Michal–Bastiani calculus on locally convex spaces

1 Introduction

This blueprint describes the mathematics formalized in the Lean library lc-analysis: the differential and integral calculus of Michal and Bastiani on locally convex spaces. Every definition and theorem below is linked to its Lean counterpart, and the dependency graph shows how the results build on each other. All the statements in this document are proved in Lean, with no sorry, and the main theorems depend only on the standard axioms of Lean.

Why Michal–Bastiani calculus

Infinite-dimensional Lie groups and manifolds of mappings are modeled on spaces that are not normable: Fréchet spaces of smooth functions such as \(C^\infty (M)\), spaces of test functions, direct limits. On such spaces the Fréchet derivative is not available, since there is no norm with which to measure the remainder. One is faced with the problem of choosing a notion of differentiability that does not depend on a norm, and that still has a chain rule, a mean value theorem and a fundamental theorem of calculus.

The answer that has become standard in infinite-dimensional Lie theory goes back to Michal (1938, [ Mic38 ] ) and Bastiani (1964, [ Bas64 ] ). It was developed further by Keller [ Kel74 ] , Hamilton [ Ham82 ] , Milnor [ Mil84 ] , Glöckner [ Glo02 ] and Neeb [ Nee06 ] , and its main reference today is the book of Glöckner and Neeb [ GN26 ] , which we follow closely; [ GN26 , 1.2.13 ] refers to its numbered result 1.2.13. We remark that this is not the convenient setting of Kriegl and Michor [ KM97 ] : smoothness is defined here through the topology of the model space, and not through smooth curves.

The idea is to ask for less than in the Banach case. A map is \(C^{1}\) if all its directional derivatives exist and depend continuously on the point and on the direction, jointly. No uniformity in the direction is required. In normed spaces this notion is weaker than Fréchet differentiability, but the two give the same smooth maps (Chapter 8).

Mathlib [ mC20 ] , the mathematical library of the Lean proof assistant [ dMU21 ] , has a complete Fréchet calculus on normed spaces, and a theory of topological vector spaces and locally convex spaces. It does not have a differential calculus beyond the normed case. The purpose of lc-analysis is to fill this gap, as a foundation for the formalization of infinite-dimensional Lie theory.

Standing assumptions

Unless otherwise stated, \(E\), \(F\) and \(G\) denote real locally convex Hausdorff topological vector spaces, \(U \subseteq E\) is open, and \(f \colon U \to F\). These spaces are not assumed to be complete in any sense; whenever some form of completeness is needed (mainly for the existence of integrals) it appears as an explicit hypothesis. The Lean statements sometimes need less than the standing assumptions: for instance, the definition of weak integrals makes sense in any topological vector space. The precise hypotheses can be read in the linked Lean declarations.

Two tools replace norm estimates throughout:

  • Hahn–Banach. Continuous linear functionals separate points from closed convex sets. In particular, a vector is determined by its images under all continuous linear functionals.

  • Local convexity. The closed convex neighborhoods of \(0\) form a basis of the topology.

Most proofs reduce a statement in \(F\) to statements about real functions, by composing with continuous linear functionals, and then the one-variable calculus of Mathlib applies. Where in the normed case a norm estimate would bound an increment, here we show instead that the increment lies in a prescribed closed convex neighborhood.

Contents

The chapters follow the folders of the library, so that each part can be read on its own.

  • Calculus. Chapter 2 defines \(C^{k}\) maps and proves the first-order theory: linearity of the differential, the mean value theorem, the chain rule, partial derivatives and the symmetry of the second differential. Chapter 3 compares the recursive definition of \(C^{k}\) with Bastiani’s definition through iterated directional derivatives, and collects the rules of calculus (bilinear maps, Taylor’s formula, the second-order chain rule).

  • Integration. Chapter 4 defines weak integrals, and Chapter 5 studies the completeness hypotheses under which they exist, including the equivalence of two definitions of Mackey completeness.

  • Mapping spaces. Chapter 6 gives spaces of \(C^{k}\) maps the compact-open \(C^{k}\) topology and proves the exponential law; Chapter 7 extends the calculus to locally convex subsets with dense interior, the domains of charts of manifolds with boundary.

  • Normed spaces. Chapter 8 compares Michal–Bastiani and Fréchet differentiability on normed spaces, and treats maps with bounded below derivative.

  • Analytic maps. Chapter 9 develops complex and real analytic maps on locally convex spaces.

  • Differential equations and integrability. Chapter 10 treats flows on Banach spaces and equations with locally convex parameters, Chapter 11 the Frobenius theorem, and Chapter 12 complex structures and the Newlander–Nirenberg theorem.

  • Spaces. Chapter 13 treats direct limits of topological spaces and \(k_\omega \)-spaces, Chapter 14 locally convex direct sums, direct limits and Silva spaces, and Chapter 15 Krein spaces.

  • Limits of the theory. Chapter 16 collects the formalized counterexamples beyond the normed setting, and Chapter 17 explains the design choices they impose.

What is new and what is not

The mathematics is classical: every result in this document appears in the cited literature, and the proofs follow it. The few places where the argument differs from the sources are pointed out in remarks. As far as we know, Michal–Bastiani calculus had not been formalized before, neither in Lean nor in another proof assistant.

How the library was written

@glaroton set the mathematical program, chose the framework and the references, fixed the design constraints and reviewed the results. The Lean proofs were written by the AI assistant Claude (Anthropic), working under his direction. Correctness does not rest on trusting either of them: Lean’s kernel checks every proof, and the statements are kept close to the literature so that they can be compared with it directly.