Bundle Homomorphisms, Subbundles, and Fiber Bundles

Bundle Homomorphisms Section Maps and the Characterization Lemma Subbundles Kernel and Image Subbundles Fiber Bundles

Bundle Homomorphisms

A vector bundle is a manifold equipped with a fiber structure, so the natural maps between vector bundles must respect both manifold structure (continuity or smoothness) and fiber structure (linearity, base point assignment). The result is the class of bundle homomorphisms — maps that descend to a map of base spaces and act linearly on each fiber. These are the morphisms in the category of vector bundles, and they are the right tool for discussing subbundles, quotient bundles, and the differential of a smooth map.

Definition: Bundle Homomorphism

Let \(\pi : E \to M\) and \(\pi' : E' \to M'\) be vector bundles. A bundle homomorphism from \(E\) to \(E'\) is a continuous map \(F : E \to E'\) for which there exists a map \(f : M \to M'\) satisfying \(\pi' \circ F = f \circ \pi\), with the property that the restriction \(F|_{E_p} : E_p \to E'_{f(p)}\) is linear for each \(p \in M\). One says that \(F\) covers \(f\), and that \(f\) is the base map of \(F\). When the bundles and \(F\) are smooth, \(F\) is called a smooth bundle homomorphism.

The base map \(f\) is uniquely determined by \(F\): the relation \(\pi' \circ F = f \circ \pi\) forces \(f(\pi(v)) = \pi'(F(v))\) for every \(v \in E\), and the surjectivity of \(\pi\) means this determines \(f\) at every point of \(M\). The next proposition records that the regularity of \(f\) is automatic.

Proposition: The Base Map Is Determined and Inherits Regularity

Let \(F : E \to E'\) be a bundle homomorphism covering \(f : M \to M'\). Then \(f\) is continuous and uniquely determined by \(F\). If the bundles and \(F\) are smooth, \(f\) is smooth.

Proof:

Uniqueness was noted in the discussion above. For regularity, observe that \(f = \pi' \circ F \circ \zeta\), where \(\zeta : M \to E\) is the zero section of \(E\). Indeed, \(\zeta(p) \in E_p\), so \(F(\zeta(p)) \in E'_{f(p)}\), and applying \(\pi'\) gives \(f(p)\). Continuity of \(f\) follows because \(\zeta\), \(F\), and \(\pi'\) are continuous; smoothness follows because in the smooth case all three are smooth (the zero section by an earlier result, \(F\) by hypothesis, \(\pi'\) by the bundle structure).

Isomorphisms

A bijective bundle homomorphism whose inverse is also a bundle homomorphism is called a bundle isomorphism; if \(F\) is also a diffeomorphism it is a smooth bundle isomorphism, and the bundles are said to be (smoothly) isomorphic. In the smooth setting, one need not verify smoothness of the inverse separately: bijectivity plus smoothness of \(F\) is enough.

Definition: Bundle Isomorphism

A bundle isomorphism is a bijective bundle homomorphism whose inverse is also a bundle homomorphism. A smooth bundle isomorphism is a bundle isomorphism that is a diffeomorphism between smooth bundles. Two bundles \(E, E'\) are (smoothly) isomorphic if there exists a (smooth) bundle isomorphism between them.

Proposition: Bijective Smooth Bundle Homomorphisms Are Isomorphisms

Let \(E\) and \(E'\) be smooth vector bundles over a smooth manifold \(M\) with or without boundary, and suppose \(F : E \to E'\) is a bijective smooth bundle homomorphism over \(M\). Then \(F\) is a smooth bundle isomorphism.

Proof:

It is enough to show that \(F^{-1}\) is smooth; bijectivity and the bundle-homomorphism property of \(F^{-1}\) (covering \(\mathrm{Id}_M\), fiberwise linear as the inverse of a linear isomorphism) are immediate from the hypotheses on \(F\). Smoothness is a local property, so it suffices to verify it on a neighbourhood of each point of \(E'\). Fix \(p \in M\); choose smooth local trivializations \(\Phi : \pi^{-1}(U) \to U \times \mathbb{R}^k\) of \(E\) and \(\Phi' : (\pi')^{-1}(U) \to U \times \mathbb{R}^k\) of \(E'\) on a common open neighbourhood \(U\) of \(p\) (such a common \(U\) exists because both bundles are trivializable in a neighbourhood of \(p\); intersect the two trivializing sets).

The composite \(\Phi' \circ F \circ \Phi^{-1} : U \times \mathbb{R}^k \to U \times \mathbb{R}^k\) is a smooth map (because \(\Phi^{-1}\), \(F\), and \(\Phi'\) are smooth) that covers \(\mathrm{Id}_U\) and is fiberwise linear, so it has the form \[ \Phi' \circ F \circ \Phi^{-1}(q, v) = \bigl(q,\, A(q) v\bigr) \] for a smooth map \(A : U \to GL(k, \mathbb{R})\) (smoothness of \(A\) follows from smoothness of the composite by evaluating on standard basis vectors, as in the transition function lemma; the fibrewise restriction is invertible because \(F\) is bijective, so the values lie in \(GL(k, \mathbb{R})\)). Because matrix inversion is a smooth map \(GL(k, \mathbb{R}) \to GL(k, \mathbb{R})\), the map \(q \mapsto A(q)^{-1}\) is smooth, and the composite \(\Phi \circ F^{-1} \circ (\Phi')^{-1}(q, w) = (q, A(q)^{-1} w)\) is smooth on \(U \times \mathbb{R}^k\). Hence \(F^{-1}\) is smooth on \((\pi')^{-1}(U)\), and since every point of \(E'\) lies above some such \(U\), \(F^{-1}\) is smooth.

The over-\(M\) case

When both bundles have the same base \(M\), one usually restricts attention to bundle homomorphisms covering the identity. These are the morphisms in the category of vector bundles over a fixed base.

Definition: Bundle Homomorphism over M

Let \(\pi : E \to M\) and \(\pi' : E' \to M\) be vector bundles over a common base \(M\). A bundle homomorphism over \(M\) is a bundle homomorphism \(F : E \to E'\) covering the identity \(\mathrm{Id}_M\); equivalently, \(\pi' \circ F = \pi\), and the restriction \(F|_{E_p} : E_p \to E'_p\) is linear for each \(p \in M\). When the bundles and \(F\) are smooth, \(F\) is a smooth bundle homomorphism over \(M\).

Examples

Proposition: Smoothly Trivial Bundles and the Product Bundle

A smooth rank-\(k\) vector bundle over \(M\) is smoothly trivial if and only if it is smoothly isomorphic over \(M\) to the product bundle \(M \times \mathbb{R}^k\).

Proof:

A smooth global trivialization \(\Phi : E \to M \times \mathbb{R}^k\) is, by definition of the trivialization data, a smooth bundle homomorphism over \(M\) (it satisfies \(\pi_M \circ \Phi = \pi\) and restricts to a vector space isomorphism on each fiber) that is also a diffeomorphism, hence a smooth bundle isomorphism. Conversely, a smooth bundle isomorphism \(F : E \to M \times \mathbb{R}^k\) over \(M\) is a diffeomorphism satisfying the conditions of a smooth global trivialization, so \(E\) is smoothly trivial.

Proposition: The Global Differential as a Bundle Homomorphism

Let \(F : M \to N\) be a smooth map between smooth manifolds with or without boundary. The global differential \(dF : TM \to TN\), defined by \(dF(v) = dF_p(v)\) for \(v \in T_pM\), is a smooth bundle homomorphism covering \(F\).

Proof:

Smoothness of \(dF\) was established when the tangent bundle was constructed: in any pair of smooth charts, the coordinate representation of \(dF\) is the Jacobian matrix of \(F\), itself smooth. The base relation \(\pi_N \circ dF = F \circ \pi_M\) holds because \(dF\) sends \(T_pM\) to \(T_{F(p)}N\), and the restriction \(dF_p : T_pM \to T_{F(p)}N\) is a linear map by construction.

For an immersed or embedded submanifold \(S \subseteq M\), the inclusion of the restricted bundle \(E|_S \hookrightarrow E\) is a smooth bundle homomorphism covering the inclusion \(S \hookrightarrow M\): its restriction to each fiber \(E_p\) (\(p \in S\)) is the identity, which is linear, and smoothness follows because \(S \hookrightarrow M\) is smooth and the local trivializations of \(E|_S\) are obtained by restricting those of \(E\).

Section Maps and the Characterization Lemma

A bundle homomorphism over \(M\) induces an operation on sections: composing a section of the domain bundle with the homomorphism produces a section of the codomain bundle. The resulting map between section spaces is not merely \(\mathbb{R}\)-linear but linear over \(C^\infty(M)\) — multiplication by a base function commutes with the operation. This stronger linearity property is, in fact, not just a consequence of the bundle-homomorphism origin but a complete characterization: every \(C^\infty(M)\)-linear map between the spaces of smooth sections arises from a bundle homomorphism. This characterization is the algebraic content of the entire section calculus: working with bundle homomorphisms and working with \(C^\infty(M)\)-linear maps on section spaces are two views of the same object.

The induced section map

Let \(F : E \to E'\) be a smooth bundle homomorphism over a smooth manifold \(M\). For any smooth section \(\sigma \in \Gamma(E)\), the composite \[ \widetilde F(\sigma)(p) = F(\sigma(p)) , \qquad p \in M , \] defines a continuous map \(\widetilde F(\sigma) : M \to E'\) satisfying \(\pi' \circ \widetilde F(\sigma)(p) = \pi'(F(\sigma(p))) = \pi(\sigma(p)) = p\), so it is a section of \(E'\); smoothness follows from smoothness of \(F\) and \(\sigma\), so \(\widetilde F(\sigma) \in \Gamma(E')\). The assignment \(\sigma \mapsto \widetilde F(\sigma)\) is the section map induced by \(F\).

The fiberwise linearity of \(F\) makes the section map \(\mathbb{R}\)-linear; pointwise multiplication by smooth functions on \(M\) commutes with \(F\) (because at each \(p\) it is multiplication by the scalar \(u(p)\), which commutes with the linear map \(F|_{E_p}\)), giving the stronger property below.

Definition: Linearity over \(C^\infty(M)\)

A map \(\mathcal{F} : \Gamma(E) \to \Gamma(E')\) between spaces of smooth sections of vector bundles over \(M\) is called linear over \(C^\infty(M)\) if for all \(u_1, u_2 \in C^\infty(M)\) and \(\sigma_1, \sigma_2 \in \Gamma(E)\), \[ \mathcal{F}(u_1 \sigma_1 + u_2 \sigma_2) = u_1 \mathcal{F}(\sigma_1) + u_2 \mathcal{F}(\sigma_2) . \]

Linearity over \(C^\infty(M)\) is strictly stronger than \(\mathbb{R}\)-linearity: the latter allows only constant scalars, the former allows scalars that vary smoothly across the base. The distinction is what cuts down general linear maps between section spaces to those induced by bundle homomorphisms.

The characterization lemma

Lemma: Bundle Homomorphism Characterization

Let \(\pi : E \to M\) and \(\pi' : E' \to M\) be smooth vector bundles over a smooth manifold \(M\) with or without boundary, and let \(\mathcal{F} : \Gamma(E) \to \Gamma(E')\) be a map between their spaces of smooth sections. Then \(\mathcal{F}\) is linear over \(C^\infty(M)\) if and only if there exists a smooth bundle homomorphism \(F : E \to E'\) over \(M\) such that \(\mathcal{F}(\sigma) = F \circ \sigma\) for all \(\sigma \in \Gamma(E)\). When such an \(F\) exists, it is uniquely determined by \(\mathcal{F}\).

Proof:

The forward direction was noted above: if \(\mathcal{F} = \widetilde F\) is the section map induced by a smooth bundle homomorphism \(F\) over \(M\), then \(\mathcal{F}\) is linear over \(C^\infty(M)\). The substance of the lemma is the converse, which we prove now. Suppose \(\mathcal{F} : \Gamma(E) \to \Gamma(E')\) is linear over \(C^\infty(M)\); we construct a smooth bundle homomorphism \(F : E \to E'\) inducing it.

Step 1: \(\mathcal{F}\) acts locally. We show that if \(\sigma_1, \sigma_2 \in \Gamma(E)\) agree on some open set \(U \subseteq M\), then \(\mathcal{F}(\sigma_1)\) and \(\mathcal{F}(\sigma_2)\) agree on \(U\). By linearity it suffices to show that if \(\tau \in \Gamma(E)\) vanishes on \(U\), then \(\mathcal{F}(\tau)\) vanishes on \(U\). Fix \(p \in U\) and choose a smooth bump function \(\psi \in C^\infty(M)\) supported in \(U\) with \(\psi(p) = 1\). Then \(\psi \tau \equiv 0\) on \(M\) (because \(\tau\) vanishes on the support of \(\psi\)), so by linearity \[ 0 = \mathcal{F}(\psi \tau) = \psi \, \mathcal{F}(\tau) , \] and evaluating at \(p\) gives \(0 = \psi(p) \, \mathcal{F}(\tau)(p) = \mathcal{F}(\tau)(p)\). Since this holds for every \(p \in U\), \(\mathcal{F}(\tau)\) vanishes on \(U\).

Step 2: \(\mathcal{F}\) acts pointwise. We show that if \(\sigma_1, \sigma_2 \in \Gamma(E)\) agree at a single point \(p\), then \(\mathcal{F}(\sigma_1)(p) = \mathcal{F}(\sigma_2)(p)\). Again it suffices to show that if \(\tau \in \Gamma(E)\) satisfies \(\tau(p) = 0\), then \(\mathcal{F}(\tau)(p) = 0\). Choose a smooth local frame \((\sigma_1, \dots, \sigma_k)\) for \(E\) on some neighbourhood \(W\) of \(p\), and write \(\tau = u^i \sigma_i\) on \(W\) for some smooth functions \(u^i \in C^\infty(W)\). The condition \(\tau(p) = 0\) gives \(u^1(p) = \dots = u^k(p) = 0\). By the extension lemma for vector bundles applied to each \(\sigma_i\) (and the extension lemma for smooth functions applied to each \(u^i\)), there exist global smooth sections \(\widetilde \sigma_i \in \Gamma(E)\) and global smooth functions \(\widetilde u^i \in C^\infty(M)\) that agree respectively with \(\sigma_i\) and \(u^i\) on some possibly smaller neighbourhood of \(p\); shrinking that neighbourhood if necessary, we may assume \(\tau = \widetilde u^i \widetilde \sigma_i\) holds on the neighbourhood by Step 1. Then by Step 1 again and linearity over \(C^\infty(M)\), \[ \mathcal{F}(\tau)(p) = \mathcal{F}(\widetilde u^i \widetilde \sigma_i)(p) = \widetilde u^i(p) \, \mathcal{F}(\widetilde \sigma_i)(p) = u^i(p) \, \mathcal{F}(\widetilde \sigma_i)(p) = 0 . \]

Step 3: defining \(F\). Steps 1 and 2 show that the value \(\mathcal{F}(\sigma)(p)\) depends only on the value \(\sigma(p) \in E_p\), not on the global section \(\sigma\) chosen to produce it. By every element of \(E\) is on a smooth global section, for each \(v \in E\) with \(p = \pi(v)\) there exists \(\widetilde v \in \Gamma(E)\) with \(\widetilde v(p) = v\). Define \[ F(v) = \mathcal{F}(\widetilde v)(p) \in E'_p ; \] Step 2 shows this is independent of the choice of \(\widetilde v\). The relation \(\pi' \circ F = \pi\) is built in (\(F(v) \in E'_p\) where \(p = \pi(v)\)), and fiberwise linearity follows from the \(\mathbb{R}\)-linearity of \(\mathcal{F}\): for \(v, w \in E_p\) and \(a, b \in \mathbb{R}\), choose \(\widetilde v, \widetilde w \in \Gamma(E)\) with the prescribed values at \(p\); then \(a \widetilde v + b \widetilde w \in \Gamma(E)\) has value \(av + bw\) at \(p\), and \(F(av + bw) = \mathcal{F}(a \widetilde v + b \widetilde w)(p) = a \mathcal{F}(\widetilde v)(p) + b \mathcal{F}(\widetilde w)(p) = aF(v) + bF(w)\). The identity \(\mathcal{F}(\sigma) = F \circ \sigma\) holds by construction.

Step 4: smoothness of \(F\). Smoothness is a local property; we show \(F\) is smooth on a neighbourhood of each point \(v_0 \in E\). Let \(p = \pi(v_0)\), and choose a smooth local frame \((\sigma_i)\) for \(E\) on a neighbourhood \(W\) of \(p\) and a smooth local frame \((\sigma'_j)\) for \(E'\) on a (possibly smaller) common neighbourhood; using the extension lemma, replace \((\sigma_i)\) by global smooth sections \((\widetilde \sigma_i)\) agreeing with \((\sigma_i)\) on some neighbourhood \(U \subseteq W\) of \(p\), and analogously for the \(\sigma'_j\). On \(U\), Step 1 gives \(\mathcal{F}(\widetilde \sigma_i)|_U = \mathcal{F}(\sigma_i)|_U\), and there exist smooth functions \(A_i^j \in C^\infty(U)\) such that \(\mathcal{F}(\widetilde \sigma_i)|_U = A_i^j \sigma'_j\), since the \((\sigma'_j)\) form a frame for \(E'\) on \(U\) and any smooth section there is uniquely expressible in this form by the local frame criterion for smoothness.

For any \(q \in U\) and \(v = v^i \sigma_i(q) \in E_q\), the global section \(v^i \widetilde \sigma_i \in \Gamma(E)\) has value \(v\) at \(q\), so by construction \[ F(v) = \mathcal{F}(v^i \widetilde \sigma_i)(q) = v^i \mathcal{F}(\widetilde \sigma_i)(q) = v^i A_i^j(q) \sigma'_j(q) \in E'_q . \] Let \(\Phi : \pi^{-1}(U) \to U \times \mathbb{R}^k\) and \(\Phi' : (\pi')^{-1}(U) \to U \times \mathbb{R}^m\) be the smooth local trivializations associated with the frames \((\sigma_i)\) and \((\sigma'_j)\). With respect to these trivializations, \[ \Phi' \circ F \circ \Phi^{-1}\bigl(q, (v^1, \dots, v^k)\bigr) = \bigl(q,\, (A_i^1(q) v^i, \dots, A_i^m(q) v^i)\bigr) , \] which is smooth in \((q, v)\) because the \(A_i^j\) are smooth in \(q\). Since \(\Phi, \Phi'\) are diffeomorphisms, \(F\) is smooth on \(\pi^{-1}(U)\). This holds in a neighbourhood of every point, so \(F\) is smooth on \(E\).

Uniqueness. If \(F_1, F_2 : E \to E'\) are smooth bundle homomorphisms over \(M\) with \(F_1 \circ \sigma = F_2 \circ \sigma\) for every \(\sigma \in \Gamma(E)\), then for each \(v \in E\) with \(p = \pi(v)\), choosing \(\sigma \in \Gamma(E)\) with \(\sigma(p) = v\) gives \(F_1(v) = F_1(\sigma(p)) = F_2(\sigma(p)) = F_2(v)\). Hence \(F_1 = F_2\).

Why \(C^\infty(M)\)-linearity matters

The characterization lemma is the algebraic backbone of the entire theory of tensor fields and related "tensorial" geometric structures. A Riemannian metric on \(M\) is defined as a \(C^\infty(M)\)-bilinear map \(g : \Gamma(TM) \times \Gamma(TM) \to C^\infty(M)\), and the lemma's content is that this algebraic definition automatically descends to a fiber-by-fiber inner product on each \(T_pM\) (a bundle homomorphism \(TM \otimes TM \to \mathbb{R}\)) — no separate verification needed. The same applies to tensor fields, almost complex structures, and symplectic forms when defined via section calculus: \(C^\infty(M)\)-linearity is exactly the algebraic condition that distinguishes such "tensorial" objects from operators that depend on the global behaviour of sections (such as differentiation, which is \(\mathbb{R}\)-linear but not \(C^\infty(M)\)-linear because of the Leibniz rule). Connections fall on the non-tensorial side of this divide — they satisfy a Leibniz rule and are not bundle homomorphisms — and are organised by a separate algebraic framework. The lemma thus separates the tensorial from the differential, and is used implicitly whenever a geometric structure is defined by a multilinear map on section spaces.

A note on notation: by virtue of the characterization, the bundle homomorphism \(F\) and its induced section map \(\widetilde F\) are often denoted by the same letter \(F\), with context disambiguating pointwise application from action on sections. The scalar multiplication \(X \mapsto aX\) on vector fields, for example, comes from the bundle homomorphism \(TM \to TM\) that multiplies each tangent vector by the constant \(a\), and the same symbol is used for both. Differential operators that involve derivation — most prominently the Lie derivative \(\mathscr{L}_X : \mathfrak{X}(M) \to \mathfrak{X}(M)\) — are not bundle homomorphisms because they are \(\mathbb{R}\)-linear but not \(C^\infty(M)\)-linear (the Leibniz rule introduces a derivative of \(u\) when one multiplies by \(u \in C^\infty(M)\)).

Subbundles

A subbundle is to a vector bundle what a subspace is to a vector space: a sub-collection of the fibers that fits together coherently. The definition packages two conditions — being a vector bundle in its own right, and sitting inside the ambient bundle as a topological vector subspace. The characterization lemma of the previous section produces subbundles from bundle homomorphisms (kernels and images, treated below); the local-frame criterion proved here produces subbundles from local-section data.

Definition: Subbundle

Given a vector bundle \(\pi_E : E \to M\), a subbundle of \(E\) is a vector bundle \(\pi_D : D \to M\) in which \(D\) is a topological vector subspace of \(E\) and \(\pi_D = \pi_E|_D\) is the restriction of \(\pi_E\) to \(D\), such that for each \(p \in M\) the subset \(D_p = D \cap E_p\) is a linear subspace of \(E_p\), with the vector space structure on \(D_p\) inherited from \(E_p\). The condition that \(D\) be a vector bundle over \(M\) forces all fibers \(D_p\) to be nonempty and of the same dimension.

Definition: Smooth Subbundle

If \(\pi_E : E \to M\) is a smooth vector bundle, a subbundle \(D \subseteq E\) is called a smooth subbundle if it is itself a smooth vector bundle and \(D\) is an embedded submanifold (with or without boundary) of \(E\).

Proposition: The Inclusion of a Smooth Subbundle Is a Bundle Homomorphism

Let \(\pi : E \to M\) be a smooth vector bundle and let \(D \subseteq E\) be a smooth subbundle. The inclusion map \(\iota : D \hookrightarrow E\) is a smooth bundle homomorphism over \(M\).

Proof:

The smoothness of \(\iota\) follows from \(D\) being an embedded submanifold of \(E\) (the inclusion of an embedded submanifold into the ambient manifold is smooth by definition). The relation \(\pi \circ \iota = \pi_D\) (where \(\pi_D = \pi|_D\) is the projection of \(D\)) holds pointwise, exhibiting \(\iota\) as a map covering the identity \(\mathrm{Id}_M\). Fiber-wise, \(\iota|_{D_p} : D_p \hookrightarrow E_p\) is the inclusion of a linear subspace, which is linear.

The local-frame criterion

Constructing a subbundle by hand requires verifying that a candidate union \(D = \bigcup_p D_p\) of linear subspaces is itself a vector bundle in the right sense — a topology must be built, charts must be checked, and the bundle structure must match the ambient one. The next criterion reduces this entire verification to producing locally a smooth basis: if smooth local sections spanning each \(D_p\) exist on a neighbourhood of every base point, the union is automatically a smooth subbundle.

Lemma: Local Frame Criterion for Subbundles

Let \(\pi : E \to M\) be a smooth vector bundle, and suppose that for each \(p \in M\) we are given an \(m\)-dimensional linear subspace \(D_p \subseteq E_p\). Then \(D = \bigcup_{p \in M} D_p \subseteq E\) is a smooth subbundle of \(E\) of rank \(m\) if and only if the following condition holds:

Each point of \(M\) has a neighbourhood \(U\) on which there exist smooth local sections \(\sigma_1, \dots, \sigma_m : U \to E\) of the ambient bundle such that \((\sigma_1(q), \dots, \sigma_m(q))\) is a basis of \(D_q\) for every \(q \in U\).

Proof:

Necessity. Suppose \(D\) is a smooth subbundle of \(E\). For each \(p \in M\), \(D\) admits a smooth local trivialization over some neighbourhood \(U\) of \(p\), and the frame associated with this trivialization produces smooth local sections \(\tau_1, \dots, \tau_m : U \to D\) whose values form a basis of \(D_q\) at each \(q \in U\). Composing with the inclusion \(\iota : D \hookrightarrow E\) (a smooth bundle homomorphism by the proposition above), the sections \(\sigma_i = \iota \circ \tau_i : U \to E\) are smooth local sections of \(E\) with the same values, and their values continue to form a basis of \(D_q\) at each \(q \in U\). The local-section condition is satisfied.

Sufficiency. Suppose the local-section condition holds. We construct a smooth vector bundle structure on \(D\) and verify that \(D\) is an embedded submanifold of \(E\), which together make \(D\) a smooth subbundle. Let \(E\) have rank \(k \ge m\).

Fix \(p \in M\) and let \(U \ni p\) be a neighbourhood on which smooth local sections \(\sigma_1, \dots, \sigma_m\) of \(E\) form a basis of \(D_q\) for each \(q \in U\). By the completion of local frames (applied to \(E\) and the partial frame \((\sigma_1, \dots, \sigma_m)\), shrinking \(U\) if necessary), there exist smooth sections \(\sigma_{m+1}, \dots, \sigma_k\) on \(U\) such that \((\sigma_1, \dots, \sigma_k)\) is a smooth local frame for \(E\) over \(U\). Let \(\Phi : \pi^{-1}(U) \to U \times \mathbb{R}^k\) be the smooth local trivialization of \(E\) associated with this frame, as given by the frame ↔ trivialization correspondence: \[ \Phi\!\left(s^1 \sigma_1(q) + \dots + s^k \sigma_k(q)\right) = \bigl(q,\, (s^1, \dots, s^k)\bigr) . \]

Under \(\Phi\), the set \(D \cap \pi^{-1}(U)\) is sent to \(\{(q, (s^1, \dots, s^m, 0, \dots, 0)) : q \in U,\ s^i \in \mathbb{R}\}\), because a vector \(s^1\sigma_1(q) + \dots + s^k\sigma_k(q) \in E_q\) lies in \(D_q\) if and only if it lies in the span of \((\sigma_1(q), \dots, \sigma_m(q))\), which is to say \(s^{m+1} = \dots = s^k = 0\). The image \(U \times \mathbb{R}^m \times \{0\}^{k-m}\) is an embedded submanifold (with or without boundary) of \(U \times \mathbb{R}^k\), so \(D \cap \pi^{-1}(U)\) is an embedded submanifold of \(\pi^{-1}(U) \subseteq E\). The collection of such neighbourhoods covers \(E\), so \(D\) is an embedded submanifold of \(E\).

The map \(\Psi : D \cap \pi^{-1}(U) \to U \times \mathbb{R}^m\) defined by \[ \Psi\!\left(s^1 \sigma_1(q) + \dots + s^m \sigma_m(q)\right) = \bigl(q,\, (s^1, \dots, s^m)\bigr) \] is the composition of the embedding \(D \cap \pi^{-1}(U) \hookrightarrow U \times \mathbb{R}^m \times \{0\}^{k-m}\) (from the argument above) with projection onto the first \(m + \dim M\) coordinates, both of which are diffeomorphisms onto their images. Hence \(\Psi\) is a diffeomorphism, fiber-wise linear, and satisfies \(\pi_U \circ \Psi = \pi|_D\). Therefore \(\Psi\) is a smooth local trivialization of \(D\) over \(U\); since every point of \(M\) lies in such a \(U\), \(D\) inherits a smooth vector bundle structure of rank \(m\) over \(M\), and the inclusion \(D \hookrightarrow E\) is the inclusion of an embedded submanifold. Thus \(D\) is a smooth subbundle of \(E\).

Examples of subbundles

(a) The span of a nowhere-vanishing vector field. Let \(M\) be a smooth manifold and \(V\) a nowhere-vanishing smooth vector field on \(M\). The set \(D \subseteq TM\) whose fiber at each \(p \in M\) is the line \(\mathrm{span}(V_p) \subseteq T_pM\) is a smooth rank-1 subbundle of \(TM\): the local-section condition is satisfied with the single section \(\sigma_1 = V\) on \(U = M\).

(b) Span of part of a global frame. Suppose \(E \to M\) is a trivial smooth vector bundle with smooth global frame \((E_1, \dots, E_k)\). For \(0 \le m \le k\), the union \(D \subseteq E\) with fiber \(D_p = \mathrm{span}(E_1|_p, \dots, E_m|_p)\) is a smooth rank-\(m\) subbundle of \(E\); the local-section condition holds globally with \(\sigma_i = E_i\) for \(i = 1, \dots, m\). When \(M = G\) is a Lie group and \((E_1, \dots, E_n)\) is the basis of left-invariant vector fields corresponding to a basis of \(\mathrm{Lie}(G)\), every \(m\)-dimensional subspace of \(\mathrm{Lie}(G)\) determines a smooth rank-\(m\) left-invariant subbundle of \(TG\); these are the fundamental examples of "distributions" in the integrability theory of vector fields.

(c) The tangent bundle of a submanifold. Let \(M\) be a smooth manifold and \(S \subseteq M\) an immersed \(k\)-submanifold. The tangent bundle \(TS\) sits inside the ambient tangent bundle \(TM|_S\) as a smooth rank-\(k\) subbundle. To see this, choose a smooth chart \((V, \varphi)\) on \(S\) and a smooth chart \((W, \psi)\) on \(M\) compatible with the immersion (so that \(S \cap W\) is parametrized by the first \(k\) coordinates of \(\psi\)). The coordinate vector fields \(\partial/\partial \varphi^i|_q\) on \(S\) extend to smooth local sections of \(TM|_S\) on \(S \cap W\), and they span \(T_qS\) at each \(q\); the local-section condition holds, and the criterion gives \(TS\) the structure of a smooth subbundle.

Kernel and Image Subbundles

The most productive source of subbundles is the kernel-and-image construction for bundle homomorphisms. For a linear map between finite-dimensional vector spaces the kernel and image are immediately subspaces; for a bundle homomorphism, taking kernel and image fiberwise gives candidate subbundles of the domain and codomain. The catch is the dimension of the kernel and the rank of the image are not, in general, constant along the base — they can jump. When they do not jump, the candidates are genuine smooth subbundles; when they do, the analysis breaks down. The relevant notion is constant rank.

The rank of a bundle homomorphism

Definition: Rank of a Bundle Homomorphism

Let \(F : E \to E'\) be a bundle homomorphism over \(M\). For each \(p \in M\), the rank of \(F\) at \(p\) is the rank of the linear map \(F|_{E_p} : E_p \to E'_p\). The map \(F\) is said to have constant rank if the rank of \(F\) at \(p\) is the same for all \(p \in M\); this common value is the rank of \(F\).

The constant-rank condition is a natural strengthening of the bundle-homomorphism axioms: it ensures that the fiber-wise kernel and image structures fit together coherently across the base. Without it, fiber dimensions vary and no bundle structure can survive.

The constant-rank theorem

Theorem: Kernel and Image as Subbundles Under Constant Rank

Let \(E\) and \(E'\) be smooth vector bundles over a smooth manifold \(M\), and let \(F : E \to E'\) be a smooth bundle homomorphism over \(M\). Define \[ \mathrm{Ker}\, F = \bigcup_{p \in M} \mathrm{Ker}(F|_{E_p}) \subseteq E, \qquad \mathrm{Im}\, F = \bigcup_{p \in M} \mathrm{Im}(F|_{E_p}) \subseteq E' . \] Then \(\mathrm{Ker}\, F\) and \(\mathrm{Im}\, F\) are smooth subbundles of \(E\) and \(E'\) respectively if and only if \(F\) has constant rank.

Proof:

Necessity. If \(\mathrm{Ker}\, F\) and \(\mathrm{Im}\, F\) are smooth subbundles, their fibers have constant dimension by the definition of a vector bundle. The fiber \((\mathrm{Im}\, F)_p\) has dimension equal to the rank of \(F\) at \(p\), and the fiber \((\mathrm{Ker}\, F)_p\) has dimension \(\dim E_p - \mathrm{rank}(F|_{E_p})\); constancy of either dimension along \(M\) forces the rank of \(F\) to be constant.

Sufficiency. Suppose \(F\) has constant rank \(r\); let \(k\) and \(k'\) denote the ranks of \(E\) and \(E'\), so the fibers of \(\mathrm{Ker}\, F\) have dimension \(k - r\) and those of \(\mathrm{Im}\, F\) have dimension \(r\). We show that each of \(\mathrm{Im}\, F\) and \(\mathrm{Ker}\, F\) satisfies the local-frame criterion for subbundles.

Image. Fix \(p \in M\), and choose a smooth local frame \((\sigma_1, \dots, \sigma_k)\) for \(E\) over a neighbourhood \(U\) of \(p\). The sections \(F \circ \sigma_1, \dots, F \circ \sigma_k\) of \(E'\) over \(U\) span \(\mathrm{Im}(F|_{E_q})\) at each \(q \in U\), and exactly \(r\) of them are linearly independent at \(p\) (the rank of \(F|_{E_p}\) being \(r\)). By relabelling the indices if necessary, we may assume \((F \circ \sigma_1, \dots, F \circ \sigma_r)\) are linearly independent at \(p\). Linear independence is an open condition, so there is a possibly smaller neighbourhood \(U_0 \subseteq U\) of \(p\) on which \((F \circ \sigma_1, \dots, F \circ \sigma_r)\) remain linearly independent at every point. By the constant-rank hypothesis, the image \(\mathrm{Im}(F|_{E_q}) = (\mathrm{Im}\, F)_q\) has dimension exactly \(r\) for every \(q \in U_0\), so \((F \circ \sigma_1, \dots, F \circ \sigma_r)\) are not only linearly independent but span \((\mathrm{Im}\, F)_q\) at each point of \(U_0\). The local-frame criterion for subbundles applies, and \(\mathrm{Im}\, F\) is a smooth rank-\(r\) subbundle of \(E'\).

Kernel. Let \(U_0\) and \((\sigma_1, \dots, \sigma_r)\) be as above. The subbundle \(V \subseteq E|_{U_0}\) spanned by \((\sigma_1, \dots, \sigma_r)\) (by the local-frame criterion, taking these sections themselves as the local frame on \(U_0\)) is a smooth rank-\(r\) subbundle of \(E|_{U_0}\) complementary to \(\mathrm{Ker}\, F\) in the following sense: the restriction \(F|_V : V \to (\mathrm{Im}\, F)|_{U_0}\) is bijective on each fiber (because \((F \circ \sigma_i)\) is a basis of \((\mathrm{Im}\, F)_q\) and \((\sigma_i)\) is a basis of \(V_q\)), and it is a smooth bundle homomorphism over \(U_0\). By the bijective-bundle-homomorphism proposition, \(F|_V\) is a smooth bundle isomorphism. Let \((F|_V)^{-1} : (\mathrm{Im}\, F)|_{U_0} \to V\) be its inverse, and define \[ \Psi : E|_{U_0} \to E|_{U_0}, \qquad \Psi(v) = v - (F|_V)^{-1}\bigl(F(v)\bigr) . \] Each term is a smooth bundle homomorphism over \(U_0\) (composition with the homomorphism \(F\) from \(E|_{U_0}\) to \((\mathrm{Im}\, F)|_{U_0}\), then with \((F|_V)^{-1}\), each preserves fibers and is linear), so \(\Psi\) is itself a smooth bundle homomorphism over \(U_0\).

We verify that \(\Psi\) takes its values in \((\mathrm{Ker}\, F)|_{U_0}\) and restricts to the identity on \((\mathrm{Ker}\, F)|_{U_0}\). For any \(v \in E|_{U_0}\), \[ F(\Psi(v)) = F(v) - F\bigl((F|_V)^{-1}(F(v))\bigr) = F(v) - F(v) = 0 , \] since \((F|_V)^{-1}(F(v)) \in V\) and \(F|_V\) is the inverse of \((F|_V)^{-1}\) on \(V\). Hence \(\Psi(v) \in (\mathrm{Ker}\, F)_{q}\) for \(q = \pi(v)\). Conversely, if \(v \in (\mathrm{Ker}\, F)_q\), then \(F(v) = 0\), so \(\Psi(v) = v - (F|_V)^{-1}(0) = v\); \(\Psi\) is the identity on \((\mathrm{Ker}\, F)|_{U_0}\). These two properties together say that \(\Psi\) is a fiberwise projection of \(E|_{U_0}\) onto \(\mathrm{Ker}\, F\) along \(V\) (idempotent, \(\Psi \circ \Psi = \Psi\)).

The image of \(\Psi\) is therefore exactly \((\mathrm{Ker}\, F)|_{U_0}\). Since \(V \oplus \mathrm{Ker}\, F\) recover \(E|_{U_0}\) fiberwise (the rank-nullity theorem gives \(\dim V_q + \dim (\mathrm{Ker}\, F)_q = r + (k - r) = k = \dim E_q\)), and the splitting is smooth because \((F|_V)^{-1}\) is smooth, \((\mathrm{Ker}\, F)|_{U_0}\) is the image of a smooth bundle homomorphism with constant rank \(k - r\). Applying the image-subbundle argument already proved to \(\Psi\) shows \((\mathrm{Ker}\, F)|_{U_0}\) is a smooth rank-\((k-r)\) subbundle of \(E|_{U_0}\). The same argument can be carried out on a neighbourhood of every point of \(M\), so \(\mathrm{Ker}\, F\) is a smooth rank-\((k-r)\) subbundle of \(E\).

The orthogonal complement and the normal bundle

The constant-rank theorem produces subbundles abstractly. For the special case of subbundles of \(T\mathbb{R}^n\) restricted to a submanifold, the Euclidean inner product gives an explicit geometric construction of a complementary subbundle: the orthogonal complement, fiber by fiber.

Lemma: Orthogonal Complement Bundles

Let \(M\) be an immersed submanifold with or without boundary in \(\mathbb{R}^n\), and let \(D\) be a smooth rank-\(k\) subbundle of the ambient tangent bundle \(T\mathbb{R}^n|_M\). For each \(p \in M\), let \(D_p^\perp\) denote the orthogonal complement of \(D_p\) in \(T_p\mathbb{R}^n\) with respect to the Euclidean dot product, and let \[ D^\perp = \{(p, v) \in T\mathbb{R}^n : p \in M,\ v \in D_p^\perp\} \subseteq T\mathbb{R}^n|_M . \] Then \(D^\perp\) is a smooth rank-\((n - k)\) subbundle of \(T\mathbb{R}^n|_M\). For each \(p \in M\), there is a smooth orthonormal frame for \(D^\perp\) on a neighbourhood of \(p\).

Proof:

Fix \(p \in M\), and let \((X_1, \dots, X_k)\) be a smooth local frame for \(D\) over some neighbourhood \(V\) of \(p\) in \(M\). Because immersed submanifolds are locally embedded, we may shrink \(V\) so that it is a single slice in some coordinate ball or half-ball \(U \subseteq \mathbb{R}^n\); the closedness of \(V\) in \(U\) (after this shrinking) allows the completion of local frames to be applied: we can extend \((X_1, \dots, X_k)\) to a smooth local frame \((\widetilde X_1, \dots, \widetilde X_n)\) for \(T\mathbb{R}^n\) over \(U\). Applying the Gram-Schmidt process for frames to \((\widetilde X_1, \dots, \widetilde X_n)\), we obtain a smooth orthonormal frame \((E_1, \dots, E_n)\) for \(T\mathbb{R}^n\) over \(U\) such that \(\mathrm{span}(E_1|_p, \dots, E_k|_p) = \mathrm{span}(X_1|_p, \dots, X_k|_p) = D_p\) for every \(p \in V\) (the Gram-Schmidt process preserves the flag of nested spans).

The sections \((E_{k+1}, \dots, E_n)\) restricted to \(V\) form a smooth orthonormal frame for \(D^\perp\): at each \(p \in V\), the vectors \(E_{k+1}|_p, \dots, E_n|_p\) are orthonormal and orthogonal to every \(E_i|_p\) (\(i \le k\)), hence to all of \(D_p\), so they lie in \(D_p^\perp\); and they span an \((n-k)\)-dimensional subspace, which by dimension count is all of \(D_p^\perp\). The local-frame criterion for subbundles applies, and \(D^\perp\) is a smooth rank-\((n - k)\) subbundle of \(T\mathbb{R}^n|_M\).

The most important special case is the normal bundle of a submanifold, defined earlier as the orthogonal complement of \(TM\) inside \(T\mathbb{R}^n|_M\). Within the present framework, it is the orthogonal complement of a particular smooth subbundle.

Corollary: The Normal Bundle as a Subbundle

Let \(M \subseteq \mathbb{R}^n\) be an immersed \(m\)-dimensional submanifold with or without boundary. The normal bundle \(NM\) is a smooth rank-\((n - m)\) subbundle of \(T\mathbb{R}^n|_M\), and each point of \(M\) has a neighbourhood over which there exists a smooth orthonormal frame for \(NM\).

Proof:

The tangent bundle \(TM\) is a smooth rank-\(m\) subbundle of \(T\mathbb{R}^n|_M\) by Example (c) above. Applying the orthogonal-complement lemma to \(D = TM\) (which has rank \(m\)) gives \(D^\perp = NM\) as a smooth rank-\((n - m)\) subbundle of \(T\mathbb{R}^n|_M\), with smooth orthonormal local frames on neighbourhoods of every point.

What constant rank buys

The constant-rank theorem produces subbundles algebraically rather than geometrically: given a bundle homomorphism with the right algebraic property (rank constant), the kernel and image sets — defined fibrewise — automatically inherit smooth subbundle structure, without any further geometric construction. This is in marked contrast to the local-frame criterion, which requires explicit production of smooth sections spanning the candidate subbundle. The two routes are complementary: when one has a bundle homomorphism in hand (as in the orthogonal-complement construction below, where the homomorphism is projection along a complementary subbundle), the kernel-and-image route applies; when one has only candidate fibers and needs to manufacture local frames, the local-frame route applies. The orthogonal-complement specialization on the next page gives the normal bundle as a tangible geometric object — the same normal bundle that supplied the tubular neighbourhood theorem — recognized now as one instance of the orthogonal-complement construction inside the broader subbundle framework.

Fiber Bundles

Vector bundles are the most useful family of locally trivial fibrations, but they are not the only one. Dropping the requirement that the fibers be vector spaces, and replacing it with the requirement that the fibers be homeomorphic to a fixed model space, gives the broader notion of a fiber bundle. Many natural geometric and topological constructions that fall outside the vector-bundle framework — covering spaces, principal bundles, frame bundles, associated bundles, and Hopf fibrations among them — fit naturally into this larger picture, and several recur in later parts of the site. We introduce only the definitions and the most immediate examples here.

Definition: Fiber Bundle

Let \(M\) and \(F\) be topological spaces. A fiber bundle over \(M\) with model fiber \(F\) is a topological space \(E\) together with a surjective continuous map \(\pi : E \to M\) such that for each \(x \in M\) there exist a neighbourhood \(U\) of \(x\) in \(M\) and a homeomorphism \(\Phi : \pi^{-1}(U) \to U \times F\) satisfying \(\pi_1 \circ \Phi = \pi\) (where \(\pi_1 : U \times F \to U\) is projection on the first factor). Such a homeomorphism is a local trivialization of \(E\) over \(U\). The space \(E\) is the total space of the bundle, \(M\) is its base, and \(\pi\) is its projection.

Definition: Smooth Fiber Bundle

If \(E\), \(M\), and \(F\) are smooth manifolds with or without boundary, \(\pi\) is a smooth map, and the local trivializations can be chosen to be diffeomorphisms, then \(\pi : E \to M\) is called a smooth fiber bundle.

Definition: Trivial Fiber Bundle and Global Trivialization

A fiber bundle is trivial if it admits a local trivialization over the entire base; such a trivialization is called a global trivialization. A smooth fiber bundle is smoothly trivial if it admits a global trivialization that is a diffeomorphism.

The definition mirrors that of a vector bundle exactly, with the linear structure on each fiber replaced by mere homeomorphism with \(F\). What is lost is the algebraic structure: there are no "sections of a fiber bundle" in the linear sense (one cannot add or scale them, since fibers carry no operations), and no bundle homomorphisms induced by linear maps on fibers. What is gained is flexibility: any topological space can serve as the model fiber, and the theory accommodates constructions where the natural fibers are not linear.

Four examples of fiber bundles

(a) Product fiber bundles. For topological spaces \(M\) and \(F\), the product \(M \times F\) with projection \(\pi_1 : M \times F \to M\) onto the first factor is a fiber bundle, the product fiber bundle; its global trivialization is the identity \(M \times F \to M \times F\). Every product fiber bundle is trivial, and conversely the trivial fiber bundles are exactly the product bundles up to global trivialization.

(b) Vector bundles as fiber bundles. Every rank-\(k\) vector bundle is a fiber bundle with model fiber \(\mathbb{R}^k\). The local trivializations of the vector bundle satisfy the projection-compatibility condition of fiber-bundle trivializations, and the additional linear structure on each fiber is the extra data that distinguishes the vector-bundle case from the general one.

(c) Möbius-band fiber bundle. Let \(\pi : E \to \mathbb{S}^1\) be the Möbius bundle, with quotient map \(q : \mathbb{R}^2 \to E\). The image of \(\mathbb{R} \times [-1, 1]\) under \(q\) is a closed sub-band of \(E\), and it is a fiber bundle over \(\mathbb{S}^1\) with model fiber \([-1, 1]\): the same covering-and-quotient local trivializations restrict to give homeomorphisms with \(U \times [-1, 1]\) on each evenly covered open arc. This fiber bundle is not trivial — the same connectedness argument that distinguished the Möbius bundle from the trivial line bundle applies: removing the central circle \(q(\mathbb{R} \times \{0\})\) leaves a connected complement, whereas the corresponding complement in the trivial bundle \(\mathbb{S}^1 \times [-1, 1]\) has two components.

(d) Covering maps as fiber bundles. Every smooth covering map \(\pi : E \to M\) is a smooth fiber bundle whose model fiber is a discrete space of the appropriate cardinality (the number of sheets). For each evenly covered open subset \(U \subseteq M\) with components of \(\pi^{-1}(U)\) labelled by a discrete set \(S\), the homeomorphism \(\Phi : \pi^{-1}(U) \to U \times S\) sending each point to its base image and its component label is a local trivialization. The connection between covering theory and bundle theory recovers, in the discrete-fiber case, results already established for covering maps earlier in the manifold series.

The broader picture

Several important classes of fiber bundles arise repeatedly in differential geometry without fitting the vector-bundle framework. Principal \(G\)-bundles are fiber bundles whose model fiber is a Lie group \(G\), with \(G\) acting freely and transitively on each fiber by right multiplication; they are the natural setting for connections and gauge theories. Associated bundles are built from a principal \(G\)-bundle and a representation of \(G\) on a model fiber \(F\), producing fiber bundles whose model fiber is \(F\) and whose structure group is \(G\); a vector bundle is the special case where the representation is on \(\mathbb{R}^k\) (or \(\mathbb{C}^k\)). The frame bundle of a vector bundle is a principal \(GL(k, \mathbb{R})\)-bundle whose fiber at each point is the set of bases of the vector bundle's fiber there. The Hopf fibration \(\mathbb{S}^3 \to \mathbb{S}^2\) is a non-trivial fiber bundle with model fiber \(\mathbb{S}^1\), one of the first historical examples of a non-trivial fiber bundle with non-discrete fibers.

The end of the manifold series

The manifold series began with the question of how to attach calculus to spaces more general than \(\mathbb{R}^n\). The answer was a hierarchy: topological manifolds, smooth manifolds, tangent vectors, vector fields, flows, Lie groups and their algebras, submanifolds with their tangent and normal bundles. The arc closes here with the recognition that all of these constructions — and many of the further structures of differential forms, Riemannian metrics, and tensor fields still to come — share a common framework: each lives on, or is a section of, a vector bundle (or, more generally, a fiber bundle) over the manifold. The bundle is the right abstraction because it captures both the local product structure that makes calculus possible and the global topology that determines what is and is not trivial. Subsequent developments — Riemannian geometry, symplectic geometry, complex geometry, gauge theory, equivariant deep learning — build on this framework, either directly through additional sections (metrics, differential forms) or by enriching it with extra structure (connections, group actions, principal bundles).