TOPOLOGY II

ARVIND NAIR

These are my notes for the second semester of the first-year topology course taught at TIFR from
January to May 2024. Sections marked (∗) were not covered (or not covered in detail).

Contents

1. Homology and cohomology of manifolds; Poincaré duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.1. (Topological) manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2. Mayer-Vietoris sequences for pairs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3. Homology in degrees ≥ n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4. Orientations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5. Fundamental class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.6. Compactly supported cohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.7. Poincaré duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.8. Duality and the cup product pairing in cohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.9. Duality between cohomology and Borel-Moore homology (∗) . . . . . . . . . . . . . . . . . . . . . . . . 10
1.10. Euler characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.11. Finite generation of homology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.12. Manifolds with boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.13. Lefschetz duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.14. Gysin homomorphisms and Poincaré dual classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.15. Cup product and intersection product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.16. Alexander duality (∗) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2. Smooth manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.1. Smooth manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2. Real analytic manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3. Complex manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4. Partitions of unity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.5. Tangent and cotangent spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.6. Functoriality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.7. Tangent and cotangent bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.8. Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.9. C k -manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.10. Inverse and implicit function theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.11. Immersions and submersions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.12. Submanifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.13. Normal forms for immersions and submersions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.14. Critical point, critical value, regular value . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.15. Transversality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.16. Whitney’s embedding theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.17. Tubular neighbourhoods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.18. Ehresmann’s fibration theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.19. Triangulation of smooth manifolds (∗) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.20. Good covers and finite generation of homology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3. Vector fields and Frobenius’ theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.1. Vector fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2. Local flows and vector fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3. A useful lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2

3.4. Lie bracket . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

3.5. Local flows and brackets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.6. Local Frobenius theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.7. Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.8. Integral manifolds and the global Frobenius theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4. Basics of Lie groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.1. Lie groups, Lie algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.2. Lie algebra of a Lie group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.3. Homomorphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.4. Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.5. Lie subgroups and Lie subalgebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.6. Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.7. Linear algebraic groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.8. Coverings of Lie groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.9. Example: SU (2), SO(3), Sp(1), quaternions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.10. Exponential map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.11. Simply connected groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.12. Example: Universal cover of SL(2, R) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.13. Closed subgroups (∗) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.14. Group actions and homogeneous manifolds (∗) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.15. Examples: Homogeneous spaces (∗) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.16. π0 and π1 of some classical groups (∗) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.17. Example: Spin(4) (∗) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.18. Hopf’s theorem and compact Lie groups (∗) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5. Differential forms and de Rham cohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.1. Vector bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.2. Pullback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.3. Presheaves and sheaves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.4. Differentials or 1-forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.5. Differential forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.6. Exterior derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.7. Lie derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.8. Frobenius theorem revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.9. Orientations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.10. Integration on oriented manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.11. Stokes’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.12. De Rham cohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.13. Poincaré lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.14. Homotopy invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.15. Mayer-Vietoris sequence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.16. De Rham theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.17. Relative de Rham cohomology (∗) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6. Sheaves and cohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
6.1. Presheaves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
6.2. Sheaves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
6.3. Stalks, sheafification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6.4. Kernels, cokernels, images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
6.5. Example: Exponential sequence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.6. Exact functors, global sections functor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.7. Pullback, pushforward . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.8. Internal Hom and tensor product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.9. Injective sheaves and injective resolutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.10. Derived functors, cohomology of sheaves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
6.11. Acyclic resolutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
 3

6.12. Flasque sheaves, canonical flasque resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

6.13. Singular cohomology and sheaf cohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.14. Fine sheaves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.15. De Rham’s theorem and integration over smooth chains (∗) . . . . . . . . . . . . . . . . . . . . . . . . 87
6.16. Functoriality (∗) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
7. Sheaves and cohomology II (∗) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
7.1. Cohomology objects, quasi-isomorphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
7.2. “Resolutions” of complexes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
7.3. Derived functors and hypercohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
7.4. Pullback maps in (hyper)cohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
7.5. Functoriality of the de Rham isomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.6. Cup products in sheaf cohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
7.7. Cup products in hypercohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
7.8. Cup products and the de Rham isomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
7.9. Composition of functors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
7.10. Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
8. Fibre bundles: Leray-Hirsch theorem, Thom isomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
8.1. Pushforward . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
8.2. Leray-Hirsch theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
8.3. Example: Homogeneous manifolds for compact Lie groups . . . . . . . . . . . . . . . . . . . . . . . . . 101
8.4. Example: Unitary groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
8.5. Example: Flag manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
8.6. Example: Grassmannians . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
8.7. Thom isomorphism theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
9. Complex manifolds and the Dolbeault theorem (∗) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
9.1. Tangent bundle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
9.2. Almost complex structures and integrability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
9.3. ¯ etc. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Differential forms: ∂, ∂,
9.4. ¯
The ∂-Poincaré lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
9.5. Dolbeault theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4

1. Homology and cohomology of manifolds; Poincaré duality

We assume the following: the basic properties of cohomology (as in the Eilenberg-Steenrod axioms),
Mayer-Vietoris sequences, the universal coefficient theorems for homology and cohomology, cross prod-
ucts and the Künneth formula, properties of the cup and cap products. All coefficient rings are assumed
to be PIDs. (1)

1.1. (Topological) manifolds.

Definition 1.1.1. A topological manifold, or just manifold for short, is a second countable Hausdorff
topological space M which is locally Euclidean, i.e. every point has a neighbourhood homeomorphic to an
open subset of Rn .

By invariance of domain (i.e. Rm homeomorphic to Rn implies n = m) the dimension of a manifold

is constant on each connected component. We will only consider manifolds all of whose connected
components are of the same dimension.
Example 1.1.2. An open subset of a manifold is a manifold. A finite product of manifolds is a manifold.
Examples of manifolds include Euclidean spaces, spheres, projective spaces (real, complex, quaternionic),
Grassmannians (real and complex), and many other spaces which occur naturally.
Example 1.1.3. Here are two examples of the kind of pathologies we want to avoid:
(1) Let M be the real line with two origins, i.e. R × {0, 1} modulo the identification of (λ, 0) and
(λ, 1) for λ 6= 0. This is clearly locally Euclidean but it is not a manifold as it is not Hausdorff. It cannot
be embedded in RN . (2)
(2) Let M = (0, 1)2 with the topology that a set U ⊂ M is open iff U ∩ ({a} × (0, 1)) is open in
{a} × (0, 1) for all a. This is locally Euclidean (every point has a neighbourhood homeomorphic to an
interval in R) but is not second countable (e.g. because it has uncountably many connected components).

The following also holds for noncompact manifolds; we only prove the compact case:
Lemma 1.1.4. A compact manifold can be embedded as a closed subset of RN for some N .

Proof. Let {(Ui , ϕi )}i=1,...,r be a covering of M by open subsets with ϕi : Ui → Rn a homeomorphism

onto an open subset of Rn . Let {fi } be a partition of unity subordinate to it. (3) Then define a mapping
F : M → Rrn+n by (f1 ϕ1 , . . . , fr ϕr , f1 , . . . , fr ), where fi ϕi has been extended by zero outside Ui (this
is continuous). It is easy to check that F is continuous and bijective, hence a homeomorphism onto its
image. 

This can be used to show that any compact manifold is a Euclidean neighbourhood retract and also
that a compact manifold has the homotopy type of a CW complex (see the Appendix of Hatcher).

1.2. Mayer-Vietoris sequences for pairs. Let X be a topological space and U, V ⊂ X open subsets.
There is a long exact M-V sequence
→ Hi (X, U ∩ V, R) → Hi (X, U, R) ⊕ Hi (X, V, R) → Hi (X, U ∪ V, R) → Hi−1 (X, U ∩ V, R) → (1.2.1)
{U,V }
This is proved by the usual argument using “small” chains. Let C∗ (X) be the complex of singular
chains spanned by singular simplices which lie in either U or V . By the “small chains argument” we
{U,V }
know that C∗ (U ∪ V ) ,→ C∗ (U ∪ V ) induces isomorphisms in homology. By the 5-lemma we get that
{(U,V )}
C∗ (X)/C∗ (U ∪ V ) −→ C∗ (X)/C∗ (U ∪ V ) induces isomorphisms in homology. Now the short exact
sequence of complexes of abelian groups
{U,V }
0 → C∗ (X, U ∩ V ) → C∗ (X, U ) ⊕ C∗ (X, V ) → C∗ (X)/C∗ (X) → 0
is termwise split and so ⊗Z R gives a short exact sequence of complexes of R-modules, which gives the
long exact sequence (1.2.1).
1
The material here is treated in Hatcher (Ch. 3) and in Appendix A of Milnor’s Characteristic Classes, which is the
reference we follow for the proof of Poincaré duality, adding a few details. Hatcher’s presentation is also modeled on
Milnor’s. The treatment in Spanier (Ch. 6) is different and perhaps less clear.
2Another description of this example: Take the quotient of the action of R∗ on R2 − {(0, 0)} by λ · (x, y) = (λx, λ−1 y).
3Recall that in a compact Hausdorff space any open cover has partitions of unity subordinate to it. In fact, this holds
for paracompact Hausdorff spaces, and our manifolds are necessarily paracompact as a second countable locally compact
and Hausdorff space is alway paracompact.
 5

Similarly, in the same setup there is a cohomology M-V sequence

→ H i (X, U ∪ V, R) → H i (X, U, R) ⊕ H i (X, V, R) → H i (X, U ∩ V, R) → H i+1 (X, U ∪ V, R) → (1.2.2)
This comes as follows: C ∗ (X, U ) ∩ C ∗ (X, V ) is the complex of singular cochains on X which vanish on
chains contained in U or V . There is a short exact sequence
0 → C ∗ (X, U ∪ V ) → C ∗ (X, U ) ∩ C ∗ (X, V ) → C ∗ (U ∪ V, U ) ∩ C ∗ (U ∪ V, V ) → 0.
But the right term is acyclic by the usual small chains type argument. So C ∗ (X, U ) ∩ C ∗ (X, V ) has
cohomology H ∗ (X, U ∪ V ). The short exact sequence of cochain complexes
0 → C ∗ (X, U ) ∩ C ∗ (X, V ) → C ∗ (X, U ) ⊕ C ∗ (X, V ) → C ∗ (X, U ∩ V ) → 0.
then gives (1.2.2) for R = Z. For general R applying ⊗Z R gives a short exact sequence of complexes
giving (1.2.2) (because this short exact sequence is termwise split).

1.3. Homology in degrees ≥ n. For a compact set K and x ∈ K the inclusion of pairs (M, M − K) ⊂
(M, M − x) induces a homomorphism
ρx : Hn (M, M − K, R) → Hn (M, M − x, R).
Note that if B is a ball centred at x (i.e. choose a neighbourhood V of x homeomorphic to Rn by a
homeomorphism taking x to 0 and then take B to be the preimage of a ball centred at 0) then excision
and deformation retraction give Hn (M, M − x, R) = Hn (B, B − x, R) = Hn (B, ∂B, R) ∼ = R is a free
cyclic R-module of rank one.
Lemma 1.3.1. If K is a compact set in an n-manifold M then
(i) Hi (M, M − K, R) = 0 for i > n
(ii) α ∈ Hn (M, M − K, R) is zero if and only if ρx (α) = 0 for all x ∈ K.

Proof. We will drop the coefficients R from the notation. A basic fact used repeatedly in the proof
is the following: If (i) and (ii) hold for compact sets K1 , K2 and for their intersection K1 ∩ K2 then
(i) and (ii) hold for K1 ∪ K2 . This follows easily from the Mayer-Vietoris sequence (1.2.1) by taking
X = M, U = M − K1 , V = M − K2 :

→ Hi+1 (M, M − K1 ∩ K2 ) → Hi (M, M − K1 ∪ K2 )

→ Hi (M, M − K1 ) ⊕ Hi (M, M − K2 ) → Hi (M, M − K1 ∩ K2 ) →
The vanishing for i > n is obvious and property (ii) is also straightforward since ρx for x ∈ Ki factors
through Hn (M, M − Ki ). Given this we proceed in several steps:
(1) M = Rn , K a compact convex set. In this case for an interior point x ∈ K there is a deformation
retraction of both (Rn , Rn − x) and (Rn , Rn − K) to a pair (Rn , Rn − B) for some large ball B ⊃ K. So
we have isomorphisms
Hi (Rn , Rn − K) = Hi (Rn , Rn − x)
so that both (i) and (ii) are obvious. Notice that in this case we have a stronger version of (ii), namely
α ∈ Hn (Rn , Rn − K) is zero if ρx (α) = 0 for some x ∈ K, i.e. it is enough to check the image under ρx
for any one point x ∈ K.
(2) M = Rn , K = K1 ∪ · · · ∪ Kr is a finite union of compact convex sets. In this case use the
Mayer-Vietoris sequence and induction on r, using the fact that (∪i<r Ki ) ∩ Kr = ∪i<r Ki ∩ Kr is a union
of r − 1 compact convex sets. Note that we get a slightly stronger version of (ii), namely if we choose
one point xi ∈ Ki for each i, then α is nonzero if and only if ρxi (α) = for all i.
(3) M = Rn and K is an arbitrary compact set. This case requires some argument. Let α ∈
Hi (M, M − K). Then there exist a compact neighbourhood Ω of K and a class α̃ ∈ Hi (M, M − Ω)
such that α̃ 7→ α under the natural map. (To prove this, represent α by a relative cycle, i.e. a chain
ξ ∈ Ci (M ) such that ∂ξ lies in M − K. Thus ∂ξ is a linear combination of simplices contained in M − K,
i.e. supp(∂ξ) ⊂ M − K. Since supp(∂ξ) is compact we can choose a neighbourhood Ω of the compact
set K which is disjoint from supp(∂ξ). Then ξ also gives a relative cycle in Ci (M, M − Ω) and the class
α̃ of ξ has the necessary property.) Let B1 , . . . , Br be a family of closed balls covering K with Bi ⊂ Ω
and Bi ∩ K 6= ∅ for all i. The map Hi (M, M − Ω) → Hi (M, M − K) factors as
Hi (M, M − Ω) → Hi (M, M − ∪Bi ) → Hi (M, M − K).
6

For i > n the middle group is zero by case (2), so α = 0. For i = n suppose that ρx (α) = 0 for all x ∈ K.
Let αB be the image of α̃ under the first homomorphism. Then we know that ρx (αB ) = ρx (α) = 0 for
x ∈ K, and by the stronger version of (ii) in case (2), we conclude that αB = 0. Then α = 0.
(4) M arbitrary, K is contained in a neighbourhood U homeomorphic to Rn . In this case the excision
Hi (M, M − K) = Hi (U, U − K) reduces it to case (3).
(5) M arbitrary, K arbitrary. In this case write K as a union of compact sets as in (4) and use the
same Mayer-Vietoris induction procedure as in (2), using (4) as the input. 

If M is compact then taking K = M we see that Hi (M, R) = 0 for i > n. For noncompact manifolds
we have a stronger statement:
Proposition 1.3.2. Let M be a noncompact manifold. Then Hi (M, R) = 0 for i ≥ n.

Proof. We drop the coefficient ring from the notation. We also assume that M is connected without loss
of generality.
i < n: An element α ∈ Hi (M ) can be represented by a cycle with support in some open set U with
compact closure U . Let K = U − U . The triple (M, M − K, M − U ) has a long exact sequence which is
the upper row of the diagram
· · · −−−−→ Hi+1 (M, M − K) −−−−→ Hi (M − K, M − U ) −−−−→ Hi (M, M − U ) −−−−→ · · ·
x

 (1.3.1)

Hi (U ) −−−−→ Hi (M )
The first vertical arrow is an isomorphism by excision of M −U . Now the outer two groups in the top
row vanish for i > n by the previous lemma, so Hi (U ) = 0 for i > n. By assumption α is in the image
of Hi (U ) → Hi (M ), so α = 0. This proves Hi (M ) = 0 for i > n.
i = n: For α ∈ Hn (M ), we represent it as coming from α ∈ Hn (U ) as in the previous case and use the
same diagram for i = n, which is now
0 −−−−→ Hn (M − K, M − U ) −−−−→ Hn (M, M − U )
x

 (1.3.2)

Hn (U ) −−−−→ Hn (M )
by the previous lemma. Consider the image ᾱ of α in Hn (M, M − U ). Evidently ρx (ᾱ) is either nonzero
for all x or zero for all x. (Given two points take a path between them and cover it by small open balls.
Now use the stronger version of (ii) in case (1) of the previous proof.) Since the cycle representing α has
compact support contained in U , it must be the case that ρx (ᾱ) = 0 for all x ∈ U , as it is certainly so
for x outside the support. By the previous lemma ᾱ is zero in Hn (M, M − U ). By the exact sequence
α = 0 in Hn (U ), and hence in Hn (M ). 

1.4. Orientations. Let M be an n-manifold.

A local R-orientation at x ∈ M is a choice of generator of the rank one free R-module Hn (M, M −
x, R).
An R-orientation for M is a choice of local R-orientations at all points, written x 7→ µx , satisfying
a continuity condition: For each x, there is a neighbourhood B of x in M such that for y ∈ B, the local
orientations µx and µy are related by the homomorphisms
ρx ρy
Hn (M, M − x, R) ← Hn (M, M − B, R) → Hn (M, M − y, R),
i.e. there is a µB ∈ Hn (M, M − B, R) with ρx (µB ) = µx and ρy (µB ) = µy .
An R-orientation of M may or may not exist, if one exists we say that M is R-orientable, and if
a specific R-orientation has been chosen we say that M is R-oriented.
The R-orientation manifold of M is M̃R = {(x, µx )|µx is an R-orientation at x}. It is naturally
a manifold and the mapping (x, µx ) 7→ x defines a map p : M̃R → M which is a covering map. (The
fibre over any point is (noncanonically) identified with the units R× in R.) An R-orientation is simply a
section of this covering map. Note that the R-orientation manifold M̃R is itself R-oriented, i.e. it comes
with an orientation.
 7

If R = Z (in which case we drop the ring from the notation) then the orientation manifold is a double
covering of M . The manifold M is orientable if and only if this is a trivial cover, i.e. M̃ = M t M , and
an orientation is a choice of component. (So if M is orientable then it has precisely 2|π0 (M )| orientations).
Since a connected M can have a nontrivial double cover only if π1 (M ) has an index 2 subgroup, we see
that if π1 (M ) has no such subgroup (e.g. if M is simply connected) then M is orientable. Thus spheres,
tori, odd-dimensional real projective spaces, all complex and quaternionic projective spaces, complex
Grassmannians etc. are orientable.
If M is oriented (i.e. Z-oriented) then it is R-oriented for any ring R since Hn (M, M − x, R) =
Hn (M, M − x, Z) ⊗Z R by the universal coefficient theorem.
If R = Z/2 then M̃Z/2 = M and any manifold is Z/2-orientable.

1.5. Fundamental class.

Proposition 1.5.1. Suppose that M is R-oriented, with local orientations µx ∈ Hn (M, M − x, R) for
each x ∈ M . Then for any compact set K ⊂ M there is a unique class µK ∈ Hn (M, M − K, R) such
that ρx (µK ) = µx for all x ∈ K.

Proof. We drop the ring R from the notation for simplicity. Notice that the previous lemma shows that
the class µK is unique if it exists, so we have only to prove existence. For K sufficiently small this is
contained in the definition of orientation, so we need to patch these together.
Suppose that K = K1 ∪ K2 where µK1 and µK2 are known to exist. The M-V sequence is
0 → Hn (M, M − K) → Hn (M, M − K1 ) ⊕ Hn (M, M − K2 ) → Hn (M, M − K1 ∩ K2 ) →
By the uniqueness (prevous lemma), µK1 and µK2 have the same image in Hn (M, M − K1 ∩ K2 ), so
(µK1 , µK2 ) 7→ 0 in the M-V sequence. Therefore there is a unique µK ∈ Hn (M, M − K) mapping to it.
It is immediate that ρx (µK ) = µx for all x ∈ K.
For a general K write K = K1 ∪ · · · ∪ Kr with each Ki small enough as in (1) . Induction on r using
the Mayer-Vietoris argument gives the existence of µK . 

In the case where M is compact and R-oriented, taking K = M we get a class µM ∈ Hn (M, R) such
that ρx (µM ) is the local R-orientation at each point x ∈ M .
Definition 1.5.2. The fundamental class of an R-oriented n-manifold is the class [M ] := µM ∈
Hn (M, R).
Note that if M is R-orientable and connected then necessarily Hn (M, R) is a free rank one R-
module with any fundamental class as a generator. (Different choices of R-orientations give different
R-fundamental classes which are multiples of each other by units of R.) For example, this shows that
even-dimensional projective spaces can only be R-oriented if R is a Z/2-algebra.
Another consequence of the existence of a fundamental class (at least with Z/2 coefficients) for
compact manifolds is that compact manifolds of different dimensions cannot be homotopic to each other.
(Intuitively this seems quite plausible, but it is not easy to prove without some machinery.)

1.6. Compactly supported cohomology. Define the compactly supported cohomology of a topo-
logical space X as the direct limit
Hci (X, R) := limK H i (X, X − K, R)
−→
where the direct limit is taken over compact subsets K ⊂ X with respect to the pullback maps induced
by (X, X − K) ⊂ (X, X − K 0 ) if K 0 ⊂ K. This can also be defined as the cohomology of the direct
limit complex limK C ∗ (X, X − K, R) since taking direct limits commutes with cohomology. The obvious
−→
map Hc∗ (X, R) → H ∗ (X, R) to cohomology is an isomorphism if X is compact. Compactly supported
cohomology is not a homotopy invariant: For example, Hci (Rn , R) is zero for i 6= n and R for i = n (take
the limit as K runs over closed balls, which are cofinal among all compacts).
Exercise 1.6.1. Check the following properties of compactly supported cohomology:
(1) If U ⊂ X is an open set then there is a map Hc∗ (U, R) → Hc∗ (X, R).
(2) If f : X → Y is a proper map then there is a pullback f ∗ : Hci (Y, R) → Hci (X, R).
(3) If X = U t Z with Z closed then there is a long exact sequence
· · · → Hci (U, R) → Hci (X, R) → Hci (Z, R) → · · ·
in compactly supported cohomology.
8

Let X = U ∪ V where U, V are open. There is a Mayer-Vietoris exact sequence

→ Hci (U ∩ V, R) → Hci (U, R) ⊕ Hci (V, R) → Hci (X, R) → . (1.6.1)
Indeed, this follows from the cohomology Mayer-Vietoris sequence for pairs (1.2.2) as follows (we drop
the coefficients R for notational simplicity): For K ⊂ U and L ⊂ V compact sets we have a long exact
sequence
→ H i (M, M − K ∩ L) → H i (M, M − K) ⊕ H i (M, M − L) → H i (M, M − K ∪ L) →
Using excisions this is the same as
→ H i (U ∩ V, U ∩ V − K ∩ L) → H i (U, U −K) ⊕ H i (V, V − L) → H i (M, M − K ∪ L) →
Now the set of compact subsets of U ∩ V which are of the form K ∩ L is cofinal among all compact
subsets, so taking limits over K and L gives the desired sequence (1.6.1) since direct limits preserve
exactness.

1.7. Poincaré duality. Recall the cap product

H i (X, A, R) × Hj (X, A ∪ B, R) → Hj−i (X, B, R)
which exists if A, B are open in A ∪ B. (We will generally use it for open sets A, B in X or for sets A, B
which can obviously be replaced by open sets without changing the (co)homology groups involved.) It
has a simple relationship with cup product, namely
ha ` b, ci = hb, a a ci
where h , i is the pairing of cohomology and homology. (On the level of chains and cochains this is the
definition of a.)
Now suppose that µK ∈ Hn (M, M − K, R) where M is an n-manifold. Then a 7→ a a µK defines a
map
a µK : H i (M, M − K, R) → Hn−i (M, R).
If M is an R-oriented manifold and µK is the class given by the previous proposition, then the maps
a µK are compatible for varying K (by uniqueness of µK and naturality of the cap product with respect
to inclusions) and so we get a map from the direct limit
PDM : Hci (M, R) → Hn−i (M, R).
In the compact case PDM ( · ) = ( · ) a [M ] is capping with the fundamental class.
Theorem 1.7.1. Let M be an R-oriented manifold of dimension n. Then PDM : Hci (M, R) →
Hn−i (M, R) is an isomorphism for all i.

Proof. The proof is again by a Mayer-Vietoris patching argument from small cases.
(1) M = Rn . Let B be a ball and µB ∈ Hn (Rn , Rn − B, R) a generator. By the universal coefficient
theorem H n (Rn , Rn − B, R) is the dual of Hn (Rn , Rn − B, R) (because Hn−1 (Rn , Rn − B, R) = 0), so it
is free of rank one over the coefficient ring with a generator e and hµB , ei = 1. Now
1 = h1 ` e, µB i = h1, e a µB i
so that e a µB is a generator of H0 (Rn , R) = R. Thus a µB maps H n (Rn , Rn − B, R) isomorphically
to H0 (Rn , R). Taking the limit over larger balls B gives the statement.
(2) (M-V principle) Suppose that M = U ∪ V and PDU : Hc∗ (U ) → Hn−∗ (U ), PDV and PDU ∩V are
known to be isomorphisms. (We drop the coefficients from now on for notational simplicity.) We have
the diagram
−−−−→ Hci (U ∩ V ) −−−−→ Hci (U ) ⊕ Hci (V ) −−−−→ Hci (M ) −−−−→ Hci+1 (U ∩ V ) −−−−→
   
PD PD ⊕(−PD ) PD PD
y U ∩V y U V y M y U ∩V
−−−−→ Hn−i (U ∩ V ) −−−−→ Hn−i (U ) ⊕ Hn−i (V ) −−−−→ Hn−i (M ) −−−−→ Hn−i−1 (U ∩ V ) −−−−→
(1.7.1)
The naturality of cap product implies that the first two squares commute. For the third square let
K ⊂ U and L ⊂ V be compact sets and consider
H i (M, M − K ∪ L) −−−−→ H i+1 (U ∩ V, U ∩ V − K ∩ L)
 

aµK∪L y
aµ
y K∩L
Hn−i (M ) −−−−→ Hn−i−1 (U ∩ V ).
 9

The excision H i+1 (M, M − K ∩ L) = H i+1 (U ∩ V, U ∩ V − K ∩ L) shows that the third square of the
previous diagram is the limit of these squares over K and L. That these squares commute up to a sign
requires an argument which we skip, referring to Hatcher, p. 246f. The 5-lemma then implies that PDM
is an isomorphism.
(3) If M is homeomorphic to an open ball then (1) applies. If M is the union of finitely many open
balls then (2) and an induction on the number of balls implies that PDM is an isomorphism.
(4) (direct limits) Suppose M is the nested union of subspaces, i.e. M = ∪α∈A Uα where A is
directed and Uα ⊂ Uβ if α ≤ β and assume further that every compact set in M is contained in
some Uα . Then PDM is an isomorphism if PDUα is an isomorphism for all α ∈ A. Indeed, this is
because Hci (M ) = lim Hci (Uα ) and Hn−i (M ) = lim Hn−i (Uα ) and the direct limit of isomorphisms is an
−→ −→
isomorphism.
(5) If M is an arbitrary open subset of Rn then (since Rn is second countable) it can be written as
the union of countably many open balls B1 , B2 , . . . . Let Ui = ∪j≤i Bj ; then Ui is a nested family of open
subsets of M and PDUi is an isomorphism for all i by (3). By (4) PDM is an isomorphism.
(6) If M is second countable then cover M by countably many open subsets {Ui }i≥1 homeomorphic
to open subsets of Rn . Let Ui0 = ∪j≤i Ui . By (2) and induction we know PDUi0 is an isomorphism for all
i. By (4) we get that PDM is an isomorphism.
(7) If M is arbitrary then we use a Zorn’s lemma argument. Let
U = {U ⊂ M open | PDU is an isomorphism}.
Since every chain in U has an upper bound (the union is one, by (4)), Zorn’s lemma says that there is
a maximal element U . If U 6= M choose x ∈ M − U . Let V be a neighbourhood of x homeomorphic to
Rn . Then by (1), (2), and (5) we know that PDU ∪V is an isomorphism, contradicting the maximality of
U. 

(Note that the theorem actually holds even without assuming M is second countable.)
The special case when M is compact is of particular importance:
Corollary 1.7.2. If M is a compact R-oriented n-manifold then a [M ] : H i (M, R) → Hn−i (M, R) is
an isomorphism for all i.

It follows that for a compact connected R-oriented n-manifold H n (M, R) is a free rank one R-module
with canonical generator ωM where ωM a [M ] = 1 ∈ H0 (M, R).
Note that if R = Z/2 then any manifold is Z/2-orientable, so that the corollary applies. In particular,
this “explains” why the homology of RP n (n even) looks nice with Z/2 coefficients.

1.8. Duality and the cup product pairing in cohomology. Let M be an R-oriented n-manifold.
There is a canonical trace homomorphism T r : Hcn (M, R) → R defined by
PD
Hcn (M, R) → H0 (M, R) → R
where H0 (M, R) = ⊕π0 (M ) R → R is given by summing components. (The trace should be thought of as
integration over the manifold. This will be made precise in the context of de Rham cohomology later.)
Corollary 1.8.1. Let M be an R-oriented manifold. The pairing
Hci (M, R) × H n−i (M, R) → R by (a, b) 7→ T r(a ` b)
is perfect modulo torsion.

Proof. Since we are working modulo torsion the pairing between homology and cohomology is perfect.
PD×id
Hci (M ) × H n−i (M ) −−−−→ Hn−i (M ) × H n−i (M )
 
 
y y
Hcn (M ) −−−−→ R
The diagram commutes since hPD(α), βi = T r(β ` α), as one checks by working with H i (M, M − K)
instead of Hci (M ) for a compact set K. 
10

Now consider the case where M is compact and oriented and R = Z. If the dimension of M is
n = 2m then this defines a perfect bilinear form on H 2 (M )/tors with values in Z. By the graded-
commutivity of cup product this is symmetric if m is even and alternating (=skew-symmetric) if m is
odd. We get as a corollary that if m is odd then H m (M )/tors must be even dimensional. (There cannot
be a nondegenerate alternating form on an odd dimensional free abelian group or vector space over a
field of characteristic zero.) The case m odd is not so interesting because any two perfect alternating
forms on Zr are equivalent, so assume that n = 2m = 4l. Then the pairing above is a perfect symmetric
pairing on H 2l (M, Z)/tors, i.e. a perfect symmetric bilinear form. Perfect symmetric bilinear forms have
many useful invariants, all of which give invariants of the manifold. For example, the signature of the
real bilinear form on H 2l (M, R) = H 2l (M, Z) ⊗ R is then an invariant of the manifold.
Now suppose that M is simply-connected and l = 1, i.e. M is a compact simply-connected (and
hence orientable) 4-manifold. Then H1 (M ) = 0 and hence H 1 (M ) = 0 (universal coefficients) and
H 3 (M ) = 0 (duality). So the only interesting cohomology is H 2 (M ), which is free (by the universal
coefficient theorem), and it has the symmetric bilinear form given by the intersection form. An amazing
theorem of M. Freedman (1982) says that there are (up to homeomorphism) at most two simply-connected
compact four manifolds (without boundary) with a given H 2 (M, Z) and intersection form, and only one
of them has a smooth structure.

1.9. Duality between cohomology and Borel-Moore homology (∗). For noncompact manifolds,
there is another Poincaré duality statement relating cohomology to something called Borel-Moore ho-
mology. For now, we will just give the statement:
Theorem 1.9.1. Let M be an R-oriented n-manifold. Then there is a fundamental class [M ] ∈
HnBM (M, R) and cap product with [M ] gives an isomorphism PD : H i (M, R) → Hn−i
BM
(M, R) for all
i.

The definition and properties of Borel-Moore homology and the proof of this theorem will be given
in a problem set.

1.10. Euler characteristics. The Euler characteristic of a space X is defined as

X
χ(X) = (−1)i rank(Hi (X, Z)).
i

We assume here that the homology groups have finite rank. The mod 2 Euler characteristic is defined as
X
χ2 (X) = (−1)i dim Hi (X, Z/2).
i

Lemma 1.10.1. If H∗ (X) is finitely generated then χ(X) = χ2 (X).

i i i
P P
Proof. First note that i (−1) dim H (X, Z/2) = i (−1) dim Hi (X, Z/2) since Z/2 is a field and
∗ i
Hi (X, Z/2) = H (X, Z/2). The universal coefficient theorem for cohomology gives short exact sequences
0 → Ext(Hi−1 (X, Z), Z/2) → H i (X, Z/2) → Hom(Hi (X, Z), Z/2) → 0.
Write Hi (X, Z) as a sum of a free summand Zr and its torsion subgroup. Each Z summand of
Hi (X, Z) contributes Z/2 to H i (X, Z/2) and Ext(Z, Z/2) = 0 to H i+1 (X, Z). So each Z summand
contributes
( (−1)i to both χ(X) and χ2 (X). Each torsion summand
( Z/m contributes Hom(Z/m, Z/2) =
Z/2 m even Z/2 m even
to H i (X, Z/2) and also Ext(Z/m, Z/2) = to H i+1 (X, Z). In either case
0 m odd 0 m odd
the contribution of a torsion summand of Hi (X, Z) to H ∗ (X, Z/2) is zero, proving the lemma. 

Corollary 1.10.2. The Euler characteristic of a compact odd-dimensional manifold is zero.

Proof. We will assume the fact (to be proved later) that H∗ (M, Z) is finitely generated for a compact
manifold. By the lemma, χ(M ) = χ2 (M ). Poincaré duality with Z/2 coefficients implies that
dim Hi (M, Z/2) = dim H i (M, Z/2) = dim Hn−i (M, Z/2).
Since n is odd, these dimensions appear with opposite signs in χ2 . 
 11

1.11. Finite generation of homology. Consider the following trivial remark: If Poincaré duality holds
for a zero-dimensional manifold X then X must be finite (because H 0 ∼ = H0 is required to be finitely
generated). This suggests that Poincaré duality should force the homology of a compact manifold to be
finitely generated. In fact:
Theorem 1.11.1. If M is a compact R-oriented manifold then H ∗ (M, R) and H∗ (M, R) are finitely
generated R-modules.

Proof. (from Spanier, Algebraic Topology) With a space X, coefficient ring R, and a coefficient R-module
M , one can associate two cochain complexes. The first is the usual cochain complex
C ∗ (X, M ) = HomR (C∗ (X, R), M ) = Hom(C∗ (X), M )
with cohomology H ∗ (X, M ). There is also
C ∗ (X, R) ⊗R M = Hom(C∗ (X), R) ⊗R M = HomR (C∗ (X, R), R) ⊗R M
with differential δ ⊗ 1. There is a map from the second complex to the first:
X  X
µ : HomR (C∗ (X, R), R) ⊗R M → HomR (C∗ (X, R), M ) by µ ϕi ⊗ mi (ξ) = hϕi , ξi mi .
i i
This will not usually induce an isomorphism in cohomology, but for a compact manifold X Poincaré
duality implies that it does. To see this, recall that the universal coefficient theorem for homology takes
the form
0 → Hk (X, R) ⊗R M → Hk (X, M ) → T or1 (Hk−1 (X, R), M ) → 0.
Using Poincaré duality isomorphisms Hk (X) ∼ = H n−k (X) this gives
0 → H k (X, R) ⊗R M → H k (X, M ) → T or1 (H k+1 (X, R), M ) → 0.
Now consider the second complex above. Let K ∗ = HomR (C∗ (X), R). The definition of T or1 gives a
short exact sequence
0 → H k (K ∗ ) ⊗R M → H k (K ∗ ⊗R M ) → T or1 (H k+1 (K ∗ ), M ) → 0.
Since H k (K ∗ ) = H k (X, R) we see that the two sequences are the same, and the natural mapping µ
induces isomorphisms in cohomology.
Now we see that this property implies finite generation of the homology. Let C∗ be a chain complex
of R-modules such that
(*) for every R-module M the natural map of cochain complexes
µ : HomR (C∗ , R) ⊗R M → HomR (C∗ , M )
induces isomorphisms in cohomology.
We will show that this implies that H∗ (C∗ ) is finitely generated. The cohomology universal coefficient
theorem gives the sequence
h
0 → Ext1 (Hk−1 (C∗ ), M ) → H k (HomR (C∗ , M )) → HomR (H k (C ∗ ), M ) → 0.
Taking M = Hk (C∗ ) the surjectivity of h and the property (∗) imply that we can write
X 
hµ ai ⊗ bi = id ∈ HomR (Hk (C∗ ), Hk (C∗ ))
i
k
for some ai ∈ H (HomR (C∗ , R)), bi ∈ Hk (C∗ ). Now apply both sides to an element b ∈ Hk (C∗ ):
X  D X  E X
b = hµ ai ⊗ bi (b) = µ ai ⊗ bi , b = hai , bi bi .
Thus the bi generate Hk (C∗ ).
This proves that H∗ (X, R) is finitely generated; it follows from the universal coefficient theorem (or
Poincaré duality) that H ∗ (X, R) is finitely generated. 

In fact the homology of a compact manifold is always finitely generated, with no assumption about
orientability. To generalize the proof above requires introducing Cech cohomology (cf. Spanier) so we
will leave it aside for now. A different, more topological way to prove finite generation is by showing that
a compact manifold is homotopy equivalent to a finite CW complex. (See the Appendix of Hatcher for
a proof of this.) Note that the theorem would follow immediately if one could show that (a) a compact
manifold has a CW structure or (b) a compact manifold has a simplicial structure. (Of course, (b)
implies (a).) But (a) is unknown in general and (b) is false in general. A theorem of Whitehead says
that (b) (and hence (a)) is true for smooth manifolds, as we shall see later.
12

1.12. Manifolds with boundary. A manifold with boundary is a second-countable Hausdorff topo-
logical space M such that every point has a neighbourhood homeomorphic to an open subset in the
half-space Rn+ = {x ∈ Rn |xn ≥ 0}. If in such a neighbourhood x is carried to a point with xn = 0
then we say that x is in the boundary of M , denoted ∂M . It is easy to check that ∂M is a manifold
of dimension n − 1. It is also easy to check that the product of manifolds with boundary is naturally a
manifold with boundary.
A collar for ∂M is a neighbourhood of ∂M in M which is homeomorphic to ∂M × [0, 1) under a
homeomorphism taking ∂M to ∂M × {0} by the identity.
Lemma 1.12.1. If M is a manifold with compact boundary ∂M then ∂M admits a collar in M .

Proof. (cf. Hatcher p. 253) Let M 0 be the manifold one gets by attaching a collar ∂M × [0, 1] to ∂M ,
with ∂M × {0} being glued to ∂M via the identity. It is enough to show that M 0 is homeomorphic to
M.
Choose a partition of unity ϕi : ∂M → [0, 1] such that each supp(ϕi ) has closure contained in an
open set Ui in M which is homeomorphic to Rn+ .
Mk := M ∪ {(x, t) ∈ ∂M × [0, 1] : t ≤ (ϕ1 + · · · + ϕk )(x)}.
For k large enough, Mk = M 0 . So it suffices to construct homeomorphisms hk : Mk−1 ∼ = Mk for each k.
Now since Uk ∼= Rn+ we can find a homeomorphism of ∂Uk × [−1, 1] onto a neighbourhood of ∂Uk in M 0 .
Define hk to be the hidentity outside ∂Uik × [−1, 1], andh inside ∂Uk × [−1,
i 1] we choose a homeomorphism
P P
which maps {x} × −1, i≤k−1 ϕk (x) onto {x} × −1, i≤k ϕi (x) linearly. This gives the desired
hk . 

(The proof works if ∂M is paracompact.)

We will say that a manifold-with-boundary is R-oriented (or R-orientable) if M − ∂M is R-oriented
(or R-orientable).
Lemma 1.12.2. If M is R-orientable then ∂M is R-orientable.

Proof. Exercise. 

Now if M is a compact R-oriented manifold with boundary then using the collar above we can write
Hn (M, ∂M, R) = Hn (M − ∂M, ∂M × (0, )), which has a generator coming from the R-orientation. Thus
there is a fundamental class [M ] ∈ Hn (M, ∂M, R).

1.13. Lefschetz duality. Let M be an R-oriented n-manifold with boundary. The following theorem
can be generalized to the noncompact case but we will be content to state the compact version:
Theorem 1.13.1. Let M be an R-oriented compact n-manifold with boundary ∂M . Then a [M ] :
H i (M, ∂M, R) → Hn−i (M, R) and a [M ] : H i (M, R) → Hn−i (M, ∂M, R) are isomorphisms.

Proof. Both homomorphisms exist because [M ] ∈ Hn (M, ∂M ). (In the cap product H i (X, A)×Hj (X, A∪
B) → Hj−i (X, B) they correspond to the choices A = ∂M, B = ∅ and A = ∅, B = ∂M .)
The existence of a collar for ∂M easily implies that Hck (M − ∂M ) = H k (M, ∂M ) and Hn−i (M ) =
Hn−i (M − ∂M ). So the first isomorphism is just Poincaré duality for the noncompact manifold M − ∂M .
For the second, look at the long exact sequences of the pair (M, ∂M )
−−−−→ H i (M, ∂M ) −−−−→ H i (M ) −−−−→ H i (∂M ) −−−−→ H i+1 (M, ∂M ) −−−−→
   
a[M ]y
 
a[M ]y a[∂M ]y
 a[M ]y


−−−−→ Hn−i (M ) −−−−→ Hn−i (M, ∂M ) −−−−→ Hn−i−1 (∂M ) −−−−→ Hn−i−1 (M ) −−−−→
The diagram commutes because the image of [M ] under the connecting homomorphism Hn−i (M, ∂M ) →
Hn−i−1 (∂M ) is a fundamental class, which we have denoted [∂M ]. Thus Poincaré duality for the
boundary and the first part of the theorem give the second part of the theorem. 
Corollary 1.13.2. Let M be a compact R-oriented n-manifold with boundary ∂M . Then
H∗ (M, R), H∗ (M, ∂M, R) and H ∗ (M, R), H ∗ (M, ∂M, R)
are finitely generated R-modules.
 13

Proof. Substitute Lefschetz duality in the proof of finite generation above. 

Corollary 1.13.3. If M is a compact n+1-dimensional manifold with boundary then χ(∂M ) = (1 +

(−1)n )χ(M ). (So χ(∂M ) is even.)

Proof. If (X, A) has finitely generated homology then χ(X, A) = χ2 (X, A) (this was proved earlier for
A = ∅ and the same proof works). Now for A ⊂ X we have χ2 (X) = χ2 (A) + χ2 (X, A) by the relative
sequence of the pair, assuming finite generation. In our case take X = M, A = ∂M and use the equality
dim Hi (M, ∂M, Z/2) = dim H i (M, ∂M, Z/2) and duality H ∗ (M, ∂M, Z/2) ∼ = Hn+1−∗ (M, ∂M, Z/2) to
get
χ2 (M ) = χ2 (∂M ) + χ2 (M, ∂M ) = χ2 (∂M ) + (−1)n+1 χ2 (M ) (1.13.1)
which gives the corollary. 

(Strictly speaking we have only proved this for M oriented as we have not proved finite generation
of H∗ (M, ∂M ) in the nonorientable case, but assuming that the corollary is proved.)
In particular a manifold of odd Euler characteristic (for example, CP 2k ) cannot be the boundary of
a compact manifold. The oriented topological cobordism ring Ωtop ∗ is defined as follows: Ωtop
n consists of
classes of oriented manifolds (without boundary) modulo the equivalence relation of cobordism: N1 and
N2 are oriented cobordant if N1 t (−N2 ) = ∂M where M is an oriented n + 1-manifold with boundary.
(Here −N2 is N2 with orientation reversed. Evidently this relation is weaker than homeomorphism.) This
is a group under taking disjoint unions (the identity is the empty set) and Ωtop top
∗ = ⊕n Ωn is a ring under
taking direct products. The corollary tells us that [N ] 6= 0 if χ(N ) is odd; in particular [CP 2k ] ∈ Ωtop
4k
is nonzero. A theorem of Sullivan says that Ωtop top
4∗ ⊗ Z[1/2] = ⊕k≥0 Ω4k ⊗ Z[1/2] is a polynomial ring
generated by the classes of even-dimensional complex projective spaces. (The analogous theorem in the
context of smooth manifolds and the corresponding ring Ω∗ is a famous theorem of Thom.) The complete
structure of the ring Ωtop
∗ is quite subtle.

1.14. Gysin homomorphisms and Poincaré dual classes. Let f : M → N be a continuous map of
R-oriented manifolds of dimension m and n respectively. Then using Poincaré duality we get a map

f! : Hck (M, R) → Hck+n−m (N, R),

defined by f! := PD−1
N ◦ f∗ ◦ PD M . which is called the Gysin homomorphism. In the case where M
and N are compact the relation to the usual pullback f ∗ is given by

f! (f ∗ (α) ` β) = α ` f! (β) (α ∈ H k (N, R), β ∈ H ∗ (M, R))

(Check this formula.) Thus f! is a H ∗ (N, R)-module homomorphism if H ∗ (M, R) is given the H ∗ (N, R)-
module structure via f ∗ .
Now consider the case where i : M ,→ N is injective (and M and N are compact). Thus i is a
continuous bijection onto its image, i.e. a homeomorphism of M with a closed subspace of N . We will
refer to M as an embedded submanifold. Then there is a defined a class

ξM := i! (1) ∈ H n−m (M, R)

which we call the Poincaré dual class to M . It has the property that if α ∈ H m (N, R) and i∗ (α) = ωM
then α ` ξM = ωN . (Note that ξM may be zero. For example if M = S 1 is the equator of N = S 2 then
ξM ∈ H 1 (S 2 , R) = 0.)
There is also a naturally defined homology class i∗ ([M ]) ∈ H m (N, R). (Intuitively, M defines a
cycle in N (for example, if one had a simplicial complex structure on N in which M is a subcomplex)
and this is the class of that cycle.) These two are dual in the sense that

ξM a [N ] = i∗ ([M ]).

(Note that [M ], [N ], and ξM depend on the choices of orientations. But for example if R = Z then
each of these is independent of the choice up to ±1.)
14

1.15. Cup product and intersection product. Let M be compact and R-oriented. Then Poincaré
duality can be used to define the intersection product
Hi (M, R) × Hj (M, R) → Hi+j−n (M, R)
by inverting the vertical maps in the following diagram:
Hi (M, R) × Hj (M, R) −−−−→ Hi+j−n (M, R)
x x
PD×PD 
 
PD
`
H n−i (M, R) × H n−j (M, R) −−−−→ H 2n−i+j (M, R)
Intuitively it is given by intersecting homology cycles (think of them, e.g. as simplicial cycles) but it is
very difficult to make this rigorous. (Note that this works only for oriented compact manifolds whereas
the cup product on cohomology is defined for arbitrary spaces.)
In some cases it is possible to give a more geometric description of the intersection product. Let M
be a compact oriented n-manifold. Suppose that iA : A ,→ M, iB : B ,→ M are closed submanifolds of
dimension a, b respectively, each of which is oriented. Suppose also that A ∩ B is a connected manifold
of dimension a + b − n and moreover, for each x ∈ A ∩ B there is a neighbourhood U of x in M and a
homeomorphism U ∼ = Rn such that
(U ∩ A, U ∩ B, U ∩ A ∩ B) ∼ = (Ra , Rb , Ra ∩ Rb ∼
= Ra+b−n ).
We say that A and B intersect transversely. The orientations of A and B and the orientation of M
together determine an orientation on A ∩ B. (Exercise: Check this.) We will not prove the following
theorem which gives a very geometric description of some cup products. (A proof this version can be
found in Dold, Lectures on Algebraic Topology, VII.11. We will prove a version later in the context of
smooth manifolds.)
Theorem 1.15.1. Let M be a compact oriented manifold and let A, B be oriented compact submanifolds
that intersect transversely (as defined above). Then ξA ` ξB = ±ξA∩B and the intersection product of
the homology classes is iA∗ ([A]) · iB∗ ([B]) = ±iA∩B∗ ([A ∩ B]).

(Here ± is a sign determined by the orientations of A and B, and can be made +1 by choosing
orientations correctly, we ignore this for now.)
For example if dim A + dim B = n then a transverse
P intersection is a finite set of points, so that
[M ] a ξA∩B ∈ H0 (M ) is a scalar. This is a sum x∈A∩B sgn(x) where sgn(x) is plus or minus one
according to whether the orientations of A and B at x give the orientation of M or its negative.
This theorem is very useful in practice because many cohomology classes (but not all!) can be
written as linear combinations of Poincaré dual classes of submanifolds. (Exercise: Use this theorem to
determine the cohomology ring of CP n .)

1.16. Alexander duality (∗). This is a duality for a closed subset X ⊂ M of a “background” manifold,
which is taken to be a sphere or Euclidean space in the classical Alexander duality. (The case of manifolds-
with-boundary can be reduced to this using a doubling construction.)
Since we are now allowing the subset X to be complicated (e.g. not a manifold), we will need Čech
cohomology. We will just state the results for now.
Let X ⊂ M be a closed subset of a compact n-manifold. Then there are isomorphisms
∼ Hn−i (M − X)
Ȟ i (M, X) =
and
Ȟ i (X) ∼
= Hn−i (M, M − X)
∗
where Ȟ denotes Čech cohomology (to be defined later).
So for example, if X is a manifold of dimension m, embedded as a closed subset of M , then
Hi (M, M − X) ∼= H n−i (X) ∼
= Hn−m+i (X)
The dual version H i (M, M − X) ∼
= H i−(n−m) (X) is usually referred to as the Gysin isomorphism.
If X is assumed to be a locally contractible subset of a manifold then Ȟ i = H i , so that we get
simpler statements which do not involve Čech cohomology. If the background manifold M is S n , then
this reduces to a statement called Alexander duality:
H̃i (S n − X) ∼
= H̃ n−1−i (X).
 15

Here we have used reduced homology. Note that the homology of S n − X depends only on X and not
on how it sits in S n .
Remark 1.16.1 (Verdier duality). Wouldn’t it be nice if there were a single statement encapsulating all
the statements above? There is, namely Verdier duality, which was formulated by Grothendieck and
proved by Verdier in the 1960s. The formulation requires more sophisticated homological algebra tools,
so maybe later.
16

2. Smooth manifolds

From now on we will only deal with C ∞ things, and the words “C ∞ ”, “smooth” and “differentiable”
will be used interchangeably. A reference for the material here is Warner’s Foundations of Differentiable
Manifolds and Lie Groups or any other book on manifolds.

2.1. Smooth manifolds. Let M be a paracompact second-countable Hausdorff space. (4) (5)
A chart (or coordinate neighbourhood) on M is a triple (U, V, ϕ) consisting of an open subset
U ⊂ M , an open subset V of Rn (for some n), and a homeomorphism ϕ : U → V . Two charts (U, V, ϕ)
and (U 0 , V 0 , ϕ0 ) are compatible if the transition functions ϕ0 ◦ ϕ−1 : ϕ(U ∩ U 0 ) → ϕ0 (U ∩ U 0 ) and
ϕ ◦ (ϕ0 )−1 : ϕ0 (U ∩ U 0 ) → ϕ(U ∩ U 0 ) are smooth functions between open subsets of Rn .
An atlas for M is a collection of charts such that the open sets cover M . Notice that a space M
which has an atlas of charts is necessarily locally compact.
A smooth atlas for M is a collection of compatible charts which cover M , i.e. a collection of charts
A = {(Uα , Vα , ϕα )}α∈A which are pairwise compatible and such that ∪α Uα = M . Any smooth atlas A
is contained in a maximal (with respect to inclusion) smooth atlas: simply include in the collection any
chart which is compatible with all the charts in A .
A smooth (or differentiable or C ∞ ) manifold is a pair (M, A ) consisting of a paracompact
second-countable Hausdorff space M and a maximal smooth atlas A on M . We will usually drop the
atlas from the notation. (Note that a smooth atlas A determines a (unique) maximal smooth atlas (as
above take the collection of all charts compatible with the atlas). The word maximal is just put in so
that things are unique, in practice in all arguments with any atlas contained in the maximal atlas will
do.)
Note that by our second countability assumption a manifold has at most countably many connected
components, and each connected component is second countable. (6) (7)
A continuous function f : M → R is called smooth (or differentiable or C ∞ ) if, for each open
set ϕα : Uα → Vα ⊂ Rn of an atlas, the composition f ◦ ϕ−1 α : Vα → R is a smooth function. (Clearly this
holds for one atlas contained in the given maximal atlas if and only if it holds for all atlases contained in
the given maximal atlas.) There is an obvious definition for a function defined on an open subset U ⊂ M
to be smooth. The assignment U 7→ { smooth functions on U } is a sheaf (of R-algebras), the sheaf of
smooth functions on M .
Let f : M → N be a continuous map between two smooth manifolds of dimension m and n
respectively. It is called a smooth map if the following holds at all points p: For each point p ∈ M ,
using charts around p and f (p) ∈ N , we can get a map between open subsets of Rm and Rn . This map
should be smooth. Again, it is clear that it suffices to check this for any one pair of atlases contained in
the maximal atlases defining M and N respectively. Clearly the composition of smooth maps is smooth.
A smooth map f : M → N is called a diffeomorphism if it has a smooth inverse, i.e. there exists
a smooth map g : N → M such that f ◦ g = g ◦ f = id. (A smooth 1-1 map need not have smooth
inverse, for example look at R → R by x 7→ x3 .)
The product of two smooth manifolds is a smooth manifold. (• Exercise.)
Smooth manifolds with smooth maps between them form a category with the obvious forgetful
functor to the category of topological manifolds and continuous maps and further to T op.

2.2. Real analytic manifolds. By replacing “smooth” everywhere by “real analytic” we get the notion
of a real analytic manifold. So a real analytic manifold is a nice topological space M with an open covering
by subsets homeomorphic to Rn such that the transition maps are real analytic functions. It makes sense
to talk about real analytic maps between such manifolds, and real analytic functions on them.
4A topological space is paracompact if every open covering admits a locally finite open refinement. (A covering is locally
finite if every point has an open neighbourhood which meets only finitely many members of the covering. A refinement of
{Uα } is a covering {Vβ } such that for each Vβ there exists α with Vβ ⊂ Uα .)
5
A topological space is second-countable if the topology admits a countable base (i.e. there is a countable collection of
open sets such that every open set is a union of members of the collection).
6Here is an example to show that the Hausdorff condition is not redundant: Take two copies of R and identify them
outside the origin. (Equivalently, take R2 − {0} and quotient by the relation (x, y) ∼ (λx, λ−1 y) for λ ∈ R∗ , with the
quotient topology.) This gives a space which is locally homeomorphic to R but is not Hausdorff.
7
Here is another funny example: Take the square (0, 1)2 and put the topology in which a set U is open if U ∩ {x = a}
is open in {x = a} = (0, 1) for every a ∈ (0, 1). This space has uncountably many connected components, each of which is
homeomorphic to R. However it is not a smooth manifold because it is not second countable.
 17

2.3. Complex manifolds. By replacing Rn by Cn and replacing “smooth” everywhere in the above by
“holomorphic” (i.e. complex analytic) we get the notion of a complex manifold. So a complex manifold
is (determined by) a nice topological space M with a covering by open sets which are homeomorphic
to open subsets of Cn such that the transition functions are holomorphic functions. Since holomorphic
functions are smooth functions of the underlying real variables M is also a smooth manifold of (real)
dimension 2n.
(Note that a complex function is differentiable once if and only if it is analytic so there is only one
notion of complex manifold, in contrast to the real case.)
Much, but not all, of what is said below about smooth manifolds also holds for real-analytic and
complex manifolds. (Exercise: Decide which statements and their proofs are valid in these categories.)

2.4. Partitions of unity. Recall that in a paracompact Hausdorff space any open covering has a con-
tinuous partition of unity subordinate to it. (Recall that if {Uα }α∈A is an open covering of a topo-
logical space X then a partition of unity subordinate
P to it is a collection of continuous functions
{fα : X → [0, 1]} such that supp(fα ) ⊂ Uα , α f α ≡ 1 and {supp(fα )}α∈B is locally finite. It fol-
lows that {int(supp(fα )}α∈A is a locally finite open cover refining {Uα }α∈A with the same indexing
set.) We will not actually use this fact, which is proved using Urysohn’s lemma. Instead, for a smooth
manifold, the existence of smooth bump functions in Rn allows us to construct a smooth partition of
unity subordinate to any covering.
Lemma 2.4.1. If {Uα }α is an open covering of a manifold then one can find a countable locally finite
refinement by coordinate charts {(Vβ , ϕβ : Vβ → Rn )}β∈B and an open covering {Wβ }β∈B refining {Vβ }
such that W β ⊂ Vβ and W β is compact for all β.

Proof. Since M is second countable it has a countable base consisting of charts. Since M is also locally
compact it has a countable base of open sets with compact closure. (If {Un } is a countable base then
for any point x there is a neighbourhood Vx of x with compact closure, and hence a Un=n(x) ⊂ Vx with
x ∈ Un . Then Un has compact closure (since U n ⊂ V x is closed) and contains x. Thus the Ui with
compact closure already form a basis at each point, hence a base.) Thus we can assume that the original
covering {Uα } in the lemma is a countable covering {Un } by charts with compact closure.
So let {Un }n≥1 be a countable covering of M consisting of charts with compact closure. Replacing
Un by ∪j≤n Uj we can assume that {Un } is a nested collection of open sets with compact closure. Since
U n is compact and is covered by the collection we must have that U n ⊂ Um for some m large enough.
Replacing Un+1 by Um we can arrange that we have a nested collection of open sets with compact closures
and U n ⊂ Un+1 . The sets Ωn = U n − Un−1 (with Ω1 = U 1 ) are compact and Un ∩ Ωm = ∅ for m > n.
(8)
We finish the proof of the lemma as follows: For each x ∈ Ωn choose a coordinate neighbourhood
(Vx,n , ϕx,n ) containing x and contained in U n+1 − Un−2 . Let Bx,n be a compact subset of ϕx,n (Vx,n )
with nonempty interior (e.g. a ball in Rn ). Then the sets ϕ−1x,n (int(Bx,n )) form a covering of Ωn and
hence we can choose a finite subcovering, which we call {Wβ }β∈B(n) (here B(n) is the finite index set).
Then the corresponding Vx,n also form a finite covering indexed by B(n). This defines the locally finite
refinement {Vβ }β∈B indexed by B = ∪n B(n) and its refinement {Wβ }β∈B as in the lemma. (9) 
Lemma 2.4.2. If B is a closed ball in Rn then there is a smooth bump function supported exactly on
B, i.e. a smooth function f : Rn → [0, 1] with {f 6= 0} = int(B).

Proof. Exercise. 
Lemma 2.4.3. A smooth manifold has a smooth partition of unity subordinate to any open cover. In
other words, if {Uα }α∈A is an open covering of M then there exist smooth functions fα : M → [0, 1] such
that
(i) supp(fα ) = closure of {fα 6= 0} is contained in Uα
(ii) {supp(f
P α )}α∈A is locally finite
(iii) f
α α ≡ 1

8Since M is the union of the compact sets Ω we see that M is σ-compact.

i
9The proof shows the following: A second-countable and locally compact Hausdorff space is paracompact. (Indeed, in
the proof just drop the requirement that Vβ is a chart and drop the Wβ entirely.) In particular, any open subset of a
smooth manifold is again a smooth manifold (since this remark ensures paracompactness).
18

(Note that (iii) ensures that {{fα 6= 0}}α is an open cover and (i) implies that it is a locally finite
refinement of {Uα }α (with the same indexing set). )

Proof. Note that if a refinement of an open covering has a partition of unity subordinate to it then so
does the original covering. Indeed, if {Uα }α∈A is refined by {Vβ }β∈B and fβ is the partitionP of unity
subordinate to {Vβ }β∈B then for any function I : B → A such that Vβ ⊂ UI(β) , gα = β∈I −1 (α) fβ
defines a partition of unity subordinate to {Vβ }β∈B . So it is enough to construct a partition of unity for
a refinement.
Starting with an arbitrary open cover {Uα }α we can take a locally finite refinement by charts
{(Vβ , ϕβ )}β and a refinement {Wβ } with W β ⊂ Vβ compact as in the previous lemma. For each β ∈ B
cover the compact subset ϕβ (W β ) of Rn by finitely many balls Bβ,1 , . . . , Bβ,r which are contained in
ϕβ (Vβ ). Choose C ∞ bump functions fβ,i for i = 1, . . . , r as in the previous lemma, i.e. with fβ,i (x) 6=
0 ⇔ x ∈ int(Bβ,i ). Let fβ = i fβ,i . This is a function from Rn to [0, ∞) which is > 0 on ϕβ (W β )
P
and which vanishes outside ∪i Bβ,i . Thus we can consider it as a smooth function on M which is zero
P Vβ and is given by fβ (ϕβ (x)) for x ∈ Vβ (and so is > 0 on Wβ ). Then supp(fβ ) ⊂ Vβ and
outside
0 < β fβ < ∞ on the whole of M (> 0 because Wβ cover M and < ∞ because {Vβ }β is locally finite)
P
and then dividing each one by β fβ gives the partition of unity subordinate to {Vβ }. By the remark
of the previous paragraph we are finished. 

We will also need to use the following later:

Lemma 2.4.4. Let M be a smooth manifold, C ⊂ M a closed subset, and g : C → R a function. Suppose
that g satisfies:
(∗) for each p ∈ C, there is a neighbourhood U of p in M and a smooth function g̃ on U such that
g̃|C = g.
Then there exists a smooth function g̃ : M → R such that g̃|C = g.

Proof. Let Uα be a covering by charts such that for each point p, the neighbourhood U in property (∗)
contains a Uα containing p. Then we have a smooth function gα on each Uα which P restricts to g on
C ∩ Uα . Let fα be a partition of unity subordinate to the Uα . The function g̃ = α gα fα then agrees
with g on C and is smooth on M . 

Another application of partitions of unity:

Proposition 2.4.5. If C ⊂ M is a closed subset of a smooth manifold then there exists a smooth function
f : M → R such that C = f −1 (0).

Proof. First consider the case where M = U is an open subset of Rn . The complement U −C is a countable
union of open balls Bi . Choose smooth functions fi : U → [0, ∞) such that fi (x) > 0 ⇔ x ∈ Bi and
such that the first i derivatives of fi are bounded on Bi (uniformly)
P by 1/2i . (The second condition can
always be achieved by multiplying by a scalar.) Now f = i fi converges uniformly on U , so that the
function f is smooth. Obviously f (x) = 0 iff x ∈ C.
Now consider the general case. Take a locally finite covering by charts {Uα } and a partition of unity
{gα } subordinate to it. For each α the intersection C ∩ Uα is a closed subset so using
P the previous case
let fα be a smooth function on Uα with fα−1 (0) = C ∩ Uα . Consider the sum f = α gα fα , which makes
sense as a smooth function on M . Then f −1 (0) = C. 

This shows that the preimage of a point under a smooth map can be arbitrarily bad. We will soon
see a condition ensuring that the preimage is good (e.g. a manifold).
(Nothing in this section holds for real-analytic or complex manifolds as there cannot be real-analytic
or holomorphic partitions of unity on them, for obvious reasons.)

2.5. Tangent and cotangent spaces. Let M be a smooth manifold and p ∈ M . A germ at p is an
equivalence class of pairs (U, f ) consisting of an open neighbourhood U of p and a smooth function f on
U , modulo the equivalence relation that (U, f ) ∼ (V, g) if there is an open neighbourhood W ⊂ U ∩ V on
which f and g agree. Let Gp be space of germs of smooth functions at p. This is an R-algebra. The ideal
mp of germs vanishing at p has codimension one, i.e. Gp /mp = R (by evaluation at p). The subspace
of germs with vanishing first partial derivatives (in some local coordinate system) at p is denote Sp and
called the space of stationary germs. (It does not depend on the coordinate system because under a
 19

local change of coordinates the first partial derivatives change by the Jacobian matrix of the coordinate
change, so they vanish in any other coordinate system).
The cotangent space to M at p is the vector space
Tp∗ M := Gp /Sp = mp /mp ∩ Sp
of germs modulo stationary germs (or germs vanishing at p modulo stationary germs vanishing at p).
The tangent space to M at p is the dual vector space
Tp M := (Tp∗ M )∗ .
Lemma 2.5.1. Let x1 , . . . , xn be local coordinates centred at p ∈ M . Any germ f ∈ mp can be written
Pn ∂f
as f = i=1 xi gi for some smooth germs gi with ∂x i
(0) = gi (0). If f is also stationary, i.e. f ∈ mp ∩ Sp
then the gi can be chosen to vanish at 0, i.e. gi ∈ mp , so that mp ∩ Sp = m2p .

Proof. Since we are dealing with germs we can always restrict to neighbourhoods of p which, in the local
coordinates, are starlike in Rn . (U ⊂ Rn is starlike if it is invariant under dilation by elements of [0, 1].)
Then one can write, for (x1 , . . . , xn ) ∈ U
Z 1
d
f (x1 , . . . , xn ) = f (tx1 , . . . , txn ) dt
0 dt
Z 1X n
∂f
= (tx1 , . . . , txn ) · xi dt
0 i=1 ∂xi
n Z 1
X ∂f
= xi · (tx1 , . . . , txn ) dt
i=1 0 ∂x i

Both assertions of the lemma follow. 

∗
(It follows that we have an equivalent description of the tangent space as: Tp M = mp /m2p . (If
you have seen the definition of Zariski tangent space in algebraic geometry you will recognize this.))
This lemma P implies of the xi give a basis of Tp∗ MP. (Indeed, for f ∈ Gp write
P that the images P
f − f (p) = i xi gi = i xi (gi − gi (p)) + i xi gi (p). The germ f (p) + i xi (gi − gi (p)) is stationary.
They are obviously linearly independent.) The dual basis of Tp M consists of the linear functionals on
∂
germs ∂x 1
, . . . , ∂x∂n defined by:
∂ ∂f
(f ) = (0).
∂xi ∂xi
Summarizing:
Lemma 2.5.2. The tangent space Tp M is a vector space of dimension n. If x1 , . . . , xn are local coordi-
∂
nates centred at p then it is spanned by the tangent vectors ∂x 1
, . . . , ∂x∂n which are defined by
∂ ∂f
(f ) = (0).
∂xi ∂xi
n o
The dimension of the cotangent space Tp∗ M is n and the dual basis to ∂
∂xi is given by the germs of
the coordinate functions xi .

A derivation on M at p is a linear map D : Gp → R which satisfies the Leibniz rule: for two germs
f and g, D(f g) = f (p)D(g) + D(f )g(p). It is clear that D must vanish on the constant germs, and also
on m2p . So it is determined by its restriction to mp and gives an element of the tangent space. Let Derp
be the space of derivations at p. We have defined an injective map
Derp → Tp M.
∂
P
It follows also that D = i ai ∂xi
in local coordinates, where ai = D(xi ). This is an isomorphism because
n o
∂
∂xi is exactly the dual basis to the basis {xi } of mp /m2p .
Another description of the tangent space is the most intuitive. A smooth map γ : R → M is called
a smooth curve in M . Suppose that γ(0) = p. Then we can define a derivation Dγ as follows: For
any smooth germ f at p, let
d
Dγ (f ) = f ◦ γ.
dt t=0
20

(i.e. pullback using the map γ and then take the derivative at 0.) We will call this derivation the tangent
vector to γ at p. If we look at smooth curves in Rn this corresponds to the usual notion of “directional
derivative along the curve γ at p”. By choosing local coordinates at p we know that all derivations are
∂ ∂
linear combinations of the ∂x i
, and ∂x i
= Dγi where γi is the curve given by γi (t) = (0, . . . , 0, t, 0, . . . , 0)
in the local coordinates. So we have:
Lemma 2.5.3. Any derivation at p is of the form Dγ for some curve γ, so that every element of the
tangent space at p is the tangent to a smooth curve.
Remark 2.5.4. It is useful to keep in mind the following fact: If V ⊂ Rn is an open set, then choosing
∂
coordinates x1 , . . . , xn , gives tangent vectors ∂x 1
, . . . , ∂x∂n form a basis for the tangent space Tp V at any
∼
point p ∈ V , and give an isomorphism Tp V = Rn . So all tangent vectors at different points in U can
be thought of as belonging to the same Rn . (You are already used to this, you often compare vectors at
different points in Rn by “translation to the origin”.)
A more invariant way of saying this is the following: If V is a finite-dimensional real vector space,
considered as a manifold via any linear isomorphism V ∼ = Rn , then there are canonical identifications
Tv V = V for any v ∈ V .

2.6. Functoriality. Given a smooth map f : M → N between smooth manifolds, there is a linear map of
vector spaces dfp : Tp M → Tf (p) N for each p ∈ M . This can be seen in any of the equivalent descriptions
above, but let us see it using derivations. Given D ∈ Derp , define a derivation df (D) ∈ Derf (p) by
a 7→ D(a ◦ f ) for a germ a at f (p). Check that this is a derivation.
What is the map dfp in more concrete terms? If you write it out in local coordinates x1 , . . . , xm
∂ ∂
near p and y1 , . . . , yn near f (p), and use the bases ∂x i
of Tp M and ∂y i
of Tf (p) N , then dfp is just the
Jacobian matrix of partial derivatives, i.e.
∂fi
(dfp )ij =
∂xj
where f = (f1 , . . . , fn ) are the component functions of f , i.e. fi = yi ◦ f .
From this local calculation and the chain rule it is clear that if g : N → X is another smooth map,
then d(g ◦ f )p = dgf (p) ◦ dfp .
Note that a diffeomorphism f : M → N induces an isomorphism dfp : Tp M → Tf (p) N , so they
must be of the same dimension. In particular, Rn and Rm are not diffeomorphic if m 6= n, so that
the dimension of a manifold is a diffeomorphism invariant. (Note that to prove that Rn and Rm are
nonhomeomorphic for n 6= m (“invariance of domain”) requires nontrivial algebraic topology arguments.)
Note that if γ : R → M is a smooth curve in M with γ(0) = p then in local coordinates xi around
p,
 
∂ X dγi ∂
dγ0 = Dγ =
∂t i
dt t=0 ∂xi

where γi (t) = xi (γ(t)).

A geometric way to think about the map dfp : Tp M → Tf (p) N is using tangents to smooth curves.
The map dfp simply sends the tangent vector to γ : (−, ) → M at γ(0) = p to the tangent vector to
f ◦ γ : (−, ) → N at f (p).

2.7. Tangent and cotangent bundles. Let T M be the set of all pairs (p, v) where p ∈ M and v ∈ Tp M
is a tangent vector at p. There is a map π : T M → M by (p, v) 7→ p. The space T M is called the tangent
bundle of M ; let us see that it is itself a smooth manifold of dimension 2n (n = dim(M )). For this we
need to put a topology (second countable, Hausdorff and paracompact) and provide an smooth atlas of
charts. Choose an atlas for M and let ϕ : U → V ⊂ Rn be a chart of the atlas P giving ∂local coordinates
x1 , . . . , xn . A tangent vector at any point of U is then of the form v = i qi (v) ∂xi . This gives a
homeomorphism π −1 (U ) ∼ = U × Rn ∼= V × Rn by sending (p, v) to (x1 (p), . . . , xn (p), q1 (v), . . . qn (v)). Put
on T M the topology generated by the open sets of the form W × W 0 where W ⊂ π −1 (U ) and W 0 ⊂ Rn
are open. (You must check that finite intersections of such sets are again of this form.) It is easy to see
that T M is second countable, Hausdorff and locally compact (and hence also paracompact). Moreover
there is an obvious atlas and the transition maps are smooth, so that the sets π −1 (U ) give a smooth
atlas for T M , making it into a smooth manifold of dimension 2n.
 21

Given a smooth map f : M → N there is an induced map on the tangent bundle T f : T M → T N

defined by T f (p, v) = (f (p), dfp (v)). It is smooth. If g : N → W is another smooth map then it satisfies
T (gf ) = T g ◦ T f . So T is a functor.
• Check that T(p,q) (M × N ) = Tp M × Tq N (natural isomorphism of vector spaces) and T (M × N ) =
T M × T N (natural diffeomorphism of manifolds).
The tangent bundle of Rn is identified with Rn × T0 Rn ∼= R2n by the remark earlier.
Similar remarks apply to the cotangent bundle, which is denoted T ∗ M .

2.8. Examples.
Hypersurfaces in Rm+1 . Let f : R1+m → R be a smooth function. When is the zero-set M :=
{f (x0 , . . . , xm ) = 0} a smooth manifold? (This is a special case of a more general situation discussed in
the next section.) It might be useful to keep in mind a situation like the function f (x, y, z) = x2 + y 2 − z 2
in R3 . For each point p ∈ M we must produce a chart containing p, so that the collection of charts form
a smooth atlas. If the partial derivative
∂f
(p) 6= 0
∂x0
then by the implicit function theorem there is a mapping g from an open neighbourhood U of (p1 , . . . , pm )
to an open neighbourhood of p0 ∈ R, i.e. a smooth function g(x1 , . . . , xm ) with f (g(x1 , . . . , xm ), x1 , . . . , xm ) =
0. Now the map U → M by
(x1 , . . . , xm ) 7→ (g(x1 , . . . , xm ), x1 , . . . , xm )
∂f ∂f
is a chart at p. On the other hand, if ∂x 0
(p) = 0 but ∂x i
(p) 6= 0 for i > 0 then we can do the same
argument with xi playing the role of x0 to produce a chart at p. This means that for every p ∈ M such
that  
∂f ∂f ∂f
∇f (p) := (p), ,..., (p) 6= (0, . . . , 0)
∂x0 ∂x1 ∂xm
we have produced a chart at p. I leave it to you to check that this is a smooth atlas. So M is a manifold
if ∇f (p) 6= 0 at every point of M .
• Show that the tangent space map dfp : Rm+1 → R is given by v 7→ ∇f (p) · v.
Now we can use this to show that several standard examples are smooth manifolds: The sphere S n
is the zero-set of f (x0 , . . . , xn ) = x20 + x21 + · · · + x2n − 1. Since ∇f = 2(x0 , . . . , xn ) is nonzero when f = 0,
this is a smooth manifold. (Note that another way to produce a smooth atlas on S n is by using the
homeomorphisms S n − {(0, 0, . . . , 1)} ∼ = Rn and S n − {(0, 0, . . . , −1)} ∼= Rn via stereographic projection.
Check that this atlas is compatible with any one defined by the implicit function argument above, so
that we get the same smooth structure because the maximal atlas containing both is the same.)
Example: Real projective spaces. Consider the real projective space RP n . One description of
it is as RP n = S n / ∼ where x ∼ −x.
• Use this description to get an atlas.
Another description is RP n = Rn+1 − {0}/R∗ or rather n + 1-tuples of numbers (x0 , . . . , xn ) 6=
(0, . . . , 0) upto scaling by nonzero reals. Let
Ui = {(x0 , x1 , . . . , xn ) ∈ Rn+1 − {0} : xi 6= 0}/R∗ .
 
Then ϕi : Ui → Rn by ϕi (x0 , . . . , xn ) = xx0i , . . . , c
xi
xi , . . . , xn
xi is a homeomorphism, so it gives a chart.
n
The Ui form an open covering of RP , so this is an atlas. Let us check that this is a smooth atlas. The
transition function ϕi ◦ ϕ−1
j is defined on the open subset {ti 6= 0} ⊂ Rn and maps to {tj 6= 0} ⊂ Rn . It
is given by multiplication by t−1
i tj , which is smooth.
Example: Complex projective spaces. The complex projective space CP n = Pn is a smooth
manifold of dimension 2n. Since Pn = Cn+1 − {0}/C∗ we can describe an atlas of charts by
Ui = {(z0 , z1 , . . . , zn ) ∈ Cn+1 − {0} : zi 6= 0}/C∗ .
 
The isomorphism Ui = ∼ Cn = R2n by (z0 , . . . , zn ) 7→ z0 , . . . , zbi , . . . , zn gives the chart. As in the
zi zi zi
previous example, the transition functions are zi /zj , so that this is a smooth atlas. (Note that in
this case the charts of the atlas are modeled on (open subsets of) Cn and the transition functions are
holomorphic. This is an example of a complex manifold of dimension n. A complex manifold of dimension
n is also a real manifold of dimension 2n.)
22

Example: Hypersurfaces in Pn . Let f be a homogeneous polynomial in n + 1 variables. Its

zero-set in Cn+1 is invariant under dilations, so it defines a subset V ⊂ Pn which is in fact closed. So
we can ask: under what condition on f is V smooth manifold? (If f has real coefficients then we can
also ask the same question for the subset of RP n defined by f .) We can also ask the question for several
homogeneous polynomials f1 , . . . , fk . These questions will be answered in the next section.
Example: Grassmannians. For F = R or C (or H) let Gr(k, F n ) be the set of k-dimensional
F -subspaces (left F -submodules in the case F = H) of F n . This is a smooth manifold of dimension
d k(n − k). Recall how to produce an atlas. Fix E ∈ Gr(k, F n ) and choose a complement E ⊥ to E, i.e.
F n = E ⊕ E ⊥ , and let πE , πE ⊥ be the projections. Then consider the subset
UE := {E 0 ∈ Gr(k, n) : E 0 ∩ E ⊥ = {0}}.
For any E 0 ∈ UE the restriction πE |E 0 is an isomorphism. So
−1
ϕE 0 := πE ⊥ ◦ (πE |E 0 ) ∈ Hom(E, E ⊥ ).
It is easy to check that E 0 7→ ϕE 0 gives an identification of UE with Hom(E, E ⊥ ) ∼
= M at(k × n − k, F ) =
F k(n−k) ∼= Rdk(n−k) . We leave it to be checked that these are the charts of a smooth atlas making
Gr(k, F n ) into a smooth manifold. (In the case F = C this also shows that Gr(k, F n ) is a smooth
complex manifold of dimension k(n − k)).)

2.9. C k -manifolds. Instead of requiring the transition functions of an atlas to be smooth, one can
require them to be only C k for some k. The formal definition is the following: A C k -atlas on a space M
is a covering of M by charts {(Uα , ϕα : Uα → Vα ⊂ Rn )}α such that if Uα ∩ Uβ 6= φ then the transition
function ϕβ ◦ ϕ−1 k n
α : ϕα (Uα ∩ Uβ ) → ϕβ (Uα ∩ Uβ ) is a C map between open subsets of R . (Recall that
C k means that all derivatives of order ≤ k exist and are continuous.) A C k -manifold is a pair of a
space M and a maximal C k -atlas on M (as before, the maximal atlas is unique). The atlas allows us to
talk about the C k functions on a C k manifold, and also C k maps between C k manifolds.
In general, if k ≥ 1 things work quite similarly to the C ∞ case. (In particular, it is a theorem
that every C k (k ≥ 1) atlas on a C k manifold contains at least one C ∞ atlas. It is also a theorem
that if two C ∞ manifolds are diffeomorphic as C k manifolds for k ≥ 1 then they are actually smoothly
diffeomorphic.)
If k = 0 a C 0 manifold is usually called a topological manifold and a C 0 map is simply a continuous
map. The world of topological manifolds is quite different from the smooth case. (For example, it is
known that there are topological manifolds which admit no smooth atlas at all! Also, two smooth
manifolds may be homeomorphic without being smoothly diffeomorphic. Unfortunately the simplest
examples are already nonelementary.)
Note that for k ≥ 1 it makes sense to talk about the tangent space to a C k manifold: It is defined as
before as germs of C k functions modulo stationary germs of C k functions. The same proof as above shows
that the tangent space Tp M to a n-dimensional C k manifold has dimension n. A C k map f : M → N
induces a linear map df : Tp M → Tf (p) N which, in local coordinates, is given by the Jacobian.
From now on we will only work in the smooth category, so that the words “manifold, function,
mapping” etc. will always refer to smooth things, unless indicated otherwise.
Remark 2.9.1. A useful fact which we will not prove is the following: Let f : M → N be a continuous
map between C k manifolds for k ≥ 1 or k = ∞. If C is a closed subset such that f is C k on some open
neighbourhood of C then there exists a continuous map F : [0, 1] × M → M such that F |{0}×M = f ,
F |{1}×M is C k , and F{t}×C = f |C for all 0 ≤ t ≤ 1. In particular, taking C to be empty, we see that
any continuous map between smooth manifolds is homotopic to a smooth map. In particular, every
homotopy class of loops in a smooth manifold contains a smooth loop.

2.10. Inverse and implicit function theorems. In addition to the chain rule we will need to use the
inverse and implicit function theorems of multivariable calculus. These show that the linear approxima-
tion to a map (i.e. its derivative) controls the behaviour of the map up to change of coordinates, under
an assumption on the derivative (basically that it should have the maximal possible rank). (For proofs
of these see e.g. the book Ordinary Differential Equations by Arnol’d.)
Inverse function theorem: Let f : U → V be a map between open subsets U, V of Rn which is
smooth, i.e. all derivatives exist and are continuous. f is given by n differentiable functions f1 , . . . , fn ,
 23

each of n variables. Let p ∈ U be a point at which the Jacobian matrix

!
∂fi
dfp =
∂xj p

is nonsingular. Then there is a neighbourhood U 0 ⊂ U of p such that f : U 0 → V = f (U 0 ) is an

isomorphism, i.e. the inverse function g : V → U 0 defined by g(y) = x if f (x) = y is smooth and has
derivative dgy = (dfg(y) )−1 for y ∈ V .

If you are given n functions f1 , . . . , fn of n + m variables x1 , . . . , xn , y1 , . . . , ym you can ask whether

in the system of n equations
f1 (x1 , . . . , xn , y1 , . . . , ym ) = 0
·
·
·
fn (x1 , . . . , xn , y1 , . . . , ym ) = 0
it is possible to solve for the variables x1 , . . . , xn in terms of the y1 , . . . , ym . The implicit function theorem
gives a sufficient condition to do this:

Implicit function theorem: Suppose we are given n differentiable functions f1 , . . . , fn of the

n + m variables x1 , . . . , xn , y1 , . . . , ym . If (p, q) ∈ Rn × Rm such that f (p, q) = 0 and if the matrix
!
∂fi
d(f |Rn ×{q} )(p,q) =
∂xj (p,q)
i=1,...,n;j=1,...,n

is nonsingular at (p, q) then there exist open neighbourhoods U of q in Rm and V of p in Rn and a

(unique, differentiable) map g : U → V such that g(q) = p and f (g(y), y) = 0 for all y ∈ U . The
derivative of g is given by
!−1 !
∂fi ∂fi
−1

dgy = − = − d(f |Rn ×{y} )(g(y),y) d(f{g(y)}×Rn )( (g(y),y)
∂xj (g(y),y) ∂yj (g(y),y)

for y ∈ U . Moreover, there is a neighbourhood W of (p, q) such that

W ∩ {(x, y) : f (x, y) = 0} = {(g(y), y) : y ∈ U },
i.e. every solution of f (x, y) = 0 near (p, q) is of the form (g(y), y).

These two theorems are equivalent: To show that the implicit function theorem implies the inverse
function theorem, suppose f : U → V is a smooth map between open subsets of Rn with dfp (=the
Jacobian matrix at p) nonsingular. Consider the mapping F : U × Rn → Rn by F (x, y) = y − f (x).
The n functions Fi (x1 , . . . , xn , y1 , . . . , yn ) = yi − fi (x1 , . . . , xn , y1 , . . . , yn ) of 2n variables satisfy the
hypotheses of the implicit function theorem. Therefore there is a mapping g = (g1 , . . . , gn ) from an open
neighbourhood of f (p) ∈ Rn to the open subset U such that F (g(y), y) = 0, i.e. f (g(y)) = y, i.e. f ◦ g is
the identity.
Conversely, given a smooth mapping f : U → V with U ⊂ Rn+m and V ⊂ Rn , we will apply the
inverse function theorem to F : U → Rn+m given by F (x, y) = f (x, n m
 y) + y where x ∈ R and y ∈ R .
∂fi
The differential dF(x,y) is block upper-triangular with the matrix ∂xj and the m × m
i=1,...,n;j=1,...,n
identity on the diagonal. Thus under the hypothesis of the implicit function theorem it is invertible at
(p, q), and so by the inverse function theorem F has a local inverse, i.e. there is a neighbourhood W of
(p, q) in which it has a smooth inverse G. Now let g(y) = G(0, y), defined for y in the open neighbourhood
U = {y ∈ Rm : (0, y) ∈ W } of q. The identities F G = id and GF = id lead to the conclusions of the
theorem.
The following immediate consequence of the inverse function theorem is useful: A bijective smooth
map f : M → N such that dfp is an isomorphism everywhere is a diffeomorphism.
Remark 2.10.1. Note that the real-analytic version of the inverse function theorem holds, i.e. are real-
analytic map with nonsingular differential at a point has real-analytic inverse in a neighbourhood of the
point. Similarly for complex-analytic. Also the version with “smooth” replaced by C r for any r ≥ 1.)
24

2.11. Immersions and submersions. Consider a smooth map f : M → N of smooth manifolds of

dimension m and n respectively. The map f is an immersion at p ∈ M if dfp : Tp M → Tf (p) N is
injective. It is a submersion at p ∈ M if dfp : Tp M → Tf (p) N is surjective. These are both open
conditions, i.e. if they hold at a point p then they must hold in a neighbourhood of the point. To
see this, if f is an immersion at p, then dfp , considered as a linear transformation from Tp M ∼ = Rm to
∼ n
Tf (p) N = R is an n × m matrix with rank m. An n × m matrix has rank m if and only if it has a
nonsingular m × m submatrix. Since determinant is continuous in the entries, that same submatrix is
nonsingular at points close to p, which means that the rank of dfq is still m if q is close to p. The same
reasoning applies to submersions, because f is a submersion at p means dfp has rank n.
Consider the example of f : Rn → R by f (x1 , . . . , xn ) = −x21 + x22 + · · · + x2n . The map dfp at
p = (x1 , . . . , xn ) is the 1 × n matrix (−2x1 , 2x2 , . . . , 2xn ), so f is a submersion at all points except
0 ∈ Rn . Note also that f −1 (t) is a manifold except at t = 0. We will see below that this happens exactly
because f is submersive at all points of f −1 (t) if t 6= 0.
We say that f : M → N is a submersion (resp. immersion) if it is a submersion (resp. immersion)
at all p ∈ M .

2.12. Submanifolds. Let M be a smooth manifold. There are various possible notions of submanifold
which need to be distinguished.
A submanifold is a pair (W, i) consisting of a smooth manifold W and a smooth map i : W → M
which is a 1-1 immersion. (This is sometimes also called an immersed submanifold.)
A submanifold i : W → M which induces a homeomorphism of M with the subspace i(W ) ⊂ N
(with the induced topology) an embedded submanifold and we say that i is an embedding.
If i : W → M is a 1-1 immersion which is proper then then i : W → M will be called a closed
submanifold. (The terminology makes sense because a proper map between manifolds is necessarily
closed, so that i(W ) is closed in N .) If a 1-1 immersion i is proper then i : W → i(W ) is a homeomor-
phism, so that a closed submanifold is embedded. In particular, if W is compact then this holds, so a
1-1 immersion i : W → M with W compact is an embedding.
Remark 2.12.1. It is important to remember that a submanifold of M is not simply a subset of M . Here
are some examples: For an irrational α ∈ R − Q consider the map i : R → R2 /Z2 by t 7→ (t, αt) mod Z2 .
This is an immersion with dense image (by Kronecker’s theorem). It is evidently not an embedded
submanifold since the induced topology on i(W ) is not the usual one. An example of an embedded
submanifold which is not closed is (0, 1) ⊂ R.
2.13. Normal forms for immersions and submersions. The inverse and implicit function theorems
imply some properties of immersions and submersions. The first one says that an immersed submanifold
can be “straightened out” by a suitable choice of local coordinates.
Proposition 2.13.1 (Local normal form for immersions). Let f : M m → N n be an immersion at p ∈ M .
Then there exist charts U at p and U 0 at f (p) and systems of coordinates x1 , . . . , xm on U and y1 , . . . , yn
on U 0 such that f : U → U 0 is given by f (x1 , . . . , xm ) = (x1 , . . . , xm , 0, . . . , 0), i.e. f looks locally like
the inclusion Rm ⊂ Rn .

Proof. Choose local charts φ : U → V ⊂ Rm and ψ : U 0 → V 0 ⊂ Rn at p and f (p) respectively, so that

φ(p) = 0 and ψ(f (p)) = 0. We want to show that by changing the local coordinates on U 0 we can make
ψ ◦ f ◦ φ−1 the map induced by the inclusion i : Rm → Rn as the first m coordinates. So we are in
the following situation: f : V → V 0 is a smooth map between open neighbourhoods of 0 in Rm and Rn
respectively. It is enough to find an open W 0 ⊂ V 0 and a diffeomorphism T : W 0 → T (W 0 ) such that
T −1 ◦ f 0 = i on W 0 .
First of all we can clearly assume that df0 : T0 V → T0 V 0 is the inclusion i : Rm → Rn . (Indeed, by
a linear change of coordinates S : Rn → Rn we can arrange this by replacing f by S −1 ◦ f .) Now define
a smooth mapping T : V × Rn−m → Rn by
T (x, y) := f (x) + (0, y)
m n−m
where (x, y) ∈ R × R are coordinates on Rn . At y = 0 we get the map f , so that f = T ◦ i. Then
dT(0,0) (x, y) = df0 (x) + (0, y) = (x, y)
so dT(0,0) is the identity T(0,0) (V × Rn−m ) → T(0,0) Rn = T0 V 0 . By the inverse function theorem, there
is a small neighbourhood W 0 ⊂ V 0 of (0, 0) on which there is a smooth inverse T −1 : W 0 → V × Rn−m
and T −1 ◦ f = i on W 0 . 
 25

This implies (and is essentially the same as) the statement that any smooth function on M extends,
locally on M , to a smooth function on N . Note that this is only true locally on M (consider the
immersion R → R2 /Z2 given by a line of irrational slope, which is not an embedding). If M is an
embedded submanifold of N , then it says that the smooth functions on M are exactly the restrictions
of smooth functions from N .
Proposition 2.13.2 (Local normal form for submersions). Let f : M m → N n be a submersion at
p ∈ M . Then there is a chart U at p and a chart U 0 of N at f (p) such that f (U ) ⊂ U 0 and f |U looks
like the projection Rm = Rm−n × Rn → Rn .

Proof. Since the question is local we may assume that f maps an open subset of Rm to an open subset of
Rn . Write Rm = Rm−n ×Rn and write (x, y) for the coordinates in each factor. Let p be a point at which
f is submersive. We may arrange by a change of coordinates in Rm that dfp |{0}×Rn is nonsingular. Define
F : Rm → Rm by F (x, y) = (x, f (x, y)). The differential dFp is nonsingular. By the inverse function
theorem we know that F is invertible in a neighbourhood W of p. Let W 0 ×V 0 be a product neighbourhood
(w.r.t. Rm = Rn−m × Rm ) in F (W ). Then F −1 : W 0 × V 0 → W is given by F −1 (x, y) = (x, g(x, y)) for
some smooth function g from m variables to n variables, and we have
(x, y) = F ◦ F −1 (x, y) = F (x, g(x, y)) = (x, f (x, g(x, y))
so that f (x, g(x, y)) = y, i.e. f F −1 (x, y) = y and for y ∈ V 0 , f −1 (y)∩W is the graph of W 0 3 x 7→ g(x, y).
In particular, in the coordinates given by F −1 , f is the projection. 

(Remark: Show that this proposition implies the implicit function theorem. This again shows that
the inverse function theorem implies the implicit function theorem.)
Remark 2.13.3. A C k map f : M → N induces a linear map df : Tp M → Tf (p) N which, in local
coordinates, is given by the Jacobian. The notions of C k immersion, C k submersion are defined in the
obvious way and they have normal forms as above (the same proofs given above work because for k ≥ 1
one can replace the word “smooth” by “C k ” everywhere in the inverse and implicit function theorems
and they remain true).

2.14. Critical point, critical value, regular value. Let f : M → N be a smooth map. A point
p ∈ M is called a critical point for f if dfp has rank < n = dim N , i.e. f is not a submersion at p. A
point q ∈ N is called a critical value if f −1 (q) contains a critical point. The complement of the set of
critical values is the set of regular values. Thus if q ∈ N is a regular value and p ∈ f −1 (q) then f is a
submersion at p. (The fibre over a regular value is allowed to be empty.) For example, if m < n then
any point is critical and hence every point of f (M ) is a critical value.
The normal form for a submersion implies the following (fill in the necessary details):
Corollary 2.14.1. If f −1 (q) 6= ∅ and f : M m → N n is a submersion at all points in f −1 (q) (i.e., q is
a regular value of f ), then f −1 (q) is an embedded submanifold of M of dimension m − n. The tangent
space to f −1 (q) at p is identified with the subspace ker(dfp : Tp M → Tq N ).
Examples 2.14.2. Let us use the corollary to get some interesting examples of manifolds as “level sets”
of smooth maps.
(i) Let f : Rn → R by f (x1 , . . . , xn ) = x21 + · · · + x2n . Then f is submersive at all points of f −1 (λ)
for λ > 0, so that the fibres f −1 (λ) are smooth manifolds of dimension n − 1, which are all
diffeomorphic to the sphere S n−1 .
(ii) For f : Rn → R given by f (x1 , . . . , xn ) = −x21 + x22 + · · · + x2n we showed that f is submersive at
all points of f −1 (t) for any t 6= 0. So f −1 (t) is a manifold for t 6= 0.
(iii) (Hypersurfaces) Let f : Rn+1 → R be a smooth function. The induced map Rn+1 → R on
tangent spaces is given by v 7→ ∇f (p) · v. (Check this.) So a level set {f = λ} is a manifold
if ∇f (p) is nonzero at all points p ∈ {f = λ}. This is exactly what we saw earlier. Note also
that since M is a closed submanifold the tangent space Tp M is a subspace of Tp Rm+1 = Rm+1 .
What is this subspace? By the corollary it is Tp M = {v ∈ Rm+1 : v · ∇f (p) = 0}.
(iv) Let GL(n, R) be the group of nonsingular n×n matrices with real entries. It is the complement in
M (n, R) ∼
2
= Rn of the closed set det = 0. So it is a manifold. To show that SL(n, R) = det−1 (1)
is a manifold we can directly check the gradient condition ∇det(g) 6= 0 for g ∈ SL(n, R). We will
instead compute the tangent map ddet and show that det : GL(n, R) → R∗ has 1 as a regular
value, which will also give us the tangent space as a subspace of M (n, R).
26

Since GL(n, R) is identified with an open subset of the real vector space M (n, R) of all n ×
n matrices with real entries, its tangent space at the identity I is naturally M (n, R). Then
TI SL(n, R) ⊂ TI GL(n, R) = M (n, R) is ker(ddet), so we must compute the differential ddet :
M (n, R) → R. To do so, for A ∈ M (n, R) choose a path through I which will have as tangent
vector a given matrix A ∈ M (n, R). There are many ways to choose one; the easiest is γ(t) =
I + tA. For small enough t this is invertible and the tangent vector to γ at I is A. To compute
the image of A under ddet we compute
d
det(I + tA) = tr(A).
dt t=0
So ddetI = tr : M (n, R) → R is surjective and TI SL(n, R) = ker(tr) is the space of trace zero
matrices.
For g ∈ SL(n, R) one can make a similar computation of the map ddetg : Tg GL(n, R) →
T1 R∗ = R by using the curve γ(t) = g + tA. Another way to do this is to use the diffeomorphism
left translation Lg−1 : GL(n, R) → GL(n, R) by h 7→ g −1 h. Then det = det ◦ Lg−1 and hence
ddetg = ddetI ◦ (dLg−1 )g must also be surjective. (This argument also computes Tg SL(n, R) as
a subspace of M (n, R).)
(iv) Consider the map f : GL(n, R) → M (n, R) given by f (x) = t xx. It is surjective onto the space of
symmetric positive definite matrices, and it is submersive at all points. (Exercise.) In particular,
the identity matrix I ∈ M (n, R) is a regular value, so the preimage f −1 (I) = O(n) is a smooth
manifold.
(v) • Show that SU (n) has a manifold structure by doing something similar to (iv).

2.15. Transversality. Let f : M → N be a smooth map, and suppose that W ⊂ N is an embedded

submanifold. We say that f is transverse along W if
Tf (p) N = dfp (Tp M ) + Tf (p) W for all p ∈ f −1 (W ).
(The sum need not be direct.)
Two special cases of this definition are: (1) if f is an embedding of M in N then we say M intersects
W transversely if f is transverse along W (2) if f is submersive at all points of f −1 (q) for q ∈ N , then
f is transverse along {q}. The appropriate generalization of the previous corollary is
Proposition 2.15.1. If f : M m → N n is a smooth map transverse along an embedded submanifold
W w ⊂ N n then f −1 (W ) is an embedded submanifold of M of dimension m − n + w.

• Prove this by generalizing the proof of the previous proposition.

Here are two simple examples illustrating the proposition: First consider the intersection in R3 of
the surface M = {z = xy} and the plane W = {z = λ}. It is easy to check that for λ 6= 0 the intersection
is transverse and M ∩ W is a manifold. For λ = 0 the intersection fails to be transverse at (0, 0, 0) (the
two tangent spaces are the same plane) and the intersection is {xy = 0 = z}, which is a pair of crossing
lines. This is not a manifold. Second consider the intersection in R3 of the manifolds given by equations
x2 + y 2 + z 2 = 1 (2-sphere) and x2 + y 2 − z 2 = λ (λ 6= 0). For λ 6= 1 the intersection is transverse and
consists of two circles. For λ = 1 the intersection is not transverse, but nonetheless it is a manifold,
though not of the dimension given in the theorem.

2.16. Whitney’s embedding theorems. With our abstract definition of manifold it is not obvious
whether every manifold can be embedded as a submanifold of RN for some N .
Theorem 2.16.1. A compact smooth manifold can be embedded in RN for some N .

Proof. We will use a partition of unity. From Lemma 2.4.1 we can assume the following: We have a
finite (because M is compact) covering by charts (Vi , φi : Vi → Rn ) and an open refinement Wi with
W i ⊂ Vi (here i = 1, . . . , r). Let {fi }i be a partition of unity subordinate to the Vi with fi |W i ≡ 1 (as
was constructed in the proof of existence of partitions of unity). Define a map Φ : M → RN =nr+r by
Φ = (f1 φ1 , . . . , fr φr , f1 , . . . , fr ).
Then dΦ = (d(f1 φ1 ), . . . , d(fr φr ), df1 , . . . , dfr ). Since, for any p ∈ M , some fi is identically one near p,
and φi is a chart near p, we see that Φ is an immersion at p. So it is an immersion. Let us see that it
is 1-1. If Φ(p) = Φ(q) then fi (p) = fi (q) for all i. In particular, there is some i for which these are both
equal to one (since the Vi also cover M ), and then p and q belong to the closure of Vi . But V i ⊂ Ui , so
p, q ∈ Ui . But then φi (p) = φi (q) and hence p = q.
 27

This proves that Φ is a 1-1 immersion of M in RN . Since M is compact it is a homeomorphism onto

its image, i.e. Φ is an embedding. 

(We will improve this to N = 2n + 1 below using Sard’s theorem.)

Whitney’s embedding theorem tells us that two ways of defining manifolds – abstractly as we have
or as submanifolds of Euclidean space – are equivalent. For example, this tells us that our manifolds are
metric spaces. (Of course this also follows from the Urysohn metrization theorem, which is much more
general.)
Recall that q ∈ N is a critical value of a smooth map f : M → N if there is a critical point in
f −1 (q), i.e. a point p ∈ f −1 (q) such that f is not submersive at p.
Theorem 2.16.2 (Sard’s theorem). The set of critical values of a smooth map f : Rm → Rn is of
measure zero.

In particular, the set of critical values has no interior. The set of regular values is of full measure,
so in particular it is nonempty. (There is another version of Sard’s theorem which says that the set of
critical values is nowhere dense, i.e. its closure has empty interior. This also implies that the set of
regular values is nonempty.)
We will only use the case m < n below so let’s only prove that case, which is very easy. If m < n
then every point is critical, so we must show that im(f ) is of measure zero.
Lemma 2.16.3. Let U ⊂ Rn be open and let f : U → Rn be smooth (C 1 is enough). If C ⊂ U has
measure zero then f (C) has measure zero.

Proof. It is enough to consider the case where the closure of C is contained in a compact set, otherwise
we could write it as a countable union of subsets with closure contained in compact sets and prove that
each has measure zero. For C with closure contained in compact set Ω ⊂ U , since df is continuous, there
is a constant µ such that
|f (x) − f (y)| ≤ µ |x − y| (x, y ∈ Ω).
(Use the mean value theorem and let µ be the supremum of kdf x k for x ∈ Ω.) For  > 0 cover C by balls
Bi (contained in Ω) such that i vol(Bi ) < /µn . By P
P
the previous estimateP f (Bi ) is contained in a ball
Bi0 of radius µ · radius(Bi ), hence f (C) ⊂ ∪i Bi0 and i vol(Bi0 ) ≤ µn i vol(Bi ) < . Thus f (C) has
measure zero. 

To prove Sard’s theorem for f : Rm → Rn with m < n apply the lemma to the composition of f
with the projection Rn → Rm , i.e. to the map Rm × Rn−m 3 (x, y) 7→ f (x). Since Rm × {0} ⊂ Rn
has measure zero we see that the image of f has measure zero. (The general version of Sard’s theorem
requires a little more work, see any standard text for a proof.)
Now we say that a subset Z ⊂ M of a smooth manifold has measure zero if there is a countable
covering by charts {(Ui , ϕi : Ui → Rn )} such that ϕi (Ui ∩ Z) has Lebesgue measure zero in Rn . (This
notion is independent of the choice of {Ui } by the lemma above.) With this definition, we have:
Corollary 2.16.4. The set of critical values of a smooth map f : M → N is of measure zero. In
particular, it has no interior.

We will use the consequence that the set of regular values is nonempty (if m < n).
Theorem 2.16.5 (Whitney). A smooth compact n-manifold can be embedded in R2n+1 and immersed
in R2n .

Proof. We start with an embedding M ⊂ RN for some N . Suppose that v is a vector in RN which
is never a tangent vector to M . (We are identifying the tangent spaces of all points of RN with RN
simultaneously.) Let π : RN → RN −1 be the orthogonal projection along v. It is easy to check that
dπp : Tp RN → Tπ(p) RN is, under the identifications Tp RN = RN and Tπ(p) RN −1 , simply the projection
π : RN → RN −1 . Then dπp is injective for p ∈ M , so that π|M : M → RN −1 is an immersion. If, in
addition, v is never a multiple of the vector p − q for p, q ∈ M , then it is also 1-1. Since M is compact
it is an embedding, i.e. a homeomorphism onto its image. So it remains to find such a vector v.
For a pair of points p, q ∈ M , p 6= q, let `(p, q) denote the line in RN spanned by p − q. Define a
map f : M × M − ∆ → RP N −1 = RN − {0}/R∗ by (p, q) → `(p, q). (Here ∆ ⊂ M × M denotes the
diagonal {(p, p) : p ∈ M }.) This is evidently a smooth map. If 2n < N − 1 then every point of image(f )
is critical, and so by Sard’s theorem image(f ) is of measure zero.
28

Define another map g : T M − M → RP N −1 by (p, v) 7→ Rv. Again, since 2n + 1 < N , image(g) is

of measure zero. So the union image(f ) ∪ image(g) ⊂ RP N −1 is of measure zero, so there is a point ṽ
in the complement. Any vector in RN projecting to v̄ is the one we want.
For the immersion we argue a little more carefully with the map g. It is easy to see that g factors
through a map X → RP N −1 where X = T M − M/R∗ is the quotient space. Now this quotient space is
naturally a manifold of dimension 2n − 1 making the map g smooth. Thus if 2n − 1 < N − 1 we know
(by Sard’s theorem) that the image is of measure zero. Projecting along a direction not in the image
will still give an immersion. (Instead of this quotient you can put a Riemannian metric on M and take
the unit sphere bundle in the tangent bundle for X and make the same argument.) 

Notice that the argument proves slightly more: If an n-manifold has a finite cover by open neighbour-
hoods homeomorphic to Rn then it has a 1-1 immersion in R2n+1 . (The projection argument nowhere
uses compactness.) This will be used below in the proof of the embedding theorem for noncompact
manifolds.
It is clear that this argument cannot work to get an embedding in R2n since we have no control
on the initial embedding. For example, if we start with M = S 1 we may in the first step end up with
an embedding like the embedding of the trefoil knot in R3 . There is no vector in R3 along which we
can project and get a 1-1 map to the plane. On the other hand, it is obvious that the circle can be
embedded in R2 . So Whitney’s theorem that M n can be embedded in R2n requires a new idea, the
famous “Whitney trick”.
Whitney proved some much stronger theorems: If k ≥ 2n + 1 then the embeddings are dense (in
a suitable topology) among all smooth maps M n → Rk . If k = 2n then the immersions are dense.
(See R. Narasimhan’s book Analysis on real and complex manifolds or his TIFR lecture notes Topics in
Analysis for some of this.)
We can now prove the noncompact version of the embedding theorem.
Theorem 2.16.6 (Whitney). A smooth manifold can be embedded in R2n+1 and immersed in R2n .

Proof. Recall that in the proof of Lemma 2.4.1 we showed that there are open sets Un with U n ⊂ Un+1
and U n compact such that M = ∪n Un . Using this we constructed countable open covers {Vβ }β∈B
and refinement {Wβ }β∈B with B = ∪n B(n) such that {Wβ }β∈B(n) is a finite cover of U n −Un−1 and
W β ⊂ Vβ ⊂ U n+1 −Un−2 if β ∈ B(n). For i = 0, 1, 2 let
Bi = ∪k≥0 B(i + 3k).
Then M has an open covering by three sets M = M0 ∪ M1 ∪ M2 where
Mi = ∪β∈Bi Vβ = ∪k≥0 Mi,k where Mi,k = ∪β∈B(i+3k) Vβ .
We will first immerse each Mi in a Euclidean space and then use the immersions to immerse M . To
immerse Mi note that on each finite union Mi,k = ∪β∈B(i+3k) Vβ one can define an immersion Φi,k using
the construction in the compact Whitney theorem (all that was required was a finite cover as in Lemma
2.4.1 and then a partition of unity as in Lemma 2.4.3). Thus Mi,k has an injective immersion Φi,k in RN
and the image, being contained in the union of the finitely many compact sets Φi,k (W β ) for β ∈ B(i+3k),
is contained in some ball. Moreover, the projection argument as in the 2n + 1 theorem says that we can
take N = 2n + 1. Translating the immersion of Mi,k by a multiple of a suitable fixed vector so that the
images of the various Φi,k for k ≥ 0 are disjoint, we get injective immersions Φi : Mi → RN =2n+1 for
i = 0, 1, 2.
Now take a partition of unity {f0 , f1 , f2 } subordinate to the covering {M0 , M1 , M2 } and consider
the map
Φ := (f0 Φ0 , f1 Φ1 , f2 Φ2 , f0 , f1 , f2 ).
It is easy to see that this is an injective immersion in R3N +3 . It only remains to see that it is an
embedding, which would follow if we show that the map Φ is closed. Now any closed subset C ⊂ M can
be written as a union C = C0 ∪ C1 ∪ C2 where Ci = ∪β∈Bi C ∩ (U i − Ui−1 ). This is a disjoint union of
compact sets, which maps under Φ to a disjoint union of compact sets which can be seen to be closed. 

Once we have a closed embedding in R3(2n+1)+3 , the same projection idea as in the compact case
reduces the dimension to 2n + 1 and 2n for an immersion. This can be improved to embedding in R2n+1
using the same argument as in the compact case:
 29

Theorem 2.16.7 (Whitney 2n + 1 embedding). A smooth n-manifold admits a closed embedding in

R2n+1 .

• Modify the previous proof to get that any smooth manifold embeds in R2n+1 . (Check the details
of this argument, in particular you have to ensure that the map is proper.)
Remark 2.16.8 (Real and complex analytic cases). It is natural to ask about a real-analytic version of
Whitney’s theorem. It is true that a real-analytic manifold has a real-analytic embedding in Euclidean
space (Grauert, Morrey), but it’s difficult. For complex analytic manifolds there can be no such theorem,
because e.g. a compact complex manifold has no everywhere-holomorphic functions, so that any map
to Cn has to be constant. A Stein manifold is a complex manifold which admits a closed embedding
in CN for some N . (This is one of many equivalent definitions.) Bishop and Narasimhan proved that a
Stein manifold of (complex) dimension n admits an embedding in C2n+1 .
Another theorem of Whitney is the following: A smooth n-manifold admits an embedding in R2n+1
such that the image is a real-analytic submanifold. Thus any smooth manifold has real-analytic structures
(but not a canonical one).

2.17. Tubular neighbourhoods. We must use the notion of vector bundle on a smooth manifold;
we will consider it again in more detail later. A (smooth) vector bundle on a smooth manifold M is
a smooth manifold E with a smooth mapping π : E → M such that there exists an open covering {Uα }α
of M and diffeomorphisms {ϕα : π −1 (Uα ) → Uα × Rr )}α such that π is given on π −1 (Uα ) by the second
projection composed with ϕα . The integer r is called the rank of the vector bundle. An example is the
tangent bundle, which is a rank n vector bundle over M . A section of π : E → M is a smooth map
s : M → E such that π ◦ s = id. The zero section is the mapping defined by s(x) = 0 for all x ∈ M .
Any linear algebra construction that can be applied to vector spaces can be applied to smooth vector
bundles, e.g. duals, tensor products, exterior powers, quotients etc. If f : M → N is a smooth map and
E is a vector bundle on N then there is a pullback bundle f ∗ E with fibres (f ∗ E)p = Ef (p) for p ∈ M .
Let i : M → N be an embedded submanifold of N . A tubular neighbourhood of M in N
is an open neighbourhood U which is diffeomorphic to a neighbourhood of the zero section in some
vector bundle on M . It is a general theorem that an embedded submanifold of a manifold has a tubu-
lar neighbourhood, with the vector bundle being the normal bundle N = i∗ T N/T M which has fibre
Tp N/di(Tp M ) at p ∈ M . We will prove this when the submanifold is compact. To simplify the notation
we will think of M as a subset of N , which causes no problems since it is an embedded submanifold.
First we deal treat the case of a manifold embedded in Rn .
Proposition 2.17.1. Let M be a smooth manifold embedded in Rn . Then M admits a tubular neigh-
bourhood in Rn .

Proof. Let Rn be given the standard inner product. Then at each point p ∈ M there is a decomposition
Tp Rn = Tp M ⊕ (Tp M )⊥ . It is straightforward to check that
N := {(p, v) ∈ M × Rn : v ∈ (Tp M )⊥ }
with the obvious map N → M is a smooth vector bundle of rank n − dim M and is isomorphic to the
normal bundle.
Now define a map f : N → Rn by (p, v) 7→ p + v. This map is the identity on the zero section
(= M ) and in the identification T(p,v) N = Tp M ⊕ Np the map df(p,v) is given by (v, w) 7→ v + w. Thus
df is a diffeomorphism isomorphism along the zero section, in particular it is a local homeomorphism
along the zero section M ⊂ N . We must show that there is a neighbourhood U of M in N on which f
is a homeomorphism. It is enough to show that there is neighbourhood on which f is injective. Now for
each x and δ > 0 let Vp,δ = {(p, v) : |x − p| + |v| < δ}. Since f is a local homeomorphism there exists
Vp,δ on which f is injective. Let δ(p) = sup {δ : f is injective on Vp,δ }. Then x 7→ δ(x) is a continuous
function. (Indeed, since Vy,δ−|x−y| ⊂ Vx,δ for all δ > 0, we have δ(y) ≥ δ(x) − |x − y|. The same holds
with x, y reversed. This shows that |δ(x) − δ(y)| ≤ |x − y|, so that δ is continuous. In fact it is Lipschitz
with Lipschitz constant one.)
Now let U = {(p, v) ∈ N : |v| < δ(p)/4}. It will suffice to show that f is injective on U . Let
(p, v), (q, w) ∈ U and f (p, v) = p + v = q + w = f (q, w) and assume δ(p) ≤ δ(q). Then |p − q| = |w − v| ≤
|w| + |v| ≤ δ(p)/2, so |p − q| + |w| = |w − v| + |w| ≤ δ(p). Thus both (p, v) and (q, w) lie in Vp,δ for some
δ < δ(p). But on such a neighbourhood the map f is injective, so (p, v) = (q, w). 
30

Theorem 2.17.2 (Tubular neighbourhood theorem). Let N be a smooth manifold and M ⊂ N a smooth
embedded compact submanifold. Then there exists a neighbourhood U of the zero section in the normal
bundle and a diffeomorphism of U with an open neighbourhood of M in N .

Proof. We first embed N in Rn and use the previous proposition to get a neighbourhood W of N which
is tubular, or more precisely, as in that proof the mapping (p, v) 7→ p + v is a diffeomorphism onto W .
Inverting this gives a map from W to the neighbourhood of N in the normal bundle of N in Rn and
then projecting to N defines a retraction r : W → N which is the identity on N . Let
N = {(p, v) ∈ M × Rn : v ∈ (Tp M )⊥ ∩ Tp N }.
(This is clearly isomorphic to the normal bundle as defined earlier.) Define a map F : N → N by (p, v) 7→
r(p + v) where r : W → N is the retraction. It is easy to check that F is a local diffeomorphism along M
and hence a local homeomorphism, so it remains to check that it is injective on some neighbourhood of
M when M is compact. Suppose not. Then we can find sequences of points (pn , vn ) and (qn , wn ) such
that (pn , vn ) 6= (qn , wn ) and vn → 0, wn → 0 but F (pn , vn ) = F (qn , wn ) for all n. By compactness of N
we have subsequences converging to p and q with F (p, 0) = F (q, 0), i.e. p = q. But then for large enough
n, (pn , vn ) and (qn , wn ) must both lie in a neighbourhood Vp,δ for some δ > 0, on which F is injective,
contradicting that F (pn , vn ) = F (qn , wn ). 

Notice that a tubular neighbourhood W of M in N admits a retraction r : W → M .

The theorem also holds if the embedded submanifold is noncompact, but requires more work.

2.18. Ehresmann’s fibration theorem. We first define the notion of fibre bundle. Let π : E → B be
a continuous mapping of topological spaces. Let F be a topological space. We say that π is a locally
trivial fibre bundle with fibre F or just fibre bundle with fibre F for short, if there is a covering
{Uα }α of B and homeomorphisms ϕα : π −1 (Uα ) → Uα × F (called local trivializations) such that
ϕα
π −1 (Uα ) −−−−→ Uα × F
 

πy
pr
y (2.18.1)

Uα Uα
commutes for all α. When E and B are smooth we can speak of a smooth fibre bundle by requiring
the local trivializations to be smooth maps.
The domain and range of a fibre bundle π : E → B are referred to as the total space and base
respectively. The space F is referred to as the fibre.
The following important theorem gives a general criterion for a smooth map to be a fibre bundle:
Theorem 2.18.1 (Ehresmann). A proper submersion of smooth manifolds is a smooth fibre bundle.

Proof. This will follow easily from the tubular neighbourhood theorem. Let f : M → N be a proper
submersion. A fibre f −1 (y) is a smooth embedded submanifold of M (by the normal form for submer-
sions). Therefore there is a neighbourhood W of f −1 (y) in M and a smooth retraction r : W → f −1 (y)
coming from the tubular neighbourhood theorem. Consider the map
(r, f ) : W → f −1 (y) × N
by w 7→ (r(w), f (w)). This is a diffeomorphism along f −1 (y) and since this is compact, there is an open
neighbourhood W 0 ⊂ W of f −1 (y) on which this map is a diffeomorphism onto its image. Since f is
proper, W 0 contains an open set of the form f −1 (U ) for some open set U ⊂ N containing y (“tube
lemma”). Then (r, f ) restricts to a diffeomorphism f −1 (U ) ∼= f −1 (y) × U . We leave it to be checked
that this gives local trivializations showing that M → N is a smooth fibre bundle. 

In particular, all the fibres of such a map are diffeomorphic to each other if the base N is connected.
Example 2.18.2. Recall that a complex hypersurface of degree d is the subset of CP n defined by a
homogeneous polynomial in n + 1 variables of degree d. Ehresmann’s theorem can be used to show that
any two smooth hypersurfaces of the same degree are diffeomorphic. (The details will be worked out in
the problem set.)
Example 2.18.3. Here is a simple example to keep in mind showing that properness is essential: Consider
the map R2 − {0} → R by (x, y) 7→ x. (How does the proof of the theorem fail in this case?)
 31

Later we will give a different proof of Ehresmann’s theorem using flows.

Remark 2.18.4. The tubular neighbourhood theorem has no analogue in the complex setting, i.e. a
complex submanifold M ⊂ N need not have an open neighbourhood W (which is again a complex
manifold) with a holomorphic retraction r : W → N . The analogue of Ehresmann’s theorem also fails:
The fibres of a smooth submersion f : M → N of complex manifolds, which are complex submanifolds
if f is holomorphic, are not biholomorphic to each other. Put another way, there are interesting families
of complex manifold structures on the same underlying smooth manifold. The simplest example is
that of complex tori of dimension one, or equivalently, elliptic curves, or equivalently, plane cubics (i.e.
hypersurfaces of degree 3 in CP 2 ).

2.19. Triangulation of smooth manifolds (∗). We will not prove the following theorem:
Theorem 2.19.1 (Whitehead). A smooth compact manifold admits a triangulation (i.e. there is a
(finite) simplicial complex with topological realization homeomorphic to the manifold).

The analogue for topological manifolds is false: In every dimension ≥ 4 there is a topological
manifold which does not admit a triangulation. In dimension 4 this was known in the 1980s (Freedman,
Casson) but in (all) dimensions ≥ 5 it was proved only recently (Manolescu, 2013).
As a corollary, we see that the (simplicial or singular) homology and cohomology groups of a compact
smooth manifold are finitely generated. We will prove this in a different way in the next subsection.

2.20. Good covers and finite generation of homology. A good cover of an n-manifold is a locally
finite cover {Ui }i∈I such that all intersections Ui0 ∩ · · · ∩ Uin for i0 , . . . , in ∈ I are homeomorphic to Rn
(or empty).
Proposition 2.20.1. Let M be a smooth manifold. Then M has a good cover.

Proof. (Weil (10)) The idea is that in an embedding in RN we can find a covering by neighbourhoods
which are
Using Whitney’s embedding theorem we may embed M in RN . For each x ∈ M , orthogonal
projection φx in RN to the n-plane Tx M , when restricted to a sufficiently small neighbourhood Ux of
x ∈ M homeomorphic to Rn , is actually a chart around x. The collection {(Ux , φx )}x is an atlas for M .
The transition function φy ◦ φ−1x : φx (Ux ) → φy (Uy ) between open subsets of Tx M and Ty M is close to
the identity (here we fix a reference linear isomorphism between Tx M and Ty M which is the orthogonal
projection to Ty M ). So we may assume that φy ◦ φ−1 x − id has sufficiently small first (Jacobian) and
second (Hessian) derivative matrices (in the sense that all entries have absolute value as small as we
want). But a diffeomorphism of Rn which has such small Jacobian and Hessian takes a convex set to a
convex set. In particular, the intersection Ux ∩ Uy is again convex. It follows that the cover {Ux }x is
good. 

In the compact case the Mayer-Vietoris sequence immediately gives:

Corollary 2.20.2. The homology and cohomology of a compact smooth manifold (with arbitrary coeffi-
cients) are finitely generated.

Recall that in the lectures on Poincaré duality we had assumed this fact for some applications, e.g.
to prove χ(M ) = χ2 (M ) etc., so now we have justified all that, at least for smooth manifolds. In fact
the corollary is true with the word “smooth” omitted, i.e. for topological manifolds, but this is harder
to prove. (I am not sure if any compact topological manifold has a good cover, so one has to prove finite
generation of homology in a different way.)

10Weil, Sur les theorèmes de de Rham, Commentarii Math. Helvetici 26 (1952)).

32

3. Vector fields and Frobenius’ theorem

3.1. Vector fields. A vector field on a smooth manifold M is a smooth section of the tangent bundle,
i.e. a smooth map V : M → T M such that π ◦ V = id. Informally, V (p) = (p, v(p)), i.e. we are given,
for each point p, a tangent vector v(p) ∈ Tp M and it “varies smoothly as we vary p”. We will usually
ignore the first factor and speak of “the vector field v.”
Let ϕ : M → M 0 be a diffeomorphism and T ϕ : T M → T M 0 the induced diffeomorphism of the
tangent bundle. Given a vector field v on M and v 0 on M 0 we can define new vector fields ϕ∗ (v) and
ϕ∗ (v 0 ), called the pushforward of v and pullback of v 0 respectively, by

ϕ∗ (v) := T ϕ ◦ v ◦ ϕ−1
ϕ∗ (v 0 ) := T ϕ−1 ◦ v 0 ◦ ϕ.

These are both sections of the respective tangent bundles, i.e. vector fields. (There is a mild abuse of
notation here as T ϕ must be applied to V (p) = (p, v(p)).) It is immediate that

ϕ∗ ϕ∗ (v) = v
ϕ∗ ϕ∗ (v 0 ) = v 0

In concrete terms we have:

ϕ∗ (v)(q) = dϕp (v(p)) for ϕ(p) = q

∗ 0
ϕ (v )(p) = dϕ−1 0
q (v (q)) for ϕ(p) = q.

(Here again we abuse notation by thinking of v(p) as a tangent vector rather than a pair (point, tangent
vector at the point).) Thinking of tangent vectors as derivations on (germs of) smooth functions, we
have:

ϕ∗ (v)(f 0 )(x0 ) = ϕ−1∗ (v(ϕ∗ (f 0 ))(x0 )

= v(f 0 ◦ ϕ)(ϕ−1 (x0 )) (x ∈ M 0 , f 0 smooth function on M 0 )
ϕ∗ (v 0 )(f )(x) = ϕ∗ (v 0 (ϕ−1∗ (f ))(x)
= v 0 (f ◦ ϕ−1 )(ϕ(x)) (x ∈ M, f smooth function on M )

where ϕ∗ (f 0 ) = f 0 ◦ ϕ is the pullback of smooth functions from M 0 to M .

• If f 0 is a smooth function on M 0 then ϕ∗ (f 0 v 0 ) = ϕ∗ (f 0 )ϕ∗ (v 0 ) (i.e. pullback of vector fields is a
module over pullback of functions)
In general, it is only possible to pushforward and pullback vector fields by diffeomorphisms, not by
general smooth maps. (There are some special classes of maps for which pullback of vector fields exists.)
Consider what a vector field looks like “locally”. Let p ∈ M and let (U, ϕ, U 0 ⊂ Rn ) be a chart
containing p. The smooth map V : M → T M restricts to a smooth map V : U → T U which is a section
of π : T U → U . Now using the diffeomorphism
n o ϕ : U → U 0 we can get a vector field ϕ∗ (v) on U 0 ⊂ Rn ,
i.e. a section of T U 0 → U 0 . The basis ∂x∂
i
gives a trivialization T U 0 = U 0 × Tϕ(p) U 0 , in terms of which
we can write
!
X ∂ ∗
X ∂
ϕ∗ (v) = ai or v=ϕ ai (3.1.1)
i
∂xi i
∂xi

for smooth functions ai = ai (x1 , . . . , xn ) ∈ C ∞ (U 0 ). This is what aPvector field looks like locally, i.e.
∂
in a coordinate chart. In the sequel we will frequently write v = i ai ∂xi as a local expression for
the vector field v; it is to be understood as meaning (3.1.1). (In other words we will usually drop the
diffeomorphism ϕ from the notation.)
Note that a vector field is determined by its action as a derivation on smooth functions (i.e., if
v, w are vector fields and v(f ) = w(f ) for all smooth functions f then v = w, e.g. because of the
local expression). In other words, a vector field is a linear differential operator of order one on smooth
functions.
 33

3.2. Local flows and vector fields. Suppose that we have an action α : R × M → M of the group R
on M such that the map α : R × M → M is smooth. Such a thing is called a flow on M . A flow gives
a vector field: For each point p let v(p) be the tangent vector to the smooth curve t 7→ α(t, p).
In the other direction, given a vector field v, a smooth curve γ : (a, b) → M is called an integral
curve of v if, at points p along the curve, the tangent vector at p is v(p). This means that for any
smooth function f ,
d
f ◦ γ = v(γ(s))(f )
dt t=s
for s ∈ (a, b), or, in different notation, that:
∂
dγs = v(γ(s)).
∂t
We shall now see that an integral curve of a vector field through a given point p always exists, at least
locally, and it is unique if we require γ(0) = p. This will be a consequence of the existence-uniqueness
theorem for ordinary differential equations.
Since the question is local, we may assume using a coordinate chart that we are in an open subset
O of Rn . Let γ = (γ1 , . . . , γn ) be a smooth curve in our open subset with γ(0) = q ∈ O. The condition
that γ is an integral curve of v is that
∂
dγs = v(γ(s))
∂t
holds for s in a neighbourhood of 0. In the local coordinates x1 , . . . , xn this is
X dγi ∂ X ∂
= ai (γ1 (s), . . . , γn (s))
i
dt t=s ∂xi
i
∂x i

which (dropping the dummy variable s), is the same as

X dγi ∂ X ∂
= ai (γ1 (t), . . . , γn (t))
i
dt ∂x i i
∂x i

with the initial conditions γi (0) = qi . (Here qi are the local coordinates of q.) Solving this equation is
equivalent to solving the following system of n ordinary differential equations with initial conditions:
dγi
= ai (γ1 (t), . . . , γn (t))
dt
γi (0) = qi .
Since the ai are smooth functions, the existence-uniqueness theorem for o.d.e. (Cauchy-Kovalevskaya
therem) says that:
(i) there exists a relatively compact neighbourhood U of p and an  (depending on U ) such that for
any q ∈ U , the system of differential equations with initial condition has a unique solution γq (t)
which is defined for t ∈ (−, )
(ii) the solution is smooth and depends smoothly on q ∈ U , i.e. the mapping
α : (−, ) × U → Rn α(t, q) = γq (t)
(“follow the integral curve starting at q for a time t”) is smooth.
(For a discussion and proof of the existence-uniqueness theorem see Arnol’d’s book.) Let us see that α
satisfies
α(t, α(s, p)) = α(s + t, p)
whenever both sides make sense. This equation is the same as γγp (s) (t) = γp (s + t). Fixing s, we see
that both these satisfy the same system of differential equations with the same initial condition, so by
uniqueness they are the same.
The map α defined here is not a flow, because α(t, p) is only defined for small t using the integral
curve. Such a thing is called a “local flow”.
A local flow (Φ, α) on a manifold M is a pair consisting of an open neighbourhood Φ of {0} × M
in R × M and a map α : Φ → M such that:
(i) Φ ∩ R × {p} is an interval containing 0
(ii) α(0, x) = x for all x ∈ M
(iii) α(s, α(t, x)) = α(s + t, x) whenever both sides make sense.
34

(The open set Φ will usually be dropped from the notation.)

The earlier discussion shows that a vector field gives rise to a local flow and vice-versa. This evidently
sets up a bijection
where (α, Φ) ∼ (α0 , Φ0 ) if α|Φ∩Φ0 = α0 |Φ∩Φ0 .

{ vector fields } ←→ { local flows } ∼
The following property of a local flow (α, Φ) associated with the vector field v is key:
 
∂
dα(t,x) = v(α(t, x)) (if (t, x) ∈ Φ). (3.2.1)
∂t
In terms of the action on smooth functions one has:
d d
v(f )(x) = α∗ (f )(x) = f (αt (x)) (3.2.2)
dt t=0 t dt t=0
The right-hand side has the following meaning: (t, x) 7→ f (αt (x)) is a smooth function of t and x, so we
can differentiate in t and set t = 0. We will also write this as:
d
v(f ) = α∗ (f ). (3.2.3)
dt t=0 t
(This expression could have a different meaning, namely consider t 7→ αt∗ (f ) as a function from R to
C ∞ (M ) and take the derivative (in the usual limit definition) at 0. This requires putting a topology on
C ∞ (M ), which can certainly be done (uniform convergence of all derivatives on compact sets), but we
will avoid this and simply take (3.2.3) to mean (3.2.2) for all x ∈ M .)
Note that v is a vector field on M and if V ⊂ M is relatively compact, then there is an  such that
(−, )×V is contained in the local flow of v, i.e. α : (−, )×V → M is defined (since any neighbourhood
of the closure of V contains a product neighbourhood).
∂
One cannot always get a (global) flow from a vector field: Take M = R, v(x) = x2 ∂x . The integral
a −1
curve starting at a ∈ R is γ(t) = 1−at . If a < 1 this is defined for t ∈ (−∞, a ). In particular, it is not
possible to choose  such that (−, ) × R is contained in the local flow. In fact, there is no  which will
work for any unbounded interval.
Lemma 3.2.1. If v is a vector field which vanishes outside a compact set, then its integral curves are
defined for all t. In particular, if M is compact then every vector field v arises from a flow α : R×M → M .

Proof. • Exercise. (Do this in two steps: First show that if M is compact there is an  such that the
local flow contains (−, ) × M . Next show that on any manifold M if a local flow contains (−, ) × M
then it extends (uniquely) to a flow R × M → M .) 

3.3. A useful lemma. Suppose we have a vector field v on a one-dimensional manifold M .
Let p ∈ M .
∂
The integral curve of v through p is a map γ : (−, ) → M with the property that dγt ∂t = v(γ(t)).
If we assume that v(p) 6= 0 then there is a neighbourhood of 0 ∈ (−, ) on which γ is a diffeomorphism
onto a neighbourhood of p, say γ : (−0 , 0 ) ∼
= W . Then in the coordinate t on W given by γ −1 : W → R,
−1
the vector field v becomes dγγ(t) (v(γ(t)) = ∂t ∂
, i.e. γ ∗ (v) = ∂t
∂
.
Given a vector field on a general manifold, locally one can choose coordinates in which it is simple:
Lemma 3.3.1. Let v be a vector field on M and p ∈ M a point such that v(p) = 6 0. Then we can find
∂
a neighbourhood U of p and local coordinates x1 , . . . , xn on U such that v = ∂x 1
on U . (More precisely,
 
0 n ∗ ∂
there is a chart (U, ϕ, U ⊂ R ) with p ∈ U such that v = ϕ ∂x1 on U .)

Proof. Let U be a neighbourhood of p in a chart on which the local flow is defined for t ∈ (−, ). Let
N be any submanifold of dimension n − 1 in U such that p ∈ N and v(p) ∈ / Tp N . (For example, choose
∂
any local coordinates in which v(p) = ∂x 1
and take N = {x 1 = 0}.) The flow map α : (, ) × N → M
is a diffeomorphism in some neighbourhood of (0, p) ∈ (−, ) × N (by the inverse function theorem).
Choose coordinates y2 , . . . , yn in N near p and set
xi = yi ◦ (projection to N ) ◦ α−1 (i ≥ 2)
−1
x1 = (projection to (−, )) ◦ α
∂
on a neighbourhood on which α is invertible. In these coordinates the vector field v is simply ∂x1 ,
because of the fact (3.2.1) that
∂ ∂ 
dα(t,x) = = v(α(t, x))
∂t ∂x1
 35

for all t, x in some neighbourhood of (0, p). 

This leads naturally to the next question: if we have k vector fields, all nonvanishing and linearly
∂
independent at p, when can we find local coordinates so that they are the ∂x i
for i = 1, . . . , k? (More
 
precisely, we should ask, does there exist a chart (U, ϕ, U ) containing p such that vi = ϕ∗ ∂x
0 ∂
i
on U .)
The answer is given by the Lie bracket of vector fields and Frobenius’ theorem.

3.4. Lie bracket. Let v and w be vector fields on M . The Lie bracket of v and w is [v, w] defined by
its action on smooth functions:
[v, w](f ) := v(w(f )) − w(v(f )).
Defined in this way it is not obvious that this is a vector field. We claim that this is actually a derivation
(linearity is clear, one must check the Leibniz rule) on germs of smooth functions at p, and that it varies
smoothly with p, so that it defines a vectorPfield. To check these statements it will suffice to check in
∂
local coordinates that it can be written as i ci ∂x i
for smooth functions ci (x1 , . . . , xn ) (i = 1, . . . , n).
Fix a point ofPM and fix local coordinates x 1 , . . . , x n in a neighbourhood of the point. Write the vector
∂ ∂
P
fields as v = i ai ∂x i
and w = b
i i ∂xi . Then
X ∂ X  ∂bi ∂ai

(vw − wv) = ci for ci = aj − bj .
i
∂xi j
∂xj ∂xj

Remark 3.4.1. On a C k (or better) manifold one has the notion of C k vector field. The bracket of two
C k vector fields is only C k−1 , as the previous local formula suggests.
If the manifold is real analytic then there is a notion of real analytic vector field, and the bracket of
two real analytic vector fields is again real analytic.
On a complex manifold there is a notion of holomorphic vector field, and the bracket of holomorphic
vector fields is holomorphic.

The bracket has the following properties, which follow from the definition:
(i) [v, v] = 0 (antisymmetry)
(ii) [u, [v, w]] + [w, [u, v]] + [v, [w, u]] = 0 (Jacobi identity)
These two conditions mean that the space of vector fields on M with the Lie bracket is a Lie algebra
over R (of infinite dimension if dim M > 0). (11) We also have:
(iii) for functions f, g and vector fields v, w, we have
[f v, gw] = f g [v, w] + f v(g) w − gw(f ) v
There is also the following very important naturality property:
(iv) If ϕ : M → M is a diffeomorphism then
ϕ∗ [v, w] = [ϕ∗ v, ϕ∗ w]
ϕ∗ [v, w] = [ϕ∗ v, ϕ∗ w]
∗
This is easy from the fact that ϕ∗ (u)(f ) = ϕ−1 (u(ϕ∗ f )). Indeed if u = [v, w] we have
∗ ∗
ϕ∗ ([v, w])(f ) = ϕ−1 ([v, w](ϕ∗ f )) = ϕ−1 (vw(ϕ∗ f ) − wv(ϕ∗ f )).
Now
∗ ∗ ∗ ∗
ϕ∗ v((ϕ∗ w)(f )) = ϕ−1 (v(ϕ∗ ((ϕ∗ w)(f ))) = ϕ−1 (v(ϕ∗ (ϕ−1 w(ϕ∗ f )))) = ϕ−1 (vw(ϕ∗ f ))
∗
(the last equality comes from ϕ−1 ϕ∗ f = f ), so that we have ϕ∗ [v, w](f ) = [ϕ∗ v, ϕ∗ w](f ) for all smooth
functions. The case of ϕ∗ follows since ϕ∗ ϕ∗ = id.
More generally, if f : M → N is a smooth map then we say vector fields v on M and v 0 on N are
f -related if we have dfp (v(p)) = v 0 (f (p)) for all p ∈ M . The same proofs show that
(v) If f : M → N is a smooth mapping and v, w are vector fields on M which are f -related to vector
fields v 0 , w0 on N respectively, then [v, w] is f -related to [v 0 , w0 ].

11 By definition, a Lie algebra over a field k is a k-vector space V with a k-bilinear operation [ , ] : V × V → V which
is antisymmetric and satisfies the Jacobi identity.
36

3.5. Local flows and brackets. Let (α, Φ) be a local flow on M . For p ∈ M let U be a relatively
compact neighbourhood of p and let  > 0 such that (−, ) × U ⊂ Φ, i.e. α(t, x) makes sense for
t ∈ (−, ) and x ∈ U . For such t, αt (−) := α(t, −) is a diffeomorphism of U onto an open set in M
(indeed, it has a smooth inverse, namely α−t ). So for |t| < , the pullback αt∗ (v) and the pushforward
(αt )∗ (v) make sense as vector fields on neighbourhoods of p. Writing things out in local coordinates
shows that these depend smoothly on t. We also have
αt∗ (v) = (α−t )∗ (v)
(from the definitions).
A more geometric picture of the bracket can be given using the local flow of one of the vector fields.
Let αt be the local flow associated with v. Then we have the alternate description:
d d
[v, w] = αt∗ (w) = − (αt )∗ (w).
dt t=0 dt t=0

(If vt is a vector field depending smoothly on a parameter t, then dv dt is the vector field which acts on
t

dvt d
smooth functions by dt (f ) = dt (vt (f )). That this is actually a vector field can be checked by a local
calculation.) Let us prove this identity. It is enough to do this in local coordinates centred at p. By the
∂ ∂a
last lemma, we may assume that coordinates have been chosen so that v = ∂x 1
, so ∂xji = 0 for all i, j.
Then the definition of the bracket gives
X ∂bi ∂
(vw − wv)(p) = .
i
∂x1 ∂xi
On the other hand, the integral curve of v through (p1 , . . . , pn ) is simply t 7→ (p1 + t, p2 , . . . , pn ), i.e.
x1 (t) = p1 + t, xi (t) = pi for i 6= 1. The local flowis given by αt (x1 , x2 , . . . , xn ) = (x1 + t, x2 , . . . , xn ).
∗
∂ ∂
= αt∗ (bi ) ∂x
∂
P
Write w = i bi ∂xi . Using the fact that αt bi ∂x i i
(• Exercise) we get:
 
X ∂ X ∂
αt∗ (w) = αt∗ bi = bi (x1 + t, x2 , . . . , xn ) .
i
∂xi i
∂xi
So then [v, w] is
d X ∂bi ∂ X ∂bi ∂
[v, w] := αt∗ (w) = (x1 + t, x2 , . . . , xn ) =
dt t=0 i
∂x1 t=0 ∂xi i
∂x1 ∂xi
and we have justified the alternate description.
There is a nicer formula for the bracket in terms of both the associated local flows:
Lemma 3.5.1. Let v, w be vector fields with local flows αt , βs respectively. Then
∂2
[v, w] = αt∗ βs∗ − βs∗ αt∗
∂s ∂t s=t=0
(as operators on smooth functions).

Proof. By definition,
∂ ∂
(βs∗ v(f ))(x) = v(f )(βs (x)) = (αt∗ f )(βs (x)) = f (αt βs (x)).
∂t t=0 ∂t t=0

Taking ∂
∂s and using w(g)(x) = ∂
∂s (βs∗ g)(x) for g = v(f ) gives
s=0 s=0
∂ ∂ ∂2
w(v(f ))(x) = f (αt βs (x)) = (βs∗ αt∗ f )(x).
∂s s=0 ∂t t=0 ∂s∂t s=t=0
This implies the lemma. 

Let α and β be two local flows on M . Let U ⊂ M . We say that α and β commute on U , if, for
p ∈ U,
α(t, β(s, p)) = β(s, α(t, p)) (equivalently, αt ◦ βs = βs ◦ αt )
whenever s, t are small enough so that both sides make sense. Note that if this equation holds for s, t in
some neighbourhood of (0, 0), then it holds for all s, t where it makes sense.
The previous lemma shows that if local flows of v and w commute then [v, w] = 0. The converse is
also true, i.e. we will show that if [v, w] = 0 then the local flows of v and w commute.
 37

Let ϕ : V → M be a diffeomorphism onto a relatively compact V 0 ⊂ M . If w is a vector field on

M , we say that w is invariant under ϕ if ϕ∗ (w) = w (or, equivalently, if ϕ∗ (w) = w). Concretely, this
means dϕa (v(a)) = v(ϕ(a)) for all a ∈ W .
Let v, w be vector fields, with (α, Φ) a local flow of v. We say that w is invariant under α if, for
any (−, ) × V ⊂ Φ we have αt∗ (w) = w for |t| <  (equivalently, αt∗ (w) = w for |t| < ).
Lemma 3.5.2. Let v and w be vector fields with local flows α and β.
(i) αt∗ (v) = v, i.e. a vector field is invariant under its own local flow.
(ii) w is invariant under α if and only if α and β commute.
(iii) For t in some interval around 0 we have:
d ∗
α (w) = [v, αt∗ (w)].
dt t
(At t = 0 this recovers the definition of [v, w].)

∂
Proof. (i) If we choose local coordinates in which v = ∂x1 then αt (x1 , . . . , xn ) = (x1 + t, x2 , . . . , xn ) and
this is immediate.
(ii) We show that if ϕ : V → M is a diffeomorphism onto a relatively compact open subset of M
then ϕ∗ (w) = w if and only if ϕ ◦ βt = βt ◦ ϕ for t in an interval around zero. (Applying this to ϕ = αs
for small enough s gives (ii).) To see this note that ϕ ◦ βt ◦ ϕ−1 is again a local flow on V 0 = ϕ(V ). For
a smooth function f on V 0 we compute:
∗
ϕ∗ (w)(f )(x) = ϕ−1 (w(ϕ∗ f ))(x)
= w(f ◦ ϕ)(ϕ−1 (x))
d
= βt∗ (f ◦ ϕ)(ϕ−1 (x))
dt t=0
d
= (f ◦ ϕ ◦ βt ◦ ϕ−1 )(x)
dt t=0

where in the third equality we have used (3.2.3). By (3.2.3) again, ϕ ◦ βt ◦ ϕ−1 induces the vector field
ϕ∗ (v) on V 0 . So if ϕ∗ (v) = v then the flows ϕ ◦ βt ◦ ϕ−1 and βt must agree for t small enough. This
proves (ii).
(iii) This is a local statement, so we P will∂ check it in a coordinate chart with local coordinates
∂
x1 , . . . , xn in which v = ∂x 1
. Write w = i bi ∂xi . Then
  X
∗
X
∗ ∂ ∂
αt (w) = αt bi = bi (x1 + t, x2 , . . . , xn ) .
i
∂x i i
∂x i

Then
d ∗ X ∂bi ∂
αt (w) = = [v, αt∗ (w)]
dt i
∂x1 (x1 +t,x2 ,...,xn ) ∂x i

by the local calculation of the bracket. 

Proposition 3.5.3. Let v, w be vector fields on M with associated local flows α and β. Then [v, w] = 0
in a neighbourhood U of p if and only if the local flows commute on U .
2
Proof. (⇐) This implication follows from the formula [v, w] = ∂s∂ ∂t s=0,t=0 αt∗ βs∗ − βs∗ αt∗ given above, but
let us prove it using the previous lemma instead. If α and β commute then by (i) w is invariant under
α, i.e. αt∗ (w) = w for |t| small enough. By (iii), for small enough |t| we have:
d ∗
0= α (w) = [v, αt∗ (w)] = [v, w].
dt t
(⇒) By (i) and (iii) of the previous lemma and properties of the bracket, for |t| small enough we
have:
d ∗
α (w) = [v, αt∗ (w)] = [αt∗ (v), αt∗ (w)] = αt∗ [v, w].
dt t
So if [v, w] = 0 then αt∗ (w) is constant in t (for t near zero), hence is equal to w (since α0 = id), i.e. w
is invariant under α. By (ii) above α and β must commute. 
38

3.6. Local Frobenius theorem. Let M be a manifold. A (smooth) subbundle of rank k of the
tangent bundle T M is a union E = ∪p∈M Ep , where Ep ⊂ Tp M is a k-dimensional subspace which
“varies smoothly in p”: More precisely, we require that for any p ∈ M there is an open set U ⊂ M
containing p and k vector fields v1 , . . . , vk such that v1 (q), . . . , vk (q) form a basis of Eq for all q ∈ U . (12)
It is easy to check that E is itself a smooth manifold of dimension n+k, in fact an embedded submanifold
of the tangent bundle.
An example is the rank one subbundle ∪p∈M Rv(p) where v is a smooth nonvanishing vector field on
M . More generally we could take the subbundle spanned by k vector fields, provided they are linearly
independent everywhere. (Not every subbundle is of this form.)
A k-dimensional smooth subbundle E ⊂ T M of the tangent bundle is called integrable (or invo-
lutive) on an open set U ⊂ M if, for any vector fields v, w such that v(p), w(p) ∈ Ep for all p ∈ U , the
bracket has the same property, i.e. [v, w](p) ∈ Ep for all p ∈ U . We say that E ⊂ T M is integrable (or
involutive) if each point has a neighbourhood on which E is integrable.
• Prove (using a partition of unity) that E is integrable if it is integrable on U = M . (This is false
in the real analytic or complex analytic setting.)
Let E ⊂ T M be a subbundle. A submanifold i : I → M is called an integral manifold of E if
dip (Tp I) = Ei(p) for all p ∈ I.
Sometimes we will suppress the notation dip and simply identify Tp I with its image in Ti(p) M ; we will
also suppress the inclusion i and think of points of I as points of M . So we will write the above condition
simply as: Tp I = Ep for all p ∈ I.
Remark 3.6.1. Note that an integral manifold is not necessarily a closed or embedded submanifold. Here
is an example: Fix α ∈ R − Q. The integral manifolds (=integral curves=orbits) of the vector field
associated with the flow on S 1 × S 1 defined by α(t, (z1 , z2 )) = (eit z1 , eiαt z2 ) are dense in S 1 × S 1 , hence
are not embedded submanifolds.
Theorem 3.6.2 (Local Frobenius theorem). A subbundle E ⊂ T M of the tangent bundle of a manifold
M is integrable if and only if it admits an integral manifold through any point of M .

This will be a consequence of two lemmas.

Lemma 3.6.3. Any integrable subbundle E ⊂ T M can be locally spanned by k commuting vector fields,
i.e. given p, there is a neighbourhood U of p and there exist vector fields w1 , . . . , wk on U such that
(1) Ex is the span of w1 (x), . . . , wk (x) for all x ∈ U and
(2) [wi , wj ](x) = 0 for all i, j and x ∈ U .

Proof. Let U be a neighbourhood of p with local coordinates

Pn x1 , . . . , xn and let v1 , . . . , vn be k vector fields
∂
which span Ex at each point of U . Write vi = j=1 aij ∂x j
for smooth functions aij = aij (x1 , . . . , xn ).
The matrix A = (aij ) has rank k near p, so some k × k submatrix of it is nonsingular. By renumbering
we may assume that A0 = (aij )i=1,...,k,j=1,...,k is the nonsingular submatrix, and by shrinking U we may
assume that A0 is nonsingular on all of U . Define
     ∂ 
w1 v1 ∂x1
 .  .  . 
 .  = A−1  .  = A−1 A  .  .
     
  0   0  
 .  .  . 
∂
wk vk ∂xn
Then w1 , . . . , wk are still a basis for E on U (since they are related to the basis v1 , . . . , vk by the
nonsingular matrix A0 ). Since A is of the form (A0 |∗), it follows that A−1 0 A is of the form (I|∗) and so
we have
∂ X ∂
wi = + ci` .
∂xi ∂x`
`>k
h i
∂ ∂ ∂f ∂ ∂
Now ∂x i
, f ∂x `
= ∂xi ∂x`
, so [wi , wj ] involves only the ∂x `
for ` > k. As E is integrable this must be
a linear combination (with smooth function coefficients) of the wi , so it must vanish.. So we have found
vector fields satisfying (1) and (2) as in the lemma. 
12In older books, the assignment p 7→ E is called a smooth k-distribution on M . Since distribution has a well-
p
established meaning in analysis it’s best to avoid this terminology.
 39

The next idea is that the local flows of these commuting vector fields, which commute among
themselves, provide coordinates.
Lemma 3.6.4. Let v1 , . . . , vk be linearly independent vector fields on V ⊂ Rn satisfying [vi , vj ] = 0 for
all i, j. For any p ∈ V , there exists a neighbourhood U of p and coordinates x1 , . . . , xn on U such that
∂
vi = ∂x i
for 1 ≤ i ≤ k (on U ).

Proof. Choose a chart W containing p and local coordinates z1 , . . . , zn centred at p. Using a linear
change of coordinates, we can arrange that
∂ ∂
v1 (p), v2 (p), . . . , vk (p), ,..., are linearly independent at p. (3.6.1)
∂zk+1 ∂zn
Let αi : (−, ) × W → Rn be the local flows associated with the vi (we shrink W so that one  works
for all αi ).
Let Bδ denote the open ball of radius δ <  in Rn . Define a map
h : Bδ → Rn
by
h(t1 , . . . , tk ; yk+1 , . . . , yn ) = α1 (t1 , α2 (t2 , . . . , αk (tk , (0, . . . , 0, yk+1 , . . . , yn )) . . . )).
We will show that this is a coordinate system at p for sufficiently small δ. For this it suffices (by the
inverse function theorem) to show that the differential dh0 : T0 Bδ → Tp V has rank n. By definition,
since α1 is the flow associated with v1 , we have (by (3.2.1)):
   
∂ ∂
dh(t1 ,...,tk ;yk+1 ,...,yn ) = (dα1 )(t1 ,h(0,t2 ,...,tk ;yk+1 ,...,yn ))
∂t1 ∂t1
= v1 (α1 (t1 , h(0, t2 , . . . , tk ; yk+1 , . . . , yn ))
= v1 (α1 (t1 , α2 (t2 , . . . , αk (tk , (0, . . . , 0, yk+1 , . . . , yn )) . . . ))
or  
∂
dh(t1 ,...,tk ;yk+1 ,...,yn ) = v1 (h(t1 , . . . , tk ; yk+1 , . . . , yn )).
∂t1
Since the vector fields vi commute among themselves, the flows αi also do so, so we could change the
order of the αi s in the definition of h and get that for any j ≤ k:
 
∂
dh(t1 ,...,tk ;yk+1 ,...,yn ) = vj (α1 (t1 , α2 (t2 , . . . , αk (tk , (0, . . . , 0, yk+1 , . . . , yn )) . . . ))
∂tj
i.e.  
∂
dh(t1 ,...,tk ;yk+1 ,...,yn ) = vj (h(t1 , . . . , tk ; yk+1 , . . . , yn )) for 1 ≤ j ≤ k. (3.6.2)
∂tj
Putting (t1 , . . . , tk ; yk+1 , . . . , yn ) = (0, . . . , 0) = 0 we have
 
∂
dh0 = vj (h(0, . . . , 0)) = vj (p) for j ≤ k.
∂tj
We also have  
∂ ∂
dh0 = for j > k
∂yj ∂zj
(since h(0, . . . , 0; yk+1 , . . . , yn ) = (0, . . . , 0, yk+1 , . . . , yn )). By (3.6.1), dh0 has rank n and hence h is
diffeomorphism onto h(Bδ ) for small enough δ by the inverse function theorem.
Now let U = h(Bδ ) and let us check that in the coordinates t1 , . . . , tk , yk+1 , . . . , yn given by h−1 :
U → Bδ , the vector fields vj are as we want. But this is precisely the equation (3.6.2). This proves the
proposition. 

Putting these together we prove the local Frobenius theorem.

Proof. Suppose first that E ⊂ T M admits integral manifolds through any point. We must show that it
is integrable. Let U be an open set and v, w vector fields with v(p), w(p) ∈ Ep for all p ∈ U . Let p ∈ U
and let i : I → M be an integral manifold through p. Now there are vector fields ṽ and w̃ on I that are
i-related to v and w. Indeed, they are defined by
dix (ṽ(x)) = v(i(x)) (x ∈ I)
40

and similarly for w̃. It is easy to see that they are smooth vector fields on I e.g. using the normal form
for immersions. Now the fact that the bracket of i-related vector fields is i-related shows that
dix ([ṽ, w̃](x) = [v, w](x) (x ∈ I).
In particular for x such that i(x) = p we get that [v, w](p) ∈ dix (Tx I) = Ep . Thus E is integrable.
Conversely, suppose E ⊂ T M is integrable and p ∈ M . There is a neighbourhood U of p on which
E is spanned by k commuting vector fields (by the first lemma above). By the second lemma, we may
assume that there are local coordinates on U centred at p in which E is the span of the vector fields
∂ ∂
∂x1 , . . . , ∂xk . Clearly for small enough c = (ck+1 , . . . , cn ) the “slices”
Uc := {(x1 , . . . , xn ) ∈ U : xk+1 = ck+1 , . . . , xn = cn }
are integral manifolds of E, and the choice c = (0, . . . , 0) gives an integral manifold through p. 

(This proof of the local Frobenius theorem is from R. Narasimhan’s book Analysis on Real and
Complex Manifolds (or see his TIFR lectures Topics in Analysis). For a slightly different proof see
Warner’s book.)

3.7. Examples. (i) Consider the subbundle in R3 − {0} given by

Ex=(x1 ,x2 ,x3 ) = {v ∈ Tx R3 : v ⊥ x}.
(Here we are using identifications Tx (R3 − {0}) = Tx R3 = R3 .) This subbundle is integrable and the
integral manifolds are spheres centred at the origin.
(ii) Consider the subbundle in R3 − {0} given by
Ex=(x1 ,x2 ,x3 ) = {v ∈ Tx R3 : v ⊥ (x2 , x3 , x1 )}.
Check that this is not integrable. This is an example of a completely nonintegrable hyperplane
field or contact structure: A subbundle of rank n − 1 such that for every pair of vector fields v, w ∈ F
which are nonzero and linearly independent near a point x, the bracket [v, w] does not lie in F . This
has a nice consequence: starting from any point p in R3 − {0}, you can get to any other point by the
following process: choose v1 ∈ F and follow its integral curve for a time t1 to get to q ∈ R3 − {0}. Then
choose v2 ∈ F in a neighbourhood of q and follow its integral curve for a time t2 . Choose v3 ...

3.8. Integral manifolds and the global Frobenius theorem. Integral manifolds of integrable sub-
bundles E ⊂ T M need not be embedded but they have the following “factoring” property which holds
if i : I ,→ M is embedded:
Theorem 3.8.1. Suppose that i : I ,→ M is an integral manifold of an integrable subbundle E ⊂ T M
and suppose that ϕ : N → M is a smooth mapping such that ϕ(N ) ⊂ i(I). Then there is a smooth
mapping ϕ̃ : N → I such that i ◦ ϕ̃ = ϕ:

N
ϕ̃
/I
ϕ
i
' 
M

Proof. The mapping ϕ̃ is already given as a set map, since i is injective, and it is given by ϕ̃(n) =
i−1 (ϕ(n)) for n ∈ N . So we have only to prove that it is smooth. The main point is actually to prove
continuity.
Let p ∈ I and let U be a neighbourhood of p in I. For q ∈ N with ϕ̃(q) = p, i.e. with ϕ(q) = i(p)
we must find a neighbourhood W of q such that ϕ̃(W ) ⊂ U , i.e. ϕ(W ) ⊂ i(U ). Choose a neighbourhood
V of i(p) in M such that i(U ) ⊂ V and V has local coordinates x1 , . . . , xm=dim M such that the slices
Vc = {xk+1 = ck+1 , . . . , xm = cm } (for c = (ck+1 , . . . , cm ) ∈ (−, )m−k )
are integral manifolds of E. We further assume the coordinates are centred at i(p), so that i(U ) ∩ V is
contained in the slice V0 = {xk+1 = · · · = xm = 0} of V . By possibly shrinking V further we can arrange
that i(U ) ∩ V = V0 .
Let us remark that each connected component of i−1 (V ) is mapped by i into in a single slice of V .
Indeed, this is because the functions x̄i = xi ◦ i for i ≥ k + 1 are locally constant on i−1 (V
 ). (If
 i>k
then dx̄i |Tp I = 0 for j ≤ k and p ∈ i−1 (V ) because ∂ ∂
∂x1 , . . . , ∂xk span dip (Tp I) and dxj ∂
∂xi = 0 if
j > k and i ≤ k).
 41

Now we remark that the connected components of i−1 (V ) map (under i) onto connected components
of i(I) ∩ V in the topology of V (i.e. of M ). To see this note that since I is second countable the set
i−1 (V ) has countably many connected components, each mapping into a single slice, hence i(I) ∩ V maps
under the projection V → Rm−k to a countable collection of points. The connected components of a
countable set in Rn are points, hence the V -connected components of i(I) ∩ V are contained in slices.
Now on a single slice the two topologies agree because the slices are embedded submanifolds, so the
assertion follows.
Let q ∈ ϕ−1 (i(p)). Let
W = connected component of ϕ−1 (V ) containing q
(this is an open neighbourhood of q in N ). To prove continuity of ϕ̃ it will suffice to show that ϕ̃(W ) ⊂ U ,
i.e. that ϕ(W ) ⊂ i(U ). By continuity of ϕ, ϕ(W ) is a connected set containing i(p), so to prove that
ϕ(W ) ⊂ i(U ) it will suffice to show that ϕ(W ) is contained in the single slice V0 = {xk+1 = · · · = xm = 0}.
But ϕ(W ) ⊂ i(I)∩V , and the components of i(I)∩V are contained in slices (as remarked earlier), so ϕ(W )
is contained in a single slice, which must be the slice V0 since i(p) ∈ ϕ(W ). So ϕ(W ) ⊂ V0 = i(U ) ∩ V
and hence ϕ̃(W ) ⊂ U .
The smoothness of ϕ̃ is left as an exercise. 

Note that the integrability of the subbundle is used essentially in the proof to have information on
all the integral manifolds in the neighbourhood V , not just on I. Also, the second countability of I
played a key role in the proof.
The following example is useful to think about: You can construct a 1-1 immersion i : R → R2 with
image the figure eight in the plane and an immersion ϕ : (0, 1) → R2 with image inside the figure eight
but which does not factor through i even continuously.
Integral manifolds of an integral subbundle through a given point are obviously not unique: an
open subset of an integral manifold is again an integral manifold. Taking the maximal integral manifold
through a point fixes this:
Theorem 3.8.2 (Frobenius theorem). Let E be an integrable subbundle of T M and let p ∈ M . There
exists a unique maximal connected integral manifold of E containing p.

Maximality means that every other connected integral manifold of E through p is contained in this
one. Uniqueness means that if i : I ,→ M and j : J ,→ M are two maximal connected integral manifolds
of E through p then there exists a diffeomorphism ϕ : I → J such that i = j ◦ ϕ. It is easy to check
that maximality implies uniqueness by the previous theorem (factorization property) so it is left as an
exercise.

Proof. Let I be the collection of all subsets of M which are unions of integral manifolds of M . This
defines a topology on M , which it is easy to check is finer (i.e. has more open sets) than the topology
on M . Write MI for M with this topology. Then MI is Hausdorff and locally connected. Write
G
MI = Mα
α∈A

as a union of connected components. Then any connected integral manifold i : I → M has i(I) ⊂ Mα
for some α. We will show that the Mα are integral manifolds; this will prove the existence part of the
theorem.
To prove that the Mα are second countable cover M by countably many neighbourhoods V with
coordinates x1 , . . . , xm such that slices Vc for c = (ck+1 , . . . , cn ) are integral manifolds of E. Since Mα
and Vc are both open in the topology I , we known that Mα ∩ Vc is open in Vc for all c = (ck+1 , . . . , cm ).
Now Mα ∩ V is open in Mα and is the disjoint union of Mα ∩ Vc , all of which are open and closed in
it, so each connected component of Mα ∩ V is open in some Vc and hence is second countable. By the
lemma below we conclude that Mα is second countable.
Now there is clearly a unique smooth manifold structure on Mα such that each integral manifold of
E contained in Mα is an open submanifold. 

Let us prove the lemma used above:

Lemma 3.8.3. Let X be a connected, locally connected Hausdorff space. Suppose that X = ∪n≥1 Xn
where Xn ⊂ X are open and each connected component of Xn is second countable. Then X is second
countable.
42

Proof. Let Cn be the collection of connected components of Xn and let C = ∪n≥1 Cn . If C ∈ C , then
there are countably many C 0 ∈ C with C ∩ C 0 6= ∅. (If there were uncountably many such then there
would exist an n for which {C 0 ∈ Cn : C 0 ∩C 6= ∅} is uncountable. But then {C ∩C 0 : C 0 ∈ Cn , C ∩C 0 6= ∅}
gives an uncountable collection of pairwise disjoint open sets in C. Since C is assumed 2nd countable
this is impossible.)
Let C0 ∈ C and set J0 = {C0 }. Inductively define countable sets Jk by
Jk = {C ∈ C : C meets an element of Jk−1 }.
Let Y = ∪k≥0 ∪C∈Jk C. This a countable union of second countable open sets, hence second countable
and open in X. If x ∈ Y there is a C ∈ C such that x ∈ C. Since C ∩ Y 6= ∅ there is a k and C 0 ∈ Jk
such that C ∩ C 0 6= ∅, whence C ∈ Jk+1 , and hence x ∈ Y . Thus Y is open and closed in X, so Y = X
is second countable. 

(The proof of the global Frobenius theorem here is originally from Chevalley’s Theory of Lie Groups,
as modified slightly in Varadarajan’s Lie Groups and Lie Algebras. Warner has a (very) slightly different
version.)
Here is a different description of the maximal connected integral manifold I through p ∈ M : A point
x ∈ M belongs to I if there exists a piecewise-smooth path joining p and x, each smooth piece of which
is an integral curve of E (in the sense that the tangent vector belongs to E pointwise). (Exercise: Show
that this is the same as the definition given above.)
Example 3.8.4. It is useful to keep in mind the example of the irrational slope flow on the torus M =
R2 /Z2 . The corresponding vector field defines a one-dimensional integrable subbundle of the tangent
bundle, and the integral manifolds are unions of segments. If we take a small square neighbourhood U
of a point then the topology induced on U by MI is homeomorphic to the following: (0, 1)2 with the
topology that a set is open if its intersection with (0, 1) × {y0 } is open for every y0 . In this topology
(0, 1)2 has uncountably many connected components, each homeomorphic to (0, 1). So it satisfies all
the conditions to be a manifold (in fact, an integral manifold) except 2nd countability. The second
countability of maximal connected integral manifolds in the theorem tells us that the intersection I ∩ U
must have countably many connected components.

Belonging to the same maximal connected integral manifold is an equivalence relation on M . The
decomposition of M into equivalence classes is called a foliation and the maximal integral manifolds are
the leaves of the foliation.
Finally, note that the local and global Frobenius theorems hold also in the real analytic and the
complex analytic settings, with the same proofs. (All that we really used was the inverse function
theorem, which holds in those settings.)
 43

4. Basics of Lie groups

We will use the Frobenius theorems to establish the basic correspondence between Lie groups and
Lie algebras, which is the starting point of Lie theory.

4.1. Lie groups, Lie algebras. A Lie group is a smooth manifold G with a group structure in which
the multiplication map µ : G × G → G by µ(g, h) = gh and the inversion map ι : G → G by ι(g) = g −1
are smooth maps. (Obviously it is enough to require that α(g, h) = gh−1 is smooth.)
Some examples: R, R∗ , C∗ , C, S 1 , GL(n, R), the group of affine transformations x 7→ Ax + b of Rn .
• If the multiplication µ : G × G → G is smooth then the inversion ι : G → G is smooth.
If G, H are Lie groups a Lie group homomorphism is a smooth map ϕ : H → G which is also a
group homomorphism. Lie groups with homomorphisms form a category.
A Lie algebra over a field k is a k-vector space g with a bilinear operation [ , ] : g × g → g which
is antisymmetric:
[x, x] = 0 for all x ∈ g
and satisfies the Jacobi identity:
[x, [y, z]] + [z, [x, y]] + [y, [z, x]] = 0 for all x, y, z ∈ g.
We will mainly take k = R or k = C here.
Examples 4.1.1. Some examples:
Rn with the trivial bracket ([x, y] = 0 for all x, y). (More generally, any vector space with the trivial
bracket.)
M (n, R) with the bracket [A, B] = AB − BA. (More generally, if A is an associative algebra over k
we can define [a, b] := ab − ba for a, b ∈ A and this is a Lie algebra.)
R3 with the bracket [v, w] = v × w (cross product of vectors) where we set v × v = 0.
Remark 4.1.2. Obviously R is the unique real Lie algebra of dimension one. There are two real Lie
algebras of dimension two (up to isomorphism), namely R2 with the trivial bracket and the Lie algebra
g = Rx + Ry with bracket [x, y] = y.

A Lie algebra homomorphism is a linear map ψ : h → g which respects the bracket, i.e.
ψ([x, y]) = [ψ(x), ψ(y)] for all x, y ∈ h. Lie algebras with homomorphisms between them form a category.

4.2. Lie algebra of a Lie group. Let G be a Lie group. It has diffeomorphisms Lg : G → G and
Rg : G → G defined by Lg (x) = gx and Rg (x) = xg.
For the moment we drop the assumption that vector fields are smooth, i.e “vector field” means
“possibly discontinuous section of π : T M → M ”. A vector field v on G is left-invariant if
L∗g v = v for all g ∈ G
(equivalently, Lg∗ (v) = v for all g ∈ G). In other notation,
v(gx) = (dLg )x (v(x)) for all g, x ∈ G.
Let g or Lie(G) denote the space of left-invariant vector fields on a Lie group G. Since G acts transitively
on itself a left-invariant vector field is determined by v(e) ∈ Te G; since the action is simply transitive
the map v 7→ v(e) gives an isomorphism of vector spaces g ∼ = Te G.
Lemma 4.2.1. (1) Left-invariant vector fields are smooth.
(2) The bracket of two left-invariant vector fields is left-invariant.
Therefore the space g of left-invariant vector fields is a Lie algebra under bracket.

Proof. (1) It is enough to show that if f ∈ C ∞ (M ) and v is left-invariant then v(f ) ∈ C ∞ (M ). (Here
v(f ) means: apply v(g) to the germ of f at g.) We have
v(f )(g) = v(g) · f = (dLg )e (v(e)) · f = v(e) · L∗g f = v(e) · f ◦ Lg .
44

Let ig : G → G × G by ig (x) = (x, g) and let jg : G → G × G by jg (x) = (g, x). Choose a smooth vector
field w on G such that w(e) = v(e). Then (0, w) is a smooth vector field on G × G and we have
((0, w)(f ◦ µ))(ie (g)) = ((0, w)(g, e))(f ◦ µ)
= (0(g))(f ◦ µ ◦ ie ) + (w(e))(f ◦ µ ◦ jg )
= w(e)(f ◦ µ ◦ jg ) (4.2.1)
= w(e) · f ◦ Lg
= v(e) · f ◦ Lg .
(Here in the second equality we have used that for tangent vectors w1 , w2 ∈ T(g1 ,g2 ) (G × G) and h ∈
C ∞ (G × G), we have (w1 , w2 )(h) = (0, w2 )(h) + (w1 , 0)(h) = w1 (h ◦ ig2 ) + w2 (h ◦ jg1 ).) Since the left-hand
side is evidently smooth as a function of g we conclude that the right-hand side is smooth in g. By the
previous identity v(f )(g) is smooth in g.
(2) Follows from [v, w] = [L∗g v, L∗g w] = L∗g [v, w]. 

Thus with any Lie group G is canonically associated a finite-dimensional Lie algebra, namely the
Lie algebra g = Lie(G) of left-invariant vector fields. Next we will show that this is functorial.
The following remark is often useful: The tangent bundle of a Lie group is trivial. (13) More precisely,
there is an isomorphism T G ∼ = G × Te G given by (g, v) 7→ (g, (dLg−1 )g (v)). (Exercise: Check that this
is an isomorphism.)

4.3. Homomorphisms. A homomorphism ϕ : G → H gives a map dϕe : Te G → Te H on tangent spaces

at the identity, and hence defines a linear map dϕ : g → h. By definition, for X ∈ g, dϕ(X) is the unique
left-invariant vector field on H with value dϕe (X(e)) at e ∈ H. Thus we have
dϕ(X)(h) = (dLh )e dϕe (X(e)) (h ∈ H).
Recall the following definition: If f : M → N is a smooth map and X, X 0 are vector fields on M and
N respectively, we say X and X 0 are f -related if dfp (X(p)) = X 0 (f (p)) for all p ∈ M . The following is
left as an exercise:
• If X is f -related to X 0 and Y is f -related to Y 0 , then [X, X 0 ] is f -related to [Y, Y 0 ].
Lemma 4.3.1. The map dϕ : g → h is a homomorphism of Lie algebras.

Proof. Let us first see that any X ∈ g is ϕ-related to dϕ(X). Note that because ϕ is a group homomor-
phism, ϕ ◦ Lg = Lϕ(g) ◦ ϕ. This gives that
dϕg (X(g)) = dϕg ((dLg )e X(e))
= d(ϕ ◦ Lg )e X(e)
= d(Lϕ(g) ◦ ϕ)e X(e)
= (dLϕ(g) )e (dϕ)e X(e)
= (dϕ(X))(ϕ(g)).
Thus X and dϕ(X) are ϕ-related.
Let X, Y ∈ g. We must show that dϕ([X, Y ]) = [dϕ(X), dϕ(Y )]. By the previous observation and
the exercise above we know that [X, Y ] and dϕ([X, Y ]) are ϕ-related. Then we have:
dϕg ([X, Y ](g)) = (dϕ[X, Y ])(ϕ(g)). (4.3.1)
Now dϕ[X, Y ] is the unique left-invariant vector field on G with value dϕe ([X, Y ](e)) at e ∈ G. On the
other hand [dϕ(X), dϕ(Y )] is a left-invariant vector field on G with this same value at e by (4.3.1). This
proves dϕ[X, Y ] = dϕ(X), dϕ(Y )]. 

We thus have a functor Lie from the category of Lie groups (and Lie group homomorphisms) to the
category of Lie algebras (and Lie algebra homomorphisms). This functor is not an equivalence, as the
following simple example shows: The Lie groups R and S 1 have the same Lie algebra, namely g = R.
13A manifold M for which the tangent bundle is isomorphic to the trivial bundle M × Rdim M is called parallelizable.
It is easy to see that an n-manifold M is parallelizable if and only if it admits n vector fields which are linearly independent
at every point (in particular, nowhere-vanishing). Since even-dimensional spheres do not admit nowhere-vanishing vector
fields they can never be Lie groups. (It is a famous theorem of Adams from the 1960s that the only parallelizable spheres
are S 1 , S 3 , S 7 .)
 45

4.4. Examples. Let us look at some simple examples. There are two one-dimensional connected Lie
groups, namely R (with addition) and S 1 (with multiplication). (This can be deduced e.g. from the
fact that there are two connected one-manifolds up to diffeomorphism.) Their Lie algebras are both the
∂
unique one-dimensional Lie algebra R. In the case S 1 the left-invariant vector field ∂θ spans the Lie
∂ 1
algebra. In the case R the Lie algebra is spanned by ∂t . The covering map R → S by t 7→ exp(2πit)
induces an isomorphism of Lie algebras.
Another example is R∗ with multiplication. In this case the Lie algebra is again R, spanned by the
∂
invariant vector field t ∂t . Note that the identity component R+ is diffeomorphic to R under the map
t
R → R+ by t 7→ e which also induces an isomorphism of Lie algebras.
Now consider the following examples of Lie subgroups in the torus T = R2 /Z2 . For α ∈ R define a
homomorphism of Lie groups R → R2 /Z2 by
ϕα (t) = t (1, α) mod Z2 .
If α = p/q is rational then ϕα factors through the covering R → S 1 and the image is a compact circle
in the torus. If α is irrational then ϕα : R → T is a non-embedded submanifold and injective group
homomorphism. In either case the Lie algebra map is R → R2 by 1 7→ (1, α).

4.5. Lie subgroups and Lie subalgebras. A Lie subgroup of G is a Lie group H which is a sub-
manifold of G by an injective immersion i : H ,→ G which is also a group homomorphism.
A Lie subalgebra of a Lie algebra g is a subspace h ⊂ g which is closed under the bracket, i.e.
[h, h] ⊂ h.
If i : H → G is a Lie subgroup then di(Lie(H)) is an subspace of Lie(G) closed under the bracket,
i.e. a Lie subalgebra. The following theorem says that this gives a bijection between Lie subgroups and
Lie subalgebras:
Theorem 4.5.1 (Lie’s first theorem). Let G be a Lie group. There is a one-to-one correspondence
between connected Lie subgroups of G and Lie subalgebras of the Lie algebra of G, given by H 7→ Lie(H).

Proof. The main point is to show that a Lie subalgebra of Lie(G) comes from a (unique) Lie subgroup.
Let h ⊂ g be a Lie subalgebra of g = Lie(G). Define a subbundle of the tangent bundle on G by
Fg = (dLg )e (h) (g ∈ G).
This is a smooth subbundle because it is (globally) spanned by dim h smooth vector fields: Choose a
basis X1 , . . . , Xdim h of h; the left-invariant vector fields vi (g) = dLg (Xi ) are smooth and span F at every
point. Note that [vi , vj ](g) = (dLg )e ([Xi , Xj ]) ∈ Fg for all g.
LetPus check that F is integrable. Every vector field v such that v(p) ∈ Fp for all p can be written
as v = i ai vi for ai ∈ C ∞ (G). Now
[ai vi , bj vj ] = ai bj [vi , vj ] + ai vi (bj )vj − bj vj (ai )vi
which lies (pointwise) in F since [vi , vj ](g) ∈ Fg for all g. Thus F is integrable.
Let i : H → G be the maximal connected integral manifold of F through e. Then
dih (Th H) = Fi(h) = (dLi(h) )e (h) (for all h ∈ H).
We must show that H is a Lie subgroup. Let us first see that it is closed under the group operation of
G. For this note that for h ∈ H, Li(h)−1 ◦ i : H → G is again an integral submanifold of F through e.
Indeed, for any x ∈ H,
d(Li(h)−1 ◦ i)x (Tx H) = (dLi(h)−1 )i(x) dix (Tx H) = (dLi(h)−1 )i(x) (dLi(x) )e (h) = (dLi(h)−1 i(x) )e (h).
By maximality of i : H → G we have that Li(h)−1 i(H) ⊂ i(H). This proves that i(h)−1 ∈ i(H) and
hence that i(H) is a subgroup of G.
To prove that i : H → G is a Lie subgroup it remains only to show that the map defined by
α(h1 , h2 ) := i−1 (i(h1 )i(h2 )−1 ) is smooth. Consider the diagram
α
H × H −−−−→ H


yi (4.5.1)
i◦α
H × H −−−−→ G
By the factorization property of integral manifolds of integrable subbundles the map α is smooth.
46

Finally we must prove uniqueness. Suppose i0 : H 0 → G is another Lie subgroup with di0 (Lie(H 0 )) =
h. Then i0 : H 0 → G is an integral manifold of the subbundle F . By maximality of H we have
i0 (H 0 ) ⊂ i(H). Consider the diagram
∃f
H 0 −−−−→ H


yi (4.5.2)
i0
H 0 −−−−→ G
A smooth f : H 0 → H exists by the factorization property. It is a local diffeomorphism at e, hence
it contains an open neighbourhood of e ∈ H in its image. Therefore it is surjective by the following
elementary fact:
• A neighbourhood of the identity in a topological group generates the connected component of
identity. (Exercise)
Hence f is a bijective homomorphism and is a local diffeomorphism everywhere (because it is a group
homomorphism). Therefore it is an isomorphism of Lie groups. 

As the example of the Lie subgroup ϕα : R → T with image the dense winding line on the torus for
α∈
/ Q shows, the Lie subgroup will not necessarily be embedded.

4.6. Example. Let us compute the Lie algebra of G = GL(n, R). Let β : g → M (n, R) be the isomor-
phism defined by evaluating a left-invariant vector field at e:

β(X) = X(e) ∈ Te M (n, R) = M (n, R).

We will show that β is a Lie algebra isomorphism if M (n, R) is given the Lie algebra structures [A, B] =
AB − BA. Let xij be the coordinate functions on G, so that xij (g) = gij . The essential property of β is
that
β(X)ij = X(e)(xij ).
We will show that β([X, Y ]) = [β(X), β(Y )] (this will be enough since β is an isomorphism of vector
spaces). For Y ∈ g we have:

Y (xij )(h) = Y (h)(xij )

= (dLh )e (Y (e))(xij ) = Y (e)(xij ◦ Lh )
!
X X
= Y (e) xik (h)xkj = xik (h)Y (e)(xkj )
k k
X
= xik (h)β(Y )kj (4.6.1)
k

where we have used the formula for matrix multiplication:

n
X
(xij ◦ Lg )(h) = xij (gh) = xik (g)xkj (h).
k=1

Using this we have:

β([X, Y ])ij = [X, Y ](e)(xij ) = X(e)Y (xij ) − Y (e)X(xij )

X
= X(e)(xik )β(Y )kj − Y (e)(xik )β(X)kj
k
= [β(X), β(Y )]ij . (4.6.2)

It follows that for the Lie algebra of any Lie subgroup of GL(n, R) the Lie bracket on the corre-
sponding subalgebra of M (n, R) is given by [X, Y ] = XY − Y X.
In more basis-invariant terms this shows that the Lie algebra of GL(V ) is identified with End(V )
with the bracket [A, B] = AB − BA.
 47

4.7. Linear algebraic groups. (14) A large class of groups (in fact, by far the most important ones),
for example groups like SO(n), Sp(2n) etc. which are defined by polynomial conditions on matrices can
be shown to be Lie groups by a direct method. We will prove this theorem first and then use it to discuss
many examples.
Let U ⊂ Rn be open. Suppose that I is an ideal in R[x1 , . . . , xn ], which we know to be finitely
generated by the Hilbert basis theorem. The set
V (I) = {x ∈ U : f (x) = 0 for all f ∈ I}
is called an algebraic subset of U . Conversely, given a set S ⊂ U let I(S) be the ideal of polynomials
in R[x1 , . . . , xn ] which vanish on S. Then a set is algebraic if and only if V (I(S)) = S. (15)
(The same definition makes sense in the complex case: If U ⊂ Cn is open then a subset S ⊂ U
is called a complex algebraic subset if the ideal I functions in C[x1 , . . . , xn ] vanishing on S has
V (I) = S, where V (I) ⊂ U is defined as before.)
While an algebraic subset need not be a manifold we will see that a large open subset of it is a
manifold. Let M be an algebraic subset of U ⊂ Rn . For x ∈ M consider the differentials dfx : Tx U =
Rn → Tf (x) R = R. Thinking of these as linear functionals on Tx U , let
r(x) = dimension of span of {dfx ∈ (Tx U )∗ : f ∈ I}.
This is finite because if I = (f1 , . . . , fN ) then the {dfi,x }i=1,...,N span the same space as {dfx }f ∈I because
of the identity
d(f g)x = f (x)dgx + g(x)dfx .
Now let
rM = maxx∈M r(x).
(The maximum exists because I is finitely generated.) Let
M reg = {x ∈ M : r(x) = rM }
be the regular locus of M . Note that the regular locus is open and nonempty.
(The same definition can be given for U ⊂ Cn , except that we we replace Tx U by the complexified
∂
tangent space (which is spanned over C by the ∂x i
and consider the elements dfx ∈ Tx U → Tf (x) C = C
as elements of HomC (Tx U, C).)
Proposition 4.7.1 (Whitney). The open subset M reg is a smooth manifold of dimension n − r and an
embedded submanifold of U . (In fact, M reg is an embedded real-analytic submanifold of U .)

(The proof below will also apply to the complex case, showing that if M is a complex algebraic
subset of U ⊂ Cn then M reg is a complex manifold of (complex) dimension n − r and hence also a real
manifold of dimension 2(n − r).)

Proof. Let p ∈ M reg . Let x1 , . . . , xn be coordinates centred at p and assume that df1,p , . . . , dfr,p are
linearly independent. Thus the r × n matrix of partials
 
∂fi
∂xj ij
has rank r at 0; by renumbering coordinates we may assume that the first r × r columns are independent.
Define a map from ϕ : U → Rn by
ϕ(x1 , . . . , xn ) = (f1 , . . . , fr , xr+1 , . . . , xn )
This is a real-analytic diffeomorphism at 0, hence there is a neighbourhood V of 0 ∈ U such that
ϕ : V → Rn is a real-analytic diffeomorphism onto an open subset of Rn . (Here we must use the real-
analytic version of the inverse function theorem.) Let y1 , . . . , yn be the coordinates on V given by ϕ and
let
V0 = {x ∈ V : y1 (ϕ(x)) = · · · = yr (ϕ(x)) = 0}.
Clearly this is a connected manifold of dimension n − r. If we show that V0 ⊂ M (in other words, the
vanishing of f1 , . . . , fr on V0 guarantees the vanishing of all f ∈ I on V0 ) then we will have shown that
14This section is a bit of a diversion in that we use a real-analytic argument, following Chapter I of Varadarajan’s
Lie Groups, Lie Algebras, and their Representations. It is the quickest way I know of to prove that the various standard
examples of matrix groups are in fact Lie groups all at once, i.e. without working out cases separately.
15Note that for an ideal J, I(V (J)) ⊃ J, but the inclusion may be strict. For example, if J = (x2 ) ⊂ R[x] then
I(V (J)) = (x). In general, I(V (J)) = rad(J) (this is known as Hilbert’s Nullstellensatz).
48

V0 = V ∩ M . Since V0 is a manifold of dimension n − r and V0 = V ∩ M is an open subset of M , this

will prove that M is a manifold at 0.
Let A = A (V ) be the algebra of all real-analytic functions on V . Let I = I · A be the ideal in A
generated by I = I(M ). For k > r the operator ∂y∂k leaves I invariant. Indeed, it is enough to prove
that
∂
I ⊂ I (4.7.1)
∂yk
for k > r. Let F ∈ I. Define an n-tuple (f1 , . . . , fn ) by fi = yi = xi if i > r and i 6= k and let fk = F .
(Recall that f1 , . . . , fr were already fixed as yi = fi for i ≤ r.) Then the Jacobians of partials are related
by      
∂fi ∂fi ∂xi
det = det · det . (4.7.2)
∂yj 1≤i,j≤n ∂xj 1≤i,j≤n ∂yj 1≤i,j≤n
∂F
Now the left-hand side is simply ∂y k
(recall that yi = fi for i ≤ r) while on the right-hand side the
 
second factor det ∂yj belongs to A . The first factor can be simplified as an r + 1 × r + 1 determinant:
∂xi

   
∂fi ∂fi
det = det
∂xj 1≤i,j≤n ∂xj 1≤i≤r,i=k,1≤j≤r,j=k
Now since fi ∈ I and F ∈ I, this vanishes on M . (If not there would be a point q ∈ M at which
df1,q , . . . , dfr,q , dFq are linearly independent, violating the maximality of r.) Thus (4.7.2) shows that the
left-hand side, namely ∂y∂k (F ), belongs to I · A = I . This proves (4.7.1) and hence that ∂y∂k I ⊂ I for
k > r.
Now it follows that for F ∈ I and any (ar+1 , . . . , an ), we have
 ar+1   an
∂ ∂
··· (F ) ∈ I .
∂yr+1 ∂yn
Since every element of I vanishes at 0, any F ∈ I has all such derivatives vanishing at 0. Since F is an
analytic function and V0 is connected, we see that F vanishes on all of V0 . In particular, elements of I
vanish on V0 , proving that V0 ⊂ M . 
Remark 4.7.2. In fact, a theorem of Nash says that any connected compact smooth manifold is diffeo-
morphic to a connected component of the nonsingular part of an algebraic subset of Rn . More precisely,
Nash (Annals of Math. 56 (1952)) proved that a connected compact smooth n-manifold can be embedded
in R2n+1 as a connected component of an algebraic subset. As a corollary, any such manifold admits a
real analytic structure, which was an earlier theorem of Whitney.

Now define a (real) linear algebraic group to be a subgroup G ⊂ GL(n, R) which is an algebraic
subset. (A complex linear algebraic group is a subgroup of GL(n, C) which is an algebraic subset). Then
we have:
Corollary 4.7.3. A linear algebraic subgroup of GL(n, R) (resp. GL(n, C)) is a closed Lie subgroup of
GL(n, R) (resp. of GL(n, C)).

Proof. We will apply Whitney’s theorem with U = GL(n, R), an open subset of M (n, R) ∼
2
= Rn with
polynomial functions in xij . If I = I(G) then for f ∈ I we also have L∗g f ∈ I for g ∈ G. It is easy to see
from this that the number r(x) defined above is the same at all points of the group because of the left
translation. So Greg = G and we are done. 

The corollary gives us a number of examples

 of Lie groups
 which are linear algebraic. For example,
 1 ∗ ∗ 
the group of 3 × 3 upper triangular matrices  1 ∗ is defined by the ideal (x21 , x31 , x32 , x11 −
1
 
1, x22 − 1, x33 − 1) and so is a linear algebraic group. (That this is a Lie group can be proved directly,
of course.) The following more interesting examples are usually referred to as the classical groups:
(i) (special linear groups) SL(n, R) = {A ∈ GL(n, R) : det(A) = 1} and SL(n, C) = {A ∈
GL(n, C) : det(A) = 1}. (The topology is the one induced from GL(n, R) or GL(n, C).)
(ii) (compact orthogonal groups) O(n) = {A ∈ GL(n, R) : AAt = At A = I} and SO(n) = O(n) ∩
SL(n, R).
 49

(iii) (compact unitary groups) U (n) = {A ∈ GL(n, C) : AA∗ = A∗ A = I} and SU (n) = U (n) ∩
SL(n, C).
(iv) (compact symplectic groups) Let H denote the quaternions, i.e. the unique four-dimensional
skew-field over R. Then Sp(g) is the group of matrices
Pg in GL(g, H) which preserve the standard
Hermitian form on p, q ∈ Hg given by H(p, q) = i=1 pi q̄i , i.e.
Sp(g) = {g ∈ GL(g, H) : gḡ t = I}.
(Here q̄ denotes the quaternion conjugation: If q = a + bi + cj + dk then q̄ = a − bi − cj − dk.)
Another way to define the compact symplectic group is as
Sp(2g) = Sp(2g, C) ∩ U (2g, C),
i.e. the group of 2g × 2g unitary matrices which preserve the standard symplectic form on C2g .
(See (vii) below for the definition of Sp(2g, C).)
(Exercise: Check equivalence of the two definitions as follows: As C-vector spaces we can
write H = C + Cj (write a + bi + cj + dk = a + bi + (c + di)j). Check that the standard Hermitian
form can be written as H = h + Bj where h is the standard Hermitian form on C2g and B is an
alternating form on C2g . Then a matrix in GL(g, H) preserving H becomes a GL(2g, C) matrix
preserving both h and B.)
(v) (orthogonal groups) Let P p + q = n for Pqp, q > 0. Let B be the symmetric bilinear form on
p
Rn given by B(x, y) = i=1 x i yi − i=p+1 xi yi . Let O(p, q) ⊂ GL(n, R) be the subgroup
{A ∈ GL(n, R) : B(Ax, Ay) = B(x, y)}. Let SO(p, q) = O(p, q) ∩ SL(n, R).
(On a real vector space of dimension n, there is up to equivalence a unique nondegenerate
quadratic form of signature (p, q) for p + q = n (Sylvester’s theorem). So if Q is any quadratic
form, the group O(Q) of matrices preserving it is conjugate in GL(n, R) O(p, q) for some p, q.)
(vi) (unitary groups)
Pp + q = n for p, q > 0. Let B be the Hermitian form on Cn given by
Let p P
q
B(z, w) = i=1 zi w̄i − i=p+1 zi w̄i . Let U (p, q) ⊂ GL(n, C) be the subgroup {A ∈ GL(n, C) :
B(Az, Aw) = B(z, w)}. Let SU (p, q) = U (p, q) ∩ SL(n, C).
(vii) (symplectic groups) Let Sp(2g, R) be the subgroup of P matrices in GL(g, R) preserving the anti-
g
symmetric form B : R2g ×R2g → R given by B(x, y) = i=1 xi yi+g −xi+g yi for x = (x1 , . . . , x2g )
and y = (y1 , . . . , y2g ).
Similarly, looking at elements of GL(2g, C) preserving B (extended linearly to C2g × C2g )
gives the complex symplectic group Sp(2g, C).
(On a vector space of even dimension over any field there is a unique nondegenerate alternating
form. So for any such form on R2g the group of matrices preserving the form is conjugate in
GL(2g, R) to Sp(2g, R). Also, any matrix preserving an alternating form must have determinant
one, so Sp(2g, R) ⊂ SL(2g, R).)
For each of these groups, the Lie algebra is a subalgebra of M (n, R) (or M (n, C) in the complex
cases) and so the bracket is given by [A, B] = AB − BA.
There are linear algebraic groups which do not fall into the above classes, but aside from finitely
many examples, any linear algebraic group which is simple (as a linear algebraic group, i.e. does not
contain any proper nontrivial linear algebraic subgroup) falls into one of the above classical examples.

4.8. Coverings of Lie groups. If M is a smooth manifold then it has a universal cover M̃ , which
again has a natural structure of smooth manifold for which M̃ → M is smooth. (It is obvious that M̃ is
Hausdorff and locally Euclidean; the only non-obvious part is that it is second-countable. This follows
from the fact that π1 (M ) is countable. (16))
Proposition 4.8.1. The universal cover G̃ of a Lie group G has a natural structure of Lie group such
that π : G̃ → G is smooth.
16Here is a proof for a second countable topological manifold: Take a countable cover {U } of M by coordinate charts
n n
homeomorphic to Rn . For each m, n the intersection Um ∩ Un has at most countably many components; choose a point in
each component and let C be the set of all such points (over all m, n). For each n and x, y ∈ C such that x, y ∈ Un choose
a path γn,x,y from x to y in Un . For a fixed basepoint p ∈ C there are countably many loops which are finite products of
paths of the form γn,x,y . It suffices to show that every element of π1 (M, p) is of this form.
For a loop γ : [0, 1] → M based at p there is an N such that when [0, 1] is subdivided into N equal intervals,
each subinterval [(k − 1)/N, k/N ] has image in some Un . Reparametrizing γ|[(k−1)/N,k/N ] we get paths γk such that
γ = γ1 · · · γN . For each k, the point γ(k/N ) lies in some component Un ∩ Um , in which we have a chosen point, call it
xk ∈ C. Now choose a path δk in Um ∩ Un from xk to γ(k/N ) and let γ̃k = δk−1 γ δk . Then [γ] = [γ̃1 ] · · · [γ̃N ]. But for
each k, γ̃k is a path in Um from xk−1 to xk , hence homotopic to the reference path γn,xk−1 ,xk . Thus γ is homotopic to a
product of paths of the form γn,x,y .
50

Proof. We have remarked above that G̃ has a smooth manifold structure for which π is smooth, so we
must define the group structure. Fix ẽ ∈ π −1 (e). Let α : G̃ × G̃ → G be the map α(x̃, ỹ) = xy −1 .
Consider the diagram of maps of pointed spaces
∃α̃
(G̃ × G̃, (ẽ, ẽ)) −−−−→ (G̃, ẽ)

π
y
α
(G̃ × G̃, (ẽ, ẽ)) −−−−→ (G, e).

By the simple connectedness of G̃ × G̃ there is a unique lift α̃ of α such that α̃(ẽ, ẽ) = ẽ. It is easy to
see that this lift is locally smooth, hence smooth. Define the inversion map ι : G̃ → G̃ by

ι(g̃) := α̃(ẽ, g̃).

Define the group operation m : G̃ × G̃ → G̃ by

m(g̃, h̃) = α̃(g̃, ι(h̃)).

Verifying the group axioms is left as an exercise. 

Lemma 4.8.2. If p : G0 → G is a covering of topological groups and G0 is connected then ker(p) is

contained in the centre of G0 .

Proof. If p : G0 → G is a covering then ker(p) is a discrete normal subgroup of G0 . For g ∈ ker(p) the
map h 7→ hgh−1 defines a continuous map G0 → ker(p), which must then have constant value e if G0 is
connected. 

Theorem 4.8.3. A homomorphism ϕ : H → G of connected Lie groups is a covering map if and only if
dϕ : Lie(H) → Lie(G) is an isomorphism (equivalently, dϕe : Te H → Te G is an isomorphism).

Proof. Suppose first that the homomorphism ϕ : H → G is a covering map. We will show first that dϕe
is injective. Indeed, if not then consider the subbundle defined by Fh := ker(dϕh ) on H. It is smooth
and integrable. Thus the integral manifolds of Fh are submanifolds of H such that dϕ is uniformly
zero on each one. Thus each integral manifold is contracted to a point in G. This contradicts the local
one-to-one property of ϕ. Next let us see that dϕe is surjective. If not then dim H < dim G and so (by
Sard’s theorem) the map ϕ : H → G cannot be surjective.
Conversely suppose that ϕ : H → G is a homomorphism of Lie groups with dϕe an isomorphism.
Then ker(ϕ) is a discrete normal subgroup of H. Let V be a neighbourhood of e in H such that

V −1 V ∩ ker(ϕ) = {e}.

We will show that ϕ(V ) is a regularly covered neighbourhood of e ∈ G. Firstly ϕ|V is one-to-one.
(Indeed, if ϕ(h1 ) = ϕ(h2 ) for h1 , h2 ∈ V then h−1
1 h2 ∈ V ∩ ker(ϕ) and hence h1 = h2 .) Since dϕh is an
isomorphism for all h ∈ H we see that ϕ|V is a diffeomorphism of V with the open neighbourhood ϕ(V )
of e in G. Next we claim that
G
ϕ−1 (ϕ(V )) = V θ.
θ∈ker(ϕ)

(Indeed, the inclusion ⊃ is obvious, so consider ⊂. Let h ∈ H with ϕ(h) ∈ ϕ(V ). Suppose ϕ(h) = ϕ(h0 )
for h0 ∈ V . Then θ := (h0 )−1 h ∈ ker(ϕ) and h = h0 (h0 )−1 h ∈ V θ. Moreover the union above is disjoint:
if h ∈ V θ1 ∩ V θ2 6= ∅ then h = h1 θ1 = h2 θ2 for h1 , h2 ∈ V and then h−1
1 h2 ∈ V
−1
V ∩ ker(ϕ) so h1 = h2
and hence θ1 = θ2 .) This proves that ϕ(V ) is a regularly covered neighbourhood since ϕ|V θ : V θ → ϕ(V )
is a diffeomorphism for each θ.FFor a general g = ϕ(h) ∈ G the neighbourhood gϕ(V ) = ϕ(hV ) is evenly
covered by the disjoint union θ∈ker(ϕ) hV θ. 

We have also shown the following: For any Lie group G, the universal cover π : G̃ → G is a Lie
group, and G = G̃/ker(π : G̃ → G). Moreover, π1 (G) = ker(π : G̃ → G).
 51

4.9. Example: SU (2), SO(3), Sp(1), quaternions. The group SU (2) of complex matrices A ∈ SL(2, C)
satisfying AA∗ = I (for A∗ = conjugate

3
transpose of A) is isomorphic to S . This follows because any
z w
such matrix can be written as where z, w ∈ C with |z|2 + |w|2 = 1, so that SU (2) is the unit
−w̄ z̄
ball in C2 . (Alternately, the natural action of SU (2) on C2 preserves the standard Hermitian form, so
it preserves the unit ball S 3 of vectors of norm one in C2 . It can be checked that the action of SU (2) is
simply transitive, so that there is an identification of SU (2) with S 3 .)
There is a double covering SU (2) → SO(3). To see this, consider the space su(2) of traceless skew-
Hermitian matrices, i.e. A satisfying A + A∗ = 0 and tr(A) = 0. This is the Lie algebra of SU (2). This
is a real vector space of dimension 3 and it has a natural bilinear form B(A, B) = − 21 tr(AB).
• The form B(−, −) is nondegenerate and positive definite.
There is an action of SU (2) on su(2) by linear transformations by (g, A) 7→ gAg −1 . In other words,
there is a homomorphism ρ : SU (2) → GL(su(2)). The action preserves the symmetric bilinear form
B(A, B) = − 21 tr(AB). In other words, the homomorphism ρ has image inside SO(qB ) where qB is the
inner product qB (v) = B(v, v) on the three-dimensional space su(2). So we have produced a Lie group
homomorphism SU (2) → SO(qB ), with kernel exactly ±I. To see the homomorphism ρ explicitly, you
must choose an orthonormal (w.r.t. qB ) basis of su(2), which gives an isomorphism su(2) ∼ = R3 . A
popular choice among physicists is the basis iσ1 , iσ2 , iσ3 where
     
0 1 0 i 1 0
σ1 = σ2 = σ3 =
1 0 −i 0 0 −1
are called the Pauli matrices. To show that ρ is surjective one has to only check that dρe is an isomor-
phism, for then the image contains a neighbourhood of e; since SO(3) is connected this neighbourhood
generates the whole group, so ρ is surjective. (17) As SU (2) = S 3 , this establishes that SO(3) is RP 3 .
(18) In physics, the fact that the group of rotational symmetries of physical space (= R3 ) is SO(3), which
has fundamental group Z/2Z, is responsible for the existence of particles of different “spin” in nature.
A third way to see the natural double cover SU (2) → SO(3) is using quaternions. Let H denote the
skew-field of quaternions, i.e. H = R + Ri + Rj + Rk where i, j, k are symbols satisfying the relations
i2 = j 2 = k 2 = −1, ij = k, jk = i, ki = j. The norm of a quaternion a+bi+cj +dk 2 2
 is (a +b +c +d
2 2 1/2
 ) .
a + bi c + di
The norm one quaternions are identified with SU (2) by a + bi + cj + dk 7→ . The
−c + di a − bi
group of norm one quaternions is topologically the 3-sphere, and acts by multiplication on quaternions
in H with zero real part, preserving the norm. This gives the map SU (2) → SO(3) again.
The group SU (2) acts naturally on C2 , and hence on CP 1 = S 2 . The isotropy of any vector is a
iθ

closed subgroup isomorphic to U (1), namely a conjugate of the group of matrices of the form e0 e−θ 0 .
The orbit map SU (2) → CP 1 is then a map S 3 → S 2 with U (1) = S 1 fibres, which is called the Hopf
fibration. (This map is interesting because it is a map S 3 → S 2 which can be shown to be not homotopic
to a constant map (cf. Hatcher’s book for a proof). This shows that π3 (S 2 ) 6= 0, and in fact π3 (S 2 ) = Z,
the Hopf map is a generator.)
Finally, note that there is still one more description of S 3 : It is homeomorphic to the compact
symplectic group Sp(1). Indeed, Sp(1) = SU (2) since both are identified with the unit quaternions.

4.10. Exponential map. Let G be a Lie group and X ∈ Lie(G) a left-invariant vector field. Let γX (t)
be the integral curve of X with γX (0) = e, defined for |t| < . So for any |t| < , the tangent vector to
γX at γX (t) is X(γX (t)). Then the tangent vector to t 7→ Lg γX (t) at gγX (t) is (dLg )γX (t) X(γX (t)) =
X(gγX (t)). Then Lg γX (t) is the integral curve of X which is at g at t = 0.
Applying this to g = γX (s) (for s small enough) we conclude that t 7→ γX (s)γX (t) is the integral
curve of X starting at γX (s). On the other hand γX (s + t) is also an integral curve of X, and at t = 0
is at γX (s). We conclude that
γX (s)γX (t) = γX (s + t) (4.10.1)
17 One can also show that ρ is surjective onto SO(q ) ∼ SO(3) by using that SO(3) is generated by rotations and
B =
showing that every rotation is in the image of SU (2).
18 Here is a direct proof that SO(3) is homeomorphic to RP 3 . Define a map D 3 → SO(3) by sending a point x to
the matrix of the rotation by |x|π around the axis through x (and sending 0 to the identity). This map is continuous and
injective on the interior of D3 , and is surjective since any element of SO(3) is a rotation around an axis (Exercise). On
the boundary ∂D3 we see that two points go to the same matrix in SO(3) if they are antipodal. This gives a bijective
continuous map from the quotient space of D3 by the relation x ∼ −x on ∂D3 to SO(3). This quotient space is RP 3 , and
since the domain is compact the map is a homeomorphism.
52

for s, t small enough that both sides make sense. But then γX can be extended to a mapping γX : R → G
which is in fact a group homomorphism. It is also clear that γX is an integral curve of X for all t. Thus
we see that the integral curves of a left-invariant vector field on a Lie group are defined for all t ∈ R.
Define the exponential map exp : Lie(G) → G by
exp(X) := γX (1).
This has the following properties:
Lemma 4.10.1. The exponential map exp : Lie(G) → G has the following properties:
(i) γX (t) = exp(tX) for all t
(ii) exp(tX)exp(sX) = exp((s + t)X) and exp(−tX) = exp(tX)−1
(iii) exp is a smooth mapping
(iv) d exp0 : Lie(G) → Lie(G) is the identity and hence exp is a local diffeomorphism at 0
(v) The flow of X ∈ Lie(G) is given by α(t, g) = g exp(tX) (i.e. αt = Rexp(tX) ).

Proof. (i) The integral curve for tX is s 7→ γX (st).

(ii) follows from (i) and (4.10.1).
(iii) Consider the smooth vector field on G × g defined by v(g, X) = (X(g), 0). The integral curve
of v through (g, X) is then t 7→ (g exp(tX), X) and the flow is βt (g, X) = (g exp(tX), X). Both these
are defined for all t, hence β1 is defined on G × g and then exp(X) = π1 ◦ β1 (e, X). But then exp(X) is
a composition of smooth maps, hence smooth.
(iv) tX is a curve in g with tangent vector X at 0; by (i) γX (t) = exp(tX) is a curve in G with
tangent vector X at e. Thus d exp0 (X) = X.
(v) restates the fact that Lg ◦ γX is the integral curve of X starting at g. 

Lemma 4.10.2. Every Lie group homomorphism R → G (i.e. every one-parameter subgroup of G) is
of the form t 7→ exp(tX) for some X ∈ Lie(G).

∂


Proof. Let Xe be the tangent vector to δ : R → G at e, i.e dδ0 ∂t = Xe . Let X be the associated
left-invariant vector field on G. It will be enough to show that δ is an integral curve for X, i.e. that
∂


X(δ(s)) = dδs ∂s . We compute:
 
∂
X(δ(s)) = d Lδ(s) X(e) = dLδ(s) dδ0
∂s
     
∂ ∂ ∂
= d(Lδ(s) ◦ δ)0 = d(δ ◦ Ls )0 = dδs
∂t ∂s ∂s
∂ ∂


where we have used that δ◦La = Lδ(a) ◦δ (because δ is a group homomorphism) and that (dLs )0 ∂s = ∂s
∂
because ∂s is translation-invariant on R. 

Proposition 4.10.3. Let ϕ : H → G be a homomorphism of Lie groups. The diagram

dϕ
Lie(H) −−−−→ Lie(G)
 
expy
 exp
y (4.10.2)
ϕ
H −−−−→ G
is commutative.

Proof. Let X ∈ Lie(H). Then t 7→ exp(tdϕ(X)) is an integral curve of dϕ(X) with tangent vector
dϕe (X(e)) at e. On the other hand t 7→ ϕ(exp(tX)) is a homomorphism R → G and as a curve has
tangent vector dϕe (X) at e. Therefore (by the previous lemma) it is an integral curve, in fact we must
have
ϕ(exp(tX)) = exp(tdϕ(X)).
Putting t = 1 gives the proposition. 

Corollary 4.10.4. Let H be a connected Lie group. If ϕ1 , ϕ2 : H → G are Lie group homomorphisms
such that dϕ1 = dϕ2 then ϕ1 = ϕ2 .
 53

Xn
Remark 4.10.5. For X ∈ M (n, R) we have the familiar exponential series eX = n≥0 n!
P
which gives
an element of GL(n, R). Once we know that the exponential series converges absolutely, t 7→ etX is a
one-parameter subgroup of G with derivative X at t = 0. By the previous proposition we conclude that
exp(X) = eX . The same applies for any subgroup of GL(n, R).
Proposition 4.10.6 (No small subgroups property). Let G be a Lie group. There is a neighbourhood
of the identity which contains no nontrivial subgroup.

Proof. Put a Euclidean norm || · || on the finite-dimensional vector space g and for  > 0 let B := {X ∈
g : ||X|| < } be the -ball around 0. Since d exp0 : g → g is an isomorphism, we know that there
exists δ such that exp is a diffeomorphism on Bδ . Now consider the neighbourhood U = exp(Bδ/2 ). Let
e 6= g ∈ U and let X ∈ Bδ/2 such that exp(X) = g. There exists n such that nX ∈ Bδ − Bδ/2 . But
then g n = exp(nX) ∈ / U since exp is bijective on Bδ . Thus there is no subgroup contained in U except
{e}. 
Remark 4.10.7. An immediate consequence of this is that any continuous homomorphism from a profinite
group Γ to a real Lie group factors through a finite quotient of Γ. For example, if Γ is a Galois group
with the Krull topology then any continuous representation of Γ on a finite-dimensional real vector space
factors through a finite quotient.
Thus, for example, a group like Gal(Q̄/Q), known as the absolute Galois group of Q does not have
a faithful finite-dimensional real or complex representation. To get such representations we have to
consider homomorphisms into GL(n, Qp ), which is an p-adic version of a Lie group. A good reference
where the theory of p-adic Lie groups and real Lie groups is developed together is Serre’s Lie Groups
and Lie Algebras.
Remark 4.10.8. One can ask if the exponential map is surjective for a connected group G. Here are some
exercises in linear algebra:
• Show that exp : M (n, C) → GL(n, C) is surjective. (Hint: Jordan normal form.)
• Show that if g ∈ SL(2, R) is in the image of exp then trace(g) ≥ −2. Conclude that exp is not
surjective.
• Show that exp : Lie(SU (n)) → SU (n) and exp : Lie(SO(n)) → SO(n) are surjective.
In fact exp is surjective for any compact Lie group G, but we will not prove this. It is also surjective for
connected Abelian Lie groups (because exp is a group homomorphism (exercise) and hence exp(g) is a
connected subgroup which contains a neighbourhood of the identity).

4.11. Simply connected groups.

Theorem 4.11.1. Let H be a simply connected Lie group and let G be a Lie group. Then for any Lie
algebra homomorphism ψ : Lie(H) → Lie(G) there is a unique Lie group homomorphism ϕ : H → G
such that dϕ = ψ.

Proof. Note that giving a Lie group homomorphism ϕ : H → G is the same as giving its graph Ĥ :=
{(h, ϕ(h)) : h ∈ H} ⊂ H × G. Then the homomorphism ϕ : H → G is the composition of the second
projection restricted to Ĥ with the isomorphism H ∼
= Ĥ. We will use this idea.
Let ψ : h → g be a Lie algebra homomorphism. Let ĥ ⊂ h × g be its graph, i.e. ĥ = {(X, ψ(X)) :
X ∈ h}. This is a Lie subalgebra of Lie(H × G) = h × g, so there is a unique Lie subgroup Ĥ ⊂ H × G
with Lie(Ĥ) = ĥ. The first projection H × G → H restricts to a homomorphism Ĥ → H which is an
isomorphism on Lie algebras. By a previous theorem it must be a covering of Lie groups. Since H is
simply connected we have that Ĥ ∼= H. Let ϕ be the restriction of H × G → G to Ĥ. This is the required
homomorphism, which is unique by the previous corollary. 

At this stage we have proved:

Theorem 4.11.2 (Lie’s third theorem). The functor Lie from the category of connected and simply
connected Lie groups to finite-dimensional real Lie algebras is fully faithful.

The following theorem of Ado is not easy (for a proof see Jacobson’s Lie Algebras):
Theorem 4.11.3 (Ado). Every finite-dimensional Lie algebra (over R) is linear, i.e. has an injective
Lie algebra homomorphism to M (n, R) for some n.
54

Applying Lie’s first theorem we see that every finite-dimensional Lie algebra is the Lie algebra of a
Lie group. This gives the strong form of Lie’s third theorem:
Theorem 4.11.4 (Cartan). The functor Lie is an equivalence of categories from connected and simply-
connected real Lie groups to finite-dimensional real Lie algebras.
Remarks 4.11.5. (1) It is usually the weak form of Lie’s third theorem that is used, because it is usually
Lie groups which arise naturally.
(2) The essential surjectivity of Lie can be proved directly, i.e. without using Ado’s theorem. Such
a proof is given in Appendix B of Knapp’s Lie Groups Beyond an Introduction.)

Let us summarize our picture of the situation: Any Lie group G is the quotient G̃/C of its simply
connected cover by a subgroup C of the centre of G̃. Different groups with the same Lie algebra come
from choosing different subgroups of the centre. So in practice many questions about Lie groups reduce
to considering simply connected groups (which can often be reduced to a purely Lie algebra question)
and then considering the effect of the centre (which is a purely “discrete” question).

4.12. Example: Universal cover of SL(2, R). Let G = SL(2, R) and let π : G̃ → G be its universal
cover. Note that π1 (G) = π1 (SO(2)) = S 1 ) ∼
= Z, so the kernel of G̃ → G is infinite cyclic. We will see
that any Lie group homomorphism ρ : G̃ → GL(n, R) must be trivial on ker(π), i.e. must factor through
π. In particular it cannot be injective. Thus not every Lie group can be thought of as a subgroup of a
general linear group.
Suppose that ρ : G̃ → GL(n, C) is a Lie group homomorphism. Let dρ : Lie(G̃) → M (n, C) be the
induced Lie algebra homomorphism. Composing with the inverse of the isomorphism dπ : Lie(G̃) →
sl(2, R) and complexifying gives a Lie algebra homomorphism
ψ := dρ ◦ (dπ)−1 ⊗ idC : sl(2, C) → M (n, C).

Since SL(2, C) is simply connected this can be integrated to give a Lie group homomorphism
ϕ : SL(2, C) → GL(n, C)
with dϕ = ψ. Then ϕ|SL(2,R) ◦ π and ρ are two Lie group homomorphisms G̃ → GL(n, C) and


d ϕ|SL(2,R) ◦ π = dϕ|SL(2,R) ◦ dπ = ψ|sl(2,R) ◦ dπ = dρ.
Thus ρ = ϕ|SL(2,R) ◦ π and thus ρ factors through π.
Exercise. Show using a similar argument that the universal cover of SL(n, R) has no injective Lie
group homomorphism to a general linear group. (In contrast, the universal cover of the subgroup
SO(n) ⊂ SL(n, R) is necessarily linear since it is a compact group, and compact groups are always linear
(as a consequence of the Peter-Weyl theorem).)

4.13. Closed subgroups (∗). The following theorem will not be used, but it is important and often
useful:
Theorem 4.13.1 (Cartan). A closed subgroup of a Lie group is a Lie subgroup. (More precisely, if
A ⊂ G is a closed subgroup then there exists a smooth submanifold structure on A such that A ⊂ G is a
Lie subgroup. This structure is unique.)

The correspondence between Lie subgroups and Lie subalgebras means that a closed subgroup must
have a Lie algebra. It is easy to define it: If A ⊂ G is a closed subgroup then a = {X ∈ Lie(G) :
exp(tX) ∈ A for all t}. The proof requires showing that exp : a → A is a homeomorphism near the
identity, so that it can be used to define charts on A. For the proof see e.g. Warner p. 110ff.
Note that this theorem can also be used to see that the examples of linear algebraic groups given
above are indeed Lie groups, since they are closed in GL(n, R) (or GL(n, C)).
(In the sequel, the phrase “closed Lie subgroup” will be used; in fact by this theorem we can just
say “closed subgroup”.)
However, we will need one property of closed Lie subgroups. Here by a closed Lie subgroup we
will mean a Lie subgroup ι : H → G such that ι(H) is closed in G. As usual, we will drop ι from the
notation and write H ⊂ G etc.
Lemma 4.13.2. A closed Lie subgroup of a Lie group is a embedded submanifold (and in fact a closed
submanifold).
 55

Proof. Let H be a Lie subgroup of G. It is an integral manifold of the subbundle defined by Lie(H) ⊂
Lie(G). Recall that in the proof of the factorization property for integral manifolds we showed that
there is a neighbourhood U of e ∈ G with coordinates x1 , . . . , xn in which the integral manifolds are
contained in slices, in which U ∩ H is given by a slice {xk+1 = · · · = xn = 0}, and under the mapping
U → Rn−k given by π(x1 , . . . , xn ) = (xk+1 , . . . , xm ) the image π(U ∩ H) is a countable subset. Now
since H is closed, the image π(U ∩ H) is locally closed. A countable locally closed set in Rn−k has at
least one isolated point. Thus there is some component of U ∩ H on which the induced topology agrees
with the topology of H. Translating using H implies this holds everywhere on H. This shows that H is
embedded. 

(The converse of this lemma also holds.)

4.14. Group actions and homogeneous manifolds (∗). We will consider only left actions: A left
action of a Lie group G on a smooth manifold M is a smooth map α : G × M → M such that (1)
α(g, α(h, x)) = α(gh, x) for all g, h ∈ G and x ∈ M and (2) α(e, x) = x for all x ∈ M .
Theorem 4.14.1. If H is a closed Lie subgroup of G then there is a manifold structure on the space of
cosets G/H such that π : G → G/H is a smooth submersion and a fibre bundle (with fibres isomorphic
to H). If H is a normal closed Lie subgroup then with this manifold structure G/H is a Lie group. The
translation action G × G/H → G/H is a transitive smooth left action.
Conversely, if G × M → M is a transitive smooth left action of G on M and x ∈ M then the orbit
mapping G/Gx → M is a diffeomorphism.

Let ϕ : H → G be a homomorphism of Lie groups. Applying this theorem to the action H × G → G

by (h, g) 7→ ϕ(h)g gives:
Corollary 4.14.2. The kernel of a Lie group homomorphism H → G is a closed normal Lie subgroup
of H. The image of a Lie group homomorphism H → G is a Lie subgroup of G.

4.15. Examples: Homogeneous spaces (∗). The previous theorem gives a large number of examples
of manifold structures on homogeneous spaces.
(1) Consider the action of SO(n) on S n−1 via its action on Rn . The isotropy subgroup of a point is
isomorphic to SO(n − 1) and hence we have SO(n)/SO(n − 1) ∼ = S n−1 .
2n−1 n
(2) Consider the action of SU (n) on S via its action on C . The isotropy subgroup of a point
is isomorphic to SU (n − 1) and hence we have SU (n)/SU (n − 1) ∼ = S 2n−1 .
4n−1 n
(3) Consider the action of Sp(n) on S via its action on H . The isotropy subgroup of a point is
isomorphic to Sp(n − 1) and hence we have Sp(n)/Sp(n − 1) ∼ = S 4n−1 .
(4) (real Grassmannians) Let G = SL(n, R) and let H be the subgroup of matrices stabilizing the
subspace Rk ⊂ Rn given by the first k coordinates. The space G/H is identified with the space
GrR (k, n) of k-dimensional subspaces of Rn and is called the real Grassmannian of k-planes in
n-space. Note that there is a map GrR (k, n) → GrR (n − k, n) given by V 7→ V ⊥ (⊥ taken w.r.t.
the standard Euclidean inner product on Rn ). This map is smooth and hence diffeomorphism
(exercise).
Another description of the Grassmannian is
GrR (k, n) = O(n)/O(k) × O(n − k) = SO(n)/S(O(k) × O(n − k))
which shows that it is compact. In this description the isomorphism GrR (k, n) ∼
= GrR (n − k, n)
n−1
becomes obvious. When k = 1 we have RP = SO(n)/O(n − 1).
(5) (complex Grassmannians) Let G = SL(n, C) and let H be the subgroup of matrices stabilizing
the subspace Ck ⊂ Cn given by the first k coordinates. The space G/H is identified with the
space of k-dimensional complex subspaces of Cn and is called the complex Grassmannian of
k-planes in Cn and denoted Gr(k, n). Once again, taking the ⊥ w.r.t. a Hermitian form on Cn
gives a diffeomorphism Gr(k, n) → Gr(n − k, n).
Another description of the complex Grassmannian is
Gr(k, n) = U (n)/U (k) × U (n − k) = SU (n)/S(U (k) × U (n − k))
which shows that it is compact. In this description the isomorphism Gr(k, n) ∼
= Gr(n − k, n)
becomes obvious. When k = 1 we have CP n−1 = SU (n)/U (n − 1).
56

(6) (symplectic or Lagrangian Grassmanians) Let V be a 2n-dimensional vector space over F = R

or F = C with a nondegenerate F -bilinear alternating form V × V → F . A subspace I ⊂ V is
called isotropic if L⊥ ⊃ L, in which case dim I ≤ n and Lagrangian if L⊥ = L. A Lagrangian
subspace is necessarily of dimension n. Let ΛF (V ) be the set of Lagrangian subspaces of v, which
is naturally a closed subspace of the Grassmannian GrF (n, V ). Then ΛF (V ) is a homogeneous
manifold for Sp(2n, F ) = {g ∈ GL(2n, F ) : (gv, gw) = (v, w) for all v, w ∈ V }. (Equivalently,
Sp(2n, F ) acts transitively on the Lagrangian subspaces of V .)
Alternate descriptions of ΛR (V ), ΛC (V ) are:
ΛR (n) = U (n)/O(n), ΛC (n) = Sp(n)/U (n)
which show that they are compact. (Alternately, they are closed subsets of GrR (n, V ) and
Gr(k, V ), hence compact.)
(7) (Stiefel manifolds) The space Stk of k-tuples of orthonormal vectors in Rn is a homogeneous
manifold for O(n). The isotropy subgroup in O(n) of the standard k-tuple is identified with
O(n − k). Thus Stk = O(n)/O(n − k) = SO(n)/SO(n − k).
The corresponding complex Stiefel manifold classifying k-tuples of Hermitian-orthonormal
vectors in Cn is U (n)/U (n − k) = SU (n)/SU (n − k).
(8) (flag manifolds) Let F = R or F = C. Let V be a vector space of dimension n over F . A (full)
flag in V is a chain of F -subspaces
{0} = V0 ⊂ V1 ⊂ · · · ⊂ Vn = V
such that dimF Vi = i. Show that the space X of full flags (naturally a closed subset of the
product of Grassmannians GrF (1, n)×· · ·×GrF (n−1, n)) is a homogeneous manifold for SL(n, F )
(or GL(n, F )).
(9) (symplectic flag manifolds) Let V be an F -vector space of dimension 2n. Let X be the set of
chains of isotropic subspaces
{0} = I0 ⊂ I1 ⊂ · · · ⊂ In
where dimF Ik = k. (So In is Lagrangian.) This is a closed subset of a product of n Grassmanni-
ans. This is a homogeneous manifold for Sp(2n, F ). Note that there is an Sp(2n, F )-equivariant
map X → Λ(n) to the Lagrangian Grassmannian.

4.16. π0 and π1 of some classical groups (∗).

The following lemma is easy:
Lemma 4.16.1. If G is a topological group and H is a closed connected subgroup such that G/H is
connected (in the quotient topology) then G is connected.
Corollary 4.16.2. SO(n), SU (n), Sp(n) are connected.

Using det : O(n) → {±1} we see that O(n) has two components. Using det : U (n) → U (1) we see
that U (n) is connected.
Proposition 4.16.3. SL(n, R) and SL(n, C) are connected.

Proof. The Gram-Schmidt orthogonalization procedure, with parameters, shows that O(n) is a deforma-
tion retract of GL(n, R), which retracts the identity component GL(n, R)0 onto SO(n). The analogue
in the Hermitian setting shows that GL(n, C) has a deformation retraction onto U (n) and SL(n, C) has
a deformation retraction onto SU (n).
Instead of using this one can show that there is a diffeomorphism GL(n, R)+ ∼
= SO(n) × P os(n, R)
where GL(n, R)+ is the invertible matrices with determinant > 0 and P os(n) is the space of positive
definite symmetric n × n real matrices. See Warner p. 131 for a proof of this. The analogue in the
complex case is that GL(n, C) is diffeomorphic to U (n) times the space of positive definite Hermitian
matrices. 

Lemma 4.16.4. If H ⊂ G is a subgroup and H 0 is the identity component of H then p : G/H → G/H 0
is a covering. (In particular, if G/H is simply-connected then H must be connected.)
 57

Proof. Since H 0 is open in H there is a neighbourhood V of e in G such that V −1 V ∩ H ⊂ H 0 . Refer

to the following diagram to fix notation:
π0
G −−−−→ G/H 0

p
y
π
G −−−−→ G/H.
Now I claim that for g ∈ G the nhbd π(gV ) of gH in G/H is evenly covered by p.
Consider the open sets π 0 (gV θ) as θ runs over a set of representatives Θ ⊂ H for the cosets of H 0 .
Each of them maps homeomorphically onto π(gV ) under p, so it will be enough to show that
G
p−1 (π(gV )) = π 0 (gV θ).
θ∈Θ
0
The inclusion ⊃ is clear, so let us prove ⊂. If p(xH ) = π(gv) = gvH for v ∈ V then xH = gvH hence
v −1 g −1 x ∈ H. Then gvH 0 v −1 g −1 x = xH 0 and hence xH 0 ∈ π 0 (gV θ). We leave the disjointness of the
union as an exercise. 
Lemma 4.16.5. Let G be a connected Lie group and H a connected closed Lie subgroup.
(i) If H is simply-connected and G/H is simply-connected then G is simply-connected.
(ii) More generally, if G/H is simply-connected then π1 (H) → π1 (G) is surjective.

Proof. (i) follows from (ii), so we prove (ii). Let π : G̃ → G be the universal cover of G. Then π induces a
map G̃/π −1 (H) → G/H which is easily seen to be a diffeomorphism (it is a bijective map which is a local
diffeomorphism). Since G̃/π −1 (H) is simply connected, by the previous lemma π −1 (H) is connected.
Thus π : π −1 (H) → H is a connected covering and hence the universal cover πH : H̃ → H of H factors
as
p π
H̃ → π −1 (H) → H.
Now p is surjective (it is a covering) and hence induces a surjection of ker(πH ) = π1 (H) onto ker(π :
π
π −1 (H) → H) = ker(π : G̃ → G) = π1 (G). (We leave as an exercise that this is the natural map
π1 (H) → π1 (G), although it is not needed.) 
Corollary 4.16.6. The groups SU (n), Sp(n), SL(n, C) are simply-connected. There is a surjection
Z/2Z → π1 (SO(n)) = π1 (SL(n, R)) for n ≥ 3.

In fact we can do more if we use the long exact sequence of homotopy groups of a fibration. (We
are assuming here that the quotient map G → G/H is a fibration. In fact it is even better; it is a fibre
bundle.) We have seen that SO(3) admits a double cover, namely SU (2) ∼
= S 3 , so that π1 (SO(3)) = Z/2.
n
The homotopy long-exact sequence for the fibration SO(n + 1) → S , with fibres isomorphic to SO(n)
is:
· · · → π3 (SO(n)) → π3 (SO(n + 1)) →π3 (S n ) → π2 (SO(n)) →
→π2 (SO(n + 1)) → π2 (S n ) → π1 (SO(n)) → π1 (SO(n + 1)) → 0
One sees that if n ≥ 3 then π1 (SO(n + 1)) = π1 (SO(3)). By induction, we conclude that
π1 (SO(n)) = Z/2 (n ≥ 3)
and more generally that πk (SO(n)) = πk (SO(n + 1)) if k + 1 < n. Since SO(3) = RP 3 , we get
π2 (SO(3)) = 0 and hence π2 (SO(4)) = 0. Using the previous assertions we see that
π2 (SO(n)) = 0 (all n).
One also sees from the sequence with n = 3 that the map π2 (S 2 ) → π1 (SO(2)) must be Z → Z by
multiplication by 2. and hence that π2 (SO(3)) = 0. Since π2 (S 1 ) = π3 (S 1 ) = 0, we get π3 (SO(3)) =
π3 (S 2 ) = Z. (The last group is generated by the Hopf map.)
• Use the fibrations SU (n) → S 2n−1 and Sp(n) → S 4n−1 to show that π2 (SU (n)) = π2 (Sp(n)) = 0.
In fact for any compact Lie group with zero-dimensional centre, π1 (G) is finite, π2 (G) is trivial, and
π3 (G) = Z.
The fact that π1 (SO(n)) = Z/2 tells us that there is another sequence of simply-connected compact
groups, namely the universal (=double) cover of SO(n). These are called the spin groups and denoted
Spin(n).
58

In the case of n ≤ 6 they are in fact already on our list above:

4.17. Example: Spin(4) (∗). There is an action of S 3 × S 3 on H given by thinking of S 3 as quaternions

of norm one:
(α, β) · θ = α θ β (θ ∈ H).
This action (evidently) preserves the norm quadratic form on H, which is positive definite. So we have
a map
S 3 × S 3 → SO(4)
which is easily seen to be a group homomorphism. Check that the kernel is precisely {±(1, 1)} = Z/2.
By computing the Lie algebra map we can check that this is nonsingular at the identity, hence a double
covering. It follows that
Spin(4) = S 3 × S 3 = SU (2) × SU (2).
There are two more isomorphisms
Spin(5) = Sp(4)
and
Spin(6) = SU (4).
But for n ≥ 7 the spin groups Spin(n) are new simply connected compact groups, not in the list of
examples above. These groups are linear, i.e. they can be embedded in GL(n, R) but this requires the
theory of Clifford algebras. They do not have a “canonical” embedding in GL(n) like the series SU (n)
and Sp(n), however.
Here is another application of the simple connectedness of SU (n): (There are other ways to prove
this result, of course.)
Proposition 4.17.1. If G is a finite group then there is a compact manifold with π1 (M ) = G.

Proof. Suppose that there is an injective homomorphism G ,→ SU (n) for some n 6= 2. Then M =
SU (n)/G is a compact manifold with π1 (M ) ∼
= G.
To show that G ,→ SU (n) it is enough to treat the case of the symmetric group Sk on k letters
(since there is an embedding G ,→ Sk by somebody’s theorem in elementary group theory). We can
embed Sk in U (k) as the matrices of the permutation of the coordinates, i.e. σ is sent to the matrix of
the linear transformation Cn → Cn defined by zi 7→ zσ(i) . The
 image
 is not in SU (k), however, so we
g
compose with the homomorphism U (k) ,→ SU (2k) by g 7→ g −1 to get an injective homomorphism
Sk ,→ SU (2k). 
Remark 4.17.2. The question of when a group can be realized as the fundamental group of a smooth
projective complex variety is very interesting. Every finite group can be, by a result of Serre. If we drop
the smoothness requirement then any finitely presented group is the fundamental group of a complex
projective variety (Simpson), but there are nontrivial restrictions in the smooth case.

4.18. Hopf ’s theorem and compact Lie groups (∗). The cohomology algebra H ∗ (G, k) of a Lie
group G with coefficients in a field k of characteristic zero, is a graded-commutative Hopf algebra of
finite dimension. By the Hopf structure theorem (proved in the exercises) we have that H ∗ (G, k) is an
exterior algebra generated by elements of odd degree (equivalently, the cohomology algebra of a product
of odd-dimensional spheres). In fact it is possible, for the compact Lie groups, to compute the degrees
directly from the structure of the Lie group. The integral cohomology of the compact Lie groups is still
not fully calculated (as far as I know).
 59

5. Differential forms and de Rham cohomology

The material in this section is in many books, e.g. Warner, Bott-Tu etc. We will use the language
of sheaves and presheaves although that will only appear formally in the next section.

5.1. Vector bundles.

A (smooth) vector bundle of rank r on a manifold M is a smooth manifold E with a smooth map
π : E → M together with:
(i) an open covering M = ∪α Uα of M , together with diffeomorphisms
ψα : π −1 (Uα ) → Uα × Rr
such that π|π−1 (Uα ) is given by the first projection
(ii) for α and β such that Uα ∩ Uβ 6= φ, a smooth map
hαβ : Uβ ∩ Uα → GL(r, R)
such that the composite
−1
ψβ ψα
Rr × Uα ∩ Uβ −→ π −1 (Uα ∩ Uβ ) −→ Rr × Uα ∩ Uβ
is given by (hαβ , id).
The manifolds E and M are called the total space and the base space, respectively. Each Uα , ψα is
called a local trivialization, the collection {(Uα , ψα )}α is an atlas for the vector bundle. The hβα are
called the transition functions of the vector bundle. The fibres
Ex := π −1 (x)
are naturally vector spaces of dimension r. There are obvious notions of vector bundle homomorphisms
and vector bundle isomorphisms. Note that the transition functions are the key data. The satisfy the
following cocycle condition: If Uα ∩ Uβ ∩ Uγ 6= ∅ then on this set we have the identity
hαβ hβγ = hαγ . (5.1.1)
The vector bundle is determined up to isomorphism by the covering of M and the corresponding transition
functions, as the following lemma shows:
Lemma 5.1.1. Let {Ex }x∈M be a collection of vector spaces of dimension r indexed by points of M .
Suppose that we are given a covering {Uα } of M and vector space isomorphisms ρα,x : Ex → Rr when
x ∈ Uα . If the functions hαβ : Uα ∩ Uβ → GL(r, R) defined by
hαβ (x) := ρα,x ◦ ρ−1
β,x ∈ GL(r, R)
`
are smooth then there is a unique structure of rank r vector bundle on E := x Ex → M for which
 
a a
EUα = Ep , ψα : Ep → Uα × Rr by ψα (v) = (p, ρα,p (v)) if v ∈ Ep 
p∈Uα p∈Uα

make an atlas and the hαβ are the transition functions.

`
Proof. First we must put a smooth manifold structure on the total space E = x Ex . The proof is
similar to the proof that the tangent bundle has a structure of a smooth manifold. Namely, for each
member of the covering Uα we have a map
a
ψα : EUα := Ex → Uα × Rr by ψα |Ex = ρα,x .
x∈Uα

We declare this to be a diffeomorphism, i.e. we give EUα the topology and smooth structure given by
ψα . We leave to be checked that the topology on E = ∪α EUα defined in this way is second countable
and paracompact. To check that we have a smooth atlas we must check that transition functions are
smooth, i.e. if Uα ∩ Uβ 6= ∅ then the map
−1
ψβ ψα
Uβ ∩ Uα × Rr −→ EUα ∩Uβ −→ Uβ ∩ Uα × Rr
is smooth. By construction this is given by (id, hαβ ) which is smooth by hypothesis. This puts a structure
of smooth manifold on E, and the map π : E → M is defined in the obvious way on each chart.
We leave the remaining verifications in the lemma to be checked. 
60

The lemma allows standard constructions of linear algebra to be extended to vector bundles. If
E → M and E 0 → M are vector bundles of rank r and r0 then there are vector bundles
E∗, E ⊗k , Λk E, E ⊗ E0, Hom(E, E 0 ), E ⊕ E0
over M which, fibrewise, are
Ex∗ , Ex⊗k , Λ k Ex , Ex ⊗ Ex0 , Hom(Ex , Ex0 ), Ex ⊕ Ex0 .
By the lemma, it is enough to give smooth transition functions. If hαβ (resp. h0αβ ) are transition
functions for E (resp. E 0 ) then the transition functions for these bundles (with respect to an atlas which
trivializes both E and E 0 ) are given by
t −1
hαβ ∈ GL(r, R), hαβ ⊗ · · · ⊗ hαβ ∈ GL(rr0 , R), ∧k hαβ ∈ GL(Λk Rr ), ...

In many books (e.g. Bott-Tu) you will find an alternate way to construct a vector bundle out
of transition functions satisfying the cocycle condition. Let {Uα } be a covering of M and let {hαβ :
Uα ∩ Uβ → GL(r, R)}α,β be a collection of smooth maps satisfying the cocycle condition (5.1.1). Define
the space .
a
E := Uα × Rr ∼
α
where the relation ∼ is as follows: For (x, v) ∈ Uα × Rr and (y, w) ∈ Uβ × Rr we have
(x, v) ∼ (y, w) ⇐⇒ x = y and v = hαβ (w).
This is an equivalence relation and the quotient defines a topological space with a map π : E → M .
• Show that E defined in this way is a smooth manifold. (This amounts to showing that E is
diffeomorphic to the construction in Lemma 5.1.1.)
Now suppose that E is a vector bundle, trivialized over {Uα } by ψα and suppose that ψα0 is another
family of trivializations. It is easy to check that the transition functions {hαβ } and {h0αβ } are related as
follows: There exist maps λα : Uα → GL(r, R) such that
hαβ = λα h0αβ λ−1
β on Uα ∩ Uβ . (5.1.2)
• Show that isomorphism classes of vector bundles on M which admit a trivialization over the cover
{Uα } are in bijection with collections of maps {hαβ : Uαβ → GL(r, R)}α,β satisfying the cocycle condition
(5.1.1) modulo the relation (5.1.2). This implies the following:
Lemma 5.1.2. Two vector bundles on M are isomorphic if and only if their transition functions with
respect to some open cover of M are equivalent (in the sense of (5.1.2)).

A homomorphism between the vector bundles π : E → M and π 0 : E 0 → M is a smooth map

f : E → E 0 such that
f
E −−−−→ E0
 

πy
 0
yπ
M M
commutes and which is fibrewise linear, i.e. Ex → Ex0 is linear. Note that the rank of f |Ex may not be
constant in x, so that the kernel and cokernel are not necessarily vector bundles. An isomorphism of
vector bundles is an invertible homomorphism. (Exs: A vector bundle homomorphism is an isomorphism
if and only if it is an isomorphism fibrewise.)
A subbundle of a vector bundle E → M is a vector bundle E 0 → M with an injective homomor-
phism E 0 → E. Put another way, a subbundle is a submanifold E 0 ⊂ E such that there exists an atlas
0
(Uα , ψα ) for E such that for each α, π −1 (Uα ) ∩ E 0 = ψα−1 (Rr × {0}). (It is clear that such an E 0 is a
vector bundle and there is an injective vector bundle homomorphism E 0 → E.) The quotient bundle
0
E/E 0 can be defined as follows: Take the union x Ex /Ex0 with the isomorphisms ρ̄α,x : Ex /Ex0 → Rr−r
`
given by
0 0
Ex /Ex0 → Rr /(Rr × {0}) ∼ = Rr−r .
and apply Lemma 5.1.1. There is an obvious exact sequence of vector bundles
0 → E 0 → E → E/E 0 → 0.
A section of a vector bundle π : E → M over U ⊂ M is a smooth map s : U → π −1 (U ) ⊂ E such
that π ◦ s = id. A section over U = M will be called a global section.
 61

Examples 5.1.3. (i) The trivial bundle of rank r, where π : Rr × M → M by projection. The transition
functions of this bundle are hαβ = Id for any chart. (It follows from (5.1.2) that a bundle is trivial (i.e.
isomorphic to the trivial bundle) if there exists a covering {Uα }α and smooth maps λα : Uα → GL(r, R)
such that the transition functions are given by hαβ = λα λ−1 β .)
(ii) The tangent bundle T M → M and the cotangent bundle T ∗ M → M are vector bundles of rank
n. This gives a vector bundle T M ⊗r ⊗ T ∗ M ⊗s for any r, s. A section of this bundle is called a tensor
field of type (r, s) on M .
The tangent bundle is trivialized by any atlas on the manifold. In an atlas {(Uα , ϕα )} the transition
functions of the tangent bundle are given by hαβ = Jac(ϕα ◦ ϕ−1 β ). Since the cotangent bundle is its
 −1
dual it has transition functions gαβ = t Jac(ϕα ◦ ϕ−1β ) in this atlas.
(iii) The top exterior power Λn T ∗ M of the cotangent bundle on a smooth n-manifold M is a line
bundle (= rank one vector bundle). The transition functions are given by hαβ = det(Jac(ϕα ◦ ϕ−1β )
−1

in an atlas.
(iv) On any manifold M (not necessarily oriented) the orientation line bundle is defined, with
respect to any atlas, by the transition functions
det(Jac(ϕα ◦ ϕ−1
β )
hαβ = (= sign of det(Jac(ϕα ◦ ϕ−1
β ))
|det(Jac(ϕα ◦ ϕ−1
β )|

i.e. it is the sign of the transition functions for Λn T ∗ M . The resulting line bundle is denoted orM .
(v) The tautological bundle on M = RP n is a line bundle whose fibre at a point in RP n is the
corresponding one-dimensional subspace of Rn+1 . Explicitly, it is the subbundle of the trivial n + 1-
dimensional bundle defined by:
E := {(v, x) ∈ Rn+1 × RP n : v ∈ x}.
with π : E → RP n by π(v, x) = x.
• Show that this is a vector bundle, i.e. find local trivializations and transition functions for it.
(iv) The Moebius bundle on RP 1 is defined as the quotient of S 1 × R by the action of Z/2 by
(x, v) 7→ (−x, −v).
• The tautological bundle over RP 1 = S 1 is isomorphic to the Moebius bundle.
(vi) The tautological complex line bundle over CP n is
E := {(v, x) ∈ Cn+1 × CP n : v ∈ x}.
This is a vector bundle in which the fibres are of complex dimension one, so as a real vector bundle it
has rank two. The associated circle bundle (in each fibre one takes the circle S 1 ⊂ C = R2 ) has total
space S 2n+1 ⊂ Cn+1 .
(vii) There is a tautological bundle of rank k over the real Grassmannian Gr(k, n). As in the case
Gr(1, n + 1) = RP n it can be realized as the subbundle {(v, E) ∈ F n × Gr(k, n) : v ∈ E} of the trivial
bundle Rn × Gr(k, n). Similarly for the complex Grassmannian.
Remark 5.1.4. The following obvious remark will be used repeatedly: A vector bundle of rank r is
isomorphic to the trivial bundle if and only if it has r sections which are linearly independent at every
point. If E → M is such a bundle
P and s1 , . . . , sr are r such sections then s1 (x), . . . , sn (x) are a basis for
Ex , so the map sending v = i ci si (x) ∈ Ex to (x, c1 , . . . , cn ) defines a map E → M × Rr which is an
isomorphism.

5.2. Pullback. A useful construction with vector bundles is the pullback. Given a smooth map f :
M → N and a vector bundle π : E → N the pullback of E by f is a vector bundle f ∗ E → M which
has (f ∗ E)p = Ef (p) . The total space is
f ∗ E := {(v, p) ∈ E × M |π(v) = f (p)}
and projection to the second factor gives f ∗ E → M . To define a vector bundle we must give an open
covering and transition functions. If {Uα }α , {hαβ }α,β are the data for E then the covering {f −1 (Uα )}α
and functions {f ∗ hαβ = hαβ ◦ f }α,β define a vector bundle structure on f ∗ E → M . (Check this.) There
is a pullback map on sections: if s is a section of E → N then f ∗ s is the section of f ∗ E defined by
(f ∗ s)(p) = s(f (p)) ∈ Ef (p) = (f ∗ E)p . (Check that this defines a smooth section of f ∗ E → M .)
62

Example 5.2.1. (i) If i : W → M is a submanifold and E is a vector bundle on M then i∗ E is a vector

bundle on W , simply the restriction of E to W .
(ii) If i : W → M is a submanifold then there is a vector bundle monomorphism T W → i∗ T M and
the quotient
NM/W = i∗ T M/T W
is called the normal bundle of W in M . If M is given a Riemannian metric then (NM/W )x can be
identified with (Tx W )⊥ in Tx M and this identifies NM/W with a subbundle of i∗ T M . However, this is
not canonical and so you should not think of NM/W as a subbundle of i∗ T M but only as a quotient.
n
(iii) There is a natural map Gr(k, n) → RP (k )−1 by sending E ⊂ Rn to ∧k E ⊂ ∧k Rn . What is the
pullback of the tautological bundle?

5.3. Presheaves and sheaves. It will be convenient to use the language of sheaves in what follows.
Let X be a topological space. Let Op(X) be the category with objects the open sets in X, and
morphisms inclusions between open sets. A presheaf of sets on X is a contravariant functor F :
Op(X) → Sets such that F (∅) is a one-element set and F (id : U → U ) is the identity map of F (U ). A
presheaf of abelian groups on X is a contravariant functor from the category Op(X) to the category of
abelian groups which takes the empty set ∅ to {e} and the morphism U ⊂ U to the identity. Concretely,
this means we have, for each open set U ⊂ X, an abelian group F (U ) and for each pair of open sets
V ⊂ U we have a homomorphism ρU V : F (U ) → F (V ), such that for W ⊂ V ⊂ U we have

ρVW ◦ ρU U
V = ρW

and moreover F (∅) = {e} and ρU U

U = id. The homomorphism ρV will be called “restriction from U to V ”.
An element s ∈ F (U ) is referred to as a section of F over U .
Some examples of presheaves:
(1) For a topological space X, the functor U 7→ C(U, R) (continuous real-valued functions on U )
with ρVU restriction of functions from V to U ⊂ V .
(2) For a smooth manifold M the functor U 7→ C ∞ (U ) with restriction of functions defines the sheaf
of smooth functions.
(3) For a complex manifold the functor U 7→ {holomorphic functions U → C} with restriction of
functions.
(4) For a smooth manifold M the functor U 7→ X (U ) = smooth vector fields on U with obvious
restriction maps.
(5) Let X be a topological space and k ≥ 0 an integer. The functor U 7→ H k (U, Z) (singular
cohomology groups of U ) defines a presheaf, where ρVU : H k (V, Z) → H k (U, Z) is the restriction
in cohomology.
(6) For X = C the functor U 7→ {bounded holomorphic functions on U } defines a presheaf.
A morphism between presheaves F and G has the obvious meaning, namely a natural transformation
of functors from F to G. Concretely this means we are given for each U ⊂ X a homomorphism F (U ) →
G(U ) and for every pair V ⊂ U , the diagram
F (U ) −−−−→ G(U )
 
ρU
  U
Vy yρV
F (V ) −−−−→ G(V )
commutes.
A presheaf F on X is a sheaf if it satisfies the following two conditions:
(S1) If s1 , s2 ∈ F (U ) and {Uα }α is an open covering of U such that ρU U
Uα (s1 ) = ρUα (s2 ) for all α, then
s1 = s2
(S2) If U is open in X, {Uα }α is an open covering of U and {sα ∈ F (Uα )}α a collection of sections
Uβ
satisfying ρU Uα ∩Uβ (sα ) = ρUα ∩Uβ (sβ ) for all α, β, then there is a section s ∈ F (U ) such that
α

sα = ρU Uα (s) for all α.

Informally, (S2) says that local sections of sheaves glue together to give sections, (S1) says that sections
of sheaves are determined by what they are locally.
 63

Among the examples of presheaves given above, (1)–(4) are sheaves. The example (5) for k ≥ 1 fails
(S1) in general (think of a manifold with nontrivial H 1 ). The example (6) also fails (S2). (19)
A morphism of sheaves is a morphism between them considered as presheaves, i.e. a natural trans-
formation between the functors.
Given a vector bundle π : E → M on a manifold M , the functor on Op(M )
U 7→ { sections of E over U }
defines a presheaf of abelian groups on M which is easily seen to be a sheaf. This is called the sheaf of
sections of E → M . It is a sheaf of modules over the sheaf of smooth functions on M .
• A homomorphism E → E 0 is the same as a section of the vector bundle Hom(E, E 0 ).
A vector bundle E → M is trivial if and only if the sheaf of sections is isomorphic, as a module over
the sheaf of smooth functions E 0 , to the sheaf E 0 ⊕ · · · ⊕ E 0 (r = rk(E) copies).
Let F be a sheaf on X. The stalk of F at x ∈ X is the direct limit
Fx := limx∈U F (U ).
−→
(In the case of smooth functions this is the space of germs of smooth functions at x.)
Lemma 5.3.1. A homomorphism F → G of sheaves is an isomorphism if and only if the induced map
on stalks Fx → Gx is an isomorphism for all x ∈ X.

We will say more about sheaves later. For the moment we take as a definition that a sequence of
sheaves E → F → G is exact if the sequences of stalks Ex → Fx → Gx is exact for all x ∈ X.

5.4. Differentials or 1-forms.

A differential or differential 1-form on a manifold M is a section of the cotangent bundle. (This
is the dual notion to the notion of vector field.)
∂
Suppose that we are in Rn with coordinates x1 , . . . , xn . The vector fields ∂x 1
, . . . , ∂x∂n form a basis
for the tangent bundle over smooth functions. Let dx1 , . . . , dxn be the dual basis of 1-forms defined by
 
∂
dxj = δij .
∂xi
Then every 1-form on Rn can be written as
X
fi dxi
i
for smooth functions fi . Thus on a smooth manifold a 1-form can be written locally in this form.
Let f : M → R be a smooth function. Then df : T M → R is a linear form on the tangent bundle,
i.e. a section of the cotangent bundle. Let us introduce the notation
E 0 := sheaf of smooth functions
E 1 := sheaf of smooth differential 1-forms.
Then f 7→ df defines a sheaf homomorphism
d : E0 → E1
The kernel is the sheaf of constants, i.e. ker(d)(U ) consists of functions that are constant on each
connected component of U . The forms in the image are called exact.
∂ai ∂a
• Show that in Rn the 1-form ω = i ai dxi is of the form df if and only if ∂x
P
j
= ∂xji for all i and j.
• Show that the differential dθ on S 1 is not of the form df for a function f . (Here dθ means the
differential which is d(θ) in a coordinate patch - the function θ does not make sense globally on S 1 .)
• Show that the differential
−y x
ω= 1 dx + 1 dy
(x2 + y2 ) 2 (x2 + y2 ) 2
on R2 − {0} is not exact. (Hint: Integrate ω over the circle.)

19An example of a presheaf which fails (S2) and (S1) is the following: Let X = {x, y} with the discrete topology. For
an open set U ⊂ X let F (U )= functions from U to R. For V ⊂ U let ρU V := 0.
6=
64

5.5. Differential forms. A differential k-form on M is a section of the kth exterior power Λk T ∗ M →
M of the cotangent bundle. Concretely, this means we have, for each point p ∈ M , an alternating
multilinear form (Tp∗ M )⊗k → R which varies smoothly in p. Informally, a k-form is a machine which
takes k vector fields and gives back a number, and it is alternating for the action of the symmetric group
on k letters on (Tp∗ M )⊗k . The assignment
E k (U ) = { differential k-forms on U } (U ⊂ M open).
defines a sheaf on M . This is a sheaf of modules over the sheaf of smooth functions E 0 .
Suppose we are in Rn with coordinates x1 , . . . , xn . The choice of coordinates fixes a trivialization of
the tangent bundle as T Rn ∼ = Rn × T0 Rn which amounts to taking the n global vector fields ∂x ∂
1
, . . . , ∂x∂n .
The dual basis of the cotangent bundle is the set of 1-forms dx1 , . . . , dxn . Thusa 
basis (over the smooth
n n
functions) of the differential k-forms on an open subset of R is given by the differential k-forms
k
dxI = dxi1 ∧ · · · ∧ dxik
where I = {i1 , . . . , ik } runs over all subsets of {1, . . . , n} of size k. Note that dxI stands for the alternating
multilinear form of k variables defined by:
dxI (X1 , . . . , Xk ) = det (dxir (Xs ))
where X1 , . . . , Xk are sections of T M , i.e. vector fields. So for example (dx1 ∧ dx2 )(X1 , X2 ) =
dx1 (X1 ) dx1 (X2 )
and more generally dxI ((Xi )i∈I ) = (dxi (Xj ))i∈I,j∈I .
dx2 (X1 ) dx2 (X2 )
If ω is a k-form on a manifold M and U is the domain of a chart with local coordinates x1 , . . . , xn
then there is a (unique) local expression
X
ω= ωI dxI
|I|=k

for smooth functions ωI . (In words, we are using the basis dxI to trivialize the kth exterior power of the
cotangent bundle and writing a general section of this bundle in terms of the basis, with coefficients in
E 0 .)
If f : M → N is a smooth manifold then there is a pullback map:
f ∗ : E k (M ) → E k (N )
Pointwise this is given by the multilinear form (f ∗ ω)p defined by:
dfp ×···×dfp ωf (p)
Tp M × · · · × Tp M −−−−−−−→ Tf (p) N × · · · × Tf (p) N −−−−→ R.
In terms of sections of vector bundles the pullback comes from the pullback of sections of ENk to sections
of f ∗ ENk followed by the sheaf homomorphism f ∗ ENk → EMk
.

5.6. Exterior derivative.

The homomorphism
d : E0 → E1
defined above extends to forms of higher order:
Theorem 5.6.1. There is a unique R-linear sheaf homomorphism d : E k → E k+1 which agrees with
d : E 0 → E 1 for k = 0 and which satisfies
(i) d(α ∧ β) = dα ∧ β + (−1)deg(α) α ∧ dβ (Leibniz rule)
(ii) d2 = 0 (flatness).
The map d commutes with pullback: If f : M → N is a map of manifolds then d(f ∗ (α)) = f ∗ (dα).

Proof. We will first prove the uniqueness of such a map (over any manifold). Then we will define d in
coordinate patches by an explicit formula. By uniqueness the definitions must agree on overlaps and
hence we have a globally defined d : E k → E k+1 . (What we are really using here is the fact that the
functor Hom(E k , E k+1 ) defined by U 7→ Hom(E k |U , E k+1 |U ) is itself a sheaf.) Then we will check (i)
and (ii) by calculation in coordinate patches.
Let U ⊂ M be open and ω ∈ E k (U ). To verify that dω is uniquely determined by the (i) and (ii) it
is enough to do so on a coordinate patch, i.e .we may assume U is contained in a chart with coordinates
 65

P
x1 , . . . , xn . Write ω = I ωI dxI . Then d dx PI = 0 (reduce to the case d dxi = 0 by repeated use of (i) and
(ii) and d dxi = 0 by (ii)) and hence dω = I dωI ∧ dxI . This proves that dω is uniquely determined.
P
To define d we cover M by charts and just use the formula dω = I dωI ∧ dxI in each coordinate
chart. If we verify the Leibniz rule and flatness conditions then by the uniqueness just proved there is
agreement on the overlaps of charts, so d : E k → E k+1 is well-defined. For both it is enough to check on
forms of the type f dxI . We check flatness:
!
X ∂f X ∂2f
d(d(f dxI )) = d(df ∧ dxI ) = d dxi ∧ dxI = dxj ∧ dxi ∧ dxI .
i
∂xi i,j
∂xj ∂xi
Now the equality of mixed partials and the relation dxi ∧ dxj = −dxj ∧ dxi show this is zero. It only
remains to check the Leibniz rule for forms of the form α = f dxI and β = gdxJ , which is left as an
exercise.
Compatibility with pullback is immediate for exact one-forms (i.e. df for f a smooth function)
and for functions. The general case follows using that (1) locally on N every differential form is a
linear combination of wedge products of exact one-forms with function coefficients, (2) f ∗ is a ring
homomorphism, and (3) the Leibniz rule holds for d. 

• A 1-form ω ∈ E 1 (Rn ) is exact (i.e. ω = df for f ∈ E 0 (Rn )) if and only if dω = 0.

The following lemma gives an expression for the exterior derivative which does not depend on local
coordinates:
Lemma 5.6.2. Let α be a k-form on M . For vector fields X0 , . . . , Xk ,
X
dα(X0 , . . . , Xk ) = (−1)i Xi α(X0 , . . . , X
ci , . . . , Xk )
i
X
+ (−1)i+j−1 α([Xi , Xj ], X0 , . . . , X
ci , . . . , X
cj , . . . , Xk ).
i<j

Proof. Check that this definition satisfies the Leibniz rule and flatness and conclude by uniqueness that
this is the same. (Alternately, a calculation in local coordinates will suffice.) 

5.7. Lie derivative. Let X be a vector field on M . The contraction operator iX : E k → E k−1 is
defined as follows: For a k-form ω we have
iX ω(X1 , . . . , Xk−1 ) := ω(X, X1 , . . . , Xk−1 ).
(When k = 0 the contraction of a function is defined to be zero.) Evidently i2X = 0. The Lie derivative
operator LX : E k → E k is defined by:
LX ω := diX ω + iX dω = (d + iX )2 ω
This is Cartan’s formula. (20) If k = 0, i.e. we have a function, then we have
LX f = iX df = Xf
(the vector field applied to the function). The formal properties of the Lie derivative are easy to verify:
Lemma 5.7.1. The Lie derivative satisfies:
(i) dLX = LX d
(ii) LX (ω1 ∧ ω2 ) = LX ω1 ∧ ω2 + (−1)deg(ω1 ) ω1 ∧ LX ω2 (Leibniz rule)
(iii) If ϕ : M → M 0 is a diffeomorphism then ϕ∗ (LX ω) = Lϕ∗ (X) (ω).

Proof. (i) follows from LX = diX + iX d and d2 = 0.

(ii) follows from the Leibniz rule for iX and d.
(iii) Exercise. 

There is a more geometric description of the Lie derivative (this is usually given as the definition):
20Cartan’s formula says that the homomorphism of complexes L : (E ∗ (M ), d) → (E ∗ (M ), d) is homotopic to the zero
X
map by the homotopy operator iX . This has a concrete consequence: If we take A∗X ⊂ E ∗ (M ) to be the subcomplex of
forms which satisfy LX ω = 0 then A∗X ⊂ E ∗ (M ) induces an isomorphism in cohomology. In other words, the complex A∗X
also computes the de Rham cohomology of M .
66

Lemma 5.7.2. Suppose X is a vector field with associated local flow α. Then
d
LX ω := αt∗ (ω).
dt t=0

The proof is left as an exercise. (Hint: Since LX = diX + iX d is a derivation (i.e. satisfies the
Leibniz rule) you can reduce to checking the equality locally on functions and exact 1-forms.)
Here is an example showing that the Lie derivative generalizes familiar objects: Let αt be a local
flow on an open subset of Rn with associated vector field X. Then one has the identity LX (dx1 ∧ · · · ∧
dxn ) = div(X) dx1 ∧ · · · ∧ dxn where div(X) is the classical divergence of a vector field, defined by
P ∂
P ∂fi
div( i fi ∂x i
) = i ∂xi
.

5.8. Frobenius theorem revisited. Suppose F is a rank k integrable subbundle of T M . Let

I 1 (F ) := {ω ∈ E 1 (M ) : ω(p) vanishes on Fp ⊂ Tp M for all p ∈ M }.
Let I (F ) be the ideal in E ∗ (M ) generated by I 1 (F ). Then we have the following slightly more general
version of Frobenius theorem (which we will not prove or use here):
Theorem 5.8.1. A subbundle F ⊂ T M has integral manifolds through every point if and only if
dI (F ) ⊂ I (F ).

Proof. Exercise. (Use Lemma 5.6.2 and the Frobenius theorem.) 

5.9. Orientations. Recall that an orientation for a real vector space is an equivalence class of bases,
under the relation that two bases are equivalent if the matrix relating them has positive determinant.
There are thus two possible orientations for a real vector space.
The top exterior power Λn=dimV V is a one-dimensional R-vector space. A choice of generator for
Λ V specifies an orientation uniquely: any generator of Λn V is of the form v1 ∧ · · · ∧ vn and it gives the
n

basis {v1 , . . . , vn }. Conversely an orientation gives a generator of Λn V (nonuniquely): If {v1 , . . . , vn } is

any basis in the equivalence class then v1 ∧ · · · ∧ vn spans Λn V . The fact that for g ∈ GL(V ) the operator
Λn g : Λn V → Λn V is multiplication by det(g) makes this well-defined.
An oriented atlas for a manifold M is a smooth atlas {(Uα , ϕα )} in which the transition functions
satisfy the condition
det(Jac(ϕβ ◦ ϕ−1
α )) > 0 on ϕ(Uα ∩ Uβ ) ⊂ Rn .
If M admits an oriented atlas it is called orientable. Two oriented atlases which are compatible (in the
sense that their union is a smooth atlas) are compatibly oriented if their union is oriented. This is an
equivalence relation on oriented atlases; an equivalence class under this relation is an orientation on
M . If we fix an orientation we call M oriented.
Here is an equivalent way to define an orientation: An orientation is the choice of an orientation for
each (Tx M )∗ (or Tx M ) which is locally constant in x. Here by locally constant we mean the following:
given x there is a chart around U with coordinates x1 , . . . , xn such that at any point in U the orientation
of (Tx M )∗ is the same as dx1 ∧ · · · ∧ dxn . It is an easy exercise to see that this is equivalent to the notion
defined above.
Just as it is convenient to think of orientations of a vector space in terms of the top exterior power
it is convenient to think of orientations in terms of sections of the bundle Λn T ∗ M . We have:
Lemma 5.9.1. An n-manifold M is orientable if and only if Λn T ∗ M has a nowhere-vanishing section.

Proof. Suppose M is orientable. Let {(Uα , ϕα )} be an (locally finite) oriented atlas. Choose a partition
of unity {fα } subordinate to the covering. Define
X
ω := fα ϕ∗α (dx1 ∧ · · · ∧ dxn ).
α

Let us check that ω is nowhere zero. Let x ∈ M and let Uα1 , . . . , Uαr be the charts containing x. Then
r
X
ω(x) = fαi (x)ϕ∗αi (dxi1 ∧ · · · ∧ dxin )
i=1
 67

where xi1 , . . . , xin are the coordinates on ϕαi (Uαi ) ⊂ Rn . Now the coefficients fαi (x) are all positive, so
it is enough to check that all the ϕ∗αi (dxi1 ∧ · · · ∧ dxin ) lie in the same component of Λn Tx∗ M − {0}. But
the change of coordinates from xi1 , . . . , xin to xj1 , . . . , xjn is ϕαj ◦ ϕ−1
αi and hence on Uαi ∩ Uαj we have
j
ϕ∗αi (dxi1 ∧ · · · ∧ dxin ) = det(Jac(ϕαi ◦ ϕ−1 ∗ j
αj )) ϕαj (dx1 ∧ · · · ∧ dxn ).

Since the atlas is oriented we know det(Jac(ϕαi ◦ ϕ−1

αj )) > 0 and we are done.
Conversely suppose ω is a nowhere-vanishing section of Λn T ∗ M . Let {(Uα , ϕα )} be an atlas. For
each α write ω on Uα as
ω = gα ϕ∗α (dx1 ∧ · · · ∧ dxn ) .
Then gα ∈ C ∞ (Uα ) is either always > 0 or always < 0. If, at x ∈ M , the gα (x) for Uα 3 x all have
the same sign then the transition functions ϕαi ◦ ϕ−1 αj are all orientation preserving at x. Therefore it is
enough to modify the atlas so that this holds. Indeed, this is easy to do: If gα < 0 on Uα then replace the
chart (Uα , ϕα ) by the chart (Uα , s ◦ ϕα ) where s(x1 , x2 , . . . , xn ) = (x2 , x1 , . . . , xn ). Now all the functions
gα are > 0 and hence the new atlas is oriented. 

A nowhere-vanishing top form on a manifold is called a volume form.

• Show that S n is orientable and RP n is orientable if and only if n is odd.
Remark 5.9.2. Suppose M is a complex manifold. Then its underlying real manifold is always orientable
and in fact comes with a preferred choice of orientation. Indeed, consider first the case of Cn or any
open subset of it with coordinates z1 = x1 + iy1 , . . . , zn = xn + iyn . The nowhere-vanishing 2n-form
n(n−1) i
n
(−1) 2 dz1 ∧ · · · ∧ dzn ∧ dz̄1 ∧ · · · ∧ dz̄n = dx1 ∧ dy1 ∧ · · · ∧ dxn ∧ dyn
2n
gives an orientation on Cn . Now any holomorphic map f : Cn → Cn or between open subsets of Cn
automatically satisfies det(Jac(f  )) ≥ 0.  (For n = 1 the Jacobian of a holomorphic function f (z) =
ux vx
u(x, y) + iv(x, hy) is Jac(f ) = which by the Cauchy-Riemann equations ux = vy , uy = −vx is
  uy vy
ux vx
equal to so det(Jac(f )) = u2x + vx2 ≥ 0. The determinant is strictly positive everywhere since
−vx ux
one cannot have ux = vx = 0 unless f is constant. For n > 1 one has to work with the generalization
of the CR conditions.) So a holomorphic atlas of a complex manifold is automatically oriented, and the
choice of the orientation above locally fixes a choice of orientation.

5.10. Integration on oriented manifolds. On Rn there is a canonical measure, namely the Lebesgue
measure, which is the unique (up to scalar multiple) translation-invariant measure, and with respect to
which you can integrate functions. On a general manifold there is no canonical measure (obviously you
cannot patch together the Lebesgue measures on charts), so we don’t know how to integrate functions.
The correct thing to integrate are top forms (on an oriented manifold).
First let us consider the case of an open subset U ⊂ Rn . Then ω ∈ Ecn (U ) can be written ω =
f dx1 ∧ · · · ∧ dxn and we make the definition
Z Z ∞ Z ∞ Z ∞
ω := ··· f dx1 dx2 · · · dxn (5.10.1)
U xn =−∞ x2 =−∞ x1 =−∞

where on the right-hand side one has the usual Lebesgue (or, in this case, Riemann) integral of a function.
(f is extended by zero outside U .)
Now if ϕ : V → U  is an orientation-preserving diffeomorphism, then, by assumption, the Jacobian
∂ϕi
determinant det ∂yj is positive on V . Then by the usual change-of-variables formula for integration
in Rn we have Z Z
ϕ∗ (ω) = ω. (5.10.2)
V U
If ϕ is orientation-reversing then this is true with a minus sign.
We will use the following simple lemma. (We could have used the earlier lemmas about partitions
of unity, but this is even easier.)
Lemma 5.10.1. Let C ⊂ M be a compact subset of a manifold and let {Uα }α be a family of open subsets
such that C ⊂ ∪α Uα . Then there exist finitely many
P smooth functions fi : M → [0, 1] such that each fi
has compact support contained in some Uαi and i fi = 1 on C.
68

Proof. For each x ∈ C choose α(x) with x ∈ Uα(x) . Let gx : M → [0, 1] be a smooth function with
compact support supp(gx ) ⊂ Uα(x) and with gx identically 1 on some open neighbourhood Vx of x. (If
ϕ : W → Rn is a chart neighbourhood with W ⊂ Uα(x) then take a bump function b which is constant
1 in a neighbourhood of ϕ(x) and with supp(b) ⊂ ϕ(W ). Then take gx = ϕ∗ b and extend it by zero
outside W .)
Q
Now {Vx }x∈C is an open cover of C hence
P has a Q
finite subcover Vx1 , . . . , Vxr . The product i (1−gxi )
is zero on Vx1 ∪ · · · ∪ Vxr and hence h = i gxi + i (1 − gxi ) is > 0 everywhere. Then the collection
fi := gxi /h is as required. 

Now let M be an oriented n-manifold, i.e. a manifold with a chosen orientation, and let ω ∈ Ecn (M )
be a compactly supported n-form. Let {(Uα , ϕα )} be an oriented atlas of charts. The lemma gives a
finite collection of functions fi partitioning unity on C with supp(fi ) compact, and for each i there is a
αi with supp(fi ) ⊂ Uαi . Define the integral of ω over M to be the finite sum:
Z XZ
ω := (ϕ−1 ∗
αi ) (fi ω). (5.10.3)
M i ϕαi (Uαi )

The integrals on the right are defined by (5.10.1).

R
Lemma 5.10.2. (i) The integral M ω is well-defined, i.e. does not depend on the choice of oriented
chart giving the fixed orientation or on the choice of partition of unity.
(ii) It depends on the choice of orientation only up to a sign, i.e. if the orientation on M is reversed
then the integral changes sign (because each term in the sum (5.10.3) changes sign).
(iii) If the orientation of M is defined by the nowhere-vanishing
R form ω0 ∈ Λn T ∗ M and ω = f ω0
for f ≥ 0, compactly supported, and not identically zero, then M ω 6= 0.

Proof. Suppose {(Uα , ϕα )} and {(Vβ , ψβ )} are two oriented atlases giving the same orientation. Let
ω ∈ Ecn (M ) be a top form with compact support. Using the lemma we can find functions {fi }ri=1 and
{gi }si=1 . Then we have
 
XZ XZ X XZ
−1 ∗ −1 ∗ 
(ϕαi ) (fi ω) = (ϕαi ) gj fi ω =
 (ϕ−1 ∗
αi ) (gj fi ω)
i Uαi i Uαi j i,j ϕαi (Uαi ∩Vβj )

while
!
XZ XZ X XZ
(ψβ−1
j
)∗ (fi ω) = (ψβ−1
i
)∗ fi gj ω = (ψβ−1
j
)∗ (gj fi ω).
j Vβj i Vβj i i,j ψβj (Uαi ∩Vβj )

Now the diffeomorphism

ϕαi ◦ ψβ−1
j
: ψβj (Uαi ∩ Vβj ) → ϕαi (Uαi ∩ Vβj )
is orientation-preserving (by the assumption that the two atlases define the same orientation) and hence
by (5.10.2) the i, j terms in each expression above are the equal. This proves the first statement.
The other statements of the lemma are easily checked. 

5.11. Stokes’s theorem.

This theorem is usually stated in the context of manifolds with boundary. To avoid the trouble of
introducing this concept we will deal with a slightly simpler situation. A domain in an n-manifold M
is an open subset Ω ⊂ M such that
(i) the (topological) boundary ∂Ω = Ω − Ω is an embedded submanifold of M of dimension n − 1
and
(ii) there is an atlas of M such that for each (U, x1 , . . . , xn ) in the atlas for which U ∩ ∂Ω 6= ∅, the
intersection Ω ∩ U is given by xn ≤ 0 (and U ∩ ∂Ω is given by xn = 0).
(This terminology is not entirely standard.) So for example the unit ball B n in Rn is a domain; its
boundary is S n−1 .
Note that the integral defined above can easily be extended to the case of a domain in an oriented
manifold. For ω ∈ Ecn (M ) and a domain Ω ⊂ M , choose an oriented atlas with property (ii) above. Next
 69

choose functions fi and charts Uαi as before. If Uαi ∩ ∂Ω 6= φ the contribution to the sum (5.10.3) is
simply defined to be the Riemann integral
Z Z xn =0 Z ∞
−1 ∗
(ϕαi ) (fi ω) = ··· g dx1 . . . dxn .
ϕαi (Uαi ∩Ω) xn =−∞ x1 =−∞

where g has been extended by zero outside ϕαi (Uαi ∩ Ω).

The following lemma will allow us to consistently orient the boundary of a domain in an oriented
manifold. The proof is left as an exercise.
Lemma 5.11.1. If ϕ : Rn → Rn is a diffeomorphism which carries Hn := {xn ≤ 0} to itself then
det(Jac(ϕ)) > 0 implies det(Jac(ϕ|∂Hn )) > 0.

If M is an oriented manifold and Ω ⊂ M is a domain then there is an induced orientation on

∂Ω. To define this, first consider the special case (M, Ω) = (Rn , Hn ). In this case if the orientation of
Rn is defined by the n-form dx1 ∧ · · · ∧ dxn then the induced orientation of ∂Hn = Rn−1 is defined by
(−1)n−1 dx1 ∧ · · · ∧ dxn−1 . In the general case suppose that M is oriented, i.e. we have fixed an oriented
atlas {(Uα , ϕα )}. We may further assume that for a chart Uα such that Uα ∩ ∂Ω 6= ∅ in the coordinates
given by ϕα the domain Uα ∩ Ω is given by xn ≤ 0. Then by the previous lemma {(Uα ∩ ∂Ω, ϕα )} is an
oriented atlas for ∂Ω. Locally in a coordinate chart U the orientation on ∂Ω ∩ U is either the induced one
defined in the previous paragraph (special case (M, Ω) = (Rn , Hn ) or its reverse. The induced orientation
is the one which locally looks like the one defined above for the special case. (21)
Theorem 5.11.2. (Stokes) Let M be an oriented smooth manifold of dimension n and let Ω be a domain.
Let ∂Ω be the boundary of Ω, with the induced orientation. Let ω be a compactly supported n − 1-form
on M . Then Z Z
dω = i∗ ω.
Ω ∂Ω

In R this is the fundamental theorem of calculus; in R2 it is Green’s theorem; in R3 it contains the

divergence theorem (Gauss’s theorem) and the classical Stokes theorem. Taking Ω = M gives:
R
Corollary 5.11.3. If ∂M = 0 then M dω = 0.

Proof. First consider the essential special case: M = Rn and Ω = Hn = {xn ≤ 0}. We assume Rn is
given the orientation dx1 ∧ · · · ∧ dxn . Write
X
ω= gi dx1 ∧ · · · ∧ dx
c i ∧ · · · ∧ dxn .
i
Then
X ∂gi
dω = (−1)i−1 dx1 ∧ · · · ∧ dxn
i
∂xi
and Z Z xn =0 Z xn−1 =∞ Z x1 =∞
X ∂gi
dω = (−1)i−1 ··· dx1 . . . dxn .
Ω i xn =−∞ xn−1 =−∞ x1 =−∞ ∂xi
Because each gi vanishes outside a compact set all terms except the i = n term are zero, giving:
Z Z xn−1 =∞ Z x1 =∞
dω = (−1)n−1 ··· gn (x1 , . . . , . . . , xn−1 , 0) dx1 . . . dxn−1
Ω −∞ x1 =−∞
Z
= (−1)n−1 i∗ gn dx1 . . . dxn−1 .
Rn−1
∗ ∗ ∗ ∗
On the other
R R i∗ ω = i gn dx1 ∧ · · · ∧ dxn−1 since i dxn = 0 and i dxj = dxj for j < n. So the
hand,
equality Ω ω = ∂Ω i ω holds if we take into account the definition of induced orientation (i.e. the fact
that dx1 ∧ · · · ∧ dxn−1 is (−1)n−1
R -oriented for the induced orientation on R
n−1
). The same argument
n
shows that if Ω = M = R then Rn dω = 0 for compactly supported ω.
Now consider a general oriented manifold M and a domain Ω ⊂ M . Choose an oriented atlas {Uα } in
which each Uα is diffeomorphic to Rn and such that if Uα ∩ ∂Ω 6= ∅ then there are coordinates on Uα such
that Uα ∩ ∂Ω is given by xn ≤ 0. (This can be done since Ω is a domain.) Let ω ∈ Ecn−1 (M ). Choose

21The purpose of the sign in the induced orientation is to avoid a sign in the statement of Stokes’s theorem. You will
see a sign in books which have a different convention about the induced orientation.
70

functions {fi } which partition unity on supp(ω) and with each supp(fi ) ⊂ Uαi
a finite collection of P
compact. Write ω = i fi ω so that
X
dω = d(fi ω).
i
Then
Z XZ XZ
dω = d(fi ω) = d(fi ω)
M i M i Uαi
XZ
= (ϕ−1 ∗
αi ) (d(fi ω))
i ϕαi (Uαi ∩Ω)
XZ
d (ϕ−1 ∗


= αi ) (fi ω) .
i ϕαi (Uαi ∩Ω)

The charts with Uαi ⊂ Ω contribute nothing by the special case above for Ω = Rn . The remaining charts
(which intersect ∂Ω) have coordinates in which Ω is given by xn ≤ 0. By the special case proved above
we have, for such i,
Z Z
−1 ∗
j ∗ (ϕ−1 ∗


d (ϕαi ) (fi ω) = αi ) (fi ω)
ϕαi (Uαi ∩Ω) ϕαi (Uαi ∩∂Ω)
Z
= (ϕ−1 ∗ ∗
αi ) i (fi ω)
ϕαi (Uαi ∩∂Ω)
Z
= i∗ (fi )i∗ ω.
Uαi ∩∂Ω

(Here j is the inclusion of ϕαi (Uαi ∩ ∂Ω) in ϕαi (Uαi ∩ Ω), i.e. j = ϕαi R◦ i ◦ ϕ−1 ∗
αi .) Since {i fi } is a
∗
partition of unity for the oriented atlas {Uαi ∩ ∂Ω}, summing over i gives ∂Ω i ω. 

5.12. De Rham cohomology. The smooth differential forms on M with the exterior derivative d form
a complex of sheaves of R-vector spaces, the de Rham complex:
0 → E 0 → E 1 → · · · → E n → 0.
The homology groups of the complex of global sections
0 → E 0 (M ) → E 1 (M ) → · · · → E n (M ) → 0
are the de Rham cohomology groups of M :
i ker(d : E i (M ) → E i+1 (M ))
HdR (M ) := .
im(d : E i−1 (M ) → E i (M ))
These are real vector spaces (not obviously finite-dimensional). Note that in degree zero we get just the
locally constant functions, so that if M is connected
0
HdR (M ) = R.
The formal properties of de Rham cohomology are summarized in the following proposition:
Proposition 5.12.1. The de Rham cohomology groups have the following properties:
∗ k
(i) The sum HdR (M ) = ⊕k≥0 HdR (M ) is a graded-commutative algebra with the product defined by
exterior product of differential forms: [ω] ∧ [τ ] = [ω ∧ τ ].
(ii) They are contravariantly functorial for smooth maps: For f : M → N there is a ring homomor-
phism
f ∗ : HdR
k
(N ) → HdRk
(M ) by [ω] 7→ [f ∗ ω].
(iii) They are diffeomorphism invariant, i.e. if f : M → N is a diffeomorphism then f ∗ is an
isomorphism.

Proof. (i) We check that the product is well-defined: Suppose ω and τ are closed forms. If ω is replaced
by ω + dη then
[(ω + dη) ∧ τ ] = [ω ∧ τ ] + [dη ∧ τ ] = [ω ∧ τ ] + [d(η ∧ τ ) + (−1)deg(η)+1 η ∧ dτ ] = [ω ∧ τ ].
Graded-commutativity then follows from graded-commutativity of the wedge product of forms.
(ii) follows immediately from f ∗ (dω) = d(f ∗ ω).
(iii) f ∗ : E ∗ (N ) → E ∗ (M ) is an isomorphism of complexes. 
 71

One can also consider the complex of sheaves of k-forms with compact support:
0 → Ec0 → Ec1 → · · · → Ecn → 0
where Eck (U ) is the space of k-forms with support in a compact set. The homology groups of the global
sections are the de Rham cohomology groups with compact support:
i ker(d : Eci (M ) → Eci+1 (M ))
HdR,c (M ) = .
im(d : Eci−1 (M ) → Eci (M ))
i i 0
Of course HdR,c (M ) = HdR (M ) if M is compact. If M is noncompact then HdR,c (M ) = 0.
∗
Proposition 5.12.2. The groups HdR,c (M ) have the following properties:
∗ k
(i) The sum HdR,c (M ) = ⊕k≥0 HdR,c (M ) is a ring under wedge of differential forms and also a
∗
module over HdR (M ).
∗
(ii) The groups HdR,c (M ) are contravariantly functorial for proper maps of manifolds.
∗
(iii) The groups HdR,c (M ) are covariantly functorial for inclusions of open subsets.
n
R
(iv) For oriented M , integration defines a nonzero map M : HdR,c (M ) → R.
(v) If an orientation on M is fixed then
Z
(ω, τ ) 7→ ω∧τ
M
defines a nonzero pairing
k n−k
HdR (M ) × HdR,c (M ) → R.

Proof. (i) is straightforward from the Leibniz rule for d.

(ii) comes from the fact that if ω ∈ Eck (N ) and f : M → N is proper then f ∗ (ω) ∈ Eck (M ).
(iii) comes as follows: if U ⊂ M is open and ω ∈ Eck (U ) then extending ω by zero outside U defines
a form i∗ (ω) ∈ Eck (M ). Clearly d(i∗ ω) = i∗ (dω) so that this gives a map in cohomology HdR,c
k
(U ) →
k
HdR,c (M ).
(iv) follows from the corollary to Stokes’s theorem.
(v) check that the pairing descends to cohomology (Stokes’s theorem Ragain). This is the same at
k n−k n
the product HdR (M ) ⊗ HdR,c (M ) → HdR,c (M ) composed with integration M . 

For a connected oriented manifold the integration map in (iv) is an isomorphism. It is a version of
Poincaré duality that for a connected oriented manifold the pairing in (v) is nondegenerate. (We will
not prove these facts here; see Bott-Tu for a proof.)

5.13. Poincaré lemma.

Lemma 5.13.1. For any smooth manifold M the projection π : M × R → M induces an isomorphism
π ∗ : HdR
k k
(M ) → HdR (M × R).

Proof. Let it : M → M × R be it (x) = (x, t). Since π ◦ it = id we know that π ∗ is injective. It remains
to prove it is surjective. A k-form on M × R can be written as ω = ω1 + dt ∧ ω2 where ω1 and ω2 involve
no dt. (Both ω1 and ω2 may depend on t.) Consider the form
Z t 
σ := ω − d ω2 (s)ds .
0
This means the following: In local coordinates on any coordinate patch of M × R, which we assume to
be of the formP(x1 , . . . , xn , t) where xi are local coordinates on a patch in M , we can write ω2 as a sum
of terms like I fI (x1 , . . . , xn , t)dxI running over subsetsI ⊂ {1, . . . , n} with |I| = k. Then the above
P Rt
integral means: In each such coordinate patch take I 0 f (x1 , . . . , xn , s)ds dxI where the integral
here is an ordinary Riemann integral. These expressions obviously agree on overlaps (as one varies the
Rt
patch in M ) and hence define a differential k-form, which is what is meant by 0 ω2 (s)ds.
k
Now dσ = dω = 0 and clearly σ and ω represent the same element of HdR (M × R). The form dt
∂
P
does not appear in σ, so if, in local coordinates, σ = |I|=k σI dxI , then dσ = 0 implies that ∂t (σI ) = 0
∗
for all I. Hence σ is actually constant in t, i.e. σ = π (τ ) for τ a closed form on M . This proves the
lemma. 
72

k
Corollary 5.13.2 (Poincaré lemma). HdR (Rn ) = 0 for k > 0 and HdR
0
(Rn ) = R.

Since any point on a manifold has neighbourhoods diffeomorphic to Rn , the Poincaré lemma says
that the following is an exact sequence of sheaves on M :
0 → RM → E 0 → E 1 → · · · → E n → 0
where RM is the constant sheaf on M (i.e. the sheaf of locally constant functions). In other words, the
de Rham complex (E • , d) is a resolution of the constant sheaf.

5.14. Homotopy invariance. Two smooth maps f, g : M → N are smoothly homotopic if there
exists a smooth map F : R × M → N such that F (0, x) = f (x) and F (1, x) = g(x). (Here R can be
replaced by any open interval containing 0, 1.)
Lemma 5.14.1. If f and g are smoothly homotopic then f ∗ = g ∗ .

Proof. Let it : M → M × R be the inclusion it (x) = (x, t) and let π : M × R → M be the projection.
Then π ◦ it = id and hence i∗t π ∗ = id. We saw earlier that π ∗ is an isomorphism, hence i∗t = (π ∗ )−1 is
independent of t. In particular i∗0 = i∗1 and hence f ∗ = i∗0 F ∗ = i∗1 F ∗ = g ∗ . 

It is a fact that in any (continuous) homotopy class of maps between two smooth manifolds there is a
smooth map, and two smooth maps in the class are smoothly homotopic. Thus we see that a (continuous)
homotopy class of maps between manifolds induces a well-defined map in de Rham cohomology.

5.15. Mayer-Vietoris sequence. Let U, V ⊂ M be open subsets (hence manifolds themselves) such
that U ∪ V = M . Let i : U ∩ V ,→ U and j : U ∩ V ,→ V be the inclusions. For each k there is a
short-exact sequence
i∗ −j ∗
0 → E k (M ) → E k (U ) ⊕ E k (V ) −→ E k (U ∩ V ) → 0.
Exactness at the first two places is obvious, so we must only check surjectivity at the end. Let ω ∈
E k (U ∩ V ). Let ρU + ρV = 1 be a partition of unity subordinate to {U, V }. Then ρU ω defines an element
of E k (V ) by extending it by zero on V − U ∩ V . Similarly ρV ω defines an element of E k (U ). Then
σ := (−ρV ω, ρU ω) ∈ E k (U t V )
goes to ω under i∗ − j ∗ .
Since the maps in the short exact sequence are compatible with differentials we have a short-exact
sequence of complexes
i∗ −j ∗
0 → E ∗ (M ) → E ∗ (U ) ⊕ E ∗ (V ) −→ E ∗ (U ∩ V ) → 0.
and hence a long exact sequence in de Rham cohomology
k k k k+1
· · · → HdR (M ) → HdR (U t V ) → HdR (U ∩ V ) → HdR (M ) → · · · .

A good cover of a manifold is a covering {Uα } such that all finite intersections ∩ki=1 Uαi are
diffeomorphic to Rn . The Mayer-Vietoris sequence, induction, and the Poincaré lemma give:
Theorem 5.15.1. The de Rham cohomology groups of a manifold with a finite good cover are finite-
dimensional.

There is a standard argument to show that a compact manifold always has a (finite) good cover,
which uses a Riemannian metric and geodesically convex sets. However, we will not give it here.
There is a Mayer-Vietoris sequence for compactly supported de Rham cohomology: Let {U, V } be
an open covering of M . The Mayer-Vietoris sequence
i i i
· · · HdR,c (U ∩ V ) → HdR,c (U t V ) → HdR,c (M ) → · · ·
comes from the following short exact sequence of complexes
i∗ −j∗
0 → Ec∗ (U ∩ V ) → Ec∗ (U t V ) −→ Ec∗ (M ) → 0
where {i : U ,→ M, j : V ,→ M } is an open cover of M . (The maps i∗ , j∗ here come from the covariant
functoriality of Eck (−) under open inclusion; verify the exactness of the sequence.)
 73

k
Example 5.15.2. It follows easily from the M-V sequence that HdR (S n ) = R for k = 0, n and is zero
n n
otherwise. We can write down an explicit generator for HdR (S ) as follows: Consider the form ω ∈
E n (Rn+1 ) defined by
n+1
X
ω= (−1)i−1 xi dx{1,...,n+1}−{i} .
i=1
Then dω = (n + 1)dx{1,...,n+1} is a multiple of the volume form on Rn+1 . Let i : S n ,→ Rn+1 be the
inclusion. Then
σ := i∗ ω ∈ E n (S n )
n
is closed and [σ] is a generator of HdR (S n ) because:
Z Z
σ= dω 6= 0
Sn B n+1
n+1
where B is the ball of radius one.
∗
Example 5.15.3. Let us write down a generator for HdR (CP n ). (At this stage we do not yet know that
∗ n n+1 2n+1
HdR (CP ) = R[x]/(x ), so assume this.) Consider the map
√p:S → CP n which is√ the restriction
n+1 n
of C − {0} → CP . For i = 0 to n √let dzi = dxi + −1dyi and dz̄i = dxi − −1dyi so that
√
dzi ∧ dz̄i = −2 −1dxi ∧ dyi . Restricting 2−1 i dzi ∧ dz̄i = i dxi ∧ dyi to S 2n+1 gives a closed real
P P
2-form σ.
• Show that there is a closed 2-form ω on CP n such that p∗ (ω) = σ.
• Show that ω is invariant under the natural action of U (n + 1) on CP n .
• Show that ω k is nonzero for any k ≤ n.

5.16. De Rham theorem. The de Rham cohomology groups of a smooth manifold depend, a priori,
on the smooth structure of the manifold. In fact, they depend only on the underlying topological space
and they compute the singular cohomology groups, tensored with R:
∗
Theorem 5.16.1. There are natural isomorphisms HdR (M ) ∼
= H ∗ (M, R) and HdR,c
∗
(M ) ∼
= Hc∗ (M, R).

Proof. For the general case this will be proved in the next section using sheaf-theoretic methods. Here
let’s prove it for compact manifolds using the existence of a finite good cover. Indeed, if {Ui }ni=1 is a
finite good cover, then a straightforward induction using the Mayer-Vietoris sequence and the Poincaré
lemma proves the theorem. 

The isomorphism of the theorem is functorial with respect to smooth maps, but we will not prove
this now.
n
Note that for an oriented manifold M , the isomorphism HdR,c (M ) ∼
= Hcn (M, R) relates the integra-
n n
R
tion map M : HdR,c (M ) → R to the map P DM : Hc (M, R) → H0 (M, R) = R used in the Poincaré
duality theorem. (In the case M is compact P DM is cap product with the fundamental class [M ] given
n
by the orientation). Thus we can conclude that if M is connected then HdR,c (M ) ∼
= R. Also, we can
apply the Poincaré duality theorem to conclude that the pairing
k n−k
HdR (M ) × HdR,c (M ) → R.
R
by (ω, τ ) 7→ M ω ∧ τ is a nondegenerate pairing. (For a purely de Rham cohomology proof of this duality
see Bott&Tu.)

5.17. Relative de Rham cohomology (∗). Let f : M → N be a smooth map. We will define de
Rham cohomology groups H ∗ (f : M → N ). When f is an embedding this will give back the relative
cohomology H ∗ (M, N, R) under a relative version of the de Rham theorem.
THIS IS STILL TO BE WRITTEN
74

6. Sheaves and cohomology

In this section we will discuss sheaves on topological spaces and their cohomology and apply them
to manifolds. Good references for the material in this section are Voisin’s Hodge Theory and Complex
Geometry, vol. I or Ramanan’s Global Calculus.
(Note that the definitions we use, which are the ones in these books, are sometimes different from
those of Warner’s book and some other books.)

6.1. Presheaves. Let X be a topological space. Let Op(X) be the category with objects the open sets
in X, and morphisms inclusions between open sets. A presheaf of sets on X is a contravariant functor
F : Op(X) → Sets such that F (∅) is a one-element set and F (id : U → U ) is the identity map of
F (U ). A presheaf of abelian groups on X is a contravariant functor from the category Op(X) to the
category of abelian groups which takes the empty set ∅ to {e} and the morphism U ⊂ U to the identity.
Concretely, this means we have, for each open set U ⊂ X, an abelian group F (U ) and for each pair of
open sets V ⊂ U we have a homomorphism ρU V : F (U ) → F (V ), such that for W ⊂ V ⊂ U we have

ρVW ◦ ρU U
V = ρW

and moreover F (∅) = {e} and ρU U

U = id. The homomorphism ρV will be called “restriction from U to V ”.
An element s ∈ F (U ) is referred to as a section of F over U .
A morphism between presheaves F and G has the obvious meaning, namely a natural transformation
of functors from F to G. Concretely this means we are given for each U ⊂ X a homomorphism F (U ) →
G(U ) and the diagram below commutes for every pair V ⊂ U :
F (U ) −−−−→ G(U )
 
ρU
  U
V y yρV
F (V ) −−−−→ G(V ).

A presheaf of rings (or algebras) has the obvious meaning: a covariant functor from Op(X) to rings
(or algebras) such that F (∅) = {0} and F (U ⊂ U ) = id.
Examples 6.1.1. Here are some examples of presheaves:
(1) The assignment U 7→ C 0 (U, R) (continuous functions from U to R with restriction of functions
defines a presheaf of abelian groups. Here R may be replaced by any topological space Y to get a presheaf
of sets.
(2) Fix an abelian group A. The assignment U 7→ A defines the constant presheaf, where all ρU
V = id.
(3) Let X = C, or more generally any complex manifold. For an open subset U ⊂ X define O(U )
V : O(U ) → O(V ) be the restriction of functions.
to be the ring of holomorphic functions on U , and let ρU
(4) Let M be a smooth manifold and let E (U ) be the smooth functions from U to R, and ρU
V is
restriction of functions. This defines a presheaf E on M .
(5) For X = C let B be the presheaf defined by B(U )= bounded holomorphic functions on U , ρU
V=
restriction of functions.
(6) Fix a point x ∈ X and an abelian group A. Define F by F (U ) = A if x ∈ U and F (U ) = {0}
if x ∈
/ U . The restriction maps are the obvious ones (identity or zero). This presheaf is called the
skyscraper presheaf at x with value A.
(7) Let X be a smooth manifold. The functor U 7→ X (U ) = smooth vector fields on U defines a
presheaf, as does the functor of sections of any smooth vector bundle on X.
(8) Let X be a topological space and k ≥ 0 an integer. The functor U 7→ H k (U, Z) (singular
cohomology groups of U ) defines a presheaf. Restriction maps are restriction maps in cohomology.

6.2. Sheaves. A sheaf on X is a presheaf F on X which further satisfies the following two conditions:
(S1) If s1 , s2 ∈ F (U ) and {Uα }α is an open covering of U such that ρU U
Uα (s1 ) = ρUα (s2 ) for all α, then
s1 = s2
(S2) If U is open in X, {Uα }α is an open covering of U and {sα ∈ F (Uα )}α a collection of sections
Uβ
satisfying ρU Uα ∩Uβ (sα ) = ρUα ∩Uβ (sβ ) for all α, β, then there is a section s ∈ F (U ) such that
α

sα = ρU Uα (s) for all α.

 75

Informally, (S2) says that local sections of sheaves glue together to give sections, (S1) says that sections
of sheaves are determined by what they are locally.
Examples 6.2.1. Among the examples of presheaves given above, (1), (3), (4), (6), (7) are sheaves and
(8) is a sheaf in the case k = 0. (2) and (5) are not sheaves because they fail to satisfy (S2). (An example
of a presheaf which fails (S2) and (S1) is the following: Let X = {x, y} with the discrete topology. For
an open set U ⊂ X let F (U )= functions from U to R. For V ⊂ U let ρU V := 0.) The presheaf (8) for
6=
k ≥ 1 fails (S1) in general, e.g. if X is a manifold or locally contractible.
• For any topological space Y , the presheaf U 7→ C 0 (U, Y ) defines a sheaf of sets on X. If Y is a(n
abelian) topological group then this defines a sheaf of (abelian) groups.
• For any smooth manifolds M and N the presheaf U 7→ C ∞ (U, N ) defines a sheaf of sets on M . If
N is a(n abelian) Lie group then this defines a sheaf of (abelian) groups.

6.3. Stalks, sheafification. Sheaves are (as we shall see) much more convenient to work with than
presheaves, although it is usually presheaves which appear “in nature”. So it is useful to have a canonical
(i.e. functorial) way to associate a sheaf to a presheaf.
Let F be a presheaf on X. The stalk of F at x ∈ X is the direct limit:
Fx := lim F (U ).
−→
U 3x
For U 3 x denote the natural map to the stalk by
ρU
x : F (U ) → Fx .
Now if F satisfies (S1) then the collection (ρU
x (s))x∈U determines the section s ∈ F (U ), in the sense that
the map Y
F (U ) → Fx
x∈U
is injective. In fact this is equivalent to (S1). This suggests how, given a presheaf, we should make a
sheaf out of it.
Q
A continuous section of F over U is an element (sx )x∈U ∈ x∈U Fx such that for each x ∈ U
22
there exists a V and σ ∈ F (V ) such that for all y ∈ V we have sy = ρU y (σ). ( ) Define the presheaf F̃
by
F̃ (U ) = { continuous sections of F over U }.
The restriction maps F̃ (U ) → F̃ (V ) are given by (sx )s∈U 7→ (sx )x∈V . There is an obvious presheaf
homomorphisms
θ : F → F̃
(since an actual section of F over U gives a continuous section).
Let us see that F̃ defined like this is a sheaf:
Lemma 6.3.1. The sheafification F 7→ F̃ has the following properties:
(1) The construction F 7→ F̃ is functorial.
(2) The map θ induces an isomorphism Fx ∼ = F̃x of stalks.
(3) If F is a sheaf then θ : F → Fe is an isomorphism.
(4) The construction has a universal property: If G is a sheaf and φ : F → G is a presheaf homo-
morphism then there is a unique sheaf homomorphism φ̃ : F̃ → G such that φ = φ̃ ◦ θ.
(5) Hom(F, G) = Hom(F̃ , G̃), i.e. sheafification is a faithful functor.
Example 6.3.2. (i) Consider the examples (2) and (5) above which were not sheaves. It is easy to see
that the sheafification of the constant presheaf A is the sheaf which assigns to U the group of locally
constant functions U → A, i.e. Aπ0 (U ) . This is called the constant sheaf with value A. The sheafification
of (5) is the sheaf of all holomorphic functions.
(ii) Note that a presheaf can have nonzero global sections even if all its stalks are zero. (This cannot
happen for a sheaf by (S1).) An example is the presheaf U 7→ H k (U ) of de Rham or singular cohomology
22Here “continuous” does not refer to a topology; it is simply a terminology for us. However there is a natural topology
` `
on the coproduct x∈X Fx such that these are exactly the continuous sections over U of the natural map x∈X Fx → X
`
which takes the stalk Fx to x. (Exercise) The space x∈X Fx is called the espace etalé of F , cf. e.g. Hartshorne, Chapter
II, Exs.
76

groups on a manifold for k > 0. By the Poincaré lemma (for de Rham cohomology) or the contractibility
of balls (for singular cohomology) this has zero stalk at all points, but it has global sections e.g. if
H k (M ) 6= 0. This presheaf fails to satisfy (S1) because any two classes in H k (M ) have the same (=zero)
restriction to small enough coordinate neighbourhoods. In this case the associated sheaf is zero.
Example 6.3.3. Here is an important example of a sheaf arising from differential equations. Let X = C∗
and fix α ∈ C. Let F be the sheaf of holomorphic solutions of the differential equation
d
z f = αf (6.3.1)
dz
i.e. over an open set U ⊂ C∗ the group F (U ) is the holomorphic functions on U satisfying the differential
equation. The restriction maps are given by restricting functions. This defines a presheaf and it is easy
to see that it is a sheaf, a subsheaf of the sheaf of all holomorphic functions on C.
(1) Suppose first that α ∈ Z. In this case f (z) = z α is a holomorphic solution to (6.3.1) which
makes sense in all of C∗ . Moreover it is clear that every solution is a multiple of this one. Therefore the
sheaf F is isomorphic to the constant sheaf C.
(2) Suppose now that α ∈ / Z. Then z α = exp(α log(z)) defines a solution of (6.3.1) in any domain
U in which we can choose a branch of the logarithm function, i.e. any domain not containing a real
half-line. Moreover in such a domain any solution is a multiple of z α , so F (U ) ∼
= C with a basis given by
z α . On the other hand F (U ) = {0} if U contains a punctured disk. Thus F is not the constant sheaf,
but it is locally constant in the sense that any point has a neighbourhood in which F is isomorphic to
the constant sheaf C. It follows that the stalks are all isomorphic to C.
This example can be generalized by looking at solutions to
 n
d
z −α f =0
dz
for n ≥ 1. For example if α ∈ Z then there are n linearly independent solutions defined on all of C∗ and
the sheaf is isomorphic to the constant sheaf Cn . For α ∈
/ Z the sheaf of solutions is locally constant and
locally isomorphic to the constant sheaf Cn .
Example 6.3.4. Consider the previous example with α ∈ / Z but with domain X = C, i.e. let G be the
sheaf of local solutions of (6.3.1) on C. Now G(U ) = {0} if 0 ∈ U . Thus the stalk at 0 is zero. The stalks
at other points are the same since i−1 G = F where i : C∗ → C is the inclusion and F is as earlier.

6.4. Kernels, cokernels, images. Let F → G be a homomorphism of sheaves. The presheaf kernel,
presheaf cokernel, and presheaf image are the functors
U 7→ ker(F (U ) → G(U )),
U 7→ im(F (U ) → G(U )),
U 7→ coker(F (U ) → G(U )).
It is easy to see that the first is already a sheaf, called the kernel sheaf and denoted ker(F → G).
However, simple examples show that the image and cokernel presheaves are not necessarily sheaves. (23)
So we define the image sheaf and the cokernel sheaf to be the sheaves associated with the image
presheaf and cokernel presheaf respectively. We will use the notation im(F → G) and coker(F → G)
for these objects as we will never need to use the presheaf image and presheaf cokernel.
In general, if F ⊂ G is a subsheaf then we will define the quotient sheaf to be the sheaf associated
with the presheaf U 7→ G(U )/F (U ). In this way the kernel and cokernel become the sheaves associated
with the presheaf kernel and cokernel. With this definition, we have
Lemma 6.4.1. (1) A homomorphism of sheaves F → G is an isomorphism if and only if it is an
isomorphism on all stalks.
(2) Two subsheaves F 0 and F 00 of F are equal if and only if Fx0 = Fx00 in Fx .
(3) A sequence F1 → F2 → F3 of sheaves is exact if and only if the corresponding sequence of stalks
F1,x → F2,x → F3,x is exact for every x ∈ X.

23Here is an example for cokernel: Let F = G = O be the sheaf of holomorphic functions on X = C∗ and let O → O
d
be the operator dz . Then every point x ∈ C∗ has a neighbourhood U for which the presheaf cokernel has sections over U
equal to {0} (any small enough neighbourhood of x will do). But the presheaf cokernel has nonzero sections over X = C∗
since e.g. z1 is not in the image of dz
d
on C∗ . Thus the presheaf cokernel is not a sheaf.
 77

6.5. Example: Exponential sequence. Here is a basic example of a short exact sequence of sheaves on
any smooth manifold X. For any abelian Lie group G let GX be the sheaf on X defined by U 7→ C ∞ (U, G).
Then there is a short exact sequence
0 → ZX → CX → C∗ X → 0
where the second map is defined by f 7→ exp(2πif ). That this map is surjective on stalks is easy to
check and comes down to the fact that for small enough open neighbourhoods of any point in C∗ the
exponential map can be inverted. This also gives an example of a short exact sequence of sheaves such
that taking sections over an open set does not necessarily give a short exact sequence of abelian groups.

6.6. Exact functors, global sections functor. Recall the notions of exact functor and left-exact
functor. A functor T : Sh(X) → AbGps or T : Sh(X) → Sh(Y ) is called exact if it takes short-exact
sequences to short-exact sequences. It is left-exact if 0 → T (A) → T (B) → T (C) is exact whenever
0 → A → B → C is exact. It is right-exact if T (A) → T (B) → T (C) → 0 is exact whenever
A → B → C → 0 is exact.
The global sections functor is the functor
Γ(X, −) : Sh(X) → AbGps
defined by taking sections over X. So in the earlier notation, Γ(X, F ) = F (X).
Lemma 6.6.1. The global sections functor is left-exact.

Proof. This is straightforward. 

Examples 6.6.2. Here are examples to show that Γ(X, −) is not an exact functor:
(1) Let X = C∗ and consider the exponential sequence on X. The map Γ(X, CX ) → Γ(X, C∗ X ) is
not since e.g. the function z has no logarithm in X.
(2) Let X = CP 1 and O the sheaf of holomorphic functions on X. Let x, y ∈ X be two distinct
points. Then there is a short exact sequence
0 → I{x,y} → O → ix∗ C ⊕ iy∗ C → 0.
(Here the second map is f 7→ (f (x), f (y)) and I{x,y} is the kernel, i.e. the ideal of functions vanishing at
x and y.) The only global sections of O are constant functions, so the map C = Γ(X, O) → Γ(X, ix∗ C ⊕
iy∗ C) = C ⊕ C is not surjective.

6.7. Pullback, pushforward. Let f : X → Y be a continuous map of topological spaces. Given a

sheaf F on X we can define a presheaf on Y by
U 7→ F (f −1 (U )).
This is already a sheaf, called the pushforward or direct image of F by f , and denoted f∗ F . (Exercise:
Check that this defines a sheaf.) Note that a special case of direct image is Γ(X, −), which is direct
image by the map X → pt.
Given a sheaf G on Y we can defind a presheaf on X by
U 7→ lim G(V ).
−→
V ⊃f (U )

This is not a sheaf in general, so we define the pullback or inverse image sheaf f −1 G to be the sheaf
associated with this presheaf. The stalks are related by (f −1 G)x = Gf (x) for x ∈ X. For any abelian
group A we have f −1 AY = AX .
(Remark: Note that if ix : {x} → X is the inclusion of a point then for an abelian group A the
skyscraper sheaf at x with value A is ix∗ A. For a sheaf F we have i−1
x F = Fx (the stalk at x).)

Lemma 6.7.1. The inverse image functor f −1 : Sh(Y ) → Sh(X) is exact. The direct image functor
f∗ : Sh(X) → Sh(Y ) is left-exact.

Proof. The first assertion follows from the fact that (f −1 F )x = Fy and the fact that a short-exact
sequence of sheaves is exact if and only if it is exact on every stalk. The second assertion is left as an
exercise. (Note that if f : X → pt is the map to a point then f∗ F = Γ(X, F ).) 
78

There are natural transformations of functors id → f∗ f −1 and f −1 f∗ → id which define homomor-

phisms Hom(f −1 F, G) ↔ Hom(F, f∗ G). The pair (f −1 , f∗ ) is an adjoint pair of functors, which means
(by definition) that there is a natural isomorphism

Hom(f −1 G, F ) = Hom(G, f∗ F )

for F ∈ Sh(X), G ∈ Sh(Y ). (Exercise: Verify this.) We say f −1 is left adjoint to f∗ or f∗ is right
adjoint to f −1 .

Lemma 6.7.2. (1) If L is an exact functor and R is right adjoint to L then R takes injective objects to
injective objects.
(2) f∗ takes injective sheaves to injective sheaves.
(3) f∗ takes flasque sheaves to flasque sheaves.

6.8. Internal Hom and tensor product. Given two sheaves F, G of abelian groups on a space X we
have the internal Hom presheaf defined by

U 7→ Hom(F, G)(U ) := Hom(F |U , G|U )

where F |U means the pullback by the inclusion U ,→ X. (24)

Lemma 6.8.1. If F and G are sheaves then Hom(F, G) is a sheaf.

Proof. Exercise. 

Something to keep in mind: In general, the natural map Hom(F, G)x → Hom(Fx , Gx ) is neither
injective nor surjective. (25)
The adjointness relation between f∗ and f −1 then has a sheaf version, which gives the previous one
after taking global sections:

Lemma 6.8.2. If f : X → Y is a continuous map of topological spaces and F ∈ Sh(X) and G ∈ Sh(Y )
then f∗ Hom(f −1 G, F ) ∼
= Hom(F, f∗ G).

Proof. Exercise. 

For F, G ∈ Sh(X) the tensor product F ⊗ G is the sheaf associated with the presheaf

U 7→ F (U ) ⊗ G(U ).

(The presheaf is not in general a sheaf, so we must sheafify.) The stalks of the tensor product are given
by (F ⊗ G)x = Fx ⊗ Gx . (Easy exercise.)
A complex of sheaves on a topological space X is a sequence of sheaves (F i )i∈Z on X and
morphisms F i → F i+1 (i ∈ Z) such that the composition of any two successive morphisms is zero. We
write:
· · · → F −1 → F 0 → F 1 → F 1 → · · ·
We say “F i appears in degree i”. Taking sections over any open set U ⊂ X gives a complex of abelian
groups · · · → Γ(U, F i ) → Γ(U, F i+1 ) → · · · .
Let F be a sheaf on X. A resolution of F is a complex of sheaves 0 → K 0 → K 1 → · · · such that
the following sequence is exact:
0 → F → K0 → K1 → · · · .

24One might think of defining it by Hom(F, G)(U ) = Hom(F (U ), G(U )) but this does not make any sense: There are
no natural restriction maps for V ⊂ U .
25 Here is an example: Let X = C and i : {0} ,→ C the inclusion. There is an injection i Z ,→ Z ; let Q be the
∗ X
cokernel. Then ZX → Q gives a nonzero element of Hom(ZX , Q)0 while Hom(Z, Q0 ) = 0 since Q0 = 0.
 79

6.9. Injective sheaves and injective resolutions. An injective sheaf is an injective object in the
abelian category of sheaves, i.e. a sheaf I ∈ Sh(X) for which the functor Hom(−, I) is exact. It is easy
to see that this is equivalent to the following property: In the diagram
0 −−−−→ A −−−−→ B


y
I
if A → B is a sheaf monomorphism then there exists a homomorphism B → I making the diagram
commute. A short exact sequence 0 → I → G → H → 0 with I injective has a splitting. (Look at the
identity I → I to get a morphism G → I which splits the sequence.)
Example 6.9.1. (i) Let I be an injective abelian group. Then the skyscraper sheaf ix∗ I is injective. Indeed,
this follows from the equality Hom(F, ix∗ I) = Hom(Fx , I) (which is a special case of the adjointness of
i−1
x and ix∗ ).
Q
Q (ii) The product of injective sheaves is injective. (Indeed, this follows from Hom(−, α Fα ) =
α Hom(−, Fα ).)

The following lemma is usually quoted as saying “Sh(X) has enough injectives”:
Lemma 6.9.2. Any sheaf admits a monomorphism to an injective sheaf.

Proof. Recall that any abelian group admits a monomorphism to an injective (=divisible) abelian group.
ThusQfor each x ∈ X the stalk Fx admits an injection jx : Fx ,→ Ix for some injective Ix . Now define
I = x∈X ix∗ Ix . This is an injective sheaf. Consider the morphism
Y Y
F ,→ ix∗ Fx ,→ ix∗ Ix = I.
x∈X x∈X
U
Q Q
defined on U by composing x∈U ρx with the product x∈U jx . This is injective on stalks and hence is
injective. 

A resolution
0 → F → I0 → I1 → · · ·
k
where each I is injective is called an injective resolution.
Lemma 6.9.3. (i) Every sheaf of abelian groups has an injective resolution.
(ii) Given two sheaves F and G and resolutions F → I • and G → J • such that J • is injective, any
homomorphism α : F → G lifts to a homomorphism I • → J • such that
0 −−−−→ F −−−−→ I 0 −−−−→ I 1 −−−−→ · · ·
  

αy

y

y (6.9.1)

0 −−−−→ G −−−−→ J 0 −−−−→ J 1 −−−−→ · · ·

commutes. Any two such liftings are homotopy equivalent, i.e. if ϕ = (ϕi : I i → J i )i and ψ = (ψ i : I i →
J i )i are two liftings then there is a homomorphism H = (H i : I i → J i−1 )i from I • to J •−1 such that
dJ H + HdI = ϕ − ψ. (Here dI = (dI : I i → I i+1 )i and dJ = (diJ ) are the morphisms in the resolutions.)
(iii) An injective resolution of a sheaf is unique up to homotopy, i.e. given two injective resolutions
F → I • and F → J • there are maps ϕ : I • → J • and ψ : J • → I • such that ϕψ and ψϕ are both
homotopic to the identity.
(iv) Given an injective homomorphism α : F ,→ G we can find injective resolutions F → I • and
F → J • and injective homomorphisms I k ,→ J k such that the diagram (6.9.1) commutes.

Proof. (i) Existence is proved in the usual way: First let

F ,→ I 0
be a monomorphism to an injective sheaf. The cokernel of F → I 0 can be embedded in injective sheaf
I 1 . The morphism
I0 → I1
is the composition I 0 → coker(F → I 0 ) → I 1 . If I 0 → I 1 is not surjective embed its cokernel in I 2 and
continue.
80

(ii) is straightforward, (iii) follows from (ii).

(iv) We will first prove the following: Given F ,→ G one can find F ,→ I and G ,→ J (injections into
injectives) such that I ,→ J and the natural map I/F → J/G is injective. Indeed, look at the diagram
φ ψ
0 −−−−→ F −−−−→ G −−−−→ Q −−−−→ 0
  
u v w
y y y
0 −−−−→ I −−−−→ J := I ⊕ K −−−−→ K −−−−→ 0
Here u : F ,→ I and w : Q ,→ K with I, K injective have been chosen. The map u : F → I extends to
a map α : G → I with αφ = u (by injectivity of I). Define J = I ⊕ K and v : G ,→ J by v = (wψ, α).
The maps in the bottom row are the obvious ones.
Now suppose F ,→ G is injective. Applying the above we can find injectives I 0 and J 0 and maps
such that all rows and columns in the diagram below are exact:
0 0 0
  
  
y y y
0 −−−−→ F −−−−→ I 0 −−−−→ coker(F → I 0 ) −−−−→ 0
  
φ v w
y y y
0 −−−−→ G −−−−→ J 0 −−−−→ coker(G → J 0 ) −−−−→ 0
Now apply the above observation to the injection coker(F → I 0 ) ,→ coker(G → J 0 ) we find I 1 ,→ J 1
and continue as before. 

6.10. Derived functors, cohomology of sheaves. Let T : Sh(X) → AbGps be a left-exact functor
to abelian groups (or any abelian category). The derived functors of T are defined as follows: Let F
be a sheaf. Take an injective resolution F → I • . Apply the functor T to get a complex of abelian groups
0 → T (I 0 ) → T (I 1 ) → · · ·
and take cohomology:


i i ker T (I i ) → T (I i+1 )
•
R T (F ) := H (T (I )) = .
im (T (I i−1 ) → T (I i ))
(We set T (I −1 ) = 0.) Note that since T is left-exact we have R0 T (F ) = T (F ).
The main example of a left-exact functor is the global sections functor Γ(X, −) : Sh(X) → AbGps.
The cohomology groups H i (X, F ) are defined as its derived functors, i.e.
H i (X, F ) := Ri Γ(X, F ).
In other words, given a sheaf F choose a resolution by injective sheaves
0 → F → I0 → I1 → · · ·
then apply Γ(X, −) to get a complex of abelian groups
0 → Γ(X, I 0 ) → Γ(X, I 1 ) → · · ·
and then take cohomology to define
ker(Γ(X, I k ) → Γ(X, I k+1 ))
H i (X, F ) :=
im(Γ(X, I k−1 ) → Γ(X, I k ))
It follows from Lemma 6.9.3 above that this definition does not depend on the choice of injective reso-
lution.
The following proposition summarizes the properties of derived functors:
Proposition 6.10.1. The objects Ri T (F ) are determined up to canonical isomorphism. The assignment
F → Ri T (F ) has the following properties:
(1) R0 T (F ) = T (F )
(2) If I is an injective sheaf then Ri T (I) = 0 for i > 0.
 81

(3) A short exact sequence

0→F →G→Q→0
of sheaves gives a long exact sequence
0 → T (F ) → T (G) → T (Q) → R1 T (F ) → R1 T (G) → R1 T (Q) → · · ·
(4) If F → G is a homomorphism of sheaves and injective resolutions F → I • and G → J • are
chosen then there is a canonical induced homomorphism Ri F → Ri G.

Proof. (1), (2), (4) follow from the definitions and Lemma 6.9.3.
(3) is proved using the last part of Lemma 6.9.3: Choose injective resolutions F → I • and G → J •
such that I • ,→ J • compatibly with F ,→ G. Then the quotients
K i := J i /I i
is a resolution of Q by injective objects. (The quotient of an injective object by an injective subobject
is injective: use the splitting property to write J i = I i ⊕ K i and then show that Hom(−, K i ) is exact).
Moreover the short exact sequences
0 → Ii → J i → Ki → 0
are split (because I i is injective). Thus applying the functor T gives split short exact sequences
0 → T (I i ) → T (J i ) → T (K i ) → 0
Thus we have a short exact sequence of complexes 0 → T (I • ) → T (J • ) → T (K • ) → 0 which splits
at each i and hence by the usual argument (used, for example, to produce the long exact sequence in
cohomology for a pair (X, A) from the short exact sequence of cochain complexes) gives a long exact
sequence in cohomology:
· · · → H i (T (I • )) → H i (T (J • )) → H i (T (K • )) → H i+1 (T (I • )) → · · · .
This is the desired long exact sequence. 
Remarks 6.10.2. (i) One can ask whether it is possible to left-derive right-exact (covariant) functors, for
example the functor Hom(F, −) : A 7→ Hom(F, A), or (in the way one does for modules over a ring in
order to define Exti or T ori ). One problem is that the category Sh(X) does not usually have enough
projectives, i.e. not every object admits a surjection P → F from a projective object. (A projective
object is one for which Hom(P, −) is exact.) (26) So we will usually derive such bifunctors by resolving
the second variable by injectives. (Another option, when working with sheaves of modules over a fixed
sheaf of rings R, is to use resolutions by sheaves of R-flat modules.)
(ii) Although we call the Ri T derived functors, they are not really functors because the objects
i
R T (F ) depend on the choice of injective resolution, thus they have only been defined up to canonical
isomorphism. Nevertheless, the property (4) means they behave essentially like functors.
We note the following fact for later use:
Lemma 6.10.3. If i : Z ,→ X is the inclusion of a closed subspace then i∗ is an exact functor and
H k (Z, F ) = H k (X, i∗ F )
for F ∈ Sh(X).
Proof. (27) The exactness of i∗ is easily checked on stalks. Now if F → I • is an injective resolution of
F then i∗ I • is a resolution of i∗ F and each i∗ I k is injective (because i∗ is the right adjoint to the exact
functor i−1 it takes injectives to injectives). Thus i∗ I • is an injective resolution of i∗ F and so one has
H k (Z, F ) = H k (Γ(Z, I • )) = H k (Γ(X, i∗ I • )) = H k (X, i∗ F )
because Γ(Z, I k ) = Γ(X, i∗ I k ) for each k. 
26Here is an example: Suppose X is Hausdorff and F is a nonzero sheaf on X. Choose x ∈ X such that F 6= 0. Then
x
ix∗ Fx admits no epimorphism from a projective sheaf. Indeed, suppose P → ix∗ Fx is such an epimorphism. Then there
is an open set U for which P (U ) → Fx is nonzero. Let y ∈ U be a point distinct from x and let j : U − {y} ,→ X be the
inclusion. Then j! j −1 F → ix∗ Fx is surjective (it is surjective on stalks). Thus the map P → ix∗ Fx lifts to P → j! j −1 F .
Applying Γ(U, −) gives that P (U ) → Fx is factors as P (U ) → Γ(U, j! j −1 F ) → Fx . But Γ(U, j! j −1 F ) = 0 so this is a
contradiction.
This shows that on any reasonable space there are no projectives in the category of sheaves. Note however that there
are often plenty of projectives in the category of presheaves.
T
27The proof here shows the following: If A → U
B → C are left-exact functors between abelian categories and T is exact
and takes injectives to injectives then Rk (U ◦ T ) = Rk U ◦ T .
82

6.11. Acyclic resolutions. Let T be a left-exact functor from Sh(X) to AbGps or Sh(Y ) as before. A
sheaf F is called T -acyclic if Ri T (F ) = 0 for i > 0.
Proposition 6.11.1. Let 0 → F → A• be a resolution of F by T -acyclic objects. Then
ker(T (Ai ) → T (Ai+1 ))
Ri T (F ) = H i (T (A• )) = ,
im(T (Ai−1 ) → T (Ai ))
i.e. the derived functors of T can be computed using T -acyclic resolutions. More precisely, if F ,→ A• is
a resolution and F ,→ I • is an injective resolution then the (canonical) induced map
H i (T (A• )) → H i (T (I • )) = Ri T (F )
is an isomorphism when A• consists of T -acyclic objects.

Proof. The proof is by induction on degree, i.e. we assume that Rk T for k ≤ i − 1 is computed by
acyclic resolutions. We want to prove the same for Ri T . Let F ∈ Sh(X) and let F → A• be an acyclic
resolution of F . Look at the short exact sequence:
0 → F → A0 → coker(d0 ) → 0.
The long exact sequence of Ri T is:
0 → T (F ) → T (A0 ) → T (coker(d0 )) → R1 T (F ) → R1 T (A0 ) → · · ·
· · · Ri−1 T (A0 ) → Ri−1 T (coker(d0 )) → Ri T (F ) → Ri T (A0 ) → · · ·
By T -acyclicity of A0 one deduces:
R1 T (F ) = coker(T (A0 ) → T (coker(d0 ))
Ri T (F ) = Ri−1 T (coker(d0 )) (i ≥ 2)
•
On the other hand applying T to the resolution F → A gives the complex
0 → T (F ) → T (A0 ) → T (A1 ) → T (A2 ) →
ker(T (A1 )→T (A2 ))
which is exact at the first two places. Thus H 1 (T (A• )) = im(T (A0 )→T (A1 )) . But the sequence

0 → coker(d0 ) → A1 → A2
is also exact and hence
0 → T (coker(d0 )) → T (A1 ) → T (A2 )
is exact and hence ker(T (A1 ) → T (A2 )) = im(T (coker(d0 ) ,→ T (A1 )) and so
im(T (coker(d0 ) ,→ T (A1 ))
H 1 (T (A• )) = = R1 T (F ).
im(T (A0 ) → T (A1 ))
Now coker(d0 ) has the resolution by T -acyclic objects
0 → coker(d0 ) → A1 → A2 → · · · .
Then the induction hypothesis gives
Ri T (F ) = Ri−1 T (coker(d0 )) = H i (T (A1 ) → T (A2 ) → · · · )
and we are done. 

6.12. Flasque sheaves, canonical flasque resolution. A sheaf is called flasque (or sometimes
flabby) if the restriction maps F (X) → F (U ) are surjective for all U (equivalently, all restrictions
F (V ) → F (U ) for U ⊂ V are surjective).
Proposition 6.12.1. A flasque sheaf is Γ-acyclic.

Proof. Let F be flasque. We can find a short-exact sequence

α β
0→F →I→G→0
Q
where I is both flasque and injective. (Indeed, the proof of Lemma 6.9.2 showed this: the sheaf x ix∗ Ix
in that proof is automatically flasque.) Suppose we show that
β
I(U ) → G(U ) is surjective for all U ⊂ X. (6.12.1)
 83

It would then follow that G is also flasque and we could use the long exact sequence
0 → H 0 (X, F ) → H 0 (X, I) →H 0 (X, G) → H 1 (X, F ) → H 1 (X, I) →
· · · → H i−1 (X, I) →H i−1 (X, G) → H i (X, F ) → H i (X, I) → · · ·
The surjectivity of I(X) → G(X) would show that H 1 (X, F ) = 0 and then the dimension shift
H i (X, F ) ∼
= H i−1 (X, G) plus flasqueness of G would allow an induction to show that H i (X, F ) = 0
for all i > 1.
Let us prove (6.12.1). Let σ ∈ G(U ). Suppose that V, W ⊂ U are open and τV ∈ I(V ) and
τW ∈ I(W ) map (under β) to ρU U
V (σ) and ρW (σ) respectively. On the overlap V ∩ W we have
β
ρVV ∩W (τV ) − ρW V U W U
V ∩W (τW ) 7→ ρV ∩W ρV (σ) − ρV ∩W ρW (σ) = 0

and hence there is ηV ∩W ∈ F (V ∩ W ) such that

α
ηV ∩W 7→ ρVV ∩W (τV ) − ρW
V ∩W (τW ) ∈ I(V ∩ W ).

Since F is flasque there is an η ∈ F (V ) restricting to ηV ∩W (i.e. with ρVV ∩W (η) = ηV ∩W ). Define

τV0 = τV − η.
Then τV0 ∈ I(V ) and τW ∈ I(W ) agree on V ∩ W and τV0 7→ ρU U
V (σ) and τW 7→ ρW (σ). Thus there is a
U
section τ ∈ I(V ∪ W ) which goes to ρV ∪W (σ), i.e we have increased the domain on which we can lift
a section σ ∈ G(U ). Now consider the set of pairs (W, τ ∈ I(W )) where τ 7→ ρU W (σ), partially ordered
according to
0
(W, τ ) 6 (W 0 , τ 0 ) ⇔ W ⊂ W 0 and ρW 0
W (τ ) = τ.
Since any increasing chain in this poset has a maximal element (by the sheaf axiom (S2)), the poset has
a maximal element (Zorn’s lemma). By the previous argument its domain must be all of U . In other
words, there is a τ ∈ I(U ) such that τ 7→ σ. This proves (6.12.1) and the lemma. (28) 

As a corollary we see that flasque resolutions compute sheaf cohomology. Looking back at the
proof of the existence of injective resolutions we see that the same idea gives the canonical flasque
resolution, usually called the Godement resolution: For a sheaf F let
Y
C 0 (F ) := ix∗ Fx
x∈X

This is flasque and functorial in F . The natural morphism of sheaves F → C 0 (F ) is injective (look at
stalks) and we have the cokernel sheaf coker(F → C 0 (F )). Define
C 1 (F ) := C 0 (coker(F → C 0 (F ))).
The map C 0 (F ) → C 1 (F ) is given by the composition
C 0 (F ) → coker(F → C 0 (F )) ,→ C 0 (coker(F → C 0 (F ))) = C 1 (F ).
Now set
C 2 (F ) := C 0 (coker(C 0 (F ) → C 1 (F )))
and C 1 (F ) → C 2 (F ) is induced by
C 1 (F ) → coker(C 0 (F ) → C 1 (F )) → C 0 (coker(C 0 (F ) → C 1 (F ))) = C 2 (F ).
Proceeding in this way gives a complex C • (F ) which is a resolution of F . Because all the terms in the
resolution come by the C 0 (−) construction they are all flasque. (29)
We thus have a canonical resolution computing the cohomology of any sheaf. Moreover, as is obvious
from the construction, it is functorial in the sheaf (i.e. in F ). From this (and Prop. 6.11.1) we see that
F 7→ Ri Γ(X, F ) = H i (X, F ) is a functor. (30)
28Note that we have really proved that if 0 → F → G → Q → 0 is exact and F and G are flasque then Q = G/F is
flasque.
29In the case where we work with sheaves of vector spaces over a fixed field, a flasque sheaf is injective. So in this case
the Godement resolution is a canonical (and functorial) injective resolution. Thus all derived functors are actually functors.
In general, there is no canonical and functorial injective resolution in an abelian category, and this can be a source of subtle
problems.
30The existence of a canonical and functorial flasque resolution means that we could have defined the cohomology
groups using it, i.e. we could have defined H i (X, F ) by H i (X, F ) := H i (Γ(X, C • (F ))). This would have the advantage
of obviously being a functor and one can directly prove its properties. (This is the approach taken in some books, e.g.
Chapter 3 of Wells, Differential Analysis on Complex Manifolds.)
84

6.13. Singular cohomology and sheaf cohomology. Let X be a topological space. For an abelian
group A we will use C ∗ (−, A) to denote the complex of singular cochains with values in A, i.e. C ∗ (−, A) =
Hom(C∗ (−), A) where C∗ (−) is the singular chain complex. For notational simplicity we restrict to
A = Z, but everything below works the same for general A.
For each i ≥ 0, the assignment
U 7→ C i (U, Z)
defines a presheaf of abelian groups on X. The sheafification is denoted C i . By functoriality of sheafi-
fication the coboundary maps δ : C i (U, Z) → C i+1 (U, Z) give sheaf homomorphisms C i → C i+1 . The
relation δ ◦ δ = 0 means we have a complex of sheaves
0 → C0 → C1 → ···
There is an inclusion ZX ,→ C 0 coming from the presheaf inclusion given by Z ,→ C 0 (U, Z) for any U as
the constant cochain, i.e. n goes to the cochain which takes the constant value n on every chain in U .
Lemma 6.13.1. For a locally contractible space X this gives a resolution of the constant sheaf ZX , i.e.
the sequence
0 → ZX → C 0 → C 1 → · · ·
is an exact sequence of sheaves.

Proof. Recall that the stalks of a presheaf and of the associated sheaf are the same. So it suffices to
prove that each x ∈ X has a sequence of neighbourhoods U for which the sequence
0 → Z → C 0 (U, Z) → C 1 (U, Z) → · · ·
is exact. This follows immediately from the local contractibility of X since this complex computes the
reduced singular cohomology of U . 

The following point-set fact is Proposition 1.14 of Ramanan’s Global Calculus:

Lemma 6.13.2. If X is a paracompact Hausdorff space in which every open set is paracompact (31) and
F is a presheaf on X satisfying (S2), then F (U ) → F # (U ) is surjective for all open U ⊂ X.

Proof. Let U ⊂ X be open. Giving an element s ∈ F # (U ) amounts to giving a cover {Ui } of U and
sections si ∈ F (Ui ) such that for each x ∈ Ui , si 7→ sx in Fx = Fx# . We would like to show that under the
assumptions of the lemma s comes from an element of F (U ). We may assume (since U is paracompact
by assumption on X) that the cover {Ui } is locally finite.
Suppose we could find a covering {Wx } of U by open subsets, indexed by x ∈ U , such that
– for all x, Wx is an open neighbourhood of x with the property that Wx ⊂ Ui for some i
– if Wx ∩ Wy 6= φ then Wx ∪ Wy ⊂ Ui for some i.
Indeed, if so then si |Wx defines a collection of sections of F which can evidently be glued to give an
element σ ∈ F (U ) which has stalks σx = sx for all x ∈ U , hence σ 7→ s ∈ F # (U ). (There is a small
point here: Given x, the Ui containing Wx may not be unique, so that if Wx ⊂ Uj then we do not know
whether to take si |Wx or sj |Wx . However, by shrinking Wx we can ensure this (since si and sj have the
same stalk sx at x), and this does not affect the other properties of {Wx }x∈U .)
To construct the covering Wx we proceed as follows. First choose a refinement {Vi }i∈I of {Ui }i∈I
with the same indexing set and such that Vi ⊂ Ui . (32) Since {Vi } is also locally finite, we can choose
for each x ∈ U a neighbourhood Wx0 such that {i ∈ I : Wx0 ∩ Vi 6= φ} is finite. Now define the following,
which is open since it is a finite intersection of opens:
   
\ \
Wx := Wx0 ∩  U −Vi  ∩  Ui  .
Wx0 ∩Vi 6=φ,x∈V
/ i Wx0 ∩Vi 6=φ

The first intersection makes sure that if Wx ∩ Vi 6= φ then x ∈ Vi , while the second makes sure that if
Wx ∩ Vi 6= φ then Wx ⊂ Ui . Now if Wx ∩ Wy 6= φ, say z ∈ Wx ∩ Wy , then there exists i such that z ∈ Vi ,
whence z ∈ Wx ∩ Vi 6= φ, so that Wx ⊂ Ui . The same argument applies to Wy , so Wx ∪ Wy ⊂ Ui .
31It is a fact, which we do not use here, that if X is a topological space such that every open subset is paracompact,
then in fact every subspace of X is paracompact. Such a space is called hereditarily paracompact.
32This is a general property of paracompact Hausdorff spaces, which can be proved e.g. using the fact that they are
normal, and on normal spaces using Urysohn’s lemma we can find such refinements. In the situation of manifolds we
actually proved this directly in the course of constructing smooth partitions of unity.
 85

We are now in the situation we wanted. The covering {Wx }x∈U allows us to glue the sections
si |Wx ∈ F (Wx ) (which by the remark made earlier we may assume is independent of i) into a section in
F (U ), which maps to s under F (U ) → F # (U ). 

The lemma implies that on such a space X, if a presheaf F satisfying (S2) is flasque (i.e. F (X) 
F (U ) for all U ), then the associated sheaf F # is flasque. This follows immediately from the commutative
diagram
F (X) −−−−→ F # (X)
 
(6.13.1)
 
y y
F (U ) −−−−→ F # (U )
In particular, on such a space X, the sheaves C i are flasque. In particular, this holds for smooth manifolds
(since any open set is again a smooth manifold).
Theorem 6.13.3. If X is a locally contractible and paracompact Hausdorff space in which any open set
is paracompact, then for any abelian group A there are canonical isomorphisms
H ∗ (X, A) ∼
sing = H ∗ (X, AX )
between the singular cohomology groups of X with coefficients in A and the sheaf cohomology groups of
the constant sheaf AX .

Proof. We will assume A = Z, the proof is the same in general. Since ZX → C • is a flasque resolution
we know that
H ∗ (X, ZX ) ∼
= H ∗ (Γ(X, C • )).
Thus it will suffice to show that
C • (X, Z) → Γ(X, C • )
induces an isomorphism in cohomology. Let
C i (X, Z)0 := {σ ∈ Γ(X, C i ) : σ 7→ 0 ∈ Cxi for all x ∈ X}.
We first show that the following sequence is short exact:
0 → C i (X, Z)0 → C i (X, Z) → Γ(X, C i ) → 0
i
The surjectivity on the right comes from the previous lemma. That Q C (X, Z)0 ⊂ ker(C i (X, Z) →
Γ(X, C )) is clear. Equality follows from the fact that Γ(X, C ) ,→ x Cx by the sheaf axiom (S1).
i i i

Thus it is enough to prove that the complex C • (X, Z)0 is acyclic, i.e. has zero cohomology. This
will be deduced from the theorem of small chains. Recall the statement: If X is a topological space,
U = {Uα }α is an open cover of X and C•U (X) denotes the complex of singular chains generated by
singular simplices with image contained in some element of U , then
C•U (X) ,→ C• (X)
is a chain homotopy equivalence. From the universal coefficient theorem we get short-exact sequences
0 −−−−→ Ext1 (Hi−1 (X, Z), Z) −−−−→ H i (X, Z) −−−−→ Hom(Hi (X, Z), Z) −−−−→ 0
  
 π∗ 
y y U y
U
0 −−−−→ Ext1 (Hi−1 i
(X, Z), Z) −−−−→ HU (X, Z) −−−−→ Hom(HiU (X, Z), Z) −−−−→ 0
where the maps are induced by C•U (X) ⊂ C• (X) or its dual. (The upper sequence is the usual one; the
lower sequence comes from exactly the same homological algebra applied to the complexes C•U (X) and
Hom(C•U (X), Z).) Hence we get the dual assertion that the morphism
πU : C • (X, Z) → CU
•
(X, Z)
induces an isomorphism in cohomology and hence ker(πU ) is acyclic. Now ker(πU ) consists precisely
of cochains which map to zero in C • (Uα , Z) for every α, so
C • (X, Z)0 = limU ker(πU )
−→
where the limit is over all coverings of X. Since taking cohomology commutes with direct limits we
conclude that C • (X, Z)0 is acyclic. 
86

Remark 6.13.4. In fact it is true for any locally contractible space that singular cohomology and sheaf
cohomology of the constant sheaf agree, i.e. the assumption that every open in X is paracompact can be
dropped. A correct proof of this is surprisingly recent, see Y. Sella, arXiv:1602.06674v3. Sella also gives
an example where Lemma 6.13.2 fails without some hypothesis, i.e. the sheaves C k are not flasque.
Corollary 6.13.5. If M is a smooth manifold then H ∗ (M, ZM ) ∼ ∗
= Hsing (M, Z).

6.14. Fine sheaves. We would like to relate de Rham cohomology to sheaf cohomology, just as was done
above for singular cohomology. The de Rham sheaves EM ∗
are certainly not flasque. Nevertheless, they
are acyclic for the global sections functor. The key property that ensures this is that E 0 (M ) contains
partitions of unity. Abstracting this we say that a sheaf R of (commutative) rings with unity(33) on
a space X admits partitions of unity if for every locally finite covering {Uα }α of X there exists a
partition of unity {fα } subordinate to it, i.e. elements fα ∈ R(X) with supp(fα ) ⊂ Uα such that
X
fα = 1.
α

(This equation is to be interpreted as making sense locally: In a neighbourhood of any point it is a finite
sum and everything makes sense.) Our main example is the sheaf of smooth functions E 0 on a smooth
manifold, which admits partitions of unity. The sheaf of holomorphic functions on any complex manifold
is an example of a sheaf of rings with unity which does not admit partitions of unity.
A fine sheaf is a sheaf of modules over a sheaf of rings R which admits partitions of unity.(34) (35)
( ) The main example is E k for k ≥ 0. More generally, the sheaf of sections of any smooth vector bundle
36

is a sheaf of E 0 -modules, hence it is fine.

Proposition 6.14.1. A fine sheaf on a paracompact space is Γ-acyclic.

Proof. Let F be the fine sheaf, which by definition is a sheaf of modules over a sheaf of rings R. We
will use the Godement resolution F → C • = C • (F ) (the canonical flasque resolution). Note that the
differentials D : C i → C i+1 are, by construction, R-module homomorphisms. By the Γ-acyclicity of
flasque sheaves we have
ker(Γ(X, C i ) → Γ(X, C i+1 ))
H i (X, F ) = .
im(Γ(X, C i−1 ) → Γ(X, C i ))
Suppose σ ∈ Γ(X, C i ) is a cocycle with i > 0. By exactness of the Godement resolution this means that
there is a covering {Uα } and elements σα ∈ C i−1 such that
σ|Uα = Dσα .
By refining the covering we may assume (since X is paracompact) that it is locally finite. Let {fα } be
a partition of unity for R subordinate to the covering. Let
X
σ̃ := fα σα .
α

Here fα σα is extended by zero outside Uα , so this defines a section in Γ(X, C i−1 ). Then
X X
Dα̃ = D(fα σα ) = fα σ|Uα = σ.
α α
i
This proves that H (X, F ) = 0 for i > 0. 

We have the following important corollary:

Corollary 6.14.2. The de Rham cohomology groups of a smooth manifold M are naturally isomorphic
to the sheaf cohomology of the constant sheaf RM .

Combining this with the identification of singular cohomology with sheaf cohomology we get:

33A sheaf of rings has the obvious meaning: a contravariant functor R : Op(X) → Rings satisfying the sheaf conditions.
We assume that there is a section 1 ∈ R(X) which restricts to the unit in R(U ) for every U .
34A sheaf of modules F over a sheaf of rings R means: Each Γ(U, F ) is a module over Γ(U, R) and the module
multiplication maps Γ(U, R) ⊗ Γ(U, F ) → Γ(U, F ) are compatible with restrictions from U to V ⊂ U .
35
A more natural definition of fine would be the following: F is fine if the sheaf of rings Hom(F, F ) admits partitions
of unity. Everything here works with this definition also.
36
There are several different definitions of fine sheaf in the literature; the one we use here is the easiest for us.
 87

Theorem 6.14.3 (de Rham’s theorem). There is a canonical isomorphism

∼ H ∗ (M, R) = H ∗ (M, Z) ⊗Z R
H ∗ (M ) = dR sing sing

between the de Rham cohomology and the singular cohomology with R-coefficients.

The isomorphism in the theorem is canonical.(37) The isomorphism is also functorial and a ring
∗ ∗
isomorphism when Hsing (M, Z) is given the cup product and HdR (M ) is given the product given by
exterior product of differential forms. These facts will be proved later.

6.15. De Rham’s theorem and integration over smooth chains (∗). We will discuss a refinement
of the de Rham theorem without giving full details, for which we refer to e.g. Warner’s book.
Let M be a smooth manifold. A smooth k-simplex is a smooth map from a neighbourhood of the
standard k-simplex ∆k in Rn to the manifold M . A smooth k-chain is a (finite) formal linear sum with
integer coefficients of smooth k-simplices. The boundary of a smooth k-chain is a smooth k − 1-chain
and we have the complex of smooth chains:
0 → C0sm (M ) → C1sm (M ) → · · ·
Taking Hom(−, Z) we get the complex of smooth cochains:
0 1 2
0 → Csm (M, Z) → Csm (M, Z) → Csm (M, Z) → · · ·
The definition can be sheafified: Let Csm
k k
be the sheaf associated with the presheaf U 7→ Csm (U, Z).
Then we have a complex of sheaves on M
0 → Csm
0
→ Csm
1
→ ···
and an inclusion ZM ,→ Csm 0
as before. To show that this is a resolution of the constant sheaf we must
prove that it is exact on stalks in degrees ≥ 1. This requires some argument which I will not give here,
referring instead to pp. 194-195 of Warner’s book.
It follows from this that Csm
•
is a resolution of the constant sheaf ZM and hence we have an
isomorphism
H k (M, ZM ) ∼ •
= H k (Csm (M, Z)).
As a corollary we also see that the surjection C • (M, Z) → Csm
•
(M, Z) induces an isomorphism in coho-
mology.
Now suppose σ is a smooth k-simplex in U ⊂ M , i.e. a smooth map σ : nbhd(∆k ) → U . If
ω ∈ E k (U ) then σ ∗ (ω) can be integrated over ∆k . There is a minor point here: The k-simplex for k > 1
is not quite a domain in Rn since its boundary is not a manifold at the vertex points. Nevertheless,
we can integrate over it, as you can easily check by going over the arguments defining integration over
domains. Moreover a slight generalization of Stokes’s theorem to this situation says that for a smooth
k-chain σ and smooth k-form ω we have:
Z Z
dω = ω|∂σ .
σ ∂σ
Thus integration defines a homomorphism of complexes of sheaves
E • → Csm
•
.
It is easy to check that this induces the de Rham isomorphism (indeed, this follows by the footnote after
Theorem 6.14.3), implying the following:
Theorem 6.15.1 (Explicit de Rham theorem). Integration over smooth chains defines a nondegenerate
pairing
i
HdR (M ) ⊗ Hising (M, R) → R
which induces the de Rham isomorphism H i (M ) ∼
= H i (M, R).
dR sing

37In general if T is a left-exact functor, given two different T -acyclic resolutions F → A• and F → B • there is a
canonical isomorphism between H ∗ (T (A• )) and H ∗ (T (B • )): Choose an injective resolution F → I • ; by Lemma 6.9.3
there are morphisms of resolutions A• → I • ← B • . By Prop. 6.11.1 these induce canonical isomorphisms H i (T (A• )) →
H i (T (I • )) ← H i (T (B • )). The isomorphism H i (T (A• )) ∼
= H i (T (B • )) is independent of the choice of I • , again by using
Lemma 6.9.3.
Moreover, if A• and B • are acyclic resolutions and A• → B • lifts the identity F → F then it induces the same
isomorphism as the previous one.
88

6.16. Functoriality (∗). The explicit version of de Rham’s theorem shows that the de Rham isomor-
phism is functorial in M , i.e.
Let f : M → N be a smooth map of smooth manifolds. If σ : nbhd(∆k ) → M is a smooth k-simplex
in M and ω ∈ E k (N ) then Z Z
f ∗ω = ω.
σ f ◦σ
k k
This shows that the isomorphism HdR (M ) → Hsing (M, R) induced by the integration pairing is functo-
rial.
(Since we did not give all the details of the explicit de Rham theorem we will give another proof of
the functoriality in the next section.)
 89

7. Sheaves and cohomology II (∗)

In this section we introduce hypercohomology, which generalizes cohomology to complexes of sheaves.

The necessary formalism is a little cumbersome but we will use it to establish that the de Rham isomor-
phism is functorial and one of rings. It is also useful in concrete situations, as we will see later. The
presentation mostly follows Voisin’s book.
In this section we fix an abelian category A, which is always either Sh(X) or AbGps. The word
“object” will usually mean object of A and “homomorphism” will mean homomorphism in A, unless
explicitly stated otherwise. All complexes K • considered below will be in nonnegative degrees, i.e.
K i = 0 for i < 0. (The proofs work for all complexes bounded below, i.e. if there is a r such that K i = 0
for i < r.)

7.1. Cohomology objects, quasi-isomorphisms. Let F • be a complex of objects of A. The ith

cohomology object of F • is:
ker(F i → F i+1 )
H i (F • ) = .
im(F i−1 → F i )
(Here if A = Sh(X) then the image and quotient are taken in the sheaf sense.) A homomorphism of
complexes of sheaves F • → G• is a quasi-isomorphism if the induced homomorphism on cohomology
objects H i (F • ) → H i (G• ) is an isomorphism for all i.
For example on a manifold M the inclusion RM → EM •
is a quasi-isomorphism. More generally,
• •
if F ,→ A is a resolution then F → A is a quasi-isomorphism, where we think of F as the complex
0 → F → 0 with F in degree 0.
A double complex K •• (in A) is a collection of objects (K p,q )(p,q)∈Z2 (in A) with homomorphisms
D2 : K p,q → K p,q+1 (the vertical differential) and D1 : K p,q → K p+1,q (the horizontal differential) such
that
D12 = 0 = D22 and D1 D2 = D2 D1 .
One should keep in mind the picture:
x x
 
 
D
−−−−→ K p,q+1 −−−−
1
→ K p+1,q+1 −−−−→
x x

D2 

D2 

D
−−−−→ K p,q 1
−−−− → K p+1,q −−−−→
x x
 
 

The total complex is the associated single complex T ot(K •,• )• with

T ot(K •,• )i = ⊕p+q=i K p,q

and differential
D|K p,q = D1 + (−1)p D2
(The sign is put so that D2 = 0.)
An example of this construction is the tensor product of complexes: If K • and L• are complexes
then the tensor product complex is defined to be the total complex

T ot(K • ⊗ L• )i := ⊕p+q=i K p ⊗ Lq

with the total differential DK ⊗ id + (−1)p id ⊗ DL . Note that there is a map

H p (K • ) ⊗ H q (L• ) −→ H p+q (T ot(K • ⊗ L• ))

of cohomology objects. We will frequently drop the “Tot” from the notation when we are dealing with
tensor products of complexes; this will not usually cause any confusion.
90

7.2. “Resolutions” of complexes.

Lemma 7.2.1. Let F • be a complex bounded below (say F k = 0 for k < 0). There is a double complex
(I •,• , D1 , D2 ) consisting of injective objects and a monomorphism of complexes F • ,→ I •,0
. . .
: : :
x x x
  
  
I 0,2 −−−−→ I 1,2 −−−−→ I 2,2 −−−−→ · · ·
x x x
  
  
D
I 0,1 −−−−
1
→ I 1,1 −−−−→ I 2,1 −−−−→ · · ·
x x x
D D 
 2  2 
D
I 0,0 −−−−
1
→ I 1,0 −−−−→ I 2,0 −−−−→ · · ·
x x x
  
i0  i1  i2 

d d d
F 0 −−−F−→ F 1 −−−F−→ F 2 −−−F−→ · · ·
such that:
(1) each column is exact (i.e. I k,• is an injective resolution of F k )
(2) the inclusion F • ,→ I •,0 is a homomorphism of complexes.

Proof. It is enough to prove the following: A complex F • admits a monomorphism to a complex of

injectives J • . Indeed, if so then we can construct the first row I •,0 of the double complex with the
inclusion F • ,→ I •,0 . Then we can embed the cokernel coker(F • → I •,0 ) in another complex of injectives
which will be the row I •,1 and proceed from there.
So let us construct the first row. We start with i0 : F 0 ,→ I 0,0 a monomorphism to an injective
object. Next consider
(i0 , −dF ) : F 0 ,→ I 0,0 ⊕ F 1 .
Choose a monomorphism
coker(i0 , −dF ) ,→ I 1,0
to an injective object. Then the homomorphisms D1 and i1 are given by:
I 0,0


y
I 0,0 ⊕ F 1 −−−−→ coker(i0 , −dF ) ,→ I 1,0
x


F1
The identity D1 ◦ i0 = i1 ◦ dF holds because i0 (f ) and dF (f ) are the same in the module coker(i0 , −dF ).
The injectivity of F 1 → I 1,0 is easy.
Now to continue we take
(ī1 , −dF ) : F1 → coker(D1 ) ⊕ F 2
where ī1 is the composition F 1 → I 1,0 → coker(D1 ) and choose a monomorphism
coker(ī1 , −dF ) ,→ I 2,0 .
The maps D1 : I 1,0 → I 2,0 and i2 : F 2 → I 2,0 are given by:
I 1,0


y
I 1,0 ⊕ F 2 −−−−→ coker(ī1 , −dF ) ,→ I 2,0
x


F2
As before we see that D1 ◦ i = i ◦ dF by construction, and i2 is injective. Also D12 = 0 by construction.
1 2
 91

Proceeding in this fashion constructs the first row. 

Lemma 7.2.2. Suppose that I •,• is a double complex of objects in A and F • ,→ I •,0 is a monomorphism
of complexes such that F k ,→ I k,• is a resolution of F k for each k. Then F • ,→ T ot(I •,• )• is a quasi-
isomorphism.

Proof. We give the proof in A = AbGps; the same proof works in general with some rephrasing (one
cannot work with elements).
First
P let us see that H p,q
k
(F • )  H k (T ot(I)• ) for all k. A class in H k (T ot(I)• ) is represented by
α = p+q=k αp,q ∈ ⊕p+q=k I with Dα = 0. Then D2 (α0,k ) = 0 and by exactness of the first column
α0,k = D2 (β) for β ∈ I0,k−1 . Then α0 = α − Dβ represents the same class as α but α0,k 0
= 0. But
0 00
then D2 (αk−1,0 ) = 0 and we can run the same argument to produce α in the same class as α and with
00
α1,k−1 = 0. Proceeding in this fashion we can assume α ∈ I k,0 . Then Dα = 0 implies D2 α = D1 α = 0.
Hence α = ik (f ) for f ∈ F k , and ik+1 (dF (f )) = D1 ik (f ) = D1 α = 0. Then the class of f maps to the
class of α. This proves surjectivity.
Next let us see that H k (F • ) → H k (T ot(I)• ) is injective for all k. If k = 0 then this follows from
the injectivity of i0 . Let k > 0 and let f ∈ F k with dF (f ) = 0 and suppose that ik (f ) = Dβ for some
β ∈ T ot(I)k−1 . Then β ∈ I k−1,0 and hence Dβ = ik (f ) implies that D2 (β) = 0 and D1 β = ik (f ). Since
the columns are exact there is a e ∈ F k−1 with ik−1 (e) = β. Then
ik dF (e) = D1 ik−1 (e) = D1 β = ik (f ).
Since ik is injective we have dF (e) = f . This proves the injectivity. 

Note that the objects in the double complex here were not required to be injective in this lemma.
Combining the two lemmas we get the first part of the following lemma.
Lemma 7.2.3. (1) If F • is a bounded below complex there is a monomorphism F • ,→ I • to a complex
of injective objects which is a quasi-isomorphism.
(2) If F • ,→ I • is as in (1) and F • → J • is any other quasi-isomorphism with J • a complex of
id
F • −−−−→ F•
 
injectives, then there exists a morphism of complexes φ : I • → J • making the diagram 
y

y
φ
I • −−−−→ J •
commute.

Proof. (2) Exercise. 

7.3. Derived functors and hypercohomology. For a left-exact functor T define the ith (hyper)derived
functor on F • as follows: Choose a monomorphic quasi-isomorphism F • ,→ I • to a complex of injectives
and set:
Ri T (F • ) := H i (T (I • )).
In the special case where T is the global sections functor these are called the hypercohomology groups
of F • and are denoted:
Hi (X, F • ) = Ri Γ(X, F • ) = H i (Γ(X, I • )).
As in the case of derived functors these depend, a priori, on the choice of F • ,→ I • . The second part
of the previous lemma can be used to show that they are well-defined up to canonical isomorphism and
that they have the same functorial property as derived functors.
Thus if F • consists of a single sheaf F placed in degree zero, then
Ri T (F • ) = Ri T (F ) and Hi (X, F • ) = H i (X, F ).
If I • is a complex of injectives then
Ri T (I • ) = H i (T (I • )) and Hi (X, I • ) = H i (Γ(X, I • )).
In particular, an exact complex of injectives has zero higher (i > 0) derived functors.
The proof of the following proposition will be given later:
92

Proposition 7.3.1. If F • → J • is a quasi-isomorphism (but not necessarily a monomorphism) with a

complex of injective objects and T is a left-exact functor then it induces canonical isomorphisms
Ri T (F • ) → Ri T (J • ) = H i (T (J • ))
for all i.

We have the following immediate corollary:

Corollary 7.3.2. If F • → K • is a quasi-isomorphism and T is a left-exact functor then it induces
canonical isomorphisms Ri T (F • ) → Ri T (K • ) for all i.

Proof. Suppose φ : K • → L• is a quasi-isomorphism of complexes. Choose a monomorphism i : L• ,→ I •

which is a quasi-isomorphism to a complex of injectives. Then i ◦ φ : K • → I • is a quasi-isomorphism to
a complex of injectives. By Prop. 7.3.1 there are canonical isomorphisms
Ri T (K • ) = H i (T (I • )) = Ri T (L• ).
This proves the corollary. 

The proof of the following proposition (which generalizes Prop. 6.11.1) will be given later:
Proposition 7.3.3. If T is a left-exact functor and F • → A• is a quasi-isomorphism to a complex of
T -acyclic objects then it induces canonical isomorphisms
Ri T (F • ) → Ri T (A• ) = H i (T (A• ))
for all i.

For example, applying the corollary to the quasi-isomorphism RM → EM

•
on a smooth manifold we
get:
H i (M, RM ) ∼
= Hi (M, EM •
) = H i (Γ(M, EM•
))
where we have also used the second proposition. This gives the isomorphism between de Rham coho-
mology and sheaf cohomology of RM .
Suppose F • is a complex of sheaves. The Godement resolution is functorial hence we get a double
complex C • (F • ) in which each column C • (F k ) is a resolution of F k . By Lemma 7.2.2 we conclude that
F • ,→ T ot(C • (F • ))• is a quasi-isomorphism. Since the terms in the Godement resolutions are flasque
we get a canonical isomorphism:
Hi (X, F • ) = Hi (X, T ot(C • (F • ))• ).
As in the case of sheaf cohomology, we see that the functoriality of Godement resolutions shows that
hypercohomology is actually a functor from complexes of sheaves to abelian groups.

7.4. Pullback maps in (hyper)cohomology. First let f : Y → X be a continuous map of topological

spaces and F a sheaf on X. Then there is an obvious map (38)
Γ(X, F ) → Γ(Y, f −1 F ).
For each i ≥ 0 there is a canonical pullback map
H i (X, F ) → H i (Y, f −1 F ) (7.4.1)
which is the previous map when i = 0. This is defined as follows: Choose injective resolutions F ,→ I • and
f −1 J • . Then f −1 I • is a resolution of f −1 F (because f −1 is exact) and hence there is a homomorphism
of complexes f −1 I • → J • , unique up to homotopy. Applying Γ we have homomorphisms of complexes
Γ(X, I • ) → Γ(Y, f −1 I • ) → Γ(Y, J • )
which induce the desired homomorphisms (7.4.1) in cohomology. It is straightforward to check that these
are independent of the choices made.
Now let F • be a complex of sheaves on X. Then there is a canonical pullback map
Hi (X, F • ) → Hi (Y, f −1 F • ) (7.4.2)
• • −1 • •
defined as follows: Choose injective resolutions F ,→ I and f F → J as in (1) of Lemma 7.2.3.
Since f −1 is exact f −1 F • ,→ f −1 I • is also a quasi-isomorphism (39), so by (2) of the same lemma there
38Indeed, a section of f −1 F over Y consists of a covering of Y and for each U in the covering an element of
limV Γ(V, F ). An element of Γ(X, F ) gives such a section trivially.
−→ open,V ⊃f (U )
39An exact functor takes quasi-isomorphisms to quasi-isomorphisms (see Lemma 7.10.2 below).
 93

exists a homomorphism f −1 I • → J • , unique up to homotopy. Thus the maps on global sections give
maps
Γ(X, I • ) → Γ(Y, f −1 I • ) → Γ(Y, J • )
and hence a canonical map

Hi (X, F • ) = H i (Γ(X, I • )) → H i (Γ(Y, J • )) = Hi (Y, f −1 F • ).

It is easy to check that this pullback map is independent of choices and is natural in homomorphisms of
complexes of sheaves F • → G• . It obviously reduces to (7.4.1) in case F • and G• are single sheaves in
degree zero.
In some cases this map is a familiar one: For example, if f : Y → X is a continuous map of
topological spaces then f −1 ZX ∼
= ZY . Thus there is a pullback map
H i (X, ZX ) → H i (Y, ZY ).

If X and Y are locally contractible then under the isomorphisms H i (−, ZX ) ∼ i

= Hsing (−, Z) this is the
∗
usual pullback map f in singular cohomology. (Exercise.)

7.5. Functoriality of the de Rham isomorphism. Let us see how this follows nicely using hyper-
cohomology. In the case f : M → N is a map of manifolds we shall see that the pullback map in either
singular or de Rham cohomology is given by the map

H i (N, RN ) → H i (N, f −1 RN ) ∼
= H i (M, RM )
i
under the respective isomorphisms Hsing (?, R) ∼
= H i (?, R? ) and HdR
i
(?) ∼
= H i (?, R? ). This will prove the
functoriality of the de Rham isomorphism.
First consider de Rham cohomology. We have a diagram
f −1 RN −−−−→ RM
 
 
y y
f −1 EN• −−−−→ EM
•

The lower map is given by pullback of differential forms from (open sets of) N to (open sets of) M . Both
vertical maps are quasi-isomorphisms (the right by the Poincaré lemma on M , the left by the Poincaré
lemma on N plus exactness of f −1 ). The upper map is an isomorphism. The commutativity of the
diagram is obvious. Taking hypercohomology gives the right square of:
Hk (N, RN ) −−−−→ Hk (M, f −1 RN ) Hk (M, RM )
  
  
y y y
∼
=
Hk (N, EN• ) −−−−→ Hk (N, f −1 EN• ) −−−−→ Hk (M, EM
•
)
The outer vertical maps give the isomorphism between sheaf cohomology and de Rham cohomology on
N and M respectively. All the maps in the right square are isomorphisms. The horizontal maps in
the left square are pullback maps. Both squares commute, thus the outer isomorphisms are compatible.
Checking that the composition across the bottom row is the same as the map induced by pullback of
differential forms from N to M is left as an exercise.
The same argument using the complex of sheaves of continuous cochains proves that the pullback

H i (N, CN• ) = Hi (N, CN• ) → Hi (M, f −1 CN• ) ∼

= Hi (M, CM
•
) = H i (M, CM
•
)

is the sheaf pullback. Now this map is clearly compatible with the usual pullback
i i
Hsing (N, Z) → Hsing (M, Z)

Thus we have:
∗
Proposition 7.5.1. The de Rham isomorphism HdR (M ) ∼ ∗
= Hsing (M, R) is functorial for smooth maps
of manifolds.
94

7.6. Cup products in sheaf cohomology. To show that the de Rham isomorphism HdR ∗
(M ) ∼=
∗ ∗
Hsing (M, R) is compatible with products (i.e. relates the product in Hsing (M, R) induced by cup product
∗
of cochains and the product in HdR (M ) induced by wedge product of differential forms) we will define
a suitable cup product in sheaf cohomology H ∗ (M, RM ), and then show that it agrees with each of
the de Rham and singular cup products under the respective isomorphisms HdR ∗
(M ) ∼
= H ∗ (M, RM ) and
∗ ∼ ∗
Hsing (M, R) = H (M, RM ). For this we need to understand how cup products work in hypercohomology.
For simplicity the case of cohomology of a single sheaf is discussed first.
Before starting with the discussion, recall (for motivation) one way to define cup products in singular
cohomology. (40) This is via the cross product (or external cup product), which is a map
× : H i (X, Z) ⊗ H j (Y, Z) → H i+j (X × Y, Z). (7.6.1)
When X = Y we can compose this with pullback by the diagonal δ : X ,→ X × X to define the cup
product
`: H i (X, Z) ⊗ H j (X, Z) → H i+j (X, Z) by α ` β = δ ∗ (α × β). (7.6.2)

Now let us define cup products on sheaf cohomology. For sheaves F, G ∈ Sh(X) we will define a
(cup) product map
H p (X, F ) ⊗ H q (X, G) → H p+q (X, F ⊗ G). (7.6.3)
Let πi : X × X → X be the projections. The map (7.6.3) is the composition
H p (X, F ) ⊗ H q (X, G)

(1)
y
H p+q (X × X, π1−1 F ⊗ π2−1 G)

(2)
y
H p+q (X × X, δ∗ (F ⊗ G))

H p+q (X, F ⊗ G)
where the maps are as follows:
(1) Let π1 , π2 be the projections of X × Y to each factor. Then there is a homomorphism
H p (X, F ) ⊗ H q (Y, G) → H p+q (X × Y, π1−1 F ⊗ π2−1 G). (7.6.4)
(This is the analogue of the cross product in singular cohomology.) To get this let F ,→ I • and
G ,→ J • be the Godement resolutions of F and G. Then π1−1 F ⊗ π2−1 G ,→ π1−1 I • ⊗ π2−1 J • is a
resolution on X × Y . (41) Thus there is a map (42)
H i (Γ(X × Y, π1−1 I • ⊗ π2−1 J • )) −→ H i (X × Y, π1−1 F ⊗ π2−1 G).
Now using the maps of complexes
π1−1 ⊗π2−1
Γ(X, I • ) ⊗ Γ(Y, J • ) −→ Γ(X × Y, π1−1 I • ) ⊗ Γ(X × Y, π2−1 J • ) −→
−→ Γ(X × Y, π1−1 I • ⊗ π2−1 J • )
and composing with the previous map we get (7.6.4).
(2) comes from the canonical homomorphism of sheaves on X × X
π1−1 F ⊗ π2−1 G −→ δ∗ (F ⊗ G)
where δ : X ,→ X × X is the diagonal.
(3) Pushforward by δ is an exact functor and hence H i (X × X, δ∗ K) = H i (X, K) (cf. Lemma
6.10.3).

40This is the approach used in §3.B of Hatcher’s book.

41To prove this look at stalks: The stalk of π −1 A ⊗ π −1 B at (x, y) is A ⊗ B .
1 2 x y
42
We are dropping the “T ot” notation in tensor products for this subsection.
 95

In the special case where R is a sheaf of rings, so that there is a map R ⊗ R → R, (7.6.3) can be
composed with H i (X, R ⊗ R) → H i (X, R) to give products
H i (X, R) ⊗ H j (X, R) −→ H i+j (X, R). (7.6.5)
In particular, if R is any coefficient ring and RX the constant sheaf then we have defined a product
H i (X, RX ) ⊗ H j (X, RX ) −→ H i+j (X, RX ). (7.6.6)
In this context we could have replaced ⊗ = ⊗Z everywhere above by ⊗R and got a product H i (X, RX )⊗R
H j (X, RX ) −→ H i+j (X, RX ).
On the other hand, if R is a sheaf of rings we may be able to resolve it by a sheaf of R-algebras
which are Γ-acyclic. This is the case, for example, if R = RM is the constant sheaf on M and we take
the de Rham resolution RM ,→ EM •
. In that case there is another, potentially different, product on
H (X, R). Again we abstract the situation a bit to make it clearer.
∗

Suppose we are given a commutative ring R (with unity) and two sheaves of R-modules F and G.
Suppose we are given Γ-acyclic resolutions F ,→ A• , G ,→ B • and F ⊗R G ,→ C • and a homomorphism
of complexes of sheaves
µ : A• ⊗R B • −→ C •
such that the diagram
F ⊗R G F ⊗R G
 
 
y y
µ
A• ⊗R B • −−−−→ C •.
commutes. Then we have a product
H p (X, F ) ⊗R H q (X, G) −→ H p+q (X, F ⊗R G) (7.6.7)
induced as follows: Taking global sections gives
Γ(X, A• ) ⊗R Γ(X, B • ) −→ Γ(X, A• ⊗R B • ) −→ Γ(X, C • )
and thus in cohomology we get (7.6.7) (using acyclicity of A• , B • , C • ).
Proposition 7.6.1. The two products (7.6.5) and (7.6.7) are the same, i.e. under the isomorphisms
H p (X, F ) = H p (Γ(X, A• )) and H p (X, G) = H p (Γ(X, B • )) (coming from Prop. 6.11.1) the two products
are the same.

This will follow from a more general statement about hypercohomology proved below.
Let us apply this proposition to singular cohomology. First consider the case of a locally contractible
space and the sheaf ZX on it. The Γ-acyclic resolution ZX ,→ CX• given by the (sheafified) singular
cochain complexes has a natural product
CX• ⊗Z CX• −→ CX•
which comes from the presheaf homomorphisms
C p (U, Z) ⊗ C q (U, Z) −→ C p+q (U, Z) α ⊗ β 7→ α ` β
i.e. cup product of cochains. This lifts ZX ⊗Z ZX → ZX (multiplication). By Prop. 7.6.1 the induced
map in cohomology (7.6.7) is the same as the sheaf-theoretically defined cup product (7.6.5) (which
comes from (7.6.3)).
Next let us apply the proposition to de Rham cohomology. We have a morphism of Γ-acyclic
resolutions
EM
•
⊗R E M
•
−→ EM
•

given by wedge product of differential forms. This lifts the map RM ⊗R RM → RM . Applying Prop.
7.6.1 we see that the induced map in cohomology (7.6.7) is the same as the sheaf-theoretically defined
cup product (7.6.5) (which comes from (7.6.3)).
Thus we have identified both products under the respective isomorphisms with sheaf cohomology.
This proves that the de Rham isomorphism is an isomorphism of rings.
96

7.7. Cup products in hypercohomology. To prove Prop. 7.6.1 it is easier to prove a more general
statement about hypercohomology. For this we first generalize the sheaf-theoretic definition of cup
products to this context.
For complexes of sheaves F • and G• we will define a product map (43)
Hp (X, F • ) ⊗ Hq (X, G• ) → Hp+q (X, F • ⊗ G• ). (7.7.1)
Let πi : X × X → X be the projections. The map (7.7.1) is the composition
Hp (X, F • ) ⊗ Hq (X, G• )

(1)
y
Hp+q (X × X, π1−1 F • ⊗ π2−1 G• )

(2)
y
Hp+q (X × X, δ∗ (F • ⊗ G• ))

Hp+q (X, F • ⊗ G• )
where the maps are as follows:
(1) Let π1 , π2 be the projections of X × Y to each factor. Then there is a homomorphism
Hp (X, F • ) ⊗ Hq (Y, G• ) → Hp+q (X × Y, π1−1 F • ⊗ π2−1 G• ). (7.7.2)
• •
To get this let F ,→ I and G ,→ J be the Godement resolutions of F and G.
Lemma 7.7.1. Then π1−1 F ⊗ π2−1 G ,→ π1−1 I • ⊗ π2−1 J • is a resolution on X × Y .
Thus there is a map
H i (Γ(X × Y, π1−1 I • ⊗ π2−1 J • )) −→ H i (X × Y, π1−1 F ⊗ π2−1 G).
Now using the map of complexes
Γ(X, I • ) ⊗ Γ(Y, J • ) −→ Γ(X × Y, π1−1 I • ) ⊗ Γ(Y, π2−1 J • ) −→ Γ(X × Y, π1−1 I • ⊗ π2−1 J • )
and composing with the previous map we get (7.7.2).
(2) comes from the canonical homomorphism
π1−1 F • ⊗ π2−1 G• −→ δ∗ (F • ⊗ G• )
where δ : X ,→ X × X is the diagonal.
(3) Pushforward by δ is an exact functor and hence Hi (X × X, δ∗ K • ) = Hi (X, K • ). (44)
This defines the cup product on hypercohomology.
Generalizing the second situation above, suppose that we have complexes F • and G• and quasi-
isomorphisms to complexes of Γ-acyclic objects F • → A• , G• → B • and F • ⊗ G• → C • and assume that
there is a map
µ : A• ⊗ B • → C •
such that the following diagram commutes:
F • ⊗ G• F • ⊗ G•
 
 
y y
µ
A• ⊗R B • −−−−→ C •.
Applying hypercohomology to this diagram gives a map

Hk (X, F • ) ⊗ Hk (X, G• ) −−−−→ Hk (X, F • ⊗ G• ) Hk (X, F • ⊗ G• )

  
  
y y y
Hk (X, A• ) ⊗ Hk (X, B • ) −−−−→ Hk (X, A• ⊗ B • ) −−−−→ Hk (X, C • )
THIS NEEDS TO BE FINISHED
43We are dropping the “T ot” notation in tensor products for this subsection, so F • ⊗ G• is actually the total complex
of the double complex (F p ⊗ Gq ).
44This follows from Corollary 7.3.2 above and Lemma 7.10.2 below.
 97

7.8. Cup products and the de Rham isomorphism. The following was proved above:
Proposition 7.8.1. The de Rham isomorphism HdR ∗
(M ) ∼ ∗
= Hsing (M, R) is an isomorphism of rings.
R
Corollary 7.8.2. If M is a compact oriented n-manifold the pairing α ⊗ β 7→ M α ∧ β defines a
k n−k
nondegenerate pairing HdR (M ) ⊗ HdR (M ) → R.
7.9. Composition of functors. Let
T U
A −→ B −→ C
be left-exact functors between abelian categories. Assume any one of three conditions:
(1) T carries injective objects to injective objects
(2) T carries injective objects to T -acyclic objects
(3) T carries U ◦ T -acyclic objects to T -acyclic objects.
Then for a resolution F → A• by either injectives (if (1) or (2) hold) or U ◦ T -acyclics (if (3) holds) we
know that
Rk (U ◦ T )(F ) = H k (U ◦ T (A• )).
•
On the other hand, T (A ) is a complex of U -acyclic objects and hence the derived functors Rk U can be
id
computed on T (A• ) using the ”resolution” T (A• ) → T (A• ), i.e.:
Rk U (T (A• )) = H k (U (T (A• )).
Thus we have that under (1) or (2) or (3) there is a natural isomorphism: (45)
Rk (U ◦ T )(F ) = Rk U (T (A• )).
We will use this observation (rather, a similar argument) in the following situation: f : Y → X is a
continuous map and we consider the composition Γ(Y, −) = Γ(X, −) ◦ f∗ :
f∗ Γ(X,−)
Sh(Y ) −→ Sh(X) −→ AbGps.
For F ∈ Sh(Y ) and A• a flasque resolution of F , since f∗ takes flasque sheaves to flasque sheaves and
flasques are acyclic, we have:
H k (Y, F ) = Hk (X, f∗ A• ).
Thus the cohomology of any sheaf on Y (or any complex of sheaves on Y ) can be computed by a complex
of sheaves on X, namely the direct image of a flasque or injective resolution. For the latter we may, for
instance, take the Godement resolution.

7.10. Proofs. We now go back and prove various unproven statements. We will need some extra nota-
tion.
If A• is a complex then the shifted complex A• [1] is defined by
i−1
A[1]i := Ai−1 and diA[1] = (−1)i dA .
It has the property that H i (A[1]• ) = H i−1 (A• ). The functor A 7→ A[1] from complexes to complexes is
called the shift functor.
If φ : A• → B • is a morphism of complexes then the cone of φ is the complex
 i
dA (−1)i φi

i i i−1
C (φ) := A ⊕ B with dC := .
di−1
B
There is a split short-exact sequence of complexes
0 → B • [1] → C(φ)• → A• → 0.
We will also need the following facts:
Lemma 7.10.1. (1) If I • is an exact complex of injective objects and T is a left-exact functor then
T (I • ) is exact.
(2) If φ : K • → L• is a quasi-isomorphism then C(φ)• is acyclic, i.e. H ∗ (C(φ)• ) = 0, and
conversely.
(3) If A• is an exact complex of T -acyclic objects for a left-exact functor T then T (A• ) is an exact
complex.
45A special case of this was seen earlier: If T is actually exact and satisfies (1) then any object is acyclic for T and
hence taking A• = F (in degree zero) we get Rk (U ◦ T )(F ) = Rk U (T (F )). This was proved in Lemma 6.10.3 (see also the
footnote there).
98

Proof. (1) If I • is an exact complex of injective objects then there is a homotopy between the identity
map of I • and the zero map, constructed as follows: Look at the diagram
d0 d1
0 −−−−→ I 0 −−−−→ I 1 −−−−→ I 2 −−−−→ · · ·
  
  
y y y
0 −−−−→ I 0 −−−−→ I 1 −−−−→ I 2 −−−−→ · · ·
Since I 0 → I 1 is a monomorphism and I 0 is injective there is a map H 1 : I 1 → I 0 extending the identity
of I 0 , i.e. with H 1 d0 = idI 0 . Now ker(H 1 ) is a complement to d0 (I 0 ), i.e. I 1 → coker(d0 ) gives an
isomorphism ker(H 1 ) ∼ = coker(d0 ). By the diagram
d1
0 −−−−→ coker(d0 ) = ker(H 1 ) −−−−→ I 2


y
I1
0 2 2 1
and injectivity of coker(d ) we get H : I → I satisfying
H 2 ◦ d1 = IdI 1 − d0 ◦ H 1 .
Continuing in this way constructs a homotopy H k : I k → I k−1 such that dk−1 H k + H k+1 dk = idk .
Now applying T to the homotopy gives a homotopy of the identity of T (I • ) with the zero map.
Thus T (I • ) is exact.
(2) The split short exact sequence 0 → B[1]• → C(φ)• → A• → 0 gives the long exact sequence
· · · −→ H i (B[1]• ) −→ H i (C(φ)• ) −→ H i (A• ) −→ H i+1 (B[1]• ) −→ · · ·
The connecting homomorphism H i (A• ) → H i+1 (B[1]• ) = H i (B • ) is easily checked to be the map
induced by φ. Since this is an isomorphism for all i (because φ is a quasi-isomorphism) we have that
H i (C(φ)• ) = 0 for all i, i.e. the cone is acyclic. The converse also clearly holds.
(3) Suppose that
d0 d1
0 −−−−→ A0 −−−−→ A1 −−−−→ A2 −−−−→ · · ·
is an exact complex of T -acyclics. The sequences
0 −−−−→ im(di−1 ) = ker(di ) −−−−→ Ai −−−−→ ker(di+1 ) −−−−→ 0
are then short-exact for all i. Induction and the long exact sequence for Ri T (−) shows that ker(di ) is
T -acyclic for all i. Thus applying T gives short-exact sequences
0 −−−−→ T (ker(di )) −−−−→ T (Ai ) −−−−→ T (ker(di+1 )) −−−−→ 0
for all i. It follows that T (ker(di )) = ker(T (di+1 )) for all i and putting these short exact sequences gives
exactness of the complex T (A• ). 

The following statements about exact functors were used above:

Lemma 7.10.2. (1) If K • is an exact complex and T is an exact functor then T (K • ) is an exact
complex.
(2) If K • → L• is a quasi-isomorphism and T is an exact functor then T (K • ) → T (L• ) is a
quasi-isomorphism.

Proof. (1) Exercise. (Break the complex up into short exact sequences and apply T , similar to the proof
of part (3) of the previous lemma.)
(2) By (2) of the previous lemma a morphism φ : K • → L• is a quasi-isomorphism if and only if the
cone C(φ)• is exact, i.e. H ∗ (C(φ)• ) = 0. By (1), T (C(φ)• ) = C(T (φ))• is exact, and hence (by (2) of
the previous lemma) T (φ) is a quasi-isomorphism. 

Proof. (of Proposition 7.3.1) Suppose that j : F • → J • is a quasi-isomorphism to a complex of injective

objects and i : F • → I • is a monomorphism which is a quasi-isomorphism to a complex of injective
objects. By (2) of Lemma 7.2.3 there is a map φ : I • → J • such that φ ◦ i = j. Then φ is also a
quasi-isomorphism. By (2) of the previous lemma the cone C(φ)• is acyclic. It also consists of injective
objects. By (1) of that lemma T (C(φ)• ) is again an exact complex. Since the short exact sequence
0 −−−−→ J[1]• −−−−→ C(φ)• −−−−→ I • −−−−→ 0
 99

is split exact, so is the sequence

0 −−−−→ T (J[1]• ) −−−−→ T (C(φ)• ) −−−−→ T (I • ) −−−−→ 0.
Using the long exact cohomology sequence of this sequence and the fact that T (C(φ)• ) is exact we get
isomorphisms
H i (T (I • )) ∼
= H i+1 (T (J[1]• )) = H i (T (J • )).
Since Ri T (I • ) = H i (T (I • )) (by definition), this proves the proposition. 

Proof. (of Proposition 7.3.3). Let F • → A• be a quasi-isomorphism to a complex of acyclic objects for
the left-exact functor T . By Corollary 7.3.2 we know that
Ri T (F • ) ∼
= Ri T (A• )
for all i. Thus we must show that for a complex of acyclic objects we have Ri T (A• ) = H i (T (A• )). Let
A• ,→ I • be a monomorphic quasi-isomorphism to a complex of injective objects. Let Q• = I • /A• . The
sequence
0 −−−−→ A• −−−−→ I • −−−−→ Q• −−−−→ 0
shows (via the long-exact sequence in cohomology) that Q• is an exact complex. Applying T gives
(because each Ai is acyclic) a short-exact sequence of complexes
0 −−−−→ T (A• ) −−−−→ T (I • ) −−−−→ T (Q• ) −−−−→ 0.
Each Qi is T -acyclic (by the long-exact sequence for Ri T (−) because Ai and I i are T -acyclic). Taking
cohomology and applying (4) of the previous lemma to T (Q• ) gives
H i (T (A• )) = H i (T (I • )) = Ri T (F • )
proving the proposition. 
100

8. Fibre bundles: Leray-Hirsch theorem, Thom isomorphism

In this section we look at the cohomology of locally trivial fibre bundles. These are usually studied
using spectral sequences, but (following Voisin’s book) we will avoid these for some types of fibre bundles
using a sheaf-theoretic argument with hypercohomology. The main result is the Leray-Hirsch theorem,
which we then apply in some examples and use to establish the Thom isomorphism.

8.1. Pushforward. We will need the following lemma, which describes the pushforward of a complex:
Lemma 8.1.1. Let f : Y → X be a continuous map and K • a complex of sheaves on Y . Then Ri f∗ K •
is the sheaf on X associated with the presheaf U 7→ Hi (f −1 (U ), K • ).

Proof. Let K • ,→ I • be a quasi-isomorphism to a complex of injective sheaves. By definition,

Ri f∗ K • = H i (f∗ I • ).
But H i (f∗ I • ) is the sheaf associated with the presheaf
U 7→ H i (Γ(U, f∗ I • )) = H i (Γ(f −1 (U ), I • )) = Hi (f −1 (U ), K • ).
This proves the lemma. 

8.2. Leray-Hirsch theorem. We will now use the formalism above to prove a useful theorem about
the homology of certain manifolds.
A map f : Y → X is locally trivial if for each x ∈ X there is a neighbourhood Ux of x in X and a
homeomorphism f −1 (Ux ) ∼ = Ux × f −1 (x) such that f |f −1 (Ux ) is given by the first projection. An example
of a locally trivial map is a fibre bundle: A map π : E → B is a fibre bundle if there exists a space
F (the fibre) and a covering {Uα } of B and homeomorphisms π −1 (Uα ) ∼ = Uα × F such that π|π−1 (Uα ) is
given by the first projection. An example of a fibre bundle is a vector bundle.
Theorem 8.2.1 (Leray-Hirsch). Suppose that X and Y are locally contractible (46) and f : Y → X is
a locally trivial map. Assume that
(1) for each x ∈ X the cohomology groups H ∗ (f −1 (x), Z) are torsion-free
(2) there exist classes α1 , . . . , αN ∈ H ∗ (Y, Z) such that for all x ∈ X the restrictions αi |f −1 (x) form
a basis of H ∗ (f −1 (x), Z).
Let A∗ be the graded subgroup of H ∗ (Y, Z) generated by α1 , . . . , αN . (Thus A∗ ∼ = H ∗ (f −1 (x), Z) for any
x). Then the map
X X
A∗ ⊗ H ∗ (X, Z) → H ∗ (Y, Z) by ci αi ⊗ ω 7→ ci αi ` f ∗ (ω)
i i
is an isomorphism of graded groups. In particular, H (Y, Z) is freely generated as a graded H ∗ (X, Z)-
∗

module by A∗ .

Proof. Let I • be a flasque resolution of the constant sheaf ZY . We know that H ∗ (Y, Z) = H ∗ (Y, ZY ) =
H ∗ (Γ(Y, I • )). Let α̃1 , . . . , α̃N ∈ Γ(Y, I • ) be cocycles representing the classes α1 , . . . , αN . We may assume
that the αi are all homogeneous of a single degree, and that the same holds for the α̃i . Then they generate
a subgroup of Γ(Y, I • ) isomorphic to A∗ and they give an inclusion
A∗X ,→ f∗ I •
which is an inclusion of complexes of sheaves. (Here AkX is the constant sheaf on X and the complex A∗X
has zero differentials.) The local triviality of f and assumption (2) in the theorem imply that this is a
quasi-isomorphism (use Lemma 8.1.1). Thus we have isomorphisms in hypercohomology
Hi (X, A∗ ) ∼
= Hi (X, f∗ I • )
X
for all i. The right-hand side is H ∗ (Y, Z) by the lemma above, while the right-hand side is
Hi (X, A∗X ) = ⊕k Hi−k (X, ZX ) ⊗ Ak = ⊕k H i−k (X, Z) ⊗ Ak
by the assumption (1) on A∗ . Thus we have H ∗ (X, Z) ⊗ A∗ =∼ H ∗ (Y, Z). 
Corollary 8.2.2 (Künneth formula). Let X and Y be locally contractible spaces and assume that
H ∗ (X, Z) is torsion-free. Then H ∗ (X × Y, Z) ∼
= H ∗ (X, Z) ⊗ H ∗ (Y, Z) (as graded Z-modules).
46The proof will use the identification of singular cohomology with sheaf cohomology for X and Y . Since this was only
proved in these notes for locally contractible spaces under the additional hypothesis that every open subset is paracompact,
we should probably add this assumption here. In any case, in the applications all spaces will be manifolds.
 101

Remarks 8.2.3. (1) The proof given here works with any ring R as coefficients.
(2) Note in particular that the map f ∗ : H ∗ (X, Z) → H ∗ (Y, Z) is an injective ring homomorphism
under the conditions of the theorem. Moreover, the image is a direct summand if the fibres are connected:
a complement is given by choosing a basis of A>0 .
(3) The Leray-Hirsch theorem does not claim that the ring structure on H ∗ (Y, Z) is the tensor
product, in fact that is not true in general. (We will see an example below.) In the product case, it is
true that the ring structure is given by the tensor product, but the proof here does not prove this.
(4) If X and Y are manifolds and the fibres are manifolds then everything holds in de Rham
cohomology. (In fact, a proper surjective smooth submersion of smooth manifolds with connected fibres
always satisfies the local triviality condition in the Leray-Hirsch theorem, by Ehresmann’s theorem.)
(5) An easy example to see that condition (2) in the theorem is not always satisfied is the Hopf
fibration S 3 → S 2 , which is locally trivial with fibre S 1 .

8.3. Example: Homogeneous manifolds for compact Lie groups. Let G be a compact connected
Lie group. The de Rham cohomology of a compact Lie group is computed by the bi-invariant differential
forms, i.e.
∗
HdR (G) = (Λ∗ Lie(G)∗ )AdG ,
where Ad : G → GL(Lie(G)) is the adjoint representation. (This is an exercise on the problem set.) If
H is a closed connected (Lie) subgroup then Lie(G)∗  Lie(H)∗ but it need not be the case that
Λk (Lie(G)∗ )Ad(G)  Λk (Lie(H)∗ )Ad(H) for all k. (8.3.1)
Now consider the fibration π : G → G/H. It is locally trivial since the smooth manifold structure on
G/H is such that π has local smooth sections. The fibre over eH is identified with H. If (8.3.1) holds
∗ ∗
then one can choose classes α1 , . . . , αN ∈ HdR (G) which map isomorphically to a basis of HdR (H).
−1
Since the fibre π (gH) = gH = Lg (H) and the classes α1 , . . . , αN may be represented by left-invariant
∗
differential forms, we see that α1 , . . . , αN restrict to a basis of HdR (π −1 (gH)) for any fibre. Thus the
Leray-Hirsch theorem applies and tells us that
H ∗ (H, R) ⊗ H ∗ (G/H, R) ∼ = H ∗ (G, R). (8.3.2)
So if (8.3.1) holds then we can conclude (8.3.2).
(An example where (8.3.1) fails is G = SU (2) ⊃ H = U (1), with k = 1. The corresponding map
S 3 = SU (2) → SU (2)/U (1) = S 2 is the Hopf fibration.)

8.4. Example: Unitary groups. Assume that (8.3.1) holds for the pair U (n − 1) ⊂ U (n) for all n ≥ 1.
Then applying the isomorphism (8.3.2) to the fibration
U (n) → U (n)/U (n − 1) = S 2n−1
and induction on n shows that
H ∗ (U (n), R) = ΛR [c1 , c3 , . . . , c2n−1 ]
is the free exterior algebra on generators in each odd degree ≤ 2n − 1. Note that the ring structure is
also determined since all products are zero by graded-commutativity of the cup product. For SU (n) the
same argument gives
H ∗ (SU (n), R) = ΛR [c3 , . . . , c2n−1 ]
(since SU (1) = {e} the induction begins with SU (2) = S 3 so the first generator is in degree 3). The
same assertions hold with Q coefficients.
We will verify (8.3.1) for the pair U (n − 1) ⊂ U (n) for n ≥ 3 later.
Remarks 8.4.1. (1) These results illustrate Hopf’s theorem, viz. the cohomology of a compact connected
Lie group is an exterior algebra on odd-degree generators.
(2) The same method can be used for the sequence of fibrations involving the compact symplectic
group: Sp(1) ⊂ Sp(2) ⊂ · · · .
(3) The method does not work for SO(n). For example, the fibration SO(3) → S 2 = SO(3)/SO(2)
is not homologically like a product, i.e. the Leray-Hirsch theorem does not apply. (This has to be the
case because according to the Hopf theorem H ∗ (SO(n), R) is also an exterior algebra on odd degree
generators. Computing the actual degrees of the exponents is not as easy as in the SU (n), Sp(n) cases;
the answer is different for even n and odd n.)
102

8.5. Example: Flag manifolds. A k-flag in Cn is a chain V1 ⊂ V2 ⊂ · · · ⊂ Vk of subspaces of Cn with

dim(Vi ) = i. The set F l(k, n) of k-flags in Cn is a homogenous manifold: The Lie group GL(n, C) acts
transitively on it and the isotropy is a certain closed (hence Lie) subgroup. This implies that it has a
manifold structure.
It will be more convenient for us to think of a k-flag as an ordered k-tuple (`1 , . . . , `k ) of orthogonal
lines in Cn . (By orthogonal we mean with respect to the standard Hermitian inner product. The relation
with the previous description is the obvious one: Vi = `1 + · · · + `i .)
For each i = 1, . . . , k there is a map
pi : F l(k, n) → CP n−1 by (`1 , . . . , `k ) 7→ `i .
2 n−1
Pulling back the generator of H (CP ) for each i = 1, . . . , k gives elements x1 , . . . , xk ∈ H 2 (F l(k, n)).
Lemma 8.5.1. The cohomology of F l(k, n) is given by:
H ∗ (F l(k, n), Z) = Z[x1 , . . . , xk ]/(xn1 , xn−1
2 , . . . , xkn−k+1 ) = H ∗ (CP n−1 × · · · × CP n−k , Z).

Proof. Forgetting the last line defines a map to F l(k − 1, n); the fibre over a k − 1 tuple (`1 , . . . , `k−1 )
consists of lines in (`1 + · · · + `k−1 )⊥ , i.e. it is isomorphic to CP n−k . It is easy to see that
CP n−k −−−−→ F l(k, n)


y
F l(k − 1, n)
n−k
is a fibre bundle with fibre CP . Now the element xk restricts to any fibre as the generator of H 2 .
Indeed, the fibre over (`1 , . . . , `k−1 ) consists of lines in the space (`1 + · · · + `k−1 )⊥ , which is a copy of
CP n−k sitting in the CP n−1 consisting of choices of `k , from which the generator xk was pulled back.
Thus the restriction to the fibre is surjective in cohomology and the Leray-Hirsch theorem applies to
F l(k, n) → F l(k − 1, n).
Let us prove the lemma by induction on k. The induction hypothesis gives that H ∗ (F l(k − 1, n), Z)
is generated by x1 , . . . , xk−1 with relations as in the theorem. The Leray-Hirsch theorem then tells us
that H ∗ (F l(k − 1, n), Z) is freely generated over H ∗ (F l(k − 1, n), Z) by xk , with the relation xn−k+1
k = 0.
This proves the lemma. 

For the space of k-flags in an infinite-dimensional space C∞ = limn Cn , viz. the space F l(k, ∞) =
−→
limn F l(k, n) the same proof shows that the cohomology is
−→
H ∗ (F l(k, ∞), Z) = Z[x1 , x2 , . . . ]
i.e. a polynomial ring in the generators defined by pulling back the generator of H 2 (CP ∞ ) under the k
maps F l(k, ∞) → CP ∞ . Notice that this is simpler than H ∗ (F l(k, n), Z). (47)

8.6. Example: Grassmannians. Consider the Grassmannian Gr(k, n) of k-dimensional subspaces of

Cn . There is a fibre bundle
F l(k, k) −−−−→ F l(k, n)

π
y
Gr(k, n)
where the vertical map π sends (`1 , . . . , `k ) to `1 + · · · + `k . The fibre over a subspace E ∈ Gr(k, n)
consists of full flags in E, which is simply F l(k, k). We have seen above that the restriction map
H ∗ (F l(k, n), Z) → H ∗ (F l(k, k), Z)
is surjective and this applies to any fibre. Thus the Leray-Hirsch theorem applies and we deduce that
the pullback
π ∗ : H ∗ (Gr(k, n), Z) → H ∗ (F l(k, n), Z)
is injective and in fact the image is a direct summand.
Now we consider the case of Gr(k, ∞), in which case we have H ∗ (Gr(k, ∞), Z) ,→ Z[x1 , . . . , xk ].
47We are being a little careless here: If F l(k, ∞) is defined as the direct limit of F l(k, n) as n → ∞, then it is not clear
that the cohomology satisfies H i (F l(k, ∞), Z) is the group H i (F l(k, n), Z), which is independent of n for n large. This
same comment applies to Gr(k, ∞) below. Nevertheless, this is okay, see e.g. Milnor-Stasheff.
 103

Lemma 8.6.1. The image of π ∗ : H ∗ (Gr(k, ∞), Z) → H ∗ (F l(k, ∞), Z) = Z[x1 , . . . , xk ] is the subalgebra
of symmetric polynomials.

Proof. First we show that the image consists of symmetric polynomials. For a permutation σ of {1, . . . , k}
there is a map σ : F l(k, n) → F l(k, n) by permuting the lines. This map induces the automorphism
xi 7→ xσ(i) on H ∗ (F l(k, n), Z). Now π ◦ σ = π because `1 + · · · + `k = `σ(1) + · · · + `σ(k) , so that σ ∗ π ∗ = π ∗
and thus the image consists of symmetric polynomials.
To show that the image is the subalgebra of symmetric polynomials we need a definition. For a free
graded Z-module A∗ = ⊕k≥0 Ak with Ai = 0 for odd i the Poincaré series is the formal power series
X
PA (t) := rank(A2k )tk ∈ Z[[t]].
k≥0
∗ ∗
For a graded tensor product A ⊗ B we have:
PA⊗B = PA PB .
Thus PH ∗ (F l(k,∞)) = (1 − t)−k . By the computation of the cohomology of F l(k, k) we have
k k
Y 1 − ti Y
PH ∗ (F l(k,k)) = (1 + t)(1 + t + t2 ) · · · (1 + t + · · · + tk−1 ) = = (1 − t)−k (1 − ti ).
1−t i=1
k=1

Thus by the isomorphism of graded groups

H ∗ (F l(k, ∞), Z) = H ∗ (F l(k, k), Z) ⊗ H ∗ (Gr(k, ∞), Z)
coming from Leray-Hirsch, the Poincaré series Q = PH ∗ (Gr(k,∞)) satisfies:
k
Y
−k
Q · (1 − t) (1 − ti ) = (1 − t)−k
i=1

and hence
k
Y 1
Q= .
i=1
1 − ti
Now we use the classical fact (see e.g. Lang’s Algebra) that the symmetric polynomials are freely
generated, as a Z-algebra, by the elementary symmetric polynomials c1 , . . . , ck where:
c1 = x 1 + · · · + x k
X
c2 = xi xj
1≤i<j≤k
(8.6.1)
..
.
ck = x1 · · · xk .
Thus the Poincaré series of the algebra of invariant polynomials is
k
Y 1
(1 + t + t2 + · · · )(1 + t2 + t4 + · · · ) · · · (1 + tk + t2k + · · · ) = .
i=1
1 − ti
This is exactly Q and hence we have proved the lemma. 

We have proved:
Theorem 8.6.2. H ∗ (Gr(k, ∞), Z) is a polynomial ring over Z in k generators c1 , . . . , ck where deg(ci ) =
2i.

Note that the generators are canonical (under π ∗ they map to the symmetric functions in x1 , . . . , xk ,
which are themselves canonically defined).
The same arguments as above, using the fact that H ∗ (RP n , Z/2Z) = (Z/2Z)[x]/(xn+1 ), give the
Z/2Z-cohomology of the infinite real Grassmannian:
Theorem 8.6.3. H ∗ (GrR (k, ∞), Z/2Z) is the polynomial ring over Z/2Z in k generators w1 , . . . , wk
where deg(wi ) = 2i.
104

These two theorems are of fundamental importance in the theory of characteristic classes, because
Gr(k, ∞) and GrR (k, ∞) are the classifying spaces for vector bundles: On any reasonable space X (e.g.
a manifold) any rank r complex (resp. real) vector bundle E is the pullback of the tautological rank r
bundle on Gr(k, ∞) (resp. GrR (k, ∞)) by a mapping hE of X to Gr(k, ∞) (resp. to GrR (k, ∞)) which
is unique up to homotopy. The Chern classes ci (E) := h∗E (ci ) ∈ H 2i (X, Z) in the complex case (resp.
Stiefel-Whitney classes wi (E) := h∗E (wi ) ∈ H 2i (X, Z/2Z) in the real case) are then important invariants
of the vector bundle. For example, a famous theorem of Thom says that a complex manifold X is the
boundary of a manifold (i.e. cobordant to the empty set) if and only if all the Chern classes of its tangent
bundle are trivial. For more on this see Milnor-Stasheff, Characteristic Classes.

8.7. Thom isomorphism theorem. A (smooth) vector bundle E → M of rank r is called oriented if
we are given a nowhere vanishing section of ∧r E → M . Informally, we have an orientation of each fibre
vector space Ex = π −1 (x) varying smoothly in x ∈ M . This also gives an orientation in the topological
sense, i.e. a class in H r (Ex , Ex − {x}). Notice that together with an orientation of M , this determines
an orientation of the manifold E.
Theorem 8.7.1 (Thom isomorphism). Let π : E → M be an oriented smooth vector bundle of rank r
on a (compact, oriented) manifold M . Then
(1) There exists a unique class u ∈ H r (E, E − M ) (called the Thom class) such that for each
x ∈ M , the image of u under H r (E, E − M ) → H r (Ex , Ex − {x}) is the orientation class of the
fibre Ex .
(2) The map
H ∗−r (M ) −→ H ∗ (E, E − M ) α 7→ π ∗ (α) ` u
is an isomorphism of H ∗ (M )-modules.

Proof. Put a metric on the vector bundle, i.e. a norm on the vector spaces Ex = π −1 (x) which varies
smoothly in x ∈ M . (Exercise: This can be constructed using a trivialization and a partition of unity.)
Let DE be the unit disk bundle, i.e. in each fibre of π we take the vectors of norm ≤ 1, and let SE
be the sphere bundle (vectors of norm = 1 in each fibre). Then DE is a smooth manifold of dimension
dim M + r with boundary ∂DE = SE . By excision and a deformation retraction we have
H ∗ (E, E − M ) ∼
= H ∗ (DE , SE ).
Then we have isomorphisms
∩[DE ] ∩[M ]
H k (DE , SE ) −→ Hdim M +r−k (DE ) ∼
= Hdim M +r−k (M ) ←− H k−r (M )
by Lefschetz duality for DE , the fact that M ,→ DE is a homotopy equivalence, and Poincaré duality
for M .
(1) Now for k = r we see that H r (E, E − M ) ∼ = H 0 (M ) = Z (assuming M is connected, which case
is enough to consider). Let u 7→ 1 under this isomorphism. The sequence of isomorphisms used above is
compatible with restriction to a single fibre of DE → M (use the local triviality to reduce to the case of
products), and this shows that u restricts to the orientation class in each fibre.
(2) Form the quotient space Q of DE in which each fibre of SE is collapsed to a point. (To show this
exists, construct it locally using a trivialization of E and then glue.) This is a fibre bundle over M with
fibre the r-sphere DE,x /SE,x ∼ = S r . The Leray-Hirsch theorem applies – check that the existence of the
class u gives the condition needed to apply that theorem. This gives that H ∗ (S r )⊗H ∗ (M ) → H ∗ (Q) is an
isomorphism, and this restricts to the isomorphism in the theorem on the summand H r (S r ) ⊗ H ∗ (M ) =
H ∗+r (M ) since H r (S r ) = H r (DE , SE ). 

Applying this to the case of the normal bundle of a submanifold, we get:

Corollary 8.7.2. Let i : M ,→ N be an embedding of oriented smooth manifolds with M compact and
let c be the codimension of M in N . Then H ∗ (N, N − M ) ∼
= H ∗−c (M ).

Proof. Use the tubular neighbourhood theorem and excision to replace the pair (N, N − M ) by the disk
bundle of the normal bundle to M in N , and apply the previous theorem. (An orientation of the normal
bundle is given by the orientations of M and N .) 
Remark 8.7.3. The Thom isomorphism also holds in the case M is not oriented (but E is an oriented
vector bundle). See Milnor-Stasheff. It also holds without any assumption if we use Z/2Z coefficients
(and the proofs above work).
 105

9. Complex manifolds and the Dolbeault theorem (∗)

9.1. Tangent bundle. Let X be a complex n-manifold, and T X its tangent bundle considered as a real
2n manifold. We can complexify this vector space and write (T X)C for the complexified tangent bundle.

9.2. Almost complex structures and integrability. The main application of the Frobenius theorem
will be to the correspondence between Lie groups and Lie algebras, discussed in the section on Lie groups.
Here we give another application of the analytic version of the Frobenius theorem, proving (following
Voisin’s book) a special case of the famous Newlander-Nirenberg theorem.
An almost-complex structure on a real 2n manifold X is a section J ∈ Γ(X, End(T X)), i.e. an
endomorphism of the (real) tangent bundle, which satisfies J 2 = −I. Thus it induces in each tangent
space a complex structure
Example 9.2.1.
Theorem 9.2.2 (Newlander-Nirenberg). Let (X, J) be a manifold with an almost-complex structure.
0,1
Then J comes from a complex structure if and only if the subbundle TX of the complexified tangent
bundle of X defined by J is integrable.

Proof. We will only prove this assuming that (X, J) are real analytic, following the idea of Weil (as in
Voisin’s book).

¯ etc.
9.3. Differential forms: ∂, ∂,
¯
9.4. The ∂-Poincaré lemma.

9.5. Dolbeault theorem.