedge\mathrm{d}p_i\wedge\mathrm{d}q_i\wedge\mathrm{d}p_i$ , which vanishes because $\mathrm{d}q_i$ appears twice in a wedge product. Only words using each index exactly once survive: the $n!$ permutations of $(1,\dots,n)$ .

Step 3. Reorder each survivor. Each block $\mathrm{d}q_i\wedge\mathrm{d}p_i$ is a form of even degree two, and even-degree forms commute through anything: moving a block past another block flips two pairs of odd factors, sign $(-1)^{2\cdot2}\cdot(-1)^{0}=+1$ . Every permutation therefore reorders to the same standard product $\mathrm{d}q_1\wedge\mathrm{d}p_1\wedge\cdots\wedge\mathrm{d}q_n\wedge\mathrm{d}p_n$ with coefficient $+1$ .

Step 4. Count: $n!$ identical surviving terms give $\omega^n=n!\,\mathrm{d}q_1\wedge\mathrm{d}p_1\wedge\cdots$ , which is equation (6). Since Derivation B showed the flow preserves $\omega$ (that is what $M^{\mathsf T}JM=J$ says in form language), it preserves $\omega^k$ for every $k$ — the integrand of each $I_k$ is transported intact, so each integral over a comoving surface is constant. The ladder stands. Why the ladder outranks the determinant: conversely, preserving the top power alone does not force preservation of $\omega$ — the determinant is one polynomial condition on $M$ , the form condition is $n(2n-1)$ . The gap between them, empty for $n=1$ , is where volume-preserving-but-not-canonical maps live for $n\ge2$ , and it is the precise mathematical home of the distinction between a merely volume-preserving integrator and a symplectic one. Two quick limiting checks: for $n=1$ the ladder collapses to its single rung $I_1=\oint\mathrm{d}q\wedge\mathrm{d}p$ = area = volume, and canonical ⇔ area-preserving-with-orientation, consistent with the one-bracket test; for $n=2$ , $\omega^2=2\,\mathrm{d}q_1\wedge\mathrm{d}p_1\wedge\mathrm{d}q_2\wedge\mathrm{d}p_2$ , matching $n!=2$ .

Limiting cases

Case	Result	Meaning
$n=1$	$\det M=1\iff\{Q,P\}=1$	volume-preserving and canonical coincide; the ladder has one rung
$n\ge2$	$\det M=1\ \not\Rightarrow\ M^{\mathsf T}JM=J$	volume is strictly weaker than symplectic — see Worked Example 2
Driven $H(q,p,t)$	Derivation B unchanged ( $S(t)$ arbitrary symmetric)	volume conserved instant by instant; energy conservation is a separate theorem
Dissipation added	flow no longer canonical; $\det M=e^{-\gamma t}<1$	contraction onto attractors — the regime the companion article dissects
Discrete canonical map	symplectic integrator step: all $I_k$ exact	volume exact, energy only approximate (shadow Hamiltonian)
Linearized flow	$M\in Sp(2n,\mathbb R)$	eigenvalues come in $\lambda,1/\lambda$ pairs — stretch is always paid for by squeeze
$\hbar$ -side image	$\mathrm{d}^nq\,\mathrm{d}^np/(2\pi\hbar)^n$ counts states	unitary evolution is the quantum Liouville; the measure quantization uses is the invariant one

Common trapA map that preserves phase-space volume must be canonical — false in two or more degrees of freedom

The implication runs one way only. Canonical ⇒ unit Jacobian is Derivation A; the converse holds only for $n=1$ , where the single bracket $\{Q,P\}$ is the determinant. For $n\ge2$ , compare the count: $\det M=1$ is one scalar condition, while $M^{\mathsf T}JM=J$ imposes $n(2n-1)$ — six for $n=2$ , fifteen for $n=3$ . The gap is populated by perfectly smooth, invertible, volume-preserving maps that shred the symplectic structure: Worked Example 2 exhibits one that inflates the area invariant $I_1$ by a factor $2.5$ while keeping the volume exactly fixed. Practical corollaries: never certify a numerical scheme as "symplectic" by checking volume (or energy!) alone — run the bracket test of Poisson Brackets as the Test for Canonicity on its update map; and never quote Liouville's theorem as evidence that a transformation is canonical. Volume is the weakest rung of the ladder, not the ladder.

Worked examples

1 · The oscillator flow map, and the integrator that fails the test (confidence builder)

For the harmonic oscillator with $m=0.50\ \mathrm{kg}$ and $\omega=2.0\ \mathrm{rad/s}$ , write the time- $t$ flow map explicitly, verify $\det M=1$ at $t=0.40\ \mathrm{s}$ , then compute the volume inflation of the explicit-Euler discretization with step $\delta t=0.010\ \mathrm{s}$ after $10^4$ steps, and repair it with a one-line change.

Step 1. The exact flow. Hamilton's equations for $H=p^2/2m+\tfrac12 m\omega^2q^2$ have the well-known solution, linear in the initial data:

\begin{pmatrix}q(t)\\p(t)\end{pmatrix}=M(t)\begin{pmatrix}q_0\\p_0\end{pmatrix},\qquad M(t)=\begin{pmatrix}\cos\omega t & \dfrac{\sin\omega t}{m\omega}\\[4pt] -m\omega\sin\omega t & \cos\omega t\end{pmatrix}

Step 2. Test it at $t=0.40\ \mathrm{s}$ , where $\omega t=0.80\ \mathrm{rad}$ , $\cos0.80=0.69671$ , $\sin0.80=0.71736$ , and $m\omega=1.0\ \mathrm{kg/s}$ :

\det M=\cos^2\omega t-\left(\frac{\sin\omega t}{m\omega}\right)(-m\omega\sin\omega t)=0.69671^2+0.71736^2=0.48540+0.51460=\mathbf{1.00000}

The identity $\cos^2+\sin^2=1$ makes the determinant exactly $1$ at every $t$ — the flow map is a rotation in the scaled phase plane, and for $n=1$ that single number certifies full canonicity.

Step 3. Discretize badly. Explicit Euler updates $q'=q+\delta t\,p/m$ , $p'=p-\delta t\,m\omega^2 q$ , a linear map with

M_{\mathrm E}=\begin{pmatrix}1 & \delta t/m\\ -m\omega^2\delta t & 1\end{pmatrix},\qquad \det M_{\mathrm E}=1+\omega^2\delta t^2=1+(2.0)^2(0.010)^2=1.0004

Step 4. Compound the crime. Each step multiplies every phase-space area by $1.0004$ ; after $N=10^4$ steps (100 s of simulated time, about 32 periods) the accumulated factor is

(1.0004)^{10^4}=e^{10^4\ln(1.0004)}=e^{3.9992}\approx\mathbf{54.6}

A blob of initial conditions ends up 55 times fatter — the monotone energy inflation every student who codes Euler on an oscillator has watched, now diagnosed as a determinant sitting $4\times10^{-4}$ off the canonical group.

Step 5. The one-line repair. Update $p$ first, then use the new $p$ in the $q$ step (symplectic Euler): $p'=p-\delta t\,m\omega^2q$ , then $q'=q+\delta t\,p'/m$ , giving

M_{\mathrm{SE}}=\begin{pmatrix}1-\omega^2\delta t^2 & \delta t/m\\ -m\omega^2\delta t & 1\end{pmatrix},\qquad \det M_{\mathrm{SE}}=(1-\omega^2\delta t^2)+\omega^2\delta t^2=\mathbf{1}\ \text{exactly}

Same cost, same first-order accuracy, but the update map now lives in the canonical group: areas are exact forever, and the energy merely breathes around a shadow value instead of ramping. Liouville's theorem is not just something the true flow obeys — it is a design constraint you can build into an algorithm.

2 · Volume preserved, symplectic structure shredded (exam level)

In two degrees of freedom, consider the linear map $Q_1=\lambda q_1$ , $P_1=\lambda p_1$ , $Q_2=q_2/\lambda^2$ , $P_2=p_2$ with $\lambda=2$ . Show it preserves phase-space volume exactly, show it is not canonical, and quantify the damage on the Poincaré ladder for an initial product region with conjugate-plane areas $A_1=0.30\ \mathrm{J\,s}$ and $A_2=0.20\ \mathrm{J\,s}$ .

Step 1. Volume: the Jacobian is diagonal, $M=\mathrm{diag}(\lambda,\ \lambda,\ 1/\lambda^2,\ 1)=\mathrm{diag}(2,\ 2,\ 0.25,\ 1)$ in the ordering $(q_1,p_1,q_2,p_2)$ , so

\det M=2\times2\times0.25\times1=\mathbf{1}

Every $4$ -volume is exactly preserved: the map satisfies Liouville's conclusion perfectly.

Step 2. Canonicity: run the fundamental-bracket test of this chapter. $\{Q_1,P_1\}=\dfrac{\partial Q_1}{\partial q_1}\dfrac{\partial P_1}{\partial p_1}=\lambda\cdot\lambda=\mathbf{4}\neq1$ , and $\{Q_2,P_2\}=\dfrac{1}{\lambda^2}\cdot1=\mathbf{0.25}\neq1$ . Two fundamental brackets fail (the cross brackets $\{Q_1,Q_2\}$ , $\{Q_1,P_2\}$ , etc. all vanish correctly), so $M^{\mathsf T}JM\neq J$ : not canonical, despite the unit determinant.

Step 3. Read the damage on the ladder. Take the initial region to be a product $D_1\times D_2$ of disks in the two conjugate planes. The map scales areas plane by plane by exactly the failed brackets: $A_1'=\lambda^2A_1=4\times0.30=1.20\ \mathrm{J\,s}$ and $A_2'=A_2/\lambda^2=0.20/4=0.050\ \mathrm{J\,s}$ . The two rungs:

I_1: \quad A_1+A_2=0.50\ \mathrm{J\,s}\ \longrightarrow\ A_1'+A_2'=1.20+0.050=\mathbf{1.25\ J\,s}\quad(\times2.5,\ \text{broken})

I_2: \quad A_1A_2=0.060\ (\mathrm{J\,s})^2\ \longrightarrow\ A_1'A_2'=1.20\times0.050=\mathbf{0.060\ (J\,s)^2}\quad(\text{preserved})

Step 4. Interpret. The map robs conjugate plane 2 to pay conjugate plane 1 — a currency exchange the top rung cannot see, because the product of the areas is deaf to transfers between the factors. Any genuinely canonical map would have held each bracket, hence both rungs, fixed. No Hamiltonian, in any amount of time, can generate this map: Derivation B guarantees the flow preserves $I_1$ , and this map does not.

Step 5. The exam moral. When asked "is this transformation canonical?", the determinant is never sufficient evidence for $n\ge2$ — compute the fundamental brackets (six independent ones for $n=2$ , counting antisymmetry). And when asked what Liouville's theorem says, remember it is the weakest consequence of canonicity: the flow conserves the whole ladder, of which volume is only the last, bluntest rung.

Origins · 1 minHenri Poincaré and how this was discovered

Henri Poincaré

1854–1912 · Paris; École Polytechnique and the Sorbonne

Poincaré met phase-space volume where it mattered most: the three-body problem. His prize memoir of 1890 for King Oscar II of Sweden — famously corrected after Phragmén spotted an error, at a cost to Poincaré exceeding the prize money — introduced the integral invariants $I_k$ as the conserved integrals of Hamiltonian flow and wielded the top one, Liouville's volume, to prove the recurrence theorem: a bounded incompressible flow must return arbitrarily close to its starting state. The same memoir contains the first glimpse of homoclinic chaos, volume-preserving yet unpredictable — a combination that seemed paradoxical then and is a commonplace of dynamics now. Poincaré treated the invariant ladder as the true skeleton of mechanics, a view vindicated a century later when Gromov's non-squeezing theorem (1985) showed symplectic maps are rigid in ways volume accounting cannot detect, founding modern symplectic topology on precisely the gap between $\det M=1$ and $M^{\mathsf T}JM=J$ that this page maps out.

History

1838

— Liouville publishes the volume identity as a lemma on linear differential equations, with no mechanics in sight

1871

— Boltzmann carries the theorem into kinetic theory, where it underwrites the ensemble methods of statistical mechanics

1890

— Poincaré's three-body memoir introduces the integral invariants and derives recurrence from volume conservation

1902

— Gibbs's Elementary Principles founds equilibrium statistical mechanics on the invariance of the phase measure

1974

— Arnold's Mathematical Methods recasts the theorem as the invariance of the top wedge power of the symplectic form

1985

— Gromov's non-squeezing theorem proves symplectic maps are strictly more rigid than volume-preserving ones, opening symplectic topology

Going further

References. Goldstein, Poole & Safko, Classical Mechanics, 3rd ed., §9.5 and §9.9 (canonical/Poincaré invariants; Liouville's theorem) · Landau & Lifshitz, Mechanics, 3rd ed., §46 (Liouville's theorem) · Arnold, Mathematical Methods of Classical Mechanics, Ch. 8 (symplectic structure and integral invariants) · Hairer, Lubich & Wanner, Geometric Numerical Integration, Ch. VI (symplectic integrators) · McDuff & Salamon, Introduction to Symplectic Topology (non-squeezing, capacities) Four key checks underpin the article: (1) $\nabla\cdot\mathbf{v}=0$ follows from Hamilton's equations via equality of mixed partials; (2) $M^{\mathsf T}JM=J\Rightarrow(\det M)^2=1$ by taking determinants with $\det J=1$ ; (3) $\partial_t\rho=-\{\rho,H\}$ reduces to $\mathrm{d}\rho/\mathrm{d}t=0$ via the convective derivative; (4) the damped-oscillator velocity-field divergence equals $-\gamma$ . No empirical numerical constants are asserted. Key numerics: the oscillator flow-map determinant $0.48540+0.51460=1.00000$ at $\omega t=0.80$ ; the Euler inflation $(1.0004)^{10^4}=e^{3.9992}\approx54.6$ against the exact symplectic-Euler determinant $1$ ; the ladder example $I_1:0.50\to1.25\ \mathrm{J\,s}$ with $I_2=0.060\ (\mathrm{J\,s})^2$ fixed. The fluid-picture derivations and statistical-mechanics consequences live in the companion article Liouville's Theorem and are deliberately not duplicated here.

Builds on: Symplectic Condition · Definition of Canonical Transformations · Poisson Brackets: Definition · Poisson-Bracket Formulation of Dynamics Leads to: Symplectic Group and Phase-Space Geometry · Poisson Brackets as a Test for Canonicity · Algebra of Poisson Brackets

The phase-space chapter already has a full Liouville article. What does this one add that it does not?

Division of labor, stated plainly: the companion Liouville's Theorem proves volume conservation from the divergence-free velocity field and then develops what the theorem is for — the Liouville equation, equilibrium ensembles, entropy, filamentation, attractors. This page proves the same theorem from the symplectic condition and develops what the theorem is made of: the unit Jacobian of every canonical map, the invariance of $\omega$ and its wedge powers, the full Poincaré ladder, and the strict gap between volume-preserving and canonical for $n\ge2$ . If you need consequences for ensembles, go there; if you need to know why symplectic integrators work, why $\det M=+1$ exactly, or why a volume check can never certify canonicity, the answers live on this side. The two proofs spend the same mathematical coin — symmetry of second derivatives against antisymmetry of the canonical structure — in different currencies.

Why is the determinant exactly +1 rather than just plus-or-minus 1?

Two independent reasons, useful in different situations. For a flow, continuity: $M(0)=I$ has determinant $+1$ , the determinant varies continuously with $t$ , and a continuous function trapped in the two-point set $\{+1,-1\}$ cannot jump. For an arbitrary canonical map — including discrete ones not obviously connected to the identity — the Pfaffian: every real antisymmetric matrix has a polynomial invariant with $\operatorname{Pf}(B^{\mathsf T}AB)=\det(B)\operatorname{Pf}(A)$ , and applying this to $M^{\mathsf T}JM=J$ gives $\det M=+1$ directly, since $\operatorname{Pf}(J)=1\neq0$ . Physical content: canonical maps preserve orientation — Hamiltonian evolution can shear and fold phase space but never mirror-reflect it. A hypothetical $\det M=-1$ map would preserve volume too, but it lies outside the canonical group.

Is volume preservation equivalent to being canonical or not? I have seen both claims.

Both circulate because the answer depends on $n$ . For one degree of freedom they coincide exactly: phase space is a plane, the only fundamental bracket is $\{Q,P\}$ , and it literally equals $\det M$ — so $\det M=1$ is the entire symplectic condition. That special case is where the folklore "canonical = area-preserving" comes from, and it is true (with orientation). For $n\ge2$ the equivalence fails, and the counting says by how much: $\det M=1$ is one scalar equation, $M^{\mathsf T}JM=J$ imposes $n(2n-1)$ — six conditions for $n=2$ . Worked Example 2 lives in the gap: unit determinant, two broken brackets, ladder rung $I_1$ inflated by $2.5$ . So quote the equivalence only in the phase plane; everywhere else, canonical is strictly stronger.

What do the intermediate invariants I_k actually measure, physically?

Project your comoving region onto each conjugate plane $(q_i,p_i)$ and add the signed shadow areas: that sum is $I_1$ , the quantity a beam physicist calls (total) emittance. The higher $I_k$ do the same with $2k$ -dimensional shadows taken over all choices of $k$ distinct conjugate pairs at once — for $n=3$ , $I_2$ sums the three 4-volumes of shadows on the $(1,2)$ , $(1,3)$ , $(2,3)$ pair-planes. Only conjugate pairings enter; shadows on non-conjugate planes like $(q_1,q_2)$ are not protected, which is exactly why a cloud may spread without bound in configuration space. The flow conserves every rung simultaneously — a far tighter corset than volume alone, and the reason "the ellipse rotates and shears but its area survives" remains true plane by plane, not just in aggregate.

Does the symplectic proof still work when H depends explicitly on time?

Yes, and the proof makes the reason unusually transparent: Derivation B differentiates $M^{\mathsf T}JM$ along the flow and uses only one property of the Hamiltonian — that its Hessian $S(t)$ is symmetric, which any $C^2$ function's second-derivative matrix is, driven or not. Energy conservation never enters. So a parametrically driven oscillator, a particle in a ramping trap, an RF-modulated accelerator lattice: all have exactly volume-preserving, fully canonical flows at every instant, while their energies wander freely. Keep the two theorems apart — Liouville's theorem needs Hamiltonian structure; energy conservation needs Hamiltonian structure plus time-translation symmetry, the $G=H$ diagonal of Symmetries and Conservation Laws in Hamiltonian Form.

How can Gromov non-squeezing be true if I am free to squeeze one coordinate as thin as I like?

You are free to squeeze — but not in a conjugate pair simultaneously. Take a ball of radius $R$ : its shadow on the $(q_1,p_1)$ plane has area $\pi R^2$ . Gromov's theorem says no canonical map can fit the ball inside a cylinder $q_1^2+p_1^2\le r^2$ with $r<R$ — the conjugate-plane shadow can never be compressed below its initial disk — even though volume conservation alone would allow it (make the cylinder long enough and its volume is infinite anyway). Squeezing $q_1$ against $q_2$ , or $q_1$ against $p_2$ , remains perfectly legal; the protection is pair-wise, a strengthening of the $I_1$ story into a genuine topological rigidity that no finite list of integral invariants implies. It is the sharpest known classical echo of the uncertainty principle: a minimal conjugate-plane footprint that survives every canonical manipulation.

Why do symplectic integrators conserve volume exactly but energy only approximately? Should it not be the other way around?

Because the two conservation laws have different mechanisms, and the integrator can only inherit one. Volume (and the whole ladder) is conserved by any canonical map — it is a property of the group the map belongs to, not of the particular Hamiltonian. A symplectic integrator's step is constructed to be an exactly canonical map (a composition of exact sub-flows, as in the splitting schemes of Poisson-Bracket Formulation of Dynamics), so it conserves volume with zero error, forever. Energy conservation, by contrast, requires the map to be the flow of that specific $H$ — and the discrete step is not; it is the exact flow of a nearby "shadow" Hamiltonian $H+O(\delta t^k)$ . Hence the signature behavior of Worked Example 1's repaired scheme: volume exact, energy oscillating boundedly around a shadow value, never drifting. If a scheme's energy drifts monotonically, its update map has left the canonical group — check its determinant and brackets.

What happens to Liouville volume in quantum mechanics?

It becomes the state count. Semiclassically, each quantum state occupies a cell of volume $(2\pi\hbar)^n$ in phase space, so the invariant measure $\mathrm{d}^nq\,\mathrm{d}^np/(2\pi\hbar)^n$ counts states — and Liouville's theorem is what makes that count independent of the canonical coordinates used and constant under evolution, the classical skeleton of unitarity. The dynamical statement transcribes through the bracket-to-commutator dictionary: the density operator obeys $i\hbar\,\partial_t\hat\rho=[\hat H,\hat\rho]$ , unitary evolution preserving $\operatorname{Tr}\hat\rho$ and all spectral content, exactly as the classical flow transports $\rho$ without compression. Even Gromov's non-squeezing has a quantum reading: the protected conjugate-plane disk of area $\pi R^2$ is the classical ancestor of $\Delta q\,\Delta p\ge\hbar/2$ . The measure, the evolution, and the rigidity all quantize along the lines this chapter drew.