Poisson Brackets as a Test for Canonicity

Someone hands you a change of variables: $Q=\sqrt{2q}\cos p$ , $P=\sqrt{2q}\sin p$ . Perhaps it fell out of a clever substitution, perhaps a paper claims it simplifies a Hamiltonian. Before you trust a single formula computed in the new chart, one question must be answered: is this transformation canonical? Hunting for a generating function could answer it — eventually, after solving implicit relations you may not be able to invert. There is a faster court of appeal. Differentiate, assemble a handful of Poisson brackets, and compare against the fundamental table. Minutes, not hours; pure calculus, no PDEs. That court is this page. The verdict rests on a theorem with two directions: if a map is canonical, the new variables must satisfy $\{Q_i,Q_j\}=0$ , $\{P_i,P_j\}=0$ , $\{Q_i,P_j\}=\delta_{ij}$ with every bracket computed in the old variables — and, conversely, if they do satisfy them, the map is canonical: every bracket of every pair of observables then comes out identical in either chart, and Hamilton's equations survive the move. Necessity makes the test able to condemn; sufficiency makes it able to certify. We prove both, tie the test to the symplectic condition $M^{\mathsf T}JM=J$ of Symplectic Condition through one matrix identity, and specialize to one degree of freedom, where the whole verdict collapses to a single Jacobian determinant — canonicity as area preservation. The second half is craft: the checking algorithm and its bookkeeping (how many brackets, which ones antisymmetry gives you for free), the trick of running the test backwards to fix free parameters in a candidate family, and the time-dependent case, where $t$ rides along as a frozen spectator. The fingerprint concept itself — why the table determines everything — is owned by Fundamental Poisson Brackets; here the fingerprint goes to work.

Key takeaways

The test: $(q,p)\to(Q,P)$ is canonical iff $\{Q_i,Q_j\}=0$ , $\{P_i,P_j\}=0$ , $\{Q_i,P_j\}=\delta_{ij}$ , with all brackets computed in the old chart $(q,p)$ .
It is the symplectic condition in component form: the matrix of new-variable brackets equals $MJM^{\mathsf T}$ , so demanding the table is demanding $MJM^{\mathsf T}=J\iff M^{\mathsf T}JM=J$ .
One degree of freedom: $\{Q,P\}_{q,p}=\det M$ , the Jacobian determinant — the test is exactly area preservation, one number decides.
Sufficiency is what makes the test certify: if the table holds, every bracket of every pair of functions is chart-independent, and Hamilton's equations keep their form (with $K=H+\partial F/\partial t$ when the map depends on time).
Bookkeeping for $n$ degrees of freedom: $2n^2-n$ independent brackets ( $6$ for $n=2$ ); checking only the diagonal pairs $\{Q_i,P_i\}=1$ is not sufficient.
Run backwards, the test designs transformations: demand $\{Q,P\}=1$ on a parametrized family and solve for the parameters.
Key idea: canonicity $\iff$ the new variables satisfy the fundamental table in the old variables $\iff$ the Jacobian is symplectic — one theorem, three scripts, and the bracket script is the fastest to compute.

The core insight

Canonicity $\iff$ the new variables satisfy $\{Q_i,Q_j\}=\{P_i,P_j\}=0,\ \{Q_i,P_j\}=\delta_{ij}$ computed in the old variables $\iff$ the Jacobian is symplectic, $M^{\mathsf T}JM=J$ .

Where this lives today

The bracket test is production infrastructure wherever long-time Hamiltonian simulation matters. Particle-accelerator design is the sharpest example: a storage ring's lattice is modeled as a composition of transfer maps, one per magnet, and tracking codes (MAD-X, PTC, and their kin) symplectify those maps — truncated Taylor maps fail the test by small residuals, and the codes correct them until $M^{\mathsf T}JM=J$ holds to machine precision — because a residual of even $10^{-8}$ per turn, compounded over the $10^{9}$ turns of a beam lifetime, manufactures phantom emittance growth and predicts beams to be unstable that are not. "Does the one-turn map pass the canonicity test" is, quite literally, a gate in the software release process. The newest customers are machine-learned dynamics models. Symplectic neural networks — architectures whose layers are compositions of exact canonical maps (generating-function layers, shear layers) — enforce $\{Q,P\}=1$ by construction, so the learned surrogate of a molecular or celestial system inherits energy behavior and long-time stability that unconstrained networks lose within a few hundred steps. Training pipelines for these models include the bracket test as a validation metric: finite-difference the learned map's Jacobian, assemble $MJM^{\mathsf T}$ , and report the deviation from $J$ as a per-layer diagnostic. A two-century-old algebraic checkpoint has become a loss term.

The physical picture

Phase space carries a built-in area gauge — the symplectic structure that Fundamental Poisson Brackets wrote as the table $\{\eta_a,\eta_b\}=J_{ab}$ . A canonical transformation is a change of variables that leaves the gauge calibrated. The test is the calibration check at the border: the new variables arrive claiming citizenship, and the checkpoint computes their brackets using the old chart's certified machinery. Match the table and the border opens — every formula of Hamiltonian mechanics transfers verbatim. Fail one entry and the map is turned away, with the failed value reporting exactly how the calibration is off.

Figure

The canonicity checkpoint — new variables are interrogated with old-chart brackets, and the verdict is a table comparison

Why must the interrogation happen in the old chart? Because that is the only certified equipment in the room. Computing $\{Q,P\}$ with respect to $(Q,P)$ returns $1$ for any invertible relabeling whatsoever — the definition applied to its own coordinates is a tautology, not a test. The content lives entirely in the cross-examination: new functions, old derivatives. In one degree of freedom the geometric meaning becomes a picture you can hold. The bracket $\{Q,P\}_{q,p}$ is exactly the Jacobian determinant of the map — the factor by which an infinitesimal cell of phase-space area gets magnified. Canonical means the magnification is $1$ everywhere: cells may shear, stretch one way and squeeze the other, rotate — but never gain or lose area. The test result is the area report.

Figure

One degree of freedom — the bracket of Q with P equals the Jacobian determinant, the area magnification of an infinitesimal cell

One more habit of mind before the theorems: the test is local and pointwise — it constrains the Jacobian at each phase-space point — but the verdict must hold identically, at every point of the domain, not merely at a special point. $\{Q,P\}=1$ at the origin and $1.3$ elsewhere is a failed test. That is why the algebra typically ends with a cancellation that removes all $(q,p)$ -dependence; watching for that cancellation is watching the proof happen.

The theorem, both directions

State it cleanly. Given new variables $Q_i(q,p,t)$ , $P_i(q,p,t)$ , define the candidate brackets — everything evaluated with old-chart derivatives:

\{Q_i,Q_j\}_{q,p}=0,\qquad \{P_i,P_j\}_{q,p}=0,\qquad \{Q_i,P_j\}_{q,p}=\delta_{ij} \qquad (1)

where: $\{\,\}_{q,p}$ — the Poisson bracket of Poisson Brackets: Definition, with derivatives taken with respect to the original $(q,p)$ ; $\delta_{ij}$ — Kronecker delta. Equation (1) holding identically on the domain is the bracket test. The theorem: (1) holds iff the transformation is canonical. Necessity is inheritance: if the map is canonical, then $(Q,P)$ is a legitimate canonical chart, the bracket's value is chart-independent (proved from the symplectic condition in Poisson Brackets: Definition), and evaluating the new coordinates' mutual brackets in the old chart must return the same values they trivially have in their own chart — the fundamental table of Fundamental Poisson Brackets. Sufficiency is the direction the test earns its keep with, and it deserves its own derivation. The pivot is a matrix identity. Stack the old variables as $\xi=(q,p)^{\mathsf T}$ , the new as $\eta=(Q,P)^{\mathsf T}$ , with Jacobian

M_{ab}=\frac{\partial\eta_a}{\partial\xi_b},\qquad \big[\{\eta_a,\eta_b\}_{\xi}\big]=MJM^{\mathsf T} \qquad (2)

where: $J$ — the symplectic matrix $\begin{pmatrix}0&I_n\\-I_n&0\end{pmatrix}$ ; the bracketed object — the full $2n\times2n$ matrix whose entries are all candidate brackets at once. The identity is the chain rule (established as the fingerprint identity in Fundamental Poisson Brackets; we do not re-derive it). Demanding (1) is demanding this matrix equal $J$ :

MJM^{\mathsf T}=J\qquad\Longleftrightarrow\qquad M^{\mathsf T}JM=J \qquad (3)

where: the equivalence of the two orderings follows from the inversion lemma $M^{-1}=-JM^{\mathsf T}J$ (with $J^2=-I$ ), proved in the companion article. So the bracket test is the symplectic condition, entry by entry — and everything the symplectic condition certifies (form-invariance of Hamilton's equations, existence of local generating functions, membership of $M$ in $Sp(2n,\mathbb R)$ ) is certified by the table.

Derivation ASufficiency — the table forces every bracket to be chart independent

We are proving the certifying direction: assume the candidate brackets satisfy table (1), and show that for arbitrary observables $f,g$ the bracket computed in the old chart equals the bracket computed in the new one, $\{f,g\}_{q,p}=\{f,g\}_{Q,P}$ — from which the form-invariance of the dynamics follows at once. The strategy is a chain-rule expansion of $\{f,g\}_{q,p}$ through the new variables, followed by a regrouping in which every coefficient turns out to be one of the candidate brackets. The thing to watch: we never assume canonicity — only the table — so no circularity contaminates the converse.

Step 1. Treat $f$ and $g$ as functions of the new variables: $f=f\big(Q(q,p),P(q,p)\big)$ , and expand the old-chart gradients by the chain rule:

\frac{\partial f}{\partial q_k}=\sum_i\Big(\frac{\partial f}{\partial Q_i}\frac{\partial Q_i}{\partial q_k}+\frac{\partial f}{\partial P_i}\frac{\partial P_i}{\partial q_k}\Big)

and identically with $q_k\to p_k$ . Nothing is special about $f$ ; the same expansion applies to $g$ .

Step 2. Insert both expansions into $\{f,g\}_{q,p}=\sum_k(\partial_{q_k}f\,\partial_{p_k}g-\partial_{p_k}f\,\partial_{q_k}g)$ and multiply out. The result is a sum of eight types of terms, each carrying two derivatives of $f,g$ with respect to new variables and two derivatives of new variables with respect to old ones.

Step 3. Regroup by pulling the $f,g$ derivatives out of the $k$ -sum. Each surviving $k$ -sum is, by definition, a Poisson bracket of two new variables in the old chart:

\{f,g\}_{q,p}=\sum_{i,j}\Big(\frac{\partial f}{\partial Q_i}\frac{\partial g}{\partial Q_j}\{Q_i,Q_j\}+\frac{\partial f}{\partial Q_i}\frac{\partial g}{\partial P_j}\{Q_i,P_j\}+\frac{\partial f}{\partial P_i}\frac{\partial g}{\partial Q_j}\{P_i,Q_j\}+\frac{\partial f}{\partial P_i}\frac{\partial g}{\partial P_j}\{P_i,P_j\}\Big)

Why this regrouping is the whole proof: the arbitrary functions $f,g$ now touch the transformation only through the candidate brackets — the same compression mechanism as the reduction theorem of Fundamental Poisson Brackets, here running in reverse.

Step 4. Impose the assumed table: the $\{Q,Q\}$ and $\{P,P\}$ families kill their terms; $\{Q_i,P_j\}=\delta_{ij}$ and its antisymmetric partner $\{P_i,Q_j\}=-\delta_{ij}$ lock $j$ to $i$ :

\{f,g\}_{q,p}=\sum_i\Big(\frac{\partial f}{\partial Q_i}\frac{\partial g}{\partial P_i}-\frac{\partial f}{\partial P_i}\frac{\partial g}{\partial Q_i}\Big)=\{f,g\}_{Q,P}

Step 5. Harvest the dynamics. Apply the chart-independence to the pairs $(Q_i,H)$ and $(P_i,H)$ : the equations of motion $\dot Q_i=\{Q_i,H\}$ , $\dot P_i=\{P_i,H\}$ hold with the bracket computed in either chart, so Hamilton's equations take their canonical form in $(Q,P)$ . For time-dependent maps the same algebra goes through at each frozen $t$ , with the new Hamiltonian picking up the generating-function term, $K=H+\partial F/\partial t$ — the structure, not the clock, is what the test certifies. Two quick consistency checks confirm the result: setting $f=Q_i$ , $g=P_j$ in the final line returns $\delta_{ij}$ on both sides, and setting $f,g=q_k,p_l$ recovers the statement that the inverse map also passes the test.

Go deeper · Derivation B · 1 minOne degree of freedom — the test is a single Jacobian determinant

Derivation BOne degree of freedom — the test is a single Jacobian determinant

We are specializing to $n=1$ , where the three bracket families collapse to one number, and proving $\{Q,P\}_{q,p}=\det M$ : the test becomes "does the map preserve oriented area?" This is worth deriving separately because it is the workhorse case — most textbook problems and most quick sanity checks live here — and because it makes the geometric meaning of a failed test quantitative. Watch the bracket and the determinant turn out to be the same expression, not merely equal values.

Step 1. In one degree of freedom, $\{Q,Q\}=\{P,P\}=0$ hold automatically by antisymmetry — any function brackets to zero with itself. The entire test is the single condition $\{Q,P\}_{q,p}=1$ .

Step 2. Write out both objects. The bracket:

\{Q,P\}_{q,p}=\frac{\partial Q}{\partial q}\frac{\partial P}{\partial p}-\frac{\partial Q}{\partial p}\frac{\partial P}{\partial q}

The Jacobian matrix and its determinant:

M=\begin{pmatrix}\partial Q/\partial q&\partial Q/\partial p\\[2pt]\partial P/\partial q&\partial P/\partial p\end{pmatrix},\qquad \det M=\frac{\partial Q}{\partial q}\frac{\partial P}{\partial p}-\frac{\partial Q}{\partial p}\frac{\partial P}{\partial q}

Identical expressions, term for term: $\{Q,P\}_{q,p}=\det M$ .

Step 3. Geometry. The determinant of a Jacobian is the local oriented-area magnification of the map. So the test verdict reads: canonical $\iff$ every infinitesimal phase-space cell keeps its area (and orientation), $\mathrm dQ\,\mathrm dP=\mathrm dq\,\mathrm dp$ . A verdict of $\{Q,P\}=c$ means areas are uniformly magnified by $c$ ; a verdict depending on $(q,p)$ means the distortion varies from point to point — either way, not canonical. Why this is Liouville in embryo: time evolution will turn out to be a canonical map (Hamiltonian as Generator of Time Evolution), so this area statement, iterated along the flow, is the germ of Liouville's theorem — the full story, including higher-dimensional volume and the Poincaré invariants, lives in Liouville's Theorem and Phase-Space Volume.

Step 4. Consistency with the matrix form: for $n=1$ , $MJM^{\mathsf T}=J$ with $J=\begin{pmatrix}0&1\\-1&0\end{pmatrix}$ reduces (multiply out the $2\times2$ product) to $(\det M)\,J=J$ , i.e. $\det M=1$ — the same verdict from the symplectic condition, as it must be. The $2\times2$ identity $MJM^{\mathsf T}=(\det M)J$ can be confirmed by direct multiplication — it is the key step linking the symplectic condition to the area-preservation statement for one degree of freedom.

The algorithm in practice

The theorem becomes a procedure. For a candidate map in $n$ degrees of freedom:

Step 1. Assemble the partial derivatives of every $Q_i$ and $P_i$ with respect to every $q_k$ and $p_k$ — the raw material, $4n^2$ partials (many often vanish by inspection).

Step 2. Compute the three families in the old chart: $\{Q_i,Q_j\}$ and $\{P_i,P_j\}$ for $i<j$ only (antisymmetry makes $i>j$ redundant and $i=j$ automatic), and $\{Q_i,P_j\}$ for all ordered pairs — the delta pattern is not symmetric in $i\leftrightarrow j$ . Independent entries: $\tfrac{n(n-1)}{2}+\tfrac{n(n-1)}{2}+n^2=2n^2-n$ . For $n=1$ : one bracket. For $n=2$ : six. For $n=3$ : fifteen.

Step 3. Simplify each bracket symbolically and check it is identically $0$ or $\delta_{ij}$ across the whole domain — not merely at a convenient point. The signature of a passing test is a trigonometric or algebraic identity ( $\cos^2+\sin^2$ , a product of a function with its reciprocal) wiping out all coordinate dependence.

Step 4. Read a failure quantitatively. $\{Q,P\}=c\neq1$ constant: an extended canonical map — rescale one variable by $1/c$ (or absorb $c$ into the Hamiltonian) and canonicity is restored. $\{Q,P\}$ coordinate-dependent, or a same-type bracket nonzero: structurally non-canonical; no global rescaling repairs it. A worked instinct: before grinding all $2n^2-n$ entries, test the most suspicious bracket first — usually a cross-entry between different degrees of freedom, or the momentum–momentum bracket when the new momenta mix coordinates (recall the multi-DOF trap of Fundamental Poisson Brackets: perfect diagonal pairs, guilty $\{P_1,P_2\}$ ). A single failed entry ends the trial; only a full acquittal needs all entries. The test also designs. Given a parametrized family of candidate maps, impose the table as an equation for the parameters — the demand that coordinate dependence cancel identically is strong enough to fix them (worked example 2). Goldstein's classic exercise family $Q=q^{\alpha}\cos\beta p$ , $P=q^{\alpha}\sin\beta p$ surrenders both exponents this way in three lines.

\{Q,P\}_{q,p}=\alpha\beta\,q^{2\alpha-1}\stackrel{!}{=}1\ \text{for all }q\quad\Longrightarrow\quad \alpha=\tfrac12,\ \beta=2 \qquad (4)

where: the exponent must vanish, $2\alpha-1=0$ , or the bracket would vary across phase space; the prefactor then sets $\beta$ . The demand "identically, not at a point" did all the work — it converted one scalar equation into two.

Go deeper · Derivation C · 1 minCertifying the oscillator action-angle map

Derivation CCertifying the oscillator action-angle map

We are running the full test on the transformation that will later anchor action-angle theory: $Q=\arctan\!\big(m\omega q/p\big)$ , $P=\dfrac{p^2+m^2\omega^2q^2}{2m\omega}$ , for the harmonic oscillator with mass $m$ and frequency $\omega$ . The goal is the verdict $\{Q,P\}_{q,p}=1$ identically, certifying the map without ever exhibiting its generating function. The strategy is direct: four partials, one bracket, and a cancellation whose cleanliness is the point — the common factor $D=p^2+m^2\omega^2q^2$ appears in both terms and exits gracefully. Watch also the dimensions: $Q$ comes out dimensionless (an angle) and $P$ carries energy/frequency = action, exactly the conjugate pairing the table certifies.

Step 1. Partials of $Q$ . With $u=m\omega q/p$ and $\mathrm d\arctan u=\mathrm du/(1+u^2)$ , and writing $D=p^2+m^2\omega^2q^2$ :

\frac{\partial Q}{\partial q}=\frac{m\omega/p}{1+u^2}=\frac{m\omega p}{D},\qquad \frac{\partial Q}{\partial p}=\frac{-m\omega q/p^2}{1+u^2}=-\frac{m\omega q}{D}

Step 2. Partials of $P$ : direct differentiation of the quadratic form,

\frac{\partial P}{\partial q}=\frac{2m^2\omega^2q}{2m\omega}=m\omega q,\qquad \frac{\partial P}{\partial p}=\frac{2p}{2m\omega}=\frac{p}{m\omega}

Step 3. Assemble the single test bracket:

\{Q,P\}=\frac{\partial Q}{\partial q}\frac{\partial P}{\partial p}-\frac{\partial Q}{\partial p}\frac{\partial P}{\partial q}=\frac{m\omega p}{D}\cdot\frac{p}{m\omega}-\Big(-\frac{m\omega q}{D}\Big)\,m\omega q=\frac{p^2+m^2\omega^2q^2}{D}=1

The numerator reassembles $D$ exactly — the cancellation is an identity, valid at every phase-space point away from the coordinate singularity $p=0$ of the arctangent branch. Why this map matters beyond the exercise: $P=D/2m\omega=E/\omega$ is the oscillator's action variable and $Q$ its angle; the certificate $\{Q,P\}=1$ is what licenses writing $H=\omega P$ and reading off the motion as $\dot Q=\omega$ , $\dot P=0$ . The test delivered in four lines what the generating-function route ( $F_1=\tfrac12 m\omega q^2\cot Q$ ) delivers only after an implicit-function slog.

Step 4. Numerical spot-check (a habit worth keeping even when the algebra is airtight). Take $m=1$ , $\omega=2$ , $q=1$ , $p=2$ in SI units: $D=4+4=8$ ; $\partial Q/\partial q=2\cdot2/8=0.500$ ; $\partial P/\partial p=2/2=1.000$ ; $\partial Q/\partial p=-2/8=-0.250$ ; $\partial P/\partial q=1\cdot2\cdot1=2.000$ . Bracket: $0.500\times1.000-(-0.250)\times2.000=0.500+0.500=1.000$ . ✓ The numerical evaluation at $(q,p)=(1,2)$ , $m=1$ , $\omega=2$ reproduces the identity $\{Q,P\}=1$ to all displayed digits, confirming the algebra.

Limiting cases

Case	Result	Meaning
One degree of freedom	test $=$ single condition $\{Q,P\}=\det M=1$	canonicity is area preservation; one number decides
Linear map $\eta=M\xi$ , $M$ constant	test $=$ matrix check $M^{\mathsf T}JM=J$	brackets are constant; canonicity is membership in $Sp(2n,\mathbb R)$
Point transformation $Q=Q(q)$ , induced $P$	passes automatically	proved in Fundamental Poisson Brackets; no checking needed
Uniform rescaling verdict $\{Q,P\}=c$	extended canonical	repairable: rescale one variable by $1/c$ , or rescale the Hamiltonian
Time-dependent map $Q(q,p,t)$ , $P(q,p,t)$	test at each frozen $t$	$t$ is a spectator in the brackets; $K=H+\partial F/\partial t$ handles the clock
Identity map	$M=I$ , trivially $J=J$	the baseline every candidate is compared against

Common trapComputing the test brackets in the new variables proves nothing — and neither does checking only the conjugate pairs

Two failure modes account for most wrong canonicity verdicts. First: evaluating $\{Q,P\}$ with respect to $(Q,P)$ . That is the definition applied to its own chart — it returns $1$ for any invertible relabeling, canonical or not; the test has content only when the brackets are computed with old-chart derivatives, because the old chart is the one whose canonicity is already certified. Second: in $n\ge2$ degrees of freedom, verifying $\{Q_i,P_i\}=1$ for each $i$ and stopping. The dangerous entries are the cross brackets — $\{Q_1,P_2\}$ , $\{P_1,P_2\}$ , $\{Q_1,Q_2\}$ — where coupling between degrees of freedom hides; a map can carry perfect diagonal pairs and still shred the symplectic structure off-diagonal. The full count is $2n^2-n$ independent brackets. Check them, or check the matrix $MJM^{\mathsf T}$ wholesale — but never a proper subset.

Worked examples

1 · Three verdicts in ten minutes (confidence builder)

Test each map for canonicity: (a) $Q=\sqrt{2q}\cos p$ , $P=\sqrt{2q}\sin p$ ; (b) $Q=q^2$ , $P=p/(2q)$ ; (c) $Q=q$ , $P=p^2$ .

Step 1. Map (a). Partials: $\partial Q/\partial q=\cos p/\sqrt{2q}$ , $\partial Q/\partial p=-\sqrt{2q}\sin p$ , $\partial P/\partial q=\sin p/\sqrt{2q}$ , $\partial P/\partial p=\sqrt{2q}\cos p$ . Bracket:

\{Q,P\}=\frac{\cos p}{\sqrt{2q}}\cdot\sqrt{2q}\cos p-\big(-\sqrt{2q}\sin p\big)\cdot\frac{\sin p}{\sqrt{2q}}=\cos^2p+\sin^2p=\mathbf{1}

Identically $1$ by the Pythagorean identity — canonical (for $q>0$ ). Note the signature cancellation: all $q$ -dependence exited through the $\sqrt{2q}\times1/\sqrt{2q}$ pairing.

Step 2. Map (b) — designed to look guilty. $\partial Q/\partial q=2q$ , $\partial Q/\partial p=0$ , $\partial P/\partial p=1/(2q)$ , $\partial P/\partial q=-p/(2q^2)$ . Bracket: $\{Q,P\}=2q\cdot\frac{1}{2q}-0=\mathbf{1}$ . Canonical. The lesson cuts the other way from the usual warnings: a map that mangles the variables' appearance (squaring a coordinate, dividing a momentum by a coordinate) can still preserve the structure perfectly — appearance is no evidence either way; only the bracket verdict counts. (Chart caveat: the map is two-to-one in $q$ , so it is canonical on each half-plane $q>0$ , $q<0$ separately.)

Step 3. Map (c). $\{Q,P\}=\frac{\partial q}{\partial q}\frac{\partial(p^2)}{\partial p}-0=\mathbf{2p}$ . Not $1$ , and worse than a constant failure — the verdict varies over phase space, so no rescaling repairs it. Not canonical, and quantitatively: near $p=1$ the map double-counts conjugate-plane area; near $p=0.1$ it undercounts by five; formulas transplanted into $(Q,P)$ would err by a position-dependent factor. Momentum relabelings $p\to f(p)$ are only canonical when paired with the compensating coordinate change $Q=q/f'(p)$ -type corrections — the naive square is not.

2 · Fixing a family's parameters with the test (exam level)

For what values of the constants $\alpha,\beta$ is $Q=q^{\alpha}\cos\beta p$ , $P=q^{\alpha}\sin\beta p$ a canonical transformation? Verify the result numerically at $q=4$ , $p=0.3$ .

Step 1. Partials of the family: $\partial Q/\partial q=\alpha q^{\alpha-1}\cos\beta p$ , $\partial Q/\partial p=-\beta q^{\alpha}\sin\beta p$ , $\partial P/\partial q=\alpha q^{\alpha-1}\sin\beta p$ , $\partial P/\partial p=\beta q^{\alpha}\cos\beta p$ .

Step 2. The single test bracket ( $n=1$ ):

\{Q,P\}=\alpha q^{\alpha-1}\cos\beta p\cdot\beta q^{\alpha}\cos\beta p-\big(-\beta q^{\alpha}\sin\beta p\big)\cdot\alpha q^{\alpha-1}\sin\beta p=\alpha\beta\,q^{2\alpha-1}\big(\cos^2\beta p+\sin^2\beta p\big)=\alpha\beta\,q^{2\alpha-1}

Step 3. Impose the verdict identically in $q$ . Coordinate dependence must cancel: $2\alpha-1=0$ , so $\alpha=\tfrac12$ . The surviving prefactor must be unity: $\alpha\beta=1$ , so $\beta=2$ . The certified member of the family is $Q=\sqrt q\,\cos2p$ , $P=\sqrt q\,\sin2p$ — $\alpha=1/2$ , $\beta=2$ , Goldstein's classic. Any other $(\alpha,\beta)$ fails, either by coordinate dependence (wrong $\alpha$ ) or by a constant factor (wrong $\beta$ , an extended canonical map repairable by rescaling).

Step 4. Numerical verification at $q=4$ , $p=0.3$ (so $2p=0.6$ rad, $\cos0.6=0.8253$ , $\sin0.6=0.5646$ ). First term: $\partial Q/\partial q\cdot\partial P/\partial p=\frac{0.8253}{2\cdot2}\times\big(2\cdot2\cdot0.8253\big)=0.2063\times3.3013=0.6812$ . Second term: $-\partial Q/\partial p\cdot\partial P/\partial q=\big(2\cdot2\cdot0.5646\big)\times\frac{0.5646}{2\cdot2}=2.2586\times0.1412=0.3188$ . Sum: $0.6812+0.3188=\mathbf{1.0000}$ . ✓ The two terms split the unit bracket in the ratio $\cos^2:\,\sin^2$ — the identity is doing exactly what Step 2 promised.

Step 5. Sanity link back to example 1(a): $\sqrt{2q}\cos p$ and $\sqrt q\cos2p$ are different certified maps — the angle can be reparametrized against the radius in more than one canonical way, and each member must be certified on its own. The test is cheap enough that there is never a reason to guess.

Origins · 1 minCarl Gustav Jacob Jacobi and how this was discovered

Carl Gustav Jacob Jacobi

1804–1851 · Königsberg

Jacobi turned canonical transformations from a bag of substitution tricks into a theory. In his Königsberg lectures on dynamics (1842–43, published posthumously as the Vorlesungen über Dynamik), he organized what survives a change of phase-space variables — brackets, the form of the equations, generating functions — and built the machinery that bears the marks of his taste for identities: the Jacobi identity anchoring the bracket algebra, the Jacobian determinant governing changes of variables, and the transformation theory that culminates in Hamilton–Jacobi mechanics, where a single well-chosen canonical transformation solves the dynamics outright. He prized Poisson's bracket theorem as, in his words, the deepest result of analytical mechanics — then out-derived Poisson by explaining why it holds, via his identity. A prodigy who lectured on differential geometry at 21 and a legendarily demanding teacher, Jacobi reshaped how mechanics is taught: the modern sequence — brackets, canonical maps, the test, generating functions, Hamilton–Jacobi — is essentially the Königsberg syllabus. His health broke in 1843; he died of smallpox in Berlin at 46, leaving analysis, number theory, and mechanics each permanently rearranged.

History

1809

— Poisson's perturbation memoir introduces the brackets whose preservation will define canonicity

1837

— Jacobi's transformation papers systematize changes of variables that keep Hamilton's form

1842

— the Königsberg lectures: bracket preservation, generating functions, and Hamilton–Jacobi theory as one edifice

1890

— Poincaré's integral invariants recast the test's content as conserved phase-space areas

1927

— Birkhoff's Dynamical Systems brings the symplectic viewpoint to the test,

M^{\mathsf T}JM=J

as group membership

Going further

References. Goldstein, Poole & Safko, Classical Mechanics, 3rd ed., §9.4–9.5 · Landau & Lifshitz, Mechanics, 3rd ed., §45 · Morin, Introduction to Classical Mechanics, Ch. 15 · Taylor, Classical Mechanics, Ch. 13 · Arnold, Mathematical Methods of Classical Mechanics, §40 Key results to keep in mind: the chain-rule collapse in the sufficiency derivation yields $\{f,g\}_{Q,P}$ ; the equivalence $MJM^{\mathsf T}=J\iff M^{\mathsf T}JM=J$ follows via $M^{-1}=-JM^{\mathsf T}J$ with $J^2=-I$ ; and the oscillator action-angle map satisfies $\{Q,P\}=1$ with $P=E/\omega$ . The parameter-fixing example gives $\alpha\beta q^{2\alpha-1}$ , fixed to $\alpha=1/2$ , $\beta=2$ , with numerical check $0.6812+0.3188=1.0000$ at $q=4$ , $p=0.3$ ; the action-angle spot check gives $0.500+0.500=1.000$ at $m=1$ , $\omega=2$ , $q=1$ , $p=2$ .

Builds on: Poisson Brackets: Definition · Fundamental Poisson Brackets · Definition of Canonical Transformations · Symplectic Condition · Algebra of Poisson Brackets Leads to: Symplectic Group and Phase-Space Geometry · Poisson-Bracket Formulation of Dynamics · Constants of Motion via Poisson Brackets · Generating Functions and the Four Types

Why does computing the bracket of Q with P in the new chart itself always give 1, even for non-canonical maps?

Because the coordinate definition of the bracket, applied to a chart's own coordinates, is a tautology: $\{Q,P\}_{Q,P}=\partial Q/\partial Q\cdot\partial P/\partial P-\partial Q/\partial P\cdot\partial P/\partial Q=1\cdot1-0\cdot0=1$ , mechanically, for any independent pair of labels. No information about the map that produced $(Q,P)$ enters. The test has content precisely because it cross-examines: new functions, old derivatives. The old chart's brackets encode the symplectic structure you already trust; asking how the new variables register on that equipment is what detects distortion. This is the single most common way the test gets misapplied on exams.

Does the test have to pass at every point, or is one point enough?

Every point of the domain. Canonicity is a property of the map, and the symplectic condition constrains the Jacobian at each point separately; a bracket that equals $1$ at the origin but $1+q^2$ elsewhere certifies nothing (worked example 1(c) fails exactly this way, with verdict $2p$ ). In practice this is why you simplify symbolically rather than plugging in numbers first: the passing signature is an identity — coordinate dependence cancels wholesale. Numbers come afterwards, as spot checks of the algebra (and are worth doing: they catch dropped signs cheaply). One legitimate shortcut exists: for linear maps the Jacobian is constant, so one matrix check $M^{\mathsf T}JM=J$ settles all points at once.

If a map passes the test, is its inverse automatically canonical too?

Yes. Passing means $MJM^{\mathsf T}=J$ with $M$ invertible (a degenerate Jacobian would already have failed, since $\det M=\pm1$ is forced). The inversion lemma $M^{-1}=-JM^{\mathsf T}J$ then lets you verify the inverse's sandwich directly: substitute it into $M^{-1}J(M^{-1})^{\mathsf T}$ and multiply out using $J^{\mathsf T}=-J$ , $J^2=-I$ , and the assumed $M^{\mathsf T}JM=J$ ; the expression collapses to $J$ . More structurally: symplectic matrices form a group, $Sp(2n,\mathbb R)$ — closed under products and inverses — so compositions and inversions of canonical maps are canonical for free. This group structure is the subject of Symplectic Group and Phase-Space Geometry, and it is why canonical transformations can be built up in stages, each stage certified separately.

How does the test handle a transformation that depends explicitly on time?

Treat $t$ as a frozen parameter: compute the brackets with $(q,p)$ -derivatives only, at fixed $t$ , and demand the table hold at each instant. Time never enters the brackets themselves — it is a spectator. What time-dependence does change is the dynamics side: the new Hamiltonian is not $H$ re-expressed but $K=H+\partial F/\partial t$ , where $F$ is the generating function of the family (see the generating-function articles). The division of labor is clean: the bracket test certifies the symplectic structure is preserved at every instant; the $\partial F/\partial t$ term handles the clock. A moving-frame transformation like $Q=q-vt$ , $P=p$ passes the frozen-time test trivially ( $\{Q,P\}=1$ ) and shifts the Hamiltonian by the boost term.

My test came out to a constant, like 3, instead of 1. Is the map useless?

Not at all — it is an extended canonical transformation, and the constant is a uniform area magnification. Two standard repairs: rescale one of the new variables ( $P\to P/3$ ) to restore the exact table, or keep the variables and compensate in the dynamics by scaling the Hamiltonian ( $K=3H$ reproduces the correct equations of motion with the distorted brackets). Both appear in practice; the first is cleaner for quantization, the second common in scaling analyses where the magnification is the point. What has no repair is a non-constant verdict: position-dependent distortion cannot be absorbed by any global rescaling, and such a map genuinely scrambles the symplectic structure.

What is the relationship between the bracket test and Lagrange brackets?

They are matrix inverses of each other, and either can run the test. The Lagrange bracket $[u,v]=\sum_i(\partial_{u}q_i\,\partial_{v}p_i-\partial_{u}p_i\,\partial_{v}q_i)$ differentiates old variables with respect to new ones — the transpose viewpoint. Assembled into matrices, $\mathcal L=M^{\mathsf T}JM$ -type objects for Lagrange versus $MJM^{\mathsf T}$ for Poisson, and the classical identity states the two bracket matrices multiply to $-I$ (sign per convention). So "all Lagrange brackets take canonical values" and "all Poisson brackets take canonical values" are equivalent verdicts. Poisson won the practical war because its derivatives run new-with-respect-to-old — the direction you actually possess when someone hands you $Q(q,p)$ , $P(q,p)$ explicitly.

Can I use the test on a numerical map, like one step of an integrator, where I have no formulas?

Yes — this is a standard diagnostic in accelerator physics and geometric integration. Finite-difference the map: apply it to a cluster of nearby initial conditions, estimate the Jacobian $M$ column by column, assemble $MJM^{\mathsf T}$ , and measure $\|MJM^{\mathsf T}-J\|$ . For a genuinely symplectic scheme (leapfrog, symplectic Euler, splitting methods) the residual sits at finite-difference noise, and — the subtle point — it does not grow with the number of composed steps, because symplectic matrices form a group. For a non-symplectic scheme (classical RK4) the one-step residual is tiny, order of the local truncation error, but compounds along the trajectory; the test applied to the $10^6$ -step map exposes what the one-step test nearly hides. Testing the long-time composed map, not the single step, is the craft point.

If the test passes, does a generating function for the map actually exist?

Locally, yes — that is part of what sufficiency buys. A passing test means the Jacobian is symplectic everywhere, which makes the relevant one-form closed ( $\mathrm d(p\,\mathrm dq-P\,\mathrm dQ)=0$ in the two-form language of Symplectic Group and Phase-Space Geometry), and closed forms are locally exact: a generating function exists in a neighborhood of any point, of whichever type ( $F_1$ – $F_4$ ) suits the variables' independence pattern there. Globally, topology can obstruct a single-valued $F$ — the action-angle map of Derivation C is the canonical example, where the angle's multivaluedness makes $F_1$ branch. The practical division: use the test to certify, use generating functions to construct; needing the certificate before the construction is exactly why the test exists.