6 Mapping spaces and the exponential law
Spaces of \(C^{k}\) maps are the models of manifolds of mappings and of the Lie groups of diffeomorphisms. In this chapter we give them the compact-open \(C^{k}\) topology and prove the results that make them usable [ GN26 , § 1.7 ] : superposition operators are continuous and smooth, pullbacks are continuous linear maps, the differential is a continuous linear map, and the exponential law \(C^{k,\ell }(U \times V, F) \cong C^{k}(U, C^{\ell }(V, F))\) holds, with an embedding in general and an isomorphism when \(V\) is locally compact. Chapter 17 explains why no better can be expected.
6.1 Superposition operators
Let \(K\) be a compact Hausdorff space, \(E\), \(F\) locally convex spaces and \(g \colon E \to F\) of class \(C^k\) on an open set \(U \subseteq E\). With the compact-open topologies, the map \(C(K, g) \colon C(K, U) \to C(K, F)\), \(\gamma \mapsto g \circ \gamma \), is of class \(C^k\), with \(\mathrm{d}C(K, g)(\gamma , \eta ) = (x \mapsto \mathrm{d}g(\gamma (x), \eta (x)))\).
Continuity follows from the continuity of the evaluation \(C(K, E) \times K \to E\). For the derivative, the difference quotients of \(g\) extend continuously to \(t = 0\) (Glöckner’s lemma), and the tube lemma over the compact space \(K\) gives convergence in the compact-open topology. The derivative is the superposition operator of \(\mathrm{d}g\) on \(C(K, E \times E) \cong C(K, E) \times C(K, E)\), and the theorem follows by induction on \(k\).
6.2 The compact-open \(C^{k}\) topology
Let \(E\), \(F\) be locally convex spaces, \(U \subseteq E\) open and \(k \in \mathbb {N}\). In Lean, an element of \(C^{k}(U, F)\) is a map \(E \to F\) that is \(C^{k}\) on \(U\) and vanishes outside \(U\); this does not change the mathematics and makes the family of differentials below point-separating.
The compact-open \(C^{k}\) topology on \(C^{k}(U, F)\) is the initial topology with respect to the linear maps \(\mathrm{d}^{(j)} \colon C^{k}(U, F) \to C(U \times E^j, F)\), \(\gamma \mapsto \mathrm{d}^{(j)}\gamma \), for \(j \leq k\), where \(C(U \times E^j, F)\) carries the compact-open topology. It makes \(C^{k}(U, F)\) a locally convex Hausdorff topological vector space.
The map \(C^{k}(U, F) \to \prod _{j \leq k} C(U \times E^j, F)\), \(\gamma \mapsto (\mathrm{d}^{(j)}\gamma )_{j \leq k}\), is a linear topological embedding.
This holds by definition of the topology; injectivity comes from \(\mathrm{d}^{(0)}\gamma = \gamma \).
Let \(U \subseteq E\) be open, \(y \in E\), and let \((g_i)\) be a net of maps \(E \to F\) converging pointwise on \(U\) to \(g\), such that each \(g_i\) has directional derivatives \(D_i(z)\) in the direction \(y\) at all \(z \in U\), with \(D_i\) continuous on \(U\). Suppose that \(D_i \to D\) in the compact-open sense on \(U\) (for every compact \(K \subseteq U\) and open \(O \supseteq D(K)\), eventually \(D_i(K) \subseteq O\)), with \(D\) continuous on \(U\). Then \(g\) has directional derivative \(D(x)\) in the direction \(y\) at every \(x \in U\).
Let \(W\) be a closed convex neighborhood of \(D(x)\). For \(|s| \leq \delta \), \(D(x + sy)\) lies in the interior of \(W\), hence so does \(D_i(x + sy)\) for \(i\) large. By the mean value inclusion, \(t^{-1}(g_i(x + ty) - g_i(x)) \in W\) for \(0 {\lt} |t| \leq \delta \), and letting \(i \to \infty \) the same holds for \(g\), since \(W\) is closed. Glöckner and Neeb use instead their Exercise 1.1.13(b), whose hint goes through weak integrals; no completeness is needed here.
The image of \(C^{k}(U, F)\) in \(\prod _{j \leq k} C(U \times E^j, F)\) is closed.
Let \(h = (h_j)\) be in the closure of the image and \(f := h_0\). By induction on \(j\) and Lemma 6.4, applied along the filter of \(\gamma \) with \((\mathrm{d}^{(j)}\gamma )_j\) close to \(h\), the iterated directional derivatives of \(f\) exist and \(\mathrm{d}^{(j)} f = h_j\). By Bastiani’s characterization (Theorem 3.5), \(f\) is \(C^{k}\).
The following maps are continuous and linear: the inclusion \(C^{l}(U, F) \to C^{k}(U, F)\) for \(k \leq l\); the pullback \(\gamma \mapsto \gamma \circ L\), \(C^{k}(U, F) \to C^{k}(L^{-1}(U), F)\), along a continuous linear map \(L \colon E_1 \to E\); the postcomposition \(\gamma \mapsto g \circ \gamma \) with a continuous linear map \(g \colon F \to G\). Moreover \(C^{k}(U, F \times G) \cong C^{k}(U, F) \times C^{k}(U, G)\) as topological vector spaces.
In each case \(\mathrm{d}^{(j)}\) of the image is obtained from \(\mathrm{d}^{(j)}\) of the argument by pre- or postcomposition with a continuous map, or by pairing, and these operations are continuous in the compact-open topology (for pairing one restricts to compact subsets of \(U \times E^j\), which are locally compact).
Let \(\gamma \colon U \to F\) be \(C^{k+1}\) and \(j \leq k\). For \(x \in U\) and \(a, v_1, \dots , v_j, w_1, \dots , w_j \in E\),
Induction on \(j\). In the inductive step, the first summand is differentiated in the direction \((v_{j+1}, 0)\) by definition of \(\mathrm{d}^{(j+2)}\gamma \), and in the direction \((0, w_{j+1})\) by linearity of \(\mathrm{d}^{(j+1)}\gamma (x; \cdot )\) in its first argument; Schwarz’s theorem moves \(w_{j+1}\) to the last slot. The other summands depend only on \(x\).
The map \(\mathrm{d}\colon C^{k+1}(U, F) \to C^{k}(U \times E, F)\), \(\gamma \mapsto \mathrm{d}\gamma \), is continuous and linear, and the topology of \(C^{k+1}(U, F)\) is initial with respect to \(\gamma \mapsto \gamma \in C(U, F)\) and \(\mathrm{d}\).
Continuity of \(\mathrm{d}\) follows from Lemma 6.7. Conversely, \(\mathrm{d}^{(j+1)}\gamma (x; y_0, \dots , y_j) = \mathrm{d}^{(j)}(\mathrm{d}\gamma )((x, y_0); (y_1, 0), \dots , (y_j, 0))\) (Lemma 3.4), so each \(\mathrm{d}^{(j+1)}\) is \(\mathrm{d}^{(j)} \circ \mathrm{d}\) followed by a precomposition.
Let \(Z\), \(E\), \(F\) be locally convex spaces, \(U \subseteq Z\) open, \(r \in \mathbb {N}\) and \(f \colon Z \times E \to F\) of class \(C^{r}\) on \(U \times E\). Then \(f_* \colon C^{r}(U, E) \to C^{r}(U, F)\), \(\gamma \mapsto f \circ (\mathrm{id}_U, \gamma )\), is continuous. In particular, for \(g \colon E \to F\) of class \(C^{k}\), the map \(\gamma \mapsto g \circ \gamma \), \(C^{k}(U, E) \to C^{k}(U, F)\), is continuous.
Glöckner and Neeb use Faà di Bruno’s formula. We argue by induction on \(r\) instead. For \(r = 0\) this is the continuity of \(\gamma \mapsto f \circ (\mathrm{id}, \gamma )\) on \(C(U, E)\), which follows by restricting to compact subsets. By the chain rule, \(\mathrm{d}(f_*\gamma )(x, v) = \mathrm{d}f((x, \gamma (x)), (v, \mathrm{d}\gamma (x, v)))\), that is, \(\mathrm{d}(f_*\gamma ) = (\mathrm{d}f)_*(\gamma \circ \mathrm{pr}_1, \mathrm{d}\gamma )\) with \((\gamma \circ \mathrm{pr}_1, \mathrm{d}\gamma ) \in C^{r-1}(U \times Z, E \times E)\) depending continuously on \(\gamma \) (Propositions 6.6 and 6.8). The induction hypothesis and Proposition 6.8 conclude.
Let \(E_1\), \(E_2\), \(F\) be locally convex spaces, \(U \subseteq E_1\) and \(V \subseteq E_2\) open, and \(f \colon E_1 \to E_2\) of class \(C^{r}\) on \(U\) with \(f(U) \subseteq V\). Then \(f^* \colon C^{r}(V, F) \to C^{r}(U, F)\), \(\gamma \mapsto \gamma \circ f\), is continuous and linear.
Glöckner and Neeb use Faà di Bruno’s formula; we argue by induction on \(r\), as for Theorem 6.9. For \(r = 0\), \(f^*\) is a precomposition. By the chain rule, \(\mathrm{d}(\gamma \circ f)(x, v) = \mathrm{d}\gamma (f(x), \mathrm{d}f(x, v))\), so \(\mathrm{d}(f^*\gamma ) = g^*(\mathrm{d}\gamma )\) with \(g(x, v) = (f(x), \mathrm{d}f(x, v))\) of class \(C^{r-1}\) on \(U \times E_1\), and Proposition 6.8 and the induction hypothesis conclude.
The space \(C^{\infty }(U, F)\) of smooth maps, with the initial topology with respect to all the \(\mathrm{d}^{(j)}\), \(j \in \mathbb {N}\) (equivalently, with respect to the inclusions into the \(C^{k}(U, F)\)), is a locally convex Hausdorff space; it embeds in \(\prod _{j \in \mathbb {N}} C(U \times E^j, F)\) with closed image, and the results above (pullbacks, superposition operators, products, the differential) hold for \(k = \infty \).
Everything follows from the finite orders, since a map into \(C^{\infty }(U, F)\) is continuous if and only if its compositions with the inclusions into the \(C^{k}(U, F)\) are. For the closed image, the truncations of a point of the closure lie in the closed images of the \(C^{k}(U, F)\), and their preimages all have the same underlying map.
6.3 Smooth superposition operators and locality
Let \(U\) be an open subset of a locally compact locally convex space \(Z\), \(O \subseteq E\) open and \(h \colon E \to F\) smooth on \(O\). If \(g \colon X \to C^{r}(U, E)\) is \(C^{k}\) on an open set \(W\) and \(g(w)(U) \subseteq O\) for \(w \in W\), then \(w \mapsto h \circ g(w)\) is \(C^{k}\) on \(W\).
By induction on \(r\). The topology of \(C^{r}(U, F)\) is initial with respect to the differentials \(\mathrm{d}^{(j)}\), and \(C^{r}(U, F)\) is closed in \(\prod _j C(U \times Z^j, F)\), so it suffices to treat each \(\mathrm{d}^{(j)}\). For \(j = 0\), restriction to the compact subsets of \(U\) reduces the claim to superposition operators on \(C(C, E)\) with \(C\) compact, which are smooth by Theorem 6.1. For \(j \ge 1\), \(\mathrm{d}(h \circ \gamma ) = \mathrm{d}h \circ (\gamma \circ \mathrm{pr}_1, \mathrm{d}\gamma )\) is the superposition by the smooth map \(\mathrm{d}h\) of an element of \(C^{r-1}(U \times Z, E \times E)\) that depends continuously and linearly on \(\gamma \), and the induction hypothesis applies on \(U \times Z\).
Glöckner and Neeb obtain the smoothness of the maps between spaces of \(C^{\ell }\) maps that they need from the exponential law [ GN26 , Proposition 4.1.34 ] . The induction above avoids it. We do not claim it is new.
For an open cover \((V_a)\) of \(U\), the restrictions \(C^{r}(U, F) \to \prod _a C^{r}(V_a, F)\) form a linear topological embedding with closed range. Hence being \(C^{k}\), or converging, in \(C^{r}(U, F)\) is a local property.
On a locally compact space \(Y\), the restrictions \(C(Y, F) \to \prod _C C(C, F)\) to the compact subsets form a closed embedding.
\(C^{k}\) maps into a product are the maps with \(C^{k}\) components, and a map into a closed subspace is \(C^{k}\) as soon as it is \(C^{k}\) into the ambient space.
(a) The compact-open topology of \(C(U \times Z^j, F)\) is initial with respect to the restrictions to the \(V_a \times Z^j\), since every compact subset of \(U\) is covered by finitely many compact subsets of the \(V_a\); a compatible family glues to a \(C^{r}\) map. (b) is the case of the cover by compact sets, and (c) follows from the definition of the differentials.
6.4 \(C^{k,\ell }\) maps and the exponential law
Let \(E_1\), \(E_2\), \(F\) be locally convex spaces, \(U \subseteq E_1\) and \(V \subseteq E_2\) open, and \(k, \ell \in \mathbb {N}\).
A map \(f \colon U \times V \to F\) is \(C^{k,\ell }\) if the mixed differentials
exist for \(i \leq k\), \(j \leq \ell \) and are continuous on \(U \times V \times E_1^i \times E_2^j\). Equivalently (by Theorem 3.5), the maps \(f(x, \cdot )\) are \(C^{\ell }\), the maps \(x \mapsto \mathrm{d}^{(0,j)} f(x, y; w)\) are \(C^{k}\), and the \(\mathrm{d}^{(i,j)} f\) are continuous.
Let \(W\) be open and let \(g\) have directional derivatives \(A\) in the direction \(a\) and \(B\) in the direction \(b\) on \(W\), with \(B\) continuous and with continuous directional derivatives \(M\) in the direction \(a\). Then \(A\) has directional derivative \(M\) in the direction \(b\).
The difference quotients \(r^{-1}(g(z + ra) - g(z))\) converge pointwise to \(A\) as \(r \to 0\), and their derivatives in the direction \(b\), \(r^{-1}(B(z + ra) - B(z))\), converge to \(M\) uniformly on compact sets. Apply Lemma 6.4. Glöckner and Neeb differentiate a weak integral in a completion instead; no completeness is needed here.
If \(f\) is \(C^{k,\ell }\), \(i \leq k\), \(x \in U\) and \(v \in E_1^i\), then \(y \mapsto \mathrm{d}^{(i,0)} f(x, y; v)\) is \(C^{\ell }\) on \(V\) with \(j\)-th differential \(\mathrm{d}^{(i,j)} f(x, \cdot ; v; \cdot )\).
By induction on \(i\), Lemma 6.16 moves one derivative in the second variable past the derivatives in the first variable.
Let \(f\) be \(C^{k,\ell }\), \(i \leq k\), \(j \leq \ell \), \(v \in E_1^i\), \(w \in E_2^j\) and \(\pi \) a permutation of \(\{ 1, \ldots , i + j\} \). The iterated directional derivative of \(f\) along \((v_1, 0), \ldots , (v_i, 0), (0, w_1), \ldots , (0, w_j)\), taken in the order prescribed by \(\pi \), exists at every point of \(U \times V\) and equals \(\mathrm{d}^{(i,j)} f(x, y; v; w)\), which is symmetric in \(v\) and in \(w\).
By induction on the number of directions: the last one is either of the form \((v, 0)\), handled by the definition of \(\mathrm{d}^{(i,j)} f\), or \((0, w)\), handled by Proposition 6.17. The symmetry in \(v\) is the symmetry of the differentials of \(x \mapsto \mathrm{d}^{(0,j)} f(x, y; w)\); the symmetry in \(w\) follows from Proposition 6.17.
\(C^{k,\ell }(U \times V, F)\) carries the initial topology with respect to the linear maps \(\mathrm{d}^{(i,j)} \colon C^{k,\ell }(U \times V, F) \to C(U \times V \times E_1^i \times E_2^j, F)\), \(i \leq k\), \(j \leq \ell \) (compact-open topology on the right). It is a locally convex Hausdorff space.
Every \(C^{k+\ell }\) map on \(U \times V\) is \(C^{k,\ell }\), with \(\mathrm{d}^{(i,j)} f(x, y; v; w) = \mathrm{d}^{(j+i)} f((x, y); (0, w_1), \dots , (0, w_j), (v_1, 0), \dots , (v_i, 0))\), and the inclusion \(C^{k+\ell }(U \times V, F) \to C^{k,\ell }(U \times V, F)\) is continuous and linear.
The formula is an identity between iterated limits; the right hand side is continuous and depends continuously on \(f\).
For \(f \in C^{k,\ell }(U \times V, F)\), the map \(f^\vee \colon U \to C^{\ell }(V, F)\), \(x \mapsto f(x, \cdot )\), is \(C^{k}\), with \(\mathrm{d}^{(j)}(\mathrm{d}^{(i)} f^\vee (x; v))(y; w) = \mathrm{d}^{(i,j)} f(x, y; v; w)\), and \(\Phi \colon C^{k,\ell }(U \times V, F) \to C^{k}(U, C^{\ell }(V, F))\), \(f \mapsto f^\vee \), is linear and a topological embedding. If \(E_2\) is locally compact (for nonempty open \(V\), this is the same as \(V\) being locally compact), \(\Phi \) is an isomorphism of topological vector spaces, with inverse \(g \mapsto ((x, y) \mapsto g(x)(y))\).
The candidate for \(\mathrm{d}^{(i)} f^\vee (x; v)\) is \(y \mapsto \mathrm{d}^{(i,0)} f(x, y; v)\), which lies in \(C^{\ell }(V, F)\) by Proposition 6.17. The difference quotients of \(x \mapsto \mathrm{d}^{(i,j)} f(x, \cdot ; v; \cdot )\) converge uniformly on compact sets, so the derivative exists in \(C^{\ell }(V, F)\), and Bastiani’s characterization shows that \(f^\vee \) is \(C^{k}\). Both \(\Phi \) and the \(\mathrm{d}^{(i,j)}\) are recovered from each other by evaluations at points of compact sets, which are jointly continuous; hence \(\Phi \) is an embedding. If \(E_2\) is locally compact, \(V \times E_2^j\) is locally compact, so for \(g \in C^{k}(U, C^{\ell }(V, F))\) the map \((x, y, v, w) \mapsto \mathrm{d}^{(j)}(\mathrm{d}^{(i)} g(x; v))(y; w)\) is continuous; it is the mixed differential \(\mathrm{d}^{(i,j)}\) of \((x, y) \mapsto g(x)(y)\), which is therefore \(C^{k,\ell }\). Glöckner and Neeb use the evaluation map and a chain rule for \(C^{r,s}\) maps (with Faà di Bruno’s formula) instead.
Every \(C^{k,k}\) map on \(U \times V\) is \(C^{k}\), and the inclusion \(C^{k,k}(U \times V, F) \to C^{k}(U \times V, F)\) is continuous and linear.
Induction on \(k\). By the rule on partial differentials, a \(C^{k+1,k+1}\) map \(f\) is \(C^{1}\) with \(\mathrm{d}f((x, y), (a, b)) = \mathrm{d}^{(1,0)} f(x, y; a) + \mathrm{d}^{(0,1)} f(x, y; b)\). As a function of \(((x, a), (y, b))\), the right hand side is \(C^{k,k}\) on \((U \times E_1) \times (V \times E_2)\): its mixed differentials are explicit sums of mixed differentials of \(f\), by Lemma 6.7 and Proposition 6.17. The induction hypothesis gives that \(\mathrm{d}f\) is \(C^{k}\), and also, with Proposition 6.8, the continuity of the inclusion. Glöckner and Neeb prove instead the explicit formula of Proposition 6.23 for \(\mathrm{d}^{(m)} f\).
If \(f\) is \(C^{m,m}\), \((x, y) \in U \times V\) and \(w \in (E_1 \times E_2)^m\), then
with the indices in increasing order.
A map on \(U \times V\) is smooth if and only if it is \(C^{k,\ell }\) for all \(k, \ell \), and \(C^{\infty }(U \times V, F) = C^{\infty ,\infty }(U \times V, F)\) as locally convex spaces.
A map \(g \colon W \to C^{\infty }(V, F)\) is \(C^{k}\) if and only if its compositions with the inclusions \(C^{\infty }(V, F) \to C^{\ell }(V, F)\) are \(C^{k}\), for all \(\ell \).
Induction on \(k\). The derivatives of the compositions have the same underlying maps, hence define a candidate in \(C^{\infty }(V, F)\), and the topology of \(C^{\infty }(V, F)\) is initial with respect to the inclusions.
For \(f \in C^{\infty }(U \times V, F)\), \(f^\vee \colon U \to C^{\infty }(V, F)\) is smooth, and \(\Phi \colon C^{\infty }(U \times V, F) \to C^{\infty }(U, C^{\infty }(V, F))\), \(f \mapsto f^\vee \), is linear and a topological embedding. If \(E_2\) is locally compact, \(\Phi \) is an isomorphism of topological vector spaces, and a map \(g \colon U \to C^{\infty }(V, F)\) is smooth if and only if \((x, y) \mapsto g(x)(y)\) is smooth.