INTRODUCTION TO PSEUDO-DIFFERENTIAL

OPERATORS
arXiv:math/9906155v1 [[Link]] 23 Jun 1999

M. S. JOSHI

1. Preface

These notes cover most of a Part III course on pseudo-differential

operators. They assume the reader is familiar with distributions partic-
ularly the Schwartz kernel theorem - the book by Friedlander provides
an excellent introduction to this topic.
The point of view taken is somewhere between that of Shubin, Mel-
rose’s unpublished notes and that of Chazarain and Pirou. All of which
provide good places to continue.
We assume familiarity with vector bundles and with the theory of
Fredholm operators between Banach spaces. For vector bundles one
source is Atiyah, [2], and for Fredholm operators the self-contained
section 19.1 of Hormander Volume 3 is a good account.

2. Introduction

Our purpose in this course is to develop a theory of variable coeffi-

cient linear, partial differential operators on manifolds. We will then
develop applications of this theory. This will allow us to “solve” ellip-
tic operators and in particular the Laplacian acting on k− forms will
allow us to prove the Hodge theorem which states that the de Rham
cohomology of a compact manifold is just the kernel of the Laplacian
which relates the Riemannian geometry of a manifold to its topology
- it will also give us an immediate proof that the de Rham cohomol-
ogy of a compact manifold is finite dimensional and an easy proof of
Poincare duality. A fundamental part of this will be the proof of el-
liptic regularity - any distributional solution of an elliptic equation is
a smooth function - which will be an easy consequence of our calcu-
lus. We will also apply the calculus to the study of the propagation of
singularities for non-elliptic equations - in particular we will show that
associated to any linear PDE there is a Hamiltonian function on phase
space that governs the propagation of singularities of solutions. This
1
2 M. S. JOSHI

theorem gives precise information about where future singularities are,

given the knowledge of where they are at some time - this presages the
more advanced theories of PDEs which use symplectic geometry more
and more which we will touch on, but not do in this course.
We will first study operators on Rn and then later shift to manifolds.
Definition 2.1. A linear partial differential operator of order m on
Rn is a map from C ∞ (Rn ) → C ∞ (Rn ) of the form
X
P u(x) = fα (x)Dxα u(x)
|α|≤m
∞ n
where fα ∈ C (R ).

Note !|α|
1 ∂ α1 ∂ αn
Dxα = ...
i ∂x1 ∂xn
and
|α| = α1 + · · · + αn .
We put in 1/i as its makes dealing with Fourier transforms easier.
αˆ
The fact that Ddα
x f = ξ f(ξ) means that for u of compact support,
we have
!n Z
1
P u(x) = eix.ξ p(x, ξ)û(ξ)dξ, (1)
2π
where
Definition 2.2. The total symbol of P is
X
p(x, ξ) = fα (x)ξ α = e−ix.ξ P (eix.ξ )
|α|≤m

and the principal symbol of P, denoted σm (P ), is

X
pm (x, ξ) = fα (x)ξ α .
|α|=m

So any differential operator can be thought of as a function on Rnx ×

Rnξ - this can be thought of as phase space or the cotangent bundle.
To work microlocally means to work on phase space rather than on
Rn and this will lead to a more coherent theory. The basic idea in
micro-local analysis is that a differential operator is approximated by a
multiplication operator on functions on phase space. To make sense of
this idea we will introduce the algebra of pseudo-differential operators.
Our starting point is (1) - p(x, ξ) is a polynomial in ξ and smooth in x
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 3

so we can think of it a finite sum of homogeneous functions pm−j (x, ξ)

where
pm−j (x, λξ) = λm−j pm−j (x, ξ)
with j running from 0 to m. We therefore define a (classical) pseudo-
differential operator, of order m, to be an operator of the form (1)
where p(x, ξ) is an infinite sum (in a sense, we will later make precise)
of terms pm−j (x, ξ) which are smooth in ξ 6= 0 and
pm−j (x, λξ) = λm−j pm−j (x, ξ) λ > 0.
We do not require smoothness up to 0 because we would be left only
with polynomials! (see if you can prove this) As before p(x, ξ) is the
total symbol and pm the principal symbol - often denoted σm (P ).
This class is then denoted ΨDOclm(Rn ). If we denote the space of
homogeneous functions of order m by H m we have that
ΨDOclm
σm : → Hm
ΨDOclm−1
is an isomorphism.
We will show that these classes form a graded algebra under compo-
sition i.e. that
′ ′
◦ : ΨDOclm(Rn ) × ΨDOclm (Rn ) → ΨDOclm+m (Rn ).

The relationship of the total symbol of the product to the symbols

of the original operators is complicated but at the principal level one
just takes a product:
σm+m′ (P Q) = σm (P )σm′ (Q).

The identity operator, I, is a pseudo-differential operator with total

symbol 1 and principal symbol 1.
So if we want to solve
P Q = I,
we need
1 = σ0 (P Q) = σm (P )σ−m (Q).
So provided σm (P ) is never zero we take Q0 to be of order −m with
principal symbol σm (P )−1 . We then have that
σ0 (P Q0 ) = σ0 (I)
and thus that
P Q0 − I = R1 ∈ ΨDOcl−1.
4 M. S. JOSHI

We now want to solve away the error term - ie we want to find Q1 such
that P Q1 = −R1 , putting σ−m−1 (Q1 ) = −σm (P )−1 σ−1 (R1 ) we get
P (Q0 + Q1 ) − I = R2 ∈ ΨDOcl−2 .
We can now repeat this argument to get Qj ∈ ΨDOcl−m−j such that
P (Q0 + Q1 + · · · + QN −1 ) − I ∈ ΨDOcl−m−N .
If we then sum (in an appropriate sense) to get Q, we have
\
P Q − I ∈ ΨDO −∞ = ΨDO l .
l
Such a Q is called a parametrix and is an inverse at the level of sin-
gularities - on a compact manifold the error will be compact and this
will imply that P has finite dimensional kernel and cokernel - that is
the image is of finite codimension.
The above argument can be applied even if σm (P ) has some zeros
by working away from the zeros - so one can construct an inverse away
from the zero set - so the hard part of the operator is the set where
σm (P ) = 0. This is called the characteristic variety, char(P ) and it
governs much of the operators behaviour. We use these ideas to define
a notion of direction and location for a singularity of a distribution u
- the wavefront set:
!
\
WF(u) = char(P )
P u∈C ∞

- this is a subset of Rnx × Rnξ − {0}. Note that if φ is a smooth function

such that φu is smooth then we have immediately that (x, ξ) 6∈ WF(u)
for x with φ(x) 6= 0. We later show that
singsupp(u) = {x : ∃ξ, (x, ξ) ∈ WF(u)}
so the wavefront set contains all the information that the singular sup-
port does. The theorem of Hörmander on the propagation of singulari-
ties states that if P u ∈ C ∞ then WF(u) is a union of complete integral
curves of the Hamiltonian associated to the energy function pm . One
interpretation of this is that the operator P is a quantised version of
the classical system pm . When P is the wave operator this means that
the singularities of solutions travel along geodesics in the manifold.
So our first task in this course will be to construct the class of pseudo-
differential operators described above and we will then develop the
applications.
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 5

3. Oscillatory integrals

Our objective is to make sense of integrals of the form

Z
eiφ(x,θ) a(x, θ)dθ
which are not absolutely convergent. We can think of these as gener-
alisations of the Fourier transform. We will need restrictions on φ, the
phase, and a, the amplitude, for the integral to make sense.
Definition 3.1. φ ∈ C ∞ (Rnx × (Rkθ − {0})) is a phase function, if φ
is homogeneous of degree one in θ, is real-valued and it has no critical
points, i.e. !
′ ∂φ ∂φ ∂φ ∂φ
dx,θ φ = ,..., , ,...,
∂x1 ∂xn ∂θ1 ∂θk
is never zero in all coordinates simultaneously.

Definition 3.2. We shall say a is a symbol of order m if a ∈ C ∞ (Rn ×

Rk ) and
|Dxα Dθβ a(x, θ)| ≤ Cα,β,K < θ >m−|β| ,
for x in K compact, all θ where
1
< θ >= (1 + |θ|2 ) 2 .
The class is denoted S m (Rn ; Rk ) or just S m .

Note that in particular if a(x, θ) is a polynomial in θ then it is a

symbol.
Note also that the estimates are really only statements about what
happens at infinity. In particular, if χ(x, θ) is compactly supported in
θ then it is an element of
\
S −∞ = S m.
m

Example 3.3. If a(x, θ) is homogeneous of degree m in θ then if we

smooth off near 0, we obtain a symbol of order m.

The following proposition is trivial but important.

Proposition 3.4. Pointwise multiplication defines a map
′ ′
S m × S m → S m+m
and
Dxα Dθβ : S m → S m−|β| .
6 M. S. JOSHI

So now lets return to our integral. For simplicity, we will assume that
a is zero in a neighbourhood of θ = 0 but this is not an important issue
as we shall see that important effects come from infinity. If a ∈ S m
and m < −k then by the dominated convergence theorem we have
that the integral converges to a function which is continuous in x. If
we differentiate the integrand with respect to x, we obtain an integral
of the same form except that the amplitude is in S m+1 . Hence we
deduce that if m + l < −k then the integral yields a function in C l . So
if a ∈ S −∞ then the integral is smooth.
What do we do in general? We regularize the integral to obtain a
distribution.
We want to make sense of the expression:
ZZ
< u, ψ >= eiφ(x,θ) a(x, θ)ψ(x)dθdx.
for any ψ ∈ C0∞ (Rn ). As this is important we shall do so in two ways.
First consider the operator, L defined by
n
X Xk
∂φ ∂ ∂φ 2 ∂
L= + |θ| .
j=1 ∂xj ∂xj l=1 ∂θl ∂θl

L has the property that

 !2 !2 
n
X k
X
∂φ ∂φ
Leiφ =  + |θ|2  eiφ .
j=1 ∂xj l=1
∂θl

Note that our hypothesis guarantees that the coefficient will not
vanish off θ = 0 and we have also that is homogeneous of degree 2 from
the homogeneity of φ. Call the coefficient χ then putting M = χ−1 L.
We have, at least formally,
ZZ ZZ
eiφ(x,θ) a(x, θ)ψ(x)dθdx = M r eiφ(x,θ) a(x, θ)ψ(x)dθdx.

Formally, we can integrate by parts to obtain

ZZ
eiφ(x,θ) (M t )r (a(x, θ)ψ(x))dθdx (2)

and
n
X k
X
∂ ∂
Mt = bj + cl +d
j=1 ∂xj l=1 ∂θl
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 7

with d, bj ∈ S −1 and cl ∈ S 0 . (They may be singular at zero but our

integrand is supported away from zero so this is irrelevant. )
The advantage of all this is that
X
(M t )r (a(x, θ)ψ(x)) = aα (x, θ)Dxα ψ(x)
|α|≤r

with aα ∈ S m−r and so if m − r < −k , the integral (2) will converge

absolutely. So we take (2) to be the definition of
Z
eiφ(x,θ) a(x, θ)dθ

and it is clear that we have a distribution of order r. Note that our

formal computation is valid if m − r < −k and so the answer will be
independent of the choice of r.
Since this is so important to everything we do and it may appear
a little artificial, let’s consider another approach as well. Let χ be a
smooth bump function i.e.

1 for |θ| < 1
χ(θ) = 
0 for |θ| > 2.

Then we define
ZZ
< u, ψ >= lim eiφ(x,θ) χ(ǫθ)a(x, θ)ψ(x)dθdx.
ǫ→0

The integral is compactly supported for any fixed ǫ > 0 so there

is no problems with it. We want to see what happens as ǫ goes to
zero. If a ∈ S m and m < −k then the integral is uniformly absolutely
convergent and so will converge to the convergent integral
ZZ
eiφ(x,θ) a(x, θ)ψ(x)dθdx.

So as before the issue is what happens for m ≥ −k. Letting M be as

above, we have
ZZ
< u, ψ >= lim eiφ(x,θ) (M t )k (χ(ǫθ)a(x, θ)ψ(x)) dθdx.
ǫ→0

Without loss of generality we can assume ǫ is within a compact set

so we have that
|Dθα χ(ǫθ)| ≤ Cα < θ >−|α|
8 M. S. JOSHI

independently of ǫ. The important point is that this estimate is uniform

in ǫ all the way to ǫ = 0 - for fixed ǫ 6= 0 the function χ(ǫθ) is, in fact,
in S −∞ but not uniformly. Thus as before we can deduce
X
(M t )r (χ(ǫθ)a(x, θ)ψ(x)) = aα (ǫ, x, θ)∂α ψ
|α|≤k

with aα ∈ S m−r (Rǫ × Rnx ; Rk ) i.e. the behaviour of aα at ∞ is uniform

in ǫ. And so taking r such that m − r < −k we see that the integral is
uniformly convergent in ǫ and must converge to
ZZ
eiφ(x,θ) (M t )r a(x, θ)ψ(x)dθdx,

as the derivatives falling on χ will disappear when ǫ → 0. So the two

approaches yield the same answer! Note for the functional analysts,
one can view this as the extension of a Frechet-space-valued function
from a dense subset.
In general, we will deal fairly freely with oscillatory integrals but it
is important to realize that when we see such an expression that this
is what we mean.
As oscillatory integrals yield distributions rather than smooth func-
tions, one thing we want to do is find their singularities.
Proposition 3.5. If
Z
u= eiφ(x,θ) a(x, θ)dθ

then
n ′
o
singsupp(u) ⊂ x : ∃θ (x, θ) ∈ supp(a) and dθ φ(x, θ) = 0

Heuristically, this says that if the phase is oscillating near x then the
behaviour for large |θ| cancels and so there is no singularity. This is
sometimes called stationary phase - or the WKB method.

Proof. Suppose
′
dθ φ(x0 , θ) 6= 0, ∀θ
then if we multiply by a bump function, ψ, supported sufficiently close
to x0 we have
Z
ψu = eiφ(x,θ) b(x, θ)dθ
′
with b a symbol such that dθ φ is non-zero on the support of b.
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 9

This means, similarly to above, we can put

′ X ∂φ ∂
M = |dθ φ|−2
j ∂θj ∂θj
and M will be non-singular on the support of b and will, as before,
have the property
Meiφ = eiφ .
Thus if f ∈ C0∞ we have
Z
< ψu, f >= lim eiφ(x,θ) (M t )r (b(x, θ)χ(ǫθ))f (x)dθdx.
ǫ→0

The important point here is that the derivatives do not fall on f as

it is independent of θ. Any terms falling on χ, will have coefficients
which are powers of ǫ and so will disappear in the limit. Thus,

Z
ψu = eiφ(x,θ) (M t )r b(x, θ)dθ.

But (M t )r b(x, θ) ∈ S m−r so we conclude that ψu is smooth as this is

true for any r.

Example 3.6. Let φ(x, y, ξ) = hx − y, ξi then φ is a phase function as

′
dx φ = ξ is non-zero for any non-zero ξ and
′
dξ φ = x − y
so the associated oscillatory integrals are singular only on the diagonal.

4. Pseudo-differential Operators

Definition 4.1. We define P to be a pseudo-differential operator on

Rn of order m, if it has a Schwartz kernel
Z
K(x, y) = eihx−y,ξi a(x, y, ξ)dξ

with a ∈ S m (Rnx × Rny ; Rnξ ). We denote this class ΨDO m(Rn ).

This is slightly different from what we defined before but we shall

see that this definition is equivalent.
We want our operators to act on distributions so first we must show
that they preserve the class of smooth functions.
10 M. S. JOSHI

Theorem 4.2. If P is a pseudo-differential operator then

P : C0∞ (Rn ) → C ∞ (Rn )
and is continuous and if P is properly supported
P : C0∞ (Rn ) → C0∞ (Rn ).

Proof. Let u be a smooth function then

ZZ
Pu = eihx−y,ξi a(x, y, ξ)u(y)dξdy.

For fixed x, this is a well-defined oscillatory integral in (y, ξ) and so,

since the estimates are locally uniform in x, this gives a continuous
function of x. The continuity is immediate from the fact that the
oscillatory integral will be estimated by a finite number of derivatives
in y of u.
The formal x derivatives are of the same form and so will also be
continuous.
To check that the formal derivatives and the actual ones agree we
evaluate the quotient - wlog we consider differentiation in x1
P u(x + he1 ) − P u(x)
=
h
ZZ i<x+he1 ,ξ>
e a(x + hej , y, ξ) − ei<x,ξ> a(x, y, ξ) −i<y,ξ>
e u(y)dξdy
h
=
ZZ  
∂ ′′ ′′
ei(<s1 ,ξ1 >+<x ,ξ > a(s1 , x′′ , y, ξ) e−i<y,ξ> u(y)dξdy
∂s1 |s=x1 +ǫh

with 0 ≤ ǫ ≤ 1. This is just the value of the formal derivative at

x1 + ǫh and as h → 0 this must converge to the value of the formal
derivative at x and so the result follows. (just use induction for the
higher terms).

To define the action of pseudo-differential operators on distributions

we need to find their adjoints. However if K(x, y) is the Schwartz kernel
of P then P t has the kernel K(y, x). So P t has a kernel
Z
ei<y−x,ξ> a(y, x, ξ)dξ

and performing the change of variables η = −ξ we have

Z
ei<x−y,η> a(x, y, −η)dη.
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 11

Thus P t is a pseudo-differential operator and is continuous on C0∞ (Rn ).

Thus we have
Theorem 4.3. If P ∈ ΨDO m (Rn ) and is properly supported then P
induces a map on D ′ (Rn ) via
< P u, φ >=< u, P tφ >, u ∈ D ′ (Rn ), φ ∈ C0∞ (Rn ).

We have that properly supported pseudo-differential operators map

distributions and preserve the space of compactly supported smooth
functions. It is therefore not surprising that they preserve the class of
smooth functions (which is a subclass of distributions.): if P is properly
supported and u is smooth then P u is a distribution. Now given any
point x0 we can put u = u1 + u2 with u1 ∈ C0∞ and u2 such that
x0 6∈ supp(P ) ◦ supp(u2 ) which implies x0 6∈ supp(P u2 ). We thus have
that P u equals P u1 near x0 and thus is smooth near x0 .
Since we have shown that pseudo-differential operators preserve smooth
functions and are singular only the diagonal, it is now easy to show they
have the additional property of pseudo-locality.
Definition 4.4. The operator P is pseudo-local if
singsupp(P u) ⊂ singsupp(u)
for any u ∈ D ′ (Rn ).

Theorem 4.5. Properly supported pseudo-differential operators are pseudo-

local.

Proof. Let P be a pseudo-differential operator with Schwartz kernel K

and u ∈ D ′(Rn ).
Given ǫ > 0, it is enough to show that the singularities of P u lie
within an ǫ neighbourhood of those of u. To see this we decompose
both K and u into pieces supported near the singularities and pieces
which are smooth. So suppose
K = K1 + K2
with the support of K1 contained in
{|x − y| < ǫ/2}
and K2 smooth and
u = u1 + u2
with supp(u1 ) within an ǫ/2 neighbourhood of singsupp(u1 ) and u2
smooth. Then
Ku = K1 u1 + K2 u1 + Ku2 .
12 M. S. JOSHI

Now from the relation of supports, we have

supp(K1 u1 ) ⊂ {x : |x − y| < ǫ, y ∈ singsupp u}
and since the singular support is smaller than the support, the first
term is OK. We also have Ku2 is smooth as u2 is.
This leaves K2 u1 . This turns out to be smooth also as K2 is and the
result will then follow. This is an important result so we will make it
a theorem.

Theorem 4.6. If K ∈ C ∞ (Rn × Rn ) and is properly supported then

K induces a map from D ′ (Rn ) to C ∞ (Rn ).

Proof. Consider the map

Ku : x 7→< K(x, y), u(y) > .
This is well-defined for each x, as K(x, y) is a smooth function in y.
We want to show the result is smooth in x.
But by the definition of a distribution, we have for some k,
X
kKu(x) − Ku(x′ )k ≤ sup |Dyα (K(x, y) − K(x′ , y))|.
y
|α|≤k

The properness of K will ensure that the supremum is over a compact

set and thus the function will certainly be continuous.
Note the same is true for the maps
Dxα : x 7→< Dxα K(x, y), u(y) >
so the formal derivatives are continuous too. So, as in the proof that
pseudo-differential operators are continuous, Ku is smooth as required.

Since we have shown that smooth kernels kill all singularities, we

will ignore them most of the time. So we shall regard two operators as
equivalent if and only their Schwartz kernels differ by a smooth term
and we write this as

P ≡ Q.

We shall call an inverse up to a smoothing term a parametrix. Our

theorem shows that such an operator will allow us to describe the
singularities of u from those of P u. Note that if K is smoothing then
P K, KP are smoothing also for any pseudo-differential operator so the
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 13

smoothing operators form an ideal in the class of pseudo-differential

operators.
The next thing we want to do is prove that the composition of two
pseudo-differential operators is a pseudo-differential operator. In order
to do this, we need a better understanding of the relationship between
the symbol a and the associated operator P.
Theorem 4.7. If P is a pseudo-differential operator of order m, then
up to smooth terms the Schwartz kernel of P can be written in the form
Z
eihx−y,ξi c(x, ξ)dξ

or in the form Z
eihx−y,ξi d(y, ξ)dξ
with c, d ∈ S m (Rn ; Rn ).

Note these are called the left and right quantizations.

Proof. So we have initially that

Z
K(x, y) = eihx−y,ξi a(x, y, ξ)dξ.

We want to eliminate the dependence on one of the base variables.

Recall from Taylor’s theorem that

X 1 α X
a(x, y, ξ) = ∂y a(x, x, ξ)(y − x)α + (y − x)α bα (x, y, ξ)
|α|≤N −1
α! |α|=N

R1
and bα = (1 − t)N −1 ∂yα a(x, (1 − t)x + ty, ξ)dt. It is immediate (and
0
important) that ∂yα a(x, x, ξ) ∈ S m and bα (x, y, ξ) ∈ S m .
Let’s consider the terms ∂yα a(x, x, ξ)(y − x)α ,

Z Z
ihx−y,ξi
e (x − y)α∂yα a(x, x, ξ)dξ = Dξα eihx−y,ξi ∂yα a(x, x, ξ)dξ

and so integrating by parts we obtain

Z Z
eihx−y,ξi (x − y)α∂yα a(x, x, ξ)dξ = eihx−y,ξi (−1)|α| Dξα ∂yα a(x, x, ξ)dξ.

(check this actually works!)

14 M. S. JOSHI

We have applying the same argument to bα that the result is true

up to a term of order m − N. To proceed further, we need to sum the
series
X 1
Dξα ∂yα a(x, x, ξ).
α α!
This sum will not converge but we can make sense of it in an asymp-
totic way. We will use the following result:
Proposition 4.8. If aj ∈ S m−j then there exists a ∈ S m such that
N
X −1
a− am−j ∈ S m−N , ∀N
j=0

The proof of this is on the example sheet 1. Note an immediate

corollary of this, which will come in handy later on,
Corollary 4.9. If Pj ∈ ΨDO m−j (Rn ) for j = 0, . . . , ∞ then there
exists P ∈ ΨDOclm−j (Rn ) such that
X
P− Pj ∈ ΨDO m−N (Rn )
j<N

for all N.

So with this proposition, we let c be an asymptotic sum of

X 1 α α
D ∂ a(x, x, ξ)
α α! ξ y
and is then clear that
Z Z
eihx−y,ξi c(x, ξ)dξ − eihx−y,ξi a(x, y, ξ)dξ ∈ ΨDO m−N , ∀N

which establishes the first half of the result.

The second half is identical, except that the Taylor expansion is in
y not x.

Note we have actually done better than the statement of our theorem
in that the proof gives us asymptotic formulaes for c, d
X 1 α α X (−1)|α|
c(x, ξ) ∼ Dξ ∂y a(x, x, ξ), d(y, ξ) ∼ Dξα ∂xα a(y, y, ξ).
α α! α α!

Note that unlike our original symbol which depended on x, y and

was definitely not unique, these symbols are essentially unique.
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 15

If !n
Z
1
K(x, y) = eihx−y,ξi a(x, ξ)dξ
2π
then this is really just a Fourier transform across x = y. Putting w =
x − y, x′ = x, we have
!n Z
′ ′ ′ 1
L(x , w) = K(x , x − w) = ei<w,ξ>a(x′ , ξ)dξ
2π
and so Z
′
a(x , ξ) = e−iw.ξ L(x′ , w)dw.
i.e. a is determined by K. A smooth error on K, for K properly sup-
ported, would translate into an error of order −∞ on a, as the Fourier
transform of a compactly supported smooth function is Schwartz.
Definition 4.10. If P has Schwartz kernel
Z
eihx−y,ξi a(x, ξ)dξ

with a ∈ S m , up to smooth terms, then the total left symbol of P ,

σL (P ), is a(x, ξ). We will regard σL (P ) as an element of S m /S −∞ .

So the total left symbol determines P up to order −∞.

The same arguments work on the other side and we make a similar
definition.
Definition 4.11. If P has Schwartz kernel
Z
eihx−y,ξi b(y, ξ)dξ

with b ∈ S m ,up to smooth terms, then the total right symbol of P ,

σR (P ), is b(y, ξ). We will regard σR (P ) as an element of S m /S −∞ .

Why use left and right symbols? If we have operators Pj ∈ ΨDO mj (Rn )
then giving P1 the left form and P2 the right one, they can be written
(up to smooth terms)
!n Z
1
P1 u = eix.ξ p(x, ξ)û(ξ)dξ
2π
and !n Z Z
1
P2 v = eix.ξ q(y, ξ)v(y)dydξ.
2π
So it is then immediate from the Fourier inversion theorem that
16 M. S. JOSHI

!n Z
1
P1 P2 v = eihx−y,ξi p(x, ξ)q(y, ξ)u(y)dydξ.
2π
So P1 P2 is a pseudo-differential operator of order m1 + m2 with total
symbol p(x, ξ)q(y, ξ). Of course, we want a more symmetric expression
in terms of outputs and inputs - we want to compute the left symbol
of the composite in terms of the left symbols of the inputs.
Theorem 4.12. If Pj ∈ ΨDO mj are properly supported and have total
symbols pj (x, ξ) then P1 P2 ∈ ΨDO m1+m2 and has total symbol
X i|α| α
a(x, ξ) ∼ D p1 (x, ξ)Dxα p2 (x, ξ).
α α! ξ

The only thing left to prove is the correctness of the asymptotic

expansion. This is just an exercise in binomial expansions - first convert
the left symbol of p2 to a right one and then convert the product to a
left one.

5. Classicality

So far we have been working with general pseudo-differential op-

erators - this class is in fact larger than we need and we can work
with a smaller class which has some nice properties and will include all
the operators we actually need. Recall that a homogeneous function
smoothed off at the origin is a symbol so we can require a symbol to
have an asymptotic expansion in homogeneous terms.
Definition 5.1. A symbol a(x, θ) of order m is classical if there exists
a sequence of functions am−j (x, θ), homogeneous of order m − j in θ
such that
N
X −1
a− (1 − φ)(θ)am−j (x, θ) ∈ S m−N
j=0
with φ a compactly supported function equal to 1 near 0.

Definition 5.2. A pseudo-differential operator is classical if its total

left (or right) symbol is classical. We denote this class ΨDOclm (Rn )

Note that the asymptotic expansion is unique and so knowing the

operator and knowing the expansion of its symbol are really the same
thing up to smooth terms which we always ignore!
It is obvious that composition preserves classicality and that differ-
ential operators are classical pseudo-differential operators. The most
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 17

important thing about classical operators is that they have a unique

principal symbol.
Definition 5.3. If P ∈ ΨDOclm (Rn ) then if P has Schwartz kernel
Z
eihx−y,ξi p(x, ξ)dξ

with asymptotic expansion

X
pm−j (x, ξ)
then the principal symbol of P is pm (x, ξ) and is denoted σm (P )(x, ξ).

We denote the space of smooth homogeneous functions on Rnx ×

(Rnθ − {0}) which are homogenous of degree m in θ by H m (Rn ; Rn − 0).
We express the fact that σm (P )(x, ξ) determines P up to one order
lower by using a short exact sequence. All this means is that the image
of one map is exactly equal to the kernel (null-space) of the next.
Theorem 5.4. The following sequence is exact
σ
→ΨDOclm−1 (Rn ) −
0− →ΨDOclm (Rn ) −→
m
H m (Rn ; Rn − 0) −
→0.

The reason the principal symbol is particularly useful is that the

formula for products is nice.
′
Theorem 5.5. If P ∈ ΨDOclm (Rn ), Q ∈ ΨDOclm (Rn ) are properly
supported and have classical expansions
X X
pm−j (x, ξ), qm′ −k (x, ξ)
P
then P Q is also classical with asymptotic expansion rm+m′ −l (x, ξ)
where
X 1 α
rm+m′ −l = Dξ pm−j .∂xα qm′ −k
j+k+|α|=l
α!
and thus
σm+m′ (P Q) = σm (P )σm′ (Q).

Proof. We know P Q is a pseudo-differential operator with asymptotic

expansion
X 1 α
Dξ p(x, ξ)∂xα q(x, ξ).
α α!

Now Dξα p(x, ξ) has asymptotic expansion

X
Dξα pm−j
18 M. S. JOSHI

and Dξα pm−j is homogeneous of degree m − j − |α| and ∂xα q(x, ξ) has
asymptotic expansion X
∂xα qm′ −k
with ∂xα qm′ −k homogeneous of degree m′ − k. The result follows by
collecting the terms at each level of homogeneity.
The statement about principal symbols now follows from considering
the top level of homogeneity.

Now we have done all this, we can start constructing parametrices -

ie inverses up to terms of order −∞.
Definition 5.6. A classical pseudo-differential operator is said to be
elliptic of order m if P ∈ ΨDOclm and σm (P ) is never zero.

Example 5.7. The Laplacian

X
∆= Dx2j
n
P
plus any first order perturbation has principal symbol, ξj2 , and is
j=1
therefore elliptic. Or more generally, if gij (x) is a Riemannian metric
on Rn and g ij (x) is the inverse matrix then the variable coefficient
Laplacian X
∆= g ij Dxi Dxj + E,
P
with E first order, has principal symbol g ij ξi ξj , and is therefore el-
liptic. On R2 , we also have that the Neumann operator
Dx + iDy
has principal symbol ξ + iη and is therefore elliptic.

Theorem 5.8. If P is a properly supported, elliptic, classical pseudo-

differential operator of order m then there exists
Q ∈ ΨDOcl−m
such that
P Q − Id, QP − Id ∈ C ∞ .

Proof. First, note that the principal and total symbol of Id are both 1.
So we know that for the result to be true
σm (P )σ−m (Q) = 1.

So we put σ−m (Q) = σm (P )−1 . The assumption of ellipticity ensures

that σ−m (Q) is smooth and thus we can pick such a Q.
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 19

Now our formula for products of symbols tells us that the lth term
in the expansion for P Q will be

X 1 α
Dξ pm−j (x, ξ)∂xα q−m−k (x, ξ) (3)
|α|+j+k=l
α!
and we need this to be equal to zero for each l ≥ 1.
So suppose that q−m , . . . , q−m−N have been chosen so that (3) is
satisified for l ≤ N then we can define
X 1 α
q−m−N −1 = −p−1 m Dξ pm−j (x, ξ)∂xα q−m−k
|α|+j+k=N +1,k<N +1
α!
and (3) will be satisfied for l = N + 1 also. Thus by induction, we can
construct q−m−k so that (3) is true for all k.
Let q have classical expansion qm−k and then we have
P Q − Id ∈ ΨDO −∞ = C ∞ .

A similar argument constructs Q′ such that

Q′ P − Id ∈ ΨDO −∞ = C ∞ .

But regarding ΨDOclm /ΨDO −∞ as a semi-group, left inverses and

right inverses must be equal and the result follows.

We can now prove elliptic regularity.

Theorem 5.9. (Weyl’s Lemma) If P ∈ ΨDO m is proper, classical
and P is elliptic of order m and u ∈ D ′ (Rn ) and
P u ∈ C∞
then
u ∈ C ∞.
More generally, for any u ∈ D ′ (Rn ),
singsupp(P u) = singsupp(u).

Proof. Let Q be a parametrix for P then QP − Id = R and Ru ∈ C ∞ .

So by pseudo-locality, theorem 4.5,
singsupp(u) = singsupp((Id +R)u)
= singsupp(QP u)
⊂ singsupp(P u)
⊂ singsupp(u).
20 M. S. JOSHI

Note this is actually a deep result and is quite hard to prove by

other means! It also relies heavily on the fact that P is elliptic and is
definitely not true for a general pseudo-differential operator.

6. Continuity

So, we have seen that elliptic operators have parametrices which

provide inverses up to smooth terms. i.e. they provide an inverse
at the singularity level. Sobolev Spaces can be used to measure the
badness of a singularity and we will examine how a pseudo-differential
operator maps between them.
Definition 6.1. We shall say u ∈ H s (Rn ) if u ∈ S ′ (Rn ) and
û(ξ)(1 + |ξ|2 )s/2 ∈ L2 (Rn ).

Note that the lower s is, the more singular u becomes.

It is trivial that
Dxj : H s → H s−1.

There are two ways to fail to be in L2 - a distribution can be too big

at ∞ or too singular at some point. The same is true of H s so, since
what we really want to measure is how singular a distribution is, we
will often work with two related spaces.
Definition 6.2.
Hcs (Rn ) = {u : u ∈ H s (Rn ) and supp(u) compact.}.
s
Hloc (Rn ) = {u : φu ∈ H s (Rn ), ∀φ ∈ C0∞ }
These are the compact and local spaces.

The theorem we want to prove is

Theorem 6.3. If P ∈ ΨDOclm (Rn ) and is properly supported then
P : Hcs → Hcs−m.

Most of the work involved is in proving the simplest case - m = 0, s =

0. ie the L2c continuity of zeroth order pseudo-differential operators.
To do this we will need to understand adjoints of pseudo-differential
operators - we already understand transposes and these are similar.
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 21

Recall, if φ, ψ ∈ C0∞ then the transpose of P , is defined by

Z Z
P φ(x)ψ(x)dx = φ(x)P t ψ(x)dx
whereas the adjoint is defined by
Z Z
P φ(x)ψ(x)dx = φ(x)P ∗ ψ(x)dx.

So these operators are pretty similar and indeed are related by

P ∗ ψ = P t ψ̄.

Thus, if P ∈ ΨDO m has total left symbol a(x, ξ) we deduce that P ∗

is pseudo-differential and has a total (right) symbol
ā(y, ξ)
and so has total left symbol b(x, ξ) with asymptotic expansion
X 1
∂xα Dξα ā(x, ξ).
α α!

So P ∗ will be classical if P is, and

σm (P ∗ ) = σm (P ).

Thus we obtain
Theorem 6.4. If P ∈ ΨDOclm is self-adjoint i.e. P = P ∗ then σm (P )
is real-valued.
If f (x, ξ) is homogeneous of order m in ξ then there exists a pseudo-
differential operator which is self-adjoint with principal symbol f if and
only if f is real-valued.

Proof. The only part of this which is not wholly obvious is that if f
is real-valued then there exists a self-adjoint operator with principal
symbol f. To see this, let σm (Q) = f some Q ∈ ΨDOclm(Rn ) and put
Q + Q∗
P = .
2

We prove continuity using an approximate square root which will

reduce to proving the result in the case where P has smooth kernel.
Theorem 6.5. If P ∈ ΨDOclm is self-adjoint and σm (P ) is positive
m/2
then there exists Q ∈ ΨDOcl which is self-adjoint such that
P − Q2 ∈ ΨDO −∞.
22 M. S. JOSHI

m/2
Proof. For a first guess we can take Q0 ∈ ΨDOcl self-adjoint with
principal symbol σm (P )1/2 which is a valid symbol by our hypothesis.
Then we have
P − Q20 ∈ ΨDO m−1 .
We use a recursive argument so now suppose we have constructed
Q0 , Q1 , . . . , QN −1 classical and self-adjoint such that
 2
N
X −1
P − Qj  = RN ∈ ΨDOclm−N
j=0
m/2−N
and we want to construct QN ∈ ΨDOcl .
Now
 2    
N
X N
X −1 N
X −1
P − Qj  = RN − QN  Qj  −  Qj  QN − Q2N .
j=0 j=0 j=0

The principal symbol of this will be

σm−N (RN ) − 2σm/2−N (QN )σm/2 (Q0 )
and to continue the recursion this has to equal zero. So noting that RN
is self-adjoint we can pick QN to be self-adjoint with principal symbol
equal to
1
σm/2 (Q0 )−1 σm−N (RN )
2
which is of course real.
So once we have Qj constructed for all j then we can pick Q′ such
that X
Q′ − Qj ∈ ΨDO m−N , ∀N.
j<N
To do this just asymptotically sum at the symbol level. Q′ will neces-
sarily be classical as all the Qj are. However Q′ is not obviously self
′ ′∗
adjoint so we put Q = Q +Q
2
and we have constructed the requisite Q.
′
Note that Q − Q is of order −∞ as each Qj is self-adjoint. So Q is
also an asymptotic sum for the Qj .
By construction Q will now be the approximate square root.

Proving L2c continuity of a properly-supported, zeroth-order pseudo-

differential operator, A, means that for each compact set K we have
to show for φ ∈ L2 , supp(φ) ⊂ K that there exists K ′ compact and CK
such that
kAφk ≤ CK kφk
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 23

and supp(Aφ) ⊂ K ′ with the norms taken into the L2 sense. In fact as
compactly supported smooth functions are dense in L2c we need only
prove this for smooth functions. Note that if we only work with smooth
functions then we establish the existence of continuous extension, Ã,
to L2c . But this will agree with A on L2c because A is weakly continuous
on distributions and agrees with à on C0∞ .
The existence of a K ′ as above is immediate from the relation of
supports so we need to establish the bound CK . Since only the be-
haviour of the kernel of A on the set πy−1 (K) will affect its mapping
properties on elements of L2 (K), we can decompose the kernel of A
into one piece which is compactly supported and another which maps
all elements of L2 (K) to zero. Thus it is enough to consider the case
supp(A) compact.
It is easier to work with squares and we have
kAφk2 =hAφ, Aφi
=hφ, A∗ Aφi.

Now A∗ A is self-adjoint and its principal symbol is |σ0 (A)|2 which

is necessarily non-negative but may have some zeroes. The fact that
the kernel of A is compactly supported means that |σ0 (A)|2 will be
compactly supported on the base and since it is homogeneous of degree
zero, we deduce from continuity that it must be bounded above. So
suppose |σ0 (A)|2 < C everywhere.
Then the operator C − A∗ A is self-adjoint and has strictly positive
principal symbol, thus by our theorem 6.5 we have
C − A∗ A = B 2 + G
with B self-adjoint and G smooth (and self-adjoint).
So we have,
kAφk2 = Ckφk2 − hφ, B 2 φi − hφ, Gφi.

The second term equals −kBφk2 and thus is necessarily negative so

it remains simply to show that
|hφ, Gφi| ≤ C ′ kφk2 .
Using Cauchy-Schwartz, it is enough to show the L2c continuity of G.
As before, is enough to consider G compactly supported and since
G is smooth, it must also be bounded (as a function) by say D.
So we have,
24 M. S. JOSHI

Z Z 2
2
kGφk = G(x, y)φ(y)dy dx
and since |G(x, y)| ≤ D and our integrals are supported in the finite
domain K × K ′ , we deduce
kGφk2 ≤ D 2 V kφk2
where V is volume of K × K ′ .
Thus we have proven,
Theorem 6.6. If A ∈ ΨDOcl0 (Rn ) is classical and properly supported
then A is continuous as a map from L2c to L2c .

The fact we are considering only compactly supported functions is

essential here - otherwise consider for example multiplication by ex -
this is clearly not continuous on L2 but is a zeroth order differential
operator.
We now wish to prove in general that properly supported elements
of ΨDOclm(Rn ) map
Hcs → Hcs−m ,
we do this by reducing to the zeroth order case.
So let Λs be a properly supported element of ΨDO s with total symbol
hξis. This is classical as
hξis = (1 + |ξ|2)s/2 = |ξ|s (1 + |ξ|−2)s/2 ,
and we can apply a binomial expansion.
We will show that Λs defines a continuous map from Hct to Hct−s .
Then if P ∈ ΨDOclm is properly supported, we have that
P = Λs (Λ−s P Λs−m) Λm−s + R
with R of order −∞. Now, (Λ−s P Λs−m) ∈ ΨDOcl0 and is properly
supported, so it defines a continuous map of L2c . So if we know that
R is continuous, it immediately follows that P is continuous as a map
from
Hcs → Hcs−m .
(Recall that H 0 = L2 .)
We can write
Λs u = Ms u + R′ u
where !n Z
1
Ms u = eix.ξ hξis û(ξ)dξ
2π
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 25

where R′ is of order −∞. Unfortunately, Ms will not be properly sup-

ported but if supp(u) ⊂ K, K compact, we have that supp(Λs u) is
contained in
K ′ = supp(Λs ) ◦ K.
So if we take ψ = 1, ψ ∈ C0∞ (Rn ) on K ′ , we have that
Λs u = ψMs ψu + ψRψu
and ψRψ will have kernel ψ(x)R(x, y)ψ(y).
Thus to show the continuity of Λs , it is enough to show three things
1. Ms : H t → H t−s
2. Multiplication by ψ is continuous map from H s to Hcs .
3. If R ∈ ΨDO −∞ and is compactly supported then R defines a
continuous map from Hcs → HcN for any N.
The most innocuous looking of these three, 2., is in fact the hardest!
Let’s start with 1. Applying Ms is equivalent to multiplying on the
Fourier transform side by
hξis,
which by definition is an isometry from H t → H t−s .
To see 3., apply the Fourier transform to obtain
Z
d
Ru(ξ) = r̃(ξ, y)u(y)dy.
Here r̃ is the partial Fourier transform. We thus deduce,
Z
d
Ru(ξ) = (2π)n r̂(ξ, η)û(−η)dη.

Of course, r̂(ξ, η) is Schwartz and so is less than or equal to

Chξi−M1 hηi−M2
for any M1 , M2 . So we conclude,
Z 1
2
d
|Ru(ξ)|hξi 2N
≤ N 2
hξi |r̂(ξ, η)| hηi −2s
dη kuks .

ˆ
The integral will be rapidly decreasing in ξ and so we have that |Ru(ξ)|hξi N
∈
2
L with norm less than or equal to Ckuks and the continuity follows.
This leaves us with the problem of proving 2. We first prove this for
s ≥ 0 and then use duality to deduce the general case.
Let Ms be the operator defined by
du(ξ) = hξis û(ξ)
Ms
26 M. S. JOSHI

this defines an isometry from

H t → H t−s
for each t. So Mψ is a continuous map on H s if and only if
Ts = Ms Mψ M−s
is a continuous map on H 0 = L2 . Since Cc∞ , is dense in L2 we need
only check this on Cc∞ . So let u ∈ Cc∞ and then Td
s u is given by
Z
hξis ψ̂(ξ − ξ ′ ) < ξ ′ >−s û(ξ ′ )dξ ′.
Using the fact that
(ξj − ξj )ψ̂(ξ − ξ ′ ) = Dd
′ ′
j ψ(ξ − ξ ),

we can rewrite this for even N as

Z
hξis d (ξ − ξ ′ ) < ξ ′ >−s û(ξ ′ )dξ ′ ,
hξ − ξ ′ i−N ψN

for some ψN ∈ C0∞ (Rn ). Noting that

hξi ≤ C < ξ ′ > hξ − ξ ′i,
for s ≥ 0 we have
Z
|Tdu| ≤C d (ξ − ξ ′ )||û(ξ ′ )|dξ ′ .
|ψ
s N

Recall that the Fourier transform is an isometry of L2 so it is enough

to estimate the L2 norm of |Tds u|. Now
Z Z Z 2
|Tdu(ξ)|2dξ ≤ d (ξ
|ψ ′ ′
− ξ )||û(ξ )|dξ ′
dξ
s N
Z Z 2
1 1
= d (ξ
|ψ ′
− ξ )| 2 d (ξ
|ψ ′ ′
− ξ )| |û(ξ )|dξ
2
′
dξ
N N

so applying Cauchy-Schwartz,
Z Z Z 
≤ d (ξ − ξ ′ )||û(ξ ′ )|2 dξ ′
|ψ d (ξ − ξ ′ )|dξ ′ dξ
|ψ
N N
Z Z
= d (ξ − ξ ′ )||û(ξ ′ )|2 dξ ′ dξkψ
|ψ b k
N N L1

putting η = ξ − ξ ′ , we have
Z
= d k2 .
|û(ξ ′ )|2 dξ ′ kψN L1

Thus we have shown,

d k kuk
kTs uk2 ≤ kψN 1 2
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 27

and so Mψ is continuous for s ≥ 0. The technique we will use to prove

the result for general s is sufficiently important that we will give it its
own section.

7. Duality For Sobolev Spaces

There is a natural pairing between H s and H −s , we define

h., .i :H s × H −s → C (4)
Z
hu, vi = û(ξ)v̂(ξ)dξ. (5)
We can think of this as the L2 pairing of hξi−sû and hξis v̂ which shows
that the integral makes sense. This pairing is linear in the first factor
and quasi-linear in the second. It is also continuous - using Cauchy-
Schwartz for the L2 pairing, we have |hu, vi| ≤ kuks kvk−s .
This means that the pairing gives a natural map from H s to the dual
space of H −s . Simply, define
T : H s → (H −s )∗ , (6)
T v(u) =< v, u > (7)
The continuity of the pairing yields that T v is in (H −s )∗ and that the
norm of T v is less than or equal to kvks .
Now if v ∈ H s then we can define u in H −s via
û(ξ) = hξi2sv̂(ξ)
and then Z
|T v(u)| = T v(u) = |v̂(ξ)|2hξi2s dξ = kvk2s
and so
kT vk = kvks .
So T is a quasi-linear isometry. It is in fact surjective. Recall that as
H is a Hilbert space, it is naturally isomorphic to (H −s )∗ . So given an
−s

element, f , of (H −s )∗ , we have from the Riesz representation theorem

an element v ∈ H −s such that
f (u) = hu, vis, ∀u ∈ H −s .

But Z
hu, vi−s = û(ξ)v̂(ξ)hξi−2s dξ.
So taking v1 in H s such that vˆ1 (ξ) = hξi−2s v̂(ξ), we have that
T v1 = f.
28 M. S. JOSHI

So T is onto.
Recall that we wanted to show that multiplication by a compactly
supported smooth function is continuous on Sobolev spaces and we had
already shown this for s ≥ 0.
The adjoint of multiplication by ψ on H s is just multiplication by ψ̄
on H −s .
So for s < 0, u ∈ H −s , v ∈ H s we deduce,
|hMψ v, ui| = |hv, Mψ ui|
≤ kvks kMψ uk−s
≤ Ckvks kuk−s .
This says precisely that as an operator the norm of Mψ v is less than
or equal to Ckvks . As the operator norm equals the original norm, we
deduce that Mψ is continuous. Note that the idea of this proof is very
important and is a standard technique.

8. Coordinate Invariance

We know that under changes of coordinates differential operators

remain differential operators and we can compute the new operator
by repeated use of the chain rule. We want to do the same thing for
pseudo-differential operators - ie prove they remain pseudo-differential
and compute the symbol of the resulting operator. It turns out that
the lower order terms of the symbol are very complicated but that the
principal symbol transforms very nicely.
What do we really mean by a change of coordinates, well suppose
we have on operator on an open subset U of Rn and the operator P
is a map from compactly supported smooth functions on U to smooth
functions on open V containing U. Then if we have another pair of
sets U1 , V1 and a map χ from V1 to V such that χ(U1 ) = U and χ is
a smooth bijection and has a smooth inverse then we can define an
operator from compactly supported smooth functions on U1 to smooth
functions on V1 by
h i
P1 f = P (f ◦ χ−1 ) ◦ χ.
That is we take the function f on U1 and make it a function on U by
pulling back by χ−1 i.e. taking the map
 ∗
χ−1 f :U →C
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 29

defined by
 ∗
χ−1 f (x) = f (χ−1 x)
and then we apply P to get an function on and then by we pull-back
the resulting functon by χ to get a function on V1 . Thus we have a
map, P1 , from C0∞ (U1 ) → C ∞ (V1 ).
So now suppose P is a pseudo-differential operator, we can write
Z
P1 u(x) = ei<χ(x)−y,ξ> a(χ(x), y, ξ)u(χ−1(y))dydξ.

Doing a change of coordinates, y = χ(z) this becomes,

Z
P1 u(x) = ei<χ(x)−χ(z),ξ> a(χ(x), χ(z), ξ)u(z)|J|dzdξ,

where J is the Jacobian of the transformation.

Now using Taylor’s theorem,
χ(x) − χ(z) = A(x, z)(x − z)
where A(x, z) is a square matrix varying smoothly with x and z and
most importantly
∂χ(x)
A(x, x) = .
∂x
This means that A is invertible on the diagonal and also therefore
nearby. (We need only worry about behaviour close to the diagonal as
the phase is clearly critical only there and so the kernel is smooth off
the diagonal.)
So our phase is
< χ(x) − χ(z), ξ >=< x − z, At (x, z)ξ > .

So if we do another change of variables, η = At (x, z)ξ,

Z
P1 u(x) = ei<x−z,η> a(χ(x), χ(z), (At )−1 η)u(z)|K||J|dzdη, (8)
where K is the determinant of this transformation. This is roughly in
the form of a pseudo-differential operator. Both |K|, |J| are smooth
functions independent of η and so will be harmless. Our problem is
thus to check that
a1 (x, z, η) = a(χ(x), χ(z), (At )−1 (x, z)η)
is a symbol.
To prove this we use a lemma which often comes in handy when
trying to establish something is a symbol.
30 M. S. JOSHI

Lemma 8.1. If Fm ⊂ C ∞ (Rnx × Rkθ ) for all m and if b ∈ Fm implies

that
∂b X X ′ ′
= aα bα + aα bα
∂xj α α
′ ′
with aα ∈ S 0 , aα ∈ S 1 and bα ∈ Fm , bα ∈ Fm−1 and
∂b X
= cβ d β
∂θk β

with cβ ∈ S 0 and dβ ∈ Fm−1 and for each element, b, of Fm and

compact K there exists CK such that
|b(x, θ)| ≤ CK < θ >m , x ∈ K, K compact
then
Fm ⊂ S m (Rn ; Rk ).

Proof. This follows easily from repeated use of the Leibniz rule for
products.

So now defining Fm = {a(χ(x), χ(z), (At )−1 (x, z)η) : a ∈ S m }, it

follows that a1 is a symbol of order m and so P1 is a pseudo-differential
operator of order m.
As usual, what we really want to know is how the principal symbol
transforms and we also want to show that P1 is classical if P is.
So if we take a to be independent of y, which we can do, then the
total symbol of P1 is
a(χ(x), (At )−1 (x, z)η)|J||K|.
We can now compute the left symbol of P1 in the usual way:
X 1  
σL (P1 ) ∼ ∂zα Dηα a(χ(x), (At )−1 (x, z)η)|J||K| .
α α!
z=x

It is now obvious that σL (P1 ) is classical and we have a way of comput-

ing the total symbol. If am is the top order part of a then the principal
symbol of P1 will be
am (χ(x), (At )−1 (x, x)η)|J(x, x)||K(x, x)|.
However |J(x, x)||K(x, x)| = 1 as these are the Jacobians of opposite
transformations and
!−1
t −1 ∂χ t
(A ) (x, x) = .
∂x
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 31

This says precisely that the principal symbol of P1 is the principal sym-
bol of P pulled back by the lifting of χ to a function on the cotangent
bundle.
We can use this to deduce the coordinate invariance of Hcs . If Ps
is a classical, properly-supported, elliptic, pseudo-differential operator
of order s then u in Hcs if and only if P u ∈ L2c . But we know L2c is
coordinate invariant and Ps will transform to an operator with the
same properties so coordinate invariance is immediate.

9. Wavefront Sets

Recall the definition of wavefront set

Definition 9.1. If u ∈ D ′ (Rn ) then
(x0 , ξ0 ) 6∈ WF(u) ⊂ Rn × (Rn − {0})
if and only if there exists C ′ > 0, and φ smooth and of compact support
such that φ(x0 ) 6= 0 and for all N
c
|φu(ξ)| ≤ CN < ξ >−N
for
ξ ξ0
− ≤ C ′.
|ξ| |ξ0 |

We can reexpress this using pseudo-differential operators but first

we need to introduce the concept of the wavefront set of a pseudo-
differential operator! (note this is not the same as the wavefront set of
its Schwarz kernel although they are closely related.)
Definition 9.2. If P ∈ ΨDO m then (x0 , ξ0 ) 6∈ WF′ (P ) if and only if
P has total left symbol p(x, ξ) such that
|Dxα Dξβ p(x, ξ)| ≤ CN,α,β < ξ >−N
for x near x0 and ξ in a conic neighbourhood of ξ0 .

Remark 9.3. In fact the estimate for α = β = 0 together with the fact
that p is a symbol is enough. Try to prove this!

The idea here is that WF′ (P )c expresses the directions in which P

kills all singularities as it is of order −∞. Note that using the right
symbol gives the same wavefront set. We trivially have that
WF′ (AB) ⊂ WF′ (A) ∩ WF′ (B).
32 M. S. JOSHI

-once everything has been killed in a certain direction, we will not be

able to recover it.
Note also that P ∈ ΨDO m will be ΨDO −∞ if and only if WF′ (P ) =
∅.
This leads to the important concept of micro-ellipticity which we will
show is equivalent to inverting up to order −∞ in a given direction.
Definition 9.4. If P ∈ ΨDOclm then P is micro-elliptic of order m at
(x0 , ξ0 ) if σm (P )(x0 , ξ0) 6= 0.

Theorem 9.5. If P ∈ ΨDOclm then P is micro-elliptic at (x0 , ξ0 ) if

and only if there exists Q ∈ ΨDOcl−m such that
(x0 , ξ0 ) 6∈ WF′ (P Q − Id), (x0 , ξ0 ) 6∈ WF′ (QP − Id).
Q is said to be a micro-parametrix for P.

Proof. Just repeat the proof of the construction of a parametrix of a

elliptic operator. The symbols involved may become singular away
from (x0 , ξ0) but if we multiply them all by
!
ξ ξ0
φ(x − x0 )φ −
|ξ| |ξ0|
where φ is a bump function supported sufficiently close to zero, they will
be smooth and retain the property of invertibility close to (x0 , ξ0 ). This
proves the result. The converse is obvious from considering principal
symbols.

So in some sense the only difficult (interesting) points are where

σm (P ) is zero.
Definition 9.6. If P ∈ ΨDOclm then the characteristic variety of P is
char(P ) = σm (P )−1 (0).

Theorem 9.7. If u ∈ D ′ (Rn ) then

WF(u) = ∩{char P : P u ∈ C ∞ }.

ˆ is rapidly decreasing
Proof. If (x0 , ξ0 ) 6∈ WF(u) there exists φ such φu
∞ n−1
in direction ξ0 so picking ψ ∈ C (S ) such that
!
ξ0
ψ 6= 0
|ξ0 |
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 33

ˆ is rapidly decreasing in all directions in supp(ψ), we pick a

and φu
pseudo-differential operator P with total symbol
!
ξ
φ(y)ψ .
|ξ|
It is then clear that
P u ∈ C ∞ and (x0 , ξ0 ) 6∈ WF(P u).
So we have proven the result one way.
So now suppose P is micro-elliptic at (x0 , ξ0 ) and P u ∈ C ∞ then
applying a micro-parametrix for P at (x0 , ξ0), we have
u + Ru ∈ C ∞
and (x0 , ξ0 ) 6∈ WF′ (R). So pick φ, ψ to be bump functions such that
!!
ξ ξ0
supp φ(x − x0 )ψ − ∩ WF′ (R) = ∅.
|ξ| |ξ0|
 
ξ ξ0
Then picking S to have total symbol φ(y − x0 )ψ |ξ|
− |ξ0 |
, we have
that WF′ (S) ∩ WF′ (R) = ∅ and so
Su ∈ C ∞ ,
(using the pseudo-locality of S.)
But this says precisely that
(x0 , ξ0 ) 6∈ WF(u).

Since we have already shown that the principal symbols of pseudo-

differential operators transform as functions on the cotangent bundle,
we have that wavefront set is coordinate invariant if we regard it as a
subset of the cotangent bundle.
Theorem 9.8. If P ∈ ΨDOclm is properly supported and u ∈ D ′ (Rn )
then
WF(P u) ⊂ WF(u) ∩ WF′ (P ).

This says that pseudo-differential operators are micro-local. This

is a stronger property than pseudo-locality but of course weaker than
locality.
34 M. S. JOSHI

Proof. To prove the first inclusion, if (x0 , ξ0 ) 6∈ WF(u) then there exists
Q′ micro-elliptic at (x0 , ξ0 ) such that Q′ u is smooth. Taking a micro-
local parametrix, Q′′ , for Q′ at (x0 , ξ0 ), we have putting Q = Q′′ Q′ that
Qu ∈ C ∞ and that the total symbol of Q is identically one near (x0 , ξ0)
i.e. Q = Id +R with (x0 , ξ0 ) 6∈ WF′ (R). Now, P Qu will of course be
smooth and
QP u = P Qu + [Q, P ]u.
′
But (x0 , ξ0 ) 6∈ WF ([Q, P ]) as
[Q, P ] =(Id +R)P − P (Id +R)
=RP − P R.
So picking S micro-elliptic at (x0 , ξ0 ) such that
WF′ (S) ∩ WF′ (R) = ∅
we have that
SQP u ∈ C ∞ .
As SQ is micro-elliptic at (x0 , ξ0 ), the first part of the result follows.
The second part is easier. If
(x0 , ξ0 ) 6∈ WF′ (P )
then we can take S micro-elliptic at (x0 , ξ0) such that
WF′ (SP ) ⊂ WF′ (P ) ∩ WF′ (S) = ∅
and so SP ∈ ΨDO −∞ This implies that
SP u ∈ C ∞
which proves the result.

Corollary 9.9. If P ∈ ΨDOclm, properly supported then

WF(u) ⊂ WF(P u) ∪ char(P ).

Proof. This is just a simple application of micro-ellipticity. If

(x0 , ξ0) 6∈ WF(P u) ∪ char(P )
then P is micro-elliptic at (x0 , ξ0) and (x0 , ξ0 ) 6∈ WF(P u). So letting
Q be a micro-local parametrix for P at (x0 , ξ0 ), we have that
(x0 , ξ0) 6∈ WF(QP u).
But
QP u = u + Ru
with
(x0 , ξ0) 6∈ WF′ (R)
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 35

and hence
(x0 , ξ0 ) 6∈ WF(Ru).
So
(x0 , ξ0 ) 6∈ WF(u)
which is what we wanted to prove.

10. The Propagation of Singularities

In the previous section, we showed that if P u ∈ C ∞ then

WF(u) ⊂ {σm (P ) = 0}.
The obvious next question is which subsets are possible? We look
at this question in the case when σm (P ) is real. The answer due to
Hörmander is that the wavefront set is a union of integral curves of
the Hamiltonian of the principal symbol of P. If pm = σm (P ) then the
Hamiltonian is the vector field on T ∗ (Rn )
n
X ∂pm ∂ ∂pm ∂
Hp m = −
j=1 ∂ξj ∂xj ∂xj ∂ξj
or if you prefer
!
∂pm ∂pm ∂pm ∂pm ∂pm
, ,..., ,− ,...,− .
∂ξ1 ∂ξ2 ∂ξn ∂x1 ∂xn

Remark 10.1. The first order differential operator and the vector field
are usually identified in the study of differential geometry. The point
being that if γ(t) is an integral curve of a vector field v = (v1 , . . . , vn )
on Rn then
d X ∂f
(f ◦ γ)(t) = vi (γ(t)) (γ(t)).
dt ∂xi

A good question is, of course, whether this is the end of the story. For
example, if the wavefront set had to always be the entire characteristic
variety or empty then the result would be true but boring. In fact,
Hörmander also showed that any closed subset of the characteristic
variety which is a union of integral curves is possible but that is beyond
the scope of this course.
Theorem 10.2. Let P be a proper, classical pseudo-differential oper-
ator and let u ∈ D ′ be such that P u ∈ C ∞ then WF(u) is a union of
integral curves of the Hamiltonian of σm (P ).
36 M. S. JOSHI

There are two basic approaches to proving this result the highbrow
way is to develop the calculus of Fourier integral operators and then
reduce to the case where P = Dx1 in which case it is easy. The lowbrow
and original proof is the method of commutators which we will use here.
The main disadvantage of this is that we will need an excursion into
functional analysis.
Before we proceed note that
H p m pm = 0
that is pm is constant along integral curves of Hpm , in particular the
characteristic variety of P is invariant under the flow. We call the
integral curves in the characteristic variety bicharacteristics. Note that
this also shows the characteristic variety is a union of bicharacteristics.
If Q is an elliptic differential operator of order 1 − m then
WF(QP u) = WF(u)
and if Q has principal symbol q then
Hqpm = qHpm + pm Hq
so in pm = 0, we have that Hqpm = qHpm that is that Hqpm is a non-
zero multiple of Hpm and thus will have the same integral curves. It is
thus enough to prove the theorem in the special case when P is of first
order.
As the WF set is a closed set it is enough, by connectedness, to show
that if a point is in the WF set then it is nearby on the bicharacteristic
too so our although our result is global it can be proved locally. We thus
fix some point x0 and let φ be compactly supported and identically one
near x0 and ψ be identically one near x0 supported within the set where
φ is identically one. Then if P u is smooth so is ψP (φu). We then have
that ψP φ has compact kernel and that it has the same bicharacteristics
in a neighbourhood of x0 - this means it is enough to consider the case
where P has a compact kernel.
′
In general two pseudo-differential operators P ∈ ΨDOclm , Q ∈ ΨDOclm
will not commute eg
P = Dx1 , Q = x1
but the principal symbols
σm+m′ (P Q) = σm (P )σm′ (Q) = σm+m′ (QP )
are equal so
′
[P, Q] = P Q − QP ∈ ΨDO m+m −1
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 37

and consulting our asymptotic expansions has principal symbol 1i {pm , qm′ }
that is  
n
1 X ∂pm ∂qm′ ∂pm ∂qm′  1
− = Hpm qm′
i j=1 ∂ξj ∂xj ∂xj ∂ξj i
which is the crucial point here.
We will prove the result by considering the operator Dt − P and
constructing operators which commute with it. If (x0 , ξ0 ) 6∈ WF(u)
then there is a Q′ ∈ ΨDOcl0 (Rn ) micro-elliptic at (x0 , ξ0) such that
Q′ u ∈ C ∞ . We try to find a family Q(t, x, Dx ) which are pseudo-
differential for each t, with total symbol varying smoothly with t such
that
(Dt − P )Q = Q(Dt − P )
′
and Q (x, Dx ) = Q(0, x, Dx ). It then follows that
(Dt − P )(Qu) ∈ C ∞
and
Qu|t=0 ∈ C ∞ .
There is a regularity theorem which we will prove that then guarantees
that Qu is smooth for all t. This means that
WF(u) ⊂ {q0 (t, x, ξ) = 0}
for each t. Let γx0 ,ξ0 (t) be the bicharacteristic through (x0 , ξ0 ) then we
will show that q0 (t, γ(x0 ,ξ0 ) (t)) is zero if and only if q0 (0, x0 , ξ0 ) is zero
and the theorem then follows.
So how do we construct Q ? The principal symbol of [Dt − P, Q] is
∂q0
− Hp1 q0 .
∂t
We want this to be zero and q0 (0, x, ξ) = q(x, ξ). This is just a standard
first order linear PDE and is solved by making q constant along the
integral curves of (1, Hp1 ) which are of the form (t, γ(−t)) where γ(t)
is an integral curve of Hp1 . So q is either 0 or non-zero all along such
an integral curve which is what we need.
We also need to construct the lower order terms of Q. If we have con-
structed the higher terms then the next term in the symbol expansion
will be of the form
∂q−j
− Hp1 q−j = function of higher order terms
∂t
this being linear we can as usual solve. Asymptotically summing we
then have Q which commutes with Dt − P up to smooth terms which
is enough.
38 M. S. JOSHI

So we have proven the result modulo the proof of the regularity

theorem. We prove the regularity result by using Sobolev spaces. Since
we have restricted to operators P which have compact kernels and real
principal symbol we have that P −P ∗ is zeroth order and has a compact
kernel - this guarantees that it is continuous on all Sobolev spaces.
We first prove an energy estimate and then use a duality argument
to construct a solution of Dt − P with the right regularity properties
which by uniqueness then agrees with the original solution which means
it has the right regularity properties too.
Proposition 10.3. Let L = Dt − P then for all s ∈ R, there exists λs
such that for λ ≥ λs , we have for
 
u ∈ C 1 ([0, T ]; H s ) ∩ C 0 [0, T ], H s+1
that
ZT
−λt
sup e ku(t, .)ks ≤ ku(0, .)ks + 2 e−λt kLu(t, .)ks dt.
t∈[0,T ]
0

Proof. It will be slightly more convenient to eliminate the 1/i so we

instead study
∂
L′ = −Q
∂t
with Q = iP so R = Q + Q∗ is continuous on H s .
We first do the case with s = 0 that is for L2 . We have
d −λt 2 d Z −2λt
ke u(t, x)k0 = e u(t, x)ū(t, x)dx
dt dt
which equals
du
−2λke−λt u(t, x)k20 + 2ℜhe−2λt , ui.
dt
Now
hQu, ui = hu, Q∗ui = −hu, Qui + hu, Rui = −hQu, ui + hu, Rui,
which implies that
2ℜhQu, ui = hu, Rui ≤ kRkkuk20
where kRk is the operator norm on L2 . So we conclude that
d −λt 2 2
2ℜhe−2λt L′ u, ui ≥ e u(t, .) + (2λ − kRk) e−λt u(t, .) .
dt 0 0
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 39

So provided λ ≥ kRk/2, we can drop the last term and then integrating
from 0 to t we have that
ZT
−λt 2 −λt 2
e u(t, .) − e u(0, .) ≤2 ℜhe−2λt L′ u, uidt.
0 0
0

Putting M = sup e−λt u(t, .) and applying Cauchy Schwartz to

t∈[0,T ] 0

ℜhe−2λt L′ u, ui, we have that

ZT
2 2
M ≤ ku(0, .)k + 2M e−λt kLu(t, .)k0 dt.
0

As ku(0, .)k0 ≤ M, we have, dividing by M, that the result follows for

L2 .
To do the general case, we conjugate by the operators Λs from section
6. We then have
Lu = Λ−s L̃Λs u
and L̃ is obtained from L by replacing P by P̃ = Λs P Λ−s . So
kLuks = kL̃Λs uk0 .

Theorem 10.4. Let L be as above and s ∈ R. For every f ∈ L1 ((0, T ); H s)

and φ ∈ H s there is then a unique solution u ∈ C([0, T ]; H s) of
Lu = f, u|t=0 = φ
and this solution satisfies the energy estimate (10.3).

Proof. We work with L′ = iL as before.

To see uniqueness that if u1 , u2 are both solutions then u = u1 − u2
satisfies the equation with f, φ = 0. We have that
∂u
= P u ∈ C([0, T ]; H s−1)
∂t
and so we have that
u ∈ C 1 ([0, T ]; H s−1) ∩ C 0 ([0, T ]; H s).
So applying Proposition 10.3 with s replaced by s − 1 we have that
sup e−λt ku(t, .)k(s−1) = 0
and thus that u is zero.
40 M. S. JOSHI

To construct a solution we use a duality argument and the Hahn-

Banach theorem. If u is a solution of the Cauchy problem and v ∈
C0∞ ((t, x) : t < T ) then
ZT Z T
∂v
hu, − + P ∗ vidt = hf, vidt + hφ(.), v(0, .)i.
∂t
0 0
The idea here is to extend the functional defined by the left hand
side on elements of the form L∗ v to the entire space of functions
L1 ((0, T ); H −s) by Hahn-Banach to obtain an element u of
  ∗
L1 (0, T ); H −s = L∞ ([0, T ]; H s)
which will then solve the problem.
To apply the Hahn-Banach theorem we will need to show that the
functional is well-defined and continuous. Now applying Prop 10.3 to
ṽ(t) = v(T ′ − t), with ’s = −s’ for any T ′ < T we have that
ZT ′
−λt
sup ke ṽ(t, .)k−s ≤ C e−λt kgk−s dt
t∈[0,T ′ ]
0
with
∂ṽ
g= + P ṽ.
∂t
So for any T ′ , kv(T − T ′ )k−s is dominated by the right hand side so
Z T
sup kvk−s ≤ C kgk−s dt.
t∈[0,T ] 0

This shows that L∗ is injective so we can define a map from L∗ (C0∞ ((t, x) :
t < T )) to C by
ZT
∗
ψ̃(L v) = hf (t, .), v(t, .)idt + hφ, v(0, .)i.
0
The estimate also shows that
 
ZT ZT
∗
kψ̃(L v)k ≤ C kφks + kf (t, .)ks dt kL∗ v(t, .)k−s dt
0 0

that is that ψ̃ is continuous on L∗ (C0∞ ((t, x) : t < T )) if one regards it

as a subspace of (L1 ((0, T ), H −s)) .
The Hahn-Banach says that one can extend ψ̃ to a continuous linear
functional on (L1 ((0, T ), H −s)) that is to an element of
L∞ ((0, T ), H s).
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 41

Call this element u and we have

Z ZT
∂v
hu, − + Q∗ vidt = hf (t, .), v(t, .)idt + hφ, v(0, .)i
∂t
0

for v ∈ C0∞ ((t, x) : t < T ). So certainly, on (0, T ) we have

∂u
− Qu = f
∂t
in a distributional sense. Now if we assume for now that f is Schwartz
then we have that
∂u
= Qu + f ∈ L∞ ((0, T ); H s−1).
∂t
So u is in C([0, T ]; H s−1) and then if we assume φ is Schwartz we have
integrating the equation again that
u ∈ C 1 ([0, T ]; H s−2)
and has the right initial value - φ.
For f, φ Schwartz the same argument would work with s replaced
by s + 2 so in that case we have a solution u′ which will equal u by
uniqueness in C 1 ([0, T ]; H s) and will satisfy the energy estimate (10.3).
To do the general case, we use the density of Schwartz space in H s
and L1 ((0, T ); H s). (Prove this!) Let fk → f in L1 ((0, T ); H s), fk
Schwartz and φk → φ, φk Schwartz. Then for each k we can use the
previously proven to case to get
uk ∈ C 1 ([0, T ]; H s ) ∩ C 0 ([0, T ]; H s+1)
solving with data (fk , φk ). But then applying (10.3) again, we deduce
that uk is Cauchy and that the limit is the desired solution u.

So why does all this allow us to prove the propagation of singu-

larities? We have reduced to the case where u ∈ E ′(Rn ) and P has
compactly supported kernel. We take Q(t, x, Dx ) commuting up to
smoothing with Dt − P as above and Q(0, x, Dx )u ∈ C0∞ . We also
assume that Q is uniformly properly supported.
As u is compactly supported, it must be in some Sobolev space - just
estimate
hu, e−ix.ξ i.
Since we have chosen Q to be a smooth family of pseudo-differential
operators, we have
Qu ∈ C ∞ ([0, T ]; H s)
42 M. S. JOSHI

and
(Dt − P )Qu ∈ C0∞ .
If we take s′ > s then the theorem says there is v satisfying the equa-
′
tion with the same initial value and v(t, .) ∈ H s . But by uniqueness
′
the two solutions agree so Qu(t, .) is in H s . But s′ was arbitrary so we
conclude that Qu(t, .) is smooth for each t which proves our result.

11. Operators on Manifolds

So having developed a theory of pseudo-differential operators on Rn

which is coordinate invariant, we can reduce the manifold case to Rn
by working locally.
Let X be a smooth manifold with an atlas {Uα , Vα , φα }.
Definition 11.1. The map P from C0∞ (X) to C ∞ (X) is a (classical)
pseudo-differential operator of order m if in any local coordinate patch
Uα , with ψ, φ ∈ C0∞ (Uα ) ψP φ defines a (classical) pseudo-differential
operator in local coordinates C0∞ (Uα ) → C0∞ (Uα ) of order m. We also
require that if φ, ψ are in C0∞ (X) with disjoint supports then φP ψ has
a smooth Schwartz kernel.

The second condition is a decree that our operator be pseudo-local.

Why do we need this? Well coordinate patches are local objects so they
say nothing about behaviour between points far away. For example if
one takes a sphere and define the map
Af (x) = f (−x)
then if one takes an atlas with small coordinate patches, it is zero locally
and hence obeys the first condition. But it is not a pseudo-differential
operator.
Since we previously proved coordinate invariance, if an operator
is pseudo-differential with respect to some atlas then it is pseudo-
differential with respect to any atlas.
We now want to extend our symbol calculus from Rn to X. The total
symbol is a nasty object - evidenced by our formulas for the change of
coordinates but the principal symbol is nice. Let SH m (T ∗ (X) − {0})
be smooth functions on T ∗ (X) − {0} which are positively homogeneous
of degree m. (Recall that the cotangent bundle is a vector bundle so
on each fibre there is a natural scalar multiplication.) To define the
principal symbol near a point x we take a coordinate chart (Uα , Vα , φα)
and a bump function ψ identically one near x and define σm (P )(x, ξ)
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 43

to be the principal symbol of φP φ in the local coordinates on the

cotangent bundle naturally induced by the local coordinates on the
manifold. We then have
Proposition 11.2. There is a natural short exact sequence,
0 → ΨDOclm−1 (X) → ΨDOclm(X) → SH m (T ∗ (X) − {0}) → 0.

The rest of our results we then take over from the Rn case.
Proposition 11.3. If P, Q are properly supported pseudo-differential
operators of order m, m′ then P Q is a pseudo-differential operator of
order m + m′ and
σm+m′ (P Q) = σm (P )σm′ (Q).

The asymptotic completeness of the calculus of pseudo-differential

operators will also hold.
Proposition 11.4. If Pj ∈ ΨDOclm−j (X) for j = 0 . . . ∞ then there
exists P ∈ ΨDOclm (X) such that
X
P− Pj ∈ ΨDOclm−N (X)
j<N

for all N.

Proof. Let fα be a partition of unity subordinate to some atlas (Uα , Vα , φα )

and let gα ∈ C0∞ (Uα ) be identically one on the support of fα . We then
have in local coordinates that
Pα,j = gα Pj fα ∈ ΨDOclm−j (Rn )
and we can asymptotically sum to get Pα . We then just put
X
P = g α Pα g α
and we are done. We use the cut-offs to ensure that the operators have
compact kernels contained in Uα × Uα . (we can assume the cover Uα is
locally finite.)

We define ellipticity as before - the principal symbol does not vanish.

We can then construct a parametrix for an elliptic operator - there are
two basic approaches. The first is work locally and patch together -
this is fiddly and unnecessarily so. The second is to use the principal
symbol map to get a Q ∈ ΨDOcl−m with
σ−m (Q) = σm (P )−1
so
P Q = I − R, R ∈ ΨDOcl−1 (X)
44 M. S. JOSHI

and then let X

Q′ ∼ Rj .
This will then be a parametrix. If one needed to, one could compute
the total symbol in any local coordinates in the same way as for Rn .
We can also define wavefront set as before.
Definition 11.5. If u ∈ D ′ (X) then
\
WF(u) = char(P ) ⊂ T ∗ (X) − {0}.
P u∈C ∞

We then have micro-locality, micro-ellipticity and so on. Propaga-

tion of singularities will also hold as its make invariant sense on the
cotangent bundle - our proof was local anyway.
The continuity on L2c (X) of properly supported operators will go over
since one can cover any compact set by a finite number of coordinate
patches and use a partition of unity.
So the theory for a general manifold is not very different from the
theory we have developed for Rn . However if we work on a compact
manifold, we can get some extra results - we can show that elliptic
operators are Fredholm either as operators on smooth functions or be-
tween Sobolev spaces. Note that elliptic regularity guarantees that the
kernel consists of smooth functions and thus that it will be independent
of which Sobolev space we work on.
We want to define Sobolev spaces on a compact manifold. We showed
previously that an element is in Hcs (Rn ) if and only if Ps u ∈ L2c (Rn )
for Ps elliptic of order s and properly supported. We also know that
C0∞ (Rn ) is dense in Hcs so we define for X a compact manifold, and
s > 0, that H s (X) is the completion of C ∞ (X) with respect to the
inner product
hu, vis = hu, viL2 + hPs u, Ps viL2 ,
where Ps is a self-adjoint elliptic pseudo-differential operator of order
s on X. We define for s < 0 by duality.
The first term is required in case Ps has non-trivial kernel (consisting
of smooth functions.) This definition makes H s a Hilbert space with
norm dependent on Ps but a different choice will give an equivalent
norm. Suppose we took Qs instead then Qs = R0 Ps + E with E
smoothing and R0 zeroth order and thus continuous on L2 . We then
have that the norm squared coming from Qs is
hu, uiL2 + hR0 Ps u, R0Ps uiL2 + 2ℜhEu, R0 Ps uiL2 + hEu, EuiL2
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 45

using the continuity of E, R0 on L2 and Cauchy-Schwartz this is less

than or equal to
C(kukL2 + kPs uk)2 .
By symmetry, we have that the norms are equivalent and thus a natural
topology.
Of course, we want that if P is order m then
P : H s → H s−m.
For zeroth order operators on L2 , this is trivial as above. We then write
P = Pm−s (Ps−m P P−s )Ps + R
with R order −∞. Provided s ≥ m this then immediately gives the
continuity. We deduce the case on negative Sobolev spaces by duality.
To stitch the two together, we can write R as a product of operators,
(up to smoothing at least) one mapping into L2 and the other mapping
out of it.
If u is supported in a coordinate patch, we have two ways to compute
its H s norm - its Sobolev norm in local coordinates and a norm on
the manifold. As the L2 norms are equivalent and Ps is continuous,
with continuous parametrix, on both sorts of Sobolev spaces, we see
that these are equivalent. This gives us an alternate definition of the
P
norm on the manifold, let φα be a finite partition of unity with φα
supported in a coordinate patch then let
X
kuks = kφα uks
α

where kφα uks is taken either in local coordinates or on the manifold as

they are equivalent anyway. This is equivalent to the original norm -
multiplication by φα is continuous as a zeroth order pseudo-differential
operator so kφα uk ≤ Ckuk and thus summing gives the inequality
P
one
way; the other way comes the triangle inequality applied to φα u.
The important thing about Sobolev spaces on a compact manifold
or in a compact set on Rn is Rellich’s compactness lemma:
Theorem 11.6. If t > s and X is a compact manifold then the inclu-
sion
i : H t (X) → H s (X)
is compact.

Proof. We first reduce to the compact set in Rn case. We need to show

a bounded sequence in H t (X) will have a convergent subsequence in
H s (X).
46 M. S. JOSHI

Let φi , i = 1 . . . k be a partition of unity subordinate to an atlas and

let un be a bounded sequence in H t (X). We then have that
φ1 u n
has a convergent subsequence from the Rn case, φ1 u1n . But then φ2 u1n
has one too and so on. The k th subsequence ukn will be such that φj ukn
is convergent for all k and thus is convergent from our result on the
equivalence of norms.
So we need only prove the result in Rn . Let un be a bounded sequence
supported in a compact set K ⊂ Rn . Now if φ is identically one on K
then φu = u for all u supported in K. So
!n
1
û = û ∗ φ̂
2π
and thus that !n
1
Dξα û = û ∗ Dξα φ̂.
2π
As
hξis ≤ Chξ − ηi|s|hηis ,
we deduce that
Z
s
|hξi Dξα û(ξ)| ≤ hηis |û(η)|hξ − ηi|s||D α φ̂(ξ − η)|dη

and thus, using Cauchy-Schwartz, is bounded by Cα kuks .

Applying this to the sequence un the first partial derivatives are
bounded uniformly, independent of n, on compact sets. Thus we have
that un is equicontinuous on compact sets and is a bounded sequence
on compact sets. So by the Arzela-Ascoli theorem there is a convergent
subsequence with respect to the uniform norm of the Fourier transform
on any compact set. Using diagonalization, we can get the same se-
quence for all compact sets - as Rn is a countable union of compact
sets.
Let vk be the resulting subsequence. We show that this is Cauchy in
H s and the result then follows.
We can write
kvk − vl k2s = I1 + I2
with Z
I1 = (1 + |ξ|2 )s |v̂k (ξ) − v̂l (ξ)|2dξ
|ξ|≤R
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 47

and Z
I2 = (1 + |ξ|2)s−t (1 + |ξ|)t|v̂k (ξ) − v̂l (ξ)|2dξ
|ξ|≥R

which is bounded by
(1 + R2 )s−t kuk − ul k2t
which will converge to zero as R → ∞ using the originial hypothesis
that the sequence is bounded in H t .
So given ǫ > 0, we pick R so that the second integral is less than ǫ/2
and then we can make the first integral arbritrarily small as we know
v̂k is uniformly convergent on |ξ| ≤ R.

Corollary 11.7. If P is a pseudo-differential operator of negative or-

der then P is a compact operator on any Sobolev space.

So given all this, we now have a real theorem,

Theorem 11.8. Let P be an elliptic operator of order m on a compact
manifold, M, then P is Fredholm as a map from H s (M) to H s−m (M).
The index of P is independent of s and depends only on the principal
symbol of P.

Proof. Let Q be a parametrix for P then P Q − I is order −∞ on

H s−m(M) and QP −I is of order −∞ on H s . Since operators of negative
order are compact, we deduce that P is invertible modulo compact
operators and thus is Fredholm. (Atkinson’s parametrix criterion.)
The index of P is the dimension of the kernel of P minus the dimen-
sion of the kernel of P ∗. By elliptic regularity, these spaces consist of
smooth functions which are thus in every Sobolev space and so will be
independent of s.
If P ′ = P + R with R of order m − 1 then R is compact as operator
from H s (M) to H s−m(M) as we can decompose into a continuous map
into H s−m+1 and the compact inclusion from there to H s−m(M). Since
a compact perturbation does not affect the index, this yields that P, P ′
have the same index and thus that the principal symbols determine the
index.

Corollary 11.9. P has finite dimensional kernel and cokernel as a

map on C ∞ (X).
48 M. S. JOSHI

Proof. The kernel on any Sobolev space is precisely the smooth func-
tions such P u = 0 so the kernel of P is finite dimensional. As the space
of smooth functions is in any Sobolev space we also have that an at
most finite dimensional subset of them are not in the image.

A self-adjoint elliptic operator will have index zero. We really need to

make more precise what we mean by this as there are many concepts
of self-adjointness. For a map on C0∞ (Rn ) there is a natural adjoint
obtained by taking the conjugate of the tranpose of the Schwartz kernel
- on a manifold we need to worry about a density term. So we say that
an operator is formally self-adjoint if
hP u, vi = hu, P vi
in the L2 pairing for all u, v ∈ C0∞ (X). Now if P is an elliptic pseudo-
differential operator of order m then we know
P : H s (X) → H s−m (X)
is Fredholm if X is a compact manifold. Its index is equal to
dim ker P − dim ker P ∗ .
But P ∗ is the adjoint as a map from
(H s−m )∗ → (H s )∗
that is from
H m−s → H −s .
The important point is that this map is the extension of P ∗ the formal
adjoint on C ∞ (X) so by elliptic regularity the kernel will be equal to
that of the formal adjoint on C ∞ (X) which for formally self-adjoint
operator is equal to that of P. Hence the index is zero.
Another important property of formally self-adjoint, elliptic opera-
tors on compact manifolds, is that they give a natural decomposition
of spaces. If we take v ∈ C ∞ (X) and w ∈ ker P (which is smooth so
independent of s) then
hP v, wiL2 = hv, P wiL2 = 0
So P (C ∞ (X)) is orthogonal in L2 to ker P but C ∞ (X) is dense in
H m (X) so we have that P (H m(X) is orthogonal to ker P. However the
codimension of P (H m(X)) is equal to the dimension of ker P so we
have that
L2 (X) = P (H m (X)) ⊕ ker P.
Invoking elliptic regularity again, the same is true for C ∞ (X). So to
summarize
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 49

Proposition 11.10. Let P be an elliptic, formally self-adjoint pseudo-

differential operator on X then P is of index zero and
C ∞ (X) = P (C ∞ (X)) ⊕ ker P.

Note that if two principal symbols were homotopic, we could with a

little effort construct a continuous family of operators quantizing them,
as index is constant on continuous families we could then deduce that
the principal symbols gave the same index and thus that index is really
a map from homotopy classes of principal symbols to Z - ie it is a
topological object. In fact, the index is determined by the homotopy
class as a map from the cosphere bundle - if we pick a Riemannian
metric this is the covectors of modulus one with respect to the dual
metric.
Proposition 11.11. Let X be a Riemannian manifold and m ∈ R.
There exists a self-adjoint operator Am of order m with principal symbol
|ξ|m
x.

Proof. Let B be an operator of order m/2 with symbol |ξ|m/2 x then

∗
A = B B has all the requisite properties. (with a smooth perturbation
one could actually make B invertible.)

Now a self-adjoint operator always has index zero so the index of

P is equal to that of A−m P. This means that the index of the princi-
pal symbol will be equal to that obtained by taking the zeroth order
operator with symbol defined by projection on to the cosphere bundle.
It is in fact a little tricky to construct a continuous family of operators
from a continuous family of principal symbols but in fact we can prove
homotopy type invariance without our doing so.
To see this note that if the pseudo-differential operators P, Q have
principal symbols such that
(1 − t)σm (P ) + tσm (Q)
is elliptic for 0 ≤ t ≤ 1 then the family
(1 − t)P + tQ
has the right principal symbols and thus P, Q have the same index. So
for a general family, using the fact that S ∗ (X) is compact, we divide
our homotopy into little pieces such that on each of these the straight
homotopy works and homotopy invariance follows.
50 M. S. JOSHI

In fact, index is a pretty boring object for scalar differential oper-

ators. On manifolds of dimension bigger than two all operators have
index zero. We won’t prove this as its topology not analysis - it follows
from the fact that all maps from S ∗ (X) to C − {0} are null homotopic
over coordinate patches and some use of Cech cohomology.
A consequence of this is that if an operator has trivial kernel then it
is bijective which is certainly not obvious.
If one considers systems of operators or more generally operators on
vector bundles then index is not necesarily zero.

12. Systems of Operators and Vector Bundles

We can consider a system of pseudo-differential operators on a man-

ifold X to be a matrix of scalar pseudo-differential operators.
Definition 12.1. A pseudo-differential operator of order m from
C ∞ (X; Ck ) → C ∞ (X; Cl )
is a map of the form
X
(P v)i = Pij vj
Pij a scalar pseudo-differential operator of order m.
The principal symbol of P is then just the the matrix of principal
symbols, σm (Pij ).

All the usual theorems go over and the matrix product of the prin-
cipal symbols is principal symbol of the product. We thus define
Definition 12.2. A system of pseudo-differential operators is elliptic
if the principal symbol is invertible everywhere.

And thus have

Proposition 12.3. An elliptic system of pseudo-differential operators
has a two-sided parametrix.

Proof. Standard proof take σ−m (Q) = σm (P )−1 and then proceed as
before.

Clearly an analagous theorem will hold if the principal symbol is left

(or right) invertible.
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 51

What we really want to do is consider operators between vector bun-

dles which is not a large step from systems. A vector bundle a natural
generalization of the situation where we have
π0 : X × Rk → X
with X a smooth manifold and π0 the projection onto X.
Recall that a vector bundle is a triple (E, X, π) with E a manifold
of dimension n + k, X a manifold of dimension n and π a projection
such that the projection is locally trivial. That is X can be covered by
open sets Uα and there are diffeomorphisms
fα : π −1 (Uα ) → Uα × Ck
with πfα = π0 and such that where two maps are defined, we have
fβ fα−1 (p, v) = (p, Gαβ (p)v)
with Gαβ a smooth family of invertible matrices. This means that there
is a natural linear structure on the fibres Ep = π −1 (p) and so they are
k dimensional vector spaces.
One can then all the things to vector bundles one can do to vector
spaces - e.g. tensor product V ⊗W, direct sum V ⊕W, homomorphisms
Hom(V, W ). One can also pull-back bundles by smooth maps. For a
discussion of these operations see Atiyah [2]. By a section of a vector
bundle E over X, we mean a smooth map f : X → E such that
πf = Id . That is to each point, we assign a point in the fibre above
it in a smoothly varying way, this class is denoted C ∞ (X, E). We can
define everything analagously for complex vector bundles.
We can define pseudo-differential operators on sections of vector-
bundles.
Definition 12.4. Let E and F be complex C ∞ vector bundles over the
manifold X then a (classical) pseudo-differential operator of order m
from sections of E to sections of F is a continuous linear map
A : C0∞ (X, E) → C ∞ (X, F )
such that for every open U ⊂ X where E, F are simultaneously triv-
ialized then with respect to the trivializations, A is given by a (clas-
sical) pseudo-differential system of order m. The class will be denoted
ΨDO m(X; E, F ).

It is of course enough to do this for a covering whilst assuming that

operators between trivializations with disjoint bases are smoothing. A
52 M. S. JOSHI

principal symbol is then defined before but will be a section of the pull-
back of Hom(E, F ) to T ∗ (X). Locally, this just means it is a smooth
matrix varying from point to point of T ∗ (X) and will be homogeneous
of order m in ξ.
One can now repeat the entire theory of pseudo-differential operators
for operators on sections of vector bundles and define Sobolev spaces
of sections and prove all the same results. We leave the details to
the enthusiastic reader. The important fact for us will be that an
elliptic operator - ie an operator with invertible principal symbol - will
be Fredholm as above. The index will then be an invariant of the
homotopy class of the map as a function from
S ∗ (X) → GL(V, W )
where GL(V, W ) is the bundle of invertible homomorphisms as a subset
of Hom(V, W ) above. Hence index theory!

13. The Hodge Theorem

So let’s apply all this theory to do some differential geometry. Recall

that over any smooth manifold, X, there are naturally the bundles of
differential forms, Ωk (X) and the natural operator d between them.
Over any coordinate patch these bundles have a natural trivial-
ization coming from the bases dxα = dxα1 ∧ dxα2 ∧ · · · ∧ dxαn with
α = (α1 , . . . αk ) with αj strictly increasing. That is locally a differental
k-form can be written as a sum
X
ω= fα (x)dxα
with α of this form. And we have that
X ∂fα
dω = dxj ∧ dxα .
α,j ∂xj

As ∧ is anti-symmetric on one-forms it follows that d2 = 0. What is

the symbol of d? Well, the term fα dxα is mapped to
X ∂fα
dxj ∧ dxα
j ∂xj
P
which has symbol given by mapping dxα → i ξj dxj ∧dxα . But a point
in the cotangent bundle is a one-form that is a sum
X
ξj dxj
so the symbol of d at (x, ξ) is just
µ 7→ iξ ∧ µ.
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 53

The interesting thing about this symbol is that the kernel of the
symbol of dk : Ωk → Ωk+1 is precisely the image of the previous one.
To see this, observe that since everything is defined invariantly, we can
pick coordinates such that the point (x, ξ) is such that ξ = (ξ1 , 0, . . . , 0)
saying something is in the kernel of the symbol is then just saying that
the coefficient of dxα is non-zero if and only if α1 = 1 but this is
precisely the same condition as being in the image.
With this in mind, we define
Definition 13.1. An elliptic complex (of order m) of vector bundles is
a sequence E0 , . . . , Ek over a compact manifold X together with pseudo-
differential operators Pj ∈ Ψm (X; Ej , Ej+1 ) such that Pj Pj−1 = 0 and
the image of the symbol of Pj is the kernel of the symbol of Pj+1 each
j with the first one injective and the last one surjective.
(ie the symbols form an exact sequence)

Note that if k = 1 then this says that the symbol is bijective every-
where that is the single operator is elliptic.
Provided each bundle is equipped with an inner product (which is
always possible), we can use this structure to give an elliptic operator
on each bundle. The inner product allows to define an adjoint operator
Pj∗ ∈ ΨDO m (X : Ej+1 , Ej ). Putting these together we obtain operators
on sections of Ej . We define
∆j = Pj∗ Pj + Pj−1Pj−1
∗
∈ ΨDO 2m(X; Ej , Ej )
for j = 0, . . . , k. These are the desired elliptic operators. If pj is the
symbol of Pj then the symbol of ∆j is dj = p∗j pj + pj−1 p∗j−1 . If dj v = 0
then
hdj v, vi = 0
and thus that
hpj v, pj vi + hp∗j−1 v, p∗j−1vi = 0.
This means that pj v = 0 which implies (by hypothesis) that v = pj−1 w
and
0 = hp∗j−1v, wi = hv, pj−1wi = hv, vi.
So v = 0 and the ellipticity follows. The operators ∆j then have elliptic,
two-sided parametrices Tj and are therefore Fredholm.
We can also use these parametrices to construct a parametrix for the
complex that is to construct operators Qj such that
Pj−1 Qj−1 + Qj Pj − I
54 M. S. JOSHI

is smoothing for each j. To see this observe that

Pj ∆j = ∆j+1 Pj
and therefore that
Pj Tj = Tj+1Pj
up to smoothing. Let Qj = Tj Pj∗ and we have up to smoothing that
∗
Pj−1Qj−1 + Qj Pj = Pj−1 Tj−1Pj−1 + Tj Pj∗ Pj
∗
= Tj (Pj−1Pj−1 + Pj∗ Pj )
=I

The Generalized Hodge Decomposition Theorem is now very easy:

Theorem 13.2. Suppose Pj , Ej , j = 1, . . . , k define an elliptic complex
on a compact manifold X then for all j
C ∞ (X, Ej ) = ker ∆j ⊕ Pj−1 C ∞ (Ej−1) ⊕ Pj∗ C ∞ (Ej+1 )
and
∗
ker ∆j = ker Pj ∩ ker Pj−1 .

Proof. To prove the last statement, we have

∗ ∗
0 = h∆j v, vi = hPj v, Pj vi + hPj−1 v, Pj−1 vi.

We know ∆j is an elliptic self-adjoint operator so we have that

C ∞ (M) = ker ∆j ⊕ ∆j (C ∞ (M)).
But ∆j = Pj∗ Pj + Pj−1Pj−1
∗
so the decomposition follows.
We need to check the three subspaces are mutually orthogonal to
complete the theorem. The orthogonality of the last two comes from
Pj Pj−1 = 0 as
hPj−1 v, Pj∗wi = hPj Pj−1 v, wi = 0.

The orthogonality of the image of Pj−1 with ker ∆j is just

∗
hPj−1 v, wi = hv, Pj−1 wi = 0
for
∗
w ∈ ker ∆j ⊂ ker Pj−1
and the remaining relation is similar.
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 55

We can apply this to the de Rham complex to deduce things about

the cohomology of a compact manifold. Recall that a j-form is closed
if and only it is in the kernel of
dj : C ∞ (X, Ωj (X)) → C ∞ (X, Ωj+1 (X)).
A j−form is exact if it is in the image of dj−1. We make similar def-
initions for co-closed and co-exact in terms of the adjoint. The fact
that d2 = 0 means that the exact forms are closed and the co-exact are
co-closed. A form that is closed and co-closed is said to be harmonic.
The theorem above says that any form is the sum of a harmonic form,
an exact form and a co-exact form.
Recall that the j-th de Rham cohomology of a manifold is defined to
be the space of closed j-forms quotiented by the space of exact j-forms.
We claim that this is naturally isomorphic to the space of harmonic j-
forms. This will follow trivially from Theorem 13.2 if we show that
the space of closed forms is equal to the space of exact forms plus the
space of harmonic forms. Now a form is harmonic if and only if it is
both closed and co-closed - same proof as ellipticity of ∆j . Using the
decomposition, it is thus enough to show that any non-zero co-exact
form is never closed, but
hdd∗ v, vi = hd∗ v, d∗vi
which can only be zero if d∗ v = 0 and the result follows. As ∆j is Fred-
holm, we have that the space of harmonic forms is finite-dimensional so
the de Rham cohomology of a compact manifold is finite dimensional
- a non-obvious fact!
NB there is a slight flaw in the above in that it was done for complex
rather than real forms but we can commute taking real parts through
the final results to prove that they hold in the real case too.
We proved all this while using any inner product on each form bundle
- if one is more careful and defines all the inner products using the
Riemannian structure of the tangent bundle then one can also use this
to prove Poincare duality. To do this we have to use several natural
pairings which are at our disposal.
First there is a natural pairing between the exterior algebra of a
vector space and the exterior algebra of its dual.
Λj (V ) × Λj (V ∗ ) → R
Defined on basis elements by α = µ1 ∧ · · · ∧ µk , β ∗ = f1 ∧ · · · ∧ fk by
(α, β ∗ ) = det(fi (µj )). (9)
56 M. S. JOSHI

This can be extended by linearity and by considering the pairing the

wedge products of an element of a basis and the wedge products of the
associated dual basis, it is non-singular. It is also intrinsic. So we have
that
Λj (V ∗ ) ∼ (Λj (V ))∗ .
If we pick a volume element, that is a non-zero element ω of Λn (V ),
then we have a pairing
Λj (V ) × Λn−j (V ) → R
(α, β) 7→ c(α, β)
where
α ∧ β = c(α, β)ω.
This is non-singular also. So given a volume form we have
Λj (V ) ∼ (Λn−j (V ))∗ . (10)
So combining (9) and (10), we have
Λj (V ) ∼ Λn−j (V ∗ ). (11)

However, we want an isomorphism with Λn−j (V ) rather than Λn−j (V ∗ ).

To do this, we assume we have an inner product, B, on V and thus a
natural isomorphism, LB , from V to V ∗ . Because the exterior algebra
is a natural object this means that we have an isomorphism
L−1
B : Λ
n−j
(V ∗ ) ∼ Λn−j (V ).
Note also that once we have chosen an orientation of V, the bilinear
form can be used to fix the choice of volume form - an orientation is a
choice of volume form up to a positive multiple. We define an oriented
basis of V to be a basis, el , such that e1 ∧ · · · ∧ en is in the orientation
class. Any orthonormal oriented basis will then give the same volume
form.
So having chosen an inner product and an orientation, we define
∗ : Λj (V ) → Λn−j (V )
to be the given isomorphism from Λj (V ) to Λn−j (V ) followed by L−1 B .
Tracing this through on any oriented orthonormal basis, v1 , . . . , vn , we
have
∗ (v1 ∧ . . . vj ) = vj+1 ∧ · · · ∧ vn , (12)
which allows us to compute ∗ on any element and we have that
∗∗ = (−1)j(n−j) Id
on elements of order j.
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 57

We can now define ( yet another ) pairing,

(α, β) = α ∧ ∗β = c(α, β)ωB (13)
where ωB is the volume form coming from the inner product.
We now want to do all this globally. Suppose we have an oriented,
compact, Riemannian manifold, X. Then the above construction will
work with V = Tp∗ (M) for any p ∈ M. So we have an inner product
on each form bundle. What is the adjoint of d, δ, with respect to this
inner product on the j-th form bundle?
δ = d∗ = (−1)j+1+j(n−j) ∗ d ∗ .
To see this observe that by Stokes’ theorem
Z
d(α ∧ ∗β) = 0

and compute d(α ∧ ∗β).

What does all this do for us? Well since
∆ = dδ + δd
it commutes with ∗ up to sign, which means that ∗ is a bijection from
the space of harmonic forms in dimension j to that in dimension n − j,
ie Poincare duality:
H n−j (X) = H j (X).

It is also worth mentioning that the isomorphism of de Rham co-

homology with the space of harmonic forms means that by making
assumptions on the metric one can prove things about the cohomology.

14. Exercises

Question 1. Construct a function, φ, with the following properties:

1. φ is a smooth, real-valued function on Rn ,
2. φ = 1 in |x| ≤ 1,
3. φ = 0 in |x| ≥ 2,
4. φ ≥ 0.
Use the function

0 x ≤ 0,
f (x) = − x1
e x > 0.
58 M. S. JOSHI

Question 2. Show C0∞ (Rn ) is dense in S(Rn ) in the sense that given
φ ∈ S(Rn ) there exists a sequence φn ∈ C0∞ (Rn ) such that
||φ − φn ||α,β → 0, ∀α, β.
Deduce that if u ∈ S ′ (Rn ) is zero on C0∞ (Rn ) then it is zero on S(Rn ).
Observe that S ′ (Rn ) then naturally forms a subspace of D ′ (Rn ).

Question 3. (Partitions of Unity) Suppose Uα is a collection of open

subsets of Rn such that ∪Uα = Rn and U¯α is compact for each α and any
infinite intersection of Uα is empty. Show that there exists a collection
of smooth functions φα such that supp(φα ) ⊂ Uα and
X
φα = 1.
α

If you know about paracompactness, what can one do when we do not

assume the intersections are locally finite.

Question 4. Suppose u ∈ D ′ (Rn ) and supp(u) is compact. Show that

the pairing
< u, e−ix.ξ >
makes sense for any ξ ∈ Cn and that it results in a function which is
smooth in ξ. Show also that it satisifies the Cauchy-Riemann equations
in each ξj .

Question 5. Euler’s Relation

Suppose f is a smooth function on Rn −{0} and that it is (positively)
homogeneous of order m. ie
f (λx) = λm f (x)
for all λ > 0. Show that
!
∂
x − m f (x) = 0.
∂x
Prove the converse.
Show that there is a one-one correspondence between the set of ho-
mogeneous functions of order m and the smooth functions on S n−1 .
Suppose f is smooth on Rn , show that f is a homogeneous polyno-
mial.
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 59

Question 6. Suppose f (x, ξ) is a smooth function on Rnx × (Rkξ − {0})

which is homogeneous of degree µ ∈ C in ξ and φ is as in Q1. Show
that
f (x, ξ)(1 − φ)(ξ)
is a symbol of order ℜµ.

Question 7. The Borel Lemma

The point of this question is to show that given any sequence {cn }
there exists a function whose Taylor series is {cn }. i.e. there exists a
smooth function f ∈ C ∞ (R) such that
∂xj f (0) = j!cj .
Let φ be as in Q1.
1. Show that if tj → ∞ then the sum
∞
X
φ(tj x)cj xj
j=0

converges for each x.

2. Show that if {tj } is picked correctly then the sum will converge
uniformly to a continuous function.
3. Show there exists {tkj } such that the sum of the k th derivatives
converges uniformly .
4. Use a diagonalization argument to show that the same {tj } can
be used for all k and deduce that then
X
cj φ(tj x)xj
is the desired function.
∞
P
5. Need the sum cj xj exist for any non-zero x ?
j=0

Question 8. Asymptotic Summation Show that if aj ∈ S m−j (Rn ; Rk )

then there exists a ∈ S m (Rn ; Rk ) such that
N
X
a− aj ∈ S m−N for all N.
j=0

Repeat the proof of the Borel Lemma but do it about infinity instead
of about 0.

Question 9. Find a formula for the total symbol of the composite of

two differential operators in terms of their total symbols. Identify a
formula for the term of each homogeneity. (reduce to simple cases)
60 M. S. JOSHI

Question 10. Define a distribution, u on Rn , to be homogeneous of

order m if !
∂
x − m u = 0.
∂x
Observe that a homogeneous distribution is tempered and show that
its Fourier transform is also homogeneous.
n
P
∂
NB x ∂x = xj ∂x∂ j
j=1

For the brave - when can a homogeneous smooth function on Rn − 0

be extended to a homogeneous distribution on Rn ?

Question 11. Suppose a ∈ S m (Rnx ; Rkθ ). Show that a ∈ Sclm then

NY
−1
!
∂
θ − (m − j) a(x, θ) ∈ S m−N ∀N.
j=0 ∂θ

Question 12. Show that if P is a pseudo-differential operator and Q is

smoothing and at least one is proper then P Q, QP is smoothing.

Question 13. If P is a properly supported operator with Schwartz ker-

nel K and defines maps from D ′ (Rn ) → D ′ (Rn ) and from C ∞ (Rn ) →
C ∞ (Rn ), show that
singsupp(P u) ⊂ singsupp(K) ◦ singsupp(u).

Decompose P and u into pieces supported near and far from their
singular supports and then use the relation for ordinary supports.

Question 14. If P is a classical pseudo-differential operator with posi-

tive principal symbol, use a symbolic argument to show there exists Q
a classical, pseudo-differential operator such that
P − Q2 ∈ ΨDO −∞.

Question 15. Show that if a ∈ S m is classical then its asymptotic ex-

pansion in homogeneous terms is unique.

Question 16. The Weyl Quantization

Show that if P ∈ ΨDOclm(Rn ) then P can be written in the form
Z  
i<x−y,ξ> x+y
e a , ξ dξ
2
up to smooth terms for some classical symbol a and express a in terms
of the left symbol.
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 61

Question 17. Exhibit a distribution, u, satisfying

 
Dt2 − Dx2 u(x, t) = 0
which is not smooth.

Question 18. Classify the solutions of Dx1 u = 0 on Rn and classify the

possible singular supports.

Question 19. Give an example of a Schwartz kernel which maps C0∞

into D ′ but not C ∞ .

Question 20. If P is a pseudo-differential operator on Rn then

P − P∗
is smoothing if and only if the Weyl symbol of P is real. (see last
problem sheet.)

Question 21. In what Sobolev classes are the following? give the an-
swer for both the H s and Hloc
s

n
1. δ(x) on R

1 x ≤ 0
2. H(x) = 
0 x<0
3. 1/x on R (defined by principal values)
4. the characteristic function of the unit disc in R2 .
5. The distribution obtained by integrating a function over the unit
sphere.

Question 22. Compute the wavefront set of each of the distributions in

Q21//. Let g ij (x) be a positive definite matrix smoothly varying with
x and define X
2 = Dt2 − g ij (x)Dxi Dxj .
i,j
Suppose 2u is smooth, give an upper bound on WF(u).
m
Question 23. If Pj ∈ ΨDOcl j show that
[P, Q] = P Q − QP ∈ ΨDOclm1 +m2 −1
and compute the principal symbol. Show that principal symbols form
a Lie Algebra under this operation. (if you know what a Lie algebra
is.)
62 M. S. JOSHI

Question 24. Try to prove the coordinate invariance of Hcs directly.

If pm (x, ξ) is a homogeneous function on T ∗ (Rn ) − {0}, define the
Hamiltonian of pm to be the vector field
X ∂pm ∂ ∂pm ∂
− .
∂ξj ∂xj ∂xj ∂ξj
Show that the Hamiltonian is coordinate invariant. Show also that the
integral curves of the Hamiltonian preserve the set pm = 0. Show that
multiplying pm by a non-zero function does not change the integral
curves within pm = 0.

Question 25. Construct a distribution on Rn of which the wavefront

set is a single ray.

Question 26. What is the wavefront set of H(x1 )δ(x′′ ) on Rn ?

Question 27. Prove the continuity of zeroth order non-classical pseudo-

differential operators on L2c .

Question 28. If P ∈ ΨDOclm with positive principal symbol prove that

P has an mth root up to smoothing operators.

Question 29. Show that the intersection of the characteristic varieties

of vector fields tangent to a submanifold is equal to the submanifold’s
conormal bundle. Deduce an upper bound for the wavefront set of a
pseudo-differential operator.

Question 30. Suppose P is a pseudo-differential operator on Rn , define

P ′ : Cc∞ (Rn × R) → C ∞ (Rn × R)
by treating the extra variable as a parameter. Show P ′ is a pseudo-
differential operator if and only if P is a differential operator.

Question 31. We shall say u ∈ D ′ (X) is in Hlocs

at (x0 , ξ0 ) if there exists
a zeroth order pseudo-differential operator micro-elliptic at (x0 , ξ0 ) such
that
s
P u ∈ Hloc (X).
s s
Show that if u ∈ Hloc at every point (x0 , ξ0 ) then u ∈ Hloc .
s
Show also that if Q is order m and u ∈ Hloc at (x0 , ξ0 ) then Qu ∈
s−m
Hloc at (x0 , ξ0 ). Deduce that if Q is micro-elliptic at (x0 , ξ0) and Qu ∈
s s+m
Hloc at (x0 , ξ0) then u ∈ Hloc at (x0 , ξ0 ).
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 63

s
Define su (x, ξ) to the supremum of s such that u ∈ Hloc at (x, ξ). Is
su continuous or semi-continuous?
Does su (x, ξ) = +∞ imply that (x, ξ) 6∈ WF(u) ?

Question 32. Show that if V is a vector bundle then there exists a

smooth metric on V ie an inner product on each fibre which varies
smoothly from point to point.

Question 33. Show that there is natural pairing

Γ(|Λ|s , X) × Γ(|Λ|1−s , X) → C
given by integration.

Question 34. If Λn is the line bundle of n-forms on a manifolds, show

there is a natural trivilization over each coordinate patch and compute
the transition functions.

Question 35. Let 2 = Dt2 − ∆ be the flat wave equation on R × R2 .

Let P be the parabola y = x2 and let u be a distribution such that
2u = 0.
Suppose
WF(u) ∩ {t = 0} ⊂ {(x, y, ξ, η, |ξ, η|) : (x, y, ξ, η) ∈ N ∗ (P )}.
Compute the wavefront set of u and sketch the projection of WF(u)
onto (x, y) for t at important times.

Question 36. Show that if 2u is smooth and singsupp(u) is compactly

supported then u is smooth.

Question 37. Give an example of a differential operator which has pe-

riodic bicharacteristics.

Question 38. * Suppose P is a zeroth order elliptic, pseudo-differential

operator on the circle with principal symbol equal to

a
+ (x) ξ > 0
σ0 (P ) = 
a− (x) ξ < 0.
Compute the index of P in terms of the winding numbers of a± as
maps from S 1 to C − {0}.
Deduce that the result holds for P of any order.
64 M. S. JOSHI

Question 39. If ω ∈ S n−1 show that the distribution

δ(t − x.ω)
is a solution of
2u = 0
and compute its wavefront set.

Question 40. Are the left and right shift operators on l2 (N) Fredholm?
If so what are their indexes?

Question 41. Let M be a subset of ΨDOcl1 and define

I (s) (M) = {u : P1 . . . Pk u ∈ Hloc
s
, ∀k, Pj ∈ M}.
Show that if u ∈ I (s) (M) then
\
WF(u) ⊂ σ1 (P )−1 (0).
P ∈M

Show that if M is the smallest subset of ΨDOcl1 containing M which

is closed under summation, multiplication by zeroth order operators
and the taking of commutators then
I (s) (M) = I (s) (M).

m
Question 42. Let Sρ,δ be the class of symbols a such that
|Dxβ Dξα a(x, ξ)| ≤ Cα,β,K < ξ >m−|α|ρ+|β|δ
for x in K compact with ρ, δ between 0 and 1. For which ρ, δ does
the theory of pseudo-differential operators on Rn work? For which ρ, δ
does the theory work on manifolds?

Question 43. Let X be a compact manifold and suppose {Γj } is a

collection of open cones in T ∗ (X) − 0 which form an open cover. Show
there exists a collection of zeroth order pseudo-differential, operators,
{Pj }, such that
X
Pj = Id
j
and
WF′ (Pj ) ⊂ Γj .
Find out what a sheaf is and deduce that the space of distributions
on X modulo the smooth functions form a sheaf over the cosphere
bundle.
 INTRODUCTION TO PSEUDO-DIFFERENTIAL OPERATORS 65
′
Question 44. Show that if aj ∈ S m−j and a ∈ S m are such that there
exists a sequence Nj such that Nj → −∞ and
X
a− aj ∈ S Nj
j<N
m
then a ∈ S and X
a− aj ∈ S m−j
j<N

References
[1] S. Alinhac et P. Gerard, Operateurs Pseudo-differentiels et le theoreme de
Nash-Moser, CNRS 1991, ISBN 2-222-04534-7
[2] M. F. Atiyah, K-Theory
[3] Chazarain and Piriou, Introduction to the Theory of Linear Partial Differential
Operators, North Holland out of print
[4] Melrose, Notes on Micro-local Analysis, available by gopher from ftp-math-
[Link]
[5] G. Friedlander, M. Joshi, Introduction to the Theory of Distributions, C.U.P.,
1999
[6] Shubin, Pseudo-differential Operators and Spectral Theory, Springer Verlag
1987, ISBN 0-387-13621-5