ESI The Erwin Schrodinger International

Institute for Mathematical Physics

Boltzmanngasse 9
A-1090 Wien, Austria

Scattering Matrix in Conformal Geometry

C. Robin Graham
Maciej Zworski

Vienna, Preprint ESI 1068 (2001) September 14, 2001

Supported by Federal Ministry of Science and Transport, Austria

Available via anonymous ftp from FTP.ESI.AC.AT
or via WWW, URL: http://www.esi.ac.at
 SCATTERING MATRIX IN CONFORMAL GEOMETRY

C. ROBIN GRAHAM AND MACIEJ ZWORSKI

1. Statement of the results

This paper describes the connection between scattering matrices on conformally
compact asymptotically Einstein manifolds and conformally invariant objects on
their boundaries at infinity. This connection is a manifestation of the general prin-
ciple that the far field phenomena on a conformally compact Einstein manifold
are related to conformal theories on its boundary at infinity. This relationship
was proposed in [5] as a means of studying conformal geometry, and the prin-
ciple forms the basis of the AdS/CFT correspondence in quantum gravity see
[16],[26],[12],[10] and references given there.
We first define the basic objects discussed here. By a conformal structure on
a compact manifold M we mean an equivalence class [h] determined by a metric
representative h:
h [h] h = e2 h , C (M ) .
Let X be a compact n + 1-manifold with X = M , or in the case M does not
bound a manifold, with X = M t M . This can be achieved trivially by putting
X = [0, 1] M .
Let x be a defining function of X in X:
x| > 0 , x|X = 0 , dx|X 6= 0 .
X
We say that g is a conformally compact metric on X with conformal infinity [h] if
g
(1.1) g = 2 , g|T X [h] ,
x
where g is a smooth metric on X. Since we can choose different defining functions,
the metric g determines only the conformal class [h]. A conformally compact metric
is said to be asymptotically hyperbolic if its sectional curvatures approach 1 at
X; this is equivalent to |dx|g = 1 on X. The basic example is hyperbolic space
Hn+1 , with boundary given by Rn in the half-space model and by Sn in the ball
model: the two being conformally equivalent.
One of the results of [5] is that given a conformal structure [h] on M , one can
construct a conformally compact metric g with conformal infinity [h] which satisfies

O(x ) for n odd
(1.2) Ric(g) + ng =
O(xn2 ) for n even.
1
2 C. ROBIN GRAHAM AND MACIEJ ZWORSKI

When n is even, the condition (1.2) is augmented by a vanishing trace condition

to the next order. We call a metric g satisfying these conditions asymptotically
Einstein. When n is odd, the condition (1.2) together with an asymptotic evenness
condition uniquely determine g mod O(x ) up to diffeomorphism. We shall call a
metric g which is asymptotically Einstein and which also satisfies the asymptotic
evenness condition, a Poincare metric associated to [h].
Our first theorem relates the scattering matrix of a Poincare metric g to the
conformally invariant powers of the Laplacian on (M, [h]). The scattering matrix
of (X, g) is a meromorphic family S(s) of pseudodifferential operators on M defined
in terms of the behaviour at infinity of solutions of [ g s(n s)]u = 0, which we
discuss in 3. The conformally invariant powers of the Laplacian are a family P k ,
k N and k n/2 if n is even, of scalar differential operators on M constructed
in [8], which we discuss in 4. These operators are natural in the sense that they
can be written in terms of covariant derivatives and curvature of a representative
metric h, and they are invariant in the sense that if h = e2 h, then
(1.3) Pbk = e(n/2k) Pk e(n/2k) .
The operator Pk has the same principal part as k (in our convention the Laplacian
is a positive operator) and equals k if h is flat.
Theorem 1. Let (M n , [h]) be a compact manifold with a conformal structure, and
let (X, g) be a Poincare metric associated to [h]. Suppose that k N and k n/2
if n is even, and that (n/2)2 k 2 is not an L2 -eigenvalue of g . If S(s) is the
scattering matrix of (X, g), and Pk the conformally invariant operator on M , then
S(s) has a simple pole at s = n/2 + k and
(1.4) ck Pk = Res s=n/2+k S(s) , ck = (1)k [22k k!(k 1)!]1 ,
where Res s=s0 S(s) denotes the residue at s0 of the meromorphic family of opera-
tors S(s).
We remark that the condition on the spectrum is automatically satisfied if k
n/2 and in general can be guaranteed by perturbing the metric g in the interior.
Since S(s) is self-adjoint for s R, as an immediate consequence of Theorem 1
and of this remark we obtain
Corollary. The conformally invariant operators, Pk , are self-adjoint.
This was previously known only for small values of k for which the operators can
be explicitly calculated. The first two of the invariant operators are the conformal
Laplacian
n2
P1 = + R
4(n 1)
and the Paneitz operator
n
P2 = 2 + T d + (n 4)(J + J 2 2|P |2 )/2.
2
 SCATTERING MATRIX IN CONFORMAL GEOMETRY 3

1
Here R denotes the scalar curvature, J = R/(2(n 1)), Pij = n2 (Rij Jhij )
where Rij is the Ricci curvature, T = (n 2)Jh 4P acting as an endomorphism
on 1-forms, |P |2 = Pij P ij , and is the adjoint of d (the divergence operator).
Another important notion of conformal geometry is Bransons Q-curvature in
even dimensions. It is a scalar function on M constructed from the curvature
tensor and its covariant derivatives, with an invariance property generalizing that
of scalar curvature in dimension two: if once again h = e2 h, then

(1.5) b = Q + Pn/2 .
en Q

There has been great progress recently in understanding the Q-curvature and its
geometric meaning in low dimensions and on conformally flat manifolds see [4]
for an example of recent work. However, in general it remains a rather mysterious
quantity its definition (given in [2] and reviewed in 4 below) in the general
case is via analytic continuation in the dimension. In dimension two it is given by
Q = R/2, and in dimension four by 6Q = R + R 2 3|Ric|2 .
If n is even, then the operator Pn/2 has no constant term, i.e. Pn/2 1 = 0. It
therefore follows from Theorem 1 that S(s)1 extends holomorphically across s = n,
so S(n)1 = limsn S(s)1 is a well-defined function on M .
Theorem 2. With the notation of Theorem 1, for n even, we have

(1.6) cn/2 Q = S(n)1 .

Theorem 2 can be used as an alternative definition of the Q-curvature. We

show in 4 that the conformal transformation law (1.5) is an easy consequence of
Theorems 1 and 2. ItRfollows from (1.5), the self-adjointness of Pn/2 , and the fact
that Pn/2 1 = 0, that M Q is a conformal
R invariant. For (M, [h]) conformally flat,
it follows from a result of [3] that M Q is a multiple of the Euler characteristic
(M ).
A specific mathematical object which appeared in the study of the AdS/CFT
correspondence is the renormalized volume of an asymptotically hyperbolic mani-
fold (X, g) see [7] for a discussion and references. It has also appeared in geomet-
ric scattering theory see [11],[15]. As shown in [9] (Lemma 5.2 and the subsequent
paragraph), a choice of metric h in the conformal class on X uniquely determines
a defining function x near X and an identification of a neighborhood of X with
X [0, ] such that g takes the form

(1.7) g = x2 (hx + dx2 ), h0 = h,

where hx is a 1-parameter family of metrics on X. The renormalized volume

is defined as the finite part in the expansion of volg ({x > }) as 0. For
4 C. ROBIN GRAHAM AND MACIEJ ZWORSKI

asymptotically Einstein metrics the expansions take a special form

volg ({x > }) = c0 n + c2 n+2 + + cn1 1 + V + o(1)

for n odd,
(1.8)
volg ({x > }) = c0 n + c2 n+2 + + cn2 2 + L log(1/) + V + o(1)

for n even.
It turns out that for asymptotically Einstein metrics g, V is independent of the
conformal representative h on the boundary at infinity when n is odd, and L is
independent of the conformal representative when n is even. The dependence of
V on the choice of h for n even is the so-called holographic anomaly see [12],[7].
Anderson [1] has recently identified V when n = 3. In an appendix to [23], Epstein
shows that for conformally compact hyperbolic manifolds, the invariants L for n
even and V for n odd are each multiples of the Euler characteristic (X).
Using the connection with the scattering matrix, we are able to identify L in
terms of the Q-curvature:
Theorem 3. Let n be even and let L be defined by (1.8). Then
Z
(1.9) L = 2cn/2 Q,
M
where cn/2 is defined in (1.4).
We should stress that despite the fact that the scattering matrix is a global
object, in some sense our results are all formal Taylor series statements at the
boundary of X. In fact, in [6] it will be shown that, using a variant of the ideas
introduced here, a direct definition of Q and proofs of Theorem 3 and the self-
adjointness of the Pk s can be given based purely on formal asymptotics, avoiding
the analytic continuation via the scattering matrix. It is nevertheless worthwhile
to proceed with the full scattering theory: as a byproduct, this allows us to clarify
certain confusing issues about the infinite rank poles of the scattering matrix at
s = n/2 + l/2, l N.
The paper is organized as follows. In 2 we present a simple one dimensional
introduction to the relevant aspects of scattering theory. The more involved theory
for asymptotically hyperbolic manifolds is then discussed in 3: using results of
Mazzeo-Melrose [19] we give direct arguments for the existence and properties of
the Poisson operator and the scattering matrix, focussing particularly on their
behaviour for s near n/2 + N/2. Although it is not relevant for our main results, at
the end of 3 we outline a method for the study of the structure of the scattering
matrix complementing the treatment in [14]. In 4 we show how the invariant
 SCATTERING MATRIX IN CONFORMAL GEOMETRY 5

operators Pk may be constructed from a Poincare metric and discuss Bransons

Q-curvature. Finally, 5 combines scattering theory with 4, providing proofs of
the main results.
In the paper, N = {1, 2, } denotes the natural numbers and N 0 = N {0}.
Linear operators are identified with their distributional (integral) kernels.
Acknowledgments. We would like to thank Rafe Mazzeo and Andras Vasy for
helpful comments. The first author thanks the Mathematical Sciences Research
Institute, where this project was started. The second author thanks the Erwin
Schrodinger Institute, where it was continued, and the National Science Foundation
for partial support under the grant DMS-9970614.

2. A simple example
As a simple-minded introduction to scattering and as an illustration of the con-
nection between residues of the scattering matrix and asymptotic expansions we
present a one-dimensional example.
Thus we consider scattering on a half line, X = [0, ) with a compactly sup-
ported real valued potential V Lcomp (X; R). The quantum Hamiltonian is given
by H = y2 +V (y) and we impose, say, the Dirichlet boundary condition at y = 0.
We are interested in the properties of generalized eigenfunctions at energies s 2 :
(H + s2 )u(y) = 0 , u(0) = 0 ,
which for large values of y and for s 6= 0 satisfy
(2.1) u(y) = A(s)esy + B(s)esy .
For s iR \ {0}, consideration of the Wronskian of u and u shows that |A(s)| 2 =
|B(s)|2 . Normalizing, we define the scattering matrix, S(s) (which in this case is
a one-by-one matrix!) by
B(s)
S(s) = .
A(s)
This defines a meromorphic function of s C, regular for s iR. The definition
shows that S(s) = S(s)1 and that for s iR, S(s) = S(s)1 . Hence for all s
we have
(2.2) S(s) = S(s)1 , S(s) = S(s) ,
so that, in particular, S(s) is self-adjoint for s real. When V is compactly supported
then L2 -eigenvalues of H, E 0, correspond to the poles of S(s), E = s 2 ,
Re s > 0. This follows easily from (2.1) and self-adjointness of H. The poles of S
for Re s < 0 correspond to resonances see [27].
When V is not compactly supported, then under relatively mild assumptions
we still have the scattering matrix for s iR but its meromorphic continuation
6 C. ROBIN GRAHAM AND MACIEJ ZWORSKI

becomes very sensitive to the behaviour of V at infinity. The first physical case in
which these difficulties1 occurred was that of Yukawa potential, V (y) = ey .
We can study the Yukawa potential scattering using simple regular singular point
analysis, not unlike what one encounters in the study of the free hyperbolic space.
We start by making a change of variables:
x = ey , H + s2 = (xx )2 + x + s2 , X = (0, 1] .
Note that the definition of the scattering matrix in the new variables is
(H + s2 )u(x) = 0 , u(1) = 0 , s iR \ {0} ,
(2.3)
u(x) = xs + S(s)xs + O(x) , x 0 .
We obtain u(x) from the two solutions of (H + s2 )G = 0:

X
s
(2.4) G (x, s) = x b j
j (s)x , bj (s) = [j!(2s + j + 1)]
1
.
j=0

These are independent provided 2s / Z. The scattering matrix is obtained by

finding a combination matching the boundary conditions:
G (1, s)b+
0 (s)
S(s) = .
G+ (1, s)b
0 (s)

For l N, G+ (x, l/2) = G (x, l/2) and b+
0 (l/2) 6= 0, b0 (l/2) = 0. Hence S(s) has
a simple pole at s = l/2, l N.
Another independent solution for s = l/2 is obtained by taking
s (G (x, s) G+ (x, s))|s=l/2
l1 !
l
X
= x 2 s b j l l
j (l/2)x 2bl (l/2)x log x + O(x ) ,
j=0

and dividing by s b
0 (l/2) gives us a solution with the prescribed leading part, as
in (2.3):
(2.5) ul/2 (x) = xl/2 + 2pl xl/2 log x + O(xl/2 ) ,
where pl = b
l (l/2)/s b0 (l/2). Comparison of the definition of the scattering ma-
trix with the expansions gives
(2.6) pl = Res s=l/2 S(s) .
Comparison of this elementary discussion with the way in which the invariant
operators Pk are formulated in 4 provides motivation for Theorem 1.
1Thisis potentially rather confusing. The false poles of the scattering matrix discussed
below almost led Heisenberg to abandoning his S-matrix formalism, until a clear explanation
was provided by Jost see [22].
 SCATTERING MATRIX IN CONFORMAL GEOMETRY 7

3. Scattering on conformally compact manifolds

We recall from 1 that the class of asymptotically hyperbolic manifolds, (X, g),
is given by conformally compact manifolds (1.1) for which all sectional curvatures
tend to 1 as the boundary is approached. That is equivalent to demanding that
(3.1) |dx|g = 1 at X ,
which depends only on g see [17]. This class of manifolds comes with a well devel-
oped scattering theory, which originated in the study of infinite volume hyperbolic
quotients \Hn+1 by Patterson, Lax-Phillips, Agmon, Guillope, Perry and others
see [24], and references given there.
The basic facts about the spectrum (g ) of the Laplacian of an asymptotically
hyperbolic metric were established by Mazzeo and Mazzeo-Melrose [17], [19], [18].
It is given by
(g ) = [(n/2)2 , ) pp (g ) , pp (g ) (0, (n/2)2 ) ,
where the pure point spectrum, pp (g ) (the set of L2 eigenvalues), is finite.
More refined statements follow from the main result of [19] which is the existence
of the meromorphic continuation of the resolvent
R(s) = (g s(n s))1 .
This traditional choice of the spectral parameter s corresponds to
n
Re s 6= s(n s) / [(n/2)2 , )
2
Re s > n = s(n s) / [0, ) .
Since g 0, the second implication and the spectral theorem show that for
Re s > n,
R(s) : L2 (X) L2 (X)
is a holomorphic family of operators. In fact R(s) is meromorphic on L 2 (X) for
Re s > n/2, and on a smaller space continues meromorphically to C:
Proposition 3.1. (Mazzeo-Melrose [19, Theorem 7.1, Lemma 6.13]) The resolvent
R(s) : C (X) xs C (X)
is meromorphic in {Re s > n/2}, with poles having finite rank residues exactly for
s such that s(n s) pp (g ). It extends meromorphically to C, with poles having
finite rank residues, and is regular for Re s = n/2, s 6= n/2.
In [19], the structure of the kernel of R(s) is given in Theorem 7.1. The mapping
property stated above is proved in Lemma 6.13 only for hyperbolic space, but
the proof is valid more generally using only the properties of R(s) established in
Theorem 7.1. The regularity for Re s = n/2, s 6= n/2 follows from the spectral
theorem and the absence of embedded eigenvalues see [18].
8 C. ROBIN GRAHAM AND MACIEJ ZWORSKI

The line Re s = n/2 corresponds to physical scattering. As in the one dimen-

sional case, for such s other than n/2, we will consider generalized eigenfunctions
u of the form
(g s(n s))u = 0
(3.2)
u = F xns + Gxs , F, G C (X) .
Note that n = 0 in the example in 2, and as there we think of F | X and G|X
as the incoming and outgoing scattering data respectively. The scattering matrix
is the operator which maps F |X to G|X . For Re s > n/2, it is more appropriate
to think of (3.2) as a boundary value problem. Then F | X represents Dirichlet
data and G|X Neumann data, so that the analytically continued scattering
matrix can be regarded as a generalized Dirichlet to Neumann map. As we shall
see, the expansion for u must in general be modified by the inclusion of a log term
if 2s n N.
There are two consequences of Greens identity which are relevant for us. In the
case Re s = n/2, the following identity is known as the pairing formula:
Proposition 3.2. Let C (X) denote functions vanishing to infinite order at X.
For Re s = n/2 and u1 , u2 satisfying
(g s(n s))ui = ri C (X),
ui = xns fi + xs gi + O(xn/2+1 ) , fi , gi C (X),
we have
Z Z

(3.3) (u1 r2 r1 u2 ) dvg = (2s n) f1 f2 g1 g2 dvh ,
X X
where, since we chose the defining function x of X, we have a natural choice in
the conformal class, h = x2 g|T X .
Proof. This is a standard application of Greens identity see [20] and the proof
of Proposition 3.3 below.
The following analogous identity for s R will be crucial in the proof of Theo-
rem 3.
Proposition 3.3. Suppose s R, s > n/2, 2s n / N. Let u1 , u2 be real-valued

and satisfy (3.2) with Fi , Gi C (X). Then
(3.4) Z Z Z
pf [hdu1 , du2 i s(n s)u1 u2 ]dvg = n F1 G2 dvh = n G1 F2 dvh ,
x> X X
where pf denotes the finite part of the divergent integral.
A special case of (3.4) was discussed by Witten [26] in a physical context, for
X = Hn+1 , using explicit formul for u and G|X in terms of F |X given by
 SCATTERING MATRIX IN CONFORMAL GEOMETRY 9

the Poisson operator and the scattering matrix. Our project originated from an
attempt to understand that discussion in the general setting.
Proof. With g in the form (1.7), Greens identity gives
Z Z
1n
[hdu1 , du2 i s(n s)u1 u2 ]dvg = u1 x u2 dvh .
x> x=

Substituting (3.2) and calculating the finite part gives

Z Z
pf [hdu1 , du2 i s(n s)u1 u2 ]dvg = [sF1 G2 + (n s)F2 G1 ]dvh .
x> X
R R
By symmetry of the left hand side, we deduce that F G dvh =
X 1 2 X
F2 G1 dvh ,
so the result follows.

Remark. If we assume that ui s have expansions (3.2) and that

(g s(n s))u2 = 0 , (g s(n s))u1 C (X) , F2 = 0 ,
then the same argument gives
Z Z
(3.5) u2 (g s(n s))u1 dvg = (n 2s) G2 F1 dvh , s R , s > n/2 ,
X X
which will be useful later.
To have the incoming and outgoing data fi , gi (or Fi |X , Gi |X in (3.2)) invari-
antly defined we introduce density bundles of [21], |N X|s , which keep track of
changes of the defining function to first order, or equivalently of the choice of the
metric in the conformal class2: somewhat informally,
f C (X, |N X|s ) f = a|dx|s , a C (X) .
We will often identify C (X, |N X|s ) with C (X) in the presence of a chosen
defining function.
Using Proposition 3.1 we will construct a family of Poisson operators P(s) for
Re s n/2, s 6= n/2, which is meromorphic for Re s > n/2 with poles only for s
such that s(n s) pp (g ), and continuous up to Re s = n/2, s 6= n/2, with the
properties:

P(s) : C (X, |N X|ns ) C (X) ,
(g s(n s))P(s) 0 ,
(3.6) n
P(s)f = xns f + o(xns ) , if Re s >
,
2
n n
P(s)f = xns f + xs g + O(xn/2+1 ) , if Re s = , s 6= ,
2 2
2These bundles are equivalent to the density bundles of conformal geometry see 4.
10 C. ROBIN GRAHAM AND MACIEJ ZWORSKI

where g C (X) and we chose trivializations of the density bundles as indicated

/ n/2 + N0 /2, then we will have
above. If in addition s
(3.7) P(s)f xns C (X) + xs C (X) .
As in the example in 2, special care will be needed at s n/2 + N/2. To
/ n/2 + N0 /2, we first construct a non-linear
construct P(s) for Re s n/2 and s
formal solution operator:
(s) : C (X) xns C (X) ,
(n + 1 2s)1 (s) is entire in s,
(3.8)
(g s(n s))(s)f C (X) ,
(s)f = xns f + O(xns+1 ) .
This is done by the usual asymptotic construction which we briefly recall. Choose
a metric h in the conformal class on the boundary; this determines a defining
function x and product identification near X so that g is given by (1.7). It is
easy to see that the Laplacian, g , can be written as
g = (xx )2 + nxx + x2 h + xE , E Diff 20 (X) ,
that is E can be locally written as a polynomial of degree 2 in x x and xyi with
coefficients in C (X), where yi are local coordinates on X.
From this we see that for fj C (X),
(3.9) (g s(n s))(xns+j fj ) = j(2s n j)xns+j fj + (xh + E)(xns+j+1 fj ) ,
with E Diff 20 (X). By an iterative procedure we obtain a formal power series
solution (see the proof of Theorem 4.2 for further details), which then gives a
solution modulo C (X) by applying Borels lemma (see for instance [13, Theorem
1.2.6]). It is here that our operator becomes non-linear: the Borel construction
depends on the sequence of functions fj .3 This gives (s) satisfying (3.8):
(s)f =f xns + p1,s f xns+1 +
(3.10)
+ pj,s f xns+j + O(xns+j+1 ) ,
where pj,s is a differential operator on X of order at most 2[j/2], such that
(n + 1 + j 2s)(n + 1 2s)1 pj,s
is polynomial in s for each j.
3The use of Borels Lemma could be avoided by truncating the expansions at a large finite
order. Either direct estimation or a Banach-Baire type of argument implies that in Proposition
3.1, C (X) can be replaced by xM C (X) for some M = M (s), which is locally constant. This
approach has the advantage that linearity and continuity are transparent. However, the use of
Borels Lemma suffices for our purposes since we are primarily interested in the dependence on
s with f fixed.
 SCATTERING MATRIX IN CONFORMAL GEOMETRY 11

We can now easily prove

Proposition 3.4. For Re s n/2, s / n/2 + N0 /2, s(n s)
/ pp (g ), there
exists a unique linear operator P(s) satisfying (3.6), (3.7).

Proof. The existence (of a possibly nonlinear operator) comes from the properties
of (s) and Proposition 3.1:

P(s) = (I R(s)(g s(n s)))(s) .

To see the uniqueness we first observe that if vj , j = 1, 2, are two solutions in

xns C (X)+xs C (X) with the same xns leading term, then the formal expansion
analysis above shows that
v1 v2 xs C (X) .
For Re s > n/2 that would imply that v1 v2 L2 (X), contradicting the assump-
tion that s(n s)
/ pp (g ). For Re s = n/2 we can apply (3.3) with u1 = v1 v2
and u2 = R(n s)f , f C (X), which shows that v1 = v2 . That P(s) is linear
and well-defined on densities follows from uniqueness.

Consider the decomposition given above:

(3.11) P(s)f = (s)f R(s)(g s(n s)))(s)f ,

for s as in Proposition 3.4. We have (s)f xns C (X) and R(s)(g s(n
s)))(s)f xs C (X). It is evident from the construction of (s) that (s)f may
have poles for s n/2 + N/2. We shall see that the contributions from these poles
in the two terms cancel one another, so that P(s)f will extend holomorphically
across s n/2 + N/2 so long as s(n s) / pp (g ).
To analyze P(s) for s near n/2 + N/2 we need to modify the formal solution
operator (s). Observe that (n s) satisfies:

(n s) : C (X) xs C (X) ,
(n s) is holomorphic in Re s > n/2,
(g s(n s))(n s)g C (X) ,
(n s)g = xs g + O(xs+1 ) .

Fix a metric in the conformal class on X and therefore a defining function as

in (1.7). For l N and s near n/2 + l/2, consider the modified formal solution
operator
def
l (s) = (s) (n s)pl,s .
12 C. ROBIN GRAHAM AND MACIEJ ZWORSKI

We certainly have that l (s) is holomorphic for s near n/2 + l/2, s 6= n/2 + l/2,
and
l (s) : C (X) xns C (X) + xs C (X) ,
(g s(n s))l (s)f C (X) ,
l (s)f = xns f + o(xns ) .
We claim that l (s) extends holomorphically across s = n/2 + l/2 as a map into

C (X). Each of (s) and pl,s has at worst a simple pole, so
lim (2s n l)l (s)f
sn/2+l/2

certainly exists. This limit corresponds to a formal solution of ( g s(ns))u = 0

for s = n/2+l/2, which is of the form xn/2+l/2+1 C (X), so by the iterative formula
(3.9), it vanishes identically.
Using
xt 1
lim = log x,
t0 t
evaluation of the limit as s n/2 + l/2 shows that
l (n/2 + l/2)f = xn/2l/2 F + xn/2+l/2 log x G
with F, G C (X), and F |X = f , G|X = 2 Ress=n/2+l/2 pl,s f .
For s near n/2 + l/2 we define
def
(3.12) Pl (s) = (I R(s)(g s(n s)))l (s) .
It is clear that if (n/2)2 (l/2)2
/ pp (g ), then Pl (s) is holomorphic across
s = n/2 + l/2. From the properties of l (s) and Proposition 3.1, it follows that
we have
Pl (n/2 + l/2) : C (X) xn/2l/2 C (X) + xn/2+l/2 log x C (X)
(g s(n s))Pl (n/2 + l/2) = 0 .
For s 6= n/2 + l/2 we have

Pl (s) : C (X) xs C (X) + xns C (X) ker (g s(n s)) ,
so from the uniqueness part of Proposition 3.4 we see that
Pl (s) = P(s) , for s near n/2 + l/2, s 6= n/2 + l/2.
We summarize the discussion of Poisson operators in the following
Proposition 3.5. There is a unique family of Poisson operators

P(s) : C (X, |N X|ns ) C (X)
 SCATTERING MATRIX IN CONFORMAL GEOMETRY 13

for Re s n/2, s 6= n/2, which is meromorphic in {Re s > n/2} with poles only
for s such that s(n s) pp (g ), and continuous up to {Re s = n/2} \ {n/2},
such that
(g s(n s))P(s) = 0,
with expansions
P(s)f = xns F + xs G / n/2 + N0 /2
if s
P(s)f = x n/2l/2
F + Gx n/2+l/2
log x if s = n/2 + l/2, l N ,
for F, G C (X) such that F |X = f .
If s = n/2 + l/2, then G|X = 2pl f , where
pl = Ress=n/2+l/2 pl,s ,
with the differential operator pl,s defined in (3.10). For l = 2k,
2k (p2k ) = ck 2k (kh ) ,
with ck as in (1.4).
The principal symbol of p2k is easily calculated from the inductive construction
of the operators pj,s ; this will be described in more detail in the proof of Proposi-
tion 4.2.
Remark. A more precise notion of a meromorphic family of asymptotic expan-
sions can be given using a Mellin transform in x. This gives a holomorphic family
of meromorphic functions with poles corresponding to the exponents in the ex-
pansions and logarithmic terms to double poles. We do not need this precise
description here as we will work with the explicit formula for P(s).
Next we define the scattering matrix
S(s) : C (X, |N X|ns ) C (X, |N X|s )
for Re s n/2, 2s n / N0 , s(n s)
/ pp (g ). According to Proposition 3.5,

for such s and for f C (X, |N X| ) we have P(s)f = xns F + xs G, so we
ns

define
S(s)f = G|X .
The construction of P(s) in Proposition 3.4 gives an explicit realization of the
scattering matrix:

(3.13) S(s)f = xs R(s)(g s(n s))(s)f |X , 2s n / N0 .
Using (3.8) and Proposition 3.1, it follows that S(s) is defined for s as indicated
above and is meromorphic in {Re s > n/2}.
Applying complex conjugation to (3.2) shows that for Re s = n/2, s 6= n/2, we
have the functional equation
S(s) = S(n s)1 .
14 C. ROBIN GRAHAM AND MACIEJ ZWORSKI

An application of the pairing formula (3.3) with uj = P(s)fj shows the unitarity
relation:
Z
n
S(s)f1 S(s)f2 f1 f2 = 0 dvh , Re s = , s 6= n/2 .
X 2
By using the Schwartz reflection principle and then unitarity to establish regu-
larity at s = n/2, we find a meromorphic extension of S(s) to the entire complex
plane, regular for s iR, and we have the following symmetries:
S(n s) = S(s)1 , S(n s) = S(s)1 .
For s R, the functional equation and unitarity show that
S(s) = S(s) ,
which also follows directly from Proposition 3.3 since S(s) is a real operator for s
real.
The construction of the Poisson kernels and Proposition 3.5 give the following:
Proposition 3.6. The scattering matrix is meromorphic in Re s > n/2 and at a
pole s0 we have

s0 / n/2 + N/2
s0
Res s=s0 S(s) =
n/2+l/2 pl s0 = n/2 + l/2 ,
where pl is as in Proposition 3.5 and
Z !
1
s0 = (2s0 n) xs0 R(s)ds xs0
2i XX
s0

is non-trivial only if s0 (n s0 ) pp (g ). Here s0 is a small circle in C about

s0 , traversed counterclockwise.
Proof. Suppose first that the resolvent is holomorphic near n/2 + l/2. Then (3.11),
(3.13), and the holomorphy of P(s) near n/2 + l/2 show that

Res s=n/2+l/2 S(s) = Res s=n/2+l/2 xs (s) |X
= Res s=n/2+l/2 pl,s = pl ,
where pl,s are as in (3.10).
When s0 is a pole of the resolvent then
PK PK
k=1 k k e 1 k=1 k k
(3.14) R(s) = + T (s) = + T (s) ,
s0 (n s0 ) s(n s) 2s0 n s s0
where T and Te are holomorphic near s0 , and k xs0 C (X) L2 (X) are the
normalized eigenfunctions, (g s0 (n s0 ))k = 0. Let us first consider the case
 SCATTERING MATRIX IN CONFORMAL GEOMETRY 15

/ n/2 + N/2. We use (3.5) to evaluate

s0
Z Z
k (g s0 (n s0 ))(s0 )f dvg = (n 2s0 ) (xs0 k )|X f dvh .
X X

Using this and (3.14) in (3.13) shows that the residue of S(s) at s 0 is the operator
P
with kernel K k=1 (x
s0
k )|X (xs0 k )|X , which by (3.14) is s0 .
If s0 is a pole of the resolvent and s0 = n/2 + l/2 then we replace (3.11) by
(3.12) and find the residue of P(s) by the same method as above:
K
X Z
Res s=n/2+l/2 P(s)f = k (xs0 k )|X f dvh .
k=1 X

Now using this instead of holomorphy of P(s) in the argument of the first paragraph
of this proof gives the desired formula for Res s=n/2+l/2 S(s).

This shows that the scattering matrix always has poles of infinite rank at n/2+k,
k N, and possibly at n/2+k1/2, k N, with the residues given by the operators
appearing as the log term coefficients in the expansions of the Poisson operators.
Uniqueness of P(s) shows that P(n)1 = 1. It then follows from Proposition 3.5
that pn 1 = 0 so that pn,s 1 extends holomorphically across s = n. Since we always
have 0
/ pp (g ), Proposition 3.6 shows that S(s)1 extends holomorphically across
s = n. Set S(n)1 = limsn S(s)1. Then we have:
Proposition 3.7. If pn,s are defined in (3.10), then
S(n)1 = lim pn,s 1 .
sn

Proof. We have P(s)1 = (s)1 + xs G(s), where the expansion of (s)1 is as in

(3.10) and G(s) C (X) satisfies G(s)|X = S(s)1. Letting s n and recalling
P(n)1 = 1 shows that limsn ((s)1 + xs G(s)) = 1. It follows that limsn pl,s 1 = 0
for 1 l < n, and limsn (pn,s 1 + S(s)1) = 0 as desired.

We next discuss the distributional kernels of the operators R(s), P(s), and S(s).
The structure of the kernel of R(s) is best understood on a resolved product space.
Since it is not central to our discussion, we will only sketch the construction. We
first recall the blow-down map of [19, Sect.3]:
: X 0 X X X ,
where X 0 X is the blow-up of X X along the boundary diagonal, illustrated
in Fig.1. The restriction to the boundary gives us a blow-up of the diagonal in
X X with the corresponding blow-down map:
: T B X X ,
16 C. ROBIN GRAHAM AND MACIEJ ZWORSKI

where T and B are the top and bottom faces in the blow-up, as shown in Fig.1.
Let , 0 be the defining functions of the top and bottom faces respectively. Then
we have
Proposition 3.8. (Mazzeo-Melrose [19, Proposition 6.2, Theorem 7.1]) The kernel
of the resolvent on the resolved space X 0 X, R(s), and away from its poles,
can be written as
R(s) = R0 (s) + R00 (s) ,
where R0 (s) is supported away from T B, and

R00 (s) = ()s (0 )s K(s) , K(s) C (X 0 X) .

Ox0

x0

T = { = 0} Ox
F
O x
0
Y
B = { = 0}

OY
Figure 1. The boundary faces of the blown-up space X 0 X. The
variable Y stands for the defining function of the diagonal in X
X, Y = y y 0 , in local coordinates.

The pairing formula (3.3) gives an expression for the kernel of P(s) in terms of
that of the resolvent, much in the same spirit as in the classical derivation of the
Poisson kernel from the Green function:
Proposition 3.9. For f C (X) and Re s n/2, s 6= n/2, we have a formal
expression

(3.15) P(s)f = (2s n)(R(s)xs )(f 0 (x)) ,

where the right hand side is understood as the restriction to the boundary of the
distributional kernel of R(s)xs paired with f .

Proof. First let Re s = n/2. For any g C (X) and f C (X) apply (3.3) with
u1 = P(s)f and u2 = R(n s)g. We also note that for Re s = n/2 we have

R(n s) = R(s) .
 SCATTERING MATRIX IN CONFORMAL GEOMETRY 17

That gives
Z Z
P(s)f g = (2s n) (xn+s R(n s)g)|X f
X
ZX

= (2s n) (R(s)xs )(f 0 (x)) g ,
X
proving the claim. That the restriction of R(s)xs is well defined as a distribution
on the boundary follows from Proposition 3.8. The result extends to Re s > n/2
by analytic continuation.
We now continue P(s) meromorphically to C so that (3.15) holds for Re s < n/2
as well.
Using the structure of the resolvent described in Proposition 3.8 we can give
a precise description of the kernel of S(s) as a pseudodifferential operator: away
from its poles, we have
S(s) 2 Re sn (X; |N X|ns , |N X|s ) ,
(3.16) (n/2 s) sn/2
(S(s)) = 2n2s (h ).
(s n/2)
To see this we follow [11] and obtain the following formula for the distributional
kernel of the scattering matrix in terms of the resolvent on the resolved product
space:

s 0 s
(3.17) S(s) = (2s n)( ) x x R(s) T B ,
(see also [14, Proposition 4.4]). This is derived using (3.11), (3.13) and Proposition
3.9, which give
S(s)f = (xs P(s)f xs (s)f )|X

= (2s n)xs (R(s)xs )(f 0 (x)) xs (s)f |X .
For Re s < n/2 and s away from poles of R(s) we see that

S(s)f = (2s n)xs (R(s)xs )(f 0 (x)) |X .
Proposition 3.8 shows that for Re s 0, and s not a pole, we have
s
x ( R0 (s)xs )(f 0 (x)) |X 0 ,
and

xs ( R00 (s)xs )(f 0 (x)) |X = xs R(s)xs )(f 0 (x)) |X ,
and this gives (3.17) for Re s 0. Strictly speaking we need more information
about the R0 (s) term than stated in Proposition 3.8 but those properties are stated
in [19, Theorem 7.1]. We shall see below that (3.17) continues meromorphically in
s.
18 C. ROBIN GRAHAM AND MACIEJ ZWORSKI

The fact that the scattering matrix is a pseudodifferential operator of order

2 Re s n follows from (3.17). As shown in Fig.1, near T B we can use y, |Y |,
= Y /|Y |, and , 0 , the defining functions of T and B, as coordinates, where
Y = y y 0 . In terms of these functions,
x = |Y | , x0 = |Y |0 ,
and hence, using Proposition 3.8,
(3.18)
S(s) = (2s n)( ) |Y |2s K(s, y, |Y |, , ) , K(s, ) C ([X X; X ]) ,
where [X X; X ] is X X blown up along the diagonal, that is the space
where instead of y, y 0 X as coordinates, we use y, |Y |, .
Hence for Re s 0 the distributional kernel of S(s) is a distribution conormal
to the diagonal and hence S(s) is a pseudo-differential operator. We then continue
|Y |2s meromorphically in s following [13, Theorem 3.2.4], and that gives the
first part of (3.16). The parametrix construction in [19] gives K, and leads to a
computation of the principal symbol see [14].

4. Conformal Geometry
In this section we show how the conformally invariant powers of the Lapla-
cian may be derived from the formal Poincare metric associated to the conformal
structure, review the definition and basic properties of Bransons Q-curvature, and
remark on the interpretation of the scattering matrix for a Poincare metric as a
family of conformally invariant pseudo-differential operators.
The invariant powers of the Laplacian act on conformal densities. The metric
bundle of a conformal manifold (M, [h]) is the ray subbundle G S 2 T M of
multiples of the metric: if h is a representative metric, then the fiber of G over
p M is {t2 h(p) : t > 0}. The space of conformal densities of weight w C is
w
E(w) = C (M ; G 2 ) ,
where by abuse of notation we have denoted by G also the line bundle associated to
the ray bundle defined above. A choice of representative h for the conformal struc-
ture induces an identification E(w) ' C (M ); if h = e2 h then the corresponding
elements of C (M ) transform by f = ew f . If X = M as in the previous sec-
tion, a conformally compact metric on X with conformal infinity [h] determines
an isomorphism
E(w) ' C (M, |N X|w ) .
The invariance property (1.3) can be reformulated as the statement that P k is an
invariantly defined operator
Pk : E(n/2 + k) E(n/2 k).
The main result of [8] is the following existence theorem.
 SCATTERING MATRIX IN CONFORMAL GEOMETRY 19

Proposition 4.1. Let k N and k n/2 if n is even. There is a conformally

invariant natural differential operator Pk : E(n/2 + k) E(n/2 k) with
principal part equal to that of k .
Such operators are not unique. The construction in [8] is in terms of the ambient
metric of [5]. We shall first give here a different construction of invariant operators
based on the Poincare metric and then show that the two constructions give the
same operators.
We begin by recalling from [5] the formal Poincare metric associated to a con-
formal structure (M, [h]). Given a representative metric h, one considers metrics
on M [0, 1] of the form (1.7). The Einstein equation Ric(g) + ng = 0 can be
calculated directly in terms of hx and the formal asymptotics of solutions studied;
see [7]. If n is odd, then there is a unique formal smooth solution h x to
Ric(g) + ng = O(x )
which is even in x. If n is even, the condition Ric(g) + ng = O(x n2 ) uniquely
determines hx mod O(xn ), which is even in x (mod O(xn )). Although in general
smooth solutions do not exist to higher orders, the condition
trg (Ric(g) + ng) = O(xn+2 )
can be satisfied and uniquely determines the h-trace of the x n coefficient in hx ;
this is the vanishing trace condition referred to in the introduction. The indicated
Taylor coefficients of hx are determined inductively from the equation and are
given by polynomial formulae in terms of h, its inverse, and its curvature tensor
and covariant derivatives thereof.
Since any asymptotically hyperbolic metric can be put uniquely in the form
(1.7) upon choosing h, it follows that the equivalence class of the solution g up
to diffeomorphism and up to terms vanishing to the indicated orders is uniquely
determined by the conformal structure. This equivalence class is called the formal
Poincare metric associated to [h]. When n is even, the higher order terms in h x
are not determined; however for simplicity in statements below, we shall restrict
consideration to hx which are smooth and even in x to all orders. If X is a manifold
with X = M , any metric on X whose restriction to a collar neighborhood of M
is in this equivalence class is called a Poincare metric associated to [h].
Let g be a Poincare metric and h a representative for the conformal infinity. If x
is a defining function such that x2 g|T M = h, and f C (M ) represents a section
of E(w), then it is a conformally invariant statement to require that a function u
on X be asymptotic to xw f . The invariant operators Pk arise from solving
(4.1) (g s(n s))u = O(x )
for u with such asymptotic behaviour. The characteristic exponents of g s(ns)
are s, n s, so generically solutions behave like xs and xns . As in the previous
20 C. ROBIN GRAHAM AND MACIEJ ZWORSKI

section, we take u asymptotic to xns f , which according to the above remarks

means that f is to be interpreted as a density of weight w = s n.
We saw in 3 that if g is asymptotically hyperbolic and s n/2 / N/2, then
for any f C (M ) there is a formal solution u mod O(x ) to (4.1) of the form
u = xns F with F C (X) and F |M = f . As we shall see below, for Poincare
metrics this holds if only s n/2 / N. In 3, solutions for s n/2 N/2 were
constructed as a limit of solutions for nearby s. Here we make a direct analysis at
the exceptional values of s and obtain the invariant operators as obstructions to
the existence of formal smooth solutions.
Proposition 4.2. Let (X, g) be a Poincare metric associated to (M, [h]) and let
f C (M ). If k N, there is a formal solution of (4.1) for s = n/2 + k of the
form
(4.2) u = xn/2k (F + Gx2k log x)
with F, G C (X) and with F |M = f . F is uniquely determined mod O(x2k )
and G is uniquely determined mod O(x ). Moreover,
(4.3) G|M = 2ck Pk f,
where Pk is a differential operator on M with principal part k .
If n is odd and k N or if n is even and k n/2, then Pk depends only on h
and defines a conformally invariant operator : E(n/2 + k) E(n/2 k).
Proof. Choose a representative metric h and write g in the form (1.7). A straight-
forward calculation shows that [g s(n s)] xns = xns+1 Ds , where
x n s ij 0
(4.4) Ds = xx2 + (2s n 1 hij h0ij )x h hij + xhx .
2 2
Here hij denotes the metric hx with x fixed, and h0ij = x hij . Therefore for fj
C (M ), one has
(4.5) Ds (fj xj ) = j(2s n j)fj xj1 + O(xj ).
If 2s n / N, then we can use (4.5) to construct a smooth solution of (4.1)
inductively. Beginning with f0 = F0 = f , define fj , Fj for j 1 by
j(2s n j)fj = (x1j Ds (Fj1 ))|x=0
(4.6)
Fj = Fj1 + fj xj .
j
P wej have Ds Fj = O(x ) so that the definition of fj makes sense. Then
By (4.5)
F = fj x is a formal solution of (4.1). Observe that since hx is even in x, Ds
maps even functions to odd and vice versa. Therefore fj = 0 for j odd. For j = 2k
even, an easy induction shows that f2k takes the form
(s n/2 k)
(4.7) f2k = ck,s Pk,s f, ck,s = (1)k ,
22k k! (s n/2)
 SCATTERING MATRIX IN CONFORMAL GEOMETRY 21

where Pk,s is a differential operator on M , the principal part of which agrees with
that of kh . Since the Taylor expansion of hx is determined (to the appropriate
order for n even) in terms of h, one sees by counting derivatives that if k N and
k n/2 if n is even, then Pk,s also depends only on h and is a natural differential
operator with coefficients which are polynomial in s.
If 2s n = l N, then the corresponding coefficient 2s n j vanishes for j = l
in the first equation of (4.6). For l odd, by parity considerations it follows that the
right hand side of this equation also vanishes, so fl can be chosen arbitrarily (for
example fl = 0 to preserve parity) and the induction continued to infinite order.
However, if 2s n = 2k is even, then the right hand side of the first equation
of (4.6) need not vanish for j = 2k and there is an obstruction to solving with
F smooth. This is of course reflected in the pole of ck,s at s = n/2 + k. This
obstruction can be incorporated into a log term and the formal solution continued
to higher order as follows. Observe that
Ds (gj xj log x) =j(2s n j)gj xj1 log x + (2s n 2j)gj xj1
(4.8)
+ O(xj log x) .
Therefore if we take
(4.9) g2k = (2k)1 (x12k Ds (F2k1 ))|x=0 ,
choose f2k arbitrarily, and set
F2k = F2k1 + g2k x2k log x + f2k x2k ,
then we have Ds F2k = O(x2k log x). Using (4.5) and (4.8), it is easily seen that the
construction can be continued to all higher orders to obtain a solution of the form
(4.2) as claimed. We have G|x=0 = g2k , which is given by a differential operator
on M of the claimed form by the same reasoning as above. In fact, one sees easily
that (4.3) holds with Pk = Pk,n/2+k and
ck = Ress=n/2+k ck,s = (1)k [22k k!(k 1)!]1 .
Suppose now we change the conformal representative h. We obtain a differ-
ent defining function x, a different product identification for X, and a different
representation (1.7). However, g and u remain unchanged, so by uniqueness we
deduce that G|x=0 must transform as a density, proving the conformal invariance
of Pk .
Remark. Equation (4.4) holds also for general asymptotically hyperbolic metrics
in the form (1.7) and can be used to explicitly compute the operators p j,s of 3.
Of course in general it no longer need be the case that hx is even in x. Therefore
log terms may also occur for 2s n odd, and the scattering matrix may have poles
for such s according to Proposition 3.6.
22 C. ROBIN GRAHAM AND MACIEJ ZWORSKI

Note that if n is even, then Pn/2 is invariantly defined from E(0) to E(n),
that is, from C (M ) to the space of volume densities. Since the constant function
1 E(0) has a smooth extension annihilated by g , we have Pn/2 1 = 0. Therefore
Pn/2 has zero constant term.
We next recall from [8] the original construction of the invariant operators via
the ambient metric. Denote by : G M the natural projection of the metric
bundle, and by g the tautological symmetric 2-tensor on G defined for (p, h) G
and X, Y T(p,h) G by g(X, Y ) = h( X, Y ). There are dilations s : G G
for s > 0 given by s (p, h) = (p, s2 h), and we have s g = s2 g. Denote by T
d
the infinitesimal dilation vector field T = ds s |s=1 . Define the ambient space
G = G (1, 1). Identify G with its image under the inclusion : G G given
by (h) = (h, 0) for h G. The dilations s and infinitesimal generator T extend
naturally to G.
The ambient metric g is a Lorentzian metric on G which satisfies the initial
condition g = g, is homogeneous in the sense that s g = s2 g, and is an asymptotic
solution of Ric(g) = 0 along G. For n odd, these conditions uniquely determine
a formal power series expansion for g up to diffeomorphism, but for n even and
n > 2, a formal power series solution exists in general only to order n/2.
An element of E(w) can be regarded as a homogeneous function of degree w on
G. One of the ways that Pk is derived in [8] is as the obstruction to extending
f E(n/2 + k), regarded as such a homogeneous function, to a smooth function
F on G, such that F is also homogeneous of degree n/2 + k and satisfies

(4.10) F = O( ),

where denotes the Laplacian in the metric h. Similarly to Proposition 4.2, the
Taylor expansion of F is formally determined to order k 1 in , but there is an
obstruction at order k which defines the operator Pk .
Proposition 4.3. Let k N with k n/2 if n is even. The conformally invariant
operator Pk defined in [8] in terms of the ambient metric agrees with the operator
of Proposition 4.2.

Proof. As described in [5], the formal Poincare metric associated to a conformal

structure can be constructed from the ambient metric and vice versa; the two
constructions are equivalent. We review this equivalence.

In the ambient space G, the equation g(T, T ) = 1 defines a hypersurface X
which lies on one side of G and which intersects exactly once each dilation orbit

on this side of G. The Poincare metric g is the pullback to X of g. The equation
Ric(g) = 0 is equivalent to Ric(g) = ng. To see this, one uses the normal form
from [5] for ambient metrics: in suitable coordinates on G, the ambient metric
 SCATTERING MATRIX IN CONFORMAL GEOMETRY 23

takes the form

(4.11) g = 2tdtd + 2dt2 + t2 h .
Here is a defining function for G G, t is homogeneous of degree 1 with respect
to the dilations on G, and h is a smooth 1-parameter family of metrics on M . In

these coordinates we have T = tt , so X= {2t2 = 1}. Introduce a new variable

x = 2 and set s = xt so that X= {s = 1}. A straightforward calculation
shows that (4.11) becomes
(4.12) g = s2 g ds2 ,
where g is given by (1.7) with hx = h . The equivalence of Ric(g) = 0 and
Ric(g) = ng is a straightforward calculation given the relationship (4.12) (see
Proposition 5.1 of [9]). Note that hx is automatically even in x.
The equations (4.1) and (4.10) are equivalent. To see this, rewrite (4.12) as
g = s2 (g ds2 /s2 ) ,
and transform under a conformal change to obtain
= s2 [g + (ss )2 + nss ] .

If F is homogeneous of degree w, we therefore have
F = s2 [g + w(w + n)]F ,

so (4.10) is equivalent to (4.1) for s = n + w and u = F | . One may recover F
X
from u via homogeneity by F = sw u = tw xw u. In order for F to be smooth up
to = 0, we require therefore that xw u be smooth up to x = 0 (and be even in
x). Thus the two extension problems are equivalent, so the normalized obstruction
operators must agree.
We next define the Q-curvature as in [2]. For this discussion we shall denote
by Pkn the operator Pk in dimension n. Fix k N. One consequence of the
construction of Proposition 4.2 is that the operator Pkn is natural in the strong
sense that Pkn f may be written as a linear combination of complete contractions
of products of covariant derivatives of the curvature tensor of a representative for
the conformal structure with covariant derivatives of f , with coefficients which are
rational in the dimension n. Also, it follows from the fact that the zeroth order
term of Ds in (4.4) has a factor of n s, that the zeroth order term of Pkn may
be written as (n/2 k)Qnk for a scalar Riemannian invariant Qnk with coefficients
which are rational in n and regular at n = 2k. (This is of course consistent with
n
the fact mentioned above that Pn/2 1 = 0.) The Q-curvature in even dimension n
n
is then defined as Q = Qn/2 .
24 C. ROBIN GRAHAM AND MACIEJ ZWORSKI

We may also consider the Q-curvature as arising in a similar way from the zeroth
order terms of the operators Pk,s with n fixed but as s varies. For the same reason
as above, the zeroth order term of Pk,s is of the form
(4.13) Pk,s 1 = (n s)Qk,s
for a scalar Riemannian invariant Qk,s which is polynomial in s. Taking s = n/2+k
and recalling that Pk = Pk,n/2+k shows that Qnk = Qk,n/2+k . In particular,
(4.14) Q = Qn/2,n .
If g is a Poincare metric with conformal infinity [h], then according to (3.16),
the scattering matrix S(s) is a family of pseudodifferential operators on M , which
is conformally invariant in the sense that it acts invariantly on conformal densities.
Even though S(s) depends on the choice of Poincare metric g, it is shown in [14]
that its full symbol depends only on the infinite jet of g at X, which, as discussed
above, is determined by [h] (to the appropriate order for n even). One can therefore
view the symbol as determined by the conformal structure, and the choice of g as
a geometrically natural means of fixing smoothing terms to obtain a globally well-
defined operator having this symbol. Peterson [25] has defined an analogous family
of symbols directly by analytic continuation from the differential operators P k .
Branson [2] derived the transformation law (1.5) for Q by analytic continuation
in the dimension. It also follows easily from Theorems 1 and 2 by analytic continua-
d = es S(s)e(ns) ,
tion in s. The conformal invariance of S(s) is equivalent to S(s)

where here S(s) is realized as an operator on C (M ) corresponding to the choice
of conformal representative. Therefore,
d = S(s)1 + S(s)(e(ns) 1).
es S(s)1
Letting s n and applying Theorems 1 and 2 immediately yields (1.5).

5. Proofs of the main results

Theorems 1 and 2 follow from the corresponding results of 3 in the special case
when g is a Poincare metric associated to the conformal structure. It follows from
(4.7) that for Poincare metrics, the differential operators p l,s of (3.10) are given by
(5.1) pl,s = 0 for l odd, p2k,s = ck,s Pk,s , k N.
Thus for the residues we obtain
(5.2) pl = 0 for l odd, p2k = ck Pk , k N.

Proof of Theorem 1. This is immediate from Proposition 3.6 and (5.2).

Observe that Proposition 3.6 and (5.2) also show that for Poincare metrics, S(s)
has no pole at s = n/2 + l/2 if l N is odd.
 SCATTERING MATRIX IN CONFORMAL GEOMETRY 25

Proof of Theorem 2. We apply Proposition 3.7. By (5.1) and (4.13), (4.14), we

obtain
S(n)1 = lim cn/2,s Pn/2,s 1 = Ress=n cn/2,s Q = cn/2 Q.
sn

We note that Proposition 3.7 and (5.1) also show that for Poincare metrics,
S(n)1 = 0 if n is odd.
Proof of Theorem 3. Let n be even and let g be a Poincare metric. We first recall
(as described in [7]) the expression for L in terms of the expansion of the volume
form of g. If h is a representative for the conformal infinity of g, we may write g
in the form (1.7), and from the fact that g is a Poincare metric it follows that
(5.3) dvg = xn1 (1 + v (2) x2 + (even powers) + v (n) xn + . . .)dvh dx,
where each v (2j) for 1 j n/2 is a smooth function on M determined by h.
Integration yields (1.8), and shows that
Z
(5.4) L= v (n) dvh .
M
Now take s near but not equal to n and apply Proposition 3.3 with u 1 = u2 =
P(s)1, which we now denote by us . Here we have used h to trivialize the density
bundle |N X|ns . We obtain
Z Z
2 2
(5.5) pf [|dus | s(n s)us ]dvg = n S(s)1dvh .
x> M
We considerR the limiting behaviour
R in this equation as s n. According to
Theorem 2, M S(s)1dvh cn/2 M Qdvh . We shall show that the left hand side in
(5.5) converges to nL/2 as s n, thereby proving Theorem 3.
Since us = P(s)1 1 as s n, the integrand in the left hand side of (5.5)

converges to 0 pointwise on X. It follows that
Z
[|dus |2 s(n s)u2s ]dvg 0
x>x0

for each fixed x0 > 0. So it suffices to consider

Z
(5.6) [|dus |2 s(n s)u2s ]dvg ,
<x<x0

for which we may use the product identification (1.7) coming from the representa-
tive metric h, provided we choose x0 small enough.
By (3.11), (3.10), (5.1), and (4.13), for s near n we have
n/2
X
ns
(5.7) us = x (1 + ck,s (n s)Qk,s x2k ) + xs S(s)1 + O(xn+3/4 ),
k=1
26 C. ROBIN GRAHAM AND MACIEJ ZWORSKI

where the power n + 3/4 can be replaced by n + a, a < 1, if we take s close to

n. In considering (5.7), recall that ck,s is regular at s = n for 1 k < n/2, but
that cn/2,s has a simple pole at s = n with residue cn/2 . Also, by Theorem 2,
S(n)1 = cn/2 Q = cn/2 Qn/2,n . Since un = 1, the error term vanishes identically
when s = n and in general is seen to be of the form O(|n s|xn+3/4 ). There is a
similar bound when we differentiate the expansion. Let y i denote local coordinates
on M and let denote any first coordinate derivative x or yi ; then we have

n/2
X
(5.8) xus = x xns (1 + ck,s (n s)Qk,s x2k ) + xs S(s)1 +O(|ns|xn+3/4 ).
k=1

We begin by considering the u2s term in (5.6). Upon squaring (5.7), one obtains
n/2
X
u2s =x 2(ns)
(1 + Ak,s x2k ) + 2xn S(s)1 + O(xn+1/2 ),
k=1

where the coefficients Ak,s are smooth functions on M , holomorphic in s near s = n,

and satisfying Ak,n = 0 for 1 k < n/2 and An/2,n = 2cn/2 Q. Multiplying by
dvg and using (5.3) gives

n/2
X
u2s dvg = xn2s1 (1 + Bk,s x2k ) + 2x1 S(s)1 + O(x1/2 ) dvh dx,
k=1

with coefficients Bk,s again holomorphic R in s, 2and with Bn/2,n = 2cn/2 Q +1/2 v (n) .
In order to evaluate limsn (n s) pf <x<x0 us dvg , observe first that the O(x )
error term is integrable, so its contribution vanishes upon letting s n. Now
Z Z
n2s+2k1 n2s+2k 1
pf x Bk,s dvh dx = x0 (n 2s + 2k) Bk,s dvh .
<x<x0 M

If k < n/2, then this approaches a finite limit as s n, so these terms also do
not contribute. The same is true for the x1 S(s)1 term. Evaluating the limit for
k = n/2 and recalling (5.4) then gives
Z Z Z
2 1
(5.9) lim (n s) pf us dvg = Bn/2,n dvh = cn/2 Q + L/2.
sn <x<x0 2 M M

For the derivative term, we have |dus |2 = (xx us )2 + hij

x (xy i us )(xy j us ). Con-
2
sider first (xx us ) . Expanding and squaring (5.8) gives
n/2
X
(5.10) 2
(xx us ) = x 2(ns)
A0k,s x2k + 2s(n s)xn S(s)1 + O(|n s|xn+1/2 ),
k=0
 SCATTERING MATRIX IN CONFORMAL GEOMETRY 27

where A0k,s are smooth functions on M , holomorphic in s, satisfying

A0k,s = O(|n s|2 ), 0 k < n/2,
(5.11)
lim (n s)1 A0n/2,s = 2ncn/2 Q.
sn

Multiplying (5.10) by dvg gives

(xx us )2 dvg =

n/2
X
xn2s1 0
Bk,s x2k + 2s(n s)x1 S(s)1 + O(|n s|x1/2 ) dvh dx,
k=0

0
where the Bk,s satisfy the same properties (5.11) as the A0k,s . Integrating and
evaluating the finite part and the limit as above yield
Z Z
2
(5.12) lim pf (xx us ) dvg . = ncn/2 Q.
sn <x<x0 M

Differentiation with respect to y does not decrease the order of vanishing in x, and
because of this one finds by a similar calculation that
Z
(5.13) lim pf hij
x (xy i us )(xy j us )dvg = 0.
sn <x<x0

Combining (5.9), (5.12), and (5.13) gives

Z
lim pf [|dus |2 s(n s)u2s ]dvg = nL/2
sn <x<x0

as desired.

References
[1] M. Anderson, L2 curvature and volume renormalization for AHE metrics on 4-manifolds,
Math. Res. Lett. 8 (2001), to appear, math.DG/0011051.
[2] T. Branson, Sharp inequalities, the functional determinant, and the complementary series,
Trans. AMS 347 (1995), 3671-3742.
[3] T. Branson, P. Gilkey, and J. Pohjanpelto, Invariants of locally conformally flat manifolds,
Trans. AMS 347 (1995), 939953.
[4] S.-Y.A. Chang, J. Qing, and P.C. Yang, Compactification of a class of conformally flat
4-manifolds, Invent. Math. 147 (2000), 65-93.
[5] C. Fefferman and C.R. Graham, Conformal invariants, in The mathematical heritage of
Elie Cartan (Lyon, 1984). Asterisque, 1985, Numero Hors Serie, 95116.
[6] C. Fefferman and C.R. Graham, in preparation.
[7] C.R. Graham, Volume and area renormalizations for conformally compact Einstein metrics,
Rend. Circ. Mat. Palermo, Ser.II, Suppl. 63 (2000), 31-42.
[8] C.R. Graham, R. Jenne, L.J. Mason, and G.A.J. Sparling Conformally invariant powers of
the Laplacian. I. Existence. J. London Math. Soc. (2) 46 (1992), 557565.
28 C. ROBIN GRAHAM AND MACIEJ ZWORSKI

[9] C.R. Graham and J. Lee, Einstein metrics with prescribed conformal infinity on the ball,
Adv. Math. 87 (1991), 186-225.
[10] C.R. Graham and E. Witten, Conformal anomaly of submanifold observables in AdS/CFT
correspondence, Nucl. Phys. B 546 (1999), 52-64, hep-th/9901021.
[11] L. Guillope and M. Zworski, Scattering asymptotics for Riemann surfaces, Ann. Math.
145(1997), 597-660.
[12] M. Henningson and K. Skenderis, The holographic Weyl anomaly, J. High Ener. Phys. 07
(1998), 023, hep-th/9806087; Holography and the Weyl anomaly, hep-th/9812032.
[13] L. Hormander, The Analysis of Linear Partial Differential Operators, Vol. I, II, Springer-
Verlag, Berlin, 1983.
[14] M. Joshi and A. Sa Barreto, Inverse scattering on asymptotically hyperbolic manifolds, Acta
Math. 184(2000), 4186.
[15] M. Joshi and A. Sa Barreto, The wave group on asymptotically hyperbolic manifolds, to
appear in J. Funct. Anal.
[16] J. Maldacena, The large-N limit of superconformal field theories and supergravity, Adv.
Theor. Math. Phys. 2(1998), 231252, hep-th/9711200.
[17] R. Mazzeo, The Hodge cohomology of a conformally compact metric, J. Differential Geom.
28(1988), 309339.
[18] R. Mazzeo, Unique continuation at infinity and embedded eigenvalues for asymptotically
hyperbolic manifolds, Am. J. Math. 113 (1991), 2545.
[19] R. Mazzeo and R. Melrose, Meromorphic extension of the resolvent on complete spaces with
asymptotically constant negative curvature, J. Funct. Anal. 75 (1987), 260-310.
[20] R.B. Melrose, Geometric Scattering Theory. Cambridge University Press, 1995.
[21] R. Melrose and M. Zworski, Scattering metrics and geodesic flow at infinity. Invent. Math.
124(1996), 389-436.
[22] R. Newton, Scattering theory of waves and particles, McGraw-Hill Book Co., New York-
Toronto-London 1966.
[23] S. J. Patterson and P. A. Perry, The divisor of Selbergs zeta function for Kleinian groups,
Duke Math. J. 106 (2001), 321-390. Appendix A by C. Epstein, An asymptotic volume
formula for convex cocompact hyperbolic manifolds.
[24] P. Perry, The Laplace operator on a hyperbolic manifold. II. Eisenstein series and the
scattering matrix, J. reine. angew. Math. 398 (1989), 67-91.
[25] L. J. Peterson, Conformally covariant pseudo-differential operators, Diff. Geom. Appl. 13
(2000), 197-211.
[26] E. Witten, Anti-de Sitter space and holography, Adv. Theor. Math. Phys. 2 (1998), 253-290,
hep-th/9802150.
[27] M. Zworski, Resonances in physics and geometry. Notices Amer. Math. Soc. 46 (1999),
319328.

Department of Mathematics, University of Washington, Box 354350, Seattle,

WA 98195
E-mail address: [email protected]

Department of Mathematics, University of California, Berkeley, CA 94720

E-mail address: [email protected]