Michal–Bastiani calculus on locally convex spaces

11 The Frobenius theorem

The local Frobenius theorem on Banach spaces, in the form of Penot [ Pen70 ] : a total differential equation \(D_1\alpha (x, y) = f(x, \alpha (x, y))\) is integrable when it satisfies the symmetry condition (Frob), and the solution is as regular as \(f\), including the real analytic case. In chart form this is the integrability of involutive split distributions [ Lan99 , Chapter VI ] . Outside Banach spaces the theorem fails, already for distributions of rank one on a Fréchet space (Chapter 16).

11.1 The local Frobenius theorem on Banach spaces

Let \(\mathbb {K} \in \{ \mathbb {R}, \mathbb {C}\} \), \(E\), \(F\) Banach spaces over \(\mathbb {K}\), \(O \subseteq E \times F\) open and \(f \colon O \to L(E, F)\) of class \(C^{n}\) over \(\mathbb {K}\), \(1 \le n \le \infty \). We follow Penot, who derives the Frobenius theorem from the implicit function theorem [ Pen70 , Théorème 1.1 ] .

The map \(P(x, y, \delta ) = \bigl(x, y, \delta - (s \mapsto f(sx, y + \int _0^s \delta )\, x)\bigr)\) on \(E \times F \times C([-1, 1], F)\) is \(C^{n}\) where the curve \(s \mapsto (sx, y + \int _0^s \delta )\) has values in \(O\). If \((0, y_0) \in O\), its derivative at \((0, y_0, 0)\) is the invertible shear \((a, b, c) \mapsto (a, b, c - f(0, y_0)\, a)\).

Proof ▶

\(P\) is built from continuous linear maps, the superposition operator of \(f\) (Lemma 10.3) and the continuous bilinear pointwise application \(C([-1,1], L(E, F)) \times C([-1,1], E) \to C([-1,1], F)\). At \(x = 0\) the term \(f(\dots )\, x\) vanishes, and so does the part of its derivative that differentiates \(f\).

Theorem 11.2 Local Frobenius theorem [ Pen70 , Théorème 1.1 and Remarque 1 ]
✓

Assume (Frob): for \(q \in O\) and \(a, b \in E\),

\[ D_1 f(q)(a)(b) + D_2 f(q)(f(q)\, a)(b) = D_1 f(q)(b)(a) + D_2 f(q)(f(q)\, b)(a). \]

For every \((x_0, y_0) \in O\) there are \(r {\gt} 0\) with \(B(x_0, r) \times B(y_0, r) \subseteq O\) and a \(C^{n}\) map \(\alpha \colon B(x_0, r) \times B(y_0, r) \to F\) with \(\alpha (x_0, y) = y\) and \(D_1 \alpha (x, y) = f(x, \alpha (x, y))\). If \(\mathbb {K} = \mathbb {C}\) and \(f\) is holomorphic, \(\alpha \) is holomorphic.

Proof ▶

Translate \(x_0\) to \(0\). By Lemma 11.1 and the inverse function theorem there is a \(C^{n}\) family \(\varphi (x, y) \in C([-1, 1], F)\) with \(P(x, y, \varphi (x, y)) = (x, y, 0)\) and \(\varphi (0, y) = 0\); put \(\alpha (x, y) = y + \int _0^1 \varphi (x, y)\). Let \(c(s) = (sx, y + \int _0^s \varphi (x, y))\) and \(z \in E\). Differentiating the fixed point identity, \(K(s) = \partial _x\bigl(\int _0^s \varphi \bigr)(x, y)\, z\) solves the linear equation \(u' = Df(c(s))(sz, u)\, x + f(c(s))\, z\) with \(u(0) = 0\). By (Frob), so does \(L(s) = s\, f(c(s))\, z\), and uniqueness for linear equations gives \(K(1) = L(1)\), which is \(D_1 \alpha (x, y)\, z = f(x, \alpha (x, y))\, z\). Penot instead shows that \(D_3 P\) maps \(K'\) and \(L'\) to the same point and that it is invertible near \(0\); and he works with curves of class \(C^{1}\) and the implicit function theorem, where we take \(\delta = \gamma '\) as unknown. We do not claim this argument is new. For \(\mathbb {K} = \mathbb {C}\), holomorphic maps are \(C^{1}\) over \(\mathbb {C}\) and \(C^{1}\) maps over \(\mathbb {C}\) are holomorphic (Theorem 9.9).

Let \(E\), \(F\) be real Banach spaces and \(f \colon O \to L(E, F)\) real analytic in the sense of Milnor, satisfying (Frob). Then the map \(\alpha \) of Theorem 11.2 is real analytic on some \(B(x_0, r) \times B(y_0, r)\).

Proof ▶

Penot uses a real analytic implicit function theorem; we argue with flows instead. The vector field \(X(s, v, c) = (1, 0, f(x_0 + sv, c)\, v)\) on \(\mathbb {R} \times E \times F\) is real analytic. Since \(\frac{d}{dt}\alpha (x_0 + tv, y) = f(x_0 + tv, \alpha (x_0 + tv, y))\, v\), the curve \(t \mapsto (t, v, \alpha (x_0 + tv, y))\) is an integral curve of \(X\) through \((0, v, y)\), so by uniqueness it is \(t \mapsto \mathrm{Fl}(t, (0, v, y))\), where \(\mathrm{Fl}\) is the local flow of \(X\) on \((-\varepsilon , \varepsilon ) \times V\). Hence \(\alpha (x, y)\) is the last component of \(\mathrm{Fl}(\varepsilon /2, (0, \tfrac {2}{\varepsilon }(x - x_0), y))\), which is real analytic by Theorem 10.6. We do not claim this argument is new.

Proposition 11.4 Involutivity in local form [ Lan99 , Chapter VI, § 1, p. 157 ]
✓

Let \(E\), \(F\) be normed spaces and \(f \colon E \times F \to L(E, F)\) differentiable at \(q\). If the vector fields \(X\), \(Y\) are differentiable at \(q\) and \(X(p), Y(p) \in \operatorname {graph} f(p)\) for \(p\) near \(q\), then \([X, Y](q) \in \operatorname {graph} f(q)\) if and only if \(Df(q)(X(q))(Y_1(q)) = Df(q)(Y(q))(X_1(q))\). Hence (Frob) at \(q\) holds if and only if the distribution \(p \mapsto \operatorname {graph} f(p)\) is involutive at \(q\).

Proof ▶

Differentiating \(X_2 = f\, X_1\) gives \(DX_2(q)\, v = Df(q)(v)\, X_1(q) + f(q)\, DX_1(q)\, v\), and similarly for \(Y\). In \([X, Y](q)_2 - f(q)\, [X, Y](q)_1\) the terms with \(DX_1\) and \(DY_1\) cancel, leaving \(Df(q)(X(q))\, Y_1(q) - Df(q)(Y(q))\, X_1(q)\). For the converse take the fields \(p \mapsto (a, f(p)\, a)\) and \(p \mapsto (b, f(p)\, b)\).

Theorem 11.5 Frobenius theorem in chart form [ Lan99 , Chapter VI, Theorems 1.1 and 1.2 ]
✓

Under the hypotheses of Theorem 11.2, every \((x_0, y_0) \in O\) is fixed by a \(C^{n}\) diffeomorphism \(\Phi (x, y) = (x, \alpha (x, y))\) between open subsets of \(O\) with \(D\Phi (p)(a, 0) = (a, f(\Phi (p))\, a)\). Thus \(D\Phi ^{-1}(z)\) maps \(\operatorname {graph} f(z)\) onto \(E \times 0\): the chart \(\Phi ^{-1}\) rectifies the distribution. If \(\mathbb {K} = \mathbb {C}\) and \(f\) is holomorphic, \(\Phi \) and \(\Phi ^{-1}\) are holomorphic.

Proof ▶

Take \(\alpha \) from Theorem 11.2. Since \(D_1\alpha = f(x, \alpha )\) and \(\alpha (x_0, \cdot ) = \mathrm{id}\), the derivative of \(\Phi \) at \((x_0, y_0)\) is the shear \((a, b) \mapsto (a, f(x_0, y_0)\, a + b)\), which is invertible. Apply the inverse function theorem and restrict to the open set where \(D\Phi \) is invertible.