18.966 di erential topology lecture notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966...

39
18.966 Differential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul Seidel in the Spring term of 2016 at MIT. The following notes were taken by Yuchen Fu. They have not been very carefully proofread and all errors are the notetakers’. The daily homework problems are included under the tag “problem.” Contents 1 Feb 03 3 2 Feb 08 4 3 Feb 10 6 4 Feb 15 7 5 Feb 16 8 6 Feb 22 10 7 Feb 24 12 8 Feb 29 13 9 March 02 15 10 March 07 17 11 March 09 18 12 March 14 19 13 March 28 20 14 March 30 22 15 April 4 24 16 April 6 26 17 April 11 27 18 April 13 29 19 April 20 30 20 April 25 32 1

Upload: others

Post on 20-Jul-2020

7 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Differential Topology Lecture Notes

May 15, 2016

This class was taught by Prof. Paul Seidel in the Spring term of 2016 at MIT. The following notes weretaken by Yuchen Fu. They have not been very carefully proofread and all errors are the notetakers’. Thedaily homework problems are included under the tag “problem.”

Contents

1 Feb 03 3

2 Feb 08 4

3 Feb 10 6

4 Feb 15 7

5 Feb 16 8

6 Feb 22 10

7 Feb 24 12

8 Feb 29 13

9 March 02 15

10 March 07 17

11 March 09 18

12 March 14 19

13 March 28 20

14 March 30 22

15 April 4 24

16 April 6 26

17 April 11 27

18 April 13 29

19 April 20 30

20 April 25 32

1

Page 2: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes CONTENTS

21 April 27 34

22 May 02 35

23 May 04 36

24 May 09 38

Page 2

Page 3: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 1 FEB 03

1 Feb 03

Definition 1. Let U, V ⊂ Rn be open subsets. A smooth map f : U → V is a diffeomorphism if it has asmooth inverse.

Example 1. (−1, 1) and R are diffeomorphic subsets of R by the map tan(πx/2). On the other hand,x3 : R→ R is not a diffeomorphism.

Theorem 1.1 (The Inverse Mapping Theorem). Let U, V ∈ Rn, f : U → V such that x0 ∈ U, y0 = f(x0) ∈V . If Dfx0 : Rn → Rn is an invertible linear map, then there are open subsets U , V such that f |U : U → Vis a diffeomorphism.

Note that in order to prove this, the inductive procedure that produces the Taylor expansion only givesthe Taylor series, which doesn’t necessarily give the inverse function. Thus we need a better control (e.g.using contraction mapping theorem).

Question 1. Is the inverse mapping theorem true in the world of holomorphic maps?

The answer is yes. One can first prove that the real inverse is holomorphic. Or one can use locallyconvergent Taylor series.

Question 2. Is the inverse mapping theorem true for complex polynomials where “open subset” is true inthe Zariski topology?

No. Consider C− 0 → C− 0 given by z2. It sends 1 to 1. The inverse (square root) is not going tobe a polynomial. This problem can be dealt with by introducing the etale topology.

Theorem 1.2 (Local Submersion). Let U ⊂ Rm, V ⊂ Rn. Suppose 0 7→ 0 by f : U → V . Further, supposethat Df0 is onto, then after shrinking U , there is a diffeomorphism ϕ : U → U (where U 3 0) mapping 0 to0 such that f ϕ is a (surjective) linear map.

Note This is usually written in a more technical form as the implicit function theorem.Note that this in particular shows that Df was surjective in a neighborhood; of course this is no problem

since being surjective is an open condition. Do we have an injective counterpart? Yes.

Theorem 1.3 (Local Immersion). Same situation as above. Suppose that Df0 is injective, then aftershrinking U and V , there is a diffeomorphism ψ : V → V mapping 0 to 0 such that ψ f is an (injective)linear map.

Both are proven by reducing to the inverse mapping theorem by adding extra dimensions to source /target.

And here’s a more generalized version, which answers the question “when does a function look like alinear one?”

Theorem 1.4 (Theorem of Constant Rank). Same situation as above. Suppose that rankDfx is constant.Then after shrinking U and V , there are diffeomorphisms ϕ : U → U,ψ : V → V (both preserving 0) suchthat ψ f ϕ is a linear map.

Definition 2. Smooth f : U → V . We say that y ∈ V is a regular value of f if Dfx is onto for all x 7→ y.

Definition 3. f is a submersion of Dfx is onto for all x→ U , i.e. all values are regular values.

Definition 4. f is an immersion if Dfx is injective for all x ∈ U .

Definition 5. If f is an immersion, injective and proper, then it’s call an embedding.

Page 3

Page 4: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 2 FEB 08

Note Proper is as a map U → V . We define it as if (xi) doesn’t have a convergent subsequence, then(f(xi)) also doesn’t have a convergent subsequence. The image of an embedding is a closed subset of V .

Example 2. Take f : (0, 2π) → C, f(t) = eit. This is not an embedding (it breaks properness). However,f : (0, π)→ z | im(z) > 0 by f(t) = eit is an embedding.

Definition 6. Let U ⊂ Rn be an open subset. A closed subset M ⊆ U is called a submanifold if around eachpoint of M there exists a local diffeomorphism which transforms M into a linear subspace.

Corollary 1. If y0 is a regular value of f : U → V then the preimage of y0 is a submanifold.

Of course, if f is a submersion, then this is true for all y0.

Note dim(f−1(y0)) = m− n where U ⊂ Rm and V ⊂ Rn.

Corollary 2. If f : U ⊂ Rm → V ⊂ Rn is an embedding, then f(U) ⊂ V is a submanifold of dimension m.

A general strategy to prove something is a submanifold is either showing it’s a preimage or showing thatit’s an image.

Note Linear maps can always be factored into a surjection followed by an injection (take the middle spaceto be the image). Correspondingly, a map of constant rank can be locally factored into a submersion followedby an immersion.

Here’s an application.

Theorem 1.5 (Local Retraction). Let U ⊂ Rn be an open subset, and f : U → U a smooth map withf(f(x)) = f(x) (it’s called a retraction), then f(U) is a submanifold of U .

Proof. We have Dff(x) Dfx = Dfx. In particular, if x ∈ f(U), Dfx is a projection. Projections have onlyeigenvalues 0 and 1. By continuity of eigenvalues, they are locally constant on f(U). But the image is notan open subset, so we can’t apply the constant rank theorem yet.

For general x ∈ U , rankDfx = rank(Dff(x) Dfx) ≤ rankDff(x). On the other hand, if x is sufficientlyclosed to f(x), then rankDfx ≥ rankDff(x) due to semicontinuity of ranks. This implies that rankDfx =rankDff(x) if x is sufficiently close to f(x). Thuse in a neighborhood of f(U), rankDf is locally constant.

Now apply theorem of constant rank to make our conclusion.

That was also another way of showing that things are submanifolds.

Problem 1. Let f : Rn → R, f(0) = 0, Df0 = 0, D2f is everywhere positive definite. Prove that f−1(y) fory > 0 is a submanifold of dimension n− 1 (show they are regular values) (some convexity property).

Problem 2. Let Uf−→ U where 0 7→ 0 be a smooth map such that f(0) = 0, f(f(. . . (f(x))) . . .) = x (pth

order) for all x. Find a local diffeomorphism ψ such that ψ f ψ−1 is linear. (Applying inverse functiontheorem won’t work). (ψ is tricky to construct. Try to average things.)

2 Feb 08

Regarding the problem 2 from last time. First thing to notice that ψ f ψ−1 is of finite order p. (This isthe start of study of nonlinear symmetry. It in fact holds for any finite group or even compact Lie group.)Also, it conjugates to Df0 (since ψ f ψ−1 = D(ψfψ−1)0 = Dψ0 Df0 Dψ−1

0 ), so we actually aim tomake ψ f ψ−1 = Df0. so what we want is ψ f = Df0 ψ. Now we need to use the finite order property.(This won’t hold in general because the dynamic of a fixed point is more complicated than a linear map,although things like Hartman-Grobman theorem says that there are some things to be said.)

ψ(x) = x won’t satisfy this, but it satisfies up to first order (derivative is okay). What aboutDf0(f−1(x))?It’s just as equally good. Now there are p different choices Df i0(f−i(x)). Now of course each gets sent tonext by Df0 f−1, so we take ψ(x) = 1/p(x+Df0(f−1(x))) + . . .+Dfp−1

0 (f1−p(x))). Then Df0(ψ(x)) =1/p(Df0(x) + Df2

0 (f−1(x)) + . . . + f(x)), and ψ(f(x)) = 1/p(f(x) + Df0(x) + . . . + Dfp−10 (f2−p(x))).

Page 4

Page 5: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 2 FEB 08

It remains to check that ψ is a local diffeomorphism by applying the inverse function theorem. Dψ0 =1/p(I +Df0Df

−10 + . . .) = I.

Of course this extends to complex geometry (nothing about real is used), but of course if your charac-teristic is p then we’re not good.

Now let’s go back to schedule. Recall this from last lecture: Let U ⊆ Rm open, f : U → U satisfiesf(f(x)) = f(x), then M = f(U) is a submanifold.

Example 3. Every matrix A ∈ GLn(R) can be uniquely written as A = PQ where P is symmetric positivedefinite, and Q is orthogonal.

In representation theory this is generalized by the Cartan decomposition. Note that AAt = PQQtP t =PP t = P 2 =⇒ P = (AAt)1/2, and correspondingly, Q = P−1A.

Taking the square root of a positive definite matrix is a differentiable (in fact smooth) (with respect toits coefficients) operation.

Three proofs:

1. Squaring is a differentiable thing and we invert it. Boring.

2. p(A) for p is a polynomial is certainly differentiable. Fix A0, Choose p to approximate square root to

high order at the eigenvalues of A0. Then p(A) ≈√A for A ≈ A0 to high order.

3. Functional calculus (Cool stuff). Remember for λ > 0, λ1/2 =1

2πi

∮γ

dz

z − λz1/2. Similarly, if γ contains

all eigenvalues of A, then A1/2 =1

2πi

∮γ

dz

z −Az1/2.

(To see this, suppose A is diagonal, then easy to see this is true (integrate on each entry). But this isindependent of basis, so this must be true in general.)

But the right hand side is clearly differentiable in A.

Hence f(A) = (AAt)1/2A is a differentiable map, and f(f(A)) = f(Q) = f(A). Hence f(GLn(R)) =

On(R) is a submanifold of GLn(R). (Since it’s closed in Rn2

, it is also a submanifold of that.) Another proofis to consider the orthogonal group as the preimage of f(X) = XXt at I and apply the implicit functiontheorem.

And here is an converse:

Application Let U ⊆ Rm open, M ⊆ U a submanifold. Then there is an open M ⊆ U ⊆ U and a smoothmap f : U → U such that f(f(x)) = f(x), f(U) = M . Locally this is no problem: we just project to thelinear subspace. But how does this make sense globally?

Definition 7. Let M ⊆ U ⊆ Rm be a submanifold. The tangent space of M at x0 ∈M is the linear subspace

TMx0⊆ Rm defined as follows: TMx0

= dcdt|t=0 | c : (−ε, ε)→ Rm, c(0) = x0, c(t) ∈M ∀t.

It’s not immediately obvious that this is a linear space; this is actually a feature, because this also workswith objects with singularities (we then get something called tangent cones), in which case they are notlinear spaces.

A second definition goes as follows: TMx0= Dϕ0(Rn×0) | ϕ : Rm → Rm isalocaldiffeomorphism, ϕ(0) =

x0,Rn × 0 7→ M locally. It’s clearly a linear space, but not clear it’s well-defined (it seems to dependson a choice of ϕ).

Third definition (Zariski tangent space):

TMx0= X ⊆ Rm | if f is a function defined near x0 and which vanishes on M , then Dfx0

(X) = 0.

All three are equivalent in the case of manifolds. In general they are different.

Example 4. If f : U → V , y ∈ V a regular value, M = f−1(y) 3 x, then TMx = ker(Dfx).

Example 5. If f : U → V is an embedding, M = f(U) and y = f(x), then TMy = im(Dfx).

Page 5

Page 6: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 3 FEB 10

Now we return to the question of retractions.

Lemma 1 (Orthogonal Retraction). If M ⊆ Rm is a submanifold, x0 ∈ M , then there exists small opensubsets x0 ∈ U ⊆ U ⊆ Rm such that for each x ∈ U there exists a unique x ∈M ∩ U such that x− x⊥TMx.

This implies the existence of global orthogonal retraction, as the map to x is unique.

Proof. Suppose dim(M) = n. Find a local parametrization i.e. some 0 ∈ V ⊆ Rn open and g : Vimmersion−−−−−−→

Rm, g(0) = x0, g(V ) = M locally near x0. Possibly after shrinking V we can find ξ1, . . . , ξm−n : V → Rnsuch that ξ1(y), . . . , ξm−n(y) is a basis for TM⊥g(y). Then define g : V × Rm−n → Rm with g(y, t) =

g(y) +∑i

tiξi(y). Then g is a local diffeomorphism by the inverse mapping theorem. So if x = g(y, t) then

set x = g(y, 0).

Definition 8. Let U ⊆ Rm, V ⊆ Rn be open, N ⊆ V a submanifold, and f : U → V a smooth map. Wesay that f is transverse to N if for all y ∈ N , x ∈ f−1(y), im(Dfx) + TNy = Rn.

Example 6. If m + dim(N) < n, then transversality says that f(U) ∩ N =. If N = y, transversalitymeans y is a regular value of f . If N = V , transversality is automatic. If f is a submersion, it’s transverseto everything.

Problem 3. If f is transverse to N , then M = f−1(N) is a submanifold of U .

3 Feb 10

(Something about Grassmannian. Didn’t write it down.)

Lemma 2 (Reduction). Take 0 ∈ U ⊆ Rm open, f : U → Rn smooth, f(0) = 0 and rank Df0 = r, then thereare local diffeomorphisms ϕ,ψ around 0 such that (ψ f ϕ−1)(x1, . . . , xm) = (x1, . . . , xr, somethingelse) and(ψ f ϕ−1)(x1, . . . , xr, 0, . . . , 0) = (x1, . . . , xr, 0, . . . , 0).

See proof of constant rank theorem in the textbook.Now let’s talk about Banach spaces. There is a notion of differentialbility (smoothness) for maps between

(open subsets of ) Banach spaces

H ⊇ U f−→ K where H,K Banach. The derivative Dfx : H → K is a (bounded) linear operator.The inverse mapping theorem still holds.

Local submersion theorem: H ⊇ U 3 x0f−→ K, and that there is a decomposition H = H1⊕H2 such

that Dfx0 |H1 : H1 → K is an isomorphism. Then there is a diffeomorphism ϕ near x0 so that f ϕ is a(surjective) linear operator.

The seemingly complicated definition roots from the fact that a subspace in a Banach space needs nothave complement.

Remark 1. If H is a Hilbert space, it is enough to ask that Dfx0is onto (then H2 = ker(Dfx0

), H1 =ker(Dfx0

)⊥).

Local immersion theorem H ⊇ U 3 x0f−→ K, and there is K = K1⊕K2 such that H

Dfx0−−−→ Kproject−−−−→

K1 is an isomorphism, then there is ψ near f(x0) so that ψ f is a linear operator.

Remark 2. If K is a Hilbert space, it is enough to ask Dfx0is injective with closed image.

Theorem of Constant Rank H ⊇ Uf−→ K and there exists H = H1 ⊕ ker(Dfx0

),K = K1 ⊕K2 such

that H1 → HDfx−−−→ K K1 is an isomorphism for x close to x0. Then there are ψ,ϕ such that ψ f ϕ is

a linear map.

Definition 9. Let H,K be Banach spaces. An operator A : H → K is Fredholm if ker(A) ⊆ H is finite-dimensional, and im(A) ⊆ K is finite-codimensional (it is then always closed – note that it’s not automaticallytrue for any finite-codimenionsal space; it is special as the image of a linear operator).

Page 6

Page 7: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 4 FEB 15

Theorem 3.1. A is Fredholm if and only if there is B : K → H such that BA = I + compactoperator andAB = I + compactoperator.

It says that A only fails to be invertible by something compact.

Example 7. If A1 is invertible, A2 is compact, then A1 + A2 is Fredholm. NOT every Fredholm operatoris of this form.

Definition 10. The Fredholm index of A is defined by dim(ker(A))− codim(im(A)).

The index is locally constant in the space of Fredholm operators.

Example 8. A1 invertible, A2 compact, then A1 +A2 has Fredholm index 0. (Consider the family A1 + εA2

that deforms this to A1.)

Corollary 3. Let U → K be a map whose derivatives are Fredholm (a “Fredholm map”). If y ∈ K is aregular value, then f−1(y) is a finite dimensional submanifold whose dimension equals index of Df .

Problem 4. Let H be a separable Hilber space. Take a map f : H → H given by f(x) = x − g(Ax) whereg is a smooth bounded map, A is a compact linear operator. Prove that if y is a regular value, then f−1(y)a finite set. If g = 0 then we clearly have one solution. So this is a pertubation of x = y. (finite = boundedand compact as a zero-dimensional subspace.)

4 Feb 15

Definition 11. A subset A ⊆ Rn has measure 0 if, for every ε > 0, there exists a countable collection Ri of

cubes/balls/parallelegrams/etc such that A ⊆⋃Ri,

∑i

vol(Ri) < ε.

A countable union of measure zero subsets is a measure zero subset.

Lemma 3. U ⊆ Rn open, f : U → Rn smooth. If A ⊆ U is a measure zero subset, then so is f(A).

Proof. It’s enough to show this for A ⊆ K1 ⊆ int(K2) ⊆ K2 ⊆ U is compact (since any A can be written asa countable union of such subsets). Any sufficiently small ball centered in K1 lies in K2. Since ‖Dfx‖ ≤ Cfor all x ∈ K2, the image of such a ball of radius ρ lies inside a ball of radius Cρ. Cover A with balls of totalvolume ≤ ε =⇒ image is covered by balls of total volume ≤ εCm.

Theorem 4.1. Rn does not have measure 0.

This implies that no ball (or open subset) has measure zero. In other words, measure zero subsets can’thave interior points.

Definition 12. A subset B ⊆ Rn is called thin in the sense of Baire (a.k.a. first Baire category) if it can

be written as B =⋃i

Ci, where the Ci are closed subsets with no interior points.

A countable union of thin subsets is a thin subset.

Theorem 4.2 (Baire). Rn is not thin.

This implies that thin sets cannot have interior points.

Remark 3. If A has measure zero and is a countable union of closed subsets, then A is thin.

Remark 4. One can write Rn = A ∪B, where A has measure 0 and B is thin.

Theorem 4.3 (Sard). U ⊆ Rm open, f : U → Rn smooth, then the set of critical values of f has measure0.

Corollary 4. The set of critical values is also thin in the sense of Baire. Why?

Page 7

Page 8: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 5 FEB 16

Write Crit(f) = x ∈ U : Dfx is not onto, CritV(f) = f(Crit(f)), then Crit(f) is closed, hence acountable union of compact subsets, so CirtV(f) is a countable union of closed subsets.

And the following is the infinite version.

Theorem 4.4 (Sard-Smale). Take H,K Banach spaces which are second countable (have countable densesubsets), U ⊂ H open, f : U → K smooth Fredholm map. Then the set of critical values is thin in the senseof Baire.

Therefore regular values are dense in this case.Special cases of Sard’s theorem:

Lemma 4 (Trivial Sard). U ⊆ Rm open, f : U → Rn, m < n, then f(U) has measure 0 (as one cannothave any regular value).

Proof. (Sketch of Proof) Similar to the previous lemma. Suppose U is bounded and ‖Df‖ bounded, thencover it by ∼ Nm cubes of size ∼ 1/N , then the image is covered by ∼ Nm cubes of size ∼ 1/N , but theirtotal volume is ∼ 1/Nn−m. (A bit more detail: consider Taylor’s expansion near the critical point. See herefor detail.)

Lemma 5 (Easy Sard). U ⊆ Rm, f : U → Rm smooth, then the set of singular values has measure zero.

Proof. Same idea except at the critical values Det(Df) = 0, hence cubes nearby get shrunk to a smallamount. In particular, suppose we have some cover by small squares, then near the singular points, thesesquares get shrunk arbitrarily small.

The difficult case of Sard’s theorem is where the source dimension is larger than the target dimension,and it in fact uses higher derivatives.

Theorem 4.5. p : Rn → R is a polynomial, then p has only finitely many critical values.

Theorem 4.6. p : Rn → R is a real analytic function, then p has at most countably many critical values.

Example 9. p : Rn → R polynomial, proper, bounded below, then level sets p−1(y) are compact hypersurfacesfor all y 0.

Roughly speaking, these level sets will all have the same topology.

Proof. Crit(p) ⊆ Rn is defined by polynomial equations (∂P/∂xi = 0) for each i, hence has only finitelymany connected components.

Second fact (Milnor’s curve selection lemma): Let X ⊆ Rn be a semialgebraic subset (defined by poly-nomial equalities and inequalities). For any x0 ∈ X\X, there is a smooth curve c : [0, ε) → Rn such thatc(0) = x0, c(t) ∈ X for all t > 0.

Now take x0 ∈ Crit(p), X = x ∈ Crit(p) | f(x) 6= f(x0). Then either x0 /∈ X, which meansf |Crit(p) is locally constant near x0, otherwise by the curve selection lemma, there exists some c such thatc(0) = x0, c(t) ∈ Crit(p) for all t, p(c(t)) 6= p(c(0)) for all t > 0, but this is a contradiction, becaused

dt(p(c(t))) = Dp(

dc

dt) = 0, since c(t) is always critical.

5 Feb 16

Theorem 5.1 (Sard). U ⊂ Rm open, f : U → Rn smooth, then the set of critical points has measure zero.

Proof. Let D = Crit(f), D1 = x ∈ U : Dfx = 0, D2 = x ∈ U : derivatives of order less than 2 vanish.First let’s prove that f(D\D1) has measure zero. Take x ∈ D\D1, let’s assume for simplicity that

(Dfx)11 = (∂1f1)x 6= 0. Consider h(x) = (f1(x), x2, . . . , xm), then Dh is invertible at x, so this is a localdiffeomorphism. Let’s take a local inverse ϕ, then (f ϕ)(x) = (x1, g(x1, . . . , xm)). Therefore the first columnof D(f ϕ) is (1, 0, . . . , 0). Critical points of f ϕ are then critical points of g where x1 is considered fixed.

Hence critical values of f ϕ equals∐x1

x1× (critical values of g with x1 considered fixed). By induction

Page 8

Page 9: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 5 FEB 16

on m, the set of critical values of g with x1 fixed has measure zero in Rn−1. By Funibi’s theorem, criticalvalues of f ϕ has measure zero. Everything until now is still local, but we can cover D\D1 with countablymany balls, so the global statement follows.

Now let’s show that f(Dk\Dk+1) has measure zero. Fix x ∈ Dk\Dk+1, we may assume∂k+1f1

∂x1∂xi2 . . . ∂xik+1

6=

0 at x. Consider h(x) =

(∂kf1

∂xi2 . . . ∂xik+1

, x2, . . . , xm

). This is a local diffeomorphism, hence has a local

inverse ϕ. Then each point of f(Dk) locally near x is a critical value of f ϕ|0×Rm−1 . By induction on m,this has measure zero.

Finally, if k is sufficiently large k > m/n, then f(Dk) has measure zero. Sketch of proof: in U closed toDk, we use cube covering it, where there are Nm subcubes of size ∼ 1/N . Under f , there are Nm images ofsize ∼ (1/N)k, then the total volume is Nm−kn, where we see now the exponent is negative.

Application 1. Let U ⊆ Rm open, f : U → Rn, N ⊆ Rn a submanifold. Then there are arbitrary sm allc ∈ Rn such that f(x) = f(x) + c is transverse to N .

Proof. In fact, we will show that this is true for almost all c, This question can be addressed locally on N .Take a local parametrization V ⊆ Rn, g : V → Rn of N . Then transversality is the samething as the mapU × V → Rn, given by (x, y) 7→ f(x)− g(y), has zero as a regular value. But that will be true if we replacef by f . Finally, cover N by countably many local charts.

Definition 13. Let U ⊆ Rm open, f : U → Rm smooth. A fixed point x = f(x) is called nondegenerate if 1is not an eigenvalue of Dfx. (The picture of a degenerate fixed point is that of a “line of fixed points”).

Lemma 6. If x is nondegenerate, then x is an isolated fixed point.

Proof. Consider g(x) = x − f(x), then g(x) = 0 for fixed point x. And x nondegenerate corresponds toDgx = 1−Dfx is an isomorphism. Now apply inverse mapping theorem.

Application 2. There are arbitrarily small c such that f(x) = f(x) + c has only nondegenerate fixed points.

(Same trick as before.)Thus in general, when modeling things, we should be able to assume that we only have nondegenerate

fixed points (just because of uncertainty in real life), and this eventually became a basic assumption indynamical systems. (Is this really all that important? Well...)

Definition 14. U ⊆ Rm open, f : U → R. A critical point Dfx = 0 is nondegenerate if the Hessian D2fxis invertible.

Lemma 7. If x is nondegenerate, then x is isolated.

Proof. Consider g(x) = Dfx. x is critial iff g(x) = 0, and x is nondegenerate iff Dg is onto. Now applyinverse function theorem.

Application 3. Given f , we can perturb it to some f = f(x) +∑i

cixi for small ci, such that it has only

nondegenerate critical values.

Application 4. Let M,N ⊆ Rn be submanifolds, then there are arbitrarily small c ∈ Rn such that forM = M + c, M ∩N is a submanifold of dimension dim(M) + dim(N)− n.

Proof: look at the situation locally, take the difference of the two manifolds, and make zero a regularvalue.

Application 5 (Bertini’s Theorem). Let X ⊆ Cn be a smooth algebraic subvariety (a complex submanifolddefined algebraically), then for all but finitely many c ∈ C, X ∩ x1 = c is a smooth algebraic subvariety.

Proof. Use Sard to show that this is true for almost all c (so it is non-empty), then use algebraic geometryto show that the set of c is constructible (either finite or everything except finite) (use elimination theory tomake the statement “exists a singular value” into polynomial conditions).

Page 9

Page 10: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 6 FEB 22

Need to Know Topology of 1-dimensional submanifolds. Namely, if M ⊆ U is a one-dimensional sub-manifold, then each connected component of M is either homeomorphic to S1 or to R.

Theorem 5.2 (Brouwer Fixed Point). Every bounded continuous map f : Rn → Rn has a fixed point.

Proof. Suppose f has no fixed point, then the same is true of all f with |f(x) − f(x)| < ε for all x and εsmall, because dist(x, f(x)) reaches a minimal value somewhere, as it is never zero and goes to infinity whenx→∞. We can therefore assume without loss of generality that f is smooth.

Consider F : R × Rn → Rn given by F (t, x) = x − α(t)f(x) − c where c is a small constant, α(t) is asmooth cutoff function that is 1 when t < 1 and 0 when t > 2. We can choose c such that 0 is a regular valueof F . Look at F−1(0): for t 0, x = c; for t < 0, x = f(x) + c. If f has no fixed point, then for smallenough c, this has no solution. But F−1(0) is a 1-dimensional manifold and is bounded in the Rn-direction(F (t, x) = 0 iff x = α(t)f(x)+c, if ‖f(x)‖ ≤ C, then ‖x‖ ≤ C+‖c‖), yet x→∞ direction is also impossible,so F−1(0) has nowhere to go.

Problem 5. f : Rn → R proper smooth funtion. Assume that all y ∈ [a, b] are regular values. Show thatthere is a diffeomorphism of Rn which takes f−1(a) to f−1(b) (of course those are compact sets because ofproperness).

Hint: By constructing a suitable flow that flows through all the intermediate sets. (Outside of a largecompact subset make it nothing / identity). All the intermediate values also need to be proper; this isimportant.

6 Feb 22

Recall Brouwer fixed point theorem. Remark: it is in fact sufficient if there are constants C < 1 and D suchthat ‖f(x)‖ ≤ C‖x‖ + D. This should be thought of as an “topological” version of the Banach fixed pointtheorem.

Corollary 5. Let f : Rm → Rm be a continuous map of the form f(x) = A(x) + g(x), A ∈ GL(n,R), g isbounded, then f is onto.

Proof. By passing from f to A−1(f) can assume A = Id. Then apply Brouwer fixed point theorem to−f(x) + x+ c = g(x) + c where c is an constant. Then there exists some x = −g(x) + c =⇒ f(x) = c.

Definition 15. U, V ⊂ Rm be open subsets, with v connected, and f : U → V a smooth proper map. Themod 2 degrees of f is then degZ/2(f) = |f−1(y)| ∈ Z/2Z for any y a regular value.

Theorem 6.1. degZ/2(f) is well-defined.

Proof. Suppose C ⊆ V is a one-dimensional submanifold, the image of an embeding R → V . Suppose f is

transverse to C, and consider f−1(C)f−→ C. Each connected component of f−1(C) is either a circle or a

copy of R.If we consider a component of f−1(C) which is a copy of R the number of preimages of a regular point

inside that component is either always even or always odd. (Use the intermediate value theorem)If we consider a component of f−1(C) which is a circle, then corresponding number is always. Outcome:

|f−1(y)| mod 2 is the same fa for all regular values.To prove the theorem, given the regular values y0, y1, one has to find a C which poasses sufficiently closed

to y0 and y1 and which is transverse to f . In fact, one can take C to be a straight line.

Theorem 6.2. U, V ⊂ Rm open, V connected , take a proper smooth map f then for R×U → R× V givenby (t, x) 7→ (t, Ft(x)), then degZ2

(ft : U → V ) is independent of t.

Remark 5. Using this and a smoothing process, one can extend action of degree to any continuous map.

Application 6. p : C → C complex polynomial of degree d, then deg(p) = d by deformation to p(z) = zd.This implies the fundamental theorem of algebra if d is odd.

Page 10

Page 11: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 6 FEB 22

Application 7. Is the fundamental theorem of algebra true for quaternions? (Answer: yes.) Considerp : H→ H where p(z) = za1za2 . . . zad + lower order terms then p has a root. (Eilenberg-Niven 1940s)

Suppose d is odd. Recall that H = A =

(a −bb a

) ⊆ Mat2×2(C). det(A) = |a|2 + |b|2 is a multiplicative

norm. This implies that p is proper.

To compute the degree, we can deform to p(z) = zd. Then A =

(ζ 0

0 ζ

), then |ζ| = 1, ζ 6= ±1, then it

has exactly d preimages

(ζ1/d 0

0 ζ1/d

). Remains to show that some A is a regular value. This follows from

e.g. directly writing out the Jacobian at A = i.

Definition 16. U, V ⊂ Rm open, V connected, f : U → V proper smooth, the degree of f is deg(f) =∑x∈f−1(y)

sgn(det(Dfx)) for y a regular value.

Theorem 6.3. deg(f) is well-defined.

Theorem 6.4. If R× U → R× V (t, x) 7→ (t, ft(x)) proper, then deg(ft) is the same for all t.

Application 8. p : C → C polynomial of degree d, then deg(p) = d ∈ Z. This implies the fundamentaltheorem of algebra for all d.

Why? For any p, the derivative is complex multiplication with p′(z) and detDpz = det

(re(z) −im(z)im(z) re(z)

)=

|z|2 ≥ 0.

Problem 6. Prove that p : H → H, p(A) = Aa1Aa2 . . . Aad + lower terms has degree d ∈ Z. (Due toinvariance, you don’t have to prove this for a general polynomial.)

Definition 17. An orientation of a finite-dimensional vector space is an equivalent class of bases v1, . . . , vnand v′1, . . . , v′n are equivalent if v′i =

∑j

Aijvj and det(A) > 0.

Note: an orientation of the space V = 0 is defined as an element of ±1.An orientation of V and W determines uniquely one of V ⊕ W . An orientation of V and W ⊂ V

determines one of V/W .More generally, given a short exact sequence 0 → V1 → V2 → V3 → 0, an orientation of any two

determines one for the third.

Definition 18. An orientation of a submanifold M ⊆ U ⊆ Rm is an orientation of each target spaceTMx ⊆ Rm, locally constant in x.

Example 10. The Mobius band does not admit an orientation.

Lemma 8. f : U → Rn, y a regular value, then f−1(y) has a natural orientation.

If x ∈ f−1(y), we have 0 → TMx → Rm Dfx−−−→ Rn → 0, where the second and third have naturalorientations. The resulting orientation of TMx is locally constant is x.

In particular, you can’t have some R3 → R such that the preimage of a point is a Mobius band.If m = n, the orientation of f−1(y) just assigns for each x the numbrer sgn(Dfx) ∈ ±1.

Lemma 9. Uf−→ Rn, N ⊆ Rm oriented submanifold, and f is transverse to N , then f−1(N) inherits an

orientation.

For x ∈ f−1(N) = M,y = f(x), we have 0 → TM → Rm Dfx−−−→ Rn/TNy → 0 by transversality. On theother hand, we have 0→ TN → Rn → Rn/TNy → 0.

Thus the argument for integer degree being constant actually makes sense.

Remark 6. Similarly, if Uf−→ U smooth, f(U) ⊆ U compact and f has nondegenerated fixed points, we

define L(f) =∑

x:f(x)=x

sgn(det(1 − Dfx)), this we define as Lefschetz fixed point number. This is 1 in

Brouwer’s case, and reasonably good deformation preserves this number.

Page 11

Page 12: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 7 FEB 24

7 Feb 24

Funny fact: the Euler characteristic of the Hawaiian earring is not well-defined even though it’s compact.Okay, how do we even define it? We can look at cohomology, but there’s a differential topological approachto it.

Recall: K ⊆ U ⊆ Rm, K compact, U open, f : U → K a continuous map. This has a well-definedLefschetz fixed point number L(f) ∈ Z.

If f is smooth and its fixed points are nondegenerate, then L(f) =∑

x:f(x)=x

sgn(deg(1 − Dfx)). For a

general smooth f , replace f by f(x) = f(x) + c to make the fixed points nondegenerate. (Note: f still landsin a compact subset of U). For continuous maps, C0-approximate by a smooth function.

This is well-defined and homotopy invariant.

Theorem 7.1. K ⊂ U ⊂ Rn, f : [0, 1]× U → K continuous, then L(f(0, •)) = L(f(1, •)).

Theorem 7.2. Let K ⊂ U , K ⊂ U be two different compact inclusions into open in Rm and Rm. Letf : U → K, f : U → K, then L(ff) = L(ff).

Lemma 10. Fix matrices A,B, then det(1−AB) = det(1−BA).

Proof. det(1−AB) det(1−BA) = det(A−ABA) = det(A) det(1−BA).

Lemma 11. A ∈ Matm×m(R), B ∈ Matm×m(R), then det(1−AB) = det(1−BA).

Proof. Eigenvectors for nonzero eigenvalues of AB and BA correspond to each other. (On the subspacespanned by the eigenvectors, AB and BA are conjugate.)

Proof of the theorem. Add constants to f and f to make the fixed points of ff , ff nondegenerate. Notethat ff(x) = x⇔ ff(y) = y where y = f(x), and check local contribution by lemma and we are done. Whycan we perturb them to make both nondegenerate? Either by a topological argument (that nondegeneracyis an open condition), or by a measure-theory argument (by Fubini’s theorem).

This allows us to see that Lefschetz number behaves like a trace in some sense.

Definition 19. A compact subset K ⊂ Rm is a neighborhood retract if there is an open U containing K,and a continuous r : U → K, r|K = id.

Recall that a subset is a submanifold iff it is a smooth neighborhood retract. On the other hand, theHawaiian earring is not a smooth retract.

Definition 20. Let r : U → K be a neighborhood retract, then we define the Euler characteristic of K asχ(K) = L(r).

Lemma 12. This is independent of the choice of U and r.

Proof. Suppose r : U → K, r : U → K are two retractions, then L(r) = L(rr) = L(rr) = L(r).

Lemma 13. r : U → K retraction, r : U → K retraction, where they may live in different Rm and Rm, andwe have a homeomorphism f : K → K, then χ(K) = χ(K).

Proof. χ(K) = L(r) = L(f−1fr) = L(f−1rfr) = L(frf−1r) = L(ff−1r) = L(r) by the theorem we provedabove.

Lemma 14. r, r same as before. Assume that there are continuous maps f : K → K, f : K → K suchthat ff , ff are homotopic to the identity, then χ(K) = χ(K). In other words, it is defined up to homotopyequivalence.

Example 11. χ(S1) = 0.

Proof. Let r(x) = x/|x| and f(x) = ζx/|x|, ζ ∈ S1\1. Then L(r) = L(f) by homotopy equivalence, but fis fixed point free.

Page 12

Page 13: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 8 FEB 29

Problem 7. Compute the Euler characteristic of Sn.

Now let’s consider the infinite dimensional case. U ⊂ H open subset of a separable Banach space, V ⊂ Kbe the same and also connected. Let f : U → V be a proper smooth Fredholm map, such that idx(Dfx) = 0for all x. Pick a regular value y, and define degZ/2(f) = |f−1(y)| ∈ Z/2. The regular value exists bySard-Smale. This is independent of y and is called the mod 2 Leray-Schauder degree.

Example 12. f : H → H, f(x) = Ax + g(Kx), A invertible, g bounded, K compact operator. Then f isonto because degZ/2 f = 1 (because we can use homotopy equivalence to kill the second part). This showsthat the differential equation always has a solution.

What about orientation and the Z-valued degree?

Lemma 15. For any Fredholm operator F : K → K, define δ(F ) to be the set of orientations of ker(F ) ⊕(K/ im(F ))∨. Then, there are narutal identifications δ(F ) = δ(F ) for any F sufficiently close to F .

More rigorously, there is a canonical double covering δ : (all F together with a point in δ(F ))→ (space ofall Fredholm operator F ). Unfortunately, this is a nontrivial Z/2-bundle, and thus it is not possible to chooseorientation for Fredholm operators in a consistent way. Topology of the latter space, which we denote by F :it has Z connected components, indexed by the index, and each connected component has the fundamentalgroup of Z/2 classified by δ. In fact, F is weakly homotopy equivalent to Z × BO (this is a theorem ofAtiyah and Janich).

Corollary 6. If F : H → K is Fredholm, the preimage of a regular value is not necessarily oriented.

Corollary 7. Z-valued degree can’t be defined for Fredholm maps of index 0.

However, these difficulties disappear if we restrict to Fredholm operator of the form A + K for A fixedinvertible, K compact. For example, for maps f(x) = Ax+g(Kx), g bounded, we can define Z-degree whichis known as the Leray-Schauder degree.

Theorem 7.3 (Sard-Smale). H,K separable Banach spaces, U ⊆ H open, F : U → K smooth Fredholm,then the set of singular values of f is thin.

Sketch of proof. Use separability to convert this to a local problem. After coordinate change, assume we havea splitting H = H ′⊕H ′′, K = K ′⊕K ′′, H ′′,K ′′ finite dimensional, where f(x′, x′′) = (Ax′, g(x′, x′′)), where

A invertible. The set of singular values of f are then∐x′

= Ax′ × singular values of g with x′ fixed.

Therefore regular values are dense. For similar reasons, f is locally closed. Thus critical values are countableunion of closed subsets.

8 Feb 29

Application 9. U ⊂ Rm open, f : U → Rn smooth map. If n ≥ 2m, then we can perturb f to get animmersion.

(Matrices that are not injective is actually not a submanifold, so we can’t just use transversality.)In fact, we are looking at perturbations f(x) + Ax, where A is a constant matrix. This will be an

immersion if 0 is a regular value of U × (Rm − 0) FA−−→ Rm where (x, ζ) 7→ Dfxζ +Aζ.(because of the dimension of the assumption and the homogeneity of FA in ζ, 0 is a regular value iff

F−1A (0) is a submanifold of dim ≤ 0 iff F−1

A (0) = ∅ (when 2m = n, we can’t get a submanifold of dim 0:consider homogeneity in ζ), which is iff Dxf +A being always injective.)

Take “all choices at once”: U × (Rm − 0) × Mat(n,m)F−→ Rn given by (x, ζ, A) 7→ Dfx(ζ) + Aζ.

Consider (where we fix x and ζ and alter α): DF(x,ζ,A)(•, •, α) = αζ. Since ζ 6= 0, αζ can be any vector, so

F is a submersion, hence 0 is a regular value of F , and take M = F−1(0).Now 0 is a regular value of FA iff A is a regular value of the projection M → Mat(m,n), which follows

from a direct linear algebraic check. Such A is almost everywhere by Sard. (This uses Sard’s theorem for amap from a submanifold, but it’s unproblematic to work in local coordinates of M .)

Page 13

Page 14: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 8 FEB 29

Application 10. U ⊂ Rm open, f : U → Rn with 2m < n, then we can perturb f to an injective immersion.

Consider the same perturbation f(x) + Ax. We look at U × U − diagonal FA−−→ Rn given by (x, y) 7→f(x) + A(x) − f(y) − A(y). We want to show that for a generic choice of A, 0 is a regular value. Consider

the universal problem (U × U − diagnoal)×Mat(m,n)F−→ Rn where (x, y,A) 7→ f(x)− f(y) + A(x− y).

Again DF(x,y,A)(•, •, α) = α(x− y) is always onto. Take M = F−1(0), then 0 is a regular value of FA iff Ais a regular value of M → Mat(m,n).

Remark 7. If m = 2n, one can perturb f to be an immersion with only transverse self-intersection. (Sameproof applies.) In fact one can prove that no triple intersection can occur.

Remark 8. Alternatively, we could consider the “universal” perturbation, i.e. consider (U×U−diagonal)×(U → Rn)

F−→ Rn given by (x, y, f) 7→ f(x) − f(y). Then DFx,y,f (•, •, ϕ) = ϕ(x) − ϕ(y) is certainly onto.In principle, M = F−1(0) “should” be a submanifold of codimension n, and one could argue by projectionM → (U → Rn), which one can see is automatically Fredholm. Technical point: this will work if we choosea suitable saparable Banach space of functions. (But (U → Rn) is not in general a Banach space, so oneneed to choose a Banach subspace on which one requires derivatives to decay fast enough.) Once we havethis, Sard-Smale applies.

Remark 9. This strategy works for algebraic geometry as well.

Problem 8. Let f : R→ R3 be an embedding. Show for a generic line projection P : R3 → R2, P f is animmersion.

Now let’s talk about partition of unity.

Definition 21. Let U ⊂ Rm be an open subset, U =⋃α∈A

Uα an open cover. A subordinate partition of unity

is a collection of smooth functions (ψα)α∈A such that

1. ψα ≥ 0.

2. supp(ψα) = x : ψα(x) 6= 0 is contained in Uα.

3. For each x ∈ U , there is an open subset V ⊂ U containing x such that α ∈ A : ψα(y) 6= 0 for somey ∈V is finite.

4.∑α∈A

ψα(x) = 1 for all x ∈ U .

Why don’t we just say at each point finitely many of them don’t vanish? Because we need functions

constructed from partition of unity (e.g.∑α∈A

ψα(x)) to be smooth.

Theorem 8.1. Subordinate partitions of unity always exist.

Application 11. U ⊂ Rm open, f : U → R continuous, g : U → (0,∞) continuous. Then there is a smoothfunction f : U → R such that |f(x)− f(x)| ≤ g(x) everywhere.

Proof. Take an open cover U =⋃α∈A

Uα such that each Uα is an open ball and Uα ⊂ U . There are constants εα

such that g(x) ≥ εα for all x ∈ Uα. Can find smooth functions fα : Uα → R such that |fα(x)−f(x)| ≤ εα for

all x ∈ Uα, then this local problem is solved by convoluting with a smooth function. Set f(x) =∑α

ψαfα for

some partition of unity, then this is defined and is smooth at all of U . Then |f(x)−f(x)| ≤∑α

ψα|fα−f | ≤∑α

ψαg = g(x).

Corollary 8. U ⊂ Rm open, then there is a smooth and proper, bounded below function f : U → R.

Page 14

Page 15: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 9 MARCH 02

By considering the level set, we now know that U is homotopy equivalent to a countable CW complex.(This is not immediately obvious–consider the complement of the Cantor set.)

More precisely: x 7→ (x, f(x)) : U → Rm+1 embeds U as a submanifold of Rm=1.Note: This is not true in complex geometry (because you don’t have enough functions). c.f. Stiefel

manifolds.

Proof. Wlog U 6= Rm, take g(x) =1

dist(x,Rm − U)+ |x|2 : U → (0,∞) and approximate it uniformly by a

smooth function. (The |x|2 part is to control the behavior at ∞ for properness.)

Corollary 9. U ⊂ Rm, V ⊂ Rn open, f : U → V continuous. Then f is homotopic to a smooth map.

Proof. Let f(x) = (f1(x), . . . , fm(x)), g(x) = dist(f(x),Rn−V ). Find smooth functions f(x) = (f1(x), . . . , fm(x))

such that |fi(x)− fi(x)| ≤ 1

2ng(x), then ‖f(x)− f(x)‖ ≤ 1

2dist(f(x),Rn− V ), Hence Ht(x) = (1− t)f(x) +

tf(x) ∈ V for all x ∈ U, t ∈ I.

9 March 02

Application 12. U an open subset, M ⊂ U a submanifold, then every smooth function on M can beextended to a smooth function on U .

Proof. Choose open cover U = Ua such that each Ua is either contained in a local coordinate chart for asubmanifold ϕa : Ua → (0, 1)m ⊂ Rm, ϕa(M) = (0, 1)k × 0m−k, or disjoint from M . Fix f : M → Rsmooth. On each Uα there is a ga : Ua → R with ga|Ua∩M = f |Ua∩M (this is the definition for f to be

smooth). Set g(x) =∑a

ψa(x)ga(x), for x ∈M , easy to verify that g(x) = f(x).

Corollary 10 (Isotopy Theorem). U ⊂ Rm, V ⊂ Rn open subsets. Take a map R × U → V given by(t, x) 7→ ft(x), which is an embedding for each t, and independent of t outside a compact subset of U . Thenthere is a family of diffeomorphisms ψt : V → V , ψ0 = id, ψt = id outside a compact subset such thatft = ψt f0.

Proof. We want to construct a time-dependent vector field Xt on V such that Xt(ft(x)) = ∂ft/∂t(x). Thiswill yield the desired family of diffeomorphisms provided that Xt = 0 outside a compact subset. But this isan extension problem for a function as in the following diagram:

(t, ∂ft/∂t(x)) R× V

Rn

and we can use the extension theorem for functions (checking from the proof that support condition can bepreserved).

A somewhat more rigorous statement is as follows. (Not sure if we really care.)

Corollary 11. R × U → R × V ((t, x) 7→ (t, ft(x)) which is an embedding. Moreover, for any compactI ⊂ R, there is a compact K ⊂ U such that ∂f/∂t = 0 for t ∈ I, x /∈ K, then there is a diffeomorphism(t, x) 7→ (t, ψt(x)) such that ψ0(x) = 0x, ψt(f0(x)) = ft(x), and for each compact I ⊂ R there is a compactL ⊂ V such that ∂ψ/∂t = 0 for t ∈ I, y /∈ L.

Problem 9. U ⊂ Rm (m ≥ 2) open connected, x1, . . . , xr ∈ U pairwise different points, y1, . . . , yr ∈ Upairwise different, then there is a difeomorphism f : U → U, f(xi) = yi. Note: this would be false for m = 1.

Now let’s prove the existence of partition of unity.

Page 15

Page 16: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 9 MARCH 02

Proof of Partition of Unity. Take an open cover U = Uα, we can find a cover Vβ such that each Vβ is

contained in some Uα,⋃Vβ is locally finite (i.e. each point has a neighborhood that intersect only finitely

many Vβ), this is given by the fact that Rn is paracompact.Each Vβ is a ball Vβ = Brβ (xβ), the balls Brβ/3(xβ) still cover U . Now for each β, choose ρβ : U → [0,∞)

smooth, such that ρβ > 0 in Brβ/3(xβ), ρβ = 0 outside Brβ (xβ). Then ρ =∑β

ρβ is smooth and everywhere

positive on U , define σbη = ρβ/ρ this is a partition of unity for Vβ . To get the necessary partition, devide

B =∐a∈A

Ba such that if β ∈ Ba, Vβ ⊂ Ua, then set ψa =∑β∈Ba

σβ .

Now let’s talk about manifolds.

Definition 22. A topological manifold is a Hausdorff space with a countable denst subset which is locallyhomeomorphic to Rn.

A differentiable atlas on a topological manifold M is a collection fα : Vα → Uα, where Uα open coverM , Vα ∈ Rm open, each is a homeomorphism, and fβ−1fα : f−1

α (Uβ)→ f−1β (Uα) is a diffeomorphism for all

(α, β). Two atlases f, g are equivalent if the identity map f−1α gβ and its inverse g−1

β fα are both smooth. (As

an counterexample, consider two atlas on R given by x 7→ x and x 7→ x3.) Equivalently, their disjoint unionis again an atlas. This is an equivalence relation.

Remark 10. Any topological manifold can have non-equivalent atlases, or no atlas at all.

(Charles: note that when dimension is less than 4, any topological manifold has a unique differentialstructure up to diffeomorphism, so the distinction only starts to appear in higher dimensions.)

Under this definition, the following concepts are canonically defined:

• Diffeomorphism of manifolds

• immersions

• submersions

• Embeddings

• Submanifolds

Note that a submanifold is itself an abstract manifold.

Example 13. RPn = set of lines through origin in Rn+1. Fix a line L ⊂ Rn+1, Rn+1 = L ⊕ H thennearby lines are graphs LA = x + Ax, x ∈ L where A : L → H is a linear map. This gives a chart(Hom(L,H) ∼= Rn) → RPn Have to check coordinate changes: in fact, the transition function between thetwo obvious charts is given by: z 7→ 1/z.

Example 14. CPn = set of complex lines through the origin in Cn+1. This is a 2n-dim real manifold. Chartconversion: z 7→ 1/z.

Example 15. HPn = set of all subspaces L ⊂ Hn+1 such that dimR L = 4, L is invariant left multiplicationwith H. Chart conversion: H(z, 1) 7→ H(1, z).

Example 16. O octonions. We can define OP1 ∼= S8 by taking two charts O and identifying th z 7→ z−1.Define a “line” in O ⊕ O to be a subspace consisting of (αz, α) : α ∈ O or (α, αz) : α ∈ O (for somez ∈ O fixed) Note: y = x−1(xy).

You can make OP2 but it doesn’t parametrize O3 so isn’t really helpful. So why do we introduce these?Consider the correspondence line+unit length point on that line, and (unit point)→ (line throught point).Then we get maps Sn → RPn, S2n+1 → CPn, S4n+3 → HPn, S15 → OP1. The fibers are correspondinglyS0, S1, S3, S7. Now notice the next map S23 → . . . doesn’t exist! These are called Hopf fibrations and theyhave special roles in algebraic topology.

Page 16

Page 17: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 10 MARCH 07

10 March 07

Definition 23. X Hausdorff space, Γ (discrete) group acting on X. We say that the action is properlydiscontinuously if ∀x1, x2 ∈ X, there are neighborhoods x1 ∈ U1 ⊂ X,x2 ∈ U2 ⊂ X such that γ ∈ Γ |γ(U1) ∩ U2 /∈ ∅ is a finite set.

It’s the same thing as saying that Γ×X → X is a proper map.

Lemma 16. If Γ acts properly discontinuously, then Γ\X is itself Hausdorff.

Example 17. X = S1, Z acting on X by (k, eit) 7→ eit+kθ, θ /∈ 2πQ =⇒ X/Z is not Hausdorff.

Proof. Take x1, x2 and neighborhoods U1, U2 as before. Take F = γ ∈ Γ : γ(U1)∩U2 = ∅. Assuming thatx1 is not in the orbit of x2, we can shrink U1, U2 so that for a single γ ∈ Γ, γ(U1)∩U2 = ∅. Repeat that forall γ ∈ F .

Lemma 17. If Γ acts properly discontinuously and freely, then X → Γ\X is a local homeomorphism (infact a covering map).

Application 13. If M is a topological manifold and Γ acts freely and properly discontinuously, Γ\M i atopological manifold.

Application 14. If H is a smooth manifold, and Γ acts freely and properly discontinuously by diffeomor-phisms, then Γ\M is naturally a smooth manifold.

Lemma 18. Let G be a (Hausdorff) topological group, and Γ ⊂ G a discrete subgroup, Then the left actionof Γ on G is free and properly continuous.

Proof. (Sketch) Take g1, g2 ∈ G, fix a neighborhood of g1g−12 containing at most one point of Γ, then translate

it back in the neighborhoods of g1 and g2.

Lemma 19. Let G,Γ be as before, and K a compact subgroup. Then Γ acts properly discontinuously onG/K.

Example 18. G = PSL2(R) = SL2(R/ ± I. K = PSO2(R) = SO2(R)/± I. Note that G acts on the upper

half plane H = z ∈ C : im(z) > 0 by A =

(a cb d

): z 7→ (az + c)/(bz + d). The stabilizer of z = i is K.

So we get a continuous map G/K → H where A 7→ Ai. This is bijective and has a continuous inverse

z 7→(

im(z) 0re(z) 1

), hence is a homeomorphism. Given a discrete subgroup Γ ⊂ G, we get a Hausdorff space

Γ\G/K = Γ\H, which is a two-dimensional differentiable manifold if Γ acts freely on G/K = H.Conjugacy classes in PSL2(R) = G:

1. A ∼(λ 00 λ−1

), 0 < λ < 1 (hyperbolic) acts as z 7→ λ2z (free)

2. A ∼(

1 00 1

), c 6= 0 (parabolic) act as z 7→ z + c (free)

3. A ∼(

cos(θ) − sin(θ)sin(θ) cos(θ)

), θ /∈ πZ (elliptic), act with a single fixed point.

Note: A elliptic iff tr(A)2 < 4.Consider for instance (N positive integer) Γ(N) = A ∈ PSL2(Z) : A ≡ ±I (mod N), Then tr(A) ≡ ±2

(mod N), so if N > 3, tr(A) /∈ 0,±1. In that case, Γ(N) acts freely, so Γ(N)\G/K = Γ(N)\H = X(N),which is a smooth Riemannian surface (modular curve).

What if Γ ⊂ G = PSL2(R) discrete, but not acting freely on H? (e.g. Γ = PSL2(Z) has points withstabilizers Z/2 and Z/3.) Given any z ∈ H with stabilizer Γz, the quotient Γ\H looks locally like Γz\Hnear x. From an analysis of z = i, we know that Γz is a finite cyclic group. Hence, the quotient looks like

Page 17

Page 18: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 11 MARCH 09

C/(kth roots of unity) ∼= C (z 7→ zk). Hence Γ\H is still a topological surface. In fact, one can make Γ\Hinto a Riemannian surface using removal of singularities. So PSL2(Z)\H ∼= C (call the map from left toright j functions), but C carries two special points (images of the orbits with Z/2 and Z/3 isotropy). In fact,PSL2(Z) ∼= Z/2× Z/3.

Example 19. Let O(n, 1) be the group of linear automorphisms of Rn+1 which preserve the quadratic formP (x) = x2

1 + . . .+x2n−x2

n+1. In fact, they preserve x : P (x) = −1 = x : x2n+1 = x2

1 + . . .+x2n + 1. (This

is a hyperboloid with two components—conveniently named “future” and “past.”)

Let SO(n, 1) = A ∈ O(n, 1) : det(A) > 0, A preserves the connected components of the hyperboloid.This acts on Hn = x : P (x) = −1, xn+1 > 0, and the stabilizer of the point x = (0, . . . , 0, 1) is the subgroupSO(n) ⊂ SO(n, 1) Hence we get for G = SO(n, 1),K = SO(n) that G/K ∼= Hn. Therefore if Γ ∈ G is adiscrete subgroup acting freely on G/K = Hn, we get a quotient n-manifold M = Γ\G/K. These are calledhyperbolic n-manifolds.

If n = 2, SO(2, 1) ∼= PSL2(R), and the action of H2 is the usual one on the upper half plane. If n = 3,SO(3, 1) ∼= PSL2(C), which are automorphisms of the Riemann sphere CP1, thought of as lying at ∞ in thelight cone.

Let K = Q(√d) ⊂ C, d < 0 be a quadratic number field. The ring of integers O = O(K) ⊂ C forms

a lattice in the complex plane. We can then set Γ = PSL2(O), which is a discrete subgroup of PSL2(C).If Γ acts freely on H3, we get a hyperbolic 3-manifold M = Γ\H = G\G/K. These are called arithmetic3-manifolds.

Remark 11. For G/K = H3, even if Γ does not act freely, Γ\H3 is still a topological 3-manifold. Reason:if Ξ ⊂ GL(3,R)+ is a finite group, R3/Ξ ∼= R3.

Problem 10. Is (x, y) ∈ C2 : x3 = y5 a topological manifold?

Problem 11. Give an example of Ξ ⊂ GL(4,R)+ finite, so that R4/Ξ is not a 4-manifold (with proof).

11 March 09

Consider n-tuples (x1, . . . , xn) in R, xi 6= xj for i 6= j, we declare two such to be equivalent if they differ byan affine automorphism of R (x′i = axi + b, a ∈ R − 0, b ∈ R). Let M0,n+1(R) be the set of equivalenceclasses.

M0,3(R) = point.M0,4(R) =three open intervals.

M0,n+1(R) =n!

2open simplices of dimension n− 3.

There is a distinguished compactification M0,n+1(R) optained by letting points run into each other andlooking at them under a microscope.

(Picture with “screens” where the individual “points” are points in M0,j(R for j ≥ 3).This is a description of M0,n+1(R) as a set (and with some imagination as a topological space).Then we see that M0,4(R) ∼= S1, where we have a nice picture.In M0,5(R), consider the closure of one connected component of M0,5(R). It’s not hard to see that

it’s a pentagon, and realize it’s obtained by gluing 12 such pentagons. But it’s not S2! Because thenumber of pentagons meeting at a point isn’t right. But if we compute the Euler characteristic, we see it’s12− 30 + 15 = −3. It turns out it is connected sum of five copies of RP2, and is Hausdorff. It is a smoothmanifold, but it’s not obviously so.

Remark: there is a complex counterpart M0,n+1(C) using points in the plane. They are called Deligne-Mumford spaces.

Remark: M0,6(R) is obtained by gluing together 60 copies of the following polytope: a cube cutting threenon-sharing-vertex edges off.

The outcome is an Eilenberg-Maclane K(π, 1), for π infinite.

Let’s talk about orientation. M manifold, y ∈ M point, take local chart U ⊂ Rn ψ−→ V ⊂ M,ψ(x) = y,

U ′ψ′−→ V ′, ψ′(x′) = y. Two such charts induce the same local orientation at y if Dx(ψ′)−1ψ has positive

Page 18

Page 19: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 12 MARCH 14

determinant. This is an equivalence relation, and there are exactly two equivlance classes called localorientations at y. A local orientation at y induces a local orientation at any point sufficiently close to y.

We define M = (y, ) : y ∈M, is a local orientation aty.Then M has a 2-to-1 quotient map to M , and Z/2 thus acts on M. and M carries a canonical topology

as a covering space, Hence M is canonically a manifold, and is called the orientation covering of M .

Definition 24. M is orientable if M → M is the trivial covering. An orientation of M is a choice ofsection M →M.

Remark 12. M is always orientable and has a canonical orientation.

Remark 13. Notion of orientable / unorientable loop in any M is then well-defined (by looking at whetherthe two sheets get swapped). Then M is orientable if all loops are orientable.

Remark 14. If there are no nontrivial double covers (e.g. M connected, π1(M) trivial) then M is orientable.

What about Sn → RPn? If this is the orientation cover then we know RPn is unorientable, since thecovering is clearly not trivial.

Example 20. U ⊂ Rm, f : U → Rn smooth. If 0 is regular value, then M = f−1(0) is always oriented.

To see this, let x ∈ Rm−n. Map it to y ∈M , then by f to 0 ∈ Rn. ψ is compatible with the orientation iff(Dψx(1, 0, . . . , 0), . . . , Dψx(0, . . . , 0, 1), v1, . . . , vn) is a positively orientated vector in Rm, where Dfy(v1) =(1, 0, . . . , 0), Dfy(v2) = (0, 1, 0, . . .), etc.

Example 21. M,N oriented, f : M → N has y as regular value, then f−1(y) is oriented.

Lemma 20. A double cover Mτ−→ M is isomorphic to the orientation cover if and only if M is oriented

and the Z/2 action reverses the orientation.

Example 22. RPn is not orientable for n even, because the antipodal map on Sn reverse orientations. It isorientable for n odd, since then antipodal map preserves orientations.

Remark 15. Suppose M ⊂ Rn+1 is a hypersurface (n-dimensional submanifold), then M ∼= y = x+v, x ∈M,v⊥TMx, ‖v‖ = f(x) for a sufficiently small f : M → (0,∞).

Problem 12. M ⊂ Rn+1 is a hypersurface then it is orientable. (Do not use algebraic topology.)

12 March 14

Orientation of manifolds.

Application 15. Let M be a complex (analytic) manifold. Then, as a real manifold, M has a canonicalorientation. If we have a complex chart U → M we get a real chart, and two such charts induce the samelocal orientation, since for A ∈ GLn(C), det

R(A) = |det

C(A)|2 > 0.

Application 16 (Blowing up a point). P = KPn for K = R,C,H or O (in this case n = 1). L =(x, λ) ∈ Kn+1 × P : x ∈ λ. Then L = P ∪ (Kn+1\0). Hence L is a manifold which is diffeomorphicto Kn+1 outside a compact subset. Take any connected manifodld Mm, m = dim(M) = dim(L). Fix a

chart Kn+1 = Rm ϕ−→ M and set BlK(M) = (M\ϕ−1(0) ∪ L)/ ∼, where the equivalence relation identifiesx ∈ Kn+1\0 ⊂ L with ϕ(x) ∈M .

The diffeomorphism type of BlK(M) remains unchanged if we smoothly deform the chart. (Local isotopyargument). If M is not orientable, any chart can be deformed to any other chart, hence the blowup iswell defined up to diffeomorphism. The same is true whenever L admits a diffeomorphism which, outsidea compact subset, is an orientation-reversing linear map of Kn+1. For instance, this is true for K = R. Inall cases, we can make it well-defined by specifying an orientation. In general, blowup is well-defined fororiented manifolds using oriented local charts.

Page 19

Page 20: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 13 MARCH 28

Example 23. BlR(Sn) = RPn, since the projection Rn+1 x1−→ R identifies RPn+1\point → L. Now glueback the remaining point that is not covered by the chart.

Example 24. Similarly, BlC(S2n) = CPn and BlH(S4n) = HPn.

Definition 25. OP2 = BlO(S16).

Example 25. BlC(CP2) yields a 4-manifold with intersection form

(±1 00 ±1

)or

(±1 00 ∓1

)depending

on the choice of chart. (The second one gives the complex oriented charts.)

Connected Sum Take M,N connected manifolds, charts ϕ : Rm → M,ψ : Rm → N , then M#N =(M\ϕ(0) ∪N\ψ(0))/ ∼ or (M\ϕ(Bm) ∪N\ψ(Bm))/ ∼ where in the first we identify ϕ(x) with ψ(x/‖x‖2),and in the second we do ϕ(x) ∼ ψ(x) for ‖x‖ = 1.

If one of the two manifolds is not orientable, or admits an orientation-reversing diffeomorphism, then theconnected sum is well-defined as an unoriented manifold. If both are oriented, choose ϕ compatibly oriented,ψ not compatibly oriented, then the connected sum is well-defined as an oriented manifold.

Example 26. BlR(Mn) = M#RPn, which is well defined since RPn admits an orientation-reversing dif-feomorphism (coming from that of Sn by flipping a coordinate).

Example 27. Let M be a complex manifold. The complex-analytic blowup is M#CPn, where the overline

indicates we reverse the given (complex) orientation. For instance, the Hirzebruch surface F1 is CP2#CP2.

Remark 16. Oriented manifolds homotopy equivalent to Sn, fixed n ≥ 5, up to diffeomorphism, along withconnected sum as the additive operation, is a finite abelian group where Sn is the identity, and orientationreverseal is the antipodal map. It is denoted by Θn e.g. Θ7 = Z/28.

Degree Let M,N be compact manifolds of dimension n, where M,N are connected. Take ϕ : M → Nsmooth, fix y ∈ N an regular value, then deg(ϕ) is defined as the number of preimages of y mod 2. This iswell-defined (independent of y) and unchanged under deformations of ϕ.

Proposition 1. if Mϕ−→ N

ψ−→ O are maps of connected n-manifold, then deg(ψϕ) = deg(ϕ) deg(ψ).

Oriented Degree M,N compact oriented connected manifolds. ϕ : M → N smooth map, ϕ(x) = y.y regular point. We say x is positive if Dxϕ has positive determinant in oriented local charts,negative

otherwise. If y is a regular value, deg(ϕ) =∑

x∈ϕ−1(y)

±1 ∈ Z. This has the same properties as before.

Example 28. An orientation-preserving diffeomorphism ϕ : M → M and orientation-reversing diffeomor-phism ϕ : M → M has deg(ϕ) = 1,deg(ψ) = −1. A constant map M → M has degree 0. (They can’t bedeformed into each other.)

Problem 13. Let M,N be compact connected complex (analytic) manifolds of the same dimension. Letϕ : M → N be a complex map. Then either ϕ(M) has no interior points, or ϕ(M) = N .

13 March 28

Basic techniques we have developed so far for the differential world (and how they apply in the algebraicsetting):

• Local structure of differential maps. (This doesn’t work very well with polynomial maps, but betterwith formal power series. This was actually what was first developed: formal coordinate changes.)

• Partition of unity: usually works for compact manifolds. (Certainly not working for algebraic setting.)

• Sard’s theorem and consequences. (It applies for algebraic setting–say, Bertini’s theorem), complexnumbers, and (trivially) finite fields.) (Check: Poonen’s paper for finite field Bertini theorem).

Page 20

Page 21: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 13 MARCH 28

Sample results:

Proposition 2. f : M → N proper injective immersion (define it in local charts), then f(M) ⊂ N is asubmanifold.

Proposition 3. f : M → N , y ∈ N regular value (defined in local chart), then f−1(y) ⊂M is a submanifold.

Proposition 4. f : M → N,P ⊂ N a submanifold such that f is transverse to P (defined locally), thenf−1(P ) is a submanifold.

Theorem 13.1 (Ehresmann). f : M → N proper, y ∈ N regular, then ∃ open subset y ∈ V ⊂ N and a

diffeomorphism f−1(V )Ψ−→ V × f−1(y) such that the following diagram commutes:

f−1(V ) V × f−1(y)

V

f

Ψ

π

Corollary 12. Let f : M → N be a proper submersion, N connected, then any two fibers f−1(y) arediffeomorphic (but not canonically so–what’s the problem? the isomorphism in Ehresmann’s theorem is byno means unique).

Such maps are “differentiable fiber bundles”. Spirit of the story: differentialbe manifolds have no moduli.Ehresmann theorem uses the first two techniques we mentioned above.

Let’s consider the spacial case of f : U → V a proper submersion, y ∈ V regular value, U ⊂ Rmopen, V ⊂ Rn open. At each point x ∈ f−1(y), Rm = ker(Dfx) ⊕ Kx (the orthogonal complement).

Dfx : Kx

∼=−→ Rn. Hence get a basis (e1(x), . . . , en(x)) of Kx by pullback. Consider f−1(y) × Rn → Rmgiven by (x, t) 7→ (x + t1e1(x) + . . . + tnen(x)). The derivative of this is an isomorphism at each point off−1(y)× 0. In fact, it gives a diffeomorphism ψ between a neighborhood of f−1(y)× 0 ⊂ f−1(y)× Rnand a neighborhood of f−1(y) ⊂ U . Now f ψ(x, 0) = y. D(f ψ)(x,0)|0×Rn is an isomorphism. It followsthat f ψ|x×Rn is a local diffeomorphism near t = 0. Now invert all those diffeomorphisms (and observethat these vary continuously as x varies) to get Ψ.

Now consider the case of f : M → V ⊂ Rn open, and M ⊂ U ⊂ Rm a submanifold for some U open, fis a proper submersion. At x ∈ f−1(y), TMx = T (f−1(y))x ⊕Kx, the second being the orthogonal in TMx.

Then Dfx : Kx

∼=−→ Rn, get a basis (ei(x)) of Kx. Recall that we can choose U so that there is a smooth

retraction r : U →M . Define ψ(x, t), x ∈ f−1(y), t ∈ Rn small, to be r(x+∑i

tiei(x)). The rest is the same

as before. Now note that we have:

Theorem 13.2 (Whitney). Every n-manifold can be embedded as a submanifold of R2n+1.

The general Ehresmann then follows (but this uses all three techniques above.)Now recall the partition of unity (we did this as a subset of Rn, and the general proof is the same):

Theorem 13.3. Let M be a manifold, (Uα), α ∈ A an open cover, then there are functions (ψβ), β ∈ B,ψb : M → [0,∞) such that supp(ψβ) is contained in some Uα, and locally only finitely many ψβ are nonzero,

and∑β

ψβ = 1.

Corollary 13. For any M , there is a proper bounded below function f : M → R.

Proof. Take (ψi), i ∈ N a countable partition of unity such that supp(ψi) is compact. Define f(x) =∞∑k=1

k · ψk(x). If x /∈ supp(ψ1) ∪ . . . ∪ supp(ψn), then f(x) ≥∞∑

k=n+1

k · ψk(x) ≥ (n+ 1)

∞∑k=n+1

ψk(x) = n+ 1,

hence properness.

Remark 17. Useful in combination with Sard to “decompose” a non-compact manifold into parts that areseparated by compact pieces. For instance, consider the universal cover of Tn#M , π1(M) = ∗. It looks likean infinite Rn with a M at each lattice point, and this sort of cut it into concentric “balls” in that Rn.

Page 21

Page 22: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 14 MARCH 30

Let’s now consider Sard’s theorem and its generalizations.

Proposition 5. f : Mm → Nn, n > 2m, then f can be perturbed to an injective immersion f , i.e. given∆N ⊂ V ⊂ N ×N , V open in N ×N , we can ask that (f(x), f(x)) ∈ V for all x. Or, more generally, givenM ×∆N ⊂ U ⊂M ×N ×N open, can ask that (x, f(x), f(x)) ∈ U for all x.

Principle of Proof. (Say for injectivity) Start with F : P ×M → N where p0 ∈ P , F (p0, x) = f(x). Every

value of P gives some way of deformation. This should be large enough that P × (M ×M) \∆MΦ−→ N ×N

given by (p, x1, x2) 7→ (F (p, x1), F (p, x2)) is transversal to ∆N . Then Φ−1(∆N ) ⊂ P × ((M ×M) \ ∆M )is a submanifold of codimension n. Take Φ−1(∆N ) → P and let p be a regular value close to p0, thenΦ−1(∆N ) ∩ p = ∅ is empty by dimension argument. Set f = F (p, •).

How to get large enough P? If M is compact, we can do partition of unity and locally allow finitely manydegrees of perturbation. If M is non-compact, we have to be careful of maintaining finite-dimensionality:this is okay because for far enough open sets, we can recycle the degrees of perturbation. Or, an easier wayis to take the space of all maps, but then we need to use Sard-Smale, and we have to be slightly carefulabout using Banach spaces (because C∞ is not Banach; but it has dense subspaces that are).

Problem 14. f : M → N smooth, M compact, show that any sufficiently small perturbation of f ishomotopic to f .

14 March 30

Theorem 14.1 (Homotopy Lifting). Let f : M → N a proper submersion. Given a family (gt)t∈R ofdiffeomorphisms of N , g0 = id, there is a family (ht)t∈R of diffeomorphisms of M , h0 = id, such that

M M

N N

ht

f f

gt

In general, diffeomorphisms aren’t linear, so it’s hard to do linear combinations (e.g. partition of unity)with them. One way to solve this is to differentiate them, get a vector field, linearly combine the vectorfields, and integrate them. Let’s therefore talk about vector fields.

Definition 26 (Vector Field). (Physicist’s Definition) Let M be a manifold. A vector field X on M is acollection of vector fields in local charts which are related by appropriate transformations.

More concretely, let ϕα : Rn ⊃ Vα → Uα ⊂ M . Let Xα : Vα → Rn be the local vector fields. Suppose

Vαϕα−−→ M,Vβ

ϕβ−−→ M , then D(ϕ−1β ϕα)x(Xα,x) = Xβ,(ϕ−1

β ϕα)(x). This style of definition has no problem

except one has to add a condition for two vector fields to be the same (namely, if you can put them togetherand still get a vector field).

Given a time-dependent vector field (Xt)t∈R on M , assuming that our vector fields are compactly sup-ported (or in some other way prevented from reaching infinity), we get a family of diffeomorphisms (ϕt)t∈R,

ϕ0 = id, such that∂ϕt∂t

(x) = Xt(ϕt(x)). (This is not quite well-defined; use local chart definition.) In fact,

once we solve it on each local chart, the overlap automatically agrees. Conversely, (ϕt) determines Xt.Now let’s prove isotopy lifting.

Proof Sketch. Start with family of diffeomorphisms (gt) on N , we get a time dependent vector field Xt onN , which yields a time-dependent vector field on M (it is this step we explain below), which finally givesdiffeomorphisms (ht).

Now, look at a local chart Rm ⊃ Uf−→ V ⊂ Rn, given Xt : V → Rn, want Yt : U → Rm such that

(∗)Dfx(Yt,x) = Xt,f(x). This way (ht) is prevented from going to infinity because we have control on (gt), andbecause the map is proper. We can locally find Y such that (∗) is satisfied by the local submersion theorem(Hmm?). Moreover, (∗) survives convex combinations, which means we can use partitions of unity.

Page 22

Page 23: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 14 MARCH 30

This tells us another way of showing that two fibers in a connected component are the same: constructa vector field moving from one to another, and lift.

Note that this gives another proof of Ehresmann’s theorem. let f : M → N , fix y ∈ Y and families of

diffeomorphisms on N : ϕ(1)t , ϕ

(n)t (so that they move in all possible directions, say each going one direction),

then lift them to M , and use it to construct a local diffeomorphism between f−1(y)× Rn and M .Now let’s talk about manifolds with boundary.

Definition 27. A topological manifold with boundary is a (Hausdorff, second countable) space locally home-omorphic to Rn or Rn−1 × [0,∞).

Definition 28. Let A ⊂ Rn be any subset. A function f : A→ R is smooth if there is an open U ⊂ Rn, A ⊂ Uand a smooth function g : U → R such that g|A = f .

This definition may be a bit eyebrow-raising because it doesn’t seem to be local at the first sight. In factit is. Here’s an alternative definition:

Definition 29. f : A → R as before. f is smooth if for each x ∈ A there is an open U ⊂ Rn containing xand a smooth g : U → R, such that g|U∩A = f |U∩A.

Problem 15. Show that the two definitions agree.

Definition 30. Let M be a topological n-manifold with boundary. A differentiable atlas is a collection of

maps (ϕα : Vα∼=−→ Uα), where (Uα is an open cover of M , the Vα are open subsets of of Rn or Rn−1× [0,∞),

and the ϕα are homeomorphism such that Vα ∩ϕ−1α (Uβ)

ϕ−1β ϕα−−−−→ Vβ ∩ϕ−1

β (Uα) is smooth for all (α, β). Notethat smoothness at the boundary point requies a local smooth extension to an open subset o Rn.

Example 29. Let M be a manifold, f : M → R smooth, and a a regular value. Then x ∈ M : f(x) ≤ ais a manifold with boundary.

Note that by restricting the charts, the boundary of a manifold with boundary is a manifold of dimensionone lower.

Theorem 14.2 (Collar Neighborhood Theorem). Let M be a differentiable manifold with boundary. Thenthere is an open subset U ⊂M containing ∂M , an open subset V ⊂ ∂M × [0,∞) containing ∂M ×0, anda diffeomorphism c : V → U such that c|∂M×0 = id∂M .

Equivalent version:

Theorem 14.3. Let M be a differentiable manifold with boundary. Then there is an open subset U ⊂ Mcontaining ∂M , and a diffeomorphism c : ∂M × [0,∞)→ U such that c|∂M×0 = id∂M .

This is obvious locally, and of course it is not unique. The idea is as usual: partition of unity. Of coursesame problem: diffeomorphism isn’t linear. Solution: go to vector fields.

Sketch of Proof. Construct a vector field pointing inwards along ∂M , by using a partition of unity to combinethem, then integrate.

Application 17. Let M and N be two manifolds with boundary, together with a diffeomorphism f : ∂M →∂N . Define O = M tN/x ∼ f(x), x ∈ ∂M, then this is a topological manifold without boundary.

Note that O doesn’t have a unique differentiable structure. (Differentiable on two sides and agree onboundary isn’t enough to say everywhere differentiable.) To get a differentiable structure, choose collar

neighborhoods ∂M × (−∞, 0] ⊃ Vc−→ M and ∂N × [0,∞) ⊃ W

d−→ N , and identify V t W/(0, x) ∼(0, f(x)) ctd−−→ O. But note that the set sites in ∂N × R by (x, t) 7→ (f(x), t) for x ∈ V , and id on x ∈ W ,then this gives a map from an open neighborhood of ∂N × 0 ⊂ ∂N ×R to O, which we can use to define thedifferentiable structure near the “scam.” This differentiable structure is unique up to diffeomorphism.

Example 30. Connected sum M#N can be viewed as gluing together (M \B) and (N \B) for a small openball B.

Page 23

Page 24: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 15 APRIL 4

Definition 31. Two compact manifolds M1,M2 are called (co)bordant if M1 t M2 is the boundary of acompact (n+ 1)-manifold.

Remark 18. Any M is cobordant to itself: M tM = ∂(M × [0, 1]).

Remark 19. Bordism is transitive, by gluing.

Thus bordism is an equivalence relation.

Definition 32. The bordism ring Ω∗ = (Ωn)n≥0 is the set of equivalence classes of compact manifold underbordism, with disjoint union as addition ([M ] + [M ] = [∅]) and product as multiplication ([M ]× [∗] = [M ]).

Example 31. Ω0∼= F2, Ω1 = 0.

We’ll see that higher bordisms will in some way serve as generalizations for numbers to higher dimensions.

15 April 4

Recall that Ω0 = Z/2, Ω1 = 0. Today we prove that Ω2 = Z/2.

Remark 20. Recall that [M ] + [M ] = 0 ∈ Ω∗, because M tM = ∂(M × I). It is also true that [M#N ](connected sum) is equal to [M ] + [N ] due to the standard bordism between these manifolds.

The classification theory of surfaces tells any closed surface is diffeomorphic to a connected sum of T 2sand RP2s. We know T 2 = ∂(S1 ×D2), hence Ω2 is either 0 (if RP2 is an embedding) or Z/2 (otherwise).

Let M be a manifold with boundary. An orientation of M determines an orientation of ∂M . Example:∂(• → •) = •−1 •+1. The general definition is by restricting charts (here, the convention is that chartsalong ∂M are modelled on [−∞, 0] × Rn−1). Why does this work? Suppose we have transition maps

(−∞, 0)×Rn−1 ϕ−→ (−∞, 0)×Rn−1 if ϕ is a positively oriented diffeomorphism (detDϕ > 0), then the same

holds for ϕ|0×Rn−1 , because at x ∈ 0 × Rn−1, (Dϕ)x =

(> 0 0∗ ∗

).

Two compact oriented manifolds M1,M2 are oriented co(bordant) if M1 t (−M2) is the boundary of acompact oriented manifold. We can likewise define the oriented bordism ring ΩSO∗ . For instance, ΩSO0 = Z,ΩSO1 = 0, ΩSO2 = 0 (by the classification of surfaces), ΩSO3 = 0 (needs a lot of work), ΩSO4 6= 0) (in fact Zbut we’ll show today that ΩSO4 → ΩO4 is nontrivial).

Example 32. Let Γ ⊂ U(n) be a finite subgroup. Assume that Γ acts freely on Cn \ 0. Then Γ actsfreely on S2n−1 ⊂ Cn, and M = S2n−1/Γ is an oriented manifold. This is called the link of the singularity

0 ∈ Cn/Γ. Then [M ] = 0 ∈ ΩSO2n−1 by resolution of singularities, as M = ∂(D2n/Γ). In particular, all odd-dimensional real projective spaces are bound (Take Γ = ±1; they are orientable and bound as orientablemanifolds).

And let’s review some basic topological invariants.

Degree f : Mm → Nm map between compact manifolds of the same dimension, N connected (or a propermap between not necessarily compact manifolds), degZ/2(f) = #f−1(y) ∈ Z/2 for y ∈ N a regular value.

If M,N are oriented, can define the Z-degree deg(f) =∑

x∈f−1(y)

sgn det(Dfx) ∈ Z where the determinant

is in oriented local charts. This is well defined and is homotopy invariant.Consider f : Mm → Nn (again, compact manifolds or proper map, N connected). How does the preimage

of a proper value vary? If we have a submersion, Ehresmann’s theorem says they are all the same; in generalthis of course is false, but:

Theorem 15.1. Let y be a regular value. Then [f−1(y)] ∈ Ωm−n is independent of y and a homotopyinvariant of f .

In this way, bordism classes generalize the notion of “numbers” as one count solutions.

Page 24

Page 25: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 15 APRIL 4

Outline of Proof. Given two regular values y0, y1, choose a path c(0) = y0, c(1) = y1. (Be careful: we can’tassume all points along this line are regular value; for instance, singular points can form a hypersurfacethat we have to cross.) We can achieve (by perturbing c) that this is transverse to f , in the sense that

[0, 1] ×M ϕ−→ N × N given by (t, x) 7→ (c(t), f(x)) is transverse to the diagonal ∆N . Then ϕ−1(∆N ) =(t, x) : c(t) = f(x) is a manifold with boundary where t = 0, 1. The boundary is f−1(y0) t f−1(y1).

Theorem 15.2. If in addition M and N are oriented, then [f−1(y)] ∈ ΩSOm−n has the same properties. Notethat the preimage acquires an orientation by working in local oriented charts.

Now let Mm a compact manifold, f : Mm → Mm (or can ask for any M , then f(M) needs to be

compact). We L(f) =∑

f(x)=x

sgndet(1−Dxf). (Assuming the det is not zero oin local charts). Check this is

well-defined: suppose transition is f = ψ−1fψ, then det(1−Df) = det(1−Dψ−1DfDψ) = det(Dψ−1Dψ−Dψ−1DfDψ) = det(1 −Df). This is independent of orientability. To define it in general, have to perturbf such that the determinant is nonzero at all fixed points. L(f) is independent of the perturbation used todefine it, and is homotopy invariant.

Example 33. If f has no fixed points, L(f) = 0.

Lemma 21. M,N compact manifolds, f, g : M → N , then L(fg) = L(gf).

Definition 33. M compact manifold. The euler characteristic χ(M) = L(idM ).

Proposition 6. χ(M) is a homotopy invariant.

Proposition 7. If f : M → N is a d to 1 covering, then χ(M) = dχ(N).

For instance, S1 must have χ = 0, because S1 is a d to 1 of itself for any d.

Proposition 8. If f : M → N is a submersion and N is connected, χ(M) = χ(N)χ(f−1(y)) for any y.

Example 34. χ(CPn−1) = n. Remember CPn−1 are lines in Cn, so any A ∈ GLn(C) acts on it, and thefixed points correspond precisely to the eigenvectors. Deform A = 1 to A = diag(λ1, . . . , λn) for pairwisedistinct λ1, . . . , λn. A computation in local coordinates shows that all n fixed points contribute +1.

Note that this proves that complex matrices have eigenvectors.

Example 35. χ(RPn−1) = 0 if n is even, and 1 if n is odd (choose matrix in GL+n (R) with just one

eigenvalue, compute in local coordinates).

Thus χ(Sn−1) = 0 if n even, 2 if n odd.

Theorem 15.3. M compact manifold, dim(M) = n odd, then χ(M) = 0.

Proof. Perturb idM to f : M → M , which will be a diffeomorphism (being a diffeomorphism is an open

condition), so L(f) =∑

x=f(x)

sgn det(1−Dfx), and

L(f−1) =∑

x=f−1(x)

sgn det(1 −Dxf−1) =∑

f(x)=x

sgn det(Dxf−1) det(Dxf − 1). Now det(Dxf

−1) ≈ 1, so we

get∑

f(x)=x

sgn det(Dxf − 1) = (−1)nL(f). But χ(M) = L(f) = L(f−1), so we get 0 when n is odd.

Theorem 15.4. M compact, dim(M) even, χ(M) odd, then M is not the boundary of a compact (n + 1)-manifold.

We’ll prove this later. (This was proved in April 11’s class.)

Page 25

Page 26: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 16 APRIL 6

16 April 6

Recall we defined basic topological invariants: degree and Lefschetz number (Euler characteristic).

Remark 21. Let M be a compact manifold with boundary. Using partitions of unity and collar neighborhood,one can define ft : M → M for t ≥ 0, f0 = id, ft(M) ⊂ M \ ∂M(t > 0). This corresponds to a vector fieldthat “pushes boundary inwards.”

Now given f : M →M , one can define L(f) = L(f |M\∂M ft) for t > 0, where ft f maps M \ ∂M to arelatively compact subset of M \ ∂M . In particular, this defines χ(M) = L(ft)(t > 0).

Intersection Number f1 : M1 → N, f2 : M2 → N , M1,M2 are compact, and dim(N) = dim(M1) +dim(M2). Perturb f1, f2 such that (f1, f2) is transvsers to the diagonal ∆N , and the intersection numberf1 Z/2 f2 ∈ Z/2 is defined as |(f1, f2)−1∆N | ∈ Z/2, i.e. how many times the images intersect.

If all manifolds are orientable, can define the Z-intersection number similarly.

Remark 22. If M is compact and oriented, and f : M →M , define γf : M →M ×M by γf (x) = (x, f(x))and δ(x) = (x, x). Then γf · δ = L(f). In particular, δ · δ = χ(M).

Note that Euler characteristic / Lefschetz number didn’t ask for orientability; in fact, one can get awaywith something slightly weaker than all manifolds to be (separately) orientable, and in that setting we canget them without orientability condition.

Problem 16. Show that there is no map f : S2 → T 2 with degZ/2(f) = 1. (Probably use intersection

number). (Note there is a map of degree one in the other direction: take a small disk of T 2 and map it toS2 \ ∗, and map the rest to ∗).

Generalizations f : M → N , N connected, f proper, [f−1(y)] ∈ Ωdim(M)−dim(N) for a regular value of y.This is an invariant that generalizeds degZ/2(f).

Remark 23. In particular, if for some regular y, [f−1(y)] ∈ Ω∗ is nontrivial, then f must be onto.

If M,N are oriented, can get [f−1(y)] ∈ ΩSOdim(M)−dim(N), which generalizes the integer degree.

Now if f1 : M1 → N, f2 : M2 → N for compact M1,M2, then after perturbing, [(f1, f2)−1(∆N )] ∈Ωdim(M1)+dim(M2)−dim(N) is an invariant. If M1,M2, N are oriented, can get the integer version as well.

What about Lefschetz number / Euler characteristic?Let f : M → N be a smooth proper submersion. Suppose g : M →M is a fiberwise self-map i.e. gf = f ,

under a suitable undegeneracy condition, Fix(g) is a submanifold of dim(N). The Dold-Lefschetz fixed pointindex is [Fix(g)] ∈ Ωdim(N). If N is oriented, so is Fix(g) and we get [Fix(g)] ∈ ΩSOdim(N). In particular, thiscan be used to define an Eulcer characterstic of f : M → N which is a bordism class of dimension dim(N).(In reality, only N necessarily needs to be a manifold; say M can be a fiber bundle or something like that.)

What is this notion of nondegeneracy? Local picture is Rm g−→ Rm, f : Rm → Rn and gf = f (why are westaying on the same chart? Note we’re only talking about near fixed points). Suppose by submersion theoremthat f(x1, . . . , xm) = (x1, . . . , xn). Then we require that we have g(x1, . . . , xm) = (x1, . . . , xn, h(x1, . . . , xm))such that 0 is a regular value of (x1, . . . , xm) 7→ (xn+1, . . . , xm)−h(x1, . . . , xm). Note that preimages of zeroare the fixed points.

This is a weaker condition that to require the fixed point is fiberwise nondegenerate (which is impossible).

Remark 24. If M,N are oriented, one can think of this as an intersection number in M ×f M . Thisgeneralizes what we said above.

Bordism with Target Space Fix a manifold X. Consider (M,f),M compact, f : M → X. Two suchpairs (M1, f1), (M2, f2) are cobordant if there is a compact manifold with boundary N and j : N → X suchthat ∂N = M1 tM2 (−M2 in the oriented case), g|M1 = f1, g|M2 = f2. (X can also be an open set or otherrelaxed conditions; for now let’s use manifold.) This is an equivalence relation, and gives abelian groupsΩ∗(X) or ΩSO∗ (X). We shall focus on the unoriented case, where twice of anything is zero.

There is no longer a ring structure but we still have the exterior product Ω∗(X)⊗Ω∗(Y )→ Ω∗(X × Y ).

Page 26

Page 27: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 17 APRIL 11

Functoriality Given ϕ : X → Y , get ϕ∗ : Ω∗(X)→ Ω∗(Y ) and this is a homotopy invariant of ϕ.

Mayer-Victoris Sequence Given a smooth function ψ : X → R, define X− = ψ−1((−∞, 1)) and X+ =ψ−1((−1,∞)), X+ ∩X− = ψ−1((−1, 1)). Then there is a long exact sequence

. . .→ Ω∗(X+ +∩X−)inclusion,−inclusion−−−−−−−−−−−−→ Ω∗(X+⊕Ω∗(X−))

inclusion,inclusion−−−−−−−−−−−→ Ω∗(X)→ Ω∗−1(X+∩X−)→ . . .

Remark 25. This is an (extraordinary) homology theory. It is a theorem that Ω∗(X) ∼= H∗(X; Ω∗) ∼=H∗(X)⊗ Ω∗. (But this is not true in the oriented case).

Inverse (shriek) Functoriality For ϕ : X → Y a proper map, we get ϕ! : Ω∗(Y )→ Ω∗+dim(X)−dim(Y )(X).

Example 36. dim(X) = dim(Y ), Y connected, ϕ!([∗ → Y ]) ∈ Ω0(X) recovers the degree of ϕ.

To define this, given f : M → Y , perturb (ϕ, f) : X ×M → Y × Y so that it is transverse to ∆Y . Thenthe preimage of the diagonal is a compact submanifold of X ×M , we map it to X by projection.

Intersection Product We have a map Ωi(X) ⊗ Ωj(X) → Ωi+j−dim(X)(X), which can be defined using

diagonal embedding Ω∗(X)⊗ Ω∗(X)→ Ω∗(X ×X)δ!−→ Ω∗−dim(X)(X) for δ the diagonal map.

Remark 26. Shriek functoriality and intersection product use the fact that the target space is a manifold(Intersection product: Poincare, cup product, poincare again) (Shriek: poincare, pullback, poincare again).

Remark 27. Every mod 2 homology can be represented as bordism; this is false for integral homologies.

17 April 11

Let’s prove the theorem we forgot to prove.

Theorem 17.1. If M is a compact even-dimensional manifold, χ(M) is odd, then M is not the boundaryof a compact manifold.

Proof. Recall that χ(M) = L(idM ). To compute it, we choose some f : M → M which is homotopic tothe identity and has nondegnerated fixed points. Then χ(M) = L(f) = |Fix(f)| mod 2. Assume M = ∂N ;define P = N ∪M N . Choose a map f : M →M as before, and extend it to a map g : N → N , such that:

• g is homotopic to the identity through maps that take ∂N = M to itself.

• g has nondegenerate fixed points.

• for some choise of collar neighborhood M ⊂ U ⊂ N ϕ−→ 0×M ⊂ [0,∞)×M , we have g(ϕ−1(r, x)) =f(x) if (r, x) is sufficiently closed to the boundary.

Now define h : P → P to be two compies of g. This is homotopic to the identity, has nondegeneratefixed points, and |Fix(h)| = 1 mod 2 (things on sides are symmetric). Contradiction (recall that P is odddimensional), thus |Fix(f)| = χ(M) mod 2 must be even.

Vector bundles

Definition 34. Let π : E → M be a smooth map. Suppose that each fiber Fx = π−1(x) comes with thestructure of a (finite-dimensional) real vector space. Suppose also that those structures are locally trivial,i.e. for any x ∈ M there is an open neighborhood U ⊂ M and a diffeomorphism Ex × U → π−1(U) that iscompactible with the fiberwise vector space structure. Then we call E a real vector bundle over M .

Remark 28. This implies that π is a submersion.

A section of π : E →M is a smooth map s : M → E, π(s(x)) = x ∀x.

Page 27

Page 28: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 17 APRIL 11

Lemma 22. The space of all sections Γ(E) is canonically a module over C∞(M), the ring of real-valuedfunctions.

A homomorphism of vector bundles E → F is a smooth map φ that takes fibre to fibre and is linear oneach fibre.

Lemma 23. A vector bundle map φ : E → F induces a C∞(M)-module map Γ(E)→ Γ(F ).

Proposition 9. The category of vector bundles of bounded rank over M is equivalent to that of finitelygenerated projective C∞(M) modules.

Example 37. The tangent bundle TM →M . The fibre is the tangent space over x. We use local charts tocheck the triviality condition.

Transition Maps Take π : F →M a vector bundle of rank r, a cover M =⋃α∈A

Uα, and local trivializations

ϕα : Rr × Uα → π−1(Uα). Then ϕ(β)−1ϕα(v, x) = (Aβα(x)v, x) for some linear map Aβα(x), and that isin turn given by some smooth map Aβα : Uα ∩ Uβ → GL(r,R). These maps satisfy the cocycle conditionAγβAβα = Aγα. Conversely, from an open cover and a collection of (Aβα) we can construct a vector bundle.

Remark 29. A collection (Aβα) is a Cech 1-cocycle with values in the topological group GL(r,R). In fact,rank one vector bundles up to isomorphism are isomorphic to H1(M,GL(r,R)).

Example 38. Consider 0 → Z → C → C∗ → 1, which induces . . . → H1(M,C) → H1(M,C∗) →H2(M,Z)→ H2(M,C)→ . . .; the first and last terms vanish, and the second is GL(1,C), which tells us theusual knowledge that the isomorphism class of complex line bundle is specified by the first Chern class.

Example 39. Transition functions for TM are given by the derivatives of coordinate changes.

Constructions with vector bundles

• Direct sum E ⊕ F , where (E ⊕ F )x = Ex ⊕ Fx. Strictly speaking, the manifold underlying E ⊕ F is(v, w) ∈ E × F : (v, w) lie over the same point of M.

• Dual space E∗, where E∗x = HomR(Ex,R).

• Tensor product E ⊗ F .

• Internal hom Hom(E,F ) = E∗ ⊗ F .

• Symmetric product Symr(E) and exterior product Λr(E).

Remark 30. More systematically, any Lie group homomorphism GL(r1,R) × . . . × GL(rk,R) → GL(r,R)gives a way of constructing vector bundle of rank r from ones of rank (r1, . . . , rk).

In particular, if E has rank r, we have the determinant line det(E) = Λr(E) which is a line bundle (rank1 vector bundle).

Lemma 24. Let φ : E → F be a map of vector bundles such that rank(φx) is locally constant. Then thereare local trivializations of E,F around any point such that φ is locally constant.

Corollary 14. If φ : E → F has a locally constant rank, there are natural vector bundles ker(φ) ⊂ E, im(φ) ⊂F , and coker(φ) = E/ im(φ).

Corollary 15. If E ⊂ F is a vector subbundle, there is a natural quotient bundle F/E with map F F/E.

Consider a short exact sequence of vector bundles

0→ Eφ−→ F

ψ−→ G→ 0

Then E ∼= ker(ψ), G ∼= coker(φ) canonically. Unfortunately we do not have an abelian category due to therestriction of locally constant rank (one can’t get rid of this by restricting morphisms–note that being locallyconstant rank is not closed under composition).

Problem 17. If 0→ Eφ−→ F

ψ−→ G→ 0 is a short exact sequence of vector bundles, then there is a canonicalisomorphism det(F ) ∼= det(E)⊗ det(G).

Page 28

Page 29: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 18 APRIL 13

18 April 13

A real line bundle is a real vector bundle π : E →M of rank 1. Fibres Ex carry structures of a 1-dimensionalreal vector space, in a way which is locally constant.

Lemma 25. A real line bundle is trivial (isomorphic to the trivial bundle R) if and only if it has a nowherevanishing section.

Proof. For any E, sections correspond to vector bundle maps R→ E. A section which is nowhere zero thuscorresponds to an injective vector bundle map.

Recall that we have E ⊗ F the tensor product of vector bundles. E∗ ⊗ F ∼= Hom(E,F ) is the bundle offiberwise linear maps. For a line bundle L∗ ⊗ L ∼= Hom(L,L) ∼= R.

Corollary 16. Isomorphism classes of line bundles with tensor form an abelian group PicR(M). This actsby tensor on the set of isomorphism classes of vector bundles.

Remark 31. We can also consider complex vector bundles (or indeed left/right quaternion ones) over thereal manifold M . Everything said so far applies to complex line bundles as well, which form an abelian groupPicC(M). Given a real vector bundle, one can form its complexification E → E ⊗R C = E ⊕ iE (two copies

of E, with√−1 acting by

(0 −II 0

)). In particular, we get a group homomorphism PicR(M)→ PicC(M).

Remark 32. In fact, PicR(M) = H1(M ;Z/2) and PicC(M) = H2(M ;Z).

Unfortunately, tensor product of quaternion line bundles is not well-defined.

Problem 18. Let D ⊂ M be a hypersurface (a codimension 1 submanifold), construct a canonical linebundle LD such that sections of LD correspond to functions M → R which vanish along D.

To any line bundle L→M one can associated canonically a double covering Lor →M , where Lor = ζ ∈L : ζ 6= 0 (in its fiber) / R+ → M . More generally, given a vector bundle E we can define Eor → M bysetting Eor = bases (ζ1, . . . , ζr) of Ex/GL+(r,R). The fibre of Eor → M over x is the set of orientationsof Ex.

Example 40. For TM , TMor = Mor is the orientation cover of M .

Definition 35. An orientation of a vector bundle E →M is a continuous section of Eor →M (for E = TM ,this is the same as an orientation of M itself).

Lemma 26. There is a canonical isomorphism Eor ∼= (det(E))or.

Proof. Take (ζ1, . . . , ζr) 7→ ζ1 ∧ . . . ∧ ζr.

Corollary 17. Given a short exact sequence of vector bundle 0→ E → F → G→ 0 an orientation of anytwo of them has an orientation of the third.

Proof. Reduce to the canonical isomorphism det(F ) ∼= det(E)⊗ det(G).

Pullback of vector bundle f : M → N , given Eπ−→ N , defined f∗E = (ζ, x) ∈ E ×M | π(ζ) = f(x),

f∗(E)x = Ef(x).

Application 18. f : M → N , the derivative Dfx : TMx → TNf(x) is part of a vector bundle map

TMDf−−→ f∗TN . f is an immersion (submersion) if this is injective (onto).

Corollary 18. Let M,N be oriented manifolds. If f : M → N is a map, and y a regular value, then f−1(y)inherits a preferred orientation.

Proof. Call F = f−1(y). Then 0→ TF → TM |FDf−−→ (f∗TN)|F → 0, where (f∗TN)|F is the trivial bundle

with fibre TNy.

Page 29

Page 30: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 19 APRIL 20

Definition 36. Let N ⊂ M be a submanifold. The normal bundle of N in M is νN = (TM |N )/TN , i.e.0→ TN → TM |N → νN → 0.

Application 19. f : M → N a map, O ⊂ N a submanifold. f is transverse to O iff at each point

x = f−1(y), y ∈ O, TMxDfx−−−→ TNy → νOy is surjective.

Corollary 19. M,N oriented, f : M → N , O ⊂ N oriented submanifold transverse to f , then f−1(O)inherits a preferred orientation.

Proof. Call P = f−1(O), then 0→ TP → TM |PprojectionDf−−−−−−−−−→ f∗(νO)→ 0. On the other hand, 0→ TO →

TN |O → νO → 0. Now chase.

Remark 33. We don’t need N and O to be oriented, it’s enough if νO is oriented (i.e. O is co-oriented).

That induces a co-orientation of P , since

νP ∼= f∗(νO)

(This is important!) Then, if M is oriented, the co-orientation of P determines an orientation.

Application of partitions of unity

Proposition 10. Given a surjective map of vector bundles φ : E → F , there is a map ψ : F → E suchthat φψ = idF . This implies ψφ is an idempotent endomorphism, so E = E0 ⊕ E1 with E0 = ker(φ) and

E1φ−→ F .

Corollary 20. Given a short exact sequence of vector bundles 0 → Eφ−→ F

π−→ G → 0 there is an isomor-phism F ∼= E ⊕G such that φ is inclusion, π is projection; in other words, the exact sequence splits.

On complex geometry this is true for complex bundles on Stein manifolds, and algebraic vector bundlesover affine varieties. Of course in neither case it is generally true.

Proof. φ is of locally constant rank, M =⋃α

Uα and trivialization of E|Uα , F |Uα so that φ is constant rank

in each trivialization, Hence we get maps Ψα : F |Uα → E|Uα , Take a subordinate partition of unity, and thecorresponding weighted sum of Ψα.

Corollary 21. If N ⊂ M is a submanifold, TM |N ∼= TN ⊕ νN (non-canonically). This does not hold inalgebraic geometry.

Theorem 18.1 (Tubular Neighborhood Theorem). Let N ⊂M be a submanifold, then there is a diffeomor-phism between a neighborhood of N in M and a neighborhood of the 0 section in νN .

Proof Sketch. embed M ⊂ Rp, choose a neighborhood M ⊂ U ⊂ Rp and retraction r : U → M . νN →TM |N → Rp×N (ζ,x)7→ζ+x−−−−−−−→ U

r−→M , where the first is from TM |N splitting, and last map needs to restrictto some neighborhood. Easy computation shows this is a local diffeomorphism near each point of N .

19 April 20

Application of partitions of unity to vector bundles

Definition 37. A euclidean metric on a vector bundle E → M is a Euclidean metric on each vector spaceEx, which varies smoothly in x (with respect to local trivializations).

Remark 34. One can find local trivializations in which the Euclidean metric is constant (equal to thestandard metric on Rn). (Use Gram-Schmidt, which makes smooth choices.)

Page 30

Page 31: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 19 APRIL 20

Application 20. Let E →M be a rank 2 oriented bundle. Define Ix : Ex → Ex to be rotation by π/2 (i.e.the unique orthogonal automorphism satisfies I2

x = −id and such that (v, Ixv) is a positive oriented basis forall v 6= 0) This makes E into a complex rank 1 vector bundle. In fact, oriented Euclidean real rank 2 bundlescorrespond to complex hermitian rank 1 bundle.

Application 21. Let 0 → Eι−→ F

φ−→ G → 0 be a short exact sequence of vector bundles. If F has aEuclidean metric, there is a preferred splitting F = ι(E) ⊕ ι(E)⊥, where the second one is the fiberwiseorthogonal complement.

Lemma 27. Every vector bundle admits a Euclidean metric.

Proof. Partition of unity. (For complex vector bundles and hermitian metrics, the same applies.)

Corollary 22. The isomorphism classes of real oriented rank 2 bundles correspond bijectively to the iso-morphism classes of complex line bundles.

Corollary 23. A real line bundle L→M is trivial iff the double cover Lor →M is trivial.

Proof. Suppose Lor is trivial and choose trivialization (orientation of L). Equip L with a Euclidean metric,then in each Lx there is a unique positively oriented vector of length 1. These vectors give you a nowherevanishing smooth section.

Remark 35. In fact, PicR(M) = double covers of M =if M is connected Hom(π1(M),Z/2).

This says in particular that PicR(M) is a Z/2-vector space; in particular L⊗L is always trivial. But weknow this: given L a Euclidean metric, L ∼= L∗. Then L⊗ L ∼= L⊗ L∗ ∼= Hom(L,L) = R.

Corollary 24. For any real vector bundle, E ∼= E∗ non-canonically. For a complex vector bundle, E∗ =HomC(E,C) ∼= E (complex conjugate). Again this is not canonical.

Sard’s Theorem and Similar Transversality Techniques

Theorem 19.1. Given a vector bundle Eπ−→ M , there is a section M

s−→ E which is transverse to the zerosection. (In fact, this can be achieved by slightly perturbing any given section.)

Proof Sketch. Construct an auxiliary manifold P and a section s of q∗E → P ×M (where q : P ×M →M);This corresponds to a family of sections of M parametrized by P , which is transverse to the zero section.Then project s−1(0)→ P and choose a regular value.

How to choose P? We can easy choose some large finite-dimensional parameter space. (e.g. Suppose Mis compact, use partition of unity, and then choose all constant sections of all directions.) Or, choose P thespace of all possible sections. (Again be careful with Banach.)

Note that in algebraic geometry we might not even have any section; but if the sheaf is generated byglobal sections, they an analogous statement still holds.

Corollary 25. If E →M is a vector bundle with rank rank(E) > dim(M) then E ∼= R⊕ F for some F .

Proof. We can find a section that is nowhere zero. (See Application 19; note that preimage of the zerosection must be empty in this case, because the map cannot possibly be surjective.) That section yields aninjective map R→ E.

Example 41. There are exactly two vector bundles of any rank r over the circle. (The trivial one and theunorientable one).

The following is the counterpart of the Whitney embedding theorem for manifolds.

Theorem 19.2. Let E → M be a vector bundle of rank r. Then there is an injective bundle map E →Rr+dim(M).

Page 31

Page 32: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 20 APRIL 25

Proof. Step 1: suppose M is compact. Then we show that for any E →M there is an injective vector bundlemap E → RN , N 0. Note that for any x ∈ M there is a vector bundle map ϕx : E → Rn which is anisomorphism on the fibre at x. That map is an ismorphism also for fibres in some neighborhood Ux ⊂ Mof x. Choose a cover M by finitely many neighborhoods, and consider (ϕx1

, . . . , ϕxk) : E → Rn ⊕ . . .⊕ Rn.This is injective on every fibre.

Step 2: Still for compact M , we show that there is an injective vector bundle map Er → Rr+dim(M). Forthat, we start with an injective vector bundle map ϕ : E → RN , N > r + dim(M) = dim(E). Consider ϕjust as a map E → RN (E → RN ×M → RN ), and take s ∈ RN \ 0 to be a regular value. Then the

preimage of s is empty, hence Eϕ−→ RN → RN/R · s is still injective.

Step 3: Given an arbitrary M and a compact K ⊂M , we can find a vector bundle map E → Rr+dim(M)

which is injective at each point of K. (Just extend in whatever way you like to the whole M).Step 4: Same as step 3, but where K = K1 ∪ K2 ∪ . . . is a union of compact subsets K, which have

pairwise disjoint open neighborhoods. Then use partition of unity.Step 5: Every M can be written as a finite union of sets K as in step 4. For this, take a proper function

f : M → [0,∞) and consider f−1([0, 1]∪ [2, 3]∪ [4, 5]∪ . . .) and f−1([1, 2]∪ [3, 4]∪ [5, 6]∪ . . .). This gives aninjective map E → RN , N 0. Then repeat step 2 to reduce it to N = rank(E) + dim(M).

Corollary 26. There is some F such that E ⊕ F ∼= Rr+dim(M).

Corollary 27. Vector bundles of bounded rank correspond to projective modules of finite rank over C∞(M,R)as a category.

20 April 25

Applications of vector fields to vector bundles Let π : E → M vector bundle. Fix a vector field Xon M .

Definition 38. A linear lift of X is a vector field Y on E which, in a local trivialization, E|U ∼= Rr × Uhas the form Yv,x = (A(x)v,Xx) where A : U → Mat(r × r,R) is smooth.

Example 42. If X = 0, Y is just a collection of linear vector fields on the fibers of E. It integrates to afibrewise linear flow.

Remark 36. In local coordinates (x1, . . . , xn), X =∑k

ζk(x1, . . . , xn)∂xk, then in local coordinates (v1, . . . , vr, x1, . . . , xn)

on E, Y =∑i,j

Aij(x1, . . . , xn)vj∂vi +∑k

ζk(x1, . . . , xn)∂xk .

Remark 37. One should check that the notion of linear lift is independent of local trivializations. A change of

trivialization (v, x)F−→ (B(x)v, x) for some B(x) ∈ GL(r,R) yields a transformation rule DFF−1(v,x)(YF−1(v,x)) =

DFF−1(v,x)(A(x)B(x)−1v,Xx) = (BAB−1v+(DB ·X)B−1v,X). Note that linearity is indeed invariant (butA = 0 is not.)

If Y1 and Y2 are linear lifts of X, then so is rY1 + (1− r)Y2 for any r : M → R.

Lemma 28. Any vector field on M admits a linear lift to E.

Proof. Partition of unity.

Remark 38. The space of all linear lifts of X is an affine space over the space of fibrewise linear vectorfields (lifts of O).

Lemma 29. Let Y be a linear lift of X. If the flow of X is well-defined on U ⊂ R ×M , the flow of Y iswell-defined over the preimage V = (idR × π)−1(U) ⊂ R× E. These flows fit into a diagram

R× E ⊃ V E

R×M ⊃ U M

id×π

ψ

π

ϕ

Page 32

Page 33: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 20 APRIL 25

Just as a reminder, the R is the time. Moreover, ψ is a linear map between fibres, by which we mean thefollowing: Let i0 : M → R×M , p : R×M →M be inclusion and projection.

Theorem 20.1. For any vector bundle E → R×M , there exists an isomorphism p∗i∗0E → E which is theidentity on the fibres over 0 ×M .

This means that we have a family of linear isomorphisms Ψt,x : E0,x → Et,x which are the identity fort = 0, and which are smooth in local trivializations.

Proof. Take a linear lift of X = ∂t, and take its flow.

Corollary 28. Given a vector bundle E → N and two homotopic maps f0, f1 : M → N , we have f∗0E = f∗1E.

Proof. Let the homotopy be R×M F−→ N such that f0 and f1 are the two endpoints. Then f∗0E∼= F ∗E|0×M

and f∗1E∼= F ∗E|1×M . Then apply the theorem to F ∗E.

Corollary 29. If M is contractible, any vector bundle over M is trivial.

Application 22. Take a vector bundle Er → Sn. Then E is trivial on each hemisphere. Choosing triv-ializations and comparing them yields a map Sn−1 → GL(r,R). The homotopy class of this map (up tomultiplication by a constant) determines the isomorphism class of the vector bundle.

Remark 39. Take n > 1 Then one can arrange that the clutching function Sn−1 → GL(r,R)+ and itshomotopy class is now unique. In fact, one can achieve that the map takes values in SO(r).

Thus vector bundles on spheres correspond to homotopy classes of SO.

Problem 19. Show that any vector bundle over S3 is trivial. (Try not to use too much homotopy theory.Use only fundamental groups if you can.) Higher rank: can split off trivial ones. Rank 1: double cover(nothing happens). Rank 2 and 3 are interesting.

Recall that a Grassmannian Gr(r,N) is the set of r-dimensional linear subspaces of RN , which canbe described also as the set of symmetric idempotent matrices of rank r and size N . (This shows it’s asubmanifold of the space of all matrices.)

The trivial vector bundle RN → Gr(r,N) carries a canonical idempotent endomorphism (namely, at A,apply A). Its image is the tautological bundle τ → Gr(r,N).

Lemma 30. If Er → Mn, then E is isomorphic to the pullback of τ by some map M → Gr(r,N) for anyN ≥ r + n.

Proof. We know that E is a direct summand of RN →M . Writing it in that way gives a map M → Gr(r,N)(mapping each point to its fibre in E), for which the statement is trivial.

Lemma 31. Consider two maps f0, f1 : M → Gr(r,N) for N > r + n. If the vector bundles f∗0 τ, f∗1 τ are

isomorphic, the maps f0, f1 must be homotopic. (Notice the condition on dimension is different from that ofabove.)

Proof. This is equivalent to the following statement: given two surjective vector bundle maps ϕ0, ϕ1 : RN →E over M , there is a surjective vector bundle map ϕ : RN → p∗E over R×M such that ϕ restricts to ϕ0, ϕ1

on the two ends. But that follows from the Sard-style arguments from last time.

Corollary 30. Over Mn, vector bundles of rank r up to isomorphism corresponds to maps M → Gr(r,N)up to homotopy, where N > n+ r.

Remark 40. We now have two descriptions of vector bundles over Sn: as πn(Gr(r,N)) for N 0 andπn−1(SO(r)), so these two must be the same. In fact, ΩGr(r,∞) ∼= O(r) where ∞ means taking the limit.

Page 33

Page 34: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 21 APRIL 27

21 April 27

Problem 20. M compact manifold, ϕ : M → M involution, which has an even (finite) number of fixedpoint. Show that M is a boundary (i.e. trivial in the bordism ring).

Consider the Grassmannian Gr(k,N) and its tautological bundle τ . There is a one-point compactification

τ+ = τ ∪ ∞. Suppose that we are given a continuous map Sn+k ϕ−→ τ+. τ+ is not a manifold, but we canperturb our map to be smooth outside a small neighborhood of ϕ−1(∞). Moreover, a further perturbationensures that ϕ is transverse to the zero section. Then M = ϕ−1(Gr(k,N) ⊂ τ+) ⊂ Sn+k is a smoothn-manifold. Moreover, its normal bundle in Sn+k is isomorphic to (ϕ|M )∗τ .

Now let’s turn this construction around.

Pontryagin-Thom construction Given any compact n-manifold M , take an embedding M → Sn+k forsufficiently large k, the normal bundle is isomorphic to the pullback of τ by some map M → Gr(k,N) forsome large N . Using the tubular neighborhood theorem, we extend this to a proper map U → τ whereU ⊂ Sn+k is a neighborhood of M . Then extend this to Sn+k → τ+ by sending all of Sn+k \ U to ∞.

Notice these embeddings are not unique; they depend on the embedding of M chosen, but we have thefollowing:

Theorem 21.1 (Pontryagin-Thom). Ωn ∼= πn+k(τ+ ⊂ Gr(k,N)) provided that k > n,N > k + n.

The RHS is independent of N (directly checkable), and is independent of k due to stable homotopytheory.

Corollary 31. Ωn is finite in each dimension.

Definition 39. Let M be a compact n-manifold. We say M is stably parallizable if TM ⊕ Rr ∼= Rr+n forsome r. In particular, spheres (more generally, hypersurfaces) are parallizable.

Suppose that M is stably parallizable. Then, for any embedding M → Rn+k, the normal bundle is alsostably trivial, since ν ⊕ TM ∼= Rn+k =⇒ ν ⊕ Rr+n ∼= ν ⊕ TM ⊕ Rr ∼= Rn+k+r. Hence, we can (afterincreasing k) embed M → Sn+k such that the normal bundle is actually trivial. Apply the Pontryagin-Thomconstruction to this embedding, then the map to Gr(k,N) is a constant, and the map U → τ lies in a singlefibre of τ . Hence the whole map Sn+k → Sk → τ+.

For instance, any stably parallelizable 4-manifold is a boundary of a 5-manifold, since the fourth stablyhomotopy group of sphere is trivial.

In general, given a map Sn+k → (τ ×X)+ for some compact X, let U be the part that doesn’t map toinfinity. Take the projection to τ and then preimage of the zero section, then we get a manifold M . Onthe other hand, by projecting to X we get a map M → X. The converse is also true, and thus we haveΩn(X) ∼= πn+k((τ ×X)+).

Invariants of vector bundles Let M be a compact manifold. E → M a rank r vector bundle. Lets : M → E be a section that is transverse to the zero section. Then Z = s−1(M ⊂ E) is a submanifold ofdimension n− r.

Definition 40. The unoriented Euler class e(E) is [Z] ∈ Ωn−r(M). This is an invariant of E (we caninterpolate sections, etc).

We have the following properties:

• e(E ⊕ R) = 0. (Consider the unit R-section.)

• e(E ⊕ F ) = e(E)e(F ) where we use the intersection product.

Example 43. M = RPn = Gr(1, n + 1), E = τ the tautological bundle, then e(E) = [RPn−1 → RPn] thehyperplane inclusion. Therefore, e(nE) = [∗ → RPn] 6= 0, thus nE is nontrivial.

Page 34

Page 35: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 22 MAY 02

(Notation remark: nE = E ⊕ E ⊕ . . .⊕ E︸ ︷︷ ︸n copies

, and ne(E) = e(E) · e(E) · . . . · e(E)︸ ︷︷ ︸ncopies

, where · is the intersection

product.)

Remark 41. If M and E are oriented, we can define the oriented Eulcer class e(E) ∈ ΩSOn−r(M).

The simplest special case is that if L is a real line bundle, e(L) ∈ Ωn−1(M).

Definition 41. The (bordism) first Stiefel-Whitney class of E →M is w1(E) = e(det(E)) ∈ Ωn−1(M).

We have the following properties:

• w1(E ⊕ R) = w1(E).

• w1(E ⊕ F ) = w1(E) + w2(F ). This is harder to prove.

How do we get higher SW classes?

Grothendieck Construction Take Er → Mn vector bundle. There is an associated bundle of realprojective spaces RP(E)→M , the fibre of this at any point x ∈M is the projective space of Ex. Moreover,this carries a topological line bundle τ → RP(E).

For any k, consider e((r+k−1)τ). Its image under RP(E)→M is a class in Ωn−k(M). All these classesare invariants of the bundle E. For line bundle E, these are just powers of e(E) = w1(E).

22 May 02

Mn compact manifold, Er → M vector bundle, RP(E) → Mn bundle of projective spaces. (A point inRP(E) is a pair (x, λ) consisting of x ∈M and a 1-d linear subspace λ ⊂ Ex).

τE → RP(E) tautological line bundle. (The fibre of τE over (x, λ) is exactly λ, it is a subbundle of thepullback of E to RP(E)).

Hence we have e(τE) = [zero set of a generic section] ∈ Ωn+r−2(RP(E)). More generally, ie(τE) =e(iτE) = [common zero set of i generic sections of τE ] ∈ Ωn+r−1−i(RP(E)).

Definition 42. ui(E) ∈ Ωn−i(M) is the image under RP(E)→M of (i+ r − 1)e(τE).

These bordism classes are invariants of the vector bundle E. (We can call them the “bordism-theoreticcomplementary Stiefel-Whitney classes”.)

Example 44. If E is a line bundle, ui(E) = ie(E).

Theorem 22.1. ui(E) = ui(E ⊕ R).

Proof. For simplicity, let’s equip E with an inner product, and E⊕R with the corresponding inner product.Then, τE⊕R has a distinguished section, which is given by projecting the unit section of R to the tautologicalline. This section vanishes exactly along RP(E) ⊂ RP(E ⊕R). Therefore, e(τE⊕R) · . . . · e(τE⊕R) (j of them)equals [RP(E) → RP(E ⊕ R)]· (j − 1 of them) = image of (j − 1 of them) under the injection map. Now,projection to M establishes the desired result.

One example: suppose M = ∂W , then we have 0→ TM → TW |M → R→ 0, so TM ⊕ R ∼= TW |M . Inorder to study ui(TM), it is therefore sufficient to study ui(TW |M ).

Given a partition n = i1 + . . . + ik, define ui1,...,ik(M) = ui1(TM) · . . . uik(TM) mapped to Ω0∼= Z/2.

These are called complementary Stiefel-Whiteney numbers. (They are invariants of manifold, and are in facthomotopy invariants.)

Theorem 22.2. If M is the boundary of a compact manifold, then ui1,...,ik(M) = 0.

Hence, in fact, ui1,...,uk : Ωn → Z/2 are group homomorphisms. Also, though we’ll not prov it, this is infact an iff statement.

Page 35

Page 36: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 23 MAY 04

Proof. Let M = ∂W , and consider ui(TM) = ui(TM ⊕ R) = ui(TW |M ), which is the intersection of thesame map to W with M . Therefore, the whole product is the boundary of a compact 1-manifold mappingto W .

Here’s an example. M = RPn, τ the tautological bundle. Let TM` be ways to move ` around in Rn+1

to first order, i.e. Hom(τ,Rn+1/τ). Hence we have 0 → Hom(τ, τ) → Hom(τ,Rn+1) → TM → 0. In otherwords, TM ⊕R ∼= τ ⊕ . . .⊕ τ = τ ⊗Rn+1. Take E = TM ⊕R = τ ⊗Rn+1, then RP(E) ∼= RPn×RPn. (Thisfollows from RP(V ) ∼= RP(V ⊗W ) for any 1-d W .) Then τE over RP(E) is τ over the first RPn tensor the τover the second one. Therefore, e(τE) = [RPn−1×RPn] + [RPn×RPn−1] ∈ Ω2n−1(RPn×RPn). Then j fold

product of e(τE) =∑p+q=j

(j

p

)[RPn−p × RPn−q] ∈ Ω2n−j(RPn × RPn). Project and get ui(E) = projection

of i + n copies of e(τE) =

(i+ n

i

)[RPn−i] + junk ∈ Ωn−i(RPn) where for the junk part we mean bordism

classes which have lower dimensional image.

For n = 2, i.e. M = RP2, u1(TM) =

(3

1

)[RP1] = [RP1], u2(TM) =

(4

2

)[RP0] = 0. Hence u2 = 0, u11 =

u1(TM) · u1(TM) = [RP1] · [RP1] = 1. This gives another proof that RP2 is not a boundary.

Problem 21. Compute ui1,...,ik for RP3 and RP4. (3: u3, u12, u111, 4: u4, u13, u22, u112, u1111).

23 May 04

Let Mn be a manifold. A stable framing of M is an isomorphism of vector bundles TM ⊕RN−n ∼= RN . Twostable framings are considered the same if they differ by extending further Rs. Stable framings that can bedeformed into each other are also equivalent.

Lemma 32. Given a stable framing over M , the set of all possible stable framings is [M,O(N)] for N 0,i.e. the homotop classes of maps M → O(N).

Example 45. ∗ has two stable framings, as π0(O(N)) = Z/2. Similarly, S1 has four stable framings,with two compatible with each orientation, since π1(O(N)) = Z/2. (π1(SO(2)) = Z, π1)SO(N)) = Z/2 forN > 2.)

The two framings of S1 are the trivial framing (the usual trivialization of TS1) and the nontrivial framing(where everywhere TM ⊕R has the same “direction”; need a picture here.) For the trivial framing, add theunit normal bundle to the trivialization TS1, and see the resulting bundle is not isomorphic to the nontrivialframing.

Recall that if M = ∂W for some manifold W , then we have TM⊕R ∼= TW |M (where the R is the “choiceof pointing outward vector”; this choice is not canonical nor unique, but any two choices can be deformedinto each other.) Therefore a stable framing of W induces a stable framing of M . We can therefore definethe stably framed bordism group Ωfr

n .

Example 46. Ωfr0 = ΩSO

0 = Z. For each line segment, the R factor points towards one end, so the framingon the point on that end becomes “+1” and the other end becomes “-1”.

Example 47. Ωfr1 = Z/2.

To show this we can appeal to the classification of 1-d manifolds. There are two framings up to orientationfor S1, and the nontrivial framing obviously bounds (one can obviously extend that nontrivial framing tothe entire disk), and two trivial framings bound a cylinder, so it remains to show that the trivial framingdoesn’t bound.

For that, take W to be a genus g oriented surface with one boundary compact. By hand, one canfind a framing of W which restricts to the nontrivial framing on ∂W = S. On the other hand, framingsof W are parametrized by [W,SO(N)] ∼= Hom(π1(W ), π1(SO(N))) (by cellular approximation), which isisomorphic to Hom(π1(W ),Z/2). On the other hand, the framing of ∂W is parametrized by [∂W,SO(N)] ∼=Hom(π1(∂W ),Z/2), and we have an obvious diagram:

Page 36

Page 37: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 23 MAY 04

[W,SO(N)] Hom(π1(W ),Z/2)

[∂W,SO(N)] Hom(π1(∂W ),Z/2)

i∗

∼=

i∗

∼=

where both sides are induced by the inclusion i∗ : ∂W →W . However, the RHS map is zero, because Z/2is abelian so we have Hom(π1(X),Z/2) = Hom(H1(X),Z/2) = H1(X;Z/2), so the map is H1(W ;Z/2) →H1(∂W ;Z/2), whose image is the kernel ofH0(∂W ;Z/2) ∼= H1(∂W ;Z/2)→ H2(W,∂W ;Z/2) ∼= H0(W ;Z/2),which is trivial.

Therefore, any two framings of W (compatible with orientation) induces the same framing of ∂W , thus∂W is always nontrivially framed.

Example 48 (Outline of the 2-d case). Ωfr2∼= Z/2.

Historical remark: Pontryagin made a mistake here by thinking this object is zero. To see how to derivethe statement above, let M be an oriented surface. A stable framing of M leads to a quadratic refinementof the mod 2 intersection pairing, by which we mean a quadratic form q : H1(M ;Z/2) → Z/2 such thatq(x+ y)− q(x)− q(y) = x · y, where the RHS is the intersection product.

Of course, in general, given any refinement of ·, any affine shifting of the refinement is again a refinement.But given a class in H1(M,Z/2), one can represent it by an embedded circle in M , and then restrict thestable framing of M to the loop, giving q(x) ∈ Ωfr

1∼= Z/2. (Reference: [Johnson 1980, Spin Structures and

Quadratic forms on Surfaces.]) Then the map Ωfr2 → Z/2 is given by the Arf invariant Arf(q).

In fact, Ωfr∗ was implicitly discovered before the normal bordism classes! Here’s why. Consider a map

Snf−→ SN−n. The preimage of a regular value is a manifold Mn ⊂ SN whose normal bundle comes with a

preferred trivialization. Hence, M comes with a preferred stable framing

TN ⊕ νM ⊕ R︸ ︷︷ ︸trivial

∼= TSN |M ⊕ R ∼= RN+1︸ ︷︷ ︸framing of SN

This yields a map from stable homotopy groups to stable framing bordism groups

[SN , SN−n ∼= πN (SN−n)]→ Ωfrn

Theorem 23.1 (Pontryagin-Thom). This is an isomorphism for N ≥ 2n− 1.

Here’s how to construct the reverse direction map: given M with a stable framing, embed it into RN ⊂ SNfor large N with trivial normal bundle. Map tubular neighborhood to RN−n (with M 7→ 0) and the rest ofthe space going to ∞.

Example 49. π4(S3) = Z/2.

One can also introduce Ωfr∗ (X), which is the bordism of stably framed manifolds M with a map M → X.

Theorem 23.2. Ωfrn(X) ∼= πN ((RN−n×X)+) for N 0. The RHS is called the stable homotopy groups of

X (or X+).

Remark 42. We have canonical forgetful maps Ωfr∗ → ΩSO

∗ → Ω∗. Unfortunately, these maps always vanishin positive degrees and are therefore uninteresting. There is an interesting alternative, the spin bordism Ωspin

∗ ,such that Ωfr

∗ → Ωspin∗ is (almost) computable and nontrivial. Unfortunately we don’t have time to cover that

in this course.

Preview for the next class: take p : Mm → Nn a proper submersion. We’ll define Ωfr∗ (N)→ Ωfr

∗+m−n(M)(note that this is a shriek map) called the Becker-Gottieb transfer map. This is a map “in the wrongdirection” and is rather difficult to construct in pure AT setting; but the differential story is more pleasant.In particular, if N is connected, Ωfr

0 (N) = Z and we get a distinguished class in Ωfrm−n(M), which are

generalizations of Euler classes.

Page 37

Page 38: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 24 MAY 09

24 May 09

Euler characteristics Let M be a compact manifold, then χ(M) ∈ Z.

Definition 43. Let’s recall its definition.

• One could embed M → RN , and let r : U → M ⊂ U a retraction of a local neighborhood, thenχ(M) = L(r) is the Lefschetz fixed point number of r.

• Or, we can define χ(M) = L(idM ). (For this, we perturb idM to a self-map f with nondegenerate fixed

points, and L(f) =∑

f(x)=x

sgn det(1−Dxf), where 1−Dxf is an endomorphism of TMx).

• χ(M) mod 2 is the image of the Euler class of the tangent bundle e(TM) ∈ Ω0(M) under the mapΩ0(M) → Ω0(∗) = Z/2. This counts the zeros mod 2 of a vector field transverse to the zero section.If ξ is such a vector field, and x ∈ M is a point at which ξ vanishes, then Dξx : TMx → TMx is an

isomorphism. Then χ(M) =∑ξ(x)=0

sgn det(Dξx).

Let’s see 2 and 3 are the same: consider a flow of short time, then the fixed points of the flow are exactlywhere ξx vanishes. Additionally, Dxf is the exponential of Dξx, so the equivalence follows.

Remark 43. If M is odd-dimensional, then χ(M) = 0. (Consider L(f) where f is close to identity; if wemultiply by L(f−1), we multiply by (−1)n, but we’re supposed to have the same value.)

Now, let p : M → N be a submersion between compact manifolds. Then each fibre p−1(y) = My is itselfa manifold, and thus family of manifolds is locally trivial (Ehresmann’s theorem). Consider the diagram

M M

N N

f

p p

id

Let x be a fixed point of f , y = p(x). Then

TMx TMx

TNy TNy

Dfx

Dpx Dpx

id

Then id −Dfx : TMx → ker(Dpx) ⊂ TMx. Locally speaking, if f(z, y) = (h(z, y), y) (fibre, base) then(id− f)(z, y) = (h(z, y)− z, 0).

We say that x is nondegenerate if im(id − Dfx) = ker(Dpx). For a generic f , this holds for all fixedpoints. Then Fix(f) ⊂ M is a submanifold of dimension dim(N). (But p|Fix(f) : Fix(f) → N is not asubmersion.) (Locally speaking: x is a nondegenerate value iff 0 is a regular value of h(z, y)− z.)

Note that the tangent space T (Fix(f)x) = ker(id − Dfx) =⇒ T (Fix(f)x) ⊕ ker(Dpx) ∼= TMx∼=

TNy ⊕ ker(Dpx). Hence, as stable vector bundles we have T (Fix(f)) ∼= p∗TN |Fix(f).Suppose that N is stably framed (TN is stably trivial) Then so is Fix(f) and we get an invariant

L(f) = [Fix(f)] ∈ Ωfrdim(N). This is called Dold’s parametrized fixed point number.

Example 50. N = S1. L(f) ∈ Ωfr1 = Z/2.

In particular, for f = id, χ(M → N) = L(id) ∈ Ωfrdim(N). This is the parametrized Euler characteristic

of Becker and Gottlieb.There is also an equivalent definition of χ(M → N) using vector fields in ker(Dp) ⊂ TM .In general (if N is not stably framed, or not even compact), we get L(f) : Ωfr

∗ (N)→ Ωfr∗ (M) (Dold fixed

point transfer) and χ(M → N) : Ωfr∗ (N)→ Ωfr

∗ (M) (Becker-Gottlieb transfer).

Page 38

Page 39: 18.966 Di erential Topology Lecture Notespeople.math.harvard.edu/~yfu/notes/966notes.pdf · 18.966 Di erential Topology Lecture Notes May 15, 2016 This class was taught by Prof. Paul

18.966 Notes 24 MAY 09

Definition of the fixed point transfer: take a class in Ωfrk (N), represented by some q : Q→ N , where Q is

stable framed. Consider the pullback q∗M → Q, where the first is (z, x) ⊂ Q×M : q(z) = p(x). q∗f actson q∗M , then consider [Fix(q∗f)] ∈ Ωfr

k (q∗M)→ Ωfrk (M).

Now let’s talk about something else. Reference: [Zhang, A Counting Formula for the Kervaire semi-characteristic, Topology (2000)] or [Zizhou, Bordism theory and the semi-characteristic, Sci. China (2002)].

Suppose M is a compact manifold, oriented connected, χ(M) = 0.

Lemma 33. If χ(M) = 0, then M admits a nowhere vanishing vector field.

Sketch of Proof. Take a vector field ξ transverse to 0. wlog all zeroes of ξ are contained in (the interiorof) a single coordinate ball Bn ⊂ M . Now look at ξ : Bn → Rn, ξ(x) 6= 0 on ∂Bn = Sn−1. Considerξ/‖ξ‖ : Sn−1 → Sn−1. If this is homotopy to a constant, then we can find another map Bn → Rn withthe same boundary values and which has no zeroes. But [Sn−1, Sn−1] = Z, and this is actually (when onecounts it out) the counting of zeros of ξ.

Lemma 34. If χ(M) = 0, M admits a nowhere zero section of T ∗M .

Fix ξ section of T ∗M which is nowhere zero. Take E = ker(ξ) ⊂ TM which is an oriented subbundleof rank (n − 1). (Now can we get a section of E that is nowhere zero?) Let η be a section of E which istransverse to the zero section. The zero-set of η is a finite union of circles. Let C be one of these circles.Then Dη : νC → E|C is an isomorphism of vector bundles over C. Note that ξ : TM → R satisfies ξ η = 0by definition.

Claim ξ∗ : TM → R is nonzero on TC. (This is iff TC is nowhere contained in E|C .) (Proof?)Then E|C → TM |C → νC is an isomorphism. Combining this and the Dν isomorphism before gives an

automorphism of E|C , hence an element of π1(GLn−1(R))→ Z/2. This defines o(C) ∈ Z/2.

Definition 44. κ(M) = #C : o(C) = 0 ∈ Z/2.

Theorem 24.1. This is an invariant of M . (This is “the next invariant” after Euler characteristic.)

Let’s do an example. Let M = N × S1 where S1 has coordinate t, ξ = dt, and E is TN pulled back toM . η a generic section of TN pulled back to M . If p(x) = 0, then η vanishes on C = x × S1, o(C) = 0,so κ(M) = χ(N) mod 2. It’s only interesting when N has dimension being a multiple of 4, but that case isinteresting. (Notice that multiplying by S1 kills the Euler characteristic.)

Theorem 24.2. κ(M) = 0 when n ≡ 2, 3 (mod 4), κ(M) =∑i

dimH2i(M ;R) when n ≡ 1 (mod 4) (the

Kervaire semi-characteristic), and sign(M)/2 when n ≡ 0 (mod 4) (the signature).

Page 39