Michal–Bastiani calculus on locally convex spaces

7 \(C^{k}\) maps on sets with dense interior

Cubes \([0, 1]^n\), half-spaces and their open subsets are not open, but they are the domains of the charts of manifolds with boundary or with corners, and of the curves on compact intervals. Glöckner and Neeb extend the calculus to locally convex subsets with dense interior [ GN26 , § 1.4 ] : a map is \(C^{1}\) on such a set \(U\) if it is \(C^{1}\) on the interior \(U^0\) and its differential extends continuously to \(U \times E\). In this chapter we prove that the first-order calculus of Chapter 2, the higher differentials of Chapter 3, and the \(C^{k,\ell }\) maps and the exponential law of Chapter 6 all survive on such sets. Manifolds with rough boundary, whose charts take values in these sets [ GN26 , Definition 3.5.1 ] , are built on this chapter in lc-manifolds.

As everywhere else, no completeness is assumed: integrals are weak integrals. In a few places the proofs depart from those of [ GN26 ] ; the remarks say where.

7.1 Locally convex subsets and \(C^{k}\) maps

Definition 7.1 Locally convex subsets with dense interior [ GN26 , Definitions 1.4.1 and 1.4.4 ]
✓

A subset \(U\) of a locally convex space \(E\) is locally convex if every point of \(U\) has a convex neighbourhood in \(U\), and it has dense interior if \(U \subseteq \overline{U^0}\). A map \(f \colon U \to F\) on such a set is \(C^{0}\) if it is continuous. It is \(C^{k+1}\) if it is continuous, \(f|_{U^0}\) has all directional derivatives, and the differential \(\mathrm{d}(f|_{U^0}) \colon U^0 \times E \to F\) extends to a map \(\mathrm{d}f \colon U \times E \to F\) which is \(C^{k}\) on \(U \times E\).

The definition is recursive, as Definition 2.3; Glöckner and Neeb state it through the iterated differentials, and the two forms agree by Theorem 7.6 below.

Let \(U \subseteq E\) be a locally convex subset with dense interior.

  1. On open sets, \(C^{k}\) in the sense of Definition 7.1 is \(C^{k}\) in the sense of Definition 2.3.

  2. The continuous extension of \(\mathrm{d}(f|_{U^0})\) is unique.

  3. \(f\) is \(C^{k+1}\) as soon as \(\mathrm{d}(f|_{U^0})\) has some \(C^{k}\) extension to \(U \times E\), and then \(\mathrm{d}f\) is that extension.

  4. If \(f\) is \(C^{1}\), then \(\mathrm{d}f(x, \cdot )\) is continuous and linear for every \(x \in U\) [ GN26 , Lemma 1.4.7 ] .

Proof ▶

(b) holds because \(U^0\) is dense in \(U\); (c) is a reformulation of the definition, and all the rules below are proved from it. (d): \(\mathrm{d}f(x, \cdot )\) is a pointwise limit of the linear maps \(\mathrm{d}f(y, \cdot )\), \(y \in U^0\), so it is linear, and it is continuous by definition.

7.2 First-order calculus

Theorem 7.3 Mean value theorem and difference quotients [ GN26 , Proposition 1.4.8, Lemma 1.4.9 ]
✓

Let \(f \colon U \to F\) be \(C^{1}\).

  1. If the segment \([x, x + h]\) lies in \(U\), then \(\lambda (f(x + h) - f(x)) = \int _0^1 \lambda (\mathrm{d}f(x + th, h))\, dt\) for every continuous linear functional \(\lambda \) on \(F\).

  2. The difference quotients \(t^{-1}(f(x + th) - f(x))\) extend continuously to \(t = 0\), with value \(\mathrm{d}f(x, h)\).

Proof ▶

For a convex \(U\) we shrink the segment towards an interior point, where Proposition 2.10 applies, and pass to the limit in a parameter-dependent integral; this gives (a) for convex sets, and (b). For a general locally convex \(U\), the function \(t \mapsto \lambda (f(x + th))\) has right derivative \(\lambda (\mathrm{d}f(x + th, h))\) on \([0, 1)\) by (b), and (a) follows from the fundamental theorem of calculus for real functions.

Remark 7.4
#

Glöckner and Neeb prove the mean value theorem first on convex sets, with an integral in a completion of \(F\), and then split a segment into finitely many pieces lying in convex neighbourhoods. The argument above avoids the completion. We do not claim it is new.

On locally convex subsets with dense interior:

  1. Chain rule. If \(f \colon U \to V\) and \(g \colon V \to G\) are \(C^{k}\), so is \(g \circ f\), with \(\mathrm{d}(g \circ f)(x, h) = \mathrm{d}g(f(x), \mathrm{d}f(x, h))\).

  2. Pairs, products of maps, compositions with continuous linear maps, sums and continuous bilinear images of \(C^{k}\) maps are \(C^{k}\); restrictions of \(C^{k}\) maps on open sets are \(C^{k}\).

  3. \(C^{k}\) is inherited by locally convex subsets with dense interior, and it is a local property.

  4. Partial differentials. A map on \(U_1 \times U_2\) is \(C^{1}\) if and only if its partial differentials exist and extend continuously, and then \(\mathrm{d}f((x, y), (h, k)) = \mathrm{d}_1 f((x, y), h) + \mathrm{d}_2 f((x, y), k)\).

Proof ▶

(a) The candidate differential \(\mathrm{d}g(f(x), \mathrm{d}f(x, h))\) is continuous, and on \(U^0\) it is the derivative of \(g \circ f\) by Theorem 7.3(b) applied to \(g\) along \(f(x + th)\); conclude with Proposition 7.2(c), by induction on \(k\). The chain rule needs only \(V\) to be locally convex with dense interior, and only density of \(U^0\) in \(U\). (b)–(d) follow from (a) and Proposition 7.2(c) as in the open case.

Items (b) and (c) are what the transition maps of manifolds with rough boundary need [ GN26 , Remark 3.5.2 ] .

7.3 Higher differentials, curves and subspaces

A map \(f \colon U \to F\) is \(C^{k}\) if and only if it is continuous, \(f|_{U^0}\) is \(C^{k}\), and the iterated differentials \(\mathrm{d}^{(j)}(f|_{U^0})\) extend continuously to \(U \times E^j\) for \(j \le k\). The extensions are unique.

Proof ▶

Necessity by induction on \(k\), from the definition. For sufficiency, [ GN26 ] uses the version of [ GN26 , Lemma 1.3.12 ] for sets with dense interior [ GN26 , Remark 1.4.12 ] . We argue by induction on \(k\) instead, with the explicit formula of Theorem 3.5 for the differentials of \(\mathrm{d}(f|_{U^0})\) on the open set \(U^0 \times E\), whose right-hand side extends continuously. We do not claim this argument is new.

  1. On a non-degenerate interval \(I \subseteq \mathbb {R}\), the \(C^{k}\) maps in the sense of Definition 7.1 are the \(C^{k}\) curves, with \(\mathrm{d}\gamma (s, r) = r\gamma '(s)\).

  2. Let \(F_0 \subseteq F\) be a sequentially closed vector subspace. A map \(U \to F_0\) is \(C^{k}\) into \(F\) if and only if it is \(C^{k}\) into \(F_0\).

Proof ▶

(a) At interior points both notions are the derivative; at the end points both ask for a continuous extension. (b) Every point of \(U\) is the limit of a sequence in \(U^0\), so the continuous extensions of the differentials take values in \(F_0\) when those on \(U^0\) do.

7.4 \(C^{k,\ell }\) maps and the exponential law

Let \(U \subseteq E_1\) and \(V \subseteq E_2\) be locally convex subsets with dense interior. A map \(f \colon U \times V \to F\) is \(C^{k,\ell }\) if it is continuous, it is \(C^{k,\ell }\) on \(U^0 \times V^0\) (Definition 6.15), and the mixed differentials \(\mathrm{d}^{(i,j)}(f|_{U^0 \times V^0})\), \(i \le k\), \(j \le \ell \), extend to continuous maps on \(U \times V \times E_1^i \times E_2^j\). For \(k, \ell \in \mathbb {N}_0 \cup \{ \infty \} \), \(f\) is \(C^{k,\ell }\) if it is \(C^{i,j}\) for all finite \(i \le k\), \(j \le \ell \). On open sets this is Definition 6.15.

Let \(f \colon U \times V \to F\) be \(C^{k,\ell }\).

  1. For every \(x \in U\), also on the boundary, \(f(x, \cdot )\) is \(C^{\ell }\) with \(\mathrm{d}^{(j)}(f(x, \cdot )) = \mathrm{d}^{(0,j)} f(x, \cdot )\); for \(y \in V\) and \(w \in E_2^j\), \(x \mapsto \mathrm{d}^{(0,j)} f(x, y, w)\) is \(C^{k}\).

  2. Recursion. \(f\) is \(C^{k+1,\ell }\) if and only if it is \(C^{0,\ell }\) and \(((x, a), y) \mapsto \mathrm{d}^{(1,0)} f(x, y, a)\) is \(C^{k,\ell }\) on \((U \times E_1) \times V\); and similarly in the second variable.

  3. A map into a product is \(C^{k,\ell }\) if and only if its components are.

  4. Chain rules. \((x, y) \mapsto g(f_1(x), f_2(y))\) is \(C^{r,s}\) if \(g\) is \(C^{r,s}\), \(f_1\) is \(C^{r}\) and \(f_2\) is \(C^{s}\); and \(g \circ f\) is \(C^{r,s}\) if \(f\) is \(C^{r,s}\) and \(g\) is \(C^{r+s}\).

  5. \(C^{k+\ell }\) maps on \(U \times V\) are \(C^{k,\ell }\).

Proof ▶

(a) Pass to the limit from \(U^0\) with a parametrized form of the mean value inclusion: a family of maps with continuous directional derivatives keeps them at accumulation points of the parameters (the role of [ GN26 , Lemma 1.4.18 ] ); conclude with Theorem 7.6. (b) From explicit formulas for the mixed differentials of the partial differentials, and the interchange of a derivative in the first variable with derivatives in the second. (c)–(e) follow from (b) by induction, with the chain rule of Theorem 7.5.

Remark 7.10
#

Glöckner and Neeb prove the chain rules by computing the mixed differentials of the composite with Faà di Bruno’s formula [ GN26 , (1.91)–(1.97) ] . The recursive characterization (b) avoids those formulas. We do not claim this argument is new.

Give \(C^{k}(V, F)\) and \(C^{k,\ell }(U \times V, F)\) the compact-open \(C^{k}\) and \(C^{k,\ell }\) topologies, as in Chapter 6. For \(f \in C^{k,\ell }(U \times V, F)\), the map \(f^\vee \colon U \to C^{\ell }(V, F)\), \(f^\vee (x) = f(x, \cdot )\), is \(C^{k}\), with

\[ \mathrm{d}^{(j)}\bigl(\mathrm{d}^{(i)}(f^\vee )(x, v)\bigr)(y, w) = \mathrm{d}^{(i,j)} f(x, y, v, w) \qquad \text{\cite[(1.98)]{GN26}}, \]

and \(\Phi \colon f \mapsto f^\vee \) is a linear topological embedding \(C^{k,\ell }(U \times V, F) \to C^{k}(U, C^{\ell }(V, F))\). If \(V\) is locally compact, \(\Phi \) is an isomorphism of topological vector spaces. The same holds for \(k = \ell = \infty \).

Proof ▶

At interior points the derivative of \(f^\vee \) is \((\mathrm{d}^{(1,0)} f)^\vee \): the difference quotients converge in \(C^{\ell }(V, F)\), uniformly on compact subsets of \(V \times E_2^j\), by the mean value inclusion with a parameter ranging over all of \(V\). The induction then runs through the recursion of Theorem 7.9(b). For \(k = \ell = \infty \), \(C^{\infty }(V, F)\) is the closed subspace of compatible families in \(\prod _j C^{j}(V, F)\), and the law follows from the finite orders.

Remark 7.12
#

Glöckner and Neeb differentiate \(f^\vee \) through the continuous extension of difference quotients of \(C^{1,0}\) maps [ GN26 , Lemma 1.7.6 ] and an induction on the order of the differentials [ GN26 , (1.99) ] . Their alternative hypothesis (b) on \(k_{\mathbb {R}}\)-spaces is also formalized (curryEquivOfKRSpace); first countable spaces are \(k_{\mathbb {R}}\)-spaces, so the law holds for Fréchet spaces.