Global Analysis

Dexter Chua

1 Introduction 1

2 Differential Operators 1

3 Sobolev Spaces 4

4 Elliptic Regularity 9

5 Hodge Theory 12

Appendix A Proof of local elliptic regularity 14

1 Introduction
The document is intended to lead to a proof of the Hodge decomposition theorem
and spectral theorem of elliptic operators. The reader is assumed to be familiar with
analysis and the arguments are not very detailed. The most glaring omission is the
proof of (local) elliptic regularity, which is only very briefly sketched.
A lot of the material is based on lecture notes by Johannes Ebert, which can be
found at [Link]

2 Differential Operators
Fix a manifold M , and Hermitian vector bundles E0 , E1 → M with inner products.
Definition. A differential operator L : Γ(M, E0 ) → Γ(M, E1 ) of order k is a C-linear
map that is local, i.e. the value of Lu near a point p ∈ M depends only on the values
of u near p, and in a coordinate chart, it is of the form
X
Lf = Aα (x)Dα f
|α|≤k

for some Aα ∈ Γ(Hom(E0 , E1 )).

1
 Note that under our definition, any differential operator of order k is also a
differential operator of order k + 1.
Just as we can define tangent vectors as derivations, we have the following
coordinate-free definition of differential operators (which we will not use), with a
similar proof:

Fact. Let L : Γ(M, E0 ) → Γ(M, E1 ) be linear. Then

1. L is a differential operator of order 0 iff [L, f ] = 0 for all f ∈ C ∞ (M ).
2. L is a differential operator of order k iff [L, f ] is a differential operator of order
k − 1 for all f ∈ C ∞ (M ).

Integration by parts implies that we have

Lemma. For any differential operator L : Γ(M, E0 ) → Γ(M, E1 ), there is a formal
adjoint L∗ : Γ(M, E1 ) → Γ(M, E0 ) such that for any fi ∈ Γ(M, Ei ), we have

(Lf0 , f1 )L2 = (f0 , L∗ f1 )L2 .

Definition. Let L be a differential operator of order k. The (principal) symbol of

L is the family of operators symk (L)(x, ξ) ∈ Hom((E0 )x , (E1 )x ) for (x, ξ) ∈ T ∗ M
given locally by X
symk (L)(x, ξ) = Aα (x)ξα .
|α|=k

Formally, if π : T ∗ M → M is the projection, then symk (L) ∈ Γ(Hom(π ∗ E0 , π∗ E1 )).

In a coordinate-free manner, if s ∈ Γ(M, E0 ) and f ∈ C ∞ (M ) with f (x) = 0,
then
1
symk (L)(x, (df )x )(s(x)) = L(f k s)(x).
k!
We say L is elliptic at x ∈ M if symk (x, ξ) is invertible for all ξ ∈ Tx∗ M \ {0}, and L
is elliptic if it is elliptic everywhere.
While the coordinate-free definition seems rather artificial, it is actually useful
when we want to do computations later on.
It will be convenient to note that the adjoint of an elliptic operator is elliptic.
More generally,
Lemma. For any operators L, L0 , we have

symk (L∗ )(x, ξ) = ±(symk (L)(x, ξ))∗

symk (L ◦ L0 )(x, ξ) = symk (L)(x, ξ) ◦ symk (L0 )(x, ξ).

Hence the composition and adjoints of elliptic operators is elliptic.

2
Example. Consider the exterior derivative d : Ωp (M ) → Ωp+1 (M ). Using the
coordinate-free definition, we compute the symbol as
sym1 (d)(x, (df )x )(ωx ) = (d(f ω))x = (df ∧ ω)x
whenever f (x) = 0. So the symbol of d is
(ξ, ω) 7→ ξ ∧ ω.
Note that for p > 0, this is not invertible. Instead, what we have is an elliptic
complex.
Definition. An elliptic complex is a sequence of vector bundles E0 , . . . , Em with
first-order differential operators
Li : Γ(M, Ei−1 ) → Γ(M, Ei )
such that Li+1 ◦ Li = 0 and for any non-zero ξ ∈ Tx∗ M , the sequence
sym1 (L1 ) sym1 (L2 )
0 π∗ E0 π ∗ E0 ··· π ∗ Em 0
is exact outside of the zero section of T ∗ M .
Exercise. The de Rham complex and Dolbeault complex are elliptic complexes.
Ultimately, we will prove Hodge decomposition for elliptic complexes, which
subsumes the Hodge decomposition of Riemannian and Kähler manifolds.
To get from an elliptic complex to an elliptic operator, we use the following linear
algebraic result:
Lemma. Let {Vi } be finite-dimensional vector spaces, and
f1 f2
0 V0 V1 ··· Vm → 0
Vi . Then f + f ∗ : V → V is an isomorphism.
L
be an exact sequence. Let V =
Proof. It suffices to show f + f ∗ is injective. Suppose (f + f ∗ )x = 0. Then
(f + f ∗ )2 x = (f f ∗ + f ∗ f )x = 0.
So we get
0 = hf f ∗ x, xi + hf ∗ f xi = hf ∗ x, f ∗ xi + hf x, f xi.
So f x = f ∗ x = 0. By exactness, x = f y for some y, and then
0 = hf ∗ f y, yi = hf y, f yi = hx, xi.
So x = 0.
L
Corollary. If (E∗ , L∗ ) is an elliptic complex, define E = Ei . Then
∗
D = L + L : Γ(M, E) → Γ(M, E)
is elliptic.

3
3 Sobolev Spaces
3.1 Local Sobolev spaces
To apply functional analytic techniques to PDE problems, we need an appropriate
Hilbert space of functions. The right notion is that of a Sobolev space.
Definition. Let U ⊆ Rn be an open set, s ∈ R (possibly negative) and f, g ∈ Cc∞ (U ).
We define
Z
(f, g)s = (1 + |ξ|2 )s fˆ(ξ)ĝ(ξ) dξ
Rn
kf ks = (f, f )s

We define H s (U ) to be the Hilbert space completion of Cc∞ (U ) under k · ks .

If no confusion arises, we will omit the (U ) in H s (U ).
This space is actually usually called H0s (U ) instead, with H s (U ) reserved for the
same definition but without the compactly supported requirement. We will only
need the compactly supported version, and will not write the 0 .
Basic properties of Fourier transforms imply that for s ∈ Z>0 , we have the
following more intuitive definition:
Lemma. If s is a non-negative integer, and f ∈ Cc∞ (U ), then k · ks is equivalent to
the norm X
kf k0s = kDα f kL2 .
|α|≤s

Lemma. For any s, Dα extends to a continuous map H s → H s−|α| .

A priori, the above definition does not let us think of elements of H s as genuine
functions. Thankfully, we have the following result:

Lemma.
1. The natural map Cc∞ (U ) ,→ L2 induces an isomorphism H 0 ∼
= L2 .
2. If s ≥ t and f ∈ Cc∞ , then kf kt ≤ kf ks . Hence there is a continuous map
H s → H t.

3. The map is injective.

Proof. Only (3) requires proof. We have to show that if fn ∈ Cc∞ is k · ks -Cauchy
and kfn kt → 0, then kfn ks → 0. By assumption, we know fˆn2 (1 + |ξ|2 )t → 0 in L1 .
So it converges to 0 almost everywhere. So fˆn2 (1 + |ξ|2 )s → 0 almost everywhere.
But we know it is Cauchy in L1 . So it converges to 0 in L1 .

4
 This lets us view the H s as nested subspaces of L2 . In fact, we get something
better.
n
Theorem (Sobolev embedding theorem). Let k an integer such that s > k + 2.
Then there exists a constant A such that for any u ∈ C ∞ (U ), we have

kukC k ≤ Akuks .

Thus, there is a continuous inclusion H s ,→ C k .

Proof. Let |α| = ` < k. For x ∈ U and u ∈ Cc∞ (U ), we have
Z  Z Z
|Dα u(x)| ∝  e ξα û(ξ) dξ  ≤ |ξα ||û(ξ)| dξ ≤ |ξ|` |û(ξ)| dξ
ixξ
 
 
Z 1/2 Z 1/2
≤ |ξ|2 (1 + |ξ|2 )−s dξ |û(ξ)|2 (1 + |ξ|2 )s dξ .

The second term is exactly kuks , and the first term is a constant, which by elementary
calculus, is finite iff s > ` + n2 .
Admittedly, the proof above does not make much sense. The following more
direct proof of a special case shows why we should expect a result along these lines
to be true:
Proof of special case. We consider the case r = 1, U = (0, 1) and k = 0. Then we
want to show that
kukC 0 ≤ Akuk1
for some constant A. In other words, if u ∈ H 1 , we want to be able to make sense of
u as a continuous function. The trick is to notice we can make sense of du
dx as an L
2

function, and hence we can write

Z t
du
u(t) = dx + C
0 dx

for some integration constant C. This can be determined by

Z 1 Z 1Z t
du
C= u(x) dx − dx dt.
0 0 0 dx

du
This lets us control sup u in terms of integrals of u and dx .

One can in fact come up with a similar proof as long as s is an integer.

Another crucial theorem is the following:

5
Theorem (Rellich compactness theorem). If U ⊆ Rn is precompact and s > t, then
the natural map H s → H t is compact.
Proof. Let uk ∈ Cc∞ (U ) be such that kuk ks ≤ 1. We have to produce a subsequence
that converges in H t . The key claim is
Claim. ûk that equicontinuous and uniformly bounded on compact sets.
Assuming this claim, Arzelá–Ascoli implies there is a subsequence of ûk that is
uniformly convergent on compact subsets. Thus, we write
Z Z !
kuk − u` k2t = + |ûk − û` |2 (1 + |ξ|2 )t dξ.
|ξ|≤R |ξ|>R

For any fixed R, the first term vanishes in the limit k, ` → ∞ by uniform convergence.
For the second term, we bound

(1 + |ξ|2 )t ≤ (1 + |ξ|2 )s (1 + R2 )t−s .

So we know that
Z
|ûk − û` |2 (1 + |ξ|2 )t dξ ≤ (1 + R2 )t−s (kuk k2s + ku` k2s ) ≤ 2(1 + R2 )t−s .
|ξ|>R

This can be made small with large R, and we are done.

To prove the claim, we use the following trick — pick a bump function a ∈ Cc∞ (Rn )
such that a|U ≡ 1. Then trivially, uk = auk , and thus we have

ûk = â ∗ ûk .

Then controlling ûk would be the same as controlling â, which is fixed. For example,
to show that ûk is bounded, we write
Z Z 1/2
2 2 −s
|ûk | = |â ∗ ûk | ≤ |â(ξ − η)ûk (η)| dη ≤ |â(ξ − η)| (1 + |η| ) dη kuk ks

by Cauchy–Schwarz. Since â is a Schwarz function, the first factor is finite and

depends continuously on ξ. So ûk is uniformly bounded on compact subsets. Equicon-
tinuity follows from similar bounds on Dj ûk = (Dj â) ∗ ûk .
If we are lazy, we will often restrict to the case where s ∈ Z. If further s ≥ 0,
then the definition of the Sobolev norm in terms of derivatives can be very useful.
For s < 0, we will exploit the following duality:

6
Lemma. The pairing
Cc∞ × Cc∞ → C
Z
(f, g) 7→ f (x)g(x) dx.

satisfies
|(f, g)| ≤ kf ks kgk−s .
Thus, it extends to a map H × H −s → C. Moreover,
s

|(f, g)|
kf ks = sup . (†)
g6=0 kgk−s
Thus, this pairing exhibits H s and H −s as duals of each other.
Note that this duality pairing is “canonical”, and doesn’t really depend on the
Hilbert space structure on H s (while the canonical isomorphism between H s and
(H s )∗ does). Later, we will see that in the global case, the Hilbert space structure is
not quite canonical, but the duality pairing here will still be canonical.
Proof. The first inequality follows from Plancherel and Cauchy–Schwarz
Z
ˆ
(f, g) = (f , ĝ) = fˆ(ξ)(1 + |ξ|2 )s/2 ĝ(ξ)(1 + |ξ|2 )−s/2 dξ.

To show (†), one direction of the inequality is clear. For the other, if f ∈ C ∞ , take
ĝ = fˆ(1 + |ξ|2 )s . For general f , approximate.
A sample application of this is to show that differential operators induce maps
between Sobolev spaces. We already said that Dα tautologically induces a map
H s → H s−|α| . In a general differential operator, we differentiate, then multiply by a
smooth function. Thus, we want to show that multiplication by a smooth function is
a bounded linear operator. This is easy to show for s ∈ Z≥0 , and duality implies the
result for negative s ∈ Z.
Lemma. Let s ∈ Z, a ∈ Cc∞ (R) and u ∈ H s (U ). Then
kauks ≤ CkakC |s| kuks .
for some constant C independent of a. Thus, if L is a differential operator of compact
support of order k, then it extends to a map L : H s+k → H s .
Proof. We only have to check this for u ∈ Cc∞ (U ). For s ≥ 0, this is straightforward,
since kauk2s is a sum of terms of the form kDα (au)k20 for |α| ≤ k, and the product
rule together with the bound kauk0 ≤ kakC 0 kuk0 implies the result.
For negative s, if s > 0, then
|(au, v)| |(u, āv)| kuk−s kāvks
kauk−s = sup = sup ≤ sup ≤ kuk−s kakC s .
v6=0 kvks kvks kvks

7
 When proving elliptic regularity, we will need a slight refinement of the above
result.
Lemma. Let s ∈ Z, a ∈ Cc∞ (R) and u ∈ H s (U ). Then

kauks ≤ kakC 0 kuks + CkakC |s|+1 kuks−1

for some constant C independent of a.

Proof. Again the case s > 0 is straightforward. If we look at a term kDα (au)k0 , the
terms in the product rule where |α| = k and the derivatives all hit u contribute to
the firs term, and the remainder go into the second.
For s < 0, we proceed by downward induction on s, using the isometry H s+2 → H s
given by extending 1 + ∆. This is an isomorphism, because in the Fourier world,
this is just multiplication by (1 + |ξ|2 ). Thus, we check the claim for u of the form
Lv. We have

kaLvks ≤ k[a, L]vks + kL(av)ks ≤ k[a, L]vks + kavks+2 .

The second term is bounded, by induction, by

kakC 0 kvks+2 + CkakC |s+2|+1 kvks+1 = kakC 0 kLvks + CkakC |s+2|+1 kLvks−1

Observing that |s + 2| ≤ |s| since s < 0, we are happy with this term. To bound the
first term, we have
[a, L]v = [a, ∆]v = −∇a · ∇v.
So we get a bound (omitting constant multiples)

k[a, L]vks = k∇a · ∇vks ≤ k∇akC |s| k∇vks ≤ kakC |s|+1 kvks+1 = kakC |s|+1 kLvks−1 .

3.2 Global Sobolev spaces

To define Sobolev spaces on manifolds, first observe the following lemma:
Lemma. Let U 0 , V 0 ⊆ Rn be open and φ : U 0 → V 0 a diffeomorphism. If U ⊆ U 0
is precompact and V = φ(U ) ⊆ V 0 , then the induced map H s (V ) → H s (U ) is a
bounded isomorphism for all s ∈ Z.
The proof is an exercise in the chain rule (for s ≥ 0) and duality (for s < 0).
From now on, we always assume s is an integer.
This means it makes sense to define

8
Definition. Let M be a compact manifold and E → M a Hermitian vector bundle.
∼
Fix an open cover {ϕi : Rn → Ui ⊆ M } and a partition of unity {ρi } subordinate to
ϕi . For any u ∈ Γ(M, E), we define the Sobolev norm by
X
kuk2s = kρi u ◦ ϕ−1 2
i ks ,
i

and define H s (M ; E) to be the completion of Γ(M, E) with respect to k · ksk .

There is an obvious inner product that gives rise to the Sobolev norm.
One should check for themselves that

Lemma. The norm k · ks is well-defined up to equivalence, i.e. does not depend on

the choice of the Ui , φi , ρi .
One sees that our local theorems generalize easily to
Lemma.

1. There are bounded inclusions H s → H t for s > t which are compact.

n
2. There are bounded inclusions H s → C k for s > 2 + k.
3. There is a natural duality pairing H s (M ; E) × H −s (M ; E) → C.

4. Any differential operator L : Γ(M, E0 ) → Γ(M, E1 ) of order k induces a

continuous map H s+k (E0 ) → H s (M ; E1 ).

4 Elliptic Regularity
4.1 Local elliptic regularity
The main theorem of elliptic regularity is the following:

Theorem (Local elliptic regularity). Let L be a differential operator of order k ≥ 1

on Rn , U ⊆ Rn precompact and L elliptic over Ū . Then
1. There is a constant A such that for all u ∈ H s+k (U ), we have

kuks+k ≤ A(kuks + kLuks ).

2. If u ∈ H r for some r is such that Lu ∈ H s for some s, then µu ∈ H s+k (U )

for every µ ∈ Cc∞ (U ).
The first part requires getting our hands dirty and proving explicit estimates. A
proof sketch will be given here, with the details carried out in the Appendix.

9
Proof sketch of (1). First prove it if L has constant coefficients, which is easy since

Lu(ξ)
c = p(ξ)û(ξ)

for some polynomial p.

Next, for arbitrary L, we show that for any x ∈ x0 , there is a neighbourhood
V ⊆ U such that the estimate holds for all u supported in V . To do so, let L0 be
the differential operator with constant coefficients that agree with L at x0 . We then
have to bound k(L − L0 )uks on the assumption that the coefficients of L − L0 can
be made arbitrarily small near the support of u.
Finally, use compactness and partitions of unity to glue these results together.
The second part follows from the first formally.
Observe the first part tells us if we already knew that u were in H s+k , then
kuks+k would be very well-behaved. This is an a priori estimate. To show that u is
actually in H s+k , we apply appropriate smooth approximations.
Fix a φ ∈ Cc∞ (Rn ) such that
Z
φ ≥ 0, φ dx = 1, φ(−x) = φ(x).

For ε > 0, we define

1 x
φε (x) = φ .
εn ε
We then define the mollifier

Fε : C ∞ (Rn ) → C ∞ (Rn )
u 7→ φε ∗ u.

The main theorem about mollifiers, which I will not prove, is that
Theorem. Fε extends to a bounded operator H s → H s with norm ≤ 1. Moreover
• Fε commutes with all differential operators with constant coefficients.
• If L is a differential operator with compact support, then [Fε , L] extends to a
map H s → H s−k+1 for all s ∈ R, and has uniformly bounded operator norm.
• For any u ∈ H s , we have Fε u ∈ C ∞ ∩ H s .
• For any u ∈ H s , we have Fε u → u in H s .
• If U ⊆ Rn is precompact and u ∈ H r (U ) for some r < s, and kFt uks is
uniformly bounded in t, then u ∈ H s (U ).
The last point is how we are going to show that u ∈ H s+k (U ), while the others
are needed to establish the required bounds.

10
Proof of (2) from (1). Inductively, we may assume µu ∈ H s+k−1 . We then bound
kFε µuks+k ≤ A(kFε µuks + kLFε µuks )
≤ A(kFε µuks + k[L, Fε ]µuks + kFε [L, µ]uks + kFε µLuks )
≤ Akµuks + A1 kµuks+k−1 + A2 kuks+k−1 + A3 kLuks ,
where for the third term, we used that [L, µ] is a differential operator of degree
k − 1.

4.2 Global elliptic regularity

Let M be a compact manifold, E0 , E1 complex vector bundles, and L an elliptic
differential operator from E0 to E1 of order k. By patching local results together,
we conclude that
Theorem (Global elliptic regularity).
1. There is a constant A such that for all s ≥ 0 and u ∈ H s+k (M ; E0 ), we have
kuks+k ≤ A(kukk + kLukk ).

2. If Lu ∈ H s , then u ∈ H s+k . In particular, if Lu = 0, then u ∈ C ∞ (M ).

This implies a lot of very nice properties about elliptic operators. First observe
the following result:
Proposition. Let U, V, W be Hilbert spaces and L : U → V bounded, K : U → W
compact. If there is an A such that
kukU ≤ C(kLukV + kKukW ),
then ker L is finite-dimensional and im L is closed.
Proof. We show that the unit ball of ker L is compact. If (un ) is a sequence in the
unit ball of ker L, then
kun − um kU ≤ AkKun − Kum k.
Since K is compact, there is a subsequence uni such that Kuni is Cauchy. So uni is
Cauchy. So we are done.
To show im L is closed, by restricting to the complement of the kernel, we may
assume L is injective. We will show that there is a c such that
kukU ≤ ckLukV .
If not, pick a sequence un with kun kU = 1 but kLukV → 0. By compactness, we
may assume that Kun is Cauchy. Then we see that un must also be Cauchy, and
the limit u must satisfy kukU = 1 and Lu = 0, a contradiction.

11
Corollary. L : H s+k → H s has finite-dimensional kernel and closed image.
The compactness of the manifold M is crucial for the inclusion H s+k → H s to
be compact.
We want to show that in fact L is Fredholm, i.e. both the kernel and the cokernel
are compact. Here we simply use duality. We need to show that
K = {v ∈ H −s : hLu, vi = 0 for all u ∈ H s+k }
is finite-dimensional. But this is exactly the kernel of L∗ . So we deduce
Corollary. L is Fredholm, and in fact
H s = ker L∗ + im(L : H s+k → H s ).
The first factor is independent of s (since all elements are C ∞ ), and taking the limit
s → ∞, we get
Γ(M, E1 ) = ker L∗ ⊕ im L.
A bit of care is actually needed to take the limit s → ∞, but we leave that for
the reader.

5 Hodge Theory
We now arrive at the main theorem we were working towards.
Theorem (Hodge Decomposition Theorem). Let (E∗ , L) be an elliptic complex with
D = L + L∗ and ∆ = D2 as before. Then
Γ(M, E) = ker D ⊕ im D.
Moreover,
1. ker D is finite-dimensional.
2. ker D = ker ∆ = ker L ∩ ker L∗
3. im ∆ = im D = im LL∗ ⊕ im L∗ L = im L ⊕ im L∗
4. ker L = im L ⊕ ker ∆, and ker ∆ → H(E∗ ) is an isomorphism.
Proof. (1) follows from regularity. (2) follows from the identities
hu, ∆ui = hDu, Dui = hLu, Lui + hL∗ u + L∗ ui, u ∈ Γ(M, Ei ).
For (3), we can also decompose Γ(M, E) = ker ∆ ⊕ im ∆. Since im ∆ ⊆ im D, they
must be equal. Moreover,
im D ⊆ im(LL∗ ) ⊕ im(L∗ L) ⊆ im L ⊕ im L∗ ,

12
but clearly im L ⊕ im L∗ ⊥ ker ∆ since

hLu + L∗ v, wi = hu + v, Dwi.

So we must have equality throughout, and im L is clearly orthogonal to im L∗ . For

(4), it is clear that ker L ⊇ im L ⊕ ker ∆, and since ker L ⊥ im L∗ , that must be an
equality.
Theorem (Spectral theorem). Let M be a closed Riemannian manifold and E → M
a Hermitian vector bundle. Let D : Γ(M, E) → Γ(M, E) be formally self-adjoint of
order k ≥ 1. Then we have an orthogonal decomposition
M
L2 (M, E) = ker(D − λ).
λ∈R

Moreover, each ker(D − λ) is finite-dimensional, and for any Λ, there are only finitely
many eigenvalues of magnitude < Λ.
The idea is to apply the spectral theorem for compact self-adjoint operators to
the inverse of D. Of course, D need not be invertible. So we do the following:

Proof. Consider the operator L = 1 + D2 : Γ(M, E) → Γ(M, E). It is then clear that
L = L∗ is elliptic and injective. So L : H 2k → L2 is invertible (since the complement
of the image is ker L∗ ), with inverse S : L2 → H 2k . Since L induces a bijection
S
between the smooth sections, so does S. Let T be the composition L2 → H 2k ,→ L2 .
Then this is compact and self-adjoint (can check this for smooth sections, and use
that its “inverse” L is formally self adjoint).
By the spectral theorem of compact self–adjoint operators (and positivity of T ),
M
L2 (M ; E) = ker(T − µ).
µ>0

Moreover, each factor is finite-dimensional, and 0 is the only accumulation point of

the spectrum.
We will show that ker(T − µ) decomposes as a sum of eigenspaces for D. We first
establish that
ker(T − µ) = im(1 − µL)⊥ = ker(1 − µL).
Since L is self-adjoint, the second equality follows by elliptic regularity. The first
equality follows from the computation

hx, (1 − µL)ui = hx, ui − µhx, Lui = hT x, Lui − µhx, Lui = h(T − µ)x, Lui,

plus the density of Γ(M, E) and surjectivity of L : Γ(M, E) → Γ(M, E).

13
 Now since D commutes with L, we know D acts as a self-adjoint operator on the
finite-dimensional vector space ker(T − µ) = ker(1 − µL). Moreover, restricted to
this subspace, we have
1
D2 = − 1.
µ
Soqby linear algebra, ker(T − µ) decomposes into eigenspaces of D of eigenvalues
± µ1 − 1, and the theorem follows.

A Proof of local elliptic regularity

We fill in the details of the proof of local elliptic regularity. We fix L a differential
operator of order k ≥ 1 on Rn , U ⊆ Rn precompact and L elliptic over Ū .
Lemma. If L has constant coefficients, i.e.
X
L= aα Dα ,
|α|≤k

then there is an A such that for all u ∈ Cc∞ (U ),

kuks+k ≤ A(kuks + kLuks ).

Proof. Observe that we have

Lu(ξ)
c = p(ξ)û(ξ).

for some polynomial p of degree at most k. By ellipticity, for some R  0 and

constant A > 0, we have
A|p(ξ)| ≥ (1 + |ξ|2 )k/2 .
So in the decomposition
Z Z Z !
2 s+k
|û(ξ)| (1 + |ξ|) dξ = + |û(ξ)|2 (1 + |ξ|)s+k dξ,
|ξ|≤R |ξ|≥R

we can bound the first term by (1 + R2 )k kuks , and we can bound the second term by
Z
2
|Lu(ξ)|
c (1 + |ξ|2 )s dξ ≤ kLuks .
|ξ|≥R

Lemma. For any fixed L and x0 ∈ U , there is some neighbourhood V ⊆ U of x and

A > 0 such that for all u ∈ Cc∞ (V ), we have

kuks+k ≤ A(kuks + kLuks ).

14
Proof. Let L0 be the differential operator with constant coefficients that agree with
L at x0 . Then L0 is also an elliptic operator, and the above applies. So for any u,
we have

kuks+k ≤ A1 (kuks + kL0 uks ) ≤ A0 (kuks + k(L − L0 )uks + kLuks ).

So we have to control the term k(L − L0 )uks . For a tiny δ  1, pick a neighbourhood
V of x0 such that the coefficients of L − L0 are bounded by δ. Then

k(L − L0 )uks ≤ δA2 kuks+k + A3 kuks+k−1 ,

where A2 , A3 are fixed, independent of u and V . By the next lemma, for any ε > 0,
we can bound
A1 A3 kuks+k−1 ≤ εkuks+k + A4 (ε)kuks .
We then deduce that

kuks+k ≤ A1 kLuks + (δA1 A2 + ε)kuks+k + (A1 + A4 (ε))kuks .

Picking δ and ε to be small enough, we are done.

Lemma. For any r < s < t and ε > 0, there exists C(ε) such that

(1 + |ξ|2 )s ≤ (1 + |ξ|2 )t ε + (1 + |ξ|2 )r C(ε)

for all ξ. Hence

kuks ≤ εkukt + C(ε)kukr .
Proof. The claim is the same as

1 ≤ (1 + |ξ|2 )t−s ε + (1 + |ξ|2 )r−s C(ε).

Observe that for any y, we always have

1 ≤ y t−s + (1/y)s−r .

Then take y = (1 + |ξ|2 )ε1/(t−s) .

Theorem. For any L, there exists A such that

kuks+k ≤ A(kuks + kLuks ).

Proof. Pick W ⊇ Ū such that L is elliptic on W , and cover W (and hence Ū ) with
finitely Vi where we have a bound as above, say

kuks+k ≤ A(kuks + kLuks−k )

15
for any u supported in the Vi ’s. Now pick a partition of unity {µi } subordinate to
{Vi }. Then
X X
kuks+k ≤ kµi uks+k ≤ C(kµi uks + kLµi uks )
X
≤ C(kµi uks + kµi Luks + k[L, µi ]uks ).

We can bound the first two by a constant multiple of kuks and kLuks . To bound
the last term, we use that [L, µi ] is a differential operator of order k − 1, and hence

k[L, µi ]uks ≤ Ckuks+k−1 ≤ εkuks+k + C(ε)kuks .

16