Infinitesimal Canonical Transformations

Write down the canonical transformation to action-angle variables for a pendulum and you get elliptic functions sprawling across half a page. Write down the transformation that advances the solar system by a million years and you cannot — no closed form exists. Finite canonical transformations are, in general, monstrously complicated objects. And yet the theory tames every one of them with a single move: stop trying to write the finished map and write its derivative at the identity instead. A transformation that moves each phase point only infinitesimally — $Q=q+\delta q$ , $P=p+\delta p$ — turns out to be completely controlled by one ordinary scalar function $G(q,p)$ on phase space. Not a matrix, not a pair of functions: one scalar, from which both displacement fields follow by differentiation. That compression is the pivot on which the whole canonical formalism turns. Feed the machinery $G=p_x$ and it slides the system through space; feed it $G=L_z$ and it rotates positions and momenta together; feed it $G=H$ and — this is the astonishing one — it advances the system through time, with Hamilton's equations falling out as the displacement formulas. Symmetries, conservation laws, Liouville's theorem, and eventually the unitary operators of quantum mechanics are all downstream of the small step you build here. This page owns the $\varepsilon$ -machinery itself: where the displacement formulas come from (a Type-2 generating function one whisker away from the identity), why the map is canonical to first order (a symmetric Hessian meeting an antisymmetric $J$ ), and what a single truncated step can and cannot do. The finite flows those steps compose into are the business of Generators of Continuous Transformations.

Key takeaways

An infinitesimal canonical transformation (ICT) is a canonical map one small step from the identity: $Q_i=q_i+\varepsilon\,\partial G/\partial p_i$ , $P_i=p_i-\varepsilon\,\partial G/\partial q_i$ , controlled entirely by one scalar generator $G(q,p,t)$ .
It descends from the Type-2 generating function $F_2=\sum_i q_iP_i+\varepsilon\,G$ — the identity plus an $\varepsilon$ -sized correction — so canonicity is inherited, not imposed.
Compactly, $\delta\eta=\varepsilon\,J\,\nabla G$ : the step is the gradient of $G$ rotated a quarter-turn by the symplectic matrix, so the flow slides along level curves of $G$ , never across them.
Any observable responds by $\delta f=\varepsilon\{f,G\}$ — the Poisson bracket is the infinitesimal action of the transformation.
The Jacobian is $M=I+\varepsilon J\,\partial^2G$ ; because the Hessian $\partial^2 G$ is symmetric, $MJM^{\mathsf T}=J+O(\varepsilon^2)$ and $\det M=1+O(\varepsilon^2$ ): canonical to first order, exactly canonical only after integrating the flow.
Quick dictionary (deep dives elsewhere): $G=p_x$ translates, $G=L_z$ rotates, $G=H$ advances time.
Key idea: every canonical transformation near the identity is $\delta q=\varepsilon\,\partial G/\partial p$ , $\delta p=-\varepsilon\,\partial G/\partial q$ — equivalently $\delta f=\varepsilon\{f,G\}$ — one scalar generator per infinitesimal map.

The core insight

Every canonical transformation near the identity is generated by a single scalar $G$ : $\delta q_i=\varepsilon\,\partial G/\partial p_i$ , $\delta p_i=-\varepsilon\,\partial G/\partial q_i$ , equivalently $\delta f=\varepsilon\{f,G\}$ — and the map is canonical to first order because a symmetric Hessian cannot survive contact with the antisymmetric $J$ .

Where this lives today

Every serious long-time simulation of a Hamiltonian system now runs on machinery this page built. Symplectic integrators — the workhorses of solar-system ephemerides, particle-accelerator tracking, and molecular dynamics — are nothing but cleverly composed generated steps: each substep is an exactly canonical map produced by a generating function (kick from the potential, drift from the kinetic term), so the composed step has determinant exactly 1 and inherits the bounded-energy behavior of worked example 2 over billions of steps. JPL's long-horizon planetary integrations and LHC beam-tracking codes both live or die on that distinction between $\det M=1.04$ and $\det M=1$ . The newest arrivals speak the same dialect: Hamiltonian neural networks and learned symplectic maps parametrize an unknown dynamics not by fitting the flow directly but by fitting a scalar — a learned generator — and pushing states with $I+\varepsilon J\,\partial^2G$ steps or exact generating-function layers, so that the learned model conserves phase-space volume by architecture rather than by training luck. One scalar controlling a canonical flow turned out to be a compression scheme good enough for machine learning to rediscover it.

The physical picture

Picture $G$ as a landscape over phase space — a surface of hills and valleys plotted above the $(q,p)$ plane. A naive guess for "the transformation $G$ generates" would push points uphill, along the gradient $\nabla G$ . The canonical prescription does something stranger and better: it rotates the gradient by ninety degrees before using it. The step is $\delta\eta=\varepsilon\,J\nabla G$ , and the quarter-turn $J$ sends "across the contours" into "along the contours." Every phase point drifts sideways along a level curve of $G$ , like weather systems circulating along isobars instead of blowing straight into the low.

Figure

The quarter-turn that makes a flow canonical — the step is the gradient of G rotated ninety degrees, so points drift along level curves of G

The sideways drift explains, in one picture, why the generator is conserved by its own flow ( $G$ never changes along its own contours — the geometric face of $\{G,G\}=0$ ), and why the flow squeezes nothing (a divergence-free velocity field carries phase-space area rigidly — Liouville's theorem in embryo, developed fully in Liouville's Theorem and Phase-Space Volume). The second thing to internalize is where canonicity comes from. An ICT is not a small map that happens to pass a test; it is manufactured canonical. Take the Type-2 generating function of the identity transformation, $F_2=\sum_i q_iP_i$ — the "do nothing" map of Type-2 Generating Function F2(q,P,t) — and add the smallest possible perturbation, $\varepsilon\,G$ . Because any $F_2$ produces a canonical transformation, the near-identity map inherits canonicity from its birth certificate. The only work left is bookkeeping: expanding the implicit Type-2 relations to first order in $\varepsilon$ (Derivation A).

Figure

Anatomy of the near-identity generating function — the identity block does nothing, the epsilon block supplies the twisted displacements

Finally, hold on to what an ICT is not: it is not a finite transformation, and truncating at first order costs something real. One step of size $\varepsilon$ is canonical only up to $O(\varepsilon^2)$ debris; the worked examples below measure that debris on real numbers. Chaining infinitely many shrinking steps into an exactly canonical finite flow is precisely the exponentiation story of Generators of Continuous Transformations.

The displacement formulas

Everything follows from asking: what is the most general canonical map one whisker from the identity? Write the answer with a small parameter $\varepsilon$ carrying the smallness, and a function $G$ carrying the shape:

Q_i=q_i+\varepsilon\,\frac{\partial G}{\partial p_i},\qquad P_i=p_i-\varepsilon\,\frac{\partial G}{\partial q_i} \qquad (1)

where: $G(q,p,t)$ — the generator, an ordinary scalar function on phase space; $\varepsilon$ — the infinitesimal parameter (an angle, a length, a time... whatever the family is parametrized by); $\partial G/\partial p_i$ , $\partial G/\partial q_i$ — evaluated at the old point. Read the structure: coordinates move according to how $G$ leans on the momenta, momenta move (with a minus sign) according to how $G$ leans on the coordinates. That criss-cross with one sign flip is the symplectic twist — the same pattern as Hamilton's equations, and not by coincidence. Stacking $(q,p)$ into the phase-space column $\eta$ and using the symplectic matrix $J$ compresses both formulas into one:

\delta\eta=\varepsilon\,J\,\nabla G,\qquad J=\begin{pmatrix}0&I_n\\-I_n&0\end{pmatrix} \qquad (2)

where: $\nabla G$ — the $2n$ -gradient of $G$ ; $J$ — the quarter-turn of Symplectic Condition fame, $J^{\mathsf T}=-J$ , $J^2=-I$ . This is the hero-diagram statement: gradient in, quarter-turned step out. Special case worth one glance: a point transformation deforming only coordinates needs $G$ linear in $p$ ; anything with $p$ -curvature shears momenta too. The single most useful consequence: any observable $f(q,p)$ — energy, angular momentum, a correlation you invented yesterday — responds to the step through one bracket,

\delta f=\varepsilon\,\{f,G\} \qquad (3)

where: $\{\cdot,\cdot\}$ — the Poisson bracket in the convention of Poisson Brackets: Definition ( $\{q,p\}=+1$ ). Equation (3) is the active reading: the state moves, and $f$ 's value at the moved state differs by $\varepsilon\{f,G\}$ . It says the Poisson bracket is not merely a clever pairing — it is the infinitesimal action of one observable on all the others. Set $f=q_i$ or $p_i$ and (3) collapses back to (1) via the fundamental brackets of Fundamental Poisson Brackets. The birth certificate of (1) is the near-identity Type-2 generating function

F_2(q,P,t)=\sum_i q_iP_i+\varepsilon\,G(q,P,t) \qquad (4)

where: $q_iP_i$ — the identity transformation's generating function; $\varepsilon G$ — the perturbation, written in the mixed arguments $(q,P)$ that Type-2 functions use. To first order the mixed arguments are a distinction without a difference — that bookkeeping is Step 4 of Derivation A, and skipping it carelessly is a classic exam wound.

Derivation AThe displacement formulas from a near-identity Type-2 generator

We are deriving equation (1) — the claim that one scalar $G$ generates the most general infinitesimal canonical transformation — from the Type-2 machinery, which guarantees canonicity from the start. The strategy: write $F_2$ as identity-plus- $\varepsilon G$ as in (4), apply the standard Type-2 relations $p_i=\partial F_2/\partial q_i$ and $Q_i=\partial F_2/\partial P_i$ , and expand to first order. The one subtlety to watch is the argument of $G$ : the Type-2 relations naturally produce $\partial G/\partial P$ , and converting $P\to p$ inside a term already multiplied by $\varepsilon$ must be justified, not waved at.

Step 1. The momentum relation. Differentiating (4) with respect to $q_i$ :

p_i=\frac{\partial F_2}{\partial q_i}=P_i+\varepsilon\,\frac{\partial G(q,P)}{\partial q_i}

Solving for the new momentum: $P_i=p_i-\varepsilon\,\partial G/\partial q_i$ , so $\delta p_i=-\varepsilon\,\partial G/\partial q_i$ with $G$ 's derivative still evaluated at $(q,P)$ .

Step 2. The coordinate relation. Differentiating (4) with respect to $P_i$ :

Q_i=\frac{\partial F_2}{\partial P_i}=q_i+\varepsilon\,\frac{\partial G(q,P)}{\partial P_i}

so $\delta q_i=+\varepsilon\,\partial G/\partial P_i$ . The identity block contributed $Q_i=q_i$ and $P_i=p_i$ ; every displacement is the $\varepsilon$ block's doing.

Step 3. The sign asymmetry is already locked in. The plus on $\delta q$ and minus on $\delta p$ came from which variable each Type-2 relation is solved for — no choice was made; the symplectic structure of the $F_2$ formalism dictated it.

Step 4. Clean up the arguments. In Steps 1–2 the derivatives of $G$ are taken at $(q,P)$ , but $P_i=p_i+O(\varepsilon)$ , so by Taylor's theorem

\frac{\partial G(q,P)}{\partial P_i}=\frac{\partial G(q,p)}{\partial p_i}+O(\varepsilon)

and the correction enters the displacement at $O(\varepsilon^2)$ — below the order we keep. Both displacements may therefore be evaluated with $G$ as a function of the old variables, giving (1) exactly as stated. Why this argument-swap is safe here and dangerous later: the substitution costs $O(\varepsilon^2)$ per step. For a single infinitesimal step that is negligible by definition; but if you chain $N=\lambda/\varepsilon$ steps to build a finite map, the accumulated cost is $N\cdot O(\varepsilon^2)=O(\lambda\varepsilon)\to0$ only because $\varepsilon\to0$ in the limit — the finite flow is exact, any finite truncation is not. *Quick check: setting $G=p_x$ gives $\delta x=\varepsilon$ , $\delta p=0$ — a rigid translation, as it must.*

Canonical to first order: the Hessian meets J

How canonical is the truncated map (1), taken at face value as a small-but-finite step? Compute its Jacobian. Differentiating (1) with respect to the old variables packs neatly into

M=\frac{\partial(Q,P)}{\partial(q,p)}=I+\varepsilon\,J\,S,\qquad S_{ab}=\frac{\partial^2 G}{\partial\eta_a\,\partial\eta_b} \qquad (5)

where: $S$ — the Hessian of $G$ , symmetric by equality of mixed partials; $I$ — the $2n\times2n$ identity. Everything now rides on one algebraic collision: a symmetric $S$ sandwiched with the antisymmetric $J$ .

Go deeper · Derivation B · 1 minFirst-order symplecticity from the symmetry of the Hessian

Derivation BFirst-order symplecticity from the symmetry of the Hessian

We are verifying that the map (1) satisfies the symplectic condition to first order — $MJM^{\mathsf T}=J+O(\varepsilon^2)$ — and locating exactly which property of $G$ makes it work. The strategy is direct: insert $M=I+\varepsilon JS$ into the symplectic sandwich and watch the first-order terms. The punchline is that they cancel because the Hessian is symmetric — smoothness of $G$ is the entire mechanism; no condition on $G$ is needed.

Step 1. Expand the sandwich, keeping first order:

MJM^{\mathsf T}=(I+\varepsilon JS)\,J\,(I+\varepsilon JS)^{\mathsf T}=J+\varepsilon\left(JSJ+J S^{\mathsf T}J^{\mathsf T}\right)+O(\varepsilon^2)

Step 2. Use the two structural facts: $S^{\mathsf T}=S$ (mixed partials commute) and $J^{\mathsf T}=-J$ . The second term becomes $JS^{\mathsf T}J^{\mathsf T}=-JSJ$ , and

MJM^{\mathsf T}=J+\varepsilon\left(JSJ-JSJ\right)+O(\varepsilon^2)=J+O(\varepsilon^2)

Why symmetric-times-antisymmetric is the crux: had $S$ any antisymmetric part, that part would survive the cancellation and break canonicity at first order. But Hessians of smooth scalars have no antisymmetric part — which is why an arbitrary smooth $G$ generates a canonical step, and why a general small map $\delta\eta=\varepsilon V(\eta)$ with $V$ not of the form $J\nabla G$ does not.

Step 3. Volume, as a corollary. $\det M=1+\varepsilon\,\mathrm{tr}(JS)+O(\varepsilon^2)$ , and $\mathrm{tr}(JS)=0$ for the same reason (the trace of an antisymmetric-times-symmetric product vanishes: $\mathrm{tr}(JS)=\mathrm{tr}(S^{\mathsf T}J^{\mathsf T})=-\mathrm{tr}(SJ)=-\mathrm{tr}(JS)$ ). Phase-space volume is preserved to first order — the ICT seed of Liouville's theorem, grown to full strength in Liouville's Theorem and Phase-Space Volume.

Step 4. Read the $O(\varepsilon^2)$ honestly. The quadratic term $\varepsilon^2\,JSJ S^{\mathsf T}J^{\mathsf T}$ does not cancel in general: a single step of finite size $\varepsilon$ is genuinely non-canonical at second order. Worked example 2 puts a number on this — the truncated map's determinant is $1+\varepsilon^2\omega^2$ , visibly bigger than 1. Quick check: $\det(I+\varepsilon J\,\partial^2G)=1+O(\varepsilon^2)$ , and the trace identity $\mathrm{tr}(JS)=0$ can be confirmed directly from the antisymmetry argument above.

The same first-order statement in bracket language — often the fastest route in practice — is the direct check that the new variables keep the fundamental brackets: $\{Q_i,P_j\}=\delta_{ij}+O(\varepsilon^2)$ , with the $O(\varepsilon)$ terms cancelling between $\partial^2G/\partial q_j\partial p_i$ and its mirror. That computation, and its promotion to a full test for finite maps, is the machinery of Poisson Brackets as a Test for Canonicity.

Go deeper · Derivation C · 1 minThe response of an arbitrary observable

Derivation CThe response of an arbitrary observable

We are deriving equation (3), $\delta f=\varepsilon\{f,G\}$ — the statement that promotes the ICT from a rule about coordinates to a rule about everything. The strategy is a first-order Taylor expansion of $f$ under the displacements (1), which reassembles itself into a Poisson bracket. The payoff is the active/passive dictionary at the end: one formula, two legitimate readings.

Step 1. Taylor-expand $f$ at the displaced point:

\delta f=f(q+\delta q,\,p+\delta p)-f(q,p)=\sum_i\left(\frac{\partial f}{\partial q_i}\,\delta q_i+\frac{\partial f}{\partial p_i}\,\delta p_i\right)+O(\varepsilon^2)

Step 2. Insert the displacement formulas (1):

\delta f=\varepsilon\sum_i\left(\frac{\partial f}{\partial q_i}\frac{\partial G}{\partial p_i}-\frac{\partial f}{\partial p_i}\frac{\partial G}{\partial q_i}\right)

Step 3. Recognize the Poisson bracket:

\delta f=\varepsilon\,\{f,G\}

Why the slot order matters: with the standard convention, the flow of $G$ acts on $f$ as $\{f,G\}$ — generator in the second slot. Writing $\{G,f\}$ flips the sign and runs the transformation backwards. When importing formulas from another text, check its $\delta f$ against its $\{q,p\}$ before trusting a single sign.

Step 4. The two readings. Active: the point moves, $f$ 's recipe stays; $\delta f$ is the change in reading at the displaced state — this is (3). Passive: relabel coordinates and re-express $f$ in the new chart; the numerical change of the functional form is $-\varepsilon\{f,G\}$ , opposite in sign. The formulas are consistent — a point pushed forward looks, from the new chart, like a chart pulled backward — but mixing the two mid-calculation manufactures sign errors that are miserable to trace. Quick check: setting $f=G$ gives $\delta G=\varepsilon\{G,G\}=0$ — the generator is invariant under its own flow, matching the level-curve picture.

Limiting cases

Case	Result	Meaning
$G=p_x$	$\delta x=\varepsilon$ , all else 0	rigid spatial translation; deep dive in Momentum and Angular Momentum as Generators
$G=L_z$	$\delta x=-\varepsilon y$ , $\delta y=\varepsilon x$ , momenta likewise	infinitesimal rotation of positions and momenta about $z$
$G=H$ , $\varepsilon=\mathrm{d}t$	$\delta q=\mathrm{d}t\,\partial H/\partial p$ , $\delta p=-\mathrm{d}t\,\partial H/\partial q$	one tick of time evolution — Hamilton's equations as an ICT; the full story is Hamiltonian as Generator of Time Evolution
$G=\mathrm{const}$	$\delta q=\delta p=0$	the identity; generators matter only up to additive constants
$G$ linear in $p$	$\delta q$ independent of $p$	point transformations (coordinate changes) sit inside the ICT family
$\varepsilon\to0$ with $\lambda=N\varepsilon$ fixed	exactly canonical finite map	the exponentiation limit — handed off to Generators of Continuous Transformations

Common trapA single epsilon step is not a canonical transformation — it only fails at second order

"Canonical to first order" gets silently upgraded, in many heads, to "canonical." It is not. Take the truncated map (1) at small-but-finite $\varepsilon$ and its Jacobian determinant is $1+O(\varepsilon^2)$ , not 1: iterate it and the $O(\varepsilon^2)$ debris compounds — worked example 2 shows a 4% energy drift in one step of a harmonic oscillator. What is exactly canonical is (a) the untruncated map generated by the full $F_2=qP+\varepsilon G$ , mixed arguments and all — for $G=H$ that map is the symplectic-Euler integrator, determinant exactly 1 — and (b) the $N\to\infty$ limit of composed shrinking steps, i.e. the true flow. Keep the three objects distinct: truncated step (first-order canonical), generated step (exactly canonical, implicit), integrated flow (exactly canonical, the physics).

Worked examples

1 · Rotating a phase point by a centimeter of angle (confidence builder)

Apply the ICT generated by $G=L_z=xp_y-yp_x$ with $\varepsilon=0.010$ rad to the phase point $(x,y,p_x,p_y)=(3.0,\ 4.0,\ 2.0,\ 1.0)$ (SI units). Compare with the exact rotation and measure the first-order error.

Step 1. Displacements from (1), using the brackets $\{x,L_z\}=-y$ , $\{y,L_z\}=x$ , $\{p_x,L_z\}=-p_y$ , $\{p_y,L_z\}=p_x$ (each verified by explicit partials of $L_z$ ): $\delta x=-\varepsilon y=-0.040$ , $\delta y=+\varepsilon x=+0.030$ , $\delta p_x=-\varepsilon p_y=-0.010$ , $\delta p_y=+\varepsilon p_x=+0.020$ .

Step 2. ICT image: $(x',y',p_x',p_y')=(\mathbf{2.960,\ 4.030,\ 1.990,\ 1.020})$ . Note momenta rotate too — the ICT turns the whole phase-space vector, not just the configuration.

Step 3. Exact rotation by $0.010$ rad: $\cos\varepsilon=0.99995$ , $\sin\varepsilon=0.0099998$ , so $x_{\rm exact}=3.0\cos\varepsilon-4.0\sin\varepsilon=2.95985$ , $y_{\rm exact}=4.02980$ , $p_{x,\rm exact}=1.98990$ , $p_{y,\rm exact}=1.01995$ .

Step 4. Errors: $x'-x_{\rm exact}=+1.5\times10^{-4}$ , $y'-y_{\rm exact}=+2.0\times10^{-4}$ — both of size $\tfrac{1}{2}\varepsilon^2\times\mathcal{O}(x,y)$ , i.e. $O(\varepsilon^2)$ exactly as advertised, about 50 parts per million of the coordinates for a 0.57° turn.

Step 5. Invariants. $L_z$ before: $3.0\times1.0-4.0\times2.0=-5.0$ ; after the ICT: $2.960\times1.020-4.030\times1.990=\mathbf{-5.0005}$ — drift $5\times10^{-4}=\varepsilon^2|L_z|$ , second order (the exact rotation preserves $L_z$ identically; the truncated step does so only to first order). Likewise $x^2+y^2$ : $25.0\to25.0025$ , again a pure $\varepsilon^2$ artifact. The lesson in numbers: first-order canonical means first-order faithful — and measurably no better.

2 · One time-step of an oscillator: truncated step versus generated step (exam level)

A harmonic oscillator has $m=1.0$ kg, $\omega=2.0\ \mathrm{s^{-1}}$ , so $H=\tfrac{p^2}{2}+2q^2$ (joules with $q$ in m, $p$ in kg m/s). Starting at $(q,p)=(1.0,\ 0.0)$ , take one step $\varepsilon=\mathrm{d}t=0.10$ s using (a) the truncated ICT with $G=H$ , (b) the exact flow, (c) the untruncated $F_2$ -generated map. Compare canonicity and energy.

Step 1. Truncated ICT (this is the explicit-Euler map): $q_1=q+\mathrm{d}t\,p/m=\mathbf{1.000}$ m, $p_1=p-\mathrm{d}t\,m\omega^2q=-0.10\times4.0\times1.0=\mathbf{-0.400}$ kg m/s.

Step 2. Exact flow for comparison: $q(0.1)=\cos(0.2)=0.98007$ m, $p(0.1)=-2\sin(0.2)=-0.39734$ kg m/s. Errors: $\Delta q=+0.0199$ (first order in $\mathrm{d}t$ ), $\Delta p=-0.0027$ . First-order accurate, as an ICT must be.

Step 3. Canonicity of the truncated step. Its Jacobian is $M=I+\mathrm{d}t\,JS$ with $S=\mathrm{diag}(m\omega^2,\,1/m)$ , i.e. rows $(1,\ \mathrm{d}t/m)$ and $(-\mathrm{d}t\,m\omega^2,\ 1)$ : $\det M=1+\mathrm{d}t^2\omega^2=\mathbf{1.04}$ . Four percent volume inflation per step — second-order small, but fatally compounding: energy after the step is $E_1=\tfrac{0.4^2}{2}+2(1.0)^2=\mathbf{2.08}$ J against $E_0=2.00$ J, a $+4.0\%$ drift matching $\det M$ exactly (for this quadratic $H$ the inflation lands squarely on the energy).

Step 4. The untruncated generated map. Keep $F_2=qP+\mathrm{d}t\,H(q,P)$ without the argument swap: the Type-2 relations give $P_1=p-\mathrm{d}t\,m\omega^2q=-0.400$ first (implicit relation solved), then $Q_1=q+\mathrm{d}t\,P_1/m=1.0-0.040=\mathbf{0.960}$ m. This is the symplectic-Euler integrator. Its Jacobian has rows $(1-\mathrm{d}t^2\omega^2,\ \mathrm{d}t/m)$ and $(-\mathrm{d}t\,m\omega^2,\ 1)$ : $\det=(1-\mathrm{d}t^2\omega^2)+\mathrm{d}t^2\omega^2=\mathbf{1}$ , exactly, at any step size.

Step 5. Read the comparison. Both maps are first-order accurate ( $E$ after the symplectic step is $\tfrac{0.16}{2}+2(0.96)^2=1.923$ J — also off by ~4% instantaneously). The difference is structural: the truncated map dilates phase area by $1.04$ per step and its energy error grows without bound, while the generated map is exactly area-preserving and its energy error stays bounded forever, oscillating about the truth (it exactly conserves a nearby "shadow" Hamiltonian — the theme picked up on the discussion board of Poisson-Bracket Formulation of Dynamics). Same order of accuracy, opposite fates — canonicity, not accuracy, is what the $F_2$ construction buys.

Origins · 1 minSophus Lie and how this was discovered

Sophus Lie

1842–1899 · Christiania (Oslo) and Leipzig

Lie came to infinitesimal transformations from a geometer's frustration: Galois had classified algebraic equations by their discrete symmetry groups, and Lie wanted the same weapon for differential equations, whose symmetries are continuous. His insight — developed in an astonishing burst between 1869 and 1873, partly in Paris conversations with Felix Klein, partly while briefly jailed in Fontainebleau as a suspected spy carrying "coded" mathematical notes — was that a continuous group is completely controlled by its behavior infinitesimally close to the identity, where it linearizes into what we now call a Lie algebra. The composition law of the group collapses into brackets of generators; the intractable global object becomes tractable local algebra. Hamiltonian mechanics turned out to be the perfect host: Lie himself showed that canonical transformations form exactly such a continuous group, with phase-space functions as generators and the Poisson bracket as the algebra — the "function groups" of his 1888–1893 Theorie der Transformationsgruppen. He did not live to see the payoff: the quantum mechanics built in 1925–26 speaks Lie's language in every line, from $[\hat x,\hat p]=i\hbar$ to the classification of particles by group representations. The near-identity step $I+\varepsilon J\,\partial^2G$ of this page is Lie theory in its natural habitat.

History

1809

— Poisson's planetary-perturbation memoir supplies the bracket that will carry infinitesimal transformations

1837

— Jacobi builds transformation theory for dynamics: generating functions and the first canonical machinery

1871

— Lie begins the theory of continuous transformation groups — global groups tamed by infinitesimal generators

1888

— Lie's Transformationsgruppen volumes formalize function groups: phase-space functions as a Lie algebra under the Poisson bracket

1918

— Noether ties one-parameter symmetry groups to conservation laws on the Lagrangian side

1925

— Dirac maps the classical generator dictionary onto quantum commutators, making ICTs the skeleton of unitary evolution

Going further

References. Goldstein, Poole & Safko, Classical Mechanics, 3rd ed., §9.4–9.6 · Landau & Lifshitz, Mechanics, 3rd ed., §40–45 · Morin, Introduction to Classical Mechanics, Ch. 15 (Hamiltonian methods) · Taylor, Classical Mechanics, Ch. 13 · Arnold, Mathematical Methods of Classical Mechanics, §40 (Hamiltonian phase flow) Key identities worth confirming directly: re-expanding $F_2=q_iP_i+\varepsilon G$ yields $\delta q=\varepsilon\partial_pG$ , $\delta p=-\varepsilon\partial_qG$ ; the bracket relations $\{q_i,G\}=\partial G/\partial p_i$ and $\{p_i,G\}=-\partial G/\partial q_i$ follow; the fundamental brackets $\{x,p_x\}=1$ and $\{L_x,L_y\}=L_z$ hold; and $\det(I+\varepsilon J\partial^2G)=1+O(\varepsilon^2)$ underpins Liouville's theorem to first order. No empirical constants appear. The numeric examples use the rotation of $(3.0,4.0,2.0,1.0)$ by $0.010$ rad and the oscillator step comparison at $m=1.0$ , $\omega=2.0$ , $\mathrm{d}t=0.10$ .

Builds on: Definition of Canonical Transformations · Type-2 Generating Function F2(q,P,t) · Poisson Brackets: Definition · Symplectic Condition Leads to: Poisson-Bracket Formulation of Dynamics · Symplectic Group and Phase-Space Geometry · Algebra of Poisson Brackets · Poisson Brackets as a Test for Canonicity

Is epsilon a time? What exactly is it a small amount of?

Epsilon is a small amount of whatever parametrizes the family of transformations — it has no fixed identity of its own. For $G=L_z$ it is an angle in radians; for $G=p_x$ a length in meters; for $G=H$ a time in seconds. Dimensional bookkeeping enforces this: $\varepsilon G$ must carry the dimensions of a generating function (action), so $\varepsilon$ always carries action divided by $[G]$ — which is automatically the natural parameter of the flow $G$ generates. The habit worth building: when you meet a new generator, first work out what its $\varepsilon$ measures; the physics of the transformation is usually legible from that alone.

Why does the momentum displacement get the minus sign and not the coordinate one?

Because of which variable each Type-2 relation is solved for — there is genuinely no convention left to choose once the $F_2$ formalism is fixed. The relation $p=\partial F_2/\partial q$ produces $p=P+\varepsilon\,\partial G/\partial q$ , which must be solved for the new $P$ , flipping the sign; the relation $Q=\partial F_2/\partial P$ hands over $Q$ directly, keeping the plus. Geometrically the asymmetry is the antisymmetry of $J$ : the quarter-turn maps $(\partial_q,\partial_p)\to(\partial_p,-\partial_q)$ , one component sign-flipped. If both signs were positive the step would be a gradient flow, $G$ would change along it, and nothing would be conserved — the minus sign is where all the conservation physics hides.

In Derivation A, why are we allowed to swap the new P for the old p inside the derivative of G?

Because the swap happens inside a term already multiplied by $\varepsilon$ . Taylor-expanding $\partial G/\partial P$ about $P=p$ gives $\partial G/\partial p$ plus a remainder proportional to $P-p=O(\varepsilon)$ ; multiplied by the overall $\varepsilon$ , the remainder lands at $O(\varepsilon^2)$ , below first order. What students trip on is applying the same logic to finite maps: there the swap is not free, and the difference between $G(q,P)$ and $G(q,p)$ is exactly what separates the exactly-canonical symplectic-Euler map from the merely first-order-canonical explicit-Euler map in worked example 2. The swap is a first-order license, not a general permission.

What is the difference between the generator G and the generating function F2?

They live on different argument spaces and different scales. $F_2(q,P)$ is a function of mixed old-new variables that defines a finite canonical transformation implicitly through its partial derivatives; $G(q,p)$ is an ordinary phase-space function that defines an infinitesimal transformation explicitly through brackets. The relation is $F_2=\sum q_iP_i+\varepsilon G$ : the generator is the first-order coefficient of the generating function's deviation from the identity. In the finite-flow limit the two reconnect — Hamilton's principal function is the $F$ whose "infinitesimal part" is $H\,\mathrm{d}t$ — which is the road to Hamiltonian as Generator of Time Evolution and, further along, Hamilton–Jacobi theory.

Does the active versus passive distinction actually change any computed number?

It changes signs, which is worse than changing numbers — the error survives dimensional analysis. Active: the state moves, observables' recipes stay; $\delta f=+\varepsilon\{f,G\}$ . Passive: states stay, the chart moves; the functional form of $f$ in the new chart shifts by $-\varepsilon\{f,G\}$ . Both are correct, self-consistent descriptions of the same geometry. The classic failure mode is computing a rotation actively for coordinates and passively for momenta, which quietly rotates the two halves of phase space in opposite directions and destroys canonicity on paper while every individual line looks fine. Pick the active convention (this page's) and translate other sources' formulas before use.

Can the generator depend explicitly on time, and what changes if it does?

Yes — nothing in the construction of the displacement formulas requires $\partial G/\partial t=0$ ; the derivation goes through at each frozen instant, with $t$ a spectator inside $F_2$ . What changes is the conservation story: an explicitly time-dependent generator can be conserved only through cancellation between $\{G,H\}$ and $\partial G/\partial t$ — the Galilean boost $G=pt-mx$ is the canonical example, conserved for a free particle even though neither term vanishes alone. The boost's flow (at frozen $t$ ) is a perfectly good ICT: it shifts $x$ by $\varepsilon t$ and $p$ by $\varepsilon m$ . The full clocked-invariant story lives in Constants of Motion via Poisson Brackets.

If one epsilon step is only first-order canonical, how can the composition of many be exactly canonical?

Because the defect per step shrinks faster than the number of steps grows. Each step of size $\varepsilon=\lambda/N$ carries an $O(\varepsilon^2)=O(\lambda^2/N^2)$ canonicity defect; composing $N$ of them accumulates at most $N\times O(\lambda^2/N^2)=O(\lambda^2/N)$ , which vanishes as $N\to\infty$ . The limit — the integrated flow — is therefore exactly canonical even though every finite truncation is not. This is the same $\lim(1+x/N)^N=e^x$ mechanism as compound interest, and it is why the exponential map of Generators of Continuous Transformations produces exact symplectomorphisms out of approximate ingredients.

Is every canonical transformation close to the identity an ICT of the form given here — could some small map need a different structure?

To first order, yes, the form is exhaustive — that is what makes the result powerful. Any smooth one-parameter family of canonical maps starting at the identity has, at first order, a displacement field $\delta\eta=\varepsilon V(\eta)$ ; the symplectic condition forces $JV$ to have a symmetric Jacobian, which (on any simply-connected region, by the Poincaré lemma) makes $JV$ a gradient: $V=J\nabla G$ for some scalar $G$ . So "one scalar generates it" is a theorem, not a convenient ansatz. The fine print: on topologically nontrivial phase spaces $G$ may exist only locally — the displacement field is always locally generated but occasionally refuses to be one global function.