Lecture 7: Consequences of Poincaré Duality

December 6, 2018

Let M be a closed (compact, without boundary) oriented n-manifold,

and let [M ] be its fundamental class.
For all k we have that the map α 7→ α ∩ [M ] induces an isomorphism

∩[M ] : H k (M ) → Hn−k (M ).

We also know that both the homology and cohomology of M are finitely
generated. Thus, the Universal Coefficient Theorem gives a short exact
sequence.

0 → Hom(Tor(Hk−1 (M )), Q/Z) → H k (M ) → Hom(Hk (M )/Tor, Z) → 0.

The first term is the torsion subgroup of H k (M ) and the last is the free
abelian group which is the quotient of H k (M ) by its torsion subgroup.
Thus, we deduce isomorphisms
∼
=
Hom(Hk (M )/Tor, Z) −→ Hn−k (M )/Tor
∼
=
Hom(TorHk−1 (M ), Q/Z) −→ Tor(Hn−k (M )).
These isomorphisms are equivalent to pairings

Hk (M )/Tor ⊗ Hn−k (M )/Tor → Z,

called the intersection pairing, and

Tor(Hk−1 (M ) ⊗ Tor(Hn−k (M )) → Q/Z,

called the linking pairing. Each of these pairings is perfect in the sense that
the adjoint of these pairings are the above isomorphisms.

1
1 Relationship to the Thom Isomorphism
Suppose that E → X is an n-dimensional vector bundle. We give this bundle
a metric (a positive definite pairing on each fiber varying continuously as
we change fibers). This is equivalent to reducing the structure group of the
bundle from GL(n, R) to O(n). Then we denote by D(E) → X the unit
disk bundle and S(E) → X the unit sphere bundle.

Theorem 1.1. (Thom Isomorphism) Suppose that the bundle π : E → X

is orientable n-dimensional vector bundle (meaning that the structure group
has been reduced from O(n) to SO(n)). Then there is a unique cohomology
class U ∈ H n (D(E), S(E)) whose restriction to any fiber is the relative
fundamental class of that fiber. Furhermore, the map

H ∗ (X) → H ∗+n (D(E), S(E))

given by
a 7→ π ∗ a ∪ U
is an isomorphism

Proof. This follows directly from a relative version of the Serre Spectral
Sequence.

Definition 1.2. The class U in the above theorem is called the Thom Class
and the isomorphism is called the Thom Isomorphism.

Now suppose that M is an n-dimensional manifold and R ⊂ M is a

closed k-dimensional submanifold. Then R has a normal bundle in M ; that
is to say there is a vector bundle ν → R and a diffeomorphism ψR from D(ν)
onto a neighborhood of R in M , the map sending the zero section of D(ν) by
the identity to R. Such a pair consisting of a bundle and a diffeomorphism
is called a tubular neighborhood
Now suppose that both M and R are oriented. Then the normal bundle
ν inherits an orientation. Let U be the Thom class of this orientation. Under
the diffeomorphism and excision we can view U as a class in H n−k (M, M \
D(E)).

Proposition 1.3. The image of U in H n−k (M ) is Poincaré dual to the

image of the fundamental class [X] in Hk (M ) under the embedding X → M .

Proof. Since Hk (D(ν)) = Z generated by [X] embedded as the zero section

it follows that U ∈ H n−k (D(ν), S(ν)) is the Poincaré dual class to [X] in

2
(D(ν), S(ν). Thus, U ∩ [D(ν), S(ν)] = [X]. Since the fundamental class of
M restricts to the relative fundamental class of (D(ν), S(ν)), it follows that
U ∩ [M ] = ±[X]. The choice of U from the orientations of M and R leads
to a sign of +1.

Notice that if S is an oriented submanifold of dimension n − k meeting

R transversally, then the algebraic intersection of R with S is h[U ], [S]i.

2 The Intersection Pairing

Unraveling the definitions we see that the pairing

ϕ : Hk (M )/Tor ⊗ Hn−k (M )/Tor → Z

is given as follows. For homology classes a ∈ Hk (M ) and b ∈ Hn−k (M ),

the pairing ϕ(a, b) is obtained (by taking Poincaré dual cohomology classes
α ∈ H n−k (M ) and β ∈ H k (M ) and forming

hα ∪ β, [M ]i.

Equivalently, the intersection a · b is given by the evaluation on b of the

Poincaré dual cohomology class to a. From the first description it follows
that the intersection pairings are signed symmetric:

ϕ(ak , bn−k ) = (−1)k(n−k) ϕ(bn−k , ak ).

Let us give a geometric description of the intersection pairing in the case

when the ambient manifold is smooth.
Given homology classes a ∈ Hk (M ) and b ∈ Hn−k (M ) we choose cycle
representatives ea and eb. We can assume that every singular simplex appear-
ing each of these cycles is a smooth map and also that any two simplices
meet transversally. This means that the only points of intersection are where
the interior of a k-simplex in e a meets the interiior of an (n − k)-simplex in
eb. At every such point x of intersection both e a and eb are local embeddings
and their tangent spaces are complementary in Tx M . We assign a sign to
each point of intersection by comparing the direct sum of orientations on the
tangent space of ea and of eb with the ambient orientation of the tangent space
of M . The sum of the signs over the (finitely many) points of intersection
gives the intersection pairing applied to (a, b).
To see that the pairing is well-defined suppose we have e a homologous
a0 both being transverse to eb. Then there is a (k + 1)-chain e
to e c with

3
∂e
c=e a0 − ea. we can suppose that e c is also transverse to eb. Then the top
dimensional simplices meet in a 1-manifold whose boundary is either in a
codimension-1 face of e c or of eb. Since eb is a cycle its codimension-1 faces
cancel out in pairs. This means that the intersection 1-manifold continues
across such faces without introducing a boundary.. A similar argument
works for the codimension-1 faces of e c that are interior to e c. Thus, we see
that the boundary of the intersection of e a0 · eb − e
c with eb is e a · eb. But the
algebraic boundary of a 1-manifold is zero. This shows that varying e a by a
boundary does not change the algebraic intersection with b. Symmetrically,
e
varying eb by a boundary does not changes its algebraic intersection with e a.
This shows that the algebriac intersection is well-defined on homology. It is
exactly the pairing produced by Poincaré Duality.
It is also clear from this description that the pairing is signed symmetric.

2.1 Middle dimensional intersection pairings: 4k + 2 case

Now suppose that M is closed, oriented and of dimension 2n. Then we have
a pairing
Hn (M )/Tor ⊗ Hn (M )/Tor → Z
that is (−1)n symmetric and unimodular, meaning that if we choose a basis
then the pairing is represented by a (−1)n symmetric matrix of determinant
±1..
Claim 2.1. If n is odd, then there is a basis in which the pairing is an
orthogonal sum of 2 × 2 matrices
 
0 1
.
−1 0
Proof. First note that by skew symmetry we have x · x = 0 for all x ∈
Hn (M ). Let x be an indivisible element, i.e., part of a basis. Then there is
a homomorphism Hn (M ) → Z sending x to 1. Hencem there is y such that
x·y =1
The 2×2 matrix giving the pairing on the span of x, y is exactly the 2×2
matrix given in the statement of the claim. Since this matrix is unimodular,
x, y generate an orthogonal direct summand of Hn (M ) and we continue by
induction.
Corollary 2.2. If M 4k+2 is a closed, orientable manifold, the the Euler
characteristic of M is even
Notice that this is not true without the orientability assumption: The
Euler characteristic of RP 2 is 1.

4
2.2 Middle Dimensional Intersection Pairings: 4k case
If M 4k is a closed, oriented manifold then the intersection pairing
H2k (M ) ⊗ H2k (M ) → Z
is symmetric. Choosing a basis it is given by a symmetric matrix of deter-
minant ±1.
The most elementary of such pairings are h1i and h−1i: one dimensional
pairings with generator x with x·x = ±1. Of course, we can take orthogonal
direct sums of these. Pairings represented by diagonal matrices with ±1’s
down the diagonal. But there are other pairings. There is the hyperbolic
pairing given by  
0 1
.
1 0
We know that this pairing is not diagonalizable since x · x is even for all
x. A form with this property is called an even form. It is an easy exercise
to show that a form is even if and only if any matrix representative for it
has only even entries down the diagonal. The parity of a pairing is even if
x · x ≡ 0 (mod 2) for all x, and otherwise the parity is odd.
Another example of an even pairing is given by the matrix associated
with the Dynkin diagram of E8 . It is an 8 × 8 matrix with basis identified
with the vertices in the Dynkin diagram for the Lie algebra E8 . This matrix
has all diagonal entries +2. All off diagonal elements are either 0 or 1, and
an off diagonal entry at position (i, j) is 1 if and only if there is a bond in
the Dynkin diagram connecting the ith and j th vertex. It turns out that
this matrix has determinant ±1 (which turns out to be equivalent to the
fact that the center of the simply connected form of E8 is the trivial group,
or equivalently that E8 has no non-simply connected form). The form is an
even, positive definite form.
Classifying non-degenerate symmetric forms over R is easy:
Claim 2.3. Let V be a finite dimensional vector space with a non-degenerate
symmetric (real linear) pairing
A : V ⊗ V → R.
(Non-degernate means that if A(v, w) = 0 for all w ∈ V then v = 0.) Then
there is a basis {e1 , . . . , ek } for V such that A(ei , ei ) = ±1 for all i, and
A(ai , aj ) = 0 for all i 6= j. That is to say, the symmetric matrix of the
pairing is diagonal with ±1s down the diagonal. The number of +1’s and
the number of −1’s that appear are invariants of the isomorphism class of
the pairing.

5
Proof. Suppose pthat V 6= 0 and choose x ∈ V . Suppose thatA(x, x) 6= 0.
Then e1 = x/ |A(x, x)| has A(e1 , e1 ) = ±1. If A(x, x) = 0, then there is
y ∈ V with A(x, y) = 1/2. If A(y, y) = 0, then A(x + y, x + y) = 1. Thus,
we can always fine x ∈ V with A(x, x) 6= 0, and hence there is an element
e1 ∈ V with A(e1 , e1 ) = ±1. Extend e1 to a basis {e1 , . . . , ek } and for every
i > 1 replace ei with ei − A(e1 , ei )e1 . After this replacement A(e1 , ei ) = 0
for all i > 1. This means that V is an orthogonal sum of he1 i and the
subspace V 0 spanned by {e2 , . . . , ek }. We then go by induction to find a
basis as required.
Arrange that A(ei , ei ) = +1 for 1 ≤ i ≤ k + and equal −1 for k + + 1 ≤
i ≤ k and let V + be the subspace of V spanned by the {ei }i=1 Y k + and V −
be the subspace spanned by {ei }ki=k+ +1 . Then the pairing is positive definite
on V + . Suppose that V 0 ⊂ V is subspace on which the pairing is positive
definite. Then V 0 ∩ V − = {0} and hence the projection of V 0 to V + is an
injection, meaning the dim(V + ) ≤ k + . Thus, the number of +1’s down the
diagonal is the the maximal dimension of any subspace on which the pairing
is positive definite.

Definition 2.4. We define the signature of a pairing to be the number of

+1’s minus the number of −1’s in any diagonalization of the pairing as in the
previous claim. The pairing is positive definite if and only if the signature
equals the rank and is negative definite if and only if the signature is equal
to minus the rank. Otherwise, the pairing is said to be indefinite. Notice
that the signature of a pairing is between minus the rank of the pairing and
plus the rank of the pairing and is congruent to the rank modulo 2.
If L is a lattice (a finitely generated free abelian group) and if A : L⊗L →
Z is a non-degenerate symmetric pairing (meaning the determinant of a
matrix representative is non-zero) on L, then the signature of the pairing is
the signature of the extension of A to a real-linear non-degenerate symmetric
pairing on L ⊗ R.
We gave an example, E8 , of an even, unimodular, symmetric, positive-
definite pairing of rank 8, and hence of signature 8. In fact we have:
Lemma 2.5. If L is a lattice with an even symmeric, unimodular pairing,
then the signature of L is congruent to 0 modulo 8.
There is a nice classification result for indefinite, unimodular pairings.
It is quite intricate to prove and we shall not discuss the proof.
Theorem 2.6. Two indefinite unimodular pairings are isomorphic (over Z)
if and only if they have the same rank, signature and parity.

6
 This result does not extend to definite pairings.
Claim 2.7. E8 ⊕ h1i and ⊕pi=1 h1i are both odd pairings of rank 9 and sig-
nature 9. They are not isomorphic,
Proof. The only thing that needs establishing to prove the claim is that
the pairings are not isomorphic. Let us consider the x in each pairing with
x · x = 1. The only solutions in E8 ⊕ h1i are the two generators of the
second factor, whereas in ⊕9i=1 h1i there are the nine basis elements and
their negatives.

For every rank n there are only finitely many isomorphism classes of
definite forms of rank n. For example, there are two even definite forms of
rank 16:

3 The linking pairing

Let M be a closed, oriented n-manifold. The linking pairing is the pairing
TorHk−1 (M ) ⊗ TorHn−k (M ) → Q/Z produced by Poincaré duality. Let us
give a geometric description along the lines of the intersection pairing. Let
a ∈ Hk−1 (M ) and b ∈ Hn−k (M ) be torsion classes. Choose representative
cycles ea and eb which we can assume are smooth and in general position.
The latter means that the cycles are disjoint. Then for some N ≥ 1 there is
a chain e c of degree k with ∂e c = Nea. We can assume that e c is smooth and
transverse to b. Thus, as in the definition of the intersection pairing we have
e
the algebraic intersection e c · eb ∈ Z. As in the case of the intersection pairing,
if we replace e c by e 0
c , a chain with the same boundary with the property that
c0 − e
e c is itself a boundary we do not change the algebraic intersection with
eb. More generally, if e c0 and e c have the same boundary, then their difference
is a cycle and the difference of their algebraic intersections with eb is the
homological intersection of the homology class represented by e c0 − e
c with b.
Since b is a torsion class, this homological intersection is zero. This proves
c · eb is independent of the choice of e
that e c with boundary Ne a. If we consider
0
c with ∂e
e c0 = N 0 e
a we see that
1 e 1 0 e
c · b = 0e
e c · b.
N N
Thus, given the disjoint cycles ea and eb representing torsion classes we have a
well-defined rational number defined by choosing e c with ∂ec being a multiple
of e
a and intersecting e
c with eb and dividing by the multiple in question. This
rational number is the linking number of the disjoint cycles e a and eb.

7
 If we vary e a by a homology, to another cycle e a0 disjoint from eb, then
this homology will have an algebraic intersection number with eb which is an
integer and it is easy to see that the difference linking number of the cycles e a0
with eb and the linking number of the cycles e a0 with eb is exactly that integer.
0
Similarly, if we vary eb by a homology to eb disjoint from e a this homology
has intersection number with e a, which is an integer and the linking number
of the cycles changes by this integer. Consequently, the homology classes a
and b have a well-defined linking pairing in Q/Z. This pairing is denoted
lk(a, b). It is the pairing produced by Poincaré duality.
The linking pairing is also signed symmetric: If a has degree k − 1 and
b has degree n − k, then

lk(a, b) = (−1)k(n−k+1) lk(b, a).

4 Homotopy Type of simply connected 4-manifolds

Let X be a simply connected space whose homology satisfies Poincaré duality
in dimension 4, meaning that there is a classes [X] ∈ H4 (X) such that ∩[X]
induces an isomorphism H ∗ (X) → H4−∗ (X). Then H2 (X) is a free abelian
group (since Tor(H2 (X)) is isomorphic to Tor(H 2 (X)) which in turn is dual
to Tor(H1 (X)) which vanishes since X is simply connected). Thus, there is a
map ∨S 2 → X from a wedge of 2-spheres to X inducing an isomorphism on
H2 . Since H3 (X) = 0 it follows that H∗ (X, ∨S 2 ) = 0 for ∗ < 4 and H4 (X) →
H4 (X, ∨S 2 ) is an isomorphism, so that H4 (X, ∨S 2 ) = Z. This means that
there is an map f : (D4 , S 3 ) → (X, ∨S 2 ) inducing an isomorphism on H4 .
Thus, we form ∨S 2 ∪D4 → X where the map to X which is given by f on D4
and the attaching map S 3 → ∨S 2 is the restriction of f to S 3 . The resulting
map is an isomorphism on homology and hence is a homotopy equivalence.
The group π3 (∨ki=1 S 2 ) is isomorphic to the group of k × k symmetric
matrices over Z. One way to see this is deform a map S 3 → ∨S 2 transverse
to a point in the interior of each S 2 . Then the preimage of each point
gives us a framed link L in S 3 and the linking number of these (and the
self-linking number using the framing) gives the symmetric matrix. The
condition that X satisfies Poincaré duality is the condition that this matrix
is unimodular. Two such X are homotopy equivalent if and only if the
pairings are isomorphic.
We have sketched a proof of the following:

Claim 4.1. Simply connected CW complexes satisfying 4-dimensional Poincaré

duality up to homotopy equivalence are classified by the isomorphism type

8
of the intersection pairing on H2 . All unimodluar, symmetric pairing come
from such CW complexes.
This leads naturally to a question:
Question: Which unimodular pairings are realized as the intersec-
tion pairing on H2 of a closed, oriented 4-manifold
We have some examples CP 2 with the orientation induced from its nat-
ural complex structure represents h1i; CP 2 with the opposite orientation
represents h−1i. S 2 × S 2 represents
 
0 1
.
1 0
We have:
Theorem 4.2. (M. Freedman) Every symmetric unimodular form is the
intersection form on H2 of a simply connected, oriented, closed topologi-
cal 4-manifold. If the form is even it is the form of a unique 4-manifold
up to homeomorphism. If it is an odd form there are exactly two non-
homeomorphic 4-manifolds realizing this form. For each odd form exactly
one of these 4-manifolds with that intersection form has the property that
its product with a circle has a smooth structure.
Theorem 4.3. (S. Donaldson) If a positive definite form is realized as the
intersection form on H2 a smooth, simply connected 4-manifold, then that
form is diagonal with +1’s down the diagonal. Thus, for example, E8 ⊕ h1i
is not the form of a smooth 4-manifold.
The contrast of these two theorems shows that the theory of topologiical
4-manfiolds and smooth 4-manifolds differ. In fact, they differ drastically.
Theorem 4.4. (R. Friedman and J. Morgan) There are topological mani-
folds with infinitely many non-diffeomorphic smooth structures. One exam-
ple is CP 2 blown up 9 times.
By contrast, in every other dimension any compact topological manifold
has at most finitely many non-diffeomorphic smooth structures.

5 Lefschetz Duality
Let M be a compact, oriented n-manifold. Lefschetz duality is equivalent
to the statement that the induced pairings
Hk (M, ∂M )/Tor ⊗ Hn−k (M )/Tor → Z

9
 and
Tor(Hk−1 (M, ∂M )) ⊗ Tor(Hn−k (M )) → Q/Z
are perfect pairings. Of course they are still perfect pairings if we reverse
the roles of relative and absolute homology. This means that

∩[M, ∂M ] : H k (M ) → Hn−k (M, ∂M )

is also an isomorphism.
Thus, Lefschetz duality tells us that the long exact sequences of homology
and cohomology are dual:

H k (M, ∂M ) −−−−→ H k (M ) −−−−→ H k (∂M ) −−−−→ H k+1 (M, ∂M ) −−−−→

   
∩[M,∂M ]y ∩[M,∂M ]y ∩[∂M ]y ∩[M,∂M ]y
   

Hn−k (M ) −−−−→ Hn−k (M, ∂M ) −−−−→ Hn−k−1 (∂M ) −−−−→ Hn−k−1 (M ) −−−−→
Proposition 5.1. Now suppose that M is a 4k + 1 compact, oriented man-
ifold with boundary. Then the signature of the intersection on H2k (∂M ) is
zero.
Proof. Consider the exact sequence
∂ ∗ i
H2k+1 (M, ∂M ) −→ H2k ∂M ) −→ H2k (M ).

Modulo torsion, the first term is dual to the last and i∗ is the adjoint of ∂
with respect to the intersection pairing on H2k (M ). That is to say ∂(a) · b =
ha, i∗ (b)i. In particular the rank of the image of ∂ is equal to the rank of the
image of i∗ . But the rank of the image of i∗ is the rank of H2k (∂M ) minus
the rank of the Ker(i∗ ), which by exactness is the rank of H2k (∂M ) minus
the rank of the image of ∂. We conclude that the rank of the image of ∂ is
equal to one-half the rank of H2k (∂M ).
Since i∗ ◦ ∂ = 0, we see that the image of ∂ is a self-annihilating sub-
space of H2k (∂M ) (meaning that any two elements in this subspace has
intersection product 0). Let L ⊂ H2k (∂M ) be the subgroup of all elements
with the property that some positive multiple of the element is in the image
of ∂. The image of L/Tor in H2k (∂M )/Tor is a direct summand which is
self-annihilating under the intersection pairing. Furthermore, the rank of
L/Tor is one-half the rank of H2k (∂M )/Tor. It follows that the pairing is
isomorphic to an orthogonal direct sum of pairings of the form
 
0 1
1 0

10
and  
0 1
.
1 1
Thus, the signature of the pairing is trivial.

Corollary 5.2. For any k ≥ 1 the manifold CP 2k is not the boundary of a

compact, oriented (4k + 1)-manifold.

Proof. The signature of the pairing on H2k (CP 2k ) is +1.

Consider the closed, oriented 5-manifold obtained by taking CP 2 × I

and gluing the ends together by the map induced by complex conjugation.
The result is a closed oriented 5-manifold. Using the linking pairing on H2
one can show that this manifold is not the boundary of a compact, oriented
6-manifold.

11