Symplectic Group and Phase-Space Geometry

The symplectic condition of the previous article certifies transformations one at a time. Step back and look at the collection: matrices satisfying $M^{\mathsf T}JM=J$ compose into each other, invert within the family, and contain the identity. They form a group — the symplectic group $\mathrm{Sp}(2n,\mathbb{R})$ — and that upgrade from "condition" to "group" changes the subject. Hamiltonian evolution stops being a differential equation you solve and becomes a path traced through the group, starting at the identity, steered by the Hamiltonian's Hessian through the group's Lie algebra $\mathfrak{sp}(2n)$ . Stability of orbits becomes spectral theory in the group: symplectic eigenvalues are locked into quadruples $\{\lambda,\,1/\lambda,\,\bar\lambda,\,1/\bar\lambda\}$ , which is why Hamiltonian systems can never simply damp down into an attractor. Globalize, and phase space itself acquires its true identity: a symplectic manifold $(\mathcal{M},\omega)$ — a smooth space equipped with a closed, nondegenerate two-form that measures oriented conjugate-pair area. Canonical transformations are its symmetries (symplectomorphisms), Liouville volume is the top power $\omega^n/n!$ , and Darboux's theorem delivers the punchline: symplectic geometry has no local invariants at all — every symplectic manifold looks locally like the standard $(q,p)$ with the standard $J$ . There is no symplectic curvature. All structure is global, and it is startlingly rigid: Gromov's non-squeezing theorem forbids a canonical map from threading a ball through a conjugate-plane hole narrower than the ball, no matter that volume would happily allow it. This page owns the group, its algebra, its spectra, and the geometry — the gateway from mechanics to symplectic topology, integrable tori, and quantization.

Key takeaways

The set of matrices obeying $M^{\mathsf T}JM=J$ forms the symplectic group $\mathrm{Sp}(2n,\mathbb{R})$ : closed under products and inverses, a Lie group of dimension $n(2n+1)$ , non-compact.
Its Lie algebra is $\mathfrak{sp}(2n)=\{A:\ A^{\mathsf T}J+JA=0\}=\{JS:\ S=S^{\mathsf T}\}$ ; linearized Hamiltonian dynamics has exactly this generator form, $A=J\,\mathrm{Hess}(H)$ .
The time- $t$ flow map of any Hamiltonian is symplectic: $\tfrac{\mathrm{d}}{\mathrm{d}t}(M^{\mathsf T}JM)=0$ along the variational equation — dynamics is a path in the group.
Symplectic spectra are rigid: eigenvalues come in quadruples $\{\lambda,1/\lambda,\bar\lambda,1/\bar\lambda\}$ , so stability lives only on the unit circle and asymptotic stability is impossible without dissipation.
Globally, phase space is a symplectic manifold $(\mathcal{M},\omega)$ , $\omega=\sum_i\mathrm{d}p_i\wedge\mathrm{d}q_i$ closed and nondegenerate; canonical transformations are the diffeomorphisms with $\phi^*\omega=\omega$ ; Liouville volume is $\omega^n/n!$ .
Darboux: no local symplectic invariants — locally everything is the standard chart; the interesting structure (tori, capacities, non-squeezing) is global.
Key idea: a transformation is canonical iff its Jacobian preserves the symplectic form, $M^{\mathsf T}JM=J$ ; the matrices obeying this constitute $\mathrm{Sp}(2n,\mathbb{R})$ , and Hamiltonian flow is a path in this group.

The core insight

A transformation is canonical iff its Jacobian preserves the symplectic form: $M^{\mathsf T} J M = J$ . The matrices obeying this constitute the symplectic group $\mathrm{Sp}(2n,\mathbb{R})$ , and Hamiltonian flow is a path in this group.

Where this lives today

The group is now laboratory equipment. In quantum optics, the Gaussian operations on a set of field modes — beam splitters, phase shifters, squeezers — implement exactly $\mathrm{Sp}(2n,\mathbb{R})$ on the quadrature vector, with the compact corner $\mathrm{U}(n)$ (beam splitters and phases) energy-conserving and the non-compact squeezing directions pumped by nonlinear crystals; LIGO's squeezed-light upgrade is a deliberate excursion along a non-compact group direction, and continuous-variable quantum computing specifies its Gaussian gates as symplectic matrices outright. On the theory side, the geometry of this page became its own field: Gromov's non-squeezing spawned symplectic capacities and Floer homology (built to prove Arnold's fixed-point conjecture), which in turn power mirror symmetry in string theory. And in the numerics of everything from exoplanet stability to protein folding, integrators are designed so their update maps sit exactly in $\mathrm{Sp}(2n)$ — worked example 2's spectrum analysis is, unchanged, how accelerator physicists read a one-turn matrix and decide whether a $27$ -kilometer ring will hold its beam for ten hours.

The physical picture

Picture the group as a curved landscape of all admissible reshufflings of phase space, with the identity — "change nothing" — as home base. Every canonical transformation you have built in this chapter is one point in that landscape. What dynamics adds is motion through it: as time runs, the flow map of a Hamiltonian system traces a continuous path $M(t)$ starting at home, its velocity at every moment dictated by the Hamiltonian's local curvature (the Hessian). Different Hamiltonians trace different paths; all paths lie inside the same landscape. Mechanics, seen from this altitude, is the study of one group and the curves it contains.

Figure

Hamiltonian flow as a path in the symplectic group — the algebra is the space of allowed velocities at the identity

Zoom out from linear maps to the arena they act on. Phase space is not a vector space with a ruler; it is a manifold carrying one geometric instrument only — the two-form $\omega$ , a signed-area meter for conjugate pairs. Feed it two displacement vectors and it returns the sum of the oriented areas of their parallelogram's shadows on the $(q_1,p_1),\dots,(q_n,p_n)$ planes. Canonical maps are exactly the motions this instrument cannot detect. Two consequences frame everything else. First, no lengths means no curvature: Darboux's theorem says any symplectic manifold is locally indistinguishable from the flat standard one — the polar opposite of Riemannian geometry, where curvature is measurable in any neighborhood. Second, no lengths does not mean no rigidity: the area meter constrains maps far beyond volume bookkeeping, and the sharpest expression is Gromov's non-squeezing theorem — the "symplectic camel" that cannot pass through the eye of a conjugate-plane needle.

Figure

Symplectic rigidity — volume-preserving maps could thread the ball through the narrow cylinder, canonical maps cannot

The group and its algebra

Package the definition once and for all:

\mathrm{Sp}(2n,\mathbb{R})=\left\{M\in\mathrm{GL}(2n,\mathbb{R})\ :\ M^{\mathsf T}JM=J\right\} \qquad (1)

where: $J$ — the standard antisymmetric form of Symplectic Condition, $J^{\mathsf T}=-J=J^{-1}$ ; $\mathrm{GL}$ — invertible matrices. The claim hidden in the word "group" — that this set closes under composition and inversion — is Derivation A, and it is the algebraic engine behind an everyday physical fact: doing one canonical change of variables after another is again a canonical change of variables, and undoing one is too. Counting freedom: the condition imposes $n(2n+1)$ independent constraints on $4n^2$ entries, leaving

\dim\mathrm{Sp}(2n,\mathbb{R})=n(2n+1) \qquad (2)

where: for one degree of freedom this gives $3$ — and indeed $\mathrm{Sp}(2,\mathbb{R})=\mathrm{SL}(2,\mathbb{R})$ , the unit-determinant $2\times2$ matrices (rotation, squeeze, shear: three knobs). For $n=2$ , dimension $10$ against $\mathrm{SL}(4)$ 's $15$ : from two freedoms on, symplectic is a genuinely smaller, stiffer family than volume-preserving. Near the identity, write $M=I+\epsilon A$ and expand the defining relation to first order:

\mathfrak{sp}(2n,\mathbb{R})=\left\{A\ :\ A^{\mathsf T}J+JA=0\right\}=\left\{JS\ :\ S=S^{\mathsf T}\right\} \qquad (3)

where: $\mathfrak{sp}$ — the Lie algebra, the space of allowed "velocities" through the group at the identity; the parametrization $A=JS$ says every generator is $J$ times a symmetric matrix. Count: symmetric $2n\times2n$ matrices have $n(2n+1)$ entries — matching (2), as a Lie algebra must. Now the punchline that ties the group to mechanics. Linearize any Hamiltonian flow: a displacement $\delta z$ between two neighboring trajectories obeys the variational equation

\frac{\mathrm{d}}{\mathrm{d}t}\,\delta z=J\,\mathrm{Hess}(H)\,\delta z,\qquad \mathrm{Hess}(H)_{ij}=\frac{\partial^{2}H}{\partial z_i\,\partial z_j} \qquad (4)

where: $\mathrm{Hess}(H)$ — symmetric by equality of mixed partials. The generator $J\,\mathrm{Hess}(H)$ is exactly of the form (3): Hamiltonian dynamics does not merely visit the symplectic group — its velocity is confined to the symplectic algebra at every instant, for every $H$ . Which Hamiltonian you have determines which path; that it stays in the group is guaranteed before you specify anything.

Derivation AThe symplectic matrices form a group

We are verifying the group axioms for the set (1) — identity, closure, inverses — because everything downstream (flows compose, transformations undo, spectra are constrained) leans on this structure. The strategy is direct matrix algebra on the defining relation; the only non-obvious ingredient is invertibility, which the determinant result supplies. Watch how mechanically the axioms fall out: the congruence form of the condition is engineered for exactly this.

Step 1. Identity: $I^{\mathsf T}JI=J$ trivially — the do-nothing map is canonical, and the set is nonempty.

Step 2. Closure. Let $M,N$ both satisfy the condition. Then

(MN)^{\mathsf T}J\,(MN)=N^{\mathsf T}\!\left(M^{\mathsf T}JM\right)\!N=N^{\mathsf T}JN=J

Why this matters physically: composing canonical transformations — polar coordinates, then action-angle, then a rotating frame — never needs re-certification. Each factor's certificate transfers to the product.

Step 3. Invertibility. Taking determinants of the condition gives $(\det M)^2=1$ (and the Pfaffian argument of Symplectic Condition sharpens this to $\det M=+1$ ), so $M^{-1}$ exists for every member.

Step 4. The inverse is symplectic. Multiply $M^{\mathsf T}JM=J$ by $(M^{\mathsf T})^{-1}$ on the left and $M^{-1}$ on the right:

J=(M^{-1})^{\mathsf T}\,J\,M^{-1}

which is the defining relation for $M^{-1}$ . Explicitly, $M^{-1}=-JM^{\mathsf T}J$ — inverses come for free, no Gaussian elimination.

Step 5. The condition is polynomial in the matrix entries, so $\mathrm{Sp}(2n,\mathbb{R})$ is a closed subgroup of $\mathrm{GL}(2n,\mathbb{R})$ — hence a Lie group, with the algebra (3) as its tangent space at the identity. Verified: the same closure argument with $J\to I$ reproduces the orthogonal group — the two classical families differ only in the symmetry of the preserved form; cross-checked against Goldstein §9.4 and Arnold Ch. 8.

Go deeper · Derivation B · 1 minHamiltonian flow stays symplectic for all time

Derivation BHamiltonian flow stays symplectic for all time

We are proving the central dynamical theorem of the chapter in its strongest form: the time- $t$ flow map of an arbitrary (even time-dependent) Hamiltonian has symplectic Jacobian for every $t$ — not just infinitesimally. The strategy is a conservation law in disguise: show that the matrix $M^{\mathsf T}JM$ has zero time derivative along the variational equation and starts at $J$ . This differential proof complements the composition-of-infinitesimals argument in Symplectic Condition, and it is the form used in practice to check numerical integrators.

Step 1. Let $M(t)=\partial z(t)/\partial z(0)$ be the Jacobian of the flow with respect to initial conditions (the variational or monodromy matrix). Differentiating the flow equation $\dot z=J\nabla_zH$ with respect to $z(0)$ gives the linear evolution

\dot M=A(t)\,M,\qquad A(t)=J\,\mathrm{Hess}(H)\big(z(t)\big)

Step 2. Check that $A$ lies in the algebra: with $S=\mathrm{Hess}(H)$ symmetric,

A^{\mathsf T}J+JA=S^{\mathsf T}J^{\mathsf T}J+J^{2}S=S-S=0

using $J^{\mathsf T}J=I$ and $J^{2}=-I$ . Why time dependence is harmless: nothing in this step used constancy of $S$ — the Hessian may vary along the trajectory and with explicit time; symmetry is all that is needed, and Hessians are always symmetric.

Step 3. Differentiate the candidate invariant:

\frac{\mathrm{d}}{\mathrm{d}t}\left(M^{\mathsf T}JM\right)=\dot M^{\mathsf T}JM+M^{\mathsf T}J\dot M=M^{\mathsf T}\!\left(A^{\mathsf T}J+JA\right)\!M=0

Step 4. Evaluate at $t=0$ : $M(0)=I$ , so $M^{\mathsf T}JM=J$ initially — and by Step 3, forever. The flow map is symplectic at all times; over a periodic orbit, the once-around $M(T)$ is the symplectic monodromy matrix whose spectrum (Derivation C) decides stability. Verified: equivalent to the statement that all fundamental brackets $\{z_i(t),z_j(t)\}$ are constants of the flow (Goldstein §9.4, Landau–Lifshitz §45); for the oscillator the explicit $M(t)$ was checked numerically at $\omega t=\pi/3$ in the Symplectic Condition article.

The spectral rigidity of symplectic maps

Because $M^{-1}=-JM^{\mathsf T}J$ , the inverse of a symplectic matrix is similar to its transpose — and that single fact handcuffs the spectrum.

Go deeper · Derivation C · 1 minEigenvalues come in quadruples

Derivation CEigenvalues come in quadruples

We are establishing the reflexive spectrum theorem: if $\lambda$ is an eigenvalue of a symplectic matrix, so are $1/\lambda$ , $\bar\lambda$ , and $1/\bar\lambda$ . This is the mathematical origin of every "Hamiltonian systems cannot be asymptotically stable" statement in mechanics. The strategy: manipulate the characteristic polynomial using the similarity $M^{-1}\sim M^{\mathsf T}$ , then add realness. Watch how the determinant $\det M=1$ enters to keep the bookkeeping exact.

Step 1. From the condition, $M^{\mathsf T}=JM^{-1}J^{-1}$ : the transpose is similar to the inverse. Similar matrices share characteristic polynomials, and every matrix shares its characteristic polynomial with its transpose. Hence $p(\lambda)\equiv\det(M-\lambda I)$ satisfies

p(\lambda)=\det(M^{\mathsf T}-\lambda I)=\det(M^{-1}-\lambda I)

Step 2. Factor the inverse out:

\det(M^{-1}-\lambda I)=\det\!\big(M^{-1}(I-\lambda M)\big)=\det(M^{-1})\,(-\lambda)^{2n}\det\!\big(M-\tfrac{1}{\lambda}I\big)

With $\det M^{-1}=1$ this is the reflexivity identity

p(\lambda)=\lambda^{2n}\,p\!\left(\frac{1}{\lambda}\right)

Why this is the whole theorem: if $p(\lambda_0)=0$ with $\lambda_0\neq0$ (symplectic matrices are invertible, so $0$ never occurs), the identity forces $p(1/\lambda_0)=0$ — reciprocals travel together.

Step 3. Add realness: $M$ has real entries, so complex eigenvalues arrive with conjugates. Combining, the spectrum is closed under both $\lambda\to1/\lambda$ and $\lambda\to\bar\lambda$ : the generic orbit is the quadruple $\{\lambda,\ 1/\lambda,\ \bar\lambda,\ 1/\bar\lambda\}$ , degenerating to pairs on the unit circle ( $\bar\lambda=1/\lambda$ ) or the real axis ( $\bar\lambda=\lambda$ ), and to singletons only at $\pm1$ .

Step 4. Read off the stability doctrine. Bounded linearized motion needs every $|\lambda|=1$ . The quadruple structure means eigenvalues cannot leak off the circle one at a time — instability sets in only when two eigenvalues collide on the circle and split into a reciprocal pair off it. This collision scenario is parametric resonance; the Krein theory of which collisions are dangerous governs everything from Mathieu equations to storage-ring lattices. Verified: consistent with Arnold Ch. 8 and the accelerator-physics trace criterion; the oscillator monodromy eigenvalues $e^{\pm i\omega T}$ and the inverted oscillator's $e^{\pm\mu T}$ instantiate the circle pair and the real reciprocal pair respectively.

From matrices to manifolds: the geometry of phase space

Everything so far was linear. The full arena is a $2n$ -dimensional symplectic manifold $(\mathcal{M},\omega)$ : a smooth space equipped with a two-form

\omega=\sum_{i=1}^{n}\mathrm{d}p_i\wedge\mathrm{d}q_i,\qquad \mathrm{d}\omega=0,\qquad \omega\ \text{nondegenerate} \qquad (5)

where: $\omega$ — the area meter, whose matrix in canonical coordinates is $J$ ; closed ( $\mathrm{d}\omega=0$ ) — the meter is consistent across neighborhoods, which is what makes Poincaré's invariants exist; nondegenerate — no direction is invisible to it, which forces even dimension and gives every energy function a well-defined flow. A canonical transformation is a symplectomorphism: a diffeomorphism $\phi$ with $\phi^{*}\omega=\omega$ — the coordinate-free wording of $M^{\mathsf T}JM=J$ , now allowed to act between curved charts. The volume form is not extra structure — it is a power of the area meter:

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

where: $\omega^n$ — the $n$ -fold wedge. Preserve $\omega$ and you preserve all its powers: Liouville's theorem is the $k=n$ member of a whole ladder — Poincaré's integral invariants $\int\omega^{k}$ , the conserved "areas" of $2k$ -dimensional surfaces for every $k$ — with the physics developed in the Liouville article of the recovery lane. Two structural theorems complete the picture. Darboux (1882): around any point of any symplectic manifold there exist coordinates in which $\omega$ takes exactly the standard form (5) — no symplectic analogue of curvature exists, so all local questions reduce to the linear theory of $\mathrm{Sp}(2n)$ . Arnold–Liouville: when a system has $n$ independent conserved quantities in involution (mutual Poisson brackets zero — see Algebra of Poisson Brackets), the motion is confined to $n$ -dimensional tori on which $\omega$ vanishes identically — Lagrangian submanifolds — and canonical action-angle coordinates straighten the flow into uniform winding. Those tori are the stage for Action-Angle Variables and the objects KAM theory famously shows mostly survive perturbation.

Limiting cases

Case	Result	Meaning
$n=1$	$\mathrm{Sp}(2,\mathbb{R})=\mathrm{SL}(2,\mathbb{R})$ , dimension 3	symplectic = unit determinant only here; rotation, squeeze, shear
$n\geq2$	$\dim\mathrm{Sp}(2n)=n(2n+1)<\dim\mathrm{SL}(2n)=4n^2-1$	symplectic strictly smaller than volume-preserving; the gap is Gromov rigidity
Compact corner	$\mathrm{Sp}(2n)\cap\mathrm{O}(2n)\cong\mathrm{U}(n)$	maps preserving areas and lengths are the unitary group — the classical seed of quantum mechanics
Spectrum on unit circle	all $\lvert\lambda\rvert=1$	stable (elliptic) linearized motion; generic for oscillators
Real reciprocal pair	$\lambda,1/\lambda$ real	hyperbolic instability: growth in one direction paid by decay in the partner
Odd dimension	no nondegenerate antisymmetric form exists	phase spaces are always even-dimensional; odd-dimensional level sets carry contact geometry instead

Common trapThe symplectic group is non-compact — canonical does not mean tame

Because the preserved form is antisymmetric, $\mathrm{Sp}(2n,\mathbb{R})$ contains unbounded squeezes: $Q=\lambda q$ , $P=p/\lambda$ is canonical for any $\lambda$ , however extreme. Nothing in canonicity prevents a blob of states from being stretched into an arbitrarily long, thin filament — that is precisely what chaos does, at exponential rate, while every symplectic invariant stays exact. Do not conflate "canonical transformation" with "small" or "benign" (a mistake encouraged by the compact rotation groups, where preserved lengths bound everything). The correct intuitions: volume exact, conjugate-pair areas exact, eigenvalues reciprocally paired — and diameters, lengths, energies of subsystems completely unprotected.

Worked examples

1 · Group operations, concretely (confidence builder)

In $\mathrm{Sp}(2,\mathbb{R})$ take the shear $M_1=\begin{pmatrix}1&0.5\\0&1\end{pmatrix}$ and the squeeze $M_2=\begin{pmatrix}2&0\\0&0.5\end{pmatrix}$ (units absorbed). Verify both are symplectic, compose them, and compute $M_1^{-1}$ from the free-inverse formula.

Step 1. Membership. For $n=1$ , symplectic $\iff\det=1$ : $\det M_1=1\times1-0.5\times0=1$ ✓; $\det M_2=2\times0.5=1$ ✓.

Step 2. Closure. $M_1M_2=\begin{pmatrix}1&0.5\\0&1\end{pmatrix}\begin{pmatrix}2&0\\0&0.5\end{pmatrix}=\begin{pmatrix}2&0.25\\0&0.5\end{pmatrix}$ , and $\det=2\times0.5-0.25\times0=1$ ✓ — the product is again symplectic, no re-check of anything else required. Note the order matters: $M_2M_1=\begin{pmatrix}2&1\\0&0.5\end{pmatrix}\neq M_1M_2$ — the group is non-abelian, which is why sequences of canonical transformations must be composed in the order performed.

Step 3. Free inverse. $M_1^{-1}=-JM_1^{\mathsf T}J$ : first $M_1^{\mathsf T}J=\begin{pmatrix}1&0\\0.5&1\end{pmatrix}\begin{pmatrix}0&1\\-1&0\end{pmatrix}=\begin{pmatrix}0&1\\-1&0.5\end{pmatrix}$ ; then

M_1^{-1}=-\begin{pmatrix}0&1\\-1&0\end{pmatrix}\begin{pmatrix}0&1\\-1&0.5\end{pmatrix}=-\begin{pmatrix}-1&0.5\\0&-1\end{pmatrix}=\begin{pmatrix}1&-0.5\\0&1\end{pmatrix}

Check: $M_1M_1^{-1}=\begin{pmatrix}1&0.5\\0&1\end{pmatrix}\begin{pmatrix}1&-0.5\\0&1\end{pmatrix}=I$ ✓. The inverse of a shear is the opposite shear — obtained with two matrix multiplications and zero elimination steps, courtesy of the group structure.

2 · A four-dimensional spectrum and the stability verdict (exam level)

A linearized Hamiltonian system with two degrees of freedom has the once-around monodromy matrix, in coordinates $(q_1,p_1,q_2,p_2)$ (each conjugate block $2\times2$ , block-diagonal):

M=\begin{pmatrix}\cos60^{\circ}&\sin60^{\circ}&0&0\\ -\sin60^{\circ}&\cos60^{\circ}&0&0\\ 0&0&e^{0.5}&0\\ 0&0&0&e^{-0.5}\end{pmatrix}=\begin{pmatrix}0.5&0.86603&0&0\\ -0.86603&0.5&0&0\\ 0&0&1.64872&0\\ 0&0&0&0.60653\end{pmatrix}

Verify symplecticity, list the spectrum, classify the stability, and exhibit a unit-determinant map on the same space that is NOT symplectic.

Step 1. Symplecticity block by block. With block-diagonal $M$ and block-diagonal $J$ , the sandwich $M^{\mathsf T}JM$ splits into the two $2\times2$ sandwiches. Block 1 (rotation): $\det=0.5^2+0.86603^2=0.25+0.75=1.000$ ✓. Block 2 (hyperbolic): $\det=1.64872\times0.60653=1.0000$ ✓ (indeed $e^{0.5}e^{-0.5}=1$ exactly). Each block passes the $n=1$ criterion, so the full $4\times4$ matrix satisfies $M^{\mathsf T}JM=J$ ✓.

Step 2. Spectrum. Block 1: $\lambda=0.5\pm0.86603\,i=e^{\pm i\pi/3}$ , modulus $1.000$ — a unit-circle pair. Block 2: $\lambda=1.64872$ and $0.60653$ — a real reciprocal pair, product $1.0000$ ✓. The four eigenvalues realize the quadruple theorem's two degenerate forms simultaneously: $\{e^{i\pi/3},e^{-i\pi/3}\}$ (circle pair) and $\{e^{0.5},e^{-0.5}\}$ (real pair).

Step 3. Stability verdict. Freedom 1 is elliptic (stable): perturbations rotate, bounded forever, winding $60^{\circ}$ per period. Freedom 2 is hyperbolic (unstable): a perturbation along the expanding eigenvector grows by $e^{0.5}=1.649$ per period — after $k$ periods, $e^{0.5k}$ — while its partner contracts by exactly the reciprocal. The orbit is a saddle-center: stable to some perturbations, unstable to others, phase-space area exactly conserved throughout ( $1.649\times0.607=1$ — growth is paid for, never created).

Step 4. The impostor. $D=\mathrm{diag}(2,\,1,\,0.5,\,1)$ has $\det D=2\times1\times0.5\times1=1$ — volume-preserving. But its first-block sandwich gives $\{Q_1,P_1\}=\partial_{q_1}Q_1\,\partial_{p_1}P_1=2\times1=2\neq1$ : it inflates the $(q_1,p_1)$ area by 2 while deflating $(q_2,p_2)$ by 2. Unit determinant, not symplectic — the $n\geq2$ trap of the warning box, realized in four numbers. No Hamiltonian flow, over any time, can produce $D$ .

Origins · 1 minVladimir Arnold and how this was discovered

Vladimir Arnold

1937–2010 · Moscow State University and Steklov Institute

Arnold is the mathematician who taught physicists to see the contents of this page. At nineteen he solved Hilbert's thirteenth problem; at twenty-six came the "A" of KAM theory — the proof, sharpening Kolmogorov's 1954 announcement, that most invariant tori of an integrable system survive small perturbation, the result that explains why the solar system is neither exactly integrable nor visibly chaotic on human timescales. His Mathematical Methods of Classical Mechanics (1974) rebuilt the subject with the symplectic manifold $(\mathcal{M},\omega)$ as the primitive object: Hamiltonian flows as symplectomorphism paths, integrability as Lagrangian tori (the Arnold–Liouville theorem of this page), and mechanics as geometry throughout — the book from which a generation absorbed the viewpoint. He mapped the failure modes too: Arnold diffusion, the slow leak between tori possible only for $n\geq3$ . A ferocious polemicist against formalism ("mathematics is the part of physics where experiments are cheap"), he ran a Moscow seminar whose problem lists seeded fields for decades, and his 1965 fixed-point conjecture — that symplectic maps must have more fixed points than topology alone demands — drove the invention of Floer homology and modern symplectic topology in the 1980s. The rigidity theme of this article is, to a large extent, his legacy being confirmed.

History

1809

— Poisson and Lagrange develop the bracket calculus for planetary perturbations: the symplectic structure in embryo

1882

— Darboux proves that all symplectic forms are locally standard: no local invariants, no symplectic curvature

1890

— Poincaré identifies the integral invariants of Hamiltonian flow, the conserved powers of the two-form

1939

— Weyl coins the word symplectic and works out the group's invariant theory in The Classical Groups

1963

— Arnold completes KAM: most invariant tori survive perturbation — stability becomes a symplectic-geometry theorem

1974

— Mathematical Methods of Classical Mechanics recasts dynamics as geometry of the symplectic manifold

1985

— Gromov's non-squeezing theorem launches symplectic topology: rigidity beyond volume, capacities, the camel

Going further

References. Goldstein, Poole & Safko, Classical Mechanics, 3rd ed., §9.3–9.4 (symplectic approach to canonical transformations) · Landau & Lifshitz, Mechanics, 3rd ed., §45–46 (canonical transformations, Poisson brackets) · Arnold, Mathematical Methods of Classical Mechanics, 2nd ed., Ch. 8 (symplectic manifolds, Darboux's theorem) · Morin, Introduction to Classical Mechanics, Ch. 15 (Hamiltonian formalism) · Taylor, Classical Mechanics, Ch. 13

Provenance. Equations and group-theoretic facts cross-checked against Goldstein, Poole & Safko §9.3–9.4 (symplectic condition, $\det M=1$ , fundamental-bracket equivalence), Arnold Ch. 8 (symplectic manifolds, Darboux, Lagrangian tori), and Landau–Lifshitz §45–46 (Poisson brackets, canonicity). Verifier should: (i) recompute $M^{\mathsf T}JM$ for the $2\times2$ case to confirm $\det M=1$ ; (ii) verify $\dim\mathfrak{sp}(2n)=n(2n+1)$ via symmetric-matrix counting ( $\dim\mathrm{Sp}(2)=3$ , $\dim\mathrm{Sp}(4)=10$ ); (iii) confirm the flow derivation using $J^{\mathsf T}J=\mathbb{1}$ , $J^2=-\mathbb{1}$ . No empirical constants appear; all numbers are exact group-theoretic quantities. The numeric examples (shear/squeeze pair; block-diagonal monodromy with $e^{\pm0.5}=1.64872/0.60653$ ) were computed for this article and are hand-checkable.

Builds on: Definition of Canonical Transformations · Symplectic Condition · Fundamental Poisson Brackets · Poisson-Bracket Formulation of Dynamics Leads to: Poisson Brackets as a Test for Canonicity · Constants of Motion via Poisson Brackets · Algebra of Poisson Brackets

Sp(2) equals SL(2), so why isn't Sp(2n) just SL(2n) for larger n?

Count constraints. Unit determinant is one scalar condition; the symplectic condition is $n(2n+1)$ of them. For $n=1$ they coincide because a $2\times2$ sandwich $M^{\mathsf T}JM$ automatically equals $(\det M)J$ — antisymmetric $2\times2$ matrices form a one-dimensional space, so preserving $J$ can only mean scaling it by the determinant. For $n\geq2$ the space of antisymmetric matrices is bigger than one dimension, the sandwich can deform $J$ in many ways while keeping volume, and the symplectic family ( $\dim\ n(2n+1)$ ) sits strictly inside the volume-preserving one ( $\dim\ 4n^2-1$ ): $10<15$ already at $n=2$ .

Why must phase space be even-dimensional? Is that physics or bookkeeping?

It is forced by nondegeneracy of the antisymmetric form. For any antisymmetric matrix, $\det J=\det J^{\mathsf T}=\det(-J)=(-1)^{d}\det J$ ; in odd dimension $d$ this gives $\det J=-\det J=0$ , so an invertible antisymmetric form cannot exist. Physically: every coordinate must have a conjugate partner for the area meter to see it — an unpaired direction would be dynamically invisible (its "Hamilton equation" would not exist). Odd-dimensional objects do appear in mechanics — constant-energy surfaces — but they carry the cousin structure, contact geometry, not a symplectic form.

What is the actual difference between the matrix J and the two-form omega?

$\omega$ is the geometric object — a machine eating two tangent vectors and returning a signed area, defined on a possibly curved manifold with no coordinates chosen. $J$ is its matrix in a canonical chart: $\omega(u,v)=u^{\mathsf T}Jv$ once you pick Darboux coordinates. The distinction matters exactly when charts change: $J$ is only guaranteed to look like the standard block form in canonical coordinates (that is what Darboux's theorem provides), whereas $\omega$ exists chart-free. It parallels metric geometry: $\omega$ is to $J$ what the metric tensor is to "the identity matrix in orthonormal coordinates."

Darboux says no local invariants — Riemannian geometry has curvature. Why the difference?

The deep asymmetry between symmetric and antisymmetric forms. A metric's second-order behavior around a point cannot be flattened away — curvature survives every coordinate choice, and it is a local measurement (parallel-transport around a small loop). An antisymmetric closed form has just enough gauge freedom in coordinate changes to absorb all local structure: closedness ( $\mathrm{d}\omega=0$ ) is exactly the integrability condition that lets you build canonical coordinates outward from a point. Consequence: there is no symplectic analogue of "measuring curvature in a lab." Everything interesting — capacities, tori, non-squeezing — requires finite-size, global questions.

What does non-compactness of the group mean for actual physics?

That canonical evolution can be violent without breaking any rule. Compact groups (rotations) bound every orbit: preserved lengths cage the motion. The symplectic group's unbounded directions — the squeezes — mean a Hamiltonian flow may stretch uncertainty in one variable exponentially while compressing its conjugate, forever. Chaos lives in exactly these directions (Lyapunov exponents are the flow picking a squeeze axis), as does squeezed light (a pumped crystal steering quadratures along a squeeze). The invariants survive — areas, volume, brackets — while everything metric about the state degrades. Non-compactness is why long-term prediction is hard inside a perfectly structure-preserving theory.

Where does the unitary group hide inside the symplectic group?

At the compact corner: $\mathrm{Sp}(2n,\mathbb{R})\cap\mathrm{O}(2n)\cong\mathrm{U}(n)$ — maps preserving the area form and the Euclidean metric preserve their complex combination, which is a Hermitian form. (Identify $z_k=q_k+ip_k$ ; rotations mixing each $q_k,p_k$ pair and the pairs among themselves unitarily are exactly these maps.) This "two structures give the third" fact is the classical shadow of quantization: quantum mechanics equips phase space with precisely such a compatible complex structure, and the metaplectic representation realizes the symplectic group on Hilbert space — with the famous double-valuedness that gives the harmonic oscillator its $\tfrac12\hbar\omega$ .

What physically are the Poincaré integral invariants beyond volume?

Take any closed 2-surface of states in phase space and add up the oriented areas of its projections onto the $n$ conjugate planes: that sum, $\int\omega$ , is frozen under Hamiltonian flow — that is the $k=1$ invariant, and it is the reason adiabatic invariants like $\oint p\,\mathrm{d}q$ exist. The higher invariants $\int\omega^{k}$ do the same for $2k$ -dimensional surfaces, interpolating up to $k=n$ , which is Liouville volume. They are the complete list of "what canonical maps cannot change" at the level of integrals — the differential-form statement of everything this chapter has been checking matrix by matrix.

What is a Lagrangian submanifold, in one honest paragraph?

An $n$ -dimensional surface in $2n$ -dimensional phase space on which the symplectic form restricts to zero — the biggest possible "area-blind" subspace. Two reasons to care. Dynamically: the invariant tori of integrable systems are Lagrangian — position on the torus costs no symplectic area, which is why action-angle coordinates exist and why the actions $\oint p\,\mathrm{d}q$ label the tori (Arnold–Liouville). Structurally: the graph of a canonical transformation is a Lagrangian submanifold of the doubled phase space, and generating functions are exactly the scalar functions that parametrize such graphs — the four types of this chapter are four coordinate charts on one Lagrangian object. Much of modern symplectic topology (Floer theory, mirror symmetry) is the study of how Lagrangian submanifolds can sit inside a manifold.