17 Limitations of the Michal–Bastiani calculus
The Michal–Bastiani calculus is defined on arbitrary locally convex spaces, needs no completeness, and its smooth maps are continuous. The price is paid elsewhere: spaces of operators and of smooth maps carry several natural topologies, none of which makes evaluation continuous in infinite dimensions, and as a consequence the category of smooth maps is not cartesian closed. This chapter collects these caveats, explains why they are not a matter of completeness, and records the convention that we follow throughout to avoid them: operator-valued data are always required to be smooth jointly (in uncurried form), never as maps into a space of operators. Nothing in this chapter is formalized; it is here to explain the design.
Throughout, \(E\) and \(F\) are locally convex spaces over \(\mathbb {K} \in \{ \mathbb {R}, \mathbb {C}\} \), and \(\mathcal{L}(E, F)\) is the space of continuous linear maps.
17.1 Topologies on spaces of operators
Three topologies on \(\mathcal{L}(E, F)\) will appear, each one the topology of uniform convergence on a family of subsets of \(E\):
\(\mathcal{L}(E, F)_s\): uniform convergence on finite sets (pointwise convergence); on normed spaces this is the strong operator topology;
\(\mathcal{L}(E, F)_c\): uniform convergence on compact sets (the compact-open topology);
\(\mathcal{L}(E, F)_b\): uniform convergence on bounded sets; on normed spaces this is the operator norm topology, since it suffices to look at the unit ball.
Since finite sets are compact and compact sets are bounded, \(s \subseteq c \subseteq b\).
A word of caution on terminology: in operator theory “strong topology” means \(s\), while in the theory of locally convex spaces the “strong topology” (on a dual space, say) is \(b\).
On an equicontinuous subset \(H \subseteq \mathcal{L}(E, F)\) (for normed \(E\), \(F\): a set bounded in operator norm) the topologies \(s\) and \(c\) coincide. Indeed, given a compact \(K \subseteq E\), a continuous seminorm \(q\) on \(F\) and \(\varepsilon {\gt} 0\), equicontinuity gives a \(0\)-neighborhood \(V\) with \(q(T v) {\lt} \varepsilon \) for \(T \in H\), \(v \in V\); covering \(K\) by finitely many translates \(x_i + V\) reduces uniform convergence on \(K\) to convergence at the finitely many points \(x_i\). In particular, for a Banach space \(E\), a sequence in \(\mathcal{L}(E, F)\) which converges in \(s\) is bounded in operator norm (uniform boundedness principle), hence converges in \(c\): on Banach spaces the two topologies can only be distinguished by nets.
If \(E\) is infinite-dimensional and \(F \ne 0\), then \(s \ne c\) on \(\mathcal{L}(E, F)\). It suffices to treat \(F = \mathbb {K}\), i.e. the dual \(E'\). Every \(s\)-neighborhood of \(0\) in \(E'\) contains the annihilator of a finite set \(\{ x_1, \dots , x_k\} \). If the polar \(K^\circ \) of a compact set \(K\) contained such an annihilator, then every \(f \in E'\) vanishing on the \(x_j\) would vanish on \(K\) (scale \(f\)), and by the Hahn–Banach theorem \(K\) would lie in the finite-dimensional span of the \(x_j\). A compact set not contained in any finite-dimensional subspace exists (for normed \(E\): \(\{ 0\} \cup \{ x_n / n\} \) with \(x_n\) linearly independent unit vectors), so \(K^\circ \) is a \(c\)-neighborhood of \(0\) which is not an \(s\)-neighborhood.
(Incomplete \(E\), sequences.) Let \(E = c_{00}\) be the space of finitely supported sequences with the supremum norm, and \(\varphi _n = n\, e_n^*\). For each \(x \in E\), \(\varphi _n(x) = 0\) for large \(n\), so \(\varphi _n \to 0\) in \(s\). On the compact set \(K = \{ 0\} \cup \{ e_m / m : m \ge 1\} \) we have \(\varphi _n(e_n / n) = 1\), so \(\varphi _n \not\to 0\) in \(c\). (By Remark 17.1, such a sequence cannot exist on a Banach space.)
(\(E = \ell ^2\), nets.) Let \(A = \{ \sqrt{n}\, e_n^* : n \ge 1\} \subseteq (\ell ^2)'\). Then \(0\) lies in the \(s\)-closure of \(A\): otherwise there would be \(x^{(1)}, \dots , x^{(k)} \in \ell ^2\) and \(\varepsilon {\gt} 0\) with \(\max _j \sqrt{n}\, |x^{(j)}_n| \ge \varepsilon \) for all \(n\), whence \(\sum _j |x^{(j)}_n|^2 \ge \varepsilon ^2 / n\), and summing over \(n\) gives a contradiction. But \(0\) is not in the \(c\)-closure of \(A\): on the compact set \(K = \{ 0\} \cup \{ e_m / \sqrt{m}\} \) each element of \(A\) takes the value \(1\).
For a Banach space \(E\), the topology \(c\) on \(E'\) is the bounded weak-\(*\) topology: the finest topology which agrees with the weak-\(*\) topology on bounded sets (Banach–Dieudonné theorem; for the statement on convex sets see [ FHHMPZ01 , Thm. 4.44 ] ).
17.2 The exponential law
For locally convex subsets \(U \subseteq E_1\) and \(V \subseteq E_2\) with dense interior, endow \(C^\infty (V, F)\) with the compact-open \(C^\infty \)-topology, and write \(f^\vee (x) = f(x, \cdot )\).
For all locally convex \(E_1\), \(E_2\), \(F\) and every \(f \in C^\infty (U \times V, F)\), the map \(f^\vee \colon U \to C^\infty (V, F)\) 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 moreover \(V\) is locally compact, or \(U \times V \times E_1 \times E_2\) is a \(k_{\mathbb {R}}\)-space, then \(\Phi \) is an isomorphism of topological vector spaces.
No completeness is assumed in Theorem 17.4. The hypothesis for the converse direction is topological: a \(k_{\mathbb {R}}\)-space is a Hausdorff space on which a real function is continuous as soon as its restrictions to compact sets are [ GN26 , Rem. A.6.15 ] . It holds when \(E_1\) and \(E_2\) are metrizable (normed spaces, complete or not; Fréchet spaces), when \(V\) is finite-dimensional, and also for \(k_\omega \)-spaces such as countable direct limits of finite-dimensional spaces, e.g. \(\mathbb {R}^{(\mathbb {N})}\) [ GN26 , Rem. 4.1.22, Prop. A.7.20 ] . So the absence of a countable basis of \(0\)-neighborhoods only threatens the converse direction, and only for spaces which are not \(k_{\mathbb {R}}\)-spaces.
17.3 Evaluation and cartesian closedness
The category would be cartesian closed if \(C^\infty (V, F)\) were an object and the evaluation \(\varepsilon \colon C^\infty (V, F) \times V \to F\) were smooth. This holds when \(V\) is finite-dimensional: then \(\varepsilon \) is \(C^{\infty , \ell }\) [ GN26 , Lemma 1.7.25 ] . It fails in every infinite-dimensional case, and already at the linear level.
If \(E\) is infinite-dimensional, the evaluation \(E'_c \times E \to \mathbb {K}\), \((f, x) \mapsto f(x)\), is not continuous. Consequently, for every open \(U \subseteq E\), the evaluation \(C^\infty (U, \mathbb {K}) \times U \to \mathbb {K}\) is not continuous for the compact-open \(C^\infty \)-topology.
Suppose the evaluation is continuous at \((0, 0)\). Then there are a compact \(K \subseteq E\), \(\delta {\gt} 0\) and an absolutely convex \(0\)-neighborhood \(W \subseteq E\) with \(|f(x)| \le 1\) whenever \(\sup _K |f| \le \delta \) and \(x \in W\). By homogeneity, \(K^\circ \subseteq \delta ^{-1} W^\circ \), and by the bipolar theorem \(W \subseteq \overline{W} = W^{\circ \circ } \subseteq \delta ^{-1} K^{\circ \circ }\), where \(K^{\circ \circ }\) is the closed absolutely convex hull of \(K\). The absolutely convex hull of a precompact set is precompact, and so is its closure; hence \(E\) has a precompact \(0\)-neighborhood and is finite-dimensional.
For the second statement, the compact-open \(C^\infty \)-topology induces on \(E' \subseteq C^\infty (U, \mathbb {K})\) the topology of \(E'_c\) (for a linear map all derivatives of order \(\ge 1\) are \(f\) itself or zero), and continuity of the evaluation at \((0, x_0)\) would give continuity of \((f, h) \mapsto f(x_0 + h) - f(x_0) = f(h)\) at \((0, 0)\).
The proposition covers Banach and Hilbert spaces: there the evaluation \(\mathcal{L}(E, F)_b \times E \to F\) is continuous [ GN26 , Rem. 2.3.9 ] , but \(b\) is not the topology which the compact-open \(C^\infty \)-topology induces on operators. For non-normable \(E\) there is no way out at all: the evaluation \(\mathcal{L}(E) \times E \to E\) is discontinuous for every vector topology on \(\mathcal{L}(E)\) (Maissen [ Mai63 , Satz 1 ] , see also [ KM97 , p. 2 ] and the notes to Ch. 3 of [ GN26 ] ).
The convenient calculus of Frölicher, Kriegl and Michor [ KM97 ] is cartesian closed. It works on Mackey complete (\(c^\infty \)-complete) spaces, and this is where a completeness hypothesis enters. The price: the exponential law holds only as an isomorphism of bornological spaces (already for \(E_1 = E_2 = \mathbb {R}\), the map \(\Phi \) is discontinuous for the topologies of the convenient setting, see the notes to Ch. 1 of [ GN26 ] ), and convenient smooth maps need not be continuous. On Fréchet spaces both calculi have the same smooth maps.
17.4 The derivative as an operator-valued map
Let \(f \colon U \to F\) be \(C^{k}\) with \(k \ge 1\), \(U \subseteq E\) open.
Into \(\mathcal{L}(E, F)_c\), no hypotheses. The map \(x \mapsto \mathrm{d}f(x) \in \mathcal{L}(E, F)_c\) is \(C^{k-1}\), for arbitrary locally convex \(E\), \(F\): apply the first part of Theorem 17.4 to \(\mathrm{d}f\) and corestrict to the closed subspace \(\mathcal{L}(E, F) \subseteq C(E, F)\) (as in the proof of [ GN26 , Prop. 2.3.17(b) ] ). Dually, for fixed \(x\) the map \(C^\infty (U, F) \to \mathcal{L}(E, F)_c\), \(f \mapsto \mathrm{d}f(x)\), is continuous linear, directly from the definition of the compact-open \(C^\infty \)-topology.
Into \(\mathcal{L}(E, F)_b\), normed spaces. If \(E\), \(F\) are normed, a Michal–Bastiani \(C^{k+1}\) map is Fréchet \(C^{k}\) (Theorem 8.1): continuity of \(x \mapsto \mathrm{d}f(x)\) in operator norm costs one derivative, and nothing for smooth maps (Corollary 8.3). On the other hand, \(f \mapsto \mathrm{d}f(x)\) is not continuous from \(C^\infty (U, F)\) to \(\mathcal{L}(E, F)_b\) when \(E\) is infinite-dimensional, since on \(E'\) the compact-open \(C^\infty \)-topology is \(c \ne b\).
Non-normable spaces. By Maissen’s theorem, there is no vector topology on \(\mathcal{L}(E)\) making evaluation continuous, so no choice of topology turns the derivative into a well-behaved operator-valued map while keeping the chain rule continuous.
17.5 The convention adopted
To stay clear of all of the above, operator-valued data are always required to be smooth in uncurried form. A smooth family of projectors on an embedded submanifold \(X \subseteq F\) is a map \(E \colon X \times F \to F\), linear in the second variable, such that \((p, w) \mapsto E_p w\) is smooth in charts (LocallyConvex.Manifolds.ProjectorField); the Christoffel symbols of a connection are smooth as functions \((x, a, b) \mapsto \Gamma _x(a, b)\) of three variables (LocallyConvex.Manifolds.Connection). No topology on \(\mathcal{L}(F)\) is ever needed. On Banach spaces this loses nothing: by Corollary 8.3, \((x, w) \mapsto E_x w\) is then Fréchet smooth, hence so is \(x \mapsto E_x\) into \(\mathcal{L}(F)\) with the operator norm; the uncurried and the operator-norm notions of smoothness agree for \(C^\infty \) data.
The convention does have visible consequences on non-normable models: the dual bundle \(T^*M = \bigcup _x (T_x M)'\) carries in general no canonical vector bundle structure [ GN26 , notes to Ch. 3 ] , so \(1\)-forms are treated as smooth maps on \(TM\) which are linear on the fibres; and the adjoint action of a Lie group is handled as the smooth map \(G \times \mathfrak {g} \to \mathfrak {g}\), not as a map into \(\mathcal{L}(\mathfrak {g})\).
- AI89
T. Ya. Azizov and I. S. Iokhvidov, Linear Operators in Spaces with an Indefinite Metric, Wiley, 1989.
- Bas64
A. Bastiani, Applications différentiables et variétés différentiables de dimension infinie, J. Analyse Math. 13 (1964), 1–114, doi:10.1007/BF02786619.
- Bog74
J. Bognár, Indefinite Inner Product Spaces, Ergebnisse der Mathematik und ihrer Grenzgebiete 78, Springer, 1974, doi:10.1007/978-3-642-65567-8.
- BS71a
J. Bochnak and J. Siciak, Polynomials and multilinear mappings in topological vector-spaces, Studia Math. 39 (1971), 59–76, doi:10.4064/sm-39-1-59-76.
- BS71b
J. Bochnak and J. Siciak, Analytic functions in topological vector-spaces, Studia Math. 39 (1971), 77–112, doi:10.4064/sm-39-1-77-112.
- FHHMPZ01
M. Fabian, P. Habala, P. Hájek, V. Montesinos Santalucía, J. Pelant and V. Zizler, Functional Analysis and Infinite-Dimensional Geometry, CMS Books Math. 8, Springer, 2001, doi:10.1007/978-1-4757-3480-5.
- GGH10
H. Glöckner, R. Gramlich and T. Hartnick, Final group topologies, Kac–Moody groups and Pontryagin duality, Israel J. Math. 177 (2010), 49–101, doi:10.1007/s11856-010-0038-5.
- GN26
H. Glöckner and K.-H. Neeb, Infinite-Dimensional Lie Groups, book manuscript, 2026, arXiv:2602.12362.
- Glo02
H. Glöckner, Infinite-dimensional Lie groups without completeness restrictions, in Geometry and Analysis on Finite- and Infinite-Dimensional Lie Groups, Banach Center Publ. 55 (2002), 43–59, doi:10.4064/bc55-0-3.
- Ham82
R. S. Hamilton, The inverse function theorem of Nash and Moser, Bull. Amer. Math. Soc. 7 (1982), 65–222, doi:10.1090/S0273-0979-1982-15004-2.
- Kel74
H. H. Keller, Differential Calculus in Locally Convex Spaces, Lecture Notes in Math. 417, Springer, 1974, doi:10.1007/BFb0070564.
- KM97
A. Kriegl and P. W. Michor, The Convenient Setting of Global Analysis, Math. Surveys Monogr. 53, Amer. Math. Soc., 1997, doi:10.1090/surv/053.
- Lan99
S. Lang, Fundamentals of Differential Geometry, Grad. Texts in Math. 191, Springer, 1999, doi:10.1007/978-1-4612-0541-8.
- Mai63
B. Maissen, Über Topologien im Endomorphismenraum eines topologischen Vektorraumes, Math. Ann. 151 (1963), 283–285, doi:10.1007/BF01470820.
- Mar72
J. E. Marsden, Darboux’s theorem fails for weak symplectic forms, Proc. Amer. Math. Soc. 32 (1972), 590–592, doi:10.1090/S0002-9939-1972-0293678-X.
- Mic38
A. D. Michal, Differential calculus in linear topological spaces, Proc. Nat. Acad. Sci. USA 24 (1938), 340–342, doi:10.1073/pnas.24.8.340.
- Mic80
P. W. Michor, Manifolds of Differentiable Mappings, Shiva Math. Series 3, 1980.
- Mil84
J. Milnor, Remarks on infinite-dimensional Lie groups, in Relativity, Groups and Topology II (Les Houches, 1983), North-Holland, 1984, 1007–1057.
- Muj86
J. Mujica, Complex Analysis in Banach Spaces, North-Holland Math. Stud. 120, North-Holland, 1986.
- Nee06
K.-H. Neeb, Towards a Lie theory of locally convex groups, Jpn. J. Math. 1 (2006), 291–468, doi:10.1007/s11537-006-0606-y; arXiv:1501.06269.
- Pen70
J.-P. Penot, Sur le théorème de Frobenius, Bull. Soc. Math. France 98 (1970), 47–80, doi:10.24033/bsmf.1694.
- Wei69
A. Weinstein, Symplectic structures on Banach manifolds, Bull. Amer. Math. Soc. 75 (1969), 1040–1041, doi:10.1090/S0002-9904-1969-12353-0.
- dMU21
L. de Moura and S. Ullrich, The Lean 4 theorem prover and programming language, CADE-28, LNCS 12699 (2021), 625–635, doi:10.1007/978-3-030-79876-5_37.
- mC20
The mathlib Community, The Lean mathematical library, CPP 2020, 367–381, doi:10.1145/3372885.3373824.