Conservation Laws in the Hamiltonian Formalism

The two-way street is now engineering practice in quantum technology. A conserved charge that commutes with the Hamiltonian partitions a quantum processor's state space into symmetry sectors, and error-mitigation schemes exploit this daily: particle-number and parity conservation are verified mid-circuit in quantum-chemistry runs, discarding shots that leak between sectors, while symmetry-protected codes hide logical qubits in subspaces that symmetry-respecting noise cannot mix. On the classical side, momentum maps and symplectic reduction — the Marsden–Weinstein construction from the timeline — underlie spacecraft attitude controllers and legged-robot gait generators, where "reduce by the symmetry, control the quotient" is a design pattern. The 2026 twist is machines discovering conservation laws rather than being told them. Hamiltonian neural networks bake $\mathrm{d}H/\mathrm{d}t = \partial H/\partial t$ into their architecture so learned dynamics conserve energy by construction, and symmetry-discovery pipelines — descendants of the Noether-network idea — scan trajectory data for functions whose bracket with the learned Hamiltonian vanishes, recovering $\ell$ and Runge–Lenz-type invariants from raw orbits with no physics hints. In these systems equation (4) is literally the loss function.

Start with the picture that makes the two-way street visible. Put a particle in a central potential and let the dynamics run: the $H$-flow carries the state around its orbit, and at every instant the number $L_z = x p_y - y p_x$ comes out the same — $L_z$ is conserved by $H$. Now run the other engine. Use $L_z$ itself as if it were a Hamiltonian: the flow it generates does not advance time, it rigidly rotates the whole orbit — positions and momenta together — and every rotated copy has exactly the same energy. That is $H$ being conserved by $L_z$. One vanishing bracket, two flows, each leaving the other's generator untouched:

The everyday version of this magic is the cyclic coordinate. Think of $H(q,p)$ as a landscape: Hamilton's equations say the momentum $p_i$ rolls downhill in the $q_i$ direction with "force" $-\partial H/\partial q_i$. If the landscape is perfectly flat along $q_i$ — the coordinate never appears in $H$, so it is cyclic, or ignorable — there is nothing to push $p_i$, and it never changes. A central potential is flat along the azimuth $\phi$, so in the $(\phi, p_\phi)$ plane every trajectory glides along a horizontal line. A pendulum, by contrast, has $\theta$ sitting inside $H$ through $-mgl\cos\theta$; the landscape tilts, the flow bends, and $p_\theta$ trades away and back every swing while only $H$ survives:

Each conservation law is a wall in phase space that the trajectory cannot cross: the motion is pinned to the level surface $f = \text{const}$, and each additional independent constant squeezes it into a thinner subset. One conserved $H$ pins the oscillator to its ellipse; with $n$ commuting constants an $n$-degree-of-freedom system is confined to an $n$-torus — the structure previewed at the end of this article and developed in the article on action-angle variables. This is also the ultimate labor-saving device: a momentum you know is constant becomes a fixed parameter, and the dynamics in the remaining variables decouples — exactly the trick that reduces planetary motion to a one-dimensional radial problem. The third picture to install is the generator ladder. Every conserved quantity earns its keep twice: it is a number you can bank, and it is an engine that drives a transformation — momentum drives rigid shifts, angular momentum drives rotations, the Hamiltonian drives time itself. Quantum mechanics does not discard the ladder; it exponentiates it, turning each classical generator into the unitary operator implementing the same symmetry on states:

Everything here flows from the canonical equations — derived and dissected in the article on Hamilton's canonical equations; here they are the engine whose structure we exploit. For $n$ degrees of freedom,

where: $q_i$ — generalized coordinates ($i = 1,\dots,n$); $p_i$ — their conjugate momenta; $H(q,p,t)$ — the Hamiltonian. Read the second equation as the whole subject in miniature: the only thing that can change a momentum is an explicit appearance of its coordinate in $H$; the antisymmetric pairing of the two equations is what makes every cancellation below work. The first harvest is structural. Suppose a coordinate $q_i$ simply fails to appear in $H$. Then there is no force conjugate to it, and its momentum is a constant of the motion for all time:

where: $q_i$ — a cyclic (ignorable) coordinate, absent from $H$; $p_i$ — its conserved conjugate momentum. This is conservation you can see in the formula for $H$ before solving anything. Cyclicity transfers faithfully from the Lagrangian: since $H = \sum_i p_i\dot q_i - L$ is a Legendre transform in the velocities only, $\partial H/\partial q_i = -\partial L/\partial q_i$, so the two formalisms agree about which momenta are conserved — the Lagrangian half of the story is told in the article on cyclic coordinates and conserved momenta. Cyclic coordinates are a special case of a criterion that covers any phase-space function — angular momenta, Runge–Lenz-type vectors, combinations nobody would guess from staring at coordinates. Any observable $f(q,p,t)$ evolves along the motion by the master equation of the previous article (time evolution of phase-space functions), quoted here rather than re-derived:

where: $\{f,H\} = \sum_i \left( \partial_{q_i} f\, \partial_{p_i} H - \partial_{p_i} f\, \partial_{q_i} H \right)$ — the Poisson bracket; $\partial f/\partial t$ — the explicit time dependence of $f$ itself. The bracket term is the dynamics — the chain rule with Hamilton's equations already substituted. Setting the left side to zero turns equation (3) into the general conservation criterion. For the common case of an observable built from $q$ and $p$ alone,

where: $f(q,p)$ — any observable with no explicit time dependence. Checking a conservation law is now pure differentiation — no trajectory is ever solved. Take $f = p_i$ and equation (4) reproduces the cyclic-coordinate rule (2), since $\{p_i, H\} = -\partial H/\partial q_i$; take $f = L_z$ in a central potential and four derivatives confirm rotational conservation (Worked Example 1 does it with numbers). Quantities that carry explicit time dependence can still be conserved — the bracket and $\partial f/\partial t$ must then cancel, a subtlety the FAQ takes up. The most important single substitution is $f = H$ itself. Antisymmetry kills the dynamical term identically, $\{H,H\} = 0$, leaving the energy theorem:

where: $\mathrm{d}H/\mathrm{d}t$ — the rate of change along a trajectory; $\partial H/\partial t$ — the explicit time dependence of the function $H$. The Hamiltonian is conserved unless someone wrote the time into it — a moving constraint, a ramped field, a driven support. Whether that conserved constant is the total energy $T+V$ is a separate structural question, settled in Derivation A below.

Which statement answers which question

Here is the pivot from bookkeeping to structure. So far the bracket $\{f,H\}$ has been read with $H$ as the actor — $H$ generates time evolution, and equation (4) asks whether $f$ survives it. But nothing in the bracket singles out $H$: any smooth phase-space function $f$ can be installed as a generator. Replace $H$ by $f$ and the time step $\mathrm{d}t$ by a small parameter $\varepsilon$ in Hamilton's equations, and you obtain the canonical flow generated by $f$,

where: $\varepsilon$ — an infinitesimal flow parameter (an angle for rotations, a length for translations, a time for the dynamics); $f$ — the generating function. These are "Hamilton's equations with $f$ in the driver's seat," and the transformation is automatically canonical — which is what makes generators such a privileged way to write symmetries. Under this flow, any other observable $g$ is dragged along, and its first-order change is again a bracket — the same chain-rule computation as equation (3), with $f$ as generator and $\varepsilon$ as clock:

where: $\delta g$ — the change of $g(q,p)$ under the $f$-flow (6). Every pairing you know is a special case: $f = H$, $\varepsilon = \mathrm{d}t$ gives the master equation (3); $f = p_x$ gives $\delta x = \varepsilon$ and nothing else — a rigid translation; $f = L_z$ rotates the system (Derivation B checks this explicitly). Now put $g = H$ and read the result with fresh eyes. The change of the Hamiltonian under the flow generated by $f$ is

where: the second equality is the antisymmetry of the bracket. Stare at the two ends of this chain: $\{f,H\} = 0$ says simultaneously that the dynamics conserves $f$ and that the $f$-flow is a symmetry of $H$. There is no separate Noether theorem to prove in Hamiltonian mechanics — the theorem is the antisymmetry of the Poisson bracket, converse included. The Lagrangian route and its total-derivative fine print live in the article on Noether's theorem; the taxonomy of transformations in the article on symmetries and continuous transformations. The dictionary this builds is the same one quantum mechanics uses, with $\{\,,\}\to[\,,\,]/i\hbar$ and generators exponentiated into unitaries:

Conservation laws are not trophies; they are tools. The moment a momentum is known to be constant, it stops being a dynamical variable and becomes a parameter, and the system effectively loses degrees of freedom. The cleanest showcase is the central-force problem, where the cyclic angle $\phi$ hands over its conserved $p_\phi = \ell$ and the planar motion collapses to one radial dimension governed by

where: $p_r$ — the radial momentum $m\dot r$; $\ell$ — the conserved angular momentum, now a fixed constant; $\ell^2/2m r^2$ — the centrifugal barrier, angular kinetic energy masquerading as a repulsive potential. Everything about orbits — turning points, stability, Kepler ellipses — is read from the effective potential $V_{\text{eff}}(r) = \ell^2/2mr^2 + V(r)$; with $E$ and $\ell$ both conserved, two constants for two degrees of freedom, the problem is integrable. Kepler's second law is nothing but $p_\phi = mr^2\dot\phi = \ell$ read as an areal velocity. The systematic engine behind this maneuver is the Routhian — a partial Legendre transform eliminating only the cyclic variables — treated in the article on cyclic coordinates and conserved momenta. The deeper geometric fact is worth flagging now: using a symmetry properly removes two dimensions per conserved momentum (fix the value of $p_\phi$, then quotient out the $\phi$-direction the symmetry moves), which is why one conserved $\ell$ turns a four-dimensional phase space into the two-dimensional $(r, p_r)$ plane. The FAQ returns to this count.

Constants of the motion do not live alone; the bracket organizes them into algebras. The first structural fact is Poisson's theorem: if $f$ and $g$ are both constants of the motion, the Jacobi identity forces their bracket to be one too:

where: the implication is immediate from the Jacobi identity $\{\{f,g\},H\} + \{\{g,H\},f\} + \{\{H,f\},g\} = 0$ with the last two terms dying by hypothesis. Occasionally this manufactures a genuinely new conservation law; more often it returns a constant you already had, or a trivial number — the reason the theorem does not make every system integrable (the FAQ dissects this). The showcase closure is angular momentum. Direct computation with the canonical brackets gives

where: $L_i$ — the Cartesian components of $\mathbf L = \mathbf r \times \mathbf p$; $\epsilon_{ijk}$ — the Levi-Civita symbol. Conserve any two components and Poisson's theorem hands you the third — the rotation group's algebra $\mathfrak{so}(3)$ living inside phase space. Note the flip side: because the components fail to commute with each other, at most one of them (conventionally $L_z$, with $L^2$) can join a maximal set of compatible constants. That word "compatible" is the doorway to integrability. A set of constants $F_1 = H, F_2, \dots, F_n$ is in involution when every pair commutes:

where: $n$ — the number of degrees of freedom; independence means the gradients $\nabla F_i$ stay linearly independent on the common level set. The Liouville–Arnold theorem then delivers the grand prize: with $n$ independent constants in involution, the motion is confined to $n$-dimensional tori and solvable by quadratures — integrability. The central-force problem ($H$ and $\ell$, $n=2$) passes; the three-body problem does not, and the gap between those two sentences is where chaos lives. The torus construction is the business of the article on action-angle variables.

Whether $H$ is conserved, whether it equals the energy, and whether some momentum is conserved are three independent switches; the table sorts the standard settings.

1 · Auditing angular momentum with the bracket (confidence builder)

A particle of mass $m = 2.0\ \mathrm{kg}$ moves in the plane; at the instant of interest $x = 3.0\ \mathrm{m}$, $y = 1.0\ \mathrm{m}$, $p_x = -2.0\ \mathrm{kg\,m/s}$, $p_y = 6.0\ \mathrm{kg\,m/s}$. Check whether $L_z$ is conserved (a) in the isotropic potential $V = \tfrac12 k (x^2 + y^2)$, $k = 8.0\ \mathrm{N/m}$, and (b) in the anisotropic $V = \tfrac12(k_x x^2 + k_y y^2)$, $k_x = 8.0$, $k_y = 2.0\ \mathrm{N/m}$ — and if not, at what rate it is changing.

Step 1. Evaluate the observable. $L_z = x p_y - y p_x = (3.0)(6.0) - (1.0)(-2.0) = \boxed{20.0\ \mathrm{kg\,m^2/s}}$.

Step 2. Set up the bracket with $H = (p_x^2 + p_y^2)/2m + V$. The four partials of $L_z$ are $\partial L_z/\partial x = p_y$, $\partial L_z/\partial y = -p_x$, $\partial L_z/\partial p_x = -y$, $\partial L_z/\partial p_y = x$, so

The two kinetic terms cancel identically — kinetic energy is rotation-invariant, so only the potential can torque.

Step 3. Isotropic case: $\partial V/\partial x = kx$, $\partial V/\partial y = ky$, so the potential terms give $ykx - xky = 0$. With numbers, the four contributions are $(6.0)(-2.0)/2.0 = -6.0$, then $+ (1.0)(8.0)(3.0) = +24.0$, then $+6.0$, then $-(3.0)(8.0)(1.0) = -24.0$ — total $\boxed{0}$: $L_z$ is conserved, exactly as the generator argument demands ($V$ depends only on $r^2$).

Step 4. Anisotropic case: $\partial V/\partial x = k_x x$, $\partial V/\partial y = k_y y$, and the potential terms no longer cancel:

Step 5. Cross-check against elementary torque: $\tau_z = x F_y - y F_x = x(-k_y y) - y(-k_x x) = -6.0 + 24.0 = 18.0\ \mathrm{N\,m}$ — agreed. In polar form this is the general lesson $\{L_z, H\} = -\partial V/\partial\phi$: angular momentum fails to be conserved by exactly the angular slope of the landscape. Meanwhile $H$ carries no explicit $t$, so the energy $E = (4.0 + 36.0)/4.0 + \tfrac12(8.0\cdot 9.0 + 2.0\cdot 1.0) = 47.0\ \mathrm{J}$ stays conserved even as $L_z$ drifts — the two laws switch on and off independently.

2 · Frozen canonical momentum, oscillating kinetic momentum (exam level)

A proton ($m = 1.67\times10^{-27}\ \mathrm{kg}$, $q = +1.60\times10^{-19}\ \mathrm{C}$) moves in a uniform magnetic field $B = 0.50\ \mathrm{T}$ along $\hat z$, described in the Landau gauge $\mathbf A = (0, Bx, 0)$, so $H = \left[p_x^2 + (p_y - qBx)^2\right]/2m$ for planar motion. At $t=0$ the proton is at the origin moving with $v = 4.0\times10^{5}\ \mathrm{m/s}$ along $+\hat x$. Identify and evaluate the conserved quantities, and follow the canonical and kinetic $y$-momenta around one orbit.

Step 1. Read the structure before computing. The coordinate $y$ appears nowhere in $H$ — cyclic — so the canonical momentum $p_y = m v_y + qBx$ is conserved. And $H$ has no explicit $t$, so the energy $H = \tfrac12 m v^2$ is conserved too. Two frozen numbers, straight off the page.

Step 2. Orbit parameters. Cyclotron frequency $\omega = qB/m = (1.60\times10^{-19})(0.50)/(1.67\times10^{-27}) = 4.79\times10^{7}\ \mathrm{rad/s}$, period $T = 2\pi/\omega = 1.31\times10^{-7}\ \mathrm{s}$, radius $r = v/\omega = mv/qB = (6.68\times10^{-22})/(8.0\times10^{-20}) = \boxed{8.35\times10^{-3}\ \mathrm{m}}$ — an 8.4-mm circle.

Step 3. Evaluate the conserved quantities at $t = 0$. Energy: $E = \tfrac12 m v^2 = \tfrac12(1.67\times10^{-27})(4.0\times10^{5})^2 = 1.34\times10^{-16}\ \mathrm{J} = \boxed{834\ \mathrm{eV}}$. Canonical momentum: $p_y = m v_y + qBx = 0 + 0 = \boxed{0}$ — and being conserved, it is zero forever.

Step 4. Watch the two momenta around the orbit. The Lorentz force curls the initial $+\hat x$ velocity toward $-\hat y$, so $x(t) = r\sin\omega t$, $y(t) = -r(1 - \cos\omega t)$, circling the guiding center $(0, -r)$. The kinetic momentum $m v_y = -mv\sin\omega t$ swings between $\pm mv = \pm 6.68\times10^{-22}\ \mathrm{kg\,m/s}$; the gauge term $qBx = qBr\sin\omega t = +mv\sin\omega t$ (since $qBr = mv$ on a cyclotron orbit) cancels it exactly, at every instant — at $\omega t = \pi/2$: $-6.68\times10^{-22} + 6.68\times10^{-22} = 0$.

Step 5. Interpret the frozen number. The conserved $p_y$ is not the "motion in $y$"; it is the location of the orbit: $p_y = qB\,x_c$ with $x_c$ the guiding-center coordinate — here $x_c = 0$, correctly the center's $x$-position. A conserved canonical momentum can encode where the motion is anchored rather than how fast anything moves; an answer reporting "$m v_y$ is conserved because $y$ is cyclic" gets the physics exactly backwards.

References. Goldstein, Poole & Safko, Classical Mechanics, 3rd ed., §8.1–8.2 (cyclic coordinates, conservation of H) and §9.5–9.6 (Poisson brackets, infinitesimal canonical transformations) · Landau & Lifshitz, Mechanics, 3rd ed., §40 (Hamilton's equations and energy conservation) and §42 (Poisson brackets) · Morin, Introduction to Classical Mechanics, Ch. 15 (Hamiltonian methods) · Taylor, Classical Mechanics, Ch. 13 (Hamiltonian mechanics) · Arnold, Mathematical Methods of Classical Mechanics, Ch. 8 & 10 (symmetry, first integrals, Liouville integrability)

Provenance. A verifier should confirm: (1) Hamilton's equations $\dot q_i=\partial H/\partial p_i,\ \dot p_i=-\partial H/\partial q_i$ (Goldstein §8.1, Landau §40); (2) the cancellation giving $\mathrm{d}H/\mathrm{d}t=\partial H/\partial t$ (Goldstein eq. 8.40); (3) Euler's theorem on homogeneous quadratic $T$ giving $H=T+V$ for scleronomic systems (Goldstein §2.7, §8.2); (4) the reduced central-force Hamiltonian $H=p_r^2/2m+\ell^2/2mr^2+V(r)$ (Goldstein Ch. 3, Taylor Ch. 8); (5) the Poisson-bracket evolution equation $\mathrm{d}f/\mathrm{d}t=\{f,H\}+\partial f/\partial t$ (Goldstein §9.6, Landau §42). No numerical constants are asserted beyond standard symbolic results.

Builds on: The Hamiltonian Function · Hamilton's Canonical Equations · Generalized Momenta · Legendre Transformation Leads to: Phase Space · Phase Portraits and Trajectories · Hamiltonian Flow and the Phase Fluid · The Pendulum Phase Portrait and Separatrix