Manifolds modeled on locally convex spaces

17 Symplectic manifolds

Symplectic forms on manifolds modeled on locally convex spaces. In infinite dimensions nondegeneracy splits into a weak and a strong notion, and many symplectic forms of interest are only weakly nondegenerate. Hamiltonian vector fields are then unique but need not exist, and the Darboux theorem fails [ Mar72 , Proposition, p. 591 ] . For strong symplectic forms on Banach manifolds Darboux’s theorem holds [ Wei69 , § 1, Theorem ] ; Moser’s proof needs flows of time-dependent vector fields, which exist on Banach manifolds but need not exist on locally convex ones, and it is formalized in Section 17.7 for Banach models. Real-valued forms require models in the lowest universe, since the space of values of a form lives in the universe of the model.

17.1 Symplectic forms

A symplectic form on \(M\) is a closed \(2\)-form \(\omega \in \Omega ^2(M, \mathbb {R})\) which is weakly nondegenerate: if \(\omega _x(v, w) = 0\) for all \(w \in T_xM\), then \(v = 0\). It is strong if moreover every continuous linear functional on \(T_xM\) is of the form \(\omega _x(v, \cdot )\), for every \(x \in M\).

In other words, \(\omega \) is weak if the flat map \(\omega ^\flat _x \colon T_xM \to T_xM'\), \(v \mapsto \omega _x(v, \cdot )\), is injective, and strong if it is bijective. We use no topology on the dual space.

Let \(V\) be a normed space and \(\omega \) a continuous skew-symmetric bilinear form on \(V\) such that \(\omega ^\flat \colon V \to V'\) is surjective. Then \(V\) is reflexive: the canonical map \(V \to V''\) is surjective. Consequently, a nonempty manifold modeled on a normed space \(E\) which carries a strong symplectic form has a reflexive model.

Proof ▶

Let \(\Phi \in V''\). The functional \(-\Phi \circ \omega ^\flat \) is continuous, so it equals \(\omega (v, \cdot )\) for some \(v \in V\). For \(\varphi = \omega (w, \cdot ) \in V'\), \(\varphi (v) = \omega (w, v) = -\omega (v, w) = \Phi (\omega ^\flat w) = \Phi (\varphi )\), and every \(\varphi \) is of this form. For the second statement, \(\omega _x\) is continuous on \(E \times E\) because its local representative is, so it is a bounded bilinear form on the normed space \(E\). No completeness is used.

In particular non-reflexive Banach spaces such as \(\ell ^1\) or \(c_0\) do not model strong symplectic manifolds. Marsden observes the converse phenomenon for cotangent bundles: the canonical form of \(T^*M\) is strong if the model of \(M\) is reflexive, and only weak otherwise [ Mar72 , p. 590 ] .

17.2 Hamiltonian and symplectic vector fields

A smooth vector field \(X\) is a Hamiltonian vector field of \(f \in C^\infty (M, \mathbb {R})\) if \(i_X \omega = \mathrm{d}f\), and a symplectic vector field if \(L_X \omega = 0\).

Let \(\omega \) be a symplectic form on \(M\).

  1. A vector field is determined by \(i_X\omega \); in particular Hamiltonian vector fields are unique.

  2. \(X\) is symplectic if and only if \(i_X\omega \) is closed. Hamiltonian fields are symplectic.

  3. If \(M\) is smoothly contractible, every symplectic vector field is Hamiltonian.

  4. If \(X\) and \(Y\) are symplectic, then \(i_{[X, Y]}\omega = \mathrm{d}(\omega (Y, X))\); so \([X, Y]\) is Hamiltonian.

  5. (Conservation of energy) If \(X\) is the Hamiltonian field of \(f\), then \(L_X f = 0\).

Proof ▶

(1) is weak nondegeneracy at each point. (2) Since \(\omega \) is closed, Cartan’s formula gives \(L_X \omega = \mathrm{d}(i_X\omega )\); and \(\mathrm{d}\mathrm{d}f = 0\). (3) is the Poincaré Lemma applied to the closed \(1\)-form \(i_X\omega \). (4) From \([L_X, i_Y] = i_{[X, Y]}\) and \(L_X\omega = 0\) we get \(i_{[X, Y]}\omega = L_X i_Y\omega = \mathrm{d}i_X i_Y \omega + i_X \mathrm{d}i_Y\omega \), and \(\mathrm{d}i_Y \omega = 0\). (5) \(L_X f = i_X \mathrm{d}f = i_X i_X \omega = 0\).

Definition 17.5 Poisson bracket
✓

For \(f\) with Hamiltonian vector field \(X_f\) and any \(g \in C^\infty (M, \mathbb {R})\), the Poisson bracket is \(\{ f, g\} = -L_{X_f} g\).

Let \(f, g, h\) have Hamiltonian vector fields \(X_f, X_g, X_h\). Then \(\{ f, g\} = \omega (X_f, X_g) = L_{X_g} f = -\{ g, f\} \); the bracket \([X_f, X_g]\) is the Hamiltonian vector field of \(-\{ f, g\} \); and the Jacobi identity \(\{ f, \{ g, h\} \} - \{ g, \{ f, h\} \} = \{ \{ f, g\} , h\} \) holds.

Proof ▶

The first identities follow from \(i_{X_g}\omega = \mathrm{d}g\) and the skew-symmetry of \(\omega \), and the statement on \([X_f, X_g]\) is part (4) of Theorem 17.4. For the Jacobi identity, \([L_{X_f}, L_{X_g}] = L_{[X_f, X_g]}\) on functions, and \(L_{[X_f, X_g]} h = i_{[X_f, X_g]} i_{X_h}\omega = -i_{X_h} i_{[X_f, X_g]}\omega = -L_{X_h} L_{X_f} g\).

17.3 Symplectic maps

Definition 17.7 Symplectic maps
✓

A smooth map \(\varphi \colon (N, \omega _0) \to (M, \omega )\) is symplectic if \(\varphi ^*\omega = \omega _0\).

Identities and compositions of symplectic maps are symplectic. If \(\varphi \) is symplectic, \(X\) and \(Y\) are \(\varphi \)-related and \(Y\) is the Hamiltonian field of \(f\), then \(X\) is the Hamiltonian field of \(f \circ \varphi \).

Proof ▶

\(\varphi ^*(i_Y\omega ) = i_X \varphi ^*\omega \) and \(\varphi ^* \mathrm{d}f = \mathrm{d}(f \circ \varphi )\).

17.4 The canonical symplectic form

Let \(H\) be a real inner product space, not necessarily complete. The Liouville form on \(H \times H\) is \(\theta _{(x, a)}(y, b) = \langle x, b \rangle \), and the canonical symplectic form is \(\omega = \mathrm{d}\theta \), \(\omega ((y, b), (y', b')) = \langle y, b' \rangle - \langle y', b \rangle \). This is the skew form induced by the symmetric form \(\langle \cdot , \cdot \rangle \) in [ Mar72 , p. 590 ] .

Proposition 17.10 The canonical symplectic form
✓

The canonical symplectic form is exact and weakly nondegenerate. If \(H\) is a Hilbert space, it is strong.

Proof ▶

If \(\omega ((y, b), \cdot ) = 0\), evaluating at \((0, y)\) and \((b, 0)\) gives \(y = b = 0\). If \(H\) is complete, a functional \(\varphi \) on \(H \times H\) is \(\varphi (y', b') = \langle a_1, y' \rangle + \langle a_2, b' \rangle \) by the Riesz representation theorem, and \(\varphi = \omega ((a_2, -a_1), \cdot )\).

If \(H\) is not complete, then \(H \times H\) is not reflexive (its bidual is the bidual of the completion), so by Theorem 17.2 it carries no strong symplectic form at all; this converse is not formalized.

17.5 The real analytic Darboux theorem

Real analyticity is in the sense of Milnor (Definition 2.2.2 of [ GN26 ] ): restrictions of complex analytic maps on open subsets of the complexification.

Let \(E\) and \(D \cong E'\) be Banach spaces and \(\Theta \colon B(p, r) \to \mathcal{L}(E, D)\) real analytic, skew-symmetric and closed, with \(\Theta (p) = 0\). On a smaller ball, \(\alpha (y) = \int _0^1 s\, \Theta (p + s(y - p))(y - p)\, \mathrm{d}s\) is real analytic, \(\mathrm{d}\alpha = \Theta \), \(\alpha (p) = 0\) and \(\mathrm{d}\alpha (p) = 0\).

Proof ▶

Differentiating under the integral sign, \(\mathrm{d}\alpha (y)(a, b)\) is the integral of \(2s\, \Theta (z_s)(a, b) + s^2\, \mathrm{d}\Theta (z_s)(y - p)(a, b) = \frac{\mathrm{d}}{\mathrm{d}s}\bigl(s^2 \Theta (z_s)(a, b)\bigr)\), \(z_s = p + s(y - p)\), by the closedness of \(\Theta \) and the skew-symmetry of \(\mathrm{d}\Theta \). For real analyticity, \(\alpha (y)\) is the value at time \(1\) of the solution of \(\dot y = 0\), \(\dot a = s\, \Theta (p + s(y - p))(y - p)\) starting at \((y, 0)\); the time-dependent field is real analytic and stationary at \((p, 0)\), and the flow up to time \(1\) near \((p, 0)\) is real analytic (Theorem 5.4 and its real analytic version). This replaces real analytic parametric integrals, which are not available.

In Theorem 17.18, if \(\Omega \colon U \to \mathcal{L}(E, E')\) is real analytic, then \(\psi \) and \(\psi ^{-1}\) can be chosen real analytic.

Proof ▶

Moser’s method with the primitive of Lemma 17.11. The field \(X_t = -\Omega _t^{-1}\alpha \) is real analytic since inversion in the Banach algebra \(\mathcal{L}(E)\) is [ GN26 , Proposition 2.2.13 ] ; the flow of \((1, X_t)\) is real analytic, so \(\psi = F_1\) is; and \(\psi ^{-1}(y) = \mathrm{Fl}(-1, (1, y))\) for the flow of the suspended field, so no real analytic inverse function theorem is needed. We do not claim this argument is new; we know no reference for the real analytic case on Banach manifolds.

Let \(M\) be a real analytic manifold modeled on a Banach space and \(\omega \) a closed real analytic real \(2\)-form [ GN26 , Remark 3.6.18 ] , strongly nondegenerate at \(x_0\). Then the Darboux chart of Theorem 17.19 can be chosen in the real analytic maximal atlas.

Proof ▶

As for Theorem 17.19, with Theorem 17.12; \(\varphi = \psi ^{-1} \circ i\) is compatible with the real analytic atlas since \(\psi \) is a real analytic diffeomorphism.

Remark 17.14 On the hypothesis
#

Real analytic forms are defined by the real analyticity of \((y, u, v) \mapsto \Omega (y)(u, v)\) [ GN26 , Remark 3.6.18 ] , while the local theorem needs \(y \mapsto \Omega (y) \in \mathcal{L}(E, E')\) real analytic for the operator norm. The second follows from the first: this is the real analytic counterpart of the theorem on operator-valued maps used in Theorem 9.5, proved in lc-analysis by extending to the complexification and applying the holomorphic version there.

17.6 The cotangent bundle

In this section \(E\) is a normed space (not necessarily complete) and \(M\) a manifold modeled on \(E\), so that the cotangent bundle \(T^*M = (TM)'\) exists, with fiber \(E' = \mathcal{L}(E, \mathbb {R})\) (Theorem 9.5); its total space is a manifold modeled on \(E \times E'\).

The canonical (Liouville) \(1\)-form \(\theta \) on \(T^*M\) is \(\theta _\lambda (z) = \lambda (T\pi \, z)\), for the projection \(\pi \colon T^*M \to M\). It is a smooth \(1\)-form: in the chart of \(T^*M\) built from a chart \(c\) of \(M\) and the trivialization over a chart \(j\), its local representative is \(((y, \mu ), (u, \psi )) \mapsto \mu (g_{cj}(y) u)\), where \(g_{cj}(y) = \mathrm{d}(j \circ c^{-1})(y)\); for \(j = c\) it is \(\theta _{(y, \mu )}(u, \psi ) = \mu (u)\).

The canonical symplectic form \(\omega = -\mathrm{d}\theta \) on \(T^*M\) is \(\omega _\lambda ((u_1, \psi _1), (u_2, \psi _2)) = \psi _2(u_1) - \psi _1(u_2)\) in the preferred chart. It is exact and weakly nondegenerate, for every normed model \(E\).

Proof ▶

In the preferred chart \(\theta \) is \(((y, \mu ), (u, \psi )) \mapsto \mu (u)\), linear in \((y, \mu )\), and the local formula for \(\mathrm{d}\) gives \(\mathrm{d}\theta ((u_1, \psi _1), (u_2, \psi _2)) = \psi _1(u_2) - \psi _2(u_1)\). If \(\omega ((u_1, \psi _1), \cdot ) = 0\), testing on \((0, \psi )\) gives \(\psi (u_1) = 0\) for all \(\psi \in E'\), so \(u_1 = 0\) by the Hahn–Banach theorem, and testing on \((u, 0)\) gives \(\psi _1 = 0\). We follow Lang’s sign, which agrees with Definition 17.9 under \(\psi \leftrightarrow \langle \cdot , b \rangle \).

Theorem 17.17 Strong nondegeneracy and reflexivity
✓

If \(M\) is nonempty, the canonical symplectic form on \(T^*M\) is strongly nondegenerate if and only if \(E\) is reflexive, i.e. the canonical map \(E \to E''\) is surjective. In particular the cotangent bundle of a manifold modeled on a non-reflexive Banach space (such as \(\ell ^1\) or \(c_0\)) is only a weak symplectic manifold; cotangent bundles of non-reflexive spaces are mentioned in [ Mar72 , p. 590 ] .

Proof ▶

A functional \(\varphi \) on \(E \times E'\) is \(\omega ((u, \psi ), \cdot )\) for \(\psi = -\varphi (\cdot , 0)\) and \(u \in E\) representing \(\varphi (0, \cdot ) \in E''\). Conversely, for \(\Phi \in E''\), write \(\Phi \circ \mathrm{pr}_2 = \omega ((u, \psi ), \cdot )\) and test on \((0, \psi ')\): \(\Phi (\psi ') = \psi '(u)\). No completeness is used. (Theorem 17.2 would give the reflexivity of \(E \times E'\).)

17.7 The Darboux theorem

In this section \(E\) is a Banach space. A closed real \(2\)-form \(\omega \) is strongly nondegenerate at \(x_0\) if \(\omega ^\flat _{x_0} \colon T_{x_0}M \to T_{x_0}M'\) is bijective.

Let \(U \subseteq E\) be open and \(\Omega \colon U \to \mathcal{L}(E, E')\) smooth for the operator norm, with \(\Omega (y)\) skew-symmetric, \(\mathrm{d}\Omega = 0\), and \(\Omega (p)\) bijective. Then there is a diffeomorphism \(\psi \) between open neighborhoods of \(p\) in \(U\) with \(\psi (p) = p\), \(\mathrm{d}\psi (p) = \mathrm{id}\) and \(\psi ^*\Omega = \Omega (p)\): \(\Omega (\psi x)(\mathrm{d}\psi (x) u, \mathrm{d}\psi (x) v) = \Omega (p)(u, v)\).

Proof ▶

Moser’s method [ Mos65 , § 4, Theorem 2 ] . Put \(\Omega _0 = \Omega (p)\), \(\Theta = \Omega - \Omega _0\) and \(\Omega _t = \Omega _0 + t\Theta \).

  1. On a ball \(B\) around \(p\), \(\Theta = \mathrm{d}\alpha \) with \(\alpha (p) = 0\) and \(\mathrm{d}\alpha (p) = 0\): the Poincaré lemma for star-shaped sets (Corollary 8.37) gives a primitive \(\eta \), from which we subtract the closed affine form \(\eta (p) + \mathrm{d}\eta (p)(y - p)\).

  2. The set \(W\) of \((t, y)\) with \(\Omega _t(y)\) invertible is open and contains \(\mathbb {R} \times \{ p\} \). On \(W\) put \(X_t(y) = -\Omega _t(y)^{-1}\alpha (y)\), so that \(\Omega _t(X_t, \cdot ) = -\alpha \), \(X_t(p) = 0\) and \(\partial _x X_t(p) = 0\).

  3. The autonomous field \(Z(t, y) = (1, X_t(y))\) on the open set \(W \subseteq \mathbb {R} \times E\), regarded as a manifold, has the integral curve \(t \mapsto (t, p)\) through \((0, p)\), defined on \(\mathbb {R}\). Since the domain of the maximal flow is open (Theorem 5.4) and \([-1, 2]\) is compact, the flow \(F_t\) of \(X_t\) with \(F_0 = \mathrm{id}\) is defined and smooth on \((-\delta , 1 + \delta ) \times V\) for a neighborhood \(V\) of \(p\); this is Weinstein’s “compactness of \([0, 1]\) and openness of the domain”.

  4. For \(x \in V\) and \(u, v \in E\), \(t \mapsto \Omega _t(F_t x)(\partial _x F_t(x) u, \partial _x F_t(x) v)\) has derivative \(0\), by the variational equation \(\partial _t \partial _x F_t = \partial _x X_t \circ \partial _x F_t\) and an algebraic identity, the pointwise form of \(\frac{\mathrm{d}}{\mathrm{d}t}(F_t^*\Omega _t) = F_t^*(\mathrm{d}\alpha + \mathrm{d}(i_{X_t}\Omega _t)) = 0\).

  5. Hence \(F_1^*\Omega = \Omega _0\), and \(\partial _x F_t(p) = \mathrm{id}\), so by the inverse function theorem \(\psi = F_1\) is a local diffeomorphism at \(p\).

No bump functions are used (they need not exist on a Banach space): the field lives on \(W\) and is never cut off. We normalize the primitive so that \(\mathrm{d}\alpha (p) = 0\), which gives \(\mathrm{d}\psi (p) = \mathrm{id}\). We do not claim this argument is new.

Let \(M\) be modeled on a Banach space and \(\omega \) a closed real \(2\)-form on \(M\), strongly nondegenerate at \(x_0\). There is a chart \(\varphi \) of the maximal atlas around \(x_0\), with the same value and differential at \(x_0\) as the preferred chart, in which \(\omega \) has constant coefficients, equal to its value at \(x_0\). In particular every strong symplectic manifold modeled on a Banach space is locally isomorphic to its model with a constant form.

Proof ▶

In the preferred chart \(i\) at \(x_0\), \(\omega \) is a smooth map \(y \mapsto \Omega (y) \in \mathcal{L}(E, E')\), continuous for the operator norm since the model is normed (lc-analysis). Apply Theorem 17.18 and put \(\varphi = \psi ^{-1} \circ i\).

Remark 17.20 Infinite-dimensional subtleties
#
  • Strong nondegeneracy is used once, to define \(X_t\), and only at \(x_0\): with completeness the open mapping theorem makes \(\Omega _0 \colon E \to E'\) a topological isomorphism, and invertible operators form an open set for the operator norm. For weakly nondegenerate forms the theorem fails, even on Hilbert space [ Mar72 , Proposition, p. 591 ] . Marsden’s counterexample is formalized in lc-analysis (LocallyConvex.Analysis.Counterexamples.Marsden.no_darboux_chart).

  • Reflexivity is a consequence, not a hypothesis (Theorem 17.2).

  • Completeness of \(E\) is used for the flows, the inverse function theorem, the smoothness of \(L \mapsto L^{-1}\) and the open mapping theorem. On non-normable models there is no vector topology on \(\mathcal{L}(E, E')\) for which evaluation is continuous, and neither the flows nor the inverse function theorem are available.

  • No Hausdorff hypothesis on \(M\) is needed: the flows live in \(W \subseteq \mathbb {R} \times E\) and the argument takes place in one chart.

Weinstein also proves a relative version for pairs of manifolds [ Wei71 , Theorem 4.1, Corollary 4.2 ] , which is not formalized.