CHAPTER 8

Elliptic boundary problems

Summary
Elliptic boundary problems are discussed, especially for operators of Dirac type.
We start with a discussion of manifolds with boundary, including functions spaces
and distributions. This leads to the ‘jumps formula’ for the relationship of the
action of a differential operator to the operation of cutting off at the boundary;
this is really Green’s formula. The idea behind Calderòn’s approach to boundary
problems is introduced in the restricted context of a dividing hypersurface in a
manifold without boundary. This includes the fundamental result on the boundary
behaviour of a pseudodifferential operator with a rational symbol. These ideas are
then extended to the case of an operator of Dirac type on a compact manifold
with boundary with the use of left and right parametrices to define the Calderòn
projector. General boundary problems defined by pseudodifferential projections are
discussed by reference to the ‘Calderòn realization’ of the operator. Local boundary
conditions, and the corresponding ellipticity conditions, are then discussed and the
special case of Hodge theory on a compact manifold with boundary is analysed in
detail for absolute and relative boundary conditions.

Introduction
Elliptic boundary problems arise from the fact that elliptic differential operators
on compact manifolds with boundary have infinite dimensional null spaces. The
main task we carry out below is the parameterization of this null space, in terms of
boundary values, of an elliptic differential operator on a manifold with boundary.
For simplicity of presentation the discussion of elliptic boundary problems here will
be largely confined to the case of first order systems of differential operators of
Dirac type. This has the virtue that the principal results can be readily stated.

Status as of 4 August, 1998

Read through Section 8.1–Section 8.2: It is pretty terse in places! Several vital
sections are still missing.

8.1. Manifolds with boundary

Smooth manifolds with boundary can be defined in very much the same was as
manifolds without boundary. Thus we start with a paracompact Hausdorff space
X and assume that it is covered by ‘appropriate’ coordinate patches with corre-
sponding transition maps. In this case the ‘model space’ is Rn,1 = [0, ∞) × Rn−1 ,
a Euclidean half-space of fixed dimension, n. As usual it is more convenient to use
179
180 8. ELLIPTIC BOUNDARY PROBLEMS

as models all open subsets of Rn,1 ; of course this means relatively open, not open
as subsets of Rn . Thus we allow any
O = O0 ∩ Rn,1 , O0 ⊂ Rn open,
as local models.
By a smooth map between open sets in this sense we mean a map with a smooth
extension. Thus if Oi for i = 1, 2 are open in Rn,1 then smoothness of a map F
means that

(8.1) F : O1 → O2 , ∃ Oi0 ⊂ Rn , i = 1, 2, open and F̃ : O10 → O20

which is C ∞ with Oi = Oi0 ∩ Rn,1 and F = F 0 |O1 .
It is important to note that the smoothness condition is much stronger than
just smoothness of F on O ∩ (0, ∞) × Rn−1 .
By a diffeomorphism between such open sets we mean an invertible smooth
map with a smooth inverse. Various ways of restating the condition that a map be
a diffeomorphism are discussed below.
With this notion of local model we define a coordinate system (in the sense of
manifolds with boundary) as a homeomorphism of open sets
Φ
X ⊃ U −→ O ⊂ Rn,1 , O, U open.
Thus Φ is assumed to exist and both Φ and Φ−1 are assumed to be continuous.
−1

The compatibility of two such coordinate systems (U1 , Φ1 , O1 ) and (U2 , Φ2 , O2 ) is

the requirement that either U1 ∩ U2 = φ or if U1 ∩ U2 6= φ then
Φ1,2 = Φ2 ◦ Φ−1
1 : Φ1 (U1 ∩ U2 ) → Φ2 (U1 ∩ U2 )

is a diffeomorphism in the sense described above. Notice that both Φ1 (U1 ∩ U2 )

and Φ2 (U1 ∩ U2 ) are open in Rn,1 . The inverse Φ1,2 is defined analogously.
A C ∞ manifold with boundary can then be formally defined as a paracompact
Hausdorff topological space which has a maximal covering by mutually compatible
coordinate systems.
An alternative definition, i.e.
characterization, of a manifold with boundary is that there exists a C ∞ manifold
X̃ without boundary and a function f ∈ C ∞ (X̃) such that df 6= 0 on {f = 0} ⊂ X̃
and n o
X = p ∈ X̃; f (p) ≥ 0 ,
with coordinate systems obtained by restriction from X̃. The doubling construction
described below shows that this is in fact an equivalent notion.

8.2. Smooth functions

As in the boundaryless case, the space of functions on a compact manifold
with boundary is the primary object of interest. There are two basic approaches to
defining local smoothness, the one intrinsic and the other extrinsic, in the style of
the two definitions of a manifold with boundary above. Thus if O ⊂ Rn,1 is open
we can simply set

C ∞ (O) = {u : O → C; ∃ ũ ∈ C ∞ (O0 ) ,
O0 ⊂ Rn open, O = O0 ∩ Rn,1 , u = ũ|O .
 8.2. SMOOTH FUNCTIONS 181

Here the open set in the definition might depend on u. The derivatives of ũ ∈
C ∞ (O0 ) are bounded on all compact subsets, K b 0. Thus
(8.2) sup |Dα u| < ∞ , O◦ = O ∩ ((0, ∞) × Rn−1 ) .
K∩O ◦

The second approach is to use (8.2) as a definition, i.e.

to set
(8.3) C ∞ (O) = {u : O◦ → C; (8.2) holds ∀ K b O and all α} .
In particular this implies the continuity of u ∈ C ∞ (O) up to any point p ∈ O ∩
({0} × Rn−1 ), the boundary of O as a manifold with boundary.
As the notation here asserts, these two approaches are equivalent. This follows
(as does much more) from a result of Seeley:
Proposition 8.1. If C ∞ (O) is defined by (8.3) and O0 ⊂ Rn is open with
O = O0 ∩ Rn,1 then there is a linear extension map
E : C ∞ (O) → C ∞ (O0 ) , Eu|O0 = u
which is continuous in the sense that for each K 0 b O0 , compact, there is some
K b O such that for each α
sup |Dα Eu| ≤ Cα,K 0 sup |Dα u| .
K0 K∩O

The existence of such an extension map shows that the definition of a diffeo-
morphism of open sets O1 , O2 , given above, is equivalent to the condition that
the map be invertible and that it, and its inverse, have components which are in
C ∞ (O1 ) and C ∞ (O2 ) respectively.
Given the local definition of smoothness, the global definition should be evident.
Namely, if X is a C ∞ manifold with boundary then
C ∞ (X) = u : X → C; (Φ−1 )∗ (u|U ) ∈ C ∞ (O) ∀ coordinate systems .


This is also equivalent to demanding that local regularity, i.e.

the regularity of (Φ−1 )∗ (u|O ), hold for any one covering by admissible coordinate
systems.
As is the case of manifolds without boundary, C ∞ (X) admits partitions of unity.
In fact the proof of Lemma 6.3 applies verbatim; see also Problem 6.3.
The topology of C ∞ (X) is given by the supremum norms of the derivatives in
local coordinates. Thus a seminorm
sup Dα (Φ−1 )∗ (u|U )
KbO

arises for each compact subset of each coordinate patch. In fact there is a countable
set of norms giving the same topology. If X is compact, C ∞ (X) is a Fréchet space,
if it is not compact it is an inductive limit of Fréchet spaces (an LF space).
The boundary of X, ∂X, is the union of the Φ−1 (O ∩ ({0} × Rn−1 )) over
coordinate systems. It is a manifold without boundary. It is compact if X is
compact. Furthermore, ∂X has a global defining function ρ ∈ C ∞ (X); that is
ρ ≥ 0, ∂X = {ρ = 0} and dρ 6= 0 at ∂X. Moreover if ∂X is compact then any such
boundary defining function can be extended to a product decomposition of X near
∂X:
(8.4) ∃ C ⊃ ∂X , open in X  > 0 and a diffeomorphism ϕ : C ' [0, )ρ × ∂X.
182 8. ELLIPTIC BOUNDARY PROBLEMS

If ∂X is not compact this is still possible for an appropriate choice of ρ. For an

outline of proofs see Problem 8.1.
Lemma 8.1. If X is a manifold with compact boundary then for any boundary
defining function ρ ∈ C ∞ (X) there exists  > 0 and a diffeomorphism (8.4).

Problem 8.1.
The existence of such a product decomposition near the boundary (which might
have several components) allows the doubling construction mentioned above to be
carried through. Namely, let
(8.5) X̃ = (X ∪ X)/∂X
be the disjoint union of two copies of X with boundary points identified. Then
consider
(8.6) C ∞ (X̃) = {(u1 , u2 ) ∈ C ∞ (X) ⊕ C ∞ (X);
(ϕ−1 )∗ (u1 |C ) = f (ρ, ·) , (ϕ−1 )∗ (u2 |C ) = f (−ρ, ·) ,
f ∈ C ∞ ((−1, 1) × ∂X)} .
This is a C ∞ structure on X̃ such that X ,→ X̃, as the first term in (8.5), is an
embedding as a submanifold with boundary, so
C ∞ (X) = C ∞ (X̃)|X .
In view of this possibility of extending X to X̃, we shall not pause to discuss
all the usual ‘natural’ constructions of tensor bundles, density bundles, bundles of
differential operators, etc. They can simply be realized by restriction from X̃. In
practice it is probably preferable to use intrinsic definitions.
The definition of C ∞ (X) implies that there is a well-defined restriction map
C ∞ (X) 3 u 7−→ u|∂X ∈ C ∞ (∂X).
It is always surjective. Indeed the existence of a product decomposition shows that
any smooth function on ∂X can be extended locally to be independent of the chosen
normal variable, and then cut off near the boundary.
There are important points to observe in the description of functions near
the boundary. We may think of C ∞ (X) ⊂ C ∞ (X ◦ ) as a subspace of the smooth
functions on the interior of X which describes the ‘completion’ (compactification if
X is compact!) of the interior to a manifold with boundary. It is in this sense that
the action of a differential operator P ∈ Diff m (X)
P : C ∞ (X) → C ∞ (X)
should be understood. Thus P is just a differential operator on the interior of X
with ‘coefficients smooth up to the boundary.’
Once this action is understood, there is an obvious definition of the space of
C ∞ functions which vanish to all orders at the boundary,
C˙∞ (X) = {u ∈ C ∞ (X); P u|∂X = 0 ∀ P ∈ Diff ∗ (X)} .
Having chosen a product decomposition near the boundary, Taylor’s theorem gives
us an isomorphism
C ∞ (X)/C˙∞ (X) ∼
M
= C ∞ (∂X) · [dρ|∂X ]k .
k≥0
 8.3. DISTRIBUTIONS 183

8.3. Distributions
It is somewhat confusing that there are three (though really only two) spaces of
distributions immediately apparent on a compact manifold with boundary. Under-
standing the relationship between them is important to the approach to boundary
problems used here.
The coarsest (as it is a little dangerous to say largest) space is C −∞ (X ◦ ),
the dual of Cc∞ (X ◦ ; Ω), just the space of distributions on the interior of X. The
elements of C −∞ (X ◦ ) may have unconstrained growth, and unconstrained order of
singularity, approaching the boundary. They are not of much practical value here
and appear for conceptual reasons.
Probably the most natural space of distributions to consider is the dual of
C ∞ (X; Ω) since this is arguably the direct analogue of the boundaryless case. We
shall denote this space
(8.7) C˙−∞ (X) = (C ∞ (X; Ω))0
and call it the space of supported distributions. The ‘dot’ is supposed to indicate
this support property, which we proceed to describe.
If X̃ is any compact extension of X (for example the double) then, as already
noted, the restriction map C ∞ (X̃; Ω) → C ∞ (X; Ω) is continuous and surjective.
Thus, by duality, we get an injective ‘extension’ map
(8.8) C˙−∞ (X) 3 u 7→ ũ ∈ C −∞ (X̃), ũ(ϕ) = u(ϕ|X ).
We shall regard this injection as an identification C˙−∞ (X) ,→ C −∞ (X̃); its range
is easily characterized.
Proposition 8.2. The range of the map (8.8) is the subspace consisting of
those ũ ∈ C −∞ (X̃) with supp ũ ⊂ X.
The proof is given below. This proposition is the justification for calling
C˙−∞ (X) the space of supported distributions; the dot is support to indicate that
this is the subspace of the ‘same’ space for X̃, i.e.
C −∞ (X̃), of elements with support in X.
This notation is consistent with C˙∞ (X) ⊂ C ∞ (X̃) being the subspace (by
extension as zero) of elements with support in X. The same observation applies to
sections of any vector bundle, so
C˙∞ (X; Ω) ⊂ C ∞ (X̃; Ω)
is a well-defined closed subspace. We set
(8.9) C −∞ (X) = (C˙∞ (X; Ω))0
and call this the space of extendible distributions on X. The inclusion map for the
test functions gives by duality a restriction map:

(8.10) RX : C −∞ (X̃) → C −∞ (X),

RX u(ϕ) = u(ϕ) ∀ ϕ ∈ C˙∞ (X; Ω) ,→ C ∞ (X̃; Ω) .
We write, at least sometimes, RX for the map since it has a large null space so
should not be regarded as an identification. In fact
n o
(8.11) Nul(RX ) = v ∈ C −∞ (X̃); supp(v) ∩ X ◦ = φ = C˙−∞ (X̃\X ◦ ) ,
184 8. ELLIPTIC BOUNDARY PROBLEMS

is just the space of distributions supported ‘on the other side of the boundary’. The
primary justification for calling C −∞ (X) the space of extendible distributions is:

Proposition 8.3. If X is a compact manifold with boundary, then the space

Cc∞ (X ◦ ; Ω) is dense in C˙∞ (X; Ω) and hence the restriction map

(8.12) C −∞ (X) ,→ C −∞ (X ◦ )

is injective, whereas the restriction map from (8.10), RX : C˙−∞ (X) −→ C −∞ (X),
is surjective.

Proof. If V is a real vector field on X̃ which is inward-pointing across the

boundary then
exp(sV ) : X̃ → X̃
is a diffeomorphism with Fs (X) ⊂ X ◦ for s > 0. Furthermore if ϕ ∈ C ∞ (X̃) then
Fs∗ ϕ → ϕ in C ∞ (X̃) as s → 0. The support property shows that Fs∗ ϕ ∈ Cc∞ (X ◦ )
if s < 0 and ϕ ∈ C˙∞ (X). This shows the density of Cc∞ (X ◦ ) in C˙∞ (X). Since
all topologies are uniform convergence of all derivatives in open sets. The same
argument applies to densities. The injectivity of (8.12) follows by duality.
On the other hand the surjectivity of (8.10) follows directly from the Hahn-
Banach theorem. 

Proof of Proposition 8.2. For ũ ∈ C −∞ (X̃) the condition that supp ũ ⊂ X

is just

(8.13) ũ(ϕ) = 0 ∀ ϕ ∈ Cc∞ ⊂ (X̃\X; Ω) ⊂ C ∞ (X̃; Ω) .

Certainly (8.13) holds if u ∈ C˙−∞ (X) since ϕ|X = 0. Conversely, if (8.13)

holds, then by continuity and the density of Cc∞ (X̃\X; Ω) in C ∞ (X̃\X ◦ ; Ω), what
follows from Proposition 8.3, ũ vanishes on C˙∞ (X\X ◦ ). 

It is sometimes useful to consider topologies on the spaces of distributions

C −∞ (X) and C˙−∞ (X). For example we may consider the weak topology. This is
given by all the seminorms u 7→ khu, φik, where φ is a test function.

Lemma 8.2. With respect to the weak topology, the subspace Cc∞ (X ◦ ) is dense
in both C˙−∞ (X) and C −∞ (X).

8.4. Boundary Terms

To examine the precise relationship between the supported and extendible dis-
tributions consider the space of ‘boundary terms’.
n o
−∞
(8.14) C˙∂X (X) = u ∈ C˙−∞ (X) ; supp(u) ⊂ ∂X .

Here the support may be computed with respect to any extension, or intrinsically
on X. We may also define a map ‘cutting off’ at the boundary:
Z
(8.15) ∞
C (X) 3 u 7→ uc ∈ C˙ −∞
(X) , uc (ϕ) = uϕ ∀ ϕ ∈ C ∞ (X ; Ω) .
X
 8.4. BOUNDARY TERMS 185

Proposition 8.4. If X is a compact manifold with boundary then there is a

commutative diagram

(8.16) C˙∞ (X)


C ∞ (X)
s
()c sss
sss
ysss 
0 / C˙−∞ (X) / C˙−∞ (X) / C −∞ (X) /0
∂X

with the horizontal sequence exact.

Proof. The commutativity of the triangle follows directly from the definitions.
The exactness of the horizontal sequence follows from the density of Cc∞ (X ◦ ; Ω) in
−∞
C˙∞ (X; Ω). Indeed, this shows that v ∈ C˙∂X (X) maps to 0 in C −∞ (X) since
v(ϕ) = 0 ∀ ϕ ∈ Cc (X ; Ω). Similarly, if u ∈ C˙−∞ (X) maps to zero in C −∞ (X)
∞ ◦

then u(ϕ) = 0 for all ϕ ∈ Cc∞ (X ◦ ; Ω), so supp(u) ∩ X ◦ = ∅. 

Note that both maps in (8.16) from C ∞ (X) into supported and extendible
distributions are injective. We regard the map into C −∞ (X) as an identification.
In particular this is consistent with the action of differential operators. Thus P ∈
Diff m (X) acts on C ∞ (X) and then the smoothness of the coefficients of P amount
to the fact that it preserves C ∞ (X), as a subspace. The formal adjoint P ∗ with
respect to the sesquilinear pairing for some smooth positive density, ν
Z
(8.17) hϕ, ψi = ϕψν ∀ ϕ, ψ ∈ C ∞ (X)
X

acts on C˙∞ (X):

(8.18) hP ∗ ϕ, ψi = hϕP ψi ∀ ϕ ∈ C˙∞ (X) , ψ ∈ C ∞ (X), P ∗ : C˙∞ (X) −→ C˙∞ (X).
However, P ∗ ∈ Diff m (X) is fixed by its action over X ◦ . Thus we do have
(8.19) hP ∗ ϕ, ψi = hϕ, P ψi ∀ ϕ ∈ C ∞ (X) , ψ ∈ C˙∞ (X) .
We define the action of P by duality. In view of the possibility of confusion,
we denote P the action on C −∞ (X) and by Ṗ the action on C˙∞ (X).
(8.20)
hP u , ϕi = hu, P ∗ ϕi ∀ u ∈ C −∞ (X) , ϕ ∈ C˙∞ (X), P : C −∞ (X) −→ C −∞ (X)
hṖ u , ϕi = hu, P ∗ ϕi ∀ u ∈ C˙−∞ (X) , ϕ ∈ C ∞ (X), Ṗ : C˙−∞ (X) −→ C˙−∞ (X).
It is of fundamental importance that (8.19) does not hold for all ϕ, ψ ∈ C ∞ (X).
This failure is reflected in Green’s formula for the boundary terms, which appears
below as the ‘Jump formula’. This is a distributional formula for the difference
−∞
(8.21) Ṗ uc − (P u)c ∈ C˙∂X , u ∈ C ∞ (X) P ∈ Diff m (X) .
Recall that a product decomposition of C ⊂ X near ∂X is fixed by an inward
pointing vector field V. Let x ∈ C ∞ (X) be a corresponding boundary defining
function, with V x = 0 near ∂X, with χV : C → ∂X the projection onto the
186 8. ELLIPTIC BOUNDARY PROBLEMS

boundary from the product neighborhood C. Then Taylor’s formula for u ∈ C ∞ (X)
becomes
X 1
(8.22) u∼ χ∗ (V k u|∂x )xk .
k! V
k
It has the property that a finite sum
N
X 1 ∗ k
uN = ϕu − ϕ χ (V u|∂X )xk
k! V
k=0
where ϕ ≡ 1 near ∂X, supp ϕ ⊂ C, satisfies
(8.23) Ṗ (uN )c = (P uN )c , P ∈ Diff m (X) , m < N .
Since (1 − ϕ)u ∈ C˙∞ (X) also satisfies this identity, the difference in (8.21) can (of
course) only depend on the V k u|∂X for k ≤ m, in fact only for k < m.
Consider the Heaviside function 1c ∈ C˙−∞ (X), detained by cutting off the
identity function of the boundary. We define distributions
(8.24) δ (j) (x) = V j+1 1c ∈ C˙−∞ , j ≥ 0 .
∂X

Thus, δ (0) (x) = δ(x) is a ‘Dirac delta function’ at the boundary. Clearly supp δ(x) ⊂
∂X, so the same is true of δ (j) (x) for every j. If ψ ∈ C ∞ (∂X) we define
(8.25) ψ · δ (j) (x) = ϕ(XV∗ ψ) · δ (j) (x) ∈ C˙−∞ (X) .
∂X
(j)
This, by the support property of δ , is independent of the cut off ϕ used to define
it.
Proposition 8.5. For each P ∈ Diff m (X) there are differential operators on
the boundary Pij ∈ Diff m−i−j−1 (∂X), i + j < m, i, j ≥ 0, such that
X
(8.26) Ṗ uc − (P u)c = (Pij (Vuj |∂X ) · δ (j) (x), ∀ u ∈ C ∞ (X),
i,j
−m
and P0m−1 = i σ(P, dx) ∈ C ∞ (∂X).
Proof. In the local product neighborhood C,
X
(8.27) P = Pl V l
0≤l≤m

where Pl is a differential operator of the order at most m − l, on X be depending on

x as a parameter. Thus the basic cases we need to analyze arise from the application
of V to powers of x :
(8.28) xl V j+1 (xp )c − (V j+1 xp )c


(
0 l+p>j
= p!(j−p)! (j−p−l) l+p≤j.
(j−p−l)! (−1)lδ
Taking the Taylor sense of the Pl ,
X
Pl ∼ xr Pl,r
r
and applying P to (8.22) gives
X
(8.29) P uc − (P u)c = (−1)r (Pl,r (V k u|∂x )) δ (l−1−r−k) (x) .
r+k<l
 8.5. SOBOLEV SPACES 187

This is of the form (8.26). The only term with l − 1 − r − k = m − 1 arises from
l − m, k = r = 0 so is the operator Pm at x = 0. This is just i−m σ(P, dx). 

8.5. Sobolev spaces

As with C ∞ functions we may define the standard (extendible) Sobolev spaces
by restriction or intrinsically. Thus, if X̃ is an extension of a compact manifold
with boundary, X, the we can define

(8.30) H m (X) = Hcm (X̃)|X, ∀ m ∈ R; H m (X) ⊂ C −∞ (X).

That this is independent of the choice of X̃ follows from the standard properties
of the Sobolev spaces, particularly their localizability and invariance under diffeo-
morphisms. The norm in H m (X) can be taken to be
n o
(8.31) kukm = inf kũkH m (X̃) ; ũ ∈ H m (X̃), u = ũX .

A more intrinsic defintion of these spaces is discussed in the problems.

There are also supported Sobolev spaces,
n o
(8.32) Ḣ m (X) = u ∈ H m (X̃); supp(u) ⊂ X ⊂ C˙−∞ (X).

Sobolev space of sections of any vector bundle can be defined similarly.

Proposition 8.6. For any m ∈ R and any compact manifold with boundary
X, H m (X) is the dual of Ḣ −m (X; Ω) with respect to the continuous extension of
the densely defined bilinear pairing
Z
(u, v) = uv, u ∈ C ∞ (X), v ∈ C˙∞ (X; Ω).
X

Both H m (X) and Ḣ m (X) are C ∞ (X)-modules and for any vector bundle over
X, H m (X; E) ≡ H m (X) ⊗C ∞ (X) C ∞ (X; E) and Ḣ m (X; E) ≡ Ḣ m (X) ⊗C ∞ (X)
C ∞ (X; E).

Essentially from the definition of the Sobolev spaces, any P ∈ Diff k (X; E1 , E2 )
defines a continuous linear map

(8.33) P : H m (X; E1 ) −→ H m−k (X; E2 ).

We write the dual (to P ∗ of course) action

(8.34) Ṗ : Ḣ m (X; E1 ) −→ Ḣ m−k (X; E2 ).

These actions on Sobolev spaces are consistent with the corresponding actions on
distributions. Thus
[ \
C −∞ (X; E) = H m (X), C ∞ (X; E) = H m (X),
m m
[ \
C˙−∞ (X; E) = Ḣ (X), C˙∞ (X; E) =
m
Ḣ m (X).
m m
188 8. ELLIPTIC BOUNDARY PROBLEMS

8.6. Dividing hypersurfaces

As already noted, the point of view we adopt for boundary problems is that
they provide a parametrization of the space of solutions of a differential opera-
tor on a space with boundary. In order to clearly indicate the method pioneered
by Calderòn, we shall initially consider the restrictive context of an operator of
Dirac type on a compact manifold without boundary with an embedded separating
hypersurface.
Thus, suppose initially that D is an elliptic first order differential operator act-
ing between sections of two (complex) vector bundles V1 and V2 over a compact
manifold without boundary, M. Suppose further that H ⊂ M is a dividing hyper-
surface. That is, H is an embedded hypersuface with oriented (i.e. trivial) normal
bundle and that M = M+ ∪ M− where M± are compact manifolds with boundary
which intersect in their common boundary, H. The convention here is that M+ is
on the positive side of H with respect to the orientation.
In fact we shall make a further analytic assumption, that
(8.35) D : C ∞ (M ; V1 ) −→ C ∞ (M ; V2 ) is an isomorphism.
As we already know, D is always Fredholm, so this implies the topological condi-
tion that the index vanish. However we only assume (8.35) to simplify the initial
discussion.
Our objective is to study the space of solutions on M+ . Thus consider the map
bH
u ∈ C ∞ (M+ ; V1 ); Du = 0 in M+
◦
C ∞ (H; V1 ), bH u = u|∂M+ .

(8.36) −→
The idea is to use the boundary values to parameterize the solutions and we can
see immediately that this is possible.
Lemma 8.3. The assumption (8.35) imples that map bH in (8.36) is injective.
Proof. Consider the form of D in local coordinates near a point of H. Let the
coordinates be x, y1 , . . . , yn−1 where x is a local defining function for H and assume
that the coordinate patch is so small that V1 and V2 are trivial over it. Then
n−1
X
D = A0 Dx + Aj Dyj + A0
j=1
0
where the Aj and A are local smooth bundle maps from V1 to V2 . In fact the
ellipticity of D implies that each of the Aj ’s is invertible. Thus the equation can
be written locally
n−1
X
Dx u = Bu, B = − A−1 −1 0
0 Dyj − A0 A .
j=1

The differential operator B is tangent to H. By assumption u vanishes when re-

stricted to H so it follows that Dx u also vanishes at H. Differentiating the equation
with respect to x, it follows that all derivatives of u vanish at H. This in turn
implies that the global section of V1 over M
(
u in M+
ũ =
0 in M−
is smooth and satisfies Dũ = 0. Then assumption (8.35) implies that ũ = 0, so
u = 0 in M+ and bH is injective. 
 8.6. DIVIDING HYPERSURFACES 189

In the proof of this Lemma we have used the strong assumption (8.35). As
we show below, if it is assumed instead that D is of Dirac type then the Lemma
remains true without assuming (8.35). Now we can state the basic result in this
setting.
Theorem 8.1. If M = M+ ∪ M− is a compact manifold without boundary
with separating hypersurface H as described above and D ∈ Diff 1 (M ; V1 , V2 ) is a
generalized Dirac operator then there is an element ΠC ∈ Ψ0 (H; V ), V = V1 |H,
satisfying Π2C = ΠC and such that
(8.37) bH : {u ∈ C ∞ (M+ ; V1 ); Du = 0} −→ ΠC C ∞ (H; V )
is an isomorphism. The projection ΠC can be chosen so that
(8.38) bH : {u ∈ C ∞ (M− ; V1 ); Du = 0} −→ (Id −ΠC )C ∞ (H; V )
then ΠC is uniquely determined and is called the Calderòn projection.
This result remains true for a general elliptic operator of first order if (8.35)
is assumed, and even in a slightly weakened form without (8.35). Appropriate
modifications to the proofs below are consigned to problems.
For first order operators the jump formula discussed above takes the following
form.
Lemma 8.4. Let D be an elliptic differential operator of first order on M, acting
between vector bundles V1 and V2 . If u ∈ C ∞ (M+ ; V1 ) satisfies Du = 0 in M+ ◦
then
1
(8.39) Duc = σ1 (D)(dx)(bH u) · δ(x) ∈ C −∞ (M ; V2 ).
i
Since the same result is true for M− , with an obvious change of sign, D defines
a linear operator
(8.40) D : u ∈ L1 (M ; V1 ); u± = u|M± ∈ C ∞ (M± ; V1 ), Du± = 0 in M± ◦

−→
1
σ(D)(dx)(bH u+ − bH u− ) · δ(x) ∈ C ∞ (H; V2 ) · δ(x).
i
To define the Calderòn projection we shall use the ‘inverse’ of this result.
Proposition 8.7. If D ∈ Diff 1 (M ; V1 , V2 ) is elliptic and satisfies (8.35) then
(8.40) is an isomorphism, with inverse ID , and
 
1
(8.41) ΠC v = bH ID σ(D)(dx)v · δ(x) , v ∈ C ∞ (H; V1 ),
i +
satisfies the conditions of Theorem 8.1.
Proof. Observe that the map (8.40) is injective, since its null space consists
of solutions of Du = 0 globally on M ; such a solution must be smooth by elliptic
regularity and hence must vanish by the assumed invertibility of D. Thus the main
task is to show that D in (8.40) is surjective.
Since D is elliptic and, by assumption, an isomorphism on C ∞ sections over
M it is also an isomorphism on distributional sections. Thus the inverse of (8.40)
must be given by D−1 . To prove the surjectivity it is enough to show that
(8.42) D−1 (w · δ(x))|M± ∈ C ∞ (M± ; V1 ) ∀ w ∈ C ∞ (H; V2 ).
There can be no singular terms supported on H since w ·δ(x) ∈ H −1 (M ; V2 ) implies
that u = D−1 (w · δ(x)) ∈ L2 (M ; V1 ).
190 8. ELLIPTIC BOUNDARY PROBLEMS

Now, recalling that D−1 ∈ Ψ−1 (M ; V2 , V1 ), certainly u is C ∞ away from H.

At any point of H outside the support of w, u is also smooth. Since we may
decompose w using a partition of unity, it suffices to suppose that w has support in
a small coordinate patch, over which both V1 and V2 are trivial and to show that
u is smooth ‘up to H from both sides’ in the local coordinate patch. Discarding
smoothing terms from D−1 we may therefore replace D−1 by any local parametrix
Q for D and work in local coordinates and with components:
(8.43) Z
0 0
Qij (wj (y) · δ(x)) = (2π)−n ei(x−x )ξ+i(y−y )·η qij (x, y, ξ, η)w(y 0 )δ(x0 )dx0 dy 0 dξdη.

For a general pseudodifferential operator, even of order −1, the result we are seeking
is not true. We must use special properties of the symbol of Q, that is D−1 .

8.7. Rational symbols

Lemma 8.5. The left-reduced symbol of any local parametrix for a generalized
Dirac operator has an expansion of the form
(8.44)
∞
X
qij (z, ζ) = g(z, ζ)−2l+1 pij,l (z, ζ) with pij,l a polynomial of degree 3l − 2 in ζ;
l=1
here g(z, ζ) is the metric in local coordinates; each of the terms in (8.44) is therefore
a symbol of order −l.
Proof. This follows by an inductive arument, of a now familiar type. First,
the assumption that D is a generalized Dirac operator means that its symbol
σ1 (D)(z, ζ) has inverse g(z, ζ)−1 σ1 (D)∗ (z, ζ); this is the princiapl symbol of Q.
Using Leibniz’ formula one concludes that for any polynomial rl of degree j
∂ζi g(z, ζ)−2l+1 rj (z, ζ) = g(z, ζ)−2l rj+1
0


(z, ζ)
where rj+1 has degree (at most) j + 1. Using this result repeatedly, and proceeding
by induction, we may suppose that q = qk0 + qk+100
where qk0 has an expansion up
00
to order k, and so may be taken to be such a sum, and qk+1 is of order at most
−k − 1. The composition formula for left-reduced symbols then shows that
00
σ1 (D)qk+1 ≡ g −2k qk+1 mod S −k−1
where qk+1 is a polynomial of degree at most 3k. Inverting σ1 (D)(ζ) as at the initial
00
step then shows that qk+1 is of the desired form, g −2k−1 rk+1 with rk+1 of degree
3k + 1 = 3(k + 1) − 2, modulo terms of lower order. This completes the proof of
the lemma. 
With this form for the symbol of Q we proceed to the proof of Proposition 8.7.
That is, we consider (8.43). Since we only need to consider each term, we shall
drop the indicies. A term of low order in the amplitude qN gives an operator with
kernel in C N −d . Such a kernel gives an operator
C ∞ (H; V2 ) −→ C N −d (M ; V1 )
with kernel in C N −d . The result we want will therefore follow if we show that each
term in the expansion of the symbol q gives an operator as in (8.42).
To be more precise, we can assume that the amplitude q is of the form
q = (1 − φ)g −2l q 0
 8.9. INVERSES 191

where q 0 is a polynomial of degree 3l − 2 and φ = φ(ξ, η) is a function of compact

support which is identically one near the origin. The cutoff function is to remove
the singularity at ζ = (ξ, η) = 0. Using continuity in the symbol topology the
integrals in x0 and y 0 can be carried out. By assumption w ∈ Cc∞ (Rn−1 ), so the
resulting integral is absolutely convergent in η. If l > 1 it is absolutely convergent
in ξ as well, so becomes
Z
Q(w(y) · δ(x)) = (2π)−n eixξ+iy·η q(x, y, ξ, η)ŵ(η)dξdη.

In |ξ| > 1 the amplitude is a rational function of ξ, decaying quadratically as

ξ → ∞. If we assume that x > 0 then the exponential factor is bounded in the half
plane =ξ ≥ 0. This means that the limit as R → ∞ over the integral in =ξ ≥ 0
over the semicircle |ξ| = R tends to zero, and does so with uniform rapid decrease
in η. Cauchy’s theorem shows that, for R > 1 the real integral in ξ can be replaced
by the contour integral over γ(R), which is, forR >> |η| given by the real interval
[−R, R] together with the semicircle of radius R in the upper half plane. If |η| > 1
the integrand is meromorphic in the upper half plane with a possible pole at the
1
singular point g(x, y, ξ, η) = 0; this is at the point ξ = ir 2 (x, y, η) where r(x, y, η)
is a positive-definite quadratic form in η. Again applying Cauchy’s theorem
Z 1
Q(w(y)δ(x) = (2π)−n+1 i exr 2 (x,y,η)+iy·η q 0 (x, y, η)ŵ(η)dη

where q 0 is a symbol of order −k + 1 in η.

1
The product exr 2 (x,y,η) q 0 (x, y, η) is uniformly a symbol of order −k+1 in x > 1,
with x derivatives of order p being uniformly symbols of order −k + 1 + p. It follows
from the properties of pseudodifferential operators that Q(w · δ(x)) is a smooth
function in x > 0 with all derivatives locally uniformly bounded as x ↓ 0.

8.8. Proofs of Proposition 8.7 and Theorem 8.1

This completes the proof of (8.42), since a similar argument applies in x < 0,
with contour deformation into the lower half plane. Thus we have shown that
(8.40) is an isomorphism which is the first half of the statement of Proposition
(8.7). Furthermore we see that the limiting value from above is a pseudodifferential
operator on H :
(8.45) Q0 w = lim D−1 (w · δ(x)), Q0 ∈ Ψ0 (H; V2 , V1 ).
x↓0

This in turn implies that ΠC , defined by (8.41) is an element of Ψ0 (H; V1 ), since it

is Q0 ◦ 1i σ(D)(dx).
Next we check that ΠC is a projection, i.e.
that Π2C = ΠC . If w = ΠC v, v ∈ C ∞ (H; V1 ), then w = bH u, u = ID 1i σ(D)(dx)v|M+ ,
so u ∈ C ∞ (M+ ; V1 ) satisfies Du = 0 in M+ ◦
. In particular, by (8.39), P uc =
1 2
i σ 1 (D)(dx)w · δ(x), which means that w = Π C w so ΠC = ΠC . This also shows
that the range of ΠC is precisely the range of bH as stated in (8.37). The same
argument shows that this choice of the projection gives (8.38). 

8.9. Inverses
Still for the case of a generalized Dirac operator on a compact manifold with
dividing hypersurface, consider what we have shown. The operator D defines a
192 8. ELLIPTIC BOUNDARY PROBLEMS

map in (8.39) with inverse

(8.46) ID : {v ∈ C ∞ (H; V1 ); ΠC v = v} −→ {u ∈ C ∞ (M+ ; V1 ); Du = 0 in M+ }.
This operator is the ‘Poisson’ operator for the canonical boundary condition given
by the Calderòn operator, that is u = ID v is the unique solution of
(8.47) Du = 0 in M+ , u ∈ C ∞ (M+ ; V1 ), ΠC bH u = v.
We could discuss the regularity properties of ID but we shall postpone this until
after we have treated the ‘one-sided’ case of a genuine boundary problem.
As well as ID we have a natural right inverse for the operator D as a map from
C ∞ (M+ ; V1 ) to C ∞ (M− ; V2 ). Namely
Lemma 8.6. If f ∈ C ∞ (M+ ; V2 ) then u = D−1 (fc )|M+ ∈ C ∞ (M+ ; V1 ) and the
map RD : f 7−→ u is a right inverse for D, i.e.
D ◦ RD = Id .
Proof. Certainly D(D−1 (fc ) = fc , so u = D−1 (fc )|M+ ∈ C −∞ (M+ ; V1 ) sat-
ifies Du = f in the sense of extendible distributions. Since f ∈ C ∞ (M+ ; V2 ) we
can solve the problem Du ≡ f in the sense of Taylor series at H, with the con-
stant term freely prescibable. Using Borel’s lemma, let u0 ∈ C ∞ (M+ ; V1 ) have the
appropriate Taylor series, with bH u0 = 0.. Then D(u0c ) − fc = g ∈ C˙∞ (M +; V2 ).
Thus u00 = D−1 g ∈ C ∞ (M ; V1 ). Since D(u0 − u00 ) = fc , the uniqueness of solutions
implies that u = (u0 − u00 )|M+ ∈ C ∞ (M+ ; V1 ). 

Of course RD cannot be a two-sided inverse to D since it has a large null space,

described by ID .
Problem 8.2. Show that, for D as in Theorem 8.1 if f ∈ C ∞ (M+ ; V2 ) and v ∈
C (H; V1 ) there exists a unique u ∈ C ∞ (M+ ; V2 ) such that Du = f in C ∞ (M+ ; V2 )
∞

and bH u = ΠC v.

8.10. Smoothing operators

The properties of smoothing operators on a compact manifold with boundary
are essentially the same as in the boundaryless case. Rather than simply point to
the earlier discussion we briefly repeat it here, but in an abstract setting.
Let H be a separable Hilbert space. In the present case this would be L2 (X)
or L2 (X; E) for some vector bundle over X, or some space H m (X; E) of Sobolev
sections. Let B = B(H) be the algebra of bounded operators on H and K = K(H)
the ideal of compact operators. Where necessary the norm on B will be written
k kB ; K is a closed subspace of B which is the closure of the ideal F = F(H) of
finite rank bounded operators.
We will consider a subspace J = J (H) ⊂ B with a stronger topology. Thus we
suppose that J is a Fréchet algebra. That is, it is a Fréchet space with countably
many norms k kk such that for each k there exists k 0 and Ck with
(8.48) kABkk ≤ Ck kAkk0 kBkk0 ∀ A, B ∈ J .
In particular of course we are supposing that J is a subalgebra (but not an ideal)
in B. To make it a topological *-subalgebra we suppose that
(8.49) kAkB ≤ CkAkk ∀ A ∈ J , ∗ : J −→ J .
 8.10. SMOOTHING OPERATORS 193

In fact we may suppose that k = 0 by renumbering the norms. The third condition
we impose on J implies that it is a subalgebra of K, namely we insist that
(8.50) F ∩ J is dense in J ,
in the Fréchet topology. Finally, we demand, in place of the ideal property, that J
be a bi-ideal in B (also called a ‘corner’) that is,
(8.51) A1 , A2 ∈ J , B ∈ B =⇒ A1 BA2 ∈ J ,
0
∀ k ∃ k such that kA1 BA2 kk ≤ CkA1 kk0 kBkB kA2 kk0 .
Proposition 8.8. The space of operators with smooth kernels acting on sec-
tions of a vector bundle over a compact manifold satisfies (8.48)–(8.52) with H =
H m (X; E) for any vector bundle E.
Proof. The smoothing operators on sections of a bundle E can be written as
integral operators
Z
(8.52) Au(x) = A(x, y)u(y) , A(x, y) ∈ C ∞ (X 2 ; Hom(E) ⊗ ΩR ) .
E
Thus J = C ∞ (X 2 ; Hom(E) ⊗ ΩR ) and we make this identification topological. The
norms are the C k norms. If P1 , . . . , pN (m) is a basis, on C ∞ (X 2 ), for the differential
operators of order m on Hom(E) ⊗ ΩL then we may take
(8.53) kAkm = sup kPj AkL∞
j
∗ ∗ ∗
for some inner products on the bundles. In fact Hom(E) = πL E ⊗ πR E from it
which follows easily that this is a basis Pj = Pj,k ⊗ Pj,R decomposing as products.
From this (8.48) follows easily since
(8.54) kABkm = sup k(PjL A) · (Pj,R B)k∞ kABkL∞ ≤ CkAkL∞ kBkL∞
j

by the compactnes of X. From this (8.53) follows with k = 0.

The density (8.50) is the density of the finite tensor product C ∞ (X; E) ⊗ C ∞
(X; E ∗ ⊗ ΩL ) in C ∞ (X 2 ; Hom(E) ⊗ ΩL ). This follows from the boundaryless case
by doubling (or directly). Similarly the bi-ideal condition (8.52) can be seen from
the regularity of the kernel. A more satisfying argument using distribution theory
follows from the next result.

Proposition 8.9. An operator A : C˙∞ (X; E) → C −∞ (X; F ) is a smoothing
operator if and only if it extends by continuity to C˙−∞ (X; E) and then has range
in C ∞ (X; F ) ,→ C −∞ (X; F ).
Proof. If A has the stated mapping property then compose with a Seeley
extension operator, then EA = Ã is a continuous linear map
à : C˙−∞ (X; E) → C ∞ (X̃; F̃ ) ,
for an extension of F to F̃ over the double X̃. Localizing in the domain to trivialize
E and testing with a moving delta function we recover the kernel of à as
Ã(x, y) = Ã · δy ∈ C ∞ (X̃; F̃ ) .
Thus it follows that à ∈ C ∞ (X̃ × X; Hom(E, F̃ ) ⊗ ΩR ). The converse is more
obvious.

 194 8. ELLIPTIC BOUNDARY PROBLEMS

Returning to the general case of a bi-ideal as in (8.48)–(8.52) we may consider

the invertibility of Id +A, A ∈ J .
Proposition 8.10. If A ∈ J , satisfying (8.48)–(8.52), then Id +A has a gen-
eralized inverse of the form Id +B, B ∈ J , with
AB = Id −πR , BA = Id −πL ∈ J ∩ F
both finite rank self-adjoint projections.
Proof. Suppose first that A ∈ J and kAkB < 1. Then Id +A is invertible in
B with inverse Id + B ∈ B,
X
(8.55) B= (−1)j Aj .
j≥1

Not only does this Neumann series converge in B but also in J since for each k
(8.56) kAj kk ≤ Ck kAkk0 kAj−2 kB kAkk0 ≤ Ck0 kAkj−2
B , j ≥ 2.
Thus B ∈ J , since by assumption J is complete (being a Fréchet space). In this
case Id +B ∈ B is the unique two-sided inverse.
For general A ∈ J we use the assumed approximability in (8.50). Then A =
A0 + A00 when A0 ∈ F ∩ J and kA00 kB ≤ CkA00 kk < 1 by appropriate choice. It
follows that Id +B 00 = (Id +A00 )−1 is the inverse for Id +A00 and hence a parameterix
for Id +A:
(8.57) (Id +B 00 )(Id +A) = Id +A0 + B 00 A0
(Id +A)(Id +B 00 ) = Id +A0 + A0 B 00
Unfinished. with both ‘error’ terms in F ∩ J . 
Lemma on
subprojec-
tions. 8.11. Left and right parametrices
Suppose that H1 and H2 are Hilbert spaces and A : H1 −→ H2 is a bounded
linear operator between them. Let J1 ⊂ B(H1 ) and J2 ⊂ B(H2 ) be bi-ideals as in
the previous section. A left parametrix for A, modulo J1 , is a bounded linear map
BL : H2 −→ H1 such that
(8.58) BL ◦ A = Id +JL , JL ∈ J1 .
Similarly a right parametrix for A, modulo J2 is a bounded linear map BR : H2 −→
H1 such that
(8.59) A ◦ BR = Id +JR , JR ∈ J2 .
Proposition 8.11. If a bounded linear operator A : H1 −→ H2 has a left
parametrix BL modulo a bi-ideal J1 , satisfying (8.48)–(8.52), then A has closed
range, null space of finite dimension and there is a generalized left inverse, differing
from the original left parametrix by a term in J1 , such that
(8.60) BL ◦ A = Id −πL , πL ∈ J1 ∩ F,
with πL the self-adjoint projection onto the null space of A.
 8.12. RIGHT INVERSE 195

Proof. Applying Proposition 8.10, Id +JL has a generalized inverse Id +J,

0 0
J ∈ J1 , such that (Id +J)(Id +JL ) = (Id −πL ), πL ∈ J1 ∩ F. Replacing BL by
0
B̃L = (Id +J)BL gives a new left parametrix with error term πL ∈ J1 ∩ F. The
0
null space of A is contained in the null space of BL ◦ A and hence in the range
of FL ; thus it is finite dimensional. Furthermore the self-dajoint projection πL
0
onto the null space is a subprojection of πL , so is also an element of J1 ∩ F.
The range of A is closed since it has finite codimension in Ran(A(Id −πL )) and
if fn ∈ Ran(A(Id −πL )) = Aun , un = (Id −πL )un , converges to f ∈ H2 , then
un = BL fn converges to u ∈ H1 with A(Id −πL )u = f. 

Proposition 8.12. If a bounded linear operator A : H1 −→ H2 has a right

parametrix BR modulo a bi-ideal J2 , satisfying (8.48)–(8.52), then it has closed
range of finite codimension and there is a generalized right inverse, differing from
the original right parametrix by a term in J2 , such that
(8.61) A ◦ BR = Id −πR , πR ∈ J2 ∩ F,
with Id −πR the self-adjoint projection onto the range space of A.
Proof. The operator Id +JR has, by Proposition 8.10, a generalized inverse
0
Id +J with J ∈ J1 . Thus BR = BR ◦ (Id +J) is a right parametrix with error term
0 0
Id −πR , πR ∈ J1 ∩ F being a self-adjoint projection. Thus the range of A contains
0
the range of Id −πR and is therefore closed with a finite-dimensional complement.
Furthemore the self-adjoint projection onto the range of A is of the form Id −πR
0
where πR is a subprojection of πR , so also in J1 ∩ F. 

The two cases, of an operator with a right or a left parametrix are sometimes
combined in the term ‘semi-Fredholm.’ Thus an operator A : H1 −→ H2 is semi-
Fredholm if it has closed range and either the null space or the orthocomplement
to the range is finite dimensional. The existence of a right or left parametrix,
modulo the ideal of compact operators, is a necessary and sufficient condition for
an operator to be semi-Fredholm.

8.12. Right inverse

In treating the ‘general’ case of an elliptic operator on compact manifold
with boundary we shall start by constructing an analogue of the right inverse in
Lemma 8.6. So now we assume that D ∈ Diff 1 (X; V1 , V2 ) is an operator of Dirac
type on a compact manifold with boundary.
To construct a right inverse for D we follow the procedure in the boundaryless
case. That is we use the construction of a pseudodifferential parametrix. In order
to make this possible we need to extend M and D ‘across the boundary.’ This
is certainly possible for X, since we may double it to a compact manifold without
boundary, 2X. Then there is not obstruction to extending D ‘a little way’ across the
boundary. We shall denote by M an open extension of X (of the same dimension)
so X ⊂ M is a compact subset and by D̃ an extension of Dirac type to M.
The extension of D to D̃, being elliptic, has a parametrix Q̃. Consider the map
(8.62) Q̃0 : L2 (X; V2 ) −→ H 1 (X; V1 ), Q̃0 f = Q̃fc |X
where fc is the extension of f to be zero outside X. Then Q̃0 is a right parametrix,
DQ̃0 = Id +E where E is an operator on L2 (X; V2 ) with smooth kernel on X 2 .
196 8. ELLIPTIC BOUNDARY PROBLEMS

Following Proposition 8.12, D has a generalized right inverse Q̃00 = Q̃0 (Id +E 0 ) up
to finite rank smoothing and
(8.63) D : H 1 (X; V1 ) ←→ L2 (X; V2 )
has closed range with a finite dimensional complement in C ∞ (X; V2 ).
Proposition 8.13. The map (8.63) maps C ∞ (X; V2 ) to C ∞ (X; V1 ), it is sur-
jective if and only if the only solution of D∗ u = 0, u ∈ C˙∞ (X; V2 ) is the trivial
solution.
Proof. The regularity statement, that Q0 C ∞ (X; V ) ⊂ C ∞ (X; V1 ) follows as
in the proof of Lemma 8.6. Thus Q0 maps C ∞ (X; V1 ) to C ∞ (X; V2 ) if and only
if any paramatrix Q̃0 does so. Given f ∈ C ∞ (X; V2 ) we may solve Du0 ≡ f in
Taylor series at the boundary, with u0 ∈ C ∞ (X; V1 ) satisfying bH u0 = 0. Then
D(u0 )c − f ∈ C˙∞ (X; V2 ) so it follows that Q0 (fc )|X ∈ C ∞ (X; V1 ).
Certainly any solution of D∗ u = 0 with u ∈ C˙∞ (X; V2 ) is orthogonal to the
range of (8.63) so the condition is necessary. So, suppose that (8.63) is not surjec-
tive. Let f ∈ L2 (X; V2 ) be in the orthocomplement to the range. Then Green’s
formula gives the pairing with any smooth section
(Dv, f )X = (Dṽ, fc )X̃ = (ṽ, D∗ fc )X̃ = 0.
This means that D∗ fc = 0 in X̃, that is as a supported distribution. Thus, f ∈
C˙∞ (X; V2 ) satisfies D∗ f = 0. 

As noted above we will proceed under the assumption that D∗ f has no such
non-trivial solutions in C˙∞ (X; V2 ). This condition is discussed in the next section.
Theorem 8.2. If unique continuation holds for D∗ then D has a right inverse
(8.64) Q : C ∞ (X : V2 ) −→ C ∞ (X; V1 ), DQ = Id
where Q = Q̃0 + E, Q̃0 f = Q̃f |X where Q̃ is a parametrix for an extension of D
across the boundary and E is a smoothing operator on X.
Proof. As just noted, unique continuation for D∗ implies that D in (8.63)
is surjective. Since the parametrix maps C ∞ (X; V2 ) to C ∞ (X; V1 ), D must be
surjective as a map from C ∞ (X; V1 ) to C ∞ (X; V2 ). The parametrix modulo finite
rank operators can therefore be corrected to a right inverse for D by the addition
of a smoothing operator of finite rank. 

8.13. Boundary map

The map b from C ∞ (X; E) to C ∞ (∂X; E) is well defined, and hence is well
defined on the space of smooth solutions of D. We wish to show that it has closed
range. To do so we shall extend the defintion to the space of square-integrable
solutions. For any s ∈ R set
(8.65) N s (D) = {u ∈ H s (X; E); Du = 0} .
Of course the equation Du = 0 is to hold in the sense of extendible distributions,
which just means in the interior of X. Thus N ∞ (D) is the space of solutions of D
smooth up to the boundary.
 8.14. CALDERÒN PROJECTOR 197

Lemma 8.7. If u ∈ N 0 (D) then

1
(8.66) Ḋuc = v · δ(x), v ∈ H − 2 (∂X; E)
1
defines an injective bounded map b̃ : N 0 (D) −→ H − 2 (∂X; E) by b̃(u) = iσ(D)(dx)v
which is an extension of b : N ∞ (D) −→ C ∞ (∂X; E) defined by restriction to the
boundary.
Proof. Certainly Ḋuc ⊂ C˙∂X
∞
(X; E) has support in the boundary, so is a sum
of products in any product decomposition of X near ∂X,
X
D(uc ) = vj · δ (j) (x).
j

Since D is a first order operator and uc ∈ L2 (X̃; E), for any local extension,
Ḋuc ∈ Ḣ −1 (X; E). Localizing so that E is trivial and the localized vj have compact
supports this means that
1
(8.67) (1 + |η|2 + |ξ|2 )− 2 vbj (η)ξ j ∈ L2 (Rn ).
If vj 6= 0 for some j > 0 this is not true even in some region |η| < C. Thus vj ≡ 0
for j > 0 and (8.66) must hold. Furthermore integration in ξ gives
Z
1
(1 + |η|2 + |ξ|2 )−1 dξ = c(1 + |η|2 )− 2 , c > 0, so
R
(8.68) Z
1
(1 + |η|2 )− 2 |v̂(η)|2 dη < 0.
Rn−1
− 12
Thus v ∈ H (∂X; E) and b̃ is well defined. The jumps formula shows it to
be an extension of b. The injectivity of b̃ follows from the assumed uniqueness of
solutions to Ḋu = 0 in X. 
1
Notice that (8.68) is actually reversible. That is if v ∈ H − 2 (∂X; E) then
v · δ(x) ∈ H −1 (X; E). This is the basis of the construction of a left parametrix for
b̃, which then shows its range to be closed.
Lemma 8.8. The boundary map b̃ in Lemma 8.7 has a continuous left paramet-
−1 0
D : H 2 (∂X; E) −→ N (D), ID ◦ b̃ = Id +G, where G has smooth kernel on
rix If
1
X × ∂X, and the range of b̃ is therefore a closed subspace of H − 2 (∂X; E).

Proof. The parametrix If D is given directly by the parametrix Q̃ for D̃, and
extension to X̃. Applying Q̃ to (8.66) gives
1
(8.69) u = If D v + Ru, ID = RX ◦ Q̃ ◦ σ(D)(dx)
f
i
−1 2
with R having smooth kernel. Since If D is bounded from H 2 (∂X; E) to L (X; E)
and R is smoothing it follows from Proposition 8.11 that the range of b̃ is closed. 

8.14. Calderòn projector

1
Having shown that the range of b̃ in Lemma 8.7 is closed in H − 2 (∂X; E) we
now deduce that there is a pseudodifferential projection onto it. The discussion
above of the boundary values of the Q̃(w · δ(x)) is local, and so applies just as well
 198 8. ELLIPTIC BOUNDARY PROBLEMS

in the present more general case. Since this is just the definition of the map If
D in
Lemma 8.8, we conclude directly that
∞
(8.70) P v = lim
◦
D v, v ∈ C (∂X; E)
If
X
0
defines P ∈ Ψ (∂X; E).
Lemma 8.9. If P is defined by (8.70) then P 2 − P ∈ Ψ−∞ (∂X; E) and there
exist A, B ∈ Ψ−∞ (∂X; E) such that P − Id = A on Ran(b̃) and Ran(P + B) ⊂
Ran(b̃).
Proof needs
clarification.
Proof. That P 2 − P ∈ Ψ−∞ (∂X; E) follows, as above, from the fact that Q̃
is a two-sided parametrix on distributions supported in X. Similarly we may use
1
the right inverse of D to construct B. If v ∈ H − 2 (∂X; E) then by construction,
0
DIf
Dv = R v

where R0 has a smooth kernel on X × ∂X. Applying the right inverse Q it follows
that u0 = If 0 0 0
D v −(Q◦R )v ∈ N (D), where Q◦R also has smooth kernel on X ×∂X.
Thus b̃(u0 ) = (P + B)v ∈ Ran(b̃) where B has kernel arising from the restriction of
the kernel of A ◦ R0 to ∂X × ∂X, so B ∈ Ψ−∞ (∂X; E). 
Now we may apply Proposition 6.11 with F = Ran(b̃) and s = − 12 to show the
existence of a Calderòn projector.
Proposition 8.14. If D is a generalized Dirac operator on X then there is an
1
element ΠC ∈ Ψ0 (∂X; E) such that Π2C = ΠC , Ran(ΠC ) = Ran(b̃) on H − 2 (∂X; E),
−∞
ΠC − P ∈ Ψ (∂X; E) where P is defined by (8.70) and Ran(ΠC ) = Ran(b) on
C ∞ (∂X; E).
Proof. The existence of psuedodifferential projection, ΠC , differing from P
by a smoothing operator and with range Ran(b̃) is a direct consequence of the
application of Proposition 6.11. It follows that Ran(b̃) ∩ C ∞ (∂X; E) is dense in
1
Ran(b̃) in the topology of H − 2 (∂X; E). Furthermore, if follows that if v ∈ Ran(b̃) ∩
C (∂X; E) then u ∈ N (D) such that b̃u = v is actually in C ∞ (X; E), i.e.
∞ 0

it is in N ∞ (D). Thus the range of b is just Ran(b̃) ∩ C ∞ (∂X; E) so Ran(b) is the

range of ΠC acting on C ∞ (∂X; E). 
In particular b̃ is just the continuous extension of b from N ∞ (D) to N 0 (D), of
which it is a dense subset. Thus we no longer distinguish between these two maps
and set b̃ = b.

8.15. Poisson operator

8.16. Unique continuation
8.17. Boundary regularity
8.18. Pseudodifferential boundary conditions
The discussion above shows that for any operator of Dirac type the ‘Calderòn
realization’ of D,
1
(8.71) DC : {u ∈ H s (X; E1 ); ΠC bu = 0} −→ H s−1 (X; E2 ), s >
2
 8.18. PSEUDODIFFERENTIAL BOUNDARY CONDITIONS 199

is an isomorphism.
We may replace the Calderòn projector in (8.71) by a more general projection
Π, acting on C ∞ (∂X, V1 ), and consider the map
(8.72) DΠ : {u ∈ C ∞ (X; V1 ); Πbu = 0} −→ C ∞ (X; V2 ).
In general this map will not be particularly well-behaved. We will be interested in
the case that Π ∈ Ψ0 (∂X; V1 ) is a pseudodifferential projection. Then a condition
for the map DΠ to be Fredholm can be given purely in terms of the relationship
between Π and the (any) Calderòn projector ΠC .
Theorem 8.3. If D ∈ Diff 1 (X; E1 , E2 ) is of Dirac type and P i ∈ Ψ0 (∂X; E1 )
is a projection then the map
D
(8.73) DΠ : {u ∈ C ∞ (X; E1 ); Π(u∂X ) = 0} −→ C ∞ (X; E2 )
is Fredholm if and only if
(8.74) Π ◦ ΠC : Ran(ΠC ) ∩ C ∞ (∂V1 ) −→ Ran(Π) ∩ C ∞ (∂E1 ) is Fredholm
and then the index of DΠ is equal to the relative index of ΠC and Π, that is the
index of (8.74).
Below we give a symbolic condition equivalent which implies the Fredholm con-
dition. If enough regularity conditions are imposed on the generalized inverse to
(8.71) then this symbolic is also necessary.
Proof. The null space of DΠ is easily analysed. Indeed Du = 0 implies that
u ∈ N , so the null space is isomorphic to its image under the boundary map:
b
{u ∈ N ; Πbu = 0} −→ {v ∈ C; Πv = 0} .
Since C is the range of ΠC this gives the isomorphism
(8.75) Nul(DΠ ) ' Nul (Π ◦ ΠC : C −→ Ran(Π)) .
In particular, the null space is finite dimensional if and only if the null space of
Π ◦ ΠC is finite dimensional.
Similarly, consider the range of DΠ . We construct a map
(8.76) τ : C ∞ (∂X; V1 ) −→ C ∞ (X; V2 )/ Ran(DΠ ).
Indeed each v ∈ C ∞ (∂X; V1 ) is the boundary value of some u ∈ C ∞ (X : V1 ), let
τ (v) be he class of DU. This is well-defined since any other extension u0 is such
that b(u − u0 ) = 0, so D(u − u0 ) ∈ Ran(DΠ ). Furthermore, τ is surjective, since DC
is surjective. Consider the null space of τ. This certainly contains the null space of
Π. Thus consider the quotient map
τ̃ : Ran(Π) −→ C ∞ (X : V2 )/ Ran(DΠ ),
which is still surjective. Then τ̃ (v) = 0 if and only if there exists v 0 ∈ C such that
Π(v − v 0 ) = 0. That is, τ̃ (v) = 0 if and only if Π(v) = Π ◦ ΠC . This shows that the
finer quotient map
(8.77) τ 0 : Ran(Π)/ Ran(Π ◦ ΠC ) ←→ C ∞ (X; V2 )/ Ran(DΠ )
is an isomorphism. This shows that the range is closed and of finite codimension if
Π ◦ ΠC is Fredholm.
The converse follows by reversing these arguments. 
200 8. ELLIPTIC BOUNDARY PROBLEMS

8.19. Gluing
Returning to the case of a compact manifold without boundary, M, with a
dividing hypersurface H we can now give a gluing result for the index.
Theorem 8.4. If D ∈ Diff 1 (M ; E1 , E2 ) is of Dirac type and M = M1 ∩ M2 is
the union of two manifolds with boundary intersecting in their common boundary
∂M1 ∩ ∂M2 = H then
(8.78) Ind(D) = Ind(Π1,C , Id −Π2,C ) = Ind(Π2,C , Id −Π1,C )
where Πi,C , i = 1, 2, are the Calderòn projections for D acting over Mi .

8.20. Local boundary conditions

8.21. Absolute and relative Hodge cohomology
8.22. Transmission condition