Hamiltonian Flow and the Phase Fluid

Every serious long-time simulation of a Hamiltonian system now respects the flow's geometry on purpose. Symplectic integrators — leapfrog/Verlet in molecular dynamics, Wisdom–Holman maps in planetary dynamics — advance the state by maps satisfying $M^{\mathsf T}JM=J$ exactly, and the payoff is a beautiful inversion: a symplectic scheme solves, essentially exactly, the flow of a nearby "shadow Hamiltonian" $\tilde H=H+O(\Delta t^k)$. Energy errors therefore oscillate boundedly instead of drifting — why MD codes hold protein simulations together for microseconds of physical time, and why planetary ephemerides stay honest over the age of the solar system, where a non-symplectic Runge–Kutta of the same order would slowly spiral Mercury somewhere it has no business being. The 2026-era twist is machine learning: Hamiltonian and symplectic neural networks bake $X_H=J\nabla H$ into the architecture, so learned dynamics conserve by construction rather than by hope. Accelerator physics, meanwhile, lives and dies by the shear clock of Derivation C. Injection mismatch, decoherence, and emittance dilution at the LHC are filamentation phenomena — blobs of $10^{11}$ protons smearing along tori at a rate set by the tune spread, with Liouville forbidding true compression and "beam cooling" defined as the art of sneaking non-Hamiltonian terms past that prohibition. The same reasoning now governs galactic dynamics: the stellar phase-space spirals Gaia mapped in the Milky Way disc are filamentation in action, used to date past perturbations — one theorem, read from tabletop pendulum to galaxy.

Imagine phase space densely packed with copies of the system, one starting from every possible state. Hamilton's equations tell each copy where to go next, exactly as a stationary velocity field tells each fluid parcel where to drift. Watching them all at once, you see a flowing fluid — the phase trajectories of the phase-portraits article, now understood as streamlines of a single global object. Here it is for a pendulum, computed point by point:

Three features repay a long stare. The arrows are everywhere tangent to the energy contours: the velocity is the gradient of $H$ rotated by 90° in the $(q,p)$ plane — the "symplectic gradient" — so the fluid can never flow uphill or downhill in $H$, only sideways along its contours. That 90° rotation is energy conservation, drawn. Second, the arrows lengthen exactly where contours crowd: in one degree of freedom $H$ behaves as the stream function of a 2D incompressible flow, and closely spaced streamlines mean fast flow, precisely as in a river narrowing between banks. Third, the flow stands still at the center and at the saddles $q=\pm\pi$ — the anatomy of the whole portrait hangs on these (they get their own article on fixed points and stability). The fluid never compresses or rarefies. Drop a dye-marked droplet of initial conditions into it and watch:

For the harmonic oscillator (left), the droplet rides around rigidly — in scaled coordinates the flow is a uniform rotation at rate $\omega$ for every amplitude, so the disc never changes shape (the harmonic-oscillator-phase-space article makes this exact). The pendulum (right) tells the general story: inner points circulate faster than outer ones because the pendulum's period grows with amplitude, so the disc shears into a long thin filament wrapping around its annulus of tori. The RK4-computed areas of the three snapshots agree to a tenth of a percent — drastically deformed, not one bit compressed. The 'no leaks, no compression' rule is forced by the pairing of $+\partial_p H$ in $\dot q$ with $-\partial_q H$ in $\dot p$: the minus sign makes the two compression terms cancel. Because the flow preserves not just volume but Poisson brackets, it is a canonical transformation for every $t$: time evolution is a continuously-generated change of canonical coordinates with $H$ as generator — the classical seed of the quantum statement that $\hat H$ generates unitary time evolution.

Collect the $2n$ coordinates into one vector — the subject becomes transparent the moment you stop treating $q$ and $p$ separately. Hamilton's equations become a single first-order flow equation,

where: $z$ — a point in $2n$-dimensional phase space; $X_H$ — the Hamiltonian vector field, the phase-fluid velocity at $z$; $n$ — the number of degrees of freedom. Read it as a weather map: phase space carries a stationary wind pattern (stationary when $H$ is time-independent), and dynamics is drifting with the wind. The components of that wind come straight from the canonical equations,

where: $H(q,p,t)$ — the Hamiltonian; $\partial H/\partial p_i=\dot q_i$ — velocity in the coordinate directions; $-\partial H/\partial q_i=\dot p_i$ — velocity in the momentum directions. The asymmetry — plus on one half, minus on the other — looks like a typo and turns out to be the engine of everything below: it is exactly what makes the divergence vanish and the flow symplectic. Those $2n$ components compress into one geometric statement. With the symplectic unit matrix $J$,

where: $J$ — the $2n\times2n$ symplectic (Poisson) matrix, antisymmetric with $J^2=-I$; $\nabla H$ — the ordinary phase-space gradient; $I_n$ — the $n\times n$ identity block. In each $(q_i,p_i)$ plane, $J$ rotates a vector by 90°: a pure gradient flow would slide downhill and stop, but the symplectic rotation makes the fluid circulate along level sets of $H$ forever. Check the tangency directly: $\nabla H\cdot X_H = \nabla H^{\mathsf T}J\,\nabla H = 0$ because $J$ is antisymmetric. Now take the divergence of the field — one line, and the most consequential line in this chapter:

where: $\nabla\cdot X_H$ — the phase-space divergence, the fractional rate at which a comoving volume grows. Each term cancels its partner by equality of mixed partials — no restriction to conservative forces or time-independence. Every Hamiltonian flow, without exception, is incompressible. Derivation A unpacks this and its integral form, Liouville's theorem. The flow also tells every observable how to change. For any $f(q,p,t)$, the chain rule along a trajectory gives

where: $\{f,H\}$ — the Poisson bracket of $f$ with the Hamiltonian, precisely the directional derivative of $f$ along $X_H$; $\partial f/\partial t$ — any explicit time dependence of $f$. Geometry says "points move along $X_H$"; algebra says "functions evolve by bracketing with $H$." The consequences — conservation laws, Poisson's theorem — live in the article on time evolution of phase-space functions. For an autonomous Hamiltonian the flow maps compose like a group, $\phi_t\circ\phi_s=\phi_{t+s}$, and the family is the exponential of its generator,

where: $\phi_t$ — the time-$t$ flow map $z_0\mapsto z(t)$; $\mathcal{L}_{X_H}$ — the Lie derivative along $X_H$, acting on functions as $\{\,\cdot\,,H\}$. For nonlinear systems the exponential is shorthand for the Lie series (whose truncations are how symplectic integrators are built); for linear systems it is a literal matrix exponential, as the first worked example shows. Setting $f=H$ in (5) gives $\dot H=\{H,H\}+\partial H/\partial t=\partial H/\partial t$: the bracket kills itself by antisymmetry, so autonomous flows conserve energy and trajectories stay on the level sets $H=\mathrm{const}$.

Which equation answers which question

Incompressibility is only the visible edge of a sharper fact. Freeze the flow at time $t$ and regard $\phi_t$ as a coordinate change sending each initial state $z_0$ to the evolved state $z(t)$. The claim: this map is canonical for every $t$ — it preserves the entire symplectic structure, not merely volume. The statement is a condition on the Jacobian of the map,

where: $M$ — the Jacobian of the flow map, describing how an infinitesimal cloud of initial conditions is linearly deformed after time $t$; $M^{\mathsf T}JM=J$ — the symplectic condition, the matrix form of "all Poisson brackets are preserved." Determinants of (7) give $(\det M)^2=1$, and continuity from $M(0)=I$ forces $\det M=+1$: volume conservation drops out as a corollary. The converse fails — for $n\geq2$ there are volume-preserving maps that are not symplectic, so (7) genuinely says more than Liouville. The equivalent coordinate-free statement is that the flow preserves the symplectic 2-form,

where: $\omega$ — the canonical 2-form phase space carries in place of a metric (introduced in the phase-space article); $\phi_t^{*}$ — the pullback by the flow. Concretely: push the edge vectors of any little parallelogram through the flow, and the sum of its projected oriented areas onto the $(q_i,p_i)$ planes is unchanged. Here is that statement with actual matrices:

Both panels are time-$t$ flow maps of real systems — the free particle's shear $q\mapsto q+pt/m$, and the linearized saddle's stretch-and-squeeze $e^{\pm\lambda t}$ — and both leave the unit square's area at exactly $1$ while mangling its shape. Finally, energy itself rides the flow trivially:

where: $\{H,H\}=0$ by antisymmetry of the bracket. For autonomous systems $H$ is constant on trajectories — the fluid is stratified into invariant energy shells, and everything the flow does happens within a shell.

Volume conservation coexists with violent deformation, and the deformation has a rate you can calculate — the subject of the third derivation, and what the filament figure and the second worked example are quantitatively about.

1 · The quarter-period flow map of an oscillator (confidence builder)

A block of mass $m=0.5\ \mathrm{kg}$ oscillates on a spring with $\omega=2.0\ \mathrm{rad/s}$. Write down the flow map $\phi_t$ explicitly, evaluate its Jacobian at $t=T/4$, and verify symplecticity and area preservation on a concrete blob: a rectangle of initial conditions with $\Delta q=0.10\ \mathrm{m}$, $\Delta p=0.20\ \mathrm{kg\,m/s}$.

Step 1. Solve the flow. Hamilton's equations $\dot q=p/m$, $\dot p=-m\omega^2q$ are linear, so the flow map is the matrix exponential $z(t)=e^{tJ\mathbb{H}}z_0$: $q(t)=q_0\cos\omega t+\dfrac{p_0}{m\omega}\sin\omega t$, $p(t)=-m\omega\,q_0\sin\omega t+p_0\cos\omega t$, with Jacobian $M(t)=\begin{pmatrix}\cos\omega t & \frac{1}{m\omega}\sin\omega t\\ -m\omega\sin\omega t & \cos\omega t\end{pmatrix}$.

Step 2. Numbers at a quarter period. $T=2\pi/\omega=3.14\ \mathrm{s}$, so $T/4=0.785\ \mathrm{s}$ and $\omega t=\pi/2$: cosines vanish, sines are 1. With $m\omega=(0.5)(2.0)=1.0\ \mathrm{kg/s}$,

in SI units — numerically the symplectic unit $J$ itself. The quarter-period map is a 90° rotation of phase space (a literal rotation in the scaled coordinates $Q=\sqrt{m\omega}\,q$, $P=p/\sqrt{m\omega}$): position and momentum trade places, the phase-space face of the fact that a quarter period converts maximal displacement into maximal momentum.

Step 3. Check symplecticity by direct multiplication: $M^{\mathsf T}JM=J^{\mathsf T}JJ=(-J)(J)(J)=-J^3=J$ since $J^2=-I$. Condition (7) holds exactly, and $\det M=(0)(0)-(1.0)(-1.0)=\boxed{1}$ — exactly, not approximately.

Step 4. Push the blob through. Corners $(q,p)=(0.30,0)$, $(0.40,0)$, $(0.40,0.20)$, $(0.30,0.20)$, area $0.10\times0.20=0.020\ \mathrm{J\,s}$. The map sends $(q,p)\mapsto(1.0\,p,\,-1.0\,q)$, so the corners land at $(0,-0.30)$, $(0,-0.40)$, $(0.20,-0.40)$, $(0.20,-0.30)$: a rectangle with $\Delta q'=0.20\ \mathrm{m}$, $\Delta p'=0.10\ \mathrm{kg\,m/s}$, area $\boxed{0.020\ \mathrm{J\,s}}$. The blob rotated and traded aspect ratio; the area survived to the last digit.

2 · Filamentation time of a pendulum ensemble (exam level)

A demonstration pendulum of length $L=1.00\ \mathrm{m}$ ($g=9.80\ \mathrm{m/s^2}$) is released repeatedly with imperfect reproducibility: the release amplitudes spread over $\theta_0=0.50$ to $0.60\ \mathrm{rad}$. Treat the runs as a blob of initial conditions. How long until the blob smears through one radian of relative phase — and how many oscillations until it has wrapped into a full ring filament? Use $\omega(\theta_0)\approx\omega_0(1-\theta_0^2/16)$.

Step 1. Base frequency and period: $\omega_0=\sqrt{g/L}=\sqrt{9.80/1.00}=3.13\ \mathrm{rad/s}$, $T_0=2\pi/\omega_0=2.01\ \mathrm{s}$.

Step 2. Frequency spread across the blob: $\omega(0.50)=\omega_0(1-0.25/16)=0.9844\,\omega_0$ and $\omega(0.60)=\omega_0(1-0.36/16)=0.9775\,\omega_0$, so $\Delta\omega=\omega_0\,\dfrac{0.36-0.25}{16}=3.13\times6.88\times10^{-3}=0.0215\ \mathrm{rad/s}$. This is $(\mathrm{d}\omega/\mathrm{d}J)\Delta J$ of Derivation C, evaluated through the directly measurable amplitude spread.

Step 3. One-radian shear time: $t_1=1/\Delta\omega=1/0.0215=\boxed{46.5\ \mathrm{s}}$ — about $46.5/2.01\approx23$ oscillations. For the first twenty-odd swings the ensemble looks like a coherent blob riding its torus; by a minute in, it is visibly a smear.

Step 4. Full wrap into a ring at relative phase $2\pi$: $t_{2\pi}=2\pi/\Delta\omega=\boxed{292\ \mathrm{s}}$ — about $145$ oscillations, five minutes of swinging. After that the ensemble covers its annulus of tori nearly uniformly: runs completely decorrelated in phase, every one still conserving its own energy, total phase-space area exactly what it was at release.

Step 5. The same physics at 11,000 revolutions per second. An injected synchrotron bunch that misses the matched orbit is precisely such a blob, and machine nonlinearity gives a tune spread $\Delta Q$ per turn. With a typical $\Delta Q\approx10^{-3}$, the bunch filaments completely in $N\approx1/\Delta Q\approx\boxed{1000\ \mathrm{turns}}$ — at the LHC's $11\ \mathrm{kHz}$ revolution frequency, about $90\ \mathrm{ms}$ after injection. The smeared bunch has larger effective emittance even though Liouville guarantees the microscopic area never grew: filamentation, not compression, is how Hamiltonian beams "blow up."

References. Goldstein, Poole & Safko, Classical Mechanics, 3rd ed., §9.4 and §9.9 (the symplectic approach to canonical transformations; Liouville's theorem) · Landau & Lifshitz, Mechanics, 3rd ed., §46 (Liouville's theorem) · Arnold, Mathematical Methods of Classical Mechanics, 2nd ed., Ch. 8–9 (symplectic manifolds, Hamiltonian flows) · Taylor, Classical Mechanics, Ch. 13 (Hamiltonian mechanics; §13.7 Liouville's theorem) · Hairer, Lubich & Wanner, Geometric Numerical Integration (symplectic integrators, shadow Hamiltonians) Three key identities underpin the article: (1) $\nabla\cdot X_H=\partial_{q_i}\partial_{p_i}H-\partial_{p_i}\partial_{q_i}H=0$; (2) $A=J\mathbb{H}$ satisfies $A^{\mathsf T}J+JA=0$ using $J^2=-I$, $J^{\mathsf T}=-J$, $\mathbb{H}^{\mathsf T}=\mathbb{H}$, hence $\frac{d}{dt}(M^{\mathsf T}JM)=0$; (3) $\{H,H\}=0$, giving energy conservation for autonomous $H$. The only numerical constant beyond standard textbook structure is the harmonic period $2\pi/\omega$.

Builds on: Hamilton's Canonical Equations · Phase Space · The Hamiltonian Function · Phase Portraits and Trajectories Leads to: Liouville's Theorem · The Harmonic Oscillator in Phase Space · The Pendulum Phase Portrait and Separatrix · Fixed Points and Stability in Phase Space