INTRODUCTION TO MICROLOCAL ANALYSIS

PETER HINTZ

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2. Schwartz functions and tempered distributions . . . . . . . . . . . . . . . . . . . . 5
2.1. Fourier transform and its inverse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2. Sobolev spaces and the Schwartz representation theorem ............. 8
2.3. The Schwartz kernel theorem ........................................ 9
2.4. Differential operators ................................................ 9
2.5. Exercises ............................................................ 11
3. Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1. Ellipticity ........................................................... 15
3.2. Classical symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3. Asymptotic summation .............................................. 16
3.4. Exercises ............................................................ 19
4. Pseudodifferential operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.1. Left/right reduction, adjoints ........................................ 23
4.2. Topology on spaces of pseudodifferential operators . . . . . . . . . . . . . . . . . . . . 26
4.3. Composition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.4. Principal symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.5. Classical operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.6. Elliptic parametrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.7. Boundedness on Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.8. Exercises ............................................................ 34
5. Pseudodifferential operators on manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.1. Local coordinate invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.2. Manifolds, vector bundles, densities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.3. Differential operators on manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.4. Definition of Ψm (M ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.5. Principal symbol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.6. Quantization ........................................................ 51
Date: January 11, 2022.
I am grateful to everybody who sends me corrections and comments via email!
1
2 PETER HINTZ

5.7. Operators acting on sections of vector bundles . . . . . . . . . . . . . . . . . . . . . . . . 52

5.8. Special classes of operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.9. Elliptic operators on compact manifolds, Fredholm theory . . . . . . . . . . . . 54
5.10. Sobolev spaces on manifolds ........................................ 57
5.11. Elliptic operators on compact manifolds, revisited . . . . . . . . . . . . . . . . . . . 60
5.12. A simple nonlinear example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.13. Commutators and symplectic geometry ............................. 62
5.14. Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6. Microlocalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
6.1. Operator wave front set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
6.2. Elliptic set, characteristic set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.3. Wave front set of distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
6.4. Pairings, products, restrictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6.5. Exercises ............................................................ 78
7. Hyperbolic evolution equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
7.1. Existence and uniqueness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
7.2. Egorov’s theorem; propagation of singularities . . . . . . . . . . . . . . . . . . . . . . . . 82
7.3. Exercises ............................................................ 84
8. Real principal type propagation of singularities . . . . . . . . . . . . . . . . . . . . 85
8.1. Positive commutator argument I: sketch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
8.2. Positive commutator argument II: construction of the commutant . . . . . 89
8.3. Positive commutator argument III: a priori estimate . . . . . . . . . . . . . . . . . . 92
8.4. Positive commutator argument IV: regularization . . . . . . . . . . . . . . . . . . . . . 92
8.5. Exercises ............................................................ 95
9. Propagation of singularities at radial points . . . . . . . . . . . . . . . . . . . . . . . . 95
9.1. Intermezzo: radial compactification of phase space . . . . . . . . . . . . . . . . . . . . 97
9.2. Radial point estimates: a simple example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
9.3. Radial point estimates: general setup ................................ 99
10. Asymptotic behavior of linear waves on de Sitter space . . . . . . . . . 105
10.1. Lorentzian geometry and wave operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
10.2. Waves on the static model of de Sitter space ........................ 107
10.3. Analysis of the spectral family P̂ (σ) ................................ 110
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

1. Introduction

Microlocal analysis is a paradigm for the study of distributions and their singularities.
Interesting distributions mostly arise in two ways:
 MICROLOCAL ANALYSIS 3

(1) as solutions of partial differential equations (PDE), and

(2) as integral kernels of operators used to localize, transform, or otherwise ‘test’ a
partial differential operator.
In these notes, we explicitly mostly focus on the first kind, and prove very general results
about solutions of linear PDE. The second kind will be present throughout, starting in §4,
though mostly not explicitly so.
Following a quick reminder on Schwartz functions and tempered distributions in §2, the
notes can be roughly divided into two parts. The first part (§§3–4) introduces pseudodif-
ferential operators (ps.d.o.s) on Rn and their basic properties. Consider for example the
Laplacian
n
X 1
∆= Dx2j , Dxj := ∂xj , (1.1)
i
j=1
which is a differential operator of order 2:
∆ ∈ Diff 2 (Rn ). (1.2)
Consider the operator L ∈ Diff 2 (Rn ) defined by
L := ∆ + 1. (1.3)
Then L : S (Rn ) → S (Rn ) is invertible (see Exercise 2.1); what kind of object is its inverse
L−1 ? Morally, it should be an operator of order −2, since composing it with L gives the
identity operator, which has order 0. And indeed, L−1 is a pseudodifferential operator of
order −2,
L−1 ∈ Ψ−2 (Rn ). (1.4)
By means of the Fourier transform and its inverse (see §2.1), we can write
Z Z
(L−1 u)(x) = (2π)−n ei(x−y)ξ (1 + |ξ|2 )−1 u(y) dy dξ (1.5)
Rn Rn

More generally, we shall define spaces of operators

Ψm (Rn ), m ∈ R, (1.6)
acting on Schwartz functions (and much larger function spaces too, such as tempered dis-
tributions), with Diff m (Rn ) ⊂ Ψm (Rn ) for m = 0, 1, 2, . . ., and forming a graded algebra:
0 0
Ψm (Rn ) ◦ Ψm (Rn ) ⊂ Ψm+m (Rn ). (1.7)
Roughly speaking, a typical element A ∈ Ψm (Rn )
is defined similarly to (1.5), but with
−1
(1 + |ξ| ) replaced by a more general symbol a(x, ξ) with |a(x, ξ)| . (1 + |ξ|2 )m/2 ; see §3
2

for the definition of symbols. In §4, we will define Ψm (Rn ) precisely, prove (1.7), as well
as the boundedness of ps.d.o.s on a variety of useful function spaces. We will also discuss
generalizations of (1.4) for elliptic (pseudo)differential operators. (Ellipticity is a notion
concerning only the principal symbol of A; the latter is, roughly speaking, the leading order
part of a, i.e. a modulo symbols of order m − 1, and ellipticity is the requirement that
the principal symbol be invertible.) In particular, we shall prove that on closed manifolds
(compact without boundary) M , every elliptic operator L ∈ Ψm (M ) is Fredholm as a map
on C ∞ (M ), or as a map L : H s (M ) → H s−m (M ) (s ∈ R); thus, we can solve the equation
Lu = f provided f satisfies a finite number of linear constraints, and then u is unique
modulo elements of the finite-dimensional space ker L.
4 PETER HINTZ

While there are many more interesting things one can say about linear elliptic operators
(index theory, Weyl’s law, degenerate or non-compact problems, etc.), we will switch gears
in the second part (§§6–8) of the notes and study non-elliptic phenomena. We begin in §6
by defining the wave front set of a distribution u ∈ S 0 (Rn ), which is a subset
WF(u) ⊂ T ∗ Rn \ o = Rn × (Rn \ {0}), (1.8)
conic in the second factor. (Here, o is the zero section of the cotangent bundle T ∗ Rn .)
Its projection onto Rn coincides with the singular support, sing supp u; roughly speaking,
WF(u) measures where and in what (co)directions u is singular. As a basic example, see
Exercise 6.2, the wave front set of the characteristic function 1Ω of a smooth domain Ω ⊂ Rn
is given by the conormal bundle of the boundary (minus the zero section)
WF(1Ω ) = N ∗ ∂Ω \ o. (1.9)

Elliptic regularity can then be microlocalized : if L ∈ Ψm (Rn ) has principal symbol `, and
if u ∈ S 0 (Rn ) is such that Lu is smooth, then WF(u) is contained in the characteristic set
Char(L) of L: roughly speaking, the set of those (x, ξ) where ` is not elliptic. For example,
the wave operator
X n
2
 = −Dt + Dx2j (1.10)
j=1

R1+n −σ 2 |ξ|2 = nj=1 ξj2 , where we write (σ, ξ) for

|ξ|2 ,
P
on has (principal) symbol ` =
t,x +
the momentum variables (dual under the Fourier transform) to (t, x). Thus,
Char() = {(t, x, σ, ξ) ∈ T ∗ R1+n \ o : σ 2 = |ξ|2 }. (1.11)
As a very concrete example, note that
u = H(t − x1 ) =⇒ u = 0, (1.12)
and indeed WF(u) ⊂ Char() by (1.9).
The theorem on the propagation of singularities, proved in §8, gives a complete descrip-
tion of the structure of WF(u) for u a distributional solution of an equation Lu = f ∈ C ∞ :
it states that WF(u) ⊂ Char(L) is invariant under the flow along the Hamiltonian vector
field of the principal symbol of L. In the case of , this flow, for time s ∈ R, maps (t, x, σ, ξ)
to (t − 2sσ, x + 2sξ, σ, ξ); use this to verify the theorem for (1.12)!
We shall prove this using the method of positive commutators, which showcases the
utility of ps.d.o.s as tools, rather than as interesting operators in their own right as in (1.4),
and exploits the link between symplectic geometry and ps.d.o.s (a form of the ‘classical–
quantum correspondence’). More importantly, this is a very flexible method, which allows
one to control solutions of PDE also in more degenerate situations—which arise frequently
in applications. We give one example concerning radial points in §9.
As an application which makes use of all these tools, we sketch the proof of resonance
expansions for solutions of linear wave equations on a spacetime of interest in General
Relativity (de Sitter space) in §10.

Acknowledgments. These notes draw material from Richard Melrose’s lecture notes
[Mel07], available under www-math.mit.edu/~rbm/iml90.pdf, the textbooks Microlocal
Analysis for Differential Operators: an Introduction by Grigis and Sjöstrand [GS94] and
Partial Differential Equations by Michael E. Taylor [Tay11], lecture notes by Jared Wunsch
 MICROLOCAL ANALYSIS 5

[Wun13], lecture notes by András Vasy [Vas18], as well as my own notes from lectures by
Rafe Mazzeo and András Vasy at Stanford University and Ingo Witt at the University of
Göttingen.
I am also grateful to the participants of my course 18.157: Introduction to Microlocal
Analysis at MIT in the spring semesters of 2019 and 2020 for many suggestions and cor-
rections. Special thanks go to Yonah Borns-Weil, Jesse Gell-Redman, and Ethan Sussman
for corrections and suggestions.

2. Schwartz functions and tempered distributions

Let k ∈ N = {1, 2, 3, . . .}. For an open set Ω ⊂ Rn , we denote by C k (Ω) the space of
k
T timesk continuously differentiable functions (with no growth restrictions), and C ∞ (Ω) =
k k
k∈N C (Ω). By Cb (Ω) ⊂ C (Ω) we denote the space of functions which are bounded,
together with their derivatives up to order k. We denote by Cck (Ω) the space of compactly
supported elements of C k (Ω). Unless otherwise noted, all functions will be complex-valued.
We use standard multiindex notation: for x = (x1 , . . . , xn ) ∈ Rn and α = (α1 , . . . , αn ) ∈
Nn0 , we set
n
Y α 1
xα := xj j , ∂xα := ∂xα11 · · · ∂xαnn , Dxα := Dxα11 · · · Dxαnn , D = ∂. (2.1)
i
j=1

When the context is clear, we shall often simply write Dα := Dxα , and Dj := Dxj . We also
put
X n Yn
|α| := αj , α! := αj !. (2.2)
j=1 j=1
We will moreover use the Japanese bracket, defined for x ∈ Rn by
hxi = (1 + |x|2 )1/2 . (2.3)
Definition 2.1. The space S (Rn ) of Schwartz functions consists of all φ ∈ C ∞ (Rn ) such
that for all k ∈ N0 ,
kφkk := sup |xα Dβ φ(x)| < ∞. (2.4)
x∈Rn
|α|+|β|≤k

Example 2.2. We have exp(−|x|2 ) ∈ S (Rn ). Moreover, we have a (continuous) inclusion

Cc∞ (Rn ) ,→ S (Rn ) with dense range. Recall that there are lots of smooth functions with
compact support; indeed, when K ⊂ U ⊂ Rn with K compact and U open and bounded,
there exists φ ∈ Cc∞ (Rn ) with φ ≡ 1 on K and φ ≡ 0 on Rn \ U .

Equipped with the countably many seminorms k · kk , S (Rn ) is a Fréchet space. Directly
from the definition, we have continuous maps
xj : S (Rn ) → S (Rn ) (φ 7→ xj φ),
(2.5)
Dj : S (R ) → S (R )
n n
(φ 7→ Dj φ).
Given a ∈ Cb∞ (Rn ), pointwise multiplication by a is also continuous. Furthermore, integra-
tion is a continuous map Z
: S (Rn ) → C. (2.6)
6 PETER HINTZ

Indeed, this follows from

Z  Z 
−n−1 n+1
  

 n φ(x) dx = 
   hxi hxi φ(x) dx ≤ Cn kφkn+1 . (2.7)
R Rn
Other useful operations are the pointwise product
S (Rn ) × S (Rn ) 3 (φ, ψ) → φψ ∈ S (Rn ), (φψ)(x) = φ(x)ψ(x), (2.8)
and the exterior product
S (Rn ) × S (Rn ) 3 (φ, ψ) → φ  ψ ∈ S (R2n ), (φ  ψ)(x, y) = φ(x)ψ(y). (2.9)
Definition 2.3. The space S 0 (Rn ) of tempered distributions is the space of all continuous
linear functionals u : S (Rn ) → C, equipped with the weak topology: the seminorms are
|u|φ := |u(φ)| for φ ∈ S (Rn ). We shall usually write hu, φi := u(φ).
Example 2.4. The δ-distribution is defined by hδ, φi := φ(0). We have δ ∈ S 0 (Rn ) since
|hδ, φi| ≤ kφk0 .

Combining (2.6) and (2.8), we can define a continuous map

Z
0
S (R ) 3 φ → Tφ ∈ S (R ), Tφ (ψ) =
n n
φ(x)ψ(x) dx. (2.10)
Rn
Proposition 2.5. The map φ 7→ Tφ is injective, and has dense range in the weak topology.

Proof. Regarding injectivity: Tφ (φ̄) = Rn |φ(x)|2 dx = 0 implies φ = 0. To prove the

R

density, it suffices to show that, given u ∈ S 0 (Rn ) and φ1 , . . . , φN ∈ S (Rn ) as well as

any  > 0, there exists φ ∈ S (Rn ) such that |hu, φj i − Tφ (φj )i| <  for all j = 1, . . . , N .
Assuming, as one may, that the φj are orthonormal with respect to the L2 (Rn ) inner
product, this holds (with ‘< ’ replaced by ‘= 0’) for φ = N
P
j=1 hu, φj iφ̄j . A better proof,
based on a mollification argument, is suggested in Exercise 2.2. 

Now on the one hand, we can extend the maps (2.5) by duality to S 0 (Rn ): indeed, for
u ∈ S 0 (Rn ) and φ ∈ S (Rn ), we define xj u, Dj u ∈ S 0 (Rn ) by
hxj u, φi := hu, xj φi, hDj u, φi := hu, −Dj φi. (2.11)
On the other hand, when u ∈ S (Rn ),
then Txj u (φ) = Tu (xj φ) and TDj u (φ) = Tu (−Dj φ),
i.e. on the image of S (R ) inside of S 0 (Rn ), the definitions 2.11 agree with the usual
n

definitions of multiplication and differentiation of Schwartz functions. The density state-

ment of Proposition 2.5 then shows that (2.11) defines the unique continuous extensions of
multiplication or differentiation from S (Rn ) to S 0 (Rn ).
Similarly, by duality and starting from (2.9), pointwise multiplication by a Schwartz
function extends in a unique manner to a continuous map on S 0 (Rn ); more generally, this
is true for multiplication by a function in Cb∞ (Rn ).
Other notions, which will be significantly refined later, are:
Definition 2.6. Let u ∈ S 0 (Rn ). Then the support, supp u, is the complement of the set
of x ∈ Rn such that there exists χ ∈ Cc∞ (Rn ), χ(x) 6= 0, such that χu = 0.
The singular support, sing supp u, is the complement of the set of x ∈ Rn such that there
exists χ ∈ Cc∞ (Rn ), χ(x) 6= 0, such that χu is smooth, i.e. χu = Tφ , φ ∈ S (Rn ).
 MICROLOCAL ANALYSIS 7

2
Example 2.7. We have supp δ = sing supp δ = {0}. For u = δ + e−|x| ∈ S 0 (Rn ), we have
supp u = Rn , but sing supp u = {0} still.

2.1. Fourier transform and its inverse. We define the Fourier transform of φ ∈ S (Rn )
by
Z
(Fφ)(ξ) = φ̂(ξ) := e−ix·ξ φ(x) dx, ξ ∈ Rn , (2.12)
Rn
and the inverse Fourier transform of ψ ∈ S (Rn ) by
Z
−1 −n
(F ψ)(x) := (2π) eix·ξ ψ(ξ) dξ, x ∈ Rn . (2.13)
Rn

As in (2.7), one finds kFφk0 ≤ CN kφkn+1 and kF −1 φk0 ≤ CN kφkn+1 . Moreover, we have
F(Dxj φ) = ξj Fφ, F(xj φ) = −Dξj Fφ,
−1 −1 −1 (2.14)
F (Dξj φ) = − xj F φ, F (ξj φ) = Dxj F −1 φ,
using integration by parts for the first and third statement; reading these from right to left
shows that
kFφkk ≤ CN kφkk+n+1 ∀ k ∈ N0 , (2.15)
hence the (inverse) Fourier transform preserves the Schwartz space:
F, F −1 : S (Rn ) → S (Rn ). (2.16)

Note then that for u, ψ ∈ S (Rn ),

Z Z Z
−ix·ξ
hFu, ψi = e u(x) dx ψ(ξ) dξ = e−ix·ξ u(x)ψ(ξ) dx dξ
Rn Rn Rn ×Rn (2.17)
= hu, Fψi.

This allows us to extend F, F −1 to maps on tempered distributions,

F, F −1 : S 0 (Rn ) → S 0 (Rn ), (2.18)
and the formulas (2.14) remain valid for φ ∈ S 0 (Rn ).

Example
R 2.8. The Fourier transform of δ is calculated by hFδ, ψi = hδ, Fψi = ψ̂(0) =
Rn ψ(x) dx, so Fδ = 1.

We recall the proof that F and F −1 are indeed inverses to each other.
Theorem 2.9. We have F ◦ F −1 = F −1 ◦ F = I on S (Rn ) and S 0 (Rn ).

Proof. Let A := F −1 F : S (Rn ) → S (Rn ). By (2.14), we have ADxj = F −1 ξj F = Dxj A

and Axj = F −1 (−Dξj )F = xj A, i.e. A commutes with differentiation along and multipli-
cation by coordinates. Given φ ∈ S (Rn ) and x0 ∈ Rn , we can write
n
X Z 1
φ(x) = φ(x0 ) + φj (x)(xj − (x0 )j ), φj (x) = (∂j φ)(x0 + t(x − x0 )) dt. (2.19)
j=1 0
8 PETER HINTZ

The fact that φj is in general not Schwartz is remedied by fixing a cutoff χ ∈ Cc∞ (Rn ),
identically 1 near x0 , and writing φ(x) = χ(x)φ(x) + (1 − χ(x))φ(x), so
n
X
φ(x) = χ(x)φ(x0 ) + φ̃j (x)(xj − (x0 )j ),
j=1 (2.20)
(1 − χ(x))φ(x)
φ̃j (x) = χ(x)φj (x) + (xj − (x0 )j ).
|x − x0 |2
Since A annihilates every term in the sum, we have (Aφ)(x0 ) = φ(x0 )(Aχ)(x0 ); note that
the constant (Aχ)(x0 ) here does not depend on φ, and not on the cutoff χ either (since the
left hand side does not involve χ at all).
The same cutoff χ can be used to evaluate Aφ at points x close to x0 ; but
Dxj (Aχ)(x) = A(Dxj χ)(x) = 0 (2.21)
for x ∈ χ−1 (1). We conclude that A = cI for some constant c ∈ C. One can find c by
explicitly evaluating
2 2 /4 2 /4 2
F(e−|x| )(ξ) = π n/2 e−|ξ| , F −1 (e−|ξ| )(x) = π n/2 e−|x| , (2.22)
so c = 1 indeed. The proof that FF −1 = I is completely analogous. 

We also recall that F is an isomorphism on L2 (Rn ); this follows from the density of
S (Rn ) in L2 (Rn ) and the following fact:
Proposition 2.10. For φ ∈ S (Rn ), we have
kFφkL2 (Rn ) = (2π)n/2 kφkL2 (Rn ) . (2.23)

Proof. Analogously to (2.17), we have

Z Z
(Fφ)(ξ)ψ̄(ξ) dξ = (2π) n
φ(x)F −1 ψ(x) dx, φ, ψ ∈ S (Rn ). (2.24)

Plugging in ψ = Fφ proves the proposition. 

2.2. Sobolev spaces and the Schwartz representation theorem. Using the Fourier
transform, we can define operators which differentiate a ‘fractional number of times’:
Definition 2.11. For s ∈ R (or s ∈ C), we let
hDis = (1 + |D|2 )s/2 : S 0 (Rn ) → S 0 (Rn ), hDis = F −1 hξis F. (2.25)

This agrees for s ∈ 2N0 with the usual definition, and for s = −2 with the operator (1.4).
What is implicitly used here is that multiplication by (1 + |ξ|2 )s is continuous on S (Rn ).
Definition 2.12. For s ∈ R, the Sobolev space of order s is defined by
H s (Rn ) := {u ∈ S 0 (Rn ) : hDis u ∈ L2 (Rn )}. (2.26)
With the norm
kukH s := khDis ukL2 = (2π)−n/2 khξis FukL2 , (2.27)
it is a Hilbert space.
Example 2.13. The δ-distribution at 0 ∈ Rn satisfies δ ∈ H s (Rn ) for all s < −n/2.
 MICROLOCAL ANALYSIS 9

Since multiplication by hxir is continuous on S 0 (Rn ) for any r ∈ R, we can more generally
define weighted Sobolev spaces,
hxir H s (Rn ) := {u ∈ S 0 (Rn ) : hxi−r u ∈ H s (Rn )}. (2.28)
These are Sobolev spaces with squared norm
kuk2hxir H s (Rn ) := khxi−r uk2H s (Rn ) . (2.29)

The second part of the following is (a version of) the Schwartz representation theorem:
Theorem 2.14. We have
\ [
S (Rn ) = hxir H s (Rn ), S 0 (Rn ) = hxir H s (Rn ). (2.30)
s,r∈R s,r∈R

Proof. See Exercises 2.3 and 2.5. 

It easily implies (using Sobolev embedding, Exercise 2.3) that every tempered distribution
is a sum of terms of the form xα Dβ a, a ∈ Cb0 (Rn ).

2.3. The Schwartz kernel theorem. The Schwartz kernel theorem is a philosophically
useful fact, establishing a 1–1 correspondence between the ‘most general’ operators in
the present context—mapping Schwartz functions to tempered distributions—and distri-
butional integral kernels, also called Schwartz kernels. To state this, we note that any
distribution K ∈ S 0 (Rn+m ) induces a bounded linear operator S (Rm ) → S 0 (Rn ) by
integration along the Rm factor, to wit
Z Z 
(OK φ)(ψ) := hK, ψ  φi = K(x, y)φ(y) dy ψ(x) dx, φ ∈ S (Rm ), ψ ∈ S (Rn ).
Rm
(2.31)
Formally, one usually writes
Z
(OK φ)(x) = K(x, y)φ(y) dy. (2.32)
Rm

Theorem 2.15. The map K 7→ OK is a bijection between S 0 (Rn+m ) and the space of
continuous linear operators S (Rm ) → S 0 (Rn ).

Proof. See Exercises 2.6 and 2.7. 

Example 2.16. The Schwartz kernel of the identity operator I on S (Rn ) is given by
K(x, y) = δ(x − y), x, y ∈ Rn . (2.33)

2.4. Differential operators. Given aα ∈ Cb∞ (Rn ) for α ∈ Nn0 , |α| ≤ m, we can define the
m-th order differential operator
X
A= aα (x)Dα . (2.34)
|α|≤m

Since multiplication by aα is continuous on S (Rn ), A defines a continuous linear operator

on S (Rn ). By duality, A extends (uniquely) to an continuous linear operator on S 0 (Rn ).
Definition 2.17. By Diff m (Rn ), we denote the space of all operators A : S (Rn ) → S (Rn )
of the form (2.34).
10 PETER HINTZ

Given A as in (2.34), let us define the full symbol of A to be

X
σ(A)(x, ξ) := aα (x)ξ α . (2.35)
|α|≤m

Then, in view of (2.14), we can write

Z
−n
(Au)(x) = (2π) eix·ξ σ(A)(x, ξ)û(ξ) dξ
Rn
Z Z  (2.36)
−n i(x−y)·ξ
= (2π) e σ(A)(x, ξ)u(y) dy dξ,
Rn Rn
which we read as an iterated integral. On the other hand, the Schwartz kernel K of A is
easily verified to be X
K(x, y) = aα (x)(Dα δ)(x − y), (2.37)
|α|≤m
so (formally) we have
Z
−n
K(x, y) = (2π) ei(x−y)·ξ a(x, ξ) dξ, (2.38)
Rn
which is indeed (rigorously) correct if one reads this as the Fourier transform of a in ξ.
Proposition 2.18. Let A ∈ Diff m (Rn ). Then A is local, that is,
supp Au ⊂ supp u, u ∈ S 0 (Rn ), (2.39)
and A is pseudolocal, that is,
sing supp Au ⊂ sing supp u, u ∈ S 0 (Rn ). (2.40)

The proof is straightforward. From the perspective of the Schwartz kernel K of A, (2.39)
is really due to the fact that K(x, y) is supported on the diagonal x = y, while (2.40) is
really due to the fact that K(x, y) is smooth away from x = y. (That is, adding to K an
element of S (R2n ) preserves (2.40), but destroys (2.39).) Since as microlocal analysts we
are interested in singularities, it is the property (2.40) which we care about most; and this
will persist when A is a pseudodifferential operator. On the other hand, the only continuous
linear operators A : S (Rn ) → S (Rn ) satisfying condition (2.39) are differential operators,
see Exercise 2.9.
We mention three features of differential operators concerning their principal symbol.
α
P
Definition 2.19. Given m ∈ N0 and a differential operator A = |α|≤m aα (x)D , its
principal symbol is defined as
X
σm (A)(x, ξ) := aα (x)ξ α , (2.41)
|α|=m

i.e. keeping only the top order terms.

Proposition 2.20. Let A ∈ Diff m (Rn ).
(1) Define the adjoint A∗ of A by Rn (A∗ u)(x)v(x) dx = Rn u(x)(Av)(x) dx, u, v ∈
R R
Cc∞ (Rn ). Then A∗ ∈ Diff m (Rn ), and the principal symbol is
σm (A∗ )(x, ξ) = σm (A)(x, ξ). (2.42)
 MICROLOCAL ANALYSIS 11

0 0
(2) Let B ∈ Diff m (Rn ). Then A ◦ B ∈ Diff m+m (Rn ), and
σm+m0 (A ◦ B)(x, ξ) = σm (A)(x, ξ)σm0 (B)(x, ξ). (2.43)
(3) Let κ : Rn → Rn be a diffeomorphism which is the identity outside of a compact
set. Define Aκ : S (Rn ) → S (Rn ) by (Aκ u)(y) := (A(u ◦ κ−1 ))(κ(y)). Then Aκ ∈
Diff m (Rn ), and the principal symbols are related via
σm (Aκ )(y, η) = σm (A) κ(y), (κ0 (y)T )−1 η .


(2.44)

Proof. Exercise 2.11. 

Thus, the principal symbol is well-defined as a function on T ∗ Rn , and is a homomorphism

from the (non-commutative) algebra Diff(R ) = m∈N0 Diff m (Rn ) into the commutative
n
S
algebra of functions a(x, ξ) which are homogeneous polynomials in ξ with coefficients in
Cb∞ (Rn ).

2.5. Exercises.
Pn 2
Exercise 2.1. Let ∆ = j=1 Dxj .

(1) Show that ∆ + 1 : S 0 (Rn ) → S 0 (Rn ) is an isomorphism.

(2) Find a non-trivial solution u ∈ C ∞ (Rn ) of (∆ + 1)u = 0. Why does this not
contradict the first part?
Exercise 2.2. We will prove in a constructive manner that Cc∞ (Rn ) ⊂ S 0 (Rn ) (or more
precisely
R the image of Cc∞ (Rn ) under the map (2.10)) is dense. Let χ ∈ Cc∞ (Rn ), with
Rn χ(x) = 1.

(1) Let φ ∈ S (Rn ). Put φ (x) = χ(x)φ(x). Show that φ → φ in S (Rn ) as  & 0.
Conclude that the space
E 0 (Rn ) := {u ∈ S 0 (Rn ) : supp u is compact} (2.45)
of compactly supported distributions is dense in S 0 (Rn ).
(2) Let ψ ∈ Cc∞ (Rn ) and put ψ̃(x) = ψ(−x). For φ ∈ S (Rn ), define the convolution of
φ with ψ̃ by
Z
(φ ∗ ψ̃)(x) := φ(x − y)ψ̃(y) dy. (2.46)
Rn

Show that φ ∗ ψ̃ ∈ S (Rn ).

Define the convolution of u ∈ S 0 (Rn ) with ψ by
hu ∗ ψ, φi := hu, φ ∗ ψ̃i. Check that this is the correct definition when u ∈ S (Rn ).
(3) Let χ (x) := −n χ(−1 x). Show that χ ∗ φ → φ in S (Rn ) as  & 0.
(4) When u ∈ E 0 (Rn ) and ψ ∈ Cc∞ (Rn ), show that u ∗ ψ ∈ Cc∞ (Rn ), that is, there
exists v ∈ Cc∞ (Rn ) so that u ∗ ψ = Tv . (Hint. Show that one can define hu ∗ ψ, φi
consistently for φ ∈ S 0 (Rn ); define a candidate for v by using δ-distributions for φ.
In order to show that u ∗ ψ = Tv , take any φ ∈ S (Rn ), apply both sides to χ ∗ φ
and let  → 0.)
(5) Combine the previous parts to conclude that Cc∞ (Rn ) ⊂ S 0 (Rn ) is dense.
Exercise 2.3. (Sobolev embedding.) Let s > n/2.
12 PETER HINTZ

(1) Prove that there exists a constant Cs < ∞ such that for φ ∈ S (Rn ), the estimate
kφkL∞ (Rn ) ≤ Cs kφkH s (Rn ) . (2.47)
holds. (Hint. Pass to the Fourier transform.) Deduce that H s (Rn ) ⊂ Cb0 (Rn ).
(2) Show more generally that H s (Rn ) ⊂ Cbk (Rn ) for s > n/2 + k.
(3) Prove the first equality in Theorem 2.14.
Exercise 2.4. Show that hxi−r H −m (Rn ) is the L2 -dual of hxir H m (Rn ). That is, show that
the sesquilinear pairing
Z
h−, −iL2 : S (R ) × S (R ) 3 (φ, ψ) 7→
n n
φ(x)ψ(x)dx (2.48)
Rn
extends by continuity and density to
h−, −iL2 : hxir H m (Rn ) × hxi−r H −m (Rn ) → C, (2.49)
and that the map

∗
hxir H m (Rn ) → hxi−r H −m (Rn ) , φ 7→ hφ, −iL2 , (2.50)
is an antilinear isomorphism.
Exercise 2.5. (Schwartz representation theorem.) Prove the second equality in Theo-
rem 2.14 as follows.
(1) Given u ∈ S 0 (Rn ), there exist C, k such that |u(φ)| ≤ Ckφkk .
(2) Let Rq = hxi−q hDi−q . Then Rq is an isomorphism on S (Rn ) and S 0 (Rn ). More-
over, for sufficiently large q, we have kRq φkk ≤ CkφkL2 (Rn ) (for some other constant
C). (Hint. Use the previous exercise. It may be convenient to take s there and q
here to be even integers.)
(3) Denoting Rq† = hDi−q hxi−q , deduce that Rq† u ∈ L2 (Rn ), and conclude that u ∈
hxiq H −q (Rn ).
Exercise 2.6. (Schwartz kernel theorem I.) Prove the injectivity claim of Theorem 2.15.
(Hint. Let K ∈ S 0 (Rn+m ) be given with OK = 0. Given φ ∈ S (Rn+m ), you need to show
that hK, φi = 0. You know that this is true when φ is a finite linear combination of exterior
products ψ1  ψ2 , ψ1 ∈ S (Rn ), ψ2 ∈ S (Rm ). Try to use the Fourier transform, or Fourier
series, to approximate φ by such linear combinations. It may help to first reduce to the
case that supp K is compact.)
Exercise 2.7. (Schwartz kernel theorem II.) Let A : S (Rn ) → S 0 (Rm ) be continuous. Prove
the surjectivity claim of Theorem 2.15 as follows.
(1) The continuity of A is equivalent to the statement that for all ψ ∈ S (Rn ) there
exists N > 1 such that |hAφ, ψi| ≤ N kφkN for all φ ∈ S (Rm ).
(2) There exist N, M ∈ R such that A extends by continuity to a bounded operator
A : hxi−M H M (Rm ) → hxiN H −N (Rn ). (2.51)
(Hint. An estimate from Exercise 2.5 will come in handy, in the form kψkk ≤
Ck kψkhxi−M H M (Rn ) for given k and sufficiently large M .)
(3) The operator
A0 := hDi−N −n/2−1 hxi−N AhDi−M −m/2−1 hxi−M (2.52)
is bounded from H −m/2−1 (Rm ) to Cb0 (Rn )
 MICROLOCAL ANALYSIS 13

(4) Evaluate A0 δy for y ∈ Rm and deduce that A0 has a Schwartz kernel K 0 ∈ Cb0 (Rn+m ).
(5) By relating the Schwartz kernels of A0 and A, prove that A = OK for some K ∈
S 0 (Rn+m ).
Exercise 2.8. Let A : S (Rn ) → S 0 (Rn ) be continuous, and denote by K ∈ S 0 (R2n ) its
Schwartz kernel. Show that K ∈ S (R2n ) if and only if A maps S (Rn ) → S (Rn ) and as
such moreover extends by continuity to a bounded map S 0 (Rn ) → S (Rn ).
Exercise 2.9 (Peetre’s Theorem). Let A : S (Rn ) → S (Rn ) be a continuous linear operator,
and suppose for all u ∈ S (Rn ), we have supp Au ⊂ supp u. Prove that A is a differential
operator. (Hint. Show that the Schwartz kernel K of A has support in the diagonal
{x = y}. Then show that it must be a locally finite linear combination of (differentiated)
δ-distributions with smooth coefficients. To prove that A is a differential operator of finite
order, exploit that K is a tempered distribution.)
Exercise 2.10. Show that the principal symbol σm (A) of A ∈ Diff m (Rn ) captures the ‘high
frequency behavior’ of A in the following sense: for x0 , ξ0 ∈ Rn , we have
σm (A)(x0 , ξ0 ) = lim λ−m (e−iλξ0 · Aeiλξ0 · )(x0 ), (2.53)
λ→∞

where eiξ0 · is the function x 7→ eiξ0 ·x .

Exercise 2.11. Prove Proposition 2.20.

3. Symbols

As a first step towards the definition of pseudodifferential operators, we generalize the

class of symbols a(x, ξ) from polynomials in ξ to more general functions:
Definition 3.1. Let m ∈ R, n, N ∈ N. Then the space of (uniform) symbols of order m
S m (Rn ; RN ) ⊂ C ∞ (Rn × RN ) (3.1)
consists of all functions a = a(x, ξ) which for all α ∈ Nn0 , β ∈ NN
0 satisfy the estimate

|∂xα ∂ξβ a(x, ξ)| ≤ Cαβ hξim−|β| . (3.2)

for some constants Cαβ . We also write
S m (RN ) := S m (R0 ; RN ) (3.3)
for symbols only depending on the symbolic variable ξ.

The gain of decay upon differentiation in the ξ-variables is often called symbolic behavior
(in ξ).
Remark 3.2. Sometimes these symbol classes are denoted S∞ m (Rn ; RN ), the subscript ‘∞’
∞
indicating the uniform boundedness in C of the ‘coefficients’, i.e. the x-variables. There
exist many generalizations and variants of the class S m (Rn ; RN ), such as: symbols of type
ρ, δ; symbols which in addition have symbolic behavior in x (these are symbols of scattering
(pseudo)differential operators); or symbols with joint symbolic behavior in (x, ξ) (symbols
of isotropic operators). See [Mel07, §4] and [Hör71, §1.1].
14 PETER HINTZ

Equipped with the norms given by the best constants in (3.2), or more concisely
kakm,k := sup max hξi−m+|β| |∂xα ∂ξβ a(x, ξ)|, (3.4)
(x,ξ)∈Rn ×RN |α|+|β|≤k

the space S m (Rn ; RN ) is a Fréchet space. Directly from the definition, we note that differ-
entiations
Dxα : S m (Rn ; RN ) → S m (Rn ; RN ),
(3.5)
Dξβ : S m (Rn ; RN ) → S m−|β| (Rn ; RN )
are continuous.
Example 3.3. Full symbols of differential operators of order m on Rn , see (2.35), lie in
S m (Rn ; Rn ). A special case of this is: given a ∈ Cb∞ (Rn ), the function (x, ξ) 7→ a(x) lies in
S 0 (Rn ; RN ) (for any N ).
Example 3.4. Let m ∈ R. Then hξim ∈ S m (Rn ; Rn ). (See Exercise 3.1.)
Proposition 3.5. Pointwise multiplication of symbols is a continuous bilinear map
0 0
S m (Rn ; RN ) × S m (Rn ; RN ) → S m+m (Rn ; RN ). (3.6)
0
Proof. This follows from the Leibniz rule: for a ∈ S m (Rn ; RN ), b ∈ S m (Rn ; RN ), and
α ∈ Nn0 , β ∈ NN
0 , we have
 X    
α β
 α β α0 β 0 α00 β 00 

|∂x ∂ξ (a · b)| = 
 (∂x ∂ ξ a)(∂x ∂ ξ b)
α0 β0
α0 +α00 =α

β 0 +β 00 =β
0 0 00 |
X
≤ Cα0 β 0 Cα00 β 00 hξim+m −|β |−|β
α0 +α00 =α
β 0 +β 00 =β
0
≤ Cαβ hξim+m −|β| . 

We note the trivial continuous inclusion

0
m ≤ m0 =⇒ S m (Rn ; RN ) ⊆ S m (Rn ; RN ), (3.7)
hence the S m (Rn ; RN ) give a filtration of the space of all symbols m∈R S m (Rn ; RN ). In
S
the other direction, we define the space of residual symbols by
\
S −∞ (Rn ; RN ) := S m (Rn ; RN ). (3.8)
m∈R
Equipped with the norms k · km,k , m, k ∈ N, this is again a Fréchet space.
Example 3.6. We have S (Rn × RN ) ⊂ S −∞ (Rn ; RN ), or more generally Cb∞ (Rn , S (RN )) ⊂
S −∞ (Rn ; RN ). Moreover, given a cutoff χ ∈ Cc∞ (RN ), its pullback along Rn ×RN 3 (x, ξ) 7→
ξ, i.e. (x, ξ) 7→ χ(ξ), is a residual symbol.

While the inclusion (3.7) never has dense range for m < m0 , there is a satisfying replace-
ment:
Proposition 3.7. Let m < m0 . Then S −∞ (Rn ; RN ) is a dense subspace of S m (Rn ; RN ) in
0
the topology of S m (Rn ; RN ). In fact, a stronger statement is true: for any a ∈ S m (Rn ; RN )
there exists a sequence aj ∈ S −∞ (Rn ; RN ) which is uniformly bounded in S m (Rn ; RN ) and
0
converges to a in the topology of S m (Rn ; RN ).
 MICROLOCAL ANALYSIS 15

Proof. Fix a cutoff function χ ∈ Cc∞ (RN ) ⊂ S −∞ (Rn ; RN ) (see Example 3.6) which is
identically 1 in |ξ| ≤ 1 and identically 0 when |ξ| ≥ 2. By Proposition 3.5, we have
aj (x, ξ) := a(x, ξ)χ(ξ/j) ∈ S −∞ (Rn ; RN ). (3.9)
To prove the proposition, it suffices to show, in view of Proposition 3.5, that
χj (ξ) := χ(ξ/j) (3.10)
is bounded in S 0 (RN ) and converges to 1 in the topology of S  (RN ) for all  > 0. Regarding
the former, we have |χj (ξ)| ≤ kχk0,0 for all j, while for |β| ≥ 1 we have ∂ξβ χj (ξ) ≡ 0 for
|ξ| ≤ 1, and
|ξ||β| ∂ξβ χj (ξ) = χβ (ξ/j), χβ (ξ) = |ξ||β| (∂ξβ χ)(ξ) ∈ Cc∞ (RN ). (3.11)
Regarding the latter, we note that supp(χj − 1) ⊂ {|ξ| ≥ j}, hence
|χ(ξ/j) − 1| ≤ j − hξi . (3.12)
For derivatives, we note that the support observation and (3.11) give
|ξ||β|− |∂ξβ (χj (ξ) − 1)| = |ξ||β|− |∂ξβ χj (ξ)| ≤ j − |χβ (ξ/j)|. (3.13)
Thus, kχj − 1k,k ≤ Ck j − → 0 as j → ∞, as desired. 

3.1. Ellipticity. We now generalize the key property of the symbol of the operator L =
∆ + 1 in (1.3).
Definition 3.8. Let m ∈ R. A symbol a ∈ S m (Rn ; RN ) is (uniformly) elliptic if there
exists a symbol b ∈ S −m (Rn ; RN ) such that ab − 1 ∈ S −1 (Rn ; RN ).
Proposition 3.9. Let m ∈ R, and a ∈ S m (Rn ; RN ). Then the following are equivalent:
(1) a is elliptic.
(2) There exist constants C, c > 0 such that
|ξ| ≥ C =⇒ |a(x, ξ)| ≥ c|ξ|m . (3.14)
(3) There exist constants C, c > 0 such that
|a(x, ξ)| ≥ c|ξ|m − C|ξ|m−1 , |ξ| ≥ 1. (3.15)

Proof. If a is elliptic, then in the notation of Definition 3.8, we have

1 − Chξi−1 ≤ |a(x, ξ)||b(x, ξ)| ≤ C|a(x, ξ)|hξi−m , (3.16)
for some constant C > 0, that is,
|a(x, ξ)| ≥ chξim − hξim−1 . (3.17)
This proves (3.15). This in turn implies (3.14) since for all c > 0, there exists C > 0 such
that |ξ|m−1 ≤ c|ξ|m for |ξ| ≥ C (indeed, this holds for C = c−1 ).
Conversely, if (3.14) holds, choose a cutoff χ ∈ C ∞ (Rn ), χ(ξ) = 0 for |ξ| ≤ 2C, χ(ξ) = 1
for |ξ| ≥ 3C, then (see Exercise 3.2)
b(x, ξ) := χ(ξ)/a(x, ξ) ∈ S −m (Rn ; RN ), (3.18)
and a(x, ξ)b(x, ξ) = χ(ξ) ∈ S −∞ (Rn ; RN ). 
16 PETER HINTZ

Note that if a ∈ S m (Rn ; RN ) is elliptic, then so is a + a0 for any a0 ∈ S m−1 (Rn ; RN ).

Thus, ellipticity is only a condition on the equivalence class
[a] ∈ S m (Rn ; RN )/S m−1 (Rn ; RN ). (3.19)
For full symbols of differential operators, we can identify [a] with the leading order, homo-
geneous of degree m, part of a. Compare with Definition 2.19 and Proposition 2.20.

3.2. Classical symbols. An important subclass of symbols mimics those of differential

operators: they are sums of homogeneous (in ξ) functions. More precisely, we call a function
a(x, ξ), defined for ξ 6= 0, (positively) homogeneous of order m ∈ C iff
a(x, λξ) = λm a(x, ξ), λ > 0. (3.20)
Definition 3.10. Let m ∈ C. Then m (Rn ; RN
Shom \ {0}) is the space of all functions
∞ n N
a(x, ξ) ∈ C (R × (R \ {0})), positively homogeneous of order m in ξ, such that for all
α, β ∈ Nn0
|∂xα ∂ξβ a(x, ξ)| ≤ Cαβ |ξ|Re m−|β| , ξ 6= 0. (3.21)
Definition 3.11. Let m ∈ C, and fix a cutoff χ ∈ Cc∞ (RN ) which is identically 1 near 0.
A symbol a ∈ S Re m (Rn ; RN ) is called a classical symbol of order m if there exist functions
m−j
am−j ∈ Shom (Rn × (RN \ {0})) such that for all J ∈ N, we have
J−1
X
a− (1 − χ)am−j ∈ S Re m−J (Rn ; RN ). (3.22)
j=0
m (Rn ; RN ). Finally, we put
The space of classical symbols of order m is denoted Scl
−∞
Scl (Rn ; RN ) := S −∞ (Rn ; RN ). (3.23)

Equipped with the seminorms of am−j and the remainders a − J−1

P
j=0 (1 − χ)am−j in the
m n N
respective spaces, Scl (R ; R ) is a Fréchet space. Proposition 3.7 fails dramatically for
classical symbols; indeed (Exercise 3.4),
S −∞ (Rn ; RN ) ⊂ Scl
m
(Rn ; RN ) is closed for any m ∈ C. (3.24)
We have the following straightforward lemma (Exercise 3.5):
Lemma 3.12. The homogeneous terms am−j in (3.22) are uniquely determined by a.

For a ∈ Sclm (Rn ; RN ) as in Definition 3.11, we can thus identify the equivalence class

[a] ∈ S Re m (Rn ; RN )/S Re m−1 (Rn ; RN ) with the leading order homogeneous part am , or
even more simply with the function Rn × SN −1 3 (x, ξ) 7→ am (x, ξ), where SN −1 = {ξ ∈
RN : |ξ| = 1} is the unit sphere. Cf. (2.41).

3.3. Asymptotic summation. There is a (general) ‘converse’ to (3.22) which is very

useful when performing iterative constructions which yield lower order corrections:
Proposition 3.13. Let aj ∈ S mj (Rn ; RN ), j ≥ 0, and suppose lim supj→∞ mj = −∞. Let
m̄j := supj 0 ≥j mj 0 , and m = m̄0 . Then there exists a symbol a ∈ S m (Rn ; RN ) such that for
all J ∈ N
J−1
X
a− aj ∈ S m̄J (Rn ; RN ). (3.25)
j=0
 MICROLOCAL ANALYSIS 17

Moreover, a is unique modulo S −∞ (Rn ; RN ).

We call a ‘the’ asymptotic sum of the aj , and write

∞
X
a∼ aj . (3.26)
j=0

Proof of Proposition 3.13. This is similar to Borel’s theorem concerning the existence of
a smooth function with prescribed Taylor series at 0. Uniqueness is clear, since any two
asymptotic sums a, a0 satisfy a−a0 ∈ S m̄J (Rn ; RN ), with m̄J → −∞, hence a−a0 is residual
indeed.
For existence, we may partially sum finitely many of the aj and thereby reduce to the
case that aj ∈ S m−j (Rn ; RN ), j ≥ 0, and m̄j = m − j. Fix a cutoff χ ∈ C ∞ (Rn ), identically
0 in |ξ| ≤ 1 and equal to 1 for |ξ| ≥ 2. With j > 0, j → 0, to be determined, we wish to
set
∞
X
a(x, ξ) := χ(j ξ)aj (x, ξ). (3.27)
j=0

This sum is locally finite, hence a ∈ C ∞ (Rn

× RN ). Choosing j more precisely, we can
arrange that
kχ(j ·)aj km−j 0 ,j 0 ≤ 2−j , j > j 0 ≥ 0. (3.28)
Indeed, for fixed j, j 0 , we can choose j > 0 such that this holds since χ(j ·)aj → 0 in
0
S m−j (Rn ; RN ) as j → 0, as in the proof of Proposition 3.7; but for any fixed j, (3.28)
gives a finite number of conditions on j , one for each 0 ≤ j 0 < j.
But then χ(j 0 ξ)aj 0 (x, ξ)+ ∞ m−j 0 (Rn ; RN ). Thus, the
P
j=j 0 +1 χ(j ξ)aj (x, ξ) converges in S
sequence (3.27) converges in S m (Rn ; RN ), and we have
J−1
X J−1
X ∞
X
a(x, ξ)− aj (x, ξ) = (1−χ(j ξ))aj (x, ξ)+ χ(j ξ)aj (x, ξ) ∈ S m−J (Rn ; RN ), (3.29)
j=0 j=0 j=J

as desired. 

The space Scl m (Rn ; RN ) can be characterized as the space of symbols in S Re m (Rn ; RN )

which are asymptotic sums of symbols which in |ξ| ≥ 1 are positively homogeneous of degree
m − j, j ∈ N0 .
For completeness and later use, we refine the previous result to ensure the continuous
dependence of a on the sequence (aj ).

Proposition 3.14. Let `S m (Rn ; RN ) = ∞ m−j (Rn ; RN ) be the space of all sequences
Q
j=0 S
(a0 , a1 , . . .) of symbols aj ∈ S m−j (R ; R ). Equip `S m with the topology generated by the
n N

seminorms k(aj )kJ := max1≤k≤J kak km−k,J . Then there exists a continuous (nonlinear)
map
X
: `S m → S m (Rn ; RN ) (3.30)
A
with the property that A ((aj )j∈N0 ) ∼ ∞
P P
j=0 aj .
18 PETER HINTZ

Remark 3.15. The topology on `S m (Rn ; RN ) is akin to e.g. the standard topology on
C ∞ (Rn ) which is given by seminorms k · kC k (B(0,k)) . To verify convergence of a sequence of
sequences of symbols in this topology, one merely needs to check that for any fixed J ∈ N,
the first J terms of the sequence converge in the respective symbol spaces.

Proof of Proposition 3.14. Fix χ ∈ C ∞ (RN ), χ(ξ) = 0 for |ξ| ≤ 1 and χ(ξ) = 1 for |ξ| ≥ 2.
As in the previous proof, we shall set, for a = (aj )j∈N0 ∈ `S m (Rn ; RN ),
X  X∞
a (x, ξ) := χ(j (a)ξ)aj (x, ξ), (3.31)
A
j=0

where j (a), as in (3.28), is chosen so that for all j ∈ N

max kχ(j (a)ξ)aj (x, ξ)km−j 0 ,j 0 ≤ 2−j , (3.32)
0≤j 0 ≤j−1

and we set 0 (a) = 1. We now need to make a concrete choice of j (a): to this effect, we
note that for |α| + |β| ≤ j 0 ≤ j − 1,

0 
 0
hξi−m+j ∂xα ∂ξβ χ(j (a)ξ)aj (x, ξ)  ≤ Cj hξi−m+j hξim−j kaj km−j,j 0 1|ξ|≥j (a)−1
(3.33)
≤ Cj j (a)kaj km−j,j ,
where Cj only depends on χ (and j of course). Therefore (3.32) holds provided we take

−1
j (a) := 2−j 1 + Cj kaj km−j,j . (3.34)
P P P∞
With this choice, A a is well-defined, and A a ∼ j=0 aj .
We now check continuity. Define
χj (q, ξ) := χ(2−j (1 + Cj q)−1 ξ). (3.35)
Fix a = (aj )j∈N0 ∈ `S m (Rn ; RN ), and fix k ∈ N0 ,  > 0. We need to show that there exist
δ > 0 and J ∈ N such that
X X
a0 ∈ `S m (Rn ; RN ), ka − a0 kJ ≤ δ =⇒ a− a0 < , (3.36)

A A m,k
which holds provided
X∞
χj (kaj km−j,j , ξ)aj − χj (ka0j km−j,j , ξ)a0j

m,k
< . (3.37)
j=0

The j-th summand can individually be estimated by

kχj (kaj km−j,j , ξ)(aj − a0j )km,k + χj (kaj km−j,j , ξ) − χj (ka0j km−j,j , ξ) a0j m,k

≤ Cj kaj − a0j km,k + kaj km−j,j − ka0j km−j,j  ka0j km,k

 
(3.38)
≤ Cj + kaj km,k + ka − a0 kmax(j,k) ka − a0 kmax(j,k) ,

which tends to zero as a0 → a in `S m (Rn ; RN ).

The tail of the sum (3.37) on the other hand is estimated simply using (3.32)
kχj (kaj km−j,j , ξ)aj km,k + kχj (ka0j km−j,j , ξ)a0j km,k ≤ 2−j + 2−j = 2−j+1 (3.39)
P∞
provided j > k. Thus, we first choose J0 ∈ N, J0 > k, such that j=J0 2−j+1 < /2,
and then δ > 0, J ∈ N such that a0 ∈ `S m (Rn ; RN ), ka − a0 kJ < δ implies that the j-th
 MICROLOCAL ANALYSIS 19

summand in (3.37) is bounded by /(2J0 ) for j = 0, . . . , J0 − 1. This achieves (3.36) and

thus finishes the proof. 

3.4. Exercises.
Exercise 3.1 (Symbols and classical symbols). (1) Let m ∈ R. Prove hξim ∈ S m (RN ).
By expanding into Taylor series in 1/|ξ|, show that indeed hξim ∈ Scl
m (RN ).
µ µ N
(2) More generally, let µ ∈ C. Show that hξi ∈ Scl (R ).
Exercise 3.2 (Inverses of elliptic symbols). (1) Show that if a ∈ S m (Rn ; RN ) satisfies (3.14),
and χ ∈ S (R ) vanishes for |ξ| ≤ 2C, then χ/a ∈ S −m (Rn ; RN ).
0 N

(2) If in addition a and χ are classical symbols, show that χ/a is classical as well.
Exercise 3.3 (Compositions of functions with symbols). (1) Let f ∈ C ∞ (R). Show that
0 n N 0 n
if a ∈ S (R ; R ), then also f ◦ a ∈ S (R ; R ). n

(2) Show that if a ∈ S 0 (Rn ; RN ) is elliptic and positive, then there exists b ∈ S 0 (Rn ; RN )
such that a − b2 ∈ S −1 (Rn ; RN ).
Exercise 3.4. Prove (3.24).
Exercise 3.5. Prove Lemma 3.12. (Hint. Use induction on j; the case j = 0 is the main
content.)

4. Pseudodifferential operators

For developing the theory of ps.d.o.s, it is useful to consider slightly more general symbols,
in the class
hx − yiw S m (Rnx × Rny ; Rnξ ) = {hx − yiw ã : ã ∈ S m (Rn × Rn ; Rn )}, (4.1)
where w ∈ R. Our immediate goal will be to make sense of the following definition.
Definition 4.1. Let m, w ∈ R, and a ∈ hx − yiw S m (Rnx × Rny ; Rnξ ). Then we define its
quantization Op(a) by
Z Z
−n
(Op(a)u)(x) := (2π) ei(x−y)·ξ a(x, y, ξ)u(y) dy dξ, u ∈ S (Rn ). (4.2)
Rn Rn

Previously, see (2.36), we only considered the special case of the left quantization of a
left symbol a ∈ S m (Rnx ; Rnξ ), independent of y:
Z Z
−n
(OpL (a)u)(x) = (2π) ei(x−y)·ξ a(x, ξ)u(y) dy dξ; (4.3)
Rn Rn
this immediately makes sense as an iterated integral for u ∈ S (Rn ), and should be thought
of as ‘differentiate first, then multiply by coefficients’. Dually, we can consider the right
quantization of a right symbol a ∈ S m (Rny ; Rnξ ),
Z Z
(OpR (a)u)(x) = (2π)−n ei(x−y)·ξ a(y, ξ)u(y) dy dξ, (4.4)
Rn Rn
which does not immediately make sense (similarly to (4.2)); this should be thought of as
‘multiply by coefficients, then differentiate’. (Take a(z, ξ) = ξ α aα (z) with aα ∈ Cb∞ (Rn )
and evaluate OpL (a)u and OpR (a)u!)
20 PETER HINTZ

The quantization map (4.2) should be read as ‘multiply (y), then differentiate (ξ), then
multiply (x)’. (Try this with a(x, y, ξ) = a1 (x)ξ α a2 (y).) We shall see below that every
operator Op(a) can be written as Op(a) = OpL (aL ) = OpR (aR ) for suitable left and right
symbols aL and aR of the same order as a, see §4.1. (You have done most of the work
for proving this for differential operators, i.e. in the case that a is a polynomial in ξ, in
Exercise 2.11.)
Lemma 4.2. Let w ∈ R, m < −n, and let a = hx − yiw ã, ã ∈ S m (Rn × Rn ; Rn ). Then the
integral (4.2) is absolutely convergent and defines a continuous operator
Op(a) : S (Rn ) → hxiw Cb0 (Rn ). (4.5)
More precisely, for N > n + |w|, there exists a constant C < ∞ such that
k Op(a)ukhxiw C 0 (Rn ) ≤ Ckãkm,0 kukN , u ∈ S (Rn ). (4.6)
b

For the proof, we need a simple lemma:

Lemma 4.3. Let w ∈ R. Then hx + yiw ≤ 2|w|/2 hxiw hyi|w| .

Proof. By the triangle and Cauchy–Schwarz inequalities, we have

1 + |x + y|2 ≤ 1 + 2|x|2 + 2|y|2 ≤ 2(1 + |x|2 )(1 + |y|2 ). (4.7)
If w > 0, then taking this to the power w/2 proves the lemma. For w = 0, the lemma is
the equality 1 = 1. For w < 0, hence −w > 0, we obtain, analogously to (4.7),
hxi−w ≤ 2−w/2 hx + yi−w hyi−w , (4.8)
which upon multiplication by hxiw hx + yiw gives the desired result. 

Proof of Lemma 4.2. Since u is Schwartz, we have |u(y)| ≤ CN kukN hyi−N for all N ∈ N0 .
Therefore, the integrand in (4.2) satisfies
|ei(x−y)·ξ a(x, y, ξ)u(y)| ≤ Chx − yiw kãkm,0 hξim · kukN hyi−N
(4.9)
≤ Chxiw · hξim hyi|w|−N · kãkm,0 kukN .
This is integrable in (y, ξ) provided m < −n and |w| − N < −n, proving the lemma. 
Proposition 4.4. Let w ∈ R and a = hx − ã ∈ yiw ã, × S −∞ (Rn Rn ; Rn ).
Then the
quantization Op(a) : S (R ) → S (R ) is continuous. In fact, for all k ∈ N0 , m ∈ R, there
n n

exists N ∈ N and a constant C such that

k Op(a)ukk ≤ Ckãkm,N kukN . (4.10)
Lemma 4.5. Differentiations Dxα and Dyα are continuous maps hx−yiw S m (Rn ×Rn ; Rn ) →
hx − yiw S m (Rn × Rn ; Rn ). More precisely,
khx − yi−w Dxα akm,k ≤ Ckhx − yi−w akm,k+|α| , (4.11)
likewise for Dyα a.

Proof. It suffices to prove the claim for Dx1 . For a(x, y, ξ) = hx−yiw ã(x, y, ξ), ã ∈ S m (Rn ×
Rn ; Rn ), we have
∂x1 a = hx − yiw (∂x1 ã) + whx − yiw−2 (x1 − y1 )ã. (4.12)
The first summand lies in hx − yiw S m (Rn × Rn ; Rn ), and the second summand even lies in
the smaller space hx − yiw−1 S m (Rn × Rn ; Rn ). 
 MICROLOCAL ANALYSIS 21

Proof of Proposition 4.4. The key is that for ξ 6= 0, the phase (x − y) · ξ has no critical
points in y. We exploit this by writing
(1 − ξ · Dy )ei(x−y)·ξ = hξi2 ei(x−y)·ξ , (4.13)
so upon integrating by parts in y, one gains decay in ξ. Concretely, for N ∈ N, we have
Z Z
−n
Op(a)u(x) = (2π) ((1 − ξ · Dy )N ei(x−y)·ξ )hξi−2N a(x, y, ξ)u(y) dy dξ
n n
ZR ZR (4.14)
−n
ei(x−y)·ξ (1 + ξ · Dy )N hξi−2N a(x, y, ξ)u(y) dy dξ.


= (2π)
Rn Rn
By the Leibniz rule, we have
X
(1 + ξ · Dy )N hξi−2N a(x, y, ξ)u(y) = aγ (x, y, ξ) · Dyγ u,


(4.15)
|γ|≤N

where X
aγ (x, y, ξ) = cγδ hξi−2N ξ δ Dy a(x, y, ξ) (4.16)
|δ|,||≤N

for some combinatorial constants cγδ . By Lemma 4.5, ãγ := hx − yi−w aγ ∈ S −∞ (Rn ×
Rn ; Rn ), and for any m ∈ R,
kãγ km−N,0 ≤ Ckãkm,2N . (4.17)
Thus, if N > m + n, Lemma 4.2 applies, giving
k Op(aγ )Dγ ukhxiw C 0 (Rn ) ≤ Ckãγ km−N,0 kDγ ukM , M > n + |w|, (4.18)
and therefore
k Op(a)ukhxiw C 0 (Rn ) ≤ Ckãkm,2N kukM , M > n + N + |w|. (4.19)

To get higher regularity and decay, let now α, β ∈ Nn0 , then

Z Z
α β −n
(Dξ + y)α ei(x−y)·ξ (ξ + Dx )β a(x, y, ξ)u(y) dy dξ


x Dx Op(a)u(x) = (2π)
n n
ZR ZR
= (2π)−n ei(x−y)·ξ (y − Dξ )α (ξ + Dx )β a(x, y, ξ)u(y) dy dξ.


Rn Rn
(4.20)
This can be expanded using the Leibniz rule; note that powers of y are acceptable since u
is Schwartz. We thus obtain
kxα Dxβ Op(a)ukhxiw C 0 (Rn ) ≤ Ckãkm,N kukN (4.21)
for N sufficiently large (depending on m, n, α, β). Thus, Op(a)u ∈ S (Rn ), finishing the
proof. 

This shows that the map

hx − yiw S −∞ (Rn × Rn ; Rn ) × S (Rn ) 3 (a, u) 7→ Op(a)u ∈ S (Rn ) (4.22)
m0
is a continuous bilinear map when putting the topology of hx − yiw S (Rn × Rn ; Rn ) on
the first factor (for any m0 ∈ R). By Proposition 3.7, it thus extends by continuity to a
continuous bilinear map
hx − yiw S m (Rn × Rn ; Rn ) × S (Rn ) 3 (a, u) 7→ Op(a)u ∈ S (Rn ). (4.23)
22 PETER HINTZ

Identifying Op(a) with its Schwartz kernel, we thus get a continuous map

Op : hx − yiw S m (Rn × Rn ; Rn ) → S 0 (Rn × Rn ), (4.24)

which is given (interpreted as a limit along a sequence of residual symbols) by

Z
−n
Op(a)(x, y) = (2π) ei(x−y)·ξ a(x, y, ξ) dξ. (4.25)
Rn

(This is of course much weaker than (4.23).)

Remark 4.6. Let χ ∈ Cc∞ (Rn ) be identically 1 near 0. Given a ∈ hx − yiw S m (Rn × Rn ; Rn ),
(the proof of) Proposition 3.7 implies that
Z Z
−n
Op(a)u(x) = lim (2π) ei(x−y)·ξ χ(ξ/j)a(x, y, ξ)u(y) dy dξ, (4.26)
j→∞ Rn Rn

with convergence in S (Rn ).

Definition 4.7. Let m ∈ R. The space of pseudodifferential operators of order m,

Ψm (Rn ), (4.27)

is the space of all operators of the form Op(a) : S (Rn ) → S (Rn ), where a ∈ hx −
yiw S m (Rn × Rn ; Rn ) and w ∈ R. (As we show in the next section, one can take w = 0. See
Exercise 4.1 for the case of differential operators.) We set
\
Ψ−∞ (Rn ) := Ψm (Rn ). (4.28)
m∈Rn

Note that a priori it is not clear that Ψ−∞ (Rn ) is equal to the space of quantizations of
residual symbols (it is certainly contained in the latter); we show this in Proposition 4.10
below.
By duality, we can define the action of A = Op(a) ∈ Ψm (Rn ) on tempered distributions:
for u, v ∈ S (Rn ) and a ∈ hx − yiw S −∞ (Rn × Rn ; Rn ), we have
ZZZ
−n
hOp(a)u, vi = (2π) ei(x−y)·ξ a(x, y, ξ)u(y)v(x) dy dξ dx
R3n
ZZZ
= (2π)−n ei(x−y)·ξ a(y, x, −ξ)v(y)u(x) dy dξ dx (4.29)
R3n
†
= hu, Op(a )vi,

where we put
a† (x, y, ξ) = a(y, x, −ξ). (4.30)
Since a 7→ a† is an isomorphism on hx − yiw S m (Rn × Rn ; Rn ), the equality

Op(a)† = Op(a† ), that is, hOp(a)u, vi = hu, Op(a† )vi (4.31)

continues to hold for a ∈ hx − yiw S m (Rn × Rn ; Rn ). By the density S (Rn ) ⊂ S 0 (Rn ), we

can thus uniquely extend, by continuity, Op(a) to an operator on S 0 (Rn ) via (4.31).
 MICROLOCAL ANALYSIS 23

4.1. Left/right reduction, adjoints. In this section, we shall prove:

Theorem 4.8. Let a ∈ hx − yiw S m (Rn × Rn ; Rn ). Then there exists a unique left symbol
aL ∈ S m (Rn ; Rn ) such that

Op(a) = Op(aL ) = OpL (aL ), (4.32)

and a unique right symbol aR ∈ S m (Rn ; Rn ) such that

Op(a) = Op(aR ) = OpR (aR ). (4.33)

The symbols aL , aR depend continuously on a. Modulo residual symbols, they are given by
asymptotic sums
X 1
∂ξα Dyα a(x, y, ξ) |y=x ,


aL (x, ξ) ∼ (4.34)
n
α!
α∈N0
X (−1)|α|
∂ξα Dxα a(x, y, ξ) |x=y .


aR (y, ξ) ∼ (4.35)
n
α!
α∈N0

(The summands are ordered by increasing |α|.)

Definition 4.9. In the notation of Theorem 4.8, we call aL , resp. aR the left, resp. right
reduction of the full symbol a. Writing A = Op(a), we write

aL =: σL (A), aR =: σR (A). (4.36)

We first consider the case ‘m = −∞’ of Theorem 4.8 and give a description of kernels of
residual operators, i.e. elements of Ψ−∞ (Rn ):

Proposition 4.10. An operator A : S (Rn ) → S 0 (Rn ) is a residual operator if and only if

its Schwartz kernel K(x, y) is smooth and satisfies

|∂xα ∂yβ K(x, y)| ≤ CαβN hx − yi−N ∀ α, β, N. (4.37)

Moreover, any such A can be written as A = OpL (aL ) = OpR (aR ) for unique symbols
aL , aR ∈ S −∞ (Rn ; Rn ).

Proof. Since A ∈ Ψ−N (Rn ) for all N ∈ R, we can write A = Op(aN ) with aN = hx−yiwN ãN ,
ãN ∈ S −N (Rn × Rn ; Rn ), for some wN ∈ R. Taking N > n, the Schwartz kernel K of A is
then given by the absolutely convergent integral
Z
−n
K(x, y) = (2π) ei(x−y)·ξ aN (x, y, ξ) dξ. (4.38)
Rn

Let M ∈ N0 . For |x − y| < 1, and α, β with |α| + |β| ≤ M , and taking N > n + M , we can
thus bound
|∂xα ∂yβ K(x, y)| ≤ Cαβ kãN k−N,M (4.39)
using the triangle inequality. This gives (4.37) in this region.
24 PETER HINTZ

x−y i(x−y)·ξ = ei(x−y)·ξ and repeated

For |x − y| > 1, we fix N = n + 1. We use ( |x−y| 2 · Dξ )e

integration by parts to deduce that

Z  x−y M 
−n  i(x−y)·ξ
 
|K(x, y)| = (2π)  e − 2
· Dξ an+1 (x, y, ξ) dξ 
Rn |x − y|
Z
−M (4.40)
≤ CM |x − y| hx − yi wn+1
hξi−n−1−M dξ
Rn
0 −M +wn+1
≤ CM hx − yi .

Since M is arbitrary, this proves (4.37) for α = β = 0. Up to k-fold derivatives in x, y are

estimated in the same way, but now working with an+1+k instead of an+1 .
For the converse, note that if K satisfies (4.37), we can define
Z
aL (x, ξ) = e−iz·ξ K(x, x − z) dz. (4.41)
Rn

Then Op(aL ) = K by the Fourier inversion formula, and the estimates (4.37) imply aL ∈
S −∞ (Rn ; Rn ). Similarly, we have K = Op(aR ) for
Z
aR (y, ξ) = e−iz·ξ K(y + z, y) dz. (4.42)
Rn


Remark 4.11. Define seminorms on the space of all K ∈ C ∞ (Rnx × Rny ) satisfying the es-
timates (4.37) to be the optimal constants: |K|αβN := supx,y∈Rn hx − yiN |∂xα ∂yβ K(x, y)|.
Then the proof of Proposition 4.10 shows that the maps K 7→ aL/R ∈ S −∞ (Rn ; Rn ) and
S −∞ (Rn ; Rn ) 3 a 7→ K = OpL/R (a) are continuous.

To handle the case of general orders m ∈ R, we first note that integration by parts in ξ
implies the equality of Schwartz kernels
Z
α −n
Op((y − x) a)(x, y) = (2π) ((−Dξ )α ei(x−y)·ξ )a(x, y, ξ) dξ
Z
(4.43)
= (2π)−n ei(x−y)·ξ Dξα a(x, y, ξ) dξ
= Op(Dξα a)(x, y),

first for a ∈ hx − yiw S −∞ (Rn × Rn ; Rn ), and then for symbols of order m by density and
continuity. The additional off-diagonal growth of (y − x)α a is the reason for working with
the more general symbol class (4.1).

Proof of Theorem 4.8. Let N ∈ N, then Taylor’s formula states

X 1
(y − x)α ∂yα a(x, y, ξ) |y=x + rN (x, y, ξ),


a(x, y, ξ) =
α!
|α|<N
X N Z 1 (4.44)
rN (x, y, ξ) = (y − x) α
(1 − t)N −1 (∂yα a)(x, x + t(y − x), ξ) dt.
α! 0
|α|=N
 MICROLOCAL ANALYSIS 25

Using the identity (4.43), we have

 
X 1
Op a − (Dα ∂ α a)|y=x  = Op(r̃N ) ∈ Ψm−N (Rn ), (4.45)
α! ξ y
|α|<N

where
X NZ 1
r̃N (x, y, ξ) = (1 − t)N −1 (Dξα ∂yα a)(x, x + t(y − x), ξ) dt. (4.46)
α! 0
|α|=N
In view of the symbolic estimates for a, the remainder here satisfies the estimate
|∂xβ ∂yγ ∂ξδ r̃N (x, y, ξ)| ≤ CβγδN hx − yiw hξim−N −|δ| , (4.47)
hence
r̃N ∈ hx − yiw S m−N (Rn × Rn ; Rn ). (4.48)
α α
for all N . Note that for |α| = k, we have Dξ ∂y a|y=x ∈ S m−k n n
(Rx ; Rξ ). Thus, we can let
m n n
b ∈ S (R ; R ) be an asymptotic sum
X 1
b∼ (Dξα ∂yα a)|y=x , (4.49)
α
α!
and then \
R := Op(a − b) ∈ Ψm−N (Rn ) = Ψ−∞ (Rn ). (4.50)
N ∈N
By Proposition 4.10, we then have R = OpL (r) for some r ∈ S −∞ (Rn ; Rn ). Therefore,
A = OpL (aL ), aL := b + r. (4.51)
The continuous dependence of aL on a follows by using the explicit asymptotic summation
procedure of Proposition 3.14 to define b, which thus depends continuously on a, and then
noting that the optimal constants for the Schwartz kernel K of R in (4.37), and thus the
S −∞ (Rn ; RN ) seminorms of r (see Remark 4.11), depend continuously on a, b.
Reduction to a right symbol is proved analogously. Instead of going through the argu-
ment, one can instead use duality as in (4.29), the idea being that the adjoint of a left
quantization is a right quantization (and vice versa). Namely, using (4.30), we write the
adjoint of Op(a) as Op(a)† = Op(a† ) = Op(a0L ) for a0L ∈ S m (Rn ; Rn ), and then
Op(a) = Op(a† )† = (Op(a0L ))† = Op((a0L )† ) = OpR (aR ), (4.52)
where aR (y, ξ) = a0L (y, −ξ). The formula for left reductions gives
X 1
a0L (x, ξ) ∼ ((−∂ξ )α Dxα a)(y, x, −ξ)|y=x , (4.53)
α
α!
yielding the asymptotic description (4.35) of aR .
It remains to prove the uniqueness of aL , aR . For this, note that a left symbol aL can be
viewed as an element aL ∈ C ∞ (Rnx ; S 0 (Rnξ )), and the Schwartz kernel of Op(aL ) is
Op(aL )(x, x − z) = (F2−1 aL )(x, z). (4.54)
Since F2 is an isomorphism of C ∞ (Rn ; S 0 (Rn )), Op(aL ) = 0 implies aL = 0. The proof for
aR is similar. 
Corollary 4.12. Let m ∈ R or m = −∞. Then Ψm (Rn ) = OpL/R (S m (Rn ; Rn )).
26 PETER HINTZ

A slight variant of (4.29) gives the first part of the following corollary; the second part
is an immediate application of Theorem 4.8.
Corollary 4.13. Let A ∈ Ψm (Rn ), then
Z Z
∗
(A u)(x)v(x) dx = u(x)(Av)(x) dx, u, v ∈ S (Rn ). (4.55)
Rn Rn

defines an operator A∗ ∈ Ψm (Rn ). If A = Op(a), then A∗ = Op(a∗ ), a∗ (x, y, ξ) = ā(y, x, ξ).

If A = OpL (aL ), then A∗ = OpL (a∗L ) with
X 1
a∗L (x, ξ) ∼ ∂ξα Dxα aL (x, ξ) (4.56)
n
α!
α∈N0

4.2. Topology on spaces of pseudodifferential operators. Let m ∈ R or m = −∞.

Since OpL : S m (Rn ; Rn ) → Ψm (Rn ) is an isomorphism of vector spaces, it is natural to
transport the Fréchet space structure of S m (Rn ; Rn ) to Ψm (Rn ) via OpL . For instance:
Lemma 4.14. Let χ ∈ Cc∞ (Rnξ ) be identically 1 near 0, and put J = Op(χ(·)),  > 0.
Then J ∈ Ψ−∞ (Rn ) is uniformly bounded in Ψ0 (Rn ) and converges to the identity operator
I = Op(1) in the topology of Ψη (Rn ) for any η > 0.

Proof. This is equivalent to the main part of (the proof of) Proposition 3.7. 

It is reassuring to note that one can equally well define the topology on Ψm (Rn ) using
the right quantization. This is a consequence of the following result.
Proposition 4.15. Let m ∈ R or m = −∞. Then the isomorphism of vector spaces
OpR : S m (Rn ; Rn ) → Ψm (Rn ) is an isomorphism of Fréchet spaces.

Proof. Right reduction σR is the inverse of OpR . By definition of the Fréchet space structure
of Ψm (Rn ), the proposition is thus equivalent to the continuity of σR ◦ OpL , which is part
of Theorem 4.8. 

4.3. Composition. Proving that composition of ps.d.o.s produces another ps.d.o. is now
straightforward:
0
Theorem 4.16. Let A ∈ Ψm (Rn ), B ∈ Ψm (Rn ). Then A ◦ B : S (Rn ) → S (Rn ) is a
pseudodifferential operator,
0
A ◦ B ∈ Ψm+m (Rn ), (4.57)
and its left symbol is given as an asymptotic sum
X 1
σL (A ◦ B) ∼ ∂ α σL (A) · Dxα σL (B). (4.58)
n
α! ξ
α∈N0

The bilinear map (A, B) 7→ A ◦ B is continuous.

Note that the symbolic expansion (4.58) is local in (x, ξ): the symbols of A and B do
not ‘interact’ at all, modulo residual terms, at distinct points in phase space Rnx × Rnξ .
 MICROLOCAL ANALYSIS 27

Proof of Theorem 4.16. Write A = OpL (a) and B = OpR (bR ). Assume first that A, B ∈
Ψ−∞ (Rn ), then for u, v ∈ S (Rn ), we have
Z
−n
Av(x) = (2π) eix·ξ a(x, ξ)v̂(ξ) dξ,
R n
Z (4.59)
−iy·ξ
Bu(ξ) =
c e bR (y, ξ)u(y) dy.
Rn

Thus,
Z Z
−n
ABu(x) = (2π) ei(x−y)·ξ a(x, ξ)bR (y, ξ)u(y) dy dξ, (4.60)
Rn Rn
giving A ◦ B = Op(c), c(x, y, ξ) = a(x, ξ)bR (y, ξ). (This is one of the reasons for considering
such general symbols!) By density and continuity, this continues to hold for A, B as in the
statement of the theorem.
To get the asymptotic expansion (4.58), let us write a = σL (A), b = σL (B), then1
X 1
∂ξα a(x, ξ)Dyα bR (y, ξ)|y=x


σL (A ◦ B)(x, ξ) ∼
α
α!
!
X 1 β γ β+γ
X (−1)|δ|
δ δ
∼ ∂ a(x, ξ) · ∂ξ Dx (∂ξ Dx b)(x, ξ)
β!γ! ξ δ! (4.61)
β,γ δ
 
X 1 β X1 X !
∼ ∂ξ a(x, ξ) · Dxβ  ∂ξ Dx b(x, ξ) (−1)|δ| 
β! 
! γ!δ!
β γ+δ=

and the observation that for  = 0, the final sum evaluates to 1, while for Nn0 3  6= 0,
X ! Y
(−1)|δ| = (1 − 1)j = 0. (4.62)
γ!δ!
γ+δ= j 6=0

This finishes the proof. 

As a simple application, we can now prove the pseudolocality of ps.d.o.s:

Proposition 4.17. Let A ∈ Ψm (Rn ). Then
sing supp Au ⊂ sing supp u, u ∈ S 0 (Rn ). (4.63)

To prove this, we record:

Lemma 4.18. A residual operator A ∈ Ψ−∞ (Rn ) is continuous as a map
A : S 0 (Rn ) → S 0 (Rn ) ∩ C ∞ (Rn ). (4.64)
More precisely, for any u ∈ S 0 (Rn ) we have Au ∈ hxiN Cb∞ (Rn ) for some N (depending on
u).

1Since these are asymptotic sums, it suffices to consider only those terms which have symbolic order
bigger than some fixed but arbitrary number; in particular, there are no convergence or rearrangement
issues.
28 PETER HINTZ

Proof. Let K denote the Schwartz kernel of A; recall that it satisfies the estimates (4.37).
For u ∈ S 0 (Rn ), we then have, for some N ∈ N,
|(Au)(x)| = |hK(x, ·), ui| ≤ CkK(x, ·)kN = C sup |y α Dyβ K(x, y)|
y∈Rn
|α|+|β|≤N

≤ C sup |hyiN Dyβ K(x, y)| = C sup hyiN hx − yi−N |hx − yiN Dyβ K(x, y)|.
y∈Rn y∈Rn
|β|≤N |β|≤N
(4.65)

Using Lemma 4.3, we see that hyiN hx − yi−N ≤ CN hxiN , hence

|(Au)(x)| ≤ ChxiN . (4.66)
Derivatives in x are estimated analogously, so Au ∈ C ∞ (Rn ), and in fact
|∂xα (Au)(x)| ≤ Cα hxiN . (4.67)
Note here that the number N above only depends on u, not on K itself. 

Proof of Proposition 4.17. Suppose x ∈ e ∈ Cc∞ (Rn ) such

/ sing supp u. There exist cutoffs χ, χ
that
χ(x) 6= 0, χ
e ≡ 1 on supp χ, χ eu ∈ Cc∞ (Rn ). (4.68)
Then
χu) + χA(1 − χ
χAu = χA(e e)u. (4.69)
Since A acts on S (Rn ), χu) ∈
we have χA(e S (Rn ).
For the second term, note that χ and
1−χ e have disjoint supports; hence we have
X 1
σL (χA ◦ (1 − χe))(x, ξ) ∼ χ(x)∂ξα σL (A)(x, ξ) · Dxα (1 − χ
e(x)) = 0, (4.70)
α
α!

which implies
e) ∈ Ψ−∞ (Rn ).
χA(1 − χ (4.71)
e)u ∈ C ∞ (Rn ), finishing the proof.
By Lemma 4.18, we conclude that χA(1 − χ 

Returning to the observation (4.71), note that if A = Op(a) has Schwartz kernel K ∈
S 0 (Rn × Rn ),
then the Schwartz kernel of χA(1 − χ
e) is χ(x)(1 − χ
e(y))K(x, y). Thus, (4.71)
can equivalently be stated as:
Proposition 4.19. The Schwartz kernel K of a pseudodifferential operator is smooth away
from the diagonal ∆ = {(x, x) : x ∈ Rn }. That is, sing supp K ⊂ ∆.

4.4. Principal symbols. Similarly to Proposition 2.20, the ‘leading order part’ of the left
or right symbol of an operator A ∈ Ψm (Rn ) has particularly simple properties.
Definition 4.20. Let m ∈ R. The principal symbol σm (A) of a ps.d.o. A ∈ Ψm (Rn ) is the
equivalence class
σm (A) := [σL (A)] ∈ S m (Rn ; Rn )/S m−1 (Rn ; Rn ). (4.72)
We shall often omit from the notation the passage to the equivalence class.
 MICROLOCAL ANALYSIS 29

Directly from the definition, this gives a short exact sequence for every m ∈ R:
σ
0 → Ψm−1 (Rn ; Rn ) → Ψm (Rn ; Rn ) −−m
→ S m (Rn ; Rn )/S m−1 (Rn ; Rn ) → 0. (4.73)
The surjectivity of σm is clear: given a representative a ∈ S m (Rn ; Rn ) of an equivalence
class of symbols, we have σm (OpL (a)) = [a].
Proposition 4.21. The principal symbol map has the following properties:
(1) σm (OpR (a)) = [a], i.e. using the right symbol in (4.72) gives the same principal
symbol map.
(2) For A ∈ Ψm (Rn ), we have σm (A∗ ) = σm (A).
0
(3) For A ∈ Ψm (Rn ), B ∈ Ψm (Rn ), we have σm+m0 (A ◦ B) = σm (A)σm0 (B).

(The behavior under changes of variables will be discussed in §5.1.) Notice that the
principal symbol map translates operator composition (a highly non-commutative opera-
tion) to the multiplication of (equivalence classes of) functions (a commutative operation),
though of course at what seems to be an enormous loss of information compared to the full
expansion (4.58) (which itself gives up information on the residual part of A ◦ B). However,
in most situations, the principal symbol, and sometimes a ‘subprincipal’ part of the full
symbol, dominate the behavior of the operator, while lower order parts are irrelevant; cf.
the discussion of ellipticity for symbols in §3.1.
0
One crucial calculation is the following. For A ∈ Ψm (Rn ), B ∈ Ψm (Rn ), note that
σm+m0 (A ◦ B) = σm (A)σm0 (B) = σm+m0 (B ◦ A), so
σm+m0 ([A, B]) = 0, [A, B] = A ◦ B − B ◦ A. (4.74)
In view of (4.73), we thus have [A, B] ∈ Ψ m+m0 −1
(Rn ), and it is natural to inquire about its
0
principal symbol as an operator of order m + m − 1. It turns out that it can be computed
solely in terms of the principal symbols of A and B:
0
Proposition 4.22. For A ∈ Ψm (Rn ), B ∈ Ψm (Rn ), we have
σm+m0 −1 (i[A, B]) = {σm (A), σm0 (B)}, (4.75)
where the Poisson bracket of a, b ∈ C ∞ (Rnx × Rnξ ) is defined as
n
X
{a, b} := (∂ξj a)(∂xj b) − (∂xj a)(∂ξj b). (4.76)
j=1

This will be the key connection between ‘quantum mechanics’ (quantizations of sym-
bols, noncommutative algebra of operators) and ‘classical mechanics’ (symbols themselves,
commutative algebra of functions), which will play a central role in §8.

Proof of Proposition 4.22. We leave it to the reader to verify that (4.75) is well-defined, i.e.
0 0
that the image of the right hand side in the quotient space S m+m −1 /S m+m −2 does not
depend on the choice of representatives of the principal symbols of A and B.
The proof is an immediate application of (4.58). Let a = σL (A), b = σL (B). Working
0
modulo S m+m −2 (Rn ; Rn ), we have
n n
1X 1X
σL (A ◦ B) ≡ ab + (∂ξj a)(∂xj b), σL (B ◦ A) ≡ ab + (∂ξj b)(∂xj a), (4.77)
i i
j=1 j=1
30 PETER HINTZ

and (4.75) follows. 

4.5. Classical operators. Following Definition 3.11, we have a subclass of classical oper-
ators:
Definition 4.23. For m ∈ C, we define the space of classical pseudodifferential operators
of order m by
Ψm n m n n
cl (R ) := OpL (Scl (R ; R )) ⊂ Ψ
Re m
(Rn ), (4.78)
equipped with the structure of a Fréchet space which makes OpL into an isomorphism. We
put Ψ−∞ n
cl (R ) := Ψ
−∞ (Rn ).

The symbol expansions in Theorem 4.16 and Corollary 4.13 imply that compositions and
adjoints of classical operators are still classical:
Proposition 4.24. Composition of ps.d.o.s restricts to a continuous bilinear map
0 m+m 0
Ψm n m n
cl (R ) × Ψcl (R ) 3 (A, B) 7→ A ◦ B ∈ Ψcl (Rn ). (4.79)
Similarly, the map
∗
Ψm n m̄ n
cl (R ) 3 A 7→ A ∈ Ψcl (R ) (4.80)
is a continuous conjugate-linear map.
m (Rn ; Rn ), we can identify the prin-
For a classical operator A = OpL (a), with a ∈ Scl
cipal symbol σRe m (A) with the homogeneous leading order part of a, as discussed after
Lemma 3.12. The corresponding short exact sequence is
0 → Ψm−1
cl (Rn ) → Ψm n m n n
cl (R ) → Shom (R ; R \ {0}) → 0. (4.81)

4.6. Elliptic parametrix. Recall Definition 3.8 and the discussion around (3.19). Then:
Definition 4.25. We call an operator A ∈ Ψm (Rn ) (uniformly) elliptic if its principal
symbol σm (A) is elliptic.

As a first, and important, application of the symbol calculus we have developed above,
we construct parametrices (approximate inverses—a term which, almost by nature, has no
precise definition, but rather depends on the context) of uniformly elliptic operators.
Theorem 4.26. Let A ∈ Ψm (Rn ) be uniformly elliptic. Then there exists an operator
B ∈ Ψ−m (Rn ) which is unique modulo Ψ−∞ (Rn ), such that
AB − I, BA − I ∈ Ψ−∞ (Rn ). (4.82)

We call an operator B satisfying (4.82) a parametrix of A.

Proof of Theorem 4.26. Let b ∈ S −m (Rn ; Rn ) be such that σm (A)b − 1 ∈ S −1 (Rn ; Rn ). Put
B0 = Op(b) ∈ Ψ−m (Rn ), then
A ◦ B0 = I − R, R ∈ Ψ−1 (Rn ). (4.83)
Indeed, this follows from σ0 (AB0 −I) = 0. We approximately invert I −R using a Neumann
series: we choose
∞
X
R0 ∼ Rj ∈ Ψ−1 (Rn ), (4.84)
j=1
 MICROLOCAL ANALYSIS 31

i.e. the left symbol of R0 is an asymptotic sum of the left symbols of Rj = R ◦ · · · ◦ R (j

times). Since (I − R)(I + N j N +1 for all N , we have
P
j=1 R ) = I − R

(I − R)(I + R0 ) = I + E, E ∈ Ψ−∞ (Rn ). (4.85)

Therefore, if we put
B := B0 (I + R0 ) ∈ Ψ−m (Rn ), (4.86)
then AB = I + E, as desired.
An analogous argument produces B 0 ∈ Ψ−m (Rn ) with B 0 A = I + E 0 , E 0 ∈ Ψ−∞ (Rn ).
But then abstract ‘group theory’ gives
B = IB = (B 0 A − E 0 )B = B 0 AB − E 0 B = B 0 (I + E) − E 0 B = B 0 + (B 0 E − E 0 B). (4.87)
Therefore B − B 0 ∈ Ψ−∞ (Rn ). In particular, any two parametrices differ by an element of
Ψ−∞ (Rn ). 

As a simple application, we prove:

Proposition 4.27. Let A ∈ Ψm (Rn ) be uniformly elliptic, and suppose
u ∈ S 0 (Rn ), Au = f ∈ C ∞ (Rn ). (4.88)
Then u ∈ C ∞ (Rn ). More precisely, we have
sing supp u = sing supp Au. (4.89)

Proof. We prove (4.89). Let B ∈ Ψ−m (Rn ) be a parametrix of A, with BA = I + R,

R ∈ Ψ−∞ (Rn ). Then by Proposition 4.17, we have
sing supp u = sing supp(BAu+Ru) = sing supp BAu ⊂ sing supp Au ⊂ sing supp u. (4.90)
Therefore, equality must hold at each step. 
Example 4.28. Examples to which Proposition 4.27 applies are the Laplacian ∆ ∈ Ψ2 (Rn )
and the Cauchy–Riemann operator ∂¯ = 21 (∂x1 + i∂x2 ) ∈ Ψ1 (R2 ), which is identified with
¯ = 0 for u ∈ S 0 (Rn ), then
C via (x1 , x2 ) 7→ x1 + ix2 . For the latter, we deduce that if ∂u
∞ n
u ∈ C (R ). In complex analysis we learn that in fact u is analytic; here we are only
developing microlocal analysis in the smooth category, hence do not directly recover this
stronger conclusion.

4.7. Boundedness on Sobolev spaces. In practice, one typically uses function spaces
other than S (Rn ) and S 0 (Rn ), such as Hölder or Lp spaces. Here, we focus on function
spaces related to L2 , in parts because they are the most natural for the study of non-elliptic
operators in §8.
As usual, we first consider residual operators:
Proposition 4.29. Let A ∈ Ψ−∞ (Rn ). Then A extends by continuity from2 S (Rn ) to a
bounded linear operator A : L2 (Rn ) → L2 (Rn ).

This will follow from the estimates (4.37) and Schur’s lemma:

2We use here that S (Rn ) ⊂ L2 (Rn ) is a dense subspace.

32 PETER HINTZ

Lemma 4.30. Let (X, µ) and (Y, ν) be measure spaces. Suppose K(x, y) is measurable on
X × Y and Z Z
|K(x, y)| dµ(x) ≤ C1 , |K(x, y)| dν(y) ≤ C2 (4.91)
X Y
for almost all y ∈ Y and x ∈ X, respectively. Let
Z
T u(x) = K(x, y)u(y) dν(y). (4.92)
Y
Then T : L2 (Y ) → L2 (X) is bounded. Quantitatively,
kT ukL2 (X) ≤ (C1 C2 )1/2 kukL2 (Y ) . (4.93)

Proof. Let u ∈ L2 (Y ) and v ∈ L2 (X), then by Cauchy–Schwarz

Z Z 
 

 K(x, y)u(y)v(x) dν(y) dµ(x)
X Y
Z 1/2 Z 1/2
≤ |K(x, y)||u(y)|2 dµ(x) dν(y) |K(x, y)||v(x)|2 dµ(x) dν(y)
X×Y X×Y
1/2 1/2
≤ C1 kukL2 (Y ) · C2 kvkL2 (X) . 

Proof of Proposition 4.29. The Schwartz kernel K of A satisfies |K(x, y)| ≤ Chx − yi−n−1 ,
hence Z Z
|K(x, y)| dx ≤ C hzi−n−1 dz < ∞, (4.94)
R n Rn
R
and likewise Rn |K(x, y)| dy < ∞. The claim then follows from Lemma 4.30. 

Using ‘Hörmander’s square root trick’, we can now prove:

Theorem 4.31. Let A ∈ Ψ0 (Rn ). Then A : L2 (Rn ) → L2 (Rn ) is bounded.

Proof. By Corollary 4.13 and Theorem 4.16, we have A∗ A ∈ Ψ0 (Rn ). With a = σ0 (A) (that
is, a is any representative of σ0 (A)), we have σ0 (A∗ A) = |a|2 , which is real, non-negative,
and bounded. Thus, for C > supx,ξ∈Rn |a|2 , the symbol C − |a|2 ∈ S 0 (Rn ; Rn ) is elliptic
and positive. By Exercise 3.3, it has an approximate square root 0 < b0 ∈ S 0 (Rn ; Rn ), so
C − |a|2 − b20 ∈ S −1 (Rn ; Rn ). Let B0 = Op(b0 ), then
C − A∗ A = B0∗ B0 + R1 , R1 ∈ Ψ−1 (Rn ). (4.95)
Assume inductively that we have found Bj ∈ Ψ−j (Rn ), j = 0, . . . , k − 1, such that
Rk := C − A∗ A − (B0 + · · · + Bk−1 )∗ (B0 + · · · + Bk−1 ) ∈ Ψ−k (Rn ). (4.96)
This holds for k = 1. We try to improve the error term by finding the next correction
Bk = Op(bk ) ∈ Ψ−k (Rn ); we compute
Rk+1 = C − A∗ A − (B0 + · · · + Bk )∗ (B0 + · · · Bk )
(4.97)
= Rk − Bk∗ (B0 + · · · + Bk−1 ) + (B0 + · · · + Bk−1 )∗ Bk + Bk∗ Bk ∈ Ψ−k (Rn ).

Thus, the requirement Rk+1 ∈ Ψ−k−1 (Rn ) is equivalent to a principal symbol condition,
bk b0 + b0 bk = σ−k (Rk ) (in S −k (Rn ; Rn )/S −k−1 (Rn ; Rn )). (4.98)
 MICROLOCAL ANALYSIS 33

Since Rk = Rk∗ , the principal symbol σ−k (Rk ) is real; hence we can take bk = 12 σ−k (Rk )/b0 ∈
S −k (Rn ; Rn ).
Finally, we let B ∈ Ψ0 (Rn ) be the asymptotic sum
∞
X
B∼ Bk . (4.99)
k=0

We have then arranged

R := C − A∗ A − B ∗ B ∈ Ψ−∞ (Rn ). (4.100)
(Thus, we have constructed a square root, modulo residual operators, of C − A∗ A.)
Given u ∈ S (Rn ), we then have
kAuk2L2 (Rn ) = hA∗ Au, uiL2 (Rn )
= Ckuk2L2 (Rn ) − kBuk2L2 (Rn ) − hRu, ui
(4.101)
≤ Ckuk2L2 (Rn ) + kRukL2 (Rn ) kukL2 (Rn )
≤ C 0 kuk2L2 (Rn )

by Proposition 4.29. Thus, A extends by continuity to a bounded operator on L2 (Rn ). 

Boundedness of ps.d.o.s on Sobolev spaces is a straightforward consequence:

Corollary 4.32. Let s, m ∈ R, and A ∈ Ψm (Rn ). Then A : H s (Rn ) → H s−m (Rn ) is
bounded.

Proof. Recall the operators hDiσ = F −1 hξiσ F for σ ∈ R from Definition 2.11; note that
hDiσ ∈ Ψσ (Rn ). Moreover, hDi−s : L2 (Rn ) → H s (Rn ) and hDis−m : H s−m (Rn ) → L2 (Rn )
are isometric isomorphisms. Now
hDis−m AhDi−s ∈ Ψ0 (Rn ) (4.102)
is bounded on L2 (Rn ) by Theorem 4.31, which is equivalent to the statement of the corollary.


In fact, this can be generalized to weighted Sobolev spaces, see (2.28):

Theorem 4.33. Let s, m, r ∈ R, and A ∈ Ψm (Rn ). Then A : hxir H s (Rn ) → hxir H s−m (Rn )
is bounded.

Proof. Since hxir hDi−s : L2 (Rn ) → hxis H s (Rn ) and hDis−m hxi−r : hxir H s (Rn ) → L2 (Rn )
are isomorphisms, we need to show that
A0 := hDis−m hxi−r ◦ A ◦ hxir hDi−s ∈ Ψ0 (Rn ). (4.103)
If a = σL (A), then the full symbol a] (x, y, ξ) of A] := hxi−r ◦A◦hxir is given by a] (x, y, ξ) =
hxi−r hyir a(x, ξ). By Lemma 4.3, we have

|a] (x, y, ξ)| ≤ 2|r|/2 hx − yi|r| |a(x, ξ)| ≤ Chx − yi|r| hξim−s , (4.104)
34 PETER HINTZ

which is the first step towards showing that a] ∈ hx − yi|r| S m (Rn × Rn ; Rn ); it remains to
consider derivatives. The essence of this is contained in
|∂yj a] (x, y, ξ)| = −rhxir hyi−r−2 yj a(x, ξ)hξi−s 
 

yj
≤ Chx − yi|r| 2 hξim (4.105)
hyi
≤ Chx − yi|r| hξim .
We conclude that A] ∈ Ψm (Rn ), hence A0 ∈ Ψ0 (Rn ), finishing the proof. 

In view of the Schwartz representation theorem, Theorem 2.14, we S thus obtain another
proof of Lemma 4.18. Indeed, a residual operator maps S 0 (Rn ) = r,s hxir H s (Rn ) into
r ∞ n r ∞ n
S S
r hxi H (R ) = r hxi Cb (R ) (using Sobolev embedding, Exercise 2.3).
We can sharpen and upgrade the elliptic regularity result, Proposition 4.27:
Corollary 4.34. Let A ∈ Ψm (Rn ) be uniformly elliptic, and suppose u ∈ hxir H −N (Rn )
for some r, N ∈ R. If Au = f ∈ hxir H s−m (Rn ), then u ∈ hxir H s (Rn ).

Proof. With B ∈ Ψ−m (Rn ) denoting a parametrix of A, so I = BA + R, R ∈ Ψ−∞ (Rn ),

we have
u = BAu + Ru = Bf + Ru, (4.106)
r s n r σ n
S
with Bf ∈ hxi H (R ) and Ru ∈ hxi σ∈R H (R ). 
Remark 4.35. It is important that the assumption on u already has the weight factor hxir .
0
Indeed, the conclusion would be false in general if we merely assumed u ∈ hxir H −N (Rn )
for some r0 < r. (Convince yourself of this. For example, take A = ∆, the Laplacian on
Rn , and u = 1.)
4.8. Exercises.
Exercise 4.1. Let m ∈ N0 , and let a ∈ S m (Rn × Rn ; Rn ) be a polynomial in the symbolic
variable ξ.
(1) Show, starting from the definition as a limit of quantizations of residual symbols,
that Op(a) ∈ Diff m (Rn ).
(2) Prove that Op(hx − yiw a) ∈ Diff m (Rn ) (which in particular entails the boundedness
of the coefficients). (Hint. Compute its Schwartz kernel.)
Exercise 4.2. Let A ∈ Ψm (Rn ), and denote by K its Schwartz kernel.
(1) Give another, direct, proof that K ∈ C ∞ ((Rn ×Rn )\∆), where ∆ = {(x, x) : x ∈ Rn }
is the diagonal. (Hint. For φ, ψ ∈ Cc∞ (Rn ) with supp φ ∩ supp ψ = ∅, rewrite the
pairing hAφ, ψi for A ∈ Ψ−∞ (Rn ) using integrations by parts as in the proof of
Proposition 4.4. Then use a density argument.)
(2) Prove that for every  > 0, N ∈ R, α, β ∈ Nn0 , there exists a constant C such that
|∂xα ∂yβ K(x, y)| ≤ C|x − y|−N , |x − y| ≥ . (4.107)
Exercise 4.3. Suppose K(x, z) ∈ C ∞ (R × (R \ {0})) satisfies K(x, λz) = λ−1 K(x, z), λ > 0,
and K(x, −z) = −K(x, z). Assume that K(x, 1) ∈ Cb∞ (Rx ). Let χ ∈ Cc∞ (Rn ) be identically
1 near 0. Show that the operator
Z
Au(x) = lim χ(x − y)K(x, x − y)u(y) dy, u ∈ Cc∞ (Rn ), (4.108)
→0 |x−y|≥
 MICROLOCAL ANALYSIS 35

is well-defined and defines an element A ∈ Ψ0cl (R). Compute its principal symbol.
Exercise 4.4. Prove Gårding’s inequality. Let A ∈ Ψ2m (Rn ), and suppose Re σ2m (A) ≥
chξi2m for some c ∈ R. Then for every  > 0 and N ∈ R, there exists a constant C such
that
RehAu, uiL2 (Rn ) ≥ (c − )kuk2H m (Rn ) − Ckuk2H −N (Rn ) , u ∈ S (Rn ). (4.109)
(Hint. Use the ‘square root trick’.) The sharp Gårding inequality states that (4.109) holds
for  = 0, but then with −N = m − 1/2; see [Hör03, Theorem 18.1.14]. (This can be further
refined to the Fefferman–Phong inequality, which allows −N = m − 1.)

The following series of exercises introduces the basic properties of scattering pseudodif-
ferential operators on Rn .
Exercise 4.5. (Scattering symbols.) For m, r1 , r2 ∈ R, define the space of symbols
S m,r1 ,r2 (Rnx × Rny ; Rnξ ) (4.110)
to consist of all a ∈ C ∞ (R3n ) such that the seminorms
kakm,r1 ,r2 ,k := sup hxi−r1 +|α1 | hyi−r2 +|α2 | hξi−m+|β| |∂xα1 ∂yα2 ∂ξβ a(x, y, ξ)| (4.111)
|α1 |+|α2 |+|β|≤k

are finite for all k ∈ N0 .3 Let

\
S −∞,−∞,−∞ (Rn × Rn ; Rn ) := S m,r1 ,r2 (Rn × Rn ; Rn ). (4.112)
m,r1 ,r2 ∈R

(1) Prove that S −∞,−∞,−∞ (Rn × Rn ; Rn ) ⊂ S m,r1 ,r2 (Rn × Rn ; Rn ) is dense in the topol-
0 0 0
ogy of S m ,r1 ,r2 (Rn × Rn ; Rn ) whenever m < m0 , r1 < r10 , r2 < r20 .
(2) Prove the following variant of Proposition 3.13: given aj ∈ S m−j,r1 −j,r2 −j (Rn ×
Rn ; Rn ), there exists a ∈ S m,r1 ,r2 (Rn × Rn ; Rn ), unique modulo S −∞,−∞,−∞ (Rn ×
Rn ; Rn ), such that a − J−1 m−J,r1 −J,r2 −J (Rn × Rn ; Rn ) for all J ∈ N.
P
j=0 aj ∈ S

Exercise 4.6. (Scattering ps.d.o.s, I: boundedness.) Let m, r1 , r2 ∈ R. Prove that

Op(a) : S (Rn ) → S (Rn ), a ∈ S m,r1 ,r2 (Rn × Rn ; Rn ). (4.113)
Prove this more generally for a ∈ hx − yiw S m,r1 ,r2 (Rn × Rn ; Rn ), w ∈ R.
Exercise 4.7. (Scattering ps.d.o.s, II: residual operators.) Let r2 ∈ R. Show that an
operator A can be written as A = Op(aN ), aN ∈ S −N,−r2 ,−N (Rn × Rn ; Rn ), for all N if and
only if its Schwartz kernel K = K(x, y) satisfies K ∈ S (Rnx × Rny ). Show that in this case,
there exist unique aL , aR ∈ S −∞,−∞ (Rn ; Rn ) such that A = OpL (aL ) = OpR (aR ).
Exercise 4.8. (Scattering ps.d.o.s, III: reduction.) We write
S m,r (Rn ; Rn ) (4.114)
for the space of a = a(x, ξ) ∈ C ∞ (R2n ) satisfying |∂xα ∂ξβ a(x, ξ)| ≤ Cαβ hxir−|α| hξim−|α| for
all α, β ∈ Nn0 .
Let A = Op(a), a ∈ S m,r1 ,r2 (Rn × Rn ; Rn ). Prove that there exists a unique left symbol
aL ∈ S m,r1 +r2 (Rn ; Rn ) such that A = OpL (aL ).
3That is, such a are symbolic not only in ξ, but also in x and y.
36 PETER HINTZ

Exercise 4.9. (Scattering ps.d.o.s, IV: algebra.) Define

Ψm,r n
sc (R ) := Op(S
m,r
(Rn ; Rn )). (4.115)
(1) Prove that A ∈ Ψm,r n ∗ m,r
sc (R ) implies A ∈ Ψsc (R ).
n
0 0
(2) Suppose A ∈ Ψm,r n m ,r
sc (R ), B ∈ Ψsc (Rn ). Prove that
0 0
A ◦ B ∈ Ψm+m
sc
,r+r
(Rn ). (4.116)
Exercise 4.10. (Scattering ps.d.o.s, V: principal symbol.) Define the principal symbol of
A = OpL (aL ) ∈ Ψm,r n
sc (R ) by
m,r
σsc (A) := [aL ] ∈ S m,r (Rn ; Rn )/S m−1,r−1 (Rn ; Rn ). (4.117)
State and prove the analogue of Proposition 4.21 for scattering ps.d.o.s.4
Exercise 4.11. (Scattering ps.d.o.s, VI: ellipticity.) Suppose A = OpL (aL ) ∈ Ψm,r n
sc (R ) is
elliptic, that is, there exists b ∈ S −m,−r n n
(R ; R ) such that aL b − 1 ∈ S −1,−1 n n
(R ; R ).
Prove that there exists B ∈ Ψ−m,−r
(1) T sc (Rn ) such that BA − I ∈ Ψ−∞,−∞
sc (Rn ; Rn ) =
m,r n
m,r∈R Ψsc (R ).
(2) Suppose u ∈ S 0 (Rn ), and Au = f ∈ S (Rn ). Prove that u ∈ S (Rn ). (Notice the
difference to the statements of Proposition 4.27 or Corollary 4.34! For example, the
Laplacian ∆ ∈ Ψ2 (Rn ) is uniformly elliptic, but ∆u = 0 for u = 1, u = x1 x2 , etc.
However, ∆ is not elliptic as an element of Ψ2,0 n
sc (R ). (Check!) What about ∆ + 1?)

Exercise 4.12. (Scattering ps.d.o.s, VII: boundedness on Sobolev spaces.)

(1) Prove that elements of Ψ0,0 n 2 n
sc (R ) are bounded as maps on L (R ).
(2) Show that Λm,r := hxir hDim ∈ Ψsc (Rn ) and Λ0m,r := hDim hxir ∈ Ψm,r
m,r n
sc (R ).
m,r n
(3) Let A ∈ Ψsc (R ). Show that for all ρ, σ ∈ R, A is a bounded operator
A : hxiρ H σ (Rn ) → hxiρ+r H σ−m (Rn ). (4.118)
Exercise 4.13. (Scattering ps.d.o.s, VIII: elliptic scattering ps.d.o.s are Fredholm.)
0 0
(1) Let m < m0 and r > r0 . Show that the inclusion hxir H m (Rn ) ,→ hxir H m (Rn ) is
compact.
(2) Let A ∈ Ψm,r n
sc (R ) be elliptic (see Exercise 4.11). Show that for any ρ, σ ∈ R, the
operator
A : hxiρ H σ (Rn ) → hxiρ+r H σ−m (Rn ) (4.119)
is a Fredholm operator.
(3) Show that the index of A in (4.119) is independent of ρ, σ.

5. Pseudodifferential operators on manifolds

We now show how the ps.d.o. algebra on Rn can be transferred to smooth manifolds by
using local coordinate charts. The key ingredient for showing that this is a reasonable thing
to do is the invariance of the class of m-th order ps.d.o.s under changes of coordinates on
Rn .
4Thus, the principal symbol is more powerful in the scattering world: it not only captures the high
frequency, i.e. large ξ, behavior of an operator, but also the large x behavior.
 MICROLOCAL ANALYSIS 37

5.1. Local coordinate invariance. We now prove the analogue of the final part of Propo-
sition 2.20 for ps.d.o.s.
Definition 5.1. Let Ω ⊂ Rn be an open set. Then
Ψm m
c (Ω) := {A ∈ Ψ (Ω) : supp KA b Ω × Ω}, (5.1)
where KA ∈ S 0 (R2n ) denotes the Schwartz kernel of A.
Theorem 5.2. Suppose Ω, Ω0 ⊂ Rn are open, and κ : Ω → Ω0 is a diffeomorphism. Given
A ∈ Ψm 0 ∗ −1 ∗ m m 0
c (Ω ), define Aκ u := κ A(κ ) (u|Ω ). Then Aκ ∈ Ψc (Ω), and the map Ψc (Ω ) 3
m
A 7→ Aκ ∈ Ψc (Ω) is bijective. Moreover,
σm (Aκ )(x, ξ) = σm (A)(κ(x), (κ0 (x)T )−1 ξ). (5.2)

Proof. We have A = OpL (a) for some a ∈ S m (Rn ; Rn ). Choose ψ ∈ Cc∞ (Ω0 ) such that
ψ(x)ψ(y) = 1 on supp KA ; thus KA (x, y) = ψ(x)KA (x, y)ψ(y), and therefore


KA = Op ψ(x)a(x, ξ)ψ(y) . (5.3)
We localize near the diagonal: for  > 0 (to be determined), let χ (x, y) ∈ C ∞ (R2n ) be such
that χ (x, y) = 1 for |x − y| <  and χ (x, y) = 0 for |x − y| > 2. Then
KA := Op(a ), a (x, y, ξ) = χ (x, y)ψ(x)ψ(y)a(x, ξ), (5.4)
is the Schwartz kernel of an operator A ∈ Ψm 0
c (Ω ), and
R := A − A (5.5)
is a ps.d.o. with Schwartz kernel supported away from the diagonal, hence R ∈ Ψ−∞ (Rn ),
and its Schwartz kernel satisfies KR ∈ Cc∞ (Ω0 × Ω0 ). We then have
Z
(R )κ u(x) = KR (κ(x), y 0 )u(κ−1 (y 0 )) dy 0
0
ZΩ (5.6)
0
= KR (κ(x), κ(y))| det κ (y)|u(y) dy.
Ω
Therefore, the Schwartz kernel of (R )κ is K(R )κ (x, y) = KR (κ(x), κ(y))| det κ0 (y)| for
x, y ∈ Ω, and 0 otherwise. Thus, (R )κ ∈ Ψ−∞
c (Ω).
It remains to show that (A )κ ∈ Ψm c (Ω). To this end, note that
ZZ
0 0
(A )κ u(x) = (2π)−n ei(κ(x)−y )ξ a (κ(x), y 0 , ξ 0 )u(κ−1 (y 0 )) dy 0 dξ 0
ZZ (5.7)
0
= (2π)−n ei(κ(x)−κ(y))ξ a (κ(x), κ(y), ξ 0 )| det κ0 (y)|u(y) dy dξ 0 ,

thus the Schwartz kernel of (A )κ is

Z
−n 0
K(A )κ (x, y) = (2π) ei(κ(x)−κ(y))ξ b (x, y, ξ 0 ) dξ 0 ,

b (x, y, ξ 0 ) = a (κ(x), κ(y), ξ 0 )| det κ0 (y)| (5.8)

= χ (κ(x), κ(y))ψ(κ(x))ψ(κ(y))a(κ(x), ξ)| det κ0 (y)|.

We Taylor expand the exponent: denoting by κj the j-th component of κ, we have
n
X Z 1
κj (x) − κj (y) = Φjk (x, y)(xk − yk ), Φjk (x, y) = ∂xk κj (y + t(x − y)), (5.9)
k=1 0
38 PETER HINTZ

and therefore
(κ(x) − κ(y)) · ξ 0 = hΦ(x, y)(x − y), ξ 0 i = hx − y, Φ(x − y)T ξ 0 i, (5.10)
where Φ(x, y) = (Φjk (x, y))j,k=1,...,n , and h·, ·i is the inner product on Rn . Note now that
Φ(x, x) = κ0 (x) (5.11)
is invertible for x ∈ Ω since κ is a diffeomorphism. For
(x, y) ∈ supp χ (κ(x), κ(y))ψ(κ(x))ψ(κ(y)), (5.12)
we have (x, y) ∈ κ−1 (supp ψ) × κ−1 (supp ψ) b Ω × Ω and |κ(x) − κ(y)| ≤ 2. Therefore, we
can choose  > 0 such that Φ(x, y) is invertible for (x, y) in the set (5.12). In (5.8), we can
then make the change of variables ξ 0 = (Φ(x, y)T )−1 ξ, so
(A )κ = Op(c ), c (x, y, ξ) = b (x, y, (Φ(x, y)T )−1 ξ)| det Φ(x, y)|−1
= a (κ(x), κ(y), (Φ(x, y)T )−1 ξ)| det Φ(x, y)|−1 | det κ0 (y)|;
(5.13)
it remains to check that c ∈ S m (Rn ×Rn ; Rn ). We can drop the (smooth) Jacobian factors.
We then compute
∂xα ∂yβ ∂ξγ a (κ(x), κ(y), (Φ(x, y)T )−1 ξ)

ββ 0 β 00 α00 +β 00 α0 β 0 γ+α00 +β 00
X
= Fαα 0 α00 γ (x, y)ξ (∂x ∂y ∂ξ a )(κ(x), κ(y), (Φ(x, y)T )−1 ξ) (5.14)
|α0 |+|α00 |≤|α|
|β 0 |+|β 00 |≤|β|

ββ β 0 00
∞ T and its inverse are uni-
for some smooth functions Fαα 0 α00 γ ∈ Cc (Ω × Ω). Since Φ(x, y)

formly bounded on supp a , there exist c, C > 0 such that c|ξ| ≤ |(Φ(x, y)T )−1 ξ| ≤ C|ξ|
on supp a . Therefore, (5.14) is bounded by a constant times hξim−|γ| on supp a , proving
c ∈ S m (Rn × Rn ; Rn ).
As for the principal symbol, we have σm (Aκ ) = σm ((A )κ ) = σm (Op(c )), which can be
read off from the (first term of the) reduction formula (4.34): using (5.11), it is given by
the equivalence class in S m (Rn ; Rn )/S m−1 (Rn ; Rn ) of
c (x, x, ξ) = b (x, x, (κ0 (x)T )−1 ξ)| det κ0 (x)|−1
= a (κ(x), κ(x), (κ0 (x)T )−1 ξ) (5.15)
0 T −1
= a(κ(x), (κ (x) ) ξ).
The proof is complete. 

5.2. Manifolds, vector bundles, densities. We shall only work with smooth manifolds:
they are locally diffeomorphic to the unit ball B(0, 1) = {x ∈ Rn : |x| < 1}. We recall the
‘hands-on’ definition of smooth manifolds:
Definition 5.3. Let n ∈ N. A smooth manifold of dimension n is a second countable,
paracompact Hausdorff space M such that
(1) for each point p ∈ M , there exist an open neighborhood Up 3 p and a homeomor-
phism Fp : Up → B(0, 1) ⊂ Rn ;
 MICROLOCAL ANALYSIS 39

(2) for all p, q ∈ M such that Up ∩ Uq 6= ∅, the transition map

Fp ◦ Fq−1 |Up ∩Uq : Fq (Up ∩ Uq ) → Fp (Up ∩ Uq ) (5.16)
is smooth (as a map between open subsets of Rn ).
Definition 5.4. Let M be a smooth manifold. A atlas on M is a collection {(Uα , Fα )} of
pairs (Uα , Fα ), with Uα 6= ∅ and M = α Uα , such that Fα : Uα → Rn is a diffeomorphism
S
onto an open subset Fα (Uα ) of Rn . A maximal atlas, or smooth structure, is an atlas with
the property that any other atlas is contained in it. An element of the5 maximal atlas is
called a (local coordinate) chart.

The ‘hands-on’ definition of vector bundles is the following.

Definition 5.5. Let M be a smooth n-dimensional manifold. A real vector bundle of rank
k over M is a triple (π, E, M ) with the following properties:
(1) E is a smooth (n + k)-dimensional manifold, and π : E → M is smooth;
(2) for each p ∈ M there exist an open neighborhood Up ⊂ M , p ∈ Up , and a diffeo-
morphism τp : π −1 (Up ) → Up × Rk such that π(τp−1 (q, v)) = q is the projection onto
the first factor;
(3) for all p, q ∈ M such that Up ∩ Uq 6= ∅, the transition map
τpq := τp ◦ τq−1 |(Up ∩Uq )×Rk : (Up ∩ Uq ) × Rk → (Up ∩ Uq ) × Rk (5.17)
takes the form τpq (r, v) = (r, Φpq (r)v), where Φpq : Up ∩ Uq → GL(k) is smooth.
The fiber of E over p ∈ M is denoted Ep := π −1 (p); it is a k-dimensional real vector space.
The zero section of E is the submanifold of E given locally in π −1 (Up ) by τp−1 (Up × {0}).
A smooth section of E is a smooth map s : M → E such that π ◦ s = IdM . The space of
smooth sections is denoted C ∞ (M ; E).

Another useful notion for later is the pullback of vector bundles:

Definition 5.6. Let M, N be smooth manifolds (not necessarily of the same dimension),
and let f : M → N be smooth. If π : E → N is a vector bundle, then the pullback of E by
f is the vector bundle
π̃ : f ∗ E → M (5.18)
given by f ∗ E = M f × π E = {(p, e) ∈ M ×E : f (p) = π(e)}, with projection map π̃(p, e) = p,
and with linear structure on (f ∗ E)p = Ef (p) equal to that on Ef (p) .

To specify a real rank k vector bundle uniquely (up to vector bundle isomorphisms), it
suffices to have the following data and conditions:
(1) a cover {Uα } of M by open non-empty subsets;
(2) for all α, β with Uαβ := Uα ∩ Uβ 6= ∅ a map ταβ : Uαβ × Rk → Uαβ × Rk of the form
ταβ (p, v) = (p, Φαβ (p)v) with Φαβ : Uαβ → GL(k) smooth;
(3) ταα (p, v) = (p, v) for all p ∈ Uα , v ∈ Rk ;
(4) the cocycle condition holds: for α, β, γ with Uαβγ := Uα ∩ Uβ ∩ Uγ 6= ∅, we have
τγβ ◦ τβα = τγα on Uαβγ × Rk .
5A maximal atlas always exists and is unique.
40 PETER HINTZ

Indeed, one can then set !

G
k
E := Uα × R / ∼, (5.19)
α
where we define the equivalence relation ∼ by
Uα × Rk 3 (p, v) ∼ (q, w) ∈ Uβ × Rk ⇐⇒ p = q, ταβ (q, w) = (p, v). (5.20)
(The cocycle condition guarantees that this is transitive, while reflexivity follows from
the cocycle condition together with ταα = Id.) We denote the equivalence classes by
[Uα , (p, v)] ∈ E. Note that the same point p can lie in two (or more) open sets Uα and Uβ ,
but typically [Uα , (p, v)] 6= [Uβ , (p, v)] for v ∈ Rk . The projection map π : E → M is simply
given by π([Uα , (p, v)]) = p. As local trivializations, we can take
τα : {[Uα , (p, v)] : p ∈ Uα } 7→ (p, v) ∈ Uα × Rk . (5.21)
Example 5.7. Taking a cover of an n-dimensional manifold M by coordinate charts Fi : Ui →
Rn , we take τij (p, v) = (p, (Fi ◦ Fj−1 )0 |Fj (p) v). The resulting vector bundle is the tangent
bundle of M , denoted T M . Note that a chart F : U → Rn induces a trivialization of TU M =
π −1 (U ) via U × Rn 3 (p, v) 7→ [U, (p, v)] ∈ TU M . A tangent vector V = [U, (p, v)] ∈ Tp M
has several interpretations.
(1) V is a directional derivative on M at p. That is, it gives a map
d 
C ∞ (M ) 3 f 7→ V (f ) := (F −1 )∗ f |U (F (p) + sv)  .


(5.22)

ds s=0
(See also Exercise 5.1.)
(2) V is the tangent vector of a smooth curve on M . This comes from the following
construction: consider the set of all smooth curves γ : I → M with γ(0) = p, where
I ⊂ R is an open interval containing 0. An equivalence relation on this set is defined
as follows: γ1 ∼ γ2 if and only if, in any local coordinate system F : U → Rn with
d d
p ∈ U , we have ds F (γ1 (s))|s=0 = ds F (γ2 (s))|s=0 . (Check that this condition is
independent of the choice of F .) Then Tp M is the set of equivalence classes of
curves. The tangent vector V is identified with the equivalence class of the curve
s 7→ F −1 (F (p) + sv).

Functorial operations on vector spaces give corresponding operations on vector bundles.

For instance, given a linear map A : V → W between to vector spaces, the adjoint is
AT : W ∗ → V ∗ ; if A is invertible, then this gives a map (AT )−1 : V ∗ → W ∗ . In the notation
of Example 5.7, we thus take
τij (p, v) = p, ((κ0ij |Fj (p) )T )−1 v , κij := Fi ◦ Fj−1 .


(5.23)
The resulting vector bundle is the cotangent bundle, denoted π : T ∗ M → M . Note that
formula (5.23) appears in (5.2) (except in the latter we also use/change local coordinates
on the base M via κij ). We recall then that given a smooth function f ∈ C ∞ (M ), we can
define its exterior derivative df ∈ C ∞ (M ; T ∗ M ) as follows: if Fi : Ui → Fi (Ui ) ⊂ Rn is a
coordinate chart, we define
τi (df (p)) := p, ∂xj ((Fi−1 )∗ f ) j=1,...,n ∈ Ui × Rn , p ∈ Ui .



(5.24)
That this indeed gives a well-defined section of T ∗ M follows from the change of variables
formula.
 MICROLOCAL ANALYSIS 41

As in Example 5.7, a local coordinate system F : U → Rn induces a trivialization of

T ∗U = π −1 (U ). The relationship between T ∗ M and T M is as follows: for any p ∈ M , we
have an isomorphism of vector spaces
Tp∗ M ∼
= Hom(Tp M, R) (5.25)
which is defined independently of any choices. This isomorphism can be described in
local trivializations of T M and T ∗ M induced by the same local coordinate system F : if
ω = [U, (p, ξ)] ∈ Tp∗ M and V = [U, (p, v)] ∈ Tp M where ξ, v ∈ Rn , we set ω(V ) = hξ, vi
(Euclidean inner product). See Exercise 5.2.
A natural choice for f in local coordinates near p is f = xk (i.e. f = Fi∗ xk ), in which
case (5.24) defines the differential dxk with τi (dxk ) = (p, (0, . . . , 1, . . . , 0))P(with the 1 in
the k-th slot). The exterior derivative df is then usually written as df = nj=1 (∂xj f )dxj ,
dropping the coordinate and trivialization maps from the notation.
Example 5.8. Let E → M and F → M denote two vector bundles.
(1) The fiberwise direct sum of vector spaces produces the vector bundle E ⊕ F → M ,
with fibers (E ⊕ F )p = Ep ⊕ Fp .
(2) Likewise, taking the fiberwise tensor product gives E ⊗ F → M , with fibers (E ⊗
F )p = Ep ⊗ Fp .
(3) The vector bundle Hom(E, F ) → M has fibers Hom(Ep , Fp ). We have Hom(E, F ) =
E ⊗ F ∗.
(4) Let q ∈ N. The fiberwise q-th exterior power of E produces the vector bundle
Λq E → M . In the special case E = T ∗ M , one often writes Λq M := Λq T ∗ M .

We discuss another important vector bundle, closely related to the top exterior power
Λn T ∗ M of the cotangent bundle of an n-dimensional manifold M :
Definition 5.9. Let α ∈ R. In the notation of Example 5.7, the α-density bundle on M is
the vector bundle
Ωα M → M (5.26)
with transition functions τij (p, v) = (p, | det κ0ij |Fj (p) |−α v), κij = Fi ◦ Fj−1 . We also write

ΩM := Ω1 M. (5.27)
Remark 5.10. Ωα M → M arises functorially from the following operation on vector spaces,
applied to T M : given a real n-dimensional vector space V , we define
Ωα V := {ω : Λn V → R : ω(µv) = |µ|α ω(v), v ∈ Λn V, µ ∈ R}. (5.28)
To see the relationship, note first that Λn V is 1-dimensional. Then, given another n-
dimensional vector space W and a map κ : V → W , let us fix bases e1 , . . . , en of V and
f1 , . . . , fn of W . Consider, as a warm-up, the top exterior powers: e1 ∧ · · · ∧ en and
f1 ∧ · · · ∧ fn are bases of Λn V and Λn W , and the map Λn κ : Λn V → Λn W is given by
e1 ∧ · · · ∧ en 7→ κ(e1 ) ∧ · · · κ(en ) = (det κ)f1 ∧ · · · ∧ fn , where det κ is the determinant of the
matrix of κ in these bases: that is, in the stated basis, Λn κ is simply multiplication by det κ.
Similarly, Ωα V and Ωα W are 1-dimensional, with basis elements ωV : µe1 ∧ · · · ∧ en 7→ |µ|α
and ωW : µf1 ∧ · · · ∧ fn 7→ |µ|α . Now, the map
Ωα κ : Ωα V → Ωα W (5.29)
42 PETER HINTZ

is given by
Ωα κ(ω)(f1 ∧ · · · ∧ fn ) = ω(κ−1 (f1 ) ∧ · · · ∧ κ−1 (fn )), (5.30)
hence Ωα κ(ωV ) = | det κ|−α ωW .

The proof of the following simple lemmas is left to the reader as a simple exercise.
Lemma 5.11. Let α, β ∈ R. Then
(1) (Ωα )∗ M = Ω−α M ,
(2) Ωα M ⊗ Ωβ M = Ωα+β M ,
(3) Ω0 M = M × R.

Proof. See Exercise 5.6. 

If x ∈ Rn denotes local coordinates on a manifold M , then a typical α-density is

|dx|α : µ∂x1 ∧ · · · ∧ ∂xn 7→ |µ|α . (5.31)
Similarly to differential forms, α-densities can be pulled back by smooth maps:
Lemma 5.12. Let f : M → N be a smooth map between smooth manifolds of the same
dimension; assume that f has no critical points. In local coordinates x, y on M, N , and
u(y) = u0 (y)|dy|α , define (f ∗ u)(x) = u0 (f (x))| det f 0 (x)|α |dx|α . Then f ∗ is a well-defined
map
f ∗ : C ∞ (N ; Ωα N ) → C ∞ (M ; Ωα M ). (5.32)
Remark 5.13. This can be formulated invariantly: if V ∈ Tp M , define its pushforward along
a smooth map f : M → N to be the tangent vector f∗ V ∈ Tf (p) N so that, as a directional
derivative, (f∗ V )(g) = V (f ∗ g) for g ∈ C ∞ (N ). Then in the notation of Lemma 5.12, the
pullback of an α-density u is given by (f ∗ u)|p (V1 ∧ · · · ∧ Vn ) = u|f (p) (f∗ V1 ∧ · · · ∧ f∗ Vn )
where n = dim M .

See Exercise 5.12.

For us, 1-densities are the most useful: sections of ΩM can be invariantly integrated. On
Rn , we write for u ∈ Cc∞ (Rn ; ΩRn ), u = u0 (x)|dx|:
Z Z
u := u0 (x) dx (5.33)
Rn Rn
Let {φi } be a partition of unity on M subordinate to a cover by coordinate systems Fi : Ui →
Fi (Ui ) ⊂ Rn with Ui compact. Define the map
Z XZ
∞
: Cc (M ; ΩM ) → R, u 7→ (Fi−1 )∗ (φi u). (5.34)
M i Rn

(Note here that (Fi−1 )∗ (φi u) ∈ Cc∞ (Rn ; ΩRn )!)

Proposition 5.14. The map (5.34) is independent of the choice of local coordinates and
the partition of unity.

Proof. First, suppose u ∈ Cc∞ (M ; ΩM ) is supported in the intersection of two coordinate

charts, with local coordinates x ∈ Rn and y ∈ Rn and transition function x = κ(y), then
u(x) = u0 (x)|dx| = u1 (y)|dy|. (5.35)
 MICROLOCAL ANALYSIS 43

But at x = κ(y), we have (Ω1 κ)|dy| = | det κ0 (y)|−1 |dx|, so |dx| = | det κ0 (y)||dy|. Therefore,
u1 (y) = u0 (κ(y))| det κ0 (y)|, and thus
Z Z Z
0
u1 (y) dy = u0 (κ(y))| det κ (y)| dy = u0 (x) dx. (5.36)
Rn Rn Rn

The proposition follows easily from this: if {ψj } is another partition ofPunity
R subordinate
to a cover by coordinate systems Gj : Vj → Gj (Uj ) ⊂ Rn , then M u = j M ψj u, and
R
Z XZ
u= ψj u
M j M
XZ
= (Fi−1 )∗ (φi ψj u)
i,j Rn
XZ
(Gj ◦ Fi−1 )∗ (G−1 ∗


= j ) (ψj φi u)
i,j Rn
XZ
= (G−1 ∗
j ) (ψj φi u)
i,j Rn
XZ
= (G−1 ∗
j ) ψj u. 
j Rn

In analogy with the case of Rn , this leads us to define distributions on a manifold as

follows:
Definition 5.15. The space D 0 (M ) consists of all continuous linear maps Cc∞ (M ; ΩM ) →
C. More generally, if E → M is a vector bundle, then D 0 (M ; E) consists of all continuous
linear functionals Cc∞ (M ; E ∗ ⊗ ΩM ) → C. The space E 0 (M ; E) consists of all continuous
linear functionals C ∞ (M ; E ∗ ⊗ ΩM ) → C.

Thus, C ∞ (M ; E) ,→ D 0 (M ; E) via the pairing

Z
Cc∞ (M ; E ∗ ∞
⊗ ΩM ) × C (M ; E) 3 (u, φ) 7→ hu(p), φ(p)i, (5.37)
M
where h·, ·i : E ∗ × E → R is the dual pairing; note that hu, φi ∈ C ∞ (M ; ΩM ) can indeed be
invariantly integrated by Proposition 5.14.
The support and singular support of a distribution are defined analogously to the local
(Rn ) case, see Definition 2.6. The space E 0 (M ; E) ⊂ D 0 (M ; E) is, as in the local theory (on
Rn ), the space of distributions with compact support. (Without further structure, there
is no natural analogue of the space of Schwartz functions or tempered distributions on a
general smooth manifold.)
Example 5.16. Let p ∈ M , then δp ∈ E 0 (M ; ΩM ) is the distribution defined by mapping
φ ∈ Cc∞ (M ) to φ(p).

In order to state the Schwartz kernel theorem in this context, we define the projections
πL : M 2 → M, (p, q) 7→ p,
2
(5.38)
πR : M → M, (p, q) 7→ q.
Then:
44 PETER HINTZ

Theorem 5.17. Let M be a smooth n-dimensional manifold, and let E, F → M be two

vector bundles. Then there is a one-to-one correspondence between continuous linear op-
erators A : Cc∞ (M ; E) → D 0 (M ; F ) and distributional Schwartz kernels K ∈ D 0 (M ×
M ; πL∗ F ⊗ πR
∗ (E ∗ ⊗ ΩM )). This correspondence is given by assigning to K the operator

OK : Cc∞ (M ; E) → D 0 (M ; F ), defined as
(OK φ)(ψ) = hK, πL∗ ψ ⊗ πR
∗
φi, φ ∈ Cc∞ (M ; E), ψ ∈ Cc∞ (M ; F ∗ ⊗ ΩM ). (5.39)

5.3. Differential operators on manifolds. Let M be a smooth n-dimensional manifold.

Before we talk about ps.d.o.s on M , let us think about differential operators.
Definition 5.18. The space of smooth vector fields on M is V(M ) := C ∞ (M ; T M ).

An element V ∈ V(M ) can be regarded as a differential operator by assigning

C ∞ (M ) 3 f 7→ V f ∈ C ∞ (M ), (V f )(p) = df (p)(V (p)) (5.40)
Definition 5.19. (1) We define Diff 0 (M ) = C ∞ (M ).
(2) We define Diff 1 (M ) as the space of all operators A : C ∞ (M ) → C ∞ (M ) of the form
Au = V u + f u with V ∈ V(M ), f ∈ C ∞ (M ).
(3) Let m ∈ N0 . Then Diff m (M ) is the space of all operators A : C ∞ (M ) → C ∞ (M )
which are of the form
XK
Au = Ak1 · · · AkNk u, Akj ∈ Diff 1 (M ), K ∈ N, Nk ≤ m. (5.41)
k=1

(Check that this agrees with the standard local coordinate definition.) Of course, differ-
ential operators also map Cc∞ (M ) → Cc∞ (M ), D 0 (M ) → D 0 (M ), E 0 (M ) → E 0 (M ). What
are the Schwartz kernels of differential operators? The Schwartz kernel KI of the identity
operator I ∈ Diff 0 (M ) should be
KI (x, y) = δ(x − y). (5.42)
We aim to make sense of this. Using the right projection πR from (5.38), we define the
right density bundle by
∗
ΩR := πR (ΩM ) (5.43)
∞ 2 ∞
Thus, integration in the second variable is a well-defined map Cc (M ; ΩR ) → Cc (M ).
More generally, the following map is well-defined:
Z
0 ∞
D (M ; ΩR ) × Cc (M ) 3 (K, u) 7→
2
K(·, y)u(y) ∈ D 0 (M ). (5.44)
M
By the Schwartz kernel theorem, every continuous map Cc∞ (M ) → D 0 (M ) is of this type!
Thus, (5.42) is well-defined as an element
KI ∈ D 0 (M 2 ; ΩR ). (5.45)
Remark 5.20. As a check, recall that KI acts on elements of6
Cc∞ (M 2 ; Ω(M 2 ) ⊗ (ΩR )∗ ) = Cc∞ (M 2 ; ΩL ), (5.46)
and indeed maps u ∈ Cc∞ (M 2 ; ΩL ) into M u(x, x), defined by Proposition 5.14. (Note that
R

restriction to the diagonal gives a map Cc∞ (M 2 ; ΩL ) → Cc∞ (M ; ΩM ) by Lemma 5.12.)

6This uses that Ω(M 2 ) = Ω ⊗ Ω .
L R
 MICROLOCAL ANALYSIS 45

Given A ∈ Diff m (M ), its Schwartz kernel KA ∈ D 0 (M 2 ; ΩR ) is then given by

KA = (πL∗ A)KI , (5.47)
where πL∗ A
denotes the lift of A to the first factor, i.e. differentiating only in the first factor
2
of M . Check that this is well-defined in the following general context: if π : E → M is a
vector bundle, then
(πL∗ A)K ∈ D 0 (M 2 ; πR
∗
E), ((πL∗ A)K)(·, y) = (AK)(·, y), y ∈ M, K ∈ D 0 (M 2 ; πR
∗
E),
(5.48)
is well-defined.

5.4. Definition of Ψm (M ). We continue to denote by M a smooth n-dimensional man-

ifold, and use the notation (5.38). The following definition captures what we would like
pseudodifferential operators on a manifold (not necessarily compact) to be: their Schwartz
kernels should, near the diagonal, be Schwartz kernels of ps.d.o.s on Rn in local coordinates,
while away from the diagonal they are simply smooth.
Definition 5.21. Let M be a smooth n-dimensional manifold. Let m ∈ R. Then Ψm (M )
is the space of linear operators
A : Cc∞ (M ) → C ∞ (M ) (5.49)
with the following properties:
(1) if φ, ψ ∈ C ∞ (M ) have supp φ ∩ supp ψ = ∅, then there exists K ∈ C ∞ (M 2 ; ΩR ) such
that Z
φA(ψu) = K(·, y)u(y), u ∈ Cc∞ (M ). (5.50a)
M
(2) if F : U → Rn is a diffeomorphism from an open set ∅ = 6 U ⊂ M to F (U ), and if
ψ ∈ Cc∞ (U ), then there exists B ∈ Ψm
c (F (U )) ⊂ Ψm (Rn ) (see Definition 5.1) such

that on U , we have


ψA(ψu) = F ∗ B (F −1 )∗ (ψu) , u ∈ Cc∞ (M ). (5.50b)

Remark 5.22. Taking as the smooth manifold M = Rn , the space Ψm (M ) defined here is
much larger than the space Ψm (Rn ) of uniform pseudodifferential operators defined in §4.
(One reason is that we do not constrain the size of Schwartz kernels away from the diagonal
∆M = {(p, p) : p ∈ M }.). To avoid confusion, one should denote the latter space by
Ψm n n
∞ (R ). When working on R , we shall, in these notes, only ever employ operators in the
uniform algebra, hence we shall right away drop the ‘∞’ subscript again!
Remark 5.23. Directly from the definition, the space Ψ−∞ (M ) consists of all operators
which have a Schwartz kernel in C ∞ (M 2 ; ΩR ). Equivalently, Ψ−∞ (M ) is the space of all
bounded linear operators E 0 (M ) → C ∞ (M ).

Ps.d.o.s act on distributions with compact support. We give a direct proof here, and
defer a ‘better’ proof in the spirit of (4.31) to later; see Corollary 5.42.
Proposition 5.24. Let A ∈ Ψm (M ). Then A extends by continuity from Cc∞ (M ) to a
bounded linear operator
A : E 0 (M ) → D 0 (M ). (5.51)
46 PETER HINTZ

Proof. Fix a cover of M by coordinate systems Fi : Ui → Fi (Ui ) ⊂ Rn with Ui compact,

and let {φi }, φi ∈ Cc∞ (Ui ), be a subordinate partition of unity. Fix φ̃i ∈ Cc∞ (Ui ) with φ̃i = 1
near supp φi . By (5.50b), we can write
φ̃i Aφi = φ̃i Aφ̃i φi = Fi∗ Bi (Fi−1 )∗ φi , Bi ∈ Ψm
c (Fi (Ui )). (5.52)

Let now u ∈ E 0 (M ), then φi u 6= 0 only for finitely many i. We then set

X X
Ãu := Fi∗ Bi (Fi−1 )∗ (φi u) + (1 − φ̃i )Aφi u. (5.53)
i i

Each one of the finitely many non-zero summands in the first sum is a pullback from Rn
of a tempered distribution with compact support, hence lies in E 0 (M ). The second (also
finite) sum lies in C ∞ (M ) by (5.50a).
For u ∈ Cc∞ (M ), we clearly have Ãu = Au. Since Cc∞ (M ) ⊂ E 0 (M ) is dense, (5.53)
defines the unique continuous extension of A to E 0 (M ) (which, of course, we call A simply,
rather than Ã). 

To get a more manageable characterization of Ψm (M ), we first prove:

Lemma 5.25. Let M be an n-dimensional manifold, and let F : U → F (U ) ⊂ Rn be a
coordinate patch. If B ∈ Ψm m n ∞ ∞
c (F (U )) ⊂ Ψ (R ), then the operator A : Cc (M ) → C (M )
defined by
Au = F ∗ B(F −1 )∗ (u|U ), u ∈ Cc∞ (M ), (5.54)
on U , and Au = 0 in M \ U , defines an element A ∈ Ψm (M ).

Proof. We first check (5.50a): given φ, ψ ∈ C ∞ (M ) with supp φ ∩ supp ψ = ∅, we have for
u ∈ Cc∞ (M )
φA(ψu) = F ∗ B 0 (F −1 )∗ (u|U ), B 0 := ((F −1 )∗ φ)B((F −1 )∗ ψ) ∈ Ψ−∞
c (F (U )) ⊂ Ψ
−∞
(Rn ),
(5.55)
where we used that supp((F −1 )∗ φ) ∩ supp((F −1 )∗ ψ) = ∅. Since the Schwartz kernel of B 0
is smooth, we obtain (5.50a). (In more detail, if KB 0 ∈ Cc∞ (F (U ) × F (U )) denotes the
Schwartz kernel of B 0 , then (5.50a) holds for K(x, y) := F ∗ (KB 0 (x, y)|dy|).)
As for (5.50b), suppose G : V → G(V ) ⊂ Rn is another coordinate patch, and let χ ∈
Cc∞ (M ), supp χ ⊂ V . Then
B1 := ((F −1 )∗ χ)B((F −1 )∗ χ) ∈ Ψm
c (F (U ∩ V )). (5.56)
Denote the change of coordinates by κ = F ◦ G−1 : G(U ∩ V ) → F (U ∩ V ), then
B2 := (B1 )κ = κ∗ B1 (κ−1 )∗ ∈ Ψm
c (G(U ∩ V )) (5.57)
by Theorem 5.2. Therefore,
χA(χu) = F ∗ B1 (F −1 )∗ u|U ∩V = G∗ κ∗ B1 (κ−1 )∗ (G−1 )∗ u|U ∩V = G∗ B2 (G−1 )∗ u|U ∩V , (5.58)
as desired. 

This already implies that there are lots of pseudodifferential operators on M , given
by locally finite sums of operators of the form (5.54). This gives almost (namely, up to
operators with smooth integral kernels) all of Ψm (M ):
 MICROLOCAL ANALYSIS 47
S
Theorem 5.26. Let M be an n-dimensional manifold, and let M = i Ui be a locally finite
open cover by coordinate charts Fi : Ui → Fi (Ui ) ⊂ Rn with Ui compact. Let A : Cc∞ (M ) →
D 0 (M ) be a linear operator. Then A ∈ Ψm (M ) if and only if there exist operators Bi ∈
Ψm ∞ 2
c (Fi (Ui )) and a section K ∈ C (M ; ΩR ) such that
X
A=K+ Fi∗ Bi (Fi−1 )∗ . (5.59)
i

Proof. If A is of the form (5.59), then A ∈ Ψm (M ) by the previous lemma. Conversely,

suppose A ∈ Ψm (M ). Let {φi }, φi ∈ Cc∞ (Ui ), be a partition of unity subordinate to {Ui },
i.e. supp φi ⊂ Ui , and i φi (x) = 1 for all x ∈ M . Choose φ̃i ∈ Cc∞ (Ui ) with supp φ̃i ⊂ Ui
P

and φ̃i ≡ 1 near supp φi . For u ∈ Cc∞ (M ), we then have

X X
Au = φ̃i Aφi u + (1 − φ̃i )Aφi u. (5.60)
i i

By definition, (1 − φ̃i )Aφi has a smooth Schwartz kernel Ki ∈ C ∞ (M 2 ; ΩR ); since supp Ki

is locally finite, we can define
X
K := Ki ∈ C ∞ (M 2 ; ΩR ). (5.61)
i

Considering a term φ̃i Aφi in the first sum in (5.60), note that
φ̃i A(φi u) = φ̃i Aφ̃i (φi u). (5.62)
But φ̃i Aφ̃i = Fi∗ Bi0 (Fi−1 )∗ φ̃i for some Bi0 ∈ Ψm c (Fi (Ui )), and therefore

φ̃i Aφi = Fi∗ Bi (Fi−1 )∗ , Bi u := Bi0 ((Fi∗ φi )u), u ∈ Cc∞ (M ), (5.63)

with Bi ∈ Ψm c (Fi (Ui )), as desired. 

When M is not compact, one can in general not compose two ps.d.o.s, even when both
are of order −∞, since they only act on Cc∞ (M ), but not on C ∞ (M ) in general, the problem
being potential growth of the Schwartz kernel away from the diagonal. The simplest cure
is to place an additional assumption on the Schwartz kernels:
Definition 5.27. We say that A ∈ Ψm (M ), with Schwartz kernel K ∈ D 0 (M 2 ; ΩR ), is
properly supported if the projection maps πL : supp K → M and πR : supp K → M are
proper, i.e. preimages of compact sets are compact.
Every ps.d.o. is the sum of a properly supported operator and a residual operator. (See
Exercise 5.9.) In other words, in situations where one does not care about order −∞ errors,
one can work entirely with properly supported operators.
Thus, properly supported operators are bounded on Cc∞ (M ) and E 0 (M ). Using partition
of unity arguments, one can show that they are also bounded on C ∞ (M ), D 0 (M ). For a
proof using a duality argument, see Corollary 5.42 below.
Remark 5.28. Complementing Remark 5.22, the subspace of Ψm (M ), M = Rn , consisting of
properly supported operators does not have a simple relationship with Ψm n
∞ (R ): on the one
hand, Schwartz kernels of elements of Ψ∞ (R ) may even have full support in Rn ×Rn , hence
m n

are not properly supported; on the other hand, properly supported elements of Ψm (M ) may
not be elements of Ψm n
∞ (R ) since membership in the latter space requires uniform bounds
off the diagonal, see e.g. Exercise 4.2.
48 PETER HINTZ

0
Theorem 5.29. Let A ∈ Ψm (M ) and B ∈ Ψm (M ), and assume at least one of A and B
0
is properly supported. Then A ◦ B : Cc∞ (M ) → C ∞ (M ) is a ps.d.o., A ◦ B ∈ Ψm+m (M ). If
both A and B are properly supported, then so is A ◦ B.

We will use the description of Theorem 5.26 for a particular kind of open cover:
Lemma 5.30. Let M be a smooth manifold. There exists a locally finite open cover {Ui }
of M such that whenever Ui ∩ Uj 6= ∅, then there exists a local coordinate chart F : U →
F (U ) ⊂ Rn with U ⊃ Ui ∪ Uj .

Proof. M is metrizable; this follows either by Urysohn’s metrization theorem, or from basic
Riemannian geometry. Denote a fixed metric on M by d, and denote metric balls by
B(p, r) = {q ∈ M : d(p, q) < r}. For each p ∈ M , let
r0 (p) := sup{r ∈ [0, 1] : B(p, r) is contained in a coordinate chart}. (5.64)
Since M is a manifold, we have r0 (p) > 0 for all p ∈ M . For p ∈ M , define the open set
 r (p) 
0
Vp := B p, . (5.65)
10
1
Suppose Vp ∩ Vq 6= 0; then d(p, q) ≤ 10 (r0 (p) + r0 (q)) ≤ 51 max(r0 (p), r0 (q)). By symmetry,
we may assume r0 (p) ≥ r0 (q). If z ∈ Vp ∪ Vq , then
 r (p) r0 (q)   r (p) r (p) r (p) 
0 0 0 0
d(p, z) < max , d(p, q) + ≤ max , + < 21 r0 (p). (5.66)
10 10 10 5 10
Therefore, Vp ∪ Vq ⊂ B(p, r02(p) ) is contained in a coordinate chart. Any locally finite
refinement {Ui } of the cover {Vp : p ∈ M } of M satisfies the conditions of the lemma. 
S
Proof of Theorem 5.29. By the previous lemma, we can fix an open cover M = i Ui of M
by coordinate charts Fi : Ui → Fi (Ui ) ⊂ Rn , with Ui compact, and so that for any i, j with
Ui ∩ Uj 6= ∅, the union Ui ∪ Uj is contained in a coordinate chart Fij : Uij → Fij (Uij ) ⊂ Rn .
Let us assume that A is properly supported. (The case that B is properly supported is
handled similarly.) Write
X
A=K+ Fi∗ Ai (Fi−1 )∗ , Ai ∈ Ψm
c (Fi (Ui )),
i
X (5.67)
B=K + 0
Fi∗ Bi (Fi−1 )∗ , Bi ∈ Ψm 0 ∞ 2
c (Fi (Ui )), K ∈ C (M ; ΩR ).
i

Since A and all the Fi∗ Ai (Fi−1 )∗ are properly supported, so is K ∈ C ∞ (M 2 ; ΩR ).

We consider the composition A ◦ B term by term and keep track of the support of the
Schwartz kernels of the various pieces.
We first prove that K ◦ K 0 ∈ Ψ−∞ (M ). We have (K ◦ K 0 )(x, y) = K(x, z)K 0 (z, y) dz;
R
for any compact set K1 ⊂ M there exists K2 ⊂ M such that in fact
Z
0
(K ◦ K )(x, y) = K(x, z)K 0 (z, y) dz, x ∈ K1 . (5.68)
K2
−1
Indeed, this holds for K2 = πL (supp K ∩ πR (K1 )). Thus, the Schwartz kernel of K ◦ K 0
∞ 2
lies in C (M ; ΩR ).
 MICROLOCAL ANALYSIS 49

Consider next the composition Ki := K◦Fi∗ Bi (Fi−1 )∗ . This maps u ∈ E 0 (M ) into C ∞ (M )

(in fact, into Cc∞ (M )); and if supp u ∩ Ui = ∅, then Ki u = 0. Thus, by Remark 5.23,
Ki ∈ C ∞ (M 2 ; ΩR ), supp Ki ⊂ M × Ui . (5.69)
−1
(In fact, supp Ki ⊂ πL (supp K ∩ πR (Ui )) × Ui is compact, but we do not need this infor-
mation.) Similarly, one shows that
Ki0 := Fi∗ Ai (Fi−1 )∗ ◦ K 0 ∈ C ∞ (M 2 ; ΩR ), supp Ki0 ⊂ Ui × M. (5.70)
(Note that its Schwartz kernel is not compactly supported since K 0 is not properly sup-
ported.) Finally, we need to consider the composition
Cij := Fi∗ Ai (Fi−1 )∗ ◦ Fj∗ Bj (Fj−1 )∗ : Cc∞ (M ) → C ∞ (M ). (5.71)
When Ui ∩ Uj = ∅, this composition is the 0 operator. When Ui ∩ Uj 6= ∅, we can use
Lemma 5.25 and write (5.71) equivalently as
Fij∗ Aij (Fij−1 )∗ ◦ Fij∗ Bij (Fij−1 )∗ = Fij∗ (Aij ◦ Bij )(Fij−1 )∗ , (5.72)
0 0
where Aij ∈ Ψm m
c (Fij (Ui )), Bij ∈ Ψc (Fij (Uj )). But then Aij ◦ Bij ∈ Ψc
m+m (F (U ∪ U )),
ij i j
m+m 0
hence (5.72) lies in Ψ (M ), with Schwartz kernel supported in Uij × Uij .
The proof is complete once we show that the supports of the Schwartz kernels of Ki , Ki0 ,
and Cij are locally finite. Take a point (p, q) ∈ M 2 , and choose i0 , j0 such that p ∈ Ui0 and
q ∈ Uj0 . Then U := Ui0 × Uj0 has non-trivial intersection with only finitely many of these
supports, as follows immediately from the local finiteness of {Ui }. 

Since every operator on a compact manifold is properly supported, we deduce:

0 0
Corollary 5.31. If M is a compact manifold, then Ψm (M ) ◦ Ψm (M ) ⊂ Ψm+m (M ).

5.5. Principal symbol. Motivated by Theorem 5.2, in particular formula (5.2), we want
to define the principal symbol of A ∈ Ψm (M ) as an equivalence class of symbols on T ∗ M .
Definition 5.32. Let M be a manifold and π : E → M a rank k vector bundle. For m ∈ R,
we define S m (E) ⊂ C ∞ (E) as the subspace of all a ∈ C ∞ (E) having the following property:
for each coordinate chart F : U → F (U ) ⊂ Rn on M on which E is trivial with trivialization
F ×Id
τ : π −1 (U ) → U × Rk −−−→ F (U ) × Rk , set
b(x, v) := (τ −1 )∗ (a|π−1 (U ) )(x, v) = a(τ −1 (x, v)) ∈ C ∞ (F (U ) × Rk ). (5.73)

Then for any φ ∈ Cc∞ (F (U )) ⊂ Cc∞ (Rn ), we have φ(x)b(x, v) ∈ S m (Rn ; Rk ).

The key to making this a useful definition is the analogue of Lemma 5.25.
Lemma 5.33. In the notation of Definition 5.32, suppose φ ∈ Cc∞ (F (U )), b ∈ S m (Rn ; Rk ).
Then a := τ ∗ (φb) ∈ S m (E).

Proof. The expression for a in another coordinate system F 0 : U 0 → F 0 (U 0 ) ⊂ Rn on M and

a trivialization of E on U 0 is
b0 (x, v) = φ(κ(x))b(κ(x), Φ(x)v), x ∈ F 0 (U ∩ U 0 ) (5.74)
50 PETER HINTZ

for some diffeomorphism κ : F 0 (U ∩ U 0 ) → F (U ∩ U 0 ) and a smooth map Φ : U 0 → U . Let

ψ ∈ Cc∞ (F 0 (U 0 )). Then χ(x) := ψ(x)φ(κ(x)) ∈ Cc∞ (F 0 (U 0 ) ∩ κ−1 (F (U ))) = Cc∞ (F 0 (U ∩ U 0 )),
and we need to check that
χ(x)b(κ(x), Φ(x)v) ∈ S m (Rn ; Rk ). (5.75)
This however follows from the same type of calculation as (5.14). 

In analogy with Theorem 5.26, we have:

S
Corollary 5.34. Let M = i Ui be a locally finite open cover by coordinate charts Fi : Ui →
Fi (Ui ) ⊂ Rn , with Ui compact, over which E has a trivialization τi : π −1 (Ui ) → Fi (Ui ) × Rk .
Let a ∈ C ∞ (E). Then a ∈ S m (E) if and only if there exist symbols bi ∈ S m (Rn ; Rk ) and
χi ∈ Cc∞ (Fi (Ui )) such that
X
a= (τi−1 )∗ (χi bi ). (5.76)
i

Now, given an operator A ∈ Ψm (M ), we expect its principal symbol to be an element of

the quotient space S m (T ∗ M )/S m−1 (T ∗ M ). We first define it locally. Let F : U → F (U ) ⊂
Rn be a coordinate chart with Ū compact, and let V ⊂ U be open with V̄ ⊂ U . Denote by
τ : TU∗ M → F (U ) × Rn the trivialization induced by F . Let χ ∈ Cc∞ (U ) be such that χ = 1
on V̄ . Then we put
aV := τ ∗ σL (F −1 )∗ χAχF ∗ |TV∗ M ∈ S m (TV∗ M ).



(5.77)
By the local (Rn ) theory, and in particular by Theorem 5.2, the equivalence class
[aV ] ∈ S m (TV∗ M )/S m−1 (TV∗ M ) (5.78)
is independent of the choice of χ, and of the coordinate system F covering a neighborhood
of V̄ . Moreover, if V 0 ⊂ V , then restriction to V 0 gives [aV ]|V 0 = [aV 0 ].
Definition 5.35. The principal symbol of A ∈ Ψm (M ) is the unique element
σm (A) ∈ S m (T ∗ M )/S m−1 (T ∗ M ) (5.79)
with the following property: if a ∈ S m (T ∗ M ) is any representative of σm (A), and V is as
above, then [a|TV∗ M ] = [aV ] ∈ S m (TV∗ M )/S m−1 (TV∗ M ).

We start by proving existence. (Effectively, we are proving that U 7→ S m (TU∗ M )/S m−1 (TU∗ M )
is a sheaf.) This follows easily from the properties of the [aV ]. Indeed, taking a locally
finite subcover {Vi } of the cover
P of M by all sets V as above, and a subordinate partition
of unity {φi }, we have a = i φi aVi ∈ S m (T ∗ M ) by Corollary 5.34; we then put
σm (A) := [a]. (5.80)
We check that this satisfies the property required in Definition 5.35. Given V open as above,
it suffices to show that for φ ∈ Cc∞ (V ), we have [φa|V ] = [φaV ]. Now φφi aVi = φφi aV + ei
for some ei ∈ S m−1 (T ∗ M ) with support in TV∗i ∩V M . Let φ̃i ∈ Cc∞ (Vi ) be equal to 1 on
supp φi , and with supp φ̃i locally finite; then
X X X
φa|V = φ̃i (φi φaVi ) = φ̃i (φi φaV + ei ) = φaV + φ̃i ei , (5.81)
i i i

∈ S m−1 (T ∗ M )).
P
as desired (since i φ̃i ei
 MICROLOCAL ANALYSIS 51

We now turn to the uniqueness part of Definition 5.35; it suffices to show that if a ∈
S m (T ∗ M ) is such that a|TV∗ M ∈ S m−1 (TV∗ M ) for open sets V ⊂ M as above, then a ∈
S m−1 (T ∗ M ). But this follows by writing a = i φi a|TV∗ M , where φi , Vi are as above.
P
i

Proposition 5.36. The principal symbol map gives a short exact sequence
σ
→ S m (T ∗ M )/S m−1 (T ∗ M ) → 0.
0 → Ψm−1 (M ) → Ψm (M ) −−m (5.82)

Proof. We only prove surjectivity of σm . Take a partition of unity {φi } subordinate to a

locally finite cover of M by coordinate charts Fi : Ui → Fi (Ui ) ⊂ Rn with Ui compact.
m ∗
P
Writing any a ∈ S (T M ) as a = i φi a, it suffices to show that there exists an operator
Ai ∈ Ψm (M ) with Schwartz kernel supported in Ui × Ui such that σm (Ai ) = [φi a], as we
can then take A = i Ai (which is a locally finite sum). This is easy: if φ̃i ∈ Cc∞ (Ui ),
P

φ̃i = 1 on supp φi , then simply take



Ai = Fi∗ Op (Fi−1 )∗ φi (x)a(x, ξ) (Fi−1 )∗ φ̃i (y) (Fi−1 )∗ .


(5.83)

An immediate consequence of the Rn result, Proposition 4.21, is:
Proposition 5.37. The principal symbol map is multiplicative: if A ∈ Ψm (M ), B ∈
0
Ψm (M ), with at least one of them properly supported, then
σm+m0 (A ◦ B) = σm (A)σm0 (B). (5.84)

The analogue of Proposition 4.22 concerning the principal symbol of commutators will
be discussed in §5.13.

5.6. Quantization. There is no completely natural way, in general, to quantize symbols

on T ∗ M to pseudodifferential operators on M . However, S we do have the following useful
construction: fix a locally finite open cover M = i Ui by coordinate charts Fi : Ui →
Fi (Ui ) ⊂ Rn with Ui compact, and let {φi }, φi ∈ Cc∞ (Ui ), be a partition of unity subordinate
to {Ui }. Fix φ̃i ∈ Cc∞ (Ui ) with φ̃i = 1 near supp φi . Given a ∈ S m (T ∗ M ), define then
X
Fi∗ OpL (Fi−1 )∗ (φi a) φ̃i (Fi−1 )∗ ,


Op(a) := (5.85)
i
where OpL : S m (T ∗ Rn ) → Ψm (Rn ) is the left quantization map. By Theorem 5.26, the
formula (5.85) defines a map
Op : S m (T ∗ M ) → Ψm (M ). (5.86)
Proposition 5.38. The map Op in (5.86) is continuous, linear, and takes values in the
subspace of properly supported operators. Moreover, Op : S m (T ∗ M ) → Ψm (M ) is surjective
modulo Ψ−∞ (M ); that is, Ψm (M ) = Op(S m (T ∗ M )) + Ψ−∞ (M ).

Proof. We only sketch a proof of the final claim. It follows from Theorem 5.26. Indeed,
in the notation
P ∗ of equation (5.59), we can write any A ∈ Ψm (M ) in the form Am :=
A − K = i Fi Bi (Fi−1 )∗ . Using a partition of unity, we can combine the symbols of the
operators Bi ∈ Ψm (Rn ) into a symbol am ∈ S m (T ∗ M ); by the coordinate invariance of
the principal symbol, we then have Am−1 := Am − Op(am ) ∈ Ψm−1 (M ). We may then
apply Theorem 5.26 to Am−1 . An inductive argument thus produces am−j ∈ S m−j (T ∗ M ),
j ∈ N, so that Am−j−1 := Am−j − Op(am−j ) ∈ Ψm−j (M ). Letting a ∈ S m (T ∗ M ) be an
52 PETER HINTZ

asymptotic sum of the amj , j ∈ N0 , we therefore have A = Am + K = Op(a) + K + R

where K ∈ C ∞ (M ; ΩR ) and R ∈ Ψ−∞ (M ) are residual operators, and thus so is their sum
K + R. 

5.7. Operators acting on sections of vector bundles. The reader might ask why we
have not discussed adjoints of A ∈ Ψm (M ) (or even A ∈ Diff m (M )) yet. Since we do not
have an invariant way of integrating functions on M , the only sensible way to define A∗ is
by Z Z
(Au)(x)v(x) = u(x)A∗ v(x), u ∈ Cc∞ (M ), v ∈ Cc∞ (M ; ΩM ), (5.87)
M M
that is, ∗
A should be an operator acting on sections of ΩM . We leave it to the reader to
define the space of m-th order differential operators Diff m (M ; E, F ) mapping sections of E
to section of F , and go straight for the pseudodifferential version.
Definition 5.39. Let M be a smooth manifold, and let πE : E → M , πF : F → M denote
two real vector bundles of rank kE , kF . Then Ψm (M ; E, F ) is the space of linear operators
A : Cc∞ (M ; E) → C ∞ (M ; F ) (5.88)
with the following properties:
(1) if φ, ψ ∈ C ∞ (M ) have supp φ ∩ supp ψ = ∅, then there exists K ∈ C ∞ (M 2 ; πL∗ F ⊗
πR∗ (E ∗ ⊗ ΩM )) such that φAψ = K.

(2) Let U ⊂ M be any open set, G : U → G(U ) ⊂ Rn a diffeomorphism, and let

−1
τE : πE (U ) → G(U ) × RkE , τF : πF−1 (U ) → G(U ) × RkF local trivializations of E, F .
Using τE , identify smooth sections of E over U with kE -tuples of smooth functions
on U , likewise for sections of F . If ψ ∈ Cc∞ (U ), then there exists a kF × kE matrix
B = (Bij ) of ps.d.o.s Bij ∈ Ψm c (G(U )) such that on U
kE 
X

ψA(ψu)(x)i = G∗ Bij (G−1 )∗ u )j , u ∈ Cc∞ (M ; E), i = 1, . . . , kF . (5.89)
j=1

In the special case F = E, we write Ψm (M ; E) = Ψm (M ; E, E).

In local coordinates and trivializations, the symbol of A ∈ Ψm (M ; E, F ) is a symbol with

values in linear maps from RkE to RkF . The invariant definition is as follows. Denote by
π : T ∗ M → M the projection. Given a vector bundle G → M , we can consider its pullback
π ∗ G → T ∗ M ; a trivialization of G over an open set U ⊂ M , so G|U ∼ = U × RkG , then
induces a trivialization of π G over TU M which is ‘constant in the fibers of T ∗ M ’, namely
∗ ∗

(π ∗ G)|TU∗ M ∼
= T ∗ U × RkG , (5.90)
by identifying (π ∗ G)(x,ξ) = Gx ∼
= RkG using the local trivialization. We then denote by
S m (T ∗ M ; π ∗ G) ⊂ C ∞ (T ∗ M ; π ∗ G) (5.91)
the space of all smooth functions which in local coordinates and in a trivialization of G
(which induces a trivialization of π ∗ G as in (5.90)) are kG -vectors of symbols on Rn of order
m. Invariantly then,
σm (A) ∈ (S m /S m−1 )(T ∗ M ; π ∗ Hom(E, F )), π : T ∗ M → M. (5.92)
 MICROLOCAL ANALYSIS 53

This means that a representative of σm (A) is a map assigning to (x, ξ) ∈ T ∗ M an element

of Hom(Ex , Fx ). We have a short exact sequence
σ
→ (S m /S m−1 )(T ∗ M ; π ∗ Hom(E, F )) → 0. (5.93)
0 → Ψm−1 (M ; E, F ) → Ψm (M ; E, F ) −−m
The results from §§5.4–5.5 carry over; moreover, one can define adjoints:
Proposition 5.40. Let M be a smooth manifold, and let E, F, G → M denote three vector
bundles.
0
(1) Let A ∈ Ψm (M ; F, G) and B ∈ Ψm (M ; E, F ), with at least one of them properly
0
supported. Then A ◦ B ∈ Ψm+m (M ; E, G), and σm+m0 (A ◦ B) = σm (A) ◦ σm0 (B).7
(2) Let A ∈ Ψm (M ; E, F ). Then the (real) adjoint AT , defined by
Z Z
(Au)v = u(AT v), u ∈ Cc∞ (M ; E), v ∈ Cc∞ (M ; F ∗ ⊗ ΩM ), (5.94)
M M
is a pseudodifferential operator,
AT ∈ Ψm (M ; F ∗ ⊗ ΩM, E ∗ ⊗ ΩM ). (5.95)
It is properly supported if A is.
Remark 5.41. (1) If E, F are complex vector bundles with a anti-linear involution
(‘complex conjugation’), then one can define the adjoint A∗ similarly to (5.94),
but with complex conjugation of the second factor; one then has
σ(A∗ ) = σ(A)∗ . (5.96)
This in particular applies to the case that E = F = M × C, so sections of E, F are
simply complex-valued functions on M , which we discussed in (5.94).
(2) An inner product on E induces an isomorphism E ∗ ∼ = E (anti-linear when the inner
product is sesquilinear). If one moreover chooses a trivialization of ΩM , e.g. from
a semi-Riemannian metric, then AT ∈ Ψm (M ; F, E) (and A∗ ∈ Ψm (M ; E, F ) in the
complex case).

A consequence of (5.95) is the following extension of Proposition 5.24:

Corollary 5.42. Let A ∈ Ψm (M ; E, F ). Then A extends to a bounded linear operator
A : E 0 (M ; E) → D 0 (M ; F ). If A is properly supported, then A also maps D 0 (M ; E) →
D 0 (M ; F ), and by restriction C ∞ (M ; E) → C ∞ (M ; F ).

Proof. AT is a bounded map Cc∞ (M ; F ∗ ⊗ ΩM ) → C ∞ (M ; E ∗ ⊗ ΩM ). Therefore we can

define A : E 0 (M ; E) → D 0 (M ; F ) by duality using (5.95); this agrees with the original
operator A when restricted to Cc∞ (M ; E).
If A is properly supported, then AT maps Cc∞ (M ; F ∗ ⊗ ΩM ) → Cc∞ (M ; E ∗ ⊗ ΩM ), hence
we can now define A : D 0 (M ; E) → D 0 (M ; F ) by duality. Now C ∞ (M ; E) ⊂ D 0 (M ; E). It
∞ ∞
P to check that Au ∈ C (M ; F ) when u ∈ C (M ; E); but using a partition of unity
remains
1 = i φi , with supp φi compact, we have
!
X X
Au = A φi u = A(φi u), (5.97)
i i

7Note that for operators acting between bundles, composition is no longer commutative on the level of
principal symbols.
54 PETER HINTZ

with convergence in D 0 (M ; F ). But since A is properly supported, the final sum is a locally
finite sum of smooth terms, hence smooth. 

5.8. Special classes of operators. Let M be a manifold, and let E, F → M denote two
vector bundles of rank kE , kF .
Definition 5.43. Let m ∈ C. The subspace Ψm cl (M ; E, F ) ⊂ Ψ
Re m (M ; E, F ) of classi-

cal pseudodifferential operators consists of those operators whose full symbol in a local
coordinate chart and in local trivializations of E, F is a kF × kE matrix of classical sym-
bols of order m. The principal symbol map on Ψm cl (M ; E, F ) records the leading order
homogeneous part,
∗ ∗
σm : Ψ m m
cl (M ; E, F ) → Shom (T M \ o; π Hom(E, F )). (5.98)

The reason this is a sensible definition is that classicality is preserved under local co-
ordinate transformations; this follows from the proof of Theorem 5.2, in particular equa-
tion (5.13). Using the Rn results such as Proposition 4.24, one easily checks that the
composition of two classical ps.d.o.s (at least one of which is properly supported) is again a
classical ps.d.o., and that taking adjoints preserves classicality as well. A class of a classical
ps.d.o.s is of course given by differential operators:
Diff m (M ; E, F ) ⊂ Ψm
cl (M ; E, F ). (5.99)
Moreover, parametrices of classical operators are again classical.
Often, operators arising in geometric problems are Laplace operators to leading order,
such as the Hodge Laplacian (5.117). A very useful generalization of this is the following.
Definition 5.44. Let m ∈ R and A ∈ Ψm (M ; E). Then A is principally scalar if its
principal symbol is multiplication by scalars on the fibers of E, that is, if there exists a
symbol a ∈ S m (T ∗ M ) such that σm (A)(x, ξ) = a(x, ξ) IdEx .

Principally scalar operators behave similarly to operators acting on scalar functions; we

shall see examples of this in §8.

5.9. Elliptic operators on compact manifolds, Fredholm theory. Let M be a com-

pact n-dimensional manifold (without boundary), and let E, F → M denote two vector
bundles.
Definition 5.45. We say that A ∈ Ψm (M ; E, F ), with principal symbol a = σm (A),
is elliptic if there exists a symbol b ∈ S −m (T ∗ M ; π ∗ Hom(F, E)) such that ab − 1 ∈
S −1 (π ∗ End(F )), ba − 1 ∈ S −1 (T ∗ M ; π ∗ End(E)).
Remark 5.46. By ‘abstract group theory’ as in (4.87), the following seemingly more general
definition is in fact equivalent to the ellipticity of A: there exist b, b0 ∈ S −m (π ∗ Hom(F, E))
such that ab − 1 ∈ S −1 (π ∗ End(F )) and b0 a − 1 ∈ S −1 (π ∗ End(E)).
Theorem 5.47. Let A ∈ Ψm (M ; E, F ) be an elliptic operator.
(1) Then A : C ∞ (M ; E) → C ∞ (M ; F ) is Fredholm, that is, ker A is finite-dimensional,
and ran A is closed and has finite codimension.
 MICROLOCAL ANALYSIS 55

(2) Let ν ∈ C ∞ (M ; ΩM ) be a smooth positive measure on M , and suppose E, F are

equipped with positive definite fiber metrics. Define the generalized inverse of A by
Gf = u if f ∈ ran A, Au = f, u ⊥ ker A in L2 (M ; E; ν),
(5.100)
Gf = 0 if f ⊥ ran A in L2 (M ; F ; ν).

Then G ∈ Ψ−m (M ; F, E), and

GA = I − πN , AG = I − πR , (5.101)
where πN ∈ Ψ−∞ (M ; E) is the L2 (M ; E; ν)-orthogonal projection to the finite-
dimensional space ker A, and πR ∈ Ψ−∞ (M ; F ) is the L2 (M ; F ; ν)-orthogonal pro-
jection to the finite-dimensional space (ran A)⊥ . In particular, if A is invertible,
then G = A−1 ∈ Ψ−m (M ; F, E).

Proof. The elliptic parametrix construction, see Theorem 4.26, works in this setting as well
(see Exercise 5.11). Thus, there exists B ∈ Ψ−m (M ; F, E) such that
R1 = AB − I ∈ Ψ−∞ (M ; F ), R2 = BA − I ∈ Ψ−∞ (M ; E). (5.102)
We show that dim ker A < ∞. First, note that
u ∈ D 0 (M ; E), Au = 0 =⇒ u = (BA − R2 )u = −R2 u ∈ C ∞ (M ; E). (5.103)
Let us look at this from the point of view that the identity map on ker A ⊂ L2 (M ; E; ν) can
be written as I = BA−R2 = −R2 . Now R2 : L2 (M ; E; ν) → C ∞ (M ; E), hence is compact as
a map L2 (M ; E; ν) → L2 (M ; E; ν) by Arzelà–Ascoli. Therefore, the unit ball in the closed
subspace ker A ⊂ L2 (M ; E; ν) is compact, thus ker A ⊂ L2 (M ; E; ν) is finite-dimensional,
as desired.
Next, we show that ran A is closed. Suppose fj = Auj → f ∈ C ∞ (M ; F ), uj ∈ C ∞ (M ; E).
We may change uj by an element of ker A to ensure that uj ⊥ ker A. We have
uj = BAuj − R2 uj = Bfj − R2 uj . (5.104)
Suppose that, along some subsequence, kuj kL2 → ∞. Then
   
uj fj uj
=B − R2 . (5.105)
kuj k kuj k kuj k
This is bounded in C ∞ (M ; E), hence we can pass to a subsequence which converges in
L2 (M ; E; ν), say uj /kuj k → u ∈ L2 (M ; E; ν). Then Au = limj→∞ fj /kuj k = 0, so u ∈
ker A, but also u ⊥ ker A by construction. Since kukL2 = 1, this is a contradiction.
Therefore, kuj kL2 is bounded. Equation (5.104) then shows that uj is bounded in
C ∞ (M ; E), hence has a subsequence converging to u ∈ C ∞ (M ; E), and
Au = lim Auj = lim fj = f. (5.106)
j→∞ j→∞

Finally, we prove that ran A has finite codimension: but this follows from (ran A)⊥ =
ker A∗ and the ellipticity of A∗ ∈ Ψm (M ; F, E). (We are fixing arbitrary choices of smooth
positive measure on M and positive definite fiber metrics on E, F here; the same argument
goes through with minor notational changes even if one does not make such choices.)
56 PETER HINTZ

Let πN : L2 (M ; E; ν) → L2 (M ; E; ν) denote the orthogonal projection onto ker A. Fixing

an orthonormal basis {u1 , . . . , uJ } ⊂ C ∞ (M ; E) of ker A, we thus have
J
X
πN = uj h−, uj iL2 (M ;E;ν) , (5.107)
j=1

which indeed has smooth Schwartz kernel. An analogous argument shows that the orthog-
onal projection πR : L2 (M ; F ; ν) → L2 (M ; F ; ν) onto (ran A)⊥ = ker A∗ ⊂ C ∞ (M ; F ) has
smooth Schwartz kernel. Therefore, πN , πR ∈ Ψ−∞ .
The statement G ∈ Ψ−m (M ; F, E) for the generalized inverse (5.100) then follows by
writing
G = G(AB − R1 )
= (I − πN )B − GR1
(5.108)
= (I − πN )B − (BA − R2 )GR1
= (I − πN )B − B(I − πR )R1 + R2 GR1 .
Indeed, the first summand lies in Ψm (M ; F, E), the second in Ψ−∞ (M ; F, E), and the last
one is a smoothing operator, hence lies in Ψ−∞ (M ; F, E) as well. 

As a typical example, we discuss the Laplace operator on a compact n-dimensional man-

ifold M , which we assume to be connected for convenience. Denote by S 2 T ∗ M the second
symmetric tensor product of T ∗ M with itself. Let g ∈ C ∞ (M ; S 2 T ∗ M ) be a Riemannian
metric, so in local coordinates
n
X 1
g= gij (x) · (dxi ⊗ dxj + dxj ⊗ dxi ), gij (x) = gji (x). (5.109)
2
i,j=1

Write g ij (x) = g −1 (x)ij and |g| = | det(gij )|. Then the (scalar) Laplace operator is
n
X
|g|−1/2 Dxi |g|1/2 g ij (x)Dxj u


∆g u = (5.110)
i,j=1

in local coordinates. Thus, ∆g ∈ Ψ2 (M ), with

n
X
σ2 (∆g )(x, ξ) = g ij (x)ξi ξj =: |ξ|2g−1 (x) . (5.111)
i,j=1

Thus, ∆g is elliptic. By Theorem 5.47, the kernel and cokernel of ∆g are finite-dimensional.
Moreover, ∆g is a symmetric operator with respect to the inner product on L2 (M ; |dg|),
where |dg| ∈ C ∞ (M ; ΩM ) is defined in local coordinates by
|dg| = |g(x)|1/2 dx. (5.112)
Thus, ker ∆g = (ran ∆g )⊥ ; and if u ∈ ker ∆g , then
Z Z
0= (∆g u)ū |dg| = |∇u|2g |dg|, (5.113)
M M
so u is constant. In the notation of Theorem 5.47, we thus have
1
πN = h·, 1i1 = πR (5.114)
vol(M )
 MICROLOCAL ANALYSIS 57

(projection onto constants).

Let us study
∆g u = f, f ∈ D 0 (M ). (5.115)
−2
Let G ∈ Ψ (M ) denote the generalized inverse of ∆g . In the notation of Theorem 5.47,
we then have
u = (G∆g + πN )u = Gf + πN u. (5.116)
This solves (5.115) if and only if f = ∆g u = ∆g Gf + ∆g πN u = (I − πR )f . This shows:
Proposition 5.48. The equation (5.115) has a solution u ∈ D 0 (M ) if and only if hf, 1i = 0,
and in this case u is unique up to additive constants. If f ∈ C ∞ (M ), then u ∈ C ∞ (M ).
Example 5.49. For the operator A = ∆g + 1 ∈ Ψ2 (M ) on a compact Riemannian manifold,
one finds ker A = 0 = (ran A)⊥ , thus one can always solve Au = f for f ∈ C ∞ (M ) or
f ∈ D 0 (M ), with solution u ∈ C ∞ (M ) or u ∈ D 0 (M ).
Example 5.50. One can define natural generalizations of ∆g which act on vector bundles
rather than functions. Let dk ∈ Diff 1 (M ; Λk T ∗ M ; Λk+1 T ∗ M ) denote the exterior derivative,
and denote by δk ∈ Diff 1 (M ; Λk T ∗ M ; Λk−1 T ∗ M ) the adjoint of dk−1 . Let dn = 0 and
δ0 = 0. Then the Hodge Laplacian in degree k is
∆k := δk+1 dk + dk−1 δk ∈ Diff 2 (M ; Λk T ∗ M ). (5.117)
Its principal symbol is scalar, i.e. at each (x, ξ) ∈ T ∗ M a multiple of the identity operator on
(π ∗ Λk T ∗ M )(x,ξ) ; in fact σ2 (∆k )(x, ξ) = |ξ|2g−1 (x) Id. (The expression (5.110) is the special
case k = 0.) Again ∆k is symmetric with respect to the fiber inner product and volume
density induced by g. Its kernel and orthocomplement of the range are finite-dimensional,
and can be identified with the singular cohomology group H k (M ; C) by Hodge theory. For
a general version of this, see Exercise 5.18.

5.10. Sobolev spaces on manifolds. We need two key facts about Sobolev spaces H s (Rn )
for the generalization of Sobolev spaces to manifolds. For an open set Ω b Rn , we define
Hcs (Ω) := {u ∈ H s (Rn ) : supp u ⊂ Ω}. (5.118)
Lemma 5.51. Sobolev spaces on Rn have the following properties.
(1) Let a ∈ Cb∞ (Rn ). Then multiplication by a is a bounded linear map H s (Rn ) →
H s (Rn ) for all s ∈ Rn .
(2) Suppose κ : Ω → Ω0 is a diffeomorphism of precompact open subsets Ω, Ω0 b Rn .
Then κ∗ : Hcs (Ω0 ) → Hcs (Ω). Here, the pullback of a distribution u ∈ D 0 (Rn ) with
support in Ω0 is defined via duality using the formula
hκ∗ u, φi = hu, | det(κ−1 )0 |(κ−1 )∗ φi, φ ∈ Cc∞ (Ω).

Proof. The ‘standard’ proof of the first claim proceeds by proving it for s ∈ N0 using the
Leibniz rule, then for all real s ≥ 0 by complex interpolation, and then for all s ∈ R by
duality. With the machinery of §4 at hand, we can instead just observe that a ∈ Ψ0 (Rn ),
and appeal to Corollary 4.32.
The second claim is clear for s = 0. We shall prove it for general s ∈ R using our
ps.d.o. machinery. Indeed, given u ∈ Hcs (Ω0 ) ⊂ E 0 (Ω0 ), we certainly have κ∗ u ∈ E 0 (Ω).
Let A ∈ Ψs (Rn ) be elliptic, and let φ, φ̃ ∈ Cc∞ (Ω) be such that φ = φ̃ = 1 on supp(κ∗ u),
58 PETER HINTZ

and such that φ̃ = 1 in a neighborhood of supp φ. By choosing A carefully (localizing its

Schwartz kernel near the diagonal—which does not affects its ellipticity property), we may
arrange that
A(κ∗ u) = φ̃Aφκ∗ u. (5.119)
Note that φ̃Aφ ∈ Ψsc (Ω). Therefore, by Theorem 5.2,
A(κ∗ u) = κ∗ (A0 u), A0 = (κ−1 )∗ φ̃Aφκ∗ ∈ Ψsc (Ω0 ). (5.120)
Therefore A0 u ∈ L2 (Ω0 ), hence κ∗ (A0 u) ∈ L2 (Ω), so A(κ∗ u) ∈ L2 (Rn ). Since A is elliptic,
Corollary 4.34 implies that κ∗ u ∈ H s (Rn ), as desired. 

The ‘local coordinate’ definition of Sobolev spaces on manifolds is then:

Definition 5.52. Let M be an n-dimensional manifold, s ∈ R. Then:
s (M ) as the space of all u ∈ D 0 (M ) such that for all coordinate charts
(1) We define Hloc
F : U → F (U ) ⊂ Rn on M , and all φ ∈ Cc∞ (U ), we have (F −1 )∗ (φu) ∈ H s (Rn ).
(2) We define Hcs (M ) = {u ∈ Hloc
s (M ) : supp u ⊂ M is compact}.

If M is compact, we write
H s (M ) = Hloc
s
(M ) = Hcs (M ). (5.121)

Lemma 5.51 shows that if u ∈ Hcs (F (U )) for some coordinate chart F : U → F (U ) ⊂ Rn

on M , then F ∗ u ∈ Hcs (M ).
Remark 5.53. One can equip Hloc s (M ) with the structure of a Fréchet space by using the
−1 ∗
seminorms k(Fi ) φi ukH s (Rn ) for any fixed countable cover of M by coordinate charts
Fi : Ui → Fi (Ui ) ⊂ Rn and a subordinate partition of unity {φi }, φi ∈ Cc∞ (Ui ). The
resulting topology is independent of the cover and the partition of unity.

The proof of Lemma 5.51 suggests a more intrinsic definition of Sobolev spaces on M .
Note first that the spaces L2loc (M ) and L2c (M ) are well-defined, independently of a choice
of integration measure on M . (On the other hand, the space L2 (M ), even as a set, is not
well-defined when M is non-compact without specified integration measure.)
Proposition 5.54. Let u ∈ D 0 (M ).
(1) Suppose u ∈ Hcs (M ). Then Au ∈ L2loc (M ) for all A ∈ Ψs (M ). If A is properly
supported, then A : Hcs (M ) → L2c (M ), Hloc
s (M ) → L2 (M ).
loc
(2) If Au ∈ Lloc (M ) for some properly supported elliptic operator A ∈ Ψs (M ), then
2
s (M ).
u ∈ Hloc

Proof. Suppose that F : U → F (U ) ⊂ Rn is a coordinate system on M , and let φ, φ̃ ∈ Cc∞ (U )

with φ̃ = 1 in a neighborhood of supp φ.
The first claim follows by writing
φAu = φAφ̃u + φA(1 − φ̃)u. (5.122)
Indeed, the second summand lies in C ∞ (M ) ⊂ L2loc (M ). The first summand can be evalu-
ated in local coordinates, and lies in L2c (M ).
 MICROLOCAL ANALYSIS 59

Turning to the second claim, we need to show that (F −1 )∗ (φu) ∈ E 0 (F (U )) lies in H s (Rn ).
Let B ∈ Ψs (Rn ) be elliptic. We can arrange for its Schwartz kernel to be supported so close
to the diagonal that
(1 − (F −1 )∗ φ̃)B((F −1 )∗ φ) = 0. (5.123)
−1 ∗ 2 n
By elliptic regularity, we need to establish B(F ) φu ∈ L (R ), which by (5.123) is
equivalent to
B 0 u ∈ L2c (M ), B 0 := φ̃F ∗ B(F −1 )∗ φ ∈ Ψs (M ). (5.124)
−s
Since A is elliptic, there exists a properly supported parametrix Q ∈ Ψ (M ) with I =
QA + R, where R ∈ Ψ−∞ (M ) is then also properly supported. Therefore,
B 0 u = B 0 (QA + R)u = (B 0 Q)(Au) + B 0 Ru. (5.125)
Now B0Q∈ Ψ0 (M ) is bounded on L2loc (M ), so (B 0 Q)(Au) ∈ L2loc (M ), while B 0 R ∈
Ψ (M ), so B 0 Ru ∈
−∞ C ∞ (M ). Therefore, B 0 u ∈ L2loc (M ). 
Corollary 5.55. Let A ∈ Ψm (M ). Then A is a bounded linear operator
s−m
A : Hcs (M ) → Hloc (M ). (5.126)
s (M ) → H s−m (M ).
If A is properly supported, then A : Hcs (M ) → Hcs−m (M ), Hloc loc

Proof. We only prove (5.126). Let Λ ∈ Ψs−m (M ) be properly supported and elliptic. By
the second part of Proposition 5.54, it suffices to show that Λ ◦ A : Hcs (M ) → L2loc (M ); but
this follows from Λ ◦ A ∈ Ψs (M ) and the first part of Proposition 5.54. 

On a compact manifold M , the space H s (M ) can be given the structure of a Hilbert

space:
Proposition 5.56. Let M be compact, and let s ∈ R. Fix a smooth positive volume density
on M . Then there exists A ∈ Ψs (M ) such that
hu, viH s (M ) := hAu, AviL2 (M ) , kuk2H s (M ) := hu, uiH s (M ) , (5.127)
gives H s (M ) the structure of a Hilbert space. The topology on H s (M ) is equal to the norm
topology of (H s (M ), k · kH s (M ) ).

Proof. Let s ≥ 0. Fix a smooth fiber metric k · k on T ∗ M , and let Λ0 ∈ Ψs/2 (M ) be an

operator with σs/2 (Λ0 )(x, ξ) = kξks/2 .8 Then Λ0 is elliptic, and so is
Λs := I + (Λ0 )∗ Λ0 ∈ Ψs (M ). (5.128)
By Theorem 5.47, Λs : C ∞ (M )
→ C ∞ (M )
is Fredholm. We claim that Λs is invertible on
C (M ). Indeed, Λs u = 0 implies kukL2 (M ) + kΛ0 uk2L2 (M ) = 0, hence u = 0. Since Λs is
∞ 2

symmetric (that is, hΛs u, f iL2 (M ) = hu, Λs f iL2 (M ) for u, f ∈ C ∞ (M )), this also shows that
Λs is surjective. The second part of Theorem 5.47 then implies that
Λ−s := Λ−1 −s
s ∈ Ψ (M ). (5.129)
Using Proposition 5.54, we conclude that Λs : H s (M ) → L2 (M ) and Λ−s : H −s (M ) →
L2 (M ) are isomorphisms.
For s ∈ R, we can thus take A = Λs . 
8Strictly speaking, one should smooth the right hand side out near ξ = 0 to get a smooth symbol; but
principal symbols only care about behavior for large ξ, hence we do not do this here.
60 PETER HINTZ

Remark 5.57. For s = 2k, k ∈ N, one can take Λ2k = (∆g + 1)k for any Riemannian metric
g on M . (In fact, this is true for any k ∈ R by a theorem of Seeley which states, as a special
case, that (∆g + 1)s ∈ Ψ2s (M ) for any s ∈ R. This operator is defined using the functional
calculus for self-adjoint operators.)

Adding vector bundles to this discussion requires only notational changes. Namely, if
E → M is a real/complex rank k vector bundle, we say that u ∈ D 0 (M ; E) lies in Hloc s (M ; E)

if and only if in local trivializations of E over coordinate charts on M , u is a k-vector of real-

valued/complex-valued elements of H s (Rn ). We let Hcs (M ; E) = Hloc s (M ; E) ∩ E 0 (M ; E)

as usual. We leave the statements and proofs of the generalizations of Proposition 5.54,
Corollary 5.55, and Proposition 5.56 to the reader.
Example 5.58. If M is n-dimensional and p ∈ M , then δp ∈ H s (M ; ΩM ) for all s < −n/2;
cf. Example (5.16).

5.11. Elliptic operators on compact manifolds, revisited. Throughout this section,

we denote by M a compact manifold.
0
Lemma 5.59. Let s0 < s. Then the inclusion H s (M ) ,→ H s (M ) is compact.

We leave the proof to the reader; it can be proved by localizing in coordinate charts and
using a suitable 9 analogue on Rn —in fact, one can use a special case of the first part of
Exercise 4.13.
We can now refine Theorem 5.47:
Theorem 5.60. Let A ∈ Ψm (M ; E, F ) be an elliptic operator. Then for any s ∈ R,
A : H s (M ; E) → H s−m (M ; F ) (5.130)
is Fredholm. Its kernel ker A is independent of s, and ker A ⊂ C ∞ (M ; E). Moreover, if
we fix a smooth positive density on M , the cokernel coker A can be identified with a subset
Y ⊂ C ∞ (M ; F ) independent of s, in the sense that f ∈ H s−m (M ; F ) lies in ran A if and
only if hf, giL2 (M ;F ) = 0 for all g ∈ Y .

Proof. If B ∈ Ψm (M ; F, E) denotes an elliptic parametrix, then AB = I + R1 and BA =

I + R2 with R1 , R2 ∈ Ψ−∞ as in (5.102). By Lemma 5.59, the errors R1 : H s (M ; F ) →
C ∞ (M ; F ) ,→ H s (M ; F ) and R2 : H s−m (M ; E) → C ∞ (M ; E) ,→ H s−m (M ; E) are compact
operators. Therefore, A is Fredholm. The regularity statement ker A ⊂ C ∞ (M ; E) is
a special case of (5.103). The solvability claim follows from Theorem 5.47 and elliptic
regularity. 

In fact, this theorem as a converse: if A ∈ Ψm (M ; E, F ) is such that (5.130) is Fredholm

for some s ∈ R, then A is elliptic.
Remark 5.61. Theorem 5.60 also shows that the index ind A = dim ker A − dim coker A is
independent of s. Simple functional analytic arguments show that ind A = ind(A + B) for
any B ∈ Ψm−1 (M ; E, F ); thus, ind A only depends on the principal symbol σm (A)! The
Atiyah–Singer index theorem gives a formula to compute ind A in terms of σm (A).

An interesting application concerns the spectral theory of symmetric ps.d.o.s.

9Note that the inclusion H s (Rn ) ,→ H s0 (Rn ) is not compact!
 MICROLOCAL ANALYSIS 61

Theorem 5.62. Fix a smooth positive volume density on M , and a positive definite fiber
inner product on E → M . Let m > 0, and let A ∈ Ψm (M ; E) be symmetric, that is,
hAu, vi = hu, Avi for u, v ∈ C ∞ (M ; E), where h·, ·i is the inner product on L2 (M ; E). Then
A is an unbounded self-adjoint operator on L2 (M ; E) with domain H m (M ; E). Its spectrum
spec A ⊂ R is discrete and accumulates only at ∞. There exists an orthonormal basis of
L2 (M ; E) consisting of eigenfunctions of A, all of which are smooth.

Proof. By [RS72, Theorem VIII.3], we need to show that A ± i : H m (M ; E) → L2 (M ; E) is

surjective. By Theorem 5.60, its range is closed, and any element u ∈ (ran(A ± i))⊥ lies in
ker(A ∓ i) ⊂ C ∞ (M ; E), so
0 = Imh(A ∓ i)u, ui = ∓ikuk2L2 (M ;E) =⇒ u = 0. (5.131)

This proves self-adjointness.

One can also argue directly: if A is given the domain D(A) = H m (M ; E), then v ∈
L2 (M ; E) lies in D(A∗ ) if and only if D(A) 3 u 7→ hAu, vi satisfies a bound |hAu, vi| ≤
CkukL2 for some C. But hAu, vi = hu, A∗ vi, hence we conclude that A∗ v ∈ L2 (M ; E), and
by elliptic regularity v ∈ H m (M ; E); thus D(A∗ ) ⊂ D(A). The converse is clear since A is
symmetric.
To prove the discreteness of the spectrum, note simply that (A + i)−1 : L2 (M ; E) →
H m (M ; E) ,→ L2 (M ; E) is a compact operator, hence its spectrum is discrete and can only
accumulate at 0. But Aφ = λφ implies (A + i)−1 φ = (λ + i)−1 φ, hence spec A can only
accumulate at ∞. 
Example 5.63. This applies to the Laplacian ∆g on any compact Riemannian manifold
(M, g), acting on functions or differential forms.
Example 5.64. There exist elliptic operators whose spectrum is the entire complex plane.
In fact, there exists an elliptic operator A ∈ Ψ1 (S1 ) with index 1 (or any other integer). By
Remark 5.61, A − λ is never invertible for any λ ∈ C.

5.12. A simple nonlinear example. As a simple (and naive, weak, and wasteful, but
instructive) nonlinear application of the elliptic theory developed thus far, we shall solve a
non-linear elliptic equation on a compact 2-dimensional manifold M . If g is a Riemannian
metric on M , we denote the Gauss curvature of M by Kg ∈ C ∞ (M ). If φ ∈ C ∞ (M ), then
the metric g 0 (x) = e2φ(x) g(x) is said to be conformal to g. The Gauss curvature of g 0 is
given by
Kg0 = e−2φ (Kg − ∆g φ). (5.132)
Proposition 5.65. Suppose (M, g) has constant Gauss curvature Kg ≡ −1. Let g̃ ∈
C ∞ (M ; S 2 T ∗ M ) be a Riemannian metric with kg − g̃kH 4 < ,  > 0 small. Then there
exists φ ∈ C ∞ (M ) such that e2φ g̃ has constant Gauss curvature −1.

This is a local version of the uniformization theorem; the conclusion holds for any metric
g̃, not necessarily close to g. The assumptions require that M is a manifold of genus at
least 2 (that is, a donut with at least two holes). For M ∼ = S2 , one can always find a
conformal multiple with constant curvature +1, and for M ∼ = T2 , one can always find one
with constant curvature 0.
62 PETER HINTZ

Proof of Proposition 5.65. We have H 4 (M ) ⊂ C 2 (M ) (in fact C 2,α (M ) for all α < 1); and
moreover the map H 4 (M ; S 2 T ∗ M ) 3 g 7→ Kg ∈ H 2 (M ; S 2 T ∗ M ) is smooth.
We want to solve the equation
−1 = Ke2φ g̃ = e−2φ (Kg̃ + ∆g̃ φ), (5.133)
or equivalently
∆g̃ φ + e2φ = Kg̃ . (5.134)
Since kKg̃ − Kg kH 2 is small, we expect φ ∈ H 4 (M ) to be small; this suggests Taylor
expanding:
Aφ = E(φ) − N (φ), A = ∆g + 2, E(φ) = Kg̃ − (∆g̃ − ∆g )φ, N (φ) = e2φ − 1 − 2φ. (5.135)
We solve this using the contraction mapping principle, i.e. by iterating the map
T : H 4 (M ) 3 φ 7→ A−1 (E(φ) − N (φ)) ∈ H 4 (M ). (5.136)
Now k∆g̃ − ∆g kL(H 4 (M ),H 2 (M )) ≤ C for some constant C, thus kE(φ)kH 2 ≤ C(1 +
kφkH 4 (M ) ). Moreover, kN (φ)kH 2 ≤ Ckφk2H 4 (M ) for kφkH 4 (M ) ≤ 1. Therefore, if kφkH 4 (M ) ≤
δ for some δ > 0, then
kT φkH 4 (M ) ≤ C 0 (C(1 + δ) + Cδ 2 ) ≤ δ, C 0 = kA−1 kL(H 2 (M ),H 4 (M )) , (5.137)
provided we take  = (δ) := δ/(10CC 0 ) and δ small enough. The map T is then a
contraction on the δ-ball in H 4 (M ), since
kT φ − T ψkH 4 (M ) ≤ C 0 Ckφ − ψkH 4 (M ) + Ckφ − ψkH 4 (M ) (kφkH 4 (M ) + kψkH 4 (M ) )

≤ C 0 (C + Cδ)kφ − ψkH 4 (M )

≤ 21 kφ − ψkH 4 (M )
(5.138)
for small enough δ > 0 and  = (δ).
Let now φ ∈ H 4 (M ), kφkH 4 (M ) ≤ δ, denote the unique fixed point of T ; then φ
solves (5.134). We rewrite this one last time as
∆g̃ φ = Kg̃ − e2φ . (5.139)
Suppose we already know φ ∈ H k (M ), k ≥ 4. Then the right hand side of this equation
k k+2 (M ). Therefore, φ ∈
T ink H (M ), ∞so by elliptic regularity, we conclude that φ ∈ H
lies
k H (M ) = C (M ), finishing the proof. 

5.13. Commutators and symplectic geometry. We tie up a loose end and describe,
invariantly, the principal symbol of the commutator of two ps.d.o.s. Key is the symplectic
structure of the cotangent bundle T ∗ M .
Definition 5.66. Let M be an n-dimensional manifold. The canonical 1-form on T ∗ M is
the section α ∈ C ∞ (T ∗ M ; T ∗ (T ∗ M )) defined by
α(x,ξ) (v) := ξ(π∗ v), x ∈ M, ξ ∈ Tx∗ M, v ∈ T(x,ξ) (T ∗ M ), (5.140)
where π : T ∗ M → M is the projection. The canonical symplectic form on T ∗ M is
ω := −dα ∈ C ∞ (T ∗ M ; Λ2 (T ∗ M )). (5.141)
 MICROLOCAL ANALYSIS 63

In local coordinatesPx ∈ Rn and corresponding canonical coordinates ξ ∈ Rn on the fibers

∗
P
of T M , we have π∗ ( k ak ∂xk + bk ∂ξk ) = k ak ∂xk , and therefore
n
X n
X
α= ξk dxk , ω= dxk ∧ dξk . (5.142)
k=1 k=1
This is a non-degenerate 2-form: contraction T (T ∗ M )
3 v 7→ v y ω = ω(v, −) ∈ T ∗ (T ∗ M )
is an isomorphism, and identifies vector fields and 1-forms on T ∗ M :
n n
yω
X X
ak ∂xk + bk ∂ξk 7−−→ −bk dxk + ak dξk . (5.143)
k=1 k=1

Definition 5.67. Let p ∈ C ∞ (T ∗ M ). Then the Hamiltonian vector field of p is the unique
Hp ∈ V(T ∗ M ) such that
Hp y ω = dp. (5.144)
∞ ∗
The Poisson bracket of p, q ∈ C (T M ) is defined as
{p, q} := Hp q = −Hq p. (5.145)

In local coordinates, we deduce from (5.143) that

n
X
Hp = (∂ξk p)∂xk − (∂xk p)∂ξk . (5.146)
k=1
Thus, the ‘ad hoc’ definition (4.76) in fact makes invariant sense on T ∗ M ! As a consequence
of the local Rn theory, Proposition 4.22, we thus deduce:
0
Corollary 5.68. Let A ∈ Ψm (M ), B ∈ Ψm (M ), at least one of which is properly supported.
0
Then [A, B] ∈ Ψm+m −1 (M ), and
σm+m0 −1 (i[A, B]) = {σm (A), σm0 (B)}. (5.147)

5.14. Exercises.
Exercise 5.1. In the notation of Example 5.7, prove that the map (5.22) is well-defined, i.e.
does not depend on the choice of coordinate system.
Exercise 5.2. Prove that the definition of the isomorphism (5.25) given in the subsequent
paragraph is independent of the choice of the local coordinate chart.
Exercise 5.3. Let V ∈ C ∞ (M, T M ) denote a vector field.
(1) For a smooth function f ∈ C ∞ (M ), define (V f )(p) := V (p)f for p ∈ M as the
directional derivative of f along V (p) (see Example 5.7). Show that V f ∈ C ∞ (M ).
Show moreover that the map f 7→ V f is a derivation, i.e. it satisfies the Leibniz
rule
V (f g) = f V (g) + gV (f ). (5.148)
∞ ∞ ∗
(2) Given f ∈ C (M ), note that df ∈ C (M, T M ). Show that df (V (p)) = V (p)f ,
where the left hand side is the dual pairing between Tp∗ M and Tp M (see (5.25)).
Exercise 5.4. Let E → M be a vector bundle. Let F → M be a subbundle of E, i.e. a
vector bundle over M with the property that Fx ⊂ Ex for all x ∈ M . Give a construction of
the quotient vector bundle E/F → M whose fibers are the quotient vector spaces Ex /Fx .
64 PETER HINTZ

Exercise 5.5. Given a smooth submanifold S ⊂ M of a manifold M , define its normal

bundle as the quotient bundle N S = TS M/T S. Show that the duality between T M and
T ∗ M induces a duality between N S and the conormal bundle N ∗ S ⊂ TS∗ M consisting of
all covectors which annihilate T S.
Exercise 5.6. Prove Lemma 5.11. (Hint. One method of proof is to analyze the transition
functions of the various bundles. The idea of another method is to prove the analogous
statement for the vector spaces Ωα V of Remark 5.10.)
Exercise 5.7. Prove Lemma 5.12 and the statement in Remark 5.13.
Exercise 5.8. (1) Let M be a manifold (compact or non-compact), andSlet u ∈ E 0 (M ).
Show that there exists s ∈ R so that u ∈ Hcs (M ). (Thus, E 0 (M ) = Hcs (M ).)
(2) Suppose M is non-compact. Show that there exists u ∈ D 0 (M ) so that u ∈ s (M )
/ Hloc
for any s ∈ R. (Thus, D 0 (M ) ) Hloc
s (M ).)
S

Exercise 5.9. Let A ∈ Ψm (M ). Show that there exists a properly supported operator
A0 ∈ Ψm (M ) with A − A0 ∈ Ψ−∞ (M ).
Exercise 5.10. Let M be a manifold, and let m ∈ R.
(1) Given a sequence of symbols aj ∈ S m−j (T ∗ M ), j ∈ N0 , show that there exists a
P −1
symbol a ∈ S m (T ∗ M ) so that for all N ∈ N0 , we have a − N
j=0 aj ∈ S
m−N (T ∗ M ).

(2) Given a sequence of operators Aj ∈ S m−j (M ), j ∈ N0 , show that there exists a

properly supported operator A ∈ Ψm (M ) so that for all N ∈ N0 , we have A −
PN −1 m−N (M ).
j=0 Aj ∈ Ψ

Exercise 5.11. Give a detailed proof of the existence of elliptic parametrices on manifolds.
That is, if A ∈ Ψm (M ) is elliptic, show that there exists a properly supported operator
B ∈ Ψ−m (M ) so that A ◦ B − I, B ◦ A − I ∈ Ψ−∞ (M ).
Exercise 5.12. Let M denote a smooth manifold, and denote by
d : C ∞ (M ; Λk T ∗ M ) → C ∞ (M ; Λk+1 T ∗ M ) (5.149)
the exterior derivative. Show that d ∈ Diff 1 (M ; Λk T ∗ M, Λk+1 T ∗ M ), and compute its
principal symbol.
Exercise 5.13. Let (M, g) denote a smooth Riemannian manifold, and denote by
∇ : C ∞ (M ; T M ) → C ∞ (M ; T ∗ M ⊗ T M ), V 7→ (∇V : X 7→ ∇X V ), (5.150)
the covariant derivative on vector fields. Show that ∇ is a first order differential operator,
and compute its principal symbol.
Exercise 5.14. Let Γ ⊂ C be a smooth, simple, closed curve. Let K ∈ C ∞ (Γ × Γ). Prove
that Z
K(t, s)
Au(t) := lim u(s) ds, u ∈ C ∞ (Γ) (5.151)
→0 |t−s|≥ t − s

is well-defined and defines an element A ∈ Ψ0cl (Γ). Here, t, s ∈ Γ ⊂ C are complex numbers,
and the division here is division by a complex number. Compute its principal symbol.
Exercise 5.15. Let (X, k · kX ), (Y, k · kY ) be two Banach spaces, and suppose A : X → Y is
a bounded linear map.
 MICROLOCAL ANALYSIS 65

(1) Suppose Z is another Banach space, and there is an inclusion (continuous injective
map) X ,→ Z which is compact. Suppose there exists a constant C > 0 such that
kukX ≤ C (kAukY + kukZ ) . (5.152)
Show that ker A ⊂ X is finite-dimensional, and that ran A ⊂ Y is closed.
(2) Suppose that, in addition, to (1), there exists a Banach space Z̃ and an inclusion
Y ∗ ,→ Z̃ which is compact. Suppose there exists C > 0 such that
kvkY ∗ ≤ C kA∗ vkX ∗ + kvkZ̃ .


(5.153)
Show that if f ∈ Y is such that v(f ) = 0 for all v ∈ ker A∗ , then there exists u ∈ X
with Au = f . Deduce that A is a Fredholm operator.
Exercise 5.16. Let M be a compact manifold, let E, F → M denote two vector bundles,
and let A ∈ Ψm (M ; E, F ).
(1) Suppose there exists a symbol b ∈ S −m (T ∗ M ; Hom(F, E)) such that bσm (A) −
1 ∈ S −1 (T ∗ M ; End(E)). Show that A : H s (M ; E) → H s−m (M ; F ) has finite-
dimensional kernel and closed range.
(2) Suppose σm (A) there exists b ∈ S −m (T ∗ M ; Hom(F, E)) such that σm (A)b − 1 ∈
S −1 (T ∗ M ; End(F )). Show that A : H s (M ; E) → H s−m (M ; F ) has closed range
and finite-dimensional cokernel.
Show also that if A has a homogeneous principal symbol σm (A) (so in particular when A
is a classical operator), the assumption in part (1) is equivalent to the injectivity of σm (A)
on T ∗ M \ o, and the assumption in part (2) to the surjectivity.
Exercise 5.17 (Helmholtz decomposition). Let (M, g) be a compact Riemannian manifold,
denote by d : C ∞ (M ) → C ∞ (M ; T ∗ M ) the exterior derivative acting on functions, and
denote by δg = d∗ its adjoint. Let ω ∈ H s (M ; T ∗ M ) be a 1-form. Prove that there exist
u ∈ H s+1 (M ) and η ∈ H s (M ; T ∗ M ) such that
ω = du + η, δg η = 0. (5.154)
(Note that d and δg are first order differential operators with smooth coefficients, and hence
they do act on distributions valued in the appropriate bundles.)
Exercise 5.18. Let M be a compact manifold, let Ei → M , i = 0, . . . , N , be complex vector
bundles, and suppose di ∈ Diff 1 (M ; Ei , Ei+1 ), i = 0, . . . , N − 1. Suppose they form a
complex of differential operators
d d dN −1
C ∞ (M ; E0 ) −→
0
C ∞ (M ; E1 ) −→
1
· · · −−−→ C ∞ (M ; EN ); (5.155)
that is, for each i < N ,
di+1 ◦ di = 0 ∈ Diff 2 (M ; Ei , Ei+2 ). (5.156)
Assume moreover that this complex is elliptic, meaning that the symbol complex
σ1 (d0 ) σ1 (d1 ) σ1 (dN −1 )
C ∞ (T ∗ M \ o; π ∗ E0 ) −−−−→ C ∞ (T ∗ M \ o; π ∗ E1 ) −−−−→ · · · −−−−−−→ C ∞ (T ∗ M \ o; π ∗ EN )
(5.157)
is exact (that is, ran σ1 (di−1 )(x, ξ) = ker σ1 (di ) for all i < N ). The goal of this exercise is
to study the cohomology groups
H i (E• ) := (ker di )/(ran di−1 ), i = 1, . . . , N − 1, (5.158)
using PDE theory.
66 PETER HINTZ

(1) Equip M with a volume density and the Ei with Hermitian fiber inner products;
define δi ∈ Diff 1 (M ; Ei , Ei−1 ) to be the adjoint of di−1 . Show that the ‘Laplacian’
∆i := di−1 ◦ δi + δi+1 ◦ di ∈ Diff 2 (M ; Ei ), 1 ≤ i ≤ N − 1, (5.159)
is elliptic and symmetric.
(2) Show that
ker ∆i = {u ∈ C ∞ (M ; Ei ) : di u = 0, δi u = 0}. (5.160)
(3) Show that the inclusion ker ∆i ,→ ker di induces an isomorphism of vector spaces
ker ∆i ∼
= H i (E• ). (5.161)
(4) Prove the Hodge theorem: if (M, g) is a compact Riemannian manifold, and ∆k ∈
Diff 2 (M ; Λk T ∗ M ) is the Hodge Laplacian on k-forms, then ker ∆k ∼
= H k (M ), where
k
H (M ) denotes the k-th de Rham cohomology group (with complex coefficients) of
M.

6. Microlocalization

We now turn to the second part of these lecture notes: finer properties of distributions,
and, closely related, non-elliptic phenomena. We develop the notion of distributional wave
front from the observation about the local nature of full symbolic expansions that we
observed e.g. after the statement of Theorem 4.16.
From now on, all ps.d.o.s shall either be properly supported, or elements of the uniform
ps.d.o. algebra on Rn .

6.1. Operator wave front set. Recall from (4.56) and (4.58) the full symbols, modulo
S −∞ , for adjoints and compositions: if A = OpL (a), B = OpL (b) are ps.d.o.s on Rn , then
X 1
σL (A∗ )(x, ξ) ∼ ∂ξα Dxα ā(x, ξ),
α!
α∈Nn
0
X 1 (6.1)
σL (A ◦ B)(x, ξ) ∼ ∂ξα a(x, ξ) · Dxα b(x, ξ).
n
α!
α∈N0

A key feature, which so far we have only exploited at the principal symbol level, is that
these formulas are local in (x, ξ). We would like to say that if A is ‘trivial’ at or near (x, ξ),
in the sense that a vanishes there, then A∗ and A ◦ B (for any B) are trivial there as well.
Unfortunately, since the expressions (6.1) are asymptotic sums, thus have no content at any
fixed point (x, ξ), the meaning of this is not immediately clear.
The correct notion of ‘triviality’ must depend on the behavior of symbols as |ξ| → ∞, and
must be insensitive to modifications by symbols of order −∞. This leads to the following
definition:
Definition 6.1. Let a ∈ S m (Rn ; RN ). Then a point (x0 , ξ0 ) ∈ Rn × (RN \ {0}) does not
lie in the essential support
ess supp a ⊂ Rnx × (RN
ξ \ {0}) (6.2)
 MICROLOCAL ANALYSIS 67

if and only if a is a symbol of order −∞ near x0 and in a conic neighborhood of ξ0 ; that is,
there exists  > 0 such that for all α ∈ Nn0 , β ∈ NN0 , k ∈ R, we have
 
α β −k
ξ ξ0 
|∂x ∂ξ a(x, ξ)| ≤ Cαβk hξi ∀ (x, ξ), |ξ| ≥ 1, |x − x0 | +  −
 < . (6.3)
|ξ| |ξ0 | 

It suffices, in fact, to assume (6.3) only for α = β = 0; the estimates for general α, β are
then automatic. See Exercise 6.1.
Remark 6.2. By definition, ess supp a is a closed subset of Rn × (RN \ {0}). Moreover,
ess supp a is conic in ξ, that is, (x, ξ) ∈ ess supp a implies (x, λξ) ∈ ess supp a for all λ > 0.
Definition 6.3. Let A = OpL (a). Then we define the operator wave front set of A as the
closed, conic set
WF0 (A) := ess supp(a) ⊂ Rnx × (Rnξ \ {0}). (6.4)

The following follows immediately from (6.1), the formulas for left/right reductions
in (4.34)–(4.35), as well as formula (5.13) in the proof of the local coordinate invariance of
ps.d.o.s:
Proposition 6.4. The operator wave front set for operators A, B ∈ Ψ(Rn ) has the following
properties:
(1) Suppose A has compactly supported Schwartz kernel. Then WF0 (A) = ∅ if and only
if A ∈ Ψ−∞ (Rn ).10
(2) Let A = OpR (a0 ). Then WF0 (A) = ess supp a0 .
(3) WF0 (A + B) ⊂ WF0 (A) ∪ WF0 (B).
(4) WF0 (A ◦ B) ⊂ WF0 (A) ∩ WF0 (B).
(5) WF0 (A∗ ) = WF0 (A).
(6) If Ω, Ω0 b Rn , κ : Ω → Ω0 is a diffeomorphism, A ∈ Ψc (Ω0 ), and Aκ = κ∗ A(κ−1 )∗ ,
then
WF0 (Aκ ) = κ∗ WF0 (A), (6.5)
∗ −1 0 T
where we define κ (x, ξ) = (κ (x), κ (x) ξ).

Of these, properties (4) and (5) were our motivation for the introduction of ess supp
above. Property (6) implies that the operator wave front set can be defined invariantly for
operators on manifolds:
Definition 6.5. Let M be a manifold and A ∈ Ψm (M ). Then WF0 (A) ⊂ T ∗ M \ o is the
closed conic subset (i.e. invariant under dilations in the fibers of T ∗ M ) given near Tp∗ M ,
p ∈ M , by WF0 (A0 ) where A0 ∈ Ψm (Rn ) is the expression for A in a local coordinate
system near p.

Properties (1) and (3)–(5) in Proposition 6.4 thus hold for ps.d.o.s on manifolds as
well; since on general manifolds we do not impose growth restrictions on symbols outside of
compact sets in the base, property (1) in fact holds without any assumption on the Schwartz
kernel of A. We leave the details of the definition of WF0 (A) for operators A ∈ Ψm (M ; E, F )
10We make the assumption on the Schwartz kernel merely to exclude scenarios where A = Op(a) has
empty wave front set, but the constants in the estimate (6.3) blow up as |x0 | gets large. An example is
given by a(x, ξ) = χ(hxihξi−1 ) where χ ∈ Cc∞ (Rn ). Indeed, a is a uniform symbol of order 0 on Rn , but
WF0 (Op(a)) = ∅ since a is locally in x a symbol of order −∞.
68 PETER HINTZ

acting between sections of vector bundles over M to the reader; in local coordinates and
trivializations, a point is in WF0 (A) if it is in the operator wave front set of at least one
entry of the matrix of ps.d.o.s on Rn representing A locally.
One thinks of WF0 (A) ⊂ T ∗ M \ o as the set in phase space where A is microlocally
non-trivial. This is a much weaker notion than having a (microlocally) elliptic principal
symbol, see §6.2.
Example 6.6. If A ∈ Ψm (M ) is elliptic, then WF0 (A) = T ∗ M \ o.
Example 6.7. Let A = |α|≤m aα (x)Dxα ∈ Diff m (Rn ). Then
P
 
[
WF0 (A) =  supp aα  × (Rn \ {0}). (6.6)
|α|≤m

Thus, differential operators never have ‘interesting’ operator wave front set.
Example 6.8. Let χ ∈ S m (Rnξ ), and consider the Fourier multiplier A = χ(D) := Op(χ).
Then WF0 (A) = Rnx × ess supp χ.
Example 6.9. We combine Examples 6.7 (for m = 0) and 6.8. Let φ ∈ Cc∞ (Rn ) and
χ ∈ S 0 (Rnξ ). Then
A := χ(D) ◦ φ(x) = OpR (χ(ξ)φ(y)) ∈ Ψ0 (Rn ) (6.7)
has WF0 (A) = (supp φ) × (ess supp χ).

Working with conic sets is a bit tedious. In most circumstances, one can simplify notation
by working on the cosphere bundle
S ∗ M := (T ∗ M \ o)/R+ , (6.8)
with fibers given by Sp∗ M = (Tp∗ M \ o)/R+ , where R+ acts by dilations in the fibers. Thus,
S ∗ M is a fiber bundle with typical fiber Sn−1 . We can identify conic subsets of T ∗ M \ o
with their image in S ∗ M . For instance, if A ∈ Ψm (M ), then for α ∈ Sp∗ M , the condition
α ∈ WF0 (A) means that (p, ξ) ∈ WF0 (A) where α = [ξ] (i.e. α = R+ ξ). Note that a
compact subset of S ∗ M is identified with a conic subset of T ∗ M \ o whose cross section
(i.e. intersection with |ξ| = 1 for some choice of fiber metric on T ∗ M ) is compact. The
projection map is denoted
π : S ∗ M → M. (6.9)
The following technical result states that one can construct partitions of unity microlo-
cally:
Lemma 6.10. Suppose S ∗ M = i Ui is an open cover. Then there exist operators Ai ∈
S
Ψ0 (M ) such that
(1) the supports of the Schwartz kernels of Ai are locally finite,
WF0 (Ai ) ⊂ Ui ,
(2) P
(3) i Ai = I.

Proof. It suffices to prove this for a locally finite refinement of the cover, which we shall
denote by {Ui } still, for which moreover each Ui lies over a coordinate chart, i.e. Ui ⊂ SV∗i M
with Fi : Vi → Fi (Vi ) ⊂ Rn a chart, and with {Vi } locally finite.
 MICROLOCAL ANALYSIS 69

Pick a partition of unity {χi } subordinate to the cover {Ui }; fix ψi , ψ̃i ∈ Cc∞ (Fi (Vi )) such
that ψi ≡ 1 near Fi (π(supp χi )), and ψ̃i ≡ 1 near supp ψi . We then put
 
A0i := Fi∗ ψ̃i Op(χi )ψi (Fi−1 )∗ . (6.10)

Then (1) and (2) are satisfied for A0iP , but rather than (3) we only have i A0i = I − R0 ,
P
R0 ∈ Ψ−1 (M ). Thus, simply let B ∼ ∞ 0 j 0
j=0 (R ) and put Ai := Ai B. (This still satisfies (1)
and (2), in the former caseP since B is properly supported, and in the latter case by part (4)
of Proposition 6.4.) then i Ai = I − R, R ∈ Ψ−∞ (M ). Replacing any single one of the
Ai by Ai + R, we are done. 
Corollary 6.11. Let K b U ⊂ S ∗ M , with U open. Then there exists A ∈ Ψ0 (M ) such
that WF0 (A) ⊂ U and WF0 (I − A) ∩ K = ∅.

We say that A is microlocally equal to I on K.

Proof of Corollary 6.11. S ∗ M = U ∪(S ∗ M \K) is an open cover of S ∗ M , hence there exists
a partition of unity I = A+B with WF0 (A) ⊂ U and WF0 (I −A)∩K = WF0 (B)∩K = ∅. 

6.2. Elliptic set, characteristic set. We next refine the notion of ellipticity of operators
and symbols in a microlocal manner analogous to ess supp and WF0 .
Definition 6.12. Let A ∈ Ψm (M ). Then the elliptic set of A,
Ell(A) ⊂ T ∗ M \ o (6.11)
(or Ellm (A) if one wants to make the order explicit), consists of all (x0 , ξ0 ) ∈ Rn ×(Rn \{0})
in a conic neighborhood of which σm (A) is elliptic; that is, in local coordinates and picking
a representative of σm (A), there exist c, C > 0 and  > 0 such that
 
m
ξ ξ0 
|σm (A)(x, ξ)| ≥ c|ξ| , |ξ| ≥ C, |x − x0 | +  −
 < . (6.12)
|ξ| |ξ0 | 
The complement of Ell(A) is the characteristic set
Char(A) := (T ∗ M \ o) \ Ell(A). (6.13)

An equivalent definition of Ell(M ), closer to Definition 3.8, is that there exists b ∈

S −m (T ∗ M ) such that σm (A)b − 1 is a symbol of order −1 in a conic neighborhood of
(x0 , ξ0 ). Note that Ell(A) is automatically open. Moreover,
0
Ellm+m0 (A ◦ B) = Ellm (A) ∩ Ellm0 (B), A ∈ Ψm (M ), B ∈ Ψm (M ). (6.14)
Example 6.13. Elliptic operators on M have elliptic set equal to T ∗ M \ o, and empty
characteristic set.
Example 6.14. On R1+n = Rt × Rnx , with canonical momentum variables σ ∈ R, ξ ∈ Rn , the
wave operator  = Dt2 − Dx2 has principal symbol σ2 () = σ 2 − ξ 2 , hence its characteristic
set is the double cone
Char() = {(t, x, σ, ξ) : σ 2 − |ξ|2 = 0, (σ, ξ) 6= (0, 0)}. (6.15)

The elliptic parametrix construction, Theorem 4.26, can be microlocalized:

70 PETER HINTZ

Proposition 6.15. Let A ∈ Ψm (M ), and suppose K ⊂ Ell(A) is a closed subset. Then

there exists a microlocal parametrix for A on K, namely, an operator B ∈ Ψ−m (M ) such
that
K ∩ WF0 (AB − I) = ∅, K ∩ WF0 (BA − I) = ∅. (6.16)

Proof. By Corollary 6.11, we can pick Q, Q̃ ∈ Ψ0 (M ) with

WF0 (I − Q) ∩ K = ∅, WF0 (Q), WF0 (Q̃) ⊂ Ell(A), WF0 (I − Q̃) ∩ WF0 (Q) = ∅. (6.17)
Let then B0 ∈ Ψ−m (M ), σ−m (B0 ) = σ0 (Q̃)/σm (A), and write
AB0 = I − R = I − QR − (I − Q)R, R ∈ Ψ0 (M ). (6.18)
0
Now WF ((I − Q)R) ∩ K = ∅, while σ0 (QR) = σ0 (Q)(1 − σ0 (Q̃)) = 0, so QR ∈ Ψ−1 (M ).
We then improve the situation near K using a Neumann series argument; that is, let
∞
X
B0 ∼ (QR)j ∈ Ψ0 (M ), (6.19)
j=0

and put B := B0 B 0 ∈ Ψ−m (M ). Then

AB = I − R0 − (I − Q)RB 0 , R0 ∈ Ψ−∞ (M ), WF0 ((I − Q)RB 0 ) ∩ K = ∅, (6.20)
as desired.
A microlocal left parametrix, say B̃, can be constructed similarly. Then, modulo opera-
tors with WF0 disjoint from K, we have
B ≡ (B̃A)B = B̃(AB) ≡ B̃. (6.21)
0
Since (6.16) is invariant under addition to B of an operator with WF disjoint from K, this
proves that any microlocal left parametrix is also a right parametrix, and vice versa. 

One would like to use this to sharpen elliptic regularity theory, Proposition 4.27, by
saying that if Au = f , then on Ell(A), u is smooth when f is. This leads to the notion of
wave front set, which we discuss next.

6.3. Wave front set of distributions. Let u ∈ S 0 (Rn ) denote a distribution, and
A ∈ Ψm (Rn ). Then Au ∈ S 0 (Rn ) is ‘trivial’ outside of WF0 (A): all information about
singularities of u is lost. Indeed, if B ∈ Ψ0 (Rn ) is such that WF0 (B) ∩ WF0 (A), we have
B(Au) ∈ C ∞ (Rn ) by part (1) of Proposition 6.4. The precise notion of ‘triviality’ here is,
directly stated on manifolds:
Definition 6.16. Let u ∈ D 0 (M ). Then α ∈ S ∗ M does not lie in the wave front set,
/ WF(u) ⊂ S ∗ M
α∈ (6.22)
if and only if there exists a neighborhood U ⊂ S ∗ M of α such that
A ∈ Ψ0 (M ), WF0 (A) ⊂ U =⇒ Au ∈ C ∞ (M ). (6.23)

By definition, WF(u) ⊂ S ∗ M is closed. We leave to the reader the simple verification

that for M = Rn , WF(u) for u ∈ S 0 (Rn ) can be defined equivalently by testing with
uniform ps.d.o.s A ∈ Ψ0 (Rn ).
We will give equivalent conditions which are easier to verify. We begin by reducing the
number of operators for which one needs to check (6.23) to one.
 MICROLOCAL ANALYSIS 71

Lemma 6.17. Let u ∈ D 0 (M ). Then α ∈/ WF(u) if and only if there exists A ∈ Ψ0 (M ),

∞
elliptic at α, such that Au ∈ C (M ).

Proof. The direction ‘⇒’ is obvious. To prove ‘⇐’, we take U := Ell(A), which by as-
sumption is a neighborhood of α. Let B ∈ Ψ0 (M ), WF0 (B) =: K ⊂ U ; we claim that
Bu ∈ C ∞ (M ). By Proposition 6.15, there exists a microlocal parametrix Q ∈ Ψ0 (M ) of A
with QA = I − R, R ∈ Ψ0 (M ), WF0 (R) ∩ K = ∅. Therefore,
Bu = B(QA + R)u = (BQ)(Au) + BRu ∈ C ∞ (M ). (6.24)
Indeed, Au ∈ C ∞ (M ), hence the first summand is smooth; and WF0 (BR) ⊂ WF0 (B) ∩
WF0 (R) ⊂ K ∩ WF0 (R) = ∅, hence BR ∈ Ψ−∞ (M ) and so BRu ∈ C ∞ (M ) as well. 
Corollary 6.18. Let u ∈ D 0 (M ). Then
\
WF(u) = Char(A). (6.25)
A∈Ψ0 (M )
Au∈C ∞ (M )

/ WF(u), then Au ∈ C ∞ (M ) for some A ∈ Ψ0 (M ) with α ∈ Ell(A), so

Proof. If α ∈
/ Char(A). Conversely, if α ∈ Ell(A) for some A ∈ Ψ0 (M ) with Au ∈ C ∞ (M ), then
α ∈
α∈/ WF(u) by Lemma 6.17. 

This leads to the following very concrete description of the wave front set, which we state
directly on Rn ; it is the same on manifolds upon localizing in a chart and transferring to
Rn .
Proposition 6.19. Let u ∈ D 0 (Rn ), and let (x0 , ξ0 ) ∈ Rn × (Rn \ {0}). Then (x0 , ξ0 ) ∈
/
WF(u) if and only if there exist φ ∈ Cc∞ (Rn ), φ(x0 ) 6= 0, and  > 0 such that
 
−N n
ξ ξ 0 
|φu(ξ)|
c ≤ CN |ξ| , ξ ∈ R , |ξ| ≥ 1,  −  < . (6.26)
|ξ| |ξ0 | 
Remark 6.20. The converse direction can be strengthened slightly: to show that (x0 , ξ0 ) ∈
WF(u) does lie in the wave front set, it suffices to show that for any  > 0 there exists
φ ∈ Cc∞ (Rn ), φ(x0 ) 6= 0, supp φ ⊂ B(x0 , ), such that the estimate (6.26) fails. (That is,
as witnesses one can take any convenient cutoffs φ with support arbitrarily close to x0 .)
Indeed, this follows from the definition of WF and (the proof of) Lemma 6.17.

An advantage of our invariant approach to WF is that it implies ‘for free’ that the hands-
on condition (6.26) gives a well-defined notion of wave front set as a subset of the cotangent
bundle. We encourage the reader to try and give a direct proof of this fact, based on the
characterization (6.26).

Proof of Proposition 6.19. Suppose (6.26) holds. Let ψ ∈ C ∞ (Sn−1 ) have support in an
-ball around ξ0 /|ξ0 |, with ψ( |ξξ00 | ) 6= 0. Let moreover χ ∈ C ∞ (Rnξ ), χ(ξ) = 0 for |ξ| ≤ 1 and
χ(ξ) = 1 for |ξ| ≥ 2. Then
 
ξ
a(ξ, y) := χ(ξ)ψ φ(y) ∈ S 0 (Rn ; Rn ) (6.27)
|ξ|
72 PETER HINTZ

and |F(OpR (a)u)(ξ)| ≤ CN hξi−N ; therefore OpR (a)u ∈ C ∞ (Rn ).11 Since a is elliptic at
(x0 , ξ0 ), this implies (x0 , ξ0 ) ∈
/ WF(u).
/ WF(u), pick B ∈ Ψ0 (Rn ), elliptic at (x0 , ξ0 ), such that Bu ∈
Conversely, if (x0 , ξ0 ) ∈
C ∞ (Rn ). We can then choose φ ∈ Cc∞ (Rn ), φ(x0 ) 6= 0, and ψ ∈ C ∞ (Sn−1 ), ψ( |ξξ00 | ) 6= 0, such
that for a(ξ, y) defined in (6.27), and A := OpR (a), we have
WF0 (A) ⊂ Ell(B). (6.28)
(Cf. Example 6.9.) By the proof of Lemma 6.17, we thus have Au ∈ C ∞ (Rn ).
We claim
that in fact
Au ∈ S (Rn ), (6.29)
which proves (6.26) upon taking the inverse Fourier transform of Au. To prove (6.29), let
φ̃ ∈ Cc∞ (Rn ) be identically 1 near supp φ. Then φ̃(Au) ∈ Cc∞ (Rn ), while (1−φ̃)Au = Op(a0 )u
where
a0 (x, y, ξ) = (1 − φ̃(x))a(ξ, y) ∈ S 0 (Rn × Rn ; Rn ). (6.30)
0 0
But a (x, y, ξ) = 0 near x = y! In fact, a is a scattering symbol of order (0, 0, 0) vanishing
near the diagonal, hence Op(a0 ) ∈ Ψ−∞,−∞ sc (Rn ) has Schwartz kernel in S (Rn × Rn ) by
Exercises 4.7–4.8, which implies Op(a )u ∈ S (Rn ).12
0 

Another important consequence of Lemma 6.17 and Corollary 6.18 is the following result
which shows that WF is a significant refinement of sing supp:
Theorem 6.21. Let u ∈ D 0 (M ), and denote by π : T ∗ M → M the projection. Then
π(WF(u)) = sing supp u. (6.31)

/ sing supp u, then there exists χ ∈ Cc∞ (M ) with χ(x0 ) 6= 0 such that χu ∈
Proof. If x0 ∈
Cc (M ). But χ is elliptic at (x0 , ξ0 ) for any 0 6= ξ0 ∈ Tx∗0 M ; hence Tx∗0 M ∩ WF(u) = ∅.
∞

To prove the converse, suppose x0 ∈ / π(WF(u)). Then for each ξ ∈ Sx∗0 M , there exists
Aξ ∈ Ψ0 (M ), elliptic at ξ, such that Au ∈ C ∞ (M ). Let Uξ := Ell(Aξ ) ∩ Sx∗0 M . Then Uξ is
an open cover of the compact set Sx∗0 M ; thus we can pick a finite subcover,
N
[
Sx∗0 M = Uξi , ξi ∈ Sx∗0 M, i = 1, . . . , N. (6.32)
i=1
But then the operator
N
X
A := A∗ξi Aξi ∈ Ψ0 (M ) (6.33)
i=1
is elliptic on Sx∗0 M , and satisfies Au ∈ C ∞ (M ). If χ ∈ Cc∞ (M ), χ(x0 ) 6= 0, is chosen to
∗ ∞
have support so close to x0 such that Ssupp χ M ⊂ Ell(A), then χu ∈ C (M ) by the proof
of Lemma 6.17. 
Corollary 6.22. Let u ∈ D 0 (M ). Then WF(u) = ∅ if and only if u ∈ C ∞ (M ).
11In fact, the Fourier transform of φu ∈ E 0 (Rn ) is analytic and polynomially bounded. Using Cauchy’s
integral formula, or Exercise 6.1 if one wants to stick to real methods, one then shows that the estimate (6.26)
holds for all derivatives ∂ξα φu(ξ), α ∈ Nn0 , as well. Thus OpR (a)u ∈ S (R ).
c n

12A direct proof proceeds by writing Op(a0 ) = Op(|x − y|−2N ∆N a0 ) and noting that |x − y|−2N .
ξ
hxi−N hyi−N on supp a0 as well as ∆N 0
ξ a ∈ S
−2N
(Rn ×Rn ; Rn ). Thus, mimicking the proof of Proposition 4.10
gives Op(a0 ) ∈ S (Rn × Rn ).
 MICROLOCAL ANALYSIS 73

It is now time to give some examples:

Example 6.23. Let δ ∈ D 0 (Rn ). We claim that WF(δ) = {(0, ξ) : ξ 6= 0} = N ∗ {0} \ o. There
are several ways to see this. For instance:
(1) Using Proposition 6.19: since supp δ = {0}, WF(δ) can at most be equal to {(0, ξ)}.
But given a cutoff φ ∈ Cc∞ (Rn ) with φ(0) 6= 0, we have φδ(ξ)
c = φ(0), which is not
rapidly decreasing in the conic neighborhood of any ξ0 6= 0; hence the claim.
(2) Ad hoc argument: since sing supp δ = {0}, WF(δ) can at most be equal to {(0, ξ)}
by Corollary 6.22. But δ ∈ / C ∞ (Rn ), hence WF(δ) 6= ∅. But δ is rotationally
symmetric, hence (0, ξ) ∈ WF(δ) implies (0, Rξ) ∈ WF(δ) for all R ∈ SO(n − 1),
and we are done.
Example 6.24. Let Ω ⊂ Rn be a smoothly bounded domain. Then WF(1Ω ) = N ∗ ∂Ω \ o.
(See Exercise 6.2.)
Example 6.25. Consider (x + i0)−1 = lim→0 (x + i)−1 ∈ D 0 (R). Then WF((x + i0)−1 ) =
{(0, ξ) : ξ > 0}. This can be proved very explicitly using F((x+i0)−1 ) = (2πi)−1 H, where H
is the Heaviside function. (This equality is proved easily by calculating the inverse Fourier
transform of H as the S 0 (R)-limit of that of H(x)e−x as  & 0.) See also Exercise 6.3.
Example 6.26. If A ∈ Ψm (Rn ) is a ps.d.o. with Schwartz kernel K, then
WF(K) = {(x, x, ξ, −ξ) : (x, ξ) ∈ WF0 (A)}. (6.34)
See Exercise 6.6.

We next study the relationship of wave front sets and PDE. We first prove that pseudo-
differential operators are microlocal :
Proposition 6.27. Let A ∈ Ψm (M ) and u ∈ D 0 (M ). Then
WF(Au) ⊂ WF0 (A) ∩ WF(u). (6.35)

In view of Theorem 6.21, this is a significant strengthening of the pseudolocality property

of ps.d.o.s, see Proposition 4.17 (which holds on manifolds as well).

/ WF0 (A), then there exists B ∈ Ψ0 (M ), elliptic at

Proof of Proposition 6.27. Suppose α ∈
0 0
α, but with WF (B) ∩ WF (A) = ∅. Thus B(Au) = (BA)u ∈ C ∞ (M ) since BA ∈ Ψ−∞ (M ).
/ WF(u), then there exists B ∈ Ψ0 (M ), elliptic at α, such that Bu ∈ C ∞ (M ). Let
If α ∈
B̃ ∈ Ψ0 (M ) be elliptic at α and with WF0 (B̃) ⊂ Ell(B). Let Q ∈ Ψ0 (M ) be a microlocal
parametrix of B on WF0 (B̃), that is, QB = I − R, R ∈ Ψ0 (M ), WF0 (B̃) ∩ WF0 (R) = ∅.
Then
B̃(Au) = B̃A(QB + R)u = B̃AQ(Bu) + (B̃AR)u. (6.36)
The first summand is smooth since Bu is; the second summand is smooth since B̃AR ∈
Ψ−∞ (M ). 

Moreover, we have the following microlocal elliptic regularity result, which substantially
sharpens Proposition 4.27:
Proposition 6.28. Let u ∈ D 0 (M ) and A ∈ Ψm (M ). Then
WF(u) ⊂ WF(Au) ∪ Char(A). (6.37)
In particular, if A is elliptic, then WF(u) = WF(Au).
74 PETER HINTZ

/ WF(Au) and α ∈ Ell(A). Then there exists B ∈ Ψ0 (M ), elliptic at α,

Proof. Suppose α ∈
such that B(Au) ∈ C ∞ (M ); but α ∈ Ell(BA) by (6.14), hence α ∈
/ WF(u) by Corollary 6.18.
The claim about elliptic A follows from (6.35) and (6.37) since Char(A) = ∅. 

The wave front set studied above is more specifically the smooth wave front set or C ∞
wave front set, as it measures the lack of smoothness of a distribution. In applications, a
more refined notion is much more useful:
Definition 6.29. Let s ∈ R, u ∈ D 0 (M ). Then the H s wave front set of u is
\
WFs (u) := Char(A). (6.38)
A∈Ψ0 (M )
Au∈Hcs (M )

That is, its complement is the set of all α ∈ S ∗ M for which there exists A ∈ Ψ0 (M ), elliptic
at α, such that Au ∈ Hcs (M ).

This is equivalent to the alternative definition paralleling Definition 6.16. We collect

results analogous to those for the C ∞ wave front set; the proofs are left to the reader.
(They are the same as those for the C ∞ wave front set, except one now one needs to keep
track of Sobolev orders.) We have
WFs (u) = ∅ ⇐⇒ u ∈ Hloc
s
(M ). (6.39)
The analogue of Proposition 6.19 is the following:
Proposition 6.30. Let u ∈ D 0 (Rn ), (x0 , ξ0 ) ∈ Rn × (Rn \ {0}). Then (x0 , ξ0 ) ∈/ WFs (u) if
and only if there exists φ ∈ Cc∞ (Rn ), φ(x0 ) 6= 0, and ψ ∈ C ∞ (Sn−1 ), ψ(ξ0 /|ξ0 |) 6= 0, such
that ξ 
hξis ψ φu(ξ)
c ∈ L2 (Rnξ ). (6.40)
|ξ|

The sharpening of Propositions 6.27 and (6.28) is:

Proposition 6.31. Let A ∈ Ψm (M ), u ∈ D 0 (M ). Then
WFs−m (Au) ⊂ WF0 (A) ∩ WFs (u), (6.41)
s s−m
WF (u) ⊂ WF (Au) ∪ Char(A). (6.42)
In particular, if A is elliptic, then WFs (u) = WFs−m (Au).

Proof. See Exercise 6.9. 

The more precise way of stating the qualitative statement (6.42) of microlocal elliptic
regularity is the following quantitative estimate,13 stated on a compact manifold for conve-
nience: if B, G ∈ Ψ0 (M ) are such that
WF0 (B) ⊂ Ell(G), WF0 (B) ⊂ Ell(A), (6.43)
then for any N ∈ R, there exists C > 0 such that
 
kBukH s (M ) ≤ C kGAukH s−m (M ) + kukH −N (M ) ; (6.44)

13One can in fact recover the estimate from (6.42) using the closed graph theorem, though this loses the
(in principle) explicit nature of the constant C as depending on seminorms of A.
 MICROLOCAL ANALYSIS 75

and this estimate holds in the strong sense that if u ∈ D 0 (M ) is such that the right hand side
is finite, then so is the left hand side, and the estimate holds.14 (Thus, this is better than an
a priori estimate, at microlocal H s -membership of u is concluded, with estimates—rather
than merely assumed and estimated.)
We end with recording the relationship between H s and C ∞ wave front set:
Proposition 6.32. Let u ∈ D 0 (M ). Then
[
WF(u) = WFs (u). (6.45)
s∈R
S s
It is easy to see that s∈R WF (u) is, in general, a proper subset of WF(u).

Proof of Proposition 6.32. Clearly WFs (u) ⊂ WF(u), implying ‘⊇’. For the converse, sup-
pose α ∈ S ∗ M has an open neighborhood U ⊂ S ∗ M such that U ∩ WFs (u) = ∅ for all
s ∈ R. Then if A ∈ Ψ0 (M ),TWF0 (A) ⊂ U , is elliptic at α and has compactly supported
Schwartz kernel, then Au ∈ s∈R Hcs (M ) = Cc∞ (M ), hence α ∈
/ WF(u). 

6.4. Pairings, products, restrictions. The wave front set allows one to give fairly precise
answers to questions such as: when is the product of two distributions well-defined? When
can distributions be restricted to submanifolds? For notational simplicity, we work on Rn ,
but all results have analogues on manifolds.
We first consider generalizations of the L2 (Rn ) inner product
Z
hu, vi = u(x)v(x) dx. (6.46)
Rn
Proposition 6.33. Suppose u, v ∈ E 0 (Rn ) satisfy WF(u) ∩ WF(v) = ∅. If A ∈ Ψ0 (Rn ) is
such that
WF(u) ∩ WF0 (A) = ∅, WF(v) ∩ WF0 (I − A) = ∅, (6.47)
then the sesquilinear form
hu, vi := hAu, vi + hu, (I − A∗ )vi (6.48)
is independent of the choice of A.

Proof. Note that (6.48) is well-defined since Au, (I − A∗ )v ∈ C ∞ (Rn ) by microlocality,

Proposition 6.27.
Suppose B ∈ Ψ0 (Rn ) satisfies the conditions on A in (6.47). Then
hu, vi0 := hBu, vi + hu, (I − B ∗ )vi (6.49)
is well-defined, too, and we want to show that the difference
hu, vi0 − hu, vi = h(A − B)u, vi − hu, (A∗ − B ∗ )vi (6.50)
vanishes. If u, v were in Cc∞ (Rn ),
this would be clear by integration by parts. Since u, v are
merely distributions, we need to be more careful and use an approximation argument.
Thus, choose vj ∈ Cc∞ (Rn ) such that vj → v in E 0 (Rn ); then
h(A − B)u, vi = lim h(A − B)u, vj i = lim hu, (A∗ − B ∗ )vj i. (6.51)
j→∞ j→∞

14On a non-compact manifold, this holds if one takes B, G with Schwartz kernels supported in K × K,
K b M , and upon replacing the final, error term by kχukH −N where χ ∈ Cc∞ (M ) is identically 1 near K.
76 PETER HINTZ

We have (A∗ − B ∗ )vj → (A∗ − B ∗ )v in D 0 (Rn ); but since u ∈ E 0 (Rn ), this is not enough to
naively take the limit in (6.51). Pick thus Q ∈ Ψ0 (Rn ) with compactly supported Schwartz
kernel, and with WF0 (Q) ∩ WF(u) = ∅ and WF0 (I − Q) ∩ (WF0 (A) ∪ WF0 (B)) = ∅, then
we can further write
hu, (A∗ − B ∗ )vj i = hu, (I − Q)(A∗ − B ∗ )vj i + hu, Q(A∗ − B ∗ )vj i. (6.52)
Since (I − Q)(A∗ − B ∗ ) ∈ Ψ−∞ (Rn ), we have (I − Q)(A∗ − B ∗ )vj → (I − Q)(A∗ − B ∗ )v
with convergence in C ∞ (Rn ), hence the first pairing converges to hu, (I − Q)(A∗ − B ∗ )vi.
In the second pairing, we can integrate Q by parts, and then
hQ∗ u, (A∗ − B ∗ )vj i → hQ∗ u, (A∗ − B ∗ )vi, j→∞ (6.53)
since Q∗ u ∈ Cc∞ (Rn ). Since (A∗ − B ∗ )v ∈ C ∞ (Rn ), we can move Q∗ back to the second
factor.
Altogether, we have proved that the limit in (6.51) is indeed equal to hu, (A∗ − B ∗ )vi,
hence (6.50) vanishes, as desired. 

We state a more precise form of Proposition 6.33 which will be useful in positive commu-
tator arguments in §§8–9. First, note that the L2 -pairing (6.46) extends to a sesquilinear
pairing
H s (Rn ) × H −s (Rn ) 3 (u, v) 7→ hu, vi, (6.54)
defined by hu, vi := hhDis u, hDi−s vi. The following is proved similarly to Proposition 6.33:
Lemma 6.34. Let s ∈ R. Suppose u ∈ H s (Rn ), v ∈ E 0 (Rn ), and suppose that WF(u) ∩
WF−s (v) = ∅. Let A ∈ Ψ0 (Rn ) be such that
WF(u) ∩ WF0 (A) = ∅, WF−s (v) ∩ WF0 (I − A) = ∅. (6.55)
Then the sesquilinear form (u, v) 7→ hAu, vi + hu, (I − A∗ )vi is independent of A.
Remark 6.35. The manifold version of (6.54) is the following: fixing a smooth density µ on
M , we have a pairing
Z
2 2
Lc (M ) × Lloc (M ) 3 (u, v) 7→ hu, vi = u(x)v(x) dµ(x). (6.56)
M

For s ∈ R then, fix an elliptic operator Λ ∈ Ψs (M ) with parametrix Λ− ∈ Ψ−s (M ), so

I = Λ− Λ + R with R ∈ Ψ−∞ (M ). Then the pairing
(u, v) 7→ hΛu, Λ∗− vi + hRu, vi (6.57)
agrees with (6.56) for u, v ∈ Cc∞ (M ), and extends by continuity to a sesquilinear pairing
−s
Hcs (M ) × Hloc (M ) → C.

For a distribution u ∈ D 0 (Rn ), we have

WF(u) = −WF(ū) := {(x, −ξ) : (x, ξ) ∈ WF(u)}. (6.58)
We thus deduce from Proposition 6.33 that we can define a pairing
(u, v) = hu, v̄i, u, v ∈ E 0 (Rn ), WF(u) ∩ (−WF(v)) = ∅. (6.59)
 MICROLOCAL ANALYSIS 77

Corollary 6.36. Let u, v ∈ E 0 (Rn ), and suppose that

WF(u) ∩ (−WF(v)) = ∅. (6.60)
Then the product uv ∈ E 0 (Rn ) given, in terms of (6.59), by
Cc∞ (Rn ) 3 φ 7→ (u, φv), (6.61)
is well-defined.

The condition (6.60) is of course much more precise than the condition sing supp u ∩
sing supp v = ∅, under which the product uv can be defined easily using a partition of unity
on Rn .

Proof of Corollary 6.36. If A ∈ Ψ0 (Rn ) has WF(u) ∩ WF0 (A) = ∅ and (−WF(v)) ∩ WF0 (I −
A) = ∅, then
|(u, φv)| ≤ |(Au, φv)| + |(u, (I − AT )(φv))|. (6.62)
Since Au ∈ C ∞ (Rn ), the first summand is clearly continuous in φ. For the second summand,
choose B ∈ Ψ0 (Rn ) such that WF0 (B) ∩ WF0 (I − AT ) = ∅ and WF0 (I − B) ∩ WF(v) = ∅,
then
v = Bv + w, w = (I − B)v ∈ C ∞ (Rn ), (6.63)
T T
hence |(u, (I − A )wφ)| ≤ CkφkC k (Rn ) by the continuity of u and the ps.d.o. (I − A )w ∈
Ψ0 (Rn ). Furthermore,
(I − AT )φBv ∈ C ∞ (Rn ) (6.64)
depends continuously on φ ∈ Cc∞ (Rn ) since (I − AT )φB ∈ Ψ−∞ (Rn ) does. 
Remark 6.37. One can put a complete locally convex topology on the space DΛ0 (Rn ) := {u ∈
D 0 (Rn ) : WF(u) ⊂ Λ}, where Λ ⊂ S ∗ Rn is closed, such that Cc∞ (Rn ) ⊂ DΛ0 (Rn ) is dense, and
such that the pairing (u, v) for u, v ∈ Cc∞ (Rn ) extends by continuity to u ∈ (DΛ0 ∩ E 0 )(Rn ),
v ∈ DΛ0 0 (Rn ) when Λ ∩ (−Λ0 ) = ∅.
Remark 6.38. One can substantially refine Corollary 6.36, e.g. by working with Sobolev
spaces and assumptions on H s wave front sets for various s. Such refinements are useful in
the study of nonlinear PDE.

Lastly, we consider restrictions to submanifolds, starting with the local model

Y = {(x, 0) : x ∈ Rk } ⊂ Rn = Rkx × Ryn−k . (6.65)
Denote by ι : Y ,→ Rnthe inclusion map. For u ∈ Cc∞ (Rn ), its restriction to Y is the
∗
distribution ι u defined by
Cc∞ (Rk ) 3 φ 7→ (ι∗ u)(φ) := (u · δ(y))(φ̃), φ̃ ∈ Cc∞ (Rn ), ι∗ φ̃ = φ. (6.66)
If u ∈ E 0 (Rn ) is such that
WF(u) ∩ {(x, 0, 0, η) : x ∈ Rk , η ∈ Rn−k } = ∅, (6.67)
then the product uδ(y) ∈ E 0 (Rn ) is well-defined by Corollary 6.36, hence we get the first
part of the following result:
Proposition 6.39. Suppose u ∈ E 0 (Rn ) satisfies (6.67). Then (6.66) defines a linear
restriction map, and
WF(ι∗ u) ⊂ {(x, ξ) ∈ Rk × (Rk \ {0}) : ∃ η ∈ Rn−k , (x, 0, ξ, η) ∈ WF(u)}. (6.68)
78 PETER HINTZ

Proof. We have ι∗ (χu) = ι∗ u for any χ ∈ Cc∞ (Rn−k ) which is identically 1 near 0. The
assumption (6.67) implies that if χ has sufficiently small support, then v = χu satisfies that
v̂(ξ, η) is rapidly decreasing in a cone around {0} × (Rn−k \ {0}). (6.69)
When u = u(x, y) ∈ Cc∞ (Rn ), the Fourier inversion formula gives
Z
∗ −(n−k)
ι u(ξ) = (F1 u)(ξ, 0) = (2π)
c û(ξ, η) dη, (6.70)
Rn−k
where F1 denotes the Fourier transform in the first argument of u. More generally then,
∗ u even
the property (6.69) ensures that the integral in (6.70) converges, and it computes ιc
for distributional u subject to (6.67) by a density argument.
To prove (6.68), we apply (6.70) to a localized version of u. Indeed, suppose (x0 , ξ0 ) ∈
Rk × (Rk \ {0}) is such that for all η ∈ Rn−k , we have (x0 , 0, ξ0 , η) ∈ / WF(u). Then
for ψ ∈ Cc∞ (Rk ) with support close to x and χ ∈ Cc∞ (Rn−k ) with support close to 0, the
Fourier transform of ψ(x)χ(y)u(x, y) is rapidly decreasing for (ξ, η) in a conic neighborhood
of (ξ0 , η) for all η ∈ Rn−k , as well as for (ξ, η) in a conic neighborhood of (0, η) by (6.67).
Therefore, there exists  > 0 such that
 
−N −N
ξ ξ0 
|ψχu(ξ, η)| ≤ CN hξi hηi ,  −
d  < , (6.71)
|ξ| |ξ0 | 
for all N . Using (6.70) for χu (which satisfies ι∗ (χu) = ι∗ u), we conclude that
 
0 −N
ξ ξ0 
∗
|ψι u(ξ)| ≤ CN hξi ,  −
[  < , (6.72)
|ξ| |ξ0 | 
/ WF(ι∗ u).
for all N , proving that (x0 , ξ0 ) ∈ 
Corollary 6.40. If ι : Y ⊂ M is a smooth submanifold, then there exists a linear restriction
map
ι∗ : {u ∈ D 0 (M ) : WF(u) ∩ N ∗ Y = ∅} → D 0 (Y ), (6.73)
∗ ∗ ∗ ∗ ∼
and WF(ι u) ⊂ T Y \ o is the image of WF(u) ∩ T M in T Y = T M/N Y . ∗ ∗
Y Y

Example 6.41. On R2x,y ,

consider ua = δ(y − ax), a ∈ R. Then the restriction of ua to
Y = {y = 0} is well-defined for a 6= 0. We have ι∗ (ua ) = |a|−1 δ(x), where ι : Y ,→ R2 is the
inclusion.

One can similarly analyze the wave front sets of general pullbacks and pushforwards of
distributions, and analyze the relationship between WF(Au) and WF(u) in terms of the
wave front set of the Schwartz kernel of A : Cc∞ (Rn ) → D 0 (Rm ). See e.g. [Hör71, §2.5].

We have now developed the main aspects of the pseudodifferential calculus. For a partial
summary of the calculus on compact manifolds, see [Wun13, §3.4].

6.5. Exercises.
Exercise 6.1. Let a ∈ S m (Rn ; RN ). Show that (x0 , ξ0 ) ∈ Rn × (RN \ {0}) does not lie in
ess supp a if and only if there exist  > 0 such that for all k ∈ R, we have
 
−k
ξ ξ0 
|a(x, ξ)| ≤ Ck hξi ∀ (x, ξ), |ξ| ≥ 1, |x − x0 | +  −  < . (6.74)
|ξ| |ξ0 | 
 MICROLOCAL ANALYSIS 79

Exercise 6.2. Let Ω ⊂ Rn be a smoothly bounded domain, and denote by

(
1, x ∈ Ω
1Ω (x) = (6.75)
0, x ∈
/Ω
its characteristic function. Compute (with proof) WF(1Ω ). (Hint. Straighten out ∂Ω
locally in suitable local coordinates. Then use the characterization of Proposition 6.19 and
Remark 6.20 for cleverly chosen cutoffs; alternatively, use Lemma 6.17 for some well-chosen
test operators A.)
/ N0 . Recall that the distribution (x ± i0)s ∈ D 0 (R) is defined
Exercise 6.3. Let s ∈ C, s ∈
as the limit
h(x ± i0)s , φi := lim h(x ± i)s , φi, (6.76)
→0
where (x ± i)s = exp(s log(x ± i)). (The logarithm here is the principal branch, i.e. it is
real-valued for real arguments, and its branch cut is along (−∞, 0].) Prove that
WF((x ± i0)s ) = {(0, ξ) : ±ξ > 0}. (6.77)
(Hint. Show that it suffices to prove this for s with (large) negative real part. When
Re s < 0, shift the contour of integration in the Fourier transform R (x + i)s e−ixξ dx to a
R
line Im x = C and analyze what happens when you let  → 0 and then C → ∞.)
Exercise 6.4. Let n ∈ N, n ≥ 2. Give an example of a distribution u ∈ D 0 (Rn ) whose wave
front set WF(u) ⊂ S ∗ Rn consists of a single point.
Exercise 6.5. (Conormal distributions.) Let a ∈ S m (Rkx ; Rζn−k ).
(1) Make sense of the oscillatory integral
Z
−(n−k)
u(x, z) := (2π) eiz·ζ a(x, ζ) dζ (6.78)
Rn−k

as a distribution on Rn = Rkx × Rn−kz in such a way that for m < −(n − k), your
definition agrees with the Riemann integral.
(2) Show that WF(u) ⊂ {(x, z, ξ, ζ) : z = 0, ξ = 0}.
(3) Prove that WF(u) = {(x, 0, 0, ζ) : (x, ζ) ∈ ess supp a}.
Exercise 6.6. Let A ∈ Ψm (M ) be a pseudodifferential operator, and denote its Schwartz
kernel by K. Prove that
WF(K) = {(x, x, ξ, −ξ) ∈ T ∗ (M × M ) : (x, ξ) ∈ WF0 (A)} (6.79)
Exercise 6.7. The following is a generalization of Exercise 6.3. Denote by Ω = {z ∈
C : Im z > 0} the upper half plane, and let F : Ω → C be holomorphic. Suppose that for
each C > 0 there exist C 0 , N ∈ R so that |F (z)| < C 0 | Im z|−N for z ∈ Ω, | Re z| < C,
Im z ∈ (0, 1].
(1) Show that the functions F = F (· + i) ∈ C ∞ (R) converge in D 0 (R) as  & 0. The
limit is denoted f := F (· + i0) ∈ D 0 (R). (Hint. Write F in terms of F1 using
the fundamental theorem of calculus for F in the imaginary direction. Using the
Cauchy–Riemann equations, show in this manner that for φ ∈ Cc∞ (R) with support
(1)
in (−C, C), one can write hF , φi = hF , φ0 i where F (1) is holomorphic in Ω and
satisfies |F (1) (z)| < C 0 | Im z|−N +1 for | Re z| < C, Im z ∈ (0, 1] when N > 1, or
80 PETER HINTZ

with F (1) continuous down to the real line when N < 1. Starting with general N ,
proceed iteratively.)
(2) Show that WF(F (· + i0)) ⊂ {(x, ξ) : ξ > 0}.
Exercise 6.8. Let u, v ∈ D 0 (Rn ), and suppose that WF(u)∩(−WF(v)) = ∅. In the following,
we shall define a distribution uv ∈ D 0 (Rn ) in a hands-on manner.
(1) Let x0 ∈ Rn . Prove that there exists an open neighborhood U ⊂ Rn of x0 with the
following property: the sets W (u) = {ξ : ∃x ∈ U, (x, ξ) ∈ WF(u)} (the projection of
WF(u) over U to the momentum variables) and W (v) satisfy W (u) ∩ (−W (v)) = ∅.
(2) For x0 ∈ Rn and U ⊂ Rn as in the first part, fix any χ ∈ Cc∞ (U ). Show that the
integral Z
ŵ(ξ) := (2π)n χ χv(ξ − η) dη
cu(η)c (6.80)
Rn
converges absolutely and defines a smooth polynomially bounded function of ξ ∈
Rn . Define then wχ2 := F −1 ŵ ∈ S 0 (Rn ). Show that when u, v ∈ C ∞ (M ), then
wχ2 = χ2 uv.
(3) Show that there exists a distribution uv ∈ D 0 (Rn ) so that χ2 uv = wχ2 for any
χ ∈ Cc∞ (U ) where U is an open neighborhood of an arbitrary point x0 ∈ Rn as in
the first part. (Hint. Given a closed conic subset W ⊂ T ∗ Rn \ o, define the space
DW0 (Rn ) = {u ∈ D 0 (Rn ) : WF(u) ⊂ W }, equipped with the locally convex topology

given by seminorms |h−, φi| for φ ∈ Cc∞ (Rn ) and kφA(−)kC k where φ ∈ Cc∞ (Rn ),
A ∈ Ψ0 (Rn ) with WF0 (A) ∩ W = ∅, and k ∈ N0 . Show that Cc∞ (Rn ) ⊂ DW 0 (Rn ) is

dense. Fixing W1 , W2 ∈ T R \ o with W1 ∩ (−W2 ) = ∅, define uv ∈ D (Rn ) for

∗ n 0

u ∈ DW 0 (Rn ) and v ∈ D 0 (Rn ) using a continuity and density argument.)

1 W2

Exercise 6.9. (1) Prove Proposition 6.31.

(2) Prove the estimate (6.44).

7. Hyperbolic evolution equations

As a neat application of the ps.d.o. machinery, we now study first order systems of
evolution equations. We work on Rn , but all results have analogues on compact manifolds.

7.1. Existence and uniqueness. Consider

(
Dt u = a(t, x, Dx )u + g(t, x), t ∈ R, x ∈ Rn ,
(7.1)
u(0, x) = f (x), x ∈ Rn ,
where
f ∈ H s (Rn ; CK ), g ∈ C 0 R; H s (Rn ; CK ) .


(7.2)
We let A(t) = Op(a(t, x, ξ)), and assume that a(t, x, ξ) ∈ C ∞ (Rt ; MatK×K (S 1 (Rn ; Rn )))
is a K × K matrix of first order symbols with smooth dependence on t; we simply write
a(t) ∈ S 1 . We further assume that (7.1) is symmetric hyperbolic, meaning
a(t, x, ξ) − a(t, x, ξ)∗ ∈ S 0 . (7.3)
Theorem 7.1. The equation (7.1) with data (7.2) has a unique solution
u ∈ C 0 R; H s (Rn ; CK ) ∩ C 1 R; H s−1 (Rn ; CK ) .



(7.4)
 MICROLOCAL ANALYSIS 81

Proof. We drop the ‘bundle’ CK from the notation. We shall obtain u as a limit of solution
u to a regularized equation (
Dt u = J AJ u + g,
(7.5)
u (0) = f,
where we use a Friedrichs mollifier
J = φ(Dx ), φ ∈ Cc∞ (Rn ), φ(0) = 1. (7.6)
Note that J ∈ Ψ−∞ (Rn ) for  > 0, and J ∈ Ψ0 (Rn ) is uniformly bounded for  ∈ (0, 1].
For  > 0, J AJ is a smooth family of bounded operators on H s (Rn ), hence solvability
of (7.5) with u ∈ C 1 (R; H s (Rn )) follows from ODE theory. We need to establish uniform
estimates on u . Let Λs = hDx is . Then
1d
ku (t)k2H s = RehΛs J iAJ u , Λs u i + RehΛs g, Λs u i (7.7)
2 dt
s s s s s s
= RehiAΛ J u , Λ J u i + Reh[Λ , iA]J u , Λ J u i + RehΛ g, Λ u i. (7.8)
Since B(t) = A(t) − A(t)∗ ∈ Ψ0 , the first term is equal to
hB(t)Λs J u , Λs J u i ≤ CkJ u k2H s ≤ Cku k2H s . (7.9)
Since [Λs , A] ∈ Ψs , the second term in (7.8) is bounded by Cku k2H s as well. Applying
Cauchy–Schwarz to the third term in (7.8), we obtain
d
ku (t)k2H s ≤ Cku (t)k2H s + Ckg(t)k2H s . (7.10)
dt
By Gronwall’s inequality, this implies the -independent estimate
 
ku (t)k2H s ≤ C(t) kf k2H s + kgk2C 0 ([0,t];H s (Rn )) . (7.11)
Therefore, for any T > 0 and I = [−T, T ],
u ∈ C 0 (I; H s (Rn )) ∩ C 1 (I; H s−1 (Rn )) (7.12)
is uniformly bounded. (Boundedness in the second space follows boundedness in the first
space and the equation (7.5).)
We can extract a subsequential limit of u very easily by using the continuous injection
C 1 (I; H s−1 (Rn )) ,→ H 1 (I; H s−1 (Rn )); the latter space is a Hilbert space, so there exists a
weak subsequential limit
u ∈ H 1 (I; H s−1 (Rn )) ,→ C 0 (I; H s−1 (Rn )) (7.13)
of u . Thus, u is a weak solution of (7.1), and thus also u ∈ C 1 (I; H s−2 (Rn )). Since
δ(t) ∈ H −1/2− (R), we also have u(0) = hu, δi = limhu , δi = f . Uniqueness of u follows
using estimates similar to (7.11) for the difference of two putative solutions.
To prove the correct regularity of u, we approximate f ∈ H s (Rn ), g ∈ C 0 (R; H s (Rn )) in
these topologies by fj ∈ H s+1 (Rn ), gj ∈ C 0 (R; H s+1 (Rn )). Then we have just constructed
a solution uj ∈ C 0 (I; H s (Rn )) ∩ C 1 (I; H s−1 (Rn )) of (7.1). Moreover,
vjk := uj − uk (7.14)
solves (7.1) with initial data fj − fk and forcing gj − gk . An estimate similar to (7.11)
thus implies that vjk → 0 in C 0 (I; H s (Rn )) as j, k → ∞. Therefore, uj is Cauchy in
C 0 (I; H s (Rn )), hence its limit u satisfies (7.4), as desired. 
82 PETER HINTZ

As a simple example, we solve the wave equation on Rn ,


2 n
u := (Dt − ∆)u = g, t ∈ R, x ∈ R ,

u(0, x) = f0 (x), x ∈ Rn , (7.15)
x ∈ Rn ,

Dt u(0, x) = f1 (x),


where
g ∈ C 0 (R; H s−1 (Rn )), (f0 , f1 ) ∈ H s (Rn ) ⊕ H s−1 (Rn ). (7.16)
Corollary 7.2. The wave equation (7.15) with data (7.16) has a unique solution
∞
\
u∈ C j (R; H s−j (Rn )). (7.17)
j=0

Proof. Let Λ = hDi. We write

U := (Λu, Dt u),
F := (Λf0 , f1 ) ∈ H s−1 (Rn ; C2 ), (7.18)
0 s−1 n 2
G := (0, g) ∈ C (R; H (R ; C )),
then the equation (7.15) is equivalent to the first order system
(
Dt U = AU + G,
(7.19)
U (0) = F,

where the operator A is given by

 
0 Λ
A= . (7.20)
Λ−1 ∆ 0
∗ ∈ Ψ0 . Therefore, by Theorem 7.1, the equation (7.19) has a
T∞− A(t)
Note that A(t)
solution U ∈ j=0 C (R; H s−1−j (Rn ; C2 )). The function u := Λ−1 U0 is the desired solution
j

of (7.15). 

7.2. Egorov’s theorem; propagation of singularities. We now study the microlocal

behavior of solutions of scalar hyperbolic equations
Dt u = A(t, x, Dx )u, (7.21)
where we assume that A ∈ C ∞ (Rt ; Ψ1cl (Rn )). Denote by a ∈ C ∞ (Rt ; Shom
1 (Rn ; Rn \ {0})) the

homogeneous principal symbol of A; we assume that

a(t, x, ξ) is real-valued. (7.22)
We denote the solution operator for (7.21) by
S(t, s) : u(s) 7→ u(t). (7.23)
By Theorem 7.1, S(t, s) ∈ L(H σ (Rn )) for all σ ∈ R; moreover, S(t, s) is invertible with
S(t, s)−1 = S(s, t).
 MICROLOCAL ANALYSIS 83

Theorem 7.3. Let P0 = Op(p0 ) ∈ Ψm (Rn ) be a ‘test operator’, and define

P (t) := S(t, 0) ◦ P0 ◦ S(0, t), t ∈ R. (7.24)
Then P (t) ∈ Ψm (Rn ) modulo a smoothing operator: there exists R ∈ C ∞ (R; C ∞ (R2n )) such
that P (t) − R(t) ∈ Ψm (Rn ). The principal symbol of P (t) is given by


σm (P (t)) C(t)(x0 , ξ0 ) = p0 (x0 , ξ0 ), (7.25)
where C(t) is the time t flow (from 0 to t) of the time-dependent Hamiltonian vector field
Ha(t,x,ξ) ; that is, C(t)(x0 , ξ0 ) = γ(t) where γ(0) = (x0 , ξ0 ) and γ 0 (s) = Ha(s) |γ(s) .

Proof. Differentiating (7.24) in t gives the equation

P 0 (t) = i[A(t, x, D), P (t)], P (0) = P0 . (7.26)

Using the symbol calculus and an asymptotic summation, we will first construct an
approximate solution Q(t) = Op(q(t)), q(t, x, ξ) ∈ S m , of this, so
Q0 (t) = i[A(t, x, D), Q(t)] + R1 (t), Q(0) = P0 , R1 ∈ C ∞ (Rt ; Ψ−∞ (Rn )). (7.27)
We make the ansatz
∞
X
q(t) ∼ qk (t), qk (t) ∈ S m−k . (7.28)
k=0
Taking the principal symbol of (7.27) then gives


∂t − Ha(t) q0 (t, x, ξ) = 0, q0 (0, x, ξ) = p0 (x, ξ). (7.29)
Thus q0 (t, C(t)(x, ξ)) = p0 (x, ξ). We leave it to the reader to check that q0 (t) ∈ S m .
Proceeding iteratively, we take qj (t) ∈ S m−j , j ≥ 1, to be the solution of a transport
equation


∂t − Ha(t) qj (t, x, ξ) = ej (t, x, ξ), (7.30)
where ej (t) ∈ S m−j is computed from the full symbol of A and q0 , . . . , qj−1 .
Having thus arranged (7.27), we now prove that for any N ∈ R, the difference R(t) =
P (t) − Q(t) maps any f ∈ H −N (Rn ) into H ∞ (Rn ). Equivalently, we will show
v(t) − w(t) ∈ H ∞ (Rn ), v(t) := S(t, 0)P0 f, w(t) := Q(t)S(t, 0)f. (7.31)
Note that v(t) and w(t) solve the equations
Dt v = A(t, x, Dx )v, v(0) = f,
(7.32)
Dt w = A(t, x, Dx )w − iR1 S(t, 0)f, w(0) = f.
Therefore, putting g(t) = iR1 S(t, 0)f ∈ C ∞ (Rt ; H ∞ (Rn )), we have
Dt (v − w) = g, (v − w)(0) = 0. (7.33)
By Theorem 7.1, (7.31) follows, finishing the proof. (This argument shows that the smooth-
ing error in fact lies in the space Ψ−∞ (Rn ) + H ∞ (R2n ), with smooth dependence on t.) 

As a simple consequence, we can track the wave front set of a solution of a scalar evolution
equation (7.21).
84 PETER HINTZ

Theorem 7.4. Suppose A is as in (7.21)–(7.22). Let u0 ∈ H −N (Rn ), and let u denote the
solution (
Dt u = A(t, x, Dx )u,
(7.34)
u(0) = u0 .
Then, with C(t) as in the statement of Theorem 7.3, we have
WF(u(t)) = C(t)WF(u0 ). (7.35)

Proof. It suffices to prove the inclusion ‘⊆’ (since switching the time direction then proves
‘⊇’). Thus, suppose α ∈ / WF(u0 ). Take an operator P0 ∈ Ψ0 (Rn ), elliptic at α, such that
∞ n
P0 u0 ∈ C (R ). Then, in the notation of Theorem 7.3,
P0 u0 = S(0, t)P (t)S(t, 0)u0 ∈ C ∞ (Rn ), (7.36)
so P (t)u(t) ∈ C ∞ (Rn ). But P (t) is elliptic at C(t)α, hence C(t)α ∈
/ WF(u(t)). 
Remark 7.5. Theorem 7.1, and thus also Theorem 7.4, can be generalized easily to the case
of initial data and forcing terms in weighted Sobolev spaces. In particular, by Theorem 2.14,
we can allow u0 in Theorem 7.4 to be any tempered distribution u0 ∈ S 0 (Rn ).
Example 7.6. For A = Dx ∈ Ψ1 (R), the solution operator is (S(t, 0)u)(x) = u(t + x), and
C(t)(x0 , ξ0 ) = (x0 + t, ξ0 ). And indeed P (t) = Op(p(t)), p(t, x, ξ) = p0 (x + t, ξ).
Example 7.7. Consider the half Klein–Gordon equation
Dt u = hDx iu on Rt × Rnx . (7.37)
(This arises from the factorization Dt2 − (∆ + 1) = (Dt − hDx i)(Dt + hDx i).) In this case,
a(x, ξ) = |ξ|, which has Hamiltonian vector field Ha = |ξ|−1 ξ · ∂x . Thus, the operator P
gets ‘transported’ along straight lines with direction determined by the momentum variable
ξ. Explicitly, (7.37) is solved by u(t) = eithDi u(0), and the wave front set statement can be
checked explicitly from this; however, the true power of Theorems 7.3 and 7.4 of course lies
in the fact that they apply to equations with non-constant coefficients as well.

7.3. Exercises.
Exercise 7.1. Consider the initial value problem for the Klein–Gordon equation,

2 t ∈ R, x ∈ Rn ,
(Dt − ∆ − 1)u = g,

u(0, x) = f0 (x), x ∈ Rn , (7.38)
x ∈ Rn .

Dt u(0, x) = f1 (x),


Here (f0 , f1 ) ∈ H s (Rn ) ⊕ H s−1 (Rn ) for some s, and g ∈ C ∞ (R; H ∞ (Rn )) is smooth.
(1) Show that (7.38) has a unique solution u ∈ ∞ j s−j (Rn )).
T
j=0 C (R; H
(2) Show that WF(u(t)) ⊂ T ∗ Rn \ o is contained in the set
[ n ξ0  o
x0 ± t , ξ0 : (x0 , ξ0 ) ∈ WF(u0 ) ∪ WF(u1 ) .
±
|ξ0 |

(Hint. Factor equation (7.38) as (Dt − hDi)(Dt + hDi)u = g.) Can you make a
more precise statement?
 MICROLOCAL ANALYSIS 85

Exercise 7.2. Suppose the K × K system of first order evolution equations

u(0) = f ∈ H s (Rn ), g ∈ C 0 R; H s (Rn ; CK ) ,


Dt u = L(t, x, Dx )u + g, (7.39)
is a symmetrizable hyperbolic system: there exists a K ×K-matrix-valued symbol S(t, x, ξ) ∈
S 0 which is positive definite and such that S(t, x, ξ)L(t, x, ξ) is symmetric modulo S 0 . Prove
that (7.39) has a unique solution
u ∈ C 0 R; H s (Rn ; CK ) ∩ C 1 R; H s−1 (Rn ; CK ) .



(7.40)
(Hint. Construct a positive definite operator S(t) ∈ Ψ0 with principal symbol S(t, x, ξ).
Mimic the proof of Theorem 7.1 and prove an -independent estimate for the quantity
d s s d 2
dt hΛ u (t), S(t)Λ u iL2 instead of dt ku (t)kH s .)

Exercise 7.3. A K ×K system of the form (7.39) is strictly hyperbolic if the principal symbol
L1 (t, x, ξ) of L is positively homogeneous of degree 1 in ξ, and if for all (t, x) ∈ R × Rn ,
ξ ∈ Rn \ {0}, L1 (t, x, ξ) has K distinct real eigenvalues. Show that a strictly hyperbolic
system is symmetrizable.
Exercise 7.4. On Rt × Rnx , consider an m-th order operator
m−1
X
L = Dym + aj (y, x, Dx )Dyj (7.41)
j=0

where aj (y, x, Dx ) ∈ Diff m−j (Rn ). We study the system

(
Lu = g ∈ C 0 (Rt ; H s−m+1 ),
(7.42)
(u, Dy u, . . . , Dym−1 u) = (u0 , u1 , . . . , um−1 ),
where uj ∈ H s−j (Rn ) for j = 0, . . . , m − 1.
(1) Reduce (7.42) to an m × m system of first order evolution equations for the Cm -
valued function (Λm−1 u, Λm−2 Dy u, . . . , Dym−1 u) where Λ = hDx i.
(2) Give a condition on the principal symbol of L which is equivalent to the strict
hyperbolicity (see Exercise 7.3) of this m × m system.
(3) Under the strict hyperbolicity assumption of part (2), prove existence and unique-
ness of a solution u ∈ ∞ j (R ; H s−j (Rn )) of (7.42).
T
j=0 C t

8. Real principal type propagation of singularities

We now free ourselves from the restrictive setting15 of equations which are explicitly given
in evolution form, and consider the propagation of singularities (wave front set)/regularity
for solutions of rather general non-elliptic (pseudo)differential equations
P u = f, P ∈ Ψm (M ), (8.1)
where M = Rn or some other manifold, and m ∈ R; we assume that the principal symbol
p(x, ξ) = σm (P )(x, ξ) (8.2)
is (positively) homogeneous of degree m and real.
15The theory of Fourier integral operators provides tools to ‘microlocally conjugate’ every real principal
type operator into the operator Dt on Rt × Rn−1 , see [Hör71, DH72, Hör09], thus this setting, with L = 0,
in fact captures the general situation. We shall however not develop this theory here.
86 PETER HINTZ

Definition 8.1. A null-bicharacteristic of P is an integral curve in Char(P ) ⊂ T ∗ M \ o of

the Hamiltonian vector field Hp .
Note that Hp p = 0; hence an integral curve of Hp with initial condition α ∈ Char(P ) is
automatically a null-bicharacteristic (and all null-bicharacteristics arise in this fashion).
Theorem 8.2. Suppose P ∈ Ψm (M ) has a real-valued homogeneous principal symbol, and
u ∈ D 0 (M ) is such that P u ∈ C ∞ (M ). Then WF(u) ⊂ Char(P ) is a union of maximally
extended null-bicharacteristics of P .
The statement WF(u) ⊂ Char(P ) is just microlocal elliptic regularity, Proposition 6.28.
The theorem asserts that within Char(P ), the wave front set of a ‘microlocal solution’ u is
invariant under the Hp -flow. Rather than calling Theorem 8.2 a result on the propagation
of singularities, one often (and more usefully) regards it as a result on the propagation of
regularity, since α ∈/ WF(u) ∩ Char(P ) implies that the entire maximally extended null-
bicharacteristic of P through α is disjoint from WF(u).
Remark 8.3. Suppose α ∈ T ∗ M \ o is such that Hp |α = 0. Then the null-bicharacteristic
through α is the constant curve α, i.e. Theorem 8.2 does not give any information at α.
One says P is of real principal type if dp 6= 0 on Char(P ); in this case, Hp never vanishes
on Char(P ).
Remark 8.4. Denote by V the generator of dilations in the fibers of T ∗ M , so V = ξ∂ξ in
local coordinates. Since WF(u) is conic, Theorem 8.2 is trivial also at radial points: these
are points α ∈ T ∗ M \ o where Hp |α = cV , c ∈ R. We shall discuss interesting classes of
radial points in §9.

Proof of Theorem 8.2. This can easily be reduced to a local result. Indeed, if χ, χ̃ ∈ Cc∞ (M )
are two cutoffs and χ̃ ≡ 1 near supp χ, then
χP χ̃u = χP u + χ[P, χ̃]u ∈ Cc∞ (M ). (8.3)
Thus, if φ ∈ Cc∞ (M ) is identically 1 near supp χ̃, we can replace P by χP χ̃ and u by φu, and
we still have P u ∈ C ∞ (M ); moreover, these replacements do not alter null-bicharacteristics
of P and the wave front set of u in χ−1 (1). Localizing in this fashion to a coordinate patch,
we can thus assume
P ∈ Ψm (Rn ), u ∈ E 0 (Rn ), P u = f ∈ Cc∞ (Rn ). (8.4)
We normalize this using Λ = hDim−1
by replacing (P, u) by (P Λ−1 , Λu);
thus we can assume
m = 1. After these replacements, the principal symbol of P is still homogeneous (of degree
1) and real.
We now add an artificial time variable t and set
ũ(t, x) = u(x), f˜(t, x) = f (x). (8.5)
Then ũ solves the equation
˜
Dt ũ = P ũ − f. (8.6)
A simple extension of the proof of Theorem 7.4 (taking into account the presence of f˜ ∈
C ∞ (R; Cc∞ (Rn ))) implies that
WF(u) = WF(ũ(t)) = C(t)WF(ũ(0)) = C(t)WF(u), (8.7)
where C(t) is the time t flow of Hp . Thus, WF(u) is invariant under the Hp -flow, proving
the theorem. 
 MICROLOCAL ANALYSIS 87

One of the main drawbacks of the above proof (apart from being rather ad hoc) is that
it ultimately rests on the solvability theory for the (auxiliary) equation (8.6). But solving
PDE is difficult, hence one should try to solve as few as possible! We shall thus present
another proof of Theorem 8.2 which is longer and looks more complicated, but is at its core
very simple, and the prototypical example of a positive commutator argument.
We will prove the following sharpening of Theorem 8.2:
Theorem 8.5. Suppose P ∈ Ψm (M ) has a real-valued homogeneous principal symbol. Let
u ∈ D 0 (M ) and f := P u ∈ D 0 (M ). Let s ∈ R. Then
WFs (u) ⊂ Char(P ) ∪ WFs−m (f ). (8.8)
Moreover, within Char(P ) \ WFs−m+1 (f ),
WFs (u) \ WFs−m+1 (f ) (8.9)
is a union of maximally extended null-bicharacteristics of P .
Remark 8.6. Note that in order to obtain the propagation of microlocal H s -regularity of
u in Char(P ), one needs to assume that f lies in H s−m+1 microlocally. This is one degree
more smoothness than what microlocal elliptic regularity requires. Thus, on Char(P ), u
‘loses’ one derivative relative to elliptic regularity.

From now on, all ps.d.o.s will have Schwartz kernels supported in a fixed compact subset
of M ×M , and all distributions are supported in a fixed compact subset K ⊂ M . This is not
a restriction in view of the arguments at the beginning of the proof of Theorem 8.2. Note
that {u ∈ Hcs (M ) : supp u ⊂ K} has the structure of a Hilbert space. We will really prove
quantitative estimates somewhat similar to those that arose in the discussion of microlocal
elliptic regularity, see (6.44). Namely, we will show:
Theorem 8.7. We use the notation of Theorem 8.5. If
B, G, E ∈ Ψ0 (M ) (8.10)
are such that
(1) WF0 (B) ⊂ Ell(G),
(2) all backward null-bicharacteristics of P starting from a point in WF0 (B) ∩ Char(P )
enter Ell(E) in finite time while remaining in Ell(G),
then the following estimate holds for any N ∈ R and a constant C = C(N ) > 0:
kBukH s ≤ C (kEukH s + kGP ukH s−m+1 + kukH −N ) . (8.11)
Moreover, this holds in the strong sense that if u ∈ E 0 (M ) is such that the right hand side
is finite, then so is the left hand side, and the estimate holds.

One reads the estimate (8.11) as follows: assuming a priori microlocal H s control of u
on Ell(E), we conclude microlocal H s control of u on Ell(B) by propagation along null-
bicharacteristics (provided P u remains microlocally in H s−m+1 along the way).
Theorem 8.5 is an immediate consequence of Theorem 8.7. Indeed, if α ∈ Char(P ) \
WFs (u), then this implies that the forward null-bicharacteristic of P with initial condition
α remains disjoint from WFs (u) as long as it does not intersect WFs−m+1 (f ). Applying
the theorem to −P gives the backward propagation of regularity. (Away from Char(P ),
the estimate (8.11) follows from microlocal elliptic regularity in view of condition (1) in
88 PETER HINTZ

Theorem 8.7, though in a weak form since we are assuming microlocal H s−m+1 control on
P u).

8.1. Positive commutator argument I: sketch. Let us consider a basic example: P =

Dx1 on Rnx , so p = ξ1 and Hp = ∂x1 . Let us take B, E, G to be cutoff functions, say
smoothed out versions of the characteristic functions of QE := [−2, −1]x1 × [−2, 2]n−1 (for
E), QB := [1, 2] × [−1, 1]n−1 (for B) and [−3, 3] × [−3, 3]n−1 (for G). Take s = 0; assume
Dx1 u = f ∈ L2 (Rn ). (8.12)
Then (8.11) asserts that u|QB ∈ L2 , provided u|QE ∈ L2 . But this is obvious! Indeed, one
can solve (8.12) explicitly, and do simple estimates. (The ‘high brow’ proof of Theorem 8.7,
see [DH72], reduces to this situation (modulo smoothing operator) using Fourier integral
operators.)
A much better proof does not require the explicit solution of (8.12). (This ‘better proof’
does use the fundamental theorem of calculus, but in the form of integration by parts, and
in a way that generalizes readily to general operators.) To explain it, let us take n = 1,
x = x1 , for simplicity, and let us estimate |u(1)|2 in terms of |u(−1)|2 and Dx u. Let
χ = 1[−1,1] denote the characteristic function of [−1, 1]. Then
Z
2 2
−|u(1)| + |u(−1)| = i[Dx , χ]u · ū dx
Z Z 
−1 (8.13)
=i Dx u · χū dx − χu · Dx u dx

= 2 ImhDx u, χui.
If Dx u = 0, we see from the first line that what really provides control of |u(1)|2 is the
fact that the cutoff χ has negative (the ‘good’ sign) derivative along Hp at 1. Since we are
proving a localized estimate, the function χ must be compactly supported, and hence it
must have a positive (the ‘bad’ sign) derivative somewhere, here at −1, which necessitates
a priori control of u there.
If Dx u 6= 0, one needs more ‘negativity’ of the commutator i[Dx , χ]; one can e..g take
Z 1
−x
2 ImhDx u, e χui = − e−x |u|2 dx − e−1 |u(1)|2 + e|u(−1)|2 , (8.14a)
−1
and estimate the left hand side using Cauchy–Schwarz by
2 ImhDx u, e−x χui ≥ −kχe−x/2 Dx uk2L2 − ke−x/2 uk2L2 ([−1,1]) . (8.14b)
Combining (8.14a)–(8.14b) gives |u(1)|2 ≤ C(|u(−1)|2 + kDx uk2L2 ), as desired. This is a
typical positive commutator argument; the function e−x χ is called the commutant.16

The proof of Theorem 8.7 will be based on similar considerations. The rough, formal
sketch goes as follows.17 We formally compute for A = Op(a) ∈ Ψ2s−m+1 , A = A∗ ,
2 ImhP u, Aui = i hAu, P ui − hP u, Aui = (i[P, A] + i(P ∗ − P )A)u, u .



(8.15)
16Strictly speaking, one should call it a negative commutator argument, which can be turned into a
positive commutator argument by switching the sign of the commutant. However, people typically use
positive commutants, and we will do the same here.
17We encourage the reader to assume at first reading that P = P ∗ .
 MICROLOCAL ANALYSIS 89

Let p = σm (P ) and p1 = σm−1 (i(P ∗ − P )). Then

σ2s i[P, A] + i(P ∗ − P )A = Hp a + p1 a.

(8.16)

Suppose we can arrange

Hp a + p1 a = −b2 + e0 (8.17)

where b ∈ S s is elliptic in the desired conclusion region, and e0 ∈ S 2s has essential support
contained in the a priori control region. Taking B = Op(b), E 0 = Op(e0 ), we then have

i[P, A] + i(P ∗ − P )A = −B ∗ B + E 0 + R, R ∈ Ψ2s−1 , (8.18)

hence (8.15) implies

kBuk2 = −2 ImhP u, Aui + hE 0 u, ui + hRu, ui. (8.19)

If P u = 0, this controls the microlocal H s norm of u on Ell(B) by that on Ell(E 0 ). (Note

that, using Lemma 6.34, hE 0 u, ui is finite by the a priori H s control on u on WF0 (E 0 ).) The
term hRu, ui is lower order, and finite provided we already have proved H s−1/2 control of
u. Thus, starting with s = −N + 1/2, we can iteratively improve the control on u by half
a derivative, until after finitely many iterations we reach the desired level of regularity. (If
P u 6= 0, we arrange for more negativity in (8.17) and estimate the first term in (8.19) using
Cauchy–Schwarz, similarly to (8.14a)–(8.14b) above.)
To make this into an honest positive commutator argument, we need to

(1) construct the commutant (this is the ‘interesting’ part of the argument), i.e. con-
struct a satisfying (8.17);
(2) regularize the argument (this is the ‘technical’ but straightforward part of the ar-
gument): we need to ensure that the integrations by parts in (8.15) and (8.19) as
well as various norms are well-defined.

8.2. Positive commutator argument II: construction of the commutant. We first

do some preliminary simplifications. Using a partition of unity argument, it suffices to work
near a single null-bicharacteristic segment

γ : [0, s0 ] 3 s 7→ γ(s) ∈ Char(P ), s0 > 0. (8.20)

We will show that γ(0) ∈/ WFs (u) and γ([0, s0 ]) ∩ WFs−m+1 (P u) = ∅ implies γ(s0 ) ∈
/
s
WF (u), with estimates.
We further simplify notation by passing to the cosphere bundle.

Lemma 8.8. Let p ∈ Shom m (T ∗ M \ o). Then H is homogeneous of degree m − 1. That is,
p
denoting by Mλ : (x, ξ) 7→ (x, λξ), λ > 0, the dilation in the fibers of T ∗ M , we have

Mλ∗ Hp = λm−1 Hp . (8.21)

90 PETER HINTZ

Proof. We work in local coordinates. Since p(x, λξ) = λm p(x, ξ), differentiation in ξ shows
m−1
that ∂ξi p ∈ Shom . Let now f ∈ C ∞ (T ∗ M ) and (x0 , ξ0 ) ∈ T ∗ M \ o, then
(Mλ∗ Hp )|(x0 ,ξ0 ) (f ) = Hp |(x0 ,λξ0 ) (f ◦ Mλ−1 )
= (∂ξ p)(x0 , λξ0 ) · ∂x (f ◦ Mλ−1 ) (x0 , λξ0 )

− (∂x p)(x0 , λξ0 ) · ∂ξ (f ◦ Mλ−1 ) (x0 , λξ0 )

(8.22)
= λm−1 (∂ξ p)(x0 , ξ0 ) · (∂x f )(x0 , ξ0 )
− λm (∂x p)(x0 , ξ0 ) · λ−1 (∂ξ f )(x0 , ξ0 )
= λm−1 Hp |(x0 ,ξ0 ) f,
as claimed. 

Fix an elliptic symbol

1
|ξ| ∈ Shom (T ∗ M \ o) (8.23)
and define
H̃p := |ξ|−m+1 Hp ∈ V(T ∗ M \ o). (8.24)
This is homogeneous of degree 0 and hence descends to a smooth vector field on the
cotangent bundle H̃p0 ∈ V(S ∗ M ).18 (Indeed, for f ∈ C ∞ (S ∗ M ), one defines H̃p0 f by
q ∗ (H̃p0 f ) := H̃p (q ∗ f ) where q : T ∗ M \ o → S ∗ M is the quotient map.) We immediately
simplify notation and denote H̃p0 by H̃p simply. An integral curve of H̃p is the image in
T ∗ M \ o of an integral curve of Hp , up to reparameterization.
Working in S ∗ M , we thus take γ in (8.20) to be an integral curve of H̃p ; we may rescale to
arrange that s0 = 1. Note that there is nothing to prove if γ is stationary, cf. Remarks 8.3–
8.4. Otherwise, H̃p is non-zero along γ, and by basic ODE theory, we can straighten out
H̃p locally: there exist local coordinates on S ∗ M
(z1 , z 0 ), z1 ∈ [−2, 2], z 0 ∈ R2n−2 , |z 0 | < 1, (8.25)
near γ([0, 1]) such that γ(0) = (0, 0), H̃p = ∂z1 , and γ(1) = (1, 0).
We now construct a commutant a; we need to ensure that a is supported in any pre-
specified neighborhood of γ([0, 1]), and that e0 is supported in any pre-specified neighbor-
hood of γ(0). Suppose  > 0 is such that
[−2, 1 + 2] × {|z 0 | < 2} ∩ WFs−m+1 (P u) = ∅,


(8.26)
[−2, 2] × {|z 0 | < 2} ∩ WFs (u) = ∅.

Fix a cutoff (in the transverse directions)

ψ ∈ Cc∞ (R2n−2 ), supp ψ ⊂ {|z 0 | < 2}, ψ(z 0 ) = 1 for |z 0 | < , (8.27)
and a ‘turn-on’ function (in the z1 direction)
χ1 ∈ C ∞ (R), supp χ1 ⊂ (−, ∞), supp(1 − χ1 ) ⊂ (−∞, ). (8.28)

18The passage from H̃ ∈ V(T ∗ M \ o) to H̃ ∈ V(S ∗ M ) does lose information, namely the fiber-radial
p p
component of H̃p . For example, the vector field ξ∂ξ ∈ V(T ∗ Rn ), which is homogeneous of degree 0, descends
to the 0 vector field on S ∗ Rn . Keeping track of the radial component will be crucial in §9.
 MICROLOCAL ANALYSIS 91

The main term of the commutant arises from

(
e−F/x , x > 0,
χ0 (x) := , (8.29)
0, x≤0

where the constant F > 1 will be chosen below. Note that

χ00 (x) = F x−2 χ0 (x), (8.30)
so for x in any fixed compact subset I b [0, ∞), and for any given C > 0, we can choose
F  1 so that χ00 ≥ Cχ0 for x ∈ I. That is, the derivative of χ0 can be made to dominate
any multiple of χ0 ; this is an important mantra in the commutant construction business.
We then set
χ(z1 ) := χ0 (1 +  − z1 ), ã := χ(z1 )χ1 (z1 )2 ψ(z 0 )2 , (8.31)
which is supported in a 2-neighborhood of γ([0, 1]). Setting p̃1 = |ξ|−m+1 p 1 ∈ C ∞ (S ∗ M )
(using the notation p1 = σm−1 (i(P ∗ − P )) from (8.16)), we then compute
H̃p ã + p̃1 ã = 2χ(z1 )χ1 (z1 )χ01 (z1 )ψ(z 0 )2 + χ0 (z1 )χ1 (z1 )2 ψ(z 0 )2 + p̃1 χ(z1 )χ1 (z1 )2 ψ(z 0 )2
= −b̃2 + ẽ0 ,
(8.32)
where
ẽ0 = 2χ(z1 )χ1 (z1 )χ01 (z1 )ψ(z 0 )2 ,
b̃ = χ1 (z1 )ψ(z 0 ) −χ0 (z1 ) − p̃1 χ(z1 )
p
(8.33)
= χ1 (z1 )ψ(z 0 ) χ(z1 ) F (1 +  − z1 )−2 − p̃1 .
p p

Note that for F sufficiently large, b̃ ∈ C ∞ (S ∗ M ). Moreover, supp ẽ0 ⊂ {|z1 |, |z 0 | < 2}, and
b̃ is positive on [, 1] × {0}.
This almost arranges (8.17): we need to put the differential order back. Thus, we set
2s−m+1
a := |ξ|2s−m+1 ã ∈ Shom (T ∗ M \ o) (8.34)
and compute
Hp a + p1 a = |ξ|2s H̃p ã + p̃2 ã , p̃2 := p̃1 + |ξ|−2s+m−1 H̃p |ξ|2s−m+1 .



(8.35)
Therefore, giving ourselves some extra room (to deal with P u 6= 0), we have
Hp a + p1 a = −|ξ|2m−2s−2 a2 − b2 + e0 (8.36)
if we set
e0 = |ξ|2s ẽ0 , (8.37)
b = |ξ|s χ1 (z1 )ψ(z 0 ) −χ0 (z1 ) − p̃2 χ(z1 ) − χ(z1 )2 χ1 (z1 )2 ψ(z 0 )2
p
(8.38)
= |ξ|s χ1 (z1 )ψ(z 0 ) χ(z1 ) F (1 +  − z1 )−2 − p̃2 − χ(z1 )χ1 (z1 )2 ψ(z 0 )2 ;
p p
(8.39)

we have e0 ∈ S 2s and b ∈ S s for sufficiently large F > 1.

92 PETER HINTZ

8.3. Positive commutator argument III: a priori estimate. Let us quantize these
symbols as in (8.18), giving A ∈ Ψ2s−m+1 , B ∈ Ψs , E 0 ∈ Ψ2s ; we can also arrange WF0 (A) =
ess supp a etc. Assuming u ∈ C ∞ (M ), integrations by parts are never a concern, and we
then have the following slight improvement over (8.19):
kBuk2 + kΛAuk2 = −2 ImhP u, Aui + hE 0 u, ui + hRu, ui, R ∈ Ψ2s−1 , (8.40)
where Λ ∈ Ψm−s−1 (M ) is elliptic with principal symbol |ξ|m−s−1 ; and WF0 (R) ⊂ WF0 (A).
Let Λ− ∈ Ψ−m+s+1 (M ) denote an elliptic parametrix of Λ, with I = Λ− Λ + R̃, R̃ ∈
Ψ−∞ (M ). Fix an operator G ∈ Ψ0 (M ) with WF0 (I − G) ∩ WF0 (A) = ∅; in particular, G is
elliptic on WF0 (A). We then have
2| ImhP u, Aui| ≤ 2| ImhGP u, (R̃ + Λ− Λ)Aui| + 2| ImhP u, (I − G∗ )Aui|
(8.41)
≤ kΛ∗− GP uk2 + kΛAuk2 + Ckuk2H −N .

Let E ∈ Ψ0 (M ) be elliptic on WF0 (E 0 ). Plugging (8.41) into (8.40), we then get the
estimate


kBukL2 ≤ C kGP ukH s−m+1 + kEukH s + kGukH s−1/2 + kukH −N . (8.42)
If s − 1/2 ≤ −N , we simply estimate kGukH s−1/2 ≤ CkukH −N , obtaining the desired
estimate (8.11). For s > −N + 1/2, we can control kGukH s−1/2 inductively. Indeed, if
WF0 (G) lies in an 2−|s|−2 neighborhood of WF0 (A), one can control kGukH s−1/2 by the
right hand side of (8.42) with E, G replaced by operators Ẽ, G̃ elliptic on WF0 (E), WF0 (G)
and with operator wave front set in an 2−|s|−1 neighborhood of WF0 (E), WF0 (G). After
finitely many iterations, we thus obtain the desired estimate


kBukL2 ≤ C kG̃P ukH s−m+1 + kẼukH s + kukH −N , (8.43)

where Ẽ, G̃ ∈ Ψ0 , with WF0 (Ẽ) in a 3-neighborhood of γ(0), and WF0 (G̃) in a 3-
neighborhood of γ([0, 1]). Starting out with  replaced by 32 , we have the desired a priori
estimate.
Remark 8.9. While these arguments required the a priori membership u ∈ C ∞ (M ) (or at
least to have sufficiently high regularity), the estimate (8.33) is highly non-trivial as an
a priori estimate, as it gives quantitative control on the microlocal H s -mass of u along
γ([0, 1]).

8.4. Positive commutator argument IV: regularization. We now regularize the ar-
gument so that u ∈ H −N together with some microlocal regularity is sufficient. By an
inductive argument as above, we may moreover assume that WFs−1/2 (u) is disjoint from
a 2-neighborhood of γ([0, 1]). The a priori assumption is that WFs (u) is disjoint from a
2-neighborhood of γ(0).
The regularization argument replaces a, b, e0 by symbols ar , br , e0r , r ∈ (0, 1], of (much)
lower symbolic order, which converge to a, b, e0 as r → 0 (or rather, to a, b, e0 multiplied by
a cutoff which cuts away the singularity at ξ = 0) in slightly weakened symbol classes. We
first deal with the symbolic construction.
For K > 1, define
φr (τ ) = (1 + rτ )−K/2 , r ∈ (0, 1]. (8.44)
 MICROLOCAL ANALYSIS 93

Thus, φr (|ξ|2 ) ∈ S −K (T ∗ M ) is uniformly bounded in S 0 (T ∗ M ), and converges to φ0 ≡ 1

in the topology of S δ (T ∗ M ) for any δ > 0. Note moreover that
rτ
τ φ0r (τ ) = fr (τ )φr (τ ), fr (τ ) = −(K/2) , (8.45)
1 + rτ
so in particular |fr (τ )| ≤ K/2. With η ∈ C ∞ (R), vanishing near 0 and identically 1 outside
[−1, 1], we then define the regularized commutant
ar = φr (|ξ|2 )η(|ξ|) · a
(8.46)
= φr (|ξ|2 )η(|ξ|) · |ξ|2s−m+1 χ(z1 )χ1 (z1 )2 ψ(z 0 )2 .

Thus, ar ∈ L∞ ((0, 1]r , S 2s−m+1 (T ∗ M )), and ar ∈ S 2s−m+1−K (T ∗ M ) for r > 0. In addition
to the terms in (8.36), the computation of Hp ar produces two extra terms: for Hp falling
on φr , we get a term involving
H̃p φr (|ξ|2 ) = f˜r φr (|ξ|2 ), f˜r := (|ξ|−2 H̃p |ξ|2 )fr (|ξ|2 );


(8.47)

note that η(|ξ|)f˜r ∈ L∞ ((0, 1]r , S 0 (T ∗ M )) is uniformly bounded. When Hp falls on η(|ξ|),
we get a symbol with compact support in ξ, which is hence of order −∞.
Using the notation of (8.32) and (8.35), we then compute
Hp ar + p1 ar = |ξ|2s φr (|ξ|2 )η(|ξ|) H̃p ã + p̃2 ã

+ f˜r φr (|ξ|2 )η(|ξ|)|ξ|m−1 a + φr (|ξ|2 )(Hp η(|ξ|))a (8.48)

= −|ξ|2m−2s−2 a2r − b2r + e0r
where
e0r = φr (|ξ|2 ) η(|ξ|) · 2|ξ|2s χ(z1 )χ1 (z1 )χ01 (z1 )ψ(z 0 )2 + (Hp η(|ξ|))a ,

br = |ξ|s φr (|ξ|2 )η(|ξ|)χ1 (z1 )ψ(z 0 ) χ(z1 )

p p
(8.49)
q
× F (1 +  − z1 )−2 − p̃2 − f˜r − φr (|ξ|2 )η(|ξ|)χ(z1 )χ1 (z1 )2 ψ(z 0 )2 .

Since f˜r is uniformly bounded, the extra term f˜r here is harmless: choosing F > 1 suffi-
ciently large makes the square root well-defined. Indeed, we have
br ∈ L∞ ((0, 1]r ; S s (T ∗ M )), e0r ∈ L∞ ((0, 1]r ; S 2s (T ∗ M )). (8.50)
Moreover, by construction, supp ar , supp br and supp e0r are contained in 2-neighborhoods
of γ([0, 1]) and γ(0), respectively.
The quantization of (8.49) requires a bit of care since we need more precision than that
afforded by a quantization which only respects principal symbols. Recalling the construction
in §5.6, we thus fix a linear continuous quantization map
Op : S m (T ∗ M ) → Ψm (M ) (8.51)
P
by Op(a) = φ̃i Op(ai )φi , where φi is a partition of unity on M subordinate to a cover
by coordinate systems, φ̃i = 1 near supp φi , and ai ∈ S m (Rn ; Rn ) is the local coordinate
expression for a. Thus, Op is a quantization map in the sense that σm (Op(a)) = [a] for
a ∈ S m (T ∗ M ), and Op is surjective modulo Ψ−∞ (M ). This definition also ensures that Op
is continuous, and WF0 ◦ Op = ess supp.
94 PETER HINTZ

Let then
Ar = Op(ar ) ∈ L∞ ((0, 1]r ; Ψ2s−m+1 (M )),
Br = Op(br ) ∈ L∞ ((0, 1]r ; Ψs (M )), (8.52)
Er0 = Op(e0r ) ∞ 2s
∈ L ((0, 1]r ; Ψ (M )).
Letting Λ = Op(hξim−s−1 ), we then have
i[P, Ar ] + i(P ∗ − P )Ar = −(ΛAr )∗ (ΛAr ) − Br∗ Br + Er0 + Rr ,
(8.53)
Rr ∈ L∞ ((0, 1]r , Ψ2s−1 (M )).
The orders of Ar , Br and Er0 , Rr are lower by K and 2K, respectively, for r > 0. Thus, if
we take K large enough (depending on s and N ), we can safely compute
2 ImhP u, Ar ui = h(i[P, Ar ] + i(P ∗ − P )Ar )u, ui
(8.54)
= −kΛAr uk2 − kBr uk2 + hEr0 u, ui + hRr u, ui.

We need to show that the final two terms are uniformly bounded for r ∈ (0, 1]. The
crucial insight is that we have uniform control on Ar etc. in the following sense:
Definition 8.10. Suppose A = {Ar } ∈ L∞ ((0, 1]r ; ΨN (M )) is a bounded family (for some
N ∈ R) of ps.d.o.s on M . Then α ∈ S ∗ M does not lie in the uniform wave front set
WF0L∞ (A) ⊂ S ∗ M if and only if there exists an operator B ∈ Ψ0 (M ), elliptic at α, such
that BAr is bounded in Ψ−∞ (M ).

(This generalizes WF0 : if Ar = A is r-independent, then WF0L∞ (A) = WF0 (A).) We

then have the following extension of microlocal elliptic regularity:
Lemma 8.11. Let A = {Ar } ∈ L∞ ((0, 1]r ; Ψm (M )). Suppose B ∈ Ψ0 (M ) is such that
WF0L∞ (A) ⊂ Ell(B). Let s, N ∈ R. Then there exists a constant C (independent of r) such
that

kAr ukH s−m ≤ C kBukH s + kukH −N . (8.55)

Proof. Writing I = QB + R with Q, R ∈ Ψ0 (M ), WF0 (R) ∩ WF0L∞ (A) = ∅, we have

Ar u = Ar QBu + Ar Ru. (8.56)
But Ar Q ∈ L∞ ((0, 1]r ; Ψm (M )) and Ar R ∈ L∞ ((0, 1]r ; Ψ−∞ (M )) are uniformly bounded;
this implies (8.55). 

By construction, we have WF0L∞ ({Ar }) ⊂ ess supp a etc. Let us thus take G ∈ Ψ0 (M ),
elliptic near WF0L∞ ({Ar }) and with WF0 (G) contained in a 2-neighborhood of γ([0, 1]), and
E ∈ Ψ0 (M ) elliptic near WF0L∞ ({Er0 }) and with WF0 (E) contained in a 2-neighborhood
of γ(0); we then conclude that
|hEr0 u, ui| ≤ C kEuk2H s + kuk2H −N ,


(8.57)
|hRr u, ui| ≤ C kGuk2H s−1/2 + kuk2H −N .

Plugging this into (8.54) and arguing as in (8.41)–(8.42), we thus obtain a uniform estimate


kBr ukL2 ≤ C kEukH s + kGP ukH s−m+1 + kGukH s−1/2 + kukH −N . (8.58)
Since the unit ball in L2 is compact, Br u has a weakly convergent subsequence with limit
v ∈ L2 . On the other hand, Br u → B0 u in D 0 (M ); hence B0 u = v ∈ L2 . Therefore,
 MICROLOCAL ANALYSIS 95

WFs (u) ∩ Ell(B0 ) = ∅, proving microlocal H s -regularity of u at γ(1), and at the same time
giving an estimate for kB0 ukL2 ≤ lim inf kBr ukL2 by the right hand side of (8.58).
The proof of Theorem 8.7 is complete.
Remark 8.12. Theorems 8.5 and 8.7 also hold for operators P ∈ Ψm (M ; E) acting on
sections of a vector bundle E, provided P has a scalar, homogeneous principal symbol, see
Definition 5.44. To extend the proof to this case, one fixes an arbitrary smooth fiber inner
product on E. The main change is that the ‘subprincipal’ symbol p1 is now endomorphism-
valued, and hence so is p̃1 ∈ C ∞ (S ∗ M ; End(π ∗ E)). This is inconsequential however
√ since
the square root in (8.33) is still well-defined (using the power series expansion for 1 − S
for S ∈ End(Ex ) with kSk ≤ 12 ).

8.5. Exercises.
Exercise 8.1. Suppose u ∈ D 0 (R2 ) solves the Keldysh equation
(xDx2 + Dy2 )u = f ∈ C ∞ (R2 ). (8.59)
Assume that
WF(u) ∩ N ∗ {x = 0} = ∅. (8.60)
∞ 2
Show that u ∈ C (R ). Show also that there exist solutions of the equation (8.59) which
are not smooth (and which thus necessarily violate (8.60)).
Exercise 8.2. Suppose u ∈ D 0 (R2 ) solves the Tricomi equation (Dx2 + xDy2 )u = f ∈ C ∞ (R2 ).
Assume that u = u(x, y) is smooth for x < −1. Show that u ∈ C ∞ (R2 ).
Exercise 8.3. Suppose u ∈ D 0 (R2 ) satisfies xu ∈ C ∞ (R2 ) and yu ∈ C ∞ (R2 ).
(1) Show that WF(u) ⊂ T0∗ R2 \ o = {(x, y; ξ, η) : (x, y) = (0, 0), (ξ, η) 6= (0, 0)}.
(2) Suppose that there exists α ∈ T0∗ R2 \ o with α ∈/ WF(u). Show that WF(u) = ∅.

9. Propagation of singularities at radial points

The propagation theorem proved in §8 is a general purpose tool for analyzing the regu-
larity of solutions of general linear PDE P u = f when P ∈ Ψm (M ) has real homogeneous
principal symbol, assuming one has information on u somewhere to begin with. In par-
ticular, in view of the (necessary) a priori control assumption of microlocal regularity of u
(encoded by the term Eu in the estimate (8.11)), one cannot, in general, control u globally
only in terms of f .
What is needed for global control of u is the existence of a subset of phase space S ∗ M
where one can get unconditional control of u. There are two main situations in which this
happens:
(1) initial value problems. Consider, as the simplest example, the forcing problem for
the wave equation on Rn ,
(
u(t, x) = g(t, x), t ∈ R, x ∈ Rn ,
(9.1)
u(0, x) = Dt u(0, x) = 0, x ∈ Rn ,
and assume that t ≥ 1 on supp g. By Corollary 7.2, equation has a unique solution
u, which is necessarily equal to 0 for t < 1. A fortiori, u is smooth there, and we
can then analyze the regularity of u for later times using Theorem 8.5. This gives
96 PETER HINTZ

more information than Corollary 7.2, since we can precisely study situations where
the forcing term g is smooth in some places but singular at others.
We remark that our discussion of hyperbolic evolution equations in §7 was based
on a product decomposition of Rn+1 into Rt ×Rnx , starting already with the function
space we used for g in (7.2); this is a sensible setting for the study of the operator
Dt −A(t) there, which is not a ps.d.o. in general, unless A(t) is a differential operator.
The wave operator in (9.1) can be analyzed both from this product perspective (§7)
and from the ‘spacetime’ perspective (§8) in which one simply views  as an operator
 ∈ Ψ2 (Rn+1 ).
(2) radial points (or other degeneracies) of P .
Let us give a simple example of an operator with radial points. Let P = x ∈ Ψ0 (Rn ),
Rn = Rx × Rn−1 y , be the multiplication operator, with principal symbol p(x, y, ξ, η) = x,
characteristic set
Σ := Char(P ) = {(x, y, ξ, η) : x = 0, (ξ, η) 6= (0, 0)} ⊂ T ∗ Rn \ o, (9.2)
and Hamiltonian vector field Hp = −∂ξ . Suppose u ∈ D 0 (Rn ) solves the ‘PDE’
P u = xu = f ∈ C ∞ (Rn ). (9.3)
Elliptic regularity (Proposition 6.27) or common sense imply that WF(u) ⊂ Σ. By the
propagation of singularities (Theorem 8.2), WF(u) is a union of maximally extended null-
bicharacteristics of P . Note that at (x, 0, ξ, 0) ∈ Σ, Hp = −∂ξ is radial; the null-bichar-
acteristic remains in the half-line {(x, 0, cξ, 0) : c > 0}, hence the propagation theorem is
vacuous there. Let us thus define the following two sets of radial points:
R± = {(0, y, ξ, 0) : ± ξ > 0} ⊂ Σ. (9.4)

Now, the general solution of the PDE (9.3) is of the form

u(x, y) = u+ (y)(x + i0)−1 + u− (y)(x − i0)−1 + ũ(x, y), (9.5)
where u± ∈ D 0 (Rn−1 ), u+ (y) + u− (y) = f (0, y) (note that u± do not need to be smooth!),
and ũ ∈ C ∞ (Rn ).
Proposition 9.1. Suppose u ∈ D 0 (Rn ), xu = f ∈ C ∞ (Rn ), and WFs0 (u) ∩ R+ = ∅ for
some s0 > − 21 . Then WF(u) ⊂ R− . Moreover, WFs (u) = ∅ for all s < − 12 .

Proof. The key observation is that (x ± i0)−1 ∈ Hloc s (Rn ) if and only if s < − 1 . The as-
2
sumption thus implies that u+ ≡ 0; thus u− (y) = f (0, y) is smooth, and the first conclusion
follows from the fact that WF(u+ (y)(x − i0)−1 ) ⊂ R− , see Example 6.25. The second
−1/2−
conclusion then follows again from the fact that (x − i0)−1 ∈ Hloc (Rn ) for all  > 0. 

This can be broken down into a concatenation of three arguments:

(1) if WFs0 (u) ∩ R+ = ∅, then WF(u) ∩ R+ = ∅, hence u is microlocally smooth in a
neighborhood of R+ ;
(2) by propagation of regularity, WF(u) ⊂ R− ;
(3) if WF(u) is disjoint from a punctured neighborhood of R− , then u is microlocally
in H s at R− for all s < − 21 .
 MICROLOCAL ANALYSIS 97

Parts (1) and (3) are special cases of a general result on the propagation of singulari-
ties/regularity at radial points proved below. A key feature is that there is a threshold
regularity: if the microlocal regularity of u exceeds a threshold (here − 12 ) at R+ , then u
is microlocally smooth at R+ (provided f is) and we can propagate H s regularity out of
R+ for s > − 21 ; on the other hand, one can conclude microlocal regularity of u below this
threshold when propagating into R− .

9.1. Intermezzo: radial compactification of phase space. We pause to describe a

convenient and intuitive point of view for understanding qualitative properties of null-
bicharacteristic flows.
Definition 9.2. The radial (or projective) compactification of Rn is the set Rn = Rn tSn−1 ,
equipped with the structure of a manifold with boundary as follows: writing 0 6= x ∈ Rn
in polar coordinates as x = rω, r > 0, ω ∈ Sn−1 , then


Rn = Rn ∪ [0, ∞)ρ × Sn−1 / ∼, (9.6)

where Rn 3 rω ∼ (r−1 , ω). Thus, ρ−1 (0) ∼

= Sn−1 is the ‘sphere at infinity’, and Rn ⊂ Rn is
the interior.
Remark 9.3. We have C ∞ (Rn ) = Scl 0 (Rn ): being smooth on Rn precisely means having a
−1 µ
Taylor expansion in ρ = r at ρ = 0. More generally, Scl (Rn ) = ρ−µ C ∞ (Rn ), in the sense
µ
that the space of restrictions of elements of ρ−µ C ∞ (Rn ) is equal to Scl (Rn ).

Convenient local coordinates near ∂Rn are projective coordinates: write x = (x1 , . . . , xn ),
and let us work in the subset of Rn where x1 >  max(|x2 |, . . . , |xn |). We then let
1 xj
ρ1 := , x̂j := , j = 2, . . . , n. (9.7)
x1 x1
Then (ρ1 , x̂2 , . . . , x̂n ) (with |x̂j | < −1 ) is system of local coordinates on Rn which by
continuity extends to a local coordinate system
[0, ∞)ρ1 × {(x̂2 , . . . , x̂n ) : |x̂j | < −1 , j = 2, . . . , n} (9.8)
on Rn . Together with the standard coordinate system on Rn , such coordinate systems
(upon permuting indices and taking  > 0 small enough) cover Rn .
Lemma 9.4. Let A ∈ GL(n, R). Then matrix-vector multiplication Rn 3 x 7→ Ax ∈ Rn
extends, by continuity, to a diffeomorphism A : Rn → Rn .

Proof. This is an easy verification in projective coordinate systems. 

This lemma allows us to define radial compactifications of vector bundles:

Definition 9.5. Let E → M be a real rank k vector bundle. Then the radial compactifi-
cation E → M is the fiber bundle obtained by radially compactifying each fiber of E. (In
local trivializations of E, the transition maps of E are the continuous extensions of those
of E using Lemma 9.4.)

As microlocal analysts, we are interested in the radially compactified cotangent bundle

T ∗ M → M. (9.9)
98 PETER HINTZ

Note that for p ∈ M , we can identify Sp∗ M with the sphere at infinity of Tp∗ M ; this embeds
S∗M ⊂ T ∗M (9.10)
as a submanifold, called fiber infinity. We can now make the relationship between homo-
geneous vector fields and vector fields on S ∗ M more precise.
Lemma 9.6. Suppose V ∈ V(T ∗ M \ o) is homogeneous of degree 0. Then V extends by
continuity to a smooth vector field
V ∈ V(T ∗ M \ o) (9.11)
which is tangent to S ∗ M .

Proof. Indeed, in local coordinates (x1 , . . . , xn , ξ1 , . . . , ξn ) on T ∗ M , this means that

n
X
V = aj (x, ξ)∂xj + bjk (x, ξ)ξk ∂ξj , (9.12)
j=1

where aj (x, λξ) = aj (x, ξ) and bjk (x, λξ) = bjk (x, ξ) for all λ > 0. Let us work in projective
coordinates
1 ξj
ρ = , ξˆj = , j = 2, . . . , n (9.13)
ξ1 ξ1
in ξ1 >  max(|ξ2 |, . . . , |ξn |). Then aj (x, ξ) = aj (x, (1, ξˆ2 , . . . , ξˆn )) is smooth down to ρ = 0,
and so is bjk . Moreover, ∂xj ∈ V(T ∗ M ). It remains to compute for 2 ≤ i, j ≤ n:
n
X
ξ1 ∂ξ1 = −ρ∂ρ − ξˆk ∂ξ̂k ,
k=2
ξ1 ∂ξj = ∂ξ̂j ,
n (9.14)
X
ξi ∂ξ1 = −ξˆi ρ∂ρ − ξˆi ξˆk ∂ξ̂k ,
k=2
ξi ∂ξj = ξˆi ∂ξ̂j .
This proves the lemma. 

9.2. Radial point estimates: a simple example. In the coordinates used in (9.2),
consider again the equation P u := xu = f and the Hamiltonian vector field Hp = −∂ξ . In
projective coordinates
η

ρ = ξ −1 , η̂ = , ξ > |η|, (9.15)
ξ
let us rescale this to the homogeneous degree 0 vector field
V = ξHp = −ξ∂ξ = ρ∂ρ + η̂∂η̂ , ρ > 0, |η̂| < −1 . (9.16)
Restricting this to a vector field on S ∗ Rn ,
the first term disappears, and we see that Hp
being radial means that V |S ∗ Rn vanishes on
∂R+ := {(x, y, ρ, η̂) : x = 0, η̂ = 0} ⊂ S ∗ Rn , (9.17)
the boundary of R+ (from (9.4)) at fiber infinity.
In our quest to prove microlocal estimates at ∂R+ via positive commutators, we therefore
need to make use of the first summand in (9.16): we need to exploit that V has non-trivial
 MICROLOCAL ANALYSIS 99

behavior in the fiber-radial direction, that is, it acts non-trivially on differential weights
ρ−s = ξ s . Concretely, consider a commutant
a = ρ−2s+1 ψ(η̂), (9.18)
where ψ ∈ Cc∞ (R) is identically 1 near 0, and satisfies xψ 0 (x) ≤ 0 for all x ∈ R. Then, in
our projective coordinate system (9.15), we have
Hp a = ρV a = ρ−2s −(2s + 1)ψ(η̂) + η̂ψ 0 (η̂) .


(9.19)
Thus, when s > − 21 , both summands have the same (indefinite) sign (namely, they are
≤ 0). Moreover, crucially, the first summand is elliptic at ∂R+ . Thus, quantizing the
calculation (9.19) as in §8, we can write
i[P, A] = −B ∗ B + R, R ∈ Ψ2s−1 , s > − 12 , (9.20)
with B ∈ Ψs elliptic at ∂R+ . Ultimately, this gives an estimate for kBukL2 , thus a mi-
s
crolocal H estimate of u at ∂R+ , without any a priori control. Notice on the other hand
that the ‘positivity’ (meaning: the ‘good’ sign, so negativity. . . ) of the first term in (9.19)
is delicate and limited; thus, error terms from the regularization argument can only be
absorbed when the amount of regularization is limited, which will be the reason for an a
priori regularity assumption at ∂R+ .
Conversely, if s < − 21 , then the two terms in (9.19) have opposite signs, but the first
summand is still elliptic at ∂R+ . Thus, assuming H s control of u on the support of
the second summand (which is contained in a punctured neighborhood of ∂R+ ), we can
conclude H s regularity of u at ∂R+ . (The situation at ∂R− is completely analogous of
course.)

9.3. Radial point estimates: general setup. We now set up the general theorem on
the propagation of singularities/regularity at (generalized) radial points.
Thus, suppose P ∈ Ψm cl (M ) is a classical operator with real homogeneous principal
−1
symbol p.19 Fix an elliptic symbol 0 6= ρ ∈ Scl (T ∗ M ) and let
p̃ := ρm p ∈ C ∞ (T ∗ M \ o), H̃p := ρm−1 Hp ∈ V(T ∗ M \ o). (9.21)
Suppose that
R ⊂ Char(P ) (9.22)
is a smooth submanifold to which H̃p is tangent. Suppose that dp̃ 6= 0 in a neighborhood
of R in S ∗ M . For the sake of definiteness, we assume that R is a source for the H̃p -flow,
in the following precise sense:
(1) Suppose ρ1,j ∈ C ∞ (S ∗ M ), j = 1, . . . , k, define R inside Char(P ), in the sense that
R = {p̃ = 0, ρ1,1 = · · · = ρ1,k = 0}, (9.23)
and dρ1,1 , . . . , dρ1,k are linearly independent at R. Let
k
X
ρ1 = ρ21,j , (9.24)
j=1

19It suffices to assume that P ∈ Ψm (M ), with real homogeneous principal symbol. The only change
is that in equation (9.27) below, β̃ ∈ S 0 is not necessarily smooth on T ∗ M ; what enters in the threshold
quantities in Theorems 9.8 and 9.9 below is then the supremum or infimum of β̃, whichever gives the stronger
requirement.
100 PETER HINTZ

which is a ‘quadratic defining function’ of R. Since H̃p is tangent to R, the deriva-

tives H̃p ρ1,j vanish at R, hence H̃p ρ1 vanishes quadratically at R. We then assume
that there exists a positive function 0 < β1 ∈ C ∞ (S ∗ M ) such that
H̃p ρ1 = β1 ρ1 + F2 + F3 , (9.25)
where F2 ≥ 0, and F3 vanishes cubically at R. (Thus, R is a source for the H̃p -flow
3/2
within Char(P ) ⊂ S ∗ M since |F3 | ≤ Cρ1 ≤ 12 β1 ρ1 near R, so H̃p ρ1 ≥ 12 β1 ρ1 ; cf.
the behavior in the η̂-variables in (9.16).)
(2) We have
H̃p ρ = β0 ρ, β0 |R > 0. (9.26)
Note that since H̃p is tangent to S ∗ M , H̃p ρ vanishes there, hence is of the stated
form with β0 ∈ C ∞ (T ∗ M ) near R. (The assumption (9.26) implies that R is a
source for the H̃p -flow also in the fiber-radial direction.)
The subprincipal part of P at R now plays a significant role, too:20
1
(3) Let p1 := σm−1 ( 2i (P − P ∗ )) and p̃1 := ρm−1 p1 . Define β̃ ∈ C ∞ (S ∗ M ) near R by
p̃1 = β0 β̃. (9.27)
Remark 9.7. Condition (1) is independent of choices, and the positivity of β0 in (9.26) does
not depend on the choice of ρ in the case that H̃p vanishes at R; in general, when H̃p is only
tangent to R, the choice of ρ does matter (but only through the derivative taken in (9.26),
not through the rescaling in (9.21).

We state the main result of this section in two forms, one qualitative (analogous to
Theorem 8.5), one quantitative (analogous to Theorem 8.7).
Theorem 9.8. (Microlocal regularity at radial sets, qualitative statement.) Let P and
R ⊂ Char(P ) ⊂ S ∗ M be as above. Let u ∈ D 0 (M ), P u = f .
(1) (Propagation out of the radial set.) Let s, s0 ∈ R, and suppose that s > s0 > m−1
2 + β̃
s0 s−m+1 s
on R. If WF (u) ∩ R = ∅ and WF (f ) ∩ R = ∅, then WF (u) ∩ R = ∅.
(2) (Propagation into the radial set.) Let s ∈ R, and suppose s < m−1 2 + β̃ on R. If
s s−m+1
WF (u) is disjoint from a punctured neighborhood of R, and if WF (f )∩R = ∅,
then WFs (u) ∩ R = ∅.

The quantitative version (and also slightly more global, though the difference can be
bridged using the propagation estimates of Theorem 8.7) is the following:
Theorem 9.9. (Microlocal regularity at radial sets, quantitative statement.) Let P and
R ⊂ Char(P ) ⊂ S ∗ M be as above. Let u ∈ D 0 (M ), P u = f .
(1) (Propagation out of R.) Let B, G ∈ Ψ0 (M ) be such that
(a) WF0 (B) ⊂ Ell(G);
(b) Ell(G) contains a neighborhood of R;
20For a simple example, consider P = xD − λ ∈ Ψ1 (R). Then P u = 0 e.g. for u = xiλ , suggesting that
x +
the threshold regularity at the radial sets T0∗ R \ o is 21 − Im λ (which is the Sobolev regularity which xiλ
barely fails to have). And indeed, 21 − Im λ = 2i 1
(P − P ∗ ) is the skew-adjoint part of P . (Any additional
terms of even lower order do not contribute to the threshold regularity.)
 MICROLOCAL ANALYSIS 101

(c) all backward null-bicharacteristics of P from WF0 (B)∩Char(P ) tend to R (that

is, ρ1 tends to 0 along them) while remaining in Ell(G).
Then for all s, s0 , N ∈ R such that s > s0 > m−12 + β̃ on R, there exists C > 0 such
s0
that if WF (u) ∩ R = ∅, then


kBukH s ≤ C kGP ukH s−m+1 + kukH −N . (9.28a)
This estimate does not hold in the usual strong sense. However, if B̃ ∈ Ψ0 (M ) is
elliptic at R, then the estimate


kBukH s ≤ C kGP ukH s−m+1 + kB̃ukH s0 + kukH −N (9.28b)
does hold in the strong sense that if all quantities on the right are finite, then so is
the left hand side, and the estimate holds.
(2) (Propagation into R.) Let B, G, E ∈ Ψ0 (M ) be such that
(a) WF0 (B) ⊂ Ell(G);
(b) all forward null-bicharacteristics of P from WF0 (B) ∩ Char(P ) are either con-
tained in R, or enter Ell(E) in finite time, all while remaining in Ell(G).
Then for all s, N ∈ R such that s < m−1 2 + β̃ on R, there exists C > 0 such that


kBukH s ≤ C kGP ukH s−m+1 + kEukH s + kukH −N . (9.29)
This estimate holds in the usual strong sense.

The proof will require a secondary regularization argument, which will use the following
lemma:
Lemma 9.10. Suppose A ∈ L∞ ((0, 1] ; Ψm ) is uniformly bounded, and A → A in Ψm+η
as  → 0, for all η > 0. Then A converges strongly to A in L(H s ; H s−m ); that is, for any
u ∈ H s , we have A u → Au in H s−m as  → 0.

Proof. If u ∈ H s+1 , then we certainly have A u → Au in H s−m . Given u ∈ H s and ρ > 0,

choose u0 ∈ H s+1 with ku − u0 kH s < ρ. Let then 0 > 0 such that for  ∈ (0, 0 ), we have
kA u0 − Au0 kH s−m < ρ. Then for such ,
kA u − AukH s−m ≤ kA (u − u0 )kH s−m + kA u0 − Au0 kH s−m + kA(u − u0 )kH s−m
(9.30)
≤ Cρ + ρ + Cρ,
where C = sup kA kL(H s ,H s−m ) . 

Proof of Theorems 9.8 and 9.9. We follow the steps of the positive commutator argument
in §§8.2–8.4.
• Construction of the commutant for part (1). With p̃ as in (9.21), the quadratic defining
function of R, ρ1 , as in (9.24)–(9.25), and the defining function of fiber infinity, ρ, as used
in (9.26), we set
a := ρ−2s+m−1 φ(p̃)2 ψ(ρ1 )2 . (9.31)
∞ ∞
Here, φ ∈ Cc (R; [0, 1]), φ(0) = 1, so φ(p̃) localizes near Char(P√); and ψ ∈ Cc ([0, ∞); [0, 1])
is 1 near 0 (so ψ(ρ1 ) localizes further near R) and satisfies −ψ 0 ψ ∈ C ∞ ([0, ∞)). (The
latter assumption only requires a bit of thought near the boundary of supp ψ. Taking ψ to
be a variant of e−1/x H(x) there does the job.) Write
H̃p p̃ = q̃ p̃, q̃ = ρ−m H̃p ρm , (9.32)
102 PETER HINTZ

where q̃ is smooth near R. We then compute the symbol of i[P, A] + i(P ∗ − P )A =

∗
i[P, A] + 2 P −P
2i A to be
Hp a + 2p1 a = ρ−m+1 (H̃p a + 2p̃1 a)

= ρ−2s β0 (−2s + m − 1 + 2β̃)φ(p̃)2 ψ(ρ1 )2
(9.33)
+ 2(H̃p ρ1 )φ(p̃)2 ψ 0 (ρ1 )ψ(ρ1 )

+ 2q̃ p̃φ0 (p̃)φ(p̃)ψ(ρ1 )2 .
When the support of ψ and φ is sufficiently small, the terms in the parenthesis here play
the following roles:
(1) the first is elliptic (and negative) at R under the assumptions on s;
(2) the second is non-positive as well, and has essential support contained in a punctured
neighborhood of R;
(3) the third is supported away from Char(P ), hence can be dealt with using elliptic
regularity.
At this point, one can already prove the estimate (9.28a) as in §8.3; we leave this to the
reader.
• Regularization of commutant for part (1). We regularize our commutant a as in §8.4,
see equations (8.44)–(8.45), though with slightly different notation. Thus, let now K > 0
and r ≥ 0, and put
rρ−1
φr (ρ) = (1 + rρ−1 )−K ; ρφ0r (ρ) = fr (ρ)φr (ρ), fr (ρ) = K φr , (9.34)
1 + rρ−1
so φr ∈ L∞ ((0, 1]r ; S 0 ), and φr ∈ S −K for r > 0. We then let
ar := φr (ρ)2 a ∈ L∞ ((0, 1]r ; S 2s−m+1 ), (9.35)
and compute

Hp ar + 2p1 ar = ρ−2s φr (ρ)2 β0 (−2s + m − 1 + 2β̃ + 2fr )φ(p̃)2 ψ(ρ1 )2
+ 2(H̃p ρ1 )φ(p̃)2 ψ 0 (ρ1 )ψ(ρ1 ) (9.36)

+ 2q̃φ0 (p̃)φ(p̃)ψ(ρ1 )2 .
Note that since 0 ≤ fr ≤ K, the amount K of regularization we can do is limited when
propagating out of the radial set R: in order to ensure that the first term is negative at R,
we need s − K > m−12 + β̃, restricting K. Fix such K > 0. For δ > 0 chosen so small that
−1
still s − K − δβ0 > m−12 + β̃ on R, we then write
Hp ar + 2p1 ar = −2δρ2s−2m+2 a2r − b2r − b21,r + hr p, (9.37)
where
q
br := ρ−s φr (ρ)φ(p̃)ψ(ρ1 ) β0 2s − m − 1 + 2β̃ + 2fr + 2δβ0−1 φr (ρ)2 φ(p̃)2 ψ(ρ1 )2 ,



q
−s
b1,r := ρ φr (ρ)φ(p̃) −2(H̃p ρ1 )ψ 0 (ρ1 )ψ(ρ1 ),
hr := 2ρ−2s+m φr (ρ)2 q̃φ0 (p̃)φ(p̃)ψ(ρ1 )2 .
(9.38)
 MICROLOCAL ANALYSIS 103

Thus, br , b1,r ∈ L∞ ((0, 1]r ; S s ) and hr ∈ L∞ ((0, 1]r , S 2s−m ), with orders reduced by K, K,
and 2K, respectively, for r > 0.
• Quantization of the symbol calculation; conclusion of the proof of part (1). Let Ar =
Op(ar ), Br = Op(br ), B1,r = Op(b1,r ), and Hr = Op(hr ), using a full quantization as
in (8.51). Put Λ = Op(ρs−m+1 ). Then (9.37) and (9.38) imply

i[P, Ar ] + i(P ∗ − P )Ar = −2δ(ΛAr )∗ (ΛAr ) − Br∗ Br − B1,r

∗
B1,r + Hr P + Rr , (9.39)

where Rr ∈ L∞ ((0, 1]r ; Ψ2s−1 ), with WF0L∞ ({Rr }) ⊂ ess supp a.

Now, recall that we are assuming WFs0 (u) ∩ R = ∅; let B̃ ∈ Ψ0 be elliptic at R and such
that B̃u ∈ H s0 . Fix K > 0 such that
m−1
+ β̃ < s − K < s0 , (9.40)
2

and choose the support of our cutoffs so small that ess supp a ⊂ Ell(B̃). Since Ar ∈
Ψ2s−m+1−2K ⊂ Ψ2s0 −m+1 for r > 0, we have Ar u ∈ H −s0 +m−1 . We want to compute


2 ImhP u, Ar ui = i hP u, Ar ui − hAr u, P ui
(9.41)
= i[P, Ar ] + i(P ∗ − P )Ar u, u .




All terms make sense individually using Lemma 6.34 (since WFs0 −m+1 (P u) ∩ WF0 (Ar ) = ∅
and since the operator in the second line lies in Ψ2s0 ). However, the integration by parts
needs to be justified, since for general u ∈ H s0 , one only has P u ∈ H s0 −m , which is
in general insufficient to justify the integration by parts. This is easily accomplished by
inserting yet another regularizer, J ∈ L∞ ((0, 1] ; Ψ0 ), with J = J∗ ∈ Ψ−∞ for  > 0, and
J → I in the topology of Ψη , η > 0, and using Lemma 9.10. Namely,


hP u, Ar ui − hAr u, P ui = lim hP u, J Ar ui − hAr u, J P ui
→0
= lim h(Ar J P − P ∗ J Ar )u, ui
→0
= lim hJ (Ar P − P ∗ Ar )u, ui + h([Ar , J ]P − [P ∗ , J ]Ar )u, ui


→0
= h(Ar P − P ∗ Ar )u, ui.
(9.42)

Note here that [Ar , J ] is uniformly bounded (in  ∈ (0, 1], for r > 0 fixed ) in Ψ2s0 −m
and converges to 0 in Ψ2s0 −m+η , η > 0, hence [Ar , J ]P u → 0 in H −s0 , and therefore
h[Ar , J ]P u, ui → 0 as  → 0; likewise, h[P ∗ , J ]Ar u, ui → 0 as  → 0 for fixed r > 0.
We proceed to rewrite the right hand side of the pairing (9.41) by plugging in (9.39). Let
G ∈ Ψ0 , WF0 (I − G) ∩ WF0L∞ ({Ar }) = ∅. Then the Peter–Paul inequality and Lemma 8.11
give the estimate

kBr uk2L2 + kB1,r uk2L2 + 2δkAr uk2H −s+m−1

≤ 2δkAr uk2H −s+m−1 + δ −1 kGP uk2H s−m+1 (9.43)
+ |hP u, Hr ui| + C kGuk2H s−1/2 + kuk2H −N


104 PETER HINTZ

for an r-independent constant C. The first term in the last line can be estimated (using
WF0L∞ ({Hr }) ∩ WF0 (I − G) = ∅) by
|hP u, Hr ui| ≤ C kGP uk2H s−m+1 + kHr uk2H −s+m−1 + kuk2H −N


(9.44)
≤ C kGP uk2H s−m+1 + kGuk2H s−1 + kuk2H −N ;

recall that Hr ∈ L∞ ((0, 1]; Ψ2s−m ). Combined with (9.43), and an iterative argument
(improving the regularity by 1/2 in each step) as usual, we finally obtain the uniform
estimate


kBr ukL2 ≤ C kGP ukH s−m+1 + kukH −N . (9.45)
(Recall that our proof of this estimate requires that B̃u ∈ H s .) As in §8.4, we thus conclude
that B0 u ∈ L2 , in particular WFs (u) ∩ R = ∅, together with an estimate of kB0 ukL2 by the
right hand side of (9.45). This proves the estimate (9.28a).
• Modifications for part (2). The propagation of microlocal regularity into a radial point
uses the same commutant; now the degree K of regularization is arbitrary. Indeed, in the
calculation (9.36), the first term (which is the main term, elliptic at R) is now positive (and
only gets more positive with more regularization), and thus has the opposite sign of the
second term. One thus now writes
Hp ar + 2p1 ar = 2δρ2s−2m+2 a2r + b2r − b21,r + hr p, (9.46)
where b1,r , hr are as in (9.38), and
q
br := ρ φr (ρ)φ(p̃)ψ(ρ1 ) β0 −2s + m − 1 + 2β̃ + 2fr − 2δβ0−1 φr (ρ)2 φ(p̃)2 ψ(ρ1 )2 .
−s

(9.47)
Upon quantizing this, we get a uniform estimate


kBr ukL2 ≤ C kGP ukH s−m+1 + kB1,r ukL2 + kGukH s−1/2 + kukH −N . (9.48)
Thus, we now have an a priori control term kB1,r ukL2 : it is uniformly bounded if WFs (u)
is disjoint from a punctured neighborhood of R. (Note that WF0L∞ (B1,0 ) is some small
positive distance away from R, hence this appears stronger than merely assuming B1,0 u ∈
L2 ; but from B1,0 u ∈ L2 , one can conclude that WFs (u) is disjoint from a punctured
neighborhood of R using the propagation of regularity, Theorem 8.5.) The study of the
limit r → 0 thus gives B0 u ∈ L2 , hence WFs (u) ∩ R = ∅, and (after an iterative argument
improving the regularity by 1/2 at each step) the uniform estimate


kB0 ukL2 ≤ C kGP ukH s−m+1 + kEukH s + kukH −N (9.49)
for E ∈ Ψ0 with Ell(E) ⊃ WF0L∞ ({B1,r }). 
Remark 9.11. Paralleling Remark 8.12, we point out that Theorems 9.8 and 9.9 apply also
to ps.d.o.s P ∈ Ψmcl (M ; E) acting between sections of vector bundles, provided P has a real
scalar principal symbol. Now, a subprincipal term of P modifies the threshold regularity;
and in fact the mere definition of β̃ in equation (9.27) requires the choice of a fiber inner
product on E. Thus, in applications, one typically needs to choose this fiber inner product
carefully in order to obtain the strongest possible conclusions under the weakest possible
assumptions in these theorems. (Note that this is still a purely symbolic calculation, hence
straightforward, even if occasionally a bit lengthy in practice.)
 MICROLOCAL ANALYSIS 105

10. Asymptotic behavior of linear waves on de Sitter space

We now show, following [Vas13] (see also [Zwo16]) how the tools developed so far can
be used for a description of the precise asymptotic (late time) behavior of solutions of
wave equations on spacetimes of interest in general relativity. Concretely, we shall consider
de Sitter space, or rather a subset of it called the static patch (or static model ) of de Sitter
space (M, g) which is a solution of Einstein’s vacuum equation with cosmological constant
Λ > 0,
Ric(g) + Λg = 0, (10.1)
where Ric denotes the Ricci curvature of g.
We first give a quick introduction to Lorentzian metrics and wave equations in §10.1
before studying the wave equation on static de Sitter space in §§10.2–10.3.

10.1. Lorentzian geometry and wave operators.

Definition 10.1. Let M be an n-dimensional manifold, n ≥ 2. Let g ∈ C ∞ (M ; S 2 T ∗ M ),
so gp = g(p), p ∈ M , is a bilinear form on Tp M depending smoothly on p. Then g is a
Lorentzian metric if gp has signature (1, n − 1) (sometimes written (+, −, . . . , −)) for all p.
We call (M, g) a Lorentzian manifold.

This means that at any p ∈ M , there exists a basis V1 , . . . , Vn of Tp M such that

g(V1 , V1 ) = 1; g(Vj , Vj ) = −1, j = 2, . . . , n; g(Vi , Vj ) = 0, i 6= j. (10.2)

Since a Lorentzian metric g is a non-degenerate bilinear form on Tp M , it induces an

isomorphism Tp M → Tp∗ M via V 7→ gp (V, −). Thus, g induces a signature (1, n − 1)
bilinear form on Tp∗ M , denoted G or g −1 ∈ C ∞ (M ; S 2 T M ) and called the dual metric.

Example 10.2. Let M = Rn = Rt × Rxn−1 . Then the Minkowski metric on M is

n−1
X
g = dt2 − dx2j . (10.3)
j=1

The dual metric is

n−1
X n−1
X
G= ∂t2 − ∂x2j = ∂t ⊗ ∂t − ∂ xj ⊗ ∂ xj (10.4)
j=1 j=1

so for instance G(dt, dt) = 1.

Definition 10.3. Let (M, g) be a Lorentzian manifold.
(1) Let p ∈ M . Then we say that a tangent vector V ∈ Tp M is
• timelike if gp (V, V ) > 0,
• spacelike if gp (V, V ) < 0,
• null or lightlike if gp (V, V ) = 0.
Likewise, one can classify covectors ζ ∈ Tp∗ M as timelike, spacelike, or null, depend-
ing on the sign of Gp (ζ, ζ).
(2) Let S ⊂ M be a smooth hypersurface. Then S is spacelike if for all p ∈ S and
ζ ∈ N ∗ S \ o, the covector ζ is timelike.
106 PETER HINTZ

Physically, massive observers (like myself) travel along timelike curves in M (curves with
timelike tangent vectors), and massless particles (think of photons) travel along lightlike
curves.
Example 10.4. In the notation of Example 10.2, the vector ∂t + v1 ∂x1 is timelike iff |v1 | < 1,
null iff |v1 | = 1, and spacelike iff |v1 | > 1. The hypersurface {t = 0} is spacelike; indeed,
its conormal bundle is spanned by dt, which is timelike. More generally, for v ∈ Rn−1 , the
hypersurface {t = v · x} is spacelike if and only if |v| < 1.

Note that the set of timelike vectors is a solid cone with the vertex removed, thus has two
connected components. A continuous choice of one of them is called a time orientation of
(M, g). This does not exist in general. If it does, there exists a smooth timelike vector field
V on M ; we then say that a timelike/null vector W is future timelike/null if g(V, W ) > 0.
(In particular, V is future timelike.)
Given a Lorentzian manifold (M, g), we define the wave operator g ∈ Diff 2 (M ) by
the same formula as the Laplace operator on a Riemannian manifold: in local coordinates
(z1 , . . . , zn ) on M , we write gij = g(∂zi , ∂zj ), g ij = G(dzi , dzj ), and |g| = | det(gij )|; then
n
X
|g|−1/2 Dzi |g|1/2 g ij Dzj u .


g u := (10.5)
i,j=1

Its principal symbol is the dual metric function

n
X
G(ζ) := σ2 (g )(ζ) = g ij ζi ζj = |ζ|2Gp , ζ ∈ Tp∗ M. (10.6)
i,j=1

Thus, the characteristic set Char(g ) = {ζ : T ∗ M \ o : G(ζ) = 0} consists of all lightlike

covectors.
Remark 10.5. Integral curves of Hp are the lift to T ∗ M of geodesics of (M, g). Recall that
for a geodesic γ : I ⊂ R → M , the squared length gγ(s) (γ 0 (s), γ 0 (s)) is constant; we then call
a geodesic with squared length 0 (i.e. γ 0 (s) is null for all s) a null-geodesic. Correspondingly,
singularities of solutions of the wave equation g u = f propagate along null-geodesics inside
of Char(g ).
Definition 10.6. Let E → M be a vector bundle over the Lorentzian manifold (M, g). We
say that P ∈ Diff 2 (M ; E) is a wave-type operator if P is principally scalar with σ2 (P ) = G,
where G(ζ) = g −1 (ζ, ζ) is the dual metric function.

Typical examples include the scalar wave operator g , or the tensor wave operator
− trg ∇2 , or modifications of such operators by first and zeroth order terms.
We record here an existence and uniqueness statement for the wave equation whose proof
we omit.
Proposition 10.7. Let P ∈ Diff 2 (M ; E) be a wave-type operator on a Lorentzian manifold
(M, g). Suppose t ∈ C ∞ (M ) is a timelike function, i.e. dt is everywhere timelike. Suppose
Ω ⊂ M is a domain with spacelike boundary, and suppose Ω0 := Ω̄ ∩ t−1 ([0, ∞)) is compact.
Then, given any f ∈ C ∞ (Ω0 ; E) such that supp f ⊂ {t ≥ 0}, there exists a unique u ∈
C ∞ (Ω; E) with supp u ⊂ {t ≥ 0} such that P u = f in Ω.
 MICROLOCAL ANALYSIS 107

This for example applies to wave-type operators on Minkowski space (Example 10.2)
for domains Ω = {t < F (x)}, where F ∈ C ∞ (Rn−1 ) satisfies |F 0 (x)| < 1 for all x, and
F (x) → −∞ when |x| → ∞. By a simple approximation argument, one can also take
functions like F (x) = 1 − c|x| when c < 1.

10.2. Waves on the static model of de Sitter space. From now on, we shall work on
a particular 3-dimensional Lorentzian manifold. (This can all be generalized significantly,
of course, but we stick to a concrete setting for simplicity of presentation.)
Definition 10.8. We define 3-dimensional de Sitter space (M, g) by
M := Rt × {x ∈ R2 : |x| < 2},
(10.7)
g := (1 − |x|2 )dt2 + (dt ⊗ (x · dx) + (x · dx) ⊗ dt) − dx2 ,
where we write x · dx = x1 dx1 + x2 dx2 , and dx2 = dx21 + dx22 .

It is easier to work with polar coordinates (r, θ) on R2 , in which

g = (1 − r2 )dt2 + (dt ⊗ r dr + r dr ⊗ dt) − dr2 − r2 dθ2 ,
(10.8)
G = g −1 = ∂t2 + (∂t ⊗ r∂r + r∂r ⊗ ∂t ) − (1 − r2 )∂r2 − r−2 ∂θ2 .
We note a few features of this spacetime:
(1) g is stationary, that is, L∂t g = 0, or more prosaically: the coefficients of g are
t-independent.
(2) dt is timelike (since |dt|2G = 1 > 0), so the level sets t−1 (t0 ), t0 ∈ R, are spacelike;
we declare dt to be future timelike;
(3) for any r0 > 1, the level set r−1 (r0 ) is spacelike (since |dr|2G = −(1 − r02 ) < 0), and
G(dr, dt) = r0 > 0, so dt and dr are both future timelike;
(4) the hypersurface
H+ := r−1 (1) (10.9)
is null (meaning dr is null there). It is called the cosmological horizon.
Thus, the geometry of (M, g) is quite interesting: consider a point p ∈ M with r(p) > 1,
and a future timelike or null vector V ∈ Tp M , ζ := gp (V, −) ∈ Tp∗ M . Then V r = dr(V ) =
Gp (dr, ζ) > 0. Therefore, any physical observer or light particle travels even further away
from r = 1. On the other hand, if r(p) < 1, there are no such restrictions.
An application of Proposition 10.7 (with Ω a smoothed out version of {(t, x) ∈ M : t ≤
T, |x| < R} for R ∈ (1, 2) and any T > 0) implies that the wave equation g u = f ∈ C ∞ (M )
with supp f ⊂ t−1 ([0, ∞)) has a unique solution u ∈ C ∞ (M ) with supp u ⊂ t−1 ([0, ∞)).
(This is true more generally for wave-type operators on (M, g).) Our aim is to describe the
asymptotic behavior of u(t, x) as t → ∞.
We denote the spatial slices of M by
X = {x ∈ R2 : |x| < 2}, (10.10)
which can be identified with t−1 (t0 ) for any t0 ∈ R.
Theorem 10.9. Let P ∈ Diff 2 (M ) be a wave-type operator on (M, g) with t-independent
coefficients, that is, [∂t , P ] ≡ 0. Then there exists a sequence of numbers (called resonances)
σj ∈ C with Im σj → −∞, and finite-dimensional spaces (of (generalized) resonant states
108 PETER HINTZ

Pkj −1
Rj ⊂ C ∞ (M ) ∩ ker P consisting of functions of the form k=0 e−iσj t tk ak (x), ak ∈ C ∞ (X),
such that the following holds:
Let f ∈ Cc∞ (M ), supp f ⊂ t−1 ([0, ∞)), and let u ∈ C ∞ (M ) denote the unique solution of
P u = f, supp u ⊂ t−1 ([0, ∞)). (10.11)
Let α ∈ R be such that − Im σj =
6 α for all j. Then there exist uj ∈ Rj and a constant
C > 0 such that for t ≥ 0
X
u(t, x) = uj (t, x) + ũ(t, x), |ũ(t, x)| ≤ Ce−αt . (10.12)
Im σj >−α

That is, modulo an error decaying at any fixed exponential rate α, u(t, x) is equal to a finite
sum of terms of the form e−iσj t tk ajk (x) with ajk ∈ C ∞ (X).

Note that |e−iσj t | = e(Im σj )t , which decays when Im σj < 0.

Remark 10.10. Compare this with the description of linear waves on a compact Riemannian
manifold (X, h), i.e. solutions u of (Dt2 − ∆h )u =P f ∈ Cc∞ (Rt × X): they can be expanded in
eigenfunctions φj , ∆h φj = λ2j φj , λj ∈ R, so u = ∞ j=0 (aj+ e
iλj t +a e−iλj t ). All frequencies
j−
here are real, so this is a sumR of oscillating, but non-decaying terms. A strong manifestation
of the lack of decay is that X |∂t u(t, x)|2 + |∂x u(t, x)|2 dx is conserved (t-independent).
Morally speaking, the reason for the decay (modulo finitely many terms) of u in Theo-
rem 10.9 is that waves can cross H+ , and once they have done so, they continue travelling
outwards and leave our (incomplete) spacetime M . The numbers σj in Theorem 10.9 are
the replacement for eigenvalues in the system of interest here were energy can ‘leak’ out,
and the spaces Rj of generalized resonant states are the replacements of eigenspaces. (In
particular, just like eigenvalues, the resonances σj cannot be computed explicitly except in
very special situations.)
Remark 10.11. For the wave operator P = g , the resonances are σj = −ij, j ∈ N0 , and
kj = 2 for j ≥ 2. For the Klein–Gordon operator P = g − m2 ,
kj = 1 for j = 0, 1, while √
the resonances are −i ± i 1 − m2 − iN0 . (See [Vas10] and [HV18, Appendix C], the latter
also including a calculation for a wave-type operator acting on symmetric 2-tensors, though
in 3 + 1 dimensions.)
Remark 10.12. If E → X is a vector bundle which, via π : M 3 (t, x) 7→ x ∈ X, lifts to
a ‘stationary’ vector bundle E := π ∗ E → M , then Theorem 10.9 remains valid for wave-
type operators P ∈ Diff 2 (M ; E) with t-independent coefficients. (This is a well-defined
notion since sections of E can be invariantly differentiated with respect to t.) In this case,
the resonant states are elements of C ∞ (X; E). Examples include the wave operator on
differential forms or symmetric 2-tensors (or other tensor bundles).

For the most part, we shall only sketch the proof of Theorem 10.9; we provide details for
the most interesting (and conceptually central) part of the argument. To begin with, it is
not hard to show an exponential bound for u: there exists C0 > 0 such that
|u(t, x)| ≤ C0 eC0 t , t ≥ 0. (10.13)
(This follows from the stationarity and linearity of P by a simple energy estimate. Morally,
we can see this as follows: we have P u = 0 for t ≥ t0  1 since f has compact support;
the estimate (10.13) then follows from an estimate of the energy E(t) := ku(t, x)kH 2 (X) +
 MICROLOCAL ANALYSIS 109

k∂t u(t, x)kH 1 (X) of the form E(t + 1) ≤ CE(t) for a constant C which, by stationarity, can
be taken to be t-independent.)
The strategy of the proof is to use spectral theory after taking the Fourier transform in t
(with a sign change relative to our previous convention, for consistency with the literature):
letting Z
û(σ, x) := eiσt u(t, x) dt, (10.14)
R

and likewise fˆ(σ, x), (formally) taking the Fourier transform of (10.11) gives
P̂ (σ)û(σ) = fˆ(σ), (10.15)
where the operator P̂ (σ) ∈ Diff 2 (X) is obtained from P = P (x, Dt , Dx ) by replacing Dt by
−σ, so
P̂ (σ) = P (x, −σ, Dx ). (10.16)
Since the leading order part of the wave-type operator P is Dt2 +2Dt rDr −(1−r2 )Dr2 −r−2 Dθ ,
the leading order part of −P̂ (σ) is (1 − r2 )Dr2 + r−2 Dθ2 ; near r = 0, this is close to the
Laplacian on R2 , and indeed it is elliptic for r < 1, but at r = 1 it degenerates, and it
becomes a hyperbolic operator in r > 1 (with r taking the role of a ‘time function’ there).
Now, since u(t, x) = 0 for t ≤ 0, the bound (10.13) implies that û(σ, x) is well-defined for
Im σ > C0 ; moreover, it implies that all resonances σj satisfy Im σj ≤ C0 . For fˆ(σ, x), the
situation is even better: since f has compact support in t, fˆ(σ, x) ∈ C ∞ (X) is holomorphic
in the full complex plane σ ∈ C.
Thus, the equation (10.15) holds true for Im σ > C0 . Suppose now we can invert P̂ (σ)
(on C ∞ (X), or suitable Sobolev spaces) for such σ; then
û(σ, x) = P̂ (σ)−1 fˆ(σ, x), (10.17)
and we therefore have
Z
−1
u(t, x) = (2π) e−iσt P̂ (σ)−1 fˆ(σ, x) dσ. (10.18)
Im σ=C0 +1

We shall prove is that P̂ (σ)−1 is a meromorphic family of operators on C ∞ (X); the

connection to Theorem 10.9 will then be:
(1) the resonances σj are then the poles of P̂ (σ)−1 ;
(2) the integers kj + 1 are the order of the pole at σ = σj .
Indeed, in the expression (10.18), we use Cauchy’s theorem to shift the integration con-
tour from Im σ = C0 + 1 to Im σ = −α, giving
X
(2π)−1 Resσ=σj e−iσt P̂ (σ)−1 fˆ(σ, x)


u(t, x) =
Im σj >−α
Z (10.19)
−1 −iσt −1
+ (2π) e P̂ (σ) fˆ(σ, x) dσ.
Im σ=−α

In the case kj = 0 and P̂ (σ) = (σ − σj )−1 P1 + holomorphic, we have

Resσ=σj e−iσt P̂ (σ)−1 fˆ(σ, x) = e−iσj t P1 (fˆ(σj )). (10.20)
110 PETER HINTZ

Thus, in this case, Rj = {e−iσj t a(x) : a(x) ∈ ran P1 }; the case of higher order poles is
similar. The second term in (10.19) is the remainder ũ(t, x) in the notation of Theorem 10.9;
note that the integrand is pointwise bounded by e−αt .
Remark 10.13. Justifying (10.19) uses that P̂ (σ)−1 fˆ(σ, x) has suitable decay as | Re σ| → ∞
with Im σ ∈ [−α, C0 + 1]. Operator norm bounds on P̂ (σ)−1 for such σ are called high
energy estimates, which can be proved by methods from semiclassical microlocal analysis.
Moreover, such estimates imply that there are only finitely many resonances in any strip
| Im σ| < C, C ∈ R.
Remark 10.14. The arguments sketched here can be used to describe in a similar manner
linear and even nonlinear waves on black hole spacetimes such as Schwarzschild–de Sitter
and Kerr–de Sitter black holes. For the high energy estimates (briefly mentioned below),
one needs an additional ingredient to deal with trapping effects. See for instance [Vas13,
WZ11, Dya11, HV18], and references therein.

The only statement we shall prove here in detail is that P̂ (σ)−1 is meromorphic (on
suitable function spaces).

10.3. Analysis of the spectral family P̂ (σ). We begin by defining the relevant function
spaces:
Definition 10.15. Suppose M is a manifold, and X ⊂ M is open. Let s ∈ R, and let
F(M ) denote a space of distributions on M , such as F(M ) = D 0 (M ) or F(M ) = Hloc s (M ).

We then define the space

F̄(X) := {u|X : u ∈ F(M )} (10.21)
of restrictions to X; its elements are called extendible distributions. We also define
Ḟ(X̄) := {u : u ∈ F(M ), supp u ⊂ X̄}. (10.22)
Its elements are called supported distributions.

Note that the kernel of F(M ) 3 u 7→ u|X ∈ F̄(X) is Ḟ(M \ X); hence we have
F̄(X) ∼= F(M )/Ḟ(M \ X). (10.23)

Recall from (10.10) that P̂ (σ) is an operator on the spatial slice X = {|x| < 2} ⊂ R2 .
Taking F = H s (R2 ) gives the function spaces
H̄ s (X), Ḣ s (X̄); (10.24)
we have P̂ (σ) : H̄ s (X) → H̄ s−2 (X) and Ḣ s (X̄) → Ḣ s−2 (X̄). Note that Ḣ s (X̄) is a closed
subspace of H s (R2 ). In view of (10.23), the space H̄ s (X) also carries the structure of a
Hilbert space. Note moreover that C¯∞ (X) = C ∞ (X̄) ⊂ H̄ s (X) and C˙∞ (X̄) ⊂ Ḣ s (X̄) are
dense.
Lemma 10.16. The L2 pairing C¯∞ (X) × C˙∞ (X) 3 (u, v) 7→ hu, vi = uv̄ dx ∈ C extends
R
by continuity to a pairing
H̄ s (X) × Ḣ −s (X̄) → C. (10.25)
It has the property that H̄ s (X) 3 u 7→ hu, −i ∈ (Ḣ −s (X̄))∗ is an isomorphism.

One says that Ḣ −s (X̄) is the dual space of H̄ s (X) relative to L2 (X).
 MICROLOCAL ANALYSIS 111

Proof of Lemma 10.16. We have hu, vi = 0 for u ∈ C˙c∞ (R2 \ X) and v ∈ C˙∞ (X̄), hence this
holds also for u ∈ Ḣ s (R2 \ X) and v ∈ Ḣ −s (X̄). By (10.23), the pairing (10.25) is therefore
well-defined.
For the final claim, note that if u ∈ H̄ s (X) is such that hu, vi = 0 for all v ∈ Ḣ −s (X̄),
write u = ũ|X , ũ ∈ H s (R2 ) and conclude that supp ũ ⊂ R2 \X, therefore u = 0. Conversely,
given ` ∈ (Ḣ −s (X̄))∗ , use Hahn–Banach to extend ` to a continuous linear functional
`˜ ∈ H −s (R2 ); then `(v)
˜ = hũ, vi for some ũ ∈ H s (R2 ), and setting u := ũ|X completes the
proof. 

Given P ∈ Diff 2 (M ) as in Theorem 10.9, there exists a constant β̃ ∈ R (explicitly

computable and given in the course of the proof) such that the following holds:
Theorem 10.17. For s ∈ R, define the function space
X s := {u ∈ H̄ s (X) : P̂ (0)u ∈ H̄ s−1 (X)}. (10.26)
1
Let α ∈ R. Then, for s > 2 + β̃ + α,
P̂ (σ) : X s → H̄ s−1 (X), σ ∈ C, Im σ > −α, (10.27)
is a Fredholm operator. Moreover,
ker P̂ (σ) ∩ X s ⊂ C¯∞ (X), (10.28)
and ranX s P̂ (σ) ⊂ H̄ s−1 (X) is the annihilator of
ker P̂ (σ)∗ ∩ Ḣ −s+1 (X) ⊂ Ḣ 1/2−β̃+Im σ− ∀  > 0. (10.29)

Note that P̂ (σ) − P̂ (0) ∈ Diff 1 (X), hence P̂ (σ) indeed maps X s → H̄ s−1 (X). In the
final statement, P̂ (σ)∗ is the formal adjoint defined by hP̂ (σ)∗ u, vi = hu, P̂ (σ)vi for u, v ∈
Cc∞ (X); it is easy to see that
P̂ (σ)∗ = P
c∗ (σ̄). (10.30)
We prove this theorem below; first, we explain why it is so useful.
Lemma 10.18. For Im σ  1, P̂ (σ) : X s → H̄ s−1 (X) is invertible.

Proof (sketch). An element u ∈ ker P̂ (σ)∩C¯∞ (X) gives rise to a solution U (t, x) = e−iσt u(x)
of P U = 0. In view of the estimate (10.13), we must have U ≡ 0 when Im σ ≥ C0 , hence
u ≡ 0. Therefore, P̂ (σ) is injective for large Im σ.
Dually, if v ∈ ker P̂ (σ)∗ , then P ∗ V = 0 for V (t, x) = eiσt v(x). Since v = 0 for r > 1
(which follows from the fact that v, extended by 0 beyond X, solves the hyperbolic equation
P̂ (σ)∗ v = 0 in r > 1), we have V = 0 for r > 1 as well. Moreover, for Im σ  1, v lies in
H 1 . One can then again use an energy estimate (for P ∗ and ‘from t = ∞’) to show that
there exists C1 ∈ R such that V ≡ 0 when Im σ > C1 , hence v ≡ 0. By Theorem 10.9, this
implies that P̂ (σ) is surjective. 
Corollary 10.19. For α ∈ R, s > 21 + β̃ + α, Im σ > −α as in Theorem 10.9, the family
P̂ (σ) : X s → H̄ s−1 (X) is a family of Fredholm operators of index 0. Its inverse extends
from Im σ  1 to a finite-meromorphic family
P̂ (σ)−1 : H̄ s−1 (X) → H̄ s (X). (10.31)
112 PETER HINTZ

The first part is clear since the index of a continuous family of Fredholm operators is
constant. For the second part, we use the following terminology:
Definition 10.20. Let X, Y denote two Banach spaces. Let Ω ⊂ C be an open set. Then
we say that B(σ) : X → Y , σ ∈ Ω, is finite-meromorphic if there exists a discrete subset
D = {σ1 , σ2 , . . .} ⊂ Ω such that:
(1) B(σ) is holomorphic on Ω \ D;
(2) near σj , there exists kj ∈ N such that
kj
X
B(σ) = (σ − σj )−k Bjk + B̃j (σ), (10.32)
k=1

where B̃j (σ) : X → Y is holomorphic near σ = σj , and Bjk : X → Y , 1 ≤ k ≤ kj , is

a finite rank operator.

Corollary 10.19 is then an immediate consequence of:

Proposition 10.21. (Analytic Fredholm Theorem.) Let X, Y be Banach spaces, let Ω ⊂
C be open and connected, and suppose A(σ) : X → Y , σ ∈ Ω, is an analytic family of
Fredholm operators. Then either A(σ) is not invertible for any σ ∈ Ω, or A(σ)−1 is finite-
meromorphic.

Proof. Suppose the (open) set Ω0 ⊂ Ω of σ for which A(σ) is invertible is non-empty; then
A(σ) has index 0 for σ ∈ Ω0 , hence for all σ ∈ Ω.
If Ω0 6= Ω, let σ0 ∈ Ω ∩ ∂Ω0 . Consider A(σ0 ) : X → Y . Let X2 = ker A(σ0 ) and
R1 = ran A(σ0 ); pick closed subspaces X1 ⊂ X and Y2 ⊂ Y with
X = X1 ⊕ X2 , Y = Y1 ⊕ Y2 . (10.33)
Since ind A(σ0 ) = 0, dim X2 = dim Y2 = N < ∞. We write A(σ) as a block matrix in the
decomposition (10.33),
 
P (σ) Q(σ)
A(σ) = , (10.34)
S(σ) T (σ)
where P (σ0 ) : X1 → Y1 is invertible, and Q, S, T = 0 at σ = σ0 . Thus, P (σ) : X1 → Y1 is
invertible for |σ − σ0 | <  for some  > 0; by the Schur complement formula (block-wise
inversion of A(σ)), A(σ) is invertible for |σ − σ0 | <  if and only if
Z(σ) := T (σ) − S(σ)P (σ)−1 Q(σ) : X2 → Y2 (10.35)
is invertible; in this case, we have
 −1
P + P −1 QZ −1 SP −1 −P −1 QZ −1

−1
A(σ) = . (10.36)
−Z −1 SP −1 Z −1
But Z(σ) is a holomorphic N × N matrix near σ0 , and invertible for some σ arbitrarily
close to σ0 . Hence, fixing a basis of X2 and Y2 , its determinant det Z(σ) is a non-zero
holomorphic function which vanishes at σ = σ0 ; hence det Z(σ)−1 is meromorphic, and so
is Z(σ)−1 . Therefore, A(σ) is invertible in a punctured neighborhood of σ0 . The conclusion
is now immediate from (10.36). 
 MICROLOCAL ANALYSIS 113

Returning to the main calculation (10.19) in our sketch of the proof of Theorem 10.9, this
justifies (modulo control for large | Re σ|) the contour shifting and the use of the residue
theorem.
Now, Theorem 10.9 will be an easy consequence of the following result:
Proposition 10.22. We have the following Fredholm estimates for P̂ (σ):
1
(1) Let s > s0 > 2 + β̃ − Im σ. Then there exists C > 0 such that for u ∈ X s ,


kukH̄ s (X) ≤ C kP̂ (σ)ukH̄ s−1 (X) + kukH̄ s0 (X) ; (10.37)
this holds in the strong sense that if all quantities on the right hand side are finite,
then so is the left hand side, and the inequality holds.
(2) Define, analogously to X s , the space
Y −s+1 := {v ∈ Ḣ −s+1 (X̄) : P̂ (σ)∗ v ∈ Ḣ −s (X̄)} (10.38)
1
Let N ∈ R and s > 2 + β̃ − Im σ. Then there exists C > 0 such that for all
v ∈ Y −s+1 ,
kvkḢ −s+1 (X̄) ≤ C kP̂ (σ)∗ vkḢ −s (X̄) + kvkḢ −N (X̄) ;


(10.39)
this holds in the strong sense.

Proof of Theorem 10.9 assuming Proposition 10.22. The estimate (10.37) together with the
compactness of the inclusion H̄ s (X) ,→ H̄ s0 (X) (exercise!) imply that dim kerH̄ s (X) P̂ (σ) <
∞, and that ranX s P̂ (σ) ⊂ H̄ s−1 (X) is closed. Moreover, since (10.37) holds in the strong
sense, it implies that if P̂ (σ)u = 0, then we can take s arbitrary and obtain u ∈ C¯∞ (X).
On the other hand, the estimate (10.39) implies dim K < ∞ where
K := kerḢ −s (X̄) P̂ (σ)∗ < ∞. (10.40)

Again, since (10.39) holds in the strong sense, we see that P̂ (σ)∗ v = 0 implies v ∈
Ḣ 1/2−β̃+Im σ− (X̄) for all  > 0.
Finally, let f ∈ H̄ s−1 (X) be such that hf, vi = 0 for all v ∈ K. We claim that there
exists u ∈ H̄ s (X) such that
P̂ (σ)u = f ∈ H̄ s−1 (X). (10.41)
(This implies that ran P̂ (σ) has finite codimension, thus finishing the proof.) This solvability
follows by a general argument from the (almost) injectivity (10.39) of the adjoint operator.
First of all, fix a closed complementary subspace L ⊂ Ḣ −s (X̄) of K; then a simple
argument by contradiction shows that there exists a constant C 0 such that
kvkḢ −s+1 (X̄) ≤ C 0 kP̂ (σ)∗ vkḢ −s (X̄) , v ∈ L. (10.42)

Therefore, we have |hv, f i| . kP̂ (σ)∗ vkḢ −s (X̄) for v ∈ L. Writing a general element v ∈
Ḣ −s+1 (X̄) as v = v1 + v2 , v1 ∈ L, v2 ∈ K, we have
|hv, f i| = |hv1 , f i| . kP̂ (σ)∗ v1 kḢ −s (X̄) = kP̂ (σ)∗ vkḢ −s (X̄) . (10.43)
Using Hahn–Banach, the (thus well-defined and bounded) functional
Ḣ −s (X̄) 3 P̂ (σ)∗ v 7→ hv, f i, v ∈ Y −s+1 , (10.44)
114 PETER HINTZ

can be extended to an element of (Ḣ −s (X̄))∗ , which is represented by an element u ∈ H̄ s (X̄)

by Lemma 10.16. In particular, for all v ∈ Cc∞ (X),
hv, f i = hP̂ (σ)∗ v, ui = hv, P̂ (σ)ui, (10.45)
which implies P̂ (σ)u = f , as desired. 

The proof of Proposition 10.22 will, of course, be microlocal. Thus, we need to analyze
the characteristic set and null-bicharacteristic flow of P̂ (σ). Recall the form (10.8) of the
dual metric G of de Sitter space; writing covectors on X = {|x| < 2} in polar coordinates
in r = |x| =
6 0 as
ξ dr + η dθ, (10.46)
we therefore have
p(r, θ, ξ, η) = σ2 (P̂ (σ)) = −(1 − r2 )ξ 2 − r−2 η 2 . (10.47)
We denote the characteristic set of P̂ (σ) by
Σ := p−1 (0) ⊂ T ∗ X \ o. (10.48)
Polar coordinates break down at r = 0, one can easily calculate in standard coordi-
nates (x1 , x2 ) on R2 (namely: by computing the form of the dual metric of (10.7)) that
p(x1 , x2 , ξ1 , ξ2 ) = −(1 − x21 )ξ12 − (1 − x22 )ξ22 − 2x1 x2 ξ1 ξ2 , which is clearly elliptic for (x1 , x2 )
near (0, 0).
Lemma 10.23. Σ is a smooth conic submanifold of T ∗ X \ o, and r ≥ 1 on Σ. It has two
connected components,
Σ = Σ+ ∪ Σ− , Σ± = {(r, θ, ξ, η) ∈ Σ : ± ξ > 0}. (10.49)

Proof. Certainly, p = 0 requires r ≥ 1 in view of (10.47). Furthermore, suppose ζ ∈ Σ is a

point at which r ≥ 1, p = 0; we need to show dp 6= 0. If we assume the contrary, dp = 0,
then ∂η p = −2r−2 η = 0 implies η = 0. Then 0 = p = −(1 − r2 )ξ 2 implies ξ = 0 (and thus
we are the zero section, hence outside of Σ) unless r = 1. If r = 1 and ξ 6= 0, however, we
have 0 = ∂r p = 2rξ 2 6= 0, a contradiction.
The final claim follows immediately from (10.47). 

In fact, since p is homogeneous, we have Σ− = −Σ+ . Moreover, the (−Hp )-flow in Σ−

is the mirror image (multiplication by −1 in the fibers of T ∗ X) of the Hp -flow on Σ+ . We
thus only study the properties of the Hp -flow in Σ+ .
Note now that ξ −1 is elliptic and positive near Σ+ ; let us thus work with projective
coordinates
1 η
ρ := , η̂ = . (10.50)
ξ ξ
We can then identify Σ+ with its boundary at fiber infinity inside T ∗ X,
Σ+ = {(r, θ, η̂) : η̂ 2 = r2 (1 − r2 )}. (10.51)
(Forgetting about the θ-variable, this thus looks like a parabola in (r, η̂) with vertex at
r = 1.)
 MICROLOCAL ANALYSIS 115

Next, we compute the Hamiltonian vector field Hp = −2(1 − r2 )ξ∂r − 2r−2 η∂θ − 2(rξ 2 +
r−3 η 2 )∂
ξ and its rescaling

H̃p := ξ −1 Hp = −2(1 − r2 )∂r − 2r−2 η̂∂θ + 2(r + r−3 η̂ 2 )(ρ∂ρ + η̂∂η̂ ), (10.52)
which on Σ+ takes the form
H̃p = −2(1 − r2 )∂r − 2r−2 η̂∂θ + 2r−1 (ρ∂ρ + η̂∂η̂ ). (10.53)
Its only critical points are at r = 1, η̂ = 0. We have thus identified the radial set
R+ := {(r = 1, θ, η̂ = 0)} ⊂ Σ+ ⊂ S ∗ X. (10.54)
Lemma 10.24. Let s 7→ γ(s) ∈ Σ+ be a null-bicharacteristic, i.e. an integral curve of H̃p ,
with γ(0) ∈
/ R+ . Then:
(1) in the backward direction, γ(s) tends to R+ as s → −∞;
(2) in the forward direction, γ(s) crosses r = 2 in finite time (in the direction of in-
creasing r).

Proof. We have r > 1 at γ(0). Note then that H̃p r = 2(r2 − 1) > 0; thus, r ◦ γ(s) is
monotonically increasing in the forward direction, and indeed H̃p r ≥ 2(r(γ(0))2 − 1) for
s ≥ 0. This implies the second statement. On the other hand, as s → −∞, r(γ(s)) → 0; in
view of (10.51), this implies γ(s) → R+ indeed. 

Thus, the only interesting place is R+ .

Lemma 10.25. R+ is a source for the H̃p -flow inside of T ∗ X. For ρ as in (10.50), we
have
β0 := ρ−1 H̃p ρ = 2 at R+ (10.55)
(cf. the definition (9.26)). Moreover, we have
 P̂ (σ) − P̂ (σ)∗ 
β̃(σ) := β0−1 ξ −1 σ1 = β̃ − Im σ (10.56)
2i
at R+ for some (σ-independent) β̃ ∈ C ∞ (R+ ) (cf. the definition (9.27)).

Proof. The calculation of β0 is trivial, and shows that R+ is a source in the fiber-radial
direction. Next, the function ρ1 = η̂ 2 is a quadratic defining function for R+ inside of Σ+ ,
and we have H̃p ρ1 = 4r−1 ρ1 .
For the calculation of β̃(σ), note that by inspection of (10.8), we have
P̂ (σ) = P̂ (0) + σ(−2rDr + R0 ) + σ 2 R1 (10.57)
near r = 1, where R0 , R1 ∈ C ∞ (X) are lower order terms, and P̂ (0) ∈ Diff 2 (X) has real
principal symbol. Thus,
 P̂ (σ) − P̂ (σ)∗   P̂ (0) − P̂ (0)∗ 
σ1 = σ1 − 2(Im σ)rξ. (10.58)
2i 2i
This implies (10.56). 

Equipped with this dynamical information, and the calculation (10.56), we are now in a
position to prove Proposition 10.22.
116 PETER HINTZ

Proof of Proposition 10.22. • Proof of the estimate (10.37). The idea is to piece together
radial point estimates, real principal type propagation estimates, and microlocal elliptic
regularity to control u solving
P̂ (σ)u = f ∈ H̄ s−1 (X). (10.59)
For clarity and simplicity, we shall not use the semiglobal results (such as Theorems 9.9
and 8.7), but rather work step by step.
For 0 < δ  1, let
Xδ = {|x| < 2 − δ} ⊂ X; (10.60)
we assume all Schwartz kernels below to have compact support in Xδ × Xδ . For s > s0 >
1 0
2 + β̃ − Im σ, and for B ∈ Ψ elliptic near R+ , we have
kBukH s . kf kH s−1 + kukH s0 , (10.61)
in the strong sense. (We may replace f by Gf , where G ∈ Ψ0 microlocalizes near WF0 (B).)
By Lemma 10.24, the H s -regularity of u can now be propagated to all of the characteristic
set over Xδ ; thus, for B+ ∈ Ψ0 elliptic near Σ+ ∩ S ∗ Xδ , we have (for any fixed N ∈ R)
kB+ ukH s . kBukH s + kf kH s−1 + kukH −N
(10.62)
. kf kH s−1 + kukH s0 .
The same estimate holds, by the same reasoning, for B− ∈ Ψ0 elliptic near Σ− ∩ S ∗ Xδ .
On the other hand, for B0 ∈ Ψ0 elliptic near S ∗ Xδ \ (Ell(B+ ) ∪ Ell(B− )), microlocal
elliptic regularity gives
kB0 ukH s . kf kH s−2 + kukH −N . (10.63)
∗
But Ell(B0 ) ∪ Ell(B− ) ∪ Ell(B+ ) ⊃ S Xδ . Fix cutoffs
χ ∈ Cc∞ (Xδ ), χ ≡ 1 on X2δ , χ̃ ∈ Cc∞ (X), χ̃ ≡ 1 on Xδ ; (10.64)
we have then proved
kχukH s . kχ̃f kH s−1 + kχ̃ukH s0 . kf kH̄ s−1 (X) + kχ̃ukH s0 . (10.65)

This gives an estimate of kukH̄ s (Xδ ) , but with an error term (the last term on the right)
which is measured on a larger set than u itself, as is typical for any microlocal estimate. To
bridge the gap, we use that in r > 1, P̂ (σ) is a hyperbolic operator (equal to (r2 − 1)Dr2 −
r−2 Dθ2 to leading order, so r becomes the ‘time’ function); note then that if
φ ∈ Cc∞ (X2δ ), φ ≡ 1 on X3δ , (10.66)
then for
ũ := (1 − φ)u, (10.67)
which is supported in r ≥ 2 − 3δ, we have
˜ f˜ := (1 − φ)f − [P̂ (σ), φ]u,
P̂ (σ)ũ = f, (10.68)
and the forcing f˜ ∈ H̄ s−1 (X), with r ≥ 2 − 3δ on supp f˜, satisfies the estimate
kf˜k s−1
H̄ (X) . kf k s−1
H̄ + kχukH s . kf k s−1
(X) H̄ + kχ̃ukH s0
(X) (10.69)
in view of (10.65). We claim that the unique solution ũ (subject to the support condition)
of (10.68) satisfies the estimate
kũkH̄ s (X) . kf˜kH̄ s−1 (X) . (10.70)
 MICROLOCAL ANALYSIS 117

One way to prove this estimate is the following: using (a slight extension of) the uniqueness
and existence theory for hyperbolic equations developed in §7, ũ can be estimated on X in
some space of distributions by the norm of f˜ on X; using that ũ vanishes, hence is smooth,
in r < 2 − 3δ, the propagation of regularity implies that ũ ∈ Hloc s (X). This is almost

what we are after, except for the loss of uniform control right at ∂X (which is a completely
artificial place!); to fix this, one proceeds as follows:
(1) one extends f˜ to an element of H s−1 on a slightly enlarged domain X−δ , and so
that the H̄ s−1 (X−δ )-norm of the extension is bounded by, say, 2 × kf˜kH̄ s−1 (X) ;
(2) one then solves (10.68) on X−δ , getting ũ ∈ Hloc s (X ) by the arguments described
−δ
just now;
(3) finally, one restricts back to X, giving ũ ∈ H̄ s (X) and the estimate (10.70) plus an
extra term kũkH̄ −N (X−δ ) coming from the use of microlocal propagation estimates;
the latter term however is bounded by some (weak) norm of f˜ by the results of §7.
Putting (10.65) together with (10.69), (10.70), and writing u = χu + (1 − χ)u = χu +
(1 − χ)ũ, we find
kukH̄ s (X) . kf kH̄ s−1 (X) + kukH̄ s0 (X) , (10.71)
as desired.
• Proof of the estimate (10.39). We study the equation
P̂ (σ)∗ v = h ∈ Ḣ −s (X̄). (10.72)
The arguments near ∂X are now slightly easier, as we are working with supported dis-
tributions which vanish on R2 \ X. Thus, letting χ, χ̃, φ be as in (10.64) and (10.66), we
have
k(1 − χ)vkḢ −s+1 (X̄) . k(1 − φ)hkḢ −s (X̄) . (10.73)
But this H −s+1 -control of v for 2 − δ < r < 2 can be propagated along Σ ∩ S{r>1}
∗ X. A
simple calculation shows that the threshold regularity at R± for P̂ (σ) is 12 − β̃(σ) (i.e. there
is a sign switch); since −s + 1 < 12 − β̃(σ), we can thus propagate H −s+1 -regularity of v
into R± . We thus control v microlocally near the full characteristic set Σ; away from Σ,
we have microlocal H −s+2 -estimates on v by microlocal elliptic regularity. Altogether, the
microlocal estimates give
kχvkH −s+1 . kχ̃hkH −s + k(1 − χ)vkḢ −s+1 (X̄) + kχ̃vkH −N . (10.74)
(The first term on the right is the forcing term of the equation (10.72), the second term is
the a priori control term needed for real principal type propagation estimates, and the last
the term is the usual weak error term in microlocal estimates.) Combined with (10.73), we
obtain the estimate
kvkḢ −s+1 (X̄) . khkḢ −s (X̄) + kvkḢ −N (X̄) , (10.75)
as desired. The proof is complete. 

We end this section with a general observation which is of critical importance when
studying perturbations of linear operators or nonlinear PDE (the two being closely related):
the microlocal estimates used above (elliptic regularity, real principal type propagation,
radial point estimates) are stable under perturbations. Let us explain this ingredient by
ingredient for a family of operators P (a) ∈ Ψm depending continuously on a parameter
118 PETER HINTZ

a ∈ A (where A is a normed vector space), |a|  1. For example, in the notation above,
the reader may take P (0) = P̂ (σ) for some fixed σ, and P (a) is any perturbation of this.
(1) (Elliptic estimates.) Suppose B, G ∈ Ψ0 are such that WF0 (B) ⊂ Ell(G)∩Ell(P (0)).
Then there exist , C such that for |a| < , we have the uniform estimate


kBukH s ≤ C kGP (a)ukH s−m + kukH −N (10.76)
(for any fixed N ∈ R), cf. (6.44). This follows from the fact that ellipticity is an
open condition, hence the microlocal parametrix construction for P (a) on WF0 (B)
can be performed with uniform control of the ps.d.o. seminorms of all operators
arising in the construction.
(2) (Real principal type propagation.) The flow of the Hamiltonian vector field Hp(a)
of P (a) depends continuously on the parameter a. In particular, if the assumptions
on the microlocalizers B, G, E in Theorem 8.7 hold for the operator P (0), then
they hold for P (a) as well when a is small, for the same microlocalizers. We claim
that the estimate (8.11) (with s, N fixed) holds uniformly for small a. The robust
way to prove this (which does not involve straightening out Hp(a) in a manner
that is continuous in a) is to take the commutant used in the positive commutator
argument for the operator P (0), and run the argument with the same commutant:
this works since positivity is an open condition, hence any square roots we took,
and any symbols which were elliptic in the arguments for P (0), will remain elliptic
for P (a) as well.
(3) (Radial point estimates.) Even if P (0) has a radial set satisfying the hypotheses
in §9.3, this is general not true anymore for P (a). However, fixing microlocalizers
as in any of the two parts of Theorem 9.9 when applied to P (0), the quantitative
estimates (9.28a), (9.28b), (9.29) (with s, s0 , N fixed) continue to hold for P (a) when
a is sufficiently small, with uniform constants C. This is again due to the stability
of the positive commutator arguments under perturbations: the same commutant
that was used for P (0) can be used for P (a) as well.

References
[DH72] Johannes J. Duistermaat and Lars Hörmander. Fourier integral operators. II. Acta Mathematica,
128(1):183–269, 1972.
[Dya11] Semyon Dyatlov. Exponential energy decay for Kerr–de Sitter black holes beyond event horizons.
Mathematical Research Letters, 18(5):1023–1035, 2011.
[GS94] Alain Grigis and Johannes Sjöstrand. Microlocal analysis for differential operators: an introduction,
volume 196. Cambridge University Press, 1994.
[Hör71] Lars Hörmander. Fourier integral operators. I. Acta mathematica, 127(1):79–183, 1971.
[Hör03] Lars Hörmander. The analysis of linear partial differential operators. I. Classics in Mathematics.
Springer-Verlag, Berlin, 2003.
[Hör09] Lars Hörmander. The analysis of linear partial differential operators. IV. Classics in Mathematics.
Springer-Verlag, Berlin, 2009.
[HV18] Peter Hintz and András Vasy. The global non-linear stability of the Kerr–de Sitter family of black
holes. Acta mathematica, 220:1–206, 2018.
[Mel07] Richard B. Melrose. Introduction to microlocal analysis. Lecture notes from courses taught at MIT,
2007.
[RS72] Michael Reed and Barry Simon. Methods of modern mathematical physics. I. Functional analysis.
Academic Press, New York-London, 1972.
[Tay11] Michael E. Taylor. Partial differential equations II. Qualitative studies of linear equations, volume
116 of Applied Mathematical Sciences. Springer, New York, second edition, 2011.
 MICROLOCAL ANALYSIS 119

[Vas10] András Vasy. The wave equation on asymptotically de Sitter-like spaces. Advances in Mathematics,
223(1):49–97, 2010.
[Vas13] András Vasy. Microlocal analysis of asymptotically hyperbolic and Kerr–de Sitter spaces (with an
appendix by Semyon Dyatlov). Invent. Math., 194(2):381–513, 2013.
[Vas18] András Vasy. A minicourse on microlocal analysis for wave propagation. In Thierry Daudé, Dietrich
Häfner, and Jean-Philippe Nicolas, editors, Asymptotic Analysis in General Relativity, volume 443
of London Mathematical Society Lecture Note Series, pages 219–373. Cambridge University Press,
2018.
[Wun13] Jared Wunsch. Microlocal analysis and evolution equations: lecture notes from 2008 CMI/ETH
summer school. In Evolution equations, volume 17 of Clay Math. Proc., pages 1–72. Amer. Math.
Soc., Providence, RI, 2013.
[WZ11] Jared Wunsch and Maciej Zworski. Resolvent estimates for normally hyperbolic trapped sets.
Annales Henri Poincaré, 12(7):1349–1385, 2011.
[Zwo16] Maciej Zworski. Resonances for asymptotically hyperbolic manifolds: Vasy’s method revisited. J.
Spectr. Theory, 2016(6):1087–1114, 2016.

Department of Mathematics, ETH Zürich, Rämistrasse 101, 8092 Zürich, Switzerland

Email address: [email protected]