Lectures on Microlocal Analysis

18.157, Spring 2022

 Contents

Versions 6
Outline and Practicalities 7

Chapter 1. Pseudodifferential operators, manifolds and compactification 9

1. Manifolds with corners 13
2. Compactification 17
3. Collar neighbourhood 21

Chapter 2. Symbols and conormal distributions at a point 25

1. Schwartz kernels 25
2. Homogeneous distributions 29
3. Topology and asymptotic summation 32
4. Integration 34
5. Wavefrontset 34
6. Restriction 35
7. Multiplicativity 36
8. Asymptotic completeness 37
9. ISm (Rn ; {0}) as a module 39
10. Action of Ψ∗ on I ∗ 40
11. Radial compactification and symbols 41

Chapter 3. The ring Ψ∗ (Rn ) 45

1. Coordinate invariance of Icm (Rn ; {0}) 45
2. Left/right invariance 47
3. Isotropic algebra 53
4. L2 boundedness 55

Chapter 4. Ellipticity and wavefront set 59

1. Ellipticity of symbols 60
2. Ellipticity of pseudodifferential operators 62
3. Wavefront set of a distribution 64

Chapter 5. Propagation of singularities 69

1. Hamiltonian mechanics 69
2. Hörmander’s Theorem 70
3. The Hamilton vector field 71
4. Construction of symbols 72
5. Proof of regularity 73
6. A question about the wave equation 75

Chapter 6. Smoothing operators and K-theory 77

3
4 CONTENTS

1. Hilbert space and operators 77

2. Schwartz smoothing algebra 79
3. Odd K-theory 81
4. The involutive Grassmannian 82
5. Semiclassical smoothing operators 86
6. The group G∞ sl 89
7. Constructing the Bott element 92
8. Quantization of the Bott element 93
9. Simplifying the quantized Bott operator 95
10. Surjectivity of the periodicity map 96
11. Manifolds with boundary 96
12. Chern character-building 97
13. Isotropic K-pop 99
14. Bott and Clifford 100

Chapter 7. Operators on manifolds 103

1. Functions and densities 103
2. Distributions 104
3. Vector bundles 105
4. Operators and kernels 106
5. Smoothing operators 107
6. Cornormal distributions at the zero section 107
7. Conormal distributions at a submanifold 110
8. Pseudodifferential operators 110
9. Elliptic operators 110
10. Self-adjoint operators 114
11. Fibre bundles 115
12. Families of pseudodifferential operators 117
13. Families index theorem 119
14. Spin and Dirac 120
15. Gerbes? 122

Chapter 8. Semiclassical quantization 125

1. Blow up 125
2. Semiclassical smoothing operators 127
3. Pull-back and push-forward 128
4. Abstract product theorem 130
5. Semiclassical pseudodifferential operators 132
6. Semiclassical index map 133
7. The Atiyah-Singer index theorem 133
8. Index formula 135
9. The Dirac case 135

Chapter 9. Manifolds with boundary 137

1. The b-generalized products 139
2. Conormality at the boundary 142
3. The b-calculus 143
4. Metrics and boundaries 148
5. Hodge theorems 149
 CONTENTS 5

6. Ellipticity and parametrices 151

7. Hodge theorem for b-metrics 154
8. Hodge theorem for scattering metrics 155
9. Atiyah-Patodi-Singer index theorem 156
10. Spectral and scattering theory 158
11. What else? 158

Bibliography 159
6 CONTENTS

Versions
v1 30 November, 2021: L1 and compactification
v2 31 January, 2022: L2 started
v2b 4 February, 2022: Some corrections from Paige
v3 7 February, 2022: L3
v3a 8 February, 2022: L3 reorganized, correction from Zach
v3b 9 February, 2022: Asymptotic completeness added to L3
v3c 9 February, 2022: Finished L3
v4 14 February, 2022: Problems 1 included, part of L5
v5 16 February, 2022: L6
v5a 16 February, 2022: More of L6
v6 17 February, 2022: Corrections from Paige, L6
v7 19 February, 2022: Most of L7
v7a 19 February, 2022: More on L7
v7b 20 February, 2022: First version of Problems 2
v7c 22 February, 2022: More handwritten notes
v7d 24 February, 2022: Corrections from Paige
v8 26 February, 2022: L8
v9 2 March, 2022: L9
v9a 2 March, 2022: More on L9, start of L10, corrections from Paige, precision
in section on Elliptic symbols
v9b 4 March, 2022: Corrections, + version fixed
v9c 5 March, 2022: Revised propagation proof
v10 6 March, 2022: Smoothing operators for L10
v10a 7 March, 2022: Corrections
v10b 8 March, 2022: Bott periodicity statement
v11 9 March, 2022: L11
v11a 9 March, 2022: Minor
v12 9 March, 2022: L12 started
v12a 11 March, 2022: Collar nhbd
v12b 11 March, 2022: Isotropic section and corrections from Paige
v12c 14 March, 2022: Bott element
v13 15 March, 2022: L12/13
v13a 19 March, 2022: Hilbert space discussion added
v13b 20 March, 2022: Unipotent Grassmannian
v14 21 March, 2022: L14
v15 28 March, 2022: L15, corrections from Paige, Problems 3
v15a 28 March, 2022: Corrections K-theory
v15b 1 April, 2022: Corrections from Paige to Chapter 6
v16 6 April, 2022: Most of L17, L16 still to come
v17 7 April, 2022: Outline to end
v18 9 April, 2022: Most of L18
v18a 12 April, 2022: L18
v19 18 April, 2022: L19 in part
v19a 18 April, 2022: more of L19
v20 20 April, 2022: L20/21
v21 21 April, 2022: L21
v22 22 April, 2022: L22 – outline of proof of index theorem
 OUTLINE AND PRACTICALITIES 7

v22a 26 April, 2022: More for L22.

v23 30 April, 2022: Much of L24.
v24 1 May, 2022: More of L24.
v24a 2 May, 2022: L24/L25
v25 2 May, 2022: L25
v25a 3 May, 2022: L25 revisions
v26 8 May, 2022: L26
v26a 10 May, 2022: more of L26
R1 11 May 2022, Chapter 1 revised.
R2 1 June 2022, Major changes, still much to do!
C.OpMa
R3 6 June 2022, Mostly Chapter 7

Outline and Practicalities

[Revised: 24 January, 2022.]
In this course I hope to cover four (types of) theorems which involve microlocal
analysis and in particular the theory of pseudodifferential operators. Namely
(1) Hörmander’s theorem on the propagation of singularities
(2) Weyl’s law for the distribution of eigenvalues
(3) The Atiyah-Singer index theorem and K-theory
(4) Hodge theory and boundaries
As a first step I will proceed to discuss the algebras of pseudodifferential op-
erators on Euclidean space and on a compact manifold and then similar algebras
(and related modules) on manifolds with boundary and for fibrations and more
157.74 (0.1) Ψ∗ (Rn ), Ψ∗ (M ), Ψ∗∗ (M )
where the upper star is an order and the lower star is some sort of structural
information.
To me the four results listed above are fundamental, and I like them! The
first two are relatively closely related and both give realization of the ‘semiclassical
limit’, the interplay between the non-commutative theory of (pseudo-)differential
operators and the more familiar behaviour of analysis of functions. The latter two
are more global but both involve the essential invertibility of (pseudo-)differential
operators.
Let me briefly indicate what these theorems are about.
Hörmander’s theorem on the propagation of singularities is a precise version,
and massive generalization, of ‘Huyghen’s Principle’. The latter describes the
spreading of the singular edge of solutions of the wave equation. The precise version
is one of the consequences of ‘microlocalization’, transferring analysis from ‘space’
to ‘phase space’ interpreted concretely as a manifold and its cotangent bundle re-
spectively.
Weyl’s asymptotic formula describes, at ‘high energy’, the number of eigen-
values of a self-adjoint elliptic operator, on a compact manifold, in terms of the
volume inside the energy surface in the cotangent bundle. The original theorem
was actually about the eigenvalues of the Dirichlet problem on a domain in R2 .
Elliptic (pseudo-)differential operators on a compact manifold are Fredholm –
they are invertible modulo finite dimensional null space and complement of the
range. The index, the difference of these two dimensions, is a very stable number
in the sense that it only depends on the ‘topology’ defined by the leading part of
8 CONTENTS

the operator and the theorem gives a formula for it. One classical version of this
is the Riemann-Roch theorem for the ∂ operator on (line bundles over) a compact
Riemann surface. This already requires some effort to understand! There is a
one-dimensional real version of the theorem, due to Toeplitz, which states that the
index of an elliptic Toeplitz operator on the circle (the projection onto the Hardy
space, consisting of the functions smooth on and holomorphic on the interior of the
disk, of multiplication by a non-vanishing smooth function) is equal to (minus) the
winding number of the function.
You probably do know the Hodge theorem for a compact manifold without
boundary as the identification of the deRham cohomology with the space of har-
monic forms. For non-compact manifolds there is no simple generalization, rather
there are many corresponding to structures ‘at infinity’ (meaning near the bound-
ary).
Clearly, each of these theorems could easily expand to take the whole semes-
ter. Still I hope to show how they can be approached using pseudodifferential
operators and ‘quantization’. In fact an alternative title for this course might be
‘Smooth quantization’. So most of the time will be devoted to preparing the back-
ground material, specifically pseudodifferential operators on Rn , pseudodifferential
operators on a manifold, families of pseudodifferential operators and then rings of
pseudodifferential operators quantizing a Lie algebroid.
I plan to give 26 one-hour lectures in the 9:30-10:30 slot on Tuesdays and Thurs-
days and leave 20 minutes for questions and discussions (even short presentations
by students); if there is sufficient interest I will organize another ‘discussion’ time,
perhaps on Wednesdays in the afternoon. There will be notes for each topic (the
precise correspondence to the individual lectures will depend on various things),
which will include topics I will not have time to cover and will certainly include
further references – to books, lecture notes and papers. With any luck at least some
of the lectures should appear on my webpage before the beginning of the semester.
Problem sets: There will be approximately 5, every two weeks. Grading may
be by discussion with me.
Grades: Graduate students are expected to participate actively. That is what
‘A’ means to me. By this I mean that I expect people to attend lectures and to ask
questions. For undergraduates this course might be heavy lifting, it is for me, so
please talk to me by early in the semester at the latest. We can discuss what you
should expect. There are no exams.
Prerequisites: I will assume familiarity with manifolds and distributions, essen-
tially as in 18.155 but plan to review pretty much everything.
Why don’t I just follow a book or my earlier lecture notes? This probably
reflects some personal failing and general dissatisfaction with how things are done!
I find it difficult to think through things without seeing some other way of ap-
proaching them. If it is not to your taste, I am sorry but that is the way it is. I
may not get to all the results listed above, but I expect to at least get to the point
where they are all within reach and that is really what I want to do – try to put
these results in a general context that maybe encourages them to be exploited (i.e.
applied) and extended.
In the interim, feel free to contact me with questions or comments.
Richard Melrose, 17 November, 2021.
 CHAPTER 1

Pseudodifferential operators, manifolds and

compactification
L1
The main aim of this course is to describe various algebras of pseudodiffer-
ential operators. Let me start with a traditional ‘crypto-historical’ description of
the ‘standard’ algebra of pseudodifferential operators on Rn . I recall notation for
functions below. Let’s assume you know about the spaces of smooth functions
on Euclidean space and the successively larger subspaces of compactly supported
functions, Schwartz functions and of functions with all derivatives bounded
157.1 (1.1) C c∞ (Rn ) ⊂ S (Rn ) ⊂ C ∞
∞
(Rn ) ⊂ C ∞ (Rn )
maybe including their topologies and duals.
For any multiindex α ∈ Nn0 , N0 = {0, 1, 2, . . . } being the non-negative integers,
the corresponding iterated partial derivative acts on each of these spaces
∂ α1 ∂ αn
157.2 (1.2) u 7−→ Dα u, Dα u(x) = i−|α| ... . . . u(x), |α| = α1 + . . . αn
∂x1 ∂xn
where the normalizing power of i is inserted to help with notation for the Fourier
transform.
These generate the commutative ring of differential operators with constant
coefficients with general element
X
157.42 (1.3) p(D) = cα Dα , cα ∈ C.
|α|≤m

This is a filtered ring which is isomorphic

157.1
to the ring of polynomials in n variables.
Similarly, each of the spaces in (1.1) is a ring, so multiplication of functions is
defined. Combining these we consider linear partial differential operators which are
given by sums
X
157.3 (1.4) P (x, D)u = pα (x)Dα u.
|α|≤m
157.1
In each case, when the coefficients are in one of the spaces (1.1), we get an operator
– a continuous linear map – on the corresponding space.
Whilst this is probably very familiar, and the operator product is given explicitly
by Leinbiz’ formula, it is very significant that these form a ring (and algebra) with
product
157.43 (1.5) X X
P (x, D)Q(x, D) = pα (x)(Dxγ qβ (x))Dα+β−γ , Q(x, D) = qβ (x)Dβ .
γ≤α,β |β|≤m0
157.3
It is worth thinking a little more about what is going on here. First note that (1.4)
157.42
is not as ‘natural’ as (1.3) in so far as we have chosen to write the ‘coefficients’,
9
 10 1. PSEUDODIFFERENTIAL OPERATORS, MANIFOLDS AND COMPACTIFICATION

157.42
the function pα (x), on the left. This is true in (1.3) as well but there the constants
commute with the 157.43
differentiation operators. Of course this is reflected in the fact
that the product (1.5) is not commutative.
Now, let’s concentrate Schwartz space. For this we have the Fourier transform
Z
157.4 (1.6) F : S (Rn ) −→ S (Rn ), F u(ξ) = û(ξ) = e−ix·ξ u(x)dx.
Rn
It is a linear isomorphism. We know that
157.5 (1.7) u ∈ S (Rn ) =⇒ F (Dα u)(ξ) = ξ α û(ξ).
The Fourier transform conjugates differentiation to multiplication. Whilst a mono-
mial such as ξ α is not in the Schartz space, but it does define an operator on it by
multiplication.
So the inverse Fourier transform, u(x) = (2π)−n eix·ξ û(ξ), allows us to write
R
Z
157.6 (1.8) Dα u(x) = (2π)−n eix·ξ ξ α û(ξ)dξ.
Rn
157.3
A linear partial differential
157.6 157.3
operator, (1.4), is given by a finite sum so we can
combine (1.8) with (1.4) and write
Z X
−n
157.7 (1.9) P u(x) = (2π) p(x, ξ)û(ξ)dξ, p(x, ξ) = pα (x)ξ α dξ.
Rn |α|≤m
n
Since û ∈ S (R ), the integral converges absolutely. If we just assume that the
coefficients are in C ∞ (Rn ) then the integral converges uniformly on compact subsets
in x ∈ Rn , with all its formal derivatives in x because of the obvious estimates
157.8 (1.10) |Dxγ p(x, ξ)| ≤ CK,γ (1 + |ξ|)m , x ∈ K b Rn , ξ ∈ Rn .
We can actually define the ‘standard’ space of pseudodifferential operators of
order m ∈ R by considering those functions a ∈ C ∞ (Rnx × Rnξ ) which satisfy the
symbol estimates
157.9 (1.11) |Dxγ Dξβ a(x, ξ)| ≤ Cβ,γ (1 + |ξ|)m−|β| , ∀ γ, β ∈ Nn0 .
157.7
Notice that p in (1.9) satisfies these estimates for an integer m if the coefficients
are in the space
∞
157.10 (1.12) C∞ (Rn ) = {f ∈ C ∞ (Rn ); sup |Dxγ f (x)| < ∞ ∀ γ},
consisting of the smooth functions with all derivatives
157.9
bounded.
m
The space of functions satifying estimates (1.11) is often written S1,0 as part
m
of a more general class of spaces Sρ,δ where the exponent m − |β| is replaced by
m − ρ|β| + δ|α|. Since we will not need the generalHormander-WeylCalculus
ρ, δ spaces – in fact there are
many variants of such estimates (see for instance [5]) – and we will already have
enough things to think about I will use157.9
the shorter notation, for the moment just
S m = S1,0
l
for the functions satisfying
157.9
(1.11).
It follows directly from (1.11) that if a ∈ S m then the direct generalization of
157.7
(1.9),
157.11 (1.13) Z
A(x, Dx )u = Au(x) = (2π)−n ∞
a(x, ξ)û(ξ)dξ =⇒ a : S (Rn ) −→ C ∞ (Rn ).
Rn
In fact much more is true:
 1. PSEUDODIFFERENTIAL OPERATORS, MANIFOLDS AND COMPACTIFICATION 11

157.12 Theorem∗ 1.1.157.91

The space of operators, Ψm (Rn ) = Ψm (Rn ) defined by sym-
bols, a, satisfying (1.11) act on S (Rn ) and form a filtered ∗-closed (∗ for adjoint
here) ring
0 0
157.13 (1.14) Ψm (Rn ) ◦ Ψm (Rn ) ⊂ Ψm+m (Rn ), ∀ m, m0 ∈ R.
iml MR1269107
This is the main content of the first chapter of [6], see alsoHormander-pseudo
Grigis-Sjöstrand [3].
Probably the first place this result appeared in this form is [4].
It is not that it is so hard to prove such a result, it is rather that I prefer to
approach it from a position of strength, so somewhat indirectly, in the sense that I
want to give a good deal of background before proving it.
Still2 it is important to see what is straightforward to prove and
what may157.11require some more thought. First let’s make sure we
do have (1.13).
157.11
Proof of (1.13). If u ∈ S (Rn ) then the product
\
a(x, ξ)û(ξ) ∈ S −∞ = SM 157.44 (1.15)
M ∈R
157.9
meaning that the estimates in (1.11) hold for all m. Indeed this
is just the product rule for differentiation. Written out fully in
terms of Leibniz’ formula
X β 
Dxγ Dξβ (a(x, ξ)û(ξ)) = Dxα Dξγ a(x, ξ) · Dξβ−γ û(ξ).
157.45 (1.16)
γ
γ≤β

Then one can apply the more obvious fact the product is rapidly
decaying in ξ :
S m · S (Rnξ ) ⊂ S m−k ∀ k ∈ R. 157.46 (1.17)
157.11
The integral (1.13) is therefore convergent. Again, if you
like to be precise, you can see that
Sm ⊂ C ∞
0
(Rn ; L1 (Rnξ )), m < −n 157.47 (1.18)
−n− 1 n
since (1 + |ξ|) ∈ L (R ) if  > 0. Now we can use standard
properties of Lebesgue (or improper Riemann) integrals to see
0
that Au ∈ C ∞ (Rn ) is a bounded continuous
157.11
function and the
same holds for all derivatives giving (1.13). 
We should check a couple of other statements, towards The-
157.12
orem 1.1. First the stronger mapping property that
A : S (Rn ) −→ S (Rn ). 157.48 (1.19)
This is a matter of getting ‘decay’. Namely we need to show
that for any monomial and any derivative
xγ Dxα Au ∈ C ∞
0
(Rn ). 157.49 (1.20)
We can approach this one step at a time, asking just about xj Au.
Note that we can certainly multiply by xj but the operator xj A
is not in general in Ψm (Rn ) (for any m) since xj a(x, ξ) is not
1The ∗ in the header of the theorem is to indicate that I will not prove it immediately but a
full proof, and more, will follow later.
2Narrowed parts of a lecture are things I don’t expect to have time to cover.
 12 1. PSEUDODIFFERENTIAL OPERATORS, MANIFOLDS AND COMPACTIFICATION

bounded
157.11
as |xj | → ∞ even for fixed ξ. However the integral in
(1.13) still converges rapidly in ξ for x in a compact set if we
replace a by xj a so
xj A : S (Rn ) −→ C ∞ (Rn ) 157.50 (1.21)
for instance.
157.50
157.51 Lemma 1.1. In the sense of operators (1.21)
[xj , A] = xj A − Axj ∈ Ψm−1 (Rn ). 157.52 (1.22)
Proof. We use ‘integration by parts’. Consider the opera-
tor Axj . The Fourier transform of xj u, u ∈ S (Rn ) is i∂ξj û so
Z
−n
Axj u = (2π) a(x, ξ)eix·ξ i∂ξj û(ξ)dξ = xj A(x, D)u + bj (x, D)u,

bj (x, ξ) = −i∂ξj a(x, ξ) ∈ S m−1 .

The rapid decay of a(x, ξ)û(ξ) in ξ means that
Z
∂ξj a(x, ξ)eix·ξ û(ξ) dξ = 0.


157.55 (1.23)

Proceeding by induction we conclude that

X
xγ A(x, D) = Bδ (x, D)xδ , Bδ (x, D) ∈ Ψm−|γ|+|φ| (R157.56
n
). (1.24)
δ≤γ


which I will denote Ψm
We will be most interested in a smaller space157.13 n
cl (R ) often
called the ring (with the composition property (1.14)) of ‘classical’ pseudodifferen-
tial operators where the symbols a have the additional property:
157.9
157.453 Definition 1.1. A symbol in the sense of (1.11) is classical (also ‘polyhomo-
geneous’) if there exists a sequence ai ∈ C ∞ (Rnx × (Rnξ \ {0})) of homogeneous
functions of degree m − i (in the ξ variables)
157.14 (1.25) ai (x, tξ) = tm−i a(x, ξ), t > 0, (x, ξ) ∈ Rn × (Rn \ {0})
such that for (any) cutoff χ ∈ C c∞ (Rnξ ) with χ = 1 near 0
N
X
157.15 (1.26) a(x, ξ) − (1 − χ(ξ))ai (x, ξ) ∈ S m−N −1 .
i=0
m
These ‘classical’ symbols form a filtered subring Scl
157.15
⊂ S m = S m . The relation-
ship (1.26) (see Problem set 1) is often written
X
157.33 (1.27) a' ai
i
and a is then said to have a complete asymptotic expansion. Such ‘asymptotic
summation’ (the existence of a given the ai ) is discussed below, it is closely related
to E. Borel’s ‘Lemma’ on Taylor series. There is no statement of convergence of the
157.15
series in (1.26) (although there is one lurking in the background) but you should be
157.15
able to see that the ai , assuming they exist are determined
157.11
by the relations ( 1.26).
When we insert such classical symbols in (1.13) (or if you prefer, restrict to
classical symbols) the resulting space of constitutes a filtered subring Ψm n
cl (R ) ⊂
 1. MANIFOLDS WITH CORNERS 13

Ψm (Rn ) which for positive integral m includes the differential operators of order m
discussed above.
These two rings have many important properties but one of the most important
157.15
is that one can recover the terms ai in (1.26) from the operator A and the leading
term defines the principal symbol, σm , as a map
157.16 (1.28)
Ψm (Rn )
σm
/ {a0 ∈ C ∞ (Rn × (Rn \ {0}) homogeneous of degree m in ξ}.
cl

and this map is surjective, multiplicative and defines a short exact sequence
0
σm+m0 (A ◦ B) = σm (A)σm0 (B), A ∈ Ψm n m n
cl (R ), B ∈ Ψcl (R )
157.17 (1.29) Ψm−1
cl (Rn ) ,→Ψm n
cl (R ) −→
{a0 ∈ C ∞ (Rn × (Rn \ {0})) homogeneous of degree m in ξ}
Here I have stuck with a cumbersome notation for the homogeneous space which
will be refined below. There are similar exact sequences for the larger algebra
Ψ∗ (Rn ) but the principal symbol lies in a quotient space.
So, we want to prove all these things and a lot more! However, I do not want to
go there directly but rather map out the territory a bit first, in particular discussing
the ‘symbol spaces’ concretely. L1-end

1. Manifolds with corners

This might appear to be a serious non-sequitor but I hope you will get used
to the idea of these sections on background material and see a bit later why I am
proceeding this way.
Both for ‘local’ analysis and the formulation of global results it is very conve-
nient to focus on manifolds with corners as our basic ‘category of spaces’ (which it
is as will be made precise later). There are several reasons to introduce these. An
immediate one is to understand the symbol spaces and their generalizations. This
I will get to next time. This allows me to introduce the spaces of conormal distri-
butions which arise as the Schwartz kernels of the pseudodifferential operators we
are interested in. Thinking about the kernels abstractly will allow us to generalize
readily later. This involves manifolds with boundary, but then products will get
you to manifolds with corners.
So, this is one of the basic settings for the course – analysis on manifolds with
corners – but only taken as far as we need for the moment. Let me start with an
explicit definition and then explain all the terms used in it. I’m assuming familiarity
with the standard definition of a manifold without boundary.
157.34 Definition 1.2. A manifold, M, is a metrizable, separable (so second count-
able) topological space with an open covering giving a (maximal) atlas of C ∞ -related
coordinate patches modeled on [0, ∞)n and with embedded boundary hypersurfaces.
I will not assume connectedness without explicitly saying so, but the definition
then requires all the components to have the same dimension.
So we are given a separable metric space, M, but the ‘metrizable’ means we
do not take the actual metric seriously, just the open sets it defines as the unions
of open balls. A coordinate patch in such a topological space is a triple (F, U, V )
consisting of a homeomorphism F : U −→ V of an open subset U ⊂ M onto a
(relatively) open subset V ⊂ [0, ∞)n . So this means there exists an open subset
 14 1. PSEUDODIFFERENTIAL OPERATORS, MANIFOLDS AND COMPACTIFICATION

V 0 ⊂ Rn such that the range V = V 0 ∩ [0, ∞)n . The coordinates on the coordinate
patch are the pull-backs of the coordinate functions xi on Rn .
To make clear what ‘C ∞ -related’ for two such coordinate patches means, we
need to define C ∞ (V ) (I will not bother with lower regularity than C ∞ ):

157.35 (1.30) C ∞ (V ) = {u : V −→ R( or C);

∃ V 0 ⊂ Rn open V = V 0 ∩ [0, ∞)n , u0 ∈ C ∞ (V 0 ) and u = u0 V
}.
So I am assuming you know about C ∞ (V 0 ) for open subsets of Rn .
Now the C ∞ -compatibility of two coordinate patches (Fi , Ui , Vi ), i = 1, 2, as
introduced above, means that either U1 ∩ U2 = ∅ or else the transition maps
F12 = F1 ◦ F2−1 : F2 (U1 ∩ U2 ) −→ F1 (U1 ∩ U2 ) and
157.36 (1.31)
F21 = F2 ◦ F1−1 : F1 (U1 ∩ U2 ) −→ F2 (U1 ∩ U2 )
are C ∞ in the sense that F12 ∗
: C ∞ (F1 (U1 ∩ U2 )) −→ C ∞ (F2 (U1 ∩ U2 )) and F21∗
:
∞ ∞
C (F2 (U1 ∩ U2 )) −→ C (F1 (U1 ∩ U2 )); this is equivalent to saying either pull-back
map is an isomorphism. This is also equivalent to saying that the pull-backs of the
coordinate functions, under either of the maps Fi , restrict to U1 ∩ U2 to be C ∞
functions of the other coordinates.
So now an atlas is a covering by such (pairwise) C ∞ -compatible coordinate
patches. If some coordinate patches are compatible with all the elements of an
atlas then the combined collection is still an atlas – they are necessarily compatible
amongst themselves as well. Hence any atlas is contained in a unique maximal atlas
– all this is as in the boundaryless case.
If we just stop at this point then M is what I call a tied manifold although
there is no general agreement on this. The missing point is the additional condition
that ‘boundary hypersurfaces are embedded’. A point in a coordinate patch is a
boundary point of codimension k if exactly k of the coordinate functions vanish on
it (note that coordinate patches map into [0, ∞)n so by fiat all coordinates are non-
negative – I will actually drop this requirement later but it makes things easier to
state initially). By considering the differential of the transition map it follows that
the codimension is well-defined at each point, it is independent of the coordinate
patch used. This means that M has a stratification, a decomposition into disjoint
pieces, based on the codimension
157.37 (1.32) M = M0 ∪ M1 ∪ · · · ∪ Mn
where the Mj can be empty (from some k > 0 onward). The points of boundary
codimension zero are the interior points of the manifold (there is a slight incon-
sistency between openness of subsets of [0, ∞)n and this, so the interior there is
(0, ∞)n , of course otherwise there would be no point in talking about the interior
of a relatively open subset).
Each Mj itself is a manifold without boundary and the closures of the compo-
nents of the Mj are called the boundary faces of codimension j; the set of these
boundary faces I will write as M j (M ). In particular the boundary faces of codimen-
sion one, the Hi ∈ M 1 (M ) are called the boundary hypersurfaces. The ‘boundary
hypersurfaces are embedded’ part of the definition is just the statement that the re-
strictions of the coordinate patches to each Hi given them C ∞ -compatible atlases.
One consequence of this is functorial, that the boundary hypersurfaces (and in
consequence all boundary faces) are themselves manifolds with corners. There are
 1. MANIFOLDS WITH CORNERS 15

several useful ways to restate this condition but note how it fails for a ‘tear-shaped
region’ in the plane.
The C ∞ functions on M are those that are C ∞ in each coordinate patch, mean-
ing
157.38 (1.33) f ∈ C ∞ (M ) ⇐⇒ (F −1 )∗ (f U
) ∈ C ∞ (V ) for each coordinate patch.
This is equivalent to the same condition for any one compatible atlas.
The direct consequence of the ‘embedded’ requirement is that the boundary
hypersurfaces have defining functions:
157.39 (1.34) Hi ∈ M 1 (M ) =⇒ ∃ ρi ∈ C ∞ (M ), ρi ≥ 0, Hi = {ρi = 0},
d((F −1 )∗ ρi )(F (p)) 6= 0 ∀ p ∈ Hi for all coordinate patches containing p.
This last condition means that for each p ∈ Hi there is a coordinate patch containing
p in which ρi is a coordinate function.
If M̃ is a manifold without boundary, i.e. M̃1 = ∅, then M ⊂ M̃ is a(n embed-
ded) submanifold if M has a covering by coordinate patches of M̃ which restrict to
give it the structure of a manifold with corners.
157.40 Theorem 1.2. For any manifold with corners there exists a manifold without
boundary M̃ of the same dimension in which M is embedded as a submanifold; if
M is compact then M̃ can be taken to be compact.
Although there is no quite canonical way of constructing such an extension, M̃ ,
all the standard constructions of the tangent, cotangent, form bundles and other
bundles associated to the frame bundle, pass over to the case of a manifold with
corners in such a way that the restrictions for an extension of this type are canonical
157.41 (1.35) T M = T M̃ M
, T ∗ M = T ∗ M̃ M
etc.
However, there are important additional structures which arise from the boundary
faces as I will discuss later.
Why work in this degree of generality? Manifolds with corners are the smooth
(i.e. C ∞ ) analogue of smooth algebraic varieties with divisors and they occur for
similar reasons. One place manifolds with corners arise is through ‘compactifica-
tion’.
Remark 1. From now on the default meaning for ‘manifold’ is a manifold with
corners.
As you are no doubt aware, it is bad form to specify a class of objects, in this
case manifolds without specifying the ‘morphisms’ between them. For manifolds
without boundary the morphisms, forming a category, are just the smooth maps.
This also works when there are corners but it is not a good idea. The problem is
that a smooth map F : M −→ N in general does not ‘respect’ the boundary. For
instance a smooth curve, so M = (0, 1) might have a single point of tangency to a
boundary hypersurface.
What we want are b-maps, where as always the b- just stands boundary. Each
boundary hypersurface is specified by the ideal of functions which vanish on it;
this is a principal ideal because we are assuming boundary hypersurfaces to be
embedded. So if ρH is a defining function for H ∈ M 1 (M ) then
157.965 (1.36) I (H) = ρH C ∞ (M ).
 16 1. PSEUDODIFFERENTIAL OPERATORS, MANIFOLDS AND COMPACTIFICATION

We do want our maps to be smooth of course, and then

157.966 (1.37) F ∗I (H) ⊂ C ∞ (M ).
157.967 Definition 1.3. A smooth map F : M −→ N between manifolds (with cor-
ners) is a b-map if for each H ∈ M 1 (N ) one of the following is true
(1)
157.968 (1.38) F ∗I (H) = {0}
(2) There is a map F# (H) : M 1 (N ) −→ N0 , not identically zero, such that
Y
157.969 (1.39) F ∗I (H) = I (K)F# (H,K) , F# (H, K) = F# (H)(K).
K∈M 1 (M )

(3)
157.970 (1.40) F ∗I (H) contains a non-vanishing element.
The third case corresponds to F# (H) as in the second, with all entries 0. This is
more evident if one chooses a defining functiona ρ0H for H and ρK for the boundary
hypersufaces of M and then the three cases correspond to F ∗ ρ0H = 0,
F (H,K)
Y
157.971 (1.41) F ∗ ρ0H = a ρK# ,
K∈M 1 (M )

and F ∗ ρ0H > 0.

157.972 Proposition 1.1. Manifolds with corners are the objects of a category with
b-maps as morphisms.
Just as for manifolds without boundary, the tangent bundle is a manifold well-
defined manifold with a smooth bundle projection
157.973 (1.42) π : T M −→ M.
157.974 Exercise 1. Check that π is a b-map!
Smooth vector fields, sections of T M, are important for several reasons, for one
they are natural operators
157.977 (1.43) V : C ∞ (M ) −→ C ∞ (M ), V (uv) = uV v + vV u.
Perhaps the most fundamental is that real vector fields can be integrated. The usual
local existence theorem fails on manifolds with corners as soon as the boundary is
non-trival – since the integral curves will often tend to ‘leave the manifold’. This
is one reason to consider the space
157.975 (1.44) Vb (M ) = {V ∈ C ∞ (M ; T M ); V is tangent to the boundary}.
157.977
Tangency to the boundary can be interpreted in terms of the action (1.43) and the
ideals I (H) for H ∈ M 1 (M ), as the statement
157.976 (1.45) V : I (H) −→ I (H) ∀ H ∈ M 1 (M ).
Tangency to the corners is then automatic.
We will consider Vb (M ) more closely below but for Vb (M ) the standard local
existence theorem holds.
 2. COMPACTIFICATION 17

157.978 Lemma 1.2. For a real-valued element V ∈ Vb (M ) and each p ∈ M there exists
a maximal integral curve

157.979 (1.46) χp : Ip −→ M, 0 ∈ Ip ⊂ R open and connected, χp (0) = p,

◦ d/dt
A(χp )∗ ( = V (χ(t)) with Ip maximal.
)

As usual, if M is compact then I = R always. In any case these integral curves

always define a smooth map on the open set
[
157.980 (1.47) F :O= (Ip × {p}) 3 (t, p) 7−→ χp (t) ∈ M.
p∈M

If M is compact this is a 1-parameter family of diffeomorphisms of M.

With a little work you can see that

157.982 Proposition 1.2. On a compact manifold the Lie algebra of the groupd of
diffeomorphisms is Vb (M ).

Note that this infinite-dimensional group is missing some properties of a Lie group,
for one the diffeomorphism obtained by integrating the Lie algebra do not contain
a neighbourhood of the identity. This can be obviated by considering parameter-
dependent vector fields.
If W is a t-depedent smooth vector field on M, so a vector field in Vb (U ) for an
157.978
open subset U ⊂ Rt × M such that V t = 0 then Lemma 1.2 yields a 1-parameter
family of diffeomorphisms, Gt defined on an appropriate open subset, and satisfying

d df
157.981 (1.48) (Gt )∗ f = (Gt )∗ (W f + ), f ∈ C ∞ (R × M ).
dt dt

This invariant formulation of the chain rule, and its generalizations, are used below
in various constructions.

2. Compactification
T1.Compactification
Although we will deal with non-compact manifolds, the ones that arise below
have some ‘structure at infinity’. One way to describe what this means is through
the notion of compactification.

157.18 Definition 1.4. A compactification of a manifold M is a compact manifold

M and a smooth injection ι : M −→ M which is a diffeomorphism to a (relatively
of course) open dense submanifold.

Here, both M and M may have corners. As always when introducing a new
notion, we should specifiy when two compactifications are to be regarded as ‘the
same’.
 18 1. PSEUDODIFFERENTIAL OPERATORS, MANIFOLDS AND COMPACTIFICATION

157.19 Definition 1.5. Two compactifications ιi : M −→ M i are equivalent if there

exists a diffeomorphism e : M 1 −→ M 2 giving a commutative diagramme

157.20 (1.49) =M 1
ι1

M e
ι2
! 
M 1.

Notice that the equivalence

157.20
map e is unique if it exists since it is fixed on an
open dense subset by (1.49). We also say that one compactification is finer than
another if there is a smooth map e giving a commutative digaramme; again it if it
exists it is determined. This defines a partial order on compactifications – as we
shall see below there can be non-comparable compactifications.
If M is compact it is a compactification of itself and it is unique in this sense
of equivalence.
We might well want more structure for the compactification – for instance if
M is a complex manifold then we might want M to be complex and all maps
to be holomorphic. There are important examples from algebraic geometry here.
Most relevant at the moment is the projective compactification of a complex vector
space W ,→ PW which I mention below but there are much more sophisticated
examples to check out. There is the Deligne-Mumford compactification of the
Riemann moduli spaces M g,n (okay I hear a complaint from someone that the
M g,n are not quite manifolds, they are orbifolds in general, but take the number
of punctures n large compared to the genus g ≥ 0). Also there is the deConcini-
deConcini-Procesi
Procesi ‘wonderful’ compactification ofCoSLG
complex adjoint Lie groups [2] (there is a
real version of this compactification in [1]). Also, compactificaiton of ‘Gravitational
Instantons’ (aren’t the Physicists good at inventing names!)
The examples I will consider immediately are more prosaic, namely of a real
finite-dimensional vector space V. This is both to illustrate the notion and for later
reference. I will discuss

o
(1) The one-point compactification(s) given by a sphere V .
p
(2) The parabolic compactification gives a closed ball V .
R
(3) The radial compactification also given by a closed ball V = V .
 2. COMPACTIFICATION 19

ι(t)

0
x

From the notation you can see that I have a preference for the radial compacti-
fication – I hope the discussion below shows why. Only the radial compactification
is really used subsequently.
These can all be constructed using variants of stereographic projection. So, let’s
start with V = Rn , i.e. choose a basis. We embed Rn into Rn+1 as the hyperplane
157.21 (1.50) Rn 3 x 7−→ (x, 1) ∈ P ⊂ Rn+1 .
In the first, case consider the the sphere So of radius 12 centred at (0, 12 ) and in the
second and third cases take the sphere SR of radius 1 centred at the origin. In both
these latter cases a point of Rn determines a unique line Lo (x) or LR (x) through
the image of x in P and the centre of the corresponding sphere then
157.22 (1.51)
Io : Rn −→ So , Io x is the other point in So ∩ Lo (x)
IR : Rn −→S+
R,
IR x is the other point in SR ∩ L1 (x) ⊂ S+
R = SR ∩ {xn+1 ≥ 0}
Ip : Rn −→Bp ⊂ Rn ,
Ip x is the projection of LR x onto the closed unit ball in Rn × {0}.

In all three cases the full orthogonal group O(n), acting on the first factor of Rn ×Rn
satisfies I• Ax = AL• x for all A ∈ O(n), effectifely reducing the discussion to the
case n = 1. Explicit formulæ for the maps are easily derived:
 
x 1
Io x = , ∈ So ⊂ Rn+1
+
1 + |x|2 1 + |x|2
 
x 1
157.23 (1.52) IR = 1 , 1 ∈ S+ n+1
R ⊂ R+
(1 + |x| ) 2 (1 + |x|2 ) 2
2
x n
Ip x = 1 ∈ R .
(1 + |x|2 ) 2
 20 1. PSEUDODIFFERENTIAL OPERATORS, MANIFOLDS AND COMPACTIFICATION

1
Thus, for the radial compactification (1 + |x|2 )− 2 is a boundary defining function
and hence |x|−1 , which is a smooth function of it away from x = 0, is a defining
function near the boundary. It follows that
 
1 x
157.27 (1.53) {|x| >  > 0} 3 x 7−→ , ∈ [0, 1) × Sn−1
|x| |x|
R
extends to a smooth product decomposition of Rn near the boundary. For the
parabolic compactification it follows similarly that
 
1 x
157.28 (1.54) {|x| >  > 0} 3 x 7−→ , ∈ [0, 1) × Sn−1
|x|2 |x|
is a product decomposition near the boundary.
It can be seen directly that
 
x
157.24 (1.55) Io = SIo where S : So \ {(0, 1), (0, 0)} −→ So \ {(0, 1), (0, 0)},
|x|2
with S(y, yn ) = (y, −yn + 1)
is equatorial reflection on So . 157.23
In all cases it is clear either geometrically, or from the forumlæ (1.52), that the
action of O(n) extends smoothly from Rn to the compactification. Similarly the
scaling action by R+ , with generator on Rn
X ∂
157.25 (1.56) xi
i
∂xi
157.24
extends smoothly. For the one-point compactification this follows from (1.55) and
in the other two cases
tx x 1
157.26 (1.57) lim 1 = 1 and lim 1 = 0.
|x|→∞ (1 + t|x|2 ) 2 2
(|x| ) 2 |x|→∞ (1 + t|x|2 ) 2
Thus in all cases the action of the conformal group O(n)×R+ extends smoothly
to the compactification.
157.29 Proposition 1.3. The action of the general linear group extends smoothly
from Rn to the radial and parabolic compactifications, but not to the one-point
compactification; the translation action of Rn extends smoothly to the radial and
the one-point compactifications, but not to the parabolic compactification and there
are smooth surjective maps, which are not diffeomorphisms, giving a commutative
diagramme
157.30 (1.58)
GL(n, R) n Rn

Sn+
RO

IR
y Ip %
Sn0 o / Bnp
Io
O(n) n (R+ × Rn ) Rn GL(n, R).
 3. COLLAR NEIGHBOURHOOD 21

Outline of proof. That the group actions extend as indicated follows by

noting that the Lie algebra of GL(n, R) consists of vector fields homogeneous of
degree 0 and similarly the translations are homogeneous of degree −1. Similar
arguments show that the groups shown are the maximal subgroups of GL(n, R)nRn
which extend to act smoothly on the one-point and parabolic compactifications. 

157.31 Corollary 1. The one-point compactification is defined for a vector space

with conformal-Euclidean structure, the radial comactification is well-defined for
an affine space and the parabolic compactification is well-defined for a vector space.
Both the radial and the parabolic compactifications have boundaryless vari-
ants, in which the bounding sphere is replaced by an embedded projective space
Sn−1 /± by doubling across the boundary. The apparent advantage of this smaller
compactification does not seem to be realized in practice.
157.32 Conjecture 1. The five compactifications are minimal in their respective cat-
egories (i.e. as manifolds with/without
157.30
boundary) among compactifications with the
invariance properties in (1.58).
Although, as noted above, it is the radial compactification which mostly appears
below, other variants are relevant. In particular none of these compactifications are
natural for products – the radial compactification of V1 × V2 is not ‘comparable’
to the products of the radial compactifications. Still, the relationship between the
radial compactification of the product of vector spaces and the product of the radial
compactifications is significant and will be examined later.

3. Collar neighbourhood
Sect.collar
157.447 Remark. Edited by Paige Dote
This theorem provides a rather precise description of a neighbourhood of a
closed embedded submanifold, Y ⊂ M where M is an n-dimensional manifold.
The usual proof exploits the geodesic flow for a metric on M, but here we give a
related approach using the notation of a radial vector field for Y. This is very closely
Sternberg
related to the linearization theorem of Sternberg ([?]). We also carry it out in the
context of manifolds with corners.
A closed p-submanifold (p- if for ‘product’ it really is the natural notion of a
submanifold in the case of a manifold with corners) is a closed subset Y ⊂ M such
that for each point y ∈ Y , there exists adapted coordinates based at y in M in
terms of which Y is linear. Precisely, in a neighbourhood Uy ⊂ M of y, there exists
coordinate functions xi ≥ 0, yj ∈ C ∞ (Uy ) for i = 1, . . . , k and j = 1, . . . , n − k with
independent differentials and
157.448 (1.59) Y ∩ Uy = {x1 = · · · = xd0 = 0, y1 = · · · = yn0 −d0 ==}.
It follows that Y is itself a submanifold, with these as adapted coordinates for Y.
Note that if d0 > 0 here then Y is actually contained in a boundary face (assum-
ing it is connected, in any case one component is). Boundary faces themselves are
clearly p-submanifolds but quite a few natural submanifolds are not. For instance
the diagonal in M 2 , where ∂M 6= ∅, is not a p-submanifold. This turns out to be
rather significant.
We note the following
 22 1. PSEUDODIFFERENTIAL OPERATORS, MANIFOLDS AND COMPACTIFICATION

Definition 1.6. A smooth vector field, V ∈ Vb (M ) is tangent to Y if V f Y

=0
for any f ∈ C ∞ (M ) with f Y = 0. A smooth vector field vanishes on Y if V f Y
=0
for all f ∈ C ∞ (M ).
In adapted coordinates for Y , V is tangent to Y if it takes the form
d
X n−d
X
V = xi ai,j ∂xj + bk ∂yk
i,j=1 k=1

where ai,j , bk are smooth coefficients.

Similarly, V vanishes on Y if
d
X d n−d
X X d
X
157.449 (1.60) V = xi ai,j ∂xj + xi bi,k ∂yj = xi Wi
i,j=1 i=1 k=1 i=1

Where the Wi are smooth as are the ai,j , bi,k .

Definition 1.7. The differentials of functions which vanish on Y define the
conormal bundle N ∗ Y ⊂ TY∗ M .
Hence, in adapted coordinates N ∗ Y has the basis
dx1 , . . . , dxd .
For each y ∈ Y, is the annihilator, in Ty∗ M, of Ty Y ⊂ Ty M. By duality,
Ny∗ Y
the normal bundle N Y = TY M/T Y is spanned by
∂x1 , . . . , ∂xd .
Theorem 1.3 (Collar, or normal, neighbourhood). For a closed embedded sub-
manifold Y ⊂ M there is a neighbourhood, T , of the zero section ON ⊂ N Y ,
a neighbourhood Ω ⊂ M of Y , and a diffeomorphism F : T → Ω satisfying the
additional conditions
I1 (1) F (0N ) = Y is the natural identification and
I2 (2) F∗ : N (0N ) → N Y is the identity.
Here, the bundle projectionI1
from N Y to Y restricts to a diffeomorphism π :
ON → Y giving meaning to (1). Then, the tangent space to 0N is mapped to T Y
as the identity. It follows that the normal bundle to 0N in N Y is mapped to the
normal bundle N Y by F. Additionally, for any vector bundle, the normal bundle to
the zero section is naturally identified with the bundle itself, I2
so F∗ lifts to a bundle
map for N Y to N Y which is required to be the identity in (2).
Adapted coordinates give such a collar neighbourhood locally, so the challenge
is to make this global.
A vector field on M which vanishes on Y induces a linear map at each point of
Y:
L(V, y) : Ny∗ Y → Ny∗ Y
through
N (V, y)ξ = d(V f )(y)
∞
157.449
where, for ξ ∈ N Y, f ∈ C (M ) has f Y = 0 and df (y) = ξ. In terms of (1.60),
d
X
N (V, y)dxj = ai,j (y)dxi .
i=1
 3. COLLAR NEIGHBOURHOOD 23

Definition 1.8. A smooth (real) vector field, R, on M is radial on Y if it

vanishes at Y and
L(V, y) = Id
for all y ∈ Y .
Lemma 1.3. There is a radial vector field for any closed embedded submanifold,
P
and the difference between two such radial vector fields is locally a sum xi Wi
where the Wi are tangent to Y .
The existence of a radial vector field follows from a ‘patching’ argument. Lo-
cally, in adapted coordinates, we have an obvious radial vector field in the Euler
field
X d
R0 = xi ∂xi .
i=1
which is the generator of the scaling action
157.961 (1.61) (x, y) 7−→ (tx, y), t > 0.
There is no constraint on a radial vector field away from Y at all. So one can
take functions ϕa ∈ C c∞ (M ) with locally finite supports
P such that each supp ϕa is
contained in an adapted coordinate patch for Y and ϕa = 1 in a neighbourhood
of Y. Then,
X
R= ϕa R0,a
a
is radial with the R0,a being the coordinate Euler vector fields.
As a smooth real vector field, there is a unique integral curve of R through
each point of M. For R0 these are curves with xi = et xi and yi = yi so that xi ↓ 0
as t → −∞. For a general radial vector field the same is true near Y in the sense
that
Lemma 1.4. If R is a radial vector field for Y then Y has an open neighbourhood
such that the integral curves of R, as t → −∞, approach R smoothly in et with a
non-vanishing limiting tangent vector at Y.
As remarked at the beginning, this is a linearization theorem in the spirit of Stern-
berg.
Proposition 1.4. Near any point of Y there are adapted coordinates in terms
of which in which a given radial vector is the Euler vector field.
Proof. This can be seen using the homotopy method of Möser. In any adapted
coordinates, the radial vector field has the form
X X
R = R0 + xi xj ai,j,p ∂xp + xi bi,k ∂yk
i,j,p i,k

for smooth coefficients ai,j,p and b . Pushing R formward under the inverse of the
157.961 i,k
scaling diffeomorphism (1.61), we obtain the one parameter family of vector fields
X X
157.450 (1.62) Rt = R0 + txi xj ai,j,p (tx)∂xp + txi bi,k (tx, y)∂yk .
i,j,p i,j
 24 1. PSEUDODIFFERENTIAL OPERATORS, MANIFOLDS AND COMPACTIFICATION

This family is smooth down to t = 0, and Rt t=0 = R0 . Since this is ’exponential

scaling’ in terms of R0 , we see that
d
157.452 (1.63) t Rt = [R0 , Rt ],
dt
which can easily be checked directly. The t-dependent vector field
1
157.451 (1.64) Wt = (Rt − R0 )
t
is smooth and vanishes at Y . It follows that integration of Wt defines a 1-parameter
family of local diffeomorphisms, Gt , defined in a neighbourhood U of the point in
Y for all t ∈ [0, 1] fixing each point of Y. Thus G0 = Id and
d ∗
G f = G∗t (Wt f ),
dt t
for all f ∈ C ∞ (U ).
The standard variation formula (really the chain rule) shows that
 
d dRt
(Gt )∗ (Rt ) = (Gt )∗ + [Wt , Rt ] .
dt dt
157.451 157.452
From (1.64), it follows that [Wt , Rt ] = − 1t [R0 , Rt ]. So by (1.63),
dRt
+ [Wt , Rt ] = 0.
dt
Thus, (Gt )∗ Rt is constant in t as a vector field near Y , but by assumption (G0 )∗ R0 =
R0 so
(G1 )∗ R = R0 .
This gives a diffeomorphism locally fixing Y which reduces R to the Euler vector
field. 
The preceding lemma is an immediate consequence since the inegral curves of
R0 are of the form x 7→ tx and y = y.
This shows the existence of a diffeomorphism as required for the Collar Neigh-
bourhood theorem, with the inverse mapping p ∈ U to the tangent vector at the
end point.
 CHAPTER 2

Symbols and conormal distributions at a point

L2
1. Schwartz kernels
Before tackling the properties of the ring Ψ∗ (Rn ) of pseudodifferential operators
on Rn , I want to look into the properties of the Schwartz kernels of these operators,
so we can get a picture of them. We can ‘guess’ (it is easy to justify) that the
Schwartz
157.9
kernel of an operator A ∈ Ψm (Rn ), defined by a symbol a satisfying
(1.11), is
Z
−n
157.94 (2.1) A(x, y) = (2π) a(x, ξ)ei(x−y)·ξ dξ.
Rn

Here I use the same letter for the operator and its Schwartz kernel – since the
Schwartz kernel theorem (which I will talk a little about later) shows that they
determine each other. 157.94
We can think of (2.1) in a couple of different ways – in general it is not a
convergent integral. We can make a (formal at this stage) linear change of variables
on R2n from (x, y) to (x, z), z = x − y and then
Z
−n
157.95 (2.2) A(x, y) = α(x, x − y) where α(x, z) = (2π) a(x, ξ)eiz·ξ dξ.
Rn

Now the integral is a partial inverse Fourier transform.

157.95
In fact, since a is smooth
in x we can interpret the definition of α in (2.2) as the inverse Fourier tranform
from ξ to z for each fixed x. This in fact is what I will do today. Alternatively one
can just check that the partial Fourier transform with respect to a decomposition
of Euclidean space into a product behaves ‘correctly’.
So, for the moment, we have dispensed with the ‘coefficients’ and just look
at the (commutative) algebra of constant-coefficient pseudodifferential operators
where the composition operation is convolution.
Recall the convolution of distributions on Rn . On cannot define the convolution
of arbitrary distributions, even arbitrary tempered distributions – this however is
an issue of ‘growth’ rather than singularities. In particular the convolution
157.57 (2.3) u ∗ v is defined if either u or v has compact support
but can be defined in other cases too. I will denote the space of distributions of
compact support as
0
157.61 (2.4) C c−∞ (Rn ) = (C ∞ (Rn )) .
So the space of distributions of compact support is actually a commutative
157.57
ring, since u ∗ v = v ∗ u and the support of a convolution as in (2.3) satisfies
157.58 (2.5) supp(u ∗ v) ⊂ supp(u) + supp(v).
25
 26 2. SYMBOLS AND CONORMAL DISTRIBUTIONS AT A POINT

It is also the case that S (Rn ) is closed under convolution and we know that the
Fourier transform satisfies
157.59 (2.6) F (u ∗ v) = F (u)F (v), u, v ∈ S (Rn ).
The ring we are interested in is contained in
157.60 (2.7) C c−∞ (Rn ) + S (Rn )
157.59
for which the identity (2.6) still holds. Note that
157.62 (2.8) F (C c−∞ (Rn ) + S (Rn )) ⊂ C ∞ (Rn ) ∩ S 0 (Rn ).
So, we are looking for are some interesting spaces of smooth functions on the
dual Rn which are closed under multiplication. You might ask, in view of the
identification of the convolution kernels here with the inverse Fourier transforms of
symbols, why is there any problem at all? There isn’t a problem for convolution as
157.59
such because of (2.6) but recall that the Fourier transform does not ‘behave well’ on
say the space L∞ (Rn ). Of course the Fourier tranform maps this to a well-defined
linear subspace of the tempered distributions – which includes for instance the delta
functions at any point – but it is quite hard, in a certain sense I think impossible,
to give a ‘direct’ characterization of the Fourier image of L∞ and the same is true
for our symbols which are modeled on L∞ in the sense that they are defined by
bounds. We will in fact ‘sandwich’ the image between spaces characterized directly
(meaning without the Fourier tranform), but this still loses information which is
rather vital to us!
In the notes related to the first lecture, I discussed the radial compactification
of a real, finite-dimensional, vector space V, to a ball V . Ignoring all the niceties,
for Euclidean space, Rn with the standard Euclidean norm, we can identify the
complement of the origin with the product
 
x
157.63 (2.9) Rn \ {0} 3 x 7−→ |x|, = (r, ω) ∈ (0, ∞) × Sn−1 .
|x|
The inversion map r −→ 1/r is a diffeomorphism of (0, ∞) to itself ‘switching the
ends’. This allows us to add the sphere at infinity of Rn setting
Rn = Rn t [0, ∞) × Sn−1 /I


157.64 (2.10)
where
 
n 1 x
157.65 (2.11) I : R \ {0} 3 x 7−→ , ∈ (0, ∞) × Sn−1
|x| |x|
identifies the complement of the origin with the interior of the second part.
Thus Rn is a compact manifold with boundary ‘obtained by introducing in-
verted polar coordinates near infinity’. The interior is Rn and the boundary is ‘the
sphere at infinity’.
This immediately gives us a ring of functions on Rn , namely
157.66 (2.12) C ∞ (Rn ) ,→ C ∞ (Rn ).
I can write inclusion here for what is really the restriction from Rn to its interior
since this map is injective.
This is the space of ‘classical symbols on Rn of order zero’ which I will write as
157.67 (2.13) 0
Scl (Rn ) = C ∞ (Rn ).
I will approach the issue of characterizing this space precisely on Rn below.
 1. SCHWARTZ KERNELS 27

T1.Compactification
As a consequence of the discussion of radial compactification in § 2, or directly,
we can see that the coordinate vector fields on Rn extend to be smooth on Rn . In
fact
157.78 Proposition 2.1. The coordinate vector fields on Rn extend to smooth vector
files on Rn and span, over C ∞ (Rn ), all the smooth vector fields which are of the
form
157.79 (2.14) ρW, W smooth and tangent to the boundary of Rn .
Here ρ ∈ C ∞ (Rn ) vanishes at the boundary.
P-Comp
See Project 1
0
157.80 Corollary 2. The space Scl (Rn ) consists of smooth functions which satisfy
the estimates
157.81 (2.15) sup |(1 + |ξ|)|α| ∂ξα a(ξ)| < ∞ ∀ α.
ξ∈Rn

0
Note that I do not say that this characterizes Scl (Rn ) = C ∞ (Rn ), because it
does not.
157.83 Definition 2.1. We denote the subspace of C ∞ (Rn ) of functions satisfying all
157.81
the estimates (2.15) by
157.84 (2.16) S 0 (Rn ) ⊃ Scl
0
(Rn ).
These are the ‘symbols with bounds’ containing the classical symbols.
More generally, consider the function
157.68 (2.17) (1 + |x|2 )z/2 on Rn , z ∈ C.
This is certainly smooth on Rn . It is rather clear that
157.69 (2.18) (1 + |x|2 )z/2 ∈ C ∞ (Rn ) iff z ∈ −N0 .
Indeed, in x 6= 0 it can be written
157.70 (2.19) t−z (1 + t2 )z/2 , t = 1/|x|.
This is smooth down to t = 0, the boundary of Rn , if and only if −z is a non-negative
integer.
We define the space of classical symbols of (complex) order z to be the products
157.71 (2.20) z
Scl (Rn ) = (1 + |x|2 )z/2C ∞ (Rn ) = (1 + |x|2 )z/2 Scl
0
(Rn ).
The space of symbols (with bounds) or real order m is similarly defined to be
157.85 (2.21) S m (Rn ) = (1 + |x|2 )m/2 S 0 (Rn ).
Why no complex order in the second case?
157.83
157.86 Exercise 2. Show that in terms of Definition 2.1
157.87 (2.22) (1 + |x|2 )is/2 ∈ S 0 (Rn ) ∀ s ∈ R.
This in turn implies that
z
157.88 (2.23) Scl (Rn ) ⊂ S Re z (Rn ) ∀ z ∈ C.
 28 2. SYMBOLS AND CONORMAL DISTRIBUTIONS AT A POINT

157.72 Definition 2.2. The space of (Schwartz-) conormal distributions on Rn , with

respect to the origin is
m+n/4
157.73 (2.24) IS (Rn ) = F −1 (S m (Rn )) .
The corresponding spaces of classical (Schwartz-) conormal distributions at the
origin where now z is allowed to be complex, are
z+n/4
157.89 (2.25) Icl,S (Rn ) = F −1 (Scl
z
(Rn )) .
So
z
157.90 (2.26) Icl,S (Rn ) ⊂ ISRe z (Rn ).
Why the weird normalization of the order with the n/4? This is part of a bigger
scheme that I hope will be explained later. It is the standard notion with the n
interpreted as the codimension of the submanifold, here the origin, with respect to
which we are defining conormality.
So, apart from the issue with the order these are just the inverse Fourier tran-
forms of our ‘classical symbols’.
157.75 Theorem 2.1. If u ∈ ISm (Rn ) ⊂ S 0 (Rn ) then
singsupp(u) ⊂ {0}
157.77 (2.27)
(1 − φ)u ∈ S (R ) if φ ∈ C c∞ (Rn ), 0 ∈
n
/ supp(1 − φ).
157.77
The conditions in (2.27) do not characterize the conormal distributions.
Proof. By definition, a smooth function157.81
on Rn is a ‘symbol with bounds’ of
order m if it satisfies all the estimates (2.15). We can reëxpress these in the form
157.96 (2.28) ξ β ∂ξα a = (1 + |ξ|)m b, b ∈ L∞ (Rn ) ∀ β with |β| ≤ |α|.
It follows that if |α| > m + n + N + 1 then
R.9 (2.29) ∂ξα a = (1 + |ξ|)−n−N −1 bN , bN ∈ L∞(Rn ).
Since then F −1 (1 + |ξ|)−n−N −1 bN is bounded with its first N derivatives it fol-

lows that the same is true for

R.10 (2.30) |x|2k u, 2k > m + n + N + 1.
157.77
From this (2.27) follows. 
I have made a rather mixed definition of classical and non-classical symbols
here. The classical ones defined in terms of the radial compactification and the
non-classical ones in terms of estimates on Rn more directly, let me try to unravel
this.
157.91 Lemma 2.1. The ‘residual symbol spaces’ are
\
157.92 (2.31) S −∞ (Rn ) = S m (Rn ) = S (Rn ) = C˙ ∞ (Rn ) ⊂ Scl
z
(Rn ) ∀ z ∈ C.
m∈R

Here I am using the notation for any manifold with corners

157.93 (2.32) C˙ ∞ (M ) = {u ∈ C ∞ (M ); u vanishes to infinite order at ∂M }.
L2-end So these are the ‘trivial’ symbols in the case of Rn .
L3
 2. HOMOGENEOUS DISTRIBUTIONS 29

Last time I talked about the symbol spaces S m (Rn ) and the space of distribu-
tions conormal at 0 defined as
m+ n
157.97 (2.33) IS 4
(Rn ; {0}) = F −1 (S m (Rn )).
Since we know that the symbol spaces form a filtered ring under multiplication we
deduce a corresponding result for convolution of the conormal spaces
0 M +M 0 − n
157.98 (2.34) ISM (Rn ; {0}) ∗ ISM (Rn ; {0}) = IS 4
(Rn ; {0}).

2. Homogeneous distributions
The quesiton naturally arises as to precisely what classical
conormal distributions ‘look like’. The answer is given, almost,
in terms of homogoneous distributions. Then the first question
is, What is a homogeneous distribution?
We know what a homogeneous function on Rn \ {0} is – it
is a function satisfying
u(tx) = tz u(x) ∀ t > 0, x ∈ Rn \ {0}. 157.983 (2.35)
We do not want to include the origin here since the degree z
could be negative, so the function need not be defined at 0. In
fact we are interested in smooth functions away from 0 which
are homogeneous. So this is straihgtforward
157.984 Lemma 2.2. The smooth functions on Rn \ {0} which are
homogeneous of degree z ∈ C and smooth outside the origin may
be identified with C ∞ (Sn−1 ) via
x
u(x) = v( )|x|z , v ∈ C ∞ (Sn−1 ). 157.985 (2.36)
|x|
When Re z > −n it follows that such a function is locally
integrable across the origin, and so defines a tempered distri-
bution. In this case we can express the condition ‘weakly’ by
noting
Z Z
u(tx)φ(x)dx = t−n u(x)φ(x/t)dx =⇒

hu, φ(x/t)i = tn+z hu, φi, ∀ t > 0, φ ∈ S (Rn ). (2.37)

157.986

157.987 Definition 2.3. A (tempered) distribution u ∈ S 0 (R157.986

n
) is
homogeneous of degree z if it satisfies the second part of (2.37).
157.986
The argument leading to (2.37) can be reversed, at least in
part.
157.988 Lemma 2.3. A distribution which is homogeneous of degree
> −n and which is smooth outside the origin is of
z with Re z157.985
the form (2.36).
157.986
Proof. The identity (2.37) can be differentiated with re-
spect to t and at t = 1 this shows that Euler’s identity holds for
homogeneous distributions
(x · ∂x − z)u = 0. 157.989 (2.38)
 30 2. SYMBOLS AND CONORMAL DISTRIBUTIONS AT A POINT

Since it is assumed that u is smooth away from the origin it fol-

157.985
lows that u must be of the form (2.36) in Rn \{0}. The difference
of u and the locally integrable function so defined is therefore a
distribution supported at the origin, and hence a sum of deriva-
tives of the delta function. These all have homogeneity −n − k
for some k ∈ N0 so cannot be homogeneous of degree z with
Re z > −n. 

To pass to the case of general z we may use the Fourier

transform.
157.991 Lemma 2.4. The Fourier transform of a distribution which
is homogeneous of degree z ∈ C is homogeneous of degree −n−z.
Proof. For φ ∈ S (Rn ) change of variable shows that
\
φ(·/t)(ξ) = tn φ̂(tξ) 157.992 (2.39)
so the identity for u ∈ S 0 (Rn ) and ψ ∈ S (Rn )

hû, ψi = hu, ψ̂i 157.993 (2.40)

implies the result. 

It follows that for Re z < 0 a distribution, v, homogeneous

of degree z is the inverse Fourier transform of one homogeneous
of degree −z − n. If v is assumed to be smooth outside the origin
then it is the sum of a symbol of degree z and a distibution of
compact support. Thus its Fourier transform is singular only at
the origin and in this case we conclude that
 
ξ
v(x) = F −1 |ξ|−n−z u( ) , u ∈ C ∞ (Sn−1 ). 157.994 (2.41)
|ξ|
157.985
Conversely the functions in (2.36) are the sums of157.994
symbols and
compactly supported distribution so v defined by (2.41) is indeed
smooth outside the origin.
Thus the space of distribution homogeneous of any given
degree z ∈ C is always isomorphic to C ∞ (Sn−1 ). However for
z = −n − k, k ∈ N0 this isomorphism takes a different form.
157.995 Proposition 2.2. For any degree z ∈ / C \ (−n − N0 ) the
space of distributions homogeneous of degree z and smooth out-
side the origin is identified with C ∞ (Sn−1 ) by restriction to Rn \
157.985
{0} and (2.36). For z = −n − k, k ∈ N0 the space of distri-
butions homogeneous of degree z and smooth outside the origin
maps surjectively, by restriction to Rn \ {0} to
Z
{u ∈ C ∞ (Sn−1 ); u(ω)ω α dω = 0 ∀ |α| = k 157.996 (2.42)
Sn−1

with null space

sp{Dxα δ(x); |α| = k}. 157.997 (2.43)
 2. HOMOGENEOUS DISTRIBUTIONS 31

In fact the proof above can be modified to identify all ho-

mogenous distributions with C −∞ (Sn−1 ) in essentially the same
way.
However, this is not quite what we really want. The aim is
to understand the structure of classical conormal distributions at
the origin. These are readily described, as we do below, in terms
of homogeneous distributions and the inverse Fourier transforms
of functions of the form
 
z ξ
a(ξ) = (−χ(ξ))|ξ| u ,
|ξ|
u ∈ C ∞ (Sn−1 ), χ ∈ C c∞ (Rn ), χ(ξ) = 1 in157.998
|ξ| < 1. (2.44)
for the case z ∈ −n − N0 such a function is, using Propo-
Except157.995
sition 2.2, the sum of a homogeneous distribution and a distri-
bution of compact support. Conversely this is 157.996
not the case for
z = −n − k, k ∈ N0 , unless u is in the space (2.42). The extra
distributions we need are quasi-homogeneous in the sense that
they satisfy
(x · ∂x − z)u ∈ S (Rn ) 157.1103 (2.45)
The new case that arises is for z ∈ N0
R.1 Lemma 2.5. The inverse Fourier transform of
bα (x) = χ(x))xα log |x|, α ∈ Nno , χ ∈ C c∞ (Rn ), χ(ξ) = 1 in |x| < R.2
1 (2.46)
is quasi-homogeneous of degree −n − |α| and of the form
ξα
cn,α + e0α , e0α ∈ S (Rn ) in |ξ| > 1. R.7 (2.47)
|ξ|n+2|α|
with cn,α 6= 0.
Proof. Consider the case α = 0. Then b0 is a function of
|x| and by the O(n)-invariance of the Fourier transform the same
is true of b̂0 . Since
x · ∂x b0 = χ(x) + (x · ∂x χ) log |x| ∈ C c∞ (Rn ) R.4 (2.48)
it follows that
(ξ · ∂ξ + n)b̂0 = e ∈ S (Rn ). R.5 (2.49)
Using polar coordinates we can integrate the Schwartz term ‘in-
wards from infinity’ and conclude that
b̂0 = cn (1 − χ(ξ))|ξ|−n + e0 , e0 ∈ S (Rn ) R.6 (2.50)
where cn 6= 0 involves the measure of the sphere. It cannot
vanish, since otherwise b0 would be smooth.
Since bα = xα b0 has compact support its inverse Fourier
R.6
transform, b̂α (−ξ) ∈ S 0 (Rn ) is smooth and from (2.50)
ξα
b̂α = Dξα b0 = cn,α + e0α , e0α ∈ S (Rn ) R.3 (2.51)
|ξ|n+2|α|
where again cn,α 6= 0 since bα is not smooth. These constants
are readily computed. 
 32 2. SYMBOLS AND CONORMAL DISTRIBUTIONS AT A POINT

Finally then we see the structure of classical conormal dis-

tributions.
157.999 Proposition 2.3. A classical conormal distribution at the
origin in Rn of order z + n4 ,
u = F −1 (a), a ∈ |ξ|zC ∞ (Rn ) 157.1102 (2.52)
has an expansion near the origin, modulo smooth terms,
X
u(x) ∼ uk (x) R.8 (2.53)
k
where all terms are smooth outside the origin and
(
is homogeneous of degree − z = −n + k if − z − n + k ∈/ N0
uk
is quasi-homogeneous of degree − z = −n + k if − z − n + k ∈ N0 .
157.1000 (2.54)
157.1000
uk become increasingly smooth as k → ∞ and
From (2.54) the R.8
the meaning of (2.53) is that, for given N the difference of u and
C N (Rn ). The ap-
the sum for k < K for K sufficiently large is in 157.1000
pearance of the quasi-homogoneous terms in (2.54) corresponds
to the fact that the polynomials of degree k do not contribute
to the singularity at the origin (nor of course to the asymptotic
expansion of the symbol).
There are other ways to construct the homogeneous, and
quasi-homogeneous distributions, for instance by analytic con-
tinuation in z.

3. Topology and asymptotic summation

The topology on the symbols space S m (Rn ) is the Fréchet topology given by
the norms defining the space
157.99 (2.55) kakm,N = sup (1 + |ξ|)−m+|β| |∂ξβ a(ξ)|, N ∈ N0 .
Rn ,|β|<N

Certainly, a ∈ S m (Rn ) if and only if a ∈ C ∞ (Rn ) and all these norms are finite.
Recall that a metric on a countably normed space, such as this, is defined by
X ku − vkm,N
157.100 (2.56) d(u, v) = 2−N .
1 + ku − vkm,N
N
157.100
So the topology is metric, generated by the open balls with respect to (2.56). I say
‘a metric’ because replacing the sequence 2−N by an positive, summable, sequence
gives the same topology.
157.114 Proposition 2.4. The spaces S m (Rn ) are Fréchet spaces, so complete with
157.100
respect to the translation-invariant distance (2.56).
If it matters to you, they are Montel spaces. They are not projective limits of
Hilbert spaces.
Proof. Convergence with respect to this distance is the same as convergence
with respect to each of the norms k · km,N (without
157.100
any uniformity in N ). Thus a
Cauchy sequence with respect to the metric (2.56) is Cauchy with respect to each
 3. TOPOLOGY AND ASYMPTOTIC SUMMATION 33

of these norms and conversely. So all the derivatives converge locally uniformly and
with respect to the distance with the limit in the space. 
There is a topology on S ∞ (Rn ) = m S m (Rn ) but I will leave you to figure it
S
out:
157.454 Exercise 3. Try to sort out (or look up) the inductive limit topology on
S ∞ (Rn ) defined by taking a set to be open if its intersection with each of the
S m (Rn ) is open and show that the inclusions S m (Rn ) −→ S ∞ (Rn ) are then con-
tinuous.
So the Fréchet topology on the symbol spaces induces a Fréchet topology
n
on ISM (Rn , {0}), since the Fourier transform identifies this with S M − 4 (Rn ). This
means we know what a continuous map into the conormal space (and also what a
smooth map into it) means.
Now to density. The intersection of the symbol spaces is the space of Schwartz
functions
\
157.101 (2.57) S (Rn ) = S −∞ (Rn ) = S m (Rn ) = S (Rn ).
m
157.102 Proposition 2.5. The ‘residual space’ S (Rn ) is dense in S m (Rn ) in the topol-
ogy of S m+ (Rn ) for any  > 0. More precisely, there exist a sequence of ‘regularizing
operators’ which are linear maps
157.103 (2.58) Φk : S ∞ (Rn ) −→ S (Rn )
such that
157.104 (2.59) a ∈ S m (Rn ) =⇒ Φk a is bounded in S m (Rn )
and Φk a → a in the topology of S m+ (Rn ) ∀  > 0.
Proof. The Φk can be defined by cut-off. Take φ ∈ C c∞ (Rn ) with φ(ξ) = 1 in
{|ξ| < 1} and set
157.105 (2.60) Φk a(ξ) = φ(ξ/k)a(ξ) ∈ S (Rn ).
The difference
157.106 (2.61) (Id −Φk )a = (1 − φ(ξ/k))a(ξ)7 ∈ S m (Rn )
since 1 − φ(ξ/k) ∈ S 0 (Rn ).
Certainly 1 − φ(ξ/k) is uniformly bounded and the derivatives are
157.107 (2.62) ∂ξβ (1 − φ(ξ/k)) = −k −|β| (∂ β φ)(ξ/k), |β| > 0.
Since this function is supported in |ξ| < Ck, for some constant C, the product
satisfies
157.108 (2.63) sup |∂ξβ (1 − φ(ξ/k))| ≤ Cβ (1 + Ck)|β| k −|β| < ∞
ξ

This shows that 1 − φ(ξ/k) is bounded with respect to all the seminorms for
S 0 (Rn ). It follows that Φk a is bounded in S m (Rn ).
The seminorms on S  (Rn ) on the difference 1−Φk , which has support in |ξ| > k,
have an extra factor of (1 + |ξ)−

157.136 (2.64) (1 + |ξ|)−+|β| ∂ξβ (1 − φ(ξ/k)) = k −|β| (1 + |ξ|)−+|β| (∂ β (1 − φ))(ξ/k)

=⇒ k1 − Φk k,N ≤ CN k −
 34 2. SYMBOLS AND CONORMAL DISTRIBUTIONS AT A POINT

where CN depends on φ and N. Thus 1−φ(ξ/k) → 0 with respect to each seminorm

on S  (Rn ) for  > 0. It follows that
157.137 (2.65) Φk a → a in S  (Rn ) ∀  > 0.
157.104
This in turn implies (2.59). 
157.137
We record the norm estimate which underlies (2.65) for use below
157.138 Lemma 2.6. For a ∈ S m (Rn ) and any m0 > m
0
157.139 (2.66) k(Id −Φk )akm0 ,N ≤ CN,m,m0 kakm,N k m−m
where the constant is independent of a and k.

4. Integration
157.109 Lemma 2.7. Integration in one of the variables, say the last, gives a continuous
linear map
Z Z 
m n m− 14 n−1
157.110 (2.67) dxn : IS (R ) −→ IS (R ), F u(x)dxn = F (u) ξn =0 .
R R

Of course we can iterate this, integrating over k variables to get a conormal distri-
bution with order decreased by k/4.

Proof. The integral is defined since integration of both Schwartz functions

and distributions of compact support is well-defined. Using density we can suppose
that a ∈ S (Rn ) and then
Z Z Z
0
157.111 (2.68) dxn (F −1 a)(x0 , xn )dxn = (2π)−n dxn eix ·ξ a(ξ)dξ
R R R n
Z
−n+1 ix0 ·ξ 0
= (2π) e a(ξ , 0)dξ 0 , x = (x0 , xn ), ξ = (ξ 0 , ξn ).
0
Rn−1
by the Fourier inversion formula in one dimension. Thus
Z
157.112 (2.69) dxnF −1 (a) = F −1 (a ξn =0 ), a ∈ S (Rn ).
R
Clearly, with Φk as defined above
157.113 (2.70) Φk (a ξn =0
) = (Φk a) ξn =0

so the general case follows. 

5. Wavefrontset
The support of a function or distribution on Rn is defined by
[ {
supp(u) = {U ⊂ Rn ; U is open and u = 0 on U } . 157.115 (2.71)

This is really a notion defined for sheaves (the theory of which I will outline
below in case you have not seen it). We define a related notion for symbols
which is to do with the directions of growth at infinity.
If V ⊂ Sn−1 is open then the set
ξ
R+ V = {ξ ∈ Rn \ {0}; ∈V} 157.116 (2.72)
|ξ|
 6. RESTRICTION 35

is what we mean by an open cone – an open subset of Rn \ {0} which is

invariant under the radial R+ action. If ψ ∈ C c∞ (V ) and φ ∈ C c∞ (R) is
identically equal to 1 near 0 then
ξ
χ(ξ) = (1 − φ)(|ξ|)ψ( ) ∈ S 0 (Rn ) ⊂ C ∞ (Rn ) 157.117 (2.73)
|ξ|
where we define it to be identically zero in |ξ| <  where φ(|ξ|) = 1 on
|ξ| < . Thus, φ is only there to cut out the singularity the homogeneous
ξ
function ψ( |ξ| ) is almost certain to have at the origin.

157.118 Lemma 2.8. An element of C ∞ (Rn ) which is homogeneous of complex de-

gree z vanishes identically unless z ∈ N0 , in which case it is necessarily a
polynomial.

Proof. The homogeneity statement is that

u(tξ) = tz u(ξ) ∀ t > 0, ξ ∈ Rn . 157.455 (2.74)
157.455
Consider the derivatives of u at the origin. From (2.74) it follows that
∂ξα u(0) = t|α| ∂ξα u(0) = tz ∂ξα u(0). 157.456 (2.75)
Thus either z = |α| or ∂ξα u(0) = 0. So, if z is not a non-negative integer
then u must vanish to infinite order at 0. But then
t−z u(tu) → 0 as t ↓ 0 =⇒ u ≡ 0. 157.457 (2.76)
If z = k then u is the sum of a polynomial and a function which vanishes to
infinite order at 0 and the same argument shows that such a homogeneous
function vaniahes identically. 
157.117
The product of a ∈ S m (Rn ) and a function χ as in (2.73) is always in
S m (Rn ). However, it might be much smaller.
157.119 Definition 2.4. The cone-support of a symbol a ∈ S m (Rn ) is the (rela-
tively) closed subset of Rn
[ {
conesupp(a) = {R+ V ; χa ∈ S (Rn ) ∀ ψ ∈ C c∞ (V )} . 157.120 (2.77)

In fact the union of all sets V for which ψa ∈ S (Sn−1 )

when ψ ∈ C c∞ (V
)
is still a V for which this holds – i.e. there is a maximal such V. Clearly
the cone-support is a cone, so the information it contains is the same as the
corresponding closed subset of the sphere. It is traditional to think of it
as a cone, partly because of the definition, but also because it has a little
content as we will see later.
157.121 Exercise 4. Show that conesupp(a) = ∅ iff a ∈ S (Rn ) = S −∞ (Rn ).
n
157.122 Definition 2.5. For u ∈ ISM (Rn ; {0}) = F −1 (S M − 4 (Rn )) we set
WF(u) = conesupp(a), u = F −1 (a) ⊂ T0∗ Rm \ {0} = Rn \ {0}.
157.123 (2.78)
It matters here that this is the inverse Fourier transform not the Fourier
transform, otherwise there is a reflection. The identification of Rn \ {0}
as the cotangent fibre at 0 on Rn might appear somewhat arbitrary but is
justified by results below on coordinate-invariance. In any case, by definition
the wavefront set (that is what WF stands for) of a conormal distribution
at the origin in Rn is a closed cone in T0∗ Rn \ {0}.

6. Restriction
157.110
What is this notion of wavefrontset good for? Notice in (2.67) that inte-
gration and restriction are dual under Fourier transform, at least in this
special case. In general we cannot expect to restrict a conormal distribution
to xn = 0 – for instance this is not reasonable for the delta function at the
origin. Dually, we cannot expect to integrate a symbol, it may just be too
large at infinity.
 36 2. SYMBOLS AND CONORMAL DISTRIBUTIONS AT A POINT

157.124 Lemma 2.9. If a ∈ S m (Rn ) then

Z
±en ∈
/ conesupp(a) =⇒ dξn a(ξ, ξn ) ∈ S m+1 (Rn−1 ). 157.125 (2.79)
R

Here en = (0, . . . , 0, 1) is the unit vector and this is not in conesupp(a)

R+ en does not meet conesupp(a). So you can
if and only if the half-line 157.125
interpret the condition in (2.79) as saying
R · en ∩ conesupp(a) = ∅. 157.126 (2.80)

Proof. The condition on conesupp(a) means that we can find a cut-off

function ψ ∈ C ∞ (Sn−1 ) on the sphere which is non-vanishing at ±en and
such that  
ξ
(1 − φ|ξ|)ψ a ∈ S (Rn ). 157.129 (2.81)
|ξ|
So this means that in a conic region
Γ = {ξ ∈ Fn ; |ξ 0 | ≤ |ξn |},  > 0 157.130 (2.82)
the symbol a is rapidly decreasing with all its derivatives. In fact we can
can assume that ψ = 1 near ±en and then write
 
ξ
a = a0 + φ(|ξ|) + (1 − φ(|ξ|))ψ ∈ S m (Rn ),
|ξ| 157.131 (2.83)
a = 0 in Γ , a − a0 ∈ S (Rn ).
Since integration certainly maps S (Rn ) into S (Rn−1 ) it suffices to consider
a0 in place of a and look at
Z
b(ξ) = dξn a(ξ 0 , ξn ). 157.132 (2.84)
R
This integral certainly exists since for each ξ 0 the integrand is supported in
|ξn | ≤ −1 |ξ 0 |. Thus from the leading symbol estimate for a we see that in
|ξ 0 | > 1
Z
|b(ξ 0 )| ≤ C (|ξ 0 | + |ξn |)m dξn . 157.133 (2.85)
|ξn |≤−1 |ξ0 |
Now, changing the variable of integration to τ = ξn /|ξ 0 | it follows that
Z
|b(ξ 0 )| ≤ C |ξ 0 |m+1 (1 + |τ |)m dτ ≤ C 0 |ξ 0 |m+1 . 157.134 (2.86)
|τ |≤−1

The same argument applies to all the ξ 0 derivatives, so

b ∈ S m+1 (Rn−1 ). 157.135 (2.87)


157.127 Corollary 3. Restriction to the coordinate hyperplane is well-defined as

a linear map

xn =0
: {u ∈ ISM (Rn ; {0});{dxn , −dxn } ∩ WF(u) = ∅}
3
157.128 (2.88)
−→ I M + 4 (Rn−1 ; {0}).

7. Multiplicativity
One of the applications of the notion of wavefront – for general distributions
not just in the conormal case – is to provide conditions under which operations
extend to distributions.
R.11 Exercise 5. Show that if Γi ⊂ Rn \ {0}, i = 1, 2, are (relatively of course)
closed cones with the property
R.12 (2.89) ξi ∈ Γi , 1, 2 =⇒ ξ1 + ξ2 6= 0
 8. ASYMPTOTIC COMPLETENESS 37

then the product of functions can be extended to

R.13 (2.90)
{u1 ∈ ISm1 (Rn ; {0}); WF(u1 ) ∪ Γ1 } × {u2 ∈ ISm2 (Rn ; {0}); WF(u2 ) ∪ Γ2 } −→ S 0 (Rn ).
Hint: One way to define the product is to use the fact that these conormal
spaces are modules over S (Rn ). So it is enough to define the pairing of the prod-
uct with 1 provided there is some continuity. If u2 ∈ S (Rn ) with the integrals
interpreted as pairings
Z Z
1
R.14 (2.91) u1 (x)u2 (x)dx = û1 (ξ)û1 (−ξ)dξ.
Rn (2π)n
R.12
Now you can see where the condition (2.89) comes from. Using a partition of unity
to divide Rn into two conic regions (and a compact part) the ‘integral’ on the right
can be divided
R.13
into parts in each of which one of the terms is Schwartz. To really
justify (2.90) one
R.13
needs to do something about continuity – find a topology on the
spaces in (2.90) with S (Rn ) dense and such that convergence is consistent with the
decomposition.
You might also ask whether the product so defined is conormal. In general it
is not.

8. Asymptotic completeness
The main interest in symbols on Rn is their behaviour ‘at infinity’ (which is the
boundary of the radial compactification). This allows for a notion of ‘convergence’
which corresponds to the ‘asymptotic completeness’ in the following sense.
157.140 Theorem 2.2. If aj ∈ S mj (Rn ) is a sequence (we think of it as a series)
of symbols with mj → −∞ as j → ∞ then there exists a symbol a ∈ S M (Rn ),
M = sup mj such that for every k
X
157.141 (2.92) a− aj ∈ S M (k) , M (k) = sup mj
j>k
j≤k

and a is determined up to an error in S −∞ (Rn ) by these conditions.

157.141
The relationship (2.92) between a and the aj is interpreted as ‘a complete asymp-
totic expansion’ and written
X
157.142 (2.93) a∼ aj .
j

Note that we are certainly not saying that the series on the right converges in
any sense (well people say it converges asymptotically, just meaning the order
mj → −∞).
I have been a little vague here about the range of j, usually one takes j ∈ N0 ,
so starting off at 0, but this is just a convention.
157.141
Proof. The ‘uniqueness’ (modulo S −∞ (Rn )) is immediate from (2.92) – given
two such ‘asymptotic sums’ a and a0 the difference satisfies
X X
157.153 (2.94) a0 − a = (a0 − aj ) − (a − aj ) ∈ S M (k) (Rn ) ∀ k =⇒
j≤k j≤k

a0 − a ∈ S −∞ (Rn ) = S (Rn ).
 38 2. SYMBOLS AND CONORMAL DISTRIBUTIONS AT A POINT

For existence, I will assume, as discussed below, without loss of generality that
the mj are strictly decreasing, just to simplify notation.
I will use the ‘approximation’ operators (Id −Φl ) discussed above, where Φl is
multiplication by φ(ξ/l) for φ ∈ C c∞ (Rn ) equal to 1 near 0. So these are cutoffs near
infinity. Here l will vary with j so we are looking for a sequence of integers
157.144 (2.95) l(j) → ∞ in N.
What we want these integers to satisfy – they depend of course on the given sequence
aj – is
X
157.145 (2.96) k(Id −Φl(j) )aj kmk ,N < ∞ ∀ k, N.
j>k

So this is a countable set of conditions we need to satisfy.

157.145
Let’s just examine one of the conditions (2.96). It makes sense, since the terms
are in S mj ⊂ S mk and mk − mj > 0 for j > k by assumption. This is often called
‘absolute summability’ of the sequence with respect to the norm. It implies that
the series is Cauchy with respect to this norm, and that is what we are after. That
is we will ensure that the series
X
157.146 (2.97) (Id −Φl(j) )aj is Cauchy with respect to k · kmk ,N .
j>k

For the moment of course just for one N and k.

157.138 157.145
We have, from Lemma 2.6, an estimate on each of these norms in (2.96)
157.147 (2.98) k(Id −Φl(j) )aj km(k),N ≤ CN,k (l(j))mk −mj kaj kmk ,N
(ultimately because the mj < mk ). Here the constant does not depend on l(j) –
the dependence is the power. To make the series converge absolutely it suffices to
arrange that
157.148 (2.99) k(Id −Φl(j) )aj km(k),N ≤ j −2
for instance. In fact convergence is a property of the ‘tail’ of the sequence – the
behaviour
157.148
of any finite number of terms is irrelevant – so it is enough to arrange
157.147
(2.99) from some j onwards. From (2.98) we see that we can ensure this by choosing
l(j) so that
157.149 (2.100) l(j) > L(N, k, j)
where for this N and k, L(N, k, j) is some explicit sequence which depends on the
norms of the aj .
Now, this shows we can choose the l(j) so that any one of the series (labelled
by k) converges absolutely with respect to any one of the norms k · kmk ,N . In fact
by a ‘diagonalization’ procedure we can ensure that all the series are Cauchy with
respect to all the norms (and hence converge in the corresponding symbol space).
To do this, just arrange all the (N, k) as a sequence, parameterized by p, and
157.149
demand that (2.100) hold for j > p.
So we can choose the integers l(j) such that each of the series
X
157.150 (2.101) (Id −Φl(j) )aj converges in S mk (Rn )
j>k
 9. ISm (Rn ; {0}) AS A MODULE 39

in the strong sense that it converse absolutely with respect to each of the seminorms.
Now set
X
157.151 (2.102) a = a0 + (Id −Φl(j) )aj ∈ S m0 (Rn ).
j≥1

This is our asymptotic sum. To check this observe that the difference with a finite
sum can be written
X X X
157.152 (2.103) a− aj = − Φl(j) aj + (Id −Φl(j) )aj .
j≤k+1 j≤k+1 j>k+1
mk+1
The last sum here is in S and the finite sum is actually of compact support, so
157.141
in S −∞ (Rn ). The last term on the left is in the same space, S mk+1 (Rn ) so (2.92)
follows.
If we do not have a strictly decreasing sequence of orders, we can rearrange the
sequence so that the order is weakly decreasing and then sum up an finite sequences
of fixed order. This reduces the problem to the strictly decreasing case and, since
157.141
we have arranged absolute convergence, we recover (2.92) in general. 

Except that the topology is a little dubious, we have shown that a series with
elements in S mj (Rn )/S −∞ (Rn ) ‘always converges’ if the mj → −∞. What this
really means is that there exist representatives of the elements in S mj (Rn ) such that
the series does converge and gives a well-defined limit in S sup mj (Rn )/S −∞ (Rn ). L3-end
L4

9. ISm (Rn ; {0}) as a module

First let’s fix some notation.
R.15 Definition 2.6. The full symbol map for conormal distributions at the origin
is
n
157.158 (2.104) σ : ISm (Rn ; {0}) 3 u −→ [F (u)] ∈ S m− 4 (Rn )/S −∞ (Rn ).
So the full symbol in this sense is the Fourier transfrom – which would be called
n
the amplitude – F (u) ∈ S m− 4 (Rn ) modulo the Schwartz space.
This means there is a short exact sequence for each m
 / ISm (Rn ; {0}) σ / S m− n4 (Rn )/S −∞ (Rn ).
R.16 (2.105) S (Rn )

The smooth functions of ‘slow growth’ form a (not very pleasant) linear space
which I will denote by O (Rn ) which is a space of multipliers on S (Rn ). A smooth
function is an element ψ ∈ O (Rn ) if for each multiindex α ∈ Nn0 there exists mα
such that
157.154 (2.106) sup(1 + |ξ|)−mα |Dξα ψ(ξ)| < ∞.
ξ

Thus multiplication gives a bilinear map

157.155 (2.107) O (Rn ) × S 0 (Rn ) −→ S 0 (Rn ) which restricts to O (Rn ) × S (Rn ) −→ S (Rn ).
157.156 Proposition 2.6. For any m, multiplication defines a map
157.157 (2.108)
X 1
O (Rn ) × ISm (Rn ; {0}) −→ ISm (Rn ; {0}) with σ(ψu) ∼ ∂xα ψ(0)Dξα σ(u).
n
α!
α∈N0
 40 2. SYMBOLS AND CONORMAL DISTRIBUTIONS AT A POINT

n
The sum of the right does determine a unique element of S m− 4 /S −∞ (Rn ) =
n
S m− 4 (Rn )/S −∞ (Rn ) as we showed last time using asymptotic summation. If we
n
take a representative a ∈ S m− 4 (Rn ) – such as the actual Fourier transform of u –
then the terms in the infinite sum are of orders m − n4 − |α| so the sum ‘converges
asymptotically’.

Proof. Since we know that u ∈ ISm (Rn ; {0}) can be written as the sum of a
compactly supported term and one in S (Rn ) on which ψ ∈ O (Rn ) acts it suffices to
suppose that both u and ψ are compactly supported. The Taylor series expansion
X 1 X
157.458 (2.109) ψ= ∂xα xα ψ(0) + xα µα (x), µα ∈ C ∞ (Rn )
α!
|α|≤N |α|=N +1

can then be multiplied by a cut-off of compact support equal to 1 on supp(u) ∪

supp(ψ) showing that
X 1 X
157.459 (2.110) ψu = ∂xα ψ(0)xα u(x) + µ0α (x)(xα u)µ0α (x), µ0α ∈ C c∞ (Rn )
α!
|α|≤N |α|=N +1

m−|α|
The terms in the first sum are are in IS (Rn ; {0}) as the inverse Fourier trans-
forms of the
i|α| α |α|
157.460 (2.111) ∂ ψ(0)∂ξ û(ξ).
α! x
Similarly the terms in the second sum consists of products in
C c∞ (Rn ) · ISm−N −1 (Rn ; {0}).
157.460
Let v ∈ ISm (Rn ; {0}) be an asymptotic sum of this series (2.111) which can be
157.459
taken to have compact support. Then from (2.110)
X
157.461 (2.112) ψu − v = vN + µ0α (x)(xα u)µ0α (x), vN ∈ ISm−N −1 (Rn ; {0}).
|α|=N +1

For for N > m + k + n

\
157.463 (2.113) ISm−N −1 (Rn ; {0}) ⊂ C k (Rn ) =⇒ ψu − v ∈ C k (Rn ) = C ∞ (Rn ).

So in fact ψu − v ∈ S (Rn ). 157.157

The asymptotic formula (2.108) is a restatement of the conclusion that ψu is
157.460
an asymptotic sum of the terms (2.111). 

10. Action of Ψ∗ on I ∗
157.464 Proposition 2.7. Pseudodifferential operators act on conormal distributions
giving a bilinear map
0 0
Ψm (Rn ) × ISm (Rn ; {0}) −→ ISm+m (Rn ; {0}) with
157.465 (2.114) X 1
σm+m0 (Au) ∼ Dxα ∂ξα (aσ(u)), A = QL (a)
α
α!

To take care of the ‘error terms’ below we need first consider a coarser mapping
property; here there is no need for precision as regards orders and regularity.
 11. RADIAL COMPACTIFICATION AND SYMBOLS 41

R.17 Lemma 2.10. If k ∈ N0 and

R.18 (2.115) m + m0 < −n − k
then
0
R.20 (2.116) P ∈ Ψm (Rn ), u ∈ ISm (Rn ; {0}) =⇒ P u ∈ C ∞
k
(Rn ),
meaning the derivatives up to order k are bounded.
Proof. We can use density of S (Rn ) in symbol spaces, with a little loss of
R.18
order. The inequality (2.115) implies that the products
R.21 (2.117) |Dxβ a(x, ξ)û(ξ)| ≤ Cβ (1 + |ξ|)−n−k .
So, in particular are actually L1 functions of ξ and so the integral
Z
R.22 (2.118) P u(x) = (2π)−n a(x, ξ)û(ξ)eix·ξ dξ
Rn
converges uniformly in x. Differentiation by x toR.21 β with |β| ≤ k still leads
any orderR.20
to absolutely convergent integrals because of (2.117) so (2.116) follows. 
157.464 157.465
Proof of Proposition (2.7). We already know a special case of (2.114) be-
cause we have an inclusion map
m+ n
R.23 (2.119) IS 4
(Rn ; {0}) −→ Ψm (Rn ), u(x) 7−→ A(x, y) = u(x − y)
in which conormal distributions act as convolution operators – this after all is one
reason I are talking about conormal distributions (although not the only one). We
will exploit this ruthlessly! 
L4-end

11. Radial compactification and symbols

Project 1 P-Comp
V2: Two corrections from Benjy.
I detected some resistance to the idea of radial compactification of Rn in class
so the main part of the first problem set is to work out some of the details. Quite
a bit of this is already in the notes.
0 First recall, for background if nothing else, the basis of projective geom-
etry (which seems to have disappeared as a subject taught to undergrad-
uates not long before I started studying Mathematics). Define complex
projective space as a quotient
P1.1 (2.120)
Pn = (Cn+1 \{0})/C∗ = S2n+1 /T, C∗ = C\{0}, S2n+1 = (R2n+2 \{0})/R+ = (Cn+1 \{0})/C+ .
Check that this is a complex manifold and that Cn is identified with an
open dense subset by the inclusion
P1.2 (2.121) Cn 3 z 7−→ (z, 1) ∈ Cn+1 \ {0} ,→ Pn
and that the complement of the image may be identified with Pn−1 .
(1) Now sort out the real (or more correctly a) real analogue of this. Take
the embedding
P1.3 (2.122) Rn 3 x 7−→ (x, 1) ∈ Rn × [0, ∞) ⊂ Rn+1
and consider the quotient map
P1.4 (2.123) ι : Rn −→ (Rn × [0, ∞) \ {0}) /R+ = Sn+ = Rn
 42 2. SYMBOLS AND CONORMAL DISTRIBUTIONS AT A POINT

mapping into the upper half-sphere (see picture below); the last equation
defines Rn . I take this as the definition of the radial compactification;
show that the embedding is given explicitly by
!
x 1
P1.5 (2.124) ι(x) = p ,p
1 + |x|2 1 + |x|2
1
and deduce that t = (1+|x|2 )− 2 ∈ C ∞ (Sn+ ) is a boundary defining function
(vanishes only on the boundary and has differential non-zero there).
(2) Derive the Taylor series of a ∈ Scl0
(Rn ) = C ∞ (Sn+ ) (this is the definition
of the space of classical symbols of order 0 from lectures) in the form
∞
X x √
P1.6 (2.125) |x|−k ak ( ), ak ∈ C ∞ (Sn−1 ), |x| > 2(⇐⇒ t < 2).
|x|
k=0

Deduce that Taylor series with remainder gives

N
X x
P1.7 (2.126) |a − |x|−k a( )| ≤ CN |x|−N −1 .
|x|
k=0

[We want similar estimates for derivatives too].

(3) Introduce projective coordinates on Rn given by 2n+1 coordinate patches
on Sn+ . The first one is x ∈ Rn defining the compactification. Then for
each k = 1, . . . , n set
P1.8 (2.127) Dk± = {x ∈ Rn ; ±xk > 0}, Ck± = {p = (p1 , . . . , pn+1 ) ∈ Sn+ ; ±pk > 0}
(note that this includes part of the boundary of Sn+ ) and show that the
diffeomorphisms
1 xj
P1.9 (2.128) Ck± 3 x 7−→ ( , ± ) ∈ (0, ∞) × Rn−1
±xk xk
extend to diffeomorphism Dk± −→ [0, ∞) × Rn−1 .
(4) Show that these projective coordinate systems give a coordinate cover of
Rn .
(5) Write out formulæ for the images of the vector fields
P1.10 (2.129) ∂xj , xl ∂xj
in these projective coordinate systems (note these span the Lie algebras
of the translation group and GL(n, R) respectively).
(6) Show that the xl ∂xj extend to be smooth on Rn (meaning smooth up to
the boundary) and that they are elements of the Lie algebra
P1.11 (2.130) Vb (Rn ) = {V a C ∞ vector field tangent to the boundary}.
Note that tangency to the boundary means V t = 0 at t = 0.
(7) Show that the images of the xl ∂xj span Vb (Rn ) over C ∞ (Rn ).
(8) Show that the ∂xj are also smooth up to the boundary of Rn and span,
over C ∞ (Rn ) the space
P1.14 (2.131) tVb (Rn ) = {W = tV, V ∈ Vb (Rn )}.
 11. RADIAL COMPACTIFICATION AND SYMBOLS 43

(9) Show that the space S 0 (Rn ) of ‘symbols with bounds’ is identified with
the space
P1.12 (2.132) {u ∈ L∞ (Rn ); V1 . . . VN u ∈ L∞ (Rn ) ∀ Vi ∈ Vb (Rn ) ∀ N }.
[Don’t get hung up on worrying about distributions on a manifold with
boundary, we will come back to this later.]
(10) Putting some of these things together show that
0
P1.13 (2.133) Scl (Rn ) ⊂ S 0 (Rn ).
(11) Deduce that an element a ∈ S 0 (Rn ) is in Scl
0
(Rn ) if and only if
X ξ
P1.15 (2.134) a∼ (1 − φ)(ξ)|ξ|−k ak ( )
|ξ|
k
where φ ∈ C c∞ (Rn ) is equal to one near 0 (to make everything smooth)
and the ak ∈ C ∞ (Sn−1 ).
 CHAPTER 3

The ring Ψ∗ (Rn )

C.PsiRn L5
I will start today with the coordinate-invariance of conormal distributions at a
point and proceed to discuss the fact that the formal adjoint of a pseudodifferen-
tial operator is also a pseudodifferential operator. These might seem to be rather
unrelated results, but as we shall see the proofs are closely related.

1. Coordinate invariance of Icm (Rn ; {0})

Since we do not want to worry about the global behaviour of diffeomorphisms
we will work locally near 0 ∈ Rn . If Ω ⊂ Rn is an open neighbourhood of 0 set
157.159 (3.1) Icm (Ω; {0}) = {u ∈ ISm (Rn ; {0}); supp(u) b Ω}.
Here of course we are thinking of Ω as an open subset of Rn but we can also think
of it as a manifold. For the conormal functions to make sense on a manifold we
need:
157.160 Proposition 3.1. If F : Ω −→ Ω0 is a diffeomorphism of open neighbourhoods
of 0 ∈ Rn with F (0) = 0 then
157.161 (3.2) F ∗ : I m (Ω0 ; {0}) −→ I m (Ω; {0}).
You should recall that the pull-back of distributions is well-defined under a dif-
feomorphism (not under a general smooth map). I will remind you of the ‘issues’
arising in the proof of this by duality – namely the need to think about densities
– below. Using the density of C c∞ (Ω) in C c−∞ (Ω) for any open set Ω and the fact
157.161
that F ∗ extends by continuity I claim that (3.2) already has meaning.
Proof. First recall that the ‘full symbol map’ is still surjective if we restrict
157.159
supports as in (3.1) since any u0 ∈ I m (Rn ; {0}) differs from an element of Icm (Ω; {0})
by an element of S (Rn ). We will use this in the proof.
First we start with a simple case, when F ∈ GL(n, R) is actually an invertible
linear map. Then there is no problem with supports.
157.162 Lemma 3.1. Under pull-back by L ∈ GL(n, R)
157.163 (3.3) L∗ : ISm (Rn ; {0}) −→ ISm (Rn ; {0}).
Proof. This corresponds to the fact the Fourier transform behaves ‘well’ under
linear change of coordinates. For u ∈ S (Rn ) it follows directly that
Z Z
−1
157.164 (3.4) F (L∗ u)(ξ) = e−ix·ξ u(Lx)dx = e−i(L y)·ξ u(y)| det L|−1 dy
Rn Rn
Z
−1 t
= e−iy·(L ) ξ u(y)| det L|−1 dy = | det L|−1 û((L−1 )t ξ).
Rn

45
 46 3. THE RING Ψ∗ (Rn )

So the only issue is the constant factor, but this does not affect the fact that
n n
F (L∗ u) ∈ S m− 4 (Rn ) if û ∈ S m− 4 (Rn ). 
157.160
Now, this allows us to simplify the general case in Proposition 3.1. Namely we
can write F = LG where L is the Jacobian matrix of F at 0 and G : Ω −→ L−1 Ω0
still has G(0) = 0 and now has Jacobian equal to the identity at 0. Replacing G by
F again we can therefore assume that
X
157.165 (3.5) F (x) = x + xi xj Gij (x) in |x| < , Gij ∈ C ∞ .
ij

There is no problem in shrinking supports to a smaller neighbourhood of 0 since

the conormal distributions are all smooth away from 0.
We can exploit the triviality of the Jacobian at 0 by observing that
X
157.166 (3.6) Ft (x) = x + t xi xj Gij (x), t ∈ [0, 1]
ij

is a smooth family of diffeomorphisms of a fixed neighbourhood of 0 with image

containing some fixed neighbourhood of 0 and with
157.167 (3.7) F0 (x) = x, F1 (x) = F (x).
This allows us to replace the problem by a deformation problem, meaning we can
get from beginning to end along a path (you might think this is actually harder).
However we can now use the variation formula (really just the chain rule) that
d ∗ X
157.168 (3.8) (Ft u) = Ft∗ (Vt u), Vt = xi xj aij,k (t, x)∂xk .
dt
k,j,i

Here Vt is the t-dependent vector field which defines Ft by integration157.168

– Ft is
the unique 1-parameter family of local diffeomorphisms which satisfies (3.8) (and
F0 = Id). It is important here that Vt vanishes to second order at 0 (meaning its
coefficients vanish quadratically at 0 of course).
So how does this help us? We need another idea, which I learnt from Jürgen
Moser in a rather different context. Namely we can suppose that u = ut actually
depends smoothly on t as a parameter
157.168
(with values in the conormal distributions).
Then the variation formula (3.8) becomes
d ∗ d
157.169 (3.9) (F ut ) = Ft∗ (Vt ut + ut ).
dt t dt
d
Now, the idea is that we try to choose ut with u1 = u so that Vt ut + dt ut = 0. We
cannot manage this directly but what we can do is to choose ut ∈ C ([0, 1]; Icm (Ω00 ))
∞

so that
d
157.170 (3.10) Vt ut + ut ∈ C ∞ ([0, 1]; C c∞ (Ω00 ));
dt
here Ω00 is some suitably 157.170
small open neighbourhood of 0.
The idea is to solve (3.10) by successive steps and the crucial point here is that
157.171 (3.11) Vt : Icm (Ω00 ; {0}) −→ Icm−1 (Ω00 ; {0}) ∀ m ∈ R.
157.156
This follows from Proposition 2.6 and the fact that
157.172 (3.12)
×xi : Icm (Ω00 ; {0}) −→ Icm−1 (Ω00 ; {0}), ∂xk : Icm (Ω00 ; {0}) −→ Icm+1 (Ω00 ; {0}).
 2. LEFT/RIGHT INVARIANCE 47

So, we look for ut as a formal, for the moment, sum

X
157.173 (3.13) ut ∼ vj , vj ∈ C ∞ ([0, 1]; I m−j (Ω00 , {0})).
j

Take RBM:Clarify
Z 1
157.174 (3.14) v0 = u, vj = Vt vj−1 , j ≥ 1.
t
Then, slightly generalizing the asymptotic summation result (see Problems 157.170
157.173
2) to
include the ‘parameter’ t ∈ [0, 1], we can find ut satisfying (3.13) and hence (3.10).
However, we do know that C c∞ is coordinate-invariant so we have proved the Propo-
sition. 
The question arises as to what the full symbol of F ∗ u might be. The answer is
that it is not so simple to write out because of the iteration. We will deduce a few
things about this complicated formula below, but for the moment notice that the
‘prinicpal symbol’ is given by a relatively simple formula in terms of the Jacobian.
157.175 (3.15)
n n
σm (F ∗ u)(ξ) = | det L|−1 σm (u)((L−1 )t ξ) ∈ S m− 4 /S m−1− 4 (Rn ), L = D0 F.
157.175
157.176 Lemma 3.2. The transformation (3.15) is that of a density on T0∗ Rn .

2. Left/right invariance
For differential operators it is conventional to write the coefficients ‘on the left’
X X
157.178 (3.16) P = pα (x)Dxα , pL (x, ξ) = pα (x)ξ α .
|α|≤m |α|≤m

However one can just as well write them on the right

X X
157.179 (3.17) P = Dxα qα (x), pR (x, ξ) = qα (x)ξ α .
|α|≤m |α|≤m

157.180 Lemma 3.3.

X 1
pL (x, ξ) = ∂xβ Dξβ pR (x, ξ),
n
β!
β∈N0
157.181 (3.18) X (−1)|β|
pR (x, ξ) = ∂xβ Dξβ pL (x, ξ).
n
β!
β∈N0

Here of course only a finite number of terms are non-zero. The formal power series
here are those of an exponential so we can write
157.182 (3.19) pL = exp(Dx · ∂x )pR , pR = exp(−Dx · ∂x )pL
to see that one is the inverse of the other.
Proof. Leibniz’ formula. 
For pseudodifferential operators we can do ‘the same thing’ but it is then not
so clear that we get the same space of operators. For a differential operator with
coefficients written on the right we see, again using the Fourier inversion formula
on S that the operator is given by the formula
Z
157.183 (3.20) F (P u)(ξ) = e−iy·ξ pR (y, ξ)u(y)dy, u ∈ S (Rn ).
Rn
 48 3. THE RING Ψ∗ (Rn )

∞
157.183
157.184 Proposition 3.2. If pR ∈ C ∞ (Rny ; S m (Rnξ )) the operator defined by (3.20) is
m n
157.181
an element of Ψ (R ) with left-reduced symbol, pL , given asymptotically by (3.18).
Proof. We proceed very much as in the proof above. Namely, the Schwartz
157.183
kernel of the operator P in (3.20) is
Z
−n
157.185 (3.21) P (x, y) = B(y, x − y), B(y, z) = (2π) pR (y, ξ)ei(x−y)·ξ dξ
Rn
where we may assume that pR is of very low order to ensure absolute convergence
of the integral (and sort the general case out by continuity). Thus the kernel is
given by introducing the coordinates y and z = x − y in R2n and taking the partial
inverse Fourier transform in z.
So this is very similar to the original ‘left-reduced’ formula except we have
switched x and y as the variable independent of z = x − y on R2n . Using the same
idea as above we can consider a 1-parameter family of ‘quantization maps’ including
RBM:Check QL is earlier left and right as extreme cases
Z
157.186 (3.22) Qt (at ) = (2π)−n at (tx + (1 − t)y, ξ)ei(x−y)·ξ dξ, t ∈ [0, 1]
Rn
and again allow a to vary smoothly with t. The full estimates we are considering
on a are therefore
157.187 (3.23) sup(1 + |ξ|)−m+|β| |∂tk ∂xα ∂ξ a(t, x, ξ)| < ∞ on [0, 1]t × Rnx × Rnξ .
So again, the claim is that the space of kernels, distributions that is, on R2n defined
157.186
by (3.22) is actually independent of t.
To see this we compute, as before, the derivative in t and note that it can be
written
d X d
157.188 (3.24) Qt (at ) = Qt (i ∂ξj ∂xj at + at )
dt j
dt
where the first term comes from the chain rule and integration by parts since
d
dt (tx + (1 − t)y) = x − y and x − y = −i∂ξ i(x − y) · ξ. So, now we want to choose
at so that
X d
157.189 (3.25) i ∂ξj ∂xj at + at is of order − ∞.
j
dt
157.189
In this case we can solve (3.25) explicitly by taking
X tk
157.192 (3.26) at (x, ξ) ∼ (Dx · ∂ξ )k a(x, ξ) = exp(tDx · ∂ξ )a
j
k!
in the sense of formal power series at t = 0.
If we choose at to be an asymptotic sum (uniform in the other variables) as in
157.192
(3.26) then the ‘error term’ is
d
157.193 (3.27) Qt (at ) = Qt (et (x, ξ)), sup(1 + |ξ|)−N |∂tk ∂xα ∂ξβ et | < ∞ ∀ k, α, β.
dt
So we can unload the last step in the proof on the following lemma. 
L5-end

Remark. Undelivered Lecture 6

I did not finish the proof of left/right equivalence last time. Let me not start
at precisely the place I left off, but instead consider the ‘residual’ operators.
 2. LEFT/RIGHT INVARIANCE 49

157.186
157.194 Lemma 3.4. For each t ∈ [0, 1] the quantization Qt , in (3.22) applied to the
residual symbols, which satisfy
157.195 (3.28) sup(1 + |ξ|)N |∂xα ∂ξβ a(x, ξ)| < ∞ ∀ N, α, β
gives the space of kernel of elements of Ψ−∞ (Rn ) are precisely those smooth func-
tions which satisfy
157.196 (3.29) sup(1 + |x − y|)N |∂xα ∂yβ A(x, y)| < ∞ ∀ N, α, β.
157.195
Proof. The functions satisfying (3.28) are exactly the elements of the space
∞
C∞ (Rnx ; S(Rnξ )). Consider the case t = 1, which is the ‘left quantization’ we started
with. Since the Fourier transform is an isomorphism of S (Rn ), the space of kernels
of elements of Ψ−∞ (Rn ) consists of the functions
157.198 (3.30) A(x, y) = B(x, x − y),
where B(x, z) is the (partial) inverse Fourier transform ξ −→ z
Z
157.199 (3.31) B(x, z) = (2π)−n eiz·ξ a(x, ξ)dξ =⇒ B ∈ C ∞ ∞
(Rnx ; S (Rnz )).

Thus, after this change of variable, the space of B’s satisfy the same estimates as
the symbols they are defined by
157.201 (3.32) sup(1 + |z|)N |∂xα ∂zβ B| < ∞.
x,z

The general quantization for t ∈ [0, 1] replaces these kernels by the

157.200 (3.33) At (x, y) = B(tx + (1 − t)y, z)
for the same space of B’s. In terms of the B’s themselves this corresponds to the
change of coordinates (x, z) −→ (X = x − (1 − t)z, z). This is invertible and the
coordinate vector fields tranform to
157.202 (3.34) ∂z = ∂z − (1 − t)∂x , ∂x = ∂X
157.201
from which it is clear that the estimates (3.32) are invariant under such transfor-
mations. Thus
∞
157.203 (3.35) Qt (C ∞ (Rn ; S (Rn ))) = Ψ−∞ (Rn ) ∀ t ∈ [0, 1].


This takes care of the residual terms. 157.184

So, going back to the proof of Proposition 3.2 we are proceeding to construct
∞
a family at ∈ C ∞ ([0, 1] × Rn ; S m (Rn )) so that
dat ∞
157.204 (3.36) + i∂ξ · ∂x at ∈ C ∞ ([0, 1]t × Rn ; S −∞ (Rnξ )), a1 ∈ C ∞
∞
(Rn ; S m (Rn )) given.
dt
∞
To do this we choose successive families vj ∈ C ∞ ([0, 1] × Rn ; S m−j (Rn )) by
(1 − t)j
157.205 (3.37) v0 = a, vj = (i∂ξ · ∂x )j a.
j!
Here I have done the integrals explicitly, so these satisfy
dv0 dvj
157.206 (3.38) = 0, + i∂ξ · ∂x vj−1 = 0 j ≥ 1, v0 = a, vj t=1
= 0, j ≥ 1.
dt dt
 50 3. THE RING Ψ∗ (Rn )

Now we choose at as an asymptotic sum of the vj ’s. This goes beyond the earlier
summation because of the presence of the parmeters t ∈ [0, 1] and x ∈ Rn . What
we want to do is to ensure that the cutoff series
X
157.207 (3.39) Φn k v j
j>k

∞
should converge absolutely with respect to the seminorms of C ∞ ([0, 1]×Rn ; S mj (Rn )).
All the terms have lower order than this. The point is that there are still only a
countable number of norms, even though they now involve the supremum over
157.207
[0, 1] × Rn as well. So absolute convergence can be ensured for each series (3.39) by
choosing the nk large enough. Again this only involves a finite number of conditions
on each nk . 157.193
Once we choose at to be such an asymptotic sum then we get (3.27) and.
following the157.184
discussion of the residual terms above, the complete the proof of
Proposition 3.2.

2.1. Composition. So, finally we are in a position to prove the multiplica-

157.12
tivity of (standard) pseudodifferential operators as in Theorem 1.1:
0 0
A ∈ Ψm (Rn ), B ∈ Ψm (Rn ) =⇒ A ◦ B ∈ Ψm+m (Rn ) and
157.197 (3.40) X 1
σ(AB) ∼ ∂ α σ(A)Dxα σ(B)
n
α! ξ
α∈N0

where σ is the left-reduced full symbol.

157.197
157.614 Remark 2. The asymptotic formula in (3.40) is one version of ‘Moyal’s for-
mula’.
Proof. First I suggest the standard proof, which I will not quite follow through.
The idea is to write A in left-reduced form and B is right-reduced form – now that
we know they are equivalent. Thus
Z
−n
157.208 (3.41) Au(x) = (2π) aL (x, ξ)eix·ξ û(ξ)dξ,
Z
F (Bv)(ξ) = bR (y, ξ)e−iy·ξ v(y)dy, u, v ∈ S (Rn ).

We can assume that the symbols themselves are of order −∞ and use density. The
composite is then
Z
157.209 (3.42) (AB)v(x) = (2π)−n ei(x−y)·ξ aL (x, ξ)bR (y, ξ)v(y)dydξ

This is almost what we want, except the ‘amplitude’ in the integral depends ex-
plicitly on both x and y, as well as ξ. So we need to show that the kernel of the
composite
Z
157.210 (3.43) K(x, y) = (2π)−n ei(x−y)·ξ aL (x, ξ)bR (y, ξ)dξ ∈ Ψm+m (Rn ).

This can be proved by an argument very similar to the left/right reduction. I leave
the details to you!
 2. LEFT/RIGHT INVARIANCE 51

Let’s return for a moment to the spaces of conormal distributions at the origin,
ISm (Rn ; {0}). We can easily define conormal distributions at another point simply
by translation. Thus, if p ∈ Rn ,
157.211 (3.44) ISm (Rn ; {p}) = {u ∈ S 0 (Rn ); u(x + p) = T−p
∗
u ∈ ISm (Rn )}.
This is made more convincing by the proof of coordinate invariance. Here Tq is
translation by q ∈ Rn , Tq x = x + q.
157.212 Exercise 6. Define, and the formulate (and prove) the coordinate-invariance
of, the spaces Icm (Ω, {p}) for p ∈ Ω ⊂ Rn open.
157.213 Lemma 3.5. The Schwartz kernels of elements of Ψm (Rn ) may be identified
with the space
∞ m+ n
157.214 (3.45) C∞ (Rny ; IS 4
(Rn ; {0})) by A(x, y) 7−→ A(x − y, y).
Proof. 

Using earlier results we have another method. What we have shown above, in
left/right reduction is that the kernel of an element of Ψm (Rn ) 

If there is a little time left today I want to introduce another algebra of pseu-
dodifferential operators. This is a sign of things to come. I have been rather hard
∞
on the ‘coefficient ring’ C ∞ (Rn ) which is involved in the ring Ψm (Rn ). What is a
‘nicer’ possibility? The one I have in mind is the symbol space itself. We can easily
introduce the space of ‘symbol-valued symbols’ (in either direction)

157.215 (3.46) S m,k (Rn ; Rn )

= {a ∈ C ∞ (R2n
x,ξ ); sup(1 + |x|)
−k+|α|
(1 + |ξ|)−m+|β| |∂xα ∂ξβ a(x, ξ)| < ∞, ∀ α, β}.
What I mean by symbol-valued symbols is that as a smooth map from the first set
of variables,
a : Rnx −→ S m (Rn ).
157.215
Moreover, if k · km,N are the seminorms on S m (Rnξ ) then the estimates (3.46) are
equivalent to
157.616 (3.47) sup(1 + |x|)−m+|α| k∂xα a(x, ·)kk,N < ∞ ∀ α.
x

These defining conditions give seminorms. Here there are two orders and differ-
entiation with respect to x lowers the second (but not the first) and differentiation
with respect to ξ lowers the first but not the second.
157.215
Directly from (3.46) we see that
∞
157.216 (3.48) k ≤ 0 =⇒ S m,k (Rn ; Rn ) ⊂ C ∞ (Rn ; S m (Rn )).
It is also the case that
(1 + |x|2 )k/2 ∈ S 0,k (Rn ; Rn ) and
157.217 (3.49) 0 0 0 0
S m,k (Rn ; Rn ) · S m ,k (Rn ; Rn ) = S m+m ,k+k (Rn ; Rn ).
Combining these two observations we see that
157.218 (3.50) (1 + |x|2 )−k/2 S m,k (Rn ; Rn ) ⊂ C ∞
∞
(Rn ; S m (Rn ).
 52 3. THE RING Ψ∗ (Rn )

So, we can quantize these double symbols using left quantization and define

157.219 (3.51) Ψm,k n n n

sc (R ) = {A : S (R ) −→ S (R );

A = (1 + |x|2 )k/2 QL ((1 + |x|2 )−k/2 a), a ∈ S m,k (Rn ; Rn )}.

MR0273463
This algebra was introduced by Shubin, [8], but I call it the scattering algebra,
which is the subscript sc, because it has a direct extension to compact manifolds
with boundary (as I hope we will see).
So, I am getting ahead of myself here:
157.220 Proposition∗ 3.3. The scattering psedudofferential operators from a double-
filtered algebra
0 0 0 0
157.221 (3.52) Ψm,k n m ,k
sc (R ) ◦ Ψsc (Rn ) ⊂ Ψsc
m+m ,k+k
(Rn )
with residual space
\
157.222 (3.53) Ψ−∞,−∞
sc (Rn ) = Ψm,k n
sc (R )
m,k

equal to the space of operators with kernels A ∈ S (R2n ).

Maybe you would like to try your hand at proving this! You can easily see
why it should be true because Moyal’s formula for the composite of two
such operators gives
X 1
σL (AB) ∼ ∂ξ σL (A) · Dx σL (B) 157.223 (3.54)
α
α!
and the individual terms here are in
0
,k0 −|α| 0
−|α|,k+k0 −|α|
S m−|α|,k (Rn ; Rn ) · S m (Rn ; Rn ) ⊂ S m+m (Rn157.224
; Rn ) (3.55)
which is decreasing in both orders. It takes a little thought to prove that
everything ‘works’ correctly; here is an outline of one approach – where I
will use Kumano-go’s double symbols.
First go through the left/right reduction argument in this case. For con-
venience I take k ≤ 0 because we can always recover the general case by
multiplying by (1 + |x|2 )k/2 . So, we want to choose a 1-parameter family of
double symbols,
at ∈ C ∞ ([0, 1]t ; S m,k (Rn ; Rn )) 157.225 (3.56)
157.189 157.205
so that the identity (3.25) holds in the new sense. Looking at (3.37) we can
see that if v0 = a is chosen in S m,k (Rn ; Rn ) then the vj ∈ C ∞ ([0, 1]; S m−j,k−j (Rn ; Rn )
have both orders decreasing. Going back to the asymptotic summation
lemma we now need to do a little more with our cutoffs. So in this case we
would consider all the series
X
(1 − φ(x/nj ))φ(ξ/nj )vj (t, x, ξ) 157.226 (3.57)
j>l

where we cut out the region where both |x| and |ξ| are less than nj . So
on the support of (1 − φ(x/nj )φ(ξ/nj )) either |x| > nj or |ξ| > nj but
the symbol lies in the space S m−j,k−j (Rn ; Rn ) with j > l. So this is small
in the symbol space S m−l,k−l (Rn ; Rn ) if we choose nj large enough. This
means we can make all the series converge by an appropriate choice of the
integers nj and then the error term and then the error term
d
Qt (at ) ∈ S (R2n ) 157.227 (3.58)
dt
is a residual operator in the new sense. So we win and we see that Qt (S m,k (Rn ; Rn ) =
Ψm,k n
sc (R ) for all t ∈ [0, 1].
 3. ISOTROPIC ALGEBRA 53

Now to Kumano-go’s result. Suppose we ‘overspecify’ the amplitude of the

pseudodifferential operator by taking a ‘triple symbol’

b(x, y, ξ) ∈ S m,k,k (Rn ; Rn ; Rn ) = S k (Rn k n m n

x ); S (Ry ); S (Rξ ). 157.228 (3.59)

This means we consider smooth functions on R3n which satisfy

sup(1 + |x|)−k+α
(1 + |y|) −k+|γ|
(1 + |ξ|) −m+|β|
|∂xα ∂yγ ∂ξβ b(x, y, ξ)| 157.229
< ∞. (3.60)

This defines a countably normed space. Notice that these are symbols ‘sep-
arately’ in all the variables, there is no joint decay.

157.230 Proposition 3.4. [Kumano-go] The ‘overspecified’ quantization map

Z
Q : b −→ ei(x−y)·ξ b(x, y, ξ)dξ ∈ Ψm,k n
sc (R ). 157.231 (3.61)
Rn

Proof. We can think of the double symbols a(x, ξ)157.231 as special cases
of the triple symbols which are indpendent of y. Then (3.61) is standard
quantization, so the range certainly contains Ψm,k n
sc (R ). To see that it con-
tains nothing more, we can use the same deformation argument as above
and try to construct a family of triple symbols bt so that the intermediate
quantization maps
Z
Qt : bt −→ ei(x−y)·ξ bt (x, tx + (1 − t)y, ξ)dξ 157.232 (3.62)
Rn

have derivative a smoothing, here meaning Schwartz, kernel. 

3. Isotropic algebra
S.Isotropic MR2650633
There is a text on this by Parmeggiani [7].
157.468 Remark. Edited by Paige Dote.
∗,∗
The calculus Ψsc (Rn ), due to Shubin, is described above. Let me introduce
yet another algebra of pseudodifferential operators on Rn . This one, as we will see
later, has direct topological applications, whereas the scattering algebra is more of
geometric significance. The symbols considered are the ‘pure symbols’ on R2n =
Rnx × Rnξ ,

S m (R2n ) = {a ∈ C ∞ (R2n ); ∀α, β ∈ Nn0 , sup(1 + |x| + |ξ|)−m+|α|+|β| |∂xα ∂ξβ a| < ∞}.
So there is no difference in behavior between the x and ξ variables. That is what
‘isotropic’ is supposed to indicate here, meaning ‘the same in all directions’.
Rather than repeat the basic constructions again we can use the properties of
∗,∗
Ψsc (Rn ) in view of the following Lemma:
157.471 Lemma 3.6. The pure symbols and symbol-valued symbols are related by the
following:

m 2n m n m n
S (R ) ⊂ S (Rx ; S (Rξ )),
 m≥0
m 2n m/2 n m/2 n
157.470 (3.63) S (R ) ⊂ S (R ; S (R )), m ≤ 0 .
 −∞ 2n
 −∞ n −∞
S (R ) = S (R ; S (Rn ))

Proof. These results follow from

max{1 + |x|, 1 + |ξ|} ≤ (1 + |x| + |ξ|) ≤ (1 + |x|)(1 + |ξ|).
 54 3. THE RING Ψ∗ (Rn )

157.470
This gives the leading estimates in the first two statements in (3.63) since a ∈
S m (R2n ) satisfies
|a(x, ξ)| ≤ (1 + |x|)m (1 + |ξ|)m , m ≥ 0
|a(x, ξ)| ≤ (1 + |x|)m/2 (1 + |ξ|)m/2 , m ≥ 0.
In S m (R2n ) an x-derivative corresponds to an extra decay factor of (1+|x|+|ξ|)−1 ≤
(1 + |x|)−1 and a ξ-derivative does as well but (1 + |x| + |ξ|)−1 ≤ (1 + |ξ|)−1 , giving
all of the ‘double’ symbol estimates. The last statement follows from the second
last. 
Thus, using left quantization we can define
(3.64) Ψm n m 2n
iso (R ) = QL (S (R(x,ξ) ))

and it follows that

(
Ψm n m,m n
iso (R ) ⊂ Ψsc (R ), m≥0
(3.65) m/2,m/2 .
Ψm n
iso (R ) ⊂ Ψsc (Rn ), m≤0
157.472 Theorem 3.1. The isotropic operators form an ∗-closed filtered algebra of op-
erators with a well-defined principal symbol map giving a short exact sequence
σm
157.473 (3.66) Ψm−1 n m n m
iso (R ) ,→ Ψiso (R )  S /S
m−1
(R2n ),
157.474 (3.67) σm+m0 (AB) = σm (A)σm0 (B)
0
for A ∈ Ψm n m n
iso (R ), B ∈ Ψiso (R ).

Proof. We know that the Ψ∗,∗ n

sc (R ) is a (bi-)filtered algebra with products
asymptotic to Moyal’s formula. Hence, in this sense,
0
A ∈ Ψm n m n
iso (R ), B ∈ Ψiso (R ) =⇒ AB = QL (c)

for some c ∈ S k (Rn ; S k (Rn )) and some k where

X 1
c∼ (Dξα a)(Dxα b).
α
α!
0
The the series here has terms in in S k−j,k−j . Now since a ∈ S m (R2n ), b ∈ S m (R2n )
the terms in the Moyal series are
1 α 0
∂ξ a · Dxα b ∈ S m+m −2|α| (R2n ).
α!
Thus, this series is asymptotic in the isotropic sense. It follows that we can choose
an asymptotic sum
0 X 1
S m+m (R2n ) 3 c̃ ∼ ∂ξα a · Dxα b.
α
α!
157.471
Hence, it follows from Lemma 3.6, despite the 12 s in the orders, that
X 1
c̃ ∼ ∂ξα a · Dxα b
α
α!
in the sense of symbol-valued symbols. From the uniqueness of the asymptotic sum
157.471
in this sense, and the last part of Lemma 3.6, it follows that
0 0
m+m
Ψm n m n
iso (R ) · Ψiso (R ) ⊂ Ψiso (Rn )
 4. L2 BOUNDEDNESS 55

as claimed. Thus the Moyal product

157.473
gives an asymptotic formula for the symbol of
157.474
a product so (3.66) follows, with (3.67) essentially being the definition of σm .
The closure under passage to adjoints is similar. 

Since these are lectures, one can ask rather impertinent questions such as: Why
would one be interested in this? One practical answer is that the algebra is closely
related to the harmonic oscillator on Rn .
Definition 3.1 (Harmonic Oscillator). The harmonic oscillator, H, on Rn is
defined as X X
H = ∆ + |x|2 , where ∆ = Dx2i = − ∂x2i .

Indeed, H ∈ Ψ2iso (Rn ) is elliptic. I hope that somewhere below I will show how
one can use the isotropic algebra to prove Thom isomorphism in K-theory. This
is a very special case of the familiar Atiyah-Singer Index Theorem that I want to
describe below. The isotropic algebra is also an integral part of the proof, at least
for the proof I have in mind.
Now, for elliptic operators, such as H, in the isotropic calculus we can deduce
the existence of a parameterix just as is done microlocally in the standard case in
the next chapter and globally for manifolds later. It is significant that the error
here is compact.
157.475 Proposition 3.5. If A ∈ Ψm n
iso (R ) is elliptic then it has a two-sided parame-
−∞ n
terix modulo the ideal ΨS (R ).
Proof. By definition, A = QL (a) where a ∈ S m (R2n ) is globally elliptic. Thus
for some δ > 0
1
|a(x, ξ)| ≥ δ(|x| + |ξ|)m in |x| + |ξ| ≥ ,
δ
1−ϕ(x,ξ) −m 2n ∞ 2n
then b(x, ξ) = a(x,ξ) ∈ S (R ) if ϕ ∈ Cc (R ) is equal to 1 on the ball
|x| + |ξ| ≤ 1δ . 

4. L2 boundedness
Project 2 P-Bounded
In this project I would like you to go through some ‘symbolic arguments’, giving
L2 boundedness of pseudodifferential operators.

4.1. Schur’s criterion. This is the same Schur as the lemma about irre-
ducibility, hence I just say ‘criterion’. This is quite a handy sufficient condition for
L2 boundedness in terms of the Schwartz kernel. It can be generalized to measure
spaces (and so manifolds), but for the moment let’s think about Rn . Then
Schur Proposition 3.6. If A : Rn −→ C is a Lebesgue measurable function which
satisfies
Z Z
P2.2 (3.68) sup |A(x, y)|dy, sup |A(x, y)|dx < ∞
x y

then the integral operator (say defined initially on C c (Rn )

Z
P2.1 (3.69) Au(x) = A(x, y)u(y)dy is a bounded operator on L2 (Rn ).
Rn
 56 3. THE RING Ψ∗ (Rn )

Proof. You might like to look it up, it is basically just a clever use of Schwarz
inequality. 
Problem 2.1
Show that if A ∈ Ψm (Rn ) with m < −n then the Schwartz kernel is continuous
and satisfies
P2.3 (3.70) sup(1 + |x − y|)N |A(x, y)| < ∞ ∀ N.
x,y

Deduce that Schur’s criterion applies and hence conclude L2 boundedness.

In fact you can push this argument so that it applies for m < 0 but not up to
m = 0 (think of the identity).
Problem 2.2
For A ∈ Ψ0 (Rn ) construct Q ∈ Ψ0 (Rn ) such that
P2.4 (3.71) Q = Q∗ , Q2 = C Id −A∗ A + E, C > 0 constant, E ∈ Ψ−1 (Rn )
‘Hint’: It is enough to choose C > sup |a|2 where A = QL (a). Then show that
1
∞
q = (C − |a|2 ) 2 ∈ C ∞ (Rn ; S 0 (Rn )) and the set Q = 12 (QL (a)) + QL (a)∗ .
Problem 2.3P2.4
Now we want to improve the ‘error’ in (3.71). Show that if E ∈ Ψ−k (Rn ),
k ≥ 1, and E ∗ = E where E = QL (e) then the choice
P2.5 (3.72)
B = QL (e/q)+QL (e/q)∗ satisfies (Q−B)2 = C Id −A∗ A−E 0 , E 0 ∈ Ψ−k−1 (Rn ), (E 0 )∗ = E 0 .
Problem 2.4
Using
P2.4
this show that we may ‘correct’ Q (by adding a lower order term) so
that (3.71) holds with E ∈ Ψ−N (Rn ) for any preassigned N. (Using asymptotic
summation this works for N = −∞.
Problem 2.5
Finally deduce L2 boundedness in the sense that A ∈ Ψ0 (Rn ) extends by con-
tinuity from A : S (Rn ) −→ S (Rn ) to a bounded operator on L2 (Rn . P2.4
‘Hint’. This whole argument is due to Hörmander. It follows from (3.71) that,
n 2
for φ ∈ S (R ), in terms of the L inner product
P2.6 (3.73) 0 ≤ hQφ, Qφi = hQ2 φ, φi = Ckφk2L2 − kAφk2L2 + hEφ, φi.
So, if we know that boundedness of E (which we do) then
1
P2.7 (3.74) kAφkL2 ≤ (C + C 0 ) 2 kukL2 .
where C 0 comes from E.
Problem 2.6: Sobolev boundedness
The Sobolev space H s (Rn ) is defined as consisting of those elements of S 0 (Rn )
(becauese we are allowing s ≤ 0 such that
P2.8 (3.75) (1 + |ξ|2 )s/2 û ∈ L2 (Rn ).
Deduce that the operator (1 + |D|2 )t = QL ((1 + |ξ|2 )t/2 ) = QR ((1 + |ξ|2 )t/2 ) ∈
Ψt (Rn ), for any t ∈ R, is an isomorphism
P2.9 (3.76) (1 + |D|2 )t/2 : H s (Rn ) −→ H s−t (Rn ).
From this, L2 boundedness and the properties of the calculus deduce that
P2.10 (3.77) A ∈ Ψm (Rn ) =⇒ A : H s (Rn ) −→ H s−m (Rn ).
 4. L2 BOUNDEDNESS 57

‘Hint’: Consider for instance (1 + |D|2 )−m+s/2 A(1 + |D|2 )−s/2 .

Problem 2.7
For anyone who has read the section on the scattering (Shubin) calculus define
the weighted Sobolev spaces
P2.11 (3.78) H s,t (Rn ) = {u ∈ S 0 (Rn ); (1 + |x|2 )t/2 u ∈ H s (Rn ).
(1) Show that for any real orders
P2.12 (3.79) A ∈ Ψm,k n s,t n
sc (R ) =⇒ A : H (R ) −→ H
s−m,t−l
(Rn ).
(2) Show that
P2.13 (3.80) F : H s,t (Rn ) −→ H t,s (Rn ), ∀ s, t.
0 0
(3) Show that, in contrast to the usual Sobolev spaces, the inclusion H s ,t (Rn ) ,→
H s,t (Rn ) for s0 > s, t0 > t is compact.
Problem 2.7
For anyone who has followed the discussion above of the isotropic calculus.
(1) Define the isotropic Sobolev spaces in terms of the operators
P2.15 (3.81) Gs = QL ((1 + |x|2 + |ξ|2 )s/2 ∈ Ψsiso (Rn )
by
P2.16 (3.82) s
Hiso (Rn ) = {u ∈ S 0 (Rn ); As u ∈ L2 (Rn ).
(2) Show that for any
 CHAPTER 4

Ellipticity and wavefront set

L6
In the actual Lecture 6 I got a little carried away but let me record here what
I tried to cover. So, I am reviewing what we have done, or in some cases partly
done, and then expanding on it a little.
) ⊂ C ∞ (Rn ) are Fréchet spaces defined by the
• Symbols:- The spaces S m (R157.99
n

finiteness of the norms in (2.55). They form a filtered157.102

(abelian) ring with
identity 1 ∈ S 0 (Rn ) and157.140
have density (Proposition 2.5) and asymptotic
completeness (Theorem 2.2) properties.
When is an element invertible? For a ∈ S m (Rn ) to have an inverse
−m
in S (Rn ) a necessary and sufficient condition is
157.233 (4.1) |a(ξ)| ≥ δ(1 + |ξ|)m ⇐⇒ a−1 ∈ S −m (Rn ).
157.233
The necessity of (4.1) follows from the bound |a−1 (ξ)| ≤ C(1 + |ξ|)−m .
Conversely this certainly implies that
1
b(ξ) = ∈ C ∞ (Rn ) 157.234 (4.2)
a(ξ)
and the derivatives are then of the form
eα
∂ α b = |α|+1 157.235 (4.3)
a
where eα is a symbol or order (m − 1)|α|. This is clear for |α| = 0 and
follows by induction since taking one more derivative shows that
a∂ξj eα − (|α| + 1)∂ξj a
∂ ξj ∂ α b = 157.346 (4.4)
a|α|+2
giving the inductive step. The symbol estimates on b follow.
m+ n
• We defined IS 4 (Rn ) = F −1 (S m (Rn )) and derived various properties of
these conormal distributions at 0.
∞
• For such a Fréchet space we can define C ∞ (Rm ; S m (Rn )) as the subspace
∞ m n
of C (R × R ) with all derivatives with respect to the first variables
bounded in terms of the seminorms on S m (Rn ). Then our definiition of
Ψm (Rn ) is in terms of their Schwartz kernels (which we identify with the
operators)
∞ m+ n
157.236 (4.5) A ∈ Ψm (Rn ) =⇒ A(x, x − z) ∈ C ∞ (Rnx ; IS 4
(Rnz )).
This corresponds to the ‘quantization map’ in terms of the partial Fourier
tranform
157.237 (4.6) ∞
C∞ (Rm ; S m (Rn )) 3 a(x, ξ) −→ QL (a) = Fξ−1 (a)(x, x − y) ∈ Ψm (Rn ).
• We showed (most of the fact that) that for each t ∈ [0, 1] the ‘intermediate
quantizations’
157.238 (4.7) Qt (a) = Fξ−1 (a)(tx + (1 − t)y, x − y) ∈ Ψm (Rn )
59
 60 4. ELLIPTICITY AND WAVEFRONT SET

give the same space of operators. For t = 1 this is ‘left’ quantization and
for t = 0 it is ‘right’ quantization where the kernel is written as
∞ m+ n
157.239 (4.8) A(x, y) = B(y, x − y), B ∈ C ∞ (Rny ; IS 4
(Rn )).
• Note that the case t = 21 is also of importance. It is called ‘Weyl quanti-
zation’ and means writing the kernel as
x+y ∞ m+ n
157.240 (4.9) A(x, y) = B( , x − y), B ∈ C ∞ (Rny ; IS 4 (Rn )).
2
It has some useful properties.
• The inverses of these quantization maps are the ‘total symbols’
∞
157.241 (4.10) σL , σR , σW : Ψm (Rn ) −→ C ∞ (Rny ; S m (Rn )).
• The right and left symbols are related asymptotically by
X 1
157.244 (4.11) σR (A) ∼ Dxα ∂ξα σL (A) = exp(Dx · ∂ξ )σL (A)
α
α!
where the exponential is to be formally expanded in Taylor series at 0.
157.248 Exercise 7. Derive a similar asymptotic relationships between σW
and σL .
• (Not discussed in lecture) The ‘formal’ (just meaning non-Hilbert space)
adjoint is defined for any continuous linear operator A : S (Rn ) −→ S 0 (Rn )
by duality
Z Z
157.242 (4.12) A∗ : S (Rn ) −→ S 0 (Rn ), (Au)vdx = A∗ vdx
Rn Rn
(where the distribution pairing is written as an integral). Then
∗
157.243 (4.13) : Ψm (Rn ) −→ Ψm (Rn ), σR (A∗ ) = σL (A)).
• The composition theorem with
X
157.246 (4.14) σL (A ◦ B) ∼ (∂ξα σL (A))(Dxα σL (B)).
α
As suggested in Lecture, check this for differential operators. In fact it is
enough
157.246
to take A = Dxγ and B = b(x) and apply Leibniz’ formula to get
(4.14).
• (Also not discussed at all). The elements of Ψm (Rn ) define by bounded
linear maps for any M on the standard Sobolev spaces
157.247 (4.15) A : H M (Rn ) −→ H M −m (Rn ).
Proof later. Here H M (Rn ) = F −1 ((1 + |ξ|)−M L2 (Rn )) is defined as usual
as the inverse Fourier transform of the weighted L2 spaces on the dual.

1. Ellipticity of symbols
S.EllipticSymbols
In the notes above the notion of ellipticity for elements of S m (Rn ) is discussed
(although I did not cover this in lectures). We want an extension of this idea
to C ∞ (RN ; S m (Rn )) (I have dropped both the boundedness assumptions on the
coefficents and the assumption that N = n since they are both irrelevant here).
¯ ∈ RN × (Rn \ {0}) if
Most importantly a symbol is said to be elliptic at (x̄, ξ)
‘it is as big as it can be’ in a cone around this point.
 1. ELLIPTICITY OF SYMBOLS 61

157.249 Definition 4.1. The elliptic set

Ell(a) = Ellm (a) ⊂ RN × (Rn \ {0}) of a ∈ C ∞ (RN ; S m (Rn ))
consists of those points x̄, ξ¯ corresponding to which there exists δ > 0 such that
ξ ξ¯
157.250 (4.16) |a(x, ξ)| > δ|ξ|m in |x| < δ, | − ¯ | < δ, |ξ| > 1/δ.
|ξ| |ξ|
I have engaged in constant-saving here! Clearly the result remains true if δ
is decreased but remains positive. The basic region we are looking at here is a
‘conic neighbourhood’ of (c̄, ξ) ¯ which is then truncated by demanding |ξ| is large
as well. So the appearances of δ can be decreased individually and the estimate
remains true. Ellipticity is a ‘local invertibility’ condition on a in the filtered symbol
algebra, as shown below. It is a conic set from the definition which only depends
¯ ξ|
on ξ/| ¯ not ξ¯ itself. Thus
157.251 (4.17) ¯ ∈ Ell(a) =⇒ (x̄, tξ)
(x̄, ξ) ¯ ∈ Ell(a), t > 0.
It is also clear that Ell(a) is open for any a ∈ C ∞ (RN ; S m (Rn )) as a subset of
RN × (Rn \ {0}), which is itself open in RN +n . Of course it could be empty, and it
0
certainly is if a ∈ C ∞ (RN ; S m (Rn )), m0 < m. So one should really write
157.253 (4.18) Ellm (a) ⊂ RN × (Rn \ {0}) defined for a ∈ C ∞ (RN ; S m (Rn ))
but we treat the ‘m’ as understood from context.
We also give a name to the complement of the elliptic set, it is called the
characteristic set of the symbol
Char(a) = RN × (Rn \ {0} \ Ell(a).


157.252 (4.19)
It is then a (relatively) closed subset.
We define a third conic set corresponding to the region where the symbol is not
locally rapidly decaying with all derivatives as follows
157.254 (4.20)
¯ ∈ RN × (RN \ {0}); ∃ δ > 0 and φ ∈ C ∞ (RN ; S (Rn ))

conesupp(a) = (x̄, ξ)
1 ξ ξ¯ {
with a = φ in |x − x̄| < δ, |ξ| > , | − ¯|<δ
δ |ξ| |ξ|
It follows that conesupp(a) is relatively closed and that
157.255 (4.21) Ell(a) ⊂ conesupp(a)
where this relation is like that between the sets {u 6= 0} and supp(u) for a smooth
function.
In terms of mutliplication of symbols it is easy to see that
Ell(ab) = Ell(a) ∩ Ell(b),
157.256 (4.22) Char(ab) = Char(a) ∪ Char(b),
conesupp(ab) ⊂ conesupp(a) ∩ conesupp(b)
¯ but have
Note that we can construct symbols which are elliptic at a point (x̄; ξ)
cone support in any conic neighbourhood
ξ ξ¯
157.259 (4.23) Cδ = {(x, ξ) ∈ RN × (Rn \ {0}); |x − x̄| < δ, | − ¯ | < σ}, δ > 0.
|ξ| |ξ|
 62 4. ELLIPTICITY AND WAVEFRONT SET

Indeed to arrange this we just need to choose a smooth function on the sphere, ψδ
ξ ξ̄ ξ
which is equal to one in | |ξ| − ω| < δ/2, where ω = |ξ̄| , is positive on | |ξ| − ω| < δ/2
ξ
and has support in | |ξ| − ω| ≤ δ in terms of the distance on the sphere. Similarly
∞
choose ψ ∈ C c (R ) positive on |x| < 1, equal to 1 in |x| < 21 and supported in
N

|x| ≤ 1 and a similar cutoff µ ∈ C ∞ (Rn ) and consider

x − x̄ ξ
157.260 (4.24) cδ (x, ξ) = (1 − µ(δξ))ψ( )ψδ ( ) ∈ C c∞ (RN ; S 0 (Rn )).
δ |ξ|
157.259 157.260
157.261 Lemma 4.1. In terms of (4.23) the ‘symbolic cutoff ’ χδ in (4.24) has
157.262 (4.25) Ell(cδ ) ⊃ Cδ0 , ∀ δ 0 < δ
(4.26) conesupp(cδ ) ⊂ Cδ ⊂ Cδ00 ∀ δ 00 > δ
157.265 (4.27) conesupp(1 − cδ ) ∩ Cδ0 /2 = ∅ ∀ δ 0 < δ.
Proof. Inspection. 
157.257 ¯ ∈ Ell(a) then there exists b ∈
Lemma 4.2. If a ∈ C ∞ (RN ; S m (Rn )) and (x̄, ξ)
∞ N −m n
C (R ; S (R )) such that
157.258 (4.28) ¯ ∈
(x̄, ξ) / conesupp(ab − 1).
157.260
Proof. Take a symbolic cut-off as in (4.24) and consider
χδ (x, ξ)
157.263 (4.29) b= .
a
For δ > 0 small enough this is well-defined since by the definition of ellipticity in
157.250
(4.16), a 6= 0 on the support of χδ (x, ξ); as usual 157.265
157.258
the quotient is extended as zero
outside this support. Then (4.28) follows from (4.27), since ba = χδ (x, ξ). So it
only remains to check that b ∈ C c∞ (RN ; S −m (Rn )). Proceeding inductively
gα,β
157.264 (4.30) ∂xα ∂ξβ b = |α|+|β|+1 , gα,β ∈ C c∞ (RN ; S (m−1)|α|+m|β| (Rn )).
a
This is certainly true for α = β = 0 and the inductive step follows by differentiating
again with respect to either variable. 
It is important to note that

157.266 (4.31) Ellm (a + e) = Ellm (a) if a ∈ C ∞ (RN ; S m (Rn ))

and e ∈ C ∞ (RN ; S m− (Rn )),  > 0.
The same is true for Char(a) whereas
157.267 (4.32) conesupp(a + e) = conesupp(a) if e ∈ C ∞ (RN ; S −∞ (Rn )).

2. Ellipticity of pseudodifferential operators

We now transfer the these notions from symbols to pseudodifferential operators.
∞
157.268 Definition 4.2. If A = QL (a) ∈ Ψm (Rn ), a ∈ C ∞ (Rn ; S m (Rn )) we set
157.269 (4.33) Ell(A) = Ellm (A) = Ellm (a),
157.270 (4.34) Char(A) = Charm (A) = Char(a),
157.271 (4.35) WF0 (A) = conesupp(a).
 2. ELLIPTICITY OF PSEUDODIFFERENTIAL OPERATORS 63

Here WF0 (A) is called the ‘operator wavefront set’ of A for reasons that should
become clearer below – it is just a name.
Recall that we defined the principal sybmol of A to be σm (A) = [a] to be the
∞ ∞
equivalence class in C ∞ (Rn ; S m (Rn ))/C ∞ (Rn ; S m−1 (Rn )) and then it is also equal
to the equivalence class of the right-reduced symbol.
157.272 Lemma 4.3. The elliptic set only depends on σm (A) and WF0 (A) depends on
a modulo symbols of order −∞ and is also equal to the cone-support of the right
reduced symbol; for the product of operators
157.273 (4.36) Ellm+m0 (AB) = Ellm (A) ∩ Ellm0 (B), WF0 (AB) ⊂ WF0 (A) ∩ WF0 (B).
157.266
Proof. The first part follows directly from (4.31) and the fact that left- and
right-reduced symbols differ by a term of order m − 1. The last part is a little
more subtle, and depends on the formula for the asymptotic expansion of the right-
reduced sybmol in terms of the left-reduced symbol a.
X 1
157.274 (4.37) Dxα ∂ξα a.
α
α!
157.254 157.274
If a is rapidly decreasing in a truncated cone, as (4.20) then all the terms in (4.37)
are rapidly decaying in the same cone, because of the locality of differential oper-
ators. 157.273
It follows that any asymptotic sum is rapidly decreasing as well. The final
part (4.36) follows similarly. 
Perhaps the most important construction associated to these definitions is ‘mi-
crolocal invertibility’ at elliptic points.
157.277 ¯ ∈ Ell(A), A ∈ Ψm (Rn ), there exists B ∈ Ψ−m (Rn )
Proposition 4.1. If (x̄, ξ)
such that
157.278 (4.38) ¯ ∈
(x̄, ξ) / WF0 (Id −AB) ∩ WF0 (Id −BA).
Proof. As the notation suggests, we start with
157.257
157.279 (4.39) B0 = QL (b), b as in Lemma4.2.
The properties of b mean that
ξ ξ¯
157.281 (4.40) WF0 (B0 ) ⊂ Cδ = {(x, ξ) ∈ Rn ξ(Rn \ {0}); |x − x̄| ≤ δ, k
− ¯ ≤ δ}
|ξ| |ξ|
where we are free to choose δ > 0. Then, from the product formula for symbols,
157.280 (4.41) B0 A = QL (cδ )) − E, E ∈ Ψ−1 (Rn ), WF0 (E) ⊂ Cδ .
Now, we can almost invert Id −E using the Neumann series. That is we can
choose
X
157.282 (4.42) F ∈ Ψ−1 (Rn ), F ∼ E k =⇒
k≥1

(Id +F )(Id −E) = Id +EL0 , 0

(Id −E)(Id +F ) = Id +ER , EL0 , ER
0
∈ Ψ−∞ (Rn ).
Define
157.283 (4.43) B = (Id +F )B0 =⇒
BA = (Id +F )(QL (cδ ) − E) = Id +EL0 − (Id +F )(Id −(QL (cδ )) = Id +E 00 ,
¯ ∈
(x̄, ξ) / WF0 (E 00 ).
 64 4. ELLIPTICITY AND WAVEFRONT SET

Similarly we can proceed on the right,

X
k
157.284 (4.44) AB0 = QL (cδ ) − ER , Fr ∼ ER
k≥1

and see again that BR = B0 (Id +FR ) satisfies

157.285 (4.45) 00
ABR = Id +ER ¯ ∈
, (x̄, ξ) / WF0 (ER
00
).
Now, it is a form of the argument which gives the ‘uniqueness of the inverse in
a group’ to see that

157.286 (4.46) B = BABR + S1 = BR + S1 − S2 ,

¯ ∈
S1 = B(Id −ABR ), S2 = (Id −BA) so (x̄, ξ) / WF0 (Si ), i = 1, 2.
157.285 157.278
¯ ∈
It follows that (x̄, ξ) / WF0 (BL − BR ) so B also satisfies (4.45) and (4.38). 

157.347 Exercise 8. [Microlocal partition of unity] Suppose that K b Rn × Sn−1

and Ua ⊂ Rn × Sn−1 , a ∈ A, is an open cover of K then there exist operators
Ai ∈ Ψ0 (Rn ), i = 1, . . . , N such that, in terms of the cones in Rn × (Rn \ {0}),
X
(4.47) WF0 (Id − Ai ) ∩ R+ K = ∅, WF0 (Ai ) ⊂ R+ Uai for some ai ∈ A.
i

3. Wavefront set of a distribution

S.WF
Now we are in a position to define the wavefront set (or wavefrontset) of a
distribution on Rn . First let’s work with compactly supported distributions and
then pass to the general case.
157.275 Definition 4.3. If u ∈ C c−∞ (Rn ) then
¯ ∈ Rn × (Rn \ {0});

157.276 (4.48) WF(u) = (x̄, ξ)
¯ ∈ Ellm (A) {
∃ A ∈ Ψm (Rn ), Au ∈ S (Rn ), (x̄, ξ) .
We can characterize the wavefront set of a distribution in a more elemenatary
way.
157.287 ¯ ∈ Rn × (Rn \ {0})
Proposition 4.2. For u ∈ C c−∞ (Rn ) and (x̄, ξ)

157.288 (4.49) ¯ ∈
(x̄, ξ) / WF(u) ⇐⇒
ξ¯
∃ φ ∈ C c∞ (Rn ), φ(x̄) 6= 0, ψ ∈ C c∞ (Sn−1 ), ψ( ¯ ) 6= 0 s.t.
|ξ|
ξ
|ψ( )F (φu)| ≤ CN (1 + |ξ|)−N ∀ N.
|ξ|
Proof. One way is straightforward, the other way depends on the construction
of microlocal invserses as above. 157.288
First, assume the right side of (4.49) holds – for some φ and ψ as indicated.
Then we can choose another cut-off χ ∈ C c∞ (Rn ) around zero in Rn and conclude
that
ξ
157.289 (4.50) (1 − χ(ξ))ψ( )F (φu)
|ξ|
 3. WAVEFRONT SET OF A DISTRIBUTION 65

is also rapidly decreasing. Now

ξ
157.290 (4.51) a = (1 − χ(ξ))ψ( )φ(x) ∈ C c∞ (Rn ; S 0 (Rn ))
|ξ|
and if A = QR (a) then
ξ
157.291 (4.52) F Au = (1 − χ(ξ))ψ( )F (φu)
|ξ|
by definition of right quantization. If follows that
157.292 (4.53) ¯ ∈
Au ∈ S (Rn ) =⇒ (x̄, ξ) / WF(u)
¯
since A is elliptic at (x̄, ξ).
Remark. Edited by Paige Dote.
For the opposite implication, we suppose that u ∈ Cc−∞ (Rn ) and A ∈ Ψ−m (Rn )
Ellm (A) are such that Au ∈ S (Rn ). Choose B ∈ Ψ−n (Rn ) as
with (x, ξ) ∈157.277 in
157.273
0
Proposition 4.1, so (x, ξ) ∈
6
157.260
WF (Id −BA). It follows from the second part of (4.36)
that, with cδ as in (4.24) (so ψ is supported near x and ψ near ξ) and Cδ = QL (cδ ))
for δ > 0 sufficiently small,
WF0 (Cδ ) ∩ WF0 (Id −BA) = ∅.
Then,
(Cδ BA)u = Cδ u mod S (Rn ),
157.274
and hence, Cδ u ∈ S (Rn ) which implies the condition on the right in (4.37). 
We can also see that ‘pseudodifferential operators are microlocal’ and combine
it with ‘microlocal elliptic regularity’ which is a partial inverse
157.293 Proposition 4.3. For any u ∈ C c−∞ (Rn ) and A ∈ Ψm (Rn )

157.294 (4.54) WF(u) ⊂ Char(A) ∪ WF(Au), WF(Au) ⊂ WF0 (A) ∩ WF(u) =⇒

WF(u) ∩ Ell(A) = WF(Au) ∩ Ell(A)

We have used the fact that A ∈ Ψm (Rn ) and WF0 (A) = ∅ =⇒ Au ∈ S (Rn )
if u ∈ Cc−∞ (Rn ). Note that it is not true that WF0 (u) = ∅ and u ∈ S 0 (Rn ) =⇒
AU ∈ S (Rn ). This is one of the ‘defects’ of the algebra Ψ∗ (Rn ).
Proof. The first statement is equivalent to saying that
(x, ξ) ∈ Ellm (A), (x, ξ) ∈
/ WF(Au) =⇒ (x, ξ) ∈
/ WF(u).
Again, we use the construction in Proposition ?? to find B ∈ Ψ−m (Rn ) with
/ WF0 (Id −BA). We can choose the cone-support of the symbol of B such
(x, ξ) ∈
that B(Au) ∈ S (Rn )m abd then it follows that (x, ξ) ∈
/ WF(u). 
So far I have limited the definition of WF(u) to elements of C c−∞ (Rn ). This
is only because Ψm (Rn ) does not act on general distributions in C −∞ (Rn ). To
overcome this, we consider a smaller algebra consisting of the properly-supported
pseudodifferential operators.
A closed set S ⊂ R2n can be considered as a relation between subsets of Rn
−1
S ◦ U = πL (S ∩ πR (U ))
 66 4. ELLIPTICITY AND WAVEFRONT SET

where πL ; πR : R2n → Rn are the two projections

πR (x, y) = y and πL (x, y) = x.
I am thinking here of the support of the kernel of an operator. Thus,
x ∈ S ◦ U ⇐⇒ ∃(x, y) ∈ S such that y ∈ U.
The relation defined by S is said to be proper if
S ◦ K b Rn ∀K b Rn
i.e. it takes compact sets to compact sets.
Definition 4.4 (Properly Supported). A pseudodifferential operator (or in-
deed any operator) is said to be properly supported if the support of its kernel and
the kernel of its adjoint define proper relations.
Lemma 4.4. The properly supported pseudodifferential operators, denoted Ψm n
P (R )
∞ n ∞ n n
form a ∗-closed filtered ring defining linear maps on C c (R ), C (R ), CmIc(R ),
and C −∞ (Rn ). We further have
−∞
157.494 (4.55) Ψm (Rn ) = Ψm n
p (R ) + Ψ (Rn ).
Proof. Let ϕ ∈ Cc∞ (Rn ) be a cutoff near 0, say ϕ(z) = 1 in |z| ≤ 1. Then,
(1 − ϕ(x − y))K(x, y)
is the kernel of an element of Ψ−∞ (Rn ) as follows from (??). This proves the
equation (1). The relationship of operators and kernels show that
Z
Au(x) = A(x, y)u(y)dy, u ∈ C c∞ (Rn

vanishes if A(x, y)u(y) vanishes (??). If A has proper support S = supp(A), then
ψ ∈ C c∞ (Rn ) with supp(ψ) ∩ (S · supp(u)) = ∅ satisfies
ψ(x)A(x, y)u(y) = 0 =⇒ supp(ψ) ∩ supp(Au) = ∅.
Thus, supp(Au) is compact if
A : C c∞ (Rn ) −→ C c∞ (Rn ).
We have assumed the same for the adjoint (and hence the transpose) from which
the remaining properties follow by duality. Then,
∞
WF(u) 63 (x, ξ) if ∃ A ∈ Ψm n n
p (R ), (x, ξ) ∈ Ellm (A), Au ∈ C (R ).

It follows that this is consistent with the properties ** defined when u ∈

C −∞ (Rn ). 
One of the important properties if the wavefront set is that it is a refinement
of the singular support.
Lemma 4.5. If u ∈ C −∞ (Rn ) then
x∈ / WF(u)∀ ξ ∈ Rn \ {0}.
/ singsupp(u) ⇐⇒ (x, ξ) ∈
Proof. In the foward direction, this is immediate, since x ∈ / singsupp u implies
∃ϕ ∈ C c∞ (Rn ) such that ϕu ∈ C c∞ (Rn ). As an element of Ψ0P (Rn ), this is elliptic at
all points (x, ξ) for any ξ ∈ Rn \ {0}.
For the opposite implication, we need a covering argument. Certainly, for each
ξ 6= 0, by **, there exists ϕξ ∈ C c∞ (Rn ) such that ϕdξ u is rapidly decreasing in a
 3. WAVEFRONT SET OF A DISTRIBUTION 67

cone around ξ with ϕξ (x) 6= 0. In fact, we know from Proposition that this remains
true if we replace ϕξ by ϕϕξ for a fixed ϕ ∈ C c∞ (Rn ) around ξ with ϕ(x) 6= 0. The
cone of rapid decay does not decrease.
Now Sn−1 is compact and the cones of rapid decay of the ϕ dξ u correspond to an
open cone of S2n−1 . This has a finite subcover and it follows that we may choose
ϕ ∈ C c∞ (Rn ) with ϕ(x) 6= 0 and supp(ϕ) ⊂ {ϕξj = 1} for this finite cover. Then
ϕϕξj = ϕ and ϕu c is rapidly decreasing in all directions of Rn , so ϕu ∈ C c∞ (Rn ) and
hence x ∈ / singsupp u. 
Remark 3. The scatting and157.494
isotropic algebras do have properly supported
subalgebras, but the analogue of (4.55) is not valid in these cases. We will see why
later.
Recall that we have insisted that WF(u) is a closed cone in Rn × (Rn \ {0}).
The results above show that this condition in in R2n is
WF(u) = (singsupp(u) × {0}) ∪ WF(u),
which gives a nice picture! L6-end
 CHAPTER 5

Propagation of singularities
L7
Microlocal ellipticity, as discussed above, shows us that if u ∈ S 0 (Rn ) and
A ∈ Ψm (Rn ) then
157.297 (5.1) WF(u) ⊂ WF(Au) ∪ Char(A).
∞
In particular if Au ∈ C (Rn ) then WF(u) ⊂ Char(A).
To see what else we might be able to say, consider a simple case, where A = D1
is differention with respect to the first variable. Letting p = ξ1 be the principal
symbol we see that
157.298 (5.2) Char(D1 ) = {(x, ξ) ∈ Rn × (Rn \ {0}); ξ1 = 0}
is a smooth hypersurface. Any solution
157.299 (5.3) D1 u = 0 =⇒ u(x, x0 ) = v(x0 )
is independent of x1 , meaning that
Z
157.300 (5.4) u(φ) = v(ψ), ψ(x0 ) = φ(x1 , x0 )dx1 .

We already know that WF(u) ⊂ {ξ1 = 0} but more is true. Namely

157.301 (5.5) WF(u) = {(x1 , x0 , 0, ξ 0 ); (x0 , ξ 0 ) ∈ WF(v)}.
So the wavefront set of u is a union of lines where x1 alone varies. These are of
course the integral curves of ∂x1 but as a vector field on Rn × (Rn \ {0}) and only
those inside {ξ1 = 0}.

1. Hamiltonian mechanics
157.301
As we shall see, something like (5.5) can be proved much more generally. To
do so we use commutator methods. The idea here is that we try to get information
about solutions of P u = 0 where P ∈ Ψm (Rn ) by taking a commutator with a
‘test operator’ B ∈ Ψk (Rn ). I will explain this a bit more fully below, but for the
moment just recall that

157.302 (5.6) P, ∈ Ψm (Rn ), B ∈ Ψk (Rn ) =⇒ [P, B] ∈ Ψm+k−1 (Rn ) and

n  
X ∂p ∂b ∂p ∂b
σm+k−1 ([P, B]) = −i −
i=1
∂ξi ∂xi ∂xi ∂ξi
Here p and b are really representatives of the principal symbols of P and B and then
we are looking at a representative of the symbol of the commutator, all modulo a
term of one order lower. In fact we will assume that
157.303 (5.7) P ∈ Ψm n k n
cl (R ), B ∈ Ψcl (R ), σm (P ) = p, σk (B) = b

69
 70 5. PROPAGATION OF SINGULARITIES

where these are now homogeneous functions of orders m and k and we just ignore
157.302
the singularities at ξ = 0, meaing we work on Rn × (Rn \ {0}. Then (5.6) gives the
homogeneous principal symbol of [P, B].
Thw important point is that we can ‘recognize’ the formula, it is the Poisson
bracket of the symbols (ignoring the i). As you will know this comes from the
symplectic form on Rn × Rn = T ∗ Rn
X
157.304 (5.8) ω= dξi ∧ dxi
i

which is actually completely independent of coordinates (which we will return to

later). Then the Hamilton vector field of p is by definition the unique vector field
Hp on Rn × (Rn \ {0}) (because p might be singular at ξ = 0) satifying
X  ∂p ∂ ∂p ∂

157.305 (5.9) ω(·, Hp ) = dp ⇐⇒ Hp = −
i
∂ξi ∂xi ∂xi ∂ξi
and the Poisson bracket of two functions is
157.306 (5.10) {p, b} = Hp b = −Hb p.
That is,
157.307 (5.11) σm+k−1 ([P, B]) = −i{p, b} = −iHp b.
One of the basic points about Hamiltonian mechanics is that
157.310 (5.12) Hp p = 0
which follows from the antisymmetry of ω. This means that
157.311 Lemma 5.1. Integral curves of Hp which have a point in {p = 0} are contained
in {p = 0}.
Here of course an integral curve is connected, it is a smooth curve defined on some
interval, I −→ Rn × Rn \ {0) which has tangent vector Hp at each point.

2. Hörmander’s Theorem
∗
157.308 Theorem 5.1. If P ∈ Ψm n
cl (R ) has real principal symbol then for any u ∈
0 n
S (R )

157.309 (5.13) WF(u) \ Char(p)

is a union of maximally extended integral curves of Hp in Char(P ) \ WF(P u).
157.312 Exercise 9. Try to see what this says for the flat wave operator
b
X
157.313 (5.14) P = Dt2 − Dx2j on Rn+1 .
j=1

Check that the characteristic variety at each point is the cone τ 2 = |ξ|2 (with the
obvious notation) and that the integral curves of Hp within Char(P ) project into
Rn to light rays.
Restated the claim of the Theorem is that a maximally extended integral curve
of Hp in the open set (Rn × (Rn \ {0})) \ WF(P u) on which p vanishes either does
not meet WF(u) or lies completely within it.
 3. THE HAMILTON VECTOR FIELD 71

Proof. We start can by making some elementary reductions before getting

to the constructive part of the proof. First note that if p is homogeneous (always
meaning in ξ of course) of degree m then Hp is homogeneous of degree m − 1. If we
−m+1
premultiply P by an elliptic operators, setting P 0 = AP where A = (1 + |d|2 ) 2 ,
then
−m+1
157.315 (5.15) p0 = σ(P 0 ) = (1 + |ξ|2 ) 2 p, Hp0 = aHp + pHa .
It follows that the integral curves of hp0 in Char(P 0 ) = Char(P ) are simply repa-
rameterizations of the integral curves of Hp . Since A is globally elliptic, WF(P 0 u) =
WF(P u) and The statement for P is equivalent to that for P 0 . Thus we can freely
assume that P is of order 1.
A smooth vector field such as Hp has two types of maximally extended integral
curves in an open set (of Rn × (Rn \ {0}). Namely constant curves valued at points
where Hp = 0 and through every other point an embedded curve, without stationary
points. This dichotomy corresponds to the vanishing of dp. For a constant curve the
statement is of course trivial. Thus in fact Hörmander’s Theorem tells us nothing
about points where
157.316 (5.16) ¯ = 0, dp(x̄, ξ)
p(x̄, ξ) ¯ = 0.
157.316
There is in fact another trivial case beyond (5.16). Namely a point is radial for
p (and also P ) if
157.317 (5.17) ¯ = 0, dp(x̄, ξ)
p(x̄, ξ) ¯ = λξ¯ · dc, λ ∈ R.
157.316 157.305
Of course (5.16) is the case where λ = 0. From (5.9) we see that at such a point
157.318 (5.18) Hp = −λξ · ∂ξ
is the radial vector field – hence the name. Since p is homogenous of degree 1 the
157.318
identiy (5.18) must then hold along the ray (x̄, R+ ξ) ¯ – which is then the integral
¯
curve through (x̄, ξ). Again the statement is trivial for the this integral curve, since
WF(u) is itself radial.

3. The Hamilton vector field

Thus, we are reduced to considering a ray on which p vanishes, maximally
¯ By
extended in (Rn × (Rn \ {0}) \ WF(P u) and through a non-radial point (x̄, ξ).
the local uniqueness of integral curves it must consist of non-radial points. Let
the maximally extended integral curve, so defined on an open (possibly infinite)
interval be
157.319 (5.19) ¯
χ : I −→ {p = 0} \ WF(P u), χ(0) = (x̄, ξ).
So we need to show is that
157.320 (5.20) χ(I) 6⊂ WF(u) =⇒ χ(I) ∩ WF(u) = ∅.
Since {t ∈ I; χ(t) ∈ WF(u)} is closed its complement is a countable union of open
intervals. Considering one of these intervals, either it is equal to I of, if not, has
at least one endpoint t̄ ∈ I with χ(t̄) ∈ WF(u). Reversing the sign of P does
not change the result, so all we have to exclude is that there is some t̄ ∈ I with
χ(t̄) ∈ WF(u) but χ(t) ∈ / WF(u) for t ∈ (t̄, t̄ + ),  > 0. Clearly we can shift the
parameterization so that t̄ = 0.
We are therefore reduced to a (micro-)local statement.
 72 5. PROPAGATION OF SINGULARITIES

¯ ∈ (Rn \ {0}) is a non-radial point for p, (x̄, ξ)

If (x̄, ξ) ¯ ∈/ WF(P u)
¯
and the integral curve of Hp starting at (x̄, ξ) immediately leaves
WF(u) then (x̄, ξ) ¯ ∈
/ WF(u).
Q
To prove this we need to look closely at Hp . First we can choose local coordi-
nates near x̄ so that x̄ = 0 ξ¯ = (0, . . . , 0, 1). [Despite appearances we will not be
using the coordinate-invariance of pseudodifferential operatos here.] By the non-
radial assumption Hp 6= λ∂ξn since that is the radial direction. So at least one of
the other coefficients is non-zero at this point, and hence nearby, it could be any
one of the xi or one of the ξj , j < n. Denote this special variable ζ, we can solve
the initial value problem
Hp yj = 0, j = 1, . . . , 2(n − 1),
157.321 (5.21) Hp Ξ = 0, Ξ ζ=0
= ξn ,
Hp τ = τ, τ ζ=0
= 0,
where the initial conditions on {ζ = 0} for the yj are all the xi and ξk /ξn , k < n
except the one that corresponds to ζ (which of course vanishes).
All the solutions exist in an open conic (because Hp is homogeneous of degree
0) neighbourhood of the base point and are homogenous of degree 0, except Ξ which
is positive and homogeneous of degree 1. Their differentials are indpendent so they
give a coordinate system in terms of which
157.322 (5.22) Hp = ∂τ .
[In fact you can do essentially this with a homogeneous symplectic (‘canonical’)
L-end transformation but we do not need it.]
L9
4. Construction of symbols
This allows us to construct appropriate classical symbols. What we want is a
cut-off near the base point with useful properties, in fact a family of them depending
on a parameter 0 < δ < δ0 . The basic function we have in mind is
157.323 (5.23) aδ (τ, y) = µ(τ /δ)µ((/2 − τ )/δ)φ(y/δ).
Here 0 ≤ φ ∈ C c∞ (R2n−2 ) is a typical cut-off, supported in |τ | ≤ 1 and strictly
positive in |τ | < 1 and 0 ≤ µ ∈ C ∞ (R) vanishes in (−∞, −1), is positive on
(−1, ∞) and is equal to 1 on [0, ∞). We require
157.362 (5.24) 0 < δ < /2.
Then
supp(aδ ) = {(τ, y); −δ ≤ τ ≤ /2 + δ, |y| ≤ δ},
157.332 (5.25)
aδ > 0 on supp(aδ0 ), δ 0 > δ.
Then consider
Z τ
157.324 (5.26) bδ (τ, y) = µ((/2 − τ )/δ)2 (µ(s/δ)φ(y/δ))2 ds ∈ C c∞ (R2n−1 )
−δ
where the missing variable is Ξ. Then
157.325 (5.27) supp(bδ ) = [−δ, /2 + δ] × {|y| ≤ δ}
and
157.326 (5.28) Hp bδ = ∂τ b = a2δ + eδ , supp(eδ ) ⊂ [/2, ] × {|y| ≤ δ}.
 5. PROOF OF REGULARITY 73

Here e comes from ∂µ ((/2 − τ )/δ)2 .

Thus bδ is homogeneous of degree 0 on Rn × (Rn \ {0}). We want to consider
similar functions homogeneous of degree 2m, for any m and also ‘regularized’. Since
∂τ Ξ = 0 we set

157.327 (5.29) bδ,m,N (τ, y, Ξ) = (1 − Φ(Ξ))2 Ξ2m Φ2 (Ξ/N )bδ (τ, y),
1
where 0 ≤ Φ ∈ C c∞ (R), Φ(s) = 1 in |s| < , Φ(s) = 0 in |s| > 1 =⇒
2
Hp bδ,m,N = a2δ,m,N + eδ,m,N ,
aδ,M,n = (1 − Φ(Ξ))Ξm Φ(Ξ/N )aδ , eδ,m,N = (1 − Φ(Ξ))2 Ξ2m Φ2 (Ξ/N )eδ .
In fact we need to go a step further. Namely rather than the vector field Hp
what actually arises below is the operator with a zeroth order term
157.328 (5.30) Hp + f, f smooth, real-valued and homogenous of degree 0.
Then we replace bδ by
Z τ
157.329 (5.31) bδ (τ, y, Ξ) = ψ(τ )2 e−q(τ,y) eq(s,y) a2δ (s, y)ds
−δ
Z τ
q(τ, y) = f (s, y)ds.
0
The support properties are unchanged and now
157.330 (5.32) (Hp + f )bδ,m,N (τ, y, Ξ) = a2δ,m,N + eδ,m,N .

5. Proof of regularity
After all this preparation, let me add a few words about L2 boundedness –
which is in Problem set 2 – and uniformity before passing to the actual proof of
the statement above..
First uniformity. In the proof below, regularity for Au where A ∈ Ψm (Rn ) is ob-
tained by looking at an approximating sequence AN ∈ Ψ−∞ (Rn ) which is bounded
in Ψm (Rn ) and converges to A in Ψm+ (Rn ) for any  > 0. This is constructed usual
sort of cutoff. Since WF0 (AN ) = ∅ for finite N we reserve the notation for the uni-
form operator wavefront set defined as previously in terms of the cone-support of
the symbol. So in this sense

157.342 (5.33) ¯ ∈
(x̄, ξ) / WF0 (A∗ ) = QL (a∗ ) ⇐⇒ (x̄, ξ)
¯ ∈/ conesupp(a∗ ) ⇐⇒
∞
cδ aN is bounded in C ∞ (Rn ; S −∞ (Rn )) as N → ∞.
157.260
Here cδ is a conic cut-off as in (4.24). It follows for instance that for two such
families
157.343 (5.34)
WF0 (A∗ ) ∩ WF0 (B∗ ) = ∅ =⇒ A∗ A0∗ ψ ∈ Ψ−∞ (Rn ) is bounded for ω ∈ C c∞ (Rn ).
So for instance A∗ A0∗ u is bounded in S (Rn ) if u ∈ C c−∞ (Rn ). Note that we do not
necessarily conclude that the product is bounded in Ψ−∞ (Rn ) since we have not
assumed any uniformity near infinity is space.
The basic result on L2 boundedness is that
157.344 (5.35) A ∈ Ψm (Rn ) : H M (Rn ) −→ H M −m (Rn ) ∀ M.
 74 5. PROPAGATION OF SINGULARITIES

In the argument below we need a little more than this. Namely for sequences AN
as above bounded in Ψ2m (Rn ) if à ∈ Ψm (Rn ) is fixed and if

157.345 (5.36) WF0 (A∗ ) ⊂ Ellm (Ã) then |hψA∗ u, uiL2 | ≤ CkÃukL2 + CkukH m0 ,
ψ ∈ C c∞ (Rn ), u ∈ C c−∞ (Rn ) ∩ H m0 (Rn )
where the constants may depend on everything except N and u.
157.365 Lemma 5.2. If u ∈ C c−∞ (Rn ) and AN u is bounded in L2 (Rn ) then Au ∈
L (Rn ) and kAukL2 ≤ lim sup kAN ukL2 .
2

Proof. This follows from the fact that L2 (Rn ) is a Hilbert space. Thus the
norm boundedness of AN u implies that it has a weakly convergent subsequence
AN u ,→ v in L2 . It follows that
157.366 (5.37) hAu, φi = limhANi u, φi = hv, φi
for all φ ∈ S (Rn ) and hence Au = v as a distribution. 

So, to the proof. We are in the setup discussed above, (x̄, ξ) ¯ is a non-radial
¯ ∈
characteristic point, (x̄, ξ) ¯ ∈
/ WF(P u) and exp(tHp )(x̄, ξ) / WF(u) for t ∈ (0, ) for
some  > 0. We can always shrink . In the local coordinates in Rn × (Rn \ {0}) we
choose 0 < δ0 < /2 such that
157.367 (5.38) [−δ0 , ] × {|y| ≤ δ0 } ∩ WF(P u) = ∅, [/2, ] × {|y| ≤ δ0 } ∩ WF(u) = ∅.
We proceed to conclude from this that
157.368 (5.39) ¯ ∈
Am,δ u ∈ L2 (Rn ) ∀ δ < δ0 , ∀ m =⇒ A0 u ∈ C ∞ (Rn ) =⇒ (x̄, ξ) / WF(u)
¯
since A0 is elliptic at (x̄, ξ).
Set Bδ,m,N = QL (bδ,m,N (τ, y, Ξ)) which is uniformly bounded in Ψ2m (Rn ) and
∗
the formal adjoint is Bδ,m,N = Bδ,m,N + Sδ,m,N where Sδ,m,N is uniformly bounded
2m−1
in Ψ (R ). Consider the L2 inner product
n

157.331 (5.40) hP u, Bδ,m,N uiL2 .

operator. As N → ∞ the cone-
This is well-defined since Bδ,m,N is a smoothing157.367
support of bδ,m,N is contained in the region in (5.38) where P u has no wavefront
set. Thus
157.333 (5.41) Bδ,m,N P u −→ Bδ,m P u ∈ C c∞ (Rn ).
We use similar notation for the other families below.
Then

157.334 (5.42) 2 Re(ihP u, Bδ,m,N uiL2 )

= ihP u, Bδ,m,N uiL2 − ihBδ,m,N u, P uiL2
= Re ih([Bδ,m,N , P ] + Bδ,m,N F iL2 + ih(Sδ,m,N )P )u, uiL2 , F = (P − P ∗ )
157.333
The last term here converges as N → ∞ for the same reason as (5.41). So we
conclude that for any m,
157.335 (5.43) ih([Bδ,m,N , P ] + Bδ,m,N F iL2 converges as N → ∞.
 6. A QUESTION ABOUT THE WAVE EQUATION 75

Taking f to be the principal symbol of F = i(P − P ∗ ) ∈ Ψ0cl (Rn ) the construc-

tion of the family Bδ,m,N means that
0
157.336 (5.44) i[Bδ,m,N , P ] + Bδ,m,N F = A2δ,m,N + Eδ,m,N + Eδ,m,N ,
0
Eδ,m,N bounded in Ψ2m−2 (Rn ),
WF0 (E 0 ) ⊂ conesupp(bδ ), WF0 (E) ⊂ conesupp(e)
157.330 157.334
with e the error in (5.32). From the L2 bounds in (5.42) we deduce that
0
157.337 (5.45) kAδ,m,N ukL2 ≤ C|hEδ,m,N u, uiL2 | + C|hEδ,m,N u, uiL2 |.
Such a bound holds uniformly in N and in 0 < δ < δ0 < 21 . By assumption the last
term is uniformly bounded since WF0 (Eδ,m,N ) is (uniformly) contained in a region
disjoint from WF(u).
By L2 boundedness we know that
0
157.338 (5.46) |hEδ,m,N u, uiL2 | ≤ kuk 1
H m− 2
1
Since u is a distribution and the supports here are compact, u ∈ H m0 − 2 (Rn ) for
some m0 . It follows that the limit
157.339 (5.47) Aδ,m0 u ∈ L2 (Rn )
157.365
using Lemma 5.2.
Now, we can iterate 157.337
this estimate, half a derivative at a time. The important
point is that157.339
we know (5.45) for all δ < δ0 . So at some stage of the iteration we
know that (5.47) holds for a given m − 12 . However, the cone-supports of the terms
A and the lower order error E 0 are such that
157.340 (5.48) for δ < δ 0 , WF0 (Eδ,m,N
0
) ⊂ Ellm (Aδ0 ,m− 12 ).
0
Again from L2 boundedness and ellipticity (and the fact that the order of Eδ,m,N
2
is the same as Aδ,m− 1 ) it follows iteratively that
2

0
157.341 (5.49) kAδ,m,N u|2L2 ≤ hEδ,m,N u, uiL2 | + C
≤ Cδ,δ0 kAδ0 ,m− 12 ukL2 H m−/ha + Cδ,δ0 kukH m0 + C
giving the inductive step. 

6. A question about the wave equation

One question in lecture was: How much simpler is it to prove this for the wave
equation? Who asked this?
If we are talking about the flat wave equation then the discussion can be sim-
plified by shifting the ‘smoothing’ from the operators to the distribution u. For
constant coefficient operators this is straightforward since we can use convolution.
In any case smoothing u is an alternative approach but not much different. Apart
from that it is a lot simpler. Note however that there are other approaches which
can be used here (and more generally). Namely one has in the flat case a forward
fundamental solution, and in the general case one can construct forward microlocal
parametrices at non-radial points (this requires real work). Then one can shift the
discussion to analysing what such operators due to singularities. This amounts to
 76 5. PROPAGATION OF SINGULARITIES

analysing the wavefront set of the (the Schwartz kernel of) the fundamental solu-
tion or parametrix. I will likely get around to including something on this as ‘the
L9-end calculus of wavefront sets’.
 CHAPTER 6

Smoothing operators and K-theory

K-theory L10
I want to take a serious look at smoothing operators. In particular I want to
introduce semiclassical (families of) operators in this context – this is another form
of ‘quantization’ closely related to what we have done so far. There are several
reasons behind this, the introduction of (topological, complex) K-theory being one.
This I will introduce by constructing a smooth classifying space, initially in the odd
case but eventually in the even case as well.
I will try to keep the line of reasoning as straight as I can here since it is easy
to get distracted by the myriad of possibilities that arise.

1. Hilbert space and operators

Let me recall here some of the basic facts, without proofs, about Hilbert
space and bounded operators. Since we do not need the non-separable case,
Hilbert space here means a separable, complex, infinite dimensional, Hilbert
space, H. One basic fact is that every such space is isometrically isomorphic
to l2 (Z).
• The most basic properties of Hilbert space are the Riesz representation
thereom, identifying the dual with the conjugate
H 0 = H, 157.521 (6.1)
the existence of orthonormal bases and the convergence of the Fourier-
Bessel series.
• Compact sets in Hilbert space are precisely those closed bounded sub-
sets with the additional propery that the Fourier-Bessel series with
respect to any (one) orthonormal basis converges uniformly. This
amounts to a characterization for l2 (Z).
• A second characterization of compact sets is that they are approx-
imable by finite-dimensional spaces – K has compact closure (is pre-
compact) if it is bounded and for every  > 0 there exists a finite
dimensional subspace F such that
sup inf d(p, q) < . 157.518 (6.2)
p∈K q∈F

• The bounded operators form a complete normed, ∗-closed, algebra

(the fundamental example of a C ∗ -algebra), B (H), with
kA∗ k = kAk, kABk ≤ kAkk|Bk.
• For any closed subspace R ⊂ H
H = R ⊕ R⊥
defines the unique self-adjoint projection with range R, PR : H −→ R,
PR∗ = P = P2.
R R
• The open mapping theorem: Any surjective bounded operator is open,
meaning they map open sets to open sets.
• Closed graph theorem: A linear map L : H −→ H with closed graph
in H × H is bounded.
• A bounded bijection has a bounded inverse.

77
78 6. SMOOTHING OPERATORS AND K-THEORY

• The resolvent of a bounded operator, B, is a holomorphic map

Res(B) = C \ Spec(B) 3 z 7−→ (B − z)−1 ∈ B (H) 157.524 (6.3)
defined on the complement of the compact set Spec(B) ⊂ C where
B − z is not a bijection.
• The resolvent identity
(B − z)−1 − (B − τ )−1 = (z − τ )(B − z)−1 (B − z)−1 , z, τ ∈ Res(B),
157.559 (6.4)
holds.
• If χ : S −→ C \ Spec(B) is a smooth, simple, positively-oriented curve
then I
1
Pχ = (B − z)−1 dz 157.558 (6.5)
2πi χ
is a projection commuting with B such that Pχ B has spectrum in the
interior of χ and (Id −Pχ )B has spectrum in the exterior.
• The group of invertibles GL(H) ⊂ B (H) is open (but not dense) be-
cause of the convergence of the Neumann series
X
(Id +B)−1 = (−1)k B k , kBk < 1. 157.520 (6.6)
k

• There is a functional calculus for each self-adjoint operator, A = A∗ ∈

B (H), defining a map of algebras
{f : Spec(A) −→ C; continuous} −→ B (H) s.t. f (A)g(A) = (f g)(A).
• There is a polar decomposition of each bounded operator as a product
1
B = AV where A = (B ∗ B) 2 and V is a partial isometry 157.523 (6.7)
• It follows that unitary subgroup U(H) ⊂ GL(H) defined by the iden-
tity B ∗ B = Id is a deformation retract of GL(H).
• Kuiper’s theorem is the statement that U(H) is (weakly) contractible
in the norm topology (every continuous map from a compact space is
homotopic to the map to the identity).
• The compact operators, K (H) ⊂ B (H), consist of the operators map-
ping bounded to pre-compact sets. They constitute the norm-closure
of the finite rank operators (those with finite-dimensional range) and
form the only non-trivial closed ideal.
• The Fredholm operators, F (H) ⊂ B (H), are defined by the require-
ments that they have finite-dimensional null space and closed range
of finite codimension. The index
ind(F ) = dim null(F ) − dim(Ran(F )⊥ ), ind : F (H) −→ Z 157.526 (6.8)
labels the components.
• A bounded operator B is Fredholm if and only if it has a parametrix
modulo compact operators (an inverse in the Calkin algebra B (H)/K (H))
A ∈ B (H) such that
BA − Id, AB − Id ∈ K (H).
• The Hilbert-Schmidt ideal, HS(H) ⊂ B (H), is defined by the condition
that for any one orthonormal basis
X
kBk2HS = |hBei , ei i|2 < ∞. 157.529 (6.9)
i
This is independent of choice and
kABkHS ≤ kAkHS kBk. 157.530 (6.10)
• The trace ideal, T (H) ⊂ B (H), is HS(H)2
– the finite span of products
of elements of HS(H). Any element is the product of two Hilbert-
Schmidt operators and
1
kT kTr = k(T ∗ T ) 4 kHS , kABkTr | ≤ kAkTr kBk. 157.535 (6.11)
 2. SCHWARTZ SMOOTHING ALGEBRA 79

• The trace functional is

X
Tr : T (H) 3 T = hT ei , ei iH ∈ C 157.531 (6.12)
i
for any orthornomal basis and
Tr(AB) = Tr(BA) if A ∈ T (H), B ∈ B (H) or A, B ∈ HS(H).
157.532 (6.13)
• Any Fredholm operator B has a parametrix A modulo trace class
errors and
ind(B) = Tr(BA − AB). 157.533 (6.14)
• A Fredholm operator has a unique generalized inverse
AB = Id −Pnull(B) , BA = Id −PRan(B)⊥ . 157.534 (6.15)
• If B ∈ F (H) and K ∈ K (H) then B + K ∈ F (H) and
ind(B + K) = ind(B). 157.536 (6.16)
• On the ring Id +T (H) the Fredholm determinant
det : Id +T (H) −→ C satisfies
157.537 (6.17)
det ((Id +T1 )(Id +T2 )) = det(Id +T1 ) det(Id +T2 ).
• The group of invertibles in Id +T (H) is the inverse image det−1 (C \
{0}) and if (−1, 1) 3 s −→ T (s) ∈ T (H) is C 1 and Id +T (0) is invert-
ible then
 
d dT
det(Id +T (s)) s=0 = det(Id +T (0)) Tr (Id +T (0))−1 (0)157.538
. (6.18)
ds ds

2. Schwartz smoothing algebra

The residual ideal in the standard pseudodifferential algebra Ψ−∞ (Rn ), is the
part that cannot be reached by the symbol calculus. As I have muttered all along,
∞
it is not a very nice algebra because of the coefficient ring C ∞ (Rn ).
Let us instead concentrate on the smaller Schwartz smoothing algebra, Ψ−∞
S (Rn )
– this is actually a two-sided ideal in Ψ−∞ (Rn ). It is also ‘very non-commutative’
in that it is simple with only the two trivial ideals. Thus
157.348 (6.19) S (R2n ) = Ψ−∞
S (Rn ) ⊂ Ψ−∞ (Rn )
where the first equality is as a space of kernels. In fact it is the residual part of
157.220 C.PsiRn
S.Isotropic
both the scattering (Shubin) and isotropic algebras in Proposition 3.3 and §3.3.
The product in this algebra is
Z
157.349 (6.20) A ◦ B(x, y) = A(x, z)B(z, y)dz.
Rn

In essence, Ψ−∞
S (Rn ) is an infinite-dimensional matrix algebra. To justify this
directly we need a ‘basis’ for S (Rn ); the standard one is the Hermite basis. This
consists of the eigenfunctions for the harmonic oscillator
n
X
157.350 (6.21) H= Dx2i + |x|2 , Heκ = (n + 2|κ|)eκ , κ ∈ Nn0 .
i=1

Here the eigenfunctions are given as products of the L2 normalized eigenfunctions

in the case n = 1
n
Y 1
157.351 (6.22) eκ (x) = fκj (xj ), (Dx2 + x2 )fj = (1 + 2j)fj , fj (x) = hj (x) exp(− x2 )
j=1
2

where the hj are the Hermite polynomials.

 80 6. SMOOTHING OPERATORS AND K-THEORY

We need the isomorphism that these provide

X
S (Rn ) −→ s (n) = {c : Nn0 −→ C; (1 + |κ|2 )N |cκ |2 < ∞ ∀ N },
κ
157.352 (6.23) Z 
u 7−→ u(x)eκ (x)dx .
Rn κ

This leads to the identification

157.353 (6.24) Ψ−∞

S (Rn ) 3 A −→ a = (Aeκ0 , eκ )
X
∈ Ψ−∞
s = {a : N2n
0 −→ C; (1 + |κ| + |κ0 |)N |aκ,κ0 | < ∞ ∀ N }
κ,κ0

with the composition becoming ‘matrix multiplication’. That is, the ring of ‘rapidly
decreasing infinite matrices’.
157.371 Lemma 6.1. The algebras Ψ−∞
S (Rn ), for different n, are isomorphic (although
not naturally so).

Proof. Expansion in terms of the Hermite basis reduces Ψ−∞ S (Rn ) to ‘matri-
ces’ meaning rapidly decreasing maps
X
157.484 (6.25) a : Nn0 × Nn0 −→ C, (1 + |α| + |β|)N |a(α, β)|2 < ∞.
α,β

number of α with |α| ≤ N is bounded by N n so rapid decay in the sense

The 157.484
of (6.25) is the same as rapid decay in j if j 7−→ αj is any ordering in which
|αj | is non-decreasing. This shows that all the algebras are isomorphic to the case
n = 1. 

157.369 Lemma 6.2. The elements of Ψ−∞ S (Rn ) act as elements of the trace ideal of
compact operators on L2 (Rn ) (or any Sobolev space) and form a ‘corner’ (not an
ideal) in the bounded operators in the sense that if B ∈ B (L2 (Rn ))
157.370 (6.26) A1 , A2 ∈ Ψ−∞ (Rn ) =⇒ A1 BA2 ∈ Ψ−∞ (Rn ).
Proof. Schur’s criterion (as discussed in Problem set 2) shows that these are
bounded operators and the fact that any sequence in S (Rn ) bounded with respect
to the seminorms has a convergent subsequence in L2 (Rn ) shows that they are
compact operators.
That these operators are in the trace ideal follows from the fact that the diag-
onal operator, with respect to the Hermite basis
X
157.539 (6.27) Tk 3 u 7−→ (1 + |α|)−k hu, eα ieα
α
1
is in the trace ideal if (T ∗ T ) 4 is in Hilbert-Schmidt, which follows if
1
X
157.540 (6.28) k(T ∗ T ) 4 k2HS = (1 + |α|)−k < ∞.
α

This holds if k > n. If A ∈ Ψ−∞

S (Rn ) it follows that ATk−1 ∈ Ψ−∞
S (Rn ) is bounded
2 n
so A ∈ T (L (R )). 
 3. ODD K-THEORY 81

Why should we be interested in Ψ−∞

S (Rn )? One important reason is topological,
arising from the associated group which plays a major rôle in the discussion of K-
theory below,
157.354 (6.29) G∞ n −∞
S (R ) = {A ∈ ΨS (Rn ); ∃ B ∈ Ψ−∞
S (Rn ) with (Id +A)◦(Id +B) = Id}.
Just as for square matrices, this one-sided inverse condition implies that Id +B is
a 2-sided inverse
157.355 (6.30) (Id +B) ◦ (Id +A) = Id .
−∞
157.356 Lemma 6.3. The group G∞ S ⊂ ΨS (Rn ) is open and dense in Ψ−∞
S (Rn ) and
the union of the subgroups GL(N, C) of finite N × N matrices with respect to the
Hermite basis is dense.
157.485 Remark 4. I habitually write G−∞S (Rn ) ⊂ Ψ−∞
S (Rn ) by removing the identity.
This really corresponds to changing the product to A ◦ B = A + B + AB now giving
a ring structure.
Proof. I did not go through this in lecture but maybe I should have
done so. We know, from Schur’s criterion, that the norm as a bounded
operator defines a continuous map
Ψ−∞ (Rn ) 3 A −→ kAkL2 . 157.390 (6.31)
So the elements of a neighbourhood of 0 in Ψ−∞ (Rn )give invertible ele-
ments Id +A ∈ B (L2 (Rn )) by Neumann seris. The inverse being
X
(Id +A)−1 = Id +B, B = (−1)k Ak . 157.391 (6.32)
k≥1

All the elements in the series are in Ψ−∞ (Rn ) but convergence is in prin-
ciple only as bounded operators – in fact the Neumann series converges in
Ψ−∞ (Rn ), i. e. in S (R2n ). To see this, expand the definition of B to see
that X
B = −A + A2 + A( (−1)k Ak )A = −A + A2 + ABA. 157.392 (6.33)
k≥1

It follows from the corner property that ABA ∈ Ψ−∞ (Rn ) and that the
series for B actually converges in this sense.
Thus, G∞ (Rn ) contains a neighbourhood N of 0 ∈ Ψ−∞ (Rn ) and hence a
neighbourhood around any point of G∞ (Rn ). 

3. Odd K-theory
The group G−∞
S (Rn ) provides an entry point to ‘complex K-theory’.
157.357 Definition 6.1. The odd K-theory, of a manifold M consists of the smooth
homotopy classes of smooth maps

157.358 (6.34) K 1 (M ) = {u : M −→ G∞
S ;
u = Id on M \ K for some K b M }/smooth homotopy.
A smooth homotopy between two elements u0 and u1 is a smooth map

157.359 (6.35) v : M × [0, 1] −→ G∞

S with
v = Id on (M \ K 0 ) × [0, 1] for some K 0 b M,
v M ×{0}
= u0 , v M ×{1}
= u1 .
 82 6. SMOOTHING OPERATORS AND K-THEORY

Note that there is no problem in understanding ‘smoothness’ here since

157.372 (6.36) C ∞ (M ; S (Rm )) = {u : M × Rm −→ C; smooth with all seminorms finite}.
157.358
So, the maps in (6.34) actually form a group which can be written
157.486 (6.37) C c∞ (M ; G−∞ (Rn ))
where as always the identity is ‘removed’. Then smooth curves in this group become
elements of
157.487 (6.38) C c∞ ([0, 1] × M ; G−∞ (Rn ))
and we can write the odd K-groups as the groups of components
157.488 (6.39) K 1 (M ) = π0 (C c∞ (M ; G−∞ (Rn ))).
What justifies such a bald definition and how can we start to understand it?
It is actually saying that G∞
S , in any of its variants, is a classifying group for odd
K-theory. You might object that K-theory is supposed 157.358 to be related to vector
157.389
bundles. It is, as we shall see below, but I assert that (6.34) (and (6.42) below)
are rather natural definitions corresponding to the 157.358
assertion that K-theory is the
topology associated to invertible matrices – since (6.34) is a definition which can
be reduced to smooth maps into GL(N, C) (stabilized in N ).
I do not want to spend too much time on this, but let me outline some of the
things which can be proved relatively easily and give the proofs later, perhaps some
of them not in lectures.
157.360 Proposition∗ 6.1. The odd K-theory of a manifold is an abelian group with
(
0 k even
157.361 (6.40) K 1 (Rk ) =
Z k odd
and for any manifold M there are natural isomorphisms
157.387 (6.41) K 1 (R2k × M ) −→ K 1 (M ).
157.387 157.361
The fundamental result (6.41) is157.361
actually a consequence of (6.40)157.387
if you know some
homotopy theory. In any case
157.361
(6.40) is ‘Bott periodicity’
157.387
and ( 6.41) is periodicity
in K-theory. Of course (6.40) is a consequence of (6.41) and the special cases k = 0
and k = 1. There are lots of competing proofs, none of them really simple as far
as I know.
157.387
I will describe a proof using semiclassical quantization to construct the
map (6.41). 157.387
The idea is that (6.41) is a prototype for the Atiyah-Singer index theorem for
families. It is an odd version, whereas the standard version is in even K-theory, but
these are closely related and the index theorems are in fact equivalent.
157.388 Definition 6.2. The even K-groups of a manifold M are the groups
157.389 (6.42) K 0 (M ) = K 1 (R × M ).

4. The involutive Grassmannian

S.IGrass
As you are probably aware, ‘K-theory’ – usually meaning the even group K 0 (M )
– for a manifold, is related to vector bundles over the manifold. I have not yet
reminded you of these and rather than proceed to do that I will simply assert for
the moment that vector bundles, like the groups GL(N, C) can be ‘stabilized’. So
I will work from the top down and start with the stabilized version.
 4. THE INVOLUTIVE GRASSMANNIAN 83

To be explicit consider the space L2 (R; C2 ) = L2 (R) ⊕ L2 (R). We need (at

least) two copies because I want to have infinite ‘upside and downside’ options.
Now we replace the identity operator considered above for the group G−∞
S (Rn ) by
the involutive matrix (meaning the square is the identity)
 
1 0
157.541 (6.43) β∞ = .
0 −1
Things related to β∞ are often given a ‘super’ quantifier but not here.
157.542 Definition 6.3. The (involutive) Grassmannian associated to β∞ is the space
of involutive operators on L2 (R; C2 ) of the form1
Υ−∞ (Rn ; C2 ) = β = β∞ + B, B ∈ Ψ−∞ (R; C2 ), β 2 = Id .

157.543 (6.44) S S

So these are the smoothing perturbations of β∞ which are involutive.

How is this related to vector bundles? Well, an involutive operator is really a
projection (by which I mean what is sometimes called an idempotent) namely
1
157.544 (6.45) P = (β + Id) ⇐⇒ P 2 = P, β = 2P − Id .
2
So the space of involutive perturbations of β∞ is identified with the space of per-
turbations
 
1 0
157.545 (6.46) P = P1 + A, P1 = , A ∈ Ψ−∞
S (R; C2 ), P 2 = P.
0 0
Such a projection is determined by its null space and range. These are the −1 and
1 eigenspaces, respectively, of the corresponding involution β.
157.546 Remark 5. Note that the ‘projection’ P need not be self-adjoint, so the range
need not be orthogonal to the null space. One could, and indeed it is more con-
ventional to do so, work with self-adjoint projections, for which this orthogonality
157.543
does hold, but I have elected not to do since the resulting ‘classifying space’ (6.44)
is cleaner. Two projections with the same range are homotopy through projections.
Consider some points in Υ−∞ S (Rn ; C2 ). To do so, let Qk be the orthogonal
projection on the span of the first k ∈ N Hermite functions in L2 (Rn ) (with respect
(i)
to an order in which |α| is increasing but it really does not matter) and let Qk ,
2 n 2
for i = 1, 2, be the corresponding projections on L (R ; C ) acting on the first
(i) (i)
and second factors. Thus the Qk commute with P1 and P2 = Id −P1 with Qk a
subprojection of Pi . Then
(2)

2(P1 + Qk ) − Id for k > 0

157.547 (6.47) βk = β∞ = 2P1 − Id for k = 0
 (1)
2(P1 − Q−k ) − Id for k < 0


are involutions.
The group G−∞S (Rn ) is a Fréchet manifold even though I have not defined the
meaning – simply because it is an open subset of the Fréchet space Ψ−∞
S (Rn ). It is
much less clear that this Grassmannian is a manifold. As a subset
157.548 (6.48) Υ−∞
S (R2 ; C) ⊂ Ψ−∞
S (Rn ; C2 ) is closed.

1If you are wondering, it is an Upsilon

 84 6. SMOOTHING OPERATORS AND K-THEORY

If it is to be smooth then it should have a decent tangent space, as a subspace of

Ψ−∞
S (Rn ; C2 ), at each point. So consider a smooth curve
157.549 (6.49) (−1, 1) 3 t 7−→ β + a(t) ∈ Υ−∞
S (Rn ; C2 ), a(t) ∈ Ψ−∞
S (Rn ; C2 ), a(0) = 0.
Certainly we must have
da da
157.550 (6.50) (β + a(t))2 = Id =⇒ β (0) + (0)β = 0.
dt dt
We can decompose any a0 = da/dt(0) ∈ Ψ−∞
S (Rn ; C2 ) as a sum of four terms,
determined by β,
157.551 (6.51) a0 = P a0 P + (Id −P )a0 P + P a0 (Id −P ) + (Id −P )a0 (Id −P ).
157.550
The condition in (6.50) reduces to
157.552 (6.52) P a0 P = 0 = (Id −P )a0 (Id −P ) =⇒ a(0) = P a0 (Id −P ) + (Id −P )a0 P.
Thus a tangent vector at β must be off-diagonal with respect to the decomposition
defined by β.
157.555 Proposition 6.2. Given  > 0 if a ∈ Ψ−∞
S (Rn ; C2 ) is sufficiently close to 0
−∞ n 2
and β ∈ ΥS (R ; C ) the resolvent
157.556 (6.53) (β + a − z Id)−1 exists in {z; |z − 1| > /2} ∩ {|z + 1| > /2}
and the contour integral
Z Z
1 −1 1
157.557 (6.54) βa = (β + a − z Id) dz − (β + a − z Id)−1 dz
2πi |z−1|= 2πi |z+1|=
gives a smooth local retraction to Υ−∞
S (Rn ; C2 ) 157.552
which is a bijection from the inter-
section with the tangent space in the sense of (6.52).
Proof. The resolvent of β is
1 1
R.24 (6.55) (β − z Id)−1 = (1 − z)−1 (β + Id) + (z + 1)−1 (β − Id)
2 2
from which it follows that proved the norm157.557 ka(0)k < δ is small (β + a − z Id)−1
is holomorphic
R2
outside the two disks in (6.54) where it is a smoothing pertur-
bation of (??). Standard arguments from the holomorphic calculus show that it
an involution. So the map from a small L2 ball in the linear space, given by
is157.551
(6.51), to Υ−∞S (Rn ; C2 ) is well-defined. As a map on bounded
157.551
operators on L2 it
is differentiable and restricted to the diagonal elements (6.51) it has derivative the
identity at a(0) = 0. Thus the Implict Function Theorem applies and the inverse
preserves regularity, showing that the retraction is indeed a local isomorphism from
the (putative) tangent space to Υ−∞ S (Rn ; C2 ). 
These retractions show that Υ−∞
S (Rn ; C2 ) is a smooth manifold locally mod-
elled on the tangent space at each point.
157.553 Theorem 6.1. The involutive Grassmannian Υ−∞ S (Rn ; C2 ) has components la-
belled by
1
157.554 (6.56) R-ind : Υ−∞
S (Rn ; C2 ) 3 β 7−→ Tr(β − β∞ )
2
and on each component G−∞
S (Rn ; C2 ) acts transitively by conjugation.
 4. THE INVOLUTIVE GRASSMANNIAN 85

The ‘relative index’ function R-ind clearly takes the value k on the elements βk in
157.547
(6.47).

Proof. The trace is well-defined, since β −β∞ is a smoothing operator. More-

over and smooth curve in Υ−∞
157.557 S (Rn ; C2 ) is locally the image of a curve in the tangent
space under (6.54), so we know that R-ind varies smoothly along a curve and has
derivative
1 dβ 1
R.25 (6.57) Tr( ) = Tr(a0 ) = 0
2 dt 2
157.551
constant on components of Υ−∞
because of (6.51). Thus R-ind is 157.547 S (Rn ; C2 ).
For the elements defined in (6.47)
R.26 (6.58) R-ind(βk ) = k.
Now, consider the action of g ∈ G−∞ S (Rn ; C2 ). Certainly gβg −1 is again an
involutive perturbation of β∞ so GS (R ; C2 ) does act on Υ−∞
−∞ n
S (Rn ; C2 ). We know
that GS (R ; C ) is path connected so it must fix components of Υ−∞
−∞ n 2
S (Rn ; C2 ).
−∞ n 2
It remains to show that any element of ΥS (R ; C ) is conjugate to one of the
βk . For a given β ∈ Υ−∞
S (Rn ; C2 ) with positive projector P = 21 (β + Id) let HP be
the range of P acting on L2 (R; C2 ), it is a closed subspace. Then the composite
operator
157.510 (6.59) P P1 : L2 (R) ⊕ {0} −→ HP
is Fredholm. This follows from the fact that
P1 P P1 = P0 + P1 (P − P1 )P1 = Id +Q on L2 (R) ⊕ {0}
is a compact (in fact Schwartz-smoothing) perturbation of the identity on the
range of P1 . Similarly P P1 P is a compact perturbation of the identity acting on
HP . Thus P P1 has finite dimensional null space and maps the orthocomplement
isomorphically onto a closed subspace of HP with a finite-dimensional (ortho)-
complement. However these two spaces may have different dimensions – that is
what R-ind(β) = Tr(P − P1 ) measures as we see below.
157.510
If the dimension of the null space of P P1 in (6.59) is larger than that of the
complement of the range then we may replace P1 by P1 (Id −Qk ) where k is the
difference and Qk is the finite-rank projector onto the first k Hermite functions.
This may not be contained in the null space but now the operator P P1 (Id −Qk )
has null space and complement of the range of the same dimension. We can then
add a finite rank smoothing operator S between these two spaces so that
157.512 (6.60) P P1 (Id −Qk ) + S = P1 + S 0 , S, S 0 smoothing
which is an isomorphism from the range of P1 (Id −Qk ) to HP . Then it follows that
157.513 (6.61) g = P1 +P2 +S 0 ∈ G−∞ (R; C2 ), P̃ = g(P1 (Id −Qk )g −1 : L2 (R; C2 ) −→ HP
is a projection with the same range as P. The null space Ran(Id −P̃ ) of P̃ is a
complement to the range and
157.514 (6.62) (Id −P ) : Ran(Id −P̃ ) −→ Ran(Id −P )
is an isomorphism since it has no null space and is surjective. It follows that
g̃ Id +(P̃ − P ) ∈ G−∞ (R2 ; C2 ) conjugates P̃ to P.
 86 6. SMOOTHING OPERATORS AND K-THEORY

If the dimension of the null space of P P1 is smaller than the range then a similar
argument follows with P1 replaced by P1 ⊕ Qk where Qk is the same projection but
now on the second factor. 157.547
This proves that the involution is conjugate to one of the options in (6.47) and
then it follows that the relative trace is either k or −k in the two cases. 

If you are prepared to swallow a bit of homotopy theory you can see why the
Grassmannian is relevant here. Consider the component containing β0 = β∞ :
157.560 (6.63) Υ−∞ n 2 −∞ n 2
0,S (R ; C ) = {β ∈ Υ0,S (R ; C ); Tr(β − β∞ ) = 0}.

The isotropy group of β∞ is

157.561 (6.64) G−∞
S (Rn ) × G−∞
S (Rn ) = {g ∈ G−∞
S (Rn ; C2 ); hβ∞ = β∞ g}
where the two factors act in the two components of C2 . Thus in fact we have shown
that there is a natural isomorphism
Υ−∞ n 2 −∞
(Rn ; C2 )/ G−∞ (Rn ) × G−∞ (Rn ) .


157.563 (6.65) 0,S (R ; C ) ≡ GS S S

These two subgroups commute with each other so we can also write the quotient
as
Υ−∞ n 2
0,S (R ; C ) ≡
(6.66)
 
157.562
GS (R ; C )/{Id} × GS (R ) /G−∞
−∞ n 2 −∞ n
S (Rn ) × {Id}.

The first quotient here is weakly contractible – since G−∞

S (Rn ; C2 ) can be retracted
−∞ n
onto {Id} × GS (R ). This means that this base component is the quotient by the
free action of G−∞
S (Rn ) on a contractible space.
157.564 Proposition 6.3. The base component of Υ−∞ n 2
0,S (R ; C ), is a classifying space
−∞
for GS (Rn ) and hence
(
−∞ n 2 Z j even
157.565 (6.67) πj (ΥS (R ; C )) =
{0} j odd

and Υ−∞ n 2
0,S (R ; C ) is a classifying space for even K-theory.

Proof. Not given here but from the long-exact sequence of Serre the homotopy
groups
157.361
can be identified from the identification of the homotopy groups of G−∞ in
(6.40). These are proved below, using some of the properties of Υ−∞ n 2
0,S (R ; C ) – but
157.565
not (6.67)! 

5. Semiclassical smoothing operators

If A ∈ S (R2n ) is the kernel of an element of Ψ−∞
S (Rn ) we can, as we did for
pseudodifferential operators, introduce
Z
157.373 (6.68) B(x, z) = A(x, x − z) and B(x, z) = (2π)−n b(x, ξ)e−iz·ξ dξ.
Rn

157.393 Definition 6.4. A semiclassical Schwartz family, A() ∈ Ψ−∞ n

sl,S (R ), is defined
by a smooth family
157.377 (6.69) a ∈ C ∞ ([0, 1] ; S (R2n ))
 5. SEMICLASSICAL SMOOTHING OPERATORS 87

by setting, for  > 0,

Z
A()u(x) = A(; x, y)u(y)dy, u ∈ S (Rn ) where
Rn
Z
157.374 (6.70) A(; x, y) = B(; x, x − y), B(; x, z) = (2π)−n a(; x, ξ)e−iz·ξ dξ
Rn
Z
−1
−n
= (2π) a(; x, η)e−i z·η dη.
Rn

For  > 0 this does not do very much but the family certainly becomes singular
at  = 0. L10-end
Said another way, an element of Ψ−∞ n
sl,S (R ) is a family of operators for  ∈ (0, 1]
L11

defined by
157.378 (6.71) A() = QL (a(; x, ξ)), a ∈ C ∞ ([0, 1]; S (R2n
x,η )).

The basic result is that these families compose to give an algebra.

157.375 Proposition 6.4. The operator product on Ψ−∞ S (Rn ), applied for  > 0 to
Z
157.376 (6.72) (A1 ◦ A2 )(; x, y) = A1 (; x, y 0 )A2 (, y 0 , y)dy 0
Rn

extends to a smooth product defining the algebra Ψ−∞ n

sl,S (R ).

Proof. Writing the composite family in terms of the kernels −n Bi (, x, x−y
 )
of the factors shows that
x−y
A1 ◦ A2 has kernel −n D(, x, ) where

y−z z−y
Z
157.489 (6.73) D(, x, t) = −n B1 (, x, t + )B2 (, z, )dz
R n  
y−z
Z
= B1 (, x, t − Z)B2 (, x + Z, Z)dZ, Z = .
Rn 
The integrand here is smooth in  with values in the Schwartz functions in the
variables x, t, Z – as follows from the fact that
157.490 (6.74) (1 + |x| + |t − Z| + |x + Z| + |Z|) is comparable to (1 + |x| + |t| + |Z|).
Thus the integral is also Schwartz. Expanding in Taylor series at  = 0 gives the
Moyal formula below.
Alternatively this follows from our earlier results on Ψ0 (Rn ). Observe that
157.379 (6.75) (1 + |ξ|) ≤ (1 + |ξ|),  ∈ (0, 1)
from which it follows that
157.380 (6.76) (0, 1) 3  −→ a(, x, ξ) ∈ S (Rn ; S 0 (Rn ))
is (uniformly) bounded. Indeed the derivatives satisfy

157.381 (6.77) ∂xα ∂ξβ a(; x, ξ) = |β| ∂xα ∂ηβ a(; x, η) η=ξ
=⇒
|∂xα ∂ξβ a(; x, ξ)| ≤ Cα,β |β|
(1 + |ξ|)−|β| ≤ Cα,β (1 + |ξ|)−|β| .
So our earlier composition result shows that uniformly for  > 0
157.382 (6.78) A1 () ◦ A2 () is bounded in Ψ0 (Rn )
 88 6. SMOOTHING OPERATORS AND K-THEORY

Moreover the left-reduced symbol of the composite is given by the asymptotic for-
mula
X
157.383 (6.79) A1 () ◦ A2 () = QL (c()), c() ∼ ∂ξα a1 (; x, ξ)Dxα a2 (; x, ξ)
α
X
= |α| ∂ηα a1 (; x, η)Dxα a2 (; x, η) η=ξ
.
α

Using Borel’s Lemma to sum the Taylor series gives c ∈ C ∞ ([0, 1]; S (R2n )) such
that
157.384 (6.80) X
c(; x, η) − |α| ∂ηα a1 (; x, η)Dxα a2 (; x, η) ∈ N +1C ∞ ([0, 1]; S (R2n )) ∀ N.
|α|≤N

The claim then is that

157.385 (6.81) E() = QL (c(; x, ξ)) − A1 () ◦ A2 () ∈ C ∞ ([0, 1]; Ψ−∞ (Rn )
dk E()
with = 0 ∀ k.
dk =0


From either proof we see that again ‘Moyal’s formula’ appears (I think it is
more historically legitimate here than before!)
Ai ∈ Ψ−∞ n ∞ 2n
sl,S (R ), Ai = QL (ai (, x, ξ)), ai ∈ C ([0, 1]; S (R )) =⇒

A1 ◦ A2 = QL (c), c ∈ C ∞ ([0, 1]; S (R2n ),

157.397 (6.82) X
c(, x, η) ' |α| ∂ηα a1 (, x, η)Dxα a2 (, x, η)
α

Equality here is in the sense of formal power series (i.e. Taylor series) at  = 0. So
this formula tells you nothing about what happens for  > 0 as is to be expected.
However, the symbol of the product is determined by this formula up to terms
vanishing to infinte order.
157.398 Lemma 6.4. If a ∈ C ∞ ([0, 1]; S (R2n )) vanishes to infinite order at  = 0 then
the semiclassical family
157.399 (6.83) A() = QL (a(, x, xξ))
is simply a smooth map [0, 1] −→ Ψ−∞ (Rn ) vanishing to infinite order at  = 0.
The analogue of the ‘principal symbol map’ for pseudodifferential operators is
played by the ‘semiclassical symbol’
A() = QL (a(, x, xξ)),  > 0 =⇒
157.394 (6.84)
σsl : Ψ−∞ n 2n
sl,S (R ) 3−→ a(0, y, η) ∈ S (Rx,η )
157.394
157.395 Proposition 6.5. The semiclassical symbol map (6.84) gives a short exact se-
quence of algebras

(6.85) Ψ−∞ n / Ψ−∞ (Rn ) σsl

/ S (R2n )
157.396 sl,S (R ) sl,S

with the commutative product on S (R2n ).

Proof. This is just the first term in the Moyal product. 
 6. THE GROUP G∞
sl 89

Make sure you understand what a semiclassical smoothing operator ‘looks like’.
Its kernel is of the form
x−y
157.408 (6.86) A (x, y) = −n B(, x, ), B ∈ C ∞ ([0, 1] ; S (R2n )

so as  ↓ 0 the kernel is ‘squashed’ around the diagonal.
There are semiclassical operators of finite order as well, I hope I will have time
to talk a little about them. A semiclassical differential operator on Rn might be of
the form
Xn
157.409 (6.87) P =  2 Dx2j + V (x).
i=1

Notice that if you set  = 0 only the zeroth order term survives. This is not the
semiclassical symbol which is instead the ‘rescaled’ symbol including the lower order
term:
157.410 (6.88) σsl (P ) = |ξ|2 + V (x).
157.409
Operators like (6.87) arose in quantum mechanices where  ' ~ is the ‘coupling
constant’ relating the frequency of spectral lines to the jump in energy between
electron shells which produces them. In practice ~ ' 1/127 is small and the idea
is to think of it as ‘very small’ and perturb from  = 0 (I don’t like to use ~ here
since it is actually a constant!) Simply setting  = 0 is a bad idea since most of
the problem disappears. What is happening here is that the problem is becoming
commutative as  ↓ 0 because of the commutation condition that
157.427a (6.89) [∂xi , xk ] = δik
so one is turning on the non-commutative product as  becomes positive.
There is a lot one can do with semiclassical operators – see for instance the
MR2952218
book of Zworski [9].

6. The group G∞
sl

As remarked above we want to consider the semiclassical group analogous to

G∞S (R n
). In fact we need to generalize the discussion above by allowing ‘smoothing
0
values’. If we consider the Schwartz smoothing operators on Rn +n then we know
that
0 0
157.400 (6.90) S (R2n +2n ) = S (R2n ; S (R2n )).
0
The smoothing operators on Rn +n can then be considered as ‘smoothing operators
with values in smoothing operators’, in either direction. We can do the same thing
for the semiclassical smoothing operators and consider the space
0
157.401 (6.91) S (R2n ; Ψ−∞ n
sl,S (R )).

Then we get a ‘stabilized’ algebra

0
157.402 (6.92) Ψ−∞ n −∞
sl,S (R ; ΨS (Rn )).
157.408
For the kernel of an element in this algebra (6.86) is replaced by
x−y 0
157.411 (6.93) A (x, y, z, t) = −n B(, x, , t, t0 ), B ∈ C ∞ ([0, 1] ; S (R2n +2n )).

 90 6. SMOOTHING OPERATORS AND K-THEORY

Observe what the form that the semiclassical symbol takes in this more general
case. It is just a symbol with values in the smoothing operators
0 0
157.406 (6.94) σsl : Ψ−∞ n −∞
sl,S (R ; ΨS (Rn )) −→ S (R2n
x,ξ ; Ψ
−∞
(Rn ))
157.396
giving a short exact sequence just like (6.85)
157.407 (6.95)
Ψ−∞ n −∞ 0
(Rn )) / Ψ−∞ (Rn ; Ψ−∞ 0
(Rn )) / S (R2n ; Ψ−∞ (Rn0 )).
sl,S (R ; ΨS sl,S S

So the product formula for the symbol in this case is multiplicative in the variables
0
(x, ξ) but with values in the non-commutative algebra Ψ−∞S (Rn ).
157.404 Proposition 6.6. The semiclassical group
0 0
−∞
157.405 (6.96) G∞ n n n −∞
sl,S (R ; S (R )) = {A ∈ Ψsl,S (R ; ΨS (Rn ));
0
∃ B ∈ Ψ−∞ n −∞
sl,S (R ; ΨS (Rn )) with (Id +A)(Id +B) = Id = (Id +B)(Id +A)}
0
is open in Ψ−∞ n −∞
sl,S (R ; ΨS (Rn )).
157.356
Proof. This is similar too, but a bit more involved than, the proof of Lemma 6.3.
0
The topology on the semiclassical algebra Ψ−∞ n
sl,S (R ; Ψ
−∞
(Rn )) comes from its
0 157.411 157.399
identification with C ∞ ([0, 1]; S (R2n+2n ) ) through either (6.93) or (6.83). From the
latter it follows that A() defines a bounded family of operators on L2 (Rn ) with
a uniform norm bound being a continuous seminorm. So if A lies in a sufficently
small neighbourhood of 0 the family, in , of operators
157.412 (6.97) (Id −A())−1 = Id +B() exists for  > 0
with B() uniformly bounded as  ↓ 0. So we only need to show that
0
157.413 (6.98) B() ∈ Ψ−∞ n −∞
sl,S (R ; ΨS (Rn )).
Proceeding ‘symbolically’ we can see that the model problem, the existence of
the inverse
0
157.414 (6.99) (Id +σsl,S (A))−1 (y, η) = Id +B0 (x, η) ∈ G−∞
S (R2n ),
has a unique solution,157.356
for A in a possibly smaller neighbourhood of 0. This indeed
follows from Lemma 6.3. Once we know that the principal symbol can be inverted,
the existence of a formal power series inverse follows from the behaviour of the Moyal
0
product. That is, we can find a sequence of elements Bk ∈ Ψ−∞ n −∞
sl,S (R ; ΨS (Rn ))
such that for any p
p
X 0
157.415 (6.100) (Id +A()(Id +B0 ()) + k Bk () ∈ Id +p+1 Ψ−∞ n −∞
sl,S (R ; ΨS (Rn )).
k=1

Using Borel’s lemma again we can sum the series and so find a ‘parametrix’ B 0 ∈
0
Ψ−∞ n −∞
sl,S (R ; ΨS (Rn )) such that
157.416 (6.101)
0
(Id +A()(Id +B 0 ()) − Id, (Id +B 0 ())(Id +A()) − Id ∈ C˙ ∞ ([0, 1]; Ψ−∞
S (Rn+n ))
(which is the same as saying a semiclassical
157.392
error vanishing to all orders at  = 0).
The corner identity as in (6.33) then shows that the difference of the inverse
and the parametric B − B 0 is uniformly bounded as a function of  ∈ (0, 1] with
0
values in Ψ−∞
S (Rn+n ). A similar argument for the derivatives with respect to 
 6. THE GROUP G∞
sl 91

shows that this difference is actually smooth and vanishes to infinite order at  = 0
157.413
so proving (6.98). 
157.407
For the semiclassical group the symbol sequence (6.95) gives the map
0 0
157.418 (6.102) σsl : G∞ n −∞
sl,S (R ; ΨS (Rn )) −→ S (R2n ; G∞ (Rn ))
which it is important to note is surjective. In fact we want a stronger lifting property
157.417 Proposition 6.7. The semiclassical group has the lifting property that for any
0
smooth symbol map u : M −→ S (R2n ; G∞ (Rn )) of compact support (reducing to
the identity outside a compact set) there is a smooth map, also of compact support
−∞ 0
ũ : M −→ G∞ n
sl,S (R ; ΨS (Rn )) giving a commutative diagramme

(6.103) G∞ n −∞
(Rn ))
0 σsl
/ S (R2n ; G∞ (Rn0 ))
157.419 sl,S (R ; ΨS O
i
u
ũ

M.
Moreover any two such lifts are smoothly homotopic.
This is another symbolic argument, just a uniform version of the sur-
Proof. 157.418
jectivity of (6.102).  RBM:Expand?

This might appear a little arcane! However a consequence of the existence of

157.419
a lift as in (6.103) is that this semiclassical group serves as a ‘bridge’ for odd K-
157.417
theory. Namely a family u ∈ C c∞ (R2n × M ; G−∞ ) as in Proposition 6.7 defines an
element of the group K 1 (R2n × M ). On the other hand the lifted group, restricted
to say  = 1 (or  = 21 if you prefer) defines an element of K 1 (M ), since it is valued
0
in G−∞S (Rn+n ). Now we are getting closer to Bott periodicity.
157.419
157.420 Proposition∗ 6.8 (Bott periodicity). The lifting construction in (6.103) de-
fines, for any manifold M and any n, a homomorphism
157.421 (6.104) K 1 (R2n × M ) −→ K 1 (M )
which is an isomorphism.
This is a ‘protypical’ index map, it is a non-geometric version of Atiyah-Singer
which is used in the proof of the general version below.
157.421
Once the existence of the map (6.104) is established there are two parts to the
proof that it is an isomorphism. First we show that it is surjective, by a computation
based on the existence of a ‘Bott element’ β ∈ K 0 (R2 ) which is a generator (so along
the way we prove that K 0 (R2 ) = Z). This is where the hard work lies – and I may
suppress some of it as far as the lectures are concerned (although it will all be in
the notes). The second part is a clever idea of157.421 Atiyah. This is based on the fact
2
that we can move
157.421
some of the factors of R in ( 6.104) into M. Indeed, once we have
the map (6.104) for n = 1 we can get a similar map by iterating – applying the
map
157.422 (6.105) q : K 1 (R2 × M ) −→ K 1 (M )
to R2n−2 × M instead of M. This leads one to think in terms of the iterated map
q q
K 1 (R4 × M ) −→ K 1 (R2 × M ) −→ K 1 (M )
 92 6. SMOOTHING OPERATORS AND K-THEORY

157.421
The way we are getting (6.104) can be 157.423
described as ‘turning on quantization’ by
writing R2n = Rnx ×Rnξ . Looking again at (6) we can think of R4 = R2x,ξ ×R2y,η where
we first turn on quantization in (x, ξ) for the first map, then turn on quantization
in (y, η) for the second map. So we conclude easily enough that this is the same
as doing both simultaneously in the sense that we get a commutative diagramme
forming the upper right triangle

157.424 (6.106) K 1 (R4 × M ) / K 1 (R2 × M )

 ' 
K 1 (R2 × M ) / K 1 (M )
157.421
where the diagonal map is (6.104) for n = 2. Atiyah’s idea has two parts. The first
is that we get a further commutative triangle on the lower left where the roles of
the two quantizations are interchanged – the second set of variables are quantized
RBM:Expand or refer first, giving the same final result.
However, we can do a little more, namely we can rotate the variables, smoothly,
by looking at x cos θ + y sin θ and dually on the other variables. This gives us a
smooth family of double quantization maps starting at the first and finishing at
the second, with the sign reversed. By homotopy invariance these must all give the
same result.
The way we use this depends on a157.422
157.419
good understanding of the lifting map in
(6.103) as giving us a right inverse to (6.105) which we can write somewhat myste-
riously as
157.425 (6.107) K 1 (M ) 3 κ 7−→ β ⊗ κ ∈ K 1 (R2 × M )
where as above, β is the Bott element of K 0 (R2 ). 157.422
K 1 (R2 × M ) is mapped to zero in K 1 (M ) by (6.105), so
So, suppose that κ ∈157.424
along the right side of (6.106). By surjectivity κ comes from β1 ⊗ κ ∈ K 1 (R4 ×
157.424
M ) along the top line of (6.106), so lifting in the first variables. Reversing the
quantization as discussed above we see that
157.426 (6.108) κ = ±q2 (β1 ⊗ κ) = ±β1 ⊗ q2 (κ) = 0
so κ = 0 and we have injectivity.
This argument depends on seeing that quantization in one set of variables
commutes with the lifting map in a different set of variables which we will see in
L11-end the lifting construction.
L12/13

7. Constructing the Bott element

I am trying to take a minimalist approach here to avoid getting bogged down
157.419
in K-theory. The
157.426
aim is to construct the lifting map in (6.103) and think of it as
giving (6.108). As already remarked, this is really an exterior product in K-theory
157.427 (6.109) K 0 (R2 ) × K 1 (M ) −→ K 1 (R2 × M )
coming from β ∈ K 0 (R2 ) which is a generator.
The standard model for even K-theory over a manifold, X, is in terms of pairs
of (compactly supported) complex vector bundles. Such a vector bundle can always
 8. QUANTIZATION OF THE BOTT ELEMENT 93

be realized as a family of projections in N × N matrices for sufficiently large N so

as a smooth map
157.515 (6.110) P : X −→ M (N, C), P (x)2 = P (x).
For the Bott element in K 0 (R2 ) this projection arises in 2 × 2 matrices and it
is convenient (but by no means necesary) to view the projection as coming from
a family of involutions. As we have seen above, if β is an involutive matrix (or
operator) meaning β 2 = Id, then
1
157.516 (6.111) P = (β + Id) is a projection and conversely β = P − (Id −P ) = 2P − Id .
2
The explicit 2 × 2 family of matrices we will to consider satisfies
 
2 2 1 0
157.428 (6.112) β : R −→ M (2, C), β = Id, β = = P1 − P2 near ∞.
0 −1
The ‘triviality near infinity’ is analogous to the group G∞ .
Explicitly in this sense the Bott element (although how it defines an element
of K 0 (R2 ) is not yet explained) is
eiθ sin(χ(r))
 
cos(χ(r))
157.430 (6.113) β(x, ξ) = −iθ .
e sin(χ(r)) − cos(χ(r))
Here (x, ξ) = r(cos θ, sin θ) are polar coordinates, χ ∈ C ∞ ([0, ∞)) is constant near
0 and ∞ and decreases monotonically from π to 0. Thus β is constant near 0 and
hence smooth and is equal to β∞ near ∞.
An easy computation shows

157.431 (6.114) β 2 = Id, β = 2Π(x, ξ) − Id,

cos2 ( 12 χ(r)) eiθ sin( 12 χ(r)) cos( 12 χ(r))
 
Π(x, ξ) = −iθ .
e sin( 21 χ(r)) cos( 12 χ(r)) sin2 ( 12 χ(r))
Thus the range of the projection Π(x, ξ) is the span of
cos( 21 χ(r))
 
157.432 (6.115) .
e−iθ sin( 21 χ(r))
This is a complex line bundle which comes from a non-trivial line bundle on S2 as
the one-point compactification of R2 .

8. Quantization of the Bott element

157.430
The way we will make use of the involution (6.113) is by quantizing it to a
semiclassical family of involutions.
Since b0 = β − β∞ is a compactly supported smooth function with values in
2 × 2 complex matrices we can apply semiclassical quantization to find
157.433 (6.116) B0 ∈ Ψ−∞ 2
sl,S (R; C ) with σsl (B0 ) = b0 = β − β∞

So, B0 is a 2 × 2 matrix of semiclassical smoothing operators on R. As with the

identity we take the semiclassical quantization of the constant matrix β∞ = P1 −P2
to be itself.
157.434 Proposition 6.9. There exists B ∈ Ψ−∞ 2
sl,S (R; C ) satisfying

157.435 (6.117) σsl (B) = b0 , (β∞ + B)2 = Id .

 94 6. SMOOTHING OPERATORS AND K-THEORY

157.433
Proof. First make any choice B0 as in (6.116). Then, from the semiclassical
symbol calculus the principal semiclassical symbol of the square satisfies
157.436 (6.118) σsl ((β∞ + B0 )2 − Id) = σsl (β∞ bo − bo β∞ + b20 ) = 0.
Thus in fact
157.437 (6.119) (β∞ + B0 )2 = Id +E1 , E1 ∈ Ψ−∞ 2
sl,S (R; C ).

Now, we proceed, by induction, to construct successive corrections Bk ‘as usual’

(cf. Project 2) so that
p
X
157.438 (6.120) (β∞ + B0 + k Bk )2 = Id +k+1 Ek+1 , Ek+1 ∈ Ψ−∞ 2
sl,S (R; C ).
k=1
p+1
Adding  Bk+1 , where Bk+1 has semiclassical symbol bk+1 what we need to
arrange in the inductive step is
157.439 (6.121) (β∞ + B)Bk+1 + Bk+1 (β∞ + B) = Ep+1 mod Ψ−∞ 2
sl,S (R; C )
⇐⇒ βbk+1 + bk+1 β = ek+1 = σsl (Ek+1 ).
This157.438
cannot be solved for an arbitrary ek+1 . However, it follows from the defi-
nition, (6.120), of the error at the previous stage of construction, that
157.440 (6.122) (β∞ + B)Ek+1 − Ek+1 (β∞ + B) ∈ Ψ−∞ 2
sl,S (R; C ) =⇒ βek+1 = ek+1 β.

This means that

157.493 (6.123) ek+1 = P ek+1 P + (Id −P )ek+1 (Id −P )
is ‘diagonal’ with respect to the two projections. Thus
1
157.441 (6.124) bk+1 = (P ek+1 P − (Id −P )ek+1 (Id −P ))
2
157.439
solves the inductive condition in (6.121).
Asymptotically summing the resulting series in k we find a semiclassical oper-
ator B 0 with the correct principal symbol such that
157.442 (6.125) (β∞ + B 0 )2 − Id ∈ ∞ Ψ−∞ 2
sl,S (R; C )

is a smooth family of smoothing operators vanishing to infinite order at  = 0.

It follows that as a family of bounded operators on L2 (R; C2 ) the spectrum is,
uniformly for 0 <  < 0 (for suitable small 0 ), concentrated very near to ±1. We
can then use a contour integral (the functional calculus for bounded operators on
a Hilbert space) to correct the quantization a little further, setting
I
1
157.443 (6.126) β∞ + B = B+ − B− , B± = ± (β∞ + B 0 − z Id)−1 dz.
2πi C ±
Here C ± are circular contours around ±1. Now an argument with resolvents shows
that the B± are commuting projections and
157.444 (6.127) (β∞ + B+ − B− )2 = Id
where B± are defined only for 0 <  < 0 but B = B+ − B− is equal to a semi-
classical family on 0 <  < 21 0 (simply by rescaling the parameter  to stretch the
interval [ 21 0 , 34 ] to [ 12 , 1]). The difference B −B 0 is a family of smoothing operators
vanishing to infinite order at  = 0. 157.433
This completes the proof of Lemma 6.116. 
 9. SIMPLIFYING THE QUANTIZED BOTT OPERATOR 95

9. Simplifying the quantized Bott operator K-theory

S.IGrass
RBM:Relate to §6.4
So this was hard work, but we need to do a little more to see what we have really
produced and refine it further. Choose an L2 -normalized element p ∈ S (R) ⊂ L2 (R)
and let Q also denote the orthogonal projection onto (0, e0 ) where e0 is the ground
state of the harmonic oscillator (it is just a convenient rank one projection).
157.477 Proposition 6.10. There is a semiclassical family B ∈ Ψ−∞ 2
sl,S (R; C ), quantiz-
157.435
ing β as in (6.117), such that
(6.128) β0 + B =1
= (Π1 + Q) − (P2 − Q).
Thus the effect of quantization is to move a one-dimensional space from the negative
to the positive eigenspace. This really is the fundamental ‘index theorem’ from this
point of view.
Proof. We first show that any (single, not semiclassical family) involution of
the form
157.517 (6.129) β0 + S, S ∈ Ψ−∞
S (R; C2 ), (β0 + S)2 = Id
can be deformed to a model. This argument does not depend on the particular β∞
nor on the dimension involved.
Where does the ‘one-dimensional’ come from. We have constructed P and
hence a projection
1
157.478 (6.130) P = (β∞ + B) + Id = Π1 + S, S ∈ Ψ−∞ 2 2
sl,S (R; C ) =⇒ P = P.
2
The range of P () for each  > 0 is a closed subspace of L2 (R2 ; C2 )
So the idea is that for  > 0 this is a projection which differs from Π1 by a
smoothing, hence trace classe, term. The relative index of these two projections is
therefore well-defined
157.479 (6.131) R() = Tr(P () − Π1 ) ∈ C ∞ ((0, 1]).
In fact, this is constant and necessarily an integer which we then proceed to com-
pute.
The constancy follows from the properties of the trace functional and projec-
tions. Differentiating
dR dB dB
157.480 (6.132) = Tr( ) = Tr( ) = 0 since
d d d
dB dB dB dB dB dB
=P + P =⇒ =P (Id −P ) + (Id −P ) P.
d d d d d d
Note that this is the same identity as used in the construction of B.
Having shown that R is constant we compute it in terms of the semiclassical
limit.
157.481 Lemma 6.5. For a semiclassical family of smoothing operators B ∈ Ψ−∞ N
sl,S (R; C ),
 Tr(B ) ∈ C ∞ ([0, 1]) and if the kernel is
157.483 (6.133) B = QL (b(, x, ξ))
then
Z Z
−1 ∂b
157.482 (6.134) Tr(B ) =  tr(b(0)) + tr( (0)) + O().
R2 R2 ∂
 96 6. SMOOTHING OPERATORS AND K-THEORY

Proof. 


10. Surjectivity of the periodicity map

Now we have a Bott element
157.388
and its quantization. Note that I defined K 0 (R2 )
1 3
to be K (R ) in Definition 6.2 so to get an element there we need to do just a little
more. Take a smooth function on R with ‘winding number one’ such as
157.445 (6.135) exp(iψ(s)), ψ ∈ C ∞ (R; R), ψ = 0 in s < −R, ψ = 2π in s > R.
Now consider the family of operators
157.446 (6.136) R 3 s −→ b (exp(iψ(s))b+ + exp(−iψ(s)b− ) .
Both factors here are semiclassical families, the second depending on s, (with
L12/13-end values in 2 × 2 matrices) which is the identity for |s| > R and

11. Manifolds with boundary

We have defined the K-theory of any manifold, including a manifold with
boundary, M. In this case there are three natural K-groups, well six including
even/odd groups. Namely
K 1 (M ) = C c∞ (M ; G−∞ )/homotopy
157.497 (6.137) K 1 (M ; ∂M ) = C c∞ (M \ ∂M ; G−∞ )/homotopy
K 1 (∂M ) = C c∞ (∂M ; G−∞ )/homotopy.
According to our definition the even groups are defined as the odd groups for R×M.
These are related by a ‘six-term sequence’
157.498 (6.138) K 0 (M ; ∂M ) / K 0 (M ) / K 0 (∂M )
O


K 1 (∂M ) o K 1 (M ) o K 1 (M ; ∂M )
The horizontal maps here are straightforward. Namely there is an inclusion
map
C c∞ (M \ ∂M ; G−∞ ) ,→ C c∞ (M ; G−∞ )
since the former are maps which are equal to the identity outside a compact set
and so can be extended (as the identity) up to the boundary. Similarly there is a
restriction map
|∂M
C c∞ (M ; G−∞ ) −→ C c∞ (∂M ; G−∞ )
and both these maps ‘descend’ through homotopies.
The end maps, the connecting homomorphisms, are as usual a little less obvious.
On the right we start from the group
C c∞ (R × ∂M ; G∞ ).
By a very simple version of the collar neighbourhood theorem a neighbourhood of
the boundary in M is diffeomorphic to a product
[0, 1) × ∂M ,→ M.
 12. CHERN CHARACTER-BUILDING 97

We can take the radial compactification of R to [0, 1] and thereby convert the map
157.503
(11) into a map
R × ∂M −→ M.
The definition in terms of rapid decay then means that there is a pull-back map
157.505 (6.139) S (R; C ∞ (∂M ; G−∞ )) ,→ C c∞ (M ; G−∞ )
which descends to give the connection homorphism on the right.
This leaves the connecting homomorphism on the left. This is where Bott
periodicity comes in. We are starting from (the components of) C c∞ (M ; G−∞ ) on
the lower left. However, by Bott perioidicty – Theorem ?? – we can instead start
from C c∞ (R2 × M ; G−∞ ). Now, separating the factors of R2 into R × R we can apply
the argument of the preceeding paragraph, giving the connecting homomorphism
on the right, to R × M instead of M, and this gives the map on the left. There is
an orientation issue here but since it amounts to a sign the choise one makes does
not affect the main result, which is:
157.498
157.499 Theorem 6.2. The six-term sequence (6.138) is exact for any manifold with
boundary.

12. Chern character-building

Project 3 P-Chern
I have laid rather heavy emphasis on the group(s) G−∞ S (Rn ) in the definition
of (complex, topological) K-theory. Indeed by fiat I have declared this to be a
classifying group for odd K-theory. Here I want you to sort out the map to deRham
cohomology leading to the Atiyah-Hirzebruch isomorphism.
In the first part I want you to explain, step by step, the meaning of the ‘odd
Chern character’
∞
X (−1)k k!
Tr (g −1 dg)2k+1


Ch(g) = k
(2πi) (2k + 1)!
k=0

This is a formal sum of forms in odd degree.

(1) Recall that G−∞S (Rn ) is an open subset of S (R2n ) and it is ‘an infinite
matrix group’. Use this to give a clear meaning to the Maurier-Cartan
form g −1 dg here.
Hint: Perhaps avoid going into a full discussion of Fréchet manifolds!
As an open subset of a Fréchet space the tangent space at each point g is
Ψ−∞
S (Rn ). Any element a ∈ Ψ−∞ S (Rn ) here defines a curve in the group
in the obvious way as g + ta and the identification with the tangent space
(with elements equivalence classes of curves) is written ‘dg’ – meaning I
think that d(g + ta)/dt is the tangent vector at g. Then g −1 acting on the
left on the group maps g to the origin. This is a concrete group in the
sense that this gives a linear map
g −1 : Ψ−∞
S (Rn ) −→ Ψ−∞
S (Rn ).
The push-forward on the tangent spaces would be denoted g∗−1 but since
the map is linear we can drop the ∗ and then
g −1 dg : Ψ−∞
S (Rn ) −→ Ψ−∞
S (Rn )
98 6. SMOOTHING OPERATORS AND K-THEORY

is the natural map from the tangent space at g to the tangent space at Id
(which is the Lie algebra).
(2) Then the formal product (g −1 dg)j is supposed to be the j-fold exterior
product, with composiition thrown in, so it is a j-multilinear, totally
antisymmetric map from j copies of the tangent space at g to the tangent
space at Id . P3.14
(3) This explains (12) in the sense that the trace functional results in each
term being a (2k + 1)-multilinear function on the tangent space at each
point, which is what a form should be.
(4) The trace property Tr([a, b]) = 0 shows that
Tr((g −1 dg)2k ) = 0
which explains why there are no even terms.
(5) The deRhan differential is easy to define on 1-forms, even with values in
an infinite dimensional space, show that
d(g −1 dg) = −g −1 dg ∧ g −1 dg
in an appropriate sense.
(6) Conclude that the terms in Ch(g) are all closed.
(7) Suppose [0, 1] 3 t 7−→ G−∞
S (Rn ) is a smooth curve, show that
d
Ch(gt ) = d Et(g)
dt
where Et is the ‘Eta’ or Chern-Simons form
∞
(−1)k−1 k!
 
−1 dg
X
−1 2k
Et(g) = Tr g (g dg)
(2πi)k (2k)! dt
k=0

Hint: Make sense of the formula

d −1 dg dg
g dg = −g −1 g −1 dg + g −1 d .
dt dt dt
(8) Now, suppose M is a compact manifold and κ : M −→ G−∞ S (Rn ) is a
smooth map. Show that the pull-backed form is closed and so defines a
map
C ∞ (M ; G−∞
S (Rn )) 3 κ 7−→ κ∗ Ch ∈ HdR
odd
(M ).
(9) Conclude that this map descends to a map
Ch : K 1 (M ) −→ HdR
odd
(M ) the odd Chern character.
(10) Contemplate why this might induce the Atiyah-Hirzebruch isomorphism
K 1 (M ) ⊗ C −→ HdR
odd
(M ).
(So this means torsion is killed in the K-group. This does not quite work
for general non-compact manifolds). P3.20
(11) Now, if you have the energy, do the even version! Start from (7) and show
−∞
that on the group C ∞ (R; GS (Rn )), which is used above to define even
K-theory,
∞
(−1)k−1 k!
 
−1 dg
X
−1 2k
Chev = Tr g (g dg) dt
(2πi)k (2k)! dt
k=0
 13. ISOTROPIC K-POP 99

defines a closed form in each even degree. [I am not sure I have the
normalizing constants correct, but here it does not matter.]
(12) Conclude, proceeding as above, that this defines a map
K 0 (M ) ⊗ C −→ HdR
ev
(M ).

13. Isotropic K-pop

Project 4 P-Iso
P3b.0 Recall to isotropic pseudodifferential algebra, for simplicity of order 0,
on R2n is defined by quantization of the classical symbols joointly in the
variables (x, ξ) :
Ψ0iso (Rn ) = QL (C ∞ (R2n ).
The product defines a smooth bilinear map
Ψ0iso (Rn ) × Ψ0iso (Rn ) −→ Ψ0iso (Rn ).
We can then define C ∞ (M ; Ψ0iso (Rn ) for any manifold M and in particular
0
S (R2n ; Ψ0iso (Rn )).
Now, show that this allows us to define the ‘stabilized algebra’
0
Ψ0iso (Rn ; Ψ−∞
S (Rn )
by using
P3.3
the product of Schwartz smoothing operators and then the prod-
uct (13) to define
0 0 0 0
Ψ0iso (Rn ; Ψ−∞
S (Rn )×Ψ0iso (Rn ; Ψ−∞
S (Rn ) ←→ S (R2n ; Ψ0iso (Rn )×Ψ0iso (Rn )) −→ Ψ0iso (Rn ; Ψ−∞
S (Rn ).
0
This is an associative algebra of bounded operators on L2 (Rn+n ) which
0
maps S (Rn+n ) to itself.
(1) One reason that this algebra is interestimg is that it has a principal symbol
map
0 0
σ0 : Ψ0iso (Rn ; Ψ−∞
S (Rn ) −→ C ∞ (S2n−1 ; Ψ−∞
S (Rn )
which is multiplicative and gives a short exact sequence

Ψ−1 n −∞
(Rn )
0
/ Ψ0 (Rn ; Ψ−∞ 0
(Rn ) / C ∞ (S2n−1 ; Ψ−∞ 0
(Rn ).
iso (R ; ΨS iso S S

(2) We need to massage this a little more by identifying

R2n−1 −→ S2n−1
as the 1-point compactification – choose your favourite point (mine is the
South Pole). Anyway, this allows us to map
0 0
S (R2n−1 ; Ψ−∞ n ∞ 2n−1
iso (R ) ,→ C (S ; Ψ−∞
S (Rn )
with image being the subspace of functions vanishing to infinite order at
the point at infinity.
0
(3) Denote by Ψ̇0iso (Rn ; Ψ−∞
S (Rn ) which is the inverse image under σiso of
0
S (R2n−1 ; Ψ−∞ n
iso (R )
100 6. SMOOTHING OPERATORS AND K-THEORY

(4) This is an algebra without identity, so add Id to get a ring which we can
denote
0
Id +Ψ̇0iso (Rn ; Ψ−∞
S (Rn ).
Check that we now a multiplicative exact sequence
Id +Ψ−1 n −∞
(Rn )
0
/ Id +Ψ̇0 (Rn ; Ψ−∞ 0
(Rn ) / Id +S (R2n−1 ; Ψ−∞ 0
(Rn ).
iso (R ; ΨS iso S S

(5) Finally we can finish the setup by defining

0
n 0 0
o0
Ė(Rn ; Ψ−∞ (R2n )) = A ∈ Id +Ψ̇0iso (Rn ; Ψ−∞
S (Rn ; σiso (A) ∈ C ∞ (R2n−1 ; G−∞
S (Rn )
and now there is a multliplicative exact sequence
Id +Ψ−1 n −∞
(Rn )
0
/ Ėiso (Rn ; Ψ−∞ 0
(Rn ) / S (R2n−1 ; G−∞ 0
(Rn ).
iso (R ; ΨS S S

14. Bott and Clifford

Project 5 P-BC
In lectures I wrote down an explicit family of involutive matrices on R2 rep-
resenting the ‘Bott’ element (in even K-theory) and showed that the semiclassical
quantization of this family is a 1-dimensional shift from the constant family. Here I
ask you to do this for R2n . Initially I assumed you knew about the structure of the
complex Clifford algebra, I have now added a brief derivation in case you do not.
(1) Deconstruct the matrix I simply wrote down in lectures
eiθ sin(χ(r))
 
cos(χ(r))
µ = −iθ .
e sin(χ(r)) − cos(χ(r))
First write it out as
µ = cos(χ(r))Z + sin(χ(r)) (cos θE1 + sin θE2 ) .
Show that these 2 × 2 matrices satisfy
Ei2 = Z 2 = Id, E1 E2 + E2 E1 = ZE1 + E1 Z = ZE2 + E2 Z = 0.
(2) Recall the complexified Clifford algebra on a real Euclidean vector space,
V, of even dimension, 2n, defined as the quotient of the infinte tensor
algebra
X
T (V ) = C ⊕ VC ⊕ (VC ⊗ VC ) ⊕ · · · = VC⊗k , VC = V ⊗ C
k

by the two-sided ideal (under tensor product) generated by the elements

I (V ) 3 ξ ⊗ η + η ⊗ ξ − 2hξ, ηi, ξ, η ∈ V.
Here h, i is the Euclidean inner product. Thus,
Cl(V ) = T (V )/I (V ).
Remind yourself that this has dimension 22n , the same as the exterior
algebra. You should also note that it is isomorphic to the 2n × 2n matrix
algebra.
The Clifford algebra is closely related to the exterior algebra, but on
Cn rather than R2n . So we are passing to the standard complex structure
on R2n in which a basis, over the complex numbers, is
fi = ei + iei+n , i = 1, . . . , n.
 14. BOTT AND CLIFFORD 101

Show that Cl(R2n ), the complexified Clifford algebra (not the Clifford
algebra on the complexification!) acts on λ∗ Cn through the formula we
saw for the Hodge-Dirac operator
√ √
cl(ek + iek+n )f α = i 2fk ∧ f α , cl(ek − iek+n )f α = −i 2ι(fk )f α , k ≤ n
where the constants are for length normalization an here α is a strictly
increasing sequence in {1, . . . , n}. This specifies the action of all basis
elements and we see that
cl(ek + iek+n ) cl(ek − iek+n ) + cl(ek − iek+n ) cl(ek + iek+n ) = 2 Id,
cl(ek + iek+n )2 = 0 = cl(ek − iek+n )2
=⇒ cl(ek ) cl(ek0 ) + cl(ek0 ) cl(ek ) = 2δkk0 Id, k, k 0 = 1, . . . , 2n.
(3) If ei is an (oriented) orthonormal basis of V, let Ei be the corresponding
elements in the Clifford algebra so
Ei Ej + Ej Ei = δij Id .
Then show that the Clifford algebra has a ‘maximal element’
Z = in(2n−1) E1 E2 . . . E2n =⇒ Z 2 = Id, ZEi + Ei Z = 0.
P3.28
(4) Okay, now observe that (1) is now
P3.35 (6.140) µ(ζ) = cos(χ(|ζ|)Z + sin(χ(r))(ζ̂ · E∗ )
where E∗ = (E1 , . . . , E2n ) is thought of as a vector with values in matrices
and · is the inner product. P3.35
(5) Now observe that the definition (6.140) extends to V = R2n to give a
smooth family of involutive matrices. Here, as before χ(r) is decreasing
from π near 0 to 0 near ∞.
(6) Review the proof in the notes to check that there is a semiclassical family
of idempotents qunatizing µ (so now to smoothing operators on Rn ).
(7) For a bonus, show that the quantization is again has relative index 1
(assuming I got the signs right which would be a pleasant accident). You
might like to do this by quantizing in two variables repeatedly, so working
by induction over n.
 CHAPTER 7

Operators on manifolds
C.OpMa L14
In this second half of the course I will, finally, get to the discussion of analysis
on manifolds. First what I hope is a reminder of the invariant description of kernels
of operators.

1. Functions and densities

The basic object defined on a smooth manifold M is the algebra of smooth
functions C ∞ (M ). Since we will want to include non-compact manifolds we need to
also consider the subspace of functions of compact support
157.566 (7.1) C c∞ (M ) ⊂ C ∞ (M ).
In general there is no good analogue of schwartz functions on a non-compact man-
ifold but on a manifold with corners the space, C˙ ∞ (M ), of functions vanishing to
infinite order at the boundary satisfies

157.567 (7.2) C˙ ∞ (M ) ⊂ C ∞ (M )
⊂

C˙c∞ (M ) ⊂ C c∞ (M )
and does correspond to the Schwartz space at least in the sense that
157.568 (7.3) C˙ ∞ (Rn ) = S (Rn ).
Recall the change-of-variable formula for the integral of functions on Euclidean
space. If F : Ω0 −→ Ω is a diffeomorphism between open subsets of Rn and
f ∈ C c∞ (Ω) then
Z Z
∂F
157.569 (7.4) F ∗ f | det |dx = f (y)dy.
Ω0 ∂x Ω
So there is no invariant integral of functions on a manifold.
To integrate we need something that transforms with the absolute value of the
157.569
determinant of the Jacobian of the diffeomorphism that appears in (7.4). This
is provided by densities. Recall that on an n-manifold the maximal degree, n−,
forms at a point, Λnm M for m ∈ M, may be defined as the linear space of totally
antisymmetric multilinear maps
157.570 (7.5) Tm M × n factors × Tm M −→ C.
Equivalently if we let λn Tm M denote the totally antisymmetric part of the n-fold
∗
tensor product of Tm M with itself (so Λnm M = λn Tm M ) then Λnm M is the space
of linear maps
157.571 (7.6) µ : λn Tm M −→ C.
103
 104 7. OPERATORS ON MANIFOLDS

The closely related space of t-densities, for t ∈ R, is

157.572 (7.7) Ωtm M = {ν : λn Tm M \ {0} −→ C; ν(sv) = |s|t ν(v), ∀ t ∈ R \ {0}}.
So these are t-absolutely homogeneous maps. That this is a vector space de-
pends on the one-dimensionality of λn Tm M. Just as for the form bundles these
one-dimensional vector spaces form a smooth bundle over M
157.573 (7.8) Ωt M −→ M.
The particularly important case t = 1 is just denoted ΩM.
So if µ ∈ C ∞ (M ; Λn M ) is a smooth n-form on an n-manifold then
157.574 (7.9) |µ|t ∈ C 0 (M ; Ωt ) is a continuous density.
A non-vanishing
157.574
n-form on M exists if and only if M is orientable and then the
t-density (7.9) becomes smooth. Note157.572
that positivity of a density is well-defined
because of the transformation law in (7.7).
157.575 Lemma 7.1. The density bundles are always trivial, i.e. there exists a smooth
positive section of Ωt M.
Proof. In local coordinates the ‘Lebesgue section’ |dx|t is smooth and positive
so summing over a partition of unity gives a global smooth positive section. 
157.576 Proposition 7.1. There is a well-defined integral
Z
157.577 (7.10) : C c∞ (M ; Ω) −→ C

for any manifold.

Of course this can be extended to locally integrable (but compactly supported)
sections of ΩM.
157.578 Lemma 7.2. There are natural isomorphisms
∗ ∗ t
157.579 (7.11) Ωt M ⊗ Ωs M = Ωs+t M, Ωt (M × N ) = πM Ωt M ⊗ π N Ω N.
So one can define an integral on C c∞ (M ) by choosing a density (probably pos-
itive) ν ∈ C ∞ (M ; Ω) and using
Z
157.580 (7.12) f ν for f ∈ C c∞ (M ).
M
It is just that there is, in general, no natural choice of ν. On a Riemannian manifold
there is an associated, smooth positive, Riemann density νg defined as |α1 ∧ . . . αn |
∗
for any orthonormal basis αi of Tm M.

2. Distributions
The existence of the integral on densities leads to consistent pairings
157.581 (7.13) C c∞ (M ) × C ∞ (M ; Ω)
O

? )R
C c∞ (M ) × C c∞ (M /
 _ ; Ω) 3 (f, g) 5
fg ∈ C


C ∞ (M ) × C c∞ (M ; Ω)
 3. VECTOR BUNDLES 105

Each of the spaces of smooth sections of Ω has a topology coming from the
isomorphism with the corresponding space of functions given by the choice of a
smooth positive section of Ω (and the fact that multiplication by a positive smooth
element f ∈ C ∞ (M ) is an isomorphism on both C c∞ (M ) and C ∞ (M )). This means
that we can define the spaces of distributions by

C c−∞ (M ) = {u : C ∞ (M ; Ω) −→ C linear and continuous}

157.582 (7.14)
C −∞ (M ) = {u : C c∞ (M ; Ω) −→ C linear and continuous}.

This is arranged so that these are ‘generalized functions’ with natural inclusions

157.583 (7.15) C c∞ (M ) ,→ C c−∞ (M ), C ∞ (M ) ,→ C −∞ (M ).

3. Vector bundles
In fact we want to generalize this in several respects. First for sections of a
vector bundle. I am already assuming you know what a vector bundle is. Just for
completness sake let me remind you that a vector bundle (either real or complex)
over a manifold M is another manifold V (often called the total space of the vector
bundle) with a surjective smooth map

157.584 (7.16) π : V −→ M

such that each fibre Vm = π −1 (m) ⊂ V has a linear space structure which is ‘smooth
and locally trivial’. This means that each point m ∈ M has a neighbourhood Om
for which there is a diffeomorphism giving a commutative diagramme

157.585 (7.17) π −1 (Om )

F / Om × KN

π π1
$ z
Om

with F linear on each fibre. Here K stands for R or C in the real or complex case
respectively. One can always assume that the Om are coordinate patches on M
157.585
and then the maps (7.17) give a special atlas on V – the transition conditions are
automatic.
There are sections of any vector bundle just as for functions (and densities)
C c∞ (M ; V ) ,→ C ∞ (M ; V ) corresponding to smooth maps u : M −→ V with π ◦ u =
IdM .
The dual bundle, V 0 , of a vector bundle V is defined, as a set, as the union of
the duals of the fibres
G
157.586 (7.18) V0 = Vm0 .
m∈m
157.585
It has a smooth structure coming from the local trivializations (7.17) by replacing
the map Fm : Vm −→ KN by (F t )−1 : Vm0 −→ KN .
There is then a pointwise pairing

157.587 (7.19) C ∞ (M ; V ) × C ∞ (M ; V 0 ) −→ C ∞ (M )
 106 7. OPERATORS ON MANIFOLDS

157.581
and this leads to pairings analogous to (7.13):
Z
hf, gim , f ∈ C c∞ (M ; V ), g ∈ C ∞ (M ; V 0 ⊗ Ω), or

f ∈ C ∞ (M ; V ), g ∈ C c∞ (M ; V 0 ⊗ Ω).
Then, noting that there are again appropriate topologies, the spaces of distribu-
tional sections are defined as the spaces of continuous linear maps
C c−∞ (M ; V ) = {u : C ∞ (M ; V 0 ⊗ Ω) −→ C; linear and continuous}
157.589 (7.20)
C −∞ (M ; V ) = {u : C c∞ (M ; V 0 ⊗ Ω) −→ C; linear and continuous}.

4. Operators and kernels

The operators we will consider on a manifold are, at worst, continuous linear
maps
157.590 (7.21) A : C c∞ (M ; V ) −→ C −∞ (M ; W )
taking smooth sections of one vector bundle to distributional sections of another.
In fact we generally will not encounter anything quite this general. One of the
important features of distributions is that these operators are again distributions.
157.591 Theorem 7.1 (Schwartz’
157.590
kernel). There is a 1-1 correpondence between con-
tinuous linear maps (7.21) and distributions
157.592 (7.22) KA ∈ C −∞ (M 2 ; πL
∗ ∗
W ⊗ πR (V 0 ⊗ Ω)).
So I am assuming you see how to define the tensor product of two bundles and the
pull-back of a bundle under a smooth map. Here πL and πR are the projections
from M 2 to M as the left and right factors.
Proof. None given, since I do not really use this result below. Still I should at
least specify what I mean by the continuity which is in terms of the weak topology
on the image. This is really the first part of the proof. Namely we can replace A,
157.592
in (7.22), by a bilinear map
157.593 (7.23) C c∞ (M ; W 0 ⊗ Ω) × C c∞ (M ; V ) ∈ (f, g) 7−→ A(g)(f )
from which A can be recovered. Continuity of A actually means separate continuity
of this bilinear form. The theorem is that such a bilinear form extends uniquely to
a distribution KA under the (bilinear) inclusion
157.594 (7.24) C c∞ (M ; W 0 ⊗ Ω) × C c∞ (M ; V ) ,→ C c∞ (M 2 ; πL
∗
(W 0 ⊗ Ω) ⊗ πR
∗
V)
given by pulling back under the projections.
In this sense the Schwartz kernel theorem is an infinite-dimensional generaliza-
tion of the identifications for finite-dimensional vector space
157.595 (7.25) {L : V −→ W ; linear} ←→ {BL : W 0 × V −→ C; bilinear} ←→ W ⊗ V 0 .
The proof is not actually so hard. 
So, the way we will use this is by specifying operators in terms of their kernels,
157.596 (7.26) (Av)(w) = KA (w ⊗ v).
Note the reversal of order which is made so that pairing corresponds to proximity.
 6. CORNORMAL DISTRIBUTIONS AT THE ZERO SECTION 107

157.602 Remark 6. I will generally identify operators and their kernels, using the same
letter to denote both.
If M is not compact we also need to think again about the ‘calculus of supports’.

5. Smoothing operators
Suppose M is a compact manifold without boundary, so all the support annoy-
ances are absent.
157.597 Definition 7.1. The space of smoothing operators between sections of two
bundles, V, W over a compact manifold, M, is defined by the space of kernels
157.598 (7.27) Ψ−∞ (M ; V, W ) = C ∞ (M 2 ; πL
∗ ∗
W ⊗ πR (V 0 ⊗ Ω)).
When W = V we abbreviate the notation to Ψ−∞ (M ; V ).
157.599 Remark 7. Smoothing operators are characterized (for M compact) by the
two conditions that
A : C −∞ (M ; V ) −→ C ∞ (M ; W ) and
157.600 (7.28)
At : C −∞ (M ; W 0 ⊗ Ω) −→ C ∞ (M ; V 0 ⊗ Ω).
For M compact it is again the case that the smoothing operators on a fixed
bundle V – so with W = V – form an algebra. The composition written in terms
of kernels is
Z
157.601 (7.29) A◦B = A(·, m)B(m, ·).
M
Make sure you understand why the integral here makes sense and gives again a
smoothing operator.
157.603 Proposition∗ 7.2. The group of invertibles in the ring Id +Ψ−∞ (M ; V ) for
a vector bundle over a compact manifold is isomorphic (but not naturally so) to
G−∞
S (Rn ).
To prove this we need a decent basis for L2 (M ; V ) – we will get this as the
eigenbasis for any self-adjoint elliptic pseudodifferential operator acting on V.
157.604 Exercise 10. Give an appropriate definition for the space Υ−∞ (M ; V ⊗ C2 )
for any complex vector bundle, V, over a compact (connected) manifold M and
check that it is isomorphic to Υ−∞
S (Rn ; C2 ).
L14-end
L15
6. Cornormal distributions at the zero section
Let W be a real vector bundle (any complex vector bundle has an underlying
real bundle) over a manifold S. We proceed to introduce the space of conormal
distributions on W, the total space of the vector bundle, relative to S appearing
as the zero section of W. We will want this to be a module over S (W ) = C˙ ∞ (W )
and the elements should be singular only at the zero section. In the case that
W = S × RN we know what we want, namely
m− 4s
157.605 (7.30) ISm (S × RN ; S × {0}) = C ∞ (S; IS (RN )).
Here s = dim S and the only the shift in the order is questionable. We already
considered such a space (with restricted coefficients) in case that S = Rn and
 108 7. OPERATORS ON MANIFOLDS

N = n in defining the kernels of pseudodifferential operators on Rn . Note that

what this says about the symbol involved is that
s N
157.606 (7.31) ISm (S × RN ; S × {0}) =⇒ u(s) = F −1 (a(s, ξ)), a ∈ S m− 4 + 4 (RN )
so it is consistent with our earlier notation.
One thing we can see immediately is that we can restrict such distributions to
a submanifold Y ⊂ S, since this just restricts the smooth map
157.617 (7.32) Y
: ISm (S × RN ; S × {0}) −→ ISm+codim Y (S × RN ; S × {0}), Y −→ S.
157.618 Remark 8. In sofar as I can remember the normalization, this is one way to
do it. First, restricting to a submanifold of S increases the (apparent) order by
1/4 for each dimension, i,e, by codim(Y, S)/4 as indicated. Secondly when ‘base
dimension’ s = dim S and ‘fibre dimension’ are equal the order is equal to the order
of the symbol.
Why is the order so normalized you might well ask? The ultimate reason is the
appearnce of ‘Lagrangian deistribution’ which do not have a strict separation of
base and fibre dimension. Together with the fact that we want a pseudodifferential
operator to have the same order as its symbol (so that it is consistent with the order
of differential operators) that determines the normalization. The normalization
is further complicated (to those like me who are numerically challenged) by the
fact that there are three dimensions as work here. As well as s and n there is
the dimension of W as a manifold, n = s + N. So, N is the codimension of the
submanifold S (the zero section) in W which is of dimension
157.606
n. It follows that the
relation ship between the orderI M and a ∈ S m in (7.31) is
0 n N
157.619 (7.33) u ∈ I M (W, S) ⇐⇒ a ∈ S M , M 0 = M + − .
4 2
Let’s hope I am right. (See Problems 5?).
We also know what happens if we subject such a conormal distribution to a
linear transformation on the fibres, even if it depends on the point in S :
157.607 (7.34)
A ∈ C ∞ (S; RN ), u ∈ ISm (S × RN ; S × {0}) =⇒ u(s, A(s, ·)) ∈ ISm (S × RN ; S × {0})
u = F −1 (a(s, ξ)) =⇒ u(s, A(s, y)) = F −1 (a(s, ((A−1 )t ξ)| det A(s)|).
157.608 Definition 7.2. If W is a real vector bundle of rank N over S, a manifold of
dimension s then we define
157.609 (7.35) ISm (W ; 0W ) ⊂ C −∞ (W )
to consist of those distributions u ∈ C −∞ (W ) such that if φ ∈ C c∞ (S) is supported
in an open set O over which W is trivial, τ : W O ≡ O × RN then
157.610 (7.36) π ∗ φu = τ ∗ (vO ), vO ∈ ISm (O × RN ; O × {0}).
157.607
We can see from (7.34) that the symbol of such a distribution, defined locally,
will transform as a density on the dual bundle. So let’s make sure that this is
well-defined.
157.613 Lemma 7.3. Suppose U is a real vector bundle over a manifold S and E is a
complex vector bundle over S then the space of symbols on (the fibres of ) U with
values in E is well-defined by reference to local trivializations.
Proof. 
 6. CORNORMAL DISTRIBUTIONS AT THE ZERO SECTION 109

157.611 Proposition 7.3. Under fibre-wise inverse Fourier transform the space of Schwartz
conormal distributions with respect to the zero section of a vector bundle is globally
isomorphic to a space of symbolic densities
dim W dimfib W
157.612 (7.37) ISm (W ; 0W ) ∈ F −1 (C ∞ (S; S m+ 4 − 2 (W 0 ; ΩW 0 ))).
Proof. 
157.620 Proposition 7.4. If W −→ S is a real vector bundle, U −→ S is a complex
vector bundle and
157.621 (7.38) F : D2 −→ D1
is a diffeomrophism between open neigbhourhoods of the zero section of W which
maps the zero identically to itself, then
157.622 (7.39)
F ∗ : {u ∈ I m (W, OW , U ); supp(u) b D1 } −→ {v ∈ I m (W, OW , U ); supp(u) b D2 }.
Proof. Since conormal distributions are smooth away from the zero section
we may shrink the domain D1 with D2 replaced by the image and the result remains
unchanged.
We also know that the conormal space is a module over C ∞ (S) so we can use
a partition of unity in S to localize to an open sets over each of which W is trivial.
So we may assume W = S × RN .
To see how much remains to be proved, consider the following factorization of
F – in a possibly smaller domain
157.623 (7.40) F = L ◦ F̃ .
Here G is the fibre-preserving diffeomorphism constructed from the differential of
F as follows. At each point s ∈ S the tangent space to W is the direct sum
157.624 (7.41) Ts W = TS S ⊕ RN
where the second part is the tangent space of the fibre. Since S is fixed pointwise
by F the differential maps is the identity on Ts S and so decomposes into the sum
of two linear maps
157.625 (7.42) F∗ RN 3 w 7−→ (L0 w, Lw) ∈ Ts S ⊕ Rn .
So L is a well-defined family of linear maps on RN . We already know the invariance
under such maps so it suffices to consider he remainder
157.623
term F̃ .
This defines the first, fibre-linear, factor in (7.40), and hence the second. By
construction this has differential at s ∈ S which is the sum of the identiy and the
‘off-diagonal’ term mapping Ts Ws to Ts S. This means that the map itself is of the
form, in any locall trivialization
157.626 (7.43) F̃ (s, v) = (s, w) + v · h(s) + E(s, v), E(s, v) = O(|v 2 ), v ∈ RN .
Then we can deform F̃ to the identity through the family of diffeormorphisms
157.627 (7.44) F̃ (s, v) = (s, w) + t(v · h(s) + E(s, v)), t ∈ [0, 1].
Now we can use the ‘homotopy method’ again. Very close to S this 1-parameter
family of diffeomorphism is generated by a 1-parameter family of vector fields
d ∗
157.628 (7.45) F u = Ft∗ (Vt u), Vt = vanishes at S and Vt vj = O(||v|)2 .
dt t
 110 7. OPERATORS ON MANIFOLDS

Now we replace u by a smooth family ut were we want to arrange that

dut
157.629 (7.46) u1 = u, + Vt ut ∈ C ∞ .
dt


7. Conormal distributions at a submanifold

8. Pseudodifferential operators
L15-end PseudoM
L16 We now have completed the definition of the basic space of sections of a complex
vector bundle conormal with respect to a closed embedded submanifold. Moreover
we have shown that there is a well-defined (principal) symbol map
157.630 (7.47)
0 1 1
σ : I m (M, S; V ) −→ S m (N ∗ S; V ⊗ Ω(N ∗ ∗S)), m0 = m + dim M − codim S.
4 2
Here Ω(N ∗ S) is a trivial real line bundle over S with fibre the space of densities on
the fibres of N ∗ S. We know that this captures the leading part of the singularity
of the elements so that

157.631 (7.48) I m−1 (M, S; V )  / I m (M, S; V ) / / S m0 (N ∗ S; V ⊗ Ω(N ∗ ∗S))

is a short exact sequence. In the process of examining the coordinate-invariance

the error terms were all local at the symbolic level. This means the two sets
157.632 (7.49) WF(u) ⊂ N ∗ S \ 0S
defined as the complement of the larges open cone where where the local full symbol
is rapidly decaying is invariantly defined.
We use this space to define the space of pseudodifferential operators acting
between any two complex bundle vector bundles over a manifold in terms of their
kernels
∗ ∗
157.633 (7.50) Ψm (M ; V, W ) = I m (M 2 , Diag; πL W ⊗ πR (V 0 ⊗ Ω).
L16-end
L17

9. Elliptic operators
EllipticM
Now we turn to the construction of a parametrix for a globally elliptic op-
erator A ∈ Ψm (M ; V, W ). Thus the principal symbol of A has a representative
a ∈ S m (T ∗ M ; π ∗ hom(V, W )) which is elliptic. This means there is an inverse
modulo a compactly supported error,

157.634 (7.51) b ∈ S −m (T ∗ M ; π ∗ hom(W, V )),

ba − Id ∈ C c∞ (T ∗ M ; hom(V )), ab − Id ∈ C c∞ (T ∗ M ; hom(W )).

157.635 Proposition 7.5. An elliptic element A ∈ Ψm (M ; V, W ) has a parametrix

B ∈ Ψ−m (M ; W, V ) modulo smoothing operators,
157.636 (7.52) BA − Id = E ∈ Ψ−∞ (M ; V ), AB − Id = E 0 ∈ Ψ−∞ (M ; W )
and any two such parametrices differ by an element of Ψ−∞ (M ; W, V ).
 9. ELLIPTIC OPERATORS 111

Proof. The surjectivity of the symbol map

157.637 (7.53) Ψ−m (M ; W, V ) / S −m (T ∗ M ; W, V )
157.634
allows us to choose an element B0 ∈ Ψ−m (M ; W, V ) with σ−m (B0 ) = b as in (7.51).
It follows from the multiplicativity and exactness of the symbol sequence that
157.638 (7.54) B0 A − Id = E1 ∈ Ψ−1 (M ; V ) and AB0 − Id = E10 ∈ Ψ−1 (M ; W ).
Now we proceed by induction to find a sequence Bj ∈ Ψ−m−j (M ; W, V ), such that
k
X
( Bj )A − Id = Ek+1 ∈ Ψ−k−1 (M ; V ) and
j=0
157.639 (7.55)
k
X
0
A( Bj ) − Id = Ek+1 ∈ Ψ−k−1 (M ; W ).
j=0

In the inductive step for Bk+1 ∈ Ψ−m−k−1 (M ; W, V ) to satisfy the first condition
157.639
in (7.55) we must have

157.640 (7.56) Bk+1 A − Ek+1 = Ek+2 ∈ Ψ−k−2 (M ; V ) =⇒

σ−m−k−1 = σ−k−1 (Ek+1 )b mod S −m−k−2 (T ∗ ; π ∗ hom(W, V )).
So not only is this possible but the choice is unique up to a addition 157.639
of a term in
Ψ−m−k−2 (M ; W, V ). Applying A on the left in the first equation in (7.55) and on
the right in the second it follows that
0
157.641 (7.57) AEk+1 = Ek+1 A.
So in fact
0 0
157.642 (7.58) aσ−k−1 (Ek+1 ) = aσ−k−1 (Ek+1 ) =⇒ σ−k−1 (Ek+1 )b = σ−k−1 (Ek+1 )b
0 0
modulo terms of order −m − k − 2. Thus in fact ABk+1 − Ek+1 = Ek+2 ∈
−k−2
Ψ (M ; W ) so Bk+1 satisfies both conditions required in the inductive step.
The asymptotic completeness of the pseudodifferential spaces means that we
can choose
X
157.643 (7.59) B∼ Bj ∈ Ψ−m (M ; W, V )
j≥0
157.636
which is then a parametrix in the sense of (7.52).
The uniqueness follows from the existence since if B 0 is any left parametrix
then
157.644 (7.60) B 0 A − Id ∈ Ψ−∞ (M ; V ) =⇒ (B 0 − B)A ∈ Ψ−∞ (M ; V ) =⇒
B 0 − B = (B 0 − B)(AB − E 0 ) = ((B 0 − B)A)B − (B 0 − B)E 0 ∈ Ψ−∞ (M ; W, V ).

The boundedness properties on Sobolev spaces now show that an elliptic oper-
ator of order m is Fredholm as an operator
157.645 (7.61) A : H s (M ; V ) −→ H s−m (M ; W ).
Recall that even acting on distributions
157.648 (7.62) A : C −∞ (M ; V ) −→ C −∞ (M ; W )
 112 7. OPERATORS ON MANIFOLDS

the null space of A is a finite dimensional subspace of C ∞ (M ; V ) since its elements

satisfyb
157.649 (7.63) Au = 0 =⇒ u = Eu, E ∈ Ψ−∞ (M ; V ).
If we choose hermitian inner products on V and W and a smooth density ν on
M then the formal adjoint of A is the unique element
Z Z
∗
157.650 (7.64) m
A ∈ Ψ (M ; W, V ) satisfying hAu, viW ν = hu, A∗ viV ν.
M M
∗
The uniqueness of A follows from this condition.
To see existence, obvserve that the inner products give (antilinear) identifica-
tions of V with V 0 and W with W 0 and define an adjoint isomorphism of bundles
∗ ∗ 0 ∗
157.651 (7.65) πL W ⊗ πR V −→ πL W 0 ⊗ πR
∗
V.
Then the Schwartz kernel is the image of the kernel
∗ ∗ ∗ 0
A = DπR ν, D ∈ I m (M 2 , Diag; πL W ⊗ πR V )
under this isomorphism and reversal of the variables so
A∗ (x, y) = (B(y, x)∗ πR
∗
ν.
157.650
It is straightforward to see that this operator satisfies (7.64).
157.646 Lemma 7.4. The index of an elliptic element A ∈ Ψm (M ; V ; W ) as an operator
157.645
(7.61) is
157.647 (7.66) ind(A) = dim (A) − dim (A∗ )
where A∗ ∈ Ψm (M ; W, V ) is the formal adjoint of A with respect to any choice of
hermitian inner products on the bundles V, W and smooth density on M.
Proof. From the defining property of A∗ it follows that if A is considered as
an operator
157.652 (7.67) A : H m (M ; V ) −→ L2 (M ; W )
then the orthocomplement of the range is (A∗) where
A∗ : L2 (M ; W ) −→ H −m (M, V ).
157.647
This gives (7.66) for the index. Changing the Sobolev order to A : H s (M ; V ) −→
H s−m (M ; W ) does not change the null space and changes the range either by
replacing it by its intersection with H s−m (M ; W ) if s > m or by the closure if
s < m. In either case (A∗ ) remains a complementary space (but not orthogonal)
since the (extension of) the L2 pairing still vanishes. 

157.663 Lemma 7.5. If A ∈ Ellm (M ; V, W ) then there exists A0 ∈ Ψ−∞ (M ; V, W ) such

that
if ind(A) > 0, A + A0 is surjective
157.664 (7.68) if ind(A) = 0, A + A0 is bijective
if ind(A) < 0, A + A0 is injective.
 9. ELLIPTIC OPERATORS 113

157.646
Proof. Suppose ind(A) ≤ 0. By Lemma 7.4 dim (A) ≤ dim (A∗ ). Thus we
can choose a smoothing operator A0 : (A) −→ (A∗ ) which is injective. Then any
element of the null space of A+A0 , satisfies Au = −A0 u. These lie in complementary
spaces so both must vanish and hence u ∈ (A) and A0 u = 0 so u = 0. If ind(A) = 0
it follows that (A∗ ) ⊂ Ran(A + A0 ) so A + A0 is a bijection. If ind(A) ≥ 0 thus
argument applies to A∗ . 
157.655 Lemma 7.6. For any real order m and any bundle V there is an elliptic and
invertible operator Qm ∈ Ψm (M ; V ) with diagonal, positive principal symbol.
Proof. Take a Riemann metric on m and consider the symbol |ξ|m Id valued
in hom(V ). Clearly this is elliptic. For the case m = 0 we can of course take the
identity. Suppose m < 0 then choose an operator Lm/2 ∈ Ψm/2 (M ; V ) with symbol
|ξ|m/2 Id . It is a compact operator on L2 (M ; V ). Taking an hermitian inner product
on V and a positive smooth density on M gives L2 (M 0 V ) an explicity Hilbert inner
product. Then the adjoint L∗m/2 ∈ Ψm/2 (M ; V ) has the same principal symbol
so Q0m = L∗m/2 Lm/2 ∈ Ψm (M ; V ) is selfadjoint with principal symbol |ξ|m Id .
It is elliptic of order 0 with (Q0m ) ⊂ C ∞ (M ; V ) finite dimensional. Choosing a
positive definite matrix on (Q0m ) and adding it to Q0m gives an invertible element
of Ψm (M ; V ) with the same principal symbol. For m > take Q−m . 
The existence of such an element shows that it really suffices to consider oper-
ators of order 0 in the discussion of the index since if A ∈ Ψm (M ; V, W ) is elliptic
then so is A0 = AQ−m ∈ Ψ0 (M ; V, W ) and
157.656 (7.69) σ0 (A0 ) = σm (A)||ξ|−m .
Corresponding to the stability of the index of Fredholm operators on a fixed
Hilbert space it also follows that
157.654 (7.70) ind(A + E) = ind(A) if A ∈ Ψm (M ; V, W ) is elliptic and
0
E ∈ Ψm (M ; V, W ), m0 < m.
Indeed, the index of (A + E)Q−m = AQ−m + EQ−m is the same as that of A + E
and EQ−m is compact from L2 (M ; V ) to L2 (M ; W ) so the index is the same as
that of AQ−m and hence A.
Said a different way, we have proved that
157.657 Lemma 7.7. The index of and elliptic element A ∈ Ψm (M ; V, W ) is determined
by any reprentative of
157.658 (7.71) σm (A)|ξ|−m ∈ S 0 (T ∗ M ; π ∗ hom(V, W )).
and so defines a map
157.659 (7.72) ind : Ell(V, W ) = {a : S 0 (T ∗ M ; π ∗ hom(V, W ));
with an inverse modulo lower order terms} −→ Z
for any bundles V and W.
Of course the bundles certainly have to have the same rank before such an elliptic
symbol exists.
We can 157.659
think of the pseudodifferential algebra (modules really) as define the
index map (7.72) through ‘quantization of the symbol’.
 114 7. OPERATORS ON MANIFOLDS

0
157.660 Lemma 7.8. If A1 ∈ Ψm (M ; V, W ) and A2 ∈ Ψm (M ; W, U ) are two elliptic
operators between bundles then the composite is elliptic and
157.653 (7.73) ind(A2 ◦ A1 ) = ind(A2 ) + ind(A1 ).
157.655
Proof. Using Lemma 7.6 We can replace A1 by AQ −m and A2 by Q−m0 A2
157.646
and reduce to the case that m = m0 = 0. From Lemma 7.4 we see that ind(A∗ ) =
− ind(A). So if A1 and A2 have opposite signs we can assume that ind(A1 ) ≥ 0
and ind(A2 ) ≤ 0 or else pass to the adjoint, which reverses the order. If A1 and
A2 have the same sign then we can assume that both are non-negative. So it
suffices to consider the case ind(A1 ) ≥ 0. Then, adding a finite rank surjective
smoothing operator mapping the null space of A onto a complement to its range
we can assume that A1 is surjective. Then the range of A2 ◦ A1 is the range of
A2 and its null space is inverse image of the null space of A2 under A1 , which has
157.653
dimension dim((A1 )) + dim((A2 ). Thus (7.73) holds. 
157.659
157.661 Corollary 4. The index map (7.72) is additive under the product
157.662 (7.74) Ell(W, U ) ◦ Ell(V, W ) −→ Ell(V, U )
(when these spaces are non-empty).

10. Self-adjoint operators

Although I will concentrate on the index theorem in these lectures it is even
more important to discuss the spectral theory of self-adjoint ellipitc operators on a
compact manifold. Indeed I am assuming that you have seen this before.
P1.16 Theorem 7.2. An elliptic pseudodifferential operator acting on sections of a
vector bundle V over a compact manifold without boundary, M , A ∈ Ψm (M ; V ),
which is symmetric, with respect to a density on M and a fibrewise Hermitian inner
product on V, is self-adjoint and, if m 6= 0, has eigenfunctions forming a complete
orthonormal basis of L2 (M ; V ).
Proof. Symmetry here means ‘formal self-adjointness’
P1.17 (7.75) Z Z
hAv, wiL2 = hAv, wiV ν = hAv, wiV ν = hv, AwiL2 ∀ v, w ∈ C ∞ (M ; V ).
M M

If m ≤ 0 it follows from the L2 boundedness of A that this identity extends by

continuity to all v, w ∈ L2 (M ; V ) and so A is a bounded self-adjoint operator.
If m > 0 then we know that A : H m (M ; V ) −→ L2 (M ; V ) is continuous. Then
A is a closed unbounded operator with domain H m (M ; V ) ⊂ L2 (M ; V ) and elliptic
regularity shows us that the adjoint of this operator, which by definition has domain
P1.18 (7.76)
Dom(A∗ ) = {w ∈ L2 (M ; V ); C ∞ (M ; V ) 3 u −→ hAu, wiL2 is bounded on L2 }
is A again as an unbounded operator.
If m < 0 then A is a compact operator and the theory of compact self-adjoint
operators applies to give the complete orthonormal basis. If m > 0 we know that A
has a unique generalized inverse determined by the L2 structure and it follows that
B ∈ Ψ−m (M ; V ) is also self-adjoint. The non-zero eigenvalues of A are the inverses
of the non-zero eigenvalues of B and the null spaces are, essentiallt by definition,
the same. 
 11. FIBRE BUNDLES 115

One of the things that I set out to do was to prove that such an elliptic operator
has ‘Weyl asymptotics’. L17-end

11. Fibre bundles

L18
The definition of homotopy of elliptic pseudodifferential operators involves a
‘family’ in the sense of a smooth map [0, 1] 3 t 7−→ A(t) ∈ Ψm (M ; V, W ). However
we really want to allow the bundles V and W to ‘vary’ as well. In making this
precise we may as well pass to the notion of a family of pseudodifferential operators
on the fibres of a fibre bundle.
157.665 Definition 7.3. A fibre bundle is a smooth surjective map between manifolds
φ : X −→ Y which is locally a product in the sense that there is a manifold
Z (the model fibre) such that each y ∈ Y has a neighbourhood U ⊂ Y with a
diffeomorphism giving a commutative diagramme

157.667 (7.77) φ−1 (U )

F / Z ×U

φ πU
# |
U.
So this is just like a vector bundle except the fibres are manifolds, not vector
spaces, and correspondingly the ‘local trivializations’ F are fibre-preserving diffeo-
morphisms. Of course the most obvious case is a product
157.668 (7.78) X =Z ×Y
which is globally trivial as a fibre bundle. You are probably familiar with the Hopf
fibration
157.669 (7.79) S2n+1 −→ S2n
given by thinking of S2n+1 as the unit sphere in Cn+1 and then taking the fibres to
be given by the multiplicative action of the circle S ⊂ C.
The ‘fibre above y ∈ Y ’, meaning φ−1 (Y ), is diffeomorphic to the model fibre
Z and is often written Zy . There is in general no natural choice of the trivialization
157.667
map F in (7.77) so no natural diffeomorphism between Z and Zy . A fibration is
how we are to interpret the notion of a ‘smooth family of manifolds’.
157.670 Proposition 7.6. If φ : X −→ Y is a surjective smooth map between compact
manifolds then it defines a fibre bundle if and only if it is a submersion, i.e. F∗ :
Tx X −→ TF (x) Y is surjective for each x ∈ X.
The compactness, or at least some additional condition, is needed here. The sur-
jectivity is less of an issue, since if φ is a submersion between compact manifolds its
image if open and closed and hence is some union of components of Y – assuming
we are not requiring manifolds to be connected!
Proof. This is a form157.667
of the Implicit Function Theorem. For a submersion,
there is a local version of (7.77) near each point of X. Namely, the surjectivity of
the differential φ∗ is equivalent to the injectivity of φ∗ : TF (p) Y −→ Tp∗ X. So if
yi are local coordinates in U ⊂ Y near F (p) then the functions φ∗ yi defined on
φ−1 (U ) have independent differentials. So they can be completed to a coordinate
system, φ∗ yi , zk on X in some neighbourhood of p. Then the fibres localy are the
surfaces φ∗ yi =const. This needs to be globalized along φ−1 (φ(p)).
 116 7. OPERATORS ON MANIFOLDS

In each of the coordinate patches constructed along the ‘base fibre’ Z(p) =
φ−1 (φ(p)), the fibre containing p, the choice of local coordinates gives vector fields,
(k) (k)
just the Vi = ∂yi in the local coordinates label by k, which satisfy φ∗ (Vi ) =
(k)
∂yi are the coordinate vector fields in the base. The Vi in different coordinate
patches will be different but if we take a partition of unity rk on X near Z(p)
(k)
and subordinate to the cover then the Vi = ρk Vi satisfy φ∗ (Vi ) = ∂yi . The
compactness of Z(p) means that this is only a finite sum and so the Vi are defined in
an open set φ−1 (U ) for some neighbourhood p ∈ U in Y. The Vi need not commute
but we can simply integrate successively with respect them starting at Z(p). So
for each q ∈ Z(p) integrating V1 gives a smooth surface in Y containing Z(p) and
projecting under φ onto the y1 access in U. Then integrating from this surface gives
a submanifold of one dimension higher projecting onto the y1 , y2 coordinate plane
in U, where this may need to be shrunk at each step. After dim Y steps we have
a smooth map from a neighbourhood of the form φ−1 (U 0 ) of Z(p) in Y to Z(p)
which
157.667
together with the projection to U 0 gives a diffeomorphism F as required in
(7.77). 
157.672 Proposition 7.7. If V −→ X is a vector bundle over the157.667 total space of a fibre
bundle then there are local trivializations of the bundle as in (7.77) over which there
are bundle isomorphisms
157.673 (7.80) V
τ / VZ × U
φ−1 (U )

 
φ−1 (U )
F / Z ×U

φ πU
# {
U
where VZ is a fixed bundle over the model fibre.
Notice that there is some constructive confusion between the total space of a vector
bundle and the vector bundle itself (which includes the projection to the base).
Proof. 
Now, the kernel of a pseudodifferential operator on Z acting from sections of
one bundle to another is a distribution on Z 2 . So we need to consider the product
of each fibre with itself.
157.671 Proposition 7.8. If φ : X −→ Y is a fibre bundle then X 2 3 (x, x0 ) −→
(φ(x), φ(x0 )) ∈ Y 2 is a fibre bundle as is the fibre product, which is the closed
embedded submanifold
157.674 (7.81) X [2] = {(x, x0 ) ∈ X 2 ; φ(x) = φ(x0 )}
and the diagonal is the closed embedded submanifold
157.675 (7.82) Diag = {(x, x) ∈ X 2 ; x ∈ X}.
Clearly the diagonal is diffeomorphic to X.
Proof. 
 12. FAMILIES OF PSEUDODIFFERENTIAL OPERATORS 117

12. Families of pseudodifferential operators

To define the space Ψm
φ (X; V, W ) of operators

157.676 (7.83) A : C ∞ (X; V ) −→ C ∞ (X; W )

for two vector bundles over the total space of a fibre bundle we have two clear
157.673
options. Use Proposition 7.80 locally in the base and then patch things together or
start globally and see what happens. As you have seen I prefer to take the latter
course, at least for the moment!
157.677 Definition 7.4. The space of fibrewise pseudodifferential operators for a fibre
bundle φ : X −→ Y between sections of bundles V and W over X is identified with
the space of conormal distributions
1
∗ ∗
157.678 (7.84) Ψm
φ (X; V, W ) = I
m− 4 dim Y
(X [2] , Diag; πL W ⊗ πR (V 0 ⊗ Ωfib ).

Here πL , πR : X [2] −→ Y are the restrictions of the two projections from X 2 and
Ω is the bundle over Y of densities on the fibres of φ.
A more standard notion for this space would be Ψm (X/Y ; V, W ) where there
is no space X/Y but it is supposed to suggest a family acting on the fibres of X. I
will not use that notation because it does not really fit with various generalizations
below.
So, we need to see that this definition does give a linear space of operators as
157.676
in (7.83) and then that they form an algebra. Again we can do this by localization
or we can think globally. What we want to do is make sense of the formula
Z
157.679 (7.85) Av(x) = A(x, x0 )v(x0 ), v ∈ C ∞ (X; V ).
fib
∗
We can certainly pull v back to X [2] under πR to give the a section of πR V over
[2]
X which is independent of x – constant on the fibres of πL . Then the product
with the kernel makes sense and using the pairing of V 0 and V gives a section
∗ ∗
of πL W ⊗ πR Ωfib . This is singular only at the diagonal and the projection πL is
transversal to the diagonal. So, we can see that the integral is well-defined and
gives a smooth section of W over X.
The tangent bundle of X has a subbundle, Tfib X ,→ T X consisting of the
vectors tangent to the fibe at each point. For each y ∈ Y the restriction if this fibre
tangent bundle is therefore just the tangent bundle of the fibre
157.683 (7.86) Tφ X Zy
= T Zy .

The dual bundle to Tφ X which I will denote Tφ∗ X is then the collection of the fibre
cotangent bundles
157.684 (7.87) Tφ∗ X Zy
= T ∗ Zy .
157.678
The adjustment in the order in (7.84) is so that:
157.680 Proposition 7.9. There is a well-defined principal symbol map giving a short
exact sequence
∗ σ ∗
157.681 (7.88) Ψm−1
φ (X; V, W ) ,→ Ψm m m
φ (X; V, W ) −→ S (Tφ X; π hom(V, W ))
 118 7. OPERATORS ON MANIFOLDS

consistent with the single operator case in the sense that for each y ∈ Y there is a
restriction map
157.682 (7.89) Ψm m
φ (X; V, W ) 3 A 7−→ A(y) ∈ Ψφ (Zy ; Vy , Wy ), σm (A(y)) = σm (A) T ∗ Zy
.

Returning to the discussion of the index of an elliptic operator observe that

157.685 Proposition 7.10. The ‘numerical’ index of an elliptic family of pseudodiffer-
ential operators ind(A(y)) is constant.
For vector bundles over a manifold, let us denote by iso(V, W ) ⊂ hom(V, W )
bundle with fibre the invertible linear maps between fibres at each point. For two
pairs of bundles Vi , Wi i = 1, 2 there is a direct sum operation
157.688 (7.90) iso(V1 , W1 ) × iso(V2 , W2 ) −→ iso(V1 ⊕ V2 , W1 ⊕ W2 ).
157.686 Proposition 7.11. Any elliptic operator is homotopic to a classical elliptic
operator and hence for any compact manifold and vector bundle the index induces
a map
157.687 (7.91) ind : C ∞ (S ∗ M ; π ∗ (iso(V, W )) −→ Z
157.688
which is additive under direct sums as in (7.90) and under products giving commu-
tative diagrammes

157.689 (7.92) C ∞ (S ∗ M ; π ∗ iso(V1 , W1 )) × C ∞ (S ∗ M ; π ∗ iso(V2 , W2 )) / Z×Z

ind × ind

⊕ +
 
C ∞ (S ∗ M ; π ∗ iso(V1 ⊕ V2 , W1 ⊕ W2 ))
ind / Z.

(7.93) C ∞ (S ∗ M ; W, U ) × C ∞ (S ∗ M ; V, W ) / Z×Z
ind × ind

◦ +
 
C ∞ (S ∗ M ; V, U )
ind / Z.

Now, we need to relate this to K-theory. Consider all of the triple (V, W, a)
where V and W are complex vector bundles over M and a is an isomorphism
between them , which we can write as a ∈ C ∞ (S ∗ M ; π ∗ iso(V, W )). We impose
three relations on this biggish set of data.
157.690 (7.94)
Bundle isomorphism invariance: (V1 , W1 , a) ' (V2 , W2 , bae−1 )
if b ∈ C ∞ (M ; W1 , W2 ), e ∈ C ∞ (M ; V1 , V2 )
157.692 (7.95)
Homotopy invariance: (V1 , W1 , a1 ) ' (V0 , W0 , a0 ) if ∃
h ∈ C ∞ (S ∗ M ×[0, 1]; iso(V, W )) where V, W are bundles over M × [0, 1]
with V M ×{0}
= V0 , V M ×{1}
= V1 , W M ×{0}
= W0 , W M ×{1}
= W1
157.691 (7.96) Stability: (V ⊕ U, W ⊕ U, a ⊕ IdU ) ' (V, W, a) for any bundle U −→ M.
Each of these is an equivalence relation – for the first and last this is straightfor-
ward, for homotopy equivalence one needs to do a littl work in concatenating two
homotopies to get a smooth one. The trick is to replace an initial homotopy by its
 13. FAMILIES INDEX THEOREM 119

pull back under a smooth bijective map [0, 1] −→ [0.1] which is constant to infinite
order at both ends. Then following one of these with another gives smooth objects
over [0, 2] which can be reparameterized to a homotopy.

157.693 Theorem 7.3. ∗ The collection of the C ∞ (S ∗ M

157.690 157.692
; π ∗ iso(V, W )) subject to the
157.691
equivalence relation combining
157.688
(7.94), (7.95) and (7.96) is an abelin group under
the additive operation (7.90) which is naturally isomorphic to K 0 (T ∗ M ) and to
which the (numerical) index map descends giving the push-forward map
π
157.694 (7.97) K 0 (T ∗ M ) −→
!
K 0 ({pt}) = Z.
Most of proof later. The identification of K 0 (T ∗ M ) with the equivalence
classes of the data in C ∞ (S ∗ M ; π ∗ iso(V, W )) uses the results in Chapter ?? and
is discussed below, as is the definition of the ‘wrong-way map’ π! . The existence of
this map, uses the notion of the K-orientation of T ∗ M. However, the fact that157.691
157.690 157.692
the
index map does descend through the equivalence relations (7.94), (7.95) and (7.96)
follows directly from the preceding discussion. 
So my basic claim is that it is ‘better’ to prove the more general families index
theorem. In the more general families case the plan is to pass through semiclassical
quantization since this has better functorial properties (whilst being equivalent to
the pseudodifferential qunatization we have been employing).

13. Families index theorem

The ‘numerical index’ that we have been discussing exactly characterizes the
invertibility properties of elliptic operators, up to smoothing perturbation as seen in
157.663
Lemma 7.5. For an elliptic family we can ask the stronger question:- When is there
a perturbation, A0 ∈ Ψ−∞ m
φ (X; V, W ) for an elliptic family A ∈ Ψφ (X; V, W ) such
that A + A0 is invertible? Clearly the vanishing of the numerical index is necessary,
and this implies the existence of such a perturbation locally near each point in the
base, but it is not sufficient to imply the existence of a global perturbation working
at every point. The obstruction is precisely the families index which, in line with
the discussion above, we will find as a map
157.695 (7.98) indφ : C ∞ (Sφ∗ X; π ∗ iso(V, W )) −→ K 0 (Y ).
To construct such a map we look for a ‘good’ pertubation and parametrix.
157.696 Proposition 7.12. If A ∈ Ψm
φ (X; V, W ) is an elliptic family then there exists
A ∈ Ψφ (X; V, W ) and a parametrix B ∈ Ψ−m
0 −∞
φ (X; W, V ) such that

157.697 (7.99) B(A + A0 ) = Id −p1 , (A + A0 )B = Id −p2

where p1 ∈ Ψ−∞ −∞
φ (X; V ) and p2 ∈ Ψφ (X; W ) are finite rank projections (idempo-
tents).
We show below that the formal difference p1 p2 defines an element
157.695
of K 0 (Y )
which is defined independent of choices and this gives the map (7.98). In fact this
defines a map
157.698 (7.100) indφ = (φπ)! : K 0 (Tφ∗ X) −→ K 0 (Y )
the identification if which is the Atiyah-Singer theorem in K-theory.
Proof. 
L18-end
L19
 120 7. OPERATORS ON MANIFOLDS

14. Spin and Dirac

Most of the geometric examples of elliptic operators, and families of them,
are first order differential operators. Perhaps the most basic is the ‘Hodge-Dirac’
operator.
If M is a compact manifold (it need not be oriented) the form bundles Λk M
are well-defined real vector bundles over M but we will allow complex coefficients
as well. Then the deRham operator is a well-defined first order differential operator
157.702 (7.101) d : C ∞ (M ; Λk ) −→ C ∞ (M ; Λk+1 ) d2 = 0.
By convention, for k < 0 or k > dim M, the form bundles are trivial, just the zero
vector space at each point. The deRham differential is not an elliptic operator,
indeed it symbol is
157.703 (7.102) σ1 (d)(x, ξ) = iξ∧ ∈ S 1 /S 0 (T ∗ M ; hom(Λk , Λk+1 ), (x, ξ) ∈ T ∗ M.
If we choose a Riemann metric on M we get, by definition, a fibrewise Euclidean
inner product on T M and hence the dual inner product on T ∗ M and by standard
constructions inner products on all the Λk M. Moreover we also get a well-defined
Riemannin density (not a volume form unless M is oriented) which I will write as
dg. This data determines a formal adjoint of d,
157.704 (7.103) δ = d∗ : C ∞ (M ; Λk+1 ) −→ C ∞ (M ; Λk ) ∀ k, δ 2 = 0.
Then the Hodge-Dirac operator is
157.705 (7.104) ð = d + δ ∈ Ell1 (M ; Λ∗ ) σ1 (ð)(x, ξ) = i(ξ ∧ −ι(ξ))
where ι(ξ) is contraction with the metrically dual tangent vector to the cotangent
vector ξ.
157.706 Theorem 7.4 (Hodge). If M is compact without boundary, the null space of ð
is naturally isomorphic to the deRham cohomology of M.
Proof. The operator ð is elliptic since
157.707 (7.105) σ1 (ð)2 = ξ ∧ ι(ξ) + ι(ξ)ξ∧ = |ξ|2 Id .
Thus ð2 = ∆ (the Laplacian, or Laplace-Beltrami operator) has scalar principal
symbol which is invertible where ξ 6= 0. As an elliptic, formally self-adjoint operator
we know that it is self-adjoint and the range has an orthogonal decomposition which
restricts to smooth sections to
157.708 (7.106) C ∞ (M ; Λ∗ ) = null(ð) ⊕ ð (C ∞ (M ; Λ∗ )) .
This is the Hodge decomposition.
Note that if h ∈ null(ð) ⊂ C ∞ (M ; Λ∗ ) then
Z Z Z
157.711 (7.107) hdh, dhi = hdh, −δhi = hd2 h, hi = 0
M M M
so dh = δh = 0.157.708
Applying (7.106) to a closed form u ∈ C ∞ (M ; Λ∗ ) gives
157.709 (7.108) u = h ⊕ (dv + δv), v ∈ C ∞ (M ; Λ∗ ), dh = δh = 0.
This defines a surjective map
157.710 (7.109) {u ∈ C ∞ (M ; Λ∗ ); du = 0} 3 u 7−→ h ∈ null(ð).
since h is mapped to itself.
 14. SPIN AND DIRAC 121

157.709
Taking the inner product of (7.108) with δv shows that δv = 0 (since u is
closed) and hence that
157.712 (7.110) u = h + dv.
If157.710
v = dw then taking the inner product with u shows that u is in the null space of
157.710
(7.109). Conversely if u is in the null space u = dv. Thus the map (7.109) descends
to an isomorphism
∗
157.713 (7.111) HdR (M ; R) −→ null(ð).

157.714 Corollary 5. The restriction of the Hodge-Dirac operator to
ð+ : C ∞ (M Λev ) −→ C ∞ (M ; Λodd )
has index the Euler characteristic of M
Xn
157.715 (7.112) ind(ð+ ) = (−1)k dim HdR
k
(M ).
k=0
L19-end
On a Riemann manifold a Clifford module is a complex vector bundle W −→ M
with Hermitian metric together with a norm-preserving bundle map
157.716 (7.113) cl : T ∗ M −→ hom(W )
satisfying the condition
157.717 (7.114) cl(ξ) cl(η) + cl(η) cl(ξ) = 2hξ, ηi Id in hom(W ).
157.718 Lemma 7.9. A Clifford module has a unitary connection ∇ satisfying
157.719 (7.115) ∇v cl(ξ)w = cl(∇c i)w + cl(ξ)∇v w,
∀ v ∈ C ∞ (M ; T M ), ξ ∈ C ∞ (M ; T ∗ M ), w ∈ C ∞ (M ; W )
where ∇v ξ is the action of the Levi-Civita connection on the cotangent bundle.
Proof. As usual the idea is to construct such a connection locally and then
patch using a partition of unity. 
Interpreting the connection as a differential operator ∇ : C ∞ (M ; W ) −→
C (M ; T ∗ ⊗ W ) and the Clifford action as a contraction map cl : T ∗ W ⊗ W −→ W
∞

the associated Dirac operator is a well-defined differential operator

157.720 (7.116) ð = cl ◦∇ : C ∞ (M ; W ) −→ C ∞ (M ; W ).
So, how do such Clifford modules arise and what is the index of the associated
Dirac operator? Let me restrict to the case that the dimension of M is even, 2k
– the odd-dimensional case is slightly different. Give a Riemann metric on M
the complexified Clifford algebra at each point, defined as the quotient of the full
∗
complexified
157.717
tensor algebra of Tm M by the ideal generated by the elements as in
(7.114), i.e.
157.721 (7.117) ξ ⊗ η + η ⊗ ξ − hξ, ηi Id
forms the fibre of a bundle
157.722 (7.118) Cl(M ) −→ M.
This is an Azumaya algebra – not only is it a complex vector bundle by the fibres
are algebras and these algebras are isomorphic to the 2k × 2k matrix algebra. A
 122 7. OPERATORS ON MANIFOLDS

local orthonormal basis gives a trivialization of the bundle consistent with these
identifications.
A Clifford module is then a vector bundle with a multiplicative bundle map
157.723 (7.119) Cl(M ) −→ hom(W ).
Such modules arise from a Spin structure or a Spin-C structure on M. Both of
these only make sense on an oriented manifold. They have to do with the groups
157.724 (7.120) Spin −C(2k) −→ Spin(2k) −→ SO(2k).
Here (for k > 1) the spin group is the universal cover of SO(2k). Since π1 (SO(2k) =
Z2 this is a double cover. One way to construct concretely is using the Clifford
algebra on R2k . An element of SO(2k) acts on the Clifford algebra157.721through its
action on T0∗ R2k = R2k which preserves the ideal generated by (7.117). Such
an algebra-preserving isomorphism of a matrix group is necessarily generated by
conjugation so for O ∈ SO(2k) there is an element LO ∈ Cl(R2k ) such that the
action of O is
157.725 (7.121) cl(Oξ) = LO cl(ξ)L−1
O .

Such an element can be constructed by factorizing O into products of reflections.

Then LO is the product of the clifford actions of the unit normal vectors to the
fixed set of the reflection and is determined by up to sign. These LO form the Spin
group. It has a reflection action, given by the kernel of Spin(2k) −→ SO(2k) and
the Spin-C group is
157.726 (7.122) Spin −C(2k) = Spin(2k) ×Z2 S.

15. Gerbes?
I do not expect to have time to discuss this in lectures. The idea is that this give
a rather systematic approach to Spin and Spin-C structures on manifolds. Since it
is relevant, at least as background, let me briefly recall the classification of real and
complex line bundles over a manifold – which is generalized by gerbes.
A real line bundle over a manifold M can be given a smooth family of fibre
metrics and so reduced to a principal Z2 bundle the ‘sphere’ in the line at each
point

157.728 (7.123) Z2 M̂


M.
Thus, M̂ is a double cover of M with the action of Z2 being to interchage the points
in each fibre.
Such a principal bundle is classified by H 1 (M ; Z2 ). This is most easily seen in
terms of Čech cohomology. To be brief about this, any open cover of a manifold
has a refinement to a ‘good’ open cover – one in which all the open sets and all
non-trivial finite intersections of them are contractible. A covering by small (with
radius below the injectivity radius) Riemannian balls satifies this.
So, one can find such a good open cover over each element of which, Ui , the
Z2 bundle has a section. Then over each intersection Uij = Ui ∩ Uj the relation
 15. GERBES? 123

between these sections gives a map

157.729 (7.124) χij : Uij −→ Z2 .
This is a Čech cocycle, since over triple intersections
157.730 (7.125) χij χjk χki = 1.
It follows (this is Čech theory) that this determines a cohomology classe
157.731 (7.126) χ ∈ Ȟ(M ; Z2 ).
The vanishing of this cohomology class is equivalent to the exactness of cocylcle
(because the cover is good), meaning the existence of smooth (i.e. continuous)
maps
157.732 (7.127) ηi : Ui −→ Z2 s.t. χij = ði ηj−1 on Uij .
Such a collection of map allows the original sections to be ‘corrected’ to a global
section – implying the Z2 cover, and hence line bundle, is trivial. Conversely given
a cocycle χij one can construct a Z2 -principal bundle from it which recover χ.
Similarly for complex line bundles over M are classified by Ȟ 2 (M ; Z). Again
one can choose a metric on the line bundle and so reduce it to a principal-S bundle
157.733 (7.128) S S


M.
Now the local trivialization over a good open cover give a map
157.734 (7.129) cij : Uij −→ S s.t. cij cjk cki = 1 over Ui ∩ Uj ∩ Uk .
This cocycle yields a cohomology class, the Chern class, c ∈ Ȟ 1 (M ; S) which 157.733
is the
same as Ȟ 2 (M ; Z). The triviality of c implies the existence of a section of (7.128)
and conversely. Moreover a prinicpal circle bundle can be constructed from a class
c ∈ Ȟ 2 (M ; Z) with this as Chern class.
The relationship between real line bundles, and their complexification (obtained
by tensoring with C) is the Bocksteim homomorphism
157.735 (7.130) H 1 (M ; Z2 ) −→ H 2 (M ; Z).
Gerbes, in particular bundle gerbes, are the next step up from line bundles. I
include a brief discussion of ‘lifting bundle gerbes’, specifically for spin and spinC
structures. L20
 CHAPTER 8

Semiclassical quantization

1. Blow up
11.4.2022.1 Definition 8.1. If S ⊂ M is a closed embedded submanifold a blow-up of M
along S, also called the blow-up (actually the radial blow-up) of S, is a manifold
with boundary [M ; S] and smooth surjective map β : [M ; S] −→ M (the blow-down
map) with the properties
(1) β : [M ; S] \ ∂[M ; S] −→ M \ S is a diffeomorphism
(2) β : ∂[M ; S] −→ S is a sphere bundle
(3) If 0 ≤ q ∈ C ∞ (M ) vanishes precisely at S and exactly to second order then
β ∗ q = x2 is the square of a boundary defining function x ∈ C ∞ ([M ; S])
(4) The Lie algebra of smooth vector fields on M which are tangent to S lift
to span, over C ∞ ([M ; S]), the Lie algebra of vector fields tangent to the
boundary.
The ‘precise second order vanishing’ is the statement that the Hessian of q at
each point s ∈ S is postive definite as a quadratic form vi vj q on Ns S, the normal
bundle to S.
If you recall the notion of compactification from early in the course you will see
some similarity. However blow-up is much more functorial.
Why blow up submanigolds? There are several reasons (but it isn’t always a
good idea!).
11.4.2022.2 Theorem 8.1. Any closed embedded submanifold has a blow-up and any two
are naturally diffeomorphic.
Recall that the normal bundle to S, N S = TS M/T S parameterizes the vector
fields at S ‘pointing into N ’. As we shall see the boundary of the blow-up is the
corresponding sphere bundle SN S = (N S \ 0S )/R+ . So the definition of [M ; S]
involves gluing this onto M \ S to get a manifold with boundary.
Proof. The existence will use the collar neighbourhood theorem, discussed in
Sect.collar
§ 3.
We start with a simple case, namely M = W is a real vector space and S = {0}
is the origin. We can get a quadratic function q = |w|2 as the square of the length
for some Eucludean norm on W. Then we ‘introduce polar coordinates’. These are
‘singular coordinates’ (whatever that means) but correspond to a smooth map
11.4.2022.3 (8.1) β : [0, ∞) × (W \ 0)/R+ −→ [0, ∞) × {w ∈ W ; |w| = 1} 3 (r, ŵ) 7−→ rŵ ∈ W.
Here the first map is a diffeomorphism where the metric is used to identify the
quotient with the unit sphere. The inverse of β, restricted to W \ 0, is
11.4.2022.4 (8.2) w 7−→ (|w|, [w]) ∈ [0, ∞) × (W \ 0/R+ ).
125
 126 8. SEMICLASSICAL QUANTIZATION

So we conclude that under change of norm from |w| to |w|0 the identity map on
W \ 0 lifts to the smooth diffeomorphism
|w|
11.4.2022.5 (8.3) β : [0, ∞) × (W \ 0/R+ ) 3 (x, [w]) 7−→ ( x, [w])
|w|0
which is the identity on the boundary. So the construction does give the set
11.4.2022.6 (8.4) (W \ 0/R+ ) t (W \ 0)
a unique topology and C ∞ structure.
Now, we should check that this has the desired properties. The first two con-
ditions are clear since β is clearly a diffeomorphism of (0, ∞) × (W \ 0)/R+ onto
W \ {0} and the boundary11.4.2022.5
is a sphere. The quadratic function corresponding to
the metric used to define (8.3) is |w|2 so the third condition holds for this metric
and any other quadratic function is the sum of some other Euclidean metric (its
Hessian) and a function vanising to third order at the boundary. The first part is
the product of |w|2 with a smooth function on the sphere so it is the square of a
defining function for the boundary and the higher order term is smooth (since β is
smooth and vanishes to third order at the boundary.
So, it remains to check the fourth property. The vector fields tangent to 0 are
those which vanish there and so the are of the form, in any linear coordinates,
X
157.736 (8.5) aij wi ∂wj , aij ∈ C ∞ (W ).
ij

The coefficient pull back to be smooth, so to show that these lift, i.e. extend
smoothly from x > 0 down to x = 0, it suffices to show this for the wi ∂wj .
These are homogeneous of degree 0 on W and under β radial scaling becomes
(x, [w]) −→ (tx, [w]), t > 0. Thus writtend in terms of the product decomposition
11.4.2022.3
(8.1)
157.737 (8.6) wi ∂wj = a([w])x∂x + V
where V is a smooth vector field and a is a smooth function on the sphere. Both
extend smoothly down to x = 0. The radial vector field w∂157.737
w lifts to x∂x and the
wi ∂wj span all vector fields on W away from 0 so the V in (8.6) span all the vector
fields on the sphere.
This completes the proof for the blow-up of 0 ∈ W. It is clear that the linear
map reversing one coordinate lifts to be smooth as the reflection in the sphere.
Then the smoothness of the lifts of the linear vector fields wi ∂wj , which form the
Lie algebra of GL(W ) shows that the action of this group on W \ {0} extends to
[W, {0}]. It follows that the action
157.738 (8.7) GL(W ) × [W, {0}] −→ [W, {0}]
is smooth.
Now we pass to the case of the zero section of a real vector bundle, U −→ M.
We define this as the union over the base of the blow-up of the fibres just defined,
with blow-down maps
G β
157.739 (8.8) [U, 0U ] = [Um , {0}] −→ U.
m∈M

gives β −1 (O) a product trivial-

A local trivialization over an open subset O ⊂ M157.738
ization as a fibre bundle and the smoothness of (8.7) means this patches smoothly
 2. SEMICLASSICAL SMOOTHING OPERATORS 127

157.739
to make (8.8) into a smooth fibre bundle with smooth map. Similarly the first
and third conditions follow directly. The vector fields tangent to 0U are given in a
spanned in a local trivialization by the
157.740 (8.9) ∂zi , wi ∂wj
where the zi are coodinates on M. All these vector fields have smooth lifts in the
product decompositon of [U ; 0U ] and so globally and clearly span the vector fields
tangent to the boundary.
Finally then we pass to the general case but this follows from the collar neigh-
bourhood theorem (or working locally if you prefer) which gives a diffeomorphism χ
from a neighbourhood Q0 of an embedded submanifold S ⊂ M to a neighbourhood
Q of the zero section of its normal bundle. This gives a C ∞ stucture to
157.741 (8.10) [M ; S] = (M \ S) ∪ SN S
as a manifold with boundary and by the lifting property it is independent of the
choice of χ. 
The discussion above is for a closed embedded submanifold of a manifold with-
out boundary, of course it applies unchanged if M has a boundary, including cor-
ners, but S does not meet the boundary. If S does meet the boundary we need to
specify the meaning of ‘embedded’. What is needed for the existence of a collar
neighbourhood theorem is the following condition.
157.742 Definition 8.2. A subset S ⊂ M of a manifold with corners is a p-submanifold
(the ‘p-’ being for ‘product’) if at each point of S of codimension k there are
‘adapted’ local coordinates
157.743 (8.11) x1 , . . . , xk , y1 , . . . , yn−k in O ⊂ M
where the xi ≥ 0 are local boundary defining functions and
157.744 (8.12) S ∩ O = {xl+i = 0, i = 1 . . . , k − l, yj+p = 0, p = 1, . . . , n − k − j}.
Here S has codimension n − l − j.
157.745 Theorem 8.2. Any closed p-submanifold of a manifold with corners has a
blow-up which is unique up to natural diffeomorphism.
Proof. Maybe it is best to prove the collar neighbourhood theorem in this
context first! 

2. Semiclassical smoothing operators

As usual we have two ways of approaching the definition of semiclassical smooth-
ing operators on a manifold, and more generally on the fibres of a fibration. We
can either use the original definition on Rn and localize or proceed globally. As you
can already see, I favour the latter appropach.
First recall the case of Euclidean space. We defined the semiclassical smoothing
operators in terms of Schwartz functions A ∈ C ∞ ([0, 1]; S (R2n )), depending on a
paramater, with the kernel being
x − y −n
157.699 (8.13) A(, x, ) dy

where I have included the measure since we know now that we should!
One reason for introducing blow-up above is that we can understand the kernels
directly as smooth sections of a bundle.
 128 8. SEMICLASSICAL QUANTIZATION

157.701 Proposition 8.1. The space of semiclassical smoothing operators on Rn have

kernels which are smooth sections of a rescaled density bundle over
157.700 (8.14) [Rn × Rn × [0, 1]; Diag ×{0}]
L20-end
L21
Proof. 

3. Pull-back and push-forward

This should really have come earlier in the course, but there is some virtue in
leaving things until you need them, even if ‘just in time’ has its drawbacks!
Essentially the definition of smoothness of a map F : M −→ N between mani-
folds means that the pull-back map is defined, linear and continuous on functions
157.748 (8.15) F ∗ : C ∞ (N ) −→ C ∞ (M ).
We can add in a vector bundle on the image space to get
157.746 (8.16) F ∗ : C ∞ (N ; W ) −→ C ∞ (M ; F ∗ W )
Indeed, the pulled-back bundle, having fibes
157.747 (8.17) (F ∗ W )m = WF (m)
over an open set O ⊂ N induces
is defined so that this is true. A trivialization of W157.748
∗ −1
a trivializaation
157.746
of F W over F (O) and then ( 8.15) applies to the coefficients to
give (8.16). 157.746
The map (8.16) has a ‘formal transpose’
157.749 (8.18) F∗ : C c−∞ (M ; W 0 ⊗ Ω) −→ C c−∞ (N ; W 0 ⊗ Ω)
since these are the dual spaces – so
157.750 (8.19) (F∗ w)(φ) = w(F ∗ φ), φ ∈ C ∞ (N ; W ), w ∈ C c−∞ (M ; W 0 ⊗ Ω).
We need compactness of the support of w (really a bit less) to make sure that the
pairing on the right is defined.
In general the push-foward map does not preserve smoothness. If you consider
157.750
the constant map F : M −→ {p} ∈ N then it follows from (8.19) that F∗ w has
support contained in {p}. Indeed, if p ∈ / supp(φ) then supp(F ∗ φ) = 0. In fact in
∗
is constant on M for this map (with F ∗ W trivial) so
general it follows that φ φ 157.750
the pairing on the right in (8.19) is just an integral. Clearly this is not always zero,
and then F∗ w is necessarily singular.
So one cannot expect too much regularity for F∗ w even if one assumes that
w ∈ C c∞ (M ; W 0 ⊗ Ω). However, one of the properties of a fibration (the map corre-
sponding to a fibre bundle) is that a version of Fubini’s Theorem holds.
157.751 Proposition 8.2. If F : M −→ N is a fibre bundle then for any vector bundle
W −→ N,
157.752 (8.20) F∗ : C c∞ (M ; W 0 ⊗ Ω) −→ C c∞ (N ; W ⊗ Ω)
is surjective.
Proof. As I 157.750
say, Fubini. The regularity of F∗ w is a local question on N since
we can see from (8.19) that if φ ∈ C ∞ (N ) then
157.753 (8.21) ψF∗ (w) = F∗ (F ∗ ψw).
 3. PULL-BACK AND PUSH-FORWARD 129

So it is enough to assume that w has support in F −1 (O) where the fibre bundle
is trivial over O ⊂ N and then use a partition of unity; we can arrange that W
is trivial over O as well. Then the diffeomorphism invariance of integration (of
densities) means that we really are reduced to the case that F −1 (O) = Z×) and
we are integrating over Z. 
Once we know this we can reverse the definition of push-forward and see that
for a fibration pull-back extends by continuity to
157.754 (8.22) F ∗ : C −∞ (N ; W ) −→ C −∞ (M ; F ∗ W ) and is injective.
Really the pull-back is ‘constant along the fibres of F ’.
We are interested in more refined version of this. In particular notice that if
S ⊂ N is a closed embedded submanifold then for a fibration
157.755 (8.23) F −1 (S) ⊂ N is closed and embedded.
157.754
157.756 Proposition 8.3. Pull back under a fibration, as in (8.22) defines a continu-
ous linear map
157.757 (8.24) F ∗ : I m (N, S; W ) −→ I m−d/4 (M ; F −1 (S); F ∗ W )
for any closed embedded submanifold of N of codimension d and any vector bundle
W over N.
Proof. 
157.752
We can, and should, ask a similar question abovt the push-foward map (8.20).
So consider a closed, embedded submanifold D ⊂ M. There are already lots of
submanifolds of M as the total space of a fibration, namely the fibres. Recall
157.758 Definition 8.3. Two embedded submanifolds D and Z in a manifold M meet
transversally if at each point intersection
157.759 (8.25) p ∈ D ∩ Z =⇒ Tp D + Tp Z = Tp M ⇐⇒ Np∗ D ∩ N ∗ )pZ = {0}.
157.760 Lemma 8.1. The transversal intesection of two embedded submanifolds is an
embedded submanifold with codimension the sum of the codimenstions.
In particular two manifolds which do not intersect ‘intersect transverally’. For two
embedded submanifolds the notation for their intersection D t Z means that they
intersect transversally.
Proof. By definition near any point of D there are local defining functions
wi which vanish on DW with independent differentials. Similarly there are local
defining functions uj for Z. At a point of intersection the transversality condition
means that the wi and uj have independent differentials. Since they vanish precisely
on the intersection locally, it is an embedded submanifold. 
157.762 Definition 8.4. We say that a smooth map F : M −→ N is transvesal to an
embedded submanifold D ⊂ M (or that the submanifold is transversal to the map)
if the differential F∗ : Tp D −→ TF (p) N is surjective for each p ∈ D.
157.761 Proposition 8.4. A submanifold D ⊂ M of the total space of a fibration is
transversal to the fibration if and only if it is transversal to each fibre; if D is closed
then
157.763 (8.26) F∗ : Icm (M, D; F ∗ W ⊗ Ω) −→ C c∞ (M ; W ⊗ Ω).
 130 8. SEMICLASSICAL QUANTIZATION

So integrating along the fibres of a fibration which is transvesal to a submanifold

‘wipes out’ the singularities of conormal distributions.

Proof. 

4. Abstract product theorem

Now, what I have been building up to here is a result which we can use to
prove composition result for operators. The case to bear in mind is one we have
already covered. Namely take the triple product M 3 and denote a general point
(m1 , m2 , m3 ). Then consider the three diagonals,
157.771 (8.27) D1 = {m1 = m2 }, D2 = {m2 = m3 }, S3 = {m1 = m3 }.
Each pair of these intersect transversally in the same ‘triple diagonal’
157.772 (8.28) T = {m1 = m2 = m3 }.
Now consider the smooth map
157.773 (8.29) φ : M 3 3 (m1 , m2 , m3 ) 7−→ (m1 , m3 ) ∈ M 2 .
So
157.774 (8.30) D3 = φ−1 (S), S = Diag = {(m, m) ∈ M 2 }.
The product formula for pseudodifferential operators involves the multiplication
of conormal distributions with respect to D1 and D2 and push forward under φ.
The first step involves a result which really goes way back to the beginning of
the course.
157.765 Lemma 8.2. If Di ⊂ M, i = 1, 2 are closed embedded submanifolds which in-
tersect transversally then the product of conormal distributions is well defined
157.766 (8.31) × : I ∗ (M, D1 ; W1 ) × I ∗ (M, D2 ; W2 ) −→ C −∞ (M ; W1 ⊗ W2 ).
There is no statement of conormality of the product, because it is not true in general
(and with our definitions so far does not make sense which is reassuring)!
157.764 Theorem 8.3. If D1 , D2 are closed embedded transversal submanifolds of the
total space of a fibration F : M −→ N to each of which the fibration restricts to be
a diffeomorphism and such that there exists a closed embedded submanifold S ⊂ N
with
157.767 (8.32) φ−1 (S) t D1 = φ−1 (S) t D2 = D1 t D2
and a bundle map
157.769 (8.33) h : W1 ⊗ W2 −→ φ∗ (W ) ⊗ ΩM
then the composite map
157.768 (8.34) φ∗ (gξ) : Icm1 (M, D1 ; W1 ) × I m2 c(M, D2 ; W2 ) −→ I m (N, S; W ⊗ W ).
Consider the three submanifolds D1 , D2 and D3 = φ−1 (S). By as-
Proof.157.767
sumption, (8.32), they intersect transversally in pairs. Since φ is assumed to be a
diffeomorphism when restricted to D1 it follows that the restriction of the fibration
157.770 (8.35) φ : D1 ∩ D2 −→ S
 4. ABSTRACT PRODUCT THEOREM 131

157.767
is a diffeomorphism. Now it follows from (8.32) that all three manifolds must have
the same codimension – since the intersection has codimension which is equal to
the sum of each157.768 pair.
To prove (8.34),
157.766
first note that away from D1 ∩ D2 the image of the product
map in (8.31) lies in the sum of I ∗ (M ; D1 , W1 ⊗ W2 ) and I ∗ (M, D2 ; W1 ⊗ W2 ) so,
157.761
after composing with the bundle map h, so Proposition 8.4 applies and shows that
the push-forward is smooth. Thus, it suffices to consider a small neighbourhood of
a point n ∈ S ⊂ N and its preimage. We may also assume that the elements of
Icm1 (M, D1 ; W1 ) and Icm2 (M, D2 ; W2 ) are supported near unique preimage n0 of n
in D1 ∩ D2 (by using a partition of unity and discarding smooth terms).
Now, it is convenient to observe that D1 and D2 and the fibration can be
brought to simultaneous normal form near such a point n0 ∈ D1 ∩ D2 . Let z =
(z 0 , z 00 ) be local coordinates in the base, N, near n in which S = {z 00 = 0}. Since
φ is assumed to be a diffeomorphism when restricted to D1 it follows that it is a
graph over N locally. So we can introduce additional variables y near n0 so that
(y, z 0 , z 00 ) form a coordinate system and
157.778 (8.36) D1 = {(y, z 0 , z 00 ); y = 0} near n0 .
Since D2 is also a graph over N near n0 it takes the form
157.779 (8.37) yi = Yi (z 0 , z 00 ) near n0 .
The codimension of D1 , the number of yi , is equal to the codimension of S ⊂ N,
i.e. the number of z 00 variables.
The assumtion that D1 ∩ D2 ⊂ φ−1 (S) means that Y (z 0 , 0) = 0 and the
transversality of D1 and D2 implies that
157.780 (8.38) dz00 Yi (z 0 , z 00 ) are linearly independent at (0, 0, 0)
(since the dz00 Yi = 0. These form a square matrix, so the Yi can be introduced
as new variables in the base, in place of the z 00 and defining S. This then is the
coordinate normal form
157.781 (8.39) D1 = {y = 0}, D2 = {z 00 − y = 0}, D3 = {z 00 = 0}.
Ignoring the bundles, it follows that the conormal distributions locally take the
form
157.782 (8.40) Z Z
00
−y)·η 0
u1 = (2π)−d a(z 0 , z 00 , η)eiy·η dη, u2 = (2π)−d b(z 0 , z 00 , η 0 )ei(z dη 0
Rd Rd
where we can assume that the symbols a and b are supported near (z 0 , z 00 ) = 0. The
push-forward of the product is then
Z
157.783 (8.41) u1 (z, y)u2 (z, y)
Z
00 0
−2d
= (2π) a(z 0 , z 00 , ξ + η 0 )b(z 0 , z 00 , η 0 )ez ·η eiy·ξ dηdξdy, ξ = η − η 0 .
R3d
Formally at least the ξ, y double integral can be interpreted as a Fourier/inverse
Fourier transform which evaluates the integrand at ξ = 0 giving
Z
00 0
−2
157.784 (8.42) φ∗ (u1 u2 ) = (2π) a(z 0 , z 00 , η 0 )b(z 0 , z 00 , η 0 )ez ·η dη 0 ∈ I ∗ (N, S).
R3d
To justify these last step we use continuity in the symbol topology as usual. 
 132 8. SEMICLASSICAL QUANTIZATION

5. Semiclassical pseudodifferential operators

The semiclassical pseudodifferential algebra quantizes the semiclassical Lie al-
gebroid. First we need to understand what this means.
157.785 Definition 8.5. A Lie algebroid on a manifold N is determined by a real
vector bundle, W, over N and a bundle map (the anchor map)
157.786 (8.43) a : W −→ T N
with the following additional properties
(1) The space of smooth sections C ∞ (N ; W ) is a Lie algebra
(2) The Lie bracket satisfies
157.787 (8.44) a([V1 , V2 ]) = [a(V1 ), a(V2 )] ∀ Vi ∈ C ∞ (N ; W ), i = 1, 2.
The space of sections C ∞ (N ; W ) is normally called the Lie algebroid.
So of course V (N ) = C ∞ (N ; T N ) is a Lie algebroid. In the compact case its
‘quantization’ is taken to be the space Ψ∗ (N ) of pseudodifferential operators.
I will leave open for the moment the precise definition of quantization, I may
suggest a definition below but it depends rather on how much one wishes to demand.
We have met a second example above, namely if φ : X −→ Y is a fibration of
compact manifolds then
157.788 (8.45) Vφ (X) = {V ∈ V (X); V φ∗ f = 0 ∀ f ∈ C ∞ (Y )}
is a Lie algebroid. The quantization is the algebra Ψφ ∗( X) of fibre-wise pseudodif-
ferential operators, which is directly involved in the definition of the Atiyah-Singer
index above. What is the bundle? If we take local coordinates yi in the base and
extend these to coordinates near a point of X by adding some zj then the elements
of Vφ (X) are locally of the form
X
157.789 (8.46) aj (y, z)∂zj .
j

So the bundle involved here is precisely the fibre tangent bundle φ T X ⊂ T X with
the anchor map being the natural inclusion (so of constant rank).
The case of immediate interest is the semiclassicl Lie algebroid which is closely
related to Vφ . Namely, take a manifold M and consider
157.790 (8.47) M [1, sl] = M × [0, 1] −→ [0, 1].
On this space we consider fibre vector fields which in addition vanish at the bound-
ary  = 0. Clearly in local coordinates, zi , on M this means the smooth vector fields
of the form
X
157.791 (8.48) ai (, z)∂zj .
i

These clearly form a Lie algebra, but what is the bundle V ? We have to construct
157.791
it. The form of (8.48) makes it rather clear that we have a local basis with elements
157.792 (8.49) ∂zi .
sl
So these give a basis for the bundle T M (which is a bundle over M [1, sl] not M
despite my notation). You might struggle a bit to think of ∂zi as an ‘entity’ rather
L21-end than the product of  and ∂zi but that is precisely what is involved here.
L22
 7. THE ATIYAH-SINGER INDEX THEOREM 133

6. Semiclassical index map

7. The Atiyah-Singer index theorem
The setting here is a smooth, compact fibre bundle
157.793 (8.50) Z X
φ

Y.
The compactness of the base is not very critical – we just need to assume that
everything is trivial outside a compact set; non-compactness of the fibres is not
much worse but see the discussion of Atiyah-Patodi-Singer below.
The basic question then, is given a smooth family of differential operators acting
between sections of vector bundles on the fibres of φ as a map
157.804 (8.51) A : C ∞ (M ; V ) −→ C ∞ (M ; W )
when is it invertible?
This is hard.
Here is a step-by-step outline of the proof of the Atiyah-Singer index theorem
(for families, in K-theory) showing what remains to be done – since I do not expect
to have enough time to do every thing in detail.
i.data (1) Definition of the index map – this is still a little incomplete. We can
think of ‘quantization data’ as a triple (V, W, a) where V and W are
vector bundles over X and
157.794 (8.52) a ∈ C ∞ (Sφ∗ X; π ∗ hom(V, W ))
is an invertible isomorphism between the lifts of V and W from X to the
fibrewise cotangent sphere bundle over X. We talked about more general
elliptic symbols earlier.
(2) We have not quite finished the proof that the K-group K 0 (Tφ∗ X) is identi-
157.794
fied with the equivalence classes of the data (8.52) under the three relations
of bundle isomorphism (over X of V and W ), homotopy and stability.
(3) Then we can quantize a to a family of pseudodifferential operators A ∈
Ψ0φ (X; V, W ) – this is the surjectivity of the symbol map. Any two such
quantizations are homotopic and there are qunatizations where the null
spaces form a vector bundle over Y.
(4) The K-group K 0 (Y ) can be identified with equivalence classes of pairs of
vector bundles over Y with the relations, bundle isomorphism, homotopy
and stability (by adding an one bundle to both).
(5) The quantizations have appropriate properties under these maps so that
∗
157.795 (8.53) ind : K 0 (Tfib X) −→ K 0 (Y )
is well-defined with the index being the difference of the null bundle and
a complement to the range for a quantization A (any one for which the
null spaces form a bundle). 157.795
(6) Now we want to deform the index map (8.53) into the semiclassical cal-
i.data
culus. To do this we generalize the data in 1. Namely we consider triples
∗ X −→ GL(N, C) and
(β1 , β1 , b) where the βi are involutive families Tfib
 134 8. SEMICLASSICAL QUANTIZATION

b ∈ C ∞ (Tfib
∗ X; hom(CN )) is such that b : Ran(P ) −→ Ran(P ) is an iso-
1 2
∗ ∗ X) with P = 1 (β + Id) the
morphism over Sfib X (the boundary of Tfib i 2 i
positive projections of the involutions. This is the semiclassical data con-
sidered above the equivalence classes under homotopy and stability again
form the group K 0 (T ∗ X. Semiclassical quantization
157.795 fib
results in the same
i.data
index map (8.53) when the data reduces to that in 1.
∗
i.escl constant and b = Id exhausts K 8 (Tfib
(7) The data where β2 is157.795 X) so we can
get the index map (8.53) from semiclassical smoothing operators – were
β1 is constant near infinity.
(8) Now the main part of the proof of the index theorem is the embedding of
φ in a trivial fibration Rk × Y for some (largish) k :

X / Rk × Y
ι×φ
157.796 (8.54) Z Rk
φ
π2
 {
Y.
This follows by choosing an embedding ι : X −→ Rk and then into Rk × Y
by adding the map φ.
(9) Now we think of X as a submanifold of Rk ×Y and contemplate its normal
bundle – so an open neighbourhood of the image. Each fibre Zy of X is
embedded in Rk so the full normal bundle is the bundle over X which
over Zy is N Zy ⊂ Rk × {y}. So we actually have a ‘tower’ of fibrations

157.797 (8.55) Rd NX
ψ

Z X
φ

Y
(10) The main idea in this proof by Atiyah-Singer (they had another one too,
using cobordism) is that we can ‘extend’ the data and quantization from
∗
Tfib X to
∗
157.798 (8.56) Tψψ N X = N X ⊕ N ∗ X −→ X.
157.797
The fibres of ψ in (8.55) are non-compact but of course they are real vector
157.798
spaces. So the fibewise cotangent bundle in (8.56) has fibres the sum of a
vector space and its dual.
∗
(11) So, we can find a family of involutions on Tψψ N X which are costant
outside a compact set and quantize (semiclassically) to have index a trivial
one-dimensional bundle over X.
(12) Then we take what is essentially the tensor product of this ‘Bott element’
(well, better to say a ‘Thom’ element) with the original family over X to
be a family for φ ◦ ψ : N X −→ Y which quantizes to the have the same
i.escl
image as a given element 7. We do this by quantizing in two steps.
(13) Now, we can arrange the support of the Thom element to be very close
to the zero section and thereby move it, using a collar map, to Rk × Y
 9. THE DIRAC CASE 135

as a family. This is ‘excision’ and gives the same index for the extended
family.
(14) The index map from Rk ×Y to Y is an isomorphism – again this is explicit
Bott periodicity.
(15) Finally then we have an extended construction of the index map giving a
commutative diagram

157.799 (8.57) K 0 (Tφ∗ X)

⊗τ
/ K 0 (T ∗ (N X)) exc / Tπ∗ (Rk × Y )
φψ 2

ind
ind
'  v ind

K 0 (Y )
(16) The final step then is to see that this diagram is the definition of the
push-forward in K-theory, the ‘Gysin’ map in this context.

8. Index formula
Of course this is not the end of the story, quite apart from the fact that there
are a few gaps in the argument – which I try to fill in below. The index in K-theory
as above precisely captures the obstruction to an ellipitic family (or semiclassical
family) have a smoothing perturbation which makes it invertible. In the case that
Y is a point K 0 (Y ) = Z and we can look for a formula for the actual ‘numerical
index’. In the case of a family we can look for a simpler obstruction to perturbative
invertibility, corresponding to the image of the the index in (let’s say deRham)
cohomology under the Chern character
157.800 (8.58) Ch : K 0 (Y ) −→ H ev (X; R).
Either of these is the index formula.
157.801 Theorem 8.4. The index of the image in cohomology is
Z
157.802 (8.59) Ch ◦ ind = φ∗ (Ch([(V, W, a]) ∧ Td) = Ch([(V, W, a]) ∧ Td

where the push-forward in cohomology if realized as integration of a (compactly

supported) form over the fibres of T ∗ fibX.
The extra factor is the Todd class, which we can see from the proof above should
be
157.803 (8.60) Td = Ch(τ )
appropriately interpreted.

9. The Dirac case

L22-end
L23
L23-end
L24
 CHAPTER 9

Manifolds with boundary

Cmb
In these last three lectures I want to go through another example of quanti-
zation, leading to an algebra of pseudodifferential operators. In fact this is better
thought of as ‘microlocalization’ of a Lie algebroid. In this case we consider a com-
pact manifold with boundary M. We have already come across the Lie algebroid of
smooth vector fields on M which are tangent to the boundary. You might like to
check what happens on a manifold with corners but for the monment I will stick
with codimension one.
Set
157.805 (9.1) Vb (M ) = {V ∈ C ∞ (M ; T M ); V is tangent to the boundary}.
There is always a boundary defining function x ∈ C ∞ (M ) and the tangency condi-
tion just requires
157.806 (9.2) V ∈ Vb (M ) ⇐⇒ V x ∈ xC ∞ (M ).
So these are the vector fields which map the ideal of functions vanishing at the
boundary (which is a primitive ideal generated by x) into itself. There is a strong
‘naturality’ case for the consideration of Vb (M ) since it is the Lie algebra of the
group of diffeomorphism of M.
If we take local coordinates near a boundary point with yi coordinates on the
boundary then locally
X
157.807 (9.3) V ∈ Vb (M ) ⇐⇒ V = a(x, y)x∂x + bi (x, y)∂yi
i

for arbitrary smooth coefficients. This means that there is a vector bundle b T M
over M with sections precisely these vector fields
157.808 (9.4) Vb (M ) = C ∞ (M ; b T M ).
Let’s think a little about the structure of the vector bundle b T M, since I am
asserting it is, in context, the appropriate replacement for the ‘ordinary’ tangent
bundle T M (to which is it isomorphic – just not naturally so). Over the interior
of M there is not much to say since these two bundles are naturally isomorphic.
Since the elements of Vb (M ) are smoth vector fields there is a completely natural
smooth vector bundle map (the anchor map of the Lie algebroid Vb (M ))
b
157.899 (9.5) T M −→ T M.
Over the boundary this has corank 1 – there is a 1-dimensional null space since
the vector field (in local coordinates) x∂x vanishes in the ordinary sense at the
boundary. So there is a 1-dimensional subbundle
b
157.900 (9.6) N ∂M ⊂ b T∂M M.
137
 138 9. MANIFOLDS WITH BOUNDARY

In fact this is a canonically trivial subbundle. The element x∂x is actually (at a
boundary point) defined independently of coordinates. Indeed it satisfies

157.901 (9.7) (x∂x )ρ = ρ + O(ρ2 )

for any defining function ρ. This just reflects the fact that any other defining func-
tion ρ = a(y)x + O(x2 ).
In this behavour b T∂M M is ‘reversed’ from T∂M .M The latter has T ∂M as a
subbundle, with the quotient being N ∂M, the normal bundle. The former has a
(trivial) subbundle with quotient naturally T ∂M.
Let’s go a little further with this analysis of tangency. The primitive ideal of
functions vanishing at the boundary, I ∂ , generated by x, leads to a Lie ideal

157.902 (9.8) I ∂ · Vb (M ) ⊂ Vb (M ), [I∂ · Vb (M ),Vb (M )] ⊂ I ∂ · Vb (M ).

This means that the space of sections of b T∂M M as a bundle over ∂M is itself a
Lie algebra. This is clear enough in local coordinates. So there is actually a Lie
algebra map

157.903 (9.9) Vb (M ) −→ C ∞ (∂M ; b T M ).

I have perhaps not emphasized enough uthat the Lie algebra structure of V (M )
in the boundaryless case, or Vb (M ) here, is what leads to the properties of the
differential operators, in particular that the leading part defines the (polynomial)
symbol map. This is fair warning that we should expect something similar at the
boundary for our, yet to be defined, b-pseudodifferential
157.903
operators. It will be the
‘indicial operator’ and arises precisely because (9.9) is a map of Lie algebras. We
can even guess it should take values in the pseudodifferential operators on the
boundary but with ‘an extra parameter’.
So what we want to find is an algebra of operators, say on C ∞ (M ), which
include the vector fields Vb (M ) and multliplication by C ∞ (M ) and which away
from the boundary should reduce to ordinary pseudodifferential operators. One
can approach this as for the semiclassical calculus. Writing, informally, a pseudo-
differential operator in terms of symbols we can try to replace a symbol

157.809 (9.10) a(x, y, ξ, η) by a(x, y, xξ, η)

where (ξ, η) are the dual variables to (x, y). This does work but there are significant
issues involved.
Proceeding formally we can plug such a symbol into the inverse Fourier trans-
form and then change variables as for the semiclassical calculus (ignoring isses of
domains and convergence)
Z
0 0
157.867 (9.11) a(x, y, xξ, η)ei(x−x )ξ+i(y−y ) dξdη
Z
x−x0 0 dτ
= a(x, y, xξ, η)ei( x )τ +i(y−y ) dη, τ = xξ.
x

So, from this point of view what we need to do is to find a space on which (x−x0 )/x
is smooth, at least where it is finite. This we can do by an appropriate blow-up.
 1. THE B-GENERALIZED PRODUCTS 139

1. The b-generalized products

The approach I have taken, from the beginning, to pseudodifferential operators
is to try to define them directly as spaces of conormal kernels. I have mentioned
in passing the problem that immediately arises on a manifold with boundary, that
the diagonal in M 2 – which is where we expect the singularities to be – does not
meet the boundaries transversally. Namely there is an obvious dependence relation
between the two defining functions x on the left and x0 on the right and one of the
defining functions for the diagonal, x − x0 , at the intersection x = x0 = 0, i.e. the
corner. There is no problem with the tangential variables.
The geometric solution to this conundrum is to do as we did for the semiclassical
caluclus and blow up the offending submanifold, in this case the corner. Thus we
define
157.810 (9.12) β : M [2, b] = [M 2 ; (∂M )2 ] −→ M 2 .
by blowing up the corner x = x0 = 0.
There is quite a lot to get used to in this new space! The result of the blow
up of the corner is that the new manifold has a new boundary hypersurfaces sep-
aratig two corners to the single corner before blow up (assuming the boundary is
connected, which I am doing implicitly here; nothing really bad happens if there
are several components). We know that the blow-up can be defined in terms of
polar coordinates, in this case since only the variables x and x0 are involved,
π
157.811 (9.13) (x, x0 ) = r(cos θ, sin θ), r ≥ 0, θ ∈ [0, ].
2
The new boundary hpersurface, here r = 0, is the ‘front face’ denoted
π
157.812 (9.14) ff(M [2, b]) = I × ∂M × ∂M, Iθ = [0, ].
2
We have also seen that we can cover a neighbourhood of the front face of a blow-up
by projective coordinates. In fact here this can be done with just one coordinate
system as far as x and x0 are concerned – of course we also need coordinates in the
two copies of ∂M – because the one variable x + x0 dominates both x and x0 over
the manifold. So the two functions
x − x0
157.813 (9.15) x + x0 and µ = ∈ [−1, 1]
x + x0
together with tangential coordinates cover the front face. The two functions 1 + µ
and 1 − µ are defining functions for the lifts of the ‘old’ boundaries x = 0 and
x0 = 0 (the lift here means the closure of the inverse image of the complement of
the centre, (∂M )2 , of blow up). Then the lifted diagonal is locally
157.814 (9.16) Diagb = {µ = 0} × Diag∂M near ff .
It follows that now it is transversal
157.813
to the boundary which it only meets in ff .
In fact the coordinates (9.15) are sometimes a bit awkward and it is simpler to
use the more obvious projective coordinates
x x0
157.815 (9.17) x0 and s = or x and t = = 1/s
x0 x
valid respectively away from the lifts of {x0 = 0} and {x = 0}. In particular either
of these simpler systems is valid near the diagonal.
 140 9. MANIFOLDS WITH BOUNDARY

The idea of this blow-up is to introduce a space of kernels which is much easier
to describe there than directly on M 2 . Note that one version of the Schwartz kernel
theorem states that, on a manifold with corners, continuous linear operators
157.816 (9.18) A : C˙ ∞ (M ) −→ C −∞ (M ) = (C˙ ∞ (M ; Ω))0
are identified with distributions
157.817 (9.19) A ∈ C −∞ (M 2 ; πR
∗
Ω).
157.816
The image space in (9.18) is the space of extendible distributions as is the space
of kernels on M 2 (apart from the density factor). These distributions are defined
on any compact manifold with corners X as the dual of C˙ ∞ (X; Ω) (they are the
analogue of tempered distributions on Rn with which they are identified for the
radial compactification).
157.818 Lemma 9.1. The blow-down map gives an isomorphism
157.820 (9.20) β ∗ : C˙ ∞ (M 2 ) −→ C˙ ∞ (M [2, b])
and in consequence the 157.816
Schwartz kernel theorem also identifies the space of contin-
uous linear operators (9.18) with
157.819 (9.21) A ∈ C −∞ (M [2, b]; πR
∗
Ω)
The point is that we are not actually changing the space of extendible distri-
butions by passing from M 2 to M [2, b], what is changing is the space of smooth
functions (which is getting bigger) and the space of conormal distributions with
respect to the diagonal (which is a pain to define on M 2 ).
Now, before going on let’s check that the passage to M [2, b] does ‘resolve’ the
Lie algebroid Vb (M ) in an appropriate sense. The vector fields in Vb (M ) acting
in M 2 are tangent to the corner x = 0 = x0
on the left (or the right) factor of M157.806
0
since they annihilate
157.807
x and satisfy (9.2). Thus they lift to be smooth on M [2, b].
157.815
In terms of (9.3) and the local coordinates (9.17), the ∂yi lift unchanged whereas
157.821 (9.22) x∂x = s∂s .
This may not seem like much of an improvement! However, the lifted diagonal is
at s = 1 so this vector field does not vanish there and we see:
157.822 Lemma 9.2. The elements of Vb (M ) lifted to M [2, b] from the left (or right)
factor of M in M 2 are transversal to the lifted diagonal Diagb ⊂ M [2, b] and so
the normal bundle to this submanifold is identified with b T M.
This is a minimal requirement for ‘resolution’ of Vb (M ).
As well as the ‘stretched double space’ there are similar replace-
ments for the higher products M k . I will invoke the stretched
triple space below. let me continue to assume that the bound-
ary of M is connected – if it has more than one component you
should proceed component by component, not thinking of inter-
action between the components which are ‘far apart’.
We can see that the boundary faces of M k consist of prod-
ucts where in each factor we have either ∂M or M. Now arrange
these in order of increasing dimension – starting at (∂M )k – and
then blow them up, one after another
M [k, b] = [M k ; (∂M )k ; M k−1 (M k ), . . . , M 2 (M k )] 157.868 (9.23)
 1. THE B-GENERALIZED PRODUCTS 141

where M p (M k )) is the collection of boundary faces of codimen-

sion p. We stop at p = 2 since boundary faces of codimension
1 are hypersurfaces
157.868
and blowing them up does nothing. Now,
what (9.23) really means is that we157.868
blow up the successive lifts
of the boundary faces. To see that (9.23) is well-defined we need
to note first that
157.869 Lemma 9.3. Under blow up of a boundary face the lift of the
other boundary faces are boundary faces.
157.868
157.870 Lemma 9.4. The sequence of blow ups in (9.23) is well-
defined since after the blow up of M p (M k ) the lifts of the ele-
ments of M p−1 are disjoint.
157.868
This means that the order at each step in (9.23) is immate-
rial.
157.871 Proposition 9.1. All the projections M k −→ M j , for j <
k, lift to be smooth maps and simple/x b-fibrations.
Here a b-fibration is a natural extension of the notion of
a fibration to the category of manifolds with corners (it is the
analogue of a Lefschetz map in algebraic geometry if that helps!)
Rather than discuss these in detail here let me just say they are
smooth surjective maps which near a point of the domain, of
codimension l take the form in appropriately chosen coordinates
in domain and range

F (x1 , . . . , xl , y1 , . . . , ym )) = (xα1 , . . . , xαk , y1 , . . . , yq ),

where the xαi are monomials in the xj
with no common factor. (9.24)
157.872
Thus each xj can appear as a positive power in at most one of
157.871
the xαi . In Lemma (9.1) the each xj at most once and as a single
power which is the meaning of ‘simple’. For such maps the αi
can be identified with disjoint subsets of {1, . . . , l}.
The importance of b-fibrations is that they have some of the
properties of fibration – to which they reduce in the absence of
boundaries.
157.873 Proposition 9.2. Under a simple b-fibration F :: M −→ N
between compact manifolds with corners
F∗ : {u ∈ C ∞ (M Ωb ; u ≡ 0 at M 2 (M )} −→ C ∞ (M ; Ω157.874
b ). (9.25)
157.874
In fact much more is true than (9.25) for a simple b-fibration.
Namely even if we don’t assume the vanishing of the Taylor series
157.874
at the corners as in (9.25), 7C ∞ (M ; Ωb pushes forward into
X
(log x)βC ∞ (M ; Ωb ). 157.875 (9.26)
finite

So the push-forward is smooth except for powers of logs of the

defining functions. These powers come from the behavour at the
 142 9. MANIFOLDS WITH BOUNDARY

157.874
corners of M, as follows from (9.25), and can be described much
more precisely. 157.873
In particular, Proposition 9.2 must apply to the projection
maps
πL,R β : M [2, b] −→ M. 157.876 (9.27)
157.871
Here it is essy to check Proposition 9.1 by hand. Away from ff
is just locally one of the projection from M 2 to M for
the map157.872
which (9.24) certainly holds. Near the front face we have one of
the two coordinate systems
x x0
s = 0 , x0 , y, y 0 or t = , x, y, y 0 . 157.877 (9.28)
x x
The left projection, to (x, y) therefore becomes either
(s, x0 , y, y 0 ) 7−→ (x = sx0 , y) or (x, y) 157.878 (9.29)
157.872
depending on the point in ff both of which satisfy (9.24).
157.872
It follows directly from (9.24) that a simple b-fibration is
locally a fibration in the interior and also at boundary points of
codimension one. This allows us to deduce
157.879 Lemma 9.5. For a compact manifold with corners
∞
(πL )∗ : Ψm
b (M ) −→ C (M ). 157.880 (9.30)
157.804
This gives a direct proof of the mapping property (8.51).

2. Conormality at the boundary

Cmb.con
 3. THE B-CALCULUS 143

3. The b-calculus
157.816
The most basic operator in (9.18) is the identity operator. We know that in
local coordinates the kernel of this is the Dirac ‘function’ at the diagonal
157.823 (9.31) δ(x − x0 )δ(y − y 0 ).
This is, as it must be, a distributional section of the right density bundle
∗
157.824 (9.32) Id ∈ I 0 (M 2 ; πR Ω).
To see that this makes invariant sense observe that it must be possible to pair the
delta ‘function’ with an element of C ∞ (M 2 ; πL
∗
Ω) since
∗ ∗
Ω(M 2 ) = πL Ω ⊗ πR Ω.
Say using a partition of unity to localize, we need to be able to make sense of the
distributional pairing, written formally as an integral
Z Z
157.826 (9.33) δ(x − x0 )δ(y − y 0 )ψ(x, y, x0 , y 0 )|dxdy| = ψ(x, y, x, y)|dxdy|
M2 M
157.824
which is indeed invariantly defined – so this is what (9.32) actually means.
Now, what happens when we look at the lift of this kernel to M [2, b] as an
extendible distribution – these form a subspace of the distributions on the interior.
It is certainly still supported at the diagonal,
157.815
so it is only a question of what it
looks like in the new coordinates say (9.17). The homogeneity of delta means it
becomes
157.827 (9.34) δ(x − x0 )δ(y − y 0 ) = (x0 )−1 δ(s − 1)δ(y − y 0 ).
We can absorb the extra singular factor of x0 into the measure to see that
∗
157.828 (9.35) Id ∈ I 0 (M [2, b], Diagb ; πR Ωb )
where Ωb is the density bundle coming from b T M so in fact in a natural way it has
a basis near the boundary
dx
157.829 (9.36) |dy| =⇒ Ωb = x−1 Ω.
x
This leads us to the definition of b-pseudodifferential operators through their
kernels.
157.830 Definition 9.1 (‘Small’ b-calculus). The space of b-pseudodifferential oper-
ators on a compact manifold with boundary, acting between sections of vector
bundles V and W is

157.831 (9.37) Ψm m
b (M ; V, W ) = {A ∈ I (M [2, b], Diagb ;
∗ ∗
πL W ⊗ πR (V 0 ⊗ Ωb )); A ≡ 0 at both lifted boundaries}.
I have not actually defined the conormal space here but it is exactly the re-
striction of the usual conormal space if one extends 157.831
across the boundary. These
distributions are smooth away from the diagonal so (9.37) makes sense since these
‘old’ boundaries do not meet the diagonal. Locally such a kernel, with the b-density
removed, just looks like
A(x, y, s, y − y 0 ) smooth in (x, y) and conormal at s = 1, y − y 0 = 0.
 144 9. MANIFOLDS WITH BOUNDARY

The symbol map for conormal distributions gives us

157.833 (9.38)
σm : Ψm m
b (M ; V, W ) −→ (S /S
m−1
)(N ∗ Diagb ; π ∗ (hom(V, W ) ⊗ Ωb ) ⊗ Ω(N ∗ )).
157.822
Recall that here N ∗ Diag is the dual of the normal bundle, so we see from Lemma 9.2
∗
M (replacing T ∗ M in the boundaryless case). The
that this is identified with b T157.833
last book-keeping bundle in (9.38) is therefore Ω(b T ∗ M ) which is the dual of Ωb so
these two factor cancel and the symbol map is as simple as we could hope giving a
short exact sequence
157.834 (9.39)

Ψm−1 (M ; V, W ) 
b
/ Ψm (M ; V, W ) σm / (S m /S m−1 )(b T ∗ M ; π ∗ hom(V, W )).
b

157.835 Theorem 9.1. The b-pseudodifferential operators, Ψm

b (M ; V, W ), define con-
tinuous linear maps
157.836 (9.40) A : C ∞ (M ; V ) −→ C ∞ (M ; W )
and form modules
157.834
over the filtered *-closed algebra Ψm
b (M ; V ) for which the symbol
sequence (9.39) is multiplicative.
Proof. First let us check that
157.831
the b-psuedodifferential operators defined by
157.836
(9.37) do indeed define operators
157.831
(9.40). This is not quite obvious, but notice that
the space of kernels in (9.37) is a module over C ∞ (M [2, b]) so we can localize as
we wish using a partition of unity. In particular we can work in (relatively) open
subsets of M over which the bundles are trivial and so we are free to ignore them
and assume that V = W = C; this simplifies the notation. If we look at a pair
of neighbourhoods which do not meet the boundary then we are in the interior
157.836
case where we know (9.40). Similarly if one of the open sets does not meet the
boundary then the operator is again locally an interior pseudodifferential operator157.836
plus a smoothing operator vanishing rapidly at the boundary from which (9.40)
follows.
So we can localize to a product of neighbourhoods of points in the boundary –
although the two open sets need not meet the diagonal. One thing that is easy to
see is then is that
157.844 (9.41) A : C˙ ∞ (M ; V ) −→ C˙ ∞ (M ; W ).
Indeed (localized) the action on u ∈ C˙ ∞ (U ) where U ⊂ M is a coordinate neigh-
bourhood of a boundary point is by pushing forward the product
∗
157.845 (9.42) (πL )∗ (A · πR u)
By assumption, u vanishes to infinite order at x0 = 0 (the boundary of the right
∗
factor) so πR u vanishes to infinite order at the preimage, which includes ff . It
follows that the product also vanishes to infinite order at ff and then we are dealing
again with an ordinary pseudodifferential operator on M 2 .
Now, it actually follows from this that
157.846 (9.43) A : C −∞ (M ; V ) −→ C −∞ (M ; W )
so we know that Au is defined if u ∈ C ∞ (M ; V ) and we just need to show that it is
157.844
smooth up to the boundary. The argument giving (9.41) is included to show that
in this case Au ‘at the boundary’ should only depend on u ‘at the boundary’.
To make this precise we can again localize as above and only the terms where
x and x0 are near 0 are not clearly C ∞ . Using the density properties of conormal
 3. THE B-CALCULUS 145

distributions and of smooth functions we can assume more about A and then use
continuity. Namely we can suppose that the kernel A is continuous (we could
assume smoothness) and has support actually disjoint from the two ‘old boundaries’.
Then the coordiantes t = x0 /x, x y and y 0 are admissible over the support after
localization
dx0
157.847 (9.44) A = A(x, y, t, y − y 0 )| 0 dy 0 |.
x
The action of U on u is then the integral
dx0
Z
157.848 (9.45) Au = A(x, y, t, y − y 0 )u(x0 , y 0 )| 0 dy 0 |
x0 ≤ x
where u has compact support down to x0 = 0. This integral certainly exists for
x > 0 since then the integrand has compact supprt. We can change the variable of
integration from x0 to t and see that
Z
dt
157.849 (9.46) Au = A(x, y, t, y − y 0 )u(x/t, y 0 )| dy 0 |.
[0,∞) t
Now the integral exists (by fiat the support in t is in [C, 1/C] for some finite C)
157.836
and the integrand is smooth in x and y, so the result is C ∞ and we have (9.40)157.831
under these assumptions on A. In fact we can pass to unrestricted A as in (9.37)
since these kernels vanish rapidly at t = 0 and t = ∞ and 157.836
the only singularity is
conormal at t = 1 – across which we are integrating. Thus (9.40) follows in general
and as a bonus we see that, as anticipated above
157.836
157.850 Corollary 6. Restriction to the boundary in (9.40) gives
Au ∂M
= (A∂ )u , A∂ ∈ Ψm (∂M ; V, W ),
∂M
157.851 (9.47) 0 Z
0 dx 0
A = Ã(x, y, t, y − y )| 0 dy | =⇒ A∂ = A(0, t, y, y − y 0 )dt|dy 0 |
x R+
so the kernel of A∂ is the integral over the fibres of β : ff −→ (∂M )2 of the kernel
of A.
We still need to prove that the product of two b-pseudodifferential operators is
b-pseudodifferential. 
Thus everything is very much as in the boundaryless case except that we have
much more structure at the boundary. You might like to reflect on the similarity
to the behaviour of the semiclassical calculus here.
algebra Vb (M ) – these of course define elements
For a moment return to the Lie157.836
of Ψ1b (M ). Certainly they satisfy (9.40) but also
157.837 (9.48) (V u) ∂M
= (V ∂M
)u ∂M
, u ∈ C ∞ (M ).
They are ‘localized at the boundary.’ However we know that the analogous state-
ment for the semiclassical calculus holds but misses important structure at the
boundary. Much the same happens here.
The Collar Neighbourhood Theorem reminds us that a neighbourhood of a
submanifold looks like a neighbourhood of the zero section of its normal bundle.
For the boundary this translates to mean that a model for M near the boundary
is the inward-point half of the normal bundle
157.838 (9.49) N + ∂M = {v ∈ T∂M ; vx ≥ 0}/T ∂M.
 146 9. MANIFOLDS WITH BOUNDARY

The choice of a boundary defining function gives a positive section and hence triv-
ialization
157.839 (9.50) N + ∂M ←→ [0, ∞)dx × ∂M
where dx defines a linear functional on the fibres of T∂M M which vanishes on the
subbundle T ∂M.
Let’s pass to the radial compactification of N + ∂M = I ×∂M where I is a closed
interval, thought of as the radial compactification of [0, ∞) – there is no natural
trivialization
157.839
but any choice of boundary defining function provides one through
(9.50). The fibre R+ action extends smoothly to the compactification (fixing the
two boundaries {0} × ∂M and {∞} × ∂M ).
157.840 Proposition 9.3. There is a natural multiplicative map, the ‘indicial map’ to
R+ -invariant operators on the normal bundle, giving a short exact sequence
 
(9.51) xΨm / Ψm (M ; V, W ) I / Ψm+ (N + ∂M ; V∂M , W∂M ).
157.841 b (M ; V, W ) b R

Proof. Let’s choose a boundary defining function, rather than try to do things
invariantly – in the end nothing will depend on this choice. Then the model at the
boundary is
157.842 (9.52) N + ∂M = I × ∂M, I = [0, ∞]x .
The space on which the kernels for the b-pseudodifferential operators are defined is
therefore
157.843 (9.53) I[2, b] = [I 2 , {0} × {0}, {∞} × {∞}].
If you consider the R+ action on both factors starting at a point in the interior of
I[2, b] you will see that it is an open interval but the closure is smooth
157.841
up to both
the front faces. An R+ -invariant operator, as on the right of (9.51) corresponds
to a kernel which is constant under this action. So in fact it is determined by
its restriction to either of the front faces. Nothing much happens at the other
boundaries since everything is required to vanish to infinite order there.
So, the invariant operators are determined uniquely by their restrictions to the
front face ff 0 over x = 0. However this face for the model space is precisely the
same as the face ff(M [2, b]) – canonically diffeomorphic to it (independent of the
157.841
choice of defining function). So the indicial map in (9.51) is restriction of the kernel
to the front face, and then its null space consists of kernels that vanish there. You
might object that x + x0 , not x = (x + x0 )/(1 + s) is the defining function for the
front face, but the kernel vanish to infinite order where s → ∞ so we can just as
well divide by x.
This does not 157.841
explain the multiplicativity of the sequence but it follows that
the null space in (9.51) is pretty clearly an ideal, so the quotient is an algebra – it
is a question of what the product is! 157.850
To approach this reconsider Corollary 6. We have already noted that the space
157.831
of kernels in (9.37) is a module over C ∞ (M [2, b]) but more is true because of the
assumption of rapid vanishing at the boundary hypersurfaces other than ff . The
quotient of defining functions from the left and right, x/x0 , is smooth except at
one these two hypersurfaces and the rapid vanishing of the kernels there quashes
the singularity from x0 = 0. In fact the same is true for any power of this quotient,
 3. THE B-CALCULUS 147

which is to say that

157.852 (9.54) (x/x0 )iz is a multiplier on Ψm
b (M ; V, W ) for all z ∈ C.

157.853 Lemma 9.6. Conjugation generates an entire holomorphic family

157.854 (9.55) C 3 z 7−→ xiz Ax−iz
of automorphisms of Ψm
b (M ; V, W ).

It follows (even without knowing the multiplicative property for the operators)
that
157.855 (9.56) (AB)∂ = A∂ B∂
157.851
and we see from (9.47) that
Z
−iz
157.856 (9.57) iz
(x Ax )∂ = A(0, t, y, y − y 0 )t−iz dt|dy 0 |
R+

is an entire family of 157.856

pseudodifferential operators on ∂M.
The integral in (9.57) is the Mellin transform of the kernel of A restricted to
ff(M [2, b]). This is the (inverse) Fourier transform with respect to the variable
− log t ∈ R. In terms of log t the kernel is conormal at 0 × {y = y 0 } and decreases
faster than any exponential at ±∞.157.841
It follows that this indicial family determines
and is determined by the
157.855
image in (9.51).
157.856
In fact the multiplicativity of the map I
then follows from (9.56) and (9.57), with the latter being a convolution represen-
tation of the R+ -invariant operators on N + ∂M . 

We also want to analyse ‘L2 boundedness’ of b-pseudodifferential operators.

To conform to the general ‘b-yoga’ we should replace ‘ordinary L2 – meaning
computed with respect to a non-vanishing smooth density on a compact manifold
with boundary – with L2b (M ) computed with respect to a non-vanishing b-density.
Since the latter is just x−1 times the former, we see that
1
157.882 (9.58) L2 (M ) = x− 2 L2b (M ).
These spaces are well-defined for sections of vector bundles.
157.883 Proposition 9.4. Elements of Ψ0b (M ; V, W ) are bounded operators
157.884 (9.59) xs L2b (M ; V ) −→ xs L2b (M ; W ) ∀ s ∈ R.
157.884
Proof.
157.853
The case of general s in (9.59) follows from the case s = 0 in view of
Lemma 9.6.
For s = 0 we note that if we divide A ∈ Ψ0b (M ; V, W ), using a cut-off, into a
part supported very near ff(M [2, b]) and a part supported away from this boundary
hypersurface then the boundedness of regular pseudodifferential operators shows the
boundedness of the second part. We can further localize and reduce to the case
that V = W = C.
The argument for boundedness in Problems2, using the symbol sequence can
be applied almost verbatim here to show that boundedness follows if we can show
the boundedness of the ‘residual term’ Ψ−∞ b (M ).
Now, consider the R+ -invariant calculus on [0, ∞] × ∂M. We can characterize
the space L2b ([0, ∞] × ∂M ) in terms of the Mellin transform applied globally in the
 148 9. MANIFOLDS WITH BOUNDARY

first variable x – since this reduces to the Fourier transform on log x ∈ R. Thus in
fact
Z
dx
157.885 (9.60) v(x, y) −→ vM (z, y) = v(x, y)xiz
R x
extends from C˙ ∞ ([0, ∞] × ∂M ) to an isomorphism
157.886 (9.61) L2b ([0, ∞] × ∂M ) −→ L2 (Rz × ∂M ), z ∈ R.
157.856
The fact that the transformed operator in (9.57) is a family of pseudodifferential
operators in Ψ0 (∂M ) which is bounded as a function of z ∈ R shows that it is
157.886
bounded on the image of (9.61) so in fact the R+ -invariant operators are bounded
157.886
on the space on the left in (9.61). For the invariant operators we can again localize
near and away from the boundary and deduce that the part localized near the
boundary is bounded on L2b ([0, ∞]157.841
× ∂M ).
In view of the exact sequence (9.51) it suffices to consider elements of xΨ−∞
b (M )
(using the preceding argument). The extra vanishing at ff shows that this follows
directly from Schur’s criterion. 

Of course we really want boundedness on Sobolev spaces, but the ones we want
here are the xs Hbm (M ) which we need to define. If we work in a fixed product
157.885
decomposition near the boundary we can use the Mellin isomorphism (9.60) as we
would for Euclidean space and define

157.887 (9.62) Hbm ([0, ∞) × ∂M ) = {v ∈ C −∞ ([0, ∞) × ∂M ); vM ∈ L1loc (Rz ; H m (∂M ),

(1 + |z|)m vM (z) ∈ L2 (R × ∂M )}, m ≥ 0
L24-end For m < 0 we can use duality
L25

4. Metrics and boundaries

There are several intersting classes of metrics on a compact manifold with
boundary.
The most ‘obvious’ one I do not have time to talk about. This
is the case of a metric smooth, and non-degenerate up to the
boundary. It is rather a standard result, not too hard to see, that
the distance from the boundary is, at least near the boundary, a
smooth defining function and so can be extended to be smooth
and positive in the interior. Then the metric takes a particular
form in terms of the product decomposition near the boundary
given by flow along the normal geodesic
g = dx2 + h + xh0 , h a metric on ∂M. 157.857 (9.63)
Here h0 is actually an x-dependent family of symmetric tensor
on ∂M but in any case can be taken to be a smooth symmetric
2-tensor on M near the boundary.
Near the boundary there is then a corresponding decompo-
sition of the form bundle on M
Λ∗ M = Λ∗ ∂M ⊕ (dx ∧ Λ∗ ∂M ) 157.858 (9.64)
 5. HODGE THEOREMS 149

in which it is reduced to two copies of the forms on ∂M. The

Hodge-Dirac operator becomes a 2 × 2 matrix of operators
 
ð∂M −∂x
ð=d+δ = 157.859 (9.65)
∂x −ð∂M
157.858
where ∂x is acting on the coefficients in (9.64).
Certainly
ð : C ∞ (M ; Λ∗ ) −→ C ∞ (M ; Λ∗ ) 157.860 (9.66)
but is not symmetric on this domain (with respect to the inner
products on forms and density induced by g.) There are two
standard boundary conditions which lead to symmetry, namely
157.858
the vanishing of one or other of the two summands in (9.64).
More formally
DomAbs = {u ∈ C ∞ (M ; Λ∗ ); i∗∂M (ιν u) = 0}
157.861 (9.67)
DomRel = {u ∈ C ∞ (M ; Λ∗ ); i∗∂M u = 0}
where ν = ∂x is the Riemannian normal vector field at the
boundary.
Then
ð : DomAbs/Rel −→ C ∞ (M ; Λ∗ ) are Fredholm 157.862 (9.68)
and lead to Hodge decompositions as in the boundaryless case.
From this one deduces the two Hodge theorems
∗
null(ð) ∩ DomAbs ≡ HdR (M ),
∗ 157.863 (9.69)
null(ð) ∩ DomRel ≡ HdR (M, ∂M ).
These two deRham theories can be identified as the cohomologies
of the ‘absolute’ and ‘relative’ sequences
d : C ∞ (M ) −→ C ∞ (M ; Λ1 ) −→ · · · −→ C ∞ (M ; Λdim M )
157.864 (9.70)
d : C˙ ∞ (M ) −→ C˙ ∞ (M ; Λ1 ) −→ · · · −→ C˙ ∞ (M ; Λdim M ).
157.862
The main work here is to prove (9.68) and discuss the cor-
responding Hodge decompositions – which ultimately are very
much as in the boudaryless case. This can be done using the
‘edge’ calculus, or Boutet de Monvel’s ‘transmission’ calculus. I
probably will not have the time/energy to include these.
There is a long exact sequence relating the two cohomol-
ogy theories and the cohomology of the boundary that we will
encounter below
... / H k−1 (∂M ) / H k (M, ∂M ) / H k (M ) / ....
157.866 (9.71)

5. Hodge theorems
157.857
Rather than the ‘regular metrics’ as in (9.63) I want to consider two classes
of metrics which are known in the geometric literature as ‘cylindrical end’ and
‘asymptotically locally Euclidean’ metrics. In fact the precise definition of these
terms is a bit vague, so instead I will use the following notation.
 150 9. MANIFOLDS WITH BOUNDARY

157.865 Definition 9.2. A b-metric on a compact manifold with boundary is a fibre

metric on b T M which near the boundary is of the form
dx2
157.881 (9.72) gb = + h + xh0
x2
on some product decomposition, where h is a Riemann metric in ∂M and h0 is a
smooth quadratic form on b T M.
A scattering metric is then of the form
dx2 h
157.888 (9.73) gsc = x−2 gb = 4
+ 2 + x−1 h0
x x
near the boundary.
Since a product decomposition near the boundary always exists, every compact
manifold with boundary has a metric of either type.
157.888
Note in particular that an example of (9.73) is a Euclidean metric on a real
vector space, written in terms of the radial compactification.
Now, the idea is that we are supposed to think of these metrics ‘categorically’.
To analyse the Hodge-Dirac operator, we decompose the form bundle on157.881
M, near
the boundary, in terms of the product decomposition. So in the case of (9.72)
dx
157.889 (9.74) Λ∗b M = Λ∗ ∂M ⊕ ∧ Λ∗ ∂M near ∂M.
x
Then
 
ð∂ −x∂x
157.890 (9.75) ðb = d + δb = + xD
x∂x −ð∂
sign change comes from having to move past the dx/x factor. The ‘error’ term here
is a b-differential operator. Thus
 
ð∂ −x∂x
157.891 (9.76) ðb ∈ Ψ1b (M ; Λ∗b M ), I(ðb ) = .
x∂x −ð∂
157.889 157.858
If you do not put the dx/x in (9.74), but use (9.64) instead you will not get a
b-differential operator.
There is an analogue of the Hodge isomorphism here.
157.881
157.892 Theorem 9.2. For a b-metric (9.72) on a compact manifold with boundary
there is a natural isomorphism
{u ∈ L2b (M ; Λkb M ); ðu = 0} −→ Im HdR
k k


157.893 (9.77) (M, ∂M ) −→ HdR (M ) .
For a scattering metric we can proceed
157.888
in a similar fashion. We ‘rescale’ the
form bundle according to the forms in (9.73) – which are the ones that pair smoothly
with the vector fields in xVb (M ) – and see that
dx
157.894 (9.78) Λksc M = x−k Λk ∂M ⊕ ∧ x−k+1 Λk−1 ∂M near ∂M.
x2
A short calculation shows that in terms of this decomposition
157.895 (9.79)
−x2 ∂x − x(k + d)
 
xð∂
ðsc = d + δsc = + x2 D on Λksc M, d = dim ∂M
x2 ∂x − xk −xð∂
where D is again a b-differential operator.
 6. ELLIPTICITY AND PARAMETRICES 151

Now you may see why I have not needed to develop the theory of ‘scattering
pseudodifferential operators’ to handle this case since in fact
 
1 ∗ ð∂ −x∂x − (k − d − 1)
157.896 (9.80) ðsc = xR, R ∈ Ψb (M ; Λ M ), I(R) = .
x∂x − k −ð∂
This means that it is b-pseudodifferential operators which are relevant for the null
space of ðsc . It is a different matter if you wish to discuss the spectral theory of
this operator (which I feel you should want to do)– which is indeed really scattering
theory.
157.888
157.897 Theorem 9.3. For a scattering metric (9.73) on a compact manifold with
boundary of dimension n there is a natural isomorphism
1
157.898 (9.81) {u ∈ x 2 n L2b (M ; Λkb M ); ðsc u = 0} −→

H 


k
(M, ∂M ) k < 21 n
1 1

2n n
Im HdR (M, ∂M ) −→ Hb2 (M ) k = 12 n

H k (M ) k > 12 n.


157.898
The L2 space in (9.81) is the metric L2 space for gsc . Of course the ‘middle dimen-
sional case’ can only occur if n is even.
Try it out for the Euclidean metric on M = Rn . It follows that there is no
L2 null space at all! This corresponds to the fact that here are no L2 harmonic
forms on Rn – their coefficients would be harmonic functions which would mean
they decay at infinity. 157.892
Let’s think about a strategy for proving Theorem 9.2. First we need to get
some way to approach the deRham cohomology in this setting.
157.904 Proposition 9.5. For  > 0 the cohomology of the deRham complex

157.905 (9.82) x Hb∞ (M )

d / x H ∞ (M ; Λ1 ) d / ... d / x H −∞ (M ; Λn )
b b
∗
is naturally isomorphic to HdR (M, ∂M ) and the cohomology of

157.906 (9.83) x− Hb∞ (M )

d / x− H ∞ (M ; Λ1 ) d / ... d / x− H −∞ (M ; Λn )
b b
∗
is naturally isomorphic to HdR (M ).
The main step is the parametrix construction giving some ‘elliptic regularity’
and a Hodge decomposition.
157.907 Proposition 9.6. On a compact manifold with boundary and for  > 0 small
enough the Hodge-Dirac operator for a b-metric has null(ð)L2b ⊂ x Hb∞ (M ; Λ∗b ) and
is Fredholm as an operator on x− Hb∞ (M ; Λ∗b ) satisfying
(9.84) x− Hb∞ (M ; Λ∗b ) = null(ð)L2b ⊕ d x− Hb∞ (M ; Λ∗b ) ⊕ δ x− Hb∞ (M ; Λ∗b ) .



157.909
L25-end
L26
6. Ellipticity and parametrices
157.907
How do we prove say Proposition 9.6? We try to construct a parametrix as
a b-pseudodifferential operator; as we shall see this does not quite work; we shall
soon see why. An extension of the bounded result above is that
 152 9. MANIFOLDS WITH BOUNDARY

157.910 Lemma 9.7. An element of Ψ0b (M ; V ) is compact as an operator on L2b (M ; V )

if and only if both its principal symbol and its normal operator vanish.
It follows immediately that the same condition is necessary and sufficient for com-
pactness on any of the weighted spaces xs L2b (M ; V ).

Proof. The important point is the sufficiency, well the necessity is important
but only to know! 

To prove that ðb is Fredholm as a map x− H 1 (M ; Λ∗b ) −→ x− H 1 (M ; Λ∗b ) we

will want to construct a right parametrix modulo compact operators on L2b (M ; Λ∗b ).
If there were to be an element B ∈ Ψ−1 ∗
b (M : Λb ) satisfying

157.911 (9.85) ðb B = Id −E, E ∈ xΨ−1 (M ; Λ∗b )

with compact remainder E then we would need to have
157.912 (9.86) σ1 (ðb )σ−1 (B) = Id, I(ðb )I(B) = Id .
The first, symbolic, statement or ‘division problem’ is just ellipticity and is straigh-
forward. The second is an issue, and is almost never possible to satisfy within the
class of operators Ψ−1 ∗
b (M ; Λb ).
What is the problem? Well we know that the indicial operator is equivalent, in
157.891
terms of information, to the indicial family and in this case we can see from (9.76)
what it is:
 
ˆ ð∂ −iz
157.913 (9.87) I(ðb )(z) = .
iz −ð∂
We are asking that the inverse of this entire family of pseudodifferential operators
on ∂M exist for all z ∈ C. This is totally unreasonable!
What we can see is that there are values of z for which this operator has null
space. In fact we can see exactly what they are. Suppose we have an eigenvector
for ð∂ (which we do!)
   
ˆ u ˆ u
157.914 (9.88) ð∂ u = λu =⇒ I(iλ) = 0, I(−iλ) =0
−u u
Then there are points of non-invertibility at
157.915 (9.89) i Spec(ð∂ ) ⊂ C.

157.916 ˆ b ) is invertible for z ∈ C \ i Spec(ð∂ ) and

Lemma 9.8. The indicial family I(ð
defines a meromorphic family
157.917 (9.90) ˆ b )−1 : C \ i Spec(ð∂ ) −→ Ψ−1 (∂M ; Λ∗b
I(ð )
∂M
157.915
with poles of order 1 at (9.89) with finite residues the orthogonal projections onto
ˆ
the mull spaces of I.
157.918 Proposition 9.7. For any w ∈ R the operator I(ðb ) has an R+ -invariant
inverse with kernel
157.919 (9.91) K(s, y, y 0 ) ∈ ff(M [2, b]; Λ∗b ⊗ πR
∗
Ωb )
which lies in sw− H ∞ near s = 0 and t−w+ H ∞ near t = 0 = 1/s for  > 0
sufficiently small.
 6. ELLIPTICITY AND PARAMETRICES 153

We are mainly interested in the case w = 0.

In fact we can be much moreCmb specific
Cmb.con
about the behaviour of this kernel which
has expansions as discussed in §9.2. There is a significant difference between the
cases that w ∈ Spec(ð∂ ) or not. We are actually interested in the case that w is an
eigenvalue, in particular 0, of the Hodge-Dirac operator on the boundary.
By elliptic regularity we know that any solution of
157.920 (9.92) I(ðb )K = δ(s − 1)δ(y − y 0 )
must be smooth away from Diagb where the delta function is supported. The Cmb
Cmb.con
solution given by the Proposition is unique. It has an expansion in the sense of §9.2
at the boundaries of ff(M [2, b]) determined by the spectrum of ð∂
X
K∼ sλi ai (y, y 0 ) as s ↓ 0
λi ≥w
157.921 (9.93) X
K∼ t−λi a0i (y, y 0 ) as s ↓ 0.
λi >−w

Proof. We are really working with the spectral theory of ð∂ here. First notice
157.920
that any solution of (9.92) has a conormal singularity at the ‘diagonal’ appearing
on the right. Morevover, using the symbol map, we can construct a parameterix
satisfying
157.922 (9.94) I(ðb )K0 = δ(s − 1)δ(y − y 0 ) − E(s, y, y 0 ), E ∈ C c∞ (ff(M [2, b]).
So it remains to ‘solve away’ the error term E which has support in the interior of
ff .
We ‘know’ (I hope) that the expansion of a smooth function such as E (valued
here in 2 × 2 matrices acting on Λ∗ ∂M ) in the y variable in terms of the eigenbasis
of ð∂M converges rapidly. So on each eigenspace, with eigenvalue λi ∈ Spec(ð∂M )
we wish to solve
 
λi −s∂s
157.924 (9.95) Ki0 (s, y 0 ) = Ei (s, y 0 ).
s∂s −λi
This ordinary differential, and R+ -invariant, equation has a unique solution
which vanishes near s = 0 and any two solutions differ by an element of the 2-
dimensional null space which is spanned by
   
u u
xλ i and x−λi
−u u
where u is the eigenvector.
As s = 1/t → ∞ the chosen solution is in the null space. Thus we can arrange
157.921
(9.93) for this one term by adding the approriate element of the null space. Sum-
ming over the eigenexpansion
157.922
gives rapid convergence and 157.921
hence we do in fact find
a unique solution to (9.94) with the desired behaviour, (9.93), with respect to a
given w ∈ R. There can only be equality there if the wright w is equal to one of the
eigenvalues but since these form a discrete set the kernel satisfies
157.926 (9.96) K ∈ xw− Hb∞ near {s = 0} and K ∈ x−w+ Hb∞ near {t = 0}.

 154 9. MANIFOLDS WITH BOUNDARY

The idea of course is that this kernel is to be the restriction to ff(M [2, b]) of the
parametrix for ðb . Now it should be clear why we cannot get157.912 a parametrix in the
space Ψ−1
b (M ; Λ∗
b ) – because the solution we have found to ( 9.86) does not decay
157.830
rapidly at the boundaries of the front face as required by Definition 9.1.
157.928 Definition 9.3. Given a pair of index sets E = (E L , E R ) we define

157.929 (9.97) Ψ−∞,E

b
∗
(M ; V, W ) = A 0,E L ,E R (M [2, b]; πL ∗
W ⊗ πR (V 0 ⊗ Ωb ))
∗ ∗
+ A E L ,E R (M 2 ; πL W ⊗ πR (V 0 ⊗ Ωb )).
Here the first space consists of C ∞ sections awayCmb.con
from the left and right boundaries
of M [2, b] which are conormal in the sense of §2 with index sets E L and E R at the
left and right boundaries.
The second term will not appear below but needs to be there if we want to capture
the structure of the generalized inverse. 157.830
In fact these operators form a module over the operators in Definition 9.1 and
we write
−∞,E
157.930 (9.98) Ψm,E m
b (M ; V, W ) = Ψb (M ; V, W ) + Ψb (M ; V, W ).

157.931 Lemma 9.9. The elements of Ψm,E

b (M ; V, W ) define bounded operators from
w1 s w2 s−m
x Hb (M ; V ) to x Hb (M ; W ) for any s ∈ R and w1 ≥ w2 provided w2 <
inf Re E L and w1 > − inf Re E R .

7. Hodge theorem for b-metrics

157.892
Using the parametrix constructed above we157.907proceed to157.909
prove Theorem 9.2.
First we need to finish the 157.918
proof of Proposition 9.6 and so (9.84). The parametrix
construced in Proposition 9.7 for w = 0 shows that
157.943 (9.99) ðb : x− Hb∞ (M ; Λ∗ ) −→ x− Hb∞ (M ; Λ∗ )
is Fredholm. A complement to the range, namely the annihilator with respect to
the metric pairing, is the null space of
157.944 (9.100) ðb : x Hb−∞ (M : Λ∗ ) −→ x Hb−∞ (M : Λ∗ ).
In fact we already know that the null space lies in x Hb∞ (M157.911
: Λ∗ ). In fact this is
elliptic regularity at the level of the ‘small calculus’ as in (9.85) – or bettwer we
know that the error can be arranged to be in Ψ−∞ ∗
b (M ; Λ ). This is not compact
but does map
157.945 (9.101) Ψ−∞ ∗  −∞
b (M : Λ ) : x Hb (M : Λ∗ ) −→ x Hb∞ (M : Λ∗ ).
We are really interested in the Hodge cohomology, the null space of ðb on the
metric space L2g (M ; Λ∗ ) = L2b (M ; Λ∗ ). The same elliptic regularity shows that this
same as the null space of ðb on Hb∞ (M ; Λ∗ ) and so contained in the null space
the 157.943
of (9.99) – so finite dimensional. Now, the form of the parametrix shows the second
157.943
‘boundary’ part of elliptic regularity, namely that the null space of (9.99) consists
of elements with expansions, i.e. is contained in A E (M ; Λ∗ ) where the index set is
157.915
as in (9.89). So we conclude that
157.946 (9.102) {u ∈ x− Hb−∞ (M ; Λ∗ ); ðb u = 0} =⇒ u − u0 ∈ x Hb∞ (M ; Λ∗ )
 8. HODGE THEOREM FOR SCATTERING METRICS 155

where u0 is smooth up to the boundary, corresponding to 0 ∈ Spec(ð∂ ). However,

/ L2b (M : Λ∗ ) – the integral is logarithmically divergent – so
u0 ∈
157.944
157.947 (9.103) {u ∈ L2 (M : Λ∗ ); ðb u = 0} ⊂ x Hb∞ (M ; Λ∗ ) is the null space of (9.100).
So157.943
now we have our Hodge decomposition, initially in the form we have derived
from (9.99)
157.948 (9.104) x− Hb∞ (M ; Λ∗ ) = (d + δ)x− Hb∞ (M ; Λ∗ ) ⊕ {u ∈ L2 (M : Λ∗ ); ðb u = 0}.
The Hodge summand here consists of the closed and coclosed eleemnts of
x Hb∞ (M ; Λ∗ ) since the positive order of decay is enough to justify the integra-
tion by parts argument
157.949 (9.105) 0 = h(d + δ)u, dui = kduk2L2 .
157.904
So, assuming we accept Proposition 9.5, we already see that the Hodge cohomology,
157.947
RBM:Proof of Proposi-
∗ 157.904
given by (9.103), is mapped into HdR (M ; ∂M ). The usual proof of injectivity works. tion 9.5 needed
In fact the extra decay means that
157.950 (9.106) v ∈ x− Hb∞ (M ; Λ∗ ), dv = 0 =⇒ hv, u =⇒L2 = 0.
This map is therefore injective to

{u ∈ x Hb∞ (M ; Λ∗ }/ x Hb∞ (M ; Λ∗ ) ∩ dx− Hb∞ (M ; Λ∗ )

157.951 (9.107)
k k


= Im HdR (M, ∂M ) −→ HdR (M )
157.893
so we 157.948
have the existence and injectivity of the map (9.77). The Hodge decomposi-
to a closed form in xHb∞ (M : Λ∗ ) gives a two-sided inverse.
tion (9.104), applied157.892
Thus Theorem 9.2 is proved.

8. Hodge theorem for scattering metrics

157.897
To prove Theorem 9.3 we proceed very much as for a b-metric, obviously with
some changes.
First, the L2 space with respect to a scattering metric is
157.940 (9.108) xn/2 L2b (M ; Λ∗ )
since the Riemannian density is a positive multiple of x−n νb . Of course the L2 space
uses the inner product from gsc on the form bundles,
157.894
which gives a positive-definite
Λ∗sc in (9.78). The Fredholm properties of ðsc
inner product on the rescaled bundle157.896
are determined by those of I(R) in (9.80).
157.918
So we apply Proposition 9.7 for w = n/2 to the elliptic b-differential operator
x−1 ðsc and conclude that
157.952 (9.109) ðsc : xn/2−1− Hb∞ (M ; Λ∗sc ) −→ xn/2− Hb∞ (M ; Λ∗sc )
is Fredholm. Just as in the b-case, the Hodge cohomology
∗
157.953 (9.110) Hsc (M ) = {u ∈ L2g (M ; Λ∗ ); ðsc u = 0} ⊂ xn/2+ Hb∞ (M ; Λ∗sc )
157.952
is a complement, to the range in (9.109).
We need to analyse the indicial roots of R = x−1 ðsc , which is to say the singular
values of
 
ð∂ −x∂x − (N + n)
157.954 (9.111) I(R) =
x∂x + N −ð∂
 156 9. MANIFOLDS WITH BOUNDARY

where N is the number operator on boundary forms. Again these are generated
by the spectrum of ð∂ but the relationship is more complicated than in the b-case
above.
Each harmonic k-form, γ, on the boundary, in the null space of ð∂ , generates
a two-dimensional space on which I(R) acts, spanned by
dx
157.955 (9.112) x−k γ, ∧ x−k γ.
x2
These each correspond to an solution of I(R)v = 0, namely
γ dx γ
157.956 (9.113) xk , ∧ xn/2−k k .
xk x2 x
The first can be the leading term of a square-integrable form only if k > n/2 and
the second only if k < n/2.
To claify square-integrability we need to analyze the other indicial roots. An
eigenform for ð∂ with non-zero eigenvector corresponds to a pair of forms, a coclosed
(k−1)-form hk−1 and a closed k-form uk with duk−1 = λuk and δuk = λek−1 . These
generate a 4-dimensional bundle invariant under I(R) spanned by
uk−1 dx uk uk dx uk−1
157.957 (9.114) k−1
, 2 ∧ k , k and 2 k−1 .
x x x x x x
Since the 2-dimensional space spanned by the first two elements maps into that
spanned by the second two and conversely, the null space of I(R) is conatained in
these two subspaces.
In the two cases the indicial roots correspond to a null vector of
   
λ −iz + (n − k − 1)) λ −iz + (n − k)
157.958 (9.115) and
iz − k + 1 −λ iz − k −λ
which occur when

157.959 (9.116) (iz − k + 1)(iz − n + k + 1) = λ2 , (iz − k)(iz + (n − k) =−→2 =⇒

(iz)2 −(n−2)iz−(n−k−1)(k−1)−λ2 = 0, (iz)2 +(n−2k+1)iz)−k(n−k+1)−λ2 = 0 =⇒
(n − 2) 1 p (n + 1) 1 p
iz = ± (n + 1 − 2(k − 1))2 − 4(n − k)(k − 1) + 4λ2 , iz = x ± (n + 1)2 + 4λ2
2 2 2 2
157.941 ˆ
Lemma 9.10. The indicial family I(R)(z) acting on Λk ∂M ⊕ Λk−1 ± M is
invertible except for
r
d+1 d+1
157.942 (9.117) z= ± λ2j,k + ( − k)2 .
2 2
where the λ2j,k are the eigenvalues of the Laplacian on ∂M acting on closed k-forms.

9. Atiyah-Patodi-Singer index theorem

157.932 Proposition 9.8. Any element A ∈ Ψm
b (M ; V, W ) which is elliptic, i.e. σ(A)
is invertible, is Fredholm as a map
157.933 (9.118) xw Hbs (M ; V ) −→ xw Hbs−m (M ; W )
for w ∈ R \ S where S ⊂ R is discrete.
 9. ATIYAH-PATODI-SINGER INDEX THEOREM 157

Proof. The main point here is that ellipticity alone does imply that the indi-
cial operator I(A) is invertible as an R+ -convolution operator such a set of weights.
The main part of this is that the indicial family
157.934 (9.119) ˆ
I(A) : C −→ Ψm (∂M ; V, W )
which is entire, has a meromorphic inverse – so with a discrete set of poles
157.935 (9.120) b -Spec(A) ⊂ C s.t. zj ∈ b -Spec(A), |zj | → ∞ =⇒ | Re zj | → ∞.
So this means there are only finiteley many poles in any strip Re z| < R.
This in turn is a form of ‘analytic Fredholm theory’. For a holomorphic family
of elliptic operators, such as we have here, defined on a connected open set, the
inverse is meromorphic there is one point at which the operator is invertible. In
157.935
this case the existence of such a point, and the bound on the set of poles in (9.120)
is a consequence of
157.936 Lemma 9.11. The indicial family I(A)(t+iτ ) for t ∈ R is a semiclassical family
down to  = ±1/τ as τ → ±∞ in R.
Proof. 
Then the parametrix construction above generalizes to yield the Fredholm prop-
157.933
ery (9.118) for
157.937 (9.121) w∈
/ Re b -Spec(A).

157.933
So the Fredholm condition for (9.118) for w in an open set with discrete comple-
ment. The index is necessarily constant on the open sets but changes as w crosses
an end-point. In fact there is a multiplicity function corresponding to the algebraic
ˆ
multiplicity of the residues at the poles of I(A),
157.938 (9.122) rank : b -Spec(A) −→ Z
and the change of the index in passing from one interval to the next is the sum
of the multiplicity of the points in b -Spec(A) with real part corresponding to the
end-point. It is elementary to see that the index is a decreasing
APS
function of w.
For Dirac operators Atiyah, Patodi and Singer ([?]) gave a formula for the
index which applies in this case. This is extensively discussed, from the present
tapsit
point of view, in [?]. You might ask, is there a formula for the index in this general
b-pseudodifferential case? The answer of course is yes!
If you have survived this far, you would certainly be tempted to ask: Is there
a families index theorem? There is, at least there is a families index formula for
the Dirac case in the literature and this can be extended to the case of families of
Fredholm b-pseudodifferential operators. However, unlike the case of one operator
discussed above, for an elliptic family of b-pseudodifferentia operators on the fibres
of a fibre bundle (so of course the fibres are compact manifolds with boundary)
there is an obstruction to this forming a Fredholm family.
157.939 Proposition 9.9. The family of indicial operators of an elliptic family of b-
pseudodifferential operators on the fibres of a compact fibre bundle define an index
class in K 1 the vanishing of which is a necessary and sufficient condition for the
existence of a Fredholm family with the same symbol.
 158 9. MANIFOLDS WITH BOUNDARY

10. Spectral and scattering theory

11. What else?
L26-end
 Bibliography

CoSLG [1] Pierre Albin, Panagiotis Dimakis, Richard Melrose, and David Vogan, Compactification of
semi-simple lie groups, 2019.
deConcini-Procesi [2] Corrado De Concini and Claudio Procesi, The invariant theory of matrices, University Lecture
Series, vol. 69, American Mathematical Society, Providence, RI, 2017. MR 3726879
MR1269107 [3] Alain Grigis and Johannes Sjöstrand, Microlocal analysis for differential operators, London
Mathematical Society Lecture Note Series, vol. 196, Cambridge University Press, Cambridge,
1994, An introduction. MR 1269107
Hormander-pseudo [4] L. Hörmander, Pseudo-differential operators, Comm. Pure Appl. Math. 18 (1965), 501–517.
Hormander-WeylCalculus [5] , The Weyl calculus of pseudo-differential operators, Comm. Pure Appl. Math. 32
(1979), 359–443.
iml [6] R. B. Melrose, Introduction to microlocal analysis, Lecture notes.
MR2650633 [7] Alberto Parmeggiani, Spectral theory of non-commutative harmonic oscillators: an introduc-
tion, Lecture Notes in Mathematics, vol. 1992, Springer-Verlag, Berlin, 2010. MR 2650633
MR0273463 [8] M. A. Šubin, Pseudodifferential operators in Rn , Dokl. Akad. Nauk SSSR 196 (1971), 316–319.
MR 0273463
MR2952218 [9] Maciej Zworski, Semiclassical analysis, Graduate Studies in Mathematics, vol. 138, American
Mathematical Society, Providence, RI, 2012. MR 2952218

159