Manifolds modeled on locally convex spaces

4 Vector fields and the Frobenius theorem

Vector fields, their Lie bracket and their action on functions as derivations, and the Frobenius theorem on Banach manifolds. The bracket is defined in charts and needs no completeness; the Frobenius theorem needs Banach models, and it fails already for distributions of rank one on a Fréchet space (see the counterexamples of [ LCA ] ). The local theorem is the one of lc-analysis, after Penot [ Pen70 ] .

4.1 Vector fields and derivations

Definition 4.1 Vector fields
✓

A vector field on \(M\) is a section \(X \colon M \to TM\) of the tangent bundle. Its local representative in a chart \(i\) is \(\widetilde{X}_i \colon i(\operatorname {dom} i) \to E\), \(\widetilde{X}_i(y) = \) the coordinates of \(X(i^{-1}y)\) in the chart \(i\).

On chart overlaps, \(\widetilde{X}_j \circ \tau = \mathrm{d}\tau (\cdot )\, \widetilde{X}_i\) for the transition map \(\tau = j \circ i^{-1}\). A vector field is \(C^{k}\) as a map \(M \to TM\) if and only if its local representatives are \(C^{k}\) (one direction in every chart of the atlas, the other from the preferred charts).

Proof ▶

The transformation law is the cocycle property of the coordinate changes. The local representative \(\widetilde{X}_i\) is the second component of the local representative of \(X\) in the charts \(i\) and \(Ti\), which gives both implications.

For maps \(X, Y\) defined on an open subset of \(E\), the local bracket is

\[ [X, Y](y) = \mathrm{d}Y(y)\, X(y) - \mathrm{d}X(y)\, Y(y). \]

Let \(\tau \) be \(C^{2}\), and let \(X_i, Y_i, X_j, Y_j\) be \(C^{1}\) with \(X_j \circ \tau = \mathrm{d}\tau \cdot X_i\) and \(Y_j \circ \tau = \mathrm{d}\tau \cdot Y_i\) near \(y\). Then \([X_j, Y_j](\tau y) = \mathrm{d}\tau (y)\, [X_i, Y_i](y)\). The local bracket satisfies the Jacobi identity for \(C^{2}\) fields, and it is \(C^{k}\) for \(C^{k+1}\) fields.

Proof ▶

Differentiating \(y \mapsto \mathrm{d}\tau (y)\, X_i(y)\) in the direction \(Y_i(y)\) gives \(\mathrm{d}^2\tau (y)(Y_i, X_i) + \mathrm{d}\tau (y)\, \mathrm{d}X_i(y)\, Y_i\). In the difference defining the bracket the second derivatives \(\mathrm{d}^2\tau (y)(Y_i, X_i)\) and \(\mathrm{d}^2\tau (y)(X_i, Y_i)\) cancel by the symmetry of the second differential. The Jacobi identity is a direct computation with the derivative of the local bracket, where again the second derivatives cancel by symmetry.

Definition 4.4 The Lie bracket
✓

The Lie bracket of the vector fields \(X\) and \(Y\) is the vector field whose value at \(x\) is the local bracket \([\widetilde{X}, \widetilde{Y}](c_x(x))\) of the local representatives in the preferred chart at \(x\).

Let \(X\), \(Y\) be \(C^{1}\) vector fields. In every chart \(j\) of the atlas, the local representative of \([X, Y]\) is the local bracket \([\widetilde{X}_j, \widetilde{Y}_j]\). The bracket is antisymmetric, it satisfies the Jacobi identity for \(C^{2}\) fields, and if \(X\) and \(Y\) are \(C^{k+1}\), then \([X, Y]\) is \(C^{k}\).

Proof ▶

The local representatives in the preferred chart at \(x\) and in \(j\) are related by Proposition 4.2, so Lemma 4.3 applies. The other statements are then read in a single chart.

Remark 4.6 No flows
#

In finite dimensions, chart independence of the bracket is usually derived from its description as the Lie derivative along the flow of \(X\). In infinite dimensions vector fields need not have flows, so this argument is not available. We prove it directly, as above: the second derivatives of the transition maps cancel by the symmetry of second derivatives.

Proof ▶

In charts, \(\widetilde{X}'_j \circ h = \mathrm{d}h \cdot \widetilde{X}_i\) with \(h = j \circ f \circ i^{-1}\). This is the hypothesis of Lemma 4.3 with \(h\) in place of \(\tau \), whose proof only uses that \(\tau \) is \(C^{2}\).

We now let vector fields act on functions. As explained in Remark 2.9, a function \(\varphi \colon M \to \mathbb {R}\) is said to be \(C^{k}\) when all its local representatives \(\varphi \circ i^{-1}\) are \(C^{k}\) on \(i(\operatorname {dom} i)\).

Definition 4.8 Derivative along a vector field
✓

For a vector field \(X\) and \(\varphi \colon M \to \mathbb {R}\), we let \(X(\varphi )(x) = \mathrm{d}(\varphi \circ c_x^{-1})(c_x(x))\, \widetilde{X}(c_x(x))\), computed in the preferred chart at \(x\). The product \(\varphi Y\) of a function and a vector field is \((\varphi Y)(x) = \varphi (x)\, Y(x)\).

Let \(\varphi , \psi \) be \(C^{1}\) functions. Then \(X(\varphi )(x) = \mathrm{d}(\varphi \circ j^{-1})(j(x))\, \widetilde{X}_j(j(x))\) in every chart \(j\) around \(x\), and the Leibniz rules

\[ X(\varphi \psi ) = X(\varphi )\, \psi + \varphi \, X(\psi ), \qquad [X, \varphi Y] = X(\varphi )\, Y + \varphi \, [X, Y] \]

hold, the second one for \(C^{1}\) vector fields \(X\), \(Y\).

Proof ▶

Chart independence follows from \(\widetilde{X}_j \circ \tau = \mathrm{d}\tau \cdot \widetilde{X}_i\) and the chain rule. The first Leibniz rule is the product rule in a chart. For the second one, \(\widetilde{(\varphi Y)}_i = (\varphi \circ i^{-1})\, \widetilde{Y}_i\), and we differentiate this product in the local bracket.

Theorem 4.10 Vector fields are determined by local functions
✓

Let \(c\) be a chart around \(x\). For \(\lambda \in E'\) we have \(X(\lambda \circ c)(x) = \lambda (\widetilde{X}_c(c(x)))\). Consequently, if \(X(\lambda \circ c)(x) = Y(\lambda \circ c)(x)\) for every continuous linear functional \(\lambda \), then \(X(x) = Y(x)\).

Proof ▶

The first statement is Theorem 4.9 in the chart \(c\), since \(\lambda \) is linear. The second follows because continuous linear functionals separate the points of the Hausdorff locally convex space \(E\) (Hahn–Banach).

Remark 4.11 No bump functions
#

Smooth bump functions need not exist on manifolds modeled on locally convex spaces. Hence a vector field is not determined in general by its action on globally defined functions, and the classical identification of vector fields with derivations of \(C^{\infty }(M)\) is not available. Theorem 4.10 is the substitute: functions defined on a chart domain suffice.

4.2 Distributions and the Frobenius theorem

A distribution \(D\) on a manifold \(M\) modeled on a normed space \(E\) assigns to each \(x \in M\) a subspace \(D_x \subseteq T_x M\). It is smooth and split of type \((E_1, E_2)\) if around each \(x_0\), read in the preferred chart \(i\) at \(x_0\) and through a linear isomorphism \(\ell \colon E \to E_1 \times E_2\), it is the field of graphs of a smooth map \(f \colon \ell (W) \to L(E_1, E_2)\): this is the local form of a split subbundle of \(TM\) [ Lan99 , Ch. VI, § 1, pp. 155–157 ] . It is involutive if \([X, Y](x) \in D_x\) for all vector fields \(X\), \(Y\) which are \(C^{1}\) near \(x\) and tangent to \(D\) near \(x\).

Remark 4.13
#

On a Banach manifold a local vector field need not extend to a global one, since there need not be smooth bump functions. Involutivity is therefore required for fields that are \(C^{1}\) only near the point; asking it only for global fields would be a weaker condition.

If \(D\) is involutive and, near \(x_0\), the field of graphs of \(f\) as above, then \(f\) satisfies Penot’s condition (Frob) on \(\ell (W)\).

Proof ▶

The fields reading \(\ell ^{-1}(a, f(\ell y)\, a)\) in the chart \(i\), extended by \(0\) outside its domain, are tangent to \(D\) and \(C^{1}\) near every point of the domain. Their bracket at \(x\) is computed in the preferred chart at \(x\) and transported to \(i\) by the transformation rule of the local bracket. On normed spaces the local bracket is the Lie bracket, which commutes with \(\ell \), and the bracket of the graph fields lies in the graph exactly under (Frob) (lc-analysis, lieBracket_mem_graph_iff).

Let \(E_1\), \(E_2\) be Banach spaces and \(D\) a smooth, split, involutive distribution of type \((E_1, E_2)\). Then around every \(x_0 \in M\) there is a chart \(\varphi \) of the maximal atlas, with \(\varphi (x_0) = i(x_0)\), in which \(D\) is the constant subspace \(\ell ^{-1}(E_1 \times 0)\).

Proof ▶

By Lemma 4.14 and the Frobenius theorem in chart form (lc-analysis, exists_frobenius_chart) there is a diffeomorphism \(\Phi (x, y) = (x, \alpha (x, y))\) fixing \(\ell (i(x_0))\) with \(D\Phi (q)(a, 0) = (a, f(\Phi q)\, a)\). Put \(\psi = \ell ^{-1} \circ \Phi \circ \ell \) and \(\varphi = \psi ^{-1} \circ i\). Since \(D\Phi (q)\) is injective and preserves first components, \(D\Phi (q)\, v \in \operatorname {graph} f(\Phi q)\) if and only if \(v \in E_1 \times 0\).

If around every point there is a chart of the maximal atlas in which \(D\) is a constant subspace \(\ell ^{-1}(E_1 \times 0)\), then \(D\) is involutive. Hence a smooth split distribution with \(E_1\), \(E_2\) Banach is involutive if and only if it admits such flat charts.

Proof ▶

Read in a flat chart, two vector fields tangent to \(D\) near \(x\) take values in the closed subspace \(\ell ^{-1}(E_1 \times 0)\), hence so do their derivatives and their local bracket. The local bracket transforms like a vector field under the change of charts, so the bracket in the preferred chart at \(x\) lies in \(D_x\). No completeness or splitting is needed for this direction.