An introduction to differential topology and surgery theory

Anthony Conway

Fall 2018

Introduction
These notes are based on a course that was taught at Durham University during the fall of 2018.
The initial goal was to provide an introduction to differential topology and, depending on the
audience, to learn some surgery theory. As it turned out, the audience was already familiar with
basic manifold theory and so surgery became the main focus of the course. Nevertheless, as a
remnant of the initial objective, the use of homology was avoided for as long as possible.
The course decomposed into two parts. The first part assumed little background and intro-
duced some basic differential topology (manifolds, tangent spaces, immersions), algebraic topology
(higher homotopy groups and vector bundles) and briefly discusses the Smale’s sphere eversion.
The second part consisted of an introduction to surgery theory and described surgery below the
middle dimension, the surgery obstruction in even dimensions and an application to knot theory
(Alexander polynomial one knots are topologically slice).
These notes no doubt still contain some inaccuracies. Hopefully, they will be polished in 2019.

1 Differential topology 2
1.1 Smooth manifolds and their tangent spaces . . . . . . . . . . . . . . . . . . . . . . 2
1.1.1 Smooth manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.2 Tangent spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.3 Immersions and embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2 Vector bundles and homotopy groups . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.1 Vector bundles: definitions and examples . . . . . . . . . . . . . . . . . . . 8
1.2.2 Some homotopy theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2.3 Vector bundles: the homotopy classification . . . . . . . . . . . . . . . . . . 15
1.3 Turning the sphere inside out . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2 Surgery theory 19
2.1 Surgery below the middle dimension . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1.1 Surgery and its effect on homotopy groups . . . . . . . . . . . . . . . . . . . 20
2.1.2 Motivating the definition of a normal map . . . . . . . . . . . . . . . . . . . 21
2.1.3 The surgery step and surgery below the middle dimension . . . . . . . . . . 24
2.2 The even-dimensional surgery obstruction . . . . . . . . . . . . . . . . . . . . . . . 26
2.2.1 The intersection and self intersection numbers of immersed spheres . . . . . 27
2.2.2 Symmetric and quadratic forms . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2.3 Surgery on hyperbolic surgery kernels . . . . . . . . . . . . . . . . . . . . . 33
2.2.4 The even quadratic L-groups . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.2.5 The surgery obstruction in the even-dimensional case . . . . . . . . . . . . 37
2.3 An application of surgery theory to knot theory . . . . . . . . . . . . . . . . . . . . 40

1
Chapter 1

Differential topology

The goal of this chapter is to discuss some classical topics and results in differential topology. As
a motivating result, we work towards the proof of Smale’s sphere eversion theorem. The chapter
is organized as follows. In Section 1.1, we introduce smooth manifolds and their tangent spaces
as well as immersions and embeddings. In Section 1.2, we define vector bundles and discuss some
classification results. Finally, in Section 1.3, we outline the proof of the sphere eversion theorem.

1.1 Smooth manifolds and their tangent spaces

This first section is introductory: it defines smooth manifolds, tangent spaces and immersions. We
will mostly follow Tu’s textbook “An introduction to manifolds” [Tu11]; other classical references
include [GP74, Hir76, Wal16].

1.1.1 Smooth manifolds

In this subsection, we define smooth manifolds and smooth maps between them. Our main refer-
ence is [Tu11, Sections 5 and 6].

A topological manifold is a second countable Hausdorff topological space M such that for all
p ∈ M , there exists an open set p ∈ U and a map ϕ : U → Rn which is a homeomorphism onto a
open subset of Rn (i.e. M is locally Euclidean). The pair (U, ϕ) is called a chart. A topological
manifold M is n-dimensional if it is locally homeomorphic to Rn .
Remark 1.1.1. For the dimension of a manifold to be well defined, we need to know that
for n 6= m, an open subset of Rn is not homeomorphic to an open subset of Rm . This is a
non-trivial fact (known as invariance of domain), a proof can be found in Hatcher [Hat02, Theo-
rem 2B.3 and Corollary 2B.4].
Example 1.1.2. We now give some examples of topological manifolds.

1. The euclidean space Rn is a topological n-manifold: it is covered by a single chart (Rn , id).
2. The unit sphere S 2 = {(x, y, z) ∈ R3 | x2 + y 2 + z 2 = 1} is a topological 2-manifold (the
same is true for S n ). Endow S 2 with the induced topology from R3 . Since R3 is Hausdorff

2
 and second countable, so is S 2 . The 2-sphere can be covered by the following six charts
U1 = {(x, y, z) ∈ S 2 | x > 0}, ϕ1 (x, y, z) = (y, z).
U2 = {(x, y, z) ∈ S 2 | x < 0}, ϕ2 (x, y, z) = (y, z),
U3 = {(x, y, z) ∈ S 2 | y > 0}, ϕ3 (x, y, z) = (x, z),
U4 = {(x, y, z) ∈ S 2 | y < 0}, ϕ4 (x, y, z) = (x, z),
U5 = {(x, y, z) ∈ S 2 | z > 0}, ϕ5 (x, y, z) = (x, y),
U6 = {(x, y, z) ∈ S 2 | z < 0}, ϕ6 (x, y, z) = (x, y).
This is clearly an open cover. We check that the ϕi are homeomorphisms onto their image.
We do this for i = 4 but the other cases are analogous. The image of the continuous map ϕ4
consists of pairs (x, z) ∈ R2 such that x2 + y 2 + z 2 √
= 1 for some y < 0. Since x < 0, a
continuous inverse is given by mapping (x, z) to (x, − 1 − x2 − z 2 , z).
Exercise 1.1.3. Show that the plus sign “+”, viewed as a subspace of R2 with the subspace
topology, is not a topological manifold. The solution can be found in [Tu11, Example 5.4].
From now on, we will be interested in topological manifolds with an additional “differentiable
structure”. We will always use the words “smooth” and “C ∞ ” interchangeably. Furthermore,
diffeomorphisms are assumed to be smooth.
Definition 1. A smooth n-dimensional manifold is a n-dimensional topological manifold whose
family A = (Ui , ϕi )i∈I of charts satisfies:
S
1. the family (Ui )i∈I is an open cover of M , meaning that M = i∈I Ui ;
2. the charts are smoothly compatible: for every i, j ∈ I, the following map is a diffeomorphism:
ϕj ◦ ϕ−1
i : ϕi (Ui ∩ Uj ) → ϕj (Ui ∩ Uj );

3. if a chart (U, ϕ) is smoothly compatible with all charts of A , then (U, ϕ) belongs to A .
A familly A of charts satisyfing the first two conditions of Definition 1 is called an atlas. The
third condition ensures that the atlas is maximal.
Proposition 1.1.4. Any atlas on a locally Euclidean space is contained in a unique maximal atlas.
A proof of Proposition 1.1.4 can be found in [Tu11, Proposition 5.19]. In practice, Proposi-
tion 1.1.4 means implies that we need not check the maximality assumption.
Example 1.1.5. We argue that the examples of Example 1.1.2 are in fact smooth manifolds.
1. The euclidean space Rn is a smooth n-manifold. An atlas is given by (Rn , id).
2. The sphere S n is a smooth n-manifold. We prove this for n = 2. The charts described in
Example 1.1.2 clearly cover S 2 and so we need only check that they are smoothly compatible.
We only show that the following map is smooth
ϕ1 ◦ ϕ−1
4 : ϕ4 (U1 ∩ U4 ) → ϕ1 (U1 ∩ U4 ).

Recall that ϕ1 (x, y, z) = (y, z) and ϕ4 (x, y, z) = (x, z). Looking at the definitions of U1
and U4 , we see that
ϕ4 (U1 ∩ U4 ) = {(x, z) ∈ R+ × R | x2 + z 2 = 1 − y 2 for some y < 0},
ϕ1 (U1 ∩ U4 ) = {(y, z) ∈ R− × R | y 2 + z 2 = 1 − x2 for some x > 0}.
√
Recall that the inverse of ϕ4 : U4 → ϕ(U4 ) is given by ϕ−1 2 2
4 (x, z) = (x, 1 − x − z , z). We
2 2 2
deduce that for (x, z) ∈ √ϕ4 (U1 ∩ U4 ), there is a y < 0 such that y =√ 1 − x − z . Thus we
have ϕ1 ◦ϕ−1 2 2 2
4 (x, z) = (− 1 − x − z , z). This map is smooth since − 1 − x − z = y < 0.
2

3
Exercise 1.1.6. Prove the following facts about smooth manifolds.

1. Any open subset of a smooth manifold is a smooth manifold. Deduce that GL(n, R) is a
smooth manifold. The solution can be found in [Tu11, Examples 5.12 and 5.15].
2. The product of two smooth manifolds is a smooth manifold. The solution can be found
in [Tu11, Example 5.17]. Deduce that the n-torus Tn := S 1 × . . . × S 1 is a smooth manifold.
Remark 1.1.7. There are topological manifolds that do not admit any smooth structure: the
first example was produced by Kervaire [Ker60]. A nice introductory account of smoothing theory
can be found in [Sco05, pages 207-224].

Next, we define the notion of a smooth map between manifolds.

Definition 2. Let M and N be smooth manifolds. A continuous map F : M → N is smooth at
p ∈ M if there are charts (U, ϕ) and (V, ψ) around p and F (p) such that ψ ◦ F ◦ ϕ−1 is smooth. F
is smooth if it smooth at every p ∈ M . A diffeomorphism is a smooth bijective map whose inverse
is also smooth.

A concrete exercise (together with a solution) involving the verification that a given map is
smooth can be found in [Tu11, Example 6.19]

1.1.2 Tangent spaces

In the Euclidean space Rn , we think of the “tangent space at p” as the set of all vectors emanating
from p. Such a “tangent vector” can be described by a vector in Rn , and so the tangent space
of Rn at p is just a copy of Rn which we attach to p. In this subsection, we define tangent space
to an arbitrary manifold at a point. We give a definition using derivations, but several equivalent
(more geometric) formulations also exist. We closely follow [Tu11, Section 8].
Let M be an n-manifold and let U, V be two open neighborhoods of a point p ∈ M . Two
functions f : U → R and g : V → R are equivalent if they agree on some subset of U and V that
contains p. A germ at p ∈ M is an equivalence class of functions f : U → R, where U is an open
set of M . The set of all germs at p is denoted by Cp∞ (M ). Observe that, if U is an open set
containing p, then Cp∞ (U ) = Cp∞ (M ): we are dealing with a local notion.
Definition 3. Let M be a manifold and let p ∈ M . A tangent vector at p (or derivation) is
a R-linear map D : Cp∞ (M ) → R that satisfies D(f g) = (Df )g(p) + f (p)Dg. The tangent space
of M at p is defined as the set of all tangent vector at p.

The upshot is that Definition 3 associates to each point p ∈ M a vector space Tp M . As we shall
see below, this vector space is in fact n-dimensional. As a consequence, Definition 3 generalizes
the concept of the “tangent space to Rn at p” that we described above. Note also that if U is an
open set containing p, then Tp U = Tp M .
Notation 1.1.8. We use r1 , . . . , rn to denote the coordinates of Rn , i.e. each ri : Rn → R
is the standard projection. Writing a chart (U, ϕ) in local coordinates consists of writing (U, ϕ)
as (U, x1 , . . . , xn ), where xi := ri ◦ ϕ : U → R. For such a local chart, the tangent vector
∂ ∞
∂xi |p : Cp (M ) → R is defined as follows:

∂
: Cp∞ (M ) → R
∂xi p
∂ ∂f
f 7→ (f ϕ−1 ) =: (p).
∂ri ϕ(p) ∂xi
∂
To check that ∂x i
p
is indeed a derivation, use the corresponding property for the partial derivatives
in R . Observe that in Euclidean space (with the chart (Rn , id)), the coordinates xi = ri are the
n
∂
“usual” coordinates in Rn . In particular, ∂x i
p
is the usual partial derivative.

4
 The next proposition describes the tangent space at a point of Rn .
Proposition 1.1.9. Given p ∈ Rn , the tangent vectors ∂r∂ 1 p , . . . , ∂r∂n p
form a basis of Tp Rn
and, in fact, the following assignement is a linear isomorphism:

Φ : Rn → Tp Rn
X ∂
v 7→ vi i p
i
∂r

Proof. We saw in Example 1.1.8 that in (Rn , id), we have xi = ri , but we keep the former notation.
The map Φ is clearly linear and so we first prove injectivity. Suppose that Φv ≡ 0. To show that
v = 0, we show that v j = 0 for each j. Applying Φv to the coordinate function xj , using Φv ≡ 0,
the definition of Φv and basic calculus in Rn , we obtain the desired result:
X ∂ X
0 = Φv (xj ) = vi i p
(xj ) = v i δij = v j .
i
∂x i

Next, we prove surjectivity. Given a tangent vector D ∈ Tp M , our goal is to find v ∈ Rn such
that D(f ) = Φv (f ) for every germ f ∈ Cp∞ (M ). We claim that setting v i := D(xi ) for i = 1, . . . , n
produces a v that satisfies this property. Given a germ f ∈ Cp∞ (M ), Taylor’s theorem with
remainder (see e.g. [Tu11, Lemma 1.4]) implies that there are smooth functions gi : Rn → R such
∂f
that gi (p) = ∂x i (p) and
n
X
f (x) = f (p) + (xi − pi )gi (x). (1.1)
i=1

Before applying D to this equality, we note that derivations vanish on constant functions: if c
is such a function, then the R-linearity of D implies that D(c) = cD(1) and the Leibniz rule
gives D(1) = D(1 · 1) = 2D(1). Applying D to the Taylor expansion displayed in (1.1), using that
derivations are R-linear, vanish on constants (so that D(f (p)) = 0 and D(pi ) = 0 for i = 1, . . . , n)
and satisfy the Leibniz rule, we obtain
X X X ∂f
D(f (x)) = (Dxi )gi (p) + (pi − pi )D(gi (x)) = (Dxi ) i (p) = Dv (f (x)).
i i i
∂x

It follows that Df = Φv (f ) for arbitrary f ∈ Cp∞ (M ) and consequently D ≡ Φv , as claimed. This

concludes the proof of the proposition.

We now define smooth maps between manifolds.

Definition 4. The differential of a smooth map F : M → N at p ∈ M is defined as

Tp F : Tp M → TF (p) N
(Xp ) 7→ (f 7→ Xp (f ◦ F )).

The following properties of the differential are left as exercises. The solutions are very short
and can be found in [Tu11, pages 88-89].
Exercise 1.1.10. Given smooth maps F : M → N, G : N → X and p ∈ M , show that

1. Tp F is linear;
2. Tp (G ◦ F ) = TF (p) (G) ◦ Tp (F ) (this is the chain rule) and Tp (idM ) = idTp M .

Deduce that if F is a diffeomorphism, then Tp F is a linear isomorphism.

Before making this definition more concrete, we show that the dimension of the tangent spaces
to an n-manifold are n-dimensional.

5
Proposition 1.1.11. Given a chart (U, ϕ) = (U, x1 , . . . , xn ) containing p, the tangent vectors
∂ ∂
∂x1 p , . . . , ∂xn p form a basis of Tp M . In particular, the vector space Tp M is n-dimensional.

Proof. Using Exercise 1.1.10, we know that since ϕ is diffeomorphism onto its image, its differential
∂ ∂
induces a linear isomorphism Tp M → Tϕ(p) Rn . We claim that Tp ϕ maps ∂x i |p to ∂r i |ϕ(p) ; indeed
∞ n
for any germ f ∈ Cϕ(p) (R ), we use the definition of the differential and the definition of the
partial derivative to obtain
∂ ∂ ∂ ∂
Tp ϕ( )(f ) = (f ◦ ϕ) = i (f ◦ ϕ ◦ ϕ−1 ) = (f ).
∂xi p ∂xi p ∂r ϕ(p) ∂ri ϕ(p)

∂ ∂
As Tp ϕ is a linear isomorphism and as Proposition 1.1.9 implies that the ∂r 1 ϕ(p) , . . . , ∂r n ϕ(p)
∂ ∂
form a basis of Tϕ(p) Rn , we deduce that ∂x1 p , . . . , ∂xn p form a basis of Tp M . This concludes the
proof of the proposition.

We conclude by deriving the expression of the differential in local coordinates.

Remark 1.1.12. Let F : M m → N n be a smooth map. Pick charts (U, x1 , . . . , xm ) and (V, y 1 , . . . , y n )
around p and F (p). To understand Tp F in local coordinates, we study its image on the basis vec-
∂ ∂ ∂ ∂
tors ∂x 1 |p , . . . , ∂xm |p of Tp M . Since ∂y 1 |F (p) , . . . , ∂y n |F (p) is a basis for TF (p) N , for each j, we
∂ k ∂ k i
P
can write Tp F ( ∂x j |p ) = k aj ∂y k |F (p) for some aj ∈ R. Evaluating both sides on y , we get
∂(y i ◦F )
∂xj |p = aij , from which we deduce that

∂ X ∂F k ∂
Tp F ( )= .
∂xj p ∂xj ∂y k F (p)
k

A more geometric approach to the tangent space and differential is presented in [Tu11, Sub-
section 8.7]. The advantage of derivations is the simplicity of the formulas and the elementary
proofs of several properties.

1.1.3 Immersions and embeddings

In this short subsection, we describe immersions and embeddings. We also introduce regular
homotopies and define eversions. References include [Tu11, Subsection 8.8] and [Ada93].
Definition 5. A smooth map f : M → N is an immersion if Tx f : Tx M → Tf (x) N is injective at
every point x ∈ M . An immersion is an embedding if it is injective and is a homeomorphism onto
its image. A submanifold is the image of an embedding.

Any map S 1 → R2 with image the figure eight “8” is an immersion but not an embedding:
the double point leads to a failure of the injectivity condition. The next example shows that an
immersion need not be embedding, even if it is injective.
Example 1.1.13. Consider the map β : (−π, π) → R2 , t 7→ (sin(2t), sin(t)) depicted in Figure 1.1
below. β is injective and, calculating derivatives, one sees that it is an immersion. Observe that
β((−π, π)) is compact (e.g. because it is bounded and contains all its limit points). Since (−π, π)
is not compact, we deduce that β can not be an embedding.

Embeddings, by definition, are injective immersions, but thanks to Example 1.1.13, we know
that the converse need not be true. The next proposition shows that the situation simplifies
considerably if the domain is compact.
Proposition 1.1.14. If M is compact, then any injective immersion f : M → N is an embedding.

6
 Figure 1.1: The image of the map β is an injective immersion but not an embedding.

Proof. Since f is an injective immersion, we need only show that f is a homeomorphism on its
image. A standard result in general topology shows that an injective continuous map from a
compact space to a Hausdorff space is a homeomorphism onto its image. Since M is compact (by
assumption) and N is Hausdorff (it is a manifold), we can apply this result, concluding the proof
of the proposition.

Very informally, Proposition 1.1.14 shows that embeddings defined on a compact manifold
should be thought of as “immersions with no multiple points”. Next, one might wonder whether
an immersion can be deformed into an embedding. To make this question more precise, we
introduce some additional terminology.
Definition 6. Two immersions f, g : M → N are regular homotopic if there is a homotopy
H : M × [0, 1] → N such that each ht := H(−, t) : M → N is an immersion.
Immersions will often be denoted as M # N . The next example describes regular homotopy
classes of immersions S 1 # R2 .
Example 1.1.15. Intuitively, any immersion S 1 # R2 with image the figure eight “8” should
not be regular homotopic to the the standardly embedded circle: if we try to remove the kink by
making it smaller, then we run into a cusp. Therefore, while both immersions are homotopic they
should not be regular homotopic.
To make this intuition more precise, recall that the winding number of an oriented closed
immersed curve f : S 1 # R2 is equal to the total number of counterclockwise turns that the unit
normal vector to the curve makes as we travel along f (S 1 ). The winding number has several
more formal definitions: for instance, it is the equal to the degree of the Gauss map of f (S 1 ). In
particular, the winding number is a regular homotopy invariant. Since the immersed curve “8”
has vanishing winding number (regardless of the orientation) but the standard embedding of S 1
has winding number ±1, these immersions are not regular homotopic.
The Whitney-Graustein theorem states that two immersions f, g : S 1 # R2 are regular ho-
motopic if and only if they have the same winding number [Whi37]. While Whitney’s original
proof can also be found in [Ada93, Chapter 0], we refer to [Gei09] for a particular short (and
enlightening) proof using contact geometry. Summarizing, the space of regular homotopy classes
of immersions S 1 # R2 is in bijective correspondence with Z.
In Section 1.3, we shall be concerned with immersions of S n−1 # Rn . Namely, Smale’s result
is concerned with the space of regular homotopy classes of immersions S n−1 → Rn .
Definition 7. An eversion of the 2-sphere is a regular homotopy from the standard embedding
ι : S 2 → R3 to the antipodal embedding −ι : S 2 → R3 .
The antipodal map on S n is a composition of (−1)n+1 reflections, each changing the sign of
one coordinate in Rn+1 . In particular, the antipodal map in S 2 ⊂ R3 is an orientation reversing
homeomorphism, while the antipodal map on S 1 ⊂ R2 is orientation preserving.
Remark 1.1.16. Suppose we imagine taking the standard embedding ι of the 2-sphere in R3 and
coloring the outside surface red and the inside blue. Since the antipodal map is orientation revers-
ing, −ι(S 2 ) represents the sphere with the outside blue and the inside red. Therefore Definition 7
truthfully corresponds to our intuitive notion of “turning the sphere inside out”.

7
 As we shall see in Section 1.3 below, Smale showed that eversions of the 2-sphere do exist.
Remark 1.1.17. Moving down one dimension, the Whitney-Graustein theorem ensures that there
is no regular homotopy between the standard embedding ι : S 1 → R2 and ι(S 1 ) with the reverse
orientation. This makes Smale’s result all the more surprising.

To show that an eversion does exist, Smale proved that the space of regular homotopy classes of
immersions S 2 # R3 is trivial. In order to understand this result, we now discuss vector bundles.

1.2 Vector bundles and homotopy groups

In Subsection 1.2.1, we introduce vector bundles and provide several examples. In Subsection 1.2.2,
we define homotopy groups and prove several basic results on the subject. Finally in Subsec-
tion 1.2.3, we discuss the homotopy classification of vector bundles.

1.2.1 Vector bundles: definitions and examples

We introduce vector bundles. In our examples, we emphasize the tangent and normal bundles
over a manifold. References include [Tu11, Section 12.3], as well as [MS74, Hus94].

Definition 8. A rank n vector bundle consists of spaces E (the total space) and B (the base
space), a continuous surjection π : E → B (the projection) such that for every x ∈ B each
fiber Ex := π −1 (x) is an n-dimensional real vector space and the following local triviality condition
holds: for every x ∈ B, there is an open neighborhood U ⊂ X containing x and homeomorphism
ϕ : U × Rn → π −1 (U ) such that for all y ∈ U , the map ϕ(y, −) : Rn → Ey is a linear isomorphism
that makes the following diagram commute:

U × Rn
ϕ
/ π −1 (U )

# {
U.

Next, we provide some basic examples of vector bundles.

Example 1.2.1. 1. The trivial rank n vector bundle over a space B is the product B × Rn .
When the base space is understand, we shall sometime denote this vector bundle by n or Rn .
2. The infinite Mobius band is a rank 1 vector bundle over the circle S 1 = [0, 1]/0 ∼ 1: the
total space is M := [0, 1] × R/ ∼, with (0, x) ∼ (1, −x) for all x ∈ R, and the projection
map M → S 1 is defined as π(x, v) = x.

Note that vector bundles are a particular case of “fiber bundles”.

Remark 1.2.2. Let F be a topological space. In Definition 8, if we replace all occurrences of Rn
by F and remove all conditions involving linearity, then we obtain the notion of a “fiber bundle
with fiber F ”. For instance, the (finite) Mobius band is an example of a [0, 1]-bundle, and vector
bundles are “linear” Rn -bundles. In the following sections, we will frequently encounter disc
bundles.

Next, we continue with further examples of vector bundles.

Example 1.2.3. 1. We describe the tangent bundle
F T M of a smooth manifold M . As a set, the
total space T M is the disjoint union T M = p∈M Tp M and the projection is the surjective
map π : T M → M mapping v ∈ Tx M to x. Next, we endow T M with the the structure of
a 2n-manifold and π : T M → M with the structure of a rank n vector bundle. Given an

8
 open set U ⊂ M , note that T U = π −1 (U ) =
F
p∈U Tp M and, for every chart (U, ϕ) of M ,
consider the map

ψ : π −1 (U ) → ϕ(U ) × Rn
v 7→ (ϕ(π(v)), Tπ(v) ϕ(v)).

The atlas A = {(Uα , ϕα )}α of the manifold M therefore gives rise to a collection of
maps {ψα : π −1 (Uα ) → ϕα (Uα )×Rn }α . We declare a set O ⊂ T M to be open if ψα (O∩ π −1 (Uα ))
is open in ϕα (Uα ) ∩ Rn for all α. It can then be checked that thisPdefines a topology on T M ,
n i ∂
that each ψα is a homeomorphism (with inverse (ϕ(p), v) 7→ i=1 v ∂xi |p ) and that the
collection {(π −1 (Uα ), ψα )}α provides a smooth atlas for T M . This collection of maps also
endows π : T M → M with the structure of a rank n vector bundle. We refer to [Tu11,
Section 12] for further details.
2. We describe the normal bundle ν(f ) of an embedding f : M m → Rn . Use h−, −i to denote
the standard inner product on Rn . Since f is an embedding, Tx f : Tx M → Tf (x) Rn is
injective. In order to simplify notations, we think of Tx M as a subspace of Rn ∼ = Tf (x) Rn
(i.e. we write Tx M instead of Tx f (Tx M )). As a set, the total space of ν(f ) is defined as

ν(f ) = {(x, v) ∈ M × Rn | v ∈ (Tx M )⊥ },

while the projection map π : ν(f ) → M is given by projection on the first component.
Next, we endow ν(f ) with the structure of an n-manifold and explain why π : ν(f ) → M
is a rank n − m vector bundle. Fix a chart (U, ϕ) containing a point p ∈ M . Next,
choose m maps vi : U → T U so that for each x ∈ U , the vi (x) are a basis for Tx U ; such vi
exist and can for instance be obtained by setting vi (x) = (Tx ϕ)−1 (bi ), where b1 , . . . , bm
denote the canonical basis of Rm . Taking x = p, we complete our basis of Tp M to a basis
v1 (p), . . . , vm (p), e1 , . . . , en−m of Rn . These vectors still form a basis for Rn ∼
= Tf (x) Rn for
all x ∈ U in a small enough neighborhood U ⊂ U of p. Set ei (x) = ei for x ∈ U 0 . Using
0 0

the Gram-Schmidt process, we can assume that v1 (p), . . . , vm (p), e1 (p), . . . , en−m (p) is an
orthonormal basis. Consider the following map:

ψ : π −1 (U 0 ) → ϕ(U ) × Rn−m
(x, v) 7→ (ϕ(x), v · e1 (x), v · en−m (x)).
Pn−m
Observe that ψ has a smooth inverse given by (y, λ) 7→ (ϕ−1 (y), i=1 λi ei (x)). This
construction can be shown to endow ν(f ) with the structure of an n-manifold and π with
structure of a rank n − m vector bundle.
For simplicity, we chose to work with manifolds embedded in N = Rn . However, the
definition of the normal bundle ν(f ) can be carried out for embeddings ϕ : M → N , with N
a Riemannian manifold. As we shall see in Exercise 1.2.8 below, ν(f ) can also be described
without making use of a metric on N .

Normal bundles will play a major role in Chapter 2, so we record the following deep fact for
later use. A proof can be found in [Bre93, Theorem II.11.14] as well as in [MS74, Theorem 11.1].
Theorem 1.2.4 (Tububular neighbourhood theorem). An embedding f : N n → M m extends to
a codimension 0 embedding ν(f ) → M of the total space of the normal bundle.

Theorem 1.2.4 establishes the existence of tubular neighborhoods around submanifolds.

Remark 1.2.5. For an embedding f : M m → N n , a tubular neighborhood of N in M consists of
a Dn−m -bundle B → V (recall Remark 1.2.2) together with an embedding ψ : B → N extending
the map taking the center of each disc to the corresponding point of M . Theorem 1.2.4 establishes
the existence of a tubular neighborhood: endow N with a Riemannian metric and take B to be
the disc bundle obtained from ν(f ) consisting of vectors of at most unit length.

9
 Next, we describe operations on vector bundles. In brief, for every operation involving vector
spaces, such as the direct sum, the dual vector space, the tensor product, or exterior powers, there
is a corresponding operation for vector bundles over fixed base space.
Example 1.2.6. Given vector bundles πE : E → B and πF : F → B, we describe the vector
bundle π : E ⊕ F → B whose fiber over a point x ∈ B is Ex ⊕ Fx . As a set, the total space of the
direct sum E ⊕ F (sometimes called Whitney sum) is
G
E⊕F = Ex ⊕ Fx ,
x∈B

and the projection is π(v, w) = x for (v, w) ∈ Ex ⊕ Fx . Alternatively, the projection can be
described as π(v, w) = πE (v) = πF (w). We show that E ⊕ F has the structure of a vector bundle.
Assume that E (resp. F ) has rank n (resp. m). Given x ∈ B, since E and B are vector bundles,
there is an open neighborhood U of x as well as homeomorphisms
−1
ϕE : πE (U ) → U × Rn ,
ϕF : πF−1 (U ) → U × Rm .

Let proj2,n (resp. proj2,m ) be the projection of U × Rn (resp. U × Rm ) on the second coordinate.
Define a local trivialisation for E ⊕ F by considering

ϕ : π −1 (U ) → U × (Rn ⊕ Rm )
(v, w) 7→ (πE (v), (proj2,n ◦ϕE (v), proj2,m ◦ϕF (w))).

This shows that E ⊕ F is a rank n + m vector bundle.

Similarly to Example 1.2.6, given vector bundles E → B and F → B, there are vector bun-
dles E ⊗ F → B and Hom(E, F ) with respective fiber over a point x ∈ B given by Ex ⊗R Fx
and HomR (Ex , Fx ). Similarly, if F is a subbundle of E (i.e. if Fx is a subspace of Ex for each x ∈ B),
then one can form the quotient bundle E/F with fiber Ex /Fx .
Given vector bundles πE : E → B and πF : F → B, a homomorphism of vector bundles is a
continuous map f : E → F such that πF f = πE (i.e. f restricts to a map on the fibers) and such
that f : Ex → Fx is linear for each x ∈ B. A homomorphism f : E → F is an isomorphism if there
exists a homomorphism g : F → E such that g ◦ f = idE and f ◦ g = idF .
Exercise 1.2.7. Let f : E → B be a homomorphism of vector bundles. Show that f is an
isomorphism if and only if f restricts to a linear isomorphism on each fiber.

The next exercise shows that the normal bundle of an embedding M → N can be defined
without endowing N with a Riemannian metric.
Exercise 1.2.8. Let Ff : M → N be an embedding. Use T N |f (M ) → M to denote the vector bundle
with underlying set x∈M Tx f (Tx M ) (more formally, this is the pullback bundle f ∗ (T N ), see
Subsection 1.2.3 below). Show that if N is endowed with a Riemanian metric (or just take N = Rn
with the standard inner product), then the quotient bundle T N |f (M ) /T M is isomorphic to the
vector bundle ν(f ) described in Example 1.2.3.

A rank n vector bundle over B is trivial if it is isomorphic to the trivial rank n bundle B × Rn .
A section of a vector bundle π : E → B is a map s : B → E such that π ◦ s = idB . Reformulating,
for each x ∈ B, the definition of a section ensures that s(x) lies in Ex . Sections of the tangent
bundle are called vector fields.
The next remark presents a useful criterion to prove that a vector bundle is trivial.
Lemma 1.2.9. A rank n vector bundle π : E → B is trivial if and only if there sections s1 , . . . , sn
such that for each x ∈ B, the vectors s1 (x), . . . , sn (x) are linearly independent.

10
Proof. Assume that the bundle E → B is trivial and choose an isomorphism Φ : B × Rn → E.
Define ri : B → B × Rn as ri (x) = (x, ei ), where ei is the i-th canonical basis vector of Rn . For
r Φ
i = 1, . . . , n the maps si : B →i B × Rn → E provide the desired sections. To prove the converse,
assume that we have n pointwise linearly independent sections si : B → Rn ; our goal is to define
a bundle isomorphism Φ : E → B × Rn . Given v ∈ E, set x =P π(v) and note that s1 (x), . . . , sn (x)
n i i
form a basis of Ex . As a consequence, we can write v = i=1 v si (x) for some v ∈ R and
1 n
define Φ(v) := (x, v , . . . , v ). It is immediate that Φ is a bundle homomorphism that restricts to
a linear isomorphism on the fibers. Exercise 1.2.7 now implies that Φ is a bundlePisomorphism
n
(alternatively, it is easy to show that the inverse of Φ is given by (x, v 1 , . . . , v n ) 7→ i=1 v i si (x)).
This concludes the proof of the lemma.

We conclude this subsection with a concrete example of these notions.

Example 1.2.10. Show that the normal bundle of the standard embedding ι : S n → Rn+1 is
trivial. This can be done either by constructing an explicit bundle isomorphism ν(ι) → S n × R
or, thanks to Lemma 1.2.9, by constructing a nowhere vanishing section s : S n → ν(ι).

1.2.2 Some homotopy theory

This subsection provides a very brief introduction to higher homotopy groups of a space. We
assume familiarity with CW-complexes and emphasis is put on the results and methods that
will be used in surgery theory. We prove the long exact sequence of the pair and sate (without
proof) Whitehead’s theorem as well as cellular approximation. These results provide means and
motivation for the idea of “killing a homotopy class”. References include [Hat02, Section 4].
Given spaces X and Y endowed with fixed basepoints x0 ∈ X and y0 ∈ Y , a map f : X → Y is
based (or pointed) if f (x0 ) = y0 . A homotopy ft : X → Y is based if ft (x0 ) = y0 for each t ∈ [0, 1].
We use [(X, x0 ), (Y, y0 )] to denote the set of based homotopy classes of based maps from (X, x0 )
to (Y, y0 ). Fix a basepoint s0 of S n .
Definition 9. The n-th homotopy group of a based space (X, x0 ) consists of based homotopy
classes of maps (S n , s0 ) → (X, x0 ), in other words,

πn (X, x0 ) := [(S n , s0 ), (X, x0 )].

The group law is defined as follows: for f, g : S n → X, the sum f + g is the composition
c
n
S → S n ∨ S n → X, where c collapses the equatorial S n−1 ⊂ S n to a point, and we choose
the basepoint of S n to lie in this S n−1 . It can be shown that (πn (X, x0 ), +) is a group [Hat02,
Section 4.1], where a map represents the zero element if and only if it is homotopic to the constant
map at x0 . For n ≥ 2, the higher homotopy groups πn (X, x0 ) are abelian [Hat02, Section 4.1],
while for n = 1 one recovers the fundamental group.
A based map f : (X, x0 ) → (Y, y0 ) gives rise to an induced map f∗ : πn (X, x0 ) → πn (Y, y0 ) via
the well defined assignment [ϕ] 7→ [f ◦ ϕ].

Exercise 1.2.11. Show that the induced maps satisfy (g ◦ f )∗ = g∗ ◦ f∗ . Deduce that if (X, y0 )
and (Y, y0 ) are homotopy equivalent (in the based sense), then πn (X, x0 ) and πn (Y, y0 ) are iso-
morphic. This implies that the homotopy groups of a contractible space are trivial.

Contrarily to the homology groups, the homotopy groups of a space are very difficult to com-
pute: for instance the homotopy groups of spheres are not known in general.
The next result assumes some familiarity with covering space theory.
Proposition 1.2.12. A covering space projection p : (X,
b xb0 ) → (X, x0 ) induces isomorphisms
∼
=
b0 ) → πn (X, x0 ) for n ≥ 2.
p∗ : πn (X, x
b

11
Proof. We start with surjectivity. A based map f : (Y, y0 ) → (X, x0 ) lifts to (X, b xb0 ) if and only if
n
f∗ (π1 (Y, y0 )) ⊂ p∗ (π1 (X, x
b b0 )) [Hat02, Proposition 1.33]. Since S is simply connected for n ≥ 2,
any based map S n → X lifts to X, b proving the surjectivity of p∗ . Next, we prove injectivity.
If ht : Y → X is a homotopy and h0 is a lift of h0 , then there exists a unique homotopy e
e ht of h0
that lifts ht [Hat02, Proposition 1.30]. Now suppose that f, g : (S n , s0 ) → (X, b xb0 ) are two based
maps such that p ◦ f, p ◦ g : S n → X are based homotopic. Let ht be the homotopy between
h0 = p ◦ f and h1 = p ◦ g. We now apply the aforementioned homotopy lifting result with e h0 = f :
the result is a homotopy e ht from f to a lift fe1 of p ◦ g. Since all maps are based, the uniqueness
of lifts [Hat02, Proposition 1.34], implies that fe1 = g. This concludes the proof of the injectivity
of p∗ and thus the proof of the proposition.

We deduce some computations of higher homotopy groups (all maps are based, although we
sometimes omit the base points from the notation).
Example 1.2.13. Here are some applications of Proposition 1.2.12:

1. The higher homotopy groups of S 1 are trivial since the universal cover of S 1 is R. The
higher homotopy groups of the n-torus Tn are trivial since the universal cover of Tn is Rn .
2. For g ≥ 2, the higher homotopy groups of the closed genus g surface Σg all vanish: the
universal cover of Σg is the hyperbolic plane.
3. The universal cover of SO(3) is SU (2) so πk (SO(3)) ∼= πk (SU (2)) for k ≥ 2. Since SU (2)
is homeomorphic to S 3 and since we will see in Corollary 1.2.20 below that π2 (S 3 ) = 0, we
deduce that π2 (SO(3)) = 0. This also follows from the fact that π2 (G) = 0 for any compact
connected Lie group.

A powerful computational tool in homology theory is the long exact sequence of a pair. We
now describe the corresponding exact sequence for homotopy groups. Recall that a pair of spaces
is a pair (X, A), where X is a space and A ⊂ X. If X is based, then we assume that A contains
the basepoint. A (based) map of pairs f : (X, A) → (Y, B) is a (based) map f : X → Y such
that f (A) ⊂ B. Recall that two maps f, g : (X, A) → (Y, B) are (based) homotopic rel A if
they are homotopic by a (based) homotopy ft : X → Y such that ft |A is independent of t. We
use [(X, A, x0 ), (Y, B, y0 )] to denote the corresponding set of based rel A homotopy classes of
based maps. Fix a basepoint s0 ∈ Dn that lies in ∂Dn .
Definition 10. The relative homotopy group of a based pair (X, A, x0 ) consists of the rel ∂Dn
homotopy classes of maps (Dn , ∂Dn , s0 ) → (X, A, x0 ):
πn (X, A, x0 ) := [(Dn , ∂Dn , s0 ), (X, A, x0 )].

The sum is once again obtained via the map c : Dn → Dn ∨ Dn that collapses the equato-
rial Dn−1 to a point. Next, we describe the maps that appear in the exact sequence of the pair.
Given a based pair (X, A, x0 ), the inclusion A → X induces a map i : πn (A, x0 ) → πn (X, x0 )
for each n. Next, thinking of πn (X, x0 ) as [(Dn , ∂Dn , x0 ), (X, x0 , x0 )], the inclusion {x0 } → A
induces a canonical map j : πn (X, x0 ) → πn (X, A, x0 ). Finally, the connecting homomorphism
∂ : πn (X, A, x0 ) → πn−1 (A, x0 ) is defined by sending a map f : (Dn , ∂Dn , x0 ) → (X, A, x0 ) to the
restriction f |∂Dn : S n−1 → A. It can be checked that ∂ is well defined.
The next theorem describes the long exact sequence of a pair (X, A). 1
Theorem 1.2.14. Given a based pair (X, A, x0 ), there is a long exact sequence
i j ∂
. . . → πn (A, x0 ) → πn (X, x0 ) → πn (X, A, x0 ) → πn−1 (A, x0 ) → . . .
1 Recall that given a collection of groups Gn and group homomorphisms fn : Gn → Gn−1 , the sequence
fn fn−1
. . . Gn → Gn−1 → Gn−2 → . . .
is exact if ker(fn−1 ) = im(fn ) for each n.

12
Proof. To prove exactness, we must establish the equalities im(∂) = ker(i) and im(i∗ ) = ker(j)
and im(j) = ker(∂). In what follows, we will show the first two equalities, but refer to [Hat02,
proof of Theorem 4.3] for the last equality. To abbreviate notations, we will often omit basepoints.
We first show that im(∂) = ker(i). To show the inclusion im(∂) ⊂ ker(i), we start with a
relative map f : (Dn , ∂Dn ) → (X, A) representing an element [f ] of πn (X, A), we must show
that (i ◦ ∂)[f ] is zero, i.e. that i ◦ ∂(f ) is nullhomotopic. By definition, ∂([f ]) is represented
by the map f | := f |∂Dn : S n−1 → A and therefore i ◦ ∂ is represented by the the composition
i ◦ f | : S n → X. By definition i ◦ f | extends over Dn and is therefore nullhomotopic, as desired.
Next, we prove the reverse inclusion, namely ker(i) ⊂ im(∂). Let f : (Dn−1 , S n−2 ) → (A, x0 )
be a representative of πn−1 (A) such that i([f ]) is zero in πn−1 (X). In other words, we know that
the composition i ◦ f : (Dn−1 , S n−2 ) → (A, x0 ) → (X, x0 ) is nullhomotopic. Thus, we obtain a
rel ∂Dn−1 homotopy H : Dn−1 × [0, 1] → X between i ◦ f and the constant map cx0 . Since this
map is rel S n−1 , each H(−, t) must map S n−1 into {x0 }. Viewing Dn as Dn−1 × [0, 1], it only
remains to show H induces a map of pairs H : (Dn , ∂Dn ) → (X, A) such that ∂([H]) = [f ]. Note
that ∂Dn = ∂(Dn−1 × [0, 1]) can be viewed as (S n−2 × [0, 1]) ∪ (Dn−1 × {0, 1}). Since we have
a homotopy rel S n−2 between the composition i ◦ f and the constant map cx0 , we see that H
maps (S n−2 × I) ∪ (Dn−1 × {1}) to {x0 }, while the restriction of H to Dn−1 × {0} is precisely f .
We have therefore proved that ∂([H]) = [f ], concluding the proof of the equality im(∂) = ker(i).
The equality im(i) = ker(j), will follow from the following “compression criterion”.
Claim. A based map f : (Dn , S n−1 ) → (X, A) represents zero in πn (X, A) if and only if it is
homotopic rel S n−1 to a map g : (Dn , S n−1 ) → (X, A) with image contained in A.

Proof. We start with the “only if” direction, namely given a map f : (Dn , ∂Dn ) → (X, A) repre-
senting zero in πn (X, A), and construct a map g with the desired properties. We know that there
is a rel ∂Dn homotopy F : Dn × [0, 1] → X between f and the constant map cx0 . In other words, F
is such that F |Dn ×{0} = f and F |Dn ×{1} = cx0 and F |S n−1 ×[0,1] ⊂ A. We set f1 := F | : Dn '
Dn × {1} ∪ S n−1 × I → A. Since the image of f1 lies in A, it is enough to show that f is homotopic
to g. The homotopy ft is given by the restriction of F to Dn ' Dn × {t} ∪ S n−1 × [0, t] → X: by
construction, we have f1 = g and f0 (which is defined on Dn × {0}) is equal to f .
Next, we prove the converse. Assume that f is homotopic rel S n−1 to a map g with image
in A. We show that g represents zero in πn (X, A), i.e. that it is nullhomotopic rel ∂Dn−1 .
As g(Dn ) ⊂ A, we have a map g : (Dn , S n−1 ) → (A, A). This map is nullhomotopic since Dn is
contractible. This concludes the proof of the claim.

We now prove the equality im(i) = ker(j). We start by showing im(i) ⊂ ker(j). Let
f : (Dn , ∂Dn ) → (A, x0 ) → (X, x0 ) be a representative of an element in im(i) ⊂ πn (X). By
definition, (j ◦ i)(f ) : Dn → X has image in A. The compression lemma therefore ensures that this
element is zero in πn (X, A). We conclude by proving the reverse inclusion, namely ker(j) ⊂ im(i).
Assume that f : (Dn , ∂Dn ) → (X, x0 ) maps to zero in the relative group πn (X, A). By the com-
pression lemma, this means that f is homotopic rel ∂Dn to a map g with image in A. But as
such a g lies in im(i), this establishes the equality im(i) = ker(j) and concludes the proof of the
theorem.

The mapping cylinder of a a map f : X → Y is defined as Mf = X ×[0, 1]∪f Y . More explicitly,

this space is (X × [0, 1]) t Y / ∼, where the equivalence relation is (x, 1) ∼ f (x) for every x ∈ X.
It is known that Mf deformation retracts onto Y . The map f : X → Y factors as f = r ◦ i,
where r : Mf → Y is the aforementioned homotopy equivalence and i : X → Mf is the canonical
inclusion x 7→ (x, 0). Since X is a subspace of Mf , we may consider the pair (Mf , X).
Definition 11. The homotopy group of a map f : X → Y is defined as

πn (f ) := πn (Mf , X).

13
 Since these homotopy groups will be frequently used in Chapter 2 below, we record some facts
for later reference.
Remark 1.2.15. Let f : X → Y be a map.

1. Elements of πn (f ) can be represented by maps (Dn , S n−1 ) → (Mf , X). Since Mf is homo-
topy equivalent to Y , it follows that elements of πn (f ) can be represented by a pair of maps
S n−1 → X and Dn → Y such that the following diagram commutes:

S n−1 /X

ι f
 
Dn / Y.

2. Given a map X → Y , the mapping cylinder construction produces a “long exact sequence of
the pair (Y, X) ” even when X is not a subspace of Y . Indeed, replacing Y by the mapping
cylinder Mf and applying Theorem 1.2.14 leads to the following exact sequence:

. . . → πn (X) → πn (Y ) → πn (f ) → . . . .

3. A map f is n-connected if it induces an isomorphism on the first n − 1 homotopy groups

and a surjection on πn . An application of the long exact sequence (recall Theorem 1.2.14)
shows that f is n-connected if and only if πk (f ) = 0 for all k ≤ n.

Recall that a map f : X → Y between CW-complexes is cellular if f (X n ) ⊂ Y n . We now state

two important theorems in homotopy theory. We refer to [Hat02, Theorem 4.5 and 4.8] for proofs,
but present some applications instead.
Theorem 1.2.16. Let X and Y be connected CW complexes.

1. Whitehead’s theorem: If f : X → Y induces an isomorphism πn (X) → πn (Y ) for each n,

then f is a homotopy equivalence.

2. Cellular approximation: every map f : X → Y is homotopic to a cellular map.

We deduce that attaching cells of dimension n + 1 or greater does not affect the first n − 1
homotopy groups.
Corollary 1.2.17. If X is obtained from A by attaching cells of dimension n + 1 or greater, then
the inclusion A → X is n-connected. In particular, (X, X n ) is n-connected.

Proof. Using the long exact sequence, we must show that (X, A) is n-connected i.e. that πi (X, A) = 0
for i < n. Let f : (Di , ∂Di ) → (X, A) be a representative of πi (X, A). Using the (relative version
of) cellular approximation, we can assume that f (Di ) ⊂ X i . Since i ≤ n, the assumption on the
pair (X, A) implies that X i ⊂ A. Applying the compression criterion, we deduce that [f ] = 0
in πn (X, A). This concludes the proof of the corollary.

Corollary 1.2.17 is useful to “kill” a given homotopy classes.

Corollary 1.2.18. Let X be a CW complex and let [f ] ∈ πn (X). There exists a space X 0
containing X such that πi (X 0 ) ∼
= πi (X) for i ≤ n − 1 and πn (X 0 ) ∼
= πn (X)/h[f ]i.

Proof. Set X 0 = X ∪f en+1 and use the Corollary 1.2.17 to deduce that πi (X 0 ) ∼ = πi (X) for
i ≤ n − 1 and that i : πn (X) → πn (X 0 ) is surjective. To see that [f ] belongs to ker(i), note that
that in X 0 , the map i◦f extends over the n+1-cell. Some additional work shows that ker(i) = hωi.
This concludes the proof of the corollary.

14
 Next, show how to kill relative classes.
Corollary 1.2.19. Let f : X → Y be a map of CW complexes and let ω ∈ πn+1 (f ). There exists
a CW complex X 0 containing X such that πi (X 0 ) ∼
= πi (X) for i ≤ n and πn+1 (f 0 ) ∼
= πn+1 (f )/hωi.

Proof. Represent ω by a pair (q, Q) with q : S n → X and Q : Dn+1 → Y . Glue a k + 1-cell to X

along q yielding X 0 and set f 0 = f ∪ Q. Since we attached a cell of dimension n + 1, we see
that πi (X 0 , X) = 0 for i ≤ n. To show that πi (f ) = πi (f 0 ) for i ≤ n, consider the following
commutative diagram in which the vertical maps are inclusion induced:

/ πi (X) f∗
/ πi (Y ) / πi (f ) / πi−1 (X) f∗
/ πi (Y ) /

ι = ι =
 f∗0    f∗0 
/ πi (X 0 ) / πi (Y ) / πi (f 0 ) / πi−1 (X 0 ) / πi (Y ) /.

For i < n − 1, we know that the ι maps are isomorphisms and so we can apply the five lemma.
For i = n, the right hand side ι map is an isomorphism while the left hand side one is only a
surjection. This is nevertheless enough to once again apply the five lemma, concluding the proof
that πi (X 0 ) ∼
= πi (X) for i ≤ n.
Using cellular approximation for pairs, we see that πn+1 (f ) → πn+1 (f 0 ) is surjective. The
fact that ω maps to zero in πk+1 (f 0 ) follows from the compression lemma and because the map
Q : Dn+1 → Y factors through X 0 = X ∪q Dn+1 . Some additional work shows that ker(i) = hωi.
This concludes the proof of the corollary.

We conclude this subsection by proving that πk (S n ) = 0 for k < n.

Corollary 1.2.20. For k < n, we have πk (S n ) = 0.

Proof. Give S n and S k are given their usual CW structures with 0-cells as basepoints. Using
cellular approximation, every basepoint-preserving map S n → S k can be homotoped fixing the
basepoint to be cellular and hence constant if k < n. This concludes the proof of the corollary.

1.2.3 Vector bundles: the homotopy classification

The goal in this subsection is state the classification of vector bundles over a fixed base space. We
choose to only sketch the proof of the main results: the goal of this course is after all to provide
an introduction to surgery theory; proving these results in details would be an ideal topic for a
course on vector bundles and characteristic classes. Additionally, we also use the classification of
vector bundles to improve our understanding of the normal and tangent bundles.
We start by discussing the pullback construction. The pullback of a vector bundle π : E → B
along a map f : X → B is the vector bundle f ∗ (E) → X with total space

f ∗ (E) = {(x, e) ∈ X × E | f (x) = π(e)}

and bundle projection (x, e) 7→ x. We will not check that f ∗ (E) is a vector bundle but instead
observe that the fiber above x ∈ X is {x} × Ef (x) . The vector space structure on this fiber is
induced by the one on Ef (x) . Note furthermore that f : X → B lifts to a map fe: f ∗ (E) → E by
setting fe(x, e) = e and one obtains the following commutative diagram:

f ∗ (E)
fe
/E

π
 
X
f
/ B.

15
It follows that fe is linear on the fibers and, conversely, it is not difficult to show that f ∗ (E) → X
satisfies the corresponding universal property: if p : F → X is a vector bundle and if ϕ : F → E is
fiberwise linear and satisfies f ◦ p = π ◦ ϕ, then p : F → X is isomorphic to f ∗ (E) → X.
The next exercise shows that we have already encountered a pullback bundle.
Exercise 1.2.21. Given an embedding f : M → N , show that the restricted bundle T N |f (M ) →
M described in Exercise 1.2.8 is isomorphic to the pullback bundle f ∗ (T N ) → M . Use Exer-
cise 1.2.8 to deduce that T M ⊕ ν(f ) ∼
= f ∗ (T N ).

The next exercise establishes some basic properties of the pullback construction.
Exercise 1.2.22. Prove the vector bundle isomorphisms id∗ (E) ∼
= E and (g◦f )∗ (E) ∼
= f ∗ (g ∗ (E)).
Show that the pullback of a trivial bundle is trivial.

We now move towards the classification of vector bundles.

Definition 12. A universal vector bundle is a rank n vector bundle γn → BO(n) that satisfies
the following property: for every rank n vector bundle E → X over a CW complex X, there is a
unique (up to homotopy) map f : X → BO(n) such that E ∼ = f ∗ (BO(n)).

Temporarily assuming the existence of classifying spaces, we use Definition 12 to provide a

homotopy theoretic interpretation of the set Vectn (X) of isomorphism classes of rank n vector
bundles over a CW complex X.
Theorem 1.2.23. For any CW complex X, the pullback construction determines a bijection
'
[X, BO(n)] → Vectn (X).

Proof. We need only check that the assignment is well defined: the bijective correspondence
immediately follows from the definition of a universal bundle, recall Definition 12. To check that
that the assignment is well defined, we need the following fact: if E → B is a vector bundle and if
f, g : X → B are homotopic, then f ∗ (E) and g ∗ (E) are isomorphic. A proof can be found in [Hat,
Therorem 1.6]; we only mention the key idea. Given a homotopy F : X ×I → B, the restrictions of
F ∗ (E) over X × {0} and X × {1} are f0∗ (E) and f1∗ (E). The result then follows from the following
statement whose proof can be found in [Hat, Proposition 1.7]: the restriction of a vector bundle
E → X × I over X × {0} and X × {1} are isomorphic if X is a manifold. This concludes the proof
sketch.

It is worth mentioning that both Definition 12 and Theorem 1.2.23 hold whenever X is a
paracompact space. We refer to [Hat, Appendix of Section 1.2] for a discussion of paracompactness.
Next, we sketch the proof of the existence of universal vector bundles.
Theorem 1.2.24. A universal rank n vector bundle exists.

Proof. We construct the bundle γn → BO(n) but do not show that it is universal; we refer the
interested reader to [Hat, Theorem 1.16] for a proof. Let Grn (Rn+k ) denote the space of n-
dimensional vector subspaces of Rn+k . Fix an integer n and observe that there is a canonical
inclusion Grn (Rn+k ) → Grn (Rn+k+1 ): an n-dimensional vector subspace of Rn+k can also be
viewed as a vector subspace of Rn+k+1 . The infinite Grassmannian BO(n) := Grn (R∞ ) is the
direct limit (or union) over k. Next, the tautological bundle over Grn (Rn+k ) has total space

γn,k := {(V, x) ∈ Grn (Rn+k ) × Rn+k | x ∈ V }.

and projection map π(V, x) = V . Some work shows that this a rank n-vector bundle [Hat,
Lemma 1.15]. Setting γn := limk γn,k provides the desired rank n bundle over BO(n). It can once
again be shown that γn → Grn is a vector bundle [Hat, Lemma 1.15], and the proof of universality
can be found in [Hat, Theorem 1.16]. This concludes our proof sketch.

16
 As a corollary of Theorem 1.2.23, we can classify vector bundles over contractible spaces.
Corollary 1.2.25. Any vector bundle over a contractible space is trivial.

Proof. This is immediate: if X ' {x0 } is contractible, then Vectn (X) corresponds bijectively to
the set [X, BO(n)] ' [x0 , BO(n)] ' {∗}.

Corollary 1.2.25 provides some further insight into the normal bundle of submanifolds of Rn .
Example 1.2.26. We investigate the relation between the tangent bundle and the normal bundle.
1. For any embedding f : M → Rn , the direct sum T M ⊕ ν(f ) is trivial: this follows from the
isomorphism f ∗ (T Rn ) ∼
= T M ⊕ ν(f ) (recall Exercise 1.2.21), from the fact that any bundle
over Rn is trivial (by Corollary 1.2.25) and since the pullback of a trivial bundle is trivial
(recall Exercise 1.2.22).
∼ r+1 , where k denotes the rank k
2. The tangent bundle of the sphere S r satisfies T S r ⊕  =
r r r+1
trivial bundle over S . To see this, embed S in R via the standard embedding f . Since
we now know that T S r ⊕ ν(f ) is trivial, the conclusion follows from Example 1.2.10:
r+1 = T S r ⊕ ν(f ) = T S r ⊕ 1 .

3. Two vector bundles E1 and E2 over B are stably isomorphic if E1 ⊕ m is isomorphic

to E2 ⊕ n ; this time k denotes the trivial rank n vector bundle over B. For instance,
we saw above that T S r is stably trivial. Use Vectstab (B) to denote the set of stable isomor-
phism classes of vector bundles over B and set BO := limn BO(n), just as in Theorem 1.2.23,
it can be shown that there is an isomorphism
Vectstab (B) ∼
= [B, BO].
This transition to homotopy allows for convenient reformulations of naturally occurring
questions: for instance, a stably trivial rank n vector bundle over S r defines an element in
ker(πr (BO(n)) → πr (BO)). If this kernel happens to be trivial, then every stably trivial
rank n vector bundle over S r is trivial.
We conclude this subection by using Theorem 1.2.23 to study vector bundles over spheres.
Example 1.2.27. Using Theorem 1.2.23, we know that rank n vector bundles over the k-sphere.
are classified by πk (BO(n)), and this latter group is known to be πk−1 (O(n)). 2 For instance, we
deduce that rank 1 vector bundles over S 1 are classified by π0 (O(1)) ∼ = Z2 : the trivial bundle and
the infinite Mobius band of Example 1.2.1 are the only examples.
We conclude by sketching a more geometric proof of the isomorphism Vectk (S n ) ∼ = πk−1 (O(n)).
Any bundle E over S n restricts to two bundles E+ and E− over the upper and lower hemi-
n n
spheres D+ and D− . Using Corollary 1.2.25, these bundles are trivial. The vector bundle E → S n
n
is recovered by gluing D+ × Rk and D− n
× Rk along the equatorial S n−1 via a clutching map
k−1
f: S → GLn (R). Conversely, a clutching map f can be used to obtain a bundle S n → E
by gluing two bundles Dn × Rk along the equatorial S n−1 . It turns out that this construction
only depends on the homotopy class of f . The isomorphism Vectk (S n ) ∼ = πk−1 (O(n)) follows
since GLn (R) deformation retracts onto O(n).

1.3 Turning the sphere inside out

In this section, we return to eversions. Namely relying on the deep Smale-Hirsch theorem, we
show that regular homotopy classes of immersions S n # Rn+1 correspond bijectively to elements
of πn (SO(n + 1)). Setting n = 2, we deduce that all immersions S 2 # R3 are regular homotopic
and, in particular, eversions exist.
2 This follows from the long exact sequence of the fibration applied to the universal O(n)-bundle over BO(n).

17
 Recall that two immersions f0 , f1 : M # N are regularly homotopic if there is a homotopy
ft : M → N between f0 and f1 such that each ft is an immersion. We use π0 (Imm(M, N )) to
denote the set of regular homotopy classes of immersions M → N . We also write π0 (Mono(M, N ))
for the set of homotopy classes of bundle monomorphisms T M → T N .
The following theorem will be referred to as the Smale-Hirsch theorem.
Theorem 1.3.1. Let M be an m-manifold and let N be an n-manifold. If m < n, then the
differential induces a bijection between regular homotopy classes of immersions M → N and
homotopy classes of bundle monomorphisms T M → T N :
'
T : π0 (Imm(M, N )) → π0 (Mono(T M, T N )).

Note that Smale originally classified regular homotopy classes of immersions S k # Rn [Sma59]
(establishing the existence of an eversion as a corollary), while Hirsch generalized this result to
arbitrary manifolds [Hir59]. We do not present Hirsch’s proof but only show how it gives rise
to the classification of regular homotopy classes of immersions S n # Rn+1 . The exposition is
inspired by [Now].
Theorem 1.3.2. Regular homotopy classes of immersions S n # Rn+1 correspond bijectivetly to
elements of πn (SO(n + 1)).

Proof. Using the Smale-Hirsch theorem (Theorem 1.3.1), we need only establish a bijective cor-
respondence between π0 (Mono(T S n , T Rn+1 )) and πn (SO(n + 1)) = [S n , SO(n + 1)]. First, note
that the space Mono(T S n , T Rn+1 ) is homotopy equivalent to

X := {f : T S n → Rn+1 | f is a linear monomorphism on each fiber of T S n }.

To see this, contract the first Rn+1 factor of T Rn+1 ∼

= Rn+1 × Rn+1 to a point. Note that
the space X can also be understood as Mono(T S , {p} × Rn+1 ) where {p} × Rn+1 is the trivial
n

rank n + 1 vector bundle over {p}. Next, we argue that X is homotopy equivalent to the space

Y := Map(S n , GL+
n+1 (R)).

Fix a basepoint b0 : T S n → Rn+1 of X. In order to define a map Φ : X → Y , we must

send an element b : T S n → Rn+1 of X to a map Φ(b) : S n → GL+ n+1 (R). By definition of X,
for any x ∈ S n , we have linear injections b0 (x), b(x) : Tx S n ,→ Rn+1 . Define Φ(b)(x) as the
unique matrix A ∈ GL+ n+1 (R) such that Ab0 (x) = b(x) and A maps a unit vector perpen-
dicular to b0 (x)(Tx S n ) into a unit vector perpendicular to b(x)(Tx S n ). In order to define the
homotopy inverse Ψ : Y → X of Φ, given an element f : S n → GL+ n+1 (R) of Y , we define a
map Ψ(f ) : T S n → Rn+1 by mapping v ∈ Tx S n to the element Ψ(f )(v) = f (x) ◦ b0 (x)(v) of Rn+1 .
It can be checked that Φ and Ψ are homotopy inverses.
Additional work shows that these bijective correspondences descend to (regular) homotopy
classes, establishing a bijection between Mono(T S n , T Rn+1 ) and πn (GL+ n +
n+1 (R)) = [S , GLn+1 (R)].
+
The Gram-Schmidt process implies that GLn+1 (R) deformation retracts onto SO(n+1) and there-
fore πn (GL+ ∼
n+1 (R)) = πn (SO(n + 1)). This concludes the proof of the theorem.

Setting n = 1, we obtain the Whitney-Graustein theorem (recall Example 1.1.15) while set-
ting n = 2, we obtain existence of a sphere eversion.
Corollary 1.3.3. The set of regular homotopy classes of immersions S 2 # R3 is trivial. In
particular the standardly embedded sphere S 2 ⊂ R3 can be turned inside out.

Proof. The first assertion follows from Theorem 1.3.2 and Example 1.2.13 which showed that π2 (SO(3))
is trivial. Since an eversion is a regular homotopy from the standard embedding ι : S 2 → R3 to
the antipodal embedding −ι : S 2 → R3 , the second assertion follows immediately: we know that
all immersions S 2 # R3 are regularly homotopic. This concludes the proof of the corollary.

18
Chapter 2

Surgery theory

One of the motivations of surgery theory is to decide whether two manifolds are diffeomorphic.
Before describing this classification program, we introduce some terminology. If M and N are two
n-manifolds, then a cobordism is an n + 1-manifold whose boundary is ∂W = M t N . A cobordism
(W, M, N ) is an h-cobordism if the inclusions M → W and N → W are homotopy equivalences.
Here are the main steps in the surgery program to decide whether two closed manifolds M and N
are diffeomorphic:

1. construct a homotopy equivalence f : M → N ;

2. construct a cobordism (W ; M, N ) and a map F : W : N → N ×[0, 1] with F |M = f, F |N = idN ;
3. modify W and F relative boundary so that F becomes a homotopy equivalence.

If M n and N n are simply connected and n ≥ 5, then this program ensures that M and N
are diffeomorphic: Smale’s h-cobordism theorem states that every h-cobordism W is diffeomor-
phic to M × [0, 1] [Sma62]. In general, the s-cobordism theorem states that W is diffeomorphic
to M × [0, 1] if and only if W is an s-cobordism (i.e. the inclusions M → W and N → W are
simple homotopy equivalences). An account of such results (and how they imply the generalized
Poincaré conjecture) can be found in [L0̈2].
The aim of this chapter is to focus on the third step of the surgery program and study the
question whether a map f : M → X “be modified into” a homotopy equivalence. The “modifi-
cation” is called “surgery” and is discussed in Section 2.1; we also define normal maps and show
that surgery below the middle dimension is always possible. In Section 2.2, we restrict to even-
dimensional manifolds and discuss the obstruction to carrying out the third step of the surgery
program.

2.1 Surgery below the middle dimension

For simplicity, we will assume that M and X are manifolds, with M closed. 1 We wish to
know whether a map f : M → X can be modified into a homotopy equivalence. By Whitehead’s
theorem (Theorem 1.2.16), it is sufficient to kill all homotopy groups πn (f ). While Corollary 1.2.18
described a procedure to kill these homotopy groups by attaching cells, the outcome need not be
a manifold. In a nutshell, surgery is the analogue of “attaching cells”, but within the category of
manifolds.
Definition 13. Let M n be a closed manifold and let φ : S r × Dn−r → M be an embedding. The
operation of removing the interior of the image of φ and attaching Dr+1 × ∂Dn−r is called surgery.
1 In general, M need not be closed and X is assumed to be a Poincaré complex.

19
The resulting manifold M 0 is called the effect of the surgery:

M 0 = M \ φ(S r × Dn−r ) ∪φ Dr+1 × ∂Dn−r .

The trace of the surgery is the following cobordism between M and M 0 :

W := M × [0, 1] ∪ϕ Dr+1 × Dn−r .

We say that W is obtained from M × [0, 1] by adding an n-dimensional (r + 1)-handle via the
attaching map φ.

Let us illustrate Definition 13 with a low dimensional example.

Example 2.1.1. We describe 0-surgery and 1-surgery on a surface M 2 . The effect of 0-surgery
is “to add a tube to M ”: namely, we remove two disks from M (i.e. S 0 × D2 ) and add the
tube D1 × S 1 . The effect of 1-surgery is to remove a cylinder S 1 × D1 and to add two disks
S 0 × D2 . Observe that if we do 1-surgery on a neighborhood of the meridian of the tube added
by a 0-surgery, then we recover the initial surface.

The next sections will show that surgery is really the “manifold analogue” of attaching cells to
a CW complex. In Subsection 2.1.1, we further discuss surgery and introduce normal maps, while
in Subsection 2.1.3 in discuss surgery below the middle dimension.

2.1.1 Surgery and its effect on homotopy groups

The aim of this subsection is to improve our understanding of surgery and its effect on (relative)
homotopy groups. We also motivate and introduce the notion of a normal map. References
include [L0̈2, Ran02, Wal70, GS99, Wal16, CLMa].
An embedding ϕ : S r × Dn−r → M n determines an embedding ϕ0 : S r × {0} → M . In order
to discuss the degree to which ϕ0 determines ϕ, we introduce some terminology: a framing of an
embedding ϕ0 : S r → M with trivial normal bundle ν(ϕ0 ) consists of an identification of ν(ϕ0 )
with S r × Rn−r .
We establish that a framed embedding S r → M n give rise to embeddings S r × Dn−r → M .
Proposition 2.1.2. Let M be an n-manifold. A framed embedding ϕ0 : S r ,→ M n determines an
embedding ϕ : S r × Dn−r ,→ M such that ϕ × 0 = ϕ0 . In particular, in order to perform a surgery
on M , we need only specify a framed embedding ϕ0 : S r ,→ M .

Proof. This is essentially Remark 1.2.5, whose content we recall. The tubular neighborhood the-
orem (Theorem 1.2.4) ensures the existence of an embedding ν(ϕ0 ) ,→ X. Combining this with
the choice of a framing ν(ϕ0 ) ∼
= S r × Rn−r therefore produces an embedding S r × Rn−r ,→ X.
The proposition now follows by taking the unit disc bundle.

In fact, the embedding ϕ : S r × Dn−r ,→ M described in Proposition 1.2.11 is determined up

to isotopy by the data of ϕ and the framing f : if ϕ0 : S r × Dn−r → M n is another embedding with
ϕ0 |∂Dr ×{0} = ϕ0 and T ϕ0 |0×T0 Dn−k = f −1 , then ϕ0 is isotopic to ϕ. Since the trace of the surgery
(M × [0, 1]) ∪ϕ h only depends on the isotopy class of ϕ (e.g [GS99, Chapter 4.1]), we deduce that
the effect of surgery is determined by the framed embedding (ϕ0 , f ).
The previous discussion implies that if we do surgery along ϕ, then we obtain a homotopy
class [ϕ0 ] ∈ πr (M ). We use this observation to investigate the effect of surgery on the homotopy
groups of M .
Proposition 2.1.3. Let W n+1 be the trace of an r-surgery obtained on M n along ϕ0 and let M 0
be the effect of this surgery. The following statements hold:

1. The inclusion M → W is r-connected and πr (W ) = πr (M )/h[ϕ0 ]i.

20
 2. The inclusion M 0 → W is n − r − 1-connected.

In particular, if k < r and k < n − r − 1, then πk (M ) ∼

= πk (M 0 ).

Proof. We first observe that W has the homotopy of M ∪ er+1 : there is a deformation retraction
of W = M × [0, 1] ∪ϕ Dr+1 × Dn−r onto M × [0, 1] ∪ϕ|Sr ×{0} Dr+1 × {0}. Therefore, up to
homotopy, attaching an (r + 1)-handle is the same as attaching an (r + 1)-cell. The first assertion
follows from Corollary 1.2.17. To prove the second assertion, observe that the manifold M is
obtained from M 0 = M \ (∂Dr+1 × Dn−r ) ∪ Dr+1 × ∂Dn−r by performing an (n − r − 1)-surgery
along Dr+1 × ∂Dn−r . The second assertion is now a consequence of the first, and the last is a
consequence of the two first. This concludes the proof of the proposition.

Proposition 2.1.3 shows that surgery on an embedding ϕ0 : S r → M kills [ϕ0 ] ∈ πr (M ).

Consequently, we can kill represent x ∈ πn (M ) provided x is represented by a framed embedding.
However, what we really want is to kill homotopy groups of a map. Consequently, we fix a smooth
map f : M → X and suppose we do surgery on an embedding ϕ : S r × Dn−r → M . Observe that
f induces a map f 0 : M 0 → X if ϕ extends to an embedding Φ : Dr+1 × Dn−r → X that makes
the following diagram commutes:

S r × Dn−r
ϕ
/M (2.1)
i f
 
Dr+1 × Dn−r
Φ / X.

Indeed, given such an extension Φ of ϕ, we can define f 0 : M 0 → X as f ∪ Φ| , and in fact f extends

to a map F := f ∪ Φ : W → X on the trace of the surgery along ϕ. Set Φ0 := Φ|Dr+1 ×{0} and note
that the pair (ϕ0 , Φ0 ) defines a homotopy class in πr (f ) which we call the homotopy class defined
by the pair (ϕ, Φ).
The next proposition describes the effect of surgery on the relative homotopy groups of f .
Proposition 2.1.4. Let f : M n → X be a map and fix an integer r ≥ 0 so that 2(r + 1) ≤ n.
Let ϕ : S r × Dn−r → M be an embedding with an extension to a map Φ : Dr+1 × Dn−r → X
making (2.1) commute. Using x ∈ πr+1 (f ) to denote the homotopy class defined by (ϕ, Φ) and f 0
the result of surgery on ϕ, then

πr+1 (f 0 ) = πr+1 (f )/hxi and πj (f 0 ) = πj (f ) for j ≤ r.

Proof. We saw in the proof of Proposition 2.1.3 that up to homotopy, the trace W of the surgery
along ϕ is obtained both by adding an (r + 1)-cell to M × [0, 1] and by adding an (n − r)-
cell to M 0 × [0, 1]. As we explained above, the assumptions of the lemma guarantee that f
extends to a map F := f ∪ Φ : W → X. Combining these two observations, Corollary 1.2.19
implies that πr+1 (F ) = πr+1 (f )/hxi, as well as πj (f ) ∼ = πj (F ) for j ≤ r and πj (f 0 ) ∼
= πj (F )
∼ 0
for j ≤ n − r − 1. It follows that πj (f ) = πj (f ) for 2j ≤ n − 1.
We can now conclude. Recall that we assume that 2(r + 1) ≤ n and that the statements of the
proposition involve πj with j ≤ r+1. For j ≤ r, we have 2j ≤ 2r ≤ n−2 ≤ n−1 and so in this case,
we have πj (f ) ∼
= π(f 0 ). For j = r + 1, we have πr+1 (F ) = πr+1 (f )/hxi and πr+1 (f 0 ) ∼
= πr+1 (F ).
It follows that πr+1 (f 0 ) = πr+1 (f )/hxi and this concludes the proof of the proposition.

2.1.2 Motivating the definition of a normal map

Next, we wish to use inductively Proposition 2.1.4 in order to kill elements of πr (f ) (with the hope
of converting f into a homotopy equivalence). To carry this out, we must be able to represent
x ∈ πr (f ) by a pair (ϕ0 , Φ0 ) with ϕ0 : S r → M a framed embedding. Arranging that ϕ0 is an
embedding follows from classical results of Whitney (that we will state in Subsection 2.1.3 below)
and we therefore focus the framing issue. References include [L0̈2, Ran02, Wal70, Wal16, CLMa].

21
 We first forget about Φ0 and establish some necessary and sufficient conditions for an embed-
ding S r → M n to have (stably) trivial normal bundle.
Lemma 2.1.5. Assume ϕ0 : S r → M n is an embedding.

1. If ν(ϕ0 ) is trivial, then ϕ∗0 (T M ) is trivial.

2. If ϕ∗0 (T M ) is stably trivial, then ν(ϕ0 ) is stably trivial.
3. If 2r < n and ν(ϕ0 ) is stably trivial, then ν(ϕ0 ) is trivial.

Proof. The first two assertions follow from the fact that ϕ∗0 (T M ) = ν(ϕ0 ) ⊕ T S r (Example 1.2.26)
together with the fact that T S r is stably trivial (Example 1.2.26)
We now prove the third assertion. Recall from the third item of Example 1.2.10 that stable iso-
morphism classes of bundles over S r are classified by [S r , BO] = πr (BO). Since we assumed ν(ϕ0 )
to be stably trivial, ν(ϕ0 ) is trivial in πr (BO). On the other hand, ν(ϕ0 ) is a rank n − r vector
bundle and therefore defines an element in πr (BO(n − r)) (recall Theorem 1.2.23). It is therefore
enough to show that for 2r < n, the kernel ker(πr (BO(n − r)) → πr (BO)) is trivial. This further
reduces to showing that for 2r < n, we have πr (BO(n − r)) ∼ = πr (BO(n − r + k)) for k > 0.
Corollary 1.2.20 implies that πj (S n−r ) = 0 for j < n − r. Using the assumption 2r < n, the
long exact sequence of the fibration S n−r → BO(n − r + 1) → BO(n − r) gives an isomorphism
πr (BO(n − r)) ∼ = πr (BO(n − r + 1)). The procedure can then be iterated. This concludes the
proof of the third assertion 2 and therefore the proof of the proposition.

Lemma 2.1.5 shows that the task of framing an embedding ϕ0 is therefore intimately tied with
the tangent bundle T M . For f : M → X and [(ϕ0 , Φ0 )] ∈ πr (f ), the next proposition provides a
sufficient condition for the triviality of ϕ∗0 (T M ).
Lemma 2.1.6. Assume that T M ∼ = f ∗ (ξ) for some vector bundle over ξ over X. If x ∈ πr+1 (f ) is
represented by a pair (ϕ, Φ) with ϕ0 : S r → M an embedding, then ϕ∗0 (T M ) is trivial, and therefore
ν(ϕ0 ) is stably trivial.

Proof. First, note that our assumption implies that ϕ∗0 (T M ) = ∼ ϕ∗ f ∗ (ξ) = (f ◦ ϕ0 )∗ (ξ). Next,
0
observe that f ◦ϕ0 : S → X is nullhomotopic since Φ0 provides an extension to Dr+1 (recall (2.1)).
r

Since f ◦ ϕ0 is nullhomotopic, we deduce that ϕ∗0 (T M ) = (f ◦ ϕ0 )∗ (ξ) is trivial. The last assertion
follows from Lemma 2.1.5. This concludes the proof of the lemma.

Given x ∈ πr (f ), Lemma 2.1.6 establishes that if T M = f ∗ (ξ), then ν(ϕ0 ) is stably trivial
(and even trivial if 2r < n by the last item of Lemma 2.1.6). Note however that the condition
T M = f ∗ (ξ) is overkill: it implies that ϕ∗0 (T M ) is trivial, while only stable triviality is needed
(recall the second item of Lemma 2.1.5).
The next definition provides a sharper condition on f that nevertheless ensures that embeddings
of spheres can be stably framed.
Definition 14. A normal map (f, b) consists of a map f : M → X together with a stable isomor-
phism f ∗ (ξ) ∼
= T M i.e. a bundle map b : T M ⊕ u → ξ ⊕ v covering f :

T M ⊕ u
b / ξ ⊕ v

 
M
f
/ X.

Before returning to framed embeddings, we provide a common reformulation of Definition 14.

2 Observethat this statement was about vector bundles over spheres; we only stated it for the normal bundle of
an embedding for the sake of concreteness.

22
Exercise 2.1.7. A normal map can also be thought of as a pair (f, b), where b is a stable
trivialisation of T M ⊕ f ∗ (η) for some bundle η over X. To see this, use the fact that for every
vector bundle η over a compact space X, there is a bundle ξ over X so that ξ ⊕ η is trivial.

The next result shows that a normal map f control the normal bundle of elements in πr (f ).
Proposition 2.1.8. Let f : M n → X be a normal map. If x ∈ πr (f ) can be represented by a
pair (ϕ0 , Φ0 ) with ϕ0 an embedding, then

1. ν(ϕ0 ) is stably framed.

2. ν(ϕ0 ) is framed, provided we additionally assume that 2r < n.

Proof. The second assertion follows from the first thanks to the third item of Lemma 2.1.5. We
now prove the first assertion. As in the proof of Lemma 2.1.5, we have ϕ∗0 (T M ) ∼ = ν(ϕ0 ) ⊕ ,
and it is therefore enough to show that ϕ∗0 (T M ) is stably framed. Since f is a normal map (and
using Exercise 2.1.7), we have a stable framing of T M ⊕ f ∗ (ξ). Pulling back by ϕ0 , we obtain a
stable framing of ϕ∗0 (T M ) ⊕ ϕ∗0 (f ∗ (ξ)). It therefore only remains to show that ϕ∗0 (f ∗ (ξ)) is trivial.
Since [(ϕ0 , Φ0 )] belongs to πr+1 (f ), we deduce that f ◦ ϕ0 factors through Dr+1 and is therefore
nullhomotopic. This implies that ϕ∗0 (f ∗ (ξ)) is trivial, concluding the proof of the proposition.

At this point, we know that below the middle dimension, we can use the normal data of f to
represent classes of πr (f ) by framed embedded spheres. As a consequence, we can perform surgery
on one element. To repeat the process, we need to ensure that the result f 0 : M 0 → X of surgery
on a map f : M → X is still a normal map.
Since this verification will occupy most of next section, we start with a reality check: we show
that homotopy equivalences are normal maps, (it is therefore conceivable that successive surgeries
on a normal map will lead to a homotopy equivalence).

Lemma 2.1.9. If f : M → X is a homotopy equivalence, then it is a normal map .

Proof. Since the manifold M is a compact, a theorem of Whitney ensures the existence of an
embedding ι : M → Rn for some large n. Let g : X → M be the homotopy inverse of f and
consider the vector bundle ξ := g ∗ (ν(in )). Using the definition of ξ, the fact that g ◦ f ∼ id, the
fact that homotopic maps induce isomorphic pullback bundles, as well as Example, we obtain

T M ⊕ f ∗ (ξ) = T M ⊕ (g ◦ f )∗ (ν(in )) = T M ⊕ ν(in ) = n ,

This shows that f is a normal map and therefore concludes the proof of the lemma.

Observe that in the proof of Lemma 2.1.9, the dimension of the euclidean space Rn in which we
embedded M was immaterial. This suggests another equivalent definition of a normal map which
makes use of the so-called “stable normal bundle”. To define this concept, note that given any two
embeddings ι1 : M ,→ Rn1 and ι1 : M ,→ Rnw , there exists N ≥ n1 , n2 such that the compositions
ι1 : M ,→ Rn1 ⊂ RN and ι2 : M ,→ Rn2 ⊂ RN are regular homotopic. This implies that the
classifying maps ν(i1 ) : M → BO(N ) and ν(i2 ) : M → BO(N ) are homotopic. As a consequence,
we obtain a well defined map ν : M → BO, and the corresponding “stable bundle” νM is called
the stable normal bundle of M .
The next exercise provides a definition of normal maps which relies on the stable normal bundle
instead of tangent bundles.
Exercise 2.1.10. A normal map can be thought of as a pair (f, b), where b : νM → ξ is a bundle
map covering f : M → X for some stable bundle ξ. To show this, use the fact that T M and νM
are stable inverses.

23
2.1.3 The surgery step and surgery below the middle dimension
At this point, we know that if f : M n → X is a normal map and x ∈ πr+1 (f ) is represented by a
pair (ϕ, Φ) with ϕ : S r × Dn−r → M an embedding, then f extends to a map F : W → X on the
trace of the surgery and πr+1 (F ) = πr+1 (f )/hxi. In order to repeat the process, we must ensure
that on the effect M 0 of the surgery, the resulting map f 0 : M 0 → X is still normal. References
include [L0̈2, Ran02, Wal70, Wal16, CLMa].
The first step is to show that for r ≤ n−2, a class x ∈ πr+1 (f ) canonically determines a regular
homotopy class of immersions S r × Dn−r → M . This dispenses with the need of systematically
assuming that “x ∈ πr (f ) is represented by (ϕ, Φ) with ϕ0 an immersion”.
Proposition 2.1.11. Let f : M n → X be a normal map and let x be an element of πr+1 (f ).

1. The relative homotopy class x determines a regular homotopy class [ϕx : S r × Dn−r # M ]
of immersions provided r ≤ n − 2.
2. If 2r < n, then the regular homotopy class [ϕx ] contains an embedding.

Furthermore, an immersion g : S r × Dn−r → M belongs to [ϕx ] if and only if g satisfies the

following three conditions:

• There is a map G : Dr+1 × Dn−r such that g can be fitted into the diagram

Sr × D
 _ n−r
g
/M (2.2)
 
f

Dr+1 × Dn−r
G / X.

• [(g|S r ×0 , GDr+1 ×0 )] = x in πr+1 (f )

• The diagram (2.2) is covered by diagram of vector bundle maps

T (S r × Dn−r ) ⊕ a
g
/ T M ⊕ a (2.3)
 
T (Dr+1 × Dn−r ) ⊕ a−1
G / ξ ⊕ a ,

where the rightmost map is induced by the bundle data of the normal map f and the leftmost
map arises from the outer normal vector field of S r = ∂Dr+1 .

Proof. Choose a pair (ϕ0 , Φ0 ) representing the element x ∈ πk+1 (f ). Assume that ϕ0 and Φ0 are
smooth (they need not be immersions) and extend these maps to obtain a commutative diagram

Sr × D
 _ n−r
ϕ
/M

i 
f

D r+1
× Dn−r
Φ / X.

Since f is a normal map, we have stable trivialisation of T M ⊕ f ∗ (ξ). Since Dr+1 × Dn−r is
contractible, we have a trivialisation of T (Dr+1 × Dn−r ) ⊕ Φ∗ (ξ). These stable trivilisations
respectively induce bundle maps T M ⊕ a → ξ ⊕ a and T (Dr+1 × Dn−r ) ⊕ a → ξ ⊕ a (for large
enough a) that respectively cover f and Φ. To get a map T (S r ×Dn−r )⊕a → T M ⊕a covering ϕ,
we must stably trivialize T (S r × Dn−r ) ⊕ ϕ∗ (T M ). We stably trivialize the first summand by
adding  in order to obtain the trivial bundle T (Dr+1 × Dn−r ). To stably trivialize the second
summand, first observe that

ϕ∗ (T M ⊕ f ∗ (ξ)) = ϕ∗ (T M ) ⊕ ϕ∗ (f ∗ (ξ)) = ϕ∗ (T M ) ⊕ i(Φ∗ (ξ)).

24
We stably trivilialize ϕ∗ (T M ⊕ f ∗ (ξ)) using the normal data of f and, for i(Φ∗ (ξ)) we once
again use that Dr+1 × Dn−r is itrivial. We have therefore obtained our stable trivialisation
of T (S r × Dn−r ) ⊕ ϕ∗ (T M ). By construction, all the resulting bundle maps then fit into the
desired commutative diagram:

T (S r × Dn−r ) ⊕ a / T M ⊕ a

 
T (Dr+1 × Dn−r ) ⊕ a−1 / ξ ⊕ a .

One can argue that the top map is a bundle monomorphism and therefore x ∈ πr+1 (f ) determines
(a homotopy class of) a stable bundle monomorphism T (S r × Dn−r ) ⊕ a → T M ⊕ a . Since we
assumed that r ≤ n − 2, a stable version of the Smale-Hirsch theorem (recall Theorem 1.3.1) states
that there is an isomorphism

π0 (Imm(T (S r × Dn−r ), T M ) ∼
= colima→∞ π0 (Mono(T (S r × Dn−r ) ⊕ a , T M ⊕ a )).

As a consequence, the relative homotopy class x ∈ πr+1 (f ) determines a regular homotopy class
of immersions ϕx : S r × Dn−r → M , concluding the proof of the first assertion.
We sketch the proof of the last assertions. If 2r < n, then the strong Whitney embedding
theorem states that any smooth map N r → M n is homotopic to an embedding. The second
assertion follows by applying this theorem to extract an embedding from the regular homotopy
class [ϕx ]. We move on to the last assertion. The construction of ϕx shows that if g is regular
homotopic to ϕx , then it must satisfy the three listed requirements. To prove the converse, the
idea is that the commutativity of the diagram in (2.3) determines the stable homotopy class of
the bundle monomorphism g and therefore the regular homotopy class of g, details can be found
in [CLMa, Theorem 3.93]. This concludes the proof of the proposition.

The next proposition shows that the normal data extends across the trace of a surgery.
Proposition 2.1.12. Let (f, b) : M → X be a normal map and let x ∈ πr+1 (f ) be a relative
homotopy class. If ϕ : S r × Dn−r → M is an embedding that lies in the regular homotopy class
determined by x, then (f, b) induces a normal map on the trace W of the surgery along ϕ.

Proof. Since ϕ belongs to the regular homotopy class determined by x, Proposition 2.1.11 ensures
the existence of a map Φ : Dr+1 × Dn−r → X that fits into the following diagram:

Sr × D
 _ n−r
ϕ
/M

i f

Dr+1 × Dn−r
Φ / X.

This diagram implies that f extends to a map F := f ∪ Φ : W → X on the trace of the surgery.
It remains to shows that F is normal i.e. that the bundle map b : T M ⊕ a → ξ ⊕ a extends to a
bundle map B : T W ⊕ b → ξ ⊕ b covering F . The bundle map b certainly extends to a bundle
map T (M × [0, 1]) ⊕ a → X ⊕ a . Since Dr+1 × Dn−r is contractible, we also obtain a bundle
map T (Dr+1 × Dn−r ) ⊕ b → ξ ⊕ b , covering Φ (just as in the proof of Proposition 2.1.11). In
order to obtain a well defined map on T W , it is necessary and sufficient that these bundle maps
agree over S r × Dn−r i.e. that the following diagram commutes:

T (S r × Dn−r ) ⊕ a / T M ⊕ a

 
T (Dr+1 × Dn−r ) ⊕ a−1 / ξ ⊕ a .

Proposition 2.1.11 ensures that this commutativity follows from the fact that ϕ belongs to the
regular homotopy class determined by x. This concludes the proof of the proposition.

25
 Motivated by Proposition 2.1.12, we make the following definition.
Definition 15. For i = 1, 2, two normal maps (fi , bi ) : Min → X are normal bordant if there
exists a cobordism W between M1 and M2 , and a normal map (F, B) : W → X such that F
restricts to fi on Mi , and such that the stable trivialisation B of T W ⊕ F ∗ (ξ) restricts the stable
trivialisation bi of T Mi ⊕ fi∗ (ξ) over Mi .

Using this terminology, if we perform surgery along a normal map f : M → X along an

embedding satisfying the condition of Proposition 2.1.12, then the effect f 0 : M 0 → X of the
surgery is normal bordant to f .
We can now prove the main result of this section, namely surgery below the middle dimension.
Theorem 2.1.13. If 2r < n, then every normal map (f, b) : M n → X is normal bordant to an
r-connected normal map.

Proof. Let x ∈ πr+1 (f ) be a relative homotopy class. Using Proposition 2.1.11, x determines a
regular homotopy class [ϕx ] of immersions. Since 2r < n, this regular homotopy class contains an
embedding ϕ : S r × Dn−r → M . We can therefore perform surgery on ϕ, and Proposition 2.1.12
shows that we obtain a normal map f 0 : M 0 → X on the effect of the surgery. Proposition 2.1.4
implies that this process does not affect πk (f ) for k ≤ r but replaces πr+1 (f ) by πr+1 (f )/hxi.
Using these facts, we proceed inductively on r. Since πr+1 (f ) is a finitely generated Z[π1 (X)]-
module [CLMa, Lemma 3.81], we can kill the Z[π1 (X)]-generators of πr+1 (f ) one by one. Since
πk (f ) = πk (f 0 ) for k ≤ r, once we have killed a certain relative homotopy class, it remains dead.
This concludes the proof of the theorem.

2.2 The even-dimensional surgery obstruction

Let us recall the current situation: our goal is to determine whether a fixed normal map f : M → X
is normal bordant to a map f 0 : M 0 → X such that πj (f ) = 0 for each j: Whitehead’s theorem
(Theorem 1.2.16) would then ensure that f 0 is a homotopy equivalence. Theorem 2.1.13 shows
that we can assume that f is r-connected with 2r < n. As we shall in see in Proposition 2.2.2
below, for 2r-dimensional manifolds, we need only focus on killing πr+1 (f ), and this will be the
goal of the remainder of the section.
In order to embark on this venture, we need to collect a few facts from homology theory. If the
reader is not familiar with the subject, he need not worry: the “degree one” condition will not play
a major role in the proofs below and homology will only used in these introductory paragraphs.
Remark 2.2.1. Given closed n-manifolds manifolds M, X, a map f : M → X is of degree k
if f∗ [M ] = k[X], where [M ] ∈ Hn (M ; Z) ∼
= Z and [X] ∈ Hn (X; Z) ∼ = Z denote the respective
fundamental classes. Because of the two following properties, we shall from now on assume that
all our normal maps are of degree 1:

1. if f is a homotopy equivalence, then it has degree ±1;

2. if f and f 0 are bordant, then they have the same degree.

The first assertion follows from the multiplicativity of degrees and its homotopy invariance:
if g : X → M is the homotopy inverse of f , then 1 = deg(idX ) = deg(g ◦ f ) = deg(g) deg(f ).
To prove the second property, note that if (W n+1 , M, M 0 ) is a cobordism and F : W → X ex-
tends f and f 0 , then 0 = F ([W ]) = f∗ ([M ]) − f∗0 ([M 0 ]); here the first equality follows from the
fact that X is n-dimensional.

The next proposition shows that for an even dimensional manifold M 2k , our only remaining
task is to kill πk+1 (f ). The proof requires some familiarity with homology theory and covering
spaces; if the reader is not familiar with these topics, he may read the statement of the proposition
and skip its proof.

26
Proposition 2.2.2. Let f : M 2k → X be a degree one map. If f is (k + 1)-connected, then f is
a homotopy equivalence.

Proof. Let Mf and X e be the universal covers of M and X. Since f induces an isomorphism
on fundamental groups, it lifts to a map fe: M f → X. e Recall from Proposition 1.2.12 that we
have πj (M ) = πj (M ) and πj (X) = πj (X) for j ≥ 2. Using the long exact sequences of (X, M )
f e
and (X,
e Mf) (recall the second item of Remark 1.2.15) and the five lemma we deduce that for all j
we have the isomorphisms
πj (fe) = πj (f ).
Since Mf is simply connected, the relative Hurewicz theorem implies that the first non-zero πn (fe)
is equal to the first non-zero Hn (fe). It is therefore enough to show that the latter all vanish: this
will imply that πj (f ) = πn (fe) = 0 for all n, and Whitehead’s theorem (Theorem 1.2.16) will allow
us to conclude that f is a homotopy equivalence.
f) → Hj (X))
We now show that all Hj (fe) vanish. Consider the surgery kernels Kj (f ) := ker(fe∗ : Hj (M e
and K j (f ) := coker(fe∗ : Hj (X)
e → Hj (M f)). Using the long exact sequence of the pair (M, X) and
Poincaré duality for M and X, we deduce that fe is a surjection and therefore Kj (f ) ∼ = Hj+1 (fe).
Combining this fact with Hurewicz’s theorem, we deduce that for j ≤ k, we have

0 = πj+1 (fe) ∼
= Hj+1 (fe) = Kj (f ). (2.4)

Since Kj (f ) = 0 for j ≤ k, some homological algebra shows that K j (f ) = 0 for j ≤ k. Since f

is of degree 1, we can show that fe∗ is in fact a split surjection and a quick diagram chase proves
that Poincaré duality restricts to the surgery kernels. Combining these observations, it follows
that Hj+1 (fe) = Kj (f ) = 0 for all j, as desired. This concludes the proof of the proposition.

Let f : M 2k → X be a degree one normal map. To show that f is normal bordant to a

homotopy equivalence, Theorem 2.1.13 and Proposition 2.2.2 show that we can assume f to be
k-connected and study πk+1 (f ). Thanks to Proposition 2.1.11, elements of πk+1 (f ) determine
regular homotopy classes of framed immersions S k → M . If we could regular homotope such
framed immersions to framed embeddings, then we would perform surgery, and f would be normal
bordant to a homotopy equivalence.

2.2.1 The intersection and self intersection numbers of immersed spheres

Given a k-connected degree one normal map f : M 2k → X, our goal is to decide whether an
element x ∈ πk+1 (f ) can be represented by a pair (ϕ, Φ) with ϕ : S k → M a framed embedding.
Since we saw in Proposition 2.1.11 that x determines the regular homotopy class of a framed
immersion, the goal of this subsection is to decide whether an immersion is regular homotopic to
an embedding. The main tools in this process are the intersections and self-intersections numbers.
References include [Ran02, Section 11], [Wal70, page 57] and [CLMa].
Fix a manifold M 2k . A double point between two maps ϕ1 : S k → M and ϕ2 : S k → M consists
of a pair (x1 , x2 ) ∈ S k × S k such that ϕ1 (x1 ) = ϕ2 (x2 ). We often refer to p := ϕ1 (x1 ) = ϕ2 (x2 ) as
the double point. A double point (x1 , x2 ) is transverse if the following linear map is an isomorphism

(Tx1 ϕ1 , Tx2 ϕ2 ) : Tx1 N ⊕ Tx2 N → Tp M.

Given two immersions ϕi : S k → M that intersect transversely in a finite set of double points (as
depicted schematically in Figure 2.1 below), our goal is to associate to each intersection point p
a sign p and an element gp ∈ π1 (M ). The sign p is the sign of the intersection of ϕ1 (S k ) and
ϕ2 (S k ) at p; we briefly recall this concept.
Remark 2.2.3. Assume that ϕ1 , ϕ2 : S k → M 2k are two immersions that intersect transversely
at p = ϕ1 (x1 ) = ϕ2 (x2 ). Pick bases v1 , . . . , vk of Tx1 S k and w1 , . . . , wk for Tx2 S k so that

27
Bp := (Tx1 ϕ1 (v1 ), . . . , Tx1 ϕ1 (vk ), Tx2 ϕ2 (w1 ), . . . , Tx2 ϕ2 (wk )) is a basis of Tp M . The algebraic
intersection of ϕ1 and ϕ2 at p is defined as the sign of the ordered basis Bp .

Figure 2.1: A sketch of two based immersions ϕ1 , ϕ2 : S k → M 2k that intersect transversely in

two double points.

In order to describe the group element gp ∈ π1 (M ), we need to fix basepoints m in M 2k and s

in S k and introduce some terminology.
Definition 16. A based immersion is a pair (ϕ, γ) with ϕ : S k → M an immersion, and γ a path
from m to ϕ(s).
From now on, all immersions are assumed to be based. To each double point between ϕ1
and ϕ2 , the element gp ∈ π1 (M ) is defined as follows: for i = 1, 2 choose paths ηi ⊂ ϕi (S k ) from
ϕi (s) to p and set
gp := γ2 · η2 · η 1 · γ 1 .
In other words, gp is loop at m that starts along the path γ2 to the basepoint ϕ2 (s) of ϕ2 (S2 ),
round ϕ2 (S2 ) to p (using η2 ), round S1 to ϕ1 (s) (using the reverse path η 1 ) and back along the
path γ1 to m. This is illustrated in Figure 2.2 below. Given a double point p, we shall sometimes
refer to εp gp as the index of p.
The first main definition of this subsection is the following.
Definition 17. The intersection number of two based immersions ϕ1 : S k → M 2k and ϕ2 : S k → M 2k
is obtained by summing the indices of all double points of ϕ1 with ϕ2 :
X
λ(ϕ1 , ϕ2 ) = p gp ∈ Z[π1 (M )].
p

The fact that λ is independent of the choice of η1 , η2 follows because S k is simply connected
for k ≥ 2. In fact, the crucial point is that for i = 1, 2 the subgroup ϕi ∗ (π1 (S k )) ⊂ π1 (M ) is
trivial (leading to a definition for the k = 1 case). In turn, this is equivalent to asking that the
immersions ϕi lifts to the universal cover of M .
Remark 2.2.4. More generally, one can define an intersection number for immersions ϕi : Ni → M
that lift to the universal cover (resp. G-covers) of M provided ϕi ∗ (π1 (Ni )) is trivial in π1 (M )
(resp G). We choose not to pursue these generalizations further and work only with immersed
spheres S k → M 2k with k ≥ 2. Curious readers are referred to [Ran02, Chapter 11].

Figure 2.2: A sketch of the group element gp ∈ π1 (M ) associated to a transverse double point p.

It can be shown that the intersection form λ is well defined on the set Ik (M ) of based regular
homotopy classes of based immersed spheres. This leads to the following observation.

28
Remark 2.2.5. If a framed based immersion ϕ is regular homotopic to a framed embedding g,
then λ(ϕ, ϕ) = 0. To see this, let g 0 be the embedded sphere obtained from g by pushing along
a nowhere-zero section of ν(g). Since we just saw that λ is invariant under regular homotopy, we
have λ(ϕ, ϕ) = λ(g, g 0 ) = 0.

Remark 2.2.5 shows that the vanishing of the intersection numbers provides a necessary condi-
tion for an immersion to be regular homotopic to an embedding. Unfortunately, the next example
shows that this condition is not sufficient.

Example 2.2.6. The immersion f of S 1 as a figure eight “8” in M = S 2 = R2 ∪ {∞} has

intersection number λ(f, f ) = 0 but it is not regular homotopic to an embedding (e.g. using the
Whitney-Graustein theorem, recall Remark 1.1.17). This example can be generalized to arbitrary
dimensions as explained in [Ran02, Proposition 7.12].

A sufficient condition to obtain an embedding is obtained by studying self intersections. The

self intersection number of a single based immersed sphere f : S k → M is defined similarly to λ.
For each transverse double point p in S k , two branches of S k cross at p and, after choosing an
order of these branches, we can define the index p gp exactly as in the definition of λ, and consider
X
f 7→ p gp .
p

Unfortunately, since there is no canonical choice of ordering for the aforementioned branches, this
quantity is not well defined in Z[π1 (M )] (switching the order of branches that meet at p changes
the sign and group element by p ↔ (−1)k p and gp ↔ gp−1 ). On the other hand, it will be well
defined in the following quotient which cancels out the effect of swapping the order of the branches:

Q(−1)k (Z[π1 (M )]) := Z[π1 (M )]/{x ∼ (−1)k x | x ∈ Z[π1 (M )]}.

Here we used the following notation: we write : Z[π1 (M )] → Z[π1 (M )] for the involution ob-
tained by extending Z-linearly the involution g 7→ g −1 of π1 (M ) to Z[π1 (M )].
We can now define the self intersection number of a based immersed sphere.

Definition 18. The self intersection number of a based immersion f : S k → M 2k is obtained by

summing the indices of all double points:
X
µ(f ) := p gp ∈ Q(−1)k (Z[π1 (M )]).
p

By construction, the choice of Q(−1)k (Z[π1 (M )]) as a target ensures that µ is well defined.
Some additional work shows that µ is unchanged under regular homotopy.
Remark 2.2.7. We have seen that both λ and µ are well defined on the set Ik (M ) of based regular
homotopy classes of based immersed spheres. In fact, Ik (M ) is a group and even a Z[π1 (M )]-
module; we deliberately avoid giving further details and instead refer to [CLMa].

The next result (often referred to as the Wall embedding theorem) shows that in high dimen-
sions, self intersections are the only obstruction for a regular homotopy class of an immersion to
contain an embedding.

Theorem 2.2.8. Let f : S k → M 2k be a based immersion. If k ≥ 3, then the following statements

are equivalent:

1. f is based regular homotopic to a based embedding.

2. µ(f ) = 0.

29
Proof. If f is based regular homotopic to a based embedding g, then (since embeddings are injective
immersions), we deduce that µ(f ) = µ(g) = 0. The proof of the converse relies on the Whitney
trick. Since µ(f ) = 0, we can find a pair of double points p, p0 with opposite signs but same group
elements: p = −p0 and gp = gp0 . Use Df to denote the double point sets of f . Since k ≥ 3, the
Whitney trick makes it possible to change f within its based regular homotopy class to a based
immersion g with Dg = Df \ {p, p0 } and µ(g) = µ(f ). The proof is concluded by induction (and
because injective immersions of compact manifolds are embeddings by Proposition 1.1.14).

We now return to our initial question. Is a given degree one normal map f : M → X normal
bordant to a degree one normal map f 0 : M 0 → X with πi (f 0 ) = 0 for i ≤ k + 1?
The next result recalls the relation between the relative homotopy group πk+1 (f ) and the
set Ik (M ) of based regular homotopy classes of based immersions S k → M .
Lemma 2.2.9. Let f : M 2k → X be a normal map. There is a map α : πk+1 (f ) → Ik (M ) that
fits into the following commutative diagram:

πk+1 (f )
α / Ik (M ) (2.5)

∂ ' w ι
πk (M ).

Here, ∂ is the connecting homomorphism in the long exact sequence of homotopy groups (recall
Theorem 1.2.14) and ι is the map that sends a regular homotopy class of an immersion to its
underlying homotopy class.

Proof. Since f is a normal map, Proposition 2.1.11 implies that x ∈ πk+1 (f ) determines a regular
homotopy class [ϕx : S k × Dn−k → M ] of immersions. Proposition 2.1.11 also implies that if ϕ
belongs to [ϕx ] then it fits into a diagram of the form

S k × Dn−k
ϕ
/M

 
f

D k+1
× Dn−k
Φ / X,

and [(ϕ|S k ×0 , Φ|Dk+1 ×0 )] = x. We set α(x) := [ϕ|S k ×0 ]. The commutativity of the diagram in (2.5)
follows readily. This concludes the proof of the lemma.

Slightly abusing notations, we also use λ and µ to denote the maps on πk+1 (f ) obtained by
respectively precomposing λ and µ with α × α and α. If we assume that f is a k-connected degree
one normal map, then Proposition 2.2.2 and Lemma 2.2.9 produce a map
α
Kk (M ) ∼
= πk+1 (f ) → Ik (M ).

Here recall that Kk (M ) := ker(Hk (Mf; Z) → Hk (X; e Z)). Note that since π1 (M ) acts on the uni-
versal cover M
f, this surgery kernel is a left Z[π1 (M )]-module. Summarizing, given a k-connected
degree one normal map f : M → X, we get a triple (Kk (M ), λ, µ).
The next proposition collects some properties of this kernel form.
Proposition 2.2.10. The kernel form (Kk (M ), λ, µ) of a k-connected degree one normal map f : M 2k → X
satisfies the following properties for all x, y, x1 , x2 , y1 , y2 ∈ Kk (M ), all r, s ∈ Z[π1 (M ] and
all g ∈ π1 (M ):

1. Kk (M ) is a finitely generated stably free Z[π1 (M )]-module: there exist free finitely generated
Z[π1 (M )]-modules F1 , F2 such that Kk (M ) ⊕ F1 = F2 .
2. λ is sesquilinear: λ(x1 , rx2 +sy2 ) = rλ(x1 , x2 )+sλ(x1 , y2 ) and λ(rx1 +sy1 , x2 ) = λ(x1 , x2 )r+
λ(y1 , x2 )s.

30
 3. λ is (−1)k -Hermitian: λ(y, x) = (−1)k λ(x, y).
4. λ is nonsingular: the map x 7→ λ(x, −) is a Z[π1 (M )]-linear isomorphism.
5. µ(gx) = gµ(x)g
6. Using pr : Z[π1 (M )] → Q(−1)k (Z[π1 (M )]) to denote the quotient map, one has

µ(x + y) − µ(x) − µ(y) = pr(λ(x, y)).

7. Let α : Kk (M ) → Ik (M ) be the map described in Lemma 2.2.9. If the normal bundle of α(x)
is trivial, then
λ(x, x) = µ(x) + (−1)k µ(x).

Proof. We only provide a proof sketch and some references. The proof of the first assertion can be
found in [CLMa, Lemma 4.56]. To show that that the adjoint x 7→ λ(x, −) of λ is an isomorphism,
the idea is to show that it coincides with the composition of Poincaré dualiy and an evaluation
map (both of which are isomorphisms): 3

Kk (M ) → K k (M ) → HomZ[π1 (M )] (Kk (M ), Z[π1 (M )]).

The remaining statements can be shown direcly on the level of the set Ik (M ) of based immersions.
The point that we have glossed over is that Ik (M ) is a Z[π1 (M )]-module (recall Remark 2.2.7)
and α is a Z[π1 (M )]-linear map. The (−1)k -Hermitian and sesquillineraty properties can checked
by hand, while the statements on µ can be found in [CLMa, Lemma 4.11]. This concludes our
proof sketch.

Summarizing, for every k-connected degree one normal map f : M 2k → X, we have associated
a triple (Kk (M ), λ, µ). Next, we review the algebra necessary to describe such objects.

2.2.2 Symmetric and quadratic forms

The goal of this short subsection is to collect some terminology on quadratic forms. References
include [Ran81, Wal70, KL05].
Rings are understood to be associative and with a unit and all modules are assumed to be
finitely generated. A ring with involution is a ring R together with an involution x 7→ x that
satisfies ab = ba, a + b = a + b and x = x. Recall that an R-module is projective if there exists an
R-module Q such that P ⊕ Q is free.
In practice, we shall mostly work with free modules. However, as the next remark shows, it is
convenient to have a class of modules that contains stably free modules.
Example 2.2.11. Stably free modules are projective (but the converse does not hold) and in
particular, if f : M 2k → X is a k-connected degree one normal map, then Kk (M ) is projective
since Proposition 2.2.10 shows that it is stably free over R = Z[π1 (M )].

Set ε = ±1. Given a projective left R-module P , a pairing λ : P × P → R is sesquilinear if it

satisfies λ(x1 , ax2 + by2 ) = aλ(x1 , x2 ) + bλ(x1 , y2 ) and λ(ax1 + by1 , x2 ) = λ(x1 , x2 )a + λ(y1 , x2 )b for
all x1 , x2 , y2 , y2 ∈ P and all a, b ∈ R. A sesquilinear pairing is ε-Hermitian if λ(y, x) = ελ(x, y) for
all x, y ∈ P . A ε-symmetric form consists of a pair (P, λ) where P is a projective left R-module
and λ is a ε-Hermitian form on P .
Example 2.2.12. If f : M 2k → X is a k-connected degree one normal map, then Proposi-
tion 2.2.10 shows that (Kk (M ), λ) is a (−1)k -symmetric form.

In order to discuss non-singularity, we start with some remarks on module structures.

3 The meaning of the overline will be explained in Subsection 2.2.2 below.

31
Remark 2.2.13. If P is a left R-module, then Homleft- R (P, R) is endowed with a right R-module
structure via the action ϕ · r(x) := ϕ(x)r. In order to obtain a left R-module, we use the following
construction: for a right R-module M , we use M to denote the left R-module with the same
underlying abelian group structure as M but with action r · m := mr. From now, we write P ∗ for
left R-module Homleft- R (P, R) endowed with the action ϕ · r(x) = ϕ(x)r.

A sesquilinear form determines an R-linear map λ• : P → P ∗ by setting λ• (x)(y) = λ(x, y). A

symmetric form (P, λ) is non-degenerate if λ• is injective and nonsingular if λ• is an isomorphism.
These remarks lead to a common reformulation of the notion of a symmetric form.
Exercise 2.2.14. Consider the map ev : P → (P ∗ )∗ , p 7→ (ϕ 7→ ϕ(p)). Show that a sesquilinear
pairing λ is Hermitian if and only ελ• coincides with the following composition:
ev φ∗
P → (P ∗ )∗ → P ∗ .

Show furthermore that if φ : P → P ∗ is R-linear and satisfies φ = φ∗ ◦ ev, then the assignment
λ(x, y) := φ(x)(y) defines an ε-Hermitian form on P .

Next, we describe a class of symmetric form which occurs frequently in surgery theory.
Example 2.2.15. We describe the standard hyperbolic symmetric form H ε (Ru ). This form has
underlying module R2u , a basis e1 , . . . , eu , f1 , . . . , fu and the symmetric form is described by
λ(ei , ej ) = 0 = λ(fi , fj ) and λ(ei , fj ) = δij .

Returning to topology, given a k-connected degree one normal map f : M 2k → X, we know

that the pair (Kk (M ), λ) is a non-singular symmetric form. We now introduce the terminology
needed to describe the self-intersection number µ. First, set

Qε (R) = R/{r − εr | r ∈ R}.

Next, abstracting the properties of self-intersections, we make the following definition.

Definition 19. A quadratic refinement of an ε-symmetric form (P, λ) is a map µ : P → Qε (R)
such that the following properties hold for all r ∈ R and all p, p1 , p2 ∈ P :

1. µ(rp) = rµ(p)r,
4
2. λ(p, p) = µ(p) + εµ(p),
3. If we use pr : R → Qε (R) to denote the canonical projection, then

µ(p1 + p2 ) − µ(p1 ) − µ(p2 ) = pr(λ(p1 , p2 )).

An ε-quadratic form is a triple (P, λ, µ) consisting of a symmetric form (P, λ) together with a
quadratic refinement µ of λ.

We now define the quadratic analogue of Example 2.2.15.

Example 2.2.16. We describe the standard hyperbolic quadratic form Hε (Ru ). This form has un-
derlying module R2u , a basis e1 , . . . , eu , f1 , . . . , fu , the symmetric form is described by λ(ei , ej ) =
0 = λ(fi , fj ) and λ(ei , fj ) = δij , the quadratic form is µ(ei ) = 0 = µ(fi ) for all i. We will
sometimes call such a basis a hyperbolic basis.

Next, we define sums and isomorphisms. Two symmetric forms (P1 , λ1 ) and (P2 , λ2 ) are
isometric if there is an R-linear isomorphism f : P1 → P2 such that λ2 (f (x), f (y)) = λ1 (x, y)
for every x, y ∈ P1 . Two quadratic forms (P1 , λ1 , µ1 ) and (P2 , λ2 , µ2 ) are isometric if there is
4 This means that µ(p) + εµ(p) ∈ R is the image of µ(p) by the map Qε (R) → R, [r] 7→ r + εr.

32
an isomorphism f : P1 → P2 such that λ2 (f (x), f (y)) = λ1 (x, y) and µ2 (f (x)) = µ1 (x) for every
x ∈ P1 . The sum of the symmetric forms (P1 , λ1 ) and (P2 , λ2 ) is the symmetric form (P1 ⊕P2 , λ1 ⊕
λ2 ), where λ1 ⊕ λ2 ((x1 , x2 ), (y1 , y2 )) = λ1 (x1 , y1 ) + λ2 (x2 , y2 ). For quadratic refinements, we set
µ1 ⊕ µ2 (x, y) = µ1 (x) + µ2 (y).
We conclude this short subsection with some additional terminology.
Definition 20. A quadratic form is hyperbolic if it is isomorphic to a standard hyperbolic form.
Two quadratic forms (P1 , λ1 , µ1 ) and (P2 , λ2 , µ2 ) are stably isomorphic if (P1 , λ1 , µ1 ) ⊕ Hε (R)u
and (P2 , λ2 , µ2 ) ⊕ Hε (R)v are isomorphic for some non-negative integers u and v.

2.2.3 Surgery on hyperbolic surgery kernels

The aim of this subsection is to provide a sufficient condition for a k-connected degree one normal
map to be normal bordant to a homotopy equivalence.
Let f : M → X be a normal map. Recall that Ir (M ) denotes the set of based regular homotopy
classes of based immersions S r # M . Recall furthermore that Ir (M ) is a Z[π1 (M )]-module (a
fact we only mentioned in passing, recall Remark 2.2.7) and that the map α : πr+1 (f ) → Ir (M )
of Lemma 2.2.9 is a Z[π1 (M )]-homomorphism. The zero element of Ir (M ) is represented by the
standard embedding S r → Rr+1 → Rr ⊂ M , where the last map is any embedding. Therefore we
can always do surgery on the zero element of πr+1 (f ).
The next lemma describes the effect of such a surgery on the surgery kernel.

Lemma 2.2.17. Let k ≥ 2 and let f : M → X be a k-connected degree one normal map. If we
do surgery on the zero element of Kk−1 (M ) ∼ = πk (f ), then the effect on the surgery kernels is to
replace (Kk (M ), λ, µ) by (Kk (M ), λ, µ) ⊕ H(−1)k (Z[π1 (M )]).

Proof. We claim that the result of this surgery is to replace M by M #S k × S k . By definition

of the 0-element of Ik (M ) (which we recalled above the statement of the lemma), we can assume
that the standardly embedded sphere S k−1 on which we are doing surgery is contained in an S 2k
summand of a connected sum decomposition M ∼ = M #S 2k . It is therefore enough to show that
the effect of surgery along a standardly embedded (framed) S k−1 ⊂ S 2k replaces S 2k by S k × S k .
Since this embedding is framed, we can write S 2k = S k−1 × Dk+1 ∪ Dk × S k . Performing the
surgery replaces this S 2k by an S k × S k , as claimed.
Next, we study the effect of this surgery on Kk (M ) Set M 0 := M #S k × S k and π := π1 (M ).
Pull back the universal of X to M 0 . Since S k × S k is simply connected, the restriction of this cover
to S k × S k is the trivial π-cover. It follows that we get a direct sum decomposition Hk (M f0 ; Z)
as Hk (M ; Z) ⊕ Z[π][S × {pt}] ⊕ Z[π][{pt} × S ]. Since the degree one normal map on the S × S k
f k k k

summand of M 0 was obtained from a degree one normal map on S 2k , we deduce that it is zero on
the Z[π]2 -summand of Hk (M f0 ; Z). Passing to surgery kernels, we obtain the direct sum decomposi-
0
tion Kk (M ) = Kk (M )⊕Z[π][S k ×{pt}]⊕Z[π][{pt}×S k ]. Using the definition of (self)-intersection
numbers, we deduce that this surgery adds a hyperbolic form to Kk (M ), concluding the proof of
the lemma.

The next theorem provides a sufficient condition for a highly connected degree one normal map
to be normal bordant to a homotopy equivalence.
Theorem 2.2.18. Let f : M 2k → X be a k-connected degree one normal map. If k ≥ 3
and (Kk (M ), λ, µ) is stably hyperbolic, then f is normal bordant to a homotopy equivalence.

Proof. Without loss of generality, we can assume that the kernel form (Kk (M ), λ, µ) is hyperbolic,
i.e. isometric to some H(−1)k (Z[π]v ): Lemma 2.2.17 shows that performing surgery on the trivial
element of πk (f ) ∼
= Kk−1 (M ) adds a hyperbolic summand to (Kk (M ), λ, µ). Since we are working
up to normal bordism, we can indeed assume that (Kk (M ), λ, µ) ∼ = H(−1)k (Z[π]v ).

33
 Endow (Kk (M ), λ, µ) with a hyperbolic basis (b1 , . . . , bv , c1 , . . . , cv ) (recall Example 2.2.16),
and note that f is a homotopy equivalence if and only if v = 0. We must therefore show that f
is normal bordant to a degree one normal map f 0 : M 0 → X such that Kk (M 0 ) has rank v − 1
and (Kk (M 0 ), λ0 , µ0 ) is still hyperbolic. The conclusion will then follow by induction.
We do surgery on the basis element bv ∈ Kk (M ) ∼ = πk+1 (f ). Note that this is possible since we
assumed k ≥ 3: since µ(bv ) = 0, the Wall embedding theorem (recall Theorem 2.2.8) ensures that
the regular homotopy class determined by bv contains a framed embedding. The normal map f
extends to the trace W of this surgery and the result is a map
(F, f, f 0 ) : (W, M, M 0 ) → (X × [0, 1], X × {0}, X × {1}).
This triple gives rise to surgery kernels Ki (W ), Ki (M ), Ki (M 0 ). These kernels admit natural
generalisations to pairs. For instance Ki (W, M ) := ker(Hj (W f, Mf) → Hj (X e × [0, 1], X
e × {0}),
0
and Ki (W, M ) is defined similarly. As in the proof of Proposition 2.2.2, similar definitions
can be made in cohomology (using cokernels instead of kernels), and the Z[π1 (M )]-modules
Ki (W, ∂W ), Ki (W, M ), Ki (W, M 0 ) satisfy Poincaré duality and fit into long exact sequences.
We already know that Ki (M ) is non-zero if and only if i = k. Recall that up to homotopy
equivalence, W is obtained by adding a (k + 1)-cell to M . As a consequence, using Proposi-
tion 2.1.4, we know that Ki (W ) = πi+1 (F ) = πi+1 (f ) = Ki (M ) vanishes for i ≤ k − 1. We
also know from Proposition 2.1.4 that Ki (M 0 ) = Ki (M ) = 0 for i ≤ k − 1, or put differ-
ently, that f 0 : M 0 → X is also a k-connected degree one normal map. Arguing as in Propo-
sition 2.2.2, we deduce that Ki (M 0 ) = 0 for i 6= k. It follows that Ki (∂W ) = 0 for i 6= k
and that Ki (W, M ), Ki (W, M 0 ) and Ki (W, ∂W ) all vanish for i ≤ k − 1. Since f is k-connected,
we also deduce that Ki (W, M ) = 0 for i = k. Duality and some homological algebra imply
that Ki (W, M ) = 0 for all i 6= k + 1 and Ki (W, M 0 ) = 0 for all i 6= k. The long exact sequence
for (W, M 0 ) implies that Ki (W ) = 0 for i 6= k. It follows that Ki (W, ∂W ) = 0 for i 6= k + 1. We
therefore get the following braid of four interlocking exact sequences:
α β

) (
Kk+1 (W, M ) Kk (M ) Kk (W, M 0 )
8 ϕ 6 6
) ' &
0 Kk+1 (W, ∂W )
5
Kk (W )
7 8 0.
ψ
(
& (
0
4
Kk (M 0 ) 70

The long exact sequences that come into play are those of the pairs (W, M ), (W, M 0 ) and of
the triples (W, ∂W, M ), (W, ∂W, M 0 ). Set π := π1 (M ) and note that Kk+1 (W, M ) ∼ = Z[π] is
freely generated by the core of the (k + 1)-handle φk+1 that we attached to M × [0, 1]: re-
call that up to homotopy equivalence, W is obtained from M by adding a (k + 1)-cell; there-
fore we have Hk (W, M ) = Z[π][φk+1 ], and the observation follows. A similar argument shows
that Kk (W, M 0 ) ∼
= Z[π] is freely generated by the core of the dual k-handle ψ k . Next, note that
the homomorphism β : Kk (M ) → Kk (W, M 0 ) is given by x 7→ λ(bv , x)ψ k : briefly, this can be seen
by noting that α(φk+1 ) = bv , using that λ can be defined by composing Poincaré duality with the
evaluation map (a fact we chose not to mention in Subsection 2.2.1), and considering the following
commutative diagram:
PD,∼ ev,∼
Kk (M )
=
/ K k (M ) =
/ HomZ[π] (Kk (M ), Z[π])
∗
β α∗
 α 
PD,∼ ev,∼
Kk (W, M 0 )
=
/ K k+1 (W, M ) =
/ HomZ[π] (Kk+1 (W, M ), Z[π]).

We claim that there exists a basis (b01 , . . . , b0v , c01 , . . . c0v−1 ) for Kk+1 (W, ∂W ) such that ϕ(b0i ) = bi
and ϕ(c0i ) = ci for each i. Observe that by exactness of the braid at the parts involving ϕ and β,

34
we obtain Kk+1 (W, ∂W ) = ∼ im(ϕ) = ker(β). Since we chose a hyperbolic basis for Kk (M ), the
previously mentioned description of β implies that β(cv ) = λ(bv , cv ) = 1 and β(ci ) = λ(bv , ci ) = 0
for i 6= v and β(bi ) = λ(bv , bi ) = 0 for all i. The claim follows.
We now turn to the Z[π]-module Kk (M 0 ). By exactness of the diagonal involving the map ψ,
we know that this Z[π]-module is isomorphic to Z[π]2(v−1) and is freely generated by the ψ(b0i )
and ψ(c0i ) for i = 1, . . . , v − 1. We outline why these elements still form a hyperbolic basis. Since
(b1 , . . . , bv , c1 , . . . , cv ) for a hyperbolic basis and since k ≥ 3, we can use the Whitney trick to
arrange that the spheres b1 , . . . , bv−1 , c1 , . . . , cv−1 do not intersect bv [Ran02, Corollary 7.30]. As
a consequence, the surgery on bv will not affect the other intersection numbers. 5 This concludes
the proof of the theorem.

Given a degree one normal map f : M → X, our goal is to define a “surgery obstruction” σ(f )
that only depends on f : M → X and not on an intermediate highly connected degree one normal
map. The first step is to describe the group to which this surgery obstruction will belong.

2.2.4 The even quadratic L-groups

A first glance at Theorem 2.2.18 suggests that given a degree one normal map f : M 2k → X,
the required “surgery obstruction” should involve the kernel form (Kk (f 0 ), λ0 , µ0 ) of a highly con-
nected f 0 that is normal bordant to f . A second glance at Theorem 2.2.18 suggests that this surgery
obstruction ought to belong to a group of “symmetric forms modulo stably hyperbolic forms”. The
aim of this subsection is define such a group. Reference include [Ran81, Wal70, KL05].

Let R be a ring. We start with two important conventions.

Remark 2.2.19. In this subsection, all R-modules are understood to be free left R-modules and
all quadratic forms are assumed to be nonsingular.

The main definition of this subsection is the following.

Definition 21. The 2k-th quadratic L-group L2k (R) is the abelian group of stable isomorphism
classes of (−1)k -quadratic forms; the sum is given by the direct sum operation, the zero element is
represented by the class of H(−1)k (R)u for any u ≥ 0, and the inverse of [(H, λ, µ)] is [(H, −λ, −µ)].

More explicitly, L0 (R) (which is equal to L4n (R) for all n) consists of stable isomorphism
classes of quadratic forms, while L2 (R) (which is equal to L4n+2 (R) for all n) consists of stable
isomorphism classes of (−1)-quadratic forms (i.e. the form λ is skew-Hermitian).
The goal of the remainder of this subsection to show that L2k (R) is indeed an abelian group.
First, note that stable isomorphism is clearly an equivalence relations which is compatible with
the sum operation. Associativity and commutativity of the direct sum are also readily verified. By
definition of the equivalence relation, the class of the hyperbolic form is indeed the zero element.
Consequently, it remains to show that the inverse of [(H, λ, µ)] is represented by (H, −λ, −µ).
This result will follow readily from the following useful proposition.
Proposition 2.2.20. Let (H, λ, µ) be an ε-quadratic form, where H is of rank 2n. If there exists
a free half rank summand G ⊂ H such that λ|G×G = 0 and µ|G = 0, then (H, λ, µ) is isomorphic
to Hε (Rn ).

Proof. Choose a basis x1 , . . . , xn of G. Since λ is nonsingular, we have an isomorphism H → H ∗ .

Since G is a summand of H, we can extend the basis x1 , . . . , xn of G to a basis x1 , . . . , x2n of H.
For i = 1, . . . , n, consider the linear map δi : H → R given by δi (xj ) = δij . These δi define
elements in H ∗ . Since λ is non-singular, its adjoint λ• : H → H ∗ is an isomorphism and therefore
5 the
key underlying point is: “in high dimensions, the Whitney trick can be used to realize algebraic intersections
geometrically.”

35
there exists elements y1 , . . . , yn of H such that λ(x, yj ) = δi (x) for all x ∈ H. In particular, we
have λ(xi , yj ) = δij for i = . . . , n and therefore (x1 , . . . , xn , y1 , . . . , yn ) forms a basis of H.
We now prove the proposition by induction on n. Suppose n = 1, write x, y instead of x1 , y1
for simplicity, and note that G = hxi and H = hx, yi. By definition of G, we know that λ(x, x) = 0
and µ(x) = 0, and by construction of y, we have λ(x, y) = 1. Set y 0 = y−ax with a a representative
of µ(y). Using the properties of λ and µ (recall Subsection 2.2.2), we check that hx, y 0 i is a
hyperbolic basis of H:

λ(x, y 0 ) = λ(x, y) − aλ(x, x) = λ(x, y) = 1,

λ(y 0 , y 0 ) = λ(y, y) − aλ(x, y) − aλ(y, x) = λ(y, y) − (a + εa) = 0,
µ(y 0 ) = µ(y) − µ(ax) − λ(y, ax) = µ(y) − εa = a − a = 0.

Next, assume inductively that the proposition is true for i ≤ n−1. If λ(yi , yj ) = 0 and µ(yj ) = 0
for all i, j, then the form is already hyperbolic. Otherwise, pick an yj for which one these equalities
does not hold. Without loss of generality, we can assume that j = 1 and define y10 = y1 − ax1 ,
where a represents µ(y1 ). The computation above shows that X := hx1 , y10 i is hyperbolic. We
set H 0 := hx1 , y10 i⊥ so that H = X ⊕ H 0 with H 0 free of rank 2(n − 1). Observe that the
summand G0 := hx2 , . . . , xn i of H 0 has half rank. We apply the inductive step to G0 ⊂ H 0 . It
follows that H 0 is hyperbolic and therefore so is H = X ⊕ H 0 . This concludes the proof of the
proposition.

We can now prove the desired result:

Proposition 2.2.21. L2k (R) is an abelian group.

Proof. As we saw in the discussion following Definition 21, we need only show that the inverse
of [(H, λ, µ)] is represented by (H, −λ, −µ). In other words, we must show that the direct
sum (H, λ, µ) ⊕ (H, −λ, −µ) is hyperbolic. This follows by applying Proposition 2.2.20 to the
diagonal
G := {(x, x) | x ∈ H} ⊂ H ⊕ H.
Indeed, we have (λ ⊕ −λ)|G×G = 0 and (µ ⊕ −µ)|G = 0. This concludes the proof of the
proposition.

It is worth mentioning that a G as in Proposition 2.2.20 is often referred to as a quadratic

lagrangian.
Next we discuss some examples of L-groups.
Example 2.2.22. We describe the even quadratic L-groups of Z.

1. The signature gives rise to an isomorphism L0 (Z) = ∼ 8Z. (see e.g. [CLMb, Theorem 7.93]
for a proof). Here recall that the signature sign(λ) of symmetric form (H, λ) is obtained
by representing λ by a Hermitian matrix A and defining sign(λ) as the number of positive
eigenvalues of A minus the number of negative eigenvalues of A. The signature of a quadratic
form is defined as the signature of its underlying symmetric form (the fact that λ admits a
quadratic refinement ensures that sign(λ) ∈ 8Z [vdB59]). It is well known that the signature
is independent of the matrix representative of λ.
2. The Arf invariant gives rise to isomorphisms L2 (Z) = ∼ Z2 and L2 (Z2 ) ∼ = Z2 (see e.g.
[CLMb, Theorem 7.105] for a proof). We briefly describe the Arf invariant of a quadratic
form (H, λ, µ). Since the Arf invariant of a quadratic form over Z is defined by first ten-
soring by Z2 (i.e. reducing λ and µ mod 2), we assume that (H, λ, µ) is a quadratic form
over Z2 . Since λ is non-singular H must be even-dimensional, say H ∼ = Z2n
2 . All non-
singular skew-symmetric forms over a field admit a hyperbolic basis. Pick such a hyperbolic

36
 basis (e1 , . . . , en , f1 , . . . , fn ) for (H, λ) and define the Arf invariant as
n
X
Arf(H, λ, µ) := µ(ei )µ(fi ).
i=1

This definition is well known to be independent of all the choices involved. As an exam-
ple, we make the easy verification that the Arf invariant of the 2-dimensional hyperbolic
form on he, f i vanishes (since the Arf invariant is additive unders direct sums, this shows
that the Arf invariant descends to a map on L2 (Z2 )): as µ(e) = 0 = µ(f ), we conclude
that Arf(H, λ, µ) = µ(e)µ(f ) = 0.

We conclude this subsection with a brief outlook on L-theory.

Remark 2.2.23. Up to now, we have only considered the even L-groups and assumed that all
modules were free. We list some other flavors of L-groups and refer to [Ran81] for details.

1. There are L-groups Lk (R) when k is odd. Instead of quadratic forms, the definition of Lk (R)
involves so-called “quadratic formations”. These L-groups are 4-periodic: Lk+4 (R) = Lk (R)
and, for instance, it is known that L1 (Z) = 0 = L3 (Z).
2. L-groups can also be defined using projective R-modules instead of free R-modules. In
this case, the resulting groups Lpn (R) are different and are related to the Ln (R) by the so-
called Ranicki-Rothenberg exact sequence [Ran81, Ran73a, Ran73b]. If R is a PID, then
projective modules are known to be free, and so Lpn (R) = Ln (R). Use R[t±1 ] to denote the
ring of Laurent polynomials with coefficients in a commutative ring R. Shaneson showed
that Ln (R[t±1 ]) ∼ p
= Ln−1 (R) ⊕ Ln (R) [Sha69]. For instance, this Shaneson splitting shows
that L0 (Z[Z]) = L0 (Z) ⊕ L3 (Z) = L0 (Z) (since Z is a PID, we dropped the decorations).
3. Given a multiplicative subset S ⊂ R, there are also relative L-groups of linking forms and
linking formations. These groups are frequently denoted Ln (R, S) and fit into a localisation
exact sequence
. . . → Ln (R) → Ln (S −1 R) → Ln (R) → Ln−1 (R) . . .

4. There are symmetric L-groups Lk (R) of (−1)k -symmetric forms/formations: one disregards
the quadratic refinements (an arbitrary symmetric form need not admit a quadratic refine-
ment!) These groups coincide with the classical “Witt groups” of Hermitian forms.
5. Quadratic and symmetric L-groups have been reformulated using particular types of chain
complexes, known as “algebraic Poincaré complexes”.

2.2.5 The surgery obstruction in the even-dimensional case

In this subsection, we define the surgery obstruction σ(f ) of a degree one normal map f : M 2k → X.
Theorem 2.2.24 shows that in high dimensions σ(f ) provides a necessary and sufficient condition to
decide whether f is normal bordant to a homotopy equivalence. References include [L0̈2, Ran02,
Wal70, Wal16, CLMa].
Let f : M 2k → X be a degree one normal map. Using surgery below the middle dimension
(recall Theorem 2.1.13), we can assume that f is normal bordant to a k-connected degree one
normal map f 0 : M 0 → X. We know from Proposition 2.2.10 that Kk (M 0 ) is a stably free Z[π1 (X)]-
module. We can therefore choose a natural number r so that Kk (M 0 ) ⊕ Z[π1 (X)]2r is free. We
deduce that (Kk (M 0 ), λ, µ) ⊕ H(−1)k (Z[π]r ) determines an element in L2k (Z[π1 (X)]).
The main definition of this subsection is the following.
Definition 22. The surgery obstruction of a degree one normal map f : M 2k → X is defined as

σ(f ) := [(Kk (M 0 ), λ0 , µ0 ) ⊕ H(−1)k (Z[π]r )] ∈ L2k (Z[π1 (X)]),

37
where f 0 : M 0 → X is a k-connected degree one normal map that is normal bordant to f , and r is
any natural number such that Kk (M 0 ) ⊕ Z[π1 (X)]2r is free.

Since L2k (Z[π1 (X)]) consists of stable isomorphism classes of quadratic forms, we deduce
that σ(f ) is independent of the choice of r. On the other hand, showing that σ(f ) is independent
of f 0 requires some more work, as we shall see below.
The next result answers the initial question of this chapter by providing a necessary and
sufficient condition for a degree one normal map to be normal bordant to a homotopy equivalence.
Theorem 2.2.24. Let k ≥ 3 and let f : M 2k → X be a degree one normal map.

1. The surgery obstruction σ(f ) ∈ L2k (Z[π1 (X)]) is well defined.

2. f is normal bordant to a homotopy equivalence if and only if σ(f ) = 0.

Proof. We start by proving the second assertion, assuming that the first has been proved. Note
that a homotopy equivalence g : M → X induces an isomorphism on all homotopy groups. It
follows that all πj (g) vanish and therefore so do all the surgery kernels. In particular, the surgery
obstruction of g vanishes. Consequently if f is normal bordant to a homotopy equivalence g, the
first assertion implies that σ(f ) = σ(g) = 0. Conversely, if σ(f ) = 0, then f is normal bordant
to a k-connected degree one normal map f 0 : M 0 → X with stably hyperbolic surgery kernel.
Theorem 2.2.18 implies that f 0 is normal bordant to a homotopy equivalence and therefore so is f .
This concludes our proof of the second assertion, and we must now prove the first.
Given two normal bordant degree one normal maps f0 : M0 → X and f1 : M1 → X, we
must show that σ(f0 ) = σ(f1 ). Let F : W 2k+1 → X × [0, 1] be a degree one normal cobordism
between f0 and f1 . Without loss of generality, we can assume that f0 , f1 , F are k-connected and
that the surgery kernels of f0 , f1 are free. An argument involving “handle substractions” shows
that Kk (W, ∂W ) can be assumed to be zero [Wal70, CLMa]. Let λ and µ denote intersection and
self-intersection numbers on ∂W . Since ∂W = M0 t −M1 , we deduce that

(Kk (∂W ), λ, µ) = (Kk (M0 ), λ0 , µ0 ) ⊕ (Kk (M1 ), −λ1 , −µ1 ).

To show that σ(f0 ) = σ(f1 ), it is therefore enough to prove that Kk (∂W ), λ, µ) is hyperbolic.
Thanks to Proposition 2.2.20, it is enough to find a lagrangian for (Kk (∂W ), λ, µ), i.e. G ⊂ Kk (∂W )
such that G = G⊥ and µ|G = 0. Since we arranged that Kk (W, ∂W ) = 0, we can consider the
following portion of the long exact sequence of the pair (W, ∂W ) for surgery kernels:
∂ i
Kk+1 (W, ∂W ) → Kk (∂W ) → Kk (W ) → 0 (2.6)

Set G := im(∂) = ker(i). We first show that G = G⊥ . This is a standard “half lives half dies”
argument. Consider the following commutative diagram:

Kk+1 (W, ∂W )
∂ / Kk (∂W ) i / Kk (W ) / 0. (2.7)
PD ∼ PD ∼ PD ∼
=
 =
 =

i∗ / K k (∂W ) δ / K k+1 (W, ∂W )
K k (W )

 i∗
 ∂∗

Kk (W )∗ / Kk (∂W )∗ / Kk+1 (W, ∂W )∗

As we have alluded to several times in the previous sections, the middle vertical composition
coincides with the adjoint λ• of λ. We write λ•r for the rightmost vertical compositions and λr
for the resulting pairing. Note that λ•r is an isomorphism: the bottom right evaluation map is an
isomorphism because Ki (W, ∂W ) = 0 for i ≤ k.
We first show that G⊥ ⊂ G i.e. that λ(∂(x), ∂(y)) = 0 for all x, y ∈ G. The diagram displayed
in (2.7) shows that λ(∂(x), ∂(y)) = λr (i ◦ ∂(x), y) for all x, y ∈ Kk+1 (W, ∂W ). The exactness

38
of the top row of (2.7) implies that i ◦ ∂(x) = 0, whence λ(∂(x), ∂(y)) = 0. We now prove
the reverse inclusion, namely G⊥ ⊂ G. Assume that a ∈ Kk (∂W ) satisfies λ(a, ∂(y)) = 0 for
all y ∈ Kk+1 (W, ∂W ); we must show that i(a) = 0. Using the diagram displayed in (2.7), we
deduce that 0 = λ(a, ∂(y)) = λr (i(a), y) for all y. The fact that λ•r is an isomorphism now implies
that i(a) = 0, as desired. This concludes the proof that G = G⊥ .
Next, we must prove that µ vanishes on G = im(∂). Given x ∈ Kk+1 (W, ∂W ), we must show
that µ(∂(x)) = 0. We shall always use the isomorphism Kk (∂W ) ∼ = Kk (M0 ) ⊕ Kk (M1 ). As f0
and f1 are k-connected, there are isomorphisms Kk (M0 ) ∼ = πk+1 (f0 ) and Kk (M1 ) ∼
= πk+1 (f1 )
(these were discussed in the proof of Proposition 2.2.2). Use ∂h to denote the connecting homo-
morphism in the long exact sequence of the pair for homotopy groups and consider the following
composition:
∂h ×∂h
Kk (M0 ) ⊕ Kk (M1 ) ∼
= πk+1 (f0 ) ⊕ πk+1 (f1 ) −→ πk (M0 ) ⊕ πk (M1 ).

Using this composition, observe that the class ∂(x) determines two (homotopy classes of) maps
u0 : S k → M0 and u1 : S k → M1 . Using Lemma 2.2.9, these can be assumed to be immersions.
Referring to [CLMb] for a discussion of signs, it follows that

µ(∂x) = µ(u0 ) + µ(u1 ).

We must therefore show that the self intersection points of the immersions u0 and u1 cancel. To
achieve this, consider the following diagram:

Kk+1 (W, ∂W )
∂ / Kk (∂W ) / Kk (W ) / 0. (2.8)
 
πk+1 (f0 ) ⊕ πk+1 (f1 ) / πk+1 (F )
∂ ×∂
 h h 
πk (M0 ) ⊕ πk (M1 )
i∗
/ πk (W )

Since the top row is exact, the definition of u0 , u1 and the commutativity of (2.8) implies that i∗ (u0 , −u1 )
vanishes in πk (W ). It follows that u0 and u1 are homotopic in W . We write this homotopy as

H : (S k × [0, 1], S k × {0, 1}) → (W, ∂W ).

To pair up the self intersections of u0 and u1 , we use H. More precisely, the set of double points
of H(S k × [0, 1]) with itself inside W is 1-dimensional and consists of circles (these are not relevant
to our purposes) and arcs. Each of these arcs pairs up a double point of u0 and with a double
point u1 of opposite sign. We deduce that µ(∂x) = µ(u0 ) + µ(u1 ) = 0, concluding the proof
that µ|G = 0, and therefore the proof that G is a Lagrangian of (Kk (∂W ), λ, µ). As we mentioned
above, this shows that σ(f0 ) and σ(f1 ) agree in L2k (Z[π1 (X)]), thus concluding the proof of the
theorem.

Let Nn (X) denote the set of normal bordism classes of degree one normal maps M n → X. The
first assertion of Theorem 2.2.24 states that the surgery obstruction gives rise to a well defined
map σ : N2k (X) → L2k (Z[π1 (X)]). In fact, the map σ fits into the so-called surgery exact sequence
which plays a crucial role in surgery theory [Wal70].
Next, we discuss variants and generalizations of Theorem 2.2.24. First, the results of this
chapter can be adapted to manifolds with boundary: surgeries are performed in the interior of
the manifold; normal maps (f, ∂f ) : (M 2k , ∂M ) → (X, ∂X) are defined as the closed case, but ∂f
is additionally required to be a homotopy equivalence (in fact, it is enough to assume that the
maps H∗ (∂ M f; Z) → H∗ (∂ X;
e Z) are isomorphisms); surgery below the middle dimension is proved
as in the closed case; the middle dimensional surgery kernel of a k-connected map still supports a
self intersection form µ and an intersection form λ. The assumption that H∗ (∂ Mf; Z) → H∗ (∂ X;
e Z)

39
is an isomorphism intervenes to ensure that λ is non-singular; the surgery obstruction can then
be defined similarly to the closed case (recall Definition 22), and the analogue of Theorem 2.2.24
holds in this setting. We refer to [Wal70, CLMa] for further details.
We conclude by mentioning two additional settings in which surgery theory works.
Remark 2.2.25. We mention two facts whose proofs each occupy several hundreds pages.

1. Due to foundational work of Kirby and Siebenmann, surgery theory works for topological
manifolds [KS77]. In particular, there is a way to make sense of normal maps and the surgery
obstruction in the topological category.
2. Making use of Freedman’s groundbreaking work on topological 4-manifolds, surgery theory
works for topological 4-manifolds, provided the group is “good” [FQ90]. For instance, in the
next subsection, we will permit ourselves to use Theorem 2.2.24 for topological 4-manifolds,
but the reader should keep in mind that this a very difficult result.

2.3 An application of surgery theory to knot theory

We give an application of surgery theory to knot theory: we prove Freedman’s result that Alexan-
der polynomial one knots are topologically slice. We assume no prior background in knot theory,
but take a “high dimensional approach” to the subject. In particular, we choose not to define the
Alexander polynomial but instead provide a convenient reformulation of the Alexander polyno-
mial one condition. References for the proof of Freedman’s theorem include [FQ90, FT05], while
introductions to knot theory include [Lic97, Rol76].
A knot K is a smooth embedding S 1 → S 3 . While knot theory studies knots up to ambient
isotopy, our application of surgery is concerned with knot concordance.
Definition 23. A knot is topologically slice if it bounds a locally flat embedded flat disc in D4 .

We will neither discuss the classical results in knot concordance (i.e. in the study of slice
knots), nor the distinction between topologically slice knots and smoothly slice knots. Instead, we
describe a result of Freedman which uses surgery theorem to provide a sufficient condition for a
knot to be topologically slice.
We do however need some basic homological facts on knot exteriors, and the next exercise is
an application of Alexander duality (or the Mayer-Vietoris exact sequence).
Exercise 2.3.1. Let K be a knot with solid torus neighborhood νK, and let XK := S 3 \ νK be
its exterior. Show that Hi (XK ; Z) = Z for i = 0, 1 and vanishes otherwise, i.e. XK is a homology
circle. Show that if K is sliced by a disc D, then the slice disc exterior ND := D4 \ νD is also a
homology circle.

We study the boundary of the slice disc exterior ND . Observe that ∂ND is obtained by gluing
a solid torus ST to XK , identifying the meridian of ST with the (0-framed) longitude of K. 6
In other words, ∂ND obtained by performing surgery on S 3 along K: we remove νK = K × D2
from S 3 and glue ST = D2 × S 1 . We refer to MK as the 0-framed surgery along K.
The next exercise describes the first homology of MK .

Exercise 2.3.2. Show that π1 (MK ) = π1 (XK )/hλK i, where hλK i denotes the normal subgroup
generated by the longitude of K. Show that H1 (MK ; Z) = H1 (XK ; Z) ∼ = is generated by the
meridian µK of K.
6 The longitude λ
K and meridian µK of K are two simple closed curves in ∂(νK) such that µK is non-separating
and bounds a disc in νK, and λK is homologous to K in νK and nullhomologous in XK .

40
 We work towards the statement Freedman’s theorem which provides a necessary condition for a
knot to be topologically slice. Using Exercise 2.3.2, H1 (MK ; Z) ∼
= ZhµK i, and we consider the map
ab
π1 (MK ) → H1 (MK ; Z) → Z mapping the meridian of K to 1. The cover M cK corresponding to
the kernel of this map is endowed with an action of Z, and therefore H∗ (M cK ; Z) has the structure
of a Z[Z]-module. It is helpful to think of Z[Z] as the ring Z[t±1 ] of Laurent polynomials with
integer coefficients. In what follows, we shall write

H∗ (MK ; Z[Z]) := H∗ (M
cK ; Z),

H ∗ (MK ; Z[Z]) := Hc∗ (M

cK ; Z).

The Z[Z]-module H1 (MK ; Z[Z]) is called the Alexander module of K. We say that K has Alexan-
der polynomial 1 if H1 (MK ; Z[Z]) = 0.
We can state Freedman’s theorem.
Theorem 2.3.3. If K has Alexander polynomial 1, then K is topologically slice.

We sketch the usual definition of the Alexander polynomial ∆K (t) for the interested reader
who might be flustered by our unconventional definition of Alexander polynomial one knots.
Remark 2.3.4. The Alexander polynomial ∆K (t) is usually defined as the order of H1 (XK ; Z[Z])
(a Mayer-Vietoris exact sequence shows that MK and XK can be used interchangeably). ∆K (t)
is only defined up to multiplication by ±tn and it is not hard to show that H1 (MK ; Z[Z]) = 0 if
and only if ∆K (t) = 1 up to multiplication by ±tn . This justifies the terminology we used above.

The remainder of this subsection is devoted to the proof of Theorem 2.3.3. Recall from Exer-
cise 2.3.1 that ND has the homology of a circle. The basic surgery theoretic idea to slice a knot K
is to start from a 4-manifold W whose boundary is MK , and to perform surgery on W in order
to get a homology circle.
In the topological category, this is nearly enough to guarantee the sliceness of K.
Proposition 2.3.5. A knot is topologically slice if MK bounds a topological 4-manifold V such
that the following conditions hold:

1. the inclusion induced map H1 (MK ; Z) → H1 (W ; Z) is an isomorphism;

2. π1 (V ) is normally generated by the image of a meridian of K;
3. H2 (V ; Z) = 0.

Proof. Recall that the 0-framed surgery MK can be obtained from the knot exterior XK = S 3 \νK
by attaching a solid torus: MK = XK ∪ (D × S 1 ), where the meridian ∂D × {operatornamept} is
identified with the longitude of K, and the longitude {0} × S 1 is identified with the meridian µK
of K. By assumption, π1 (V ) is normally generated by the meridian of K. Adding a 4-dimensional
2-handle D2 × D2 to V along this meridian therefore leads to a simply-connected 4-manifold N
whose boundary is S 3 . In particular, we have H1 (N ; Z) = 0.
We show that Hi (N, Z) = 0 for i > 0. We already know that H1 (N ; Z) = 0, and H4 (N ; Z) = 0
since N has non-empty boundary. The third assumption states that H2 (V ; Z) = 0, and a Mayer-
Vietoris argument gives H2 (N ; Z) = H2 (V ; Z) = 0. Using the first assumption and the long exact
sequence of the pair (V, ∂V ), we deduce that H1 (V, ∂V ; Z) = 0. It then follows from Poincaré
duality and the universal coefficient theorem that H3 (V ; Z) = 0. A Mayer-Vietoris argument
shows that H3 (N ; Z) = H3 (V ; Z) = 0.
We have shown that N is a Z-homology 4-ball with boundary S 3 . Freedman’s work now implies
that N is homeomorphic to D4 [FQ90]. By construction, D ⊂ N is the desired slice disk for K.
This concludes the proof of the proposition.

Next, we outline the remainder of the proof of Theorem 2.3.3:

41
 1. We construct a space X ' S 1 such that (X, MK ) is a4-dimensional Poincaré pair, i.e. there
is a relative homology class [X, MK ] ∈ H4 (X, MK ; Z) such that the relative cap product
with [X, MK ] yields the following isomorphisms:

∩ [X, MK ] : H i (X; MK ; Z[Z]) → H4−i (X; Z[Z]),

∩ ∂[X, MK ] : H i (MK ; Z[Z]) → H3−i (MK ; Z[Z]),

where ∂ : H4 (X, MK ; Z) → H3 (MK ; Z) is the connecting homomorphism in the long exact

sequence of the pair (X, MK ).
2. We construct a topological 4-manifold W with boundary MK and a degree one normal map

f : (W, MK ) → (X, MK ).

3. We prove Theorem 2.3.3: starting from f , we construct another degree one normal map
f 0 : (W 0 , MK ) → (X 0 , MK ) such that ∂W 0 = ∂W = MK , and f 0 has vanishing surgery ob-
struction. Theorem 2.3.3 (and Remark 2.2.25) then produces manifold V with boundary MK
which satisfies the three properties of Proposition 2.3.5, thus proving that K is slice.

We now carry out this program, step by step.

Proposition 2.3.6. If a knot K has Alexander polynomial 1, then there is CW-complex X that
is homotopy equivalent to S 1 and such that (X, MK ) is a 4-dimensional Poincaré pair.

Proof. Recall from Exercise 2.3.2 that H1 (MK ; Z) = ∼ Z is generated by the meridian µK of K.
Consider the map ϕ : H1 (MK ; Z) → Z, µK → 1. Some algebraic topology gives rise to an iso-
morphism HomZ (H1 (MK ; Z), Z) ∼ = [X, S 1 ] (here [X, S 1 ] refers to homotopy classes of maps and
not to a relative homology class), and we therefore obtain a map MK → S 1 corresponding to ϕ.
Let X be the mapping cylinder of this map. We have thus obtained an inclusion ι : MK ,→ X,
where X is homotopy equivalent to S 1 .
We must now show that (X, MK ) is a 4-dimensional Poincaré pair. Consider the long ex-
act sequence of (X, MK ) with Z coefficients. Since X is homotopy equivalent to S 1 , we know
that Hi (X; Z) = 0 for i ≥ 2. The connecting homomorphism in the long exact sequence therefore
∼
=
induces an isomorphism ∂ : H4 (X, MK ; Z) → H3 (MK ; Z). Since MK is closed, H3 (MK ; Z) = ∼ Z is
generated by the fundamental class [MK ] of MK . We set [X, MK ] := ∂ −1 ([MK ]) and check that
capping with this class gives rise to the required isomorphisms:

∩ [X, MK ] : H i (X; MK ; Z[Z]) → H4−i (X; Z[Z]), (2.9)

i
∩ ∂[X, MK ] : H (MK ; Z[Z]) → H3−i (MK ; Z[Z]).

The second isomorphism holds by definition of [X, MK ]: by definition, we have ∂[X, MK ] = [MK ]
and, since MK is a manifold, it satisfies Poincaré duality. We must therefore only check the
first isomorphism. Since the Z-cover of S 1 is R, we deduce that Hi (X; Z[Z]) = 0 for i 6= 0
and H i (X; Z[Z]) = 0 for i 6= 1. It follows that the bottom and top horizontal maps in the
following diagram are isomorphisms:
∼
/ H 3 (MK ; Z[Z]) = / H 4 (X, MK ; Z[Z]) /
∼
= ∩[MK ] ∩[X,M ]
 ∼  K

/ H0 (MK ; Z[Z]) = / H0 (X; Z[Z]) /

The commutativity of this diagram implies that (2.9) is an isomorphism for i = 4. We now focus
on the cases i < 4: here, since H4−i (X; Z[Z]) = 0, it is enough to show that H i (X, MK ; Z[Z]) = 0.
Consider the long exact sequence of the pair (X, MK ) in cohomology:

→ H i−1 (X; Z[Z]) → H i−1 (MK ; Z[Z]) → H i (X, MK ; Z[Z]) → H i (X; Z[Z]) → H i (MK ; Z[Z]) →

42
For i = 0, we see that H 0 (X; Z[Z]) = 0 implies H 0 (X, MK ; Z[Z]) = 0. For i = 1, we use the
exactness together with two facts: H 0 (MK ; Z[Z]) ∼
= H3 (MK ; Z[Z]) = 0 (since Z-coverings are
non-compact), and H 1 (X; Z[Z]) → H 1 (MK ; Z[Z]) is an isomorphism (since MK → X induces
an isomorphism on the fundamental groups). For i = 2, we recall that H 2 (X; Z[Z]) = 0, and
use a second time that H 1 (X; Z[Z]) → H 1 (MK ; Z[Z]) is an isomorphism. Finally, for i = 3, we
have H 3 (X; Z[Z]) = 0 and H 2 (MK ; Z[Z]) ∼= H1 (MK ; Z[Z]) = 0 by duality and the Alexander
polynomial one condition. We have therefore shown that (2.9) is an isomorphism for each i,
concluding the proof of the proposition.

We move on to the second step of our program.

Proposition 2.3.7. If a knot K has Alexander polynomial 1 and (X, MK ) is the Poincaré pair
constructed in Proposition 2.3.6, then there exists a degree one normal map of pairs

f : (W, MK ) → (X, MK ).

Proof. Our goal is to find a 4-manifold W with ∂W = MK , a map f : W → X that extends the
map ι : MK → X and a stable trivialisation of T W ⊕ f ∗ (ξ), where ξ is a vector bundle over X.
Let ξ 0 be a trivial bundle over S 1 . Since X ' S 1 , this bundle pulls back to a trivial vector bundle ξ
over X. As a consequence, the difficulty is to find a pair (W 4 , f ) with ∂W = MK , a map f that
extends ι : MK → X and a stable trivialization of T W .
Given a space Y and an integer n, there is a so-called framed bordism group Ωfr n (Y ) whose
elements are equivalence classes of triples (M, θ, f ), where M is an n-manifold, θ is a stable
trivialisation of T M , and f : M → Y is a map. A triple (M, θ, f ) represents zero in Ωfr n (Y ) if M
bounds an (n + 1)-manifold W , the map f extends to W , and the stable trivialisation θ extends
to a stable trivialisation of T W . If we disregard the space Y and the maps to it, then we get a
group Ωfr n . Since we already have a map ι : MK → X, the idea of the proof is to define a stable
trivialisation θ of T MK such that [MK , θ, ι] is zero in Ωfr 3 (X): by definition of this group and
recalling the previous paragraph, this will immediately yield a normal nullbordism W → X.
We first endow MK with a stable trivialisation of its tangent bundle. Since MK is an orientable
3-manifold, its tangent bundle is trivial, and we can choose (stable) framing θ0 of T MK . The
pair (MK , θ0 ) therefore represents an element in Ωfr 0
3 . Some work shows that θ can be modified,
in order to produce a (stable) framing θ so that (MK , θ) represents zero in Ωfr 3 [FQ90, proof of
Lemma 11.6.B], as well as [CP14, proof of Lemma 3.3] (note that a nullbordant stable framing
of MK can also be constructed explicitly). We must now argue that (MK , θ, ι) represents the zero
element in Ωfr 3 (X).
Recalling that X is homotopy equivalent to S 1 , and using the Atiyah-Hirzebruch spectral
∼
=
sequence [CF64], it is known that there is an isomorphism Ωfr 1 fr fr
3 (S ) → Ω3 ⊕ Ω2 . The image
fr ∼
of [MK , θ, ι] in the first summand is [MK , θ], while its image in Ω2 = Z2 can be shown to be the
so-called “Arf invariant” of K:
∼
=
Ωfr fr fr
3 (X) → Ω3 ⊕ Ω2
[MK , θ, ι] 7→ ([MK , θ], Arf(K)).

It is a fairly well known fact in knot theory that the value of Arf(K) is determined mod 8 by the
value of the Alexander polynomial at −1 [Lic97, Theorem 10.7]. In particular, if K has Alexander
polynomial 1, then K has vanishing Arf invariant. We have therefore established that (MK , θ, ι)
represents zero in Ωfr3 (X) and, as we explained above, this is enough to guarantee the existence of
the desired normal nullbordism f : W → X.
It only remains to prove that f : W → X has degree one. We must show that f maps the
fundamental class [W, MK ] to the class [X, MK ] defined in Proposition 2.3.6. Since the map

43
f : (W, MK ) → (X, MK ) extends ι : MK → X, the following diagram commutes:

H4 (W, MK )
∂ / H3 (MK )
f =
 ∗ 
H4 (X, MK )
∂ / H3 (MK ).

We established in Proposition 2.3.6 that the bottom map is an isomorphism, and we defined the
class [X, MK ] as ∂ −1 [MK ]. This shows that f has degree one and concludes the proof of the
proposition.

We can now conclude the proof of Theorem 2.3.3.

Proof of Theorem 2.3.3. Let f : (W, MK ) → (X, MK ) be the degree one normal map constructed
in Proposition 2.3.7. Using surgery below the middle dimension, we can arrange that f : W → X
is 2-connected. We saw in Theorem 2.2.24 (and Remark 2.2.25) that since Z is a good group, f is
normal bordant to a homotopy equivalence if and only if the surgery obstruction σ(f ) ∈ L4 (Z[Z])
is zero. Using Remark 2.2.23, we know that L4 (Z[Z]) is isomorphic to L4 (Z) = 8Z and is detected
by the usual signature.
Since we are in the topological category, Freedman proved that there is a simply connected
closed manifold E8 whose intersection form is the E8 form. In fact, there is a degree one normal
map E8 → S 4 whose surgery obstruction generates L4 (Z) ∼ = 8Z. Taking the connected sum of
the degree one normal map f : W → X with |σ(f )| copies of ±E8 → S 4 , we obtain a new degree
one normal map f 0 : W 0 → X 0 , where W 0 still has boundary MK (since E8 is closed), but f 0 has
vanishing surgery obstruction. We deduce that (W 0 , MK ) → (X 0 , MK ) is normal bordant to a
homotopy equivalence (V, MK ) → (X 0 , MK ).
Since MK = ∂V , in order to conclude the proof of the theorem, it only remains to show
that V satisfies the homological properties of Theorem 2.3.5. Since V is homotopy equivalent
to X 0 , it has the same homology and homotopy groups as X 0 = X#S 4 , where X ' S 1 . This
immediately implies that H2 (V ; Z) = 0. Since, by construction, the map ι : MK → X ' S 1
induces the isomorphism H1 (MK ; Z) → H1 (S 1 ; Z) = π1 (S 1 ), µK → 1 on homology, we deduce
that the group π1 (V ) = π1 (X) ∼= π1 (S 1 ) ∼
= Z is (normally) generated by the meridian of K and
that H1 (MK ; Z) → H1 (V ; Z) is an isomorphism. We have therefore proved that MK bounds a
topological 4-manifold V which satisfies the three properties listed in Proposition 2.3.5. Applying
this proposition, we deduce that K is topologically slice, concluding the proof of the theorem.

44
Bibliography

[Ada93] Masahisa Adachi. Embeddings and immersions, volume 124 of Translations of Mathe-
matical Monographs. American Mathematical Society, Providence, RI, 1993. Translated
from the 1984 Japanese original by Kiki Hudson.
[Bre93] Glen E. Bredon. Topology and geometry, volume 139 of Graduate Texts in Mathematics.
Springer-Verlag, New York, 1993.
[CF64] P. E. Conner and E. E. Floyd. Differentiable periodic maps. Ergebnisse der Mathematik
und ihrer Grenzgebiete, N. F., Band 33. Academic Press Inc., Publishers, New York;
Springer-Verlag, Berlin-Göttingen-Heidelberg, 1964.
[CLMa] Diarmuid Crowley, Wolfgang Lueck, and Tibor Macko. Surgery theory: Foundations
(09-2015 notes). http://www.mat.savba.sk/ macko/surgery-book.html.
[CLMb] Diarmuid Crowley, Wolfgang Lueck, and Tibor Macko. Surgery theory: Foundations
(11-2018 notes). http://www.mat.savba.sk/ macko/surgery-book.html.
[CP14] Jae Choon Cha and Mark Powell. Nonconcordant links with homology cobordant zero-
framed surgery manifolds. Pacific J. Math., 272(1):1–33, 2014.
[FQ90] Michael H. Freedman and Frank Quinn. Topology of 4-manifolds, volume 39 of Princeton
Mathematical Series. Princeton University Press, Princeton, NJ, 1990.
[FT05] Stefan Friedl and Peter Teichner. New topologically slice knots. Geom. Topol., 9:2129–
2158, 2005.
[Gei09] Hansjörg Geiges. A contact geometric proof of the Whitney-Graustein theorem. Enseign.
Math. (2), 55(1-2):93–102, 2009.
[GP74] Victor Guillemin and Alan Pollack. Differential topology. Prentice-Hall, Inc., Englewood
Cliffs, N.J., 1974.
[GS99] Robert E. Gompf and András I. Stipsicz. 4-manifolds and Kirby calculus, volume 20
of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI,
1999.
[Hat] Allen Hatcher. Vector bundles and k-theory. https://pi.math.cornell.edu/
~hatcher/VBKT/VB.pdf.
[Hat02] Allen Hatcher. Algebraic topology. Cambridge University Press, Cambridge, 2002.
[Hir59] Morris W. Hirsch. Immersions of manifolds. Trans. Amer. Math. Soc., 93:242–276, 1959.
[Hir76] Morris W. Hirsch. Differential topology. Springer-Verlag, New York-Heidelberg, 1976.
Graduate Texts in Mathematics, No. 33.
[Hus94] Dale Husemoller. Fibre bundles, volume 20 of Graduate Texts in Mathematics. Springer-
Verlag, New York, third edition, 1994.

45
 [Ker60] Michel A. Kervaire. A manifold which does not admit any differentiable structure.
Comment. Math. Helv., 34:257–270, 1960.
[KL05] Matthias Kreck and Wolfgang Lück. The Novikov conjecture, volume 33 of Oberwolfach
Seminars. Birkhäuser Verlag, Basel, 2005. Geometry and algebra.

[KS77] Robion C. Kirby and Laurence C. Siebenmann. Foundational essays on topological

manifolds, smoothings, and triangulations. Princeton University Press, Princeton, N.J.;
University of Tokyo Press, Tokyo, 1977. With notes by John Milnor and Michael Atiyah,
Annals of Mathematics Studies, No. 88.
[L0̈2] Wolfgang Lück. A basic introduction to surgery theory. In Topology of high-dimensional
manifolds, No. 1, 2 (Trieste, 2001), volume 9 of ICTP Lect. Notes, pages 1–224. Abdus
Salam Int. Cent. Theoret. Phys., Trieste, 2002.
[Lic97] W. B. Raymond Lickorish. An introduction to knot theory, volume 175 of Graduate
Texts in Mathematics. Springer-Verlag, New York, 1997.
[MS74] John W. Milnor and James D. Stasheff. Characteristic classes. Princeton University
Press, Princeton, N. J.; University of Tokyo Press, Tokyo, 1974. Annals of Mathematics
Studies, No. 76.
[Now] Tahl Nowik. Lecture notes: immersions of surfaces in 3-space. http://u.math.biu.
ac.il/~tahl/notes.pdf.

[Ran73a] A. A. Ranicki. Algebraic L-theory. I. Foundations. Proc. London Math. Soc. (3), 27:101–
125, 1973.
[Ran73b] A. A. Ranicki. Algebraic L-theory. II. Laurent extensions. Proc. London Math. Soc. (3),
27:126–158, 1973.
[Ran81] A. Ranicki. Exact sequences in the algebraic theory of surgery, volume 26 of Mathematical
Notes. Princeton University Press, Princeton, N.J.; University of Tokyo Press, Tokyo,
1981.
[Ran02] Andrew Ranicki. Algebraic and geometric surgery. Oxford Mathematical Monographs.
The Clarendon Press, Oxford University Press, Oxford, 2002. Oxford Science Publica-
tions.

[Rol76] Dale Rolfsen. Knots and links. Publish or Perish, Inc., Berkeley, Calif., 1976. Mathe-
matics Lecture Series, No. 7.
[Sco05] Alexandru Scorpan. The wild world of 4-manifolds. American Mathematical Society,
Providence, RI, 2005.

[Sha69] Julius L. Shaneson. Wall’s surgery obstruction groups for G × Z. Ann. of Math. (2),
90:296–334, 1969.
[Sma59] Stephen Smale. The classification of immersions of spheres in Euclidean spaces. Ann.
of Math. (2), 69:327–344, 1959.
[Sma62] S. Smale. On the structure of manifolds. Amer. J. Math., 84:387–399, 1962.

[Tu11] Loring W. Tu. An introduction to manifolds. Universitext. Springer, New York, second
edition, 2011.
[vdB59] F. van der Blij. An invariant of quadratic forms mod 8. Nederl. Akad. Wetensch. Proc.
Ser. A 62 = Indag. Math., 21:291–293, 1959.

46
[Wal70] C. T. C. Wall. Surgery on compact manifolds. Academic Press, London-New York, 1970.
London Mathematical Society Monographs, No. 1.
[Wal16] C. T. C. Wall. Differential topology, volume 156 of Cambridge Studies in Advanced
Mathematics. Cambridge University Press, Cambridge, 2016.

[Whi37] Hassler Whitney. On regular closed curves in the plane. Compositio Math., 4:276–284,
1937.

47