7 Submanifolds and immersions
In finite dimensions all reasonable notions of submanifold and immersion agree, up to some topological phenomena such as the figure eight. In infinite dimensions they split apart, already for Banach models and more so for locally convex ones. In this chapter we first construct embedded submanifolds, defined by slice charts onto a closed subspace of the model, and show that they are manifolds in their own right. Then we introduce three notions of immersion (weak, plain, split) and the corresponding notions of submanifold given by an injective map (weakly immersed, immersed, initial), and we prove the implications among them. Finally we show that for a Banach domain an immersion reflects smoothness of continuous maps. The chapter closes with a list of the infinite-dimensional subtleties; some are formalized as counterexamples in Chapter 18, others are only discussed. The references are [ GN26 , Def. 3.1.10, Def. 3.3.22, Rem. 3.3.23, Lemma 3.3.24 ] , [ Nee06 , § II.6, § IV.4 ] , [ KMW24 ] and [ Lar26 , Cap. 2 ] .
Throughout, \(E\) and \(F\) are real locally convex Hausdorff spaces, not necessarily complete, and \(M\), \(N\) are manifolds modeled on \(E\) and \(F\) (Definition 2.2).
7.1 Embedded submanifolds
The whole construction rests on one calculus fact: a map with values in a closed subspace is as smooth into the subspace as it is into the ambient space.
Let \(T \subseteq F\) be a closed subspace, \(D\) a locally convex space, \(V \subseteq D\) open, and \(h \colon V \to F\) a \(C^{k}\) map with \(h(V) \subseteq T\). Then \(h\) is \(C^{k}\) as a map \(V \to T\).
Induction on \(k\). The difference quotients of \(h\) lie in \(T\), so their limit \(\mathrm{d}h(x)(w)\) lies in \(T\) because \(T\) is closed; the directional derivative of the corestriction is the corestriction of \(\mathrm{d}h\), and the inductive hypothesis applies to \(\mathrm{d}h\) on \(V \times D\). This is [ GN26 , Lemma 1.3.19 ] .
Let \(N \subseteq M\) and let \(S \subseteq E\) be a subspace. A chart \(\varphi \colon U \to E\) is a slice chart for \(N\) with model \(S\) if \(\varphi \) belongs to the maximal atlas of \(M\) and \(\varphi (U \cap N) = \varphi (U) \cap S\), that is, \(q \in N\) iff \(\varphi (q) \in S\) for \(q \in U\).
A subset \(N \subseteq M\) is an embedded submanifold modeled on the subspace \(S \subseteq E\) if \(S\) is closed and every point of \(N\) lies in the domain of a slice chart for \(N\) with model \(S\).
This is [ GN26 , Def. 3.1.10 ] . No complement of \(S\) is required.
The slice charts must be taken in the maximal atlas, i.e. among the charts compatible with the atlas of \(M\), and not in the atlas itself. For instance, when \(M = E\) the atlas only contains the identity, which is a slice chart only for subspaces.
A slice chart \(\varphi \colon U \to E\) induces a chart of \(N\) with values in \(S\), namely the restriction \(\varphi |_{U \cap N} \colon U \cap N \to \varphi (U) \cap S\), whose inverse is the restriction of \(\varphi ^{-1}\) to \(\varphi (U) \cap S\).
In Lean the induced chart is a total function, and a base point of \(N\) is used as a junk value outside the target of the inverse.
If \(S\) is closed, the transition map between two induced charts is smooth, and so is its inverse; hence it belongs to the Michal–Bastiani groupoid of \(S\).
The transition map is the restriction to \(S\) of the transition map \(\psi \circ \varphi ^{-1}\) of \(M\), which is smooth because both charts are in the maximal atlas. It takes values in \(S\), so it is smooth into \(S\) by Lemma 7.1. The inverse is handled by symmetry.
Let \(N \subseteq M\) be an embedded submanifold modeled on \(S\). The induced charts form an atlas of \(N\), with the topology induced by \(M\), and with it \(N\) is a manifold modeled on \(S\).
The atlas consists of all induced charts; at each point we choose one given by the definition. Compatibility is Lemma 7.6.
The inclusion \(N \hookrightarrow M\) of an embedded submanifold is smooth.
In an induced chart and a chart \(c\) of \(M\), the local representative of the inclusion is \((c \circ \varphi ^{-1}) \circ \iota _S\), with \(\iota _S \colon S \to E\) the inclusion: a smooth transition map of \(M\) composed with a continuous linear map.
The simplest example is a closed subspace of the model space itself.
A closed subspace \(S \subseteq F\) is an embedded submanifold of \(F\) modeled on \(S\), with the identity as slice chart. If \(S\) is complemented, it is a split submanifold.
The identity of \(F\) satisfies \(q \in S \iff \mathrm{id}(q) \in S\).
When \(S\) is complemented, Glöckner and Neeb call \(N\) a split submanifold (Definition 7.26). In the Banach literature (e.g. Bourbaki) “submanifold” usually means split, and non-split ones are called quasi-submanifolds. For locally convex models the splitting condition is much less useful [ GN26 , Rem. 3.1.11 ] . On the other hand, the usual criteria to produce submanifolds (regular values, split immersions) rest on the inverse or implicit function theorems, which are not available for general locally convex models. In this library embedded submanifolds are therefore produced by exhibiting slice charts by hand (see Chapter 18).
7.2 The tangent space of a submanifold of a vector space
When the ambient manifold is a locally convex space \(F\) itself, the tangent space of an embedded submanifold can be described as a subspace of \(F\). This is the description used for fields of projectors and the induced connections in the chapter on connections.
Let \(X \subseteq F\) be an embedded submanifold modeled on \(S\) and \(p \in X\). Choose a slice chart \(\varphi \) at \(p\). The tangent space of \(X\) at \(p\) is
If \(\varphi , \varphi '\) are slice charts at \(p \in X\), then \(\mathrm{d}\varphi (p)(w) \in S\) iff \(\mathrm{d}\varphi '(p)(w) \in S\). Hence \(T_p X\) can be read in any slice chart, and it contains \(\mathrm{d}\varphi ^{-1}(\varphi (p))(s)\) for every \(s \in S\).
The transition map \(\tau = \varphi ' \circ \varphi ^{-1}\) maps the slice into the slice, so by closedness of \(S\) its derivative at \(\varphi (p)\) maps \(S\) into \(S\); the same holds for \(\tau ^{-1}\). The chain rule then gives the equivalence. The last claim follows since \(\mathrm{d}\varphi \circ \mathrm{d}\varphi ^{-1}\) is the identity.
A curve \(\beta \colon \mathbb {R}\to X\) which is smooth as a curve in \(F\) is a \(C^{k}\) curve of the manifold \(X\), for every \(k\).
In a slice chart \(\varphi \) the local representative is the corestriction to \(S\) of \(\varphi \circ \beta \), which is smooth by Lemma 7.1.
For a closed subspace \(S \subseteq F\), \(T_p S = S\) at every \(p \in S\).
Read the tangent space in the identity chart, whose derivative is the identity.
7.3 Immersions
Let \(f \colon M \to N\) be smooth. We read its differential at \(x\) in the preferred charts \(c\) at \(x\) and \(c'\) at \(f(x)\), as the map \(v \mapsto \mathrm{d}(c' \circ f \circ c^{-1})(c(x))(v)\), \(E \to F\). It is continuous and linear.
For charts \(e, e'\) of the maximal atlas, the derivative of \(e \circ e'^{-1}\) at the image point is a left inverse of the derivative of \(e' \circ e^{-1}\); in particular the latter is injective. The differential of a smooth map, read in the preferred charts, is continuous.
Differentiate \((e \circ e'^{-1}) \circ (e' \circ e^{-1}) = \mathrm{id}\) near the point, by the chain rule.
A smooth map \(f \colon M \to N\) is a weak immersion if its differential is injective at every point. No condition is imposed on its range.
A smooth map \(f \colon M \to N\) is an immersion if, at every point, its differential has closed range and is a topological embedding, i.e. a topological isomorphism onto a closed subspace of \(F\).
A smooth map \(f \colon M \to N\) is a split immersion if, at every point, its differential has a continuous linear left inverse \(r \colon F \to E\). Equivalently, the differential is a topological isomorphism onto a complemented subspace.
Every split immersion is an immersion, and every immersion is a weak immersion.
Let \(D\) be the differential and \(r\) a continuous left inverse. The range of \(D\) is the fixed set of the continuous map \(D \circ r\), hence closed, and \(D\) is an embedding since \(r\) is a continuous left inverse. The second implication is immediate: embeddings are injective.
The three definitions are stated in the preferred charts. Since the derivatives of transition maps are linear homeomorphisms (Lemma 7.15), the conditions do not depend on this choice; the chart independence is not stated separately in Lean.
The linear inclusion \(S \hookrightarrow F\) of a closed subspace is an immersion. It is a split immersion if and only if \(S\) is complemented.
In the identity charts the differential is the inclusion itself, a topological embedding with closed range \(S\). A continuous left inverse \(r \colon F \to S\) gives the continuous projection \(\iota _S \circ r\) onto \(S\), and conversely.
[ GN26 , Def. 3.3.25(b) ] calls \(f\) an immersion when it has, locally, the form of a continuous linear map with a continuous linear left inverse; this forces the range of the differential to be complemented. We do not use it as the definition of an immersion: it is much more restrictive than Definition 7.17. For Banach models, an immersion with complemented range has this local form by the inverse function theorem; for locally convex models that theorem is not available.
7.4 Submanifolds given by an injective map
An injective map \(f \colon M \to N\) exhibits its image as a submanifold of \(N\), with the manifold structure transported from \(M\). This structure is extra data: the figure eight carries two such structures. Initial submanifolds are the exception, since their structure is unique [ GN26 , Lemma 3.3.24 ] ; this uniqueness is not formalized.
The map \(f\) exhibits its image as a weakly immersed submanifold if it is an injective weak immersion, and as an immersed submanifold if it is an injective immersion.
The map \(f\) exhibits its image as an initial submanifold if it is an injective weak immersion and, for every manifold \(P\) modeled on a locally convex space \(H\) and every map \(g \colon P \to M\), the map \(g\) is smooth as soon as \(f \circ g\) is smooth.
This is [ GN26 , Def. 3.3.22 ] : smooth maps into \(N\) with values in the image lift to smooth maps into \(M\). The topology of \(M\) may be finer than the one induced by \(N\) [ GN26 , Rem. 3.3.23 ] . In Lean the test spaces \(H\) and \(P\) range over the universe of the model spaces.
Immersed submanifolds and initial submanifolds are weakly immersed.
The first claim is Proposition 7.19; the second holds by definition.
7.5 Embedded and split submanifolds as immersions
An embedded submanifold \(X \subseteq M\) modeled on \(S\) is split if \(S\) is complemented in \(E\) [ GN26 , Def. 3.1.10 ] .
Let \(X \subseteq M\) be an embedded submanifold modeled on \(S\) and \(p \in X\). In the preferred charts, the differential of the inclusion at \(p\) is the restriction to \(S\) of a continuous linear automorphism \(L\) of \(E\).
Near the point, the local representative is \((c \circ \varphi ^{-1}) \circ \iota _S\), where \(\varphi \) is the slice chart and \(c\) the preferred chart of \(M\). Take \(L\) the derivative of \(c \circ \varphi ^{-1}\); its inverse is the derivative of \(\varphi \circ c^{-1}\) by Lemma 7.15.
The inclusion of an embedded submanifold is an immersion; hence it exhibits \(X\) as an immersed submanifold.
With \(L\) as in Lemma 7.27, the range \(L(S) = (L^{-1})^{-1}(S)\) is closed, and \(L|_S\) is an embedding since \(L^{-1}\) is a continuous left inverse.
The inclusion of a split submanifold is a split immersion.
A continuous projection of \(E\) onto \(S\), composed with \(L^{-1}\), is a continuous left inverse of \(L|_S\).
The inclusion of an embedded submanifold exhibits it as an initial submanifold: a map into \(X\) is smooth as soon as it is smooth into \(M\).
In a slice chart \(\varphi \), the local representative of a map \(g \colon P \to X\) is the corestriction to \(S\) of the local representative of \(g\) as a map into \(M\), read in \(\varphi \). The latter is smooth, and so is the former by Lemma 7.1.
Summing up, for a subset \(X \subseteq M\) and an injective map \(f \colon M \to N\),
None of these implications can be reversed, and initial and immersed submanifolds are not comparable; see Section 7.7.
7.6 Immersions with a Banach domain
In this section \(E\) and \(F\) are normed spaces and \(E\) is complete, so that \(M\) is a Banach manifold; \(F\) need not be complete. Recall that a manifold modeled on an incomplete normed space is not a Banach manifold.
Let \(E\) be a Banach space, \(F\) a normed space, and \(f \colon M \to N\) an immersion. Let \(P\) be a manifold modeled on a locally convex space \(H\) and \(g \colon P \to M\) continuous. If \(f \circ g\) is smooth, then \(g\) is smooth.
Work in charts, with \(J\) the local representative of \(f\) at \(g(x)\). For Banach spaces, Michal–Bastiani smooth maps are Fréchet smooth, so \(J'\) is continuous in norm. The differential \(J'(y_0)\) is an embedding, hence bounded below, and a bound with half the constant persists on a neighborhood \(V\) of \(y_0\). On the open set where the local representative \(\gamma \) of \(g\) lands in \(V\), \(J \circ \gamma \) is \(C^{k}\), and the calculus result LocallyConvex.Analysis.ContDiffMBOn.of_comp_of_boundedBelow of [ LCA ] gives that \(\gamma \) is \(C^{k}\): the derivative of \(\gamma \) is obtained as the limit of difference quotients, which are Cauchy by the lower bound.
Let \(E\) be a Banach space and \(F\) a normed space. An injective immersion \(f \colon M \to N\) which is a homeomorphism onto its image exhibits its image as an initial submanifold.
If \(f \circ g\) is smooth, it is continuous, so \(g\) is continuous because \(f\) is a topological embedding. Then \(g\) is smooth by Theorem 7.31.
No inverse function theorem is used, and no complement of the range of the differential is needed. On the other hand, the image of a non-split immersion need not have slice charts: whether such an immersion, when it is a homeomorphism onto its image, is an embedded submanifold is open. Completeness of \(E\) is used, to produce the derivative of the lift as a limit of Cauchy difference quotients; completeness of \(F\) is not needed. For Fréchet or more general locally convex model spaces the statement is open, since the lower bound of the differential need not persist near a point. Continuity of \(g\) cannot be dropped: the figure eight (Remark 7.37) is an injective immersion that is not initial, because a smooth curve in the plane can lift discontinuously.
7.7 Subtleties in infinite dimensions
We collect the phenomena that separate the notions above. For each we say whether it is formalized (in Chapter 18) or only discussed.
The differential can be injective without having closed range, already between Banach spaces: the inclusion \(\ell ^1 \to \ell ^2\) is continuous, linear and injective, and its range is dense and not closed. Hence it is a weak immersion and not an immersion. This is formalized (not_isImmersion_lpInclusion, Chapter 18).
A closed subspace need not be complemented, already in Banach spaces: by Phillips’ lemma [ Phi40 ] , \(c_0\) has no complement in \(\ell ^\infty \). So \(c_0 \subseteq \ell ^\infty \) is an embedded submanifold that is not split, and the inclusion is an immersion that is not a split immersion (compare Proposition 7.21). This is formalized, with Whitley’s proof [ Whi66 ] (Phillips.not_isSplitImmersion_c0, Chapter 18).
The universal property of Definition 7.24 alone does not give an injective differential. By Joris’ theorem [ Jor82 ] (\(g^2\) and \(g^3\) smooth imply \(g\) smooth), the cusp \(t \mapsto (t^2, t^3)\) has it, and its differential vanishes at \(0\). This is why injectivity of the differential is part of the definition [ GN26 , p. 336 ] ; Karshon, Miyamoto and Watts call the universal property alone a diffeological submanifold [ KMW24 ] .
Initial does not mean embedded: the dense line in the torus is initial (not formalized). Embedded does not follow from injective immersion either. The figure eight \(t \mapsto (\sin t, \sin 2t)\), \(t \in (-\pi , \pi )\), is an injectively immersed curve that is not initial: the curve \(s \mapsto (\sin s, -\sin 2s)\) is smooth in \(\mathbb {R}^2\), stays in the figure eight, and crosses the origin along the other branch, so read in the parameter it jumps from \(\pi \) to \(0\) to \(-\pi \). This is formalized, with the parameter \(\theta = 2\arctan t\) (Lemniscate.isImmersedSubmanifold_lemniscate and Lemniscate.not_isInitialSubmanifold_lemniscate, Chapter 18).
For an injective weak immersion, smooth maps into the ambient manifold that land in the image may fail to lift:
continuously, because two pieces of the image accumulate on each other (the figure eight; this occurs in finite dimensions and is formalized);
continuously, because there are uncountably many pieces (\(\mathbb {R}\times \mathbb {R}_{\mathrm{discrete}} \to \mathbb {R}^2\)); this is excluded for second countable domains, which is why Glöckner and Neeb assume separability in [ GN26 , Thm. 9.2.15 ] (not formalized);
continuously, because the map is not even locally a topological embedding (the linear inclusion \(\ell ^1 \to \ell ^2\); infinite dimensions only; that it is not an immersion is formalized, the failure of initiality is only discussed);
continuously, but not smoothly, even when the map is a homeomorphism onto its image (infinite dimensions only; not formalized). Let \(A\) act diagonally on the \(n\)-th block of \(\ell ^2(\mathbb {N}, \mathbb {R}^2)\) by a scalar \(a_n \le e^{-n}\) and let \(j \colon \ell ^2 \to \ell ^2 \times \mathbb {R}\), \(j(x) = (Ax, \| x\| ^2)\). Its differential \(v \mapsto (Av, 2\langle x, v\rangle )\) is injective, with non-closed range at \(0\), and \(j\) is a homeomorphism onto its image (weak convergence plus convergence of norms). The curve \(y(t) = \bigoplus _n n^{-2}(\cos nt, \sin nt)\) has constant norm and \(A \circ y\) is smooth, so \(j \circ y\) is smooth; but \(y\) is \(C^{1}\) and not \(C^{2}\) in \(\ell ^2\). So \(j\) is topologically embedded, weakly immersed and not initial. In finite dimensions a weak immersion whose lifts are continuous is initial, so this phenomenon has no finite-dimensional analogue. Compare Corollary 7.32, where the closed range of the differential rules it out.
Let \(A = i\sum _n \lambda _n p_n\) be a skew-adjoint operator on a Hilbert space \(\mathcal{H}\), with distinct real \(\lambda _n\) and pairwise orthogonal projections \(p_n\) (of any rank) summing to \(1\). Its stabilizer in \(U(\mathcal{H})\) is the block-diagonal group \(K = \bigoplus _n U(p_n\mathcal{H})\), a split Lie subgroup via the conditional expectation \(X \mapsto \sum _n p_n X p_n\). The orbit map \(U(\mathcal{H})/K \to B(\mathcal{H})\), \(uK \mapsto uAu^*\), is a weak immersion with differential \(X \mapsto [X, A]\), and it is initial if and only if there are finitely many \(p_n\). If there are finitely many \(p_n\), the spectrum is finite and the orbit is a closed embedded split submanifold [ BL23 ] ; with two projections of infinite rank it is a Grassmannian. If there are infinitely many \(p_n\), the spectrum is infinite, so \([\, \cdot \, , A]\) does not have closed range [ BL23 ] , and the orbit is not initial. Indeed, for orbits of Banach–Lie group actions with split stabilizer, initiality forces closed range: along \(s \mapsto \exp (s\xi )\cdot A\) every derivative is bounded by a constant times \(\| [\xi , A]\| \), and bump curves along directions with \(\| [\xi , A]\| \) tiny are smooth in \(B(\mathcal{H})\) while their lifts to \(U(\mathcal{H})/K\) are continuous but not \(C^{1}\).
Let \(A\) be a compact injective diagonal operator on \(\ell ^2\) and
with Hilbert tensor products. Then \(j\) is injective, its differential \(v \mapsto (Av, x \otimes v + v \otimes x, \dots )\) is injective, and at \(0\) its range \(A(\ell ^2) \times 0 \times 0\) is not closed. Nevertheless \(j\) is a homeomorphism onto its image, the image is closed, and \(j\) is initial with respect to all test manifolds modeled on Fréchet spaces. The argument, which is only discussed, runs as follows. Let \(f \colon P \to \ell ^2\) with \(j \circ f\) smooth.
Along a smooth curve \(c\), each coordinate \(\langle f \circ c, e\rangle \) has smooth square and cube, namely \(\langle (f \circ c) \otimes (f \circ c), e \otimes e\rangle \) and its cubic analogue; so it is smooth by Joris’ theorem [ Jor82 ] . A curve in a Banach space all of whose scalar components are smooth is smooth, by uniform boundedness and Taylor’s formula.
Smoothness along curves gives Michal–Bastiani smoothness when \(P\) is modeled on a Fréchet space. First, the directional derivatives \(\mathrm{d}^k f(p)(v, \dots , v)\) are locally bounded: otherwise there are \(p_m \to p_0\) and small \(w_m\) with \(\| \mathrm{d}^k f(p_m)(w_m, \dots , w_m)\| {\gt} m\). With \(t_m = 2^{-m}\), \(r_m = 2^{-m-2}\) and a bump \(\beta \) supported in \([-1,1]\), equal to \(1\) on \([-\tfrac 12, \tfrac 12]\), the curve \(c(t) = p_0 + \sum _m \beta ((t - t_m)/r_m)\, (p_m - p_0 + (t - t_m)w_m)\) is smooth if \(p_m - p_0\) and \(w_m\) decay fast enough in every seminorm. Near \(t_m\) it is affine, \(c(t) = p_m + (t - t_m)w_m\), so \((f \circ c)^{(k)}(t_m) = \mathrm{d}^k f(p_m)(w_m, \dots , w_m)\) is unbounded, a contradiction. Local bounds then give multilinearity (Boman’s theorem on finite-dimensional subspaces) and joint continuity of the derivatives (mean value inequality).
So for this \(j\) the closed-range condition fails at \(0\) although \(j\) is initial, and initial and immersed submanifolds are not comparable. Whether \(j\) is initial with respect to test manifolds modeled on non-metrizable spaces is open: the argument above uses sequences.
Glöckner and Neeb call an initial Lie subgroup strictly initial when the induced map of Lie algebras is injective [ GN26 , Def. 9.1.5 ] , a condition that Neeb includes in the definition [ Nee06 ] . Not every subgroup carries an initial Lie subgroup structure: \(\ell ^\infty \subseteq \mathbb {R}^{\mathbb {N}}\) does not [ GN26 , Ex. 9.1.8 ] .