Patrick Lei

All notes

Algebraic topology II

Lectures by Francesco Lin, Columbia, Spring 2021.

Algebraic Topology
Spring 2021
Notes by Patrick Lei
Lectures by Francesco Lin

Columbia University

Disclaimer

These notes were taken during lecture using the vimtex package of the editor neovim. Any errors are mine and not the instructor’s. In addition, my notes are picture-free (but will include commutative diagrams) and are a mix of my mathematical style and that of the instructor. If you find any errors, please contact me at plei@math.columbia.edu.

Acknowledgements

I would like to thank Nicolás Vilches and Alex Xu for pointing out errors and typos in these notes. I would also like to thank Alex for graciously offering to teach everyone about the Dirac operator. Finally, I would like to thank Google for Gemini 3 Pro allowing me to upscale the original hand-drawn figures to Tikz.

1 Spectral Sequences

Our goal in this chapter will be to compute the homotopy groups of spheres, but we are not algebraic topologists so we don’t actually care about that. What we know from the basic theory is that πi⁢(Sn)=0 if i<n and πn⁢(Sn)=ℤ, and this isomorphism is given by the degree. We also know that π3⁢(S2)=ℤ and is generated by the Hopf fibration S1↪S3→S2. Note that this map is the attaching map of the 4-cell of ℂ⁢ℙ2.

There is an analogous decomposition ℍ⁢ℙ2=pt∪D4∪D8. Then we obtain a fibration f:S7→S4, which is an S3-fibration. Using the long exact sequence of the fibration, we see that π7⁢(S4)⊇ℤ.

Next, we may consider 𝕆⁢ℙ2, the octonionic projective plane. This has a cell decomposition pt∪D8∪D16 and from this we obtain a fibration S7↪S15→S8 and compute that π15⁢(S8)⊇ℤ. We should note, however, that 𝕆⁢ℙ2 is not the set of lines in 𝕆3 because the octonions do not form an associative algebra. There is enough associativity to define some sort of 𝕆⁢ℙ2 but not 𝕆⁢ℙn for n≥3. In fact, we will prove that no space with the expected cohomology ring exists.

More generally, we can study f:S2⁢n−1→Sn as follows: Consider the mapping cone Sn∪fD2⁢n≕Cf. Then the cohomology is

Then choose generators α∈Hn,β∈H2⁢n. We then have α2=H⁡(f)⁢β∈H2⁢n for some integer H⁡(f), called the Hopf invariant.

Example 1.0.1.

The attaching maps of ℂ⁢ℙ2,ℍ⁢ℙ2,𝕆⁢ℙ2 all have H⁡(f)=1. This implies that H∗=ℤ⁡[α]/(α3).

We know that if f,g are homotopic, then H⁡(f)=H⁡(g) because the mapping cones are homotopy equivalent. Then we obtain a homomorphism H:π2⁢n−1⁢(Sn)→ℤ. Note that when n is odd, we have α2=−α2 by graded commutativity, and so H⁡(f)=0.

Example 1.0.2.

If n=2⁢k is even, there exists f:S4⁢k−1→S2⁢k with H⁡(f)=2.

Now given a pointed space (X,e) define J2⁢(X)=X×X/(x,e)∼(e,x). This is somewhere between the product and the smash product, and so J2⁢(S2⁢k) is a CW complex with a single cell in dimensions 0,2⁢k,4⁢k.

Exercise 1.0.3.

The attaching map of the 4⁢k-cell has H⁡(f)=2.

Corollary 1.0.4.

π4⁢k−1⁢(S2⁢k) admits a ℤ-summand. This is because H surjects onto either 2⁢ℤ or ℤ, and so it splits.

We have the following results:

Theorem 1.0.5 (Serre).

πi⁢(Sn) is a finitely generated abelian group. It has rank 1 if i=n or i=2⁢n−1 and n is even.

Theorem 1.0.6 (Adams).

If [f]∈π2⁢n−1⁢(Sn) with H⁡(f)=1, then n=2,4,8.

Remark 1.0.7.

This is related to the following. Suppose ℝn admits the structure of a division algebra. This is a bilinear map ∗:ℝn⊗ℝn→ℝn that is invertible for a≠0. Then n=1,2,4,8. We will prove this result using K-theory.

1.1 The Simplest Case

Here, we will consider πn+1⁢(Sn) for n≥3. By the Freudenthal suspension theorem, we have maps

Therefore the group we need to compute is π4⁢(S3). The key strategy is to exploit the interaction between homotopy and homology. Here, we will consider the Hurewicz map

Theorem 1.1.1 (Hurewicz).

Suppose n≥2 and X is (n−1)-connected. Then H~0⁢(X)=H1⁢(X)=⋯=Hn−1⁢(X)=0 and h:πn⁢(X)≃Hn⁢(X).

Remarks 1.1.2.

When n=1, H1⁢(X) is the abelianization of π1⁢(X). Also, there is a relative version of the Hurewicz theorem.

Sketch of Proof.

We can assume X is a CW complex with a single 0-cell and no cells in dimension 1,…,n−1. Then we can replace X with Xn+1, so we can write

By homotopy excision, we have

and this is exactly Cn+1⁢(X)→Cn⁢(X)→0. □

Our strategy to compute π4⁢(S3) will be to construct a space whose homology is π4⁢(S3). Recall that π3⁢(K⁡(ℤ,3))=ℤ by definition, and so consider f:S3→K⁡(ℤ,3) be a generator. Then f∗:π3⁢(S3)→∼π3⁢(K⁡(ℤ,3)) is an isomorphism. Now we can turn f into a fibration F↪S3→K⁡(ℤ,3). Considering the long exact sequence of the fibration, we have

and therefore π0⁢(F)=π1⁢(F)=π2⁢(F)=π3⁢(F)=0. We also have πn⁢(F)≅πn⁢(S3) for n≥4. In particular, we have π4⁢(S3)=π4⁢(F)=H4⁢(F).If we can understand F in some reasonable way, then we will be done.

If we turn F↪S3 into a fibration, then the homotopy fiber is now a K⁡(ℤ,2)=ℂ⁢ℙ∞. Now we have a fibration and we know the homology of both S3 and ℂ⁢ℙ∞, so the question is to compute H4 from the information we have.

Question 1.1.3.

Given F↪E→B a fibration, is there a relationship between H∗⁢(B),H∗⁢(E),H∗⁢(F)?

Recall that in the case when E≅F×B this relation is given by the Künneth formula. Here, we see that C∗⁢(E)=C∗⁢(F)⊗C∗⁢(B) as chain complexes, and so we reduce the problem to homological algebra. This gives us

Theorem 1.1.4.

Let R be a principal ideal domain. Then there is a natural short exact sequence

and this sequence splits (but not naturally).

Example 1.1.5.

We can compute that H∗⁢(S1×S2)=ℤ in dimensions 0,1,2,3.

However, we note that the Kunneth theorem does not hold for fibrations in general. To see this, consider the Hopf fibration.

Example 1.1.6.

Consider the fibration K⁡(ℤ,n−1)↪∗→K⁡(ℤ,n). Both the base and fiber have nontrivial homology, but clearly the total space is contractible, so it has trivial homology. In both examples, H∗⁢(E) is smaller than what we would get from Künneth.

1.2 Spectral Sequences

Definition 1.2.1.

A spectral sequence is a sequence (Er,dr), where Er is an R-module and dr:Er→Er is a differential. In addition, we require that Er+1=H∗⁢(Er,dr).

Remark 1.2.2.

Note that (Er,dr) determines Er+1, but not dr+1 in general.

Now assume that dr=0 for r≫0 so that Er=Er+1=⋯≕E∞.

Definition 1.2.3.

We say that (Er,dr)⇒G, or (Er,dr) converges to G, if G admits a filtration

such that ⨁Gi/Gi−1≅E∞.

Remark 1.2.4.

This says that G is recovered by (Er,dr) up to extension problems.

Example 1.2.5.

Consider a short exact sequence 0→A∗→C∗→C∗/A∗→0 of chain complexes. Then we will see that there is a spectral sequence with E1=H∗⁢(A)⊕H∗⁢(C/A) and (Er,dr)⇒H∗⁢(C).

First, consider the long exact sequence in homology. If we consider the boundary homomorphism δi, we obtain a long exact sequence

Now define E1=H∗⁢(A)⊕H∗⁢(C/A) with d1=⨁δi. Then we have

as desired. Then we let d2=0.

This tells us that if H∗⁢(A)=H∗⁢(C/A)=0, then H∗⁢(C)=0. Also, if H∗⁢(C)=0, then H∗⁢(A)≅H∗⁢(C/A) with a shift.

Theorem 1.2.6 (Serre).

Let F↪E→B be a fibration and let π1⁢(B)=1. Then there is a spectral sequence (Er,dr) such that

  • •

    E2≅H∗⁢(B;H∗⁢(F));

  • •

    (Er,dr)⇒H∗⁢(E).

Note here that we need to define what it means to converge in this setting. Also, note that E3 is smaller than E2, so this formalizes the notion that H∗⁢(E) is smaller than what we naively expect.

Remark 1.2.7.

We actually do not know what d2 is, so it is impossible to compute the homology in general.

Fortunately, there is more structure:

  • •

    E∗,∗r is bigraded.

  • •

    dr has bidegree (−r,r−1).

  • •

    Ep,q2=Hp⁢(B;Hq⁢(F)).

  • •

    Ep,qr=Ep,q∞ for r≫0 depending on p,q.

  • •

    Hn⁢(E) has a filtration with associated graded ⨁Ei,n−i∞.

Example 1.2.8.

Consider the example S1↪∗→ℂ⁢ℙ∞. Then we know that Ep,q2=Hp⁢(B,Hq⁢(S1)) and therefore the spectral sequence is

0123401H0⁢(B)H1⁢(B)H2⁢(B)H3⁢(B)…H0⁢(B)H1⁢(B)H2⁢(B)H3⁢(B)…

This gives us δi:H2+i⁢(B)→Hi⁢(B), and so d3 has degree (−3,2) and thus it has to be zero. By degree reasons, we see that E4=E3 and d4=0, so E∞=E3. Then, the total space is a point, and so writing the E3-page

0123401coker⁡δ0coker⁡δ1coker⁡δ2H0⁢(B)H1⁢(B)ker⁡δ0ker⁡δ1ker⁡δ2

we see that H0⁢(B)=ℤ,H1⁢(B)=0 and that ker⁡δi=coker⁡δi=0. This recovers the usual computation of the homology of ℂ⁢ℙ∞.

Example 1.2.9.

We want to compute H∗⁢(Ω⁢S2). Here, we will use the fiber sequence Ω⁢S2↪∗→S2. Here, we know that Ep,q2=Hp⁢(S2,Hq⁢(F)), so the E2-page looks like

0120123H0⁢(F)H1⁢(F)H2⁢(F)H3⁢(F)H0⁢(F)H1⁢(F)H2⁢(F)H3⁢(F)g0g0g1g1g2g2

Then the E3-page looks like

0120123H0⁢(F)coker⁡g0coker⁡g1coker⁡g2ker⁡g0ker⁡g1ker⁡g2ker⁡g3

and the differential has degree (−3,2), so d3=0. This implies that E∞=E3, and so the associated graded pieces are ℤ,0,0,0. We see that H0⁢(F)=ℤ and each gi is an isomorphism. This tells us that Hi⁢(Ω⁢S2)≅ℤ for all i.

Example 1.2.10.

We can enhance this example to Ω⁢Sn↪∗→Sn. In this case, the first possible nontrivial differential is dn, which has degree (−n,n−1). This gives us En+1=E∞ for degree reasons, so we can compute

because δi:Hi⁢(F)→Hi+n−1⁢(F) is an isomorphism.

Recall that our goal was to study ℂ⁢ℙ∞↪F→S3. However, we cannot compute this yet, so we will need to introduce even more structure. To do this, we will need to do some homological algebra.

Considered a filtered chain complex, which is an abelian group C with a map d:C→C such that d2=0. To this, we attach a filtration 0=F−1⁢C⊆F0⁢C⊆⋯⊆Fn⁢C=C such that d⁡(Fi⁢C)⊆Fi⁢C.

Remark 1.2.11.

(Fi⁢C,D) is a subcomplex of C.

Now if X is a topological space with filtration X−1=∅⊂X0⊂⋯⊂Xn=X. Then if C∗⁢(X), we can choose Fi⁢C=C∗⁢(Xi), and this is a filtration.

Given (C,d) and a filtration, we can associate two objects:

  1. 1.

    Gr∗⁢C=⨁Fi⁢C/Fi−1⁢C, the associated graded complex. Then we may consider the homology of this complex, which is a graded object.

  2. 2.

    Given a subcomplex (Fi⁢C,d)↪(C,d), we have a map H⁡(Fi⁢C,d)→H⁡(C,d). The image of this is Fi⁢H⁢(C,d), so we get a filtration on the homology H⁡(C). So then we obtain a graded

Now we want to consider therelationship between the two. Note that the first one is bigger, because x∈Gri⁢C is a cycle if dx∈Fi−1⁢C, while x∈C is a cycle if dx=0.

Theorem 1.2.12 (Leray).

There exists a spectral sequence (E∗r,dr) such that

  • •

    E∗1≅H∗⁢(Gr∗⁢C,d0).

  • •

    The spectral sequence converges to H⁡(C), or more precisely, E∗∞≅Gr∗⁢H⁢(C).

Proof.

Consider the group

Notice that

so all the groups are as expected. For the differential, we define dr:E∗r→Er∗ and in fact dr:Epr→Ep−rr and thus has degree −r. Choose α∈Epr and choose a∈Fp⁢C∩d−1(Fp−r⁢C) a representative. Then dda=0 implies that da∈d−1(0)⊂d−1(Fp−2⁢r⁢C) and therefore da∈Fp−r⁢C∩d−1⁢(Fp−2⁢r⁢C). Therefore da defines an element in Ep−rr, so we set drα=[da]∈Ep−rr.

The things that need to be checked are that this is well-defined, (dr)2=0, and that H∗⁢(E∗r,dr)≅E∗r+1 canonically. All of this is painful homological algebra and is omitted. □

However, we will need something even more painful. Now we will considered a filter graded chain complex (C∗,d), where d has degree −1. Now for a filtration, we can consider the bigraded

Also, we note that Hm⁢(C) is naturally filtered with Fp⁢Hm⁢(C)=Im(Hm⁢(Fp⁢C)→Hm⁢(C)).

Theorem 1.2.13 (Leray).

There is a spectral sequence (E∗,∗r,dr) such that

  • •

    E∗,∗1≅H∗,∗⁢(Gr⁢C).

  • •

    The spectral sequence converges to H⁡(C), where

Proof.

We write

The differential has degree (−r,r−1). Note that it decreases the filtration by r and the total degree by 1. For α∈Ep,qr, choose a representative a∈Fp⁢Cp+q with da∈Fp−r⁢Cp+q−1. Then d2a=0 and therefore da∈d−1(0)⊆d−1(Fp−2⁢r⁢Cp+q−1), so we set drα=[da]∈Ep−r,q+r−1r. □

Returning to topology, consider a filtration ∅=X−1⊆X0⊆⋯⊆Xn=X, which givs a filtration on C∗⁢(X). This now defines a spectral sequence by setting Fp⁢Cp+q⁢(X)=Cp+q⁢(Xp). The E0 page of this is simply

and dp,q0 is ∂:Cp+1→Cp+q−1.

The E1-page of the spectral sequence is

and d1:Hp+1⁢(Xp,Xp−1)→Hp+q−1⁢(Xp−1,Xp−2) is the connecting homomorphism for the triple (Xp,Xp−1,Xp−2).

The E∞-page of the spectral sequence is

which is the associated graded of Hp+q⁢(X) for the filtration Xp↪X.

Example 1.2.14.

Consider the cellular homology of X. Set Xp=Xp, the p-skeleton. Then we note that Ep,q1 is the homology Hp+1⁢(Xp,Xp−1), and now if 𝒞p⁢(X) is the cellular complex, the E2-page is precisely the cellular homology of X and is the same as the E∞-page. This proves that cellular homology is the same as singular homology.

Remark 1.2.15.

This discussion works for infinite filtrations as long as X=⋃Xn and X has the weak topology induced by the filtration.

If you are interested in working through the pain1 of spectral sequences, the book A User’s Guide to Spectral Sequences is a good resource. In this course, we will not open this Pandora’s box.

Proof of Serre Spectral Sequence.

Consider F↪E→𝜋B and assume π1⁢(B)=1. For simplicity, assume that B is a CW complex and π is a fiber bundle. Now B has a filtration by skeleta

and therefore E has a filtration induced by pulling back π. Therefore there is a spectral sequence converging to H∗⁢(E) with

We want to compute the E2-page, so by excision we have

By contractibility of ei, so

Unfortunately, this identification is not canonical, which is why we need the assumption that B is simply-connected. In general, for a path γ, transport (homotopy lifting) gives us a map γ∗:H∗⁢(F0)≃H∗⁢(F1) depending only on the relative homotopy class. For example, if we consider S1→2S1, then the two paths between p0,p1 give different identifications. Therefore, we have a canonical identification

Finally, we note that dp,q1:Ep,q1→Ep−1,q1 is exactly the cellular boundary map ∂×1Hq⁢(F), so Ep,q2=Hp⁢(B;Hq⁢(F)), as desired. □

Remark 1.2.16.

The theorem holds provided the action of π1⁢(B) on H∗⁢(F) is trivial, which means that the fibration is homologically simple.

Example 1.2.17.

Consider a sphere bundle Sn↪E→B. This is homologically simple if and only if it is orientable.

Even more generally, we need to use homology with local coefficients. Here, we will take

and dp,q1 will have components given by composing δ with transport. An alternative interpretation is to consider the universal cover B~→B. Now π1⁢(B) acts on 𝒞∗⁢(B~) and on H∗⁢(F), so we obtain modules over the group ring ℤ⁡[π1⁢(B)]. Therefore we have

Now we will use groups on the edges of each page to compute H∗⁢(F)→H∗⁢(E).H∗⁢(E)→H∗⁢(B). Degree reasons tell us that Hn⁢(F)=E0,n2↠E0,n3↠⋯↠E0,n∞⊆Hn⁢(E).

Proposition 1.2.18.

E0,n∞⊆Hn⁢(E) is the image of Hn⁢(F)→Hn⁢(E).

Similarly, we can consider the groups H0⁢(B),…,Hn⁢(B) on the bottom of the E2 page. This tells us that Hn⁢(B)=En,02⊇⋯⊇En,0∞, which is a quotient of Hn⁢(E).

Proposition 1.2.19.

En,0∞⊆Hn⁢(B) is the image of Hn⁢(E)→Hn⁢(B).

Returning to algebra, suppose we have a chain map f∗:(C∗,d)→(C~∗,d~) that preserves filtration. This gives a morphism of spectral sequences f∗,∗r:(E∗,∗r,dr)→(E~∗,∗r,d~r) such that f1,f∞ are the maps on the associated graded objects and fr is a chain map such that the induced map on homology is fr+1.

Example 1.2.20.

An example here is filtered spaces with f:X→X~ such that f⁡(Xp)⊆X~p.

Note that the propositions follow by looking at the maps on spectral sequences induced by

Now consider the map dm:Em,0m→E0,m−1m. Note that E0,m−1m is a quotient of Hm−1⁢(F) and Em,0m is a subgroup of Hm⁢(B), so we have dm:Hm⁢(B)⇢Hm−1⁢(F), which is the analogue of πm⁢(B)→∂xπm−1⁢(F).

Definition 1.2.21.

We call this the transgression map.

Proposition 1.2.22.

The transgression is given by Hm⁢(B)=Hm⁢(B,pt)⇢Hm⁢(E,F)→Hm−1⁢(F).

Example 1.2.23.

Let Xn be a closed smooth oriented manifold. Then there is a fibration Sn−1↪S⁢X→X, where S⁢X is the unit sphere bundle. This is homologically simple, and then we have dn:Hn⁢(X)→Hn−1⁢(Sn−1), which is multiplication by χ⁡(X).

Remark 1.2.24.

We can prove this result if we know the relationship between χ and zeroes of vector fields.

Example 1.2.25.

Consider the fibration Ω⁢X→∗→X. If X is (n−1)-connected, then Ω⁢X is (n−2)-connected. Then Hn⁢(X),Hn−1⁢(Ω⁢X) are the first possibly nonzero homology groups. Now we consider the spectral sequence

012340123ℤHn−1⁢(Ω⁢X)Hn⁢(X)

and thus the map dr=τ:Hr⁢(X)→Hr−1⁢(Ω⁢X) is an isomorphism for r≤2⁢n−2.

Remark 1.2.26.

We can interpret this very explicitly as π:Σ⁢Ω⁢X→X, and this is the inverse of

1.3 Spectral Sequences in Cohomology

Theorem 1.3.1.

Let F↪E→B and π⁡(B)=1. Then there exists (Er∗,∗,dr)⇒H∗⁢(E) and E2p,q=Hp⁢(B,Hq⁢(F)) and dr has degree (r,1−r). Furthermore,

  1. 1.

    There is a multiplication Erp,q⊗Erp′,q′→Erp+p′,q+q′

  2. 2.

    dr:Er∗,∗→Er∗,∗ is a derivation, which means

  3. 3.

    The multiplication on Er+1∗,∗ is induced by the one on Er∗,∗.

  4. 4.

    The multiplication on E∞ is compatible with the one on H∗⁢(E). This means that E∞ is the associated graded of the cohomology.

Example 1.3.2.

Consider ℂ⁢ℙ∞↪F→S3. We want to compute H4⁢(F), which will compute π4⁢(S3). Now the E2-page of the spectral sequence for cohomology looks like

01230123456ℤx3ℤx2ℤxℤ1ℤyℤx⁢yℤx2⁢yℤx3⁢y

and now d3⁢(x)=y because H2⁢(F)=H3⁢(F)=0. But then

This tells us that the E4-page is

012301234ℤℤ2ℤ3

and therefore H1⁢(F)=⋯=H4⁢(F)=0 and H5⁢(F)=ℤ2. By the universal coefficients theorem, we see that H4⁢(F)=ℤ2.

Corollary 1.3.3.

For n≥3, we have πn+1⁢(Sn)=ℤ/2⁢ℤ and the group is generated by the suspension of the Hopf map.

Theorem 1.3.4.

H∗⁢(S⁢U⁢(n),ℤ)≅⋀ℤ[x3,…,x2⁢n−1] and |xi|=i.

Proof.

Let n=2. We know that S⁢U⁢(2)≅S3. Now for general n, consider the fiber bundle S⁢U⁢(n−1)↪S⁢U⁢(n)→S2⁢n−1. Now assume that H∗⁢(S⁢U⁢(n−1))=⋀ℤ[x3,…,x2⁢n−3]. Then the E2⁢n−1-page of the spectral sequence is

012345670123456H∗⁢(S⁢U⁢(n−1))H∗⁢(S⁢U⁢(n−1))

and therefore we see that d2⁢n−1⁢(x3)=⋯=d2⁢n−1⁢(x2⁢n−3)=0 by grading reasons. In particular, we see that d2⁢n−1⁢(x3⁢x5)=0, and in general we see that d2⁢n−1=0. Therefore we see that E∞=H∗⁢(S⁢U⁢(n−1))⊗S∗⁢(S2⁢n−1)=⋀ℤ[x3,…,x2⁢n−1].

To conclude, we know that E∞=Gr⁢H∗⁢(E), so for some xj∈E∞, we can choose a lift in H∗⁢(E). Because H∗ is torsion free and for degree reasons, the lifts satisfy the desired identities. □

Remark 1.3.5.

The ring E∞ does not determine H∗⁢(E). There are S2↪E→S2 and S2↪E′→S2 with H∗⁢(E)≅H∗⁢(E′). For example, we can consider ℙ1×ℙ1 and Bl1⁡ℙ2 (the intersection forms are different).

Example 1.3.6.

Let char⁡k=0. Then H∗⁢(S⁢O⁢(2⁢m+1),k)≅H∗⁢(S3×S7×⋯×S4⁢m−1) and H∗⁢(S⁢O⁢(2⁢m),k)≅H∗⁢(S3×⋯×S4⁢m−5×S2⁢m−1). However, we have

as groups, but not as rings.

To study this, we will consider the Stiefel manifold V⁡(n,k) of k-orthonormal frames in ℝn. Note that V(n.k)=SO(n)/SO(n−k). In the simplest case, V⁡(n,1)=Sn−1 and V⁡(n,2)=S⁡(Sn−1), the unit sphere bundle of Sn−1. Now we have Sn−1↪V⁡(n,2)→Sn−1, so we have a spectral sequence

0123012ℤℤℤℤ

Observe that a finite CW complex might not have even finitely generated homotopy groups. For example, we have π2⁢(S1∨S2)≅⨁i∈ℤℤ.

Theorem 1.3.7 (Serre).

Let X be a 1-connected CW complex such that Hq⁢(X,ℤ) is finitely generated for all q. Then πq⁢(X) is finitely generated for all q.

The key ingredients to the proof are:

  1. 1.

    Show that for K⁡(G,n) with G finitely generated (resp. finite), then H∗ is finitely generated (resp. finite).

  2. 2.

    Inductively, kill homotopy groups using fibrations. This generalizes F↪S3→K⁡(ℤ3).

Consider the Whitehead tower of X, which is a sequence ⋯→X3→X2→X1→X, where Xn is n-connected and Xn→X is an isomorphism in πi for i≥n+1. Also, we require Xn+1→Xn to be a fibration with fiber K⁡(πn+1⁢(X),n). For example, we had F=(S3)3 and S3=(S3)2. We have πn+1⁢(X)=πn+1⁢(Xn)=Hn+1⁢(Xn).

To construct the Whitehead tower, we work inductively. Suppose we have Xn. Then Kπn+1⁢(X),n+1 is obtained by attaching cells to Xn, so we have a map Xn→K⁡(πn+1⁢(X),n+1) that is an isomorphism on πn+1. Turning this into a fibration, we take Xn+1 to be the homotopy fiber.

Lemma 1.3.8.

Consider a fibration F↪E→B such that π1⁢(B)=1. Then if Hq⁢(F),Hp⁢(B) are finitely generated, then Hn⁢(E) is finitely generated for all n.

Proof.

Consider the Serre spectral sequence. Then Ep,q2 is finitely generated for all p,q, so because ℤ is Noetherian, then Ep,q∞ must also be finitely generated. This implies that Hn⁢(E) has a filtration by finitely generated objects, so it must be finitely generated. □

We will apply this to K⁡(πn+1⁢(X))↪Xn+1→Xn.

Proposition 1.3.9.

Hp⁢(K⁡(G,n)) is finitely generated for p>0 whenever G is finitely generated.

Proof.

Write G=⨁ℤr⊕⨁ℤ/pi⁢ℤ. Now up to homotopy, we have

By Künneth, we may assume that G is cyclic, so consider the fibration K⁡(G,n−1)↪∗→K⁡(G,n). Note that K⁡(ℤ,1)=S1 and K⁡(ℤm,1) is an infinite lens space, so only the inductive step remains.

Consider Ep,q2=Hp⁢(K⁡(G,n),Hq⁢(K⁡(G,n−1))) and assume HM⁢(K⁡(G,n)) is not finitely generated but Hi⁢(K⁡(G,n)) is finitely generated for i<M. But then EM,03 is not finitely generated (it is the kernel of a map to something that is finitely generated), so EM,0∞ is not finitely generated, a contradiction. □

Remark 1.3.10.

This holds more generally for classes for classes of groups called Serre classes, for example finite abelian p-groups.

1.4 Homotopy groups of spheres

We know that πm⁢(Sn) is a finitely generated abelian group.

Theorem 1.4.1 (Serre).

The rank of πq⁢(Sn) is 1 is q=n or q=2⁢n−1 for n even and 0 otherwise.

The key computation is the rational homology H∗⁢(K⁡(ℤ,n);ℚ). This is given by

This is true for K⁡(ℤ,1)=S1,K⁡(ℤ,2)=ℂ⁢ℙ∞. We will do the case when n is even, and consider K⁡(ℤ,n−1)↪∗→K⁡(ℤ,n). Then the Serre spectral sequence in cohomology is

012340123y1x

and then dn⁢(x⁢y)=dn⁢(x)⋅y+x⁢dn⁢(y)=x2.

Lemma 1.4.2.

Let X be 1-connected and suppose Hq⁢(X,ℤ) is finitely generated for all q. Also, suppose that H∗⁢(X,ℚ)≅H∗⁢(Sm,ℚ) for m odd. Then

Proof.

Note that Hm⁢(X,ℚ)=ℚ, so [X,K⁡(ℤ,m)]=Hm⁢(X,ℤ)=ℤ⊕torsion. Then there exists f:X→K⁡(ℤ,m) such that f∗⁢(ιm)=1, where ιm generates Hm⁢(K⁡(ℤ,m),ℤ)=ℤ. This tells us that f∗:H∗⁢(K⁡(ℤ,m),ℤ)≃H∗⁢(X,ℚ) is an isomorphism. Now if F is the homotopy fiber of f, then Hq⁢(F,ℚ)=0 for q>0, so Hq⁢(F,ℤ) for all q>0 (by finite generation), and therefore πq⁢(F) is finite for all q. Now f∗:πq⁢(X)→πq⁢(K⁡(ℤ,m)) fits in the long exact sequence with πi⁢(F), so f∗ has finite kernel and cokernel, and thus rank⁡πq⁢(X)=rank⁡πq⁢(K⁡(ℤ,m)). □

Proof of Serre.

We may assume that n is even. Consider K⁡(ℤ,n−1)↪(Sn)n→Sn. This is the nth stage of the Whitehead tower. Consider the Serre spectral sequence

012340123ℚℚℚ

and note that dn is an isomorphism because Hn−1⁢((Sn)n)=0. But then we see that H∗⁢((Sn)n)=H∗⁢(S2⁢n−1,ℚ), so the desired result follows. □

An “easy” generalization is

Theorem 1.4.3 (Cartan-Serre).

Let X be 1-connected with finitely generated homology groups such that

Then rank⁡πq⁢(X) is the number of generators xi,yj in degree q.

Example 1.4.4.

Consider X=S⁢U⁢(n),S⁢O⁢(n).

Remark 1.4.5.

In general, H∗⁢(X,ℚ)≅H∗⁢(X′,ℚ) does not imply that rank⁡πq⁢(X)=rank⁡πq⁢(X′).

Now the point of rational homotopy theory is to compute rank⁡πq⁢(X) using H∗⁢(X,ℚ) and extra information (for example Massey products).

We have computed the ranks of the homotopy groups of spheres, so now we will consider the torsion. The tools we have developed are enough to show that

Theorem 1.4.6.

Let n≥3. For any prime p, the group πi⁢(Sn) has no p-torsion for i<n+2⁢p−3 and the p-primary part of πn+2⁢n−3⁢(Sn) is ℤ/p⁢ℤ.

Corollary 1.4.7.

In the stable range, πn+2⁢(Sn) is a 2-group and πn+3⁢(Sn) is the direct sum of a 2-group and ℤ/3⁢ℤ.

The idea of the proof is induction on the Whitehead tower with the fibration K⁡(πn−1⁢(X),n)↪Xn+1→Xn and Hn+2⁢(Xn+1)=πn+2⁢(X). We can study this group one prime at a time because H∗⁢(K⁡(ℤ/p,1),ℤ/p′) vanishes if p≠p′ and is very interesting if p=p′. Then there is a large gap in cohomology depending on p. If we study the ℤ-cohomology of K⁡(ℤ/p,2), then we may consider the fibration K⁡(ℤ/p,1)↪∗→K⁡(ℤ/p,2). Then we know that H∗⁢(K⁡(ℤ/p,1),ℤ)=ℤ⁡[x]/(p⁢x) where |x|=2, and so the Serre spectral sequence of the fibration is

0123012341xx2y3x⁢y3

and then d⁡(xm)=m⁢xm−1⁢y3 and thus d⁡(xp)=0 and d⁡(xi)≠0 for i<p. For more detail, see Fuchs-Fomenko.

Now we will compute the 2-primary part of these homotopy groups. Consider the fibration K⁡(ℤ,n−1)↪(Sn)n→Sn and the next step k⁡(ℤ/2,n)↪(Sn)n+1→(Sn)n. We want to compute the 2-primary part of Hn+2⁢((Sn)n+1). This means we need to understand H∗⁢(K⁡(ℤ/2,n);ℤ/2). We can understand it as cohomology operations

1.5 Cohomolgy Operations

Definition 1.5.1.

A cohomology operation φ between Hn⁢(−,G)→Hn⁢(−,K) is a natural transformation between the two functors viewed as 𝖢𝖶→𝖲𝖾𝗍.

Example 1.5.2.

Let R be a ring. Then the cup product Hn⁢(−,R)→H2⁢n⁢(−,R) given by α↦α∪α is a cohomology operation. Note that this is not a homomorphism in general.

Now Hn⁢(−,G)=[−,K⁡(G,n)] and so natural transformations are just homotopy classes of maps K⁡(G,n)→K⁡(K,m), which form the group Hm⁢(K⁡(G,n),K).

Example 1.5.3 (Bockstein homomorphism).

Consider a short exact sequence 0→A→B→C→0. Then we get a short exact sequence

of chain complexes, and this induces a long exact sequence in cohomology. Then the map

is a cohomology operation, called the Bockstein homomorphism.

An interesting case of this is the exact sequence 0→ℤ/p→ℤ/p2→ℤ/p→0. Therefore we obtain an interesting map

so there is an interesting element in Hn+1⁢(K⁡(ℤ/p,n),ℤ/p). This is actually useful in the homotopy classification of 3-dimensional lens spaces L⁡(p,q). In fact, we can show that L⁡(p,q)≃L⁡(p,q′) if and only if q′≡±k2⁢q(modp) for some k. For example, L⁡(5,1)≄L⁡(5,2).

To show this, consider the map

This is well-defined up to a choice of sign. There is a standard generator α∈H1⁢(L⁡(p,q)), which is the image of the arc (1,0)→(e2⁢π⁢i/p,0). Then Q⁡(α∗)=q, where α∗ is the Poincaré dual of α. Then if we have a homotopy equivalence L⁡(p,q)→𝑓L⁡(p,q′), we see that αp,q↦k⁢αp,q′ and then Q⁡(αp,q)=±k2⁢Q⁢(αp,q′) by naturality.

Example 1.5.4 (Steenrod square).

For all n, we will construct a cohomology operation Sqi:Hn⁢(−,ℤ/2)→Hn+i⁢(−,ℤ/2). Some properties are

  • •

    Steenrod squares are additive.

  • •

    We have

  • •

    Sqi:Hn⁢(−,ℤ/2)→Hn+1⁢(−,ℤ/2) is the Bockstein homomorphism for

  • •

    Sqi⁡(α∪β)=∑j+k=iSqj⁡(α)∪Sqk⁡(β) (the Cartan relation).

We will take these to be axioms for the Steenrod squares.

Theorem 1.5.5.

There exist unique cohomology operations with these properties.

Remark 1.5.6.

Define the total squaring operation by

Here, the Cartan relation says that Sq is multiplicative.

Example 1.5.7.

Let X=ℝ⁢ℙ∞. Then H∗⁢(X,ℤ/2)=ℤ/2⁢[α]. Then

This implies that Sq⁡(αk)=Sq⁡(α)2=αk⁢(1+α)k. In particular, we see that Sqi⁡(αk)=(ki)⁢αk+i.

Proposition 1.5.8.

The Steenrod squares Sqi commute with the suspension isomorphisms Σ:Hn⁢(X,ℤ/2)→Hn+1⁢(Σ⁢X,ℤ/2).

This is interesting because Σ⁢X has trivial cup products.

Example 1.5.9.

Consider f:S15→S8 with H⁡(f)=1. Then Σn⁢f:S15+n→S8+n is nontrivial. Consider the mappinc cone Cf=S8∪fD18 where α2=β, where α is the 8-cell and β is the 16-cell. Then we see that CΣ⁢f=Σ⁢Cf, so by dimension reasons, all cup products are trivial. However, Sq8⁡(Σ⁢α)=Σ⁢Sq8⁡(α)=Σ⁢β. Otherwise, if Σ⁢f is trivial, then CΣ⁢f=S9∪S17.

Proof of Proposition 1.5.8.

Note that Σ is obtained by

This implies that

Now we will consider the following question. How many pointwise linearly independent vector fields v1,…,vk−1 can there be on Sn−1. When are there n−1?

Some basic obervations are the following:

  • •

    If n−1 is even, then χ⁡(Sn−1)=2, so any vector field has a zero.

  • •

    By Gram-Schmidt, we can assume that v1,…,vk−1 is orthonormal at each point.

Theorem 1.5.10 (Steenrod-Whitehead).

If n=2r⁢(2⁢s+1), then Sn−1 has at most 2r−1 linearly independent vector fields.

Example 1.5.11.

If n−1 is even, then r=0, so there are no linearly independent vector fields.

Corollary 1.5.12.

If Sn−1 is parallelizable, then n is a power of 2.

Remark 1.5.13.

Adams gave a complete solution. We will see using K-theory that if Sn−1 is parallelizable, then n=1,2,4,8.

Sketch of Proof.

Let Vn,k be the Stiefel manifold of orthonormal k-frames in ℝn. Then we have the natural map

Now if v1⁢(x),…,vk−1⁢(x) are orthonormal vector fields on Sn−1, then we get a section f of p given by

Now the question is reduced to that of the existence of a section of p. We will discuss obstructions using Sqi. Recall that Vn,k=S⁢O⁢(n)/S⁢O⁢(n−k). Now consider

This induces a map ℝ⁢ℙn−1/ℝ⁢ℙn−k−1↪S⁢O⁢(n)/S⁢O⁢(n−k)=Vn,k. Now if 2⁢k−1≤n, there is a cell decomposition of Vn,k for which ℝ⁢ℙn−1/ℝ⁢ℙn−k−1 is the (n−1)-skeleton. Now suppose there is a section f of p:Vn,k→Sn−1. Then f∗⁢p∗:Hn−1⁢(Sn−1)≃Hn−1⁢(Sn−1) is an isomorphism. After homotopy, we can make f a cellular map, so f~:Sn−1→(Vn,k)n−1=ℝ⁢ℙn−1/ℝ⁢ℙn−k−1. Therefore f∗ factors through H∗⁢(ℝ⁢ℙn−1/ℝ⁢ℙn−k−1) and induces an isomorphism in degree n−1 (using ℤ/2⁢ℤ-coefficients). In degree n−k, f∗ induces a map ℤ/2→0. But now we have a commutative diagram

and now we see that Sqk−1 cannot be an isomorphism. Computing Sqk−1 using naturality, we see that k=2r+1 works. □

1.5.1 Construction of Steenrod Squares

Here, we are following Chapter 4.L in Hatcher. The idea is that the cup product is commutative on H∗⁢(−,ℤ/2). However, this is not commutative on the chain level, so Sqi will measure this failure. We work at the level of spaces.

Let X be a pointed space and consider X∧X with the swap map T:X∧X→X∧X. Now we consider the homotopy quotient of T. For example, we know that ℤ/2⁢ℤ acts freely on S∞ with quotient ℝ⁢ℙ∞. Then we define Γ∞⁢X=S∞×(X∧X)/(ℤ/2⁢ℤ). This has a projection to ℝ⁢ℙ∞ with fiber X∧X. Then the inclusion s:ℝ⁢ℙ∞→Γ∞⁢X of ℝ⁢ℙ∞ at the basepoint of X∧X is a section, so set ⋀∞X=Γ∞⁢X/s⁡(ℝ⁢ℙ∞) (doing this kills many unnecessary cells in low dimension).

Now given α∈Hn⁢(X,ℤ2), we can associate λ⁡(α)∈H2⁢n⁢(⋀∞X;ℤ) such that

Then S∞×X↪S∞×(X∧X) induces a map

But now given such λ⁡(α), this maps to a sum of the form

Remark 1.5.14.

The assignments ⋀∞(−),Γ∞⁢(−) are functorial.

Now because Hn⁢(−,ℤ2)=[−,K⁡(ℤ2,n)], we only need to construct λ⁡(ι) for the nontrivial

Here, we will set Kn≔K⁡(ℤ2,n), and therefore we need to construct the map ⋀∞Kn→K2⁢n step by step. Consider the map

Then if we consider Kn∧Kn→𝑇Kn∧Kn→ι⊗ιK2⁢n, this must be homotopic to ι⊗ι by some homotopy ht. Now we can use ht to define

Because this is basepoint preserving, it descends to ⋀1Kn→K2⁢n. But now ⋀∞Kn is obtained from ⋀1Kn by attaching cells of dimension stricly larger than 2⁢n+1, so we can extend and obtain a map λ⁡(ι):⋀∞Kn→K2⁢n.

Remark 1.5.15.

Note that λ⁡(ι) is uniquely determined because the map

is injective.

There is an alternative easier way. Consider α∈Hn⁢(X,ℤ2) and α×α∈H2⁢n⁢(X×X,ℤ2). This gives us a classifying map X×X→𝑓K2⁢n, and again let T be the swap map on X×X. Now we see that f∼f∘T with homotopy ft, so now we have a map S1×X×X→K2⁢n such that (s,x1,x2),(−s,x2,x1) map to the same point. This map extends to D2 and therefore to S2. This now gives us a map S∞×X×X→K2⁢n which is ℤ2-invariant, and precomposing with S∞×X→S∞×X×X gives us a map ℝ⁢ℙ∞×X→K2⁢n.

Theorem 1.5.16 (Cartan).

For any α,β, we have λ⁡(α∪β)=λ⁡(α)∪λ⁡(β).

1.5.2 Stable Cohomology Operations

Definition 1.5.17.

A stable cohomology operation (SCO) of degree r is a map

for all n that commutes with suspension maps, i.e. the diagram

Example 1.5.18.

The Steenrod squares and compositions of Steenrod squares are all stable cohomology operations.

Lemma 1.5.19.

Let φ be a stable cohomology operation. Then the diagram

commutes.

Proof.

The left and right squares are clear because they come from maps of spaces. At δ∗, we consider

Because A⊂X is a cofibration, then X∪C⁢A≅X/A. Then we may collapse the X to obtain Σ⁢A, and then we see that δ∗ is given by Hn⁢(A)→ΣHn+1⁢(Σ⁢A)→p∗Hn+1⁢(X/A), and both arrows in the composition commute with φ. □

This gives us the following slogan:

Stable cohomology operations commute with transgression.

Proposition 1.5.20.

Let φ be a stable cohomology operation of degree e. If α∈Hm⁢(F) is transgressive, then so is φ⁡(α)∈Hm+r (which means that d2φ⁡(α)=⋯=dm+rφ⁡(α)=0). If τ(α)∈Em+1m+1,0=Hm+1(B)/∼ is represented by β∈Hn+1⁢(B), then τ(φ(α))∈Hm+r+1(B)/∼.

Proof.

The transgression map is given by

Then φ commutes with δ∗ and p∗. □

Remark 1.5.21.

If G=H, then we can compose stable cohomology operations. This is a noncommutative algebra!

Theorem 1.5.22.

The Steenrod algebra

is generated as an algebra by {Sqi}i∈ℕ.

This is a very complicated algebra. For example, if a<2⁢b, we have the Adem relations

For a proof of this, see Hatcher.

Example 1.5.23.

For any b, we have Sq1⁡Sqb=(b−1)⁢Sqb+1. For example, Sq1⁡Sq2=Sq3. We say that Sq3 is decomposable. This means it can be written in terms of elements of lower degree.

Example 1.5.24.

The elements of the form Sq2k are indecomposable. Consider the action on H∗⁢(ℝ⁢ℙ∞). Then we see that

but Sqi⁡(α2k)=0 for 0<i<2k. The idea is that (2ki)≡0mod2 for all i.

We want to interpret stable cohomology operations using K⁡(G,n). Then if φ={φn}, we see that

But then stability tells us that Hn⁢(Σ⁢K⁢(ℤ2,n−1),ℤ2)=ℤ2, and if we consider the nontrivial ι:Σ⁢K⁢(ℤ2,n−1)→K⁡(ℤ/2,n), then we obtain a map

Then stability is equivalent to the fact that φn↦φn−1. Alternatively, we may use the suspension-loop adjunction to see that

Proposition 1.5.25.

For any r, we have

We will now compute H∗⁢(K⁡(ℤ2,n),ℤ2). We will use the fibration K⁡(ℤ2,n+1)↪∗→K⁡(ℤ2,n) and hope to proceed by induction.

Example 1.5.26.

Let n=2. Then we know K⁡(ℤ2,1)=ℝ⁢ℙ∞. Then the spectral sequence is given by

0123012341αα2α3α2eSq1⁡e

Then we see that d2(α2)=2⁢α⁢d2(α)=0. But then d3(α2)≠0, so we see that d3(Sq1⁡α)=Sq1⁡d2(α)=Sq1⁡e. Inductively, we see that d5(α4)=Sq2⁡Sq1⁡e and d2(α⁢e)=d2(α)⋅e=e2. In the end, we will see that

Definition 1.5.27.

A set I=(i1,…,ik) (possibly empty) is admissible if i1≥2⁢i2,i2≥2⁢i3,…. Then we will write SqI=Sqi1∘Sqi2∘⋯∘Sqik.

Note that if I is not admissible, then SqI can be simplified using Adem relations.

Definition 1.5.28.

Define the excess of an admissible I to be

For example, if exc⁡(I)=1, then I=(2k,2k−1,…,4,2,1). If exc⁡(I)=0, then I=∅.

Theorem 1.5.29 (Serre).

We have

Example 1.5.30.

We have H∗⁢(K⁡(ℤ2,1))=ℤ2⁢[e1]. When n=3, the generators are

When n increases, the description becomes more and more complicated.

Theorem 1.5.31 (Borel).

The set {τ⁡((SqI⁡en−1)2k)∣exc⁡(I)<n−1} are generators of H∗⁢(K⁡(ℤ2,n)). Now τ⁡(en−1)=en.

Example 1.5.32.

When n=2, we have (Sq0⁡e2)2=e22=Sq2⁡e2 and has excess 2. Similarly, we see that (Sq0⁡e2)4=Sq4⁡Sq2⁡e2 and has excess 2. We can also see that (Sq2⁡Sq1⁡e2)2=Sq5⁡Sq2⁡Sq1⁡(e2), and this also has excess 2.

Theorem 1.5.33 (Borel).

Let F↪E→B be a fibration with π1⁢(B)=1 and H~∗⁢(E)=0. Assume that

  1. 1.

    H∗⁢(F,ℤ2) is generated as an algebra by elements ai∈Hmi,m1≤m2≤⋯ which are transgressive.

  2. 2.

    The set {ai1⁢⋯⁢aik∣i1<i2<⋯<ik} is an additive basis for H∗⁢(F,ℤ2).

Then H∗⁢(B,ℤ2) is a polynomial algebra generated by {τ⁡(ai)}.

Proof.

Let Erp,q be the Serre spectral sequence. We will construct an algebraic model E~rp,q for it. Set

where |bi|=|ai|+1. Then define d~r to satisfy

To do this, we set

This tells us that d~r∘d~s=0 for all r,s and thus all differentials are well-defined on E~r. Of course, we observe that E∞∗,∗=ℤ2. Now there is a natural map Φ:E~2p,q→E2p,q given by

This induces a map Φr of spectral sequences, so Φ∞ is an isomorphism. But now if ℤ2⁢[b1,…]→H∗⁢(B,ℤ2) is not surjective, then Er∗,∗ cannot kill anything not in the image. On the other hand, if it is not injective, then E~r∗,∗ cannot kill anything in the kernel. Therefore the map is an isomorphism. □

Now we will compute H∗⁢(K⁡(ℤ,n),ℤ2). We know that K⁡(ℤ,2)↪∗→K⁡(ℤ,3) is a fibration, so consider the Serre spectral sequence

012345601234561xx2x3ex⁢ee2

Then we know that x2=Sq2⁡x,x4=Sq4⁡x2, so we find generators Sq2⁡(e),Sq4⁡Sq2⁡e,…. In addition, we note that d3(x⁢e)=e2. Also note that Sq1⁡(e) is not there!

Theorem 1.5.34 (Serre).

We have the identity

1.6 Computation of Homotopy Groups

We will compute πn+2⁢(Sn) for n large (the stable homotopy group). If we consider the fibration K⁡(ℤ,n−1)↪(Sn)n→Sn and K⁡(ℤ2,n)↪(Sn)n+1→(Sn)n, we want to compute Hn+2⁢((Sn)n+1;ℤ)=πn+2⁢(Sn). First, we will consider cohomology with ℤ2 coefficients. The Serre spectral sequence is given by

01234501234567891seSq2⁡eSq3⁡eSq4⁡eSq5⁡ee⁢s

If we write Sq2⁡e≕u,Sq3⁡e≕v, we can compute Hq⁢((Sn)n;ℤ2) to be

Table 1.1: Hq⁢((Sn)n;ℤ2)
q n+1 n+2 n+3 n+4
generator u Sq1⁡u b Sq1⁡v

This now gives us the Serre spectral sequence for the next fibration

0123456789012345678uSq1⁡uvSq1⁡veSq1⁡eSq2⁡eSq3⁡e

But this implies that Hn+2⁢((Sn)n+1;ℤ2)=ℤ2=⟨w⟩ and Sq1⁡w≠0. By the universal coefficients theorem, this implies that Hn+2⁢((Sn)n+1;ℤ)≅ℤ/2k⁢ℤ. To obtain k=1, we will consider the Bockstein exact sequence

This implies that Hn+2⁢(−,ℤ4)=ℤ2, so Hn+2⁢(−,ℤ)=ℤ2 by the universal coefficients theorem.

In the next case, we want to compute πn+3⁢(Sn). We can consider K⁡(ℤ2,n+1)↪(Sn)n+2→(Sn)n+1 and we want Hn+3⁢((Sn)n+2;ℤ)=πn+3⁢(Sn). In the Serre spectral sequence, we need to compute Sq1⁡w,Sq2⁡w, where w has degree n+2. But this will tell us that Hn+3⁢((Sn)n+2;ℤ2)=ℤ2, so by the universal coefficients theorem, we have

To compute the order, we will have to redo all computations with coefficients in ℤ modulo odd torsion. We should be able to obtain

Theorem 1.6.1 (Rokhlin).

We have the identity πn+3⁢(Sn)≅ℤ/24⁢ℤ.

For example, we can compute H∗⁢(K⁡(ℤ2,2);ℤ2). We have

Table 1.2: H∗⁢(K⁡(ℤ2,2);ℤ2)
0 1 2 3 4 5
1 ⋅ e Sq1⁡e e2 e⋅Sq1⁡eSq2⁡Sq1⁡e

By Cartan, we can compute Sq1⁡(e2)=e∪Sq1⁡e+Sq1⁡e∪e=0. By Bockstein, we see that H4⁢(K⁡(ℤ2,2);ℤ)=ℤ/4⁢ℤ⊕odd stuff.

Remark 1.6.2.

We used Whitehead towers, but there is a dual approach using Postnikov towers. The Postnikov tower looks like

and satisfies the properties

  • •

    πi⁢(Zn)=0 for i≥n+1;

  • •

    X→Zn is an isomorphism in πi for i≥n;

  • •

    We have a fibration K⁡(πn⁢(X),n)↪Zn→Zn−1.

Hatcher uses Postnikov towers to compute πn+3⁢(Sn), and if you are particularly masochistic, Mosher-Tanpora compute the stable homotopy groups up to πn+7⁢(Sn). Unfortunately, this approach will get stuck eventually, so we cannot compute all of the stable homotopy groups of spheres.

2 Characteristic Classes and K-theory

2.1 Obstruction Theory

Here is a very basic question: Can we extend maps?

Question 2.1.1 (Extension Problem).

Consider f:A→Y and A⊆X. Does there exist a map X→Y extending f?

Question 2.1.2 (Homotopy problem).

Let f0,f1:X→Y such that f0,f1 are homotopic on A. Can we extend this homotopy to a homotopy on Y?

Note that this is a special case of an extension problem

Remark 2.1.3.

This is different from the homotopy extension property, which is given by X×{0}∪A×I⊂X×I and is the definition of a cofibration.

Question 2.1.4 (Lifting problem).

Let p:E→B be a fiber bundle. If we have f:X→B can we lift f to X→E?

For the extension problem, we will assume that the action of π1⁢(Y) on πn⁢(Y) is trivial. A special case of this is if we have f:Xn→Y, can we extend it to Xn+1? If we attach a single cell Xn∪φen+1, then we may consider the map f∘φk:∂en+1→Y. This extends to en+1=Dn+1 if and only if [f∘φe]=0 in πn⁢(Y). This gives us a map en+1→cf⁢(en+1)∈πn⁢(Y), so we obtain the obstruction cocycle cf∈𝒞n+1⁢(X,πn⁢(Y)).

Lemma 2.1.5.

f:Xn→Y extends to Xn+1 if and only if cf=0.

Lemma 2.1.6.

We have cf∈𝒞n+1⁢(X;πn⁢(Y)) is a cocycle, which means δ⁢cf=0.

Proof.

Assume n>0. By taking the universal cover, assume π1⁢(X)=1. We need to show that cf∘∂=0, where ∂:Cn+2→Cn+1. But this is given by the commutative diagram

Definition 2.1.7.

Let f:Xn→Y. Then define the obstruction class 𝒪f≔[cf]∈Hn+1⁢(X,πn⁢(Y)).

Theorem 2.1.8 (Fundamental theorem of obstruction theory).

Let f:Xn→Y. Then 𝒪f=0 if and only if f|Xn−1 can be extended to Xn+1.

Here, we allow ourselves to change f on Xn∖Xn−1. Now suppose f,g:Xn→Y agree on Xn−1. Then given an n-cell en⊂Xn, we can consider Sn=Dn∪Dn, and performing f on the top and g on the bottom gives an element in πn⁢(Y). This gives us df,g∈Cn⁢(X,πn⁢(Y)), the difference cochain.

Remark 2.1.9.

Observe that df,g=0 if and only if f is homotopic to g with a homotopy fixing Xn−1.

Lemma 2.1.10.

Given f:Xn→Y and d∈𝒞n⁢(C,πn⁢(Y)), there exists g:Xn→Y such that g|Xn−1=f|Xn−1 and df,g=d.

Proof.

In the center of the disk, simply attach a new disk with the desired d⁡(e)∈πn⁢(Y). □

Lemma 2.1.11.

Consider df,g∈𝒞n⁢(X,πn⁢(Y)) and the obstruction classes cf,cg∈𝒞n+1⁢(X,πn⁢(Y)). Then

Proof.

For simplicity, suppose f,g only differ on σn. Given en+1, we want to show that

Up to homotopy, we may assume that in Figure 2.1, the shaded part is φ−1⁢(σn) and φ is a homeomorphism on the shaded component. In particular, f∘φ,g∘φ only differ on the shaded part, so the number of components of the shaded part is [e:σ].

∂en+1≅SnXn⊃σσφ
Figure 2.1: Attaching disks

□

Proof of fundamental theorem of obstruction theory.

Let 𝒪f=[cf]=0. Then cf=δ⁢d for some d∈𝒞n⁢(X,πn⁢(Y)). But then there exists g:Xn→Y such that g|Xn−1=f|Xn−1 with df,g=−d. But then

so g extends to Xn+1.

In the other direction, assume g:Xn+1→Y with g|Xn−1=f|Xn−1. Therefore cg=0, and

so [cf]=0. □

Now we will consider the relative case. Suppose (X,A) is a CW-pair and we have f:Xn∪A→Y. Can we extend this to Xn+1∪A? Here, we obtain an obstruction class 𝒪f∈Hn+1⁢(X,A;πn⁢(Y)). We have a relative version of the fundamental theorem in this case.

Returning to the homotopy problem, consider f,g:X→Y such that f|Xn−1∼g|Xn−1. Can we extend this homotopy to Xn? We are considering the extension problem for

We are considering (n+1)-cells of the form en×I, where en is an n-cell in X. Assume for simplicity that f|Xn−1=g|Xn−1. Then the obstruction is exactly df,g∈𝒞n⁢(X,πn⁢(Y)). Also, if cf=cg=0, then δ⁢df,g=cg−cf=0.

Theorem 2.1.12.

Let f,g:X→Y agree on Xn−1. Then Hn⁢(X,πn⁢(Y))∋[df,g]=0 if and only if f,g are homotopic on Xn, with the homotopy fixing Xn−2.

Example 2.1.13 (Cohomology of K⁡(G,n)).

Consider [X,K⁡(G,n)]→Hn⁢(X,G) by f↦[df,const]. Here, we can homotope f such that f|Xn−1 is constant because πi⁢(K⁡(G,n))=0 for i<n. Surjectivity is obvious from the previous discussion, so we will prove injectivity.

If [df,const]=0∈Hn⁢(X,G), then f|Xn is nullhomotopic. Because πi⁢(K⁡(G,n))=0 for i>n, f is nullhomotopic.

2.2 Primary obstruction class

We will change our notation. Denote the fiber bundle by F↪E→B. In this new notation, we want to extend a section s:Bn→E to Bn+1. If en+1 is a cell, then E|en+1≅en+1×F. We have s:∂en+1→F and we want to extend this to en+1. Now we have the map en+1↦𝒞s⁢(en+1)∈πn⁢(F). We know that cs∈𝒞n+1⁢(B,πn⁢(F)). Now we know s extends to Bn+1 if and only if cs=0, δ⁢cs=0, and the obstruction class 𝒪s∈Hn+1⁢(B,πn⁢(F)) is zero if and only if s|Bn−1 can be extended to Bn+1.

Now we assume that F is (n−1)-connected. Therefore there exists a section s:Bn→E with obstruction class 𝒪s∈Hn+1⁢(B,πn⁢(F)).

Theorem 2.2.1.

Given s,s′:Bn→E, we have 𝒪s=𝒪s′. In particular, this is an invariant of the bundle!

Definition 2.2.2.

Consider a fiber bundle ξ:F↪E→B with F being (n−1)-connected. Then the primary obstruction class is defined by be 𝒪⁡(ξ)∈Hn+1⁢(B,πn⁢(F)).

Theorem 2.2.3.

The fiber bundle ξ always admits a section over Bn and it admits a section over Bn+1 if and only if 𝒪⁡(ξ)=0.

Proof of Theorem 2.2.1.

First we note that a homotopy of s does not change Cs∈𝒞n+1⁢(B,πn⁢(F)). The homotopy extension property holds for sections where (X,A) is a CW pair with s:X→E and st:A×I→E a homotopy of sections. Here, we can extend to s~t:X×I→E.

Now suppose s,s′:Bn→E. Of course, s∼s′ on B0. We now show that if s∼s′ on Bk, then s∼s′ on Bk+1 for 0≤k<n−1. To see this, the homotopy extension property tells us that we may assume s=s′ on Bk after extending the homotopy. Then, we have ds,s′∈𝒞k+1⁢(B,πk+1⁢(F)). But we see that ds,s′=0, so s∼s′ on Bk+1. We now obtain s,s′:Bn→E such that s∼s′ on Bn−1. We may assume that s=s′ on Bn−1. Then we note that ds⁢s′∈𝒞n⁢(B,πn⁢(F)), so cs′−cs=δ⁢ds,s′ and thus 𝒪s=𝒪s′. □

The key property is that if we have f:B′→B and consider the pullback bundle f∗⁢ξ, then 𝒪⁡(f∗⁢ξ)=f∗⁢𝒪⁢(ξ). This tells us that 𝒪⁡(ξ) is a characteristic class, or in other words, a natural cohomological invariant. Consider the basic example: Suppose ξ:Sn−1↪E→B is an oriented bundle. Then Sn−1 is (n−2)-connected, so we have an obstruction class 𝒪⁡(ξ)∈Hn⁢(B,ℤ).

Definition 2.2.4.

This obstruction class is called the Euler class e⁡(ξ) of ξ. Then e⁡(ξ)=0 if and only if ξ has a section over Bn.

Here we have a special case. If η:ℝn↪E→B is an oriented vector bundle and ξ is the associated sphere bundle Sn−1→S⁡(E)→B. Then e⁡(E) is the Euler class of the associated sphere bundle. Therefore e⁡(E)=0 if and only if there exists a nonvanishing section of η over Bn.

Theorem 2.2.5.

Let Mn be a smooth oriented compact manifold. Then e⁡(TM)∈Hn⁢(M,ℤ)≅ℤ is χ⁡(M).

Corollary 2.2.6 (Poincare-Hopf).

M admits a nowhere vanishing vector field if and only if χ⁡(M)=0.

Proof.

Note that Mn−1≅M∖{finite number of points} (one for each top-cell). To compute 𝒪⁡(ξ) we can use a vector field v with only isolated zeroes. Now we may define the local degree at x to be the degree of the map Sε→Sn−1 defined by y↦v⁡(y). The local degree behaves like in Figure 2.2

deg=1deg=−1
Figure 2.2: Example of local degrees

and therefore ℤ∋𝒪⁡(ξ)=∑local degrees by definition. Now we simply need to construct an explicit nice vector field. Because M is smooth, it admits a smooth triangulation.1 The vector field is constructed by placing a zero at the center of every face and then for the center of a given positive dimensional face, the vector field points outward towards the vertices of that face. But then we have

Remark 2.2.7.

Let Ek→Mn be a smooth vector bundle with M compact. Then e⁡(E) can be interpreted as follows. Choose a generic section s:M→E. Then s−1⁢(0) is a smooth manifold of dimension n−k, and so we can write e⁡(E)=P⁢D⁢(s−1⁢(0)).

We will now relate the primary obstruction class and transgression. We will consider the Serre spectral sequence with coefficients in πn⁢(F)=Hn⁢(F). We know we have the transgression map τ:Hn⁢(F,πn⁢(F))→Hn+1⁢(B,πn⁢(F)) and the fundamental class ι∈Hn⁢(F,πn⁢(F)). This is defined by for all x∈πn⁢(F),⟨ι,h⁡(x)⟩=x.

Theorem 2.2.8.

The primary obstruction class is given by τ⁡(ι)=𝒪⁡(ξ)∈Hn+1⁢(B,πn⁢(F)).

Proof of this is given in a series of exercises in Fuchs-Fomenko 23.5.

Now we will consider the Thom isomorphism. If E→B is a rank n vector bundle. Then we may associate the disk bundle D⁡(E) and the sphere bundle S⁡(E). Now we define the Thom space to be T⁡(E)=D⁡(E)/S⁡(E) (equivalently, the one-point compactification of E).

Theorem 2.2.9 (Thom isomorphism).

Suppose E is an oriented vector bundle. Then there exists a unique U∈Hn⁢(T⁡(E);ℤ)=Hn⁢(D⁡(E),S⁡(E)) such that U|F¯ is the generator of Hn⁢(F¯,ℤ)=ℤ (given by orientation). We call U the Thom class. Furthermore, we have

is an isomorphism for all k. Finally, the map

sends the Thom class U to the Euler class e⁡(E).

Proof.

Consider the Serre spectral sequence for (Dn,Sn−1)↪(D⁡(E),S⁡(E))→B. Then the Euler class is the Thom class by the prior discussion about transgression. Alternatively, we may construct the Thom class locally and patch, and this can be found in Chapter 10 of Milnor-Stasheff. □

2.3 Vector Bundles

We may consider the functor

We may also consider the functors Vectℝk,Vectℝ+,k of real vector bundles and oriented real vector bundles. This functor is representable!

Theorem 2.3.1 (Brown representability).

Let h:𝗁𝖢𝖶∗op→𝖲𝖾𝗍 be a functor. If

  1. 1.

    For X=⋁Xα with iα:Xα↪X, then ∏iα∗:h⁡(X)≃∏h⁡(Xα) is an isomorphism;

  2. 2.

    If X=A∪B is a union of subcomplexes and a∈h⁡(A),b∈h⁡(B) retrict to the same element in h⁡(A∩B), then there exists x∈h⁡(X) restricting to a,b,

then there exists a CW complex K and u∈h⁡(k) such that

is a bijection.

Note that for Vectℂk, the condition of being homotopy invariant is nontrivial to decide. We want ft:X×I→Y such that f0∗⁢η≃f1∗⁢η. The idea is that for ξ≔ft∗⁢η→X×I, we need ξ|X×0=f0∗⁢η,ξ|X×1=f1∗⁢η.

Proposition 2.3.2.

Let X be a CW complex (or in general X is paracompact). Then if ξ is a vector bundle over X×I, we have ξ|X×0≃ξ|X×1.

Recall that X is paracompact if X is Hausdorff and any open cover {Uα} admits a locally finite refinement {Vβ}. This is equivalent to every open cover admitting a partition of unity subordinate to it.

Remark 2.3.3.

We may also assume the indices β are countable.

Example 2.3.4.

Any η→X paracompact admits a Hermitian metric. To see this, simply trivialize over the charts Uα and then glue using our partition of unity.

Now we note that if E→X×[a,b] is trivial over [a,c] and [c,b], then E is trivial. This implies that any η→Dk is trivial. Fix trivializations h1:X×[a,c]×ℂk→E1, h2:X×[c,b]×ℂk→E2. These do not necessarily agree on X×{c}×ℂk, but we simply consider φ=h1⁢h2−1 and change h2 by φ to get the desired trivialization.

The second observation is that if η→X×I then there exists an open cover {Uα} such that η|Uα×I is trivial. To see this, we simply trivialize over UX,i×[ti−1,ti] and then glue the trivializations.

Now we return to the proof that Vectℂk is homotopy invariant. Given a partition of unity φi subordinate to Uα (with e⁢t⁢a|Uα×I trivial), we write ψm:φ1+⋯+φm:X→[0,1]. Then define Xm to be the graph of ψm. We know supp⁡φm+1⊆Uα for some α, so p:Xm+1→Xm lifts to hm:η|Xm+1→η|Xm. We know η is trivial on Uα×I, so we have

Then we know X0 is X×{0}, so using local finiteness of the partition of unity, we obtain the desired result.

We now want to find a concrete space that represents this functor. Define the Grassmannian

of k-planes in ℂn. This has a tautological bundle ℂk↪γn,kℂ→Gn,kℂ given by γ={(V,v)∣v∈V}. Of course, if m<n, we know Gm,kℂ↪Gn,kℂ, so we define Gkℂ≔G∞,kℂ=⋃nGn,kℂ. Because the tautological bundles are well behaved under this, we obtain a tautological bundle γk→Gkℂ.

Theorem 2.3.5.

The bundle γk→Gkℂ is a universal bundle. This means that [X,Gkℂ]≃Vectℂk⁡(X), where f↦f∗⁢γk.

Remark 2.3.6.

We can also write Gkℂ=B⁢U⁢(k) as the classifying space of U⁡(k).

Proof of Theorem.

Let η→X. Then η≅f∗⁢γk if and only if there exists g:η→ℂ∞ which is linear injective on each fiber ηx. Given p:η→X, choose a countable cover with p−1⁢(Ui) trivial and φi a partition of unity subordinate to Ui. Then define

and consider the map (φi∘p)⋅gi:η→ℂn. This defines g:η→(ℂn)∞=ℂ∞.

Now we need to show that if g0,g1:η→ℂ∞ that are injective on fibers, then there exist gt:η×I→ℂ∞ linear injective on fibers at each time. We define

This is injective at each t and moves g0 to the odd entries. Similarly, we can move g1 to the even entries, so we can write gt=(1−t)⁢g0+t⁢g1. □

A characteristic class is a natural transformation Vectℂk⁡(X)→𝑐H∗⁢(X,ℤ). By the Yoneda lemma, characteristic classes are in bijection with the cohomology of Gkℂ. We now study basic facts of the Grassmannians:

  • •

    There is a transitive action of U⁡(n) on Gn,kℂ. The stabilizer of ℂk⊂ℂn is U⁡(k)×U⁡(n−k).

  • •

    Clearly we have Gn,1=ℂ⁢ℙn−1.

  • •

    The Grassmannian Gn,kℂ has a nice cell decomposition generalizing ℂ⁢ℙn=pt∪D2∪D4∪D6∪⋯.

Example 2.3.7.

Consider the Grassmannian G4,2 of 2-planes in ℂ4. Then the interior of cells is described by dimensions of intersections with ℂ1⊂ℂ2⊂ℂ4⊂ℂ4. The cells are given by:

0-cell:

This is just {ℂ2} corresponding to (1,2,2).

2-cell:

This is the set {ℂ1⊂V⊂ℂ3} corresponding to (1,1,2).

4-cells:

These are {ℂ1⊂V} corresponding to (1,1,1) and {V⊂ℂ3} corresponding to (0,1,2).

6-cell:

This is given by {dimV∩ℂ2>0} corresponding to (0,1,1).

8-cell:

This is everything else, which corresponds to (0,0,1).

For an alternative description, a 2 plane corresponds to the row reduced echelon form of a 2×4 matrix. Then we have the following correspondence:

0-cell:

This is the matrix [00100001].

2-cell:

This is the matrix [01∗00001].

4-cells:

These are the matrices [1∗∗00001],[010∗001∗].

6-cell:

This is the matrix [1∗0∗001∗].

8-cell:

This is the matrix [10∗∗01∗∗].

These cells are called Schubert cells. These behave nicely with the inclusions and Gn,k has (nk) cells. The number of 2⁢i-cells is given by the number of partitions of i into at most k integers that are each at most n−k. The cells are of even dimension, so we can compute the group structure on the cohomology, but not the ring structure.

Theorem 2.3.8.

The ring struction on cohomology of the Grassmannian is given by H∗⁢(Gkℂ,ℤ)≅ℤ⁡[c1,…,ck], where deg⁡ci=2⁢i.

Definition 2.3.9.

The characteristic classes of E→B associated to the ci are the Chern classes ci⁢(E). By convention, cj⁢(E)=0 for j>dimE.

Proof.

We need many fibrations. Consider the spaces Gn,k of k-planes in ℂn and Fn,k of k-flags in ℂn. It is clear that Fn,k also parameterizes ordered k-tuples of orthogonal complex lines. Of course there is a fibration Fk,k↪Fn,k→Gn,k. We can also understand H∗⁢(Gkℂ) inductively as well.

Now we will show that H∗⁢(F∞,k)≅ℤ⁡[x1,…,xk], where deg⁡xi=2. Each xi is the pullback of the generator of H2⁢(ℂ⁢ℙ∞) under the maps (ℓ1,…,ℓk)↦ℓi. We will use the fibration ℂ⁢ℙn−k↪Fn,k→Fn,k−1. By induction, we have a fibration

and H∗⁢(F∞,k−1)≅ℤ⁡[x1,…,xk−1]. Now we use the Leray-Hirsch theorem, but first we need to show that H∗⁢(E)↠H∗⁢(F). This is because we have the map F∞,k→ℂ⁢ℙ∞ given by (ℓ1,…,ℓk)→ℓk and this restricts to the generator of ℂ⁢ℙ∞ udner ℂ⁢ℙ∞↪F∞,k. Therefore, H∗⁢(F∞,k) is a free ℤ⁡[x1,…,xk−1]-module with additive basis 1,xk,xk2,…. Now we observe that

is a homotopy equivalence by Hurewicz.

Now we need to consider the fibration Fk,k↪F∞,k→𝜌G∞,k. This induces a surjective map in cohomology, so by Leray-Hirsch, H∗⁢(F∞,k) is a free H∗⁢(G∞,k)-module with basis 1,… and thus ρ∗:H∗⁢(G∞,k)→H∗⁢(F∞,k) is injective. The map F∞,k→G∞,k sends a k-tuple of lines to their sum. We also know that ρ is Sk-invariant, so on cohomology, we have an injection

where σi is the ith elementary symmetric polynomial. But then by a dimension count, this injection is surjective. □

Theorem 2.3.10 (Splitting principle).

Let En→X be a line bundle. Then there exists X~→𝑓X such that f∗⁢E is a direct sum of line bundles and f∗:H∗⁢(X,ℤ)→H∗⁢(X~,ℤ) is injective.

This gives us the following slogan:

When working with Chern classes, we can assume E is a direct sum of line bundles E≅L1⊕⋯⊕Lk.

Proof.

Set X~=Fk⁢(E)→X. Then the fiber over x∈X is the space of k-flags in Ex. Therefore we have a fibration Fk,k↪X~→X. Now X~ has a tautological splitting, where over the point (ℓ1,…,ℓk) we have the vector space ℓ1⊕⋯⊕ℓk=Ex. This pulls back E as a direct sum of line bundles. □

Now here are some properties of Chern classes:

  1. 1.

    For any E, we have c0⁢(E)=1.

  2. 2.

    If γ is the tautological bundle over ℂ⁢ℙ1, then c1⁢(ℂ⁢ℙ1)=−x and thus γ=𝒪⁡(−1).

  3. 3.

    For vector bundles E,F, we have Ck⁢(E⊕F)=∑ck⁢(E)∪Ck−i⁢(F). Equivalently, if we define the Chern class c⁡(E)=1+c1⁢(E)+c2⁢(E)+⋯, then c⁡(E⊕F)=c⁡(E)⁢c⁡(F).

Proof.

We prove this for the universal bundles. We have the bundle γk×γℓ over Gk×Gℓ, and this is given by a classifying map Gk×Gℓ→Gk+ℓ. Now under the map (ℂ⁢ℙ∞)k×(ℂ⁢ℙ∞)ℓ→Gk×Gℓ, we know γk×γℓ pulls back to a direct sum of tautological bundles, and the same is true for γk+ℓ pulled back to (ℂ⁢ℙ∞)k+ℓ. Therefore, we have

In fact, the three properties determine the {ci} uniquely.

Remark 2.3.11.
  • •

    If ℂ is the trivial bundle, then ci⁢(E⊕ℂ)=ci⁢(E).

  • •

    If E¯ is the complex conjugate bundle of E, then ci⁢(E¯)=(−1)i⁢ci⁢(E). Note that this is also the dual bundle.

Example 2.3.12.

We have c⁡(T⁢ℂ⁢ℙn)=(1+x)n+1=1+n⁢x+(n2)⁢x2+⋯∈ℤ⁡[x]/xn+1. To describe this, we consider the tautological bundle γ→ℂ⁢ℙn+1. Clearly TL⁢ℂ⁢ℙn=Homℂ⁡(L,L⟂), where the perp is taken in ℂn+1. Therefore, T⁢ℂ⁢ℙn≡Homℂ⁡(γ,γ⟂) and γ⊕γ⟂=ℂn+1. This gives us

and therefore

We will now consider Chern classes as obstructions. Let ℂn↪E→B be a rank n bundle. Then we can take orthonormal frames to obtain Vn,kℂ↪Vk⁢(E)→B. Then sections of Vk⁢(E)→B are k-tuples of sections of E which are orthonomal at each point.

Lemma 2.3.13.

We have

Therefore, the first obstruction to finding a section of Vk⁢(E) over the 2⁢(n−k)+2 skeleton lives in H2⁢(n−k)+2⁢(B,ℤ). In fact, this is the Chern class. The Chern class cj⁢(E)∈H2⁢j⁢(B,ℤ) is the obstruction to finding k=(n+1−j) linearly independent sections over the 2⁢(n+1−j)-skeleton. Now observe that Vn,1=S2⁢n−1. Therefore, if dimℂE=n, we have cn⁢(E)=e⁡(E).

2.4 Real Vector Bundles

Now consider a vector bundle ℝn↪E→B.

Lemma 2.4.1.

For i<n−k, we have πn⁢(Vn,kℝ)=0. In addition,

Therefore we obtain obstruction classes 𝒪i∈Hi⁢(B,Z~) or 𝒪i∈Hi⁢(B,ℤ2). This is an obstruction to finding n+1−i linearly independent sections on the i-skeleton. Reducing everything modulo 2, we have

Definition 2.4.2.

The Stiefel-Whitney classes of E are defined by wi≔𝒪imod2.

Here are some facts about the Stiefel-Whitney classes:

  • •

    If γ→ℝ⁢ℙ1 is the tautological class, then w1⁢(γ)=1∈H1⁢(ℝ⁢ℙ1,ℤ2).

  • •

    These correspond to H∗⁢(Gkℝ;ℤ2)=ℤ2⁢[w1,…,wk], where |wi|=i. Note that it is harder to compute the cohomology in this case than in the complex case.

  • •

    We have w1⁢(E)=0 if and only if E is orientable.

  • •

    If dimE=n and E is orientable, then wn⁢(E)=e⁡(E)mod2.

  • •

    They are related to Sq∙ on T⁡(E) (and in fact this is the definition given in Milnor-Stasheff).

Now we will assume that our real vector bundles are oriented. Recall that Gkℝ=B⁢O⁢(k) is the space of k-planes in ℝ∞. Then we set Gk+=B⁢S⁢O⁢(k) to be the space of oriented k-planes in ℝ∞. There is a double cover Gk+→Gkℝ. Recall that every oriented bundle has an Euler class!

Remark 2.4.3.

Suppose E→B is an oriented odd-dimensional vector undle. Then the map (b,v)↦(b,−v) reverses orientation on the sphere bundle, so e⁡(E)=−e⁡(E) and thus 2⁢e⁢(E)=0.

Recall that H∗⁢(S⁢O⁢(n),ℚ) varies greatly depending on the parity of n. We should obtain the same thing for the classifying spaces.

Theorem 2.4.4.

Over ℚ, we have H∗⁢(G2⁢m+1+)⁢ℚ⁢[p1,…,pm], where |pi|=4⁢i. In the even case, we have H∗⁢(G2⁢m+)=ℚ⁡[p1,…,pm,e]/(e2=pm).

We are mostly interested in even-dimensional manifolds, so we will mostly work in the second case.

Definition 2.4.5.

The characteristic classes corresponding to p1,…,pm are called the Pontryagin classes of E.

Concretely, we can define the Pontryagin classes in terms of the Chern classes. If E→B is a real vector bundle, we may take E⊗ℂ→B the corresponding complex vector bundle. Then we have

Remark 2.4.6.

Recall that E⊗ℂ is always isomorphic to its dual. Complex conjugation gives an antilinear involution, so c2⁢i+1 is 2-torsion.

Example 2.4.7.

We will compute the Pontryagin classes of T⁢ℂ⁢ℙn.

Lemma 2.4.8.

If E is already a complex vector bundle, then E⊗ℂ≅E⊕E∗ as complex vector bundles.

Proof.

Define by i the complex structure on E. Then the map (v,0)↦(v,−i⁢v),(0,w)↦(w,i⁢w) is a ℂ-linear isomorphism. □

In particular, if E is complex, then the total Pontryagin class p⁡(E)=1+p1⁢(E)+p2⁢(E)+⋯ can be computed using

For T⁢ℂ⁢ℙn, we have

so we obtain

Now we will consider Pontryagin roots. The splitting principle for E2⁢m→B says that E can always be thought of as a direct sum of oriented 2-plane bundles analogously to Chern classes. Note that an oriented 2-plane bundle is the same as a complex line bundle (because i rotates by π2 counterclockwise). In the case where we actually have E=L1⊕L2⊕⋯⊕Ln, we obtain

These have Chern roots ±x1,±x2,…,±xm, so we have

This tells us that the xi2 are the Pontryagin roots. Also, e⁡(E)=x1⁢⋯⁢xm, so e2=pm.

Remark 2.4.9.

This is related to the fact that det:𝔰⁢𝔬⁢(2⁢m)→ℝ has an S⁢O⁢(2⁢m)-invariant square root, called the Pfaffian.

2.4.1 Relation to Steenrod problem and oriented bordism

Let Xn be a closed smooth manifold and α∈Hk⁢(X,ℤ).

Question 2.4.10.

Is there an oriented smooth manifold Mk and continuous f:M→X such that f∗⁢[M]=α?

Definition 2.4.11.

A pair (M,f) is singular manifold representing α.

Now define

with addition given by disjoint union. In addition, we set −[M,f]=[M¯,f]. This is well-defined, and in fact there is a map

The group Ωk⁢(X) is called the k-th bordism group. Clearly Ωk is functorial by postcomposition.

Remark 2.4.12.

The functors Ω∗ give a generalized homology theory. Note that Ωk⁢(pt) is very complicated. These are given by

Theorem 2.4.13.

If M4⁢k=∂W4⁢k+1, then the signature of M is 0.

Example 2.4.14.

In Ω4⁢k⁢(pt), [ℂ⁢ℙ2⁢k]≠0.

Lemma 2.4.15.

If B:V⊗V→ℝ is symmetric and nondegenerate and W⊂V is isotropic and dimW=12⁢dimV, then the signature of B is 0.

Remark 2.4.16.

Actually, it is easy to see that B=(0W10)⊕dimW.

Lemma 2.4.17.

Let ι:M↪W be the inclusion onto the boundary. Then ⟨ι∗⁢(c),[M]⟩=0 for all c∈H4⁢k⁢(W).

Proof.

This is given by ⟨i∗⁢(c),[M]⟩=⟨c,i∗⁢[M]⟩=0 because in the exact sequence

we have [W.M]↦[M], so [M]=0. □

Proof of Theorem.

Consider the commutative diagram

We know that W=Im(ι∗:H2⁢k(W)→H2⁢k(M)) is isotropic in V=H2⁢k⁢(M), so we need to show that it is half-dimensional. But then Im(ι∗)=ker⁡(ι∗) under identification. But then ι∗ is dual to ι∗, so dimIm(ι∗)=dimIm(ι∗)=dimV−dimker⁡(ι∗) and thus W is half-dimensional. □

Recall that a spectrum 𝔼 is a sequence (En,jn:ΣEn→En+1). Then we have the associated homology theory

where the colimit is over the sequence of maps given by suspension followed by jℓ.

Example 2.4.18.

Consider En=K⁡(ℤ,n) and let jn be adjoint to Ω⁢K⁢(ℤ,n+1)=K⁡(ℤ,n). Then the associated homology theory is Hn⁢(−,ℤ).

Thom described the spectrum representing Ωn. Nowadays we call this the Thom spectrum M⁢S⁢O. Here, we write M⁢S⁢O⁢(k)=T⁡(γk→B⁢S⁢O⁢(k)), where B⁢S⁢O⁢(k) is the oriented Grassmannian. Then there is a natural map

In particular, we have ιk∗⁢γk+1=ℝ¯⊕γk. Then we have a map T⁡(ℝ⊕γk)→T⁡(γk+1), but in the homework we prove that T⁡(ℝ⊕E)=Σ⁢T⁢(E).

Theorem 2.4.19 (Thom 1954).

The spectrum M⁢S⁢O represents the homology theory Ω∗.

As a consequence, we have Ωn⁢(pt)=lim→⁡πn+ℓ⁢(M⁢S⁢O⁢(ℓ)). In particular, this is finitely generated. Tensoring with ℚ, we may use techniques of Serre to prove that

Therefore Ωn⁢(pt) is finitely generated with rank 0 when 4∤n and the number of partitions of i when n=4⁢i.

Remark 2.4.20.

The generators of Ω4⁢i⁢(pt)⊗ℚ are ℂ⁢ℙ2⁢i1×⋯×ℂ⁢ℙ2⁢ik, where I=(i1,…,ik) is a partition of i. Alternatively, we may define

This is called the I-Pontryagin number. Then the map

is an isomorphism after tensoring with ℚ.

Recall that the signature gives a homomorphism Ω4⁢i⁢(pt)→ℤ. Therefore it is a linear combination of Pontryagin numbers. We may compute exactly which combination using the Hirzebruch signature theorem. Later, we will prove this as a consequence of the Atiyah-Singer index theorem. We have

  1. 1.

    sign⁡(M4)=13⁢p1.

  2. 2.

    sign⁡(M8)=145⁢(7⁢p2−p12).

  3. 3.

    sign⁡(M12)=1945⁢(62⁢p3−13⁢p1⁢p2+2⁢p13).

  4. 4.

    sign⁡(M16)=114175⁢(…).

An application of this formula is in Milnor’s 1956 paper On manifolds homeomorphic to the 7-sphere.

Theorem 2.4.21 (Milnor 1956).

There exist M7 which are homeomorphic but not diffeomorphic to S7.

The idea to distinguish smooth structures is this. We know H∗⁢(M7,ℤ)≅H∗⁢(S7,ℤ). Then because Ω7=0 there exists W8 such that M7=∂W8. Then we have an intersection form on H4⁢(W,M,ℤ)/torsion≅H4⁢(W)/torsion. In particular, the signature of W is an integer and have a number q⁡(W)=⟨p12⁢(W),[W,M]⟩.

Lemma 2.4.22.

The number λ=2⁢q⁢(W)−sign⁢(W)mod7 is is independent of the choice of W. Therefore λ is an invariant of M.

Proof.

Consider two bounding manifolds W1,W2. Then we can glue W1,W2 to obtain a closed manifold X. Then sign⁡(X)=sign⁡(W1)−sign⁡(W2) and ⟨p12⁢(X),[X]⟩=q⁡(W1)−q⁡(W2). Then the Hirzebruch formla gives us

and now basic arithmetic gives us the desired result. □

Remark 2.4.23.

This is a diffeomorphism invariant of M. Then λ⁡(S7)=0 because S7 bounds S8 and gluing them together gives S8.

Examples 2.4.24.

We will give examples of M7 homeomorphic to S7 but not diffeomorphic to it. We will construct these as sphere bundles of oriented rank 4 bundles over S4. Consider ℝ4↪E→S4. These can be described using clutching functions. If we take S4=D+4∪D−4, and we know E|D±=DI×ℝ4. Therefore it suffices to describe the gluing along the equator, which gives us a map S3→S⁢O⁢(4). Now oriented vector bundles are described by π3⁢(S⁢O⁢(4))=ℤ⊕ℤ.

Very explicitly, we can identify S4 with the unit quaternions and ℝ4=ℍ. Then fg⁢j:S3→S⁢O⁢(4) is defined by

For example, the trivial map f0,0 is given by S4×ℝ4. The unit sphere is S4×S3. Similarly, f1,0 gives γ→ℍ⁢ℙ1 and gives us the unit sphere S7. Now for k odd, we will choose h,j such that h+j=1,h−j=k. Then define Mk to be the unit sphere bundle for fh⁢j.

We will see that Mk is homeomorphic to S7 but λ⁡(Mk)=k2−1. Thus M3 is homeomorphic but not diffeomorphic to S7. First, we will write an explicit f:Mk→ℝ with exactly one maximum and minimum. Then using gradient flow, we have Mk=D7∪φD7 via some φ:∂D7→∂D7. Because φ extends radially to D7, we have a homeomorphism Mk≅S7. To compute λ⁡(Mk), note that M7 bounds the unit disk bundle of E→S4.

2.5 Topological K-theory

Let X be a finite CW complex (more generally a compact Hausdorff space). Then let Vect⁡(X) denote the set of isomorphism classes of (complex) vector bundles on X.

Remark 2.5.1.

The rank may vary on each component of X.

Then Vect⁡(X) is a unital semiring with operations ⊕,⊗ and unit the trivial line bundle ℂ¯.

Remark 2.5.2.

Unlike ℕ, Vect⁡(X) is not cancellative. This means that if a+b=c+b, then a≠c in general.

Example 2.5.3.

Consider T⁢S2⊕ℝ¯=ℝ3¯. However, T⁢S2≠ℝ2.

We may define ℤ={(a,b)∈ℕ2}/(a+b′=a′+b). This only works because ℕ is cancellative, but if we modify the condition to be (a,b)∼(a′,b′) if a+b′+e=a′+b+e for some e, then this defines the Grothendieck ring.

Definition 2.5.4.

Define the K-theory ring K⁡(X) to be the Grothendieck ring of Vect⁡(X).

This defines a functor from 𝖿𝗂𝗇𝖢𝖶op→𝖱𝗂𝗇𝗀.

Example 2.5.5.

We know that Vect⁡(pt)=ℕ, so K⁡(pt)=ℤ. This is because vector bundles on a point are the same as vector spaces.

Thus if we consider the inclusion pt↪X, we obtain the (virtual) rank map K⁡(X)→K⁡(pt)=ℤ. This is locally constant. Note that if X is finite, then given E→X, there is E′→X such that E⊕E′≅ℂ¯n for some n. Here are some consequences:

  1. 1.

    If α∈K⁡(X), then we can write α=a−ℂN for some a,N. Here, if α=E−F, then we choose F⊕F′≅ℂN, and thus we get α=E+F′−ℂN.

  2. 2.

    If α=a+ℂN and β=b−ℂM, then α=β if and only if they have the same rank everywhere and a,b are stably equivalent (which means that a⊕ℂA=b+ℂB for some A,B).

    To see this, we have a+ℂM+ℰ=b+ℂN+ℰ for some ℰ. Then we simply add ℰ′ such that ℰ+ℰ′=ℂL. Now we have a+ℂM+L=b+ℂN+L, so we are done.

Now we would like to throw away the information about stably equivalent bundles. If (X,x0) is a pointed space, we define the reduced K-theory K~⁢(X,x0)≔ker⁡(K⁡(X)→K⁡(x0))=ℤ. This depends only on the component of x0 and is a non-unital ring.

Proposition 2.5.6.

We can identify K~⁢(X,x0) with vector bundles on X up to stable equivalence.

Now stabilization corresponds to the inclusion B⁢U⁢(k)↪B⁢U⁢(k+1), so a bundle up to stabilization has classifying space B⁢U≔B⁢U⁢(∞)=⋃B⁢U⁢(k). Therefore, we have

Corollary 2.5.7.

K~⁢(X)=[X,B⁢U].

Remark 2.5.8.

This is false for noncompact spaces in general.

We may also take the maps

Now we define U≔U⁡(∞)=⋃kU⁡(k). Then we have a fibration U↪E⁢U→B⁢U, where E⁢U=Vℂ⁢(∞,∞). We also have fibrations

Because πr⁢(Vℂ⁢(n,k))=0 for r≤2⁢(n−k), we see that Vℂ⁢(∞,∞) is contractible.

Example 2.5.9.

It is easy to see that

Similarly, we have

If γ is the tautological class (here, take S2=ℂ⁢ℙ1), then this is generated by γ−1. This is because c1⁢(γ)=−1. Therefore K⁡(S2)=K~⁢(S2)⊕ℤ=ℤ⊕ℤ is generated by γ,1. The map is given by a↦(c1⁢(a),dim(a)). Now that (γ⊗γ)⊕1≅γ⊕γ. In particular, (γ−1)2=0. Thus we have K⁡(S2)≅ℤ⁡[γ]/(γ−1)2.

Now we would like to extend this to a generalized cohomology theory.

Lemma 2.5.10.

Let (X,A) be a finite CW pair. Then K~⁢(X/A)→K~⁢(X)→K~⁢(A) is exact (because A⊂X is a cofibration).

Because A⊂X is a cofibration, we know that X/A≃X∪C⁢A. Now the exact sequence extends to

by our discussion of cofiber sequences from last semester. This gives us a long exact sequence

Now we can manually define a cohomology theory as follows:

Definition 2.5.11.

For q>0, define K~−q⁢(X)≔K~⁢(Σq⁢X). Analogously there is an unreduced theory K−q⁢(X,A)=K~⁢(Σq⁢(X/A)). This gives K−q⁢(X)≔K−q⁢(X,∅)=K~⁢(Σq⁢(X⊔pt)).

Example 2.5.12.

We have K−q⁢(pt)=K~⁢(Σq⁢(S0))=K~⁢(Sq)=πq⁢(B⁢U).

The K-groups satisfy a miraculous identity:

Theorem 2.5.13 (Bott periodicity, version B).

For all spaces X, K~⁢(Σ2⁢X)≅K~⁢(X).

Therefore K−q is 2-periodic and we will consider K0,K1 as the interesting cases. The original statement of Bott periodicity is

Theorem 2.5.14 (Bott periodicity, version C).

The homotopy groups K~⁢(Sq)=πq⁢(B⁢U)=πq−1⁢(U) are 2-periodic. Therefore πk⁢(U)=ℤ is k is odd and 0 if k is even.

The most useful form of this result is

Theorem 2.5.15 (Bott periodicity, version A).

For all spaces X, we have K⁡(X×S2)=K⁡(X)⁢[γ]/(γ−1)2. Here, γ is the pullback of the tautological bundle γ→ℂ⁢ℙ1=S2.

Concretely, we have a map

Francesco says that proving this result is difficult online, so we only give a sketch.

Sketch of proof.

We will use generalized clutching functions. Choose a bundle E→X. This gives a bundle E→X×D±2. But of course D2 is contractible, so we can glue them together using some f:X×S1→Aut⁡(E). Thus f⁡(x,z)∈G⁢L⁢(E(x,z)), so we obtain a class [E,f].

The key point is that stably, any bundle on X×S2 has the form [E,1]⊕[E′,z]. □

Example 2.5.16.

We can represent γ→S2 as [pt,f⁡(x,z)=z].

Proof of version B.

Consider the long exact sequence

Therefore φ is surjective because K⁡(X∨S2)=K~⁢(X∨S2)⊕ℤ=K~⁢(X)⊕K~⁢(S2)⊕ℤ. This argument also works for φ−1, and thus

By version A, we see that x=α1⊗ℂ+α2⊗γ=α⊗(γ−1)+β⊗γ. If if x∈ker⁡φ, , then its restrictions to both X,S2 are trivial. Restricting to X, we require that 0=α⁢dim(γ−1)+β⁢dimγ=β. Restricting to S2, we see that 0=dimα⋅(γ−1), so α∈K~⁢(X). Therefore the map K~⁢(X)→K~⁢(Σ2⁢X) given by α↦α⊗(γ−1) is an isomorphism. □

Corollary 2.5.17.

The group K~⁢(S2⁢n)=ℤ is generated by (γ−1)⊗n and K~⁢(S2⁢n+1)=0.

Now we want to consider K~⁢(X) in the noncompact case. The options are

  1. 1.

    The Grothendieck ring of Vect⁡(X);

  2. 2.

    The set of maps [X,B⁢U];

  3. 3.

    The limit lim←⁡K⁡(Y) for compact Y.

Unfortunately these are not the same. We are interested in the case where X is Hausdorff and locally compact. For example, if X is a CW complex such that every point is contained in only finitely many cells or if E→X is a vector bundle to a compact Hausdorff space, then E is Hausdorff and locally compact. In our case, the one-point-compactification X+ is Hausdorff.

Definition 2.5.18.

If X is locally compact and Hausdorff, define K⁡(X)≔K~⁢(X+,∞).

This is a (unital) ring and is functorial under proper maps. Also, if U⊂X is open and relatively compact with U+=X/X∖U=X+/X+∖U, we obtain a map X+→U+. In particular, we obtain a map K⁡(U)→K⁡(X). Then we have

where U ranges over the relatively compact subsets of X. Of course, if X is compact, we recover the original K-theory.

There is an alternative point of view in terms of complexes. Let

be a complex of bundles over X. We define the support supp⁡𝔼 to be the locus where the sequence is not exact.

Example 2.5.19.

If we consider the bundle 0→ℂ→𝑧ℂ→0, then the support of this complex is {0}.

Define C⁡(X) to be the homotopy classes of complexes with compact support, where 𝔼0∼E1 if there exists 𝔾→X×I with compact support and 𝔾|X×{i}=𝔼i. Now this contains the set C∅⁢(X) of complexes with empty support.

Proposition 2.5.20 (Segal).

If X is Hausdorff and locally compact, then K⁡(X)≅C⁡(X)/Cϕ⁢(X) as rings.

The basic idea is that if E−F∈K⁡(X)=K~⁢(X,∞), then the dimensions of E,F are equal at ∞, so we can consider the complex 0→E→𝛼F→0 where α is a chosen isomorphism near ∞ and extended arbitrarily with compact support to the entire X.

Remark 2.5.21.

We have the inclusion supp⁡(𝔼⊗𝔼′)⊆supp⁡(𝔼)∩supp⁡(E′). Therefore, the multiplication is well-defined.

Remark 2.5.22.

The sequence 0→ℂ→𝑧ℂ→0 in K⁡(ℝ2)=K~⁢(S2) corresponds to γ−1.

Example 2.5.23.

We have K⁡(ℝ2⁢n)=K~⁢(S2⁢n)=ℤ and this is generated by (γ−1)⊗n. The idea is to tensor multiple copies of 0→ℂ→ziℂ→0, and the generator is the Koszul complex

Now recall the Thom isomorphism in cohomology. If ℝn↪E→B is oriented, then Hk⁢(B,ℤ)→UEHk+n⁢(T⁢E,ℤ)=Hck+n⁢(E,ℤ) is an isomorphism. The Thom class uE restricts to the generator of Hcn⁢(F,ℤ) for each fiber. In K-theory, if we have ℂn→E→𝑓B, then we can use define the Koszul complex of f∗⁢E. Restricting to a fiber F, we obtain the Koszul complex of F=ℂn. In particular, the support is the zero section of E→B, which is B. Thus if 𝔽 is a compactly supported complex on B, f∗⁢𝔽⊗⋀(E) is a compactly supported complex on E. Therefore we obtain a map K⁡(B)≃K⁡(E).

Theorem 2.5.24.

This is the K-theoretic Thom isomorphism.

Proof.

We can prove this locally using Bott periodicity and then patch to obtain the global result. □

Remark 2.5.25.

The Thom isomorphism holds under weaker assumptions. This is when E admits a spinc structure by Atiyah-Bott-Shapiro.

Now suppose X is a finite CW complex with only even-dimensional cells. Now we have an exact sequence

By induction, we can obtain K⁡(X)=ℤ#⁡(cells). Here, we note that K1⁢(X2⁢n)=K1⁢(X2⁢n−2)=0 and K1⁢(X2⁢n,X2⁢n−2) is just ℤ#⁡(2⁢n-cells), so this is very easy. In particular, we have K⁡(ℂ⁢ℙn)=ℤn+1. This can also be done using the Leray-Hirsch theorem, and in fact the splitting principle holds. For all E→X, there exists Y→𝑓X such that f∗:K⁡(X)→K⁡(Y) is injective and f∗⁢E is a direct sum of line bundles.

2.6 K-theory and cohomology

Now we will discuss the relationship between K-theory and cohomology. For a space X, we have the two rings K⁡(X),H∗⁢(X). Can we find a morphism between them? The natural idea is to consider characteristic classes. We can attempt to use the total Chern class, but this is not a homomorphism!

Definition 2.6.1.

Given a vector bundle E→B, where B is a finite CW complex, define the Chern character by

Here xi are the Chern roots of E. We can actually rewrite this as

If we write x1k+⋯+xnk=sk⁢(σ1,…,σn), where sk is the Newton polynomial and σi are the elementary symmetric polynomials, we obtain

Proposition 2.6.2.

The Chern character defines a homomorphism ch:K⁡(X)→H2⁣∗⁢(X,ℤ).

Proof.

This is obvious for line bundles because ex⁢ey=ex+y, and now we are done by splitting. □

Theorem 2.6.3.

The Chern character chℚ:K⁡(X)⊗ℚ→H2⁣∗⁢(X,ℚ) is an isomorphism for any finite CW complex X.

Note that this result is false over the integers. In addition, the map

is an isomorphism.

Proof.

We induct on the skeleta. This is obvious for the case of a point. So now consider the long exact sequences

Now we need to show that the maps on the relative groups are isomorphisms by the five lemma. Because Xn/Xn−1=⋁Sn, we need to show this for spheres. In general, this is because

commutes. Returning to the case of a sphere, we have

and also

is the generator. Now we simply need to show that chℚ are isomorphisms for S1,S2. □

Remark 2.6.4.

The torsion of K⁡(X) and H2⁣∗⁢(X,ℤ) can be very different. For example, we have K⁡(ℝ⁢ℙm−1)=ℤ⊕ℤ/2m−1⁢ℤ, while H2⁣∗⁢(ℝ⁢ℙm−1)=ℤ⊕(ℤ/2⁢ℤ)m−1.

In general, K-theory is very hard to compute. One can compute it using equivariant K-theory, or we may use the Atiyah-Hirzebruch spectral sequence. Here, if X is a finite CW complex, then we have a spectral sequence with

In fact, this is true for any generalized cohomology theory.

Corollary 2.6.5.

The Chern character ch:K⁡(S2⁢n)→H2⁣∗⁢(S2⁢n,ℚ) actually restricts to ch:K~⁢(S2⁢n)→H2⁢n⁢(S2⁢n;ℚ) which is an isomorphism onto H2⁢n⁢(S2⁢n,ℤ).

An application of this2 is the following: If x∈H2⁢n⁢(S2⁢n,ℤ) is given by x=cn⁢(E), then (n−1)!∣x. This is because c1⁢(E)=⋯=cn−1⁢(E)=0, so ch⁡(E)=dimE+cn⁢(E)(n−1)!∈H∗⁢(S2⁢n,ℤ).

Theorem 2.6.6 (Borel-Serre).

If T⁢S2⁢n admits a complex structure J, then n=1,3.

Proof.

Suppose we have a complex structure. Then cn⁢(T⁢S2⁢n,J)=e⁡(T⁢S2⁢n)=χ⁡(S2⁢n)=2, so (n−1)!∣2 and thus n=1,2,3. We know T⁢S4 has no complex structure by Wu’s theorem. □

Remark 2.6.7.

When n=1, then S2=ℂ⁢ℙ1 is a complex manifold, so we have holomorphic charts. When n=3, we can construct J using the octonions (viewing S6 in the purely imaginary octonions), and the problem of whether S6 admits a complex structure is open.

Here is an important point. For E→X a vector bundle on a finite CW complex, we have a diagram which does not commute:

On one hand, we have ch⁡(x⋅λE)=ch⁡(x)⁢ch⁡(λE) and on the other hand we have uE⋅ch⁡(x). In general, ch⁡(λE)≠uE! Assume there exists t∈H2⁣∗⁢(X,ℚ) such that t⋅ch⁡(λE)=uE. Formally, we obtain t=uE|Xch⁡(λE)|X. If E has Chern roots x1,…,xn, then we have uE|X=e⁡(E)=x1⁢⋯⁢xn. On the other hand, λE|X=1−E+⋀2E−⋀3E+⋯=(1−x1)⁢(1−x2)⁢⋯⁢(1−xn). Applying the Chern character, we have ch⁡(λE|X)=(1−ex1)⁢(1−ex2)⁢⋯⁢(1−exn), and thus

Definition 2.6.8.

The Todd class td⁡(E) is defined by

Later, when we do index theory, this will be very important.

Remark 2.6.9.

We have the identity xex−1=∑Bk⁢xkk!, so this is the exponential generating function for the Bernoulli numbers.

Theorem 2.6.10 (Adams).

If there exists f:S4⁢n−1→S2⁢n with H⁡(f)=1, then n=1,2,4.

We will interpret H⁡(f) using K-theory. Consider the mapping cone Cf=S2⁢n∪fD4⁢n. then we have an exact sequence

and here we have (γ−1)2⁢n↦α,β↦(γ−1)n. Then because β2↦0, we can write β2=H⁡(f)⁢α.

The tool we will use to study this are Adams operations. These are natural transformations ψk:K⁡(X)→K⁡(X) satisfying the following properties:

  1. 1.

    For L a line bundle, ψk⁢(L)=L⊗k.

  2. 2.

    For any k,ℓ, ψk∘ψℓ=ψk⁢ℓ.

  3. 3.

    If p is prime, then ψp⁢(α)≡αp(modp). This means that ψp⁢(α)−αp=p⁢β for some β.

The way we will define these is by guessing a formula using splitting and then write it in an invariant way. We know that if E=L1⊕⋯⊕Ln, we want

Introducing a formal variable ti, this becomes t1k+⋯+tnk. Then we obtain

so we have

and thus we may define

The properties are easily checked using the splitting principle! For example, the third property is Fermat’s little theorem. Now this induces ψk:K~⁢(X)→K~⁢(X).

Proposition 2.6.11.

The map ψk:K~⁢(S2⁢n)→K~⁢(S2⁢n) is multiplication by kn.

Proof.

Observe that ∗:K~⁢(X)⊗K~⁢(Y)→K~⁢(X∧Y) is given by α⊗β↦pX∗⁢(α)⋅pY∗⁢(β). Now ψk commutes with multiplication, so we simply need to compute this for K~⁢(S2). But now we have

where α=γ−1. Now assume the result is true for S2⁢n−2. Then K~⁢(S2)⊗K~⁢(S2⁢n−2)→K~⁢(S2⁢n) is the Bott periodicity map, so because ψk commutes with multiplication, we obtain the desired result. □

Proof of Adams’ theorem.

Let f:S4⁢n−1→S2⁢n satisfy H⁡(f)=1. Write Cf=S2⁢n∪fD4⁢n. Now we have an exact sequence

and this is given by (γ−1)2⁢n↦α,β↦(γ−1)n. Then we know β2=H⁡(f)⁢α. For every k∈ℕ, we have ψk⁢(α)=k2⁢n⁢α and ψk⁢(β)=kn⁢β+uk⁢α for uk∈ℤ. Because multiplication of Adams operations is commutative, we have ψk⁢ψℓ=ψℓ⁢ψk. Therefore, for all k,ℓ, we have

Because H⁡(f)=1, we know

and thus u2 is odd. Setting k=3,ℓ=2, we have

and thus 2n∣3n−1, which implies n=1,2,4. □

Corollary 2.6.12.

If Sn−1 is parallelizable, then n−1=0,1,3,7. This is equivalent to ℝn being a division algebra. This means that there exists μ:ℝn→ℝn that is continuous and linear in each variable such that μ⁡(x,y)=0 implies that one of x,y=0. Thus n=1,2,4,8.

Sketch of proof.

If ℝn is a division algebra, then it is a division algebra with unit. Next, it is clear that Sn−1 is a unital H-space. But now this implies n=2⁢m and then there exists S4⁢m−1→S2⁢m with Hopf invariant 1. □

To go from ℝn being a divison algebra to Sn−1 being an H-space, let μ be the multiplication on ℝn. Then the multiplication on Sn−1 is simply (x,y)↦μ⁡(x,y)‖μ⁡(x,y)‖. To go from Sn−1 being parallelizable to being an H-space, choose an orthonormal frame v1,…,vn−1 of T⁢Sn−1. Then for all x, (x,v1⁢(x),…,vn−1⁢(x)) is an orthonormal basis of ℝn. We may assume that at e1, this is (e1,…,en), so we obtain a matrix αx∈S⁢O⁢(n) with αe1=id. Now the H-space structure is x∗y=αx⁢(y) and this is clearly unital with unit e1.

Remark 2.6.13.

Adams used similar techniques to provide sharp upper bounds on the number of linearly independent vector fields on Sn−1. To construct the vector fields, we need to use the theory of modules over Clifford algebras. The Clifford algebras are all associative, and the first few examples are ℝ,ℂ,ℍ. The category of modules over these are periodic, and this is some sort of algebraic Bott periodicity.

3 The Atiyah-Singer Index Theorem

This is a result that relates analysis and topology. First, we will discuss the relation between analysis and K-theory.

3.1 Fredholm Operators

Let L:ℝn→ℝm be a linear map. Then define the index ind⁡(L)≔dimker⁡(L)−dimcoker⁡(L). In finite dimensions, this is just

by rank-nullity. In infinite dimensions, things are more interesting.

Definition 3.1.1.

Let X,Y be Hilbert spaces and L:X→Y be a bounded operator. Then L is Fredholm if dimker⁡L<∞, ImL is closed, and dimcoker⁡L<∞. Now we may define the index ind⁡(L)≔dimker⁡L−dimcoker⁡L∈ℤ.

Remark 3.1.2.

For X,Y Hilbert spaces, the condition that ImL is closed is automatic. This is a consequence of the open mapping theorem.

Example 3.1.3.

Consider L:ℓ2⁢(ℕ)→ℓ2⁢(ℕ) be the left shift (x1,x2,…)↦(x2,x3,…). Then ind⁡(L)=dimker⁡L−dimcoker⁡L=1−0=1. Similarly, we may consider the right shift R:ℓ2⁢(ℕ)→ℓ2⁢(ℕ) given by (x1,x2,…)↦(0,x1,x2,…), and ind⁡(R)=0−1=−1.

Theorem 3.1.4.

The space Fred⁡(X,Y)⊆B⁡(X,Y) is open and the index ind:Fred⁡(X,Y)→ℤ is locally constant.

Remark 3.1.5.

If {Lt}t∈[0,1] is a family of operators, then the index is constant, but of course the dimensions of the kernel and cokernel can vary wildly. Thus, index should be a topological quantity in some sense.

Lemma 3.1.6.

Let T:X→Y be invertible. Then if p:X→Y has small enough norm, T+p is invertible.

Thus being invertible is an open condition!

Proof.

Write (T+p)=T⁡(1+T−1⁢p). Then T−1⁢p has small norm, so we can write

and this sum converges. Therefore we have

Proof of Theorem 3.1.4.

Let T:X→Y be Fredholm. Then we know X=C⊕ker⁡T and Y=ImT+D, where dimker⁡T<∞ and dimD<∞. Now we can write T in the block matrix form

Now suppose p=(abcd) has small norm. Then we have

and T′+a is invertible. Then by Gaussian elimination using

we see that

so we have

Theorem 3.1.7 (Atiyah-Jänich).

Let H be a separable complex Hilbert space. Then there is a natural bijection

for all finite CW complexes X.

Remark 3.1.8.

Let S,T:H→H be Fredholm. Then S∘T is also Fredholm. In particular, [X,Fred⁡(H)] is naturally a semigroup and the map [X,Fred⁡(H)]→K⁡(X) is a map of semigroups.

Corollary 3.1.9.

There is a weak homotopy equivalence Fred⁡(H)∼B⁢U×ℤ.

Remark 3.1.10.

The map dim:K⁡(X)→ℤ corresponds to the index.

To construct the map [X,Fred⁡(H)]→K⁡(X), note that an element of [X,Fred⁡(H)] is a family {Tx}x∈X, where Tx:H→H. Then because X is compact, there exists a finite dimensional V⊂H such that Im(Tx)+V=H for all x. Therefore Tx−1⁢(V)⊂H, so ⋃Tx−1⁢(V)→X is a vector bundle. Now we set

Note that the vector bundle has dimension ind⁡Tx

The key point in the proof of Atiyah-Jänich is the fact that the space G⁢L⁢(H) of bounded invertible operators on H is weakly contractible. This is called Kuiper’s theorem. Note that this is in contrast to G⁢L⁢(∞)≃U⁡(∞), which is not contractible.

Definition 3.1.11.

A bounded operator K:X→Y between Hilbert spaaces is compact if it sends bounded sets to precompact sets.

Example 3.1.12.

Any finite rank K is compact. Also, the operator

is compact.

Theorem 3.1.13.

Let T,K:X→Y. Then if T is Fredholm and K is compact, then T+K is Fredholm and ind⁡(T)=ind⁡(T+K).

Sketch of proof.

We want to show that dimker⁡(I+K)<∞, and this is the same as the unit sphere being compact. Here, if ‖xn‖=1 and (I+K)⁢(xn)=0, then the sequence yn={−K⁢xn} has a convergent subsequence, and thus the unit sphere is compact.

To show that the cokernel is finite dimensional, we use Schauder’s theorem. Here, if K:X→Y is compact, then K∗:Y∗→X∗ is also compact. □

3.2 Index and topological invariants

The goal of this section is to prove that if M is a closed oriented smooth manifold, then χ⁡(M),sign⁡(M) are naturally interpreted as indices of differential operators. Recall the de Rham complex of C∞-forms

This has cohomology Hp⁢(M,ℝ). Note from this point of view that a cohomology class does not have a preferred representative.

A goal of Hodge theory is to find a preferred representative when M is equipped with a Riemannian metric g. Intuitively, each Ωi has an L2 inner product

so we obtain an L2 norm. Now a good candidate for the preferred representative of a cohomology class is the shortest element! It is not obvious that this even exists.

Now let (V,⟨−,−⟩) be an oriented real Euclidean vector space of dimension n. Then we define the Hodge star

to satisfy the property that for any oriented orthonormal basis e1,…,en, we have

In a more invariant manner, we have vol⁡(W)↦vol⁡(W⟂). Now if α,β∈⋀iV, then α∧∗β is the inner product ⟨α,β⟩⁢dvol. On a manifold, if α,β∈Ωi, we obtain

Returning to the de Rham complex, we would like to write an adjoint d∗:Ωp→Ωp−1. We set

Proposition 3.2.1.

The morphism d∗ is the L2 adjoint of d, so if γ∈Ωp−1,α∈Ωp, then

Proof.

Consider γ∧∗α∈Ωn−1. Then we have

Then the resired result follows from Stokes, which gives us ∫d(γ∧∗α)=0. □

Now suppose that α minimizes ‖−‖L2 in [α]∈Hp. Then any other element differs by an exact form η∈Ωp−1, so we have

so we must impose d∗α=0. In particular, we must have (d+d∗)⁢α=0.

Definition 3.2.2.

The Hodge Laplacian is the operator Ω≔(d+d∗)2:Ωi→Ωi. Note that (d+d∗)2=dd∗+d∗d.

This satisfies the following properties:

  1. 1.

    Δ:Ω0→Ω0 is the negative of the usual Laplacian −∇⁣⋅⁡∇⁡f=−∑d2dxi2.

  2. 2.

    The operator Δ is self-adjoint. Here, note that d+d∗ is self-adjoint.

  3. 3.

    Δ is nonnegative. Here, note that ⟨α,Δ⁢α⟩=‖(d+d∗)⁢α‖2≥0.

  4. 4.

    We have ker⁡Δ=ker⁡(d+d∗) on Ωi.

Theorem 3.2.3 (Hodge).

There is an orthogonal decomposition

Here, ℋ=ker⁡(Δ)=ker⁡(d+d∗) is the space of Harmonic forms. In addition, we have ℋp≅Hp⁢(M,ℝ), and in particular, each cohomology class admits a unique harmonic representative.

Remark 3.2.4.

Note that d+d∗ preserves the natural ℤ/2⁢ℤ grading on Ω∗. Therefore, we may consdier d+d∗:Ωeven→Ωodd, so we may consider its index. Therefore we have

Assuming the Hodge theorem, we therefore obtain ind⁡(d+d∗)=χ⁡(M). In general, the Atiyah-Singer index theorem tells us that this is e⁡(T⁢M).

Now let M4⁢k be oriented. We have an intersection form QM:H2⁢k⁢(M,ℝ)⊗H2⁢k⁢(M,ℝ)→⌣H4⁢k=ℝ. For de Rham cohomology, we may write

Now we will realize the signature as the index of d+d∗ with respect to a different splitting of Ω∗⁢(M). Set ε2=i2⁢k+p⁢(p−1)∗:Ωp→Ω4⁢k−p. Then it is easy to see that ε2=1, so we write Ω∗=Ω+⊕Ω−. Also ε=∗ on Ω2⁢k and ε anticommutes with d+d∗. Therefore, we can consider

This has kernel containing β+ε⁡(β) for β∈ℋℓ, where ℓ<2⁢k as well as β∈ℋ2⁢k such that β=∗β. On the other hand, the cokernel consists of forms of the form β−ε⁡(β) for β∈ℋℓ with ℓ<2⁢k and β∈ℋ2⁢k such that β=−∗β. Therefore, we may consider only the forms in ℋ2⁢k, and we see that ind⁡(d+d∗)=dimℋ+−dimℋ−, where ℋ+ is the space of self-dual forms and ℋ− is the space of anti-self-dual forms. But now we see that

when β∈ℋ+, and similarly we see that Qm is negative definite on ℋ−. Because ℋ2⁢k=ℋ+⊕ℋ−, we see that ind⁡(d+d∗)=sign⁡(M). The Atiyah-Singer index theorem will express this index in terms of Pontryagin classes, and this will give us the Hirzebruch signature formula.

3.3 Elliptic Operators

Let Mn be a smooth manifold and Em,Fℓ be vector bundles over M.

Definition 3.3.1.

A differential operator (with C∞ coefficients) of order r is an operator

that looks in charts U⊆M, where E=U×ℝm,F=U×Rℓ, like an operator of the form

where α is a multi-index and aα⁢(x):ℝm→ℝℓ is a matrix smooth in x. Here, we write

and |α| is the L1-norm.

Definition 3.3.2.

A differential operator L of order r is elliptic if for all x∈U and ξ∈ℝn∖{0} and we substitute Dα↝ξα=ξ1α1⁢⋯⁢ξnαn, the polynomial

is an isomorphism.

Example 3.3.3.

The Laplacian Δ=−∂2∂x12−⋯−∂2∂xn2 of order 2 is elliptic. Here, we have

and thus Δ is elliptic.

Example 3.3.4.

The wave operator ∂2∂x12−∂2∂x22 is not elliptic. Here, note that

and this vanishes for ξ=(1,1).

Example 3.3.5.

Consider ∂¯=∂∂x+i⁢∂∂y acting on ℂ-valued functions. Then we have

and this is an isomorphism for ξ≠0.

More intrisically, consider L:Γ⁡(E)→Γ⁡(F). Then L is elliptic at x if and only if for all u∈Γ⁡(E) with u⁡(x)≠0, and φ∈C∞⁢(M) with φ⁡(x)=0 such that dφ⁡(x)=ξ∈Tx∗⁢M∖{0}, we have L⁡(φr⁢u)≠0. Here, we can compute L⁡(φr⁢u)=PL.x⁢(ξ)⁢u.

The key point is that if L is elliptic, then solutions to L⁢u=0 are very nice. For example, if u∈C1 and ∂¯⁡u=0, then u∈C∞. Similarly, if u∈C2 and Δ⁢u=0, then u∈C∞. On the other hand, this is false for the wave equation. The correct functional analysis setup for this is Sobolev spaces. Now consider E→M a Euclidean vector bundle on a Riemannian manifold. Then fix a compatible connection ∇ on E and k∈ℕ. For u∈Γ⁡(E), we will define

Definition 3.3.6.

The Lk2-norm of u is defined by

The completion of Γ⁡(E) with respect to this norm is the Hilbert space Lk2⁢(E).

Example 3.3.7.

The space L12⁢(ℝ) contains functions that are continuous and are well-behaved. On the other hand, L2⁢(ℝ) contains discontinuous functions. Also, functions in L12⁢(ℝ2) are not necessarily continuous.

Remark 3.3.8.

If M is compact, then the space Lk2⁢(E) is independent of the choices made as a topological vector space.

Lemma 3.3.9.

Let M be commpact and L:Γ⁡(E)→Γ⁡(F) be a linear differential operator of order r. Then for all k there exists Ck>0 such that

In particular, this means we can take L:Lk+r2→Lk2⁢(F), and this is bounded. There is another estimate, which is only true for elliptic operators.

Theorem 3.3.10 (Elliptic estimate).

If L is elliptic, then there exists Dk such that

Remark 3.3.11.

The ‖u‖L2 is there to take care of ker⁡L. Also, this estimate is false for Ck norms.

We will simply give the idea of the proof in the simplest case Δ:C∞⁢(Tn,ℂ)→C∞⁢(Tn,ℂ). Recall the Fourier transform

This is an isometry up to factors of 2⁢π. Under this correspondence, the differential operator f↦ddxj corresponds to {an¯}↦{−i⁢nj⁢an¯}. Now the Sobolev norm ‖f‖Lk22 corresponds to the norm

Finally, f↦Δ⁢f corresponds to {an¯}↦{|n¯|2⁢an¯}. Therefore, we see that

and thus if n¯≠0, we see that |n¯|2⁢(1+|n¯|2)k≥Dk⁢(1+|n¯|2)k+2. This gives us the result

In general, PL,x⁢(ξ) is a degree r polynomial in ξ that only vanishes at 0. This gives us an estimate of the form

Here are some nice results about Sobolev spaces:

Sobolev embedding theorem:

If p>n/2, we have Lk+p2↪Ck.

Rellich compactness:

If k1<k2, then Lk22↪Lk12 is a compact operator if M is compact. This is an analog of Arzela-Ascoli for Sobolev spaces.

Theorem 3.3.12.

If L is an elliptic operator of order r for a compact manifold M, then for all k∈ℂ,

is Fredholm. In addition, ker⁡L and coker⁡L=ker⁡L∗ consist of smooth configurations. Therefore ind⁡L is independent of k.

Corollary 3.3.13.

Let {Lt}t∈[0,1]:Γ⁡(E)→Γ⁡(F) be a family of elliptic operators of order r. Then ind⁡(Lt)∈ℤ is constant.

Therefore if an operator depends on some choice from a connected space (for example metrics), then its index is independent of the choice. Therefore, one hopes that the index depends only on topology.

Lemma 3.3.14.

Consider L:X→Y and K:X→Z such that X is compact and ‖x‖X≤C⁡(‖L⁢x‖Y+‖K⁢x‖Z). Then dimker⁡L<∞ and ImL is closed.

Now Fredholmness of elliptic operators follows from taking K to be the embedding of Lk+r2 inside L2. The second part of Theorem 3.3.12 follows from elliptic regularity. Here, we know that if u∈Lk+r2, then L⁢u∈Lk2. Now if L⁢u∈Lk+12, then u∈Lk+r+12. In particular, L⁢u=0 implies that u∈Lk2 for all k, and therefore u∈C∞ by Sobolev embedding.

We observe that ind⁡L depends only on the family of principal symbols PL,x⁢(ξ). Thus all lower order terms do not matter. Here, the lower order terms give an operator Lk+r2→Lk+12↪Lk2, and this is compact. Now we see that the index only depends on M,E,F,{PL,x}. This leads us to the following problem.

Question 3.3.15 (Gelfand).

Write a formula for ind⁡L in terms of topological data (for example characteristic classes).

Example 3.3.16.

The operators d+d∗:Ωeven→Ωodd, d+d∗:Ω+→Ω− are elliptic. We simply need to show that the symbols are injective, so let x∈X,φ⁡(x)=0, and dφ⁡(x)=ξ∈Tx∗⁢M∖{0}. Then if α∈Ωp, we have

Similarly, we obtain d∗(φ⁢α)=iξ⁢(α)⁢(x), the contraction. Therefore we have

and this is nonzero whenever ξ,α are nonzero. Therefore d+d∗ is elliptic.

Remark 3.3.17.

The Hodge theorem requires some extra work.

  1. 1.

    First one needs to show that a generalized solution in L2 exists (this is by general nonsense such as Hahn-Banach and Riesz representation).

  2. 2.

    Now use elliptic regularity to get that the solution is smooth.

3.4 The index theorem

First, we will discuss symbols and K-theory. We will always denote our elliptic differential operator by D. Now for all x∈X and ξ≠0∈Tx∗⁢X, we know PD,x⁢(ξ):Ex→Fx is an isomorphism. Now we can put everything together to obtain a bundle map σ⁡(D):π∗⁢E→π∗⁢F, where π:T⁢X→X is the projection. Now we know σ⁡(D) is an isomorphism outside of the zero section of T⁢X→X, which is compact. This gives us a well-defined element σ⁡(D)∈K⁡(T⁢X).

Remark 3.4.1.

The inclusion of X as the zero section of T⁢X gives us a map K⁡(T⁢X)→K⁡(X), and we have σ⁡(D)↦E−F.

Remark 3.4.2.

An orientation of X gives an orientation of T⁢X. Here, locally we have T⁢U=U×ℝn, and if b1,…,bn is an oriented basis of ℝn, then we obtain an oriented basis {(b1,0),(0,b1),…}.

We are finally ready to state

Theorem 3.4.3 (Atiyah-Singer index theorem).

Let X,D be as above. Then we have

Here, ch⁡(σ⁡(D))∈Hc×⁢(T⁢X,ℚ) and td⁡(T⁢X⊗ℂ)∈H∗⁢(X,ℚ). Then we can multiply them together to obtain an element of Hc∗⁢(T⁢X,ℚ), and then evaluate it on [T⁢X].

Note that this form of the theorem is somewhat inconvenient because σ⁡(D) may be very complicated. In some special cases, we may be able to simplify. First, let ψ:Hp⁢(X)→Hcp+n⁢(T⁢X) be the Thom isomorphism. Then for b∈Hc2⁢n⁢(T⁢X), we have

Here, the sign comes from our weird orientation convention. Note that if ψ−1⁢(b)=a, then b=a⌣U, where U is the Thom class. Thus b|X=a⌣e⁡(T⁢X). Formally, we may write a=b|xe⁡(T⁢X). Assuming that this is actually possible, we obtain ch⁡(σ⁡(D))|X=ch⁡(E−F), so we obtain

Remark 3.4.4.

This does not depend on σ⁡(D). In fact, it depends only on E,F,X.

For example, this makes sense when e⁡(T⁢X) is not a zero-divisor. This also makes sense if we can interpret everything in terms of classifying spaces (for example if e⁡(T⁢X) comes from Hn⁢(B⁢S⁢O⁢(n)) where n is even). Thus if b is the pullback of a class on B⁢S⁢O⁢(n), then we are still fine, and here we call D a universal operator.

Example 3.4.5.

Consider d+d∗ on the complexified tangent bundles. Also assume n=2⁢ℓ is even. We need to compute ch⁡(Ωeven−Ωodd) and td⁡(T⁢X⊗ℂ). Now T⁢X⊗ℂ has Chern roots x1,−x1,…,xℓ,−xℓ. Therefore, we have

On the other hand, we have

Finishing the computation, we obtain

Corollary 3.4.6.

This proves that χ⁡(X)=e⁡(T⁢X) in even dimension. In fact, this is the most complicated proof of this fact that Francesco knows.

Remark 3.4.7.

Our machinery is also good to work directly with complexes of differential operators

For example, we want to consider the de Rham complex. Here, we need that the associated symbol complex is exact for all ξ≠0. Here, we need to replace the index of D with the alternating sum of dimensions of cohomologies and the Chern character of the alternating sum of the bundles.

Example 3.4.8.

Consider d+d∗:Ω+→Ω−. If V=V~⊗ℂ, where V~ is oriented of dimension 2⁢m, then ⋀∗V=⋀+V⊕⋀−V. Here, we have ⋀+(V⊕W)=(⋀+V⊗⋀+W)⊕(⋀−V⊗⋀−W) and ⋀−(V⊕W)=(⋀+V⊗⋀−W)⊕(⋀−V⊗⋀+W). This implies that ⋀+−⋀− behaves nicely under direct sums. Now we use the splitting principle, so we may assume T⁢X=P1⊕⋯⊕Pℓ is a sum of oriented plane bundles. If P=L with Chern root x, then ⋀+Pℂ≅1+L¯ and ⋀−Pℂ≅1+L. Therefore

Now we see that

Finally, we see that

Now note that xitanh⁡(xi) is a power series in the even powers of xi, so we obtain

If we actually compute the Lj, then we obtain the Hirzebruch signature formula.

Corollary 3.4.9 (Hirzebruch signature formula).

Let X4⁢k be an oriented manifold. Then sign⁡(X)=Lk⁢(p1⁢(T⁢X),…,pk⁢(T⁢X))⁢[X].

Now let X be a complex compact manifold. Now we know T⁢Xℂ=T⁢X⊗ℂ=T⁢X⊕T⁢X¯, so we have T⁢Xℂ∗=⋀1,0⊕⋀0,1. Therefore, we have ⋀0,q=⋀q(⋀0,1). Now if we set Ω0,q=C∞⁢(⋀0,q), we obtain the Dolbeaut complex

where ∂¯2=0 because we have a complex manifold. Now it is easy to see that ∂¯ is elliptic. This will imply that the complex has finite-dimensional cohomology, but in fact we have

Theorem 3.4.10 (Dolbeaut).

We have ker⁡∂p¯/Im⁢∂¯p−1=H0⁢(X,𝒪X) and has dimension h0,i.

Therefore the index is ind⁡(∂¯)=∑(−1)i⁢h0,i=χ⁡(X,𝒪X). Using the index theorem, we can compute the index as

If T⁢X has roots x1,…,xn, we obtain that ⋀0,1=T∗⁢X¯, so it has the same Chern roots. Therefore, ch⁢(∑(−1)i⁢⋀0,i)=∏(1−exi). Next, it is easy to see that

Finally, we have e⁡(T⁢X)=cn⁢(T⁢X)=x1⁢⋯⁢xn. In particular, we obtain

As a corollary, we obtain

Theorem 3.4.11 (Hirzebruch-Riemann-Roch).

Let X be a compact complex manifold. Then

More generally, if V→X is a vector bundle, then χ⁡(X,𝒪V)=(ch⁡(V)⁢td⁡(X))⁢[X].

Remark 3.4.12.

The original proof of Hirzebruch is for projective varieties.

Example 3.4.13.

If S6 has a complex structure with structure sheaf 𝒪S6, then we obtain

so some of h0,1,h0,2,h0.3≥1.

Remark 3.4.14.

The remaining interesting case of a “universal operator” is the Dirac operator. If you would like to know more, contact Alex Xu at axu@math.columbia.edu.

3.5 Sketch of proof of the index theorem

First, we will give a short history of the proof. The original proof announcement and sketch was in 1963 using cobordism theory. This was inspired by Hirzebruch’s proof of the Hirzebruch-Riemann-Roch theorem in 1956. The drawback is that this does not generalize to the equivariant setting. The published proof in 1968 was heavily inspired by Grothendieck’s 1957 proof of the Grothendieck-Riemann-Roch theorem.

First, if X is a point, then ellipticity is a vacuous assumption, so the index theorem is always true. Now the difficult part is to build the machinery to deal with functoriality. The first thing is to define the topological index t−ind⁡(D) of D (in contrast with the analytic index ind⁡(D)). Here, we have a commutative diagram

First, suppose that S⊆M is an embedded compact submanifold with normal bundle N→S. We can identify this with an open neighborhood of S in M. If N is a complex vector bundle, then K⁡(S)=K⁡(N) by the Thom isomorphism. Now we will call the map

where the second arrow is induced by the open inclusion. Now we have functoriality for embeddings with complex normal bundle.

Now suppose that π:E→X is a real vector bundle. Then T⁢E≅π∗⁢T⁢X⊕π∗⁢E. If E=T⁢X, then T⁢(T⁢X)=π∗⁢T⁢X⊕π∗⁢T⁢X, and this is naturally a complex vector bundle T⁢X=π∗⁢T⁢X⊗ℂ. Therefore any inclusion i:X↪Y induces an inclusion di:T⁢X↪T⁢Y. Now the normal bundle N to T⁢X↪T⁢Y is complex. As before, we obtain a map i!:K(TX)→K(TY).

One needs to check that

  • •

    This construction is well-defined.

  • •

    X=X induces the identity.

  • •

    If we have X⁢↪𝑖⁢Y⁢↪𝑗⁢Z, then (j∘i)!=j!∘i! (functoriality).

Example 3.5.1.

Consider j:pt→ℝn. In this case, T⁡(pt)=pt, so N=ℝn. Then j!:K(pt)→K(Tℝn)=K(ℝ2⁢n) is the Thom isomorphism.

Definition 3.5.2.

The map B:K⁡(T⁢X)→ℤ is the composition

for any inclusion i:X⊂ℝn.

Remark 3.5.3.

This does not depend on i:X⊂ℝn.

Here are some obvious properties.

  1. 1.

    If X is a point, then B⁡(a)=dima.

  2. 2.

    If i:X⊂Y is an inclusion of compact smooth manifolds, then B(i!a)=B(a) for all a∈K⁡(T⁢X).

Proposition 3.5.4.
  • •

    There exists a unique homomorphism B:K⁡(T⁢X)→ℤ satisfying the above properties.

  • •

    Furthermore, for all a∈K⁡(T⁢X), B⁡(a)=(−1)n⁢ch⁢(a)⁢td⁢(T⁢X⊗ℂ)⁢[T⁢X].

Given this, all that is left is to prove that the diagram

commutes. This is the hard part of the proof. The basic strategy is to find τ:K⁡(T⁢X)→{elliptic operators} which is the inverse of σ to index, where ind⁡(τ⁢σ⁢D)=ind⁡(D). Then we need to check that K⁡(T⁢X)→ind∘τℤ satisfies the two axioms for B. The problem with this strategy is that such a τ does not exist because σ is not surjective.

Proof of proposition.

Suppose B′ is another map satisfying the properties. Consider X⊂ℝk⊂Sk Then the diagram

commutes. Then the desired result follows from staring at the diagram.

For the second part, consider Xn⊆N⊂ℝk. Here, N→X is a rank r≔k−n-bundle. Thus T⁢N→T⁢X is a rank r complex bundle. Now we have a chain of morphisms

Now we have b≔B⁡(a)=(ϕ′)−1⁢(h⁡(ϕ⁡(a))), and thus h⁡(ϕ⁡(a))=ϕ′⁢(b)=b⋅λk. If we take the Chern character of everything, we obtain the commutative diagram

Now on the right, we have ch⁡(b⁢λk)=(−1)k⁢b⁢σk, and therefore

by the commutativity of the middle square. Now

where ψ is the Thom isomorphism in cohomology. Notice that T⁢X⊕N=ℝk, so T⁢(T⁢X)⊕T⁢N→T⁢X is trivial. Therefore td⁡(T⁢N)⋅td⁡(π∗⁢(T⁢X⊗ℂ))=1, and thus we have

Substituting this in b=(−1)k⁢ch⁢(ϕ⁡(a))⁢[T⁢N], we obtain the desired result. □

Now we will see why σ is not surjective and how to fix this. There are topological obstructions. For example, consider X=S1=ℝ/2⁢π⁢ℤ. Then T⁢X=S1×ℝ is a cylinder. We will choose coordinates x on S1 and ξ on ℝ. Now choose a∈K⁡(T⁢S1). We obtain E0→α⁡(x,ξ)E1 with supp⁡α compact. Now if we fixe trivializations E0=E1=ℂn, then for ξ≫0, the map α⁡(−,ξ):S1→G⁢Ln⁢(ℂ). Then det⁡α⁡(−,ξ):S1→ℂ×. This only depends (up to homotopy) on the sign of ξ. Now S1→ℂ× has a well-defined degree, so we have

Now if a=σ⁡(D), then α⁡(x,ξ)=∑|β|=raβ⁢(x)⁢ξβ, and thus α⁡(x,−ξ)=±α⁡(x,ξ), so deg⁡a=0. On the other hand, we note that if

then deg⁡(a)=1, so it cannot come from a differential operator. This tells us that elliptic differential operators are not the right framework to implement our strategy because they are too rigid. Recall that a differential operator on ℝn of order r is a linear map

Here we will set

To define a pseudodifferential operator, we will create a much more complicated definition. Recall the Fourier transform

Then recall that Di⁢û⁢(ξ)=ξi⁢û⁢(ξ), so if we set p⁡(x,ξ)=∑aα⁢(x)⁢ξα, then we have

Using Fourier inversion, we obtain

Definition 3.5.5.

An operator P:Cc∞⁢(ℝn)→C∞⁢(ℝn) is called a pseudodifferential operator of order r if there exists some p⁡(x,ξ) such that |Dxβ⁢Dξα⁢p⁢(x,ξ)|≤Cα,β⁢(1+|ξ|)r−|α| uniformly in compact subsets in ℝn such that

Also, for all x and all ξ≠0, the principal symbol

exists.

Definition 3.5.6.

A pseudodifferential operator P is elliptic if σP⁢(x,ξ) is an isomorphism for all ξ≠0.

Warning 3.5.7.

Pseudodifferential operators are not local. This means that (P⁢u)⁢(x) is not determined by u in a neighborhood of x. Thus it takes some work to define pseudodifferential operators on manifolds.

However, once we have such a definition, we have the following properties:

  • •

    If P is an elliptic pseudodifferential operator of order r, then P:Lk+r2→Lk2 is Fredholm.

  • •

    The index of P only depends on σP⁢(x,ξ) restricted to the unit tangent bundle. This means the index does not depend on the order.

Example 3.5.8.

Consider L2⁢(S1)=ℓ2⁢(ℤ). Now consider the operator P given by

is a pseudodifferential operator of order 0. Here, σP⁢(x,ξ)=1 if ξ>0 and σP⁢(x,ξ)=0 if ξ<0. For the order 1 version of this, choose

Example 3.5.9.

Set

This defines an order 1 elliptic pseudodifferential operator. The order 0 version of this is given by

Here, ind⁡P=−1.

Now if we consider the diagram

we have now defined τ:K⁡(T⁢X)→{pseudodifferential operators}. Now we want to show that ind⁡(t⁡(a))=B⁡(a). We need to show that ind∘τ satisfies the two axioms. For X=pt, clearly ind∘τ gives the dimension, but the hard part is functoriality. The complication is that we need to work in an equivariant setting. Historically, what Atiyah and Singer did was prove an equivariant version of the index theorem directly.

Some interesting consequences of the equivariant version of the index theorem are fixed point formulas. For example, we obtain the following statement:

Proposition 3.5.10.

Suppose X is a compact complex manifold and G acts on X holomorphically. Then g∈G gives an operator on Hi⁢(X,𝒪X), and consider the quantity

called the holomorphic Lefschetz number. Then if all fixed points of g are isolated,

where p ranges over the fixed points of g.